On Mon, 2010-12-06 at 09:27 -0500, Josef Bacik wrote: > On Sat, Dec 04, 2010 at 01:58:07PM -0800, Mike Fedyk wrote: > > On Fri, Dec 3, 2010 at 1:45 PM, Josef Bacik <josef@xxxxxxxxxx> wrote: > > > On Wed, Dec 01, 2010 at 09:21:36AM -0500, Josef Bacik wrote: > > >> Hello, > > >> > > >> Various people have complained about how BTRFS deals with subvolumes recently, > > >> specifically the fact that they all have the same inode number, and there's no > > >> discrete seperation from one subvolume to another. Christoph asked that I lay > > >> out a basic design document of how we want subvolumes to work so we can hash > > >> everything out now, fix what is broken, and then move forward with a design that > > >> everybody is more or less happy with. I apologize in advance for how freaking > > >> long this email is going to be. I assume that most people are generally > > >> familiar with how BTRFS works, so I'm not going to bother explaining in great > > >> detail some stuff. > > >> > > >> === What are subvolumes? === > > >> > > >> They are just another tree. In BTRFS we have various b-trees to describe the > > >> filesystem. A few of them are filesystem wide, such as the extent tree, chunk > > >> tree, root tree etc. The tree's that hold the actual filesystem data, that is > > >> inodes and such, are kept in their own b-tree. This is how subvolumes and > > >> snapshots appear on disk, they are simply new b-trees with all of the file data > > >> contained within them. > > >> > > >> === What do subvolumes look like? === > > >> > > >> All the user sees are directories. They act like any other directory acts, with > > >> a few exceptions > > >> > > >> 1) You cannot hardlink between subvolumes. This is because subvolumes have > > >> their own inode numbers and such, think of them as seperate mounts in this case, > > >> you cannot hardlink between two mounts because the link needs to point to the > > >> same on disk inode, which is impossible between two different filesystems. The > > >> same is true for subvolumes, they have their own trees with their own inodes and > > >> inode numbers, so it's impossible to hardlink between them. > > >> > > >> 1a) In case it wasn't clear from above, each subvolume has their own inode > > >> numbers, so you can have the same inode numbers used between two different > > >> subvolumes, since they are two different trees. > > >> > > >> 2) Obviously you can't just rm -rf subvolumes. Because they are roots there's > > >> extra metadata to keep track of them, so you have to use one of our ioctls to > > >> delete subvolumes/snapshots. > > >> > > >> But permissions and everything else they are the same. > > >> > > >> There is one tricky thing. When you create a subvolume, the directory inode > > >> that is created in the parent subvolume has the inode number of 256. So if you > > >> have a bunch of subvolumes in the same parent subvolume, you are going to have a > > >> bunch of directories with the inode number of 256. This is so when users cd > > >> into a subvolume we can know its a subvolume and do all the normal voodoo to > > >> start looking in the subvolumes tree instead of the parent subvolumes tree. > > >> > > >> This is where things go a bit sideways. We had serious problems with NFS, but > > >> thankfully NFS gives us a bunch of hooks to get around these problems. > > >> CIFS/Samba do not, so we will have problems there, not to mention any other > > >> userspace application that looks at inode numbers. > > >> > > >> === How do we want subvolumes to work from a user perspective? === > > >> > > >> 1) Users need to be able to create their own subvolumes. The permission > > >> semantics will be absolutely the same as creating directories, so I don't think > > >> this is too tricky. We want this because you can only take snapshots of > > >> subvolumes, and so it is important that users be able to create their own > > >> discrete snapshottable targets. > > >> > > >> 2) Users need to be able to snapshot their subvolumes. This is basically the > > >> same as #1, but it bears repeating. > > >> > > >> 3) Subvolumes shouldn't need to be specifically mounted. This is also > > >> important, we don't want users to have to go around mounting their subvolumes up > > >> manually one-by-one. Today users just cd into subvolumes and it works, just > > >> like cd'ing into a directory. > > >> > > >> === Quotas === > > >> > > >> This is a huge topic in and of itself, but Christoph mentioned wanting to have > > >> an idea of what we wanted to do with it, so I'm putting it here. There are > > >> really 2 things here > > >> > > >> 1) Limiting the size of subvolumes. This is really easy for us, just create a > > >> subvolume and at creation time set a maximum size it can grow to and not let it > > >> go farther than that. Nice, simple and straightforward. > > >> > > >> 2) Normal quotas, via the quota tools. This just comes down to how do we want > > >> to charge users, do we want to do it per subvolume, or per filesystem. My vote > > >> is per filesystem. Obviously this will make it tricky with snapshots, but I > > >> think if we're just charging the diff's between the original volume and the > > >> snapshot to the user then that will be the easiest for people to understand, > > >> rather than making a snapshot all of a sudden count the users currently used > > >> quota * 2. > > >> > > >> === What do we do? === > > >> > > >> This is where I expect to see the most discussion. Here is what I want to do > > >> > > >> 1) Scrap the 256 inode number thing. Instead we'll just put a flag in the inode > > >> to say "Hey, I'm a subvolume" and then we can do all of the appropriate magic > > >> that way. This unfortunately will be an incompatible format change, but the > > >> sooner we get this adressed the easier it will be in the long run. Obviously > > >> when I say format change I mean via the incompat bits we have, so old fs's won't > > >> be broken and such. > > >> > > >> 2) Do something like NFS's referral mounts when we cd into a subvolume. Now we > > >> just do dentry trickery, but that doesn't make the boundary between subvolumes > > >> clear, so it will confuse people (and samba) when they walk into a subvolume and > > >> all of a sudden the inode numbers are the same as in the directory behind them. > > >> With doing the referral mount thing, each subvolume appears to be its own mount > > >> and that way things like NFS and samba will work properly. > > >> > > >> I feel like I'm forgetting something here, hopefully somebody will point it out. > > >> > > >> === Conclusion === > > >> > > >> There are definitely some wonky things with subvolumes, but I don't think they > > >> are things that cannot be fixed now. Some of these changes will require > > >> incompat format changes, but it's either we fix it now, or later on down the > > >> road when BTRFS starts getting used in production really find out how many > > >> things our current scheme breaks and then have to do the changes then. Thanks, > > >> > > > > > > So now that I've actually looked at everything, it looks like the semantics are > > > all right for subvolumes > > > > > > 1) readdir - we return the root id in d_ino, which is unique across the fs > > > 2) stat - we return 256 for all subvolumes, because that is their inode number > > > 3) dev_t - we setup an anon super for all volumes, so they all get their own > > > dev_t, which is set properly for all of their children, see below > > > > > > [root@test1244 btrfs-test]# stat . > > > File: `.' > > > Size: 20 Blocks: 8 IO Block: 4096 directory > > > Device: 15h/21d Inode: 256 Links: 1 > > > Access: (0555/dr-xr-xr-x) Uid: ( 0/ root) Gid: ( 0/ root) > > > Access: 2010-12-03 15:35:41.931679393 -0500 > > > Modify: 2010-12-03 15:35:20.405679493 -0500 > > > Change: 2010-12-03 15:35:20.405679493 -0500 > > > > > > [root@test1244 btrfs-test]# stat foo > > > File: `foo' > > > Size: 12 Blocks: 0 IO Block: 4096 directory > > > Device: 19h/25d Inode: 256 Links: 1 > > > Access: (0700/drwx------) Uid: ( 0/ root) Gid: ( 0/ root) > > > Access: 2010-12-03 15:35:17.501679393 -0500 > > > Modify: 2010-12-03 15:35:59.150680051 -0500 > > > Change: 2010-12-03 15:35:59.150680051 -0500 > > > > > > [root@test1244 btrfs-test]# stat foo/foobar > > > File: `foo/foobar' > > > Size: 0 Blocks: 0 IO Block: 4096 regular empty file > > > Device: 19h/25d Inode: 257 Links: 1 > > > Access: (0644/-rw-r--r--) Uid: ( 0/ root) Gid: ( 0/ root) > > > Access: 2010-12-03 15:35:59.150680051 -0500 > > > Modify: 2010-12-03 15:35:59.150680051 -0500 > > > Change: 2010-12-03 15:35:59.150680051 -0500 > > > > > > So as far as the user is concerned, everything should come out right. Obviously > > > we had to do the NFS trickery still because as far as VFS is concerned the > > > subvolumes are all on the same mount. So the question is this (and really this > > > is directed at Christoph and Bruce and anybody else who may care), is this good > > > enough, or do we want to have a seperate vfsmount for each subvolume? Thanks, > > > > > > > What are the drawbacks of having a vfsmount for each subvolume? > > > > Why (besides having to code it up) are you trying to avoid doing it that way? > > It's the having to code it up that way thing, I'm nothing if not lazy. And, anything that uses the mount table, exposed from the kernel, will grind a system to a halt with only a few thousand mounts, not to mention that user space utilities, like df, du ..., will become painful to use for more than a hundred or so entries. > > Josef > -- > To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
