[PATCH 2/3] namei: implement AT_THIS_ROOT chroot-like path resolution

[Christian Brauner]

On Sat, Oct 6, 2018 at 10:56 PM Florian Weimer <fw at deneb.enyo.de> wrote:
>
> * Aleksa Sarai:
>
> > On 2018-10-01, Andy Lutomirski <luto at amacapital.net> wrote:
> >> >>> Currently most container runtimes try to do this resolution in
> >> >>> userspace[1], causing many potential race conditions. In addition, the
> >> >>> "obvious" alternative (actually performing a {ch,pivot_}root(2))
> >> >>> requires a fork+exec which is *very* costly if necessary for every
> >> >>> filesystem operation involving a container.
> >> >>
> >> >> Wait. fork() I understand, but why exec? And actually, you don't need
> >> >> a full fork() either, clone() lets you do this with some process parts
> >> >> shared. And then you also shouldn't need to use SCM_RIGHTS, just keep
> >> >> the file descriptor table shared. And why chroot()/pivot_root(),
> >> >> wouldn't you want to use setns()?
> >> >
> >> > You're right about this -- for C runtimes. In Go we cannot do a raw
> >> > clone() or fork() (if you do it manually with RawSyscall you'll end with
> >> > broken runtime state). So you're forced to do fork+exec (which then
> >> > means that you can't use CLONE_FILES and must use SCM_RIGHTS). Same goes
> >> > for CLONE_VFORK.
> >>
> >> I must admit that I’m not very sympathetic to the argument that “Go’s
> >> runtime model is incompatible with the simpler solution.”
> >
> > Multi-threaded programs have a similar issue (though with Go it's much
> > worse). If you fork a multi-threaded C program then you can only safely
> > use AS-Safe glibc functions (those that are safe within a signal
> > handler). But if you're just doing three syscalls this shouldn't be as
> > big of a problem as Go where you can't even do said syscalls.
>
> The situation is a bit more complicated.  There are many programs out
> there which use malloc and free (at least indirectly) after a fork,
> and we cannot break them.  In glibc, we have a couple of subsystems
> which are put into a known state before calling the fork/clone system
> call if the application calls fork.  The price we pay for that is a
> fork which is not POSIX-compliant because it is not async-signal-safe.
> Admittedly, other libcs chose different trade-offs.
>
> However, what is the same across libcs is this: You cannot call the
> clone system call directly and get a fully working new process.  Some
> things break.  For example, for recursive mutexes, we need to know the
> TID of the current thread, and we cannot perform a system call to get
> it for performance reasons.  So everyone has a TID cache for that.
> But the TID cache does not get reset when you bypass the fork
> implementation in libc, so you end up with subtle corruption bugs on
> TID reuse.

Sure, but recursive mutexes etc. are very specific use-case.
I'd even go so far to say that if you use mutexes + threads and then
also fork in those threads you're hosed anyway. If you don't things get a little
cleaner assuming you don't call library functions that use mutexes
internally.
Event then you might (sometimes at least) still get around most problems
with atfork handlers (thought I really don't like him). But you know more
about this then I do. :)

>
> So I'd say that in most cases, the C situation is pretty much the same
> as the Go situation.  If I recall correctly, the problem for Go is
> that it cannot call setns from Go code because it fails in the kernel
> for multi-threaded processes, and Go processes are already
> multi-threaded when user Go code runs.

That is true for *some* namespaces (user, mount) but not for all.
For example, setns(CLONE_NEWNET) would be fine from go.
But the go runtime thinks it's clever to clone a new thread in between
entry and exit of a syscall. If you switch namespaces you might end
up with a new thread that belongs to the wrong namespace which is
very problematic.
So you can either rely on calling some go magic that locks
you to a specific os thread but that does only work in later go versions or
you go the constructor route, i.e. you e.g. implement a (dummy)
subcommand that you can call and that triggers the execution of a
C function that is marked with __attribute__((constructor)) that runs
before the go runtime and in which you can do setns(), fork() and
friends (somewhat) safely. This has very
bad performance and is a nasty hack but it's really unavoidable.