[syzbot] general protection fault in legacy_parse_param

[Christian Brauner]

On Mon, Aug 30, 2021 at 10:41:29AM -0700, Casey Schaufler wrote:
> On 8/30/2021 9:57 AM, Christian Brauner wrote:
> > On Mon, Aug 30, 2021 at 09:40:57AM -0700, Casey Schaufler wrote:
> >> On 8/30/2021 7:25 AM, Casey Schaufler wrote:
> >>> On 8/30/2021 5:23 AM, Christian Brauner wrote:
> >>>> On Fri, Aug 27, 2021 at 07:11:18PM -0700, syzbot wrote:
> >>>>> syzbot has bisected this issue to:
> >>>>>
> >>>>> commit 54261af473be4c5481f6196064445d2945f2bdab
> >>>>> Author: KP Singh <kpsingh at google.com>
> >>>>> Date:   Thu Apr 30 15:52:40 2020 +0000
> >>>>>
> >>>>>     security: Fix the default value of fs_context_parse_param hook
> >>>>>
> >>>>> bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=160c5d75300000
> >>>>> start commit:   77dd11439b86 Merge tag 'drm-fixes-2021-08-27' of git://ano..
> >>>>> git tree:       upstream
> >>>>> final oops:     https://syzkaller.appspot.com/x/report.txt?x=150c5d75300000
> >>>>> console output: https://syzkaller.appspot.com/x/log.txt?x=110c5d75300000
> >>>>> kernel config:  https://syzkaller.appspot.com/x/.config?x=2fd902af77ff1e56
> >>>>> dashboard link: https://syzkaller.appspot.com/bug?extid=d1e3b1d92d25abf97943
> >>>>> syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=126d084d300000
> >>>>> C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=16216eb1300000
> >>>>>
> >>>>> Reported-by: syzbot+d1e3b1d92d25abf97943 at syzkaller.appspotmail.com
> >>>>> Fixes: 54261af473be ("security: Fix the default value of fs_context_parse_param hook")
> >>>>>
> >>>>> For information about bisection process see: https://goo.gl/tpsmEJ#bisection
> >>>> So ok, this seems somewhat clear now. When smack and 
> >>>> CONFIG_BPF_LSM=y
> >>>> is selected the bpf LSM will register NOP handlers including
> >>>>
> >>>> bpf_lsm_fs_context_fs_param()
> >>>>
> >>>> for the
> >>>>
> >>>> fs_context_fs_param
> >>>>
> >>>> LSM hook. The bpf LSM runs last, i.e. after smack according to:
> >>>>
> >>>> CONFIG_LSM="landlock,lockdown,yama,safesetid,integrity,tomoyo,smack,bpf"
> >>>>
> >>>> in the appended config. The smack hook runs and sets
> >>>>
> >>>> param->string = NULL
> >>>>
> >>>> then the bpf NOP handler runs returning -ENOPARM indicating to the vfs
> >>>> parameter parser that this is not a security module option so it should
> >>>> proceed processing the parameter subsequently causing the crash because
> >>>> param->string is not allowed to be NULL (Which the vfs parameter parser
> >>>> verifies early in fsconfig().).
> >>> The security_fs_context_parse_param() function is incorrectly
> >>> implemented using the call_int_hook() macro. It should return
> >>> zero if any of the modules return zero. It does not follow the
> >>> usual failure model of LSM hooks. It could be argued that the
> >>> code was fine before the addition of the BPF hook, but it was
> >>> going to fail as soon as any two security modules provided
> >>> mount options.
> >>>
> >>> Regardless, I will have a patch later today. Thank you for
> >>> tracking this down.
> >> Here's my proposed patch. I'll tidy it up with a proper
> >> commit message if it looks alright to y'all. I've tested
> >> with Smack and with and without BPF.
> > Looks good to me.
> > On question, in contrast to smack, selinux returns 1 instead of 0 on
> > success. So selinux would cause an early return preventing other hooks
> > from running. Just making sure that this is intentional.
> >
> > Iirc, this would mean that selinux causes fsconfig() to return a
> > positive value to userspace which I think is a bug; likely in selinux.
> > So I think selinux should either return 0 or the security hook itself
> > needs to overwrite a positive value with a sensible errno that can be
> > seen by userspace.
> 
> I think that I agree. The SELinux and Smack versions of the
> hook are almost identical except for setting rc to 1 in the
> SELinux case. And returning 1 makes no sense if you follow
> the callers back. David Howells wrote both the SELinux and
> Smack versions. David - why are they different? which is correct?

The documentation for fs_context_parse_param notes:

 * @fs_context_parse_param:
 *	Userspace provided a parameter to configure a superblock.  The LSM may
 *	reject it with an error and may use it for itself, in which case it
 *	should return 0; otherwise it should return -ENOPARAM to pass it on to
 *	the filesystem.
 *	@fc indicates the filesystem context.
 *	@param The parameter

So we should simply make selinux return 0 on top of your patch when it
has consumed the option.