Re: Can we move device drivers into user-space?

[Jidong Xiao <jidong.xiao@gmail.com>]

On Fri, Feb 24, 2012 at 11:54 AM, Greg KH <gregkh@linuxfoundation.org> wrote:
> On Fri, Feb 24, 2012 at 11:38:54AM -0500, Jidong Xiao wrote:
>> On Fri, Feb 24, 2012 at 10:38 AM, Greg KH <greg@kroah.com> wrote:
>> > On Fri, Feb 24, 2012 at 10:19:36AM -0500, Jidong Xiao wrote:
>> >> On Wed, Feb 22, 2012 at 11:56 PM, Jidong Xiao <jidong.xiao@gmail.com> wrote:
>> >> > Hi,
>> >> >
>> >> > I am just curious. Since the concept user-space device drivers has
>> >> > been proposed for several years, and some related projects and
>> >> > research papers have demonstrated the feasibility of of moving device
>> >> > drivers into use space. In particular, this paper:
>> >> >
>> >> > Tolerating Malicious Device Drivers in Linux.
>> >> > http://pdos.csail.mit.edu/papers/sud:usenix10.pdf
>> >> >
>> >> > In this paper, existing device driver code need not to be changed,
>> >> > which should help the idea to be applied in practice.
>
> Please note, that one of the strengths of Linux is that we CAN change
> driver code, and we do, which makes implementations like this nice from
> an academic point of view, but unrealistic from a real-world point of
> view.
>
>> >> Actually, my major concern is, since UIO has been accepted, then why
>> >> don't we move all the rest device drivers into user space as well. As
>> >> I understand, currently, some of device drivers are running on user
>> >> space, while the other (or say the majority of) device drivers are
>> >> running on kernel space, so why don't we maintain a consistent device
>> >> drivers infrastructure, say, either all in user space, or all in
>> >> kernel space. (Sure some critical device drivers still need to be kept
>> >> in kernel space.)
>> >
>> > Feel free to create patches to do so, and handle all of the userspace
>> > changes needed in order to implement this.
>> >
>> > I think you haven't thought through the true reason we have device
>> > drivers, and why Linux isn't a microkernel...
>> >
>> > And I'd take exception to your "advantage:" line above, I don't believe
>> > that is true at all.
>> >
>> > Best of luck with your work,
>> >
>> Although I was asking "can we" do something, I am not actually
>> strongly in favor of either move or not move, as indeed it costs too
>> much to do the moving job.
>
> Then why even discuss this at all?  What is your goal here?  If it is to
> have others do work based on an idea you pointed out, you are in the
> wrong place.
>
>> But when you say "handle all of the userspace changes needed in order
>> to implement this", if the device driver can be moved without
>> modification (like the paper above shows), there should not be much
>> changes required for user space applications.
>
> The paper shows one such implementation that purports to not need
> userspace changes.  As we have yet to see any code, I remain
> unconvinced.
>
>> Also, if user space device drivers is a bad idea, why drivers/uio has
>> been created and merged into mainline kernel?
>
> UIO fits a real need for some types of devices, why wouldn't it be
> merged?  You are trying to say it is to be used for all drivers, which
> is totally missing the point.
>
>> Regarding "And I'd take exception to your "advantage:" line above, I
>> don't believe that is true at all", do you agree that a significant
>> portion of kernel crash incidents are due to bugs in drivers?
>
> No I do not.  If you refer to the references from the paper where they
> make that claim, they are talking about a different operating system
> than Linux.  But, by virtue of the fact that the majority of the code
> running in your kernel is drivers, yes, odds are drivers will have a
> small majority of the bugs overall, given the percentages involved.
> However, the bugs-per-line-of-code for Linux drivers is _much_ less than
> other operating systems, especially given the fact that Linux drivers
> require much less lines of code overall than other operating systems
> (30% at the most for the majority of device types.)  So I would like to
> see real numbers backing up your claim before I agree with it.
>
Hi, Greg,

These two studies support my point. If the first one is too old, then
the second one should be more convincing. To save your time, you can
take a look at their conclusion first.

An Empirical Study of Operating Systems Errors
http://www.stanford.edu/~engler/metrics-sosp-01.pdf

Faults in Linux: Ten Years Later
http://pagesperso-systeme.lip6.fr/Suman.Saha/src/asplos11.pdf

Regards
Jidong