Re: [PATCH 1/2] prctl: more prctl(PR_SET_MM_*) checks

[Cyrill Gorcunov <gorcunov@gmail.com>]

On Thu, May 28, 2015 at 01:12:14AM +0300, Alexey Dobriyan wrote:
> On Thu, May 28, 2015 at 01:02:29AM +0300, Cyrill Gorcunov wrote:
> > On Thu, May 28, 2015 at 12:47:57AM +0300, Alexey Dobriyan wrote:
> > > Individual prctl(PR_SET_MM_*) calls do some checking to maintain
> > > consistent view of mm->arg_start et al fields, but not enough.
> > > In particular PR_SET_MM_ARG_START/PR_SET_MM_ARG_END/PR_SET_MM_ENV_START/PR_SET_MM_ENV_END
> > > only check that address lies in existent VMA, but doesn't check that
> > > start address is lower that end address _at all_.
> > > 
> > > Consolidate all consistency checks, so there will be no difference in
> > > the future between PR_SET_MM_MAP and individual PR_SET_MM_* calls.
> > > 
> > > The program below makes both ARGV and ENVP areas reverted,
> > > makes /proc/$PID/cmdline show garbage (doesn't oops by luck).
> > 
> > Why should it oops?
> 
> Anything can happen if you constantly write code like this
> 
> 	unsigned long len = mm->arg_end - mm->arg_start;
> 
> and expect result to be positive.

It won't 'cause proc code will limit it (this where 'luck'
comes from :). Anyway, switching to validate() helper will
make code more robust, so thank you!