Snapshots

[Pradeep P. <pp...@ee...>]

Hi,
   The final piece in the puzzle is in. I have checked in the basic 
framework for doing snapshots. I haven't yet decided on the user 
intreface (may be .snapshots like NetAPP does), but you can access the 
snapshots using lfsread program.
   I want to explain a little bit about the design.
   I have toyed around with two different ways of implementing 
snapshots, theoritically actually same. To support any snapshots, one 
has to make a copy of the inode map (in LFS case IFILE) and fs 
modifications are done to the new IFILE only. Also, writes to existing 
blocks (both inode and data) should be re-directed to new blocks (this 
is almost trivial in LFS).
    Now, how do we make a copy of inode map? We can simply make a new 
IFILE file, copy the contents and start using the new file or We can 
divide the IFILE into snapshots. For example, the snapshot index (snapi) 
can be used to encode within the inode number (say first two bits). So, 
the operations are roughly as follows

1. Implementation using IFILE copy
    *) the on-disk/in-memory inode structure has snapi.
    *) user calls LFS_SNAP_CREATE ioctl
    *) make a copy of IFILE, ++snapi, set the new IFILE inode pointers
    *) when a file is written its snapi is compared to current snapi, if
       they differ new block is created.
    *) to access the snapshot just switch the IFILE. Temporarily
       switching IFILE back and forth is not that trivial though.

2. Implementation using inode encoding
    *) For each inode, inode number is derived with a simple function
       like below

	static inline ino_t INUMBER(struct super_block *sb, ino_t ino)
	{   	__u32 snapi = CURR_SNAPI(sb);

     		snapi = snapi << 30;
     		return ino | snapi;
	}

     *) user calls LFS_SNAP_CREATE ioctl
     *) make a copy of existing inodes to correponding new inodes. For
        example inode 1 is copied to 0x4001
     *) switch the root inode to the new inode.
     *) here comes the tricky part, the directory entries still contain
        old inodes. so a readdir always ORs the inode number with current
        snapi.
     *) to access the snapshot, just switch to the old root inode

I have actually switched back and forth, but finally decided to stick 
with the first implementation as the dentry contents magic is not 
required. It actually doesn't matter other than at the level of coding 
complexity/clarity. I also kept a local copy of the second 
implementation in case I need it.

Comments are welcome.

lfsprogs/lfssnap.c can be used to make snapshot. lfsprogs/lfsread can be 
used to read the contents of inodes etc. and see the snapshot.

-- 
Pradeep Padala
http://ppadala.blogspot.com