inverse map 页框逆映射

inverse map页框逆映射

在页回收中，逆映射是一个有意思，并且很有技巧的一个功能．

逆映射:就是从给定的page来找到所有的引用者，特别的是PTE页表项．
有了逆映射就能安心释放指定page了．不然就会杯具了．

逆映射的场景：共享页框最常见的是共享库，Ｃ库基本上映射到所有的用户进程了．其次比如共享内存．

**kernel的演进：**
１. 2.4时代是遍历所有的进程要找到此page的引用者的PTE.简单粗暴．
效率上有点弱，功能上还是达到的．
2.　2.5时代改进了，牺牲了空间换效率，为每个页分配一个list来记录引用者．因为每个page都要增加了size空间，其实大部分页是不存在共享的．据说被抱怨．
3.　2.6又改进了，提出所谓的基于对象的方法．核心的思想就是要一个不用的low bit表示页类型，然后将一个指针地址理解成不同的对象，然后转换这个地址为对应的对象处理。这样就省下一个list大小。

**2种共享页，或者是２种对象**
１.　anonymous 匿名页，这个名字其实真取得不好．MAP_PRIVATE, stack * or brk vma (with NULL file) 
2.　file map　文件映射页．


**page/anon_vma/address_space**
struct page {
    
    union {
        atomic_t _mapcount;    /* Count of ptes mapped in mms,    
    };
    union {
        struct {
        struct address_space *mapping;    /* If low bit clear, points to
                         * inode address_space, or NULL.
                         * If page mapped as anonymous
                         * memory, low bit is set, and
                         * it points to anon_vma object:
                         * see PAGE_MAPPING_ANON below.
                         */
        };
 ｝

};

struct anon_vma {
    spinlock_t lock;    /* Serialize access to vma list */
    struct list_head head;    /* List of private "related" vmas */
};

struct vm_area_struct {
    struct mm_struct * vm_mm;    /* The address space we belong to. */
    
    /*
     * A file's MAP_PRIVATE vma can be in both i_mmap tree and anon_vma
     * list, after a COW of one of the file pages.    A MAP_SHARED vma
     * can only be in the i_mmap tree.  An anonymous MAP_PRIVATE, stack
     * or brk vma (with NULL file) can only be in an anon_vma list.
     */
    struct list_head anon_vma_node;    /* Serialized by anon_vma->lock */
    struct anon_vma *anon_vma;    /* Serialized by page_table_lock */

    
};


struct address_space {
    struct inode        *host;        /* owner: inode, block_device */    
    struct prio_tree_root    i_mmap;        /* tree of private and shared mappings */
    struct list_head    i_mmap_nonlinear;/*list VM_NONLINEAR mappings */
    
} ;

**页逆映射**
page->mapping[0]来区分是何种对象．
如果是anonymous对象，则mapping指向anon_vma这个对象，这个对象list所有的vma_area_struct.
vma_area_struct->vm_mm->PGD就能找到PTE了．

如果是file对象，则mapping指向file address space对象．这个address_space.i_mmap记录有所进程的共享映射信息vma_area_struct．这个也就能找到所有进程的PTE信息了． 特别要说的是i_mmap是棵优先查找树． over.

**func**
int try_to_unmap(struct page *page, int migration)
{
    int ret;

    BUG_ON(!PageLocked(page));

    if (PageAnon(page))
        ret = try_to_unmap_anon(page, migration);
    else
        ret = try_to_unmap_file(page, migration);

    if (!page_mapped(page))
        ret = SWAP_SUCCESS;
    return ret;
}

static __always_inline int PageAnon(struct page *page)
{
    page = compound_head(page);
    return ((unsigned long)page->mapping & PAGE_MAPPING_ANON) != 0;
}

try_to_unmap_one()释放一页
{
address = vma_address(page, vma);算出虚拟地址
    if (address == -EFAULT)
        goto out;

    pte = page_check_address(page, mm, address, &ptl);由虚拟地址算出pte
    if (!pte)
        goto out;

    ．．．后续的页表操作　，　略
}