Re: [PATCH 01/17] VFS: Implement the pioctl() system call

[David Howells <dhowells@redhat.com>]

Christoph Hellwig <hch@infradead.org> wrote:

> > Implement the pioctl() system call.  This is used to support a number of AFS
> > functions, and could also be used for Coda and other filesystems.
> 
> Umm, adding a new system call multiplexer without any structure is a
> serious no-go.  And this one is much worse than ioctl, which with a
> fixed [fd,cmd,arg] tuple seems like a stronhold of sanity compred to this
> monster with multiple arguments and a path that may or may not be there.

Ummm...  pioctl() has lots of structure.  Standard argument/reply block
definition, for example: you get one blob of argument, you may return one blob
of argument, you must structure your blobs such that 32-bit/64-bit
compatibility problems don't occur.  It's _much_ more structured than ioctl,
for example.

The main annoyance with it, as you noted, is the fact that people have treated
the path as being optional.

> I think you'd be better of writing tools that use a sane interface than
> adding a big pile of crap like this to the kernel.

Name a single sane interface that can do all that pioctl() can?  There isn't
one.  You can emulate almost all of pioctl() in userspace by a combination of
getxattr, lgetxattr, setxattr, lsetxattr, add_key, keyctl_read, and when all
else fails, open/open-NOFOLLOW + ioctl [IF not a dev file, and IF there are no
collisions between ioctl numbers and pioctl numbers].  In other words, a mess.

Now, assuming I do produce such a userspace library - that does not address
the other requirement: that of using a common set of binaries to manipulate
both OpenAFS and kAFS without the need for recompilation.  I presume you
advocate making OpenAFS change to suit your requirements?

David