[llvm-dev] Implementing stack probes

Fri Jul 1 02:51:17 PDT 2016

In our case we have to implement our libc as this is for an embedded multicore device, and this why I need to understand the semantics of the mechanism.  What the compiler inserts and what the runtime library support expects have to work together and I do not understand the mechanism in order to implement them.

However, from what you say, the “canary” is just a value (could be anything) that is planted in the function’s own stack frame, and on return if it is altered then the check decides that a stack overflow occurred and aborts.

But this means that a stack overflow that corrupts an adjacent data region which will be visited at another time will not be detected unless by coincidence something writes to that region during the lifetime of the function.  It also means that if the function’s own stack is corrupted by overlapping writes, then its own return address may also be clobbered along with the canary, and there is no guarantee it will resume control at all.

This seems like a very error prone approach.  Or am I misunderstanding the semantics?

In a simple runtime context like ours, the stack is a simple area of memory.  We have a programmer specified stack begin address, and the programmer also determines how much stack to allocate to the program (specified at link-time).

What I would really like to do, is test if there is enough room on the stack before I use the reservation, and abort if there is not.  Our stack grows from high address to low, so currently the function prologue does something like:

SP -= #bytes to reserve

execute code

SP += #bytes to reserve

return

I would like:

SP -= #bytes to reserve

OPTIONAL: update SP high-water mark

IF: SP < __top_of_stack THEN abort

execute code

SP += #bytes to reserve

return

The symbol ‘__top_of_stack’ can be provided by the linker when laying out memory for the program.  This approach would work well as there is no chance that the function’s stack frame will clobber data as the test and abort happen before the reserved stack is used.

I don’t see how the ‘__stack_chk_’ approach can do this, or ensure that the frame does not corrupt adjacent memory.

Thanks,

            MartinO

From: David Majnemer [mailto:david.majnemer at gmail.com] 
Sent: 30 June 2016 18:00
To: Martin J. O'Riordan <martin.oriordan at movidius.com>
Cc: LLVM Developers <llvm-dev at lists.llvm.org>
Subject: Re: [llvm-dev] Implementing stack probes

On Thu, Jun 30, 2016 at 8:29 AM, Martin J. O'Riordan via llvm-dev <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> > wrote:

I am trying to implement stack probes for our SHAVE target, and I see that the compiler injects references to ‘__stack_chk_guard’ and ‘__stack_chk_fail’.  The code that gets generated is horribly wrong, but in order to understand how to fix it I was wondering if there is a clear statement of how the mechanism is supposed to work?

__stack_chk_guard is loaded and the resulting value is stored on the stack.  The location on the stack contains the "canary".  Before the function returns, the canary and __stack_chk_guard are compared again.  If they compare unequal, __stack_chk_fail is called.  Typically, the implementation of __stack_chk_fail is expected to abort the program.

I believe your libc is responsible for implementing these symbols. 

The variable ‘__stack_chk_guard’ appears to be a pointer to an unsigned integer.  Where is this supposed to reside, and what value should it contain?  And the function ‘__stack_chk_fail’ is called when the test fails - presumably this just aborts.

We have done nothing to support these hooks, so the junk instructions that come out are unsurprising.  I am presuming that we need to handle their lowering in a target specific way, but before I can do that I am wondering if there is a clear definition of the semantics of how the probe is supposed to work?

Thanks,

            MartinO

_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> 
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160701/ba06ffb6/attachment.html>