共享内存的实现---Linux内核笔记

系统调用mmap()通过映射一个普通文件实现共享内存。系统V则是通过映射特殊文件系统shm中的文件实现进程间的共享内存通信。也就是说，每个共享内存区域对应特殊文件系统shm中的一个文件（这是通过shmid_kernel结构联系起来的）。

struct shmid_kernel /* private to the kernel */
{ 
 struct kern_ipc_perm shm_perm;
 struct file *  shm_file;                          //关键是通过它来达到共享。
 int   id;
 unsigned long  shm_nattch;
 unsigned long  shm_segsz;
 time_t   shm_atim;
 time_t   shm_dtim;
 time_t   shm_ctim;
 pid_t   shm_cprid;
 pid_t   shm_lprid;
};

sys_shmget创建一个shm对象，保存在全局数组shm_ids中，同进创建一个shm文件。
sys_shmat通过调用do_mmap在当前进程中创建一段线性地址vm_area

struct vm_area_struct {
 struct mm_struct * vm_mm; /* The address space we belong to. */
 unsigned long vm_start; 线性地址的开始地址
 unsigned long vm_end;  线性地址的结束地址

......

struct file *vm_file;  与其关联的文件（对共享内存来说，对应shmid_kernel->shm_file）

}

 

比如进程A调用shmat(1, NULL, 0, NULL)映射的共享内存线性地址为vm_A, 进程B调用shmat(1, NULL, 0, NULL)映射的共享内存线性地址为vm_B, 当进程A第一次访问VM_A中的地址是，会产生一个缺页中断(因为物理地址还没映射)，内核会调用do_no_page为其映射一页物理地址，下面我们看看进程B是如何把它映射到同一页物理地址的：

static int do_no_page(struct mm_struct * mm, struct vm_area_struct * vma,
 unsigned long address, int write_access, pte_t *page_table)
{
 struct page * new_page;
 pte_t entry;

 if (!vma->vm_ops || !vma->vm_ops->nopage)
  return do_anonymous_page(mm, vma, page_table, write_access, address);
 spin_unlock(&mm->page_table_lock);

 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, 0);  // 映射一页物理地址，
....................................

}

 static int shmem_getpage(struct inode *inode, unsigned long idx, struct page **pagep, enum sgp_type sgp)
{
 struct address_space *mapping = inode->i_mapping;   //shm_file对应的地址空间
 struct shmem_inode_info *info = SHMEM_I(inode);
 struct shmem_sb_info *sbinfo;
 struct page *filepage = *pagep;
 struct page *swappage;
 swp_entry_t *entry;
 swp_entry_t swap;
 int error = 0;

 if (idx >= SHMEM_MAX_INDEX) {
  error = -EFBIG;
  goto failed;
 }

 /*
  * Normally, filepage is NULL on entry, and either found
  * uptodate immediately, or allocated and zeroed, or read
  * in under swappage, which is then assigned to filepage.
  * But shmem_readpage and shmem_prepare_write pass in a locked
  * filepage, which may be found not uptodate by other callers
  * too, and may need to be copied from the swappage read in.
  */
repeat:
 if (!filepage)
  filepage = find_lock_page(mapping, idx);    //通过vm中页号查找file->i_node->mapping是否有同样的页
 if (filepage && Page_Uptodate(filepage))     //找到就返回，没有则映射一个新页
  goto done;

.............................

}

mapping的定义:

 struct address_space {
 struct list_head clean_pages;  /* list of clean pages */
 struct list_head dirty_pages;     /* list of dirty pages */
 struct list_head locked_pages; /* list of locked pages */
 unsigned long  nrpages; /* number of total pages */
 struct address_space_operations *a_ops; /* methods */
 struct inode  *host;  /* owner: inode, block_device */
 struct vm_area_struct *i_mmap; /* list of private mappings */
 struct vm_area_struct *i_mmap_shared; /* list of shared mappings */
 spinlock_t  i_shared_lock;  /* and spinlock protecting it */
 int   gfp_mask; /* how to allocate the pages */
};

clean_pages、dirty_pages、locked_pages是当前地址空间已映射的page，find_lock_page(mapping, idx)通过查找这些page确认是否已映射物理内存，查找是通过page->index与idx的对应关系来的。

typedef struct page {
 struct list_head list;  /* ->mapping has some page lists. */
 struct address_space *mapping;  /* 当前页属于那个地址空间 */
 unsigned long index;  /* 当前页在地址空间的相对位置. */
 struct page *next_hash;  /* Next page sharing our hash bucket in
        the pagecache hash table. */
 atomic_t count;   /* Usage count, see below. */
 unsigned long flags;  /* atomic flags, some possibly
        updated asynchronously */
 struct list_head lru;  /* Pageout list, eg. active_list;
        protected by pagemap_lru_lock !! */
 struct page **pprev_hash; /* Complement to *next_hash. */
 struct buffer_head * buffers; /* Buffer maps us to a disk block. */