the-tree/mm/cma.c

0001 // SPDX-License-Identifier: GPL-2.0-or-later
0002 /*
0003  * Contiguous Memory Allocator
0004  *
0005  * Copyright (c) 2010-2011 by Samsung Electronics.
0006  * Copyright IBM Corporation, 2013
0007  * Copyright LG Electronics Inc., 2014
0008  * Written by:
0009  *  Marek Szyprowski <m.szyprowski@samsung.com>
0010  *  Michal Nazarewicz <mina86@mina86.com>
0011  *  Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
0012  *  Joonsoo Kim <iamjoonsoo.kim@lge.com>
0013  */
0014
0015 #define pr_fmt(fmt) "cma: " fmt
0016
0017 #ifdef CONFIG_CMA_DEBUG
0018 #ifndef DEBUG
0019 #  define DEBUG
0020 #endif
0021 #endif
0022 #define CREATE_TRACE_POINTS
0023
0024 #include <linux/memblock.h>
0025 #include <linux/err.h>
0026 #include <linux/mm.h>
0027 #include <linux/sizes.h>
0028 #include <linux/slab.h>
0029 #include <linux/log2.h>
0030 #include <linux/cma.h>
0031 #include <linux/highmem.h>
0032 #include <linux/io.h>
0033 #include <linux/kmemleak.h>
0034 #include <trace/events/cma.h>
0035
0036 #include "cma.h"
0037
0038 struct cma cma_areas[MAX_CMA_AREAS];
0039 unsigned cma_area_count;
0040 static DEFINE_MUTEX(cma_mutex);
0041
0042 phys_addr_t cma_get_base(const struct cma *cma)
0043 {
0044     return PFN_PHYS(cma->base_pfn);
0045 }
0046
0047 unsigned long cma_get_size(const struct cma *cma)
0048 {
0049     return cma->count << PAGE_SHIFT;
0050 }
0051
0052 const char *cma_get_name(const struct cma *cma)
0053 {
0054     return cma->name;
0055 }
0056
0057 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
0058                          unsigned int align_order)
0059 {
0060     if (align_order <= cma->order_per_bit)
0061         return 0;
0062     return (1UL << (align_order - cma->order_per_bit)) - 1;
0063 }
0064
0065 /*
0066  * Find the offset of the base PFN from the specified align_order.
0067  * The value returned is represented in order_per_bits.
0068  */
0069 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
0070                            unsigned int align_order)
0071 {
0072     return (cma->base_pfn & ((1UL << align_order) - 1))
0073         >> cma->order_per_bit;
0074 }
0075
0076 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
0077                           unsigned long pages)
0078 {
0079     return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
0080 }
0081
0082 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
0083                  unsigned long count)
0084 {
0085     unsigned long bitmap_no, bitmap_count;
0086     unsigned long flags;
0087
0088     bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
0089     bitmap_count = cma_bitmap_pages_to_bits(cma, count);
0090
0091     spin_lock_irqsave(&cma->lock, flags);
0092     bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
0093     spin_unlock_irqrestore(&cma->lock, flags);
0094 }
0095
0096 static void __init cma_activate_area(struct cma *cma)
0097 {
0098     unsigned long base_pfn = cma->base_pfn, pfn;
0099     struct zone *zone;
0100
0101     cma->bitmap = bitmap_zalloc(cma_bitmap_maxno(cma), GFP_KERNEL);
0102     if (!cma->bitmap)
0103         goto out_error;
0104
0105     /*
0106      * alloc_contig_range() requires the pfn range specified to be in the
0107      * same zone. Simplify by forcing the entire CMA resv range to be in the
0108      * same zone.
0109      */
0110     WARN_ON_ONCE(!pfn_valid(base_pfn));
0111     zone = page_zone(pfn_to_page(base_pfn));
0112     for (pfn = base_pfn + 1; pfn < base_pfn + cma->count; pfn++) {
0113         WARN_ON_ONCE(!pfn_valid(pfn));
0114         if (page_zone(pfn_to_page(pfn)) != zone)
0115             goto not_in_zone;
0116     }
0117
0118     for (pfn = base_pfn; pfn < base_pfn + cma->count;
0119          pfn += pageblock_nr_pages)
0120         init_cma_reserved_pageblock(pfn_to_page(pfn));
0121
0122     spin_lock_init(&cma->lock);
0123
0124 #ifdef CONFIG_CMA_DEBUGFS
0125     INIT_HLIST_HEAD(&cma->mem_head);
0126     spin_lock_init(&cma->mem_head_lock);
0127 #endif
0128
0129     return;
0130
0131 not_in_zone:
0132     bitmap_free(cma->bitmap);
0133 out_error:
0134     /* Expose all pages to the buddy, they are useless for CMA. */
0135     if (!cma->reserve_pages_on_error) {
0136         for (pfn = base_pfn; pfn < base_pfn + cma->count; pfn++)
0137             free_reserved_page(pfn_to_page(pfn));
0138     }
0139     totalcma_pages -= cma->count;
0140     cma->count = 0;
0141     pr_err("CMA area %s could not be activated\n", cma->name);
0142     return;
0143 }
0144
0145 static int __init cma_init_reserved_areas(void)
0146 {
0147     int i;
0148
0149     for (i = 0; i < cma_area_count; i++)
0150         cma_activate_area(&cma_areas[i]);
0151
0152     return 0;
0153 }
0154 core_initcall(cma_init_reserved_areas);
0155
0156 void __init cma_reserve_pages_on_error(struct cma *cma)
0157 {
0158     cma->reserve_pages_on_error = true;
0159 }
0160
0161 /**
0162  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
0163  * @base: Base address of the reserved area
0164  * @size: Size of the reserved area (in bytes),
0165  * @order_per_bit: Order of pages represented by one bit on bitmap.
0166  * @name: The name of the area. If this parameter is NULL, the name of
0167  *        the area will be set to "cmaN", where N is a running counter of
0168  *        used areas.
0169  * @res_cma: Pointer to store the created cma region.
0170  *
0171  * This function creates custom contiguous area from already reserved memory.
0172  */
0173 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
0174                  unsigned int order_per_bit,
0175                  const char *name,
0176                  struct cma **res_cma)
0177 {
0178     struct cma *cma;
0179
0180     /* Sanity checks */
0181     if (cma_area_count == ARRAY_SIZE(cma_areas)) {
0182         pr_err("Not enough slots for CMA reserved regions!\n");
0183         return -ENOSPC;
0184     }
0185
0186     if (!size || !memblock_is_region_reserved(base, size))
0187         return -EINVAL;
0188
0189     /* alignment should be aligned with order_per_bit */
0190     if (!IS_ALIGNED(CMA_MIN_ALIGNMENT_PAGES, 1 << order_per_bit))
0191         return -EINVAL;
0192
0193     /* ensure minimal alignment required by mm core */
0194     if (!IS_ALIGNED(base | size, CMA_MIN_ALIGNMENT_BYTES))
0195         return -EINVAL;
0196
0197     /*
0198      * Each reserved area must be initialised later, when more kernel
0199      * subsystems (like slab allocator) are available.
0200      */
0201     cma = &cma_areas[cma_area_count];
0202
0203     if (name)
0204         snprintf(cma->name, CMA_MAX_NAME, name);
0205     else
0206         snprintf(cma->name, CMA_MAX_NAME,  "cma%d\n", cma_area_count);
0207
0208     cma->base_pfn = PFN_DOWN(base);
0209     cma->count = size >> PAGE_SHIFT;
0210     cma->order_per_bit = order_per_bit;
0211     *res_cma = cma;
0212     cma_area_count++;
0213     totalcma_pages += (size / PAGE_SIZE);
0214
0215     return 0;
0216 }
0217
0218 /**
0219  * cma_declare_contiguous_nid() - reserve custom contiguous area
0220  * @base: Base address of the reserved area optional, use 0 for any
0221  * @size: Size of the reserved area (in bytes),
0222  * @limit: End address of the reserved memory (optional, 0 for any).
0223  * @alignment: Alignment for the CMA area, should be power of 2 or zero
0224  * @order_per_bit: Order of pages represented by one bit on bitmap.
0225  * @fixed: hint about where to place the reserved area
0226  * @name: The name of the area. See function cma_init_reserved_mem()
0227  * @res_cma: Pointer to store the created cma region.
0228  * @nid: nid of the free area to find, %NUMA_NO_NODE for any node
0229  *
0230  * This function reserves memory from early allocator. It should be
0231  * called by arch specific code once the early allocator (memblock or bootmem)
0232  * has been activated and all other subsystems have already allocated/reserved
0233  * memory. This function allows to create custom reserved areas.
0234  *
0235  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
0236  * reserve in range from @base to @limit.
0237  */
0238 int __init cma_declare_contiguous_nid(phys_addr_t base,
0239             phys_addr_t size, phys_addr_t limit,
0240             phys_addr_t alignment, unsigned int order_per_bit,
0241             bool fixed, const char *name, struct cma **res_cma,
0242             int nid)
0243 {
0244     phys_addr_t memblock_end = memblock_end_of_DRAM();
0245     phys_addr_t highmem_start;
0246     int ret = 0;
0247
0248     /*
0249      * We can't use __pa(high_memory) directly, since high_memory
0250      * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
0251      * complain. Find the boundary by adding one to the last valid
0252      * address.
0253      */
0254     highmem_start = __pa(high_memory - 1) + 1;
0255     pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
0256         __func__, &size, &base, &limit, &alignment);
0257
0258     if (cma_area_count == ARRAY_SIZE(cma_areas)) {
0259         pr_err("Not enough slots for CMA reserved regions!\n");
0260         return -ENOSPC;
0261     }
0262
0263     if (!size)
0264         return -EINVAL;
0265
0266     if (alignment && !is_power_of_2(alignment))
0267         return -EINVAL;
0268
0269     /* Sanitise input arguments. */
0270     alignment = max_t(phys_addr_t, alignment, CMA_MIN_ALIGNMENT_BYTES);
0271     if (fixed && base & (alignment - 1)) {
0272         ret = -EINVAL;
0273         pr_err("Region at %pa must be aligned to %pa bytes\n",
0274             &base, &alignment);
0275         goto err;
0276     }
0277     base = ALIGN(base, alignment);
0278     size = ALIGN(size, alignment);
0279     limit &= ~(alignment - 1);
0280
0281     if (!base)
0282         fixed = false;
0283
0284     /* size should be aligned with order_per_bit */
0285     if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
0286         return -EINVAL;
0287
0288     /*
0289      * If allocating at a fixed base the request region must not cross the
0290      * low/high memory boundary.
0291      */
0292     if (fixed && base < highmem_start && base + size > highmem_start) {
0293         ret = -EINVAL;
0294         pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
0295             &base, &highmem_start);
0296         goto err;
0297     }
0298
0299     /*
0300      * If the limit is unspecified or above the memblock end, its effective
0301      * value will be the memblock end. Set it explicitly to simplify further
0302      * checks.
0303      */
0304     if (limit == 0 || limit > memblock_end)
0305         limit = memblock_end;
0306
0307     if (base + size > limit) {
0308         ret = -EINVAL;
0309         pr_err("Size (%pa) of region at %pa exceeds limit (%pa)\n",
0310             &size, &base, &limit);
0311         goto err;
0312     }
0313
0314     /* Reserve memory */
0315     if (fixed) {
0316         if (memblock_is_region_reserved(base, size) ||
0317             memblock_reserve(base, size) < 0) {
0318             ret = -EBUSY;
0319             goto err;
0320         }
0321     } else {
0322         phys_addr_t addr = 0;
0323
0324         /*
0325          * All pages in the reserved area must come from the same zone.
0326          * If the requested region crosses the low/high memory boundary,
0327          * try allocating from high memory first and fall back to low
0328          * memory in case of failure.
0329          */
0330         if (base < highmem_start && limit > highmem_start) {
0331             addr = memblock_alloc_range_nid(size, alignment,
0332                     highmem_start, limit, nid, true);
0333             limit = highmem_start;
0334         }
0335
0336         /*
0337          * If there is enough memory, try a bottom-up allocation first.
0338          * It will place the new cma area close to the start of the node
0339          * and guarantee that the compaction is moving pages out of the
0340          * cma area and not into it.
0341          * Avoid using first 4GB to not interfere with constrained zones
0342          * like DMA/DMA32.
0343          */
0344 #ifdef CONFIG_PHYS_ADDR_T_64BIT
0345         if (!memblock_bottom_up() && memblock_end >= SZ_4G + size) {
0346             memblock_set_bottom_up(true);
0347             addr = memblock_alloc_range_nid(size, alignment, SZ_4G,
0348                             limit, nid, true);
0349             memblock_set_bottom_up(false);
0350         }
0351 #endif
0352
0353         if (!addr) {
0354             addr = memblock_alloc_range_nid(size, alignment, base,
0355                     limit, nid, true);
0356             if (!addr) {
0357                 ret = -ENOMEM;
0358                 goto err;
0359             }
0360         }
0361
0362         /*
0363          * kmemleak scans/reads tracked objects for pointers to other
0364          * objects but this address isn't mapped and accessible
0365          */
0366         kmemleak_ignore_phys(addr);
0367         base = addr;
0368     }
0369
0370     ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
0371     if (ret)
0372         goto free_mem;
0373
0374     pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
0375         &base);
0376     return 0;
0377
0378 free_mem:
0379     memblock_phys_free(base, size);
0380 err:
0381     pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
0382     return ret;
0383 }
0384
0385 #ifdef CONFIG_CMA_DEBUG
0386 static void cma_debug_show_areas(struct cma *cma)
0387 {
0388     unsigned long next_zero_bit, next_set_bit, nr_zero;
0389     unsigned long start = 0;
0390     unsigned long nr_part, nr_total = 0;
0391     unsigned long nbits = cma_bitmap_maxno(cma);
0392
0393     spin_lock_irq(&cma->lock);
0394     pr_info("number of available pages: ");
0395     for (;;) {
0396         next_zero_bit = find_next_zero_bit(cma->bitmap, nbits, start);
0397         if (next_zero_bit >= nbits)
0398             break;
0399         next_set_bit = find_next_bit(cma->bitmap, nbits, next_zero_bit);
0400         nr_zero = next_set_bit - next_zero_bit;
0401         nr_part = nr_zero << cma->order_per_bit;
0402         pr_cont("%s%lu@%lu", nr_total ? "+" : "", nr_part,
0403             next_zero_bit);
0404         nr_total += nr_part;
0405         start = next_zero_bit + nr_zero;
0406     }
0407     pr_cont("=> %lu free of %lu total pages\n", nr_total, cma->count);
0408     spin_unlock_irq(&cma->lock);
0409 }
0410 #else
0411 static inline void cma_debug_show_areas(struct cma *cma) { }
0412 #endif
0413
0414 /**
0415  * cma_alloc() - allocate pages from contiguous area
0416  * @cma:   Contiguous memory region for which the allocation is performed.
0417  * @count: Requested number of pages.
0418  * @align: Requested alignment of pages (in PAGE_SIZE order).
0419  * @no_warn: Avoid printing message about failed allocation
0420  *
0421  * This function allocates part of contiguous memory on specific
0422  * contiguous memory area.
0423  */
0424 struct page *cma_alloc(struct cma *cma, unsigned long count,
0425                unsigned int align, bool no_warn)
0426 {
0427     unsigned long mask, offset;
0428     unsigned long pfn = -1;
0429     unsigned long start = 0;
0430     unsigned long bitmap_maxno, bitmap_no, bitmap_count;
0431     unsigned long i;
0432     struct page *page = NULL;
0433     int ret = -ENOMEM;
0434
0435     if (!cma || !cma->count || !cma->bitmap)
0436         goto out;
0437
0438     pr_debug("%s(cma %p, count %lu, align %d)\n", __func__, (void *)cma,
0439          count, align);
0440
0441     if (!count)
0442         goto out;
0443
0444     trace_cma_alloc_start(cma->name, count, align);
0445
0446     mask = cma_bitmap_aligned_mask(cma, align);
0447     offset = cma_bitmap_aligned_offset(cma, align);
0448     bitmap_maxno = cma_bitmap_maxno(cma);
0449     bitmap_count = cma_bitmap_pages_to_bits(cma, count);
0450
0451     if (bitmap_count > bitmap_maxno)
0452         goto out;
0453
0454     for (;;) {
0455         spin_lock_irq(&cma->lock);
0456         bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
0457                 bitmap_maxno, start, bitmap_count, mask,
0458                 offset);
0459         if (bitmap_no >= bitmap_maxno) {
0460             spin_unlock_irq(&cma->lock);
0461             break;
0462         }
0463         bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
0464         /*
0465          * It's safe to drop the lock here. We've marked this region for
0466          * our exclusive use. If the migration fails we will take the
0467          * lock again and unmark it.
0468          */
0469         spin_unlock_irq(&cma->lock);
0470
0471         pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
0472         mutex_lock(&cma_mutex);
0473         ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
0474                      GFP_KERNEL | (no_warn ? __GFP_NOWARN : 0));
0475         mutex_unlock(&cma_mutex);
0476         if (ret == 0) {
0477             page = pfn_to_page(pfn);
0478             break;
0479         }
0480
0481         cma_clear_bitmap(cma, pfn, count);
0482         if (ret != -EBUSY)
0483             break;
0484
0485         pr_debug("%s(): memory range at %p is busy, retrying\n",
0486              __func__, pfn_to_page(pfn));
0487
0488         trace_cma_alloc_busy_retry(cma->name, pfn, pfn_to_page(pfn),
0489                        count, align);
0490         /* try again with a bit different memory target */
0491         start = bitmap_no + mask + 1;
0492     }
0493
0494     trace_cma_alloc_finish(cma->name, pfn, page, count, align);
0495
0496     /*
0497      * CMA can allocate multiple page blocks, which results in different
0498      * blocks being marked with different tags. Reset the tags to ignore
0499      * those page blocks.
0500      */
0501     if (page) {
0502         for (i = 0; i < count; i++)
0503             page_kasan_tag_reset(page + i);
0504     }
0505
0506     if (ret && !no_warn) {
0507         pr_err_ratelimited("%s: %s: alloc failed, req-size: %lu pages, ret: %d\n",
0508                    __func__, cma->name, count, ret);
0509         cma_debug_show_areas(cma);
0510     }
0511
0512     pr_debug("%s(): returned %p\n", __func__, page);
0513 out:
0514     if (page) {
0515         count_vm_event(CMA_ALLOC_SUCCESS);
0516         cma_sysfs_account_success_pages(cma, count);
0517     } else {
0518         count_vm_event(CMA_ALLOC_FAIL);
0519         if (cma)
0520             cma_sysfs_account_fail_pages(cma, count);
0521     }
0522
0523     return page;
0524 }
0525
0526 bool cma_pages_valid(struct cma *cma, const struct page *pages,
0527              unsigned long count)
0528 {
0529     unsigned long pfn;
0530
0531     if (!cma || !pages)
0532         return false;
0533
0534     pfn = page_to_pfn(pages);
0535
0536     if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count) {
0537         pr_debug("%s(page %p, count %lu)\n", __func__,
0538                         (void *)pages, count);
0539         return false;
0540     }
0541
0542     return true;
0543 }
0544
0545 /**
0546  * cma_release() - release allocated pages
0547  * @cma:   Contiguous memory region for which the allocation is performed.
0548  * @pages: Allocated pages.
0549  * @count: Number of allocated pages.
0550  *
0551  * This function releases memory allocated by cma_alloc().
0552  * It returns false when provided pages do not belong to contiguous area and
0553  * true otherwise.
0554  */
0555 bool cma_release(struct cma *cma, const struct page *pages,
0556          unsigned long count)
0557 {
0558     unsigned long pfn;
0559
0560     if (!cma_pages_valid(cma, pages, count))
0561         return false;
0562
0563     pr_debug("%s(page %p, count %lu)\n", __func__, (void *)pages, count);
0564
0565     pfn = page_to_pfn(pages);
0566
0567     VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
0568
0569     free_contig_range(pfn, count);
0570     cma_clear_bitmap(cma, pfn, count);
0571     trace_cma_release(cma->name, pfn, pages, count);
0572
0573     return true;
0574 }
0575
0576 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
0577 {
0578     int i;
0579
0580     for (i = 0; i < cma_area_count; i++) {
0581         int ret = it(&cma_areas[i], data);
0582
0583         if (ret)
0584             return ret;
0585     }
0586
0587     return 0;
0588 }