Re: [PATCH 0/2] Fix hardlinks in overlay

[Alexander Morozov]

Sequence of calls is kern_path(filename, LOOKUP_FOLLOW, &path) ->
filename_lookup(AT_FDCWD, getname_kernel(name), flags, path, NULL) ->
path_lookup(nameidata *nd, flags | LOOKUP_RCU, path) -> path_init(nd,
flags). In that function if I understand correctly nd->path is set to
current_thread_info()->task->fs->pwd and I don't see any further
tricks with dentry, so I suppose that
current_thread_info()->task->fs->pwd->dentry already points to
overlayfs dentry.
Let me know what we can do with it.

Thanks!
- Alex

On Mon, Mar 7, 2016 at 2:49 AM, Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
> On Fri, Oct 9, 2015 at 9:34 PM, Alexander Morozov
> <alexandr.morozov@xxxxxxxxxx> wrote:
>> There were reports that overlay doesn't work very well with unix-sockets.
>> In particular you can't access unix-socket through hardlink on overlay fs.
>> Problem is that overlay creates different inodes for hardlinks and code in
>> net/unix/af_unix.c relies on inodes for unix-socket lookup. I think this
>> affects any code which relies on inodes from kern_path. There is helper
>> d_backing_inode, which I think supposed to get inodes from underlying fs
>> (for example ext4), but in current implementation it does nothing.  These
>> patches made on top of v4.3-rc4 of main linux tree (master is broken for my
>> ubuntu VM), but I tested that they applying on master and there was no
>> changes to overlay since v4.3-rc4.
>>
>> Alexander Morozov (2):
>>   fs/overlay: move update and instantiate dentry code to function
>>   fs/overlay: use same inodes for hardlinks
>
> Problem is with lookup: how do we go from backing dentry to overlayfs
> dentry so the inode can be shared?
>
> Thanks,
> Miklos