[PATCH 5/8] arm64: Recover kretprobe modified return address in stacktrace

[Masami Hiramatsu]

On Wed, 13 Oct 2021 11:13:51 +0100
Mark Rutland <mark.rutland at arm.com> wrote:

> On Fri, Oct 08, 2021 at 09:28:58PM +0900, Masami Hiramatsu wrote:
> > Since the kretprobe replaces the function return address with
> > the kretprobe_trampoline on the stack, stack unwinder shows it
> > instead of the correct return address.
> > 
> > This checks whether the next return address is the
> > __kretprobe_trampoline(), and if so, try to find the correct
> > return address from the kretprobe instance list.
> > 
> > With this fix, now arm64 can enable
> > CONFIG_ARCH_CORRECT_STACKTRACE_ON_KRETPROBE, and pass the
> > kprobe self tests.
> > 
> > Signed-off-by: Masami Hiramatsu <mhiramat at kernel.org>
> > ---
> >  arch/arm64/Kconfig                  |    1 +
> >  arch/arm64/include/asm/stacktrace.h |    2 ++
> >  arch/arm64/kernel/stacktrace.c      |    3 +++
> >  3 files changed, 6 insertions(+)
> > 
> > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > index 5c7ae4c3954b..edde5171ffb2 100644
> > --- a/arch/arm64/Kconfig
> > +++ b/arch/arm64/Kconfig
> > @@ -11,6 +11,7 @@ config ARM64
> >  	select ACPI_PPTT if ACPI
> >  	select ARCH_HAS_DEBUG_WX
> >  	select ARCH_BINFMT_ELF_STATE
> > +	select ARCH_CORRECT_STACKTRACE_ON_KRETPROBE
> >  	select ARCH_ENABLE_HUGEPAGE_MIGRATION if HUGETLB_PAGE && MIGRATION
> >  	select ARCH_ENABLE_MEMORY_HOTPLUG
> >  	select ARCH_ENABLE_MEMORY_HOTREMOVE
> > diff --git a/arch/arm64/include/asm/stacktrace.h b/arch/arm64/include/asm/stacktrace.h
> > index 8aebc00c1718..8f997a602651 100644
> > --- a/arch/arm64/include/asm/stacktrace.h
> > +++ b/arch/arm64/include/asm/stacktrace.h
> > @@ -9,6 +9,7 @@
> >  #include <linux/sched.h>
> >  #include <linux/sched/task_stack.h>
> >  #include <linux/types.h>
> > +#include <linux/llist.h>
> >  
> >  #include <asm/memory.h>
> >  #include <asm/ptrace.h>
> > @@ -59,6 +60,7 @@ struct stackframe {
> >  #ifdef CONFIG_FUNCTION_GRAPH_TRACER
> >  	int graph;
> >  #endif
> > +	struct llist_node *kr_cur;
> 
> As with the fgraph bits above, please make this depedn on the relevant
> Kconfig, i.e.
> 
> | #ifdef CONFIG_KRETPROBES
> | 	struct llist_node *kr_cur;
> | #endif
> 
> >  };

OK.

> >  
> >  extern int unwind_frame(struct task_struct *tsk, struct stackframe *frame);
> > diff --git a/arch/arm64/kernel/stacktrace.c b/arch/arm64/kernel/stacktrace.c
> > index 8982a2b78acf..f1eef5745542 100644
> > --- a/arch/arm64/kernel/stacktrace.c
> > +++ b/arch/arm64/kernel/stacktrace.c
> > @@ -129,6 +129,8 @@ int notrace unwind_frame(struct task_struct *tsk, struct stackframe *frame)
> >  		frame->pc = ret_stack->ret;
> >  	}
> >  #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
> > +	if (is_kretprobe_trampoline(frame->pc))
> > +		frame->pc = kretprobe_find_ret_addr(tsk, (void *)frame->fp, &frame->kr_cur);
> 
> Please ifdef this, like the CONFIG_FUNCTION_GRAPH_TRACER bits above.
> i.e.
> 
> | #ifdef CONFIG_KRETPROBES
> | 	if (is_kretprobe_trampoline(frame->pc))
> | 		frame->pc = kretprobe_find_ret_addr(tsk, (void *)frame->fp, &frame->kr_cur);
> | #endif

OK.

> 
> What does kretprobe_find_ret_addr() do when it can't find the original
> address? I couldn't spot it in this series or in mainline.

If it couldn't find, it returns NULL.
Hmm, should we check it and keep frame->pc if the return value is NULL?
(anyway, it must not happen. If it happens, that task can not continue to run.)

> As a future thing, I'd like to be able to have monotonicity and
> completeness checks as part of the unwind, i.e. checking that we consume
> the kretprobe address *in-order*, and can identify whether we've skipped
> any, so that we can identify when unwinding has gone wrong. Does it do
> that today?

Good question. No today, but is easy to do since we have the loop cursor
(frame->kr_cur).

unsigned long kretprobe_find_ret_addr(struct task_struct *tsk, void *fp,
                                      struct llist_node **cur)
{
...
        do {
                ret = __kretprobe_find_ret_addr(tsk, cur);
                if (!ret)
                        break;
                ri = container_of(*cur, struct kretprobe_instance, llist);
        } while (ri->fp != fp);
...

Change this loop to;

	ri = container_of(*cur, struct kretprobe_instance, llist);
	prev_fp = ri->fp;
	do {
		ret = __kretprobe_find_ret_addr(tsk, cur);
		if (WARN_ON_ONCE(!ret))
			return ERR_PTR(-ENOENT);
		ri = container_of(*cur, struct kretprobe_instance, llist);
	} while (ri->fp == prev_fp);
	if (ri->fp != fp)
		return ERR_PTR(-EILSEQ);

Then, we can detect that wrong sequence from stacktrace side.

> 
> It'd be nice if it could signal failure reliably (without causing a
> BUG() or similar), e.g. by returning an error code.

As above, -EILSEQ is OK?

> 
> >  	frame->pc = ptrauth_strip_insn_pac(frame->pc);
> >  
> > @@ -224,6 +226,7 @@ noinline notrace void arch_stack_walk(stack_trace_consume_fn consume_entry,
> >  {
> >  	struct stackframe frame;
> >  
> > +	memset(&frame, 0, sizeof(frame));
> 
> Please initialise stackframe::kr_cur in start_backtrace, where we
> initialize all the other fields in struct stackframe, i.e. just after
> the CONFIG_FUNCTION_GRAPH_TRACER bit, have:
> 
> | #ifdef CONFIG_KRETPROBES
> | 	frame->kr_cur = NULL;
> | #endif

OK, let me update it.

Thank you!

> 
> Thanks,
> Mark.
> 
> >  	if (regs)
> >  		start_backtrace(&frame, regs->regs[29], regs->pc);
> >  	else if (task == current)
> > 


-- 
Masami Hiramatsu <mhiramat at kernel.org>