Skip to contentSkip to conversation

ARM : unwinder : Prevent data abort due to stack overflow

PATCH Dave Martin 1 participant 1 message Jan 7, 2014

lore
Share & tools
lore
#0 lore raw
Quoted 25 lines · Anurag Aggarwal
Quoted 6 lines

You could try adding some debug printks to see how the backtrace fails. You could also try adding a few hand-crafted assembler functions with appropriate code and unwind directives to trigger different kinds of backtrace failure. You might have to add a way to artificially limit sp_high to check the cases where you run out of stack in the middle of popping multiple registers.

I added a a printk statement

(quoted diff) +5 -0
+               if (*vsp >= (unsigned long *)ctrl->sp_high) {
+                       printk(KERN_ERR "Stack Overflow Detected, vsp = %lx",
+                               (unsigned long)*vsp);
+                       return -URC_FAILURE;
+               }

I ran a many test cases to try and get the above print in the dmesg log.

I tried the following things :

1) Calling unwind_backtrace from diffrenet locations in the kernel, I added the unwind call
in some irq, fork, exit and some sysfs entries call.
2) I limited the value of sp_high in unwind_frame() itself, I tried many values of sp_high,
varrying from (low + sizeof(ctrl.vrs)/4) to (low + 4*sizeof(ctrl.vrs)).

Quoted 8 lines

When running the above cases I was able to see the above printk quiet a few times in dmesg log.

So, the error condition is being handled.

If you have some test cases for verifying the unwinder, please share the same. They might help
in thorough testing of unwinder.

I think that sounds OK to give us reasonable confidence that the code is working correctly.

Go ahead and add my Reviewed-by on the patch, if you're still waiting for it.

Cheers
---Dave