Linux内核最新的连续内存分配器(CMA)——避免预留大块内存

http://blog.csdn.net/21cnbao/article/details/7309757

在我们使用ARM等嵌入式Linux系统的时候，一个头疼的问题是GPU，Camera，HDMI等都需要预留大量连续内存，这部分内存平时不用， 但是一般的做法又必须先预留着。目前，Marek Szyprowski和Michal Nazarewicz实现了一套全新的Contiguous Memory Allocator。通过这套机制，我们可以做到不预留内存，这些内存平时是可用的，只有当需要的时候才被分配给Camera，HDMI等设备。下面分析 它的基本代码流程。

声明连续内存

内核启动过程中arch/arm/mm/init.c中的arm_memblock_init()会调用dma_contiguous_reserve(min(arm_dma_limit, arm_lowmem_limit));

该函数位于：drivers/base/dma-contiguous.c

/**
 * dma_contiguous_reserve() - reserve area for contiguous memory handling
 * @limit: End address of the reserved memory (optional, 0 for any).
 *
 * This function reserves memory from early allocator. It should be
 * called by arch specific code once the early allocator (memblock or bootmem)
 * has been activated and all other subsystems have already allocated/reserved
 * memory.
 */
void __init dma_contiguous_reserve(phys_addr_t limit)
{
        unsigned long selected_size = 0;

        pr_debug("%s(limit %08lx)\n", __func__, (unsigned long)limit);

        if (size_cmdline != -1) {
                selected_size = size_cmdline;
        } else {
#ifdef CONFIG_CMA_SIZE_SEL_MBYTES
                selected_size = size_bytes;
#elif defined(CONFIG_CMA_SIZE_SEL_PERCENTAGE)
                selected_size = cma_early_percent_memory();
#elif defined(CONFIG_CMA_SIZE_SEL_MIN)
                selected_size = min(size_bytes, cma_early_percent_memory());
#elif defined(CONFIG_CMA_SIZE_SEL_MAX)
                selected_size = max(size_bytes, cma_early_percent_memory());
#endif
        }

        if (selected_size) {
                pr_debug("%s: reserving %ld MiB for global area\n", __func__,
                         selected_size / SZ_1M);

                dma_declare_contiguous(NULL, selected_size, 0, limit);
        }
};

其中的size_bytes定义为：

static const unsigned long size_bytes = CMA_SIZE_MBYTES * SZ_1M; 默认情况下，CMA_SIZE_MBYTES会被定义为16MB，来源于CONFIG_CMA_SIZE_MBYTES=16

->

int __init dma_declare_contiguous(struct device *dev, unsigned long size,
                                  phys_addr_t base, phys_addr_t limit)
{
        ...
        /* Reserve memory */
        if (base) {
                if (memblock_is_region_reserved(base, size) ||
                    memblock_reserve(base, size) < 0) {
                        base = -EBUSY;
                        goto err;
                }
        } else {
                /*
                 * Use __memblock_alloc_base() since
                 * memblock_alloc_base() panic()s.
                 */
                phys_addr_t addr = __memblock_alloc_base(size, alignment, limit);
                if (!addr) {
                        base = -ENOMEM;
                        goto err;
                } else if (addr + size > ~(unsigned long)0) {
                        memblock_free(addr, size);
                        base = -EINVAL;
                        base = -EINVAL;
                        goto err;
                } else {
                        base = addr;
                }
        }

        /*
         * Each reserved area must be initialised later, when more kernel
         * subsystems (like slab allocator) are available.
         */
        r->start = base;
        r->size = size;
        r->dev = dev;
        cma_reserved_count++;
        pr_info("CMA: reserved %ld MiB at %08lx\n", size / SZ_1M,
                (unsigned long)base);

        /* Architecture specific contiguous memory fixup. */
        dma_contiguous_early_fixup(base, size);
        return 0;
err:
        pr_err("CMA: failed to reserve %ld MiB\n", size / SZ_1M);
        return base;
}

由此可见，连续内存区域也是在内核启动的早期，通过__memblock_alloc_base()拿到的。

另外：

drivers/base/dma-contiguous.c里面的core_initcall()会导致cma_init_reserved_areas()被调用：

static int __init cma_init_reserved_areas(void)
{
        struct cma_reserved *r = cma_reserved;
        unsigned i = cma_reserved_count;

        pr_debug("%s()\n", __func__);

        for (; i; --i, ++r) {
                struct cma *cma;
                cma = cma_create_area(PFN_DOWN(r->start),
                                      r->size >> PAGE_SHIFT);
                if (!IS_ERR(cma))
                        dev_set_cma_area(r->dev, cma);
        }
        return 0;
}
core_initcall(cma_init_reserved_areas);

cma_create_area()会调用cma_activate_area(),cma_activate_area()函数则会针对每个page调用：

init_cma_reserved_pageblock(pfn_to_page(base_pfn));

这个函数则会通过set_pageblock_migratetype(page, MIGRATE_CMA)将页设置为MIGRATE_CMA类型的：

#ifdef CONFIG_CMA
/* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
void __init init_cma_reserved_pageblock(struct page *page)
{
        unsigned i = pageblock_nr_pages;
        struct page *p = page;

        do {
                __ClearPageReserved(p);
                set_page_count(p, 0);
        } while (++p, --i);

        set_page_refcounted(page);
        set_pageblock_migratetype(page, MIGRATE_CMA);
        __free_pages(page, pageblock_order);
        totalram_pages += pageblock_nr_pages;
}
#endif

同时其中调用的__free_pages(page, pageblock_order);最终会调用到__free_one_page(page, zone, order, migratetype);
相关的page会被加到MIGRATE_CMA的free_list上面去：

list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);

申请连续内存

申请连续内存仍然使用标准的arch/arm/mm/dma-mapping.c中定义的dma_alloc_coherent()和dma_alloc_writecombine()，这二者会间接调用drivers/base/dma-contiguous.c中的

struct page *dma_alloc_from_contiguous(struct device *dev, int count,
                                       unsigned int align)

 

->

struct page *dma_alloc_from_contiguous(struct device *dev, int count,
                                       unsigned int align)
{
       ...

       for (;;) {
                pageno = bitmap_find_next_zero_area(cma->bitmap, cma->count,
                                                    start, count, mask);
                if (pageno >= cma->count) {
                        ret = -ENOMEM;
                        goto error;
                }

                pfn = cma->base_pfn + pageno;
                ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA);
                if (ret == 0) {
                        bitmap_set(cma->bitmap, pageno, count);
                        break;
                } else if (ret != -EBUSY) {
                        goto error;
                }
                pr_debug("%s(): memory range at %p is busy, retrying\n",
                         __func__, pfn_to_page(pfn));
                /* try again with a bit different memory target */
                start = pageno + mask + 1;
        }
       ...

}

->

int alloc_contig_range(unsigned long start, unsigned long end,

                       unsigned migratetype)

需要隔离page，隔离page的作用通过代码的注释可以体现：

/*
 * What we do here is we mark all pageblocks in range as
 * MIGRATE_ISOLATE.  Because of the way page allocator work, we
 * align the range to MAX_ORDER pages so that page allocator
 * won't try to merge buddies from different pageblocks and
 * change MIGRATE_ISOLATE to some other migration type.
 *
 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
 * migrate the pages from an unaligned range (ie. pages that
 * we are interested in).  This will put all the pages in
 * range back to page allocator as MIGRATE_ISOLATE.
 *
 * When this is done, we take the pages in range from page
 * allocator removing them from the buddy system.  This way
 * page allocator will never consider using them.
 *
 * This lets us mark the pageblocks back as
 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
 * MAX_ORDER aligned range but not in the unaligned, original
 * range are put back to page allocator so that buddy can use
 * them.
 */

ret = start_isolate_page_range(pfn_align_to_maxpage_down(start),
                               pfn_align_to_maxpage_up(end),
                               migratetype);

简单地说，就是把相关的page标记为MIGRATE_ISOLATE，这样buddy系统就不会再使用他们。

/*
 * start_isolate_page_range() -- make page-allocation-type of range of pages
 * to be MIGRATE_ISOLATE.
 * @start_pfn: The lower PFN of the range to be isolated.
 * @end_pfn: The upper PFN of the range to be isolated.
 * @migratetype: migrate type to set in error recovery.
 *
 * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
 * the range will never be allocated. Any free pages and pages freed in the
 * future will not be allocated again.
 *
 * start_pfn/end_pfn must be aligned to pageblock_order.
 * Returns 0 on success and -EBUSY if any part of range cannot be isolated.
 */
int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
                             unsigned migratetype)
{
        unsigned long pfn;
        unsigned long undo_pfn;
        struct page *page;

        BUG_ON((start_pfn) & (pageblock_nr_pages - 1));
        BUG_ON((end_pfn) & (pageblock_nr_pages - 1));

        for (pfn = start_pfn;
             pfn < end_pfn;
             pfn += pageblock_nr_pages) {
                page = __first_valid_page(pfn, pageblock_nr_pages);
                if (page && set_migratetype_isolate(page)) {
                        undo_pfn = pfn;
                        goto undo;
                }
        }
        return 0;
undo:
        for (pfn = start_pfn;
             pfn < undo_pfn;
             pfn += pageblock_nr_pages)
                unset_migratetype_isolate(pfn_to_page(pfn), migratetype);

        return -EBUSY;
}

 

接下来调用__alloc_contig_migrate_range()进行页面隔离和迁移:

static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
{
        /* This function is based on compact_zone() from compaction.c. */

        unsigned long pfn = start;
        unsigned int tries = 0;
        int ret = 0;

        struct compact_control cc = {
                .nr_migratepages = 0,
                .order = -1,
                .zone = page_zone(pfn_to_page(start)),
                .sync = true,
        };
        INIT_LIST_HEAD(&cc.migratepages);

        migrate_prep_local();

        while (pfn < end || !list_empty(&cc.migratepages)) {
                if (fatal_signal_pending(current)) {
                        ret = -EINTR;
                        break;
                }

                if (list_empty(&cc.migratepages)) {
                        cc.nr_migratepages = 0;
                        pfn = isolate_migratepages_range(cc.zone, &cc,
                                                         pfn, end);
                        if (!pfn) {
                                ret = -EINTR;
                                break;
                        }
                        tries = 0;
                } else if (++tries == 5) {
                        ret = ret < 0 ? ret : -EBUSY;
                        break;
                }

                ret = migrate_pages(&cc.migratepages,
                                    __alloc_contig_migrate_alloc,
                                    0, false, true);
        }

        putback_lru_pages(&cc.migratepages);
        return ret > 0 ? 0 : ret;
}

其中的函数migrate_pages()会完成页面的迁移，迁移过程中通过传入的__alloc_contig_migrate_alloc()申请新的page，并将老的page付给新的page：

int migrate_pages(struct list_head *from,
                new_page_t get_new_page, unsigned long private, bool offlining,
                bool sync)
{
        int retry = 1;
        int nr_failed = 0;
        int pass = 0;
        struct page *page;
        struct page *page2;
        int swapwrite = current->flags & PF_SWAPWRITE;
        int rc;

        if (!swapwrite)
                current->flags |= PF_SWAPWRITE;

        for(pass = 0; pass < 10 && retry; pass++) {
                retry = 0;

                list_for_each_entry_safe(page, page2, from, lru) {
                        cond_resched();

                        rc = unmap_and_move(get_new_page, private,
                                                page, pass > 2, offlining,
                                                sync);

                        switch(rc) {
                        case -ENOMEM:
                                goto out;
                        case -EAGAIN:
                                retry++;
                                break;
                        case 0:
                                break;
                        default:
                                /* Permanent failure */
                                nr_failed++;
                                break;
                        }
                }
        }
        rc = 0;
...
}

其中的unmap_and_move()函数较为关键，它定义在mm/migrate.c中

/*
 * Obtain the lock on page, remove all ptes and migrate the page
 * to the newly allocated page in newpage.
 */
static int unmap_and_move(new_page_t get_new_page, unsigned long private,
            struct page *page, int force, bool offlining, bool sync)
{
    int rc = 0;
    int *result = NULL;
    struct page *newpage = get_new_page(page, private, &result);
    int remap_swapcache = 1;
    int charge = 0;
    struct mem_cgroup *mem = NULL;
    struct anon_vma *anon_vma = NULL;

    ...

    /* charge against new page */
    charge = mem_cgroup_prepare_migration(page, newpage, &mem);
    ...

    if (PageWriteback(page)) {
        if (!force || !sync)
            goto uncharge;
        wait_on_page_writeback(page);
    }
    /*
     * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
     * we cannot notice that anon_vma is freed while we migrates a page.
     * This get_anon_vma() delays freeing anon_vma pointer until the end
     * of migration. File cache pages are no problem because of page_lock()
     * File Caches may use write_page() or lock_page() in migration, then,
     * just care Anon page here.
     */
    if (PageAnon(page)) {
        /*
         * Only page_lock_anon_vma() understands the subtleties of
         * getting a hold on an anon_vma from outside one of its mms.
         */
        anon_vma = page_lock_anon_vma(page);
        if (anon_vma) {
            /*
             * Take a reference count on the anon_vma if the
             * page is mapped so that it is guaranteed to
             * exist when the page is remapped later
             */
            get_anon_vma(anon_vma);
            page_unlock_anon_vma(anon_vma);
        } else if (PageSwapCache(page)) {
            /*
             * We cannot be sure that the anon_vma of an unmapped
             * swapcache page is safe to use because we don't
             * know in advance if the VMA that this page belonged
             * to still exists. If the VMA and others sharing the
             * data have been freed, then the anon_vma could
             * already be invalid.
             *
             * To avoid this possibility, swapcache pages get
             * migrated but are not remapped when migration
             * completes
             */
            remap_swapcache = 0;
        } else {
            goto uncharge;
        }
    }

    ...
    /* Establish migration ptes or remove ptes */
    try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);

skip_unmap:
    if (!page_mapped(page))
        rc = move_to_new_page(newpage, page, remap_swapcache);

    if (rc && remap_swapcache)
        remove_migration_ptes(page, page);

    /* Drop an anon_vma reference if we took one */
    if (anon_vma)
        drop_anon_vma(anon_vma);

uncharge:
    if (!charge)
        mem_cgroup_end_migration(mem, page, newpage, rc == 0);
unlock:
    unlock_page(page);

move_newpage:
    ...
}

通过unmap_and_move()，老的page就被迁移过去新的page。

接下来要回收page，回收page的作用是，不至于因为拿了连续的内存后，系统变得内存饥饿：

->

/*
 * Reclaim enough pages to make sure that contiguous allocation
 * will not starve the system.
 */
__reclaim_pages(zone, GFP_HIGHUSER_MOVABLE, end-start);

->

/*
 * Trigger memory pressure bump to reclaim some pages in order to be able to
 * allocate 'count' pages in single page units. Does similar work as
 *__alloc_pages_slowpath() function.
 */
static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count)
{
        enum zone_type high_zoneidx = gfp_zone(gfp_mask);
        struct zonelist *zonelist = node_zonelist(0, gfp_mask);
        int did_some_progress = 0;
        int order = 1;
        unsigned long watermark;

        /*
         * Increase level of watermarks to force kswapd do his job
         * to stabilise at new watermark level.
         */
        __update_cma_watermarks(zone, count);

        /* Obey watermarks as if the page was being allocated */
        watermark = low_wmark_pages(zone) + count;
        while (!zone_watermark_ok(zone, 0, watermark, 0, 0)) {
                wake_all_kswapd(order, zonelist, high_zoneidx, zone_idx(zone));

                did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
                                                      NULL);
                if (!did_some_progress) {
                        /* Exhausted what can be done so it's blamo time */
                        out_of_memory(zonelist, gfp_mask, order, NULL);
                }
        }

        /* Restore original watermark levels. */
        __update_cma_watermarks(zone, -count);

        return count;
}

释放连续内存

内存释放的时候也比较简单，直接就是：

arch/arm/mm/dma-mapping.c：

void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)

->

arch/arm/mm/dma-mapping.c:

static void __free_from_contiguous(struct device *dev, struct page *page,
                                   size_t size)
{
        __dma_remap(page, size, pgprot_kernel);
        dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT);
}

->

bool dma_release_from_contiguous(struct device *dev, struct page *pages,
                                 int count)
{
        ...
        free_contig_range(pfn, count);
        ..

}

->

void free_contig_range(unsigned long pfn, unsigned nr_pages)
{
        for (; nr_pages--; ++pfn)
                __free_page(pfn_to_page(pfn));
}

将page交还给buddy。

内核内存分配的migratetype

内核内存分配的时候，带的标志是GFP_，但是GFP_可以转化为migratetype：

static inline int allocflags_to_migratetype(gfp_t gfp_flags)
{
        WARN_ON((gfp_flags & GFP_MOVABLE_MASK) == GFP_MOVABLE_MASK);

        if (unlikely(page_group_by_mobility_disabled))
                return MIGRATE_UNMOVABLE;

        /* Group based on mobility */
        return (((gfp_flags & __GFP_MOVABLE) != 0) << 1) |
                ((gfp_flags & __GFP_RECLAIMABLE) != 0);
}

之后申请内存的时候，会对比迁移类型匹配的free_list：

page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
                zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
                preferred_zone, migratetype);

另外，笔者也编写了一个测试程序，透过它随时测试CMA的功能：

/*
 * kernel module helper for testing CMA
 *
 * Licensed under GPLv2 or later.
 */

#include <linux/module.h>
#include <linux/device.h>
#include <linux/fs.h>
#include <linux/miscdevice.h>
#include <linux/dma-mapping.h>

#define CMA_NUM  10
static struct device *cma_dev;
static dma_addr_t dma_phys[CMA_NUM];
static void *dma_virt[CMA_NUM];

/* any read request will free coherent memory, eg.
 * cat /dev/cma_test
 */
static ssize_t
cma_test_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
{
    int i;

    for (i = 0; i < CMA_NUM; i++) {
        if (dma_virt[i]) {
            dma_free_coherent(cma_dev, (i + 1) * SZ_1M, dma_virt[i], dma_phys[i]);
            _dev_info(cma_dev, "free virt: %p phys: %p\n", dma_virt[i], (void *)dma_phys[i]);
            dma_virt[i] = NULL;
            break;
        }
    }
    return 0;
}

/*
 * any write request will alloc coherent memory, eg.
 * echo 0 > /dev/cma_test
 */
static ssize_t
cma_test_write(struct file *file, const char __user *buf, size_t count, loff_t *ppos)
{
    int i;
    int ret;

    for (i = 0; i < CMA_NUM; i++) {
        if (!dma_virt[i]) {
            dma_virt[i] = dma_alloc_coherent(cma_dev, (i + 1) * SZ_1M, &dma_phys[i], GFP_KERNEL);

            if (dma_virt[i]) {
                void *p;
                /* touch every page in the allocated memory */
                for (p = dma_virt[i]; p <  dma_virt[i] + (i + 1) * SZ_1M; p += PAGE_SIZE)
                    *(u32 *)p = 0;

                _dev_info(cma_dev, "alloc virt: %p phys: %p\n", dma_virt[i], (void *)dma_phys[i]);
            } else {
                dev_err(cma_dev, "no mem in CMA area\n");
                ret = -ENOMEM;
            }
            break;
        }
    }

    return count;
}

static const struct file_operations cma_test_fops = {
    .owner =    THIS_MODULE,
    .read  =    cma_test_read,
    .write =    cma_test_write,
};

static struct miscdevice cma_test_misc = {
    .name = "cma_test",
    .fops = &cma_test_fops,
};

static int __init cma_test_init(void)
{
    int ret = 0;

    ret = misc_register(&cma_test_misc);
    if (unlikely(ret)) {
        pr_err("failed to register cma test misc device!\n");
        return ret;
    }
    cma_dev = cma_test_misc.this_device;
    cma_dev->coherent_dma_mask = ~0;
    _dev_info(cma_dev, "registered.\n");

    return ret;
}
module_init(cma_test_init);

static void __exit cma_test_exit(void)
{
    misc_deregister(&cma_test_misc);
}
module_exit(cma_test_exit);

MODULE_LICENSE("GPL");
MODULE_AUTHOR("Barry Song <[email protected]>");
MODULE_DESCRIPTION("kernel module to help the test of CMA");
MODULE_ALIAS("CMA test");

申请内存：

# echo 0 > /dev/cma_test

释放内存：

# cat /dev/cma_test

参考链接：

[1] http://www.spinics.net/lists/arm-kernel/msg160854.html

[2] http://www.spinics.net/lists/arm-kernel/msg162063.html

[3] http://lwn.net/Articles/447405/