[PATCH] arm: Add unwinding annotations for 64bit division functions

On Thu, Sep 22, 2011 at 01:13:01PM +0100, Jon Medhurst (Tixy) wrote:

On Thu, 2011-09-22 at 12:57 +0100, Catalin Marinas wrote:

quoted

On Thu, Sep 22, 2011 at 12:06:46PM +0100, Jon Medhurst (Tixy) wrote:

quoted

On Thu, 2011-09-22 at 10:48 +0100, Catalin Marinas wrote:

quoted

We could improve things a bit in the unwinder and assume
that if the fault address is the same as the .fnstart address, the
return value is always in LR and the SP not affected (that's unwinding
bytecode 0xb0). For a few instructions into the function prologue we
can't reliably get the unwinding information.

That would help make it possible to unwind out of kprobes handlers to
the probed function. The kprobes code itself would need work as well,
and possibly the undef handler. Do we think it is worthwhile to do
this? 

Does kprobes need to trace beyond the probed function? If not, you get
the address of the probed function via pt_regs anyway, so no need for
unwinding beyond that.

To be honest, I'm not very sure how kprobes get used in the real world.
Though, if stack unwinding from their handlers currently doesn't work
and people had a usecase for it, we would expect them to complain.

The unwinding fix should be simple (I haven't tested it yet):

8<-----------------------------
ARM: Ignore the unwinding information for the first instruction in a function

From: Catalin Marinas <catalin.marinas@arm.com>

When backtracing from the first instruction of a function, the prologue
has not been executed and the unwinding information is not valid. This
patch checks for this case and just assumes that the return address is
in LR.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
---
 arch/arm/kernel/unwind.c |   10 ++++++++++
 1 files changed, 10 insertions(+), 0 deletions(-)

diff --git a/arch/arm/kernel/unwind.c b/arch/arm/kernel/unwind.c
index d2cb0b3..946face 100644
--- a/arch/arm/kernel/unwind.c
+++ b/arch/arm/kernel/unwind.c

@@ -293,6 +293,16 @@ int unwind_frame(struct stackframe *frame)
 		return -URC_FAILURE;
 	}
 
+	/*
+	 * Check for backtrace on the first instruction of a function. The
+	 * prologue has not been executed yet and the unwinding information is
+	 * not valid. Assume that the return address is in LR.
+	 */
+	if (idx.addr == frame->pc) {
+		frame->pc = frame->lr;
+		return URC_OK;
+	}
+
 	ctrl.vrs[FP] = frame->fp;
 	ctrl.vrs[SP] = frame->sp;
 	ctrl.vrs[LR] = frame->lr;

-- 
Catalin