Re: [PATCH 0/2] Lustre debugfs fixes

From: Greg Kroah-Hartman
Date: Sun Feb 07 2016 - 21:22:28 EST


On Sun, Feb 07, 2016 at 06:51:09PM -0500, Oleg Drokin wrote:
>
> On Feb 7, 2016, at 4:39 PM, Greg Kroah-Hartman wrote:
>
> > On Sat, Feb 06, 2016 at 02:01:49AM -0500, green@xxxxxxxxxxxxxx wrote:
> >> From: Oleg Drokin <green@xxxxxxxxxxxxxx>
> >>
> >> These two patches tie some loose ends from the Lustre debugfs conversion,
> >> but while investigating them I also accumulated some questions
> >> that would be good to get answers for.
> >>
> >> 1. Unlike procfs, debugfs does not really guard your back and if root
> >> comes in and tries to write to a readonly file (or read a write-only one),
> >> it's allowed (as are permission changes too) as long as the appropriate write
> >> (or read) method is provided.
> >> So apparently there's whole class of bugs related to this, sample
> >> exhibits are in e.g. acpi_ec_add_debugfs creating a totally noop module
> >> parameter to control writes that does not really prevent any writes
> >> (patch submitted separately).
> >> But also things like wil_debugfs_create_iomem_x32 where when called from
> >> e.g. wil6210_debugfs_init_offset, some read-only attributed get a generic
> >> write method that would write straight to hardware registers (who knows
> >> what would happen when you write there, possibly they are readonly, but
> >> you are not getting an error).
> >> At first it looked like an easy way to catch this would be to just check
> >> for RO/WO mode with write/read handler set, but this is thwarted by
> >> the simple attribute defines that always assign read and write methods,
> >> but do the check internally for the get/set method instead.
> >> But also some fault injection code that sets readonly access on some files,
> >> but provides a fully functional write method that works as desired.
> >>
> >> Would it make sense to redo the simple-attribute framework to easy such
> >> cases detection (and also update writeable attributes to have permissions
> >> reflecting this) and have a correspinding kernel debug compile option
> >> to check for these?
> >
> > If a developer provides the write hooks for a debugfs file, and
> > userspace changes the permissions to write to it, why would you prevent
> > this? Perhaps this is what is intended.
>
> Well, it works differently for procfs where you cannot really change the permissions.
> I understand the developer might envision permission changes (or even ignoring of
> permissions by root) and there is such a code out there even.
> But DEFINE_SIMPLE_ATTRIBUTE code for one always provides both read and write methods
> and then internally checks if get and set are available and returns -EACCESS if not.
> That's totally ok too (other than -EACCESS that I think it not permitted as an errno
> from read/write with EINVAL being used for that), I am just trying to figure out if
> there is a way to more automatically detect cases where the write or read access is
> not desired and is left there by mistake.

I don't really know, the SIMPLE_ATTRIBUTE() stuff was just a way to try
to reduce a ton of boiler-plate code that was all over the place. I
doubt anyone thought of this when it was added (I know I didn't.)

> > Remember, debugfs is only for debugging stuff, never rely on it for
> > actual device/system use.
>
> Yes, I understand that. But debugfs is on by default pretty much anywhere

But it's root-only, thankfully. If only ftrace wouldn't use it, distros
could turn it off entirely :(

> and if there's a bug that lets you write to some hardware register you
> were supposed to only read and get some elevated privileges, or you
> read some write-only file and crash the kernel - that's bad too.

I agree, again, which is why it's mounted root-only. Also, bad things
happen when you remove debugfs files while someone has a file open :)

thanks,

greg k-h