mm/cma.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Contiguous Memory Allocator
   4  *
   5  * Copyright (c) 2010-2011 by Samsung Electronics.
   6  * Copyright IBM Corporation, 2013
   7  * Copyright LG Electronics Inc., 2014
   8  * Written by:
   9  *      Marek Szyprowski <m.szyprowski@samsung.com>
  10  *      Michal Nazarewicz <mina86@mina86.com>
  11  *      Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
  12  *      Joonsoo Kim <iamjoonsoo.kim@lge.com>
  13  */
  14
  15 #define pr_fmt(fmt) "cma: " fmt
  16
  17 #ifdef CONFIG_CMA_DEBUG
  18 #ifndef DEBUG
  19 #  define DEBUG
  20 #endif
  21 #endif
  22 #define CREATE_TRACE_POINTS
  23
  24 #include <linux/memblock.h>
  25 #include <linux/err.h>
  26 #include <linux/mm.h>
  27 #include <linux/mutex.h>
  28 #include <linux/sizes.h>
  29 #include <linux/slab.h>
  30 #include <linux/log2.h>
  31 #include <linux/cma.h>
  32 #include <linux/highmem.h>
  33 #include <linux/io.h>
  34 #include <linux/kmemleak.h>
  35 #include <trace/events/cma.h>
  36
  37 #include "cma.h"
  38
  39 struct cma cma_areas[MAX_CMA_AREAS];
  40 unsigned cma_area_count;
  41 static DEFINE_MUTEX(cma_mutex);
  42
  43 phys_addr_t cma_get_base(const struct cma *cma)
  44 {
  45         return PFN_PHYS(cma->base_pfn);
  46 }
  47
  48 unsigned long cma_get_size(const struct cma *cma)
  49 {
  50         return cma->count << PAGE_SHIFT;
  51 }
  52
  53 const char *cma_get_name(const struct cma *cma)
  54 {
  55         return cma->name ? cma->name : "(undefined)";
  56 }
  57
  58 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
  59                                              unsigned int align_order)
  60 {
  61         if (align_order <= cma->order_per_bit)
  62                 return 0;
  63         return (1UL << (align_order - cma->order_per_bit)) - 1;
  64 }
  65
  66 /*
  67  * Find the offset of the base PFN from the specified align_order.
  68  * The value returned is represented in order_per_bits.
  69  */
  70 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
  71                                                unsigned int align_order)
  72 {
  73         return (cma->base_pfn & ((1UL << align_order) - 1))
  74                 >> cma->order_per_bit;
  75 }
  76
  77 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
  78                                               unsigned long pages)
  79 {
  80         return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
  81 }
  82
  83 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
  84                              unsigned int count)
  85 {
  86         unsigned long bitmap_no, bitmap_count;
  87
  88         bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
  89         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
  90
  91         mutex_lock(&cma->lock);
  92         bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
  93         mutex_unlock(&cma->lock);
  94 }
  95
  96 static int __init cma_activate_area(struct cma *cma)
  97 {
  98         int bitmap_size = BITS_TO_LONGS(cma_bitmap_maxno(cma)) * sizeof(long);
  99         unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
 100         unsigned i = cma->count >> pageblock_order;
 101         struct zone *zone;
 102
 103         cma->bitmap = kzalloc(bitmap_size, GFP_KERNEL);
 104
 105         if (!cma->bitmap) {
 106                 cma->count = 0;
 107                 return -ENOMEM;
 108         }
 109
 110         WARN_ON_ONCE(!pfn_valid(pfn));
 111         zone = page_zone(pfn_to_page(pfn));
 112
 113         do {
 114                 unsigned j;
 115
 116                 base_pfn = pfn;
 117                 for (j = pageblock_nr_pages; j; --j, pfn++) {
 118                         WARN_ON_ONCE(!pfn_valid(pfn));
 119                         /*
 120                          * alloc_contig_range requires the pfn range
 121                          * specified to be in the same zone. Make this
 122                          * simple by forcing the entire CMA resv range
 123                          * to be in the same zone.
 124                          */
 125                         if (page_zone(pfn_to_page(pfn)) != zone)
 126                                 goto not_in_zone;
 127                 }
 128                 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
 129         } while (--i);
 130
 131         mutex_init(&cma->lock);
 132
 133 #ifdef CONFIG_CMA_DEBUGFS
 134         INIT_HLIST_HEAD(&cma->mem_head);
 135         spin_lock_init(&cma->mem_head_lock);
 136 #endif
 137
 138         return 0;
 139
 140 not_in_zone:
 141         pr_err("CMA area %s could not be activated\n", cma->name);
 142         kfree(cma->bitmap);
 143         cma->count = 0;
 144         return -EINVAL;
 145 }
 146
 147 static int __init cma_init_reserved_areas(void)
 148 {
 149         int i;
 150
 151         for (i = 0; i < cma_area_count; i++) {
 152                 int ret = cma_activate_area(&cma_areas[i]);
 153
 154                 if (ret)
 155                         return ret;
 156         }
 157
 158         return 0;
 159 }
 160 core_initcall(cma_init_reserved_areas);
 161
 162 /**
 163  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
 164  * @base: Base address of the reserved area
 165  * @size: Size of the reserved area (in bytes),
 166  * @order_per_bit: Order of pages represented by one bit on bitmap.
 167  * @name: The name of the area. If this parameter is NULL, the name of
 168  *        the area will be set to "cmaN", where N is a running counter of
 169  *        used areas.
 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  * @name: The name of the area. See function cma_init_reserved_mem()
 233  * @res_cma: Pointer to store the created cma region.
 234  *
 235  * This function reserves memory from early allocator. It should be
 236  * called by arch specific code once the early allocator (memblock or bootmem)
 237  * has been activated and all other subsystems have already allocated/reserved
 238  * memory. This function allows to create custom reserved areas.
 239  *
 240  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 241  * reserve in range from @base to @limit.
 242  */
 243 int __init cma_declare_contiguous(phys_addr_t base,
 244                         phys_addr_t size, phys_addr_t limit,
 245                         phys_addr_t alignment, unsigned int order_per_bit,
 246                         bool fixed, const char *name, struct cma **res_cma)
 247 {
 248         phys_addr_t memblock_end = memblock_end_of_DRAM();
 249         phys_addr_t highmem_start;
 250         int ret = 0;
 251
 252         /*
 253          * We can't use __pa(high_memory) directly, since high_memory
 254          * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
 255          * complain. Find the boundary by adding one to the last valid
 256          * address.
 257          */
 258         highmem_start = __pa(high_memory - 1) + 1;
 259         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 260                 __func__, &size, &base, &limit, &alignment);
 261
 262         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 263                 pr_err("Not enough slots for CMA reserved regions!\n");
 264                 return -ENOSPC;
 265         }
 266
 267         if (!size)
 268                 return -EINVAL;
 269
 270         if (alignment && !is_power_of_2(alignment))
 271                 return -EINVAL;
 272
 273         /*
 274          * Sanitise input arguments.
 275          * Pages both ends in CMA area could be merged into adjacent unmovable
 276          * migratetype page by page allocator's buddy algorithm. In the case,
 277          * you couldn't get a contiguous memory, which is not what we want.
 278          */
 279         alignment = max(alignment,  (phys_addr_t)PAGE_SIZE <<
 280                           max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
 281         if (fixed && base & (alignment - 1)) {
 282                 ret = -EINVAL;
 283                 pr_err("Region at %pa must be aligned to %pa bytes\n",
 284                         &base, &alignment);
 285                 goto err;
 286         }
 287         base = ALIGN(base, alignment);
 288         size = ALIGN(size, alignment);
 289         limit &= ~(alignment - 1);
 290
 291         if (!base)
 292                 fixed = false;
 293
 294         /* size should be aligned with order_per_bit */
 295         if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
 296                 return -EINVAL;
 297
 298         /*
 299          * If allocating at a fixed base the request region must not cross the
 300          * low/high memory boundary.
 301          */
 302         if (fixed && base < highmem_start && base + size > highmem_start) {
 303                 ret = -EINVAL;
 304                 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
 305                         &base, &highmem_start);
 306                 goto err;
 307         }
 308
 309         /*
 310          * If the limit is unspecified or above the memblock end, its effective
 311          * value will be the memblock end. Set it explicitly to simplify further
 312          * checks.
 313          */
 314         if (limit == 0 || limit > memblock_end)
 315                 limit = memblock_end;
 316
 317         if (base + size > limit) {
 318                 ret = -EINVAL;
 319                 pr_err("Size (%pa) of region at %pa exceeds limit (%pa)\n",
 320                         &size, &base, &limit);
 321                 goto err;
 322         }
 323
 324         /* Reserve memory */
 325         if (fixed) {
 326                 if (memblock_is_region_reserved(base, size) ||
 327                     memblock_reserve(base, size) < 0) {
 328                         ret = -EBUSY;
 329                         goto err;
 330                 }
 331         } else {
 332                 phys_addr_t addr = 0;
 333
 334                 /*
 335                  * All pages in the reserved area must come from the same zone.
 336                  * If the requested region crosses the low/high memory boundary,
 337                  * try allocating from high memory first and fall back to low
 338                  * memory in case of failure.
 339                  */
 340                 if (base < highmem_start && limit > highmem_start) {
 341                         addr = memblock_phys_alloc_range(size, alignment,
 342                                                          highmem_start, limit);
 343                         limit = highmem_start;
 344                 }
 345
 346                 if (!addr) {
 347                         addr = memblock_phys_alloc_range(size, alignment, base,
 348                                                          limit);
 349                         if (!addr) {
 350                                 ret = -ENOMEM;
 351                                 goto err;
 352                         }
 353                 }
 354
 355                 /*
 356                  * kmemleak scans/reads tracked objects for pointers to other
 357                  * objects but this address isn't mapped and accessible
 358                  */
 359                 kmemleak_ignore_phys(addr);
 360                 base = addr;
 361         }
 362
 363         ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
 364         if (ret)
 365                 goto free_mem;
 366
 367         pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
 368                 &base);
 369         return 0;
 370
 371 free_mem:
 372         memblock_free(base, size);
 373 err:
 374         pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
 375         return ret;
 376 }
 377
 378 #ifdef CONFIG_CMA_DEBUG
 379 static void cma_debug_show_areas(struct cma *cma)
 380 {
 381         unsigned long next_zero_bit, next_set_bit, nr_zero;
 382         unsigned long start = 0;
 383         unsigned long nr_part, nr_total = 0;
 384         unsigned long nbits = cma_bitmap_maxno(cma);
 385
 386         mutex_lock(&cma->lock);
 387         pr_info("number of available pages: ");
 388         for (;;) {
 389                 next_zero_bit = find_next_zero_bit(cma->bitmap, nbits, start);
 390                 if (next_zero_bit >= nbits)
 391                         break;
 392                 next_set_bit = find_next_bit(cma->bitmap, nbits, next_zero_bit);
 393                 nr_zero = next_set_bit - next_zero_bit;
 394                 nr_part = nr_zero << cma->order_per_bit;
 395                 pr_cont("%s%lu@%lu", nr_total ? "+" : "", nr_part,
 396                         next_zero_bit);
 397                 nr_total += nr_part;
 398                 start = next_zero_bit + nr_zero;
 399         }
 400         pr_cont("=> %lu free of %lu total pages\n", nr_total, cma->count);
 401         mutex_unlock(&cma->lock);
 402 }
 403 #else
 404 static inline void cma_debug_show_areas(struct cma *cma) { }
 405 #endif
 406
 407 /**
 408  * cma_alloc() - allocate pages from contiguous area
 409  * @cma:   Contiguous memory region for which the allocation is performed.
 410  * @count: Requested number of pages.
 411  * @align: Requested alignment of pages (in PAGE_SIZE order).
 412  * @no_warn: Avoid printing message about failed allocation
 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                        bool no_warn)
 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         size_t i;
 425         struct page *page = NULL;
 426         int ret = -ENOMEM;
 427
 428         if (!cma || !cma->count)
 429                 return NULL;
 430
 431         pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
 432                  count, align);
 433
 434         if (!count)
 435                 return NULL;
 436
 437         mask = cma_bitmap_aligned_mask(cma, align);
 438         offset = cma_bitmap_aligned_offset(cma, align);
 439         bitmap_maxno = cma_bitmap_maxno(cma);
 440         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 441
 442         if (bitmap_count > bitmap_maxno)
 443                 return NULL;
 444
 445         for (;;) {
 446                 mutex_lock(&cma->lock);
 447                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 448                                 bitmap_maxno, start, bitmap_count, mask,
 449                                 offset);
 450                 if (bitmap_no >= bitmap_maxno) {
 451                         mutex_unlock(&cma->lock);
 452                         break;
 453                 }
 454                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 455                 /*
 456                  * It's safe to drop the lock here. We've marked this region for
 457                  * our exclusive use. If the migration fails we will take the
 458                  * lock again and unmark it.
 459                  */
 460                 mutex_unlock(&cma->lock);
 461
 462                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 463                 mutex_lock(&cma_mutex);
 464                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
 465                                      GFP_KERNEL | (no_warn ? __GFP_NOWARN : 0));
 466                 mutex_unlock(&cma_mutex);
 467                 if (ret == 0) {
 468                         page = pfn_to_page(pfn);
 469                         break;
 470                 }
 471
 472                 cma_clear_bitmap(cma, pfn, count);
 473                 if (ret != -EBUSY)
 474                         break;
 475
 476                 pr_debug("%s(): memory range at %p is busy, retrying\n",
 477                          __func__, pfn_to_page(pfn));
 478                 /* try again with a bit different memory target */
 479                 start = bitmap_no + mask + 1;
 480         }
 481
 482         trace_cma_alloc(pfn, page, count, align);
 483
 484         /*
 485          * CMA can allocate multiple page blocks, which results in different
 486          * blocks being marked with different tags. Reset the tags to ignore
 487          * those page blocks.
 488          */
 489         if (page) {
 490                 for (i = 0; i < count; i++)
 491                         page_kasan_tag_reset(page + i);
 492         }
 493
 494         if (ret && !no_warn) {
 495                 pr_err("%s: alloc failed, req-size: %zu pages, ret: %d\n",
 496                         __func__, count, ret);
 497                 cma_debug_show_areas(cma);
 498         }
 499
 500         pr_debug("%s(): returned %p\n", __func__, page);
 501         return page;
 502 }
 503
 504 /**
 505  * cma_release() - release allocated pages
 506  * @cma:   Contiguous memory region for which the allocation is performed.
 507  * @pages: Allocated pages.
 508  * @count: Number of allocated pages.
 509  *
 510  * This function releases memory allocated by cma_alloc().
 511  * It returns false when provided pages do not belong to contiguous area and
 512  * true otherwise.
 513  */
 514 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
 515 {
 516         unsigned long pfn;
 517
 518         if (!cma || !pages)
 519                 return false;
 520
 521         pr_debug("%s(page %p)\n", __func__, (void *)pages);
 522
 523         pfn = page_to_pfn(pages);
 524
 525         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
 526                 return false;
 527
 528         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 529
 530         free_contig_range(pfn, count);
 531         cma_clear_bitmap(cma, pfn, count);
 532         trace_cma_release(pfn, pages, count);
 533
 534         return true;
 535 }
 536
 537 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
 538 {
 539         int i;
 540
 541         for (i = 0; i < cma_area_count; i++) {
 542                 int ret = it(&cma_areas[i], data);
 543
 544                 if (ret)
 545                         return ret;
 546         }
 547
 548         return 0;
 549 }