Re: Is ndo_do_ioctl still acceptable?

[Stephen Hemminger <stephen@networkplumber.org>]

On Thu, 12 Nov 2015 05:59:03 +0100
"Jason A. Donenfeld" <Jason@zx2c4.com> wrote:

> Hi David & Folks,
> 
> Soon I will submit a virtual tunnel device driver to LKML for review.
> It uses rtnl_link_register to create a virtual network interface,
> which then handles encryption, authentication, and some other things,
> amongst various configured peers.
> 
> Right now the device is configurable via Netlink. It receives new
> peers and configuration via a rtnl_link_ops->changelink function, and
> it reports information back to userspace via a
> rtnl_link_ops->fill_info function.
> 
> Configuration works fine, though it is rather cumbersome to do this
> all via Netlink.
> 
> Reporting information back to userspace does not work fine. The reason
> is that sometimes there's too much information to report back to
> userspace than what can fit in a single preallocated Netlink skb. And
> since rtnl_link_ops->fill_info doesn't receive any information from
> userspace, I'm unable to use it to send back information in smaller
> pieces.
> 
> I realize I could register a whole new rtnl packet family and related
> set of functions with rtnl_register, such as what's done at the bottom
> of `net/core/rtnetlink.c`. This is extremely cumbersome and invasive
> though. It would require adding a new protocol family (like the
> already existing rtnl_register-ified functions for PF_UNSPEC and
> PF_BRIDGE), and I don't have enough clout to confidently submit a
> patch that augments `include/linux/socket.h` with a new PF/AF define.
> This seems very invasive and not appropriate for my driver.
> 
> What I'd really like to do is just implement ndo_do_ioctl. It seems to
> me that this gives me a precise interface to do exactly what I want in
> the cleanest and easiest to read possible way. I could have differing
> ioctls for differing things. I could write memory back to userspace in
> proper chunks, with the proper size. It's clear and straightforward
> how to do it, and what the completed result looks like. It doesn't
> require invasive changes into other parts of the kernel, as this would
> be self-contained. It's hard to imagine a better interface to use than
> ndo_do_ioctl.
> 
> But. But the word on the street is that kernel hipsters hate ioctls
> and espouse the use of netlink everywhere with religious fervor, and
> will burn at the stake any submissions I might send that go anywhere
> near using ndo_do_ioctl rather than (the most likely ill-fitting for
> the task) netlink. That, and the maintainers of the `ip` tool will be
> upset too (even though they do already make use of several ioctls
> instead of netlink). I'm told everybody will leer and jeer at me if I
> use ndo_do_ioctl instead of netlink.
> 
> Except ndo_do_ioctl is *so* perfectly fitting here for my use case!
> 
> So what's the verdict on this? Do these aforementioned kernel hipsters
> not really matter so much, and ndo_do_ioctl is actually perfectly
> fine? Or must I really affix netlink onto my forthcoming submission?

The problem is ioctl's are device specific, and therefore create dependency
on the unique features supported by your device.
Also doing security validation on ioctl's is much harder.

The question always comes up, why is this new API not something general?
And if you are dumping such a huge mound of information that only your driver
could love, then why are you doing it? Is there anything in there that
really matters? 

If all you are really doing is dumping statistics then look at ethtool.
If you are dealing with lots of virtual function devices, look how existing
netlink info is trimmed.

Don't overanalyze this.