Fix for apr_global_mutex_create() crashes with mod_security 2.9.2

[blappm]

• Call modsecurity_init() for the first invocation too.
• Provide names/files for apr_global_mutex_create() to prevent segfaults in some cases:

[Thu Nov 8 10:58:40 2018] pid 16679 mod_backtrace main() is at 7f949bb6bd30
/usr/lib64/apache2/mod_backtrace.so(+0xddd)[0x7f949bc77ddd]
/usr/sbin/httpd2-prefork(ap_run_fatal_exception+0x4a)[0x7f949bb8b0ba]
/usr/sbin/httpd2-prefork(+0x4f446)[0x7f949bb8b446]
/lib64/libpthread.so.0(+0xf850)[0x7f949a46c850]
/usr/lib64/libapr-1.so.0(apr_global_mutex_lock+0x11)[0x7f949aad6a31]
/usr/lib64/apache2/mod_security2.so(+0x7155e)[0x7f949365655e]
/usr/lib64/apache2/mod_security2.so(+0x72213)[0x7f9493657213]
/usr/lib64/apache2/mod_security2.so(+0x75981)[0x7f949365a981]
/usr/lib64/apache2/mod_security2.so(+0x77ff0)[0x7f949365cff0]
/usr/lib64/apache2/mod_security2.so(+0x7a566)[0x7f949365f566]
/usr/lib64/apache2/mod_security2.so(+0x7bbb7)[0x7f9493660bb7]
/usr/lib64/apache2/mod_security2.so(+0x5be57)[0x7f9493640e57]
/usr/lib64/apache2/mod_security2.so(+0x5b8df)[0x7f94936408df]
/usr/sbin/httpd2-prefork(ap_run_post_read_request+0x4a)[0x7f949bb72cfa]
/usr/sbin/httpd2-prefork(ap_read_request+0xd55)[0x7f949bb75f35]
/usr/sbin/httpd2-prefork(+0x52250)[0x7f949bb8e250]
/usr/sbin/httpd2-prefork(ap_run_process_connection+0x83)[0x7f949bb89d33]
/usr/sbin/httpd2-prefork(+0x5a8de)[0x7f949bb968de]
/usr/sbin/httpd2-prefork(+0x5abca)[0x7f949bb96bca]
/usr/sbin/httpd2-prefork(ap_mpm_run+0x425)[0x7f949bb97055]

[zimmerle]

Hi @blappm,

Can you give more details on the crash and on your env?

[blappm]

Yes, of course.

The crashes happend on Suse Linux 12, SLES12, SP3 and SP4, but on different Apache versions (2.12 and 2.34). We've seen it on apache installed on ~10 systems, while 20 systems were running fine, maybe because there was less traffic. The special thing here is that apache-prefork from Suse is compiled with libpthread.so, even if it is not running threaded. This may result in the wrong locking mechanism beeing used, since flock/fcntl are not used because the filename in apr_global_mutex_create() is left empty.

There are other people reporting the similar crashes:

#564

From the HA-Proxy Archive I've also found the comment below, but I think the patch
is wrong for apache prefork, it only works for the threaded apache-worker:

https://www.mail-archive.com/[email protected]/msg25681.html

`Modsecurity bugs:

• When the audit_log is used with the directive "SecAuditLogType Serial", in
  some systems, the APR mutex initialisation silently fails, this causes a
  segmentation fault. For my own usage, I have a patched version of modsec where
  I use another mutex than "APR_LOCK_DEFAULT" like "APR_LOCK_PROC_PTHREAD"

  • rc = apr_global_mutex_create(&msce->auditlog_lock, NULL, APR_LOCK_DEFAULT, mp);
  • rc = apr_global_mutex_create(&msce->auditlog_lock, NULL, APR_LOCK_PROC_PTHREAD, mp);

`

[zimmerle]

Hum.. I've got the change to use the tmp-file-names, but what about this change?

If i recall correctly this will be executed twice: while checking the configuration fro syntax error and while effective loading it. We may want to init ModSec only at the second one. Right?

[blappm]

I was not sure about that. What about moving it then just after "first_time = 1;" ? Then it will really only be called once at the beginning.

[zimmerle]

Not sure, we may not need to change it at all. I've created this branch v2/dev/pull_1957. It contains all the modifications except this one.

[zimmerle]

Hi @blappm,

Thank you for the detailed explanation. I've made one more question along the code, about modsecurity_init(modsecurity, mp) being moved outside an else.

[zimmerle]

merged ;)

[emphazer]

finally! thx a lot for this one. we noticed that issue several times the past years.