[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

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

From: Doug Moore
Subject: Re: [Libunwind-devel] use_prev_instr breaking unwind - how to stop it?
Date: Sun, 26 Mar 2017 11:43:02 -0500
User-agent: Internet Messaging Program (IMP) H5 (6.1.7)

Thanks for those links.

You ask
Would unw_init_remote() work for your use?
I tried, but the resulting binary crashed. That may be because unw_init_remote does not include "c->validate = 0;" as does unw_init_local. Even if it did work, I don't think I could stand the performance penalty of using remote instead of local.

In any case, the implementations of local and remote clearly are not compliant with the online documentation for unw_init_local, which states:

From a behavioral point of view, the call:

    ret = unw_init_local(&cursor, &ucontext);

is equivalent to:

    ret = unw_init_remote(&cursor, unw_local_addr_space,

which is untrue because of the different values of use_prev_instr and validate assigned by the two.

The rationale from 2010 includes

This seems almost impossible to run into with local unwinding - it's not possible for unw_getcontext() to have signal trampoline as the caller.

but I am trying local unwinding with a context delivered by a signal handler, not by unw_getcontext, and so stepping, for me, fails as you described in 2010:

2) Failure to trace at function entry; .eh_frame information exists and is correct. I suspect fetch_proc_info() should use "ip", not "--ip" to locate the FDE. In all cases I examined there is no adjacent preceding FDE, so lookup by ip-1 would come up empty.

At the moment, I'm defining a third flavor of unw_init to get me a cursor initialized to let me avoid this problem. I'd rather be staying within the documented libunwind interface. Can libunwind change to help me?


Quoting Lassi Tuura <address@hidden>:

The rationale from back in 2010 is in
plus surrounding thread, with earlier context on unwind failures in

There's also related earlier context in

As I mentioned in the first linked message, it seems improbable that
unw_init_local() could ever be invoked from a signal frame itself. At that
time, your use case of single stepping was mentioned as a counter-example.
Would unw_init_remote() work for your use?

On Fri, Mar 24, 2017 at 11:23 PM, Doug Moore <address@hidden> wrote:

Your change of Apr 24 2010 introduced the second argument to common_init,
and used it to pass 1 (use_prev_instr set) from unw_init_local and 0
(use_prev_instr clear) from unw_init_remote.  I don’t know how those
choices were made, but for my purposes, the choice for unw_init_local was
wrong, and I have to change that decision in the copy of libunwind that
we’ll use here.  I’d welcome an explanation of that decision.


On Mar 24, 2017, at 11:37 AM, Lassi Tuura <address@hidden> wrote:

What platform are you on? fetch_proc_info() should auto-detect signal
frame and set use_prev_instr accordingly using the (kernel-provided) frame
annotations detected in parse_cie(). I don't know if that works for
anything but linux (and possibly x86_64 only at that). Looks like there may
be some support for freebsd.

On Thu, Mar 23, 2017 at 8:45 PM, Doug Moore <address@hidden> wrote:

When unw_init_local uses common_init to set use_prev_instr, and then
fetch_proc_info decrements ip, based on use_prev_instr, unw_step fails for

I’m putting a temp break point at the start of _dl_runtime_resolve, and
when gdb gets there, I’m sending a signal to trigger a signal handler that
invokes unw_step.  What happens, sadly, is that the instruction pointer is
decremented, so the eh_frames search doesn’t find the right frame, so the
step fails.

If only I could clear that use_prev_instr field.  So far, I don’t have
any better choices than to hack the libunwind source and maintain my own
local branch.  Are there any better ideas?

Libunwind-devel mailing list


reply via email to

[Prev in Thread] Current Thread [Next in Thread]