Re: bug#8230: touch dumps core on solaris 10

[Bruno Haible]

Regarding <http://debbugs.gnu.org/cgi/bugreport.cgi?bug=8230>:
Eric Blake wrote:
> If you compiled with debugging, then it might help to show a gdb session
> of rpl_fcntl in lib/fcntl.c while stepping enough to prove the loop, as
> well as a backtrace that leads to that loop, might be worthwhile.

After compiling with debugging on Solaris 9, and running on Solaris 10,
I get this stack trace from the end of the endless recursion:

#3  0x00019e98 in futimens (fd=0, times=0x0) at futimens.c:36
#4  0xff2b6310 in futimesat () from /lib/libc.so.1
#5  0x0001fd38 in fdutimens (fd=0, file=0x0, timespec=0x0) at utimens.c:339
#6  0x00019e98 in futimens (fd=0, times=0x0) at futimens.c:36
#7  0xff2b6310 in futimesat () from /lib/libc.so.1
#8  0x0001fd38 in fdutimens (fd=0, file=0x0, timespec=0x0) at utimens.c:339
#9  0x00019e98 in futimens (fd=0, times=0x0) at futimens.c:36
#10 0xff2b6310 in futimesat () from /lib/libc.so.1

and this one from the beginning:

#0  futimens (fd=0, times=0x0) at futimens.c:36
#1  0xff2b6310 in futimesat () from /lib/libc.so.1
#2  0x0001fd38 in fdutimens (fd=0, file=0x0, timespec=0x0) at utimens.c:339
#3  0x00019e98 in futimens (fd=0, times=0x0) at futimens.c:36
#4  0x00013c6c in fdutimensat (fd=0, dir=-3041965, 
    file=0xffbffd4a <error reading variable>, ts=0x0, atflag=0)
    at fdutimensat.c:48
#5  0x00012be8 in touch (file=0xffbffd4a <error reading variable>)
    at touch.c:168
#6  0x00013b48 in main (argc=2, argv=0xffbffc5c) at touch.c:434

So the situation is this:
  - On Solaris 9, futimesat() exists in libc, but futimens() doesn't.
  - On Solaris 10, both exist, and futimesat() invokes futimens(), but the
    latter is redefined in the executable with code from gnulib.

In gnulib, we have invocations
  fdutimensat -> futimens -> fdutimens -> {futimesat,utimes}
and Solaris 10 libc has an invocation
                 futimens <-------------- futimesat

There is also a second cycle: gnulib has invocations
  utimensat -> utimens -> futimesat
and Solaris 10 libc has an invocation
  utimensat <------------ futimesat

I propose to change gnulib so that it uses
  #define futimens rpl_futimens
  #define utimensat rpl_utimensat
also on Solaris 9 (with the usual idioms for the C++ and the gnulib namespace).
I have verified that when these two macro definitions are added to config.h,
the 'touch' program no longer goes into an infinite recursion.

Bruno
-- 
In memoriam Eulogius of Córdoba 
<http://en.wikipedia.org/wiki/Eulogius_of_Córdoba>