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 #include "internal.h"
  43
  44 struct cma cma_areas[MAX_CMA_AREAS];
  45 unsigned cma_area_count;
  46 static DEFINE_MUTEX(cma_mutex);
  47
  48 phys_addr_t cma_get_base(const struct cma *cma)
  49 {
  50         return PFN_PHYS(cma->base_pfn);
  51 }
  52
  53 unsigned long cma_get_size(const struct cma *cma)
  54 {
  55         return cma->count << PAGE_SHIFT;
  56 }
  57
  58 const char *cma_get_name(const struct cma *cma)
  59 {
  60         return cma->name ? cma->name : "(undefined)";
  61 }
  62
  63 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
  64                                              unsigned int align_order)
  65 {
  66         if (align_order <= cma->order_per_bit)
  67                 return 0;
  68         return (1UL << (align_order - cma->order_per_bit)) - 1;
  69 }
  70
  71 /*
  72  * Find the offset of the base PFN from the specified align_order.
  73  * The value returned is represented in order_per_bits.
  74  */
  75 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
  76                                                unsigned int align_order)
  77 {
  78         return (cma->base_pfn & ((1UL << align_order) - 1))
  79                 >> cma->order_per_bit;
  80 }
  81
  82 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
  83                                               unsigned long pages)
  84 {
  85         return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
  86 }
  87
  88 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
  89                              unsigned int count)
  90 {
  91         unsigned long bitmap_no, bitmap_count;
  92
  93         bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
  94         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
  95
  96         mutex_lock(&cma->lock);
  97         bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
  98         mutex_unlock(&cma->lock);
  99 }
 100
 101 static int __init cma_activate_area(struct cma *cma)
 102 {
 103         int bitmap_size = BITS_TO_LONGS(cma_bitmap_maxno(cma)) * sizeof(long);
 104         unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
 105         unsigned i = cma->count >> pageblock_order;
 106         struct zone *zone;
 107
 108         cma->bitmap = kzalloc(bitmap_size, GFP_KERNEL);
 109
 110         if (!cma->bitmap)
 111                 return -ENOMEM;
 112
 113         do {
 114                 unsigned j;
 115
 116                 base_pfn = pfn;
 117                 if (!pfn_valid(base_pfn))
 118                         goto err;
 119
 120                 zone = page_zone(pfn_to_page(base_pfn));
 121                 for (j = pageblock_nr_pages; j; --j, pfn++) {
 122                         if (!pfn_valid(pfn))
 123                                 goto err;
 124
 125                         /*
 126                          * In init_cma_reserved_pageblock(), present_pages
 127                          * is adjusted with assumption that all pages in
 128                          * the pageblock come from a single zone.
 129                          */
 130                         if (page_zone(pfn_to_page(pfn)) != zone)
 131                                 goto err;
 132                 }
 133                 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
 134         } while (--i);
 135
 136         mutex_init(&cma->lock);
 137
 138 #ifdef CONFIG_CMA_DEBUGFS
 139         INIT_HLIST_HEAD(&cma->mem_head);
 140         spin_lock_init(&cma->mem_head_lock);
 141 #endif
 142
 143         return 0;
 144
 145 err:
 146         pr_err("CMA area %s could not be activated\n", cma->name);
 147         kfree(cma->bitmap);
 148         cma->count = 0;
 149         return -EINVAL;
 150 }
 151
 152 static int __init cma_init_reserved_areas(void)
 153 {
 154         int i;
 155         struct zone *zone;
 156         pg_data_t *pgdat;
 157
 158         if (!cma_area_count)
 159                 return 0;
 160
 161         for_each_online_pgdat(pgdat) {
 162                 unsigned long start_pfn = UINT_MAX, end_pfn = 0;
 163
 164                 zone = &pgdat->node_zones[ZONE_MOVABLE];
 165
 166                 /*
 167                  * In this case, we cannot adjust the zone range
 168                  * since it is now maximum node span and we don't
 169                  * know original zone range.
 170                  */
 171                 if (populated_zone(zone))
 172                         continue;
 173
 174                 for (i = 0; i < cma_area_count; i++) {
 175                         if (pfn_to_nid(cma_areas[i].base_pfn) !=
 176                                 pgdat->node_id)
 177                                 continue;
 178
 179                         start_pfn = min(start_pfn, cma_areas[i].base_pfn);
 180                         end_pfn = max(end_pfn, cma_areas[i].base_pfn +
 181                                                 cma_areas[i].count);
 182                 }
 183
 184                 if (!end_pfn)
 185                         continue;
 186
 187                 zone->zone_start_pfn = start_pfn;
 188                 zone->spanned_pages = end_pfn - start_pfn;
 189         }
 190
 191         for (i = 0; i < cma_area_count; i++) {
 192                 int ret = cma_activate_area(&cma_areas[i]);
 193
 194                 if (ret)
 195                         return ret;
 196         }
 197
 198         /*
 199          * Reserved pages for ZONE_MOVABLE are now activated and
 200          * this would change ZONE_MOVABLE's managed page counter and
 201          * the other zones' present counter. We need to re-calculate
 202          * various zone information that depends on this initialization.
 203          */
 204         build_all_zonelists(NULL);
 205         for_each_populated_zone(zone) {
 206                 if (zone_idx(zone) == ZONE_MOVABLE) {
 207                         zone_pcp_reset(zone);
 208                         setup_zone_pageset(zone);
 209                 } else
 210                         zone_pcp_update(zone);
 211
 212                 set_zone_contiguous(zone);
 213         }
 214
 215         /*
 216          * We need to re-init per zone wmark by calling
 217          * init_per_zone_wmark_min() but doesn't call here because it is
 218          * registered on core_initcall and it will be called later than us.
 219          */
 220
 221         return 0;
 222 }
 223 pure_initcall(cma_init_reserved_areas);
 224
 225 /**
 226  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
 227  * @base: Base address of the reserved area
 228  * @size: Size of the reserved area (in bytes),
 229  * @order_per_bit: Order of pages represented by one bit on bitmap.
 230  * @name: The name of the area. If this parameter is NULL, the name of
 231  *        the area will be set to "cmaN", where N is a running counter of
 232  *        used areas.
 233  * @res_cma: Pointer to store the created cma region.
 234  *
 235  * This function creates custom contiguous area from already reserved memory.
 236  */
 237 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
 238                                  unsigned int order_per_bit,
 239                                  const char *name,
 240                                  struct cma **res_cma)
 241 {
 242         struct cma *cma;
 243         phys_addr_t alignment;
 244
 245         /* Sanity checks */
 246         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 247                 pr_err("Not enough slots for CMA reserved regions!\n");
 248                 return -ENOSPC;
 249         }
 250
 251         if (!size || !memblock_is_region_reserved(base, size))
 252                 return -EINVAL;
 253
 254         /* ensure minimal alignment required by mm core */
 255         alignment = PAGE_SIZE <<
 256                         max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
 257
 258         /* alignment should be aligned with order_per_bit */
 259         if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
 260                 return -EINVAL;
 261
 262         if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
 263                 return -EINVAL;
 264
 265         /*
 266          * Each reserved area must be initialised later, when more kernel
 267          * subsystems (like slab allocator) are available.
 268          */
 269         cma = &cma_areas[cma_area_count];
 270         if (name) {
 271                 cma->name = name;
 272         } else {
 273                 cma->name = kasprintf(GFP_KERNEL, "cma%d\n", cma_area_count);
 274                 if (!cma->name)
 275                         return -ENOMEM;
 276         }
 277         cma->base_pfn = PFN_DOWN(base);
 278         cma->count = size >> PAGE_SHIFT;
 279         cma->order_per_bit = order_per_bit;
 280         *res_cma = cma;
 281         cma_area_count++;
 282         totalcma_pages += (size / PAGE_SIZE);
 283
 284         return 0;
 285 }
 286
 287 /**
 288  * cma_declare_contiguous() - reserve custom contiguous area
 289  * @base: Base address of the reserved area optional, use 0 for any
 290  * @size: Size of the reserved area (in bytes),
 291  * @limit: End address of the reserved memory (optional, 0 for any).
 292  * @alignment: Alignment for the CMA area, should be power of 2 or zero
 293  * @order_per_bit: Order of pages represented by one bit on bitmap.
 294  * @fixed: hint about where to place the reserved area
 295  * @name: The name of the area. See function cma_init_reserved_mem()
 296  * @res_cma: Pointer to store the created cma region.
 297  *
 298  * This function reserves memory from early allocator. It should be
 299  * called by arch specific code once the early allocator (memblock or bootmem)
 300  * has been activated and all other subsystems have already allocated/reserved
 301  * memory. This function allows to create custom reserved areas.
 302  *
 303  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 304  * reserve in range from @base to @limit.
 305  */
 306 int __init cma_declare_contiguous(phys_addr_t base,
 307                         phys_addr_t size, phys_addr_t limit,
 308                         phys_addr_t alignment, unsigned int order_per_bit,
 309                         bool fixed, const char *name, struct cma **res_cma)
 310 {
 311         phys_addr_t memblock_end = memblock_end_of_DRAM();
 312         phys_addr_t highmem_start;
 313         int ret = 0;
 314
 315         /*
 316          * We can't use __pa(high_memory) directly, since high_memory
 317          * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
 318          * complain. Find the boundary by adding one to the last valid
 319          * address.
 320          */
 321         highmem_start = __pa(high_memory - 1) + 1;
 322         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 323                 __func__, &size, &base, &limit, &alignment);
 324
 325         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 326                 pr_err("Not enough slots for CMA reserved regions!\n");
 327                 return -ENOSPC;
 328         }
 329
 330         if (!size)
 331                 return -EINVAL;
 332
 333         if (alignment && !is_power_of_2(alignment))
 334                 return -EINVAL;
 335
 336         /*
 337          * Sanitise input arguments.
 338          * Pages both ends in CMA area could be merged into adjacent unmovable
 339          * migratetype page by page allocator's buddy algorithm. In the case,
 340          * you couldn't get a contiguous memory, which is not what we want.
 341          */
 342         alignment = max(alignment,  (phys_addr_t)PAGE_SIZE <<
 343                           max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
 344         base = ALIGN(base, alignment);
 345         size = ALIGN(size, alignment);
 346         limit &= ~(alignment - 1);
 347
 348         if (!base)
 349                 fixed = false;
 350
 351         /* size should be aligned with order_per_bit */
 352         if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
 353                 return -EINVAL;
 354
 355         /*
 356          * If allocating at a fixed base the request region must not cross the
 357          * low/high memory boundary.
 358          */
 359         if (fixed && base < highmem_start && base + size > highmem_start) {
 360                 ret = -EINVAL;
 361                 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
 362                         &base, &highmem_start);
 363                 goto err;
 364         }
 365
 366         /*
 367          * If the limit is unspecified or above the memblock end, its effective
 368          * value will be the memblock end. Set it explicitly to simplify further
 369          * checks.
 370          */
 371         if (limit == 0 || limit > memblock_end)
 372                 limit = memblock_end;
 373
 374         /* Reserve memory */
 375         if (fixed) {
 376                 if (memblock_is_region_reserved(base, size) ||
 377                     memblock_reserve(base, size) < 0) {
 378                         ret = -EBUSY;
 379                         goto err;
 380                 }
 381         } else {
 382                 phys_addr_t addr = 0;
 383
 384                 /*
 385                  * All pages in the reserved area must come from the same zone.
 386                  * If the requested region crosses the low/high memory boundary,
 387                  * try allocating from high memory first and fall back to low
 388                  * memory in case of failure.
 389                  */
 390                 if (base < highmem_start && limit > highmem_start) {
 391                         addr = memblock_alloc_range(size, alignment,
 392                                                     highmem_start, limit,
 393                                                     MEMBLOCK_NONE);
 394                         limit = highmem_start;
 395                 }
 396
 397                 if (!addr) {
 398                         addr = memblock_alloc_range(size, alignment, base,
 399                                                     limit,
 400                                                     MEMBLOCK_NONE);
 401                         if (!addr) {
 402                                 ret = -ENOMEM;
 403                                 goto err;
 404                         }
 405                 }
 406
 407                 /*
 408                  * kmemleak scans/reads tracked objects for pointers to other
 409                  * objects but this address isn't mapped and accessible
 410                  */
 411                 kmemleak_ignore_phys(addr);
 412                 base = addr;
 413         }
 414
 415         ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
 416         if (ret)
 417                 goto err;
 418
 419         pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
 420                 &base);
 421         return 0;
 422
 423 err:
 424         pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
 425         return ret;
 426 }
 427
 428 #ifdef CONFIG_CMA_DEBUG
 429 static void cma_debug_show_areas(struct cma *cma)
 430 {
 431         unsigned long next_zero_bit, next_set_bit;
 432         unsigned long start = 0;
 433         unsigned int nr_zero, nr_total = 0;
 434
 435         mutex_lock(&cma->lock);
 436         pr_info("number of available pages: ");
 437         for (;;) {
 438                 next_zero_bit = find_next_zero_bit(cma->bitmap, cma->count, start);
 439                 if (next_zero_bit >= cma->count)
 440                         break;
 441                 next_set_bit = find_next_bit(cma->bitmap, cma->count, next_zero_bit);
 442                 nr_zero = next_set_bit - next_zero_bit;
 443                 pr_cont("%s%u@%lu", nr_total ? "+" : "", nr_zero, next_zero_bit);
 444                 nr_total += nr_zero;
 445                 start = next_zero_bit + nr_zero;
 446         }
 447         pr_cont("=> %u free of %lu total pages\n", nr_total, cma->count);
 448         mutex_unlock(&cma->lock);
 449 }
 450 #else
 451 static inline void cma_debug_show_areas(struct cma *cma) { }
 452 #endif
 453
 454 /**
 455  * cma_alloc() - allocate pages from contiguous area
 456  * @cma:   Contiguous memory region for which the allocation is performed.
 457  * @count: Requested number of pages.
 458  * @align: Requested alignment of pages (in PAGE_SIZE order).
 459  * @gfp_mask:  GFP mask to use during compaction
 460  *
 461  * This function allocates part of contiguous memory on specific
 462  * contiguous memory area.
 463  */
 464 struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align,
 465                        gfp_t gfp_mask)
 466 {
 467         unsigned long mask, offset;
 468         unsigned long pfn = -1;
 469         unsigned long start = 0;
 470         unsigned long bitmap_maxno, bitmap_no, bitmap_count;
 471         struct page *page = NULL;
 472         int ret = -ENOMEM;
 473
 474         if (!cma || !cma->count)
 475                 return NULL;
 476
 477         pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
 478                  count, align);
 479
 480         if (!count)
 481                 return NULL;
 482
 483         mask = cma_bitmap_aligned_mask(cma, align);
 484         offset = cma_bitmap_aligned_offset(cma, align);
 485         bitmap_maxno = cma_bitmap_maxno(cma);
 486         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 487
 488         if (bitmap_count > bitmap_maxno)
 489                 return NULL;
 490
 491         for (;;) {
 492                 mutex_lock(&cma->lock);
 493                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 494                                 bitmap_maxno, start, bitmap_count, mask,
 495                                 offset);
 496                 if (bitmap_no >= bitmap_maxno) {
 497                         mutex_unlock(&cma->lock);
 498                         break;
 499                 }
 500                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 501                 /*
 502                  * It's safe to drop the lock here. We've marked this region for
 503                  * our exclusive use. If the migration fails we will take the
 504                  * lock again and unmark it.
 505                  */
 506                 mutex_unlock(&cma->lock);
 507
 508                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 509                 mutex_lock(&cma_mutex);
 510                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
 511                                          gfp_mask);
 512                 mutex_unlock(&cma_mutex);
 513                 if (ret == 0) {
 514                         page = pfn_to_page(pfn);
 515                         break;
 516                 }
 517
 518                 cma_clear_bitmap(cma, pfn, count);
 519                 if (ret != -EBUSY)
 520                         break;
 521
 522                 pr_debug("%s(): memory range at %p is busy, retrying\n",
 523                          __func__, pfn_to_page(pfn));
 524                 /* try again with a bit different memory target */
 525                 start = bitmap_no + mask + 1;
 526         }
 527
 528         trace_cma_alloc(pfn, page, count, align);
 529
 530         if (ret && !(gfp_mask & __GFP_NOWARN)) {
 531                 pr_err("%s: alloc failed, req-size: %zu pages, ret: %d\n",
 532                         __func__, count, ret);
 533                 cma_debug_show_areas(cma);
 534         }
 535
 536         pr_debug("%s(): returned %p\n", __func__, page);
 537         return page;
 538 }
 539
 540 /**
 541  * cma_release() - release allocated pages
 542  * @cma:   Contiguous memory region for which the allocation is performed.
 543  * @pages: Allocated pages.
 544  * @count: Number of allocated pages.
 545  *
 546  * This function releases memory allocated by alloc_cma().
 547  * It returns false when provided pages do not belong to contiguous area and
 548  * true otherwise.
 549  */
 550 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
 551 {
 552         unsigned long pfn;
 553
 554         if (!cma || !pages)
 555                 return false;
 556
 557         pr_debug("%s(page %p)\n", __func__, (void *)pages);
 558
 559         pfn = page_to_pfn(pages);
 560
 561         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
 562                 return false;
 563
 564         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 565
 566         free_contig_range(pfn, count);
 567         cma_clear_bitmap(cma, pfn, count);
 568         trace_cma_release(pfn, pages, count);
 569
 570         return true;
 571 }
 572
 573 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
 574 {
 575         int i;
 576
 577         for (i = 0; i < cma_area_count; i++) {
 578                 int ret = it(&cma_areas[i], data);
 579
 580                 if (ret)
 581                         return ret;
 582         }
 583
 584         return 0;
 585 }