Kerberos authentication, Active Directory, and PostgreSQL

Turner, Ian wrote:

While trying to connect our PostgreSQL database to our Kerberos realm, we encountered the obscure message "Invalid message length". Tracking this down, we discovered that it was emitted by src/backend/libpq/pqcomm.c in response to a rather large Kerberos message. The root cause is as follows, and a patch is below.

The code in src/backend/libpq/auth.c contains a hard-coded limit on the size of GSS messages, and in particular on the message containing the client's Kerberos ticket for the postgres server. The limit was 2,000 bytes, which is normally adequate for tickets based on TGTs issued by Unix KDCs. However, TGTs issued by Windows domain controllers contain an authorization field known as the PAC (privilege attribute certificate), which contains the user's Windows permissions (group memberships etc.). The PAC is copied into all tickets obtained on the basis of this TGT (even those issued by Unix realms which the Windows realm trusts), and can be several K in size. Thus, GSS authentication was failing with a "invalid message length" error. We simply upped the limit to 32k, which ought to be sufficient.

The patch is quite brief:

--- postgresql-8.4-8.4.1/src/backend/libpq/auth.c       2009-06-25 12:30:08.000000000 +0100
+++ postgresql-8.4-8.4.1-fixed/src/backend/libpq/auth.c 2009-09-15 20:27:01.000000000 +0100
@@ -166,6 +166,8 @@
#endif

static int     pg_GSS_recvauth(Port *port);
+
+#define GSS_MAX_TOKEN_LENGTH (32767)
#endif   /* ENABLE_GSS */

@@ -937,7 +939,7 @@

/* Get the actual GSS token */
initStringInfo(&buf);
-               if (pq_getmessage(&buf, 2000))
+               if (pq_getmessage(&buf, GSS_MAX_TOKEN_LENGTH))
{
/* EOF - pq_getmessage already logged error */
pfree(buf.data);

Please let me know if anything additional is required in order to get this fix into the next release.

The corresponding limit in pg_SSPI_recvauth() probably needs to be
raised too..

pq_getmessage() doesn't necessarily need a limit, we could accept
arbitrarily long tokens. Although I guess we want to avoid simple
denial-of-service attacks exhausting backend memory.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com

2009/10/13 Heikki Linnakangas <heikki.linnakangas@enterprisedb.com>:

Turner, Ian wrote:

While trying to connect our PostgreSQL database to our Kerberos realm, we encountered the obscure message "Invalid message length". Tracking this down, we discovered that it was emitted by src/backend/libpq/pqcomm.c in response to a rather large Kerberos message. The root cause is as follows, and a patch is below.

The code in src/backend/libpq/auth.c contains a hard-coded limit on the size of GSS messages, and in particular on the message containing the client's Kerberos ticket for the postgres server. The limit was 2,000 bytes, which is normally adequate for tickets based on TGTs issued by Unix KDCs. However, TGTs issued by Windows domain controllers contain an authorization field known as the PAC (privilege attribute certificate), which contains the user's Windows permissions (group memberships etc.). The PAC is copied into all tickets obtained on the basis of this TGT (even those issued by Unix realms which the Windows realm trusts), and can be several K in size. Thus, GSS authentication was failing with a "invalid message length" error. We simply upped the limit to 32k, which ought to be sufficient.

The patch is quite brief:

--- postgresql-8.4-8.4.1/src/backend/libpq/auth.c       2009-06-25 12:30:08.000000000 +0100
+++ postgresql-8.4-8.4.1-fixed/src/backend/libpq/auth.c 2009-09-15 20:27:01.000000000 +0100
@@ -166,6 +166,8 @@
 #endif

 static int     pg_GSS_recvauth(Port *port);
+
+#define GSS_MAX_TOKEN_LENGTH (32767)
 #endif   /* ENABLE_GSS */

@@ -937,7 +939,7 @@

                /* Get the actual GSS token */
                initStringInfo(&buf);
-               if (pq_getmessage(&buf, 2000))
+               if (pq_getmessage(&buf, GSS_MAX_TOKEN_LENGTH))
                {
                        /* EOF - pq_getmessage already logged error */
                        pfree(buf.data);

Please let me know if anything additional is required in order to get this fix into the next release.

The corresponding limit in pg_SSPI_recvauth() probably needs to be
raised too..

Probably, but ont entirely certainly. Given how SSPI works.

But for consistency that would certainly be a good idea :-)

pq_getmessage() doesn't necessarily need a limit, we could accept
arbitrarily long tokens. Although I guess we want to avoid simple
denial-of-service attacks exhausting backend memory.

Yeah.
FWIW, the default max token size on Win2k is ~8Kb. In some service
pack and then in Win2003, it was increased to 12Kb. But it is possible
to increase that by a registry key on the domain controller - and I
read somewhere that Win2008 actually will increase this size
dynamically.

Actually, I found a note that said it's recommended to never increase
it about 65535 - so perhaps we should put our limit at that instead od
32767?

--
Magnus Hagander
Me: http://www.hagander.net/
Work: http://www.redpill-linpro.com/

On Mon, 2009-10-12 at 20:42 -0400, Turner, Ian wrote:

--- postgresql-8.4-8.4.1/src/backend/libpq/auth.c       2009-06-25 12:30:08.000000000 +0100
+++ postgresql-8.4-8.4.1-fixed/src/backend/libpq/auth.c 2009-09-15 20:27:01.000000000 +0100
@@ -166,6 +166,8 @@
#endif

static int     pg_GSS_recvauth(Port *port);
+
+#define GSS_MAX_TOKEN_LENGTH (32767)
#endif   /* ENABLE_GSS */

A small wish in case we go with this: The constant should be named
something like PG_...; otherwise it looks like we are defining or
overriding an official symbol from the GSS API.

@@ -937,7 +939,7 @@

/* Get the actual GSS token */
initStringInfo(&buf);
-               if (pq_getmessage(&buf, 2000))
+               if (pq_getmessage(&buf, GSS_MAX_TOKEN_LENGTH))
{
/* EOF - pq_getmessage already logged error */
pfree(buf.data);

To wit, the above hunk looks quite misleading in isolation.

Peter Eisentraut <peter_e@gmx.net> writes:

A small wish in case we go with this: The constant should be named
something like PG_...; otherwise it looks like we are defining or
overriding an official symbol from the GSS API.

I'd be inclined to just s/2000/32767/ and not bother with a symbol,
misleadingly named or otherwise. If the value were actually being
used in more than one place, it'd be a different story, but to be
so used it would likely need a completely different name.

regards, tom lane

Peter Eisentraut wrote:

On Mon, 2009-10-12 at 20:42 -0400, Turner, Ian wrote:

--- postgresql-8.4-8.4.1/src/backend/libpq/auth.c       2009-06-25 12:30:08.000000000 +0100
+++ postgresql-8.4-8.4.1-fixed/src/backend/libpq/auth.c 2009-09-15 20:27:01.000000000 +0100
@@ -166,6 +166,8 @@
#endif

static int     pg_GSS_recvauth(Port *port);
+
+#define GSS_MAX_TOKEN_LENGTH (32767)
#endif   /* ENABLE_GSS */

A small wish in case we go with this: The constant should be named
something like PG_...; otherwise it looks like we are defining or
overriding an official symbol from the GSS API.

Agreed. In my first glance at the patch I thought GSS_MAX_TOKEN_LENGTH
was defined in a system include file.

--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://enterprisedb.com

+ If your life is a hard drive, Christ can be your backup. +

Tom Lane:

I'd be inclined to just s/2000/32767/ and not bother with a symbol,

Heikki Linnakangas:

The corresponding limit in pg_SSPI_recvauth() probably needs to be
raised too..

Magnus Hagander:

Actually, I found a note that said it's recommended to never increase
it [above] 65535 - so perhaps we should put our limit at that instead
[of] 32767?

Perhaps the thing to do is to use a symbol, set its value to 65535, and use the same symbol for all of GSS, SSPI, and KRB5.

--Ian

Magnus Hagander wrote:

FWIW, the default max token size on Win2k is ~8Kb. In some service
pack and then in Win2003, it was increased to 12Kb. But it is possible
to increase that by a registry key on the domain controller - and I
read somewhere that Win2008 actually will increase this size
dynamically.

Actually, I found a note that said it's recommended to never increase
it about 65535 - so perhaps we should put our limit at that instead od
32767?

Yeah, setting it at 65535 seems like a good idea then. I'm tempted to
backport this, although it's not strictly speaking a bug fix. Any
objections?

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com

Heikki Linnakangas <heikki.linnakangas@enterprisedb.com> writes:

Magnus Hagander wrote:

Actually, I found a note that said it's recommended to never increase
it about 65535 - so perhaps we should put our limit at that instead od
32767?

Yeah, setting it at 65535 seems like a good idea then. I'm tempted to
backport this, although it's not strictly speaking a bug fix. Any
objections?

Why isn't it a bug fix? +1 for backport ...

regards, tom lane

2009/10/13 Tom Lane <tgl@sss.pgh.pa.us>:

Heikki Linnakangas <heikki.linnakangas@enterprisedb.com> writes:

Magnus Hagander wrote:

Actually, I found a note that said it's recommended to never increase
it about 65535 - so perhaps we should put our limit at that instead od
32767?

Yeah, setting it at 65535 seems like a good idea then. I'm tempted to
backport this, although it's not strictly speaking a bug fix. Any
objections?

Why isn't it a bug fix?  +1 for backport ...

Yeah, +1 there.

--
Magnus Hagander
Me: http://www.hagander.net/
Work: http://www.redpill-linpro.com/

Magnus Hagander wrote:

2009/10/13 Tom Lane <tgl@sss.pgh.pa.us>:

Heikki Linnakangas <heikki.linnakangas@enterprisedb.com> writes:

Magnus Hagander wrote:

Actually, I found a note that said it's recommended to never increase
it about 65535 - so perhaps we should put our limit at that instead od
32767?

Yeah, setting it at 65535 seems like a good idea then. I'm tempted to
backport this, although it's not strictly speaking a bug fix. Any
objections?

Why isn't it a bug fix? +1 for backport ...

Yeah, +1 there.

Ok, committed.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com

Tom Lane wrote:

Peter Eisentraut <peter_e@gmx.net> writes:

A small wish in case we go with this: The constant should be named
something like PG_...; otherwise it looks like we are defining or
overriding an official symbol from the GSS API.

I'd be inclined to just s/2000/32767/ and not bother with a symbol,
misleadingly named or otherwise. If the value were actually being
used in more than one place, it'd be a different story, but to be
so used it would likely need a completely different name.

It is used in two places in that file: pg_GSS_recvauth() and
pg_SSPI_recvauth(). (the original patch neglected SSPI)

I'll rename it to PG_MAX_AUTH_TOKEN_LENGTH, unless someone has a better
suggestion.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com

I'll rename it to PG_MAX_AUTH_TOKEN_LENGTH, unless someone has a better
suggestion.

If we are not changing this for all authentication schemes, then the name should probably reflect that this is for GSS and SSPI only (not even KRB5).

--Ian

"Turner, Ian" <Ian.Turner@deshaw.com> writes:

I'll rename it to PG_MAX_AUTH_TOKEN_LENGTH, unless someone has a better
suggestion.

If we are not changing this for all authentication schemes, then the name should probably reflect that this is for GSS and SSPI only (not even KRB5).

Then we'd have to rename the symbol anytime we applied it to some new
auth scheme. The original naming complaint reflected a concern that
the symbol looked like it was supplied by the system headers, rather
than being of Postgres origin. Heikki's suggestion deals with that,
and I think it's fine as-is.

regards, tom lane

The original naming complaint reflected a concern that
the symbol looked like it was supplied by the system headers, rather
than being of Postgres origin. Heikki's suggestion deals with that,
and I think it's fine as-is.

OK, fine with me.

--Ian