知识屋:更实用的电脑技术知识网站
所在位置:首页 > 操作系统 > linux

Linux内存管理之slab机制(初始化)

发布时间:2014-09-05 16:36:53作者:知识屋

一、内核启动早期初始化
start_kernel()->mm_init()->kmem_cache_init()
 
执行流程:
 
1,初始化静态initkmem_list3三链;
 
2,初始化cache_cache的nodelists字段为1中的三链;
 
3,根据内存情况初始化每个slab占用的页面数变量slab_break_gfp_order;
 
4,将cache_cache加入cache_chain链表中,初始化cache_cache;
 
5,创建kmalloc所用的general cache:
 
1)cache的名称和大小存放在两个数据结构对应的数组中,对应大小的cache可以从size数组中找到;
 
2)先创建INDEX_AC和INDEX_L3下标的cache;
 
3)循环创建size数组中各个大小的cache;
 
6,替换静态本地cache全局变量:
 
1) 替换cache_cache中的arry_cache,本来指向静态变量initarray_cache.cache;
 
2) 替换malloc_sizes[INDEX_AC].cs_cachep的local cache,原本指向静态变量initarray_generic.cache;
 
7,替换静态三链
 
1)替换cache_cache三链,原本指向静态变量initkmem_list3;
 
2)替换malloc_sizes[INDEX_AC].cs_cachep三链,原本指向静态变量initkmem_list3;
 
8,更新初始化进度
 
www.zhishiwu.com
/*
 * Initialisation.  Called after the page allocator have been initialised and
 * before smp_init().
 */ 
void __init kmem_cache_init(void) 

    size_t left_over; 
    struct cache_sizes *sizes; 
    struct cache_names *names; 
    int i; 
    int order; 
    int node; 
    /* 在slab初始化好之前,无法通过kmalloc分配初始化过程中必要的一些对象
    ,只能使用静态的全局变量
    ,待slab初始化后期,再使用kmalloc动态分配的对象替换全局变量*/ 
 
    /* 如前所述,先借用全局变量initkmem_list3表示的slab三链
    ,每个内存节点对应一组slab三链。initkmem_list3是个slab三链数组,对于每个内存节点,包含三组
    :struct kmem_cache的slab三链、struct arraycache_init的slab 三链、struct kmem_list3的slab三链
    。这里循环初始化所有内存节点的所有slab三链*/ 
    if (num_possible_nodes() == 1) 
        use_alien_caches = 0; 
    /*初始化所有node的所有slab中的三个链表*/ 
    for (i = 0; i < NUM_INIT_LISTS; i++) { 
        kmem_list3_init(&initkmem_list3[i]); 
        /* 全局变量cache_cache指向的slab cache包含所有struct kmem_cache对象,不包含cache_cache本身
        。这里初始化所有内存节点的struct kmem_cache的slab三链为空。*/ 
        if (i < MAX_NUMNODES)  
            cache_cache.nodelists[i] = NULL; 
    } 
    /* 设置struct kmem_cache的slab三链指向initkmem_list3中的一组slab三链,
    CACHE_CACHE为cache在内核cache链表中的索引,
    struct kmem_cache对应的cache是内核中创建的第一个cache
    ,故CACHE_CACHE为0 */ 
    set_up_list3s(&cache_cache, CACHE_CACHE); 
 
    /*
     * Fragmentation resistance on low memory - only use bigger
     * page orders on machines with more than 32MB of memory.
     */ 
     /* 全局变量slab_break_gfp_order为每个slab最多占用几个页面
     ,用来抑制碎片,比如大小为3360的对象
     ,如果其slab只占一个页面,碎片为736
     ,slab占用两个页面,则碎片大小也翻倍
     。只有当对象很大
     ,以至于slab中连一个对象都放不下时
     ,才可以超过这个值
     。有两个可能的取值
     :当可用内存大于32MB时
     ,BREAK_GFP_ORDER_HI为1
     ,即每个slab最多占用2个页面
     ,只有当对象大小大于8192时
     ,才可以突破slab_break_gfp_order的限制
     。小于等于32MB时BREAK_GFP_ORDER_LO为0。*/ 
    if (totalram_pages > (32 << 20) >> PAGE_SHIFT) 
        slab_break_gfp_order = BREAK_GFP_ORDER_HI; 
 
    /* Bootstrap is tricky, because several objects are allocated
     * from caches that do not exist yet:
     * 1) initialize the cache_cache cache: it contains the struct
     *    kmem_cache structures of all caches, except cache_cache itself:
     *    cache_cache is statically allocated.
     *    Initially an __init data area is used for the head array and the
     *    kmem_list3 structures, it's replaced with a kmalloc allocated
     *    array at the end of the bootstrap.
     * 2) Create the first kmalloc cache.
     *    The struct kmem_cache for the new cache is allocated normally.
     *    An __init data area is used for the head array.
     * 3) Create the remaining kmalloc caches, with minimally sized
     *    head arrays.
     * 4) Replace the __init data head arrays for cache_cache and the first
     *    kmalloc cache with kmalloc allocated arrays.
     * 5) Replace the __init data for kmem_list3 for cache_cache and
     *    the other cache's with kmalloc allocated memory.
     * 6) Resize the head arrays of the kmalloc caches to their final sizes.
     */ 
 
    node = numa_node_id(); 
 
    /* 1) create the cache_cache */ 
    /* 第一步,创建struct kmem_cache所在的cache,由全局变量cache_cache指向
    ,这里只是初始化数据结构
    ,并未真正创建这些对象,要待分配时才创建。*/ 
    /* 全局变量cache_chain是内核slab cache链表的表头*/ 
    INIT_LIST_HEAD(&cache_chain); 
     
    /* 将cache_cache加入到slab cache链表*/ 
    list_add(&cache_cache.next, &cache_chain); 
 
    /* 设置cache着色基本单位为cache line的大小:32字节*/ 
    cache_cache.colour_off = cache_line_size(); 
    /*  初始化cache_cache的local cache,同样这里也不能使用kmalloc
    ,需要使用静态分配的全局变量initarray_cache */ 
    cache_cache.array[smp_processor_id()] = &initarray_cache.cache; 
    /* 初始化slab链表,用全局变量*/ 
    cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; 
 
    /*
     * struct kmem_cache size depends on nr_node_ids, which
     * can be less than MAX_NUMNODES.
     */ 
     /* buffer_size保存slab中对象的大小,这里是计算struct kmem_cache的大小
     ,nodelists是最后一个成员
     ,nr_node_ids保存内存节点个数,UMA为1
     ,所以nodelists偏移加上1个struct kmem_list3 的大小即为struct kmem_cache的大小*/ 
    cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) + 
                 nr_node_ids * sizeof(struct kmem_list3 *); 
#if DEBUG 
    cache_cache.obj_size = cache_cache.buffer_size; 
#endif 
    /* 将对象大小与cache line大小对齐*/ 
    cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, 
                    cache_line_size()); 
    /* 计算对象大小的倒数,用于计算对象在slab中的索引*/ 
    cache_cache.reciprocal_buffer_size = 
        reciprocal_value(cache_cache.buffer_size); 
 
    for (order = 0; order < MAX_ORDER; order++) { 
        /* 计算cache_cache中的对象数目*/ 
        cache_estimate(order, cache_cache.buffer_size, 
            cache_line_size(), 0, &left_over, &cache_cache.num); 
        /* num不为0意味着创建struct kmem_cache对象成功,退出*/ 
        if (cache_cache.num) 
            break; 
    } 
    BUG_ON(!cache_cache.num); 
     /* gfporder表示本slab包含2^gfporder个页面*/ 
    cache_cache.gfporder = order; 
      /* 着色区的大小,以colour_off为单位*/ 
    cache_cache.colour = left_over / cache_cache.colour_off; 
    /* slab管理对象的大小*/ 
    cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + 
                      sizeof(struct slab), cache_line_size()); 
 
    /* 2+3) create the kmalloc caches */ 
    /* 第二步,创建kmalloc所用的general cache
    ,kmalloc所用的对象按大小分级
    ,malloc_sizes保存大小,cache_names保存cache名*/ 
    sizes = malloc_sizes; 
    names = cache_names; 
 
    /*
     * Initialize the caches that provide memory for the array cache and the
     * kmem_list3 structures first.  Without this, further allocations will
     * bug.
     */ 
    /* 首先创建struct array_cache和struct kmem_list3所用的general cache
    ,它们是后续初始化动作的基础*/ 
    /* INDEX_AC是计算local cache所用的struct arraycache_init对象在kmalloc size中的索引
    ,即属于哪一级别大小的general cache
    ,创建此大小级别的cache为local cache所用*/ 
    sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, 
                    sizes[INDEX_AC].cs_size, 
                    ARCH_KMALLOC_MINALIGN, 
                    ARCH_KMALLOC_FLAGS|SLAB_PANIC, 
                    NULL); 
    /* 如果struct kmem_list3和struct arraycache_init对应的kmalloc size索引不同
    ,即大小属于不同的级别
    ,则创建struct kmem_list3所用的cache,否则共用一个cache */ 
    if (INDEX_AC != INDEX_L3) { 
        sizes[INDEX_L3].cs_cachep = 
            kmem_cache_create(names[INDEX_L3].name, 
                sizes[INDEX_L3].cs_size, 
                ARCH_KMALLOC_MINALIGN, 
                ARCH_KMALLOC_FLAGS|SLAB_PANIC, 
                NULL); 
    } 
    /* 创建完上述两个general cache后,slab early init阶段结束,在此之前
    ,不允许创建外置式slab */ 
    slab_early_init = 0; 
 
    /* 循环创建kmalloc各级别的general cache */ 
    while (sizes->cs_size != ULONG_MAX) { 
        /*
         * For performance, all the general caches are L1 aligned.
         * This should be particularly beneficial on SMP boxes, as it
         * eliminates "false sharing".
         * Note for systems short on memory removing the alignment will
         * allow tighter packing of the smaller caches.
         */ 
         /* 某级别的kmalloc cache还未创建,创建之,struct kmem_list3和
         struct arraycache_init对应的cache已经创建过了*/ 
        if (!sizes->cs_cachep) { 
            sizes->cs_cachep = kmem_cache_create(names->name, 
                    sizes->cs_size, 
                    ARCH_KMALLOC_MINALIGN, 
                    ARCH_KMALLOC_FLAGS|SLAB_PANIC, 
                    NULL); 
        } 
#ifdef CONFIG_ZONE_DMA 
        sizes->cs_dmacachep = kmem_cache_create( 
                    names->name_dma, 
                    sizes->cs_size, 
                    ARCH_KMALLOC_MINALIGN, 
                    ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| 
                        SLAB_PANIC, 
                    NULL); 
#endif 
        sizes++; 
        names++; 
    } 
    /* 至此,kmalloc general cache已经创建完毕,可以拿来使用了*/ 
    /* 4) Replace the bootstrap head arrays */ 
    /* 第四步,用kmalloc对象替换静态分配的全局变量
    。到目前为止一共使用了两个全局local cache
    ,一个是cache_cache的local cache指向initarray_cache.cache
    ,另一个是malloc_sizes[INDEX_AC].cs_cachep的local cache指向initarray_generic.cache
    ,参见setup_cpu_cache函数。这里替换它们。*/ 
    { 
        struct array_cache *ptr; 
        /* 申请cache_cache所用local cache的空间*/ 
        ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); 
 
        BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); 
        /* 复制原cache_cache的local cache,即initarray_cache,到新的位置*/ 
        memcpy(ptr, cpu_cache_get(&cache_cache), 
               sizeof(struct arraycache_init)); 
        /*
         * Do not assume that spinlocks can be initialized via memcpy:
         */ 
        spin_lock_init(&ptr->lock); 
        /* cache_cache的local cache指向新的位置*/ 
        cache_cache.array[smp_processor_id()] = ptr; 
        /* 申请malloc_sizes[INDEX_AC].cs_cachep所用local cache的空间*/ 
        ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); 
 
        BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) 
               != &initarray_generic.cache); 
        /* 复制原local cache到新分配的位置,注意此时local cache的大小是固定的*/ 
        memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), 
               sizeof(struct arraycache_init)); 
        /*
         * Do not assume that spinlocks can be initialized via memcpy:
         */ 
        spin_lock_init(&ptr->lock); 
 
        malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = 
            ptr; 
    } 
    /* 5) Replace the bootstrap kmem_list3's */ 
    /* 第五步,与第四步类似,用kmalloc的空间替换静态分配的slab三链*/ 
    { 
        int nid; 
      /* UMA只有一个节点*/ 
        for_each_online_node(nid) { 
            /* 复制struct kmem_cache的slab三链*/ 
            init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); 
            /* 复制struct arraycache_init的slab三链*/ 
            init_list(malloc_sizes[INDEX_AC].cs_cachep, 
                  &initkmem_list3[SIZE_AC + nid], nid); 
            /* 复制struct kmem_list3的slab三链*/ 
            if (INDEX_AC != INDEX_L3) { 
                init_list(malloc_sizes[INDEX_L3].cs_cachep, 
                      &initkmem_list3[SIZE_L3 + nid], nid); 
            } 
        } 
    } 
    /* 更新slab系统初始化进度*/ 
    g_cpucache_up = EARLY; 

辅助操作
 
1,slab三链初始化
 
www.zhishiwu.com
static void kmem_list3_init(struct kmem_list3 *parent) 

    INIT_LIST_HEAD(&parent->slabs_full); 
    INIT_LIST_HEAD(&parent->slabs_partial); 
    INIT_LIST_HEAD(&parent->slabs_free); 
    parent->shared = NULL; 
    parent->alien = NULL; 
    parent->colour_next = 0; 
    spin_lock_init(&parent->list_lock); 
    parent->free_objects = 0; 
    parent->free_touched = 0; 

2,slab三链静态数据初始化
 
www.zhishiwu.com
 /*设置cache的slab三链指向静态分配的全局变量*/ 
static void __init set_up_list3s(struct kmem_cache *cachep, int index) 

    int node; 
    /* UMA只有一个节点*/ 
    for_each_online_node(node) { 
        /* 全局变量initkmem_list3是初始化阶段使用的slab三链*/ 
        cachep->nodelists[node] = &initkmem_list3[index + node]; 
        /* 设置回收时间*/ 
        cachep->nodelists[node]->next_reap = jiffies + 
            REAPTIMEOUT_LIST3 + 
            ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 
    } 

3,计算每个slab中对象的数目
 
www.zhishiwu.com
/*
 * Calculate the number of objects and left-over bytes for a given buffer size.
 */ 
 /*计算每个slab中对象的数目。*/ 
 /*
1)        gfporder:slab由2gfporder个页面组成。
2)        buffer_size:对象的大小。
3)        align:对象的对齐方式。
4)        flags:内置式slab还是外置式slab。
5)        left_over:slab中浪费空间的大小。
6)        num:slab中的对象数目。
*/ 
static void cache_estimate(unsigned long gfporder, size_t buffer_size, 
               size_t align, int flags, size_t *left_over, 
               unsigned int *num) 

    int nr_objs; 
    size_t mgmt_size; 
    /* slab大小为1<<order个页面*/ 
    size_t slab_size = PAGE_SIZE << gfporder; 
 
    /*
     * The slab management structure can be either off the slab or
     * on it. For the latter case, the memory allocated for a
     * slab is used for:
     *
     * - The struct slab
     * - One kmem_bufctl_t for each object
     * - Padding to respect alignment of @align
     * - @buffer_size bytes for each object
     *
     * If the slab management structure is off the slab, then the
     * alignment will already be calculated into the size. Because
     * the slabs are all pages aligned, the objects will be at the
     * correct alignment when allocated.
     */ 
    if (flags & CFLGS_OFF_SLAB) { 
        /* 外置式slab */ 
        mgmt_size = 0; 
        /* slab页面不含slab管理对象,全部用来存储slab对象*/ 
        nr_objs = slab_size / buffer_size; 
        /* 对象数不能超过上限*/ 
        if (nr_objs > SLAB_LIMIT) 
            nr_objs = SLAB_LIMIT; 
    } else { 
        /*
         * Ignore padding for the initial guess. The padding
         * is at most @align-1 bytes, and @buffer_size is at
         * least @align. In the worst case, this result will
         * be one greater than the number of objects that fit
         * into the memory allocation when taking the padding
         * into account.
         *//* 内置式slab,slab管理对象与slab对象在一起
         ,此时slab页面中包含:一个struct slab对象,一个kmem_bufctl_t数组,slab对象。
         kmem_bufctl_t数组大小与slab对象数目相同*/ 
        nr_objs = (slab_size - sizeof(struct slab)) / 
              (buffer_size + sizeof(kmem_bufctl_t)); 
 
        /*
         * This calculated number will be either the right
         * amount, or one greater than what we want.
         *//* 计算cache line对齐后的大小,如果超出了slab总的大小,则对象数减一*/ 
        if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size 
               > slab_size) 
            nr_objs--; 
 
        if (nr_objs > SLAB_LIMIT) 
            nr_objs = SLAB_LIMIT; 
        /* 计算cache line对齐后slab管理对象的大小*/ 
        mgmt_size = slab_mgmt_size(nr_objs, align); 
    } 
    *num = nr_objs;/* 保存slab对象数目*/ 
    /* 计算浪费空间的大小*/ 
    *left_over = slab_size - nr_objs*buffer_size - mgmt_size; 

辅助数据结构与变量
 
Linux内核中将所有的通用cache以不同的大小存放在数组中,以方便查找。其中malloc_sizes[]数组为cache_sizes类型的数组,存放各个cache的大小;cache_names[]数组为cache_names结构类型数组,存放各个cache大小的名称;malloc_sizes[]数组和cache_names[]数组下标对应,也就是说cache_names[i]名称的cache对应的大小为malloc_sizes[i]。
 
www.zhishiwu.com
/* Size description struct for general caches. */ 
struct cache_sizes { 
    size_t          cs_size; 
    struct kmem_cache   *cs_cachep; 
#ifdef CONFIG_ZONE_DMA 
    struct kmem_cache   *cs_dmacachep; 
#endif 
}; 
/*
 * These are the default caches for kmalloc. Custom caches can have other sizes.
 */ 
struct cache_sizes malloc_sizes[] = { 
#define CACHE(x) { .cs_size = (x) }, 
#include <linux/kmalloc_sizes.h> 
    CACHE(ULONG_MAX) 
#undef CACHE 
}; 
www.zhishiwu.com
/* Must match cache_sizes above. Out of line to keep cache footprint low. */ 
struct cache_names { 
    char *name; 
    char *name_dma; 
}; 
 
static struct cache_names __initdata cache_names[] = { 
#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, 
#include <linux/kmalloc_sizes.h> 
    {NULL,} 
#undef CACHE 
}; 
二、内核启动末期初始化
 
1,根据对象大小计算local cache中对象数目上限;
 
2,借助数据结构ccupdate_struct操作cpu本地cache。为每个在线cpu分配cpu本地cache;
 
3,用新分配的cpu本地cache替换原有的cache;
 
4,更新slab三链以及cpu本地共享cache。
 
第二阶段代码分析
 
Start_kernel()->kmem_cache_init_late()
 
www.zhishiwu.com
/*Slab系统初始化分两个部分,先初始化一些基本的,待系统初始化工作进行的差不多时,再配置一些特殊功能。*/ 
void __init kmem_cache_init_late(void) 

    struct kmem_cache *cachep; 
    /* 初始化阶段local cache的大小是固定的,要根据对象大小重新计算*/ 
    /* 6) resize the head arrays to their final sizes */ 
    mutex_lock(&cache_chain_mutex); 
    list_for_each_entry(cachep, &cache_chain, next) 
        if (enable_cpucache(cachep, GFP_NOWAIT)) 
            BUG(); 
    mutex_unlock(&cache_chain_mutex); 
 
    /* Done! */ 
    /* 大功告成,general cache终于全部建立起来了*/ 
    g_cpucache_up = FULL; 
 
    /* Annotate slab for lockdep -- annotate the malloc caches */ 
    init_lock_keys(); 
 
    /*
     * Register a cpu startup notifier callback that initializes
     * cpu_cache_get for all new cpus
     */ 
     /* 注册cpu up回调函数,cpu up时配置local cache */ 
    register_cpu_notifier(&cpucache_notifier); 
 
    /*
     * The reap timers are started later, with a module init call: That part
     * of the kernel is not yet operational.
     */ 

www.zhishiwu.com
/* Called with cache_chain_mutex held always */ 
/*local cache 初始化*/ 
static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) 

    int err; 
    int limit, shared; 
 
    /*
     * The head array serves three purposes:
     * - create a LIFO ordering, i.e. return objects that are cache-warm
     * - reduce the number of spinlock operations.
     * - reduce the number of linked list operations on the slab and
     *   bufctl chains: array operations are cheaper.
     * The numbers are guessed, we should auto-tune as described by
     * Bonwick.
     */ /* 根据对象大小计算local cache中对象数目上限*/ 
    if (cachep->buffer_size > 131072) 
        limit = 1; 
    else if (cachep->buffer_size > PAGE_SIZE) 
        limit = 8; 
    else if (cachep->buffer_size > 1024) 
        limit = 24; 
    else if (cachep->buffer_size > 256) 
        limit = 54; 
    else 
        limit = 120; 
 
    /*
     * CPU bound tasks (e.g. network routing) can exhibit cpu bound
     * allocation behaviour: Most allocs on one cpu, most free operations
     * on another cpu. For these cases, an efficient object passing between
     * cpus is necessary. This is provided by a shared array. The array
     * replaces Bonwick's magazine layer.
     * On uniprocessor, it's functionally equivalent (but less efficient)
     * to a larger limit. Thus disabled by default.
     */ 
    shared = 0; 
    /* 多核系统,设置shared local cache中对象数目*/ 
    if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) 
        shared = 8; 
 
#if DEBUG 
    /*
     * With debugging enabled, large batchcount lead to excessively long
     * periods with disabled local interrupts. Limit the batchcount
     */ 
    if (limit > 32) 
        limit = 32; 
#endif 
    /* 配置local cache */ 
    err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); 
    if (err) 
        printk(KERN_ERR "enable_cpucache failed for %s, error %d./n", 
               cachep->name, -err); 
    return err; 

www.zhishiwu.com
/* Always called with the cache_chain_mutex held */ 
/*配置local cache、shared local cache和slab三链*/ 
static int do_tune_cpucache(struct kmem_cache *cachep, int limit, 
                int batchcount, int shared, gfp_t gfp) 

    struct ccupdate_struct *new; 
    int i; 
 
    new = kzalloc(sizeof(*new), gfp); 
    if (!new) 
        return -ENOMEM; 
    /* 为每个cpu分配新的struct array_cache对象*/ 
    for_each_online_cpu(i) { 
        new->new[i] = alloc_arraycache(cpu_to_node(i), limit, 
                        batchcount, gfp); 
        if (!new->new[i]) { 
            for (i--; i >= 0; i--) 
                kfree(new->new[i]); 
            kfree(new); 
            return -ENOMEM; 
        } 
    } 
    new->cachep = cachep; 
    /* 用新的struct array_cache对象替换旧的struct array_cache对象
    ,在支持cpu热插拔的系统上,离线cpu可能没有释放local cache
    ,使用的仍是旧local cache,参见__kmem_cache_destroy函数
    。虽然cpu up时要重新配置local cache,也无济于事。考虑下面的情景
    :共有Cpu A和Cpu B,Cpu B down后,destroy Cache X,由于此时Cpu B是down状态
    ,所以Cache X中Cpu B的local cache未释放,过一段时间Cpu B又up了
    ,更新cache_chain 链中所有cache的local cache,但此时Cache X对象已经释放回
    cache_cache中了,其Cpu B local cache并未被更新。又过了一段时间
    ,系统需要创建新的cache,将Cache X对象分配出去,其Cpu B仍然是旧的
    local cache,需要进行更新。
    */ 
    on_each_cpu(do_ccupdate_local, (void *)new, 1); 
 
    check_irq_on(); 
    cachep->batchcount = batchcount; 
    cachep->limit = limit; 
    cachep->shared = shared; 
    /* 释放旧的local cache */ 
    for_each_online_cpu(i) { 
        struct array_cache *ccold = new->new[i]; 
        if (!ccold) 
            continue; 
        spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 
        /* 释放旧local cache中的对象*/ 
        free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); 
        spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 
        /* 释放旧的struct array_cache对象*/ 
        kfree(ccold); 
    } 
    kfree(new); 
    /* 初始化shared local cache 和slab三链*/ 
    return alloc_kmemlist(cachep, gfp); 

更新本地cache
 
www.zhishiwu.com
/*更新每个cpu的struct array_cache对象*/ 
static void do_ccupdate_local(void *info) 

    struct ccupdate_struct *new = info; 
    struct array_cache *old; 
 
    check_irq_off(); 
    old = cpu_cache_get(new->cachep); 
     /* 指向新的struct array_cache对象*/ 
    new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; 
        /* 保存旧的struct array_cache对象*/ 
    new->new[smp_processor_id()] = old; 

www.zhishiwu.com
/*初始化shared local cache和slab三链,初始化完成后,slab三链中没有任何slab*/ 
static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) 

    int node; 
    struct kmem_list3 *l3; 
    struct array_cache *new_shared; 
    struct array_cache **new_alien = NULL; 
 
    for_each_online_node(node) { 
         /* NUMA相关*/ 
                if (use_alien_caches) { 
                        new_alien = alloc_alien_cache(node, cachep->limit, gfp); 
                        if (!new_alien) 
                                goto fail; 
                } 
 
        new_shared = NULL; 
        if (cachep->shared) { 
            /* 分配shared local cache */ 
            new_shared = alloc_arraycache(node, 
                cachep->shared*cachep->batchcount, 
                    0xbaadf00d, gfp); 
            if (!new_shared) { 
                free_alien_cache(new_alien); 
                goto fail; 
            } 
        } 
        /* 获得旧的slab三链*/ 
        l3 = cachep->nodelists[node]; 
        if (l3) { 
            /* 就slab三链指针不为空,需要先释放旧的资源*/ 
            struct array_cache *shared = l3->shared; 
 
            spin_lock_irq(&l3->list_lock); 
            /* 释放旧的shared local cache中的对象*/ 
            if (shared) 
                free_block(cachep, shared->entry, 
                        shared->avail, node); 
            /* 指向新的shared local cache */ 
            l3->shared = new_shared; 
            if (!l3->alien) { 
                l3->alien = new_alien; 
                new_alien = NULL; 
            }/* 计算cache中空闲对象的上限*/ 
            l3->free_limit = (1 + nr_cpus_node(node)) * 
                    cachep->batchcount + cachep->num; 
            spin_unlock_irq(&l3->list_lock); 
            /* 释放旧shared local cache的struct array_cache对象*/ 
            kfree(shared); 
            free_alien_cache(new_alien); 
            continue;/*访问下一个节点*/ 
        } 
         /* 如果没有旧的l3,分配新的slab三链*/ 
        l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); 
        if (!l3) { 
            free_alien_cache(new_alien); 
            kfree(new_shared); 
            goto fail; 
        } 
         /* 初始化slab三链*/ 
        kmem_list3_init(l3); 
        l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 
                ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 
        l3->shared = new_shared; 
        l3->alien = new_alien; 
        l3->free_limit = (1 + nr_cpus_node(node)) * 
                    cachep->batchcount + cachep->num; 
        cachep->nodelists[node] = l3; 
    } 
    return 0; 
 
fail: 
    if (!cachep->next.next) { 
        /* Cache is not active yet. Roll back what we did */ 
        node--; 
        while (node >= 0) { 
            if (cachep->nodelists[node]) { 
                l3 = cachep->nodelists[node]; 
 
                kfree(l3->shared); 
                free_alien_cache(l3->alien); 
                kfree(l3); 
                cachep->nodelists[node] = NULL; 
            } 
            node--; 
        } 
    } 
    return -ENOMEM; 

看一个辅助函数
 
 
/*分配struct array_cache对象。*/ 
static struct array_cache *alloc_arraycache(int node, int entries, 
                        int batchcount, gfp_t gfp) 

    /* struct array_cache后面紧接着的是entry数组,合在一起申请内存*/ 
    int memsize = sizeof(void *) * entries + sizeof(struct array_cache); 
    struct array_cache *nc = NULL; 
    /* 分配一个local cache对象,kmalloc从general cache中分配*/ 
    nc = kmalloc_node(memsize, gfp, node); 
    /*
     * The array_cache structures contain pointers to free object.
     * However, when such objects are allocated or transfered to another
     * cache the pointers are not cleared and they could be counted as
     * valid references during a kmemleak scan. Therefore, kmemleak must
     * not scan such objects.
     */ 
    kmemleak_no_scan(nc); 
     /* 初始化local cache */ 
    if (nc) { 
        nc->avail = 0; 
        nc->limit = entries; 
        nc->batchcount = batchcount; 
        nc->touched = 0; 
        spin_lock_init(&nc->lock); 
    } 
    return nc; 

源代码中涉及了slab的分配、释放等操作在后面分析中陆续总结。slab相关数据结构、工作机制以及整体框架在分析完了slab的创建、释放工作后再做总结,这样可能会对slab机制有更好的了解。当然,从代码中看运行机制会更有说服了,也是一种习惯

摘自 bullbat的专栏
(免责声明:文章内容如涉及作品内容、版权和其它问题,请及时与我们联系,我们将在第一时间删除内容,文章内容仅供参考)
收藏
  • 人气文章
  • 最新文章
  • 下载排行榜
  • 热门排行榜