Hi,
On Mon, Sep 14, 2020 at 09:39:10PM +0200, Jann Horn wrote:
> On Sun, Sep 13, 2020 at 6:56 PM John Wood <john.wood@xxxxxxx> wrote:
> > On Fri, Sep 11, 2020 at 02:01:56AM +0200, Jann Horn wrote:
> > > On Fri, Sep 11, 2020 at 1:49 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> > > > On Thu, Sep 10, 2020 at 01:21:06PM -0700, Kees Cook wrote:
> > > > [...]
> > > > I don't think this is the right place for detecting a crash -- isn't
> > > > this only for the "dumping core" condition? In other words, don't you
> > > > want to do this in get_signal()'s "fatal" block? (i.e. very close to the
> > > > do_coredump, but without the "should I dump?" check?)
> > > >
> > > > Hmm, but maybe I'm wrong? It looks like you're looking at noticing the
> > > > process taking a signal from SIG_KERNEL_COREDUMP_MASK ?
> > > >
> > > > (Better yet: what are fatal conditions that do NOT match
> > > > SIG_KERNEL_COREDUMP_MASK, and should those be covered?)
> > > >
> > > > Regardless, *this* looks like the only place without an LSM hook. And it
> > > > doesn't seem unreasonable to add one here. I assume it would probably
> > > > just take the siginfo pointer, which is also what you're checking.
> > >
> > > Good point, making this an LSM might be a good idea.
> > >
> > > > e.g. for include/linux/lsm_hook_defs.h:
> > > >
> > > > LSM_HOOK(int, 0, task_coredump, const kernel_siginfo_t *siginfo);
> > >
> > > I guess it should probably be an LSM_RET_VOID hook? And since, as you
> > > said, it's not really semantically about core dumping, maybe it should
> > > be named task_fatal_signal or something like that.
> >
> > If I understand correctly you propose to add a new LSM hook without return
> > value and place it here:
> >
> > diff --git a/kernel/signal.c b/kernel/signal.c
> > index a38b3edc6851..074492d23e98 100644
> > --- a/kernel/signal.c
> > +++ b/kernel/signal.c
> > @@ -2751,6 +2751,8 @@ bool get_signal(struct ksignal *ksig)
> > do_coredump(&ksig->info);
> > }
> >
> > + // Add the new LSM hook here
> > +
> > /*
> > * Death signals, no core dump.
> > */
>
> It should probably be in the "if (sig_kernel_coredump(signr)) {"
> branch. And I'm not sure whether it should be before or after
> do_coredump() - if you do it after do_coredump(), the hook will have
> to wait until the core dump file has been written, which may take a
> little bit of time.
But if the LSM hook is placed in the "if (sig_kernel_coredump(signr)) {"
branch, then only the following signals will be passed to it.
SIGQUIT, SIGILL, SIGTRAP, SIGABRT, SIGFPE, SIGSEGV, SIGBUS, SIGSYS,
SIGXCPU, SIGXFSZ, SIGEMT
The above signals are extracted from SIG_KERNEL_COREDUMP_MASK macro, and
are only related to coredump.
So, if we add a new LSM hook (named task_fatal_signal) to detect a fatal
signal it would be better to place it just above the if statement.
diff --git a/kernel/signal.c b/kernel/signal.c
index a38b3edc6851..406af87f2f96 100644
--- a/kernel/signal.c
+++ b/kernel/signal.c
@@ -2736,6 +2736,8 @@ bool get_signal(struct ksignal *ksig)
*/
current->flags |= PF_SIGNALED;
+ // Place the new LSM hook here
+
if (sig_kernel_coredump(signr)) {
if (print_fatal_signals)
print_fatal_signal(ksig->info.si_signo);
This way all the fatal signals are caught and we also avoid the commented
delay if a core dump is necessary.
Thanks,
John Wood
