剖析zfs的写磁盘的操作

很少有和zfs相关的代码分析，中文的材料基本没有，我来剖析一下zfs的写磁盘的底层的实现吧。

在zio 的流水线当中，有一个流水线的名称为vdev_disk_io_start。

在vdev_disk_io_start（zio *zio）函数中，做了如下的事情：

1.判断zio的请求类型(zio->io_type)

2.根据类型，将zio 传递给__vdev_disk_physio(vd->vd_bdev, zio, zio->io_data,

                   zio->io_size, zio->io_offset, flags);

在函数 __vdev_disk_physio函数当中，涉及到dio_request 结构体，为了submit_bio的时候使用的对单个请求的包装。

dio_request 请求如下

/*
 * Virtual device vector for disks.
 */
typedef struct dio_request {
    struct completion    dr_comp;    /* Completion for sync IO */
    atomic_t        dr_ref;        /* References */
    zio_t            *dr_zio;    /* Parent ZIO */
    int            dr_rw;        /* Read/Write */
    int            dr_error;    /* Bio error */
    int            dr_bio_count;    /* Count of bio's */
        struct bio        *dr_bio[0];    /* Attached bio's */
} dio_request_t;

可以看到dr_bio[]变量为可变的。并且在这个函数当中，给bio进行赋值

对于zio传下来的参数，可以告诉vdev层，请求的io偏移量，请求的buf指针。以及请求的类型

error = __vdev_disk_physio(vd->vd_bdev, zio, zio->io_data,
zio->io_size, zio->io_offset, flags);

static int
__vdev_disk_physio(struct block_device *bdev, zio_t *zio, caddr_t kbuf_ptr,
                   size_t kbuf_size, uint64_t kbuf_offset, int flags)
{
        dio_request_t *dr;
    caddr_t bio_ptr;
    uint64_t bio_offset;
    int bio_size, bio_count = 16;
    int i = 0, error = 0;

    ASSERT3U(kbuf_offset + kbuf_size, <=, bdev->bd_inode->i_size);

retry:
    dr = vdev_disk_dio_alloc(bio_count);
    if (dr == NULL)
        return ENOMEM;
...
    dr->dr_zio = zio;
    dr->dr_rw = flags;
/*bio_ptr为真正的存放数据的内存指针*/
    bio_ptr    = kbuf_ptr;
    bio_offset = kbuf_offset;
    bio_size   = kbuf_size;
    for (i = 0; i <= dr->dr_bio_count; i++) {
...
if (bio_size <= 0)
break;
        dr->dr_bio[i] = bio_alloc(GFP_NOIO,
                                  bio_nr_pages(bio_ptr, bio_size));
        if (dr->dr_bio[i] == NULL) {
            vdev_disk_dio_free(dr);
            return ENOMEM;
        }

        /* Matching put called by vdev_disk_physio_completion */
        vdev_disk_dio_get(dr);

        dr->dr_bio[i]->bi_bdev = bdev;
        dr->dr_bio[i]->bi_sector = bio_offset >> 9;  //根据提供的偏移量，缩小为以扇区为单位。
        dr->dr_bio[i]->bi_rw = dr->dr_rw;
        dr->dr_bio[i]->bi_end_io = vdev_disk_physio_completion;
        dr->dr_bio[i]->bi_private = dr;

        /* Remaining size is returned to become the new size */
        bio_size = bio_map(dr->dr_bio[i], bio_ptr, bio_size); 在这里给每个bio的bi_vect变量赋值。bio_size 如果小于0，那么将跳出循环。传进bio_ptr,将数据以页的形式加入到bio当中。

        /* Advance in buffer and construct another bio if needed */
        bio_ptr    += dr->dr_bio[i]->bi_size;
        bio_offset += dr->dr_bio[i]->bi_size;
    }

    /* Extra reference to protect dio_request during submit_bio */
    vdev_disk_dio_get(dr);
    if (zio)
        zio->io_delay = jiffies_64;

    /* Submit all bio's associated with this dio */
    for (i = 0; i < dr->dr_bio_count; i++)
        if (dr->dr_bio[i])
            submit_bio(dr->dr_rw, dr->dr_bio[i]);//同一个request，bio的请求类型都是相同的，并将生成的bio提交给内核来处理。

    /*
     * On synchronous blocking requests we wait for all bio the completion
     * callbacks to run.  We will be woken when the last callback runs
     * for this dio.  We are responsible for putting the last dio_request
     * reference will in turn put back the last bio references.  The
     * only synchronous consumer is vdev_disk_read_rootlabel() all other
     * IO originating from vdev_disk_io_start() is asynchronous.
     */
    if (vdev_disk_dio_is_sync(dr)) {
        wait_for_completion(&dr->dr_comp);
        error = dr->dr_error;
        ASSERT3S(atomic_read(&dr->dr_ref), ==, 1);
    }

    (void)vdev_disk_dio_put(dr);

    return error;
}

到这里，所有的处理工作都已经结束了。然后该执行zio的其他的流水线的函数了。