Re: x86: warning in unwind_get_return_address

[Josh Poimboeuf]

On Thu, Jan 05, 2017 at 09:17:00AM -0600, Josh Poimboeuf wrote:
> On Thu, Jan 05, 2017 at 03:59:52PM +0100, Dmitry Vyukov wrote:
> > On Thu, Jan 5, 2017 at 3:49 PM, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> > > On Tue, Dec 27, 2016 at 05:38:59PM +0100, Dmitry Vyukov wrote:
> > >> On Thu, Dec 22, 2016 at 6:17 AM, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> > >> > On Wed, Dec 21, 2016 at 01:46:36PM +0100, Andrey Konovalov wrote:
> > >> >> On Wed, Dec 21, 2016 at 12:36 AM, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> > >> >> >
> > >> >> > Thanks.  Looking at the stack trace, my guess is that an interrupt hit
> > >> >> > while running in generated BPF code, and the unwinder got confused
> > >> >> > because regs->ip points to the generated code.  I may need to disable
> > >> >> > that warning until we figure out a better solution.
> > >> >> >
> > >> >> > Can you share your .config file?
> > >> >>
> > >> >> Sure, attached.
> > >> >
> > >> > Ok, I was able to recreate with your config.  The culprit was generated
> > >> > code, as I suspected, though it wasn't BPF, it was a kprobe (created by
> > >> > dccpprobe_init()).
> > >> >
> > >> > I'll make a patch to disable the warning.
> > >>
> > >> Hi,
> > >>
> > >> I am also seeing the following warnings:
> > >>
> > >> [  281.889259] WARNING: kernel stack regs at ffff8801c29a7ea8 in
> > >> syz-executor8:1302 has bad 'bp' value ffff8801c29a7f28
> > >> [  833.994878] WARNING: kernel stack regs at ffff8801c4e77ea8 in
> > >> syz-executor1:13094 has bad 'bp' value ffff8801c4e77f28
> > >>
> > >> Can it also be caused by bpf/kprobe?
> > >
> > > This is a different warning.  I suspect it's due to unwinding the stack
> > > of another CPU while it's running, which is still possible in a few
> > > places.  I'm going to have to disable all these warnings for now.
> > 
> > 
> > I also have the following diff locally. These loads trigger episodic
> > KASAN warnings about stack-of-bounds reads on rcu stall warnings when
> > it does backtrace of all cpus.
> > If it looks correct to you, can you please also incorporate it into your patch?
> 
> Ok, will do.
> 
> BTW, I think there's an issue with your mail client.  Your last two
> replies to me didn't have me on To/Cc.

Would you mind testing if the following patch fixes it?  It's based on
the latest linus/master.


diff --git a/arch/x86/kernel/unwind_frame.c b/arch/x86/kernel/unwind_frame.c
index 4443e49..05adf9a 100644
--- a/arch/x86/kernel/unwind_frame.c
+++ b/arch/x86/kernel/unwind_frame.c
@@ -6,6 +6,21 @@
 
 #define FRAME_HEADER_SIZE (sizeof(long) * 2)
 
+/*
+ * This disables KASAN checking when reading a value from another task's stack,
+ * since the other task could be running on another CPU and could have poisoned
+ * the stack in the meantime.
+ */
+#define UNWIND_READ_ONCE(state, x)			\
+({							\
+	unsigned long val;				\
+	if (state->task == current)			\
+		val = READ_ONCE(x);			\
+	else						\
+		val = READ_ONCE_NOCHECK(x);		\
+	val;						\
+})
+
 static void unwind_dump(struct unwind_state *state, unsigned long *sp)
 {
 	static bool dumped_before = false;
@@ -48,7 +63,8 @@ unsigned long unwind_get_return_address(struct unwind_state *state)
 	if (state->regs && user_mode(state->regs))
 		return 0;
 
-	addr = ftrace_graph_ret_addr(state->task, &state->graph_idx, *addr_p,
+	addr = UNWIND_READ_ONCE(state, *addr_p);
+	addr = ftrace_graph_ret_addr(state->task, &state->graph_idx, addr,
 				     addr_p);
 
 	return __kernel_text_address(addr) ? addr : 0;
@@ -162,7 +178,7 @@ bool unwind_next_frame(struct unwind_state *state)
 	if (state->regs)
 		next_bp = (unsigned long *)state->regs->bp;
 	else
-		next_bp = (unsigned long *)*state->bp;
+		next_bp = (unsigned long *)UNWIND_READ_ONCE(state, *state->bp);
 
 	/* is the next frame pointer an encoded pointer to pt_regs? */
 	regs = decode_frame_pointer(next_bp);
@@ -207,6 +223,16 @@ bool unwind_next_frame(struct unwind_state *state)
 	return true;
 
 bad_address:
+	/*
+	 * When dumping a task other than current, the task might actually be
+	 * running on another CPU, in which case it could be modifying its
+	 * stack while we're reading it.  This is generally not a problem and
+	 * can be ignored as long as the caller understands that unwinding
+	 * another task will not always succeed.
+	 */
+	if (state->task != current)
+		goto the_end;
+
 	if (state->regs) {
 		printk_deferred_once(KERN_WARNING
 			"WARNING: kernel stack regs at %p in %s:%d has bad 'bp' value %p\n",