图解slub
slub确实是一个复杂的东西，究竟如何能把这个东西讲明白也是很头大的事情。今天我尝试用图示来解释一下slub的运作过程。或许会对你的理解有点帮助。
既然是内存管理难免离不开创建、分配、回收。那我就从这几个方面展示一下slub的动态情况。
创建
所谓创建就是创建kmem_cache这个结构体，可以认为这个结构体就是slub分配时的一个中心节点。你可以在这个节点上获得所有这个slub的信息。
比如在上一篇中看到的object_size,size,oo等。这些数据在创建后就不变了，主要用来存储本slub的大小信息，所以就不在这篇文章中列出。
在这篇文章中我们主要关注的是
* cpu_slab
* node可以看到，在slub中又分了两级来管理内存：per cpu层级和NUMA node层级。可以认为这是分层思想的又一次运用。
下图中显示了kmem_cache刚创建时的样子，可以看到什么都是空的，一切都刚刚开始。
kmem_cache
+------------------------------+
|name |
| (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| |tid | #cpu
| | (unsigned long) |
| |freelist (void **) | NULL
| |page (struct page *) | NULL
| |partial (struct page *) | NULL
+---+--------------------------+
|node[MAX_NUMNODES] |
| (struct kmem_cache_node*) |
| +--------------------------+
| |list_lock |
| | (spinlock_t) |
| |nr_partial | 0
| | (unsigned long) |
| |partial | empty
| | (struct list_head) |
+---+--------------------------+分配
创建完成后就是要分配了。由于采用了分层的设计，分配时根据每个层次是否存在内存分成了几种情况。
在系统运行过程中，按照一下顺序查看是否拥有可用内存。
* cpu_slab->freelist
* cpu_slab->partial
* node->partial原理也很简单，就是先看最方便拿到的空间中是否有内存。
接下来我就按照不同的情况，给大家看看slub运作时的动态。
哪都没内存
按照以上的顺序，这个过程应该是以上三个空间中都没有内存的情况。不过这也是slub刚创建后，第一次分配时的情况。所以我干脆就把这个当作第一种情形了。
做的工作也很简单，就是从buddy系统中拿到一个page，装到cpu_slab->freelist中。这一步在allocate_slab()中执行。
这种情形的初始状态是大家都为空，但是下图中我仍然画了两个状态。目的是为了突出page->freelist指针的变化。
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist |
| | (void **) |
| | |
| |page +--|--->+--------+--------+--------+--------+
| | (struct page *) | | | | | | |
| | +-------------------|--+ +--------+--------+--------+--------+
| | |freelist ---+ |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| |partial | = NULL
| | (struct page *) |
+---+--------------------------+
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist ------|----+ = page->freelist
| | (void **) | |
| | | v
| |page | +--------+--------+--------+--------+
| | (struct page *) | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| |partial | = NULL
| | (struct page *) |
+---+--------------------------+freelist上有内存
这种情形也是被称之为Fast path的情形，因为这中情况下，只要移动freelist指针即可。这步在slab_alloc_node()中执行。
在下图中，freelist就从a移动到了A。
需要注意的是freelist中保存的是一个链表而不是线性表，所以指针的变化不一定是后移。
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | | a A
| |freelist ------|-------------+..................
| | (void **) | | .
| | | v v
| |page | +--------+--------+--------+--------+
| | (struct page *) | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| | |
| |partial | = NULL
| | (struct page *) |
+---+--------------------------+partial有内存
以此类推，接下来的情形就是partial上有内存。这时，我们就需要从partial上找到可用的page，装到freelist上。这步在___slab_alloc() -> get_freelist()执行。
下面两个图分别显示了装之前和之后的变化，可以看到原来在partial中的page A，安装后到了cpu_slub->page。
值得注意的是在partial上的page，frozen也是1，但是inuse则反映真实的分配情况而不是等于objects了。
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist | = NULL
| | (void **) |
| |page | = NULL
| | (struct page *) |
| | |
| |partial ------|--->+-----------+ +-----------+
| | (struct page *) | |A next|---->|B next|
| | | +-----------+ +-----------+
| | | |pobjectsA | = |pobjectsB | + objectsA
| | | | | | |
| | | |inuse < | |inuse < |
| | | | objectsA | | objectsB |
| | | |frozen = 1 | |frozen = 1 |
| | | +-----------+ +-----------+
+---+--------------------------+
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist ------|----+ = page->freelist
| | (void **) | |
| | | v
| |page | +--------+--------+--------+--------+
| | (struct page *) A | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| | |
| |partial ------|--->+-----------+
| | (struct page *) | |B next|
| | | +-----------+
| | | |pobjectsB | = objectsB
| | | | |
| | | |inuse < |
| | | | objectsB |
| | | |frozen = 1 |
| | | +-----------+
+---+--------------------------+node上有内存
看完了freelist/partial，最后找到在cpu_slub->node上有内存。那只能从node上把page拿下来，装到freelist上。这个过程在new_slab_objects()->get_partial()->get_partial_node()中执行。
在这里需要注意，在某些情况下，不仅会装到freelist上，还会多拿几个page装到partial上。节省了下次又找不到的情况。在下图中显示的page B就从node链表装到了partial上。而且在node上的page中frozen为0。
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist | = NULL
| | (void **) |
| |page | = NULL
| | (struct page *) |
| | |
| |partial | = NULL
| | (struct page *) |
| | |
+---+--------------------------+
|node[MAX_NUMNODES] |
| (struct kmem_cache_node*) |
| +--------------------------+
| |nr_partial = 3 |
| | (unsigned long) |
| |partial ----|--->+-----------+ +-----------+ +-----------+
| | (struct list_head) | |A lru|---->|B lru|---->| lru|
+---+--------------------------+ +-----------+ +-----------+ +-----------+
|inuse < | |inuse < |
| objects | | objects |
|frozen = 0 | |frozen = 0 |
+-----------+ +-----------+
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist ------|----+ = page->freelist
| | (void **) | |
| | | v
| |page | +--------+--------+--------+--------+
| | (struct page *) A | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| | |
| |partial ------|--->+-----------+
| | (struct page *) | |B next|
| | | +-----------+
| | | |pobjects | = objects
| | | | |
| | | |inuse < |
| | | | objects |
| | | |frozen = 1 |
| | | +-----------+
| | |
+---+--------------------------+
|node[MAX_NUMNODES] |
| (struct kmem_cache_node*) |
| +--------------------------+
| |nr_partial = 1 |
| | (unsigned long) |
| |partial ----|--->+-----------+
| | (struct list_head) | |C lru|
+---+--------------------------+ +-----------+回收
接下来看看回收。回收总的来说就是分配的另一面，但是在情形上和分配略有不同。
因为slub的设计是一个链表，所以不论需要释放内存对应的页是在哪个地方，都是在freelist链表头上插入一个对象。所以单个对象的回收没有太大区别。区别在于释放对象后要把对应的页搬移到哪个地方。
也就如何在下面三个地方之间移动页。
* cpu_slab->freelist
* cpu_slab->partial
* node->partial移动freelist
这个情况很简单，并不涉及到页的移动，而只要移动cpu_slab->freelist就好了。
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | | obj
| |freelist ------|-------------+..................
| | (void **) | | .
| | | v v
| |page | +--------+--------+--------+--------+
| | (struct page *) | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| | |
| |partial | = NULL
| | (struct page *) |
+---+--------------------------+替换cpu_slab->page
这个其实是发生在分配时候的情形，当cpu_slab->page不满足分配需求时，我们要把这个页面退回到node->partial上。
这个过程有deactivate_slab()来实现。
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist ------|----+ = page->freelist
| | (void **) | |
| | | v
| |page | +--------+--------+--------+--------+
| | (struct page *) A | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| | |
| |partial ------|--->+-----------+
| | (struct page *) | |B next|
| | | +-----------+
| | | |pobjects | = objects
| | | | |
| | | |inuse < |
| | | | objects |
| | | |frozen = 1 |
| | | +-----------+
| | |
+---+--------------------------+
|node[MAX_NUMNODES] |
| (struct kmem_cache_node*) |
| +--------------------------+
| |nr_partial = 1 |
| | (unsigned long) |
| |partial ----|--->+-----------+
| | (struct list_head) | |C lru|
+---+--------------------------+ +-----------+
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist | = NULL
| | (void **) |
| |page | = NULL
| | (struct page *) |
| | |
| |partial ------|--->+-----------+
| | (struct page *) | |B next|
| | | +-----------+
| | | |pobjects | = objects
| | | | |
| | | |inuse < |
| | | | objects |
| | | |frozen = 1 |
| | | +-----------+
| | |
+---+--------------------------+
|node[MAX_NUMNODES] |
| (struct kmem_cache_node*) |
| +--------------------------+
| |nr_partial = 1 |
| | (unsigned long) |
| |partial ----|--->+-----------+ +-----------+
| | (struct list_head) | |A lru|---->|C lru|
+---+--------------------------+ +-----------+ +-----------+
|inuse < |
| objects |
|frozen = 0 |
+-----------+解冻cpu_slab->partial
当检测到cpu_slab->partial上的对象足够多时，就会将链表上的页面都退换给node->partial。值得注意的是这个过程会把cpu_slab->partial清空。
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist ------|----+ = page->freelist
| | (void **) | |
| | | v
| |page | +--------+--------+--------+--------+
| | (struct page *) A | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| | |
| |partial ------|--->+-----------+ +-----------+
| | (struct page *) | |B next|---->|C next|
| | | +-----------+ +-----------+
| | | |pobjectsA | = |pobjectsB | + objectsA
| | | | | | |
| | | |inuse < | |inuse < |
| | | | objectsA | | objectsB |
| | | |frozen = 1 | |frozen = 1 |
| | | +-----------+ +-----------+
| | |
+---+--------------------------+
|node[MAX_NUMNODES] |
| (struct kmem_cache_node*) |
| +--------------------------+
| |nr_partial = 1 |
| | (unsigned long) |
| |partial ----|--->+-----------+
| | (struct list_head) | |D lru|
+---+--------------------------+ +-----------+
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist ------|----+ = page->freelist
| | (void **) | |
| | | v
| |page | +--------+--------+--------+--------+
| | (struct page *) A | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| | |
| |partial | = NULL
| | (struct page *) |
| | |
+---+--------------------------+
|node[MAX_NUMNODES] |
| (struct kmem_cache_node*) |
| +--------------------------+
| |nr_partial = 1 |
| | (unsigned long) |
| |partial ----|--->+-----------+ +-----------+ +-----------+
| | (struct list_head) | |B lru|---->|C lru|---->|D lru|
+---+--------------------------+ +-----------+ +-----------+ +-----------+
|inuse < | |inuse < |
| objects | | objects |
|frozen = 0 | |frozen = 0 |
+-----------+ +-----------+从node->partial上取出一个加入cpu_slab->partial
在某些情况下，我们也会将node->partial链表中的一个页放到cpu_slab->partial上。
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist ------|----+ = page->freelist
| | (void **) | |
| | | v
| |page | +--------+--------+--------+--------+
| | (struct page *) A | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| | |
| |partial ------|--->+-----------+
| | (struct page *) | |B next|
| | | +-----------+
| | | |pobjectsB | = objectsB
| | | | |
| | | |inuse < |
| | | | objectsB |
| | | |frozen = 1 |
| | | +-----------+
| | |
+---+--------------------------+
|node[MAX_NUMNODES] |
| (struct kmem_cache_node*) |
| +--------------------------+
| |nr_partial = 3 |
| | (unsigned long) |
| |partial ----|--->+-----------+ +-----------+
| | (struct list_head) | |C lru|---->|D lru|
+---+--------------------------+ +-----------+ +-----------+
|inuse < |
| objects |
|frozen = 0 |
+-----------+
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist ------|----+ = page->freelist
| | (void **) | |
| | | v
| |page | +--------+--------+--------+--------+
| | (struct page *) A | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| | |
| |partial ------|--->+-----------+ +-----------+
| | (struct page *) | |C next|---->|B next|
| | | +-----------+ +-----------+
| | | |pobjectsC | = |pobjectsB | + objectsC
| | | | | | |
| | | |inuse < | |inuse < |
| | | | objectsC | | objectsB |
| | | |frozen = 1 | |frozen = 1 |
| | | +-----------+ +-----------+
| | |
+---+--------------------------+
|node[MAX_NUMNODES] |
| (struct kmem_cache_node*) |
| +--------------------------+
| |nr_partial = 1 |
| | (unsigned long) |
| |partial ----|--->+-----------+
| | (struct list_head) | |C lru|
+---+--------------------------+ +-----------+将一个满页加到node->partial上
当一个页上的所有空间都被分配，这个页将不存在任何地方。（除非配置了CONFIG_SLUB_DEBUG）。当释放满页上的一个对象时，这个页就会再次出现在node->partial上。
下图中为了清晰起见，我将配置了CONFIG_SLUB_DEBUG时的node->full链表头显示出。用来表示页的移动情况。
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist ------|----+ = page->freelist
| | (void **) | |
| | | v
| |page | +--------+--------+--------+--------+
| | (struct page *) A | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| | |
| |partial ------|--->+-----------+
| | (struct page *) | |B next|
| | | +-----------+
| | | |pobjectsB | = objectsB
| | | | |
| | | |inuse < |
| | | | objectsB |
| | | |frozen = 1 |
| | | +-----------+
| | |
+---+--------------------------+
|node[MAX_NUMNODES] |
| (struct kmem_cache_node*) |
| +--------------------------+
| |nr_partial = 3 |
| | (unsigned long) |
| |partial ----|--->+-----------+ +-----------+
| | (struct list_head) | |C lru|---->|D lru|
| +--------------------------+ +-----------+ +-----------+
| |full ----|--->+-----------+
| | (struct list_head) | |E lru|
+---+--------------------------+ +-----------+
kmem_cache
+------------------------------+
|name (char *) |
+------------------------------+
|cpu_slab | * per cpu variable pointer
| (struct kmem_cache_cpu*) |
| +--------------------------+
| |stat[NR_SLUB_STAT_ITEMS] |
| | (unsigned) |
| +--------------------------+
| |tid (unsigned long) |
| | |
| |freelist ------|----+ = page->freelist
| | (void **) | |
| | | v
| |page | +--------+--------+--------+--------+
| | (struct page *) A | | | | | |
| | +----------------------+ +--------+--------+--------+--------+
| | |freelist = NULL |
| | |inuse = objects |
| | |frozen = 1 |
| | +----------------------+
| | |
| |partial ------|--->+-----------+
| | (struct page *) | |B next|
| | | +-----------+
| | | |pobjectsB | = objectsB
| | | | |
| | | |inuse < |
| | | | objectsB |
| | | |frozen = 1 |
| | | +-----------+
| | |
+---+--------------------------+
|node[MAX_NUMNODES] |
| (struct kmem_cache_node*) |
| +--------------------------+
| |nr_partial = 3 |
| | (unsigned long) |
| |partial ----|--->+-----------+ +-----------+ +-----------+
| | (struct list_head) | |E lru|---->|C lru|---->|D lru|
| +--------------------------+ +-----------+ +-----------+ +-----------+
| |full | = NULL
| | (struct list_head) |
+---+--------------------------+Last updated
Was this helpful?