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 <linux/kmemleak.h>
  39 #include <trace/events/cma.h>
  40
  41 #include "cma.h"
  42
  43 struct cma cma_areas[MAX_CMA_AREAS];
  44 unsigned cma_area_count;
  45 static DEFINE_MUTEX(cma_mutex);
  46
  47 phys_addr_t cma_get_base(const struct cma *cma)
  48 {
  49         return PFN_PHYS(cma->base_pfn);
  50 }
  51
  52 unsigned long cma_get_size(const struct cma *cma)
  53 {
  54         return cma->count << PAGE_SHIFT;
  55 }
  56
  57 const char *cma_get_name(const struct cma *cma)
  58 {
  59         return cma->name ? cma->name : "(undefined)";
  60 }
  61
  62 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
  63                                              unsigned int align_order)
  64 {
  65         if (align_order <= cma->order_per_bit)
  66                 return 0;
  67         return (1UL << (align_order - cma->order_per_bit)) - 1;
  68 }
  69
  70 /*
  71  * Find the offset of the base PFN from the specified align_order.
  72  * The value returned is represented in order_per_bits.
  73  */
  74 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
  75                                                unsigned int align_order)
  76 {
  77         return (cma->base_pfn & ((1UL << align_order) - 1))
  78                 >> cma->order_per_bit;
  79 }
  80
  81 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
  82                                               unsigned long pages)
  83 {
  84         return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
  85 }
  86
  87 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
  88                              unsigned int count)
  89 {
  90         unsigned long bitmap_no, bitmap_count;
  91
  92         bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
  93         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
  94
  95         mutex_lock(&cma->lock);
  96         bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
  97         mutex_unlock(&cma->lock);
  98 }
  99
 100 static int __init cma_activate_area(struct cma *cma)
 101 {
 102         int bitmap_size = BITS_TO_LONGS(cma_bitmap_maxno(cma)) * sizeof(long);
 103         unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
 104         unsigned i = cma->count >> pageblock_order;
 105         struct zone *zone;
 106
 107         cma->bitmap = kzalloc(bitmap_size, GFP_KERNEL);
 108
 109         if (!cma->bitmap)
 110                 return -ENOMEM;
 111
 112         WARN_ON_ONCE(!pfn_valid(pfn));
 113         zone = page_zone(pfn_to_page(pfn));
 114
 115         do {
 116                 unsigned j;
 117
 118                 base_pfn = pfn;
 119                 for (j = pageblock_nr_pages; j; --j, pfn++) {
 120                         WARN_ON_ONCE(!pfn_valid(pfn));
 121                         /*
 122                          * alloc_contig_range requires the pfn range
 123                          * specified to be in the same zone. Make this
 124                          * simple by forcing the entire CMA resv range
 125                          * to be in the same zone.
 126                          */
 127                         if (page_zone(pfn_to_page(pfn)) != zone)
 128                                 goto not_in_zone;
 129                 }
 130                 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
 131         } while (--i);
 132
 133         mutex_init(&cma->lock);
 134
 135 #ifdef CONFIG_CMA_DEBUGFS
 136         INIT_HLIST_HEAD(&cma->mem_head);
 137         spin_lock_init(&cma->mem_head_lock);
 138 #endif
 139
 140         return 0;
 141
 142 not_in_zone:
 143         pr_err("CMA area %s could not be activated\n", cma->name);
 144         kfree(cma->bitmap);
 145         cma->count = 0;
 146         return -EINVAL;
 147 }
 148
 149 static int __init cma_init_reserved_areas(void)
 150 {
 151         int i;
 152
 153         for (i = 0; i < cma_area_count; i++) {
 154                 int ret = cma_activate_area(&cma_areas[i]);
 155
 156                 if (ret)
 157                         return ret;
 158         }
 159
 160         return 0;
 161 }
 162 core_initcall(cma_init_reserved_areas);
 163
 164 /**
 165  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
 166  * @base: Base address of the reserved area
 167  * @size: Size of the reserved area (in bytes),
 168  * @order_per_bit: Order of pages represented by one bit on bitmap.
 169  * @name: The name of the area. If this parameter is NULL, the name of
 170  *        the area will be set to "cmaN", where N is a running counter of
 171  *        used areas.
 172  * @res_cma: Pointer to store the created cma region.
 173  *
 174  * This function creates custom contiguous area from already reserved memory.
 175  */
 176 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
 177                                  unsigned int order_per_bit,
 178                                  const char *name,
 179                                  struct cma **res_cma)
 180 {
 181         struct cma *cma;
 182         phys_addr_t alignment;
 183
 184         /* Sanity checks */
 185         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 186                 pr_err("Not enough slots for CMA reserved regions!\n");
 187                 return -ENOSPC;
 188         }
 189
 190         if (!size || !memblock_is_region_reserved(base, size))
 191                 return -EINVAL;
 192
 193         /* ensure minimal alignment required by mm core */
 194         alignment = PAGE_SIZE <<
 195                         max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
 196
 197         /* alignment should be aligned with order_per_bit */
 198         if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
 199                 return -EINVAL;
 200
 201         if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
 202                 return -EINVAL;
 203
 204         /*
 205          * Each reserved area must be initialised later, when more kernel
 206          * subsystems (like slab allocator) are available.
 207          */
 208         cma = &cma_areas[cma_area_count];
 209         if (name) {
 210                 cma->name = name;
 211         } else {
 212                 cma->name = kasprintf(GFP_KERNEL, "cma%d\n", cma_area_count);
 213                 if (!cma->name)
 214                         return -ENOMEM;
 215         }
 216         cma->base_pfn = PFN_DOWN(base);
 217         cma->count = size >> PAGE_SHIFT;
 218         cma->order_per_bit = order_per_bit;
 219         *res_cma = cma;
 220         cma_area_count++;
 221         totalcma_pages += (size / PAGE_SIZE);
 222
 223         return 0;
 224 }
 225
 226 /**
 227  * cma_declare_contiguous() - reserve custom contiguous area
 228  * @base: Base address of the reserved area optional, use 0 for any
 229  * @size: Size of the reserved area (in bytes),
 230  * @limit: End address of the reserved memory (optional, 0 for any).
 231  * @alignment: Alignment for the CMA area, should be power of 2 or zero
 232  * @order_per_bit: Order of pages represented by one bit on bitmap.
 233  * @fixed: hint about where to place the reserved area
 234  * @name: The name of the area. See function cma_init_reserved_mem()
 235  * @res_cma: Pointer to store the created cma region.
 236  *
 237  * This function reserves memory from early allocator. It should be
 238  * called by arch specific code once the early allocator (memblock or bootmem)
 239  * has been activated and all other subsystems have already allocated/reserved
 240  * memory. This function allows to create custom reserved areas.
 241  *
 242  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 243  * reserve in range from @base to @limit.
 244  */
 245 int __init cma_declare_contiguous(phys_addr_t base,
 246                         phys_addr_t size, phys_addr_t limit,
 247                         phys_addr_t alignment, unsigned int order_per_bit,
 248                         bool fixed, const char *name, struct cma **res_cma)
 249 {
 250         phys_addr_t memblock_end = memblock_end_of_DRAM();
 251         phys_addr_t highmem_start;
 252         int ret = 0;
 253
 254         /*
 255          * We can't use __pa(high_memory) directly, since high_memory
 256          * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
 257          * complain. Find the boundary by adding one to the last valid
 258          * address.
 259          */
 260         highmem_start = __pa(high_memory - 1) + 1;
 261         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 262                 __func__, &size, &base, &limit, &alignment);
 263
 264         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 265                 pr_err("Not enough slots for CMA reserved regions!\n");
 266                 return -ENOSPC;
 267         }
 268
 269         if (!size)
 270                 return -EINVAL;
 271
 272         if (alignment && !is_power_of_2(alignment))
 273                 return -EINVAL;
 274
 275         /*
 276          * Sanitise input arguments.
 277          * Pages both ends in CMA area could be merged into adjacent unmovable
 278          * migratetype page by page allocator's buddy algorithm. In the case,
 279          * you couldn't get a contiguous memory, which is not what we want.
 280          */
 281         alignment = max(alignment,  (phys_addr_t)PAGE_SIZE <<
 282                           max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
 283         base = ALIGN(base, alignment);
 284         size = ALIGN(size, alignment);
 285         limit &= ~(alignment - 1);
 286
 287         if (!base)
 288                 fixed = false;
 289
 290         /* size should be aligned with order_per_bit */
 291         if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
 292                 return -EINVAL;
 293
 294         /*
 295          * If allocating at a fixed base the request region must not cross the
 296          * low/high memory boundary.
 297          */
 298         if (fixed && base < highmem_start && base + size > highmem_start) {
 299                 ret = -EINVAL;
 300                 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
 301                         &base, &highmem_start);
 302                 goto err;
 303         }
 304
 305         /*
 306          * If the limit is unspecified or above the memblock end, its effective
 307          * value will be the memblock end. Set it explicitly to simplify further
 308          * checks.
 309          */
 310         if (limit == 0 || limit > memblock_end)
 311                 limit = memblock_end;
 312
 313         /* Reserve memory */
 314         if (fixed) {
 315                 if (memblock_is_region_reserved(base, size) ||
 316                     memblock_reserve(base, size) < 0) {
 317                         ret = -EBUSY;
 318                         goto err;
 319                 }
 320         } else {
 321                 phys_addr_t addr = 0;
 322
 323                 /*
 324                  * All pages in the reserved area must come from the same zone.
 325                  * If the requested region crosses the low/high memory boundary,
 326                  * try allocating from high memory first and fall back to low
 327                  * memory in case of failure.
 328                  */
 329                 if (base < highmem_start && limit > highmem_start) {
 330                         addr = memblock_alloc_range(size, alignment,
 331                                                     highmem_start, limit,
 332                                                     MEMBLOCK_NONE);
 333                         limit = highmem_start;
 334                 }
 335
 336                 if (!addr) {
 337                         addr = memblock_alloc_range(size, alignment, base,
 338                                                     limit,
 339                                                     MEMBLOCK_NONE);
 340                         if (!addr) {
 341                                 ret = -ENOMEM;
 342                                 goto err;
 343                         }
 344                 }
 345
 346                 /*
 347                  * kmemleak scans/reads tracked objects for pointers to other
 348                  * objects but this address isn't mapped and accessible
 349                  */
 350                 kmemleak_ignore_phys(addr);
 351                 base = addr;
 352         }
 353
 354         ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
 355         if (ret)
 356                 goto err;
 357
 358         pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
 359                 &base);
 360         return 0;
 361
 362 err:
 363         pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
 364         return ret;
 365 }
 366
 367 #ifdef CONFIG_CMA_DEBUG
 368 static void cma_debug_show_areas(struct cma *cma)
 369 {
 370         unsigned long next_zero_bit, next_set_bit;
 371         unsigned long start = 0;
 372         unsigned int nr_zero, nr_total = 0;
 373
 374         mutex_lock(&cma->lock);
 375         pr_info("number of available pages: ");
 376         for (;;) {
 377                 next_zero_bit = find_next_zero_bit(cma->bitmap, cma->count, start);
 378                 if (next_zero_bit >= cma->count)
 379                         break;
 380                 next_set_bit = find_next_bit(cma->bitmap, cma->count, next_zero_bit);
 381                 nr_zero = next_set_bit - next_zero_bit;
 382                 pr_cont("%s%u@%lu", nr_total ? "+" : "", nr_zero, next_zero_bit);
 383                 nr_total += nr_zero;
 384                 start = next_zero_bit + nr_zero;
 385         }
 386         pr_cont("=> %u free of %lu total pages\n", nr_total, cma->count);
 387         mutex_unlock(&cma->lock);
 388 }
 389 #else
 390 static inline void cma_debug_show_areas(struct cma *cma) { }
 391 #endif
 392
 393 /**
 394  * cma_alloc() - allocate pages from contiguous area
 395  * @cma:   Contiguous memory region for which the allocation is performed.
 396  * @count: Requested number of pages.
 397  * @align: Requested alignment of pages (in PAGE_SIZE order).
 398  * @no_warn: Avoid printing message about failed allocation
 399  *
 400  * This function allocates part of contiguous memory on specific
 401  * contiguous memory area.
 402  */
 403 struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align,
 404                        bool no_warn)
 405 {
 406         unsigned long mask, offset;
 407         unsigned long pfn = -1;
 408         unsigned long start = 0;
 409         unsigned long bitmap_maxno, bitmap_no, bitmap_count;
 410         size_t i;
 411         struct page *page = NULL;
 412         int ret = -ENOMEM;
 413
 414         if (!cma || !cma->count)
 415                 return NULL;
 416
 417         pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
 418                  count, align);
 419
 420         if (!count)
 421                 return NULL;
 422
 423         mask = cma_bitmap_aligned_mask(cma, align);
 424         offset = cma_bitmap_aligned_offset(cma, align);
 425         bitmap_maxno = cma_bitmap_maxno(cma);
 426         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 427
 428         if (bitmap_count > bitmap_maxno)
 429                 return NULL;
 430
 431         for (;;) {
 432                 mutex_lock(&cma->lock);
 433                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 434                                 bitmap_maxno, start, bitmap_count, mask,
 435                                 offset);
 436                 if (bitmap_no >= bitmap_maxno) {
 437                         mutex_unlock(&cma->lock);
 438                         break;
 439                 }
 440                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 441                 /*
 442                  * It's safe to drop the lock here. We've marked this region for
 443                  * our exclusive use. If the migration fails we will take the
 444                  * lock again and unmark it.
 445                  */
 446                 mutex_unlock(&cma->lock);
 447
 448                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 449                 mutex_lock(&cma_mutex);
 450                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
 451                                      GFP_KERNEL | (no_warn ? __GFP_NOWARN : 0));
 452                 mutex_unlock(&cma_mutex);
 453                 if (ret == 0) {
 454                         page = pfn_to_page(pfn);
 455                         break;
 456                 }
 457
 458                 cma_clear_bitmap(cma, pfn, count);
 459                 if (ret != -EBUSY)
 460                         break;
 461
 462                 pr_debug("%s(): memory range at %p is busy, retrying\n",
 463                          __func__, pfn_to_page(pfn));
 464                 /* try again with a bit different memory target */
 465                 start = bitmap_no + mask + 1;
 466         }
 467
 468         trace_cma_alloc(pfn, page, count, align);
 469
 470         /*
 471          * CMA can allocate multiple page blocks, which results in different
 472          * blocks being marked with different tags. Reset the tags to ignore
 473          * those page blocks.
 474          */
 475         if (page) {
 476                 for (i = 0; i < count; i++)
 477                         page_kasan_tag_reset(page + i);
 478         }
 479
 480         if (ret && !no_warn) {
 481                 pr_err("%s: alloc failed, req-size: %zu pages, ret: %d\n",
 482                         __func__, count, ret);
 483                 cma_debug_show_areas(cma);
 484         }
 485
 486         pr_debug("%s(): returned %p\n", __func__, page);
 487         return page;
 488 }
 489
 490 /**
 491  * cma_release() - release allocated pages
 492  * @cma:   Contiguous memory region for which the allocation is performed.
 493  * @pages: Allocated pages.
 494  * @count: Number of allocated pages.
 495  *
 496  * This function releases memory allocated by alloc_cma().
 497  * It returns false when provided pages do not belong to contiguous area and
 498  * true otherwise.
 499  */
 500 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
 501 {
 502         unsigned long pfn;
 503
 504         if (!cma || !pages)
 505                 return false;
 506
 507         pr_debug("%s(page %p)\n", __func__, (void *)pages);
 508
 509         pfn = page_to_pfn(pages);
 510
 511         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
 512                 return false;
 513
 514         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 515
 516         free_contig_range(pfn, count);
 517         cma_clear_bitmap(cma, pfn, count);
 518         trace_cma_release(pfn, pages, count);
 519
 520         return true;
 521 }
 522
 523 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
 524 {
 525         int i;
 526
 527         for (i = 0; i < cma_area_count; i++) {
 528                 int ret = it(&cma_areas[i], data);
 529
 530                 if (ret)
 531                         return ret;
 532         }
 533
 534         return 0;
 535 }