retitle 641470 psi3: segfault in testsuite severity 641470 serious tags 641470 +patch thanks
Hi, On Tue, Sep 13, 2011 at 06:10:59PM +0200, Georg wrote: > Package: psi3 > Version: 3.4.0-2 > Severity: important > > The contents of the input file, corresponding to the example in the user > manual: > > % 6-31G** H2O Test optimization calculation > psi: ( > label = "6-31G** SCF H2O" > jobtype = opt > wfn = ccsd > reference = rhf > dertype = first > basis = "6-31G**" > num_threads = 4 > memory = (2048 MB) > zmat = ( > o > h 1 roh > h 1 roh 2 ahoh > ) > zvars = ( > roh 0.96031231 > ahoh 104.09437511 > ) > ) > > The console messages: > > The PSI3 Execution Driver > PSI3 will perform a RHF CCSD optimisation via analytic gradients. > The following programs will be executed: > input > REPEAT 40 > cints > cscf > transqt2 > ccsort > ccenergy > cchbar > cclambda > ccdensity > oeprop > transqt --backtr > cints --deriv1 > optking --opt_step > END > psiclean > > input > REPEAT 40 > CYCLE 1 > cints > cscf > transqt2 > ccsort > ccenergy > cchbar > cclambda > ccdensity > oeprop > transqt --backtr > cints --deriv1 > optking --opt_step > > Command optking --opt_step was terminated with signal 11 I had another look at psi3 yesterday, trying to increase test suite coverage to CCSD(T) (test case cc13a) and CASSCF (test case casscf-opt) analytical gradients, when I noticed that the test suite already dies in the cc1 test case for the current package, see e.g.: https://buildd.debian.org/status/fetch.php?pkg=psicode&arch=i386&ver=3.4.0-3&stamp=1332980543 |...Testing cc1... |make[3]: Entering directory |`/build/buildd-psicode_3.4.0-3-i386-A_7Xen/psicode-3.4.0/tests/cc1' |Psi3 failed! |make[3]: *** [cc1.test] Error 1 |make[3]: Leaving directory |`/build/buildd-psicode_3.4.0-3-i386-A_7Xen/psicode-3.4.0/tests/cc1' |make[2]: *** [quicktests] Error 1 The reason appears to be the same as with this bug: if a longer sequence of commands is run, a segfault will happen eventually. After much analyzing and debugging, I finally found the problem today: the Debian-specific patch 07_464867_move_executables.patch adds "/usr/lib/psi" to the environment so the sub-commands to be executed are found. However, this is done every time a sub-command is run and eventually the maximum environment variable length is reached, at which point other environment variables get overwritten. Once $HOME gets overwritten (and thus removed from the environment), the configuration file handling in ~/.psirc in src/lib/libciomr/psi_start.cc and src/lib/libpsio/ipv1_config.cc segfaults due to this code: userhome = getenv("HOME"); cfname = (char *) malloc((10+strlen(userhome))*sizeof(char)); Probably psi3 should check for the return value of getenv, but certainly we should not overwrite the environment. The attached patch to 07_464867_move_executables.patch fixes this problem by checking for the presence of /usr/lib/psi before adding it to environment variable. Both the test suite (cc1/cc13a) and the provided example in this bug are running fine for me with this patch. Cheers, Michael
Index: patches/07_464867_move_executables.patch =================================================================== --- patches/07_464867_move_executables.patch (Revision 4169) +++ patches/07_464867_move_executables.patch (Arbeitskopie) @@ -3,16 +3,20 @@ are found. Bug-Debian: http://bugs.debian.org/464867 ---- psicode-3.4.0.orig/src/bin/psi3/runcmd.cc -+++ psicode-3.4.0/src/bin/psi3/runcmd.cc -@@ -14,6 +14,11 @@ +Index: psicode-3.4.0/src/bin/psi3/runcmd.cc +=================================================================== +--- psicode-3.4.0.orig/src/bin/psi3/runcmd.cc 2013-04-08 13:53:13.446693968 +0200 ++++ psicode-3.4.0/src/bin/psi3/runcmd.cc 2013-04-08 13:54:54.059211598 +0200 +@@ -14,6 +14,13 @@ void runcmd(int *errcod, char *cmd) { + /* Debian places the helper executables into /usr/lib/psi */ + char *path = getenv("PATH"); -+ strcat(path, ":/usr/lib/psi"); -+ putenv(path); ++ if (strstr(path, "/usr/lib/psi") == NULL) { ++ strcat(path, ":/usr/lib/psi"); ++ putenv(path); ++ } + /* i = system(cmd);
