Re: [Libunwind-devel] use_prev_instr breaking unwind - how to stop it?

[Doug Moore]

Can’t we fix this whole problem by having getcontext subtract from the pc to 
start with, so that nobody ever has to decrement the pc on the first call to 
unw_step?

Doug



> On Mar 29, 2017, at 3:48 PM, Doug Moore <address@hidden> wrote:
> 
> For linux, it’s not using the ip directly to determine is_signal_frame, but 
> it’s using information from processing dwarf data, which it can only find if 
> it has the right ip in the first place.
> 
> It seems like the init_cursor caller should be passing an argument to declare 
> whether this is a signal handler frame or not.  I’d expect her to know that.  
> It’s a lot more reliable that having libunwind try to figure it out.
> 
> Doug
> 
> 
>> On Mar 29, 2017, at 3:33 PM, Doug Moore <address@hidden> wrote:
>> 
>>> But I think we need to know the IP to detect signal frames? 
>> 
>> 
>> We need to know an IP, yes.  It’s not clear that we need the 
>> decremented-by-1 IP, though.  Especially since we’re deciding whether to 
>> decrement it based on whether it’s in a signal frame or not.
>> 
>> I looked at the x86_64 implementations of unw_is_signal_frame for freebsd 
>> and linux.
>> 
>> For freebsd, unw_is_signal_frame uses the (unmodified) ip in the cursor 
>> argument to determine whether this is a signal frame.  So, for freebsd, we 
>> could call unw_is_signal_frame early, passing the cursor we had already, and 
>> use that to determine whether or not to use ip or ip-1 for searching 
>> dwarf/eh_frame structures.
>> 
>> For linux, unw_is_signal_frame does not look at the ip; it returns a value 
>> under the assumption that another function, either tdep_fetch_frame or 
>> tdep_cache_frame, has already set that value.  And tdep_fetch_frame does not 
>> look at the ip field of the cursor passed to it either.  Really, I’m not 
>> sure yet where the ip gets examined in the linux version of 
>> unw_is_signal_frame.
>> 
>> Doug
>> 
>> 
>>> On Mar 29, 2017, at 10:12 AM, Dave Watson <address@hidden> wrote:
>>> 
>>> On 03/27/17 05:36 PM, Doug Moore wrote:
>>>> It looks like the call to tdep_fetch_frame in fetch_proc_info would 
>>>> determine whether or not this was a signal frame, and thus whether or not 
>>>> we needed to use the previous instruction for the lookup.
>>>> 
>>>> Of course, we call tdep_fetch_frame only after we’ve already tried that 
>>>> lookup.  But perhaps some of the work done to identify signal frames could 
>>>> be moved up a bit, so that a better decision could be made about whether 
>>>> or not to use the previous instruction.
>>>> 
>>>> Does that make any sense?
>>> 
>>> But I think we need to know the IP to detect signal frames? 
>>> 
>>> My thought was that if IP-1 lookup failed, we could try IP lookup, and
>>> if it was a signal frame, use that instead (probably with verify=1)
>>> 
>>> !DSPAM:10223,58dbcee539347198113515!
>>> 
>>> 
>> 
>