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 const char *cma_get_name(const struct cma *cma)
  57 {
  58         return cma->name ? cma->name : "(undefined)";
  59 }
  60
  61 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
  62                                              unsigned int align_order)
  63 {
  64         if (align_order <= cma->order_per_bit)
  65                 return 0;
  66         return (1UL << (align_order - cma->order_per_bit)) - 1;
  67 }
  68
  69 /*
  70  * Find the offset of the base PFN from the specified align_order.
  71  * The value returned is represented in order_per_bits.
  72  */
  73 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
  74                                                unsigned int align_order)
  75 {
  76         return (cma->base_pfn & ((1UL << align_order) - 1))
  77                 >> cma->order_per_bit;
  78 }
  79
  80 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
  81                                               unsigned long pages)
  82 {
  83         return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
  84 }
  85
  86 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
  87                              unsigned int count)
  88 {
  89         unsigned long bitmap_no, bitmap_count;
  90
  91         bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
  92         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
  93
  94         mutex_lock(&cma->lock);
  95         bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
  96         mutex_unlock(&cma->lock);
  97 }
  98
  99 static int __init cma_activate_area(struct cma *cma)
 100 {
 101         int bitmap_size = BITS_TO_LONGS(cma_bitmap_maxno(cma)) * sizeof(long);
 102         unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
 103         unsigned i = cma->count >> pageblock_order;
 104         struct zone *zone;
 105
 106         cma->bitmap = kzalloc(bitmap_size, GFP_KERNEL);
 107
 108         if (!cma->bitmap) {
 109                 cma->count = 0;
 110                 return -ENOMEM;
 111         }
 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 not_in_zone;
 130                 }
 131                 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
 132         } while (--i);
 133
 134         mutex_init(&cma->lock);
 135
 136 #ifdef CONFIG_CMA_DEBUGFS
 137         INIT_HLIST_HEAD(&cma->mem_head);
 138         spin_lock_init(&cma->mem_head_lock);
 139 #endif
 140
 141         return 0;
 142
 143 not_in_zone:
 144         pr_err("CMA area %s could not be activated\n", cma->name);
 145         kfree(cma->bitmap);
 146         cma->count = 0;
 147         return -EINVAL;
 148 }
 149
 150 static int __init cma_init_reserved_areas(void)
 151 {
 152         int i;
 153
 154         for (i = 0; i < cma_area_count; i++) {
 155                 int ret = cma_activate_area(&cma_areas[i]);
 156
 157                 if (ret)
 158                         return ret;
 159         }
 160
 161         return 0;
 162 }
 163 core_initcall(cma_init_reserved_areas);
 164
 165 /**
 166  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
 167  * @base: Base address of the reserved area
 168  * @size: Size of the reserved area (in bytes),
 169  * @order_per_bit: Order of pages represented by one bit on bitmap.
 170  * @res_cma: Pointer to store the created cma region.
 171  *
 172  * This function creates custom contiguous area from already reserved memory.
 173  */
 174 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
 175                                  unsigned int order_per_bit,
 176                                  const char *name,
 177                                  struct cma **res_cma)
 178 {
 179         struct cma *cma;
 180         phys_addr_t alignment;
 181
 182         /* Sanity checks */
 183         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 184                 pr_err("Not enough slots for CMA reserved regions!\n");
 185                 return -ENOSPC;
 186         }
 187
 188         if (!size || !memblock_is_region_reserved(base, size))
 189                 return -EINVAL;
 190
 191         /* ensure minimal alignment required by mm core */
 192         alignment = PAGE_SIZE <<
 193                         max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
 194
 195         /* alignment should be aligned with order_per_bit */
 196         if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
 197                 return -EINVAL;
 198
 199         if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
 200                 return -EINVAL;
 201
 202         /*
 203          * Each reserved area must be initialised later, when more kernel
 204          * subsystems (like slab allocator) are available.
 205          */
 206         cma = &cma_areas[cma_area_count];
 207         if (name) {
 208                 cma->name = name;
 209         } else {
 210                 cma->name = kasprintf(GFP_KERNEL, "cma%d\n", cma_area_count);
 211                 if (!cma->name)
 212                         return -ENOMEM;
 213         }
 214         cma->base_pfn = PFN_DOWN(base);
 215         cma->count = size >> PAGE_SHIFT;
 216         cma->order_per_bit = order_per_bit;
 217         *res_cma = cma;
 218         cma_area_count++;
 219         totalcma_pages += (size / PAGE_SIZE);
 220
 221         return 0;
 222 }
 223
 224 /**
 225  * cma_declare_contiguous() - reserve custom contiguous area
 226  * @base: Base address of the reserved area optional, use 0 for any
 227  * @size: Size of the reserved area (in bytes),
 228  * @limit: End address of the reserved memory (optional, 0 for any).
 229  * @alignment: Alignment for the CMA area, should be power of 2 or zero
 230  * @order_per_bit: Order of pages represented by one bit on bitmap.
 231  * @fixed: hint about where to place the reserved area
 232  * @res_cma: Pointer to store the created cma region.
 233  *
 234  * This function reserves memory from early allocator. It should be
 235  * called by arch specific code once the early allocator (memblock or bootmem)
 236  * has been activated and all other subsystems have already allocated/reserved
 237  * memory. This function allows to create custom reserved areas.
 238  *
 239  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 240  * reserve in range from @base to @limit.
 241  */
 242 int __init cma_declare_contiguous(phys_addr_t base,
 243                         phys_addr_t size, phys_addr_t limit,
 244                         phys_addr_t alignment, unsigned int order_per_bit,
 245                         bool fixed, const char *name, struct cma **res_cma)
 246 {
 247         phys_addr_t memblock_end = memblock_end_of_DRAM();
 248         phys_addr_t highmem_start;
 249         int ret = 0;
 250
 251         /*
 252          * We can't use __pa(high_memory) directly, since high_memory
 253          * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
 254          * complain. Find the boundary by adding one to the last valid
 255          * address.
 256          */
 257         highmem_start = __pa(high_memory - 1) + 1;
 258         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 259                 __func__, &size, &base, &limit, &alignment);
 260
 261         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 262                 pr_err("Not enough slots for CMA reserved regions!\n");
 263                 return -ENOSPC;
 264         }
 265
 266         if (!size)
 267                 return -EINVAL;
 268
 269         if (alignment && !is_power_of_2(alignment))
 270                 return -EINVAL;
 271
 272         /*
 273          * Sanitise input arguments.
 274          * Pages both ends in CMA area could be merged into adjacent unmovable
 275          * migratetype page by page allocator's buddy algorithm. In the case,
 276          * you couldn't get a contiguous memory, which is not what we want.
 277          */
 278         alignment = max(alignment,  (phys_addr_t)PAGE_SIZE <<
 279                           max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
 280         if (fixed && base & (alignment - 1)) {
 281                 ret = -EINVAL;
 282                 pr_err("Region at %pa must be aligned to %pa bytes\n",
 283                         &base, &alignment);
 284                 goto err;
 285         }
 286         base = ALIGN(base, alignment);
 287         size = ALIGN(size, alignment);
 288         limit &= ~(alignment - 1);
 289
 290         if (!base)
 291                 fixed = false;
 292
 293         /* size should be aligned with order_per_bit */
 294         if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
 295                 return -EINVAL;
 296
 297         /*
 298          * If allocating at a fixed base the request region must not cross the
 299          * low/high memory boundary.
 300          */
 301         if (fixed && base < highmem_start && base + size > highmem_start) {
 302                 ret = -EINVAL;
 303                 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
 304                         &base, &highmem_start);
 305                 goto err;
 306         }
 307
 308         /*
 309          * If the limit is unspecified or above the memblock end, its effective
 310          * value will be the memblock end. Set it explicitly to simplify further
 311          * checks.
 312          */
 313         if (limit == 0 || limit > memblock_end)
 314                 limit = memblock_end;
 315
 316         if (base + size > limit) {
 317                 ret = -EINVAL;
 318                 pr_err("Size (%pa) of region at %pa exceeds limit (%pa)\n",
 319                         &size, &base, &limit);
 320                 goto err;
 321         }
 322
 323         /* Reserve memory */
 324         if (fixed) {
 325                 if (memblock_is_region_reserved(base, size) ||
 326                     memblock_reserve(base, size) < 0) {
 327                         ret = -EBUSY;
 328                         goto err;
 329                 }
 330         } else {
 331                 phys_addr_t addr = 0;
 332
 333                 /*
 334                  * All pages in the reserved area must come from the same zone.
 335                  * If the requested region crosses the low/high memory boundary,
 336                  * try allocating from high memory first and fall back to low
 337                  * memory in case of failure.
 338                  */
 339                 if (base < highmem_start && limit > highmem_start) {
 340                         addr = memblock_alloc_range(size, alignment,
 341                                                     highmem_start, limit,
 342                                                     MEMBLOCK_NONE);
 343                         limit = highmem_start;
 344                 }
 345
 346                 if (!addr) {
 347                         addr = memblock_alloc_range(size, alignment, base,
 348                                                     limit,
 349                                                     MEMBLOCK_NONE);
 350                         if (!addr) {
 351                                 ret = -ENOMEM;
 352                                 goto err;
 353                         }
 354                 }
 355
 356                 /*
 357                  * kmemleak scans/reads tracked objects for pointers to other
 358                  * objects but this address isn't mapped and accessible
 359                  */
 360                 kmemleak_ignore_phys(addr);
 361                 base = addr;
 362         }
 363
 364         ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
 365         if (ret)
 366                 goto free_mem;
 367
 368         pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
 369                 &base);
 370         return 0;
 371
 372 free_mem:
 373         memblock_free(base, size);
 374 err:
 375         pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
 376         return ret;
 377 }
 378
 379 #ifdef CONFIG_CMA_DEBUG
 380 static void cma_debug_show_areas(struct cma *cma)
 381 {
 382         unsigned long next_zero_bit, next_set_bit, nr_zero;
 383         unsigned long start = 0;
 384         unsigned long nr_part, nr_total = 0;
 385         unsigned long nbits = cma_bitmap_maxno(cma);
 386
 387         mutex_lock(&cma->lock);
 388         pr_info("number of available pages: ");
 389         for (;;) {
 390                 next_zero_bit = find_next_zero_bit(cma->bitmap, nbits, start);
 391                 if (next_zero_bit >= nbits)
 392                         break;
 393                 next_set_bit = find_next_bit(cma->bitmap, nbits, next_zero_bit);
 394                 nr_zero = next_set_bit - next_zero_bit;
 395                 nr_part = nr_zero << cma->order_per_bit;
 396                 pr_cont("%s%lu@%lu", nr_total ? "+" : "", nr_part,
 397                         next_zero_bit);
 398                 nr_total += nr_part;
 399                 start = next_zero_bit + nr_zero;
 400         }
 401         pr_cont("=> %lu free of %lu total pages\n", nr_total, cma->count);
 402         mutex_unlock(&cma->lock);
 403 }
 404 #else
 405 static inline void cma_debug_show_areas(struct cma *cma) { }
 406 #endif
 407
 408 /**
 409  * cma_alloc() - allocate pages from contiguous area
 410  * @cma:   Contiguous memory region for which the allocation is performed.
 411  * @count: Requested number of pages.
 412  * @align: Requested alignment of pages (in PAGE_SIZE order).
 413  *
 414  * This function allocates part of contiguous memory on specific
 415  * contiguous memory area.
 416  */
 417 struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align,
 418                        gfp_t gfp_mask)
 419 {
 420         unsigned long mask, offset;
 421         unsigned long pfn = -1;
 422         unsigned long start = 0;
 423         unsigned long bitmap_maxno, bitmap_no, bitmap_count;
 424         struct page *page = NULL;
 425         int ret = -ENOMEM;
 426
 427         if (!cma || !cma->count)
 428                 return NULL;
 429
 430         pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
 431                  count, align);
 432
 433         if (!count)
 434                 return NULL;
 435
 436         mask = cma_bitmap_aligned_mask(cma, align);
 437         offset = cma_bitmap_aligned_offset(cma, align);
 438         bitmap_maxno = cma_bitmap_maxno(cma);
 439         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 440
 441         if (bitmap_count > bitmap_maxno)
 442                 return NULL;
 443
 444         for (;;) {
 445                 mutex_lock(&cma->lock);
 446                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 447                                 bitmap_maxno, start, bitmap_count, mask,
 448                                 offset);
 449                 if (bitmap_no >= bitmap_maxno) {
 450                         mutex_unlock(&cma->lock);
 451                         break;
 452                 }
 453                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 454                 /*
 455                  * It's safe to drop the lock here. We've marked this region for
 456                  * our exclusive use. If the migration fails we will take the
 457                  * lock again and unmark it.
 458                  */
 459                 mutex_unlock(&cma->lock);
 460
 461                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 462                 mutex_lock(&cma_mutex);
 463                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
 464                                          gfp_mask);
 465                 mutex_unlock(&cma_mutex);
 466                 if (ret == 0) {
 467                         page = pfn_to_page(pfn);
 468                         break;
 469                 }
 470
 471                 cma_clear_bitmap(cma, pfn, count);
 472                 if (ret != -EBUSY)
 473                         break;
 474
 475                 pr_debug("%s(): memory range at %p is busy, retrying\n",
 476                          __func__, pfn_to_page(pfn));
 477                 /* try again with a bit different memory target */
 478                 start = bitmap_no + mask + 1;
 479         }
 480
 481         trace_cma_alloc(pfn, page, count, align);
 482
 483         if (ret && !(gfp_mask & __GFP_NOWARN)) {
 484                 pr_info("%s: alloc failed, req-size: %zu pages, ret: %d\n",
 485                         __func__, count, ret);
 486                 cma_debug_show_areas(cma);
 487         }
 488
 489         pr_debug("%s(): returned %p\n", __func__, page);
 490         return page;
 491 }
 492
 493 /**
 494  * cma_release() - release allocated pages
 495  * @cma:   Contiguous memory region for which the allocation is performed.
 496  * @pages: Allocated pages.
 497  * @count: Number of allocated pages.
 498  *
 499  * This function releases memory allocated by alloc_cma().
 500  * It returns false when provided pages do not belong to contiguous area and
 501  * true otherwise.
 502  */
 503 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
 504 {
 505         unsigned long pfn;
 506
 507         if (!cma || !pages)
 508                 return false;
 509
 510         pr_debug("%s(page %p)\n", __func__, (void *)pages);
 511
 512         pfn = page_to_pfn(pages);
 513
 514         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
 515                 return false;
 516
 517         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 518
 519         free_contig_range(pfn, count);
 520         cma_clear_bitmap(cma, pfn, count);
 521         trace_cma_release(pfn, pages, count);
 522
 523         return true;
 524 }
 525
 526 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
 527 {
 528         int i;
 529
 530         for (i = 0; i < cma_area_count; i++) {
 531                 int ret = it(&cma_areas[i], data);
 532
 533                 if (ret)
 534                         return ret;
 535         }
 536
 537         return 0;
 538 }