background

Read the fucking source code! --By Lu Xun
A picture is worth a thousand words. --By Gorky

Description:

Kernel Version: 4.14
ARM64 processor, Contex-A53, binuclear
Use tools: Source Insight 3.5, Visio

1 Overview

In the previous series, to analyze the Buddy Systempage frame allocation, Slub分配器the small memory object allocation, these addresses are assigned contiguous physical memory. When the memory fragmentation, contiguous physical memory allocation becomes difficult, you can use the vmapmechanism will not contiguous physical memory page frames mapped into contiguous virtual address space. vmallocThe allocation is based on this mechanism to achieve.

Remember this picture below right?

vmap/vmallocThe area is in VMALLOC_START ~ VMALLOC_ENDbetween.

Adventure open it.

2. Data Structure

2.1 vmap_area/vm_struct

The two data structure is relatively simple, direct on the code:

struct vm_struct {
    struct vm_struct    *next;
    void            *addr;
    unsigned long       size;
    unsigned long       flags;
    struct page     **pages;
    unsigned int        nr_pages;
    phys_addr_t     phys_addr;
    const void      *caller;
};

struct vmap_area {
    unsigned long va_start;
    unsigned long va_end;
    unsigned long flags;
    struct rb_node rb_node;         /* address sorted rbtree */
    struct list_head list;          /* address sorted list */
    struct llist_node purge_list;    /* "lazy purge" list */
    struct vm_struct *vm;
    struct rcu_head rcu_head;
};

struct vmap_areaA region that describes a virtual address from the structural body va_start/va_endcan also be seen. Meanwhile, the structure experience by rb_nodehanging on a red-black tree, by listhanging on the list.
struct vmap_areaThe vmfield is a struct vm_structstructure for managing the mapping between the virtual and physical address page, you can struct vm_structform a linked list, maintain multi-stage mapping.

Relationship as shown below:

2.2 red-black tree

Red-black tree, is essentially a binary search tree, it adds coloring properties related to the basis of the binary search tree, red-black trees to enhance the search, insert, delete efficiency. In the red-black tree, the node is already sorted, for the elements of each node are left before the node, after the elements are right node.
Red-black tree must satisfy the following four rules:

Each node is not red is black;
Red-black tree roots must be black;
Red node child node must be black;
Each path from the node to the child nodes contain the same number of black nodes, while the node count the number of black, black nodes can be considered a null pointer;

It is defined as follows:

struct rb_node {
    unsigned long  __rb_parent_color;
    struct rb_node *rb_right;
    struct rb_node *rb_left;
} __attribute__((aligned(sizeof(long))));
    /* The alignment might seem pointless, but allegedly CRIS needs it */

由于内核会频繁的进行vmap_area的查找，红黑树的引入就是为了解决当查找数量非常多时效率低下的问题，在红黑树中，搜索元素，插入，删除等操作，都会变得非常高效。至于红黑树的算法操作，本文就不再深入分析，知道它的用途即可。

3. vmap/vunmap分析

3.1 vmap

vmap函数，完成的工作是，在vmalloc虚拟地址空间中找到一个空闲区域，然后将page页面数组对应的物理内存映射到该区域，最终返回映射的虚拟起始地址。

整体流程如下：

操作流程比较简单，来一个样例分析，就清晰明了了：

vmap调用中，关键函数为alloc_vmap_area，它先通过vmap_area_root二叉树来查找第一个区域first vm_area，然后根据这个first vm_area去查找vmap_area_list链表中满足大小的空间区域。

在alloc_vmap_area函数中，有几个全局的变量：

static struct rb_node *free_vmap_cache;
static unsigned long cached_hole_size;
static unsigned long cached_vstart;
static unsigned long cached_align;

用于缓存上一次分配成功的vmap_area，其中cached_hole_size用于记录缓存vmap_area对应区域之前的空洞的大小。缓存机制当然也是为了提高分配的效率。

3.2 vunmap

vunmap执行的是跟vmap相反的过程：从vmap_area_root/vmap_area_list中查找vmap_area区域，取消页表映射，再从vmap_area_root/vmap_area_list中删除掉vmap_area，页面返还给伙伴系统等。由于映射关系有改动，因此还需要进行TLB的刷新，频繁的TLB刷新会降低性能，因此将其延迟进行处理，因此称为lazy tlb。

来看看逆过程的流程：

4. vmalloc/vfree分析

4.1 vmalloc

vmalloc用于分配一个大的连续虚拟地址空间，该空间在物理上不连续的，因此也就不能用作DMA缓冲区。vmalloc分配的线性地址区域，在文章开头的图片中也描述了：VMALLOC_START ~ VMALLOC_END。

直接分析调用流程：

从过程中可以看出，vmalloc和vmap的操作，大部分的逻辑操作是一样的，比如从VMALLOC_START ~ VMALLOC_END区域之间查找并分配vmap_area，比如对虚拟地址和物理页框进行映射关系的建立。不同之处，在于vmap建立映射时，page是函数传入进来的，而vmalloc是通过调用alloc_page接口向Buddy System申请分配的。

vmalloc VS kmalloc
到现在，我们应该能清楚vmalloc和kmalloc的差异了吧，kmalloc会根据申请的大小来选择基于slub分配器或者基于Buddy System来申请连续的物理内存。而vmalloc则是通过alloc_page申请order = 0的页面，再映射到连续的虚拟空间中，物理地址不连续，此外vmalloc可以休眠，不应在中断处理程序中使用。
与vmalloc相比，kmalloc使用ZONE_DMA和ZONE_NORMAL空间，性能更快，缺点是连续物理内存空间的分配容易带来碎片问题，让碎片的管理变得困难。

4.2 vfree

直接上代码：

void vfree(const void *addr)
{
    BUG_ON(in_nmi());

    kmemleak_free(addr);

    if (!addr)
        return;
    if (unlikely(in_interrupt()))
        __vfree_deferred(addr);
    else
        __vunmap(addr, 1);
}

如果在中断上下文中，则推迟释放，否则直接调用__vunmap，所以它的逻辑基本和vunmap一致，不再赘述了。

[Original] (xii) Linux memory management of vmap and vmalloc