Re: [PATCH] pidfd: Stop taking cred_guard_mutex

[Eric W. Biederman]

Jann Horn <jannh@xxxxxxxxxx> writes:

> On Tue, Mar 10, 2020 at 9:00 PM Jann Horn <jannh@xxxxxxxxxx> wrote:
>> On Tue, Mar 10, 2020 at 8:29 PM Eric W. Biederman <ebiederm@xxxxxxxxxxxx> wrote:
>> > Jann Horn <jannh@xxxxxxxxxx> writes:
>> > > On Tue, Mar 10, 2020 at 7:54 PM Eric W. Biederman <ebiederm@xxxxxxxxxxxx> wrote:
>> > >> During exec some file descriptors are closed and the files struct is
>> > >> unshared.  But all of that can happen at other times and it has the
>> > >> same protections during exec as at ordinary times.  So stop taking the
>> > >> cred_guard_mutex as it is useless.
>> > >>
>> > >> Furthermore he cred_guard_mutex is a bad idea because it is deadlock
>> > >> prone, as it is held in serveral while waiting possibly indefinitely
>> > >> for userspace to do something.
> [...]
>> > > If you make this change, then if this races with execution of a setuid
>> > > program that afterwards e.g. opens a unix domain socket, an attacker
>> > > will be able to steal that socket and inject messages into
>> > > communication with things like DBus. procfs currently has the same
>> > > race, and that still needs to be fixed, but at least procfs doesn't
>> > > let you open things like sockets because they don't have a working
>> > > ->open handler, and it enforces the normal permission check for
>> > > opening files.
>> >
>> > It isn't only exec that can change credentials.  Do we need a lock for
>> > changing credentials?
> [...]
>> > If we need a lock around credential change let's design and build that.
>> > Having a mismatch between what a lock is designed to do, and what
>> > people use it for can only result in other bugs as people get confused.
>>
>> Hmm... what benefits do we get from making it a separate lock? I guess
>> it would allow us to make it a per-task lock instead of a
>> signal_struct-wide one? That might be helpful...
>
> But actually, isn't the core purpose of the cred_guard_mutex to guard
> against concurrent credential changes anyway? That's what almost
> everyone uses it for, and it's in the name...

Having been through all of the users nope.

Maybe someone tried to repurpose for that.  I haven't traced through
when it went the it was renamed from cred_exec_mutex to
cred_guard_mutex.

The original purpose was to make make exec and ptrace deadlock.  But it
was seen as being there to allow safely calculating the new credentials
before the point of now return.  Because if a process is ptraced or not
affects the new credential calculations.  Unfortunately offering that
guarantee fundamentally leads to deadlock.

So ptrace_attach and seccomp use the cred_guard_mutex to guarantee
a deadlock.

The common use is to take cred_guard_mutex to guard the window when
credentials and process details are out of sync in exec.  But there
is at least do_io_accounting that seems to have the same justification
for holding __pidfd_fget.

With effort I suspect we can replace exec_change_mutex with task_lock.
When we are guaranteed to be single threaded placing exec_change_mutex
in signal_struct doesn't really help us (except maybe in some races?).

The deep problem is no one really understands cred_guard_mutex so it is
a mess.  Code with poorly defined semantics is always wrong somewhere
for someone.  Which is part of why I am attacking this and having the
conversations to make certain I understand what is going on.

I see your point about commit_creds making a process undumpable.  So in
practice it really is only exec that changes creds in a way that
ptrace_may_access will allow the process to be inspected.

So I guess for now the practical non-regressing course is to change
everything to my exec_change_mutex, removing the deadlock.  Then we
figure out how to cleanly deal with the races inspecting a process with
changing credentials has.

Eric