Re: [PATCH v2 0/3] ubsan: Split out bounds checker

[Kees Cook]

On Fri, Nov 22, 2019 at 10:07:29AM +0100, Dmitry Vyukov wrote:
> On Thu, Nov 21, 2019 at 7:15 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> >
> > v2:
> >     - clarify Kconfig help text (aryabinin)
> >     - add reviewed-by
> >     - aim series at akpm, which seems to be where ubsan goes through?
> > v1: https://lore.kernel.org/lkml/20191120010636.27368-1-keescook@xxxxxxxxxxxx
> >
> > This splits out the bounds checker so it can be individually used. This
> > is expected to be enabled in Android and hopefully for syzbot. Includes
> > LKDTM tests for behavioral corner-cases (beyond just the bounds checker).
> >
> > -Kees
> 
> +syzkaller mailing list
> 
> This is great!

BTW, can I consider this your Acked-by for these patches? :)

> I wanted to enable UBSAN on syzbot for a long time. And it's
> _probably_ not lots of work. But it was stuck on somebody actually
> dedicating some time specifically for it.

Do you have a general mechanism to test that syzkaller will actually
pick up the kernel log splat of a new check? I noticed a few things
about the ubsan handlers: they don't use any of the common "warn"
infrastructure (neither does kasan from what I can see), and was missing
a check for panic_on_warn (kasan has this, but does it incorrectly).

I think kasan and ubsan should be reworked to use the common warn
infrastructure, and at the very least, ubsan needs this:

diff --git a/lib/ubsan.c b/lib/ubsan.c
index e7d31735950d..a2535a62c9af 100644
--- a/lib/ubsan.c
+++ b/lib/ubsan.c
@@ -160,6 +160,17 @@ static void ubsan_epilogue(unsigned long *flags)
 		"========================================\n");
 	spin_unlock_irqrestore(&report_lock, *flags);
 	current->in_ubsan--;
+
+	if (panic_on_warn) {
+		/*
+		 * This thread may hit another WARN() in the panic path.
+		 * Resetting this prevents additional WARN() from panicking the
+		 * system on this thread.  Other threads are blocked by the
+		 * panic_mutex in panic().
+		 */
+		panic_on_warn = 0;
+		panic("panic_on_warn set ...\n");
+	}
 }
 
 static void handle_overflow(struct overflow_data *data, void *lhs,

> Kees, or anybody else interested, could you provide relevant configs
> that (1) useful for kernel,

As mentioned in the other email (but just to keep the note together with
the other thoughts here) after this series, you'd want:

CONFIG_UBSAN=y
CONFIG_UBSAN_BOUNDS=y
# CONFIG_UBSAN_MISC is not set

> (2) we want 100% cleanliness,

What do you mean here by "cleanliness"? It seems different from (3)
about the test tripping a lot?

> (3) don't
> fire all the time even without fuzzing?

I ran with the bounds checker enabled (and the above patch) under
syzkaller for the weekend and saw 0 bounds checker reports.

> Anything else required to
> enable UBSAN? I don't see anything. syzbot uses gcc 8.something, which
> I assume should be enough (but we can upgrade if necessary).

As mentioned, gcc 8+ should be fine.

-- 
Kees Cook