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;
  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 void __init cma_activate_area(struct cma *cma)
  97 {
  98         unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
  99         unsigned i = cma->count >> pageblock_order;
 100         struct zone *zone;
 101
 102         cma->bitmap = bitmap_zalloc(cma_bitmap_maxno(cma), GFP_KERNEL);
 103         if (!cma->bitmap)
 104                 goto out_error;
 105
 106         WARN_ON_ONCE(!pfn_valid(pfn));
 107         zone = page_zone(pfn_to_page(pfn));
 108
 109         do {
 110                 unsigned j;
 111
 112                 base_pfn = pfn;
 113                 for (j = pageblock_nr_pages; j; --j, pfn++) {
 114                         WARN_ON_ONCE(!pfn_valid(pfn));
 115                         /*
 116                          * alloc_contig_range requires the pfn range
 117                          * specified to be in the same zone. Make this
 118                          * simple by forcing the entire CMA resv range
 119                          * to be in the same zone.
 120                          */
 121                         if (page_zone(pfn_to_page(pfn)) != zone)
 122                                 goto not_in_zone;
 123                 }
 124                 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
 125         } while (--i);
 126
 127         mutex_init(&cma->lock);
 128
 129 #ifdef CONFIG_CMA_DEBUGFS
 130         INIT_HLIST_HEAD(&cma->mem_head);
 131         spin_lock_init(&cma->mem_head_lock);
 132 #endif
 133
 134         return;
 135
 136 not_in_zone:
 137         bitmap_free(cma->bitmap);
 138 out_error:
 139         cma->count = 0;
 140         pr_err("CMA area %s could not be activated\n", cma->name);
 141         return;
 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                 cma_activate_area(&cma_areas[i]);
 150
 151         return 0;
 152 }
 153 core_initcall(cma_init_reserved_areas);
 154
 155 /**
 156  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
 157  * @base: Base address of the reserved area
 158  * @size: Size of the reserved area (in bytes),
 159  * @order_per_bit: Order of pages represented by one bit on bitmap.
 160  * @name: The name of the area. If this parameter is NULL, the name of
 161  *        the area will be set to "cmaN", where N is a running counter of
 162  *        used areas.
 163  * @res_cma: Pointer to store the created cma region.
 164  *
 165  * This function creates custom contiguous area from already reserved memory.
 166  */
 167 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
 168                                  unsigned int order_per_bit,
 169                                  const char *name,
 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
 201         if (name)
 202                 snprintf(cma->name, CMA_MAX_NAME, name);
 203         else
 204                 snprintf(cma->name, CMA_MAX_NAME,  "cma%d\n", cma_area_count);
 205
 206         cma->base_pfn = PFN_DOWN(base);
 207         cma->count = size >> PAGE_SHIFT;
 208         cma->order_per_bit = order_per_bit;
 209         *res_cma = cma;
 210         cma_area_count++;
 211         totalcma_pages += (size / PAGE_SIZE);
 212
 213         return 0;
 214 }
 215
 216 /**
 217  * cma_declare_contiguous_nid() - reserve custom contiguous area
 218  * @base: Base address of the reserved area optional, use 0 for any
 219  * @size: Size of the reserved area (in bytes),
 220  * @limit: End address of the reserved memory (optional, 0 for any).
 221  * @alignment: Alignment for the CMA area, should be power of 2 or zero
 222  * @order_per_bit: Order of pages represented by one bit on bitmap.
 223  * @fixed: hint about where to place the reserved area
 224  * @name: The name of the area. See function cma_init_reserved_mem()
 225  * @res_cma: Pointer to store the created cma region.
 226  * @nid: nid of the free area to find, %NUMA_NO_NODE for any node
 227  *
 228  * This function reserves memory from early allocator. It should be
 229  * called by arch specific code once the early allocator (memblock or bootmem)
 230  * has been activated and all other subsystems have already allocated/reserved
 231  * memory. This function allows to create custom reserved areas.
 232  *
 233  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 234  * reserve in range from @base to @limit.
 235  */
 236 int __init cma_declare_contiguous_nid(phys_addr_t base,
 237                         phys_addr_t size, phys_addr_t limit,
 238                         phys_addr_t alignment, unsigned int order_per_bit,
 239                         bool fixed, const char *name, struct cma **res_cma,
 240                         int nid)
 241 {
 242         phys_addr_t memblock_end = memblock_end_of_DRAM();
 243         phys_addr_t highmem_start;
 244         int ret = 0;
 245
 246         /*
 247          * We can't use __pa(high_memory) directly, since high_memory
 248          * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
 249          * complain. Find the boundary by adding one to the last valid
 250          * address.
 251          */
 252         highmem_start = __pa(high_memory - 1) + 1;
 253         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 254                 __func__, &size, &base, &limit, &alignment);
 255
 256         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 257                 pr_err("Not enough slots for CMA reserved regions!\n");
 258                 return -ENOSPC;
 259         }
 260
 261         if (!size)
 262                 return -EINVAL;
 263
 264         if (alignment && !is_power_of_2(alignment))
 265                 return -EINVAL;
 266
 267         /*
 268          * Sanitise input arguments.
 269          * Pages both ends in CMA area could be merged into adjacent unmovable
 270          * migratetype page by page allocator's buddy algorithm. In the case,
 271          * you couldn't get a contiguous memory, which is not what we want.
 272          */
 273         alignment = max(alignment,  (phys_addr_t)PAGE_SIZE <<
 274                           max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
 275         if (fixed && base & (alignment - 1)) {
 276                 ret = -EINVAL;
 277                 pr_err("Region at %pa must be aligned to %pa bytes\n",
 278                         &base, &alignment);
 279                 goto err;
 280         }
 281         base = ALIGN(base, alignment);
 282         size = ALIGN(size, alignment);
 283         limit &= ~(alignment - 1);
 284
 285         if (!base)
 286                 fixed = false;
 287
 288         /* size should be aligned with order_per_bit */
 289         if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
 290                 return -EINVAL;
 291
 292         /*
 293          * If allocating at a fixed base the request region must not cross the
 294          * low/high memory boundary.
 295          */
 296         if (fixed && base < highmem_start && base + size > highmem_start) {
 297                 ret = -EINVAL;
 298                 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
 299                         &base, &highmem_start);
 300                 goto err;
 301         }
 302
 303         /*
 304          * If the limit is unspecified or above the memblock end, its effective
 305          * value will be the memblock end. Set it explicitly to simplify further
 306          * checks.
 307          */
 308         if (limit == 0 || limit > memblock_end)
 309                 limit = memblock_end;
 310
 311         if (base + size > limit) {
 312                 ret = -EINVAL;
 313                 pr_err("Size (%pa) of region at %pa exceeds limit (%pa)\n",
 314                         &size, &base, &limit);
 315                 goto err;
 316         }
 317
 318         /* Reserve memory */
 319         if (fixed) {
 320                 if (memblock_is_region_reserved(base, size) ||
 321                     memblock_reserve(base, size) < 0) {
 322                         ret = -EBUSY;
 323                         goto err;
 324                 }
 325         } else {
 326                 phys_addr_t addr = 0;
 327
 328                 /*
 329                  * All pages in the reserved area must come from the same zone.
 330                  * If the requested region crosses the low/high memory boundary,
 331                  * try allocating from high memory first and fall back to low
 332                  * memory in case of failure.
 333                  */
 334                 if (base < highmem_start && limit > highmem_start) {
 335                         addr = memblock_alloc_range_nid(size, alignment,
 336                                         highmem_start, limit, nid, true);
 337                         limit = highmem_start;
 338                 }
 339
 340                 if (!addr) {
 341                         addr = memblock_alloc_range_nid(size, alignment, base,
 342                                         limit, nid, true);
 343                         if (!addr) {
 344                                 ret = -ENOMEM;
 345                                 goto err;
 346                         }
 347                 }
 348
 349                 /*
 350                  * kmemleak scans/reads tracked objects for pointers to other
 351                  * objects but this address isn't mapped and accessible
 352                  */
 353                 kmemleak_ignore_phys(addr);
 354                 base = addr;
 355         }
 356
 357         ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
 358         if (ret)
 359                 goto free_mem;
 360
 361         pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
 362                 &base);
 363         return 0;
 364
 365 free_mem:
 366         memblock_free(base, size);
 367 err:
 368         pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
 369         return ret;
 370 }
 371
 372 #ifdef CONFIG_CMA_DEBUG
 373 static void cma_debug_show_areas(struct cma *cma)
 374 {
 375         unsigned long next_zero_bit, next_set_bit, nr_zero;
 376         unsigned long start = 0;
 377         unsigned long nr_part, nr_total = 0;
 378         unsigned long nbits = cma_bitmap_maxno(cma);
 379
 380         mutex_lock(&cma->lock);
 381         pr_info("number of available pages: ");
 382         for (;;) {
 383                 next_zero_bit = find_next_zero_bit(cma->bitmap, nbits, start);
 384                 if (next_zero_bit >= nbits)
 385                         break;
 386                 next_set_bit = find_next_bit(cma->bitmap, nbits, next_zero_bit);
 387                 nr_zero = next_set_bit - next_zero_bit;
 388                 nr_part = nr_zero << cma->order_per_bit;
 389                 pr_cont("%s%lu@%lu", nr_total ? "+" : "", nr_part,
 390                         next_zero_bit);
 391                 nr_total += nr_part;
 392                 start = next_zero_bit + nr_zero;
 393         }
 394         pr_cont("=> %lu free of %lu total pages\n", nr_total, cma->count);
 395         mutex_unlock(&cma->lock);
 396 }
 397 #else
 398 static inline void cma_debug_show_areas(struct cma *cma) { }
 399 #endif
 400
 401 /**
 402  * cma_alloc() - allocate pages from contiguous area
 403  * @cma:   Contiguous memory region for which the allocation is performed.
 404  * @count: Requested number of pages.
 405  * @align: Requested alignment of pages (in PAGE_SIZE order).
 406  * @no_warn: Avoid printing message about failed allocation
 407  *
 408  * This function allocates part of contiguous memory on specific
 409  * contiguous memory area.
 410  */
 411 struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align,
 412                        bool no_warn)
 413 {
 414         unsigned long mask, offset;
 415         unsigned long pfn = -1;
 416         unsigned long start = 0;
 417         unsigned long bitmap_maxno, bitmap_no, bitmap_count;
 418         size_t i;
 419         struct page *page = NULL;
 420         int ret = -ENOMEM;
 421
 422         if (!cma || !cma->count || !cma->bitmap)
 423                 return NULL;
 424
 425         pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
 426                  count, align);
 427
 428         if (!count)
 429                 return NULL;
 430
 431         mask = cma_bitmap_aligned_mask(cma, align);
 432         offset = cma_bitmap_aligned_offset(cma, align);
 433         bitmap_maxno = cma_bitmap_maxno(cma);
 434         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 435
 436         if (bitmap_count > bitmap_maxno)
 437                 return NULL;
 438
 439         for (;;) {
 440                 mutex_lock(&cma->lock);
 441                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 442                                 bitmap_maxno, start, bitmap_count, mask,
 443                                 offset);
 444                 if (bitmap_no >= bitmap_maxno) {
 445                         mutex_unlock(&cma->lock);
 446                         break;
 447                 }
 448                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 449                 /*
 450                  * It's safe to drop the lock here. We've marked this region for
 451                  * our exclusive use. If the migration fails we will take the
 452                  * lock again and unmark it.
 453                  */
 454                 mutex_unlock(&cma->lock);
 455
 456                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 457                 mutex_lock(&cma_mutex);
 458                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
 459                                      GFP_KERNEL | (no_warn ? __GFP_NOWARN : 0));
 460                 mutex_unlock(&cma_mutex);
 461                 if (ret == 0) {
 462                         page = pfn_to_page(pfn);
 463                         break;
 464                 }
 465
 466                 cma_clear_bitmap(cma, pfn, count);
 467                 if (ret != -EBUSY)
 468                         break;
 469
 470                 pr_debug("%s(): memory range at %p is busy, retrying\n",
 471                          __func__, pfn_to_page(pfn));
 472                 /* try again with a bit different memory target */
 473                 start = bitmap_no + mask + 1;
 474         }
 475
 476         trace_cma_alloc(pfn, page, count, align);
 477
 478         /*
 479          * CMA can allocate multiple page blocks, which results in different
 480          * blocks being marked with different tags. Reset the tags to ignore
 481          * those page blocks.
 482          */
 483         if (page) {
 484                 for (i = 0; i < count; i++)
 485                         page_kasan_tag_reset(page + i);
 486         }
 487
 488         if (ret && !no_warn) {
 489                 pr_err("%s: alloc failed, req-size: %zu pages, ret: %d\n",
 490                         __func__, count, ret);
 491                 cma_debug_show_areas(cma);
 492         }
 493
 494         pr_debug("%s(): returned %p\n", __func__, page);
 495         return page;
 496 }
 497
 498 /**
 499  * cma_release() - release allocated pages
 500  * @cma:   Contiguous memory region for which the allocation is performed.
 501  * @pages: Allocated pages.
 502  * @count: Number of allocated pages.
 503  *
 504  * This function releases memory allocated by cma_alloc().
 505  * It returns false when provided pages do not belong to contiguous area and
 506  * true otherwise.
 507  */
 508 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
 509 {
 510         unsigned long pfn;
 511
 512         if (!cma || !pages)
 513                 return false;
 514
 515         pr_debug("%s(page %p)\n", __func__, (void *)pages);
 516
 517         pfn = page_to_pfn(pages);
 518
 519         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
 520                 return false;
 521
 522         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 523
 524         free_contig_range(pfn, count);
 525         cma_clear_bitmap(cma, pfn, count);
 526         trace_cma_release(pfn, pages, count);
 527
 528         return true;
 529 }
 530
 531 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
 532 {
 533         int i;
 534
 535         for (i = 0; i < cma_area_count; i++) {
 536                 int ret = it(&cma_areas[i], data);
 537
 538                 if (ret)
 539                         return ret;
 540         }
 541
 542         return 0;
 543 }