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                 cma->count = 0;
 105                 return -ENOMEM;
 106         }
 107
 108         WARN_ON_ONCE(!pfn_valid(pfn));
 109         zone = page_zone(pfn_to_page(pfn));
 110
 111         do {
 112                 unsigned j;
 113
 114                 base_pfn = pfn;
 115                 for (j = pageblock_nr_pages; j; --j, pfn++) {
 116                         WARN_ON_ONCE(!pfn_valid(pfn));
 117                         /*
 118                          * alloc_contig_range requires the pfn range
 119                          * specified to be in the same zone. Make this
 120                          * simple by forcing the entire CMA resv range
 121                          * to be in the same zone.
 122                          */
 123                         if (page_zone(pfn_to_page(pfn)) != zone)
 124                                 goto err;
 125                 }
 126                 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
 127         } while (--i);
 128
 129         mutex_init(&cma->lock);
 130
 131 #ifdef CONFIG_CMA_DEBUGFS
 132         INIT_HLIST_HEAD(&cma->mem_head);
 133         spin_lock_init(&cma->mem_head_lock);
 134 #endif
 135
 136         return 0;
 137
 138 err:
 139         kfree(cma->bitmap);
 140         cma->count = 0;
 141         return -EINVAL;
 142 }
 143
 144 static int __init cma_init_reserved_areas(void)
 145 {
 146         int i;
 147
 148         for (i = 0; i < cma_area_count; i++) {
 149                 int ret = cma_activate_area(&cma_areas[i]);
 150
 151                 if (ret)
 152                         return ret;
 153         }
 154
 155         return 0;
 156 }
 157 core_initcall(cma_init_reserved_areas);
 158
 159 /**
 160  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
 161  * @base: Base address of the reserved area
 162  * @size: Size of the reserved area (in bytes),
 163  * @order_per_bit: Order of pages represented by one bit on bitmap.
 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                                  struct cma **res_cma)
 171 {
 172         struct cma *cma;
 173         phys_addr_t alignment;
 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         /* ensure minimal alignment required by mm core */
 185         alignment = PAGE_SIZE <<
 186                         max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
 187
 188         /* alignment should be aligned with order_per_bit */
 189         if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
 190                 return -EINVAL;
 191
 192         if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
 193                 return -EINVAL;
 194
 195         /*
 196          * Each reserved area must be initialised later, when more kernel
 197          * subsystems (like slab allocator) are available.
 198          */
 199         cma = &cma_areas[cma_area_count];
 200         cma->base_pfn = PFN_DOWN(base);
 201         cma->count = size >> PAGE_SHIFT;
 202         cma->order_per_bit = order_per_bit;
 203         *res_cma = cma;
 204         cma_area_count++;
 205         totalcma_pages += (size / PAGE_SIZE);
 206
 207         return 0;
 208 }
 209
 210 /**
 211  * cma_declare_contiguous() - reserve custom contiguous area
 212  * @base: Base address of the reserved area optional, use 0 for any
 213  * @size: Size of the reserved area (in bytes),
 214  * @limit: End address of the reserved memory (optional, 0 for any).
 215  * @alignment: Alignment for the CMA area, should be power of 2 or zero
 216  * @order_per_bit: Order of pages represented by one bit on bitmap.
 217  * @fixed: hint about where to place the reserved area
 218  * @res_cma: Pointer to store the created cma region.
 219  *
 220  * This function reserves memory from early allocator. It should be
 221  * called by arch specific code once the early allocator (memblock or bootmem)
 222  * has been activated and all other subsystems have already allocated/reserved
 223  * memory. This function allows to create custom reserved areas.
 224  *
 225  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 226  * reserve in range from @base to @limit.
 227  */
 228 int __init cma_declare_contiguous(phys_addr_t base,
 229                         phys_addr_t size, phys_addr_t limit,
 230                         phys_addr_t alignment, unsigned int order_per_bit,
 231                         bool fixed, struct cma **res_cma)
 232 {
 233         phys_addr_t memblock_end = memblock_end_of_DRAM();
 234         phys_addr_t highmem_start;
 235         int ret = 0;
 236
 237 #ifdef CONFIG_X86
 238         /*
 239          * high_memory isn't direct mapped memory so retrieving its physical
 240          * address isn't appropriate.  But it would be useful to check the
 241          * physical address of the highmem boundary so it's justifiable to get
 242          * the physical address from it.  On x86 there is a validation check for
 243          * this case, so the following workaround is needed to avoid it.
 244          */
 245         highmem_start = __pa_nodebug(high_memory);
 246 #else
 247         highmem_start = __pa(high_memory);
 248 #endif
 249         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 250                 __func__, &size, &base, &limit, &alignment);
 251
 252         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 253                 pr_err("Not enough slots for CMA reserved regions!\n");
 254                 return -ENOSPC;
 255         }
 256
 257         if (!size)
 258                 return -EINVAL;
 259
 260         if (alignment && !is_power_of_2(alignment))
 261                 return -EINVAL;
 262
 263         /*
 264          * Sanitise input arguments.
 265          * Pages both ends in CMA area could be merged into adjacent unmovable
 266          * migratetype page by page allocator's buddy algorithm. In the case,
 267          * you couldn't get a contiguous memory, which is not what we want.
 268          */
 269         alignment = max(alignment,  (phys_addr_t)PAGE_SIZE <<
 270                           max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
 271         if (fixed && base & (alignment - 1)) {
 272                 ret = -EINVAL;
 273                 pr_err("Region at %pa must be aligned to %pa bytes\n",
 274                         &base, &alignment);
 275                 goto err;
 276         }
 277         base = ALIGN(base, alignment);
 278         size = ALIGN(size, alignment);
 279         limit &= ~(alignment - 1);
 280
 281         if (!base)
 282                 fixed = false;
 283
 284         /* size should be aligned with order_per_bit */
 285         if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
 286                 return -EINVAL;
 287
 288         /*
 289          * If allocating at a fixed base the request region must not cross the
 290          * low/high memory boundary.
 291          */
 292         if (fixed && base < highmem_start && base + size > highmem_start) {
 293                 ret = -EINVAL;
 294                 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
 295                         &base, &highmem_start);
 296                 goto err;
 297         }
 298
 299         /*
 300          * If the limit is unspecified or above the memblock end, its effective
 301          * value will be the memblock end. Set it explicitly to simplify further
 302          * checks.
 303          */
 304         if (limit == 0 || limit > memblock_end)
 305                 limit = memblock_end;
 306
 307         if (base + size > limit) {
 308                 ret = -EINVAL;
 309                 pr_err("Size (%pa) of region at %pa exceeds limit (%pa)\n",
 310                         &size, &base, &limit);
 311                 goto err;
 312         }
 313
 314         /* Reserve memory */
 315         if (fixed) {
 316                 if (memblock_is_region_reserved(base, size) ||
 317                     memblock_reserve(base, size) < 0) {
 318                         ret = -EBUSY;
 319                         goto err;
 320                 }
 321         } else {
 322                 phys_addr_t addr = 0;
 323
 324                 /*
 325                  * All pages in the reserved area must come from the same zone.
 326                  * If the requested region crosses the low/high memory boundary,
 327                  * try allocating from high memory first and fall back to low
 328                  * memory in case of failure.
 329                  */
 330                 if (base < highmem_start && limit > highmem_start) {
 331                         addr = memblock_alloc_range(size, alignment,
 332                                                     highmem_start, limit,
 333                                                     MEMBLOCK_NONE);
 334                         limit = highmem_start;
 335                 }
 336
 337                 if (!addr) {
 338                         addr = memblock_alloc_range(size, alignment, base,
 339                                                     limit,
 340                                                     MEMBLOCK_NONE);
 341                         if (!addr) {
 342                                 ret = -ENOMEM;
 343                                 goto err;
 344                         }
 345                 }
 346
 347                 /*
 348                  * kmemleak scans/reads tracked objects for pointers to other
 349                  * objects but this address isn't mapped and accessible
 350                  */
 351                 kmemleak_ignore(phys_to_virt(addr));
 352                 base = addr;
 353         }
 354
 355         ret = cma_init_reserved_mem(base, size, order_per_bit, res_cma);
 356         if (ret)
 357                 goto free_mem;
 358
 359         pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
 360                 &base);
 361         return 0;
 362
 363 free_mem:
 364         memblock_free(base, size);
 365 err:
 366         pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
 367         return ret;
 368 }
 369
 370 /**
 371  * cma_alloc() - allocate pages from contiguous area
 372  * @cma:   Contiguous memory region for which the allocation is performed.
 373  * @count: Requested number of pages.
 374  * @align: Requested alignment of pages (in PAGE_SIZE order).
 375  *
 376  * This function allocates part of contiguous memory on specific
 377  * contiguous memory area.
 378  */
 379 struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align)
 380 {
 381         unsigned long mask, offset;
 382         unsigned long pfn = -1;
 383         unsigned long start = 0;
 384         unsigned long bitmap_maxno, bitmap_no, bitmap_count;
 385         struct page *page = NULL;
 386         int ret;
 387
 388         if (!cma || !cma->count)
 389                 return NULL;
 390
 391         pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
 392                  count, align);
 393
 394         if (!count)
 395                 return NULL;
 396
 397         mask = cma_bitmap_aligned_mask(cma, align);
 398         offset = cma_bitmap_aligned_offset(cma, align);
 399         bitmap_maxno = cma_bitmap_maxno(cma);
 400         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 401
 402         for (;;) {
 403                 mutex_lock(&cma->lock);
 404                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 405                                 bitmap_maxno, start, bitmap_count, mask,
 406                                 offset);
 407                 if (bitmap_no >= bitmap_maxno) {
 408                         mutex_unlock(&cma->lock);
 409                         break;
 410                 }
 411                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 412                 /*
 413                  * It's safe to drop the lock here. We've marked this region for
 414                  * our exclusive use. If the migration fails we will take the
 415                  * lock again and unmark it.
 416                  */
 417                 mutex_unlock(&cma->lock);
 418
 419                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 420                 mutex_lock(&cma_mutex);
 421                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA);
 422                 mutex_unlock(&cma_mutex);
 423                 if (ret == 0) {
 424                         page = pfn_to_page(pfn);
 425                         break;
 426                 }
 427
 428                 cma_clear_bitmap(cma, pfn, count);
 429                 if (ret != -EBUSY)
 430                         break;
 431
 432                 pr_debug("%s(): memory range at %p is busy, retrying\n",
 433                          __func__, pfn_to_page(pfn));
 434                 /* try again with a bit different memory target */
 435                 start = bitmap_no + mask + 1;
 436         }
 437
 438         trace_cma_alloc(pfn, page, count, align);
 439
 440         pr_debug("%s(): returned %p\n", __func__, page);
 441         return page;
 442 }
 443
 444 /**
 445  * cma_release() - release allocated pages
 446  * @cma:   Contiguous memory region for which the allocation is performed.
 447  * @pages: Allocated pages.
 448  * @count: Number of allocated pages.
 449  *
 450  * This function releases memory allocated by alloc_cma().
 451  * It returns false when provided pages do not belong to contiguous area and
 452  * true otherwise.
 453  */
 454 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
 455 {
 456         unsigned long pfn;
 457
 458         if (!cma || !pages)
 459                 return false;
 460
 461         pr_debug("%s(page %p)\n", __func__, (void *)pages);
 462
 463         pfn = page_to_pfn(pages);
 464
 465         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
 466                 return false;
 467
 468         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 469
 470         free_contig_range(pfn, count);
 471         cma_clear_bitmap(cma, pfn, count);
 472         trace_cma_release(pfn, pages, count);
 473
 474         return true;
 475 }