Re: [Libunwind-devel] [PATCH 6/6] Fixes for interation end conditions of non-dwarf walker on x86_64.

[Konstantin Belousov]

On Wed, May 16, 2012 at 08:28:35AM -0700, Arun Sharma wrote:
> On Tue, May 15, 2012 at 3:53 AM, Konstantin Belousov
> <address@hidden> wrote:
> 
> >> I applied only the first part of this patch.
> > It seems that 1a6ea3da607d35c03733b9db26b64b71cdbfcfa7 was applied as a 
> > whole.
> 
> I forgot to specify the -a flag to git commit --amend.
> 
> Reverted the second half now (although with an incorrect commit
> message and savannah doesn't let me amend). dwarf_get() is different
> from dwarf_step().
> 
> >
> > I do not understand your note about dwarf_get() returning 0. Isn't 
> > dwarf_get()
> > API is defined to return 0 on success and (non-zero) UNW_ error on failure ?
> 
> Old code was passing through the return value of dwarf_get() (ret=0 =>
> success) and the caller was interpreting it as (ret=0 => stop
> unwinding).
Yes. And this is exactly the case with syscall stub unwinding on FreeBSD.
Currently, libc-generated stubs do not provide dwarf unwind info, so
dwarf_step() returns -UNW_ENOINFO and special step code is activated.

FreeBSD Gos-freebsd.c unw_handle_signal_frame() does record non-null
location for %rip in the cursor->dwarf.loc[RIP], see Gos-freebsd.c:140.

Then, the success of dwarf_get() causes unwinding stop on the first frame
for a thread that is inside syscall.

I.e., with change reverted, I get:
sirion% ./pstack $$                             ~/build/bsd/pstack/build-x86_64
2210: /usr/local/bin/zsh (osrel 900505)
Thread 100138:
 0x8014d167c _sigsuspend+0xc

while with it still applied
sirion% ./pstack $$                             ~/build/bsd/pstack/build-x86_64
2210: /usr/local/bin/zsh (osrel 900505)
Thread 100138:
 0x8014d167c _sigsuspend+0xc
 0x463478 signal_suspend+0x48
 0x43b669 scanjobs+0x3d9
 0x43b768 waitjobs+0x28
 0x4243f3 execstring+0x3ca3
 0x424e09 execlist+0x489
 0x42519b execode+0x9b
 0x436222 loop+0xc2
 0x43747d zsh_main+0x65d
 0x40e0ee _start+0x8e

(pstack is pstack(1) that utilizes libunwind).
> 
> test-ptrace doesn't fail everywhere as a result of your patch. It
> failed for me on Ubuntu 12.04, but passed on Centos5.
Quite strange, I do not have access to any Linux machine to debug.

> 
> This requires a bit more debug. Is this change solving a real world
> problem for you?
> 
>  -Arun

pgpx8tspsmjSx.pgp
Description: PGP signature