Re: trouble with NGPT + Linux 2.5.40 + glibc 2.2.93

Thanks! Uwe for your findings. I also noticed this on RH8.0.

To fix this, I commented out the libc_internal_tsd_get and
libc_internal_tsd_set routines in pth_tcb.c (everything after line 350).
With that, the initial core dump problem is gone. However,
the program hangs at final exit() for maxnativethreads > 1.
Looks like the LDT part is not working properly with new Glibc
(NATIVE_SELF, etc), so I also commented out that in useldt.h
and use the pth_get_native_descr() directly, which fixes the exit()
problem.
I think we need to look at the glibc 2.3 changes closely and need
to make appropriate changes in NGPT to use new features like TLS, etc.

- - - - -
Saurabh Desai
POSIX Threading for Linux
IBM  Linux Technology Center
e-mail: [email protected] OR [email protected]
phone: 512-838-2655, T/L: 678-2655




"Dr. Uwe Girlich" <[email protected]>@www-124.ibm.com on 10/11/2002
04:05:53 AM

Sent by:    [email protected]


To:    pthreads-devel <[email protected]>
cc:
Subject:    Re: [pthreads-devel] trouble with NGPT + Linux 2.5.40 + glibc
       2.2.93


Hello!

I finally found the cause of the SEGFAULT on RedHat 8.0 during _init() of
NGPT,
which makes NGPT not usable at all under RedHat 8.0:

During our basic init process, we call sysconf() to get the number of
processors in the machine. Besides many other things, this does an sscanf
(),
which is a locale-dependent function. Locales are thread specific, so it
calls
        decimal = _NL_CURRENT (LC_NUMERIC, DECIMAL_POINT);
which checks for __libc_internal_tsd_get. This variable is 0 in the
standard C
library but this weak symbol will be overwritten by our own NGPT
__libc_internal_tsd_get and it points now to libc_internal_tsd_get(), which
will be called. This function uses pth_get_current() and this finally has
no
idea on any current thread, because we are still in the basic init step. So
pth_get_current() returns 0 and libc_internal_tsd_get() returns NULL too
and
_NL_CURRENT() will access an invalid memory area.

We could simply correct this, if we set all entry points like
__libc_internal_tsd_get/libc_internal_tsd_get() to 0 and activate them
only after our basic initialisation. Or we could ignore the number of
processors for now and wait until the first pthread_create().

But also other functions like strtol() are locale dependent and should not
be used during init but NGPT uses this function to parse the environment
variables.

And BTW, the _init function of GLIBC was called already. At least this is
correct.

Bye, Uwe

--
Dr. Uwe Girlich              email: [email protected]
Philosys Software GmbH       www: www.philosys.de
Edisonstrasse 6              phone: +49 89 321407-44
D-85716 Unterschleissheim    fax: +49 89 321407-12

_______________________________________________
pthreads-devel mailing list
[email protected]
http://www-124.ibm.com/developerworks/oss/mailman/listinfo/pthreads-devel