| Message ID | 20200702232638.2946421-5-keescook@chromium.org |
|---|---|
| State | Not Applicable |
| Delegated to: | BPF Maintainers |
| Headers | show |
| Series | Refactor kallsyms_show_value() users for correct cred | expand |
On Thu, Jul 2, 2020 at 4:26 PM Kees Cook <keescook@chromium.org> wrote: > > The kprobe show() functions were using "current"'s creds instead > of the file opener's creds for kallsyms visibility. Fix to use > seq_file->file->f_cred. Side note: I have a distinct - but despite that possibly quite incorrect - memory that I've discussed with somebody several years ago about making "current_cred()" simply warn in any IO context. IOW, we could have read and write just increment/decrement a per-thread counter, and have current_cred() do a WARN_ON_ONCE() if it's called with that counter incremented. The issue of ioctl's is a bit less obvious - there are reasons to argue those should also use open-time credentials, but on the other hand I think it's reasonable to pass a file descriptor to a suid app in order for that app to do things that the normal user cannot. But read/write are dangerous because of the "it's so easy to fool suid apps to read/write stdin/stdout". So pread/pwrite/ioctl/splice etc are things that suid applications very much do on purpose to affect a file descriptor. But plain read/write are things that might be accidental and used by attack vectors. If somebody is interested in looking into things like that, it might be a good idea to have kernel threads with that counter incremented by default. Just throwing this idea out in case somebody wants to try it. It's not just "current_cred", of course. It's all the current_cred_xxx() users too. But it may be that there are a ton of false positives because maybe some code on purpose ends up doing things like just *comparing* current_cred with file->f_cred and then that would warn too. Linus
On Thu, Jul 02, 2020 at 06:00:17PM -0700, Linus Torvalds wrote: > On Thu, Jul 2, 2020 at 4:26 PM Kees Cook <keescook@chromium.org> wrote: > > > > The kprobe show() functions were using "current"'s creds instead > > of the file opener's creds for kallsyms visibility. Fix to use > > seq_file->file->f_cred. > > Side note: I have a distinct - but despite that possibly quite > incorrect - memory that I've discussed with somebody several years ago > about making "current_cred()" simply warn in any IO context. > > IOW, we could have read and write just increment/decrement a > per-thread counter, and have current_cred() do a WARN_ON_ONCE() if > it's called with that counter incremented. That does sound familiar. I can't find a thread on it, but my search abilities are poor. :) So an increment/decrement in all the IO-related syscalls, or were you thinking of some other location? > The issue of ioctl's is a bit less obvious - there are reasons to > argue those should also use open-time credentials, but on the other > hand I think it's reasonable to pass a file descriptor to a suid app > in order for that app to do things that the normal user cannot. > > But read/write are dangerous because of the "it's so easy to fool suid > apps to read/write stdin/stdout". > > So pread/pwrite/ioctl/splice etc are things that suid applications > very much do on purpose to affect a file descriptor. But plain > read/write are things that might be accidental and used by attack > vectors. So probably just start with read/write and tighten it over time, if we find other clear places, leaving ioctl/pread/pwrite/splice alone. > If somebody is interested in looking into things like that, it might > be a good idea to have kernel threads with that counter incremented by > > Just throwing this idea out in case somebody wants to try it. It's not > just "current_cred", of course. It's all the current_cred_xxx() users > too. But it may be that there are a ton of false positives because > maybe some code on purpose ends up doing things like just *comparing* > current_cred with file->f_cred and then that would warn too. Yeah ... and I think the kthread test should answer that question.
On Thu, Jul 02, 2020 at 06:00:17PM -0700, Linus Torvalds wrote: > If somebody is interested in looking into things like that, it might > be a good idea to have kernel threads with that counter incremented by > default. With 67 kthreads on a booted system, this patch does not immediately blow up... And it likely needs some beautification. (Note that current_cred_*() calls current_cred() under the hood, so AFAICT, only current_cred() needs coverage.) diff --git a/include/linux/cred.h b/include/linux/cred.h index 18639c069263..a624847cb0ce 100644 --- a/include/linux/cred.h +++ b/include/linux/cred.h @@ -295,7 +295,10 @@ static inline void put_cred(const struct cred *_cred) * since nobody else can modify it. */ #define current_cred() \ - rcu_dereference_protected(current->cred, 1) +({ \ + WARN_ON_ONCE(current->warn_on_current_cred); \ + rcu_dereference_protected(current->cred, 1); \ +}) /** * current_real_cred - Access the current task's objective credentials diff --git a/include/linux/sched.h b/include/linux/sched.h index b62e6aaf28f0..21ab1b81aa40 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -652,6 +652,7 @@ struct task_struct { /* Per task flags (PF_*), defined further below: */ unsigned int flags; unsigned int ptrace; + unsigned int warn_on_current_cred; #ifdef CONFIG_SMP struct llist_node wake_entry; diff --git a/kernel/fork.c b/kernel/fork.c index 142b23645d82..2e181b9bfd3f 100644 --- a/kernel/fork.c +++ b/kernel/fork.c @@ -2527,8 +2527,12 @@ pid_t kernel_thread(int (*fn)(void *), void *arg, unsigned long flags) .stack = (unsigned long)fn, .stack_size = (unsigned long)arg, }; + pid_t pid; - return _do_fork(&args); + pid = _do_fork(&args); + if (pid == 0) + current->warn_on_current_cred = 1; + return pid; } #ifdef __ARCH_WANT_SYS_FORK
On Fri, Jul 3, 2020 at 8:50 AM Kees Cook <keescook@chromium.org> wrote: > > With 67 kthreads on a booted system, this patch does not immediately > blow up... Did you try making read/write inc/dec that thing too? Or does that just blow up with tons of warnings? Linus
On Sun, Jul 05, 2020 at 01:10:54PM -0700, Linus Torvalds wrote: > On Fri, Jul 3, 2020 at 8:50 AM Kees Cook <keescook@chromium.org> wrote: > > > > With 67 kthreads on a booted system, this patch does not immediately > > blow up... > > Did you try making read/write inc/dec that thing too? Or does that > just blow up with tons of warnings? I hadn't gotten to that yet. I need to catch up on other stuff first, and I thought I'd give people time to scream if they tested this themselves. :)
On Thu, 2 Jul 2020 16:26:37 -0700 Kees Cook <keescook@chromium.org> wrote: > The kprobe show() functions were using "current"'s creds instead > of the file opener's creds for kallsyms visibility. Fix to use > seq_file->file->f_cred. This looks good to me. Acked-by: Masami Hiramatsu <mhiramat@kernel.org> Thanks! > > Cc: stable@vger.kernel.org > Fixes: 81365a947de4 ("kprobes: Show address of kprobes if kallsyms does") > Fixes: ffb9bd68ebdb ("kprobes: Show blacklist addresses as same as kallsyms does") > Signed-off-by: Kees Cook <keescook@chromium.org> > --- > kernel/kprobes.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/kernel/kprobes.c b/kernel/kprobes.c > index d4de217e4a91..2e97febeef77 100644 > --- a/kernel/kprobes.c > +++ b/kernel/kprobes.c > @@ -2448,7 +2448,7 @@ static void report_probe(struct seq_file *pi, struct kprobe *p, > else > kprobe_type = "k"; > > - if (!kallsyms_show_value(current_cred())) > + if (!kallsyms_show_value(pi->file->f_cred)) > addr = NULL; > > if (sym) > @@ -2540,7 +2540,7 @@ static int kprobe_blacklist_seq_show(struct seq_file *m, void *v) > * If /proc/kallsyms is not showing kernel address, we won't > * show them here either. > */ > - if (!kallsyms_show_value(current_cred())) > + if (!kallsyms_show_value(m->file->f_cred)) > seq_printf(m, "0x%px-0x%px\t%ps\n", NULL, NULL, > (void *)ent->start_addr); > else > -- > 2.25.1 >
diff --git a/kernel/kprobes.c b/kernel/kprobes.c index d4de217e4a91..2e97febeef77 100644 --- a/kernel/kprobes.c +++ b/kernel/kprobes.c @@ -2448,7 +2448,7 @@ static void report_probe(struct seq_file *pi, struct kprobe *p, else kprobe_type = "k"; - if (!kallsyms_show_value(current_cred())) + if (!kallsyms_show_value(pi->file->f_cred)) addr = NULL; if (sym) @@ -2540,7 +2540,7 @@ static int kprobe_blacklist_seq_show(struct seq_file *m, void *v) * If /proc/kallsyms is not showing kernel address, we won't * show them here either. */ - if (!kallsyms_show_value(current_cred())) + if (!kallsyms_show_value(m->file->f_cred)) seq_printf(m, "0x%px-0x%px\t%ps\n", NULL, NULL, (void *)ent->start_addr); else
The kprobe show() functions were using "current"'s creds instead of the file opener's creds for kallsyms visibility. Fix to use seq_file->file->f_cred. Cc: stable@vger.kernel.org Fixes: 81365a947de4 ("kprobes: Show address of kprobes if kallsyms does") Fixes: ffb9bd68ebdb ("kprobes: Show blacklist addresses as same as kallsyms does") Signed-off-by: Kees Cook <keescook@chromium.org> --- kernel/kprobes.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)