Re: Segfault in __c_f_f_c during strace of nptl application.

[Mike Frysinger <vapier@gentoo.org>]

[-- Attachment #1: Type: Text/Plain, Size: 1778 bytes --]

On Wednesday 09 September 2009 10:39:30 Carlos O'Donell wrote:
> On Sun, Jun 21, 2009 at 11:20 AM, John David
> 
> Anglin<dave@hiauly1.hia.nrc.ca> wrote:
> > Looking at my email archive, I see the real cause involves kernel memory
> > maps:
> >
> >  > > > On Wed, May 06, 2009 at 01:39:49PM -0400, John David Anglin wrote:
> >  > > > > > The tombstone is:
> >  > > > > >
> >  > > > > > do_page_fault() pid=10205 command='strace' type=15
> > address=0x407d2f18 > > > > > vm_start = 0x4068d000, vm_end = 0x4068f000
> >  > > > >
> >  > > > > So, the pointer passed to __canonicalize_funcptr_for_compare is
> > outside > > > > the vm range.
> 
> The strace problem is a compiler flaw.
> 
> This is the problem:
> * Strace examines the applications syscall.
> * Strace extracts, via PTRACE, application addresses, addresses that
> don't exist in the strace address space (and should not exist).
> * Strace compares extracted address to a constant SIG_ERR.
> * Compiler generates a call to __c_f_f_c, which dereference the
> extracted address and strace faults.
> 
> Strace and the application have completely different address spaces,
> and __c_f_f_c can't assume that an address is in the current address
> space.
> 
> The solution is to detect that a comparison between two pointers is a
> comparison between pointer and small constant, and avoid calling
> __c_f_f_c for both.
> 
> The workaround is to cast both long. I tested this and it works. I'll
> submit this to debian as the fix.

and include this explanation in a comment right above the cast ? ;)

also, the strace list is active currently, so posting a patch there should get 
it merged (and since it's an important bugfix, it should get added before the 
next release).
-mike

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 836 bytes --]