Re: [Libunwind-devel] Libunwind support for NULL IP

[Prabhat Verma]

Hello All,

Special thanks to Arun for all the help. 

I incorrectly assumed that unw_init_local makes a cursor out of the passed 
context and returns it in the output buffer. But I see that modifying the 
registers in the cursor (using unw_set_reg) results in the context being 
modified too. 

Is this the expected behavior? If not, why and under what circumstances can 
this happen? Or is it that the context and the generated local cursor is 
essentially the same memory?

I am asking this because in our codebase we provide a const guarantee for the 
passed context but I have to cast it away to get unw_init_local to work in the 
face of bad IP (the original post). For now, as a workaround, I created a 
container to hold a copy of the passed context and use this copy instead. I 
will fire up gdb and dig into libunwind codebase when I get the chance but am 
hoping someone will provide me with the answer quicker. 

All help highly appreciated. 

Regards,
Prabhat

-----Original Message-----
From: Prabhat Verma 
Sent: Thursday, April 05, 2012 5:24 PM
To: 'Arun Sharma'
Cc: address@hidden
Subject: RE: [Libunwind-devel] Libunwind support for NULL IP

Great suggestion and it works like a charm!
Inside my SEGV handler: 
ucontext_t* temp_uc = reinterpret_cast<ucontext_t *> (<the context passed by 
the OS>); size_t rsp(*reinterpret_cast<size_t *> 
(temp_uc->uc_mcontext.gregs[REG_RSP]));
check if rsp is valid and if it is, let libunwind do the heavy lifting... 

On a side note, you may disagree, but IMO there is value in moving this inside 
of libunwind. Many uses still rely on function pointers for state -machines and 
this will be a useful capability when things go wild....

Thanks a ton Arun and apologies for the trouble :) 

Regards,
Prabhat


-----Original Message-----
From: address@hidden [mailto:address@hidden On Behalf Of Arun Sharma
Sent: Thursday, April 05, 2012 12:04 PM
To: Prabhat Verma
Cc: address@hidden
Subject: Re: [Libunwind-devel] Libunwind support for NULL IP

On Wed, Apr 4, 2012 at 8:45 PM, Arun Sharma <address@hidden> wrote:

> If you want all of this to work without -fno-omit-frame-pointer, 
> things become a bit more complicated. We'll somehow have to
> (heuristically?) unwind up to the point where IP is valid. libunwind 
> knows what to do after that.

Here's one way to make it work:

In the signal handler, you can compare the faulting address to %rip and infer 
that this was a SIGSEGV caused by a bad function pointer.
Then you can special case this (i.e.  unwind the inner most frame "manually" -- 
by changing %rip in the  ucontext to the return address pointed to by %rsp). 
libunwind should be able to do the rest of the work for you.

Why not have this code inside of libunwind? The interfaces do so would be icky 
and this is best done in the signal handler, rather than another library.

 -Arun