Invalid program counters and unwinding

Florian Weimer fweimer@redhat.com
Tue Jun 26 11:35:00 GMT 2018


On 06/26/2018 01:25 PM, Jakub Jelinek wrote:
> On Tue, Jun 26, 2018 at 01:01:06PM +0200, Florian Weimer wrote:
>> On 06/26/2018 12:56 PM, Nathan Sidwell wrote:
>>> On 06/26/2018 05:26 AM, Florian Weimer wrote:
>>>
>>>> So it looks to me that the caller of _Unwind_Find_FDE needs to
>>>> ensure that the PC is a valid element of the call stack.  Is this a
>>>> correct assumption?
>>>
>>> I thought this was an (implicit?) requirement. You're unwinding a stack
>>> to deliver an exception up it.  Are there use cases where that is not
>>> the case?
>>
>> We have something approaching this scenario.
>>
>> pthread_cancel in glibc unwinds the stack using DWARF information until
>> encounters a frame without unwind information, when it switches to longjmp
>> to get past that obstacle.
>>
>> However, at the point of transition from a valid DWARF frame into the
>> wilderness (without unwind data), we should still have accurate information
>> on the caller's PC, so _Unwind_Find_FDE will reliably fail to find any
>> unwind data for it.  It's not a random pointer somewhere else, so I think
>> even the pthread_cancel case is fully supported.
> 
> The usual ways to get bogus PCs in the frames is:
> 1) stack corruption
> 2) setcontext/swapcontext with uninitialized or corrupted ucontext_t
> 3) bogus unwind info (compiler or linker etc. bug)

But if that happens, all bets are off, and we could still get a crash 
with the current implementation.

And any approach which does not inhibit concurrent dlclose will only 
make things worse if there are such concurrent calls, which is perhaps 
an unusual combination.

> At least for unwinding, I think we don't and shouldn't care, we assume only
> valid programs.  For cases like _Unwind_Backtrace when used to print info in
> case of fatal signal or stack corruption, it is more questionable, but at
> least the current implmentation doesn't care either.

At least glibc no longer tries to print a backtrace from a corrupted stack.

Thanks,
Florian



More information about the Gcc mailing list