Re: [PATCH -V18 04/13] vfs: Allow handle based open on symlinks

["Aneesh Kumar K. V" <aneesh.kumar@linux.vnet.ibm.com>]

On Mon, 23 Aug 2010 11:52:47 +1000, Neil Brown <neilb@suse.de> wrote:
> On Mon, 23 Aug 2010 06:54:03 +0530
> "Aneesh Kumar K. V" <aneesh.kumar@linux.vnet.ibm.com> wrote:
> 
> > On Mon, 23 Aug 2010 09:06:04 +1000, Neil Brown <neilb@suse.de> wrote:
> > > On Sat, 21 Aug 2010 01:13:52 -0600
> > > Andreas Dilger <andreas.dilger@oracle.com> wrote:
> > > 
> > > > On 2010-08-20, at 18:09, Neil Brown <neilb@suse.de> wrote:
> > > > > How about a new AT flag:  AT_FILE_HANDLE
> > > > > 
> > > > > Meaning is that the 'dirfd' is used only to identify a filesystem (vfsmnt) and
> > > > > the 'name' pointer actually points to a filehandle fragment interpreted in
> > > > > that filesystem.
> > > > > 
> > > > > One problem is that there is no way to pass the length...
> > > > > Options:
> > > > >   fragment is at most 64 bytes nul padded at the end
> > > > >   fragment is hex encoded and nul terminated
> > > > >   ??
> > > > > 
> > > > > I think I prefer the hex encoding, but I'm hoping someone else has a better
> > > > > idea.
> > > > 
> > > > That makes it ugly for the kernel to stringify and parse the file handles. 
> > > 
> > > We already parse filenames into components separated by '/'.  Is HEX decoding
> > > that much more ugly.
> > > 
> > > Filehandles are currently passed between the kernel and mountd as HEX
> > > strings, so at least there is some precedent.
> > > 
> > > > 
> > > > How about for AT_FILE_HANDLE THE FIRST __u32 (maybe with an extra __u32 for alignment) is the length and the rest of the binary file handle follows this?  In fact, doesn't the handle itself already encode the length in the header?
> > > 
> > > That part of a filehandle that nfsd gives to the filesystem is one byte out
> > > of a 4-byte header, plus the tail of the filehandle after the part that
> > > identifies the filesystem.
> > > This 'one byte' does imply the length, but it doesn't necessarily encode it.
> > > Rather it is a 'type'.  So it cannot really be used to determine the length
> > > at the point when the filehandle would need to be copied from userspace into
> > > the kernel.
> > > 
> > > 
> > > I don't think there is any precedent for passing a 4-byte length followed by
> > > a binary string, while there is plenty of precedent for passing a
> > > nul-terminated ASCII string.
> > > 
> > > [[ Following this approach I would like to avoid any filehandle-specific
> > > syscalls altogether.
> > > Just use a *at syscall with AT_FILE_HANDLE for filehandle lookup, and use
> > > getxattr('system:linux.file_handle') to get the filehandle for a given path.
> > > 
> > > Ofcourse we would need to at *at versions of the *xattr syscalls, but that is
> > > probably a good idea anyway.
> > > ]]
> > 
> > There are at* syscalls that doesn't take the additional flags as the
> > argument, like openat, readlinkat. How will handle based open and
> > readlink work with the above interface ?
> >
> 
> Bother... you are right.
> 
> I had remembered that at the time that all that *at calls were added there was
> discussion about how you always need some flags, particularly in the context
> of adding O_CLOEXEC and (I thought) a flag to allow non-sequential allocation
> of fds.
> I had thought that they all got 'flags' arguments as a result, but it seems
> not.
> For openat you could squeeze something into the current 'flags' arg
> (O_FILE_HANDLE), but for readlinkat, symlinkat at least there is no such
> option.  Sad really.
> 

IMHO that is really bad overloading of flags value. I also find  hex
encoded handle in pathname argument of *at syscalls confusing. The above
discussion also hint that we would need a new syscall for open,
readlink and setxattr, So how about me posting the new series which
remove open on symlink patch and add a bunch of syscalls to allow
operation on symlinks based on handle ?

-aneesh