Kerberos 5 breakage.

It appears that compiling with Kerberos 5 support turned on against MIT
Kerberos 1.0.5 produces some breakage.

Anyone seen this?

(I'm keen to use Kerberos 5 as its a deployed company wide and Sybase does
not support it so it would make PostgreSQL look really good if this
worked.)

Last I heard, Kerberos worked, but that was a while ago. Feel free to
send in some patches.

-- 
Bruce Momjian                          |  830 Blythe Avenue
maillist@candle.pha.pa.us              |  Drexel Hill, Pennsylvania 19026
  +  If your life is a hard drive,     |  (610) 353-9879(w)
  +  Christ can be your backup.        |  (610) 853-3000(h)

On Tue, 19 May 1998, Bruce Momjian wrote:

Last I heard, Kerberos worked, but that was a while ago. Feel free to
send in some patches.

I've compiled with kerberos 4 compatibility mode libraries in kerberos 5
and it appears to compile, link and run but I've not got a good testbed
for kerberos 4.

While Kerberos 5 authentication and authorization is nice, I'd like to
investigate the possibility of adding encryption as well.

I've got to complete the setup of a test enviornment for this before I can
start on the code.

/*
Matthew N. Dodd | A memory retaining a love you had for life
winter@jurai.net | As cruel as it seems nothing ever seems to
http://www.jurai.net/~winter | go right - FLA M 3.1:53
*/

"Matthew N. Dodd" <winter@jurai.net> writes:

I've compiled with kerberos 4 compatibility mode libraries in kerberos 5
and it appears to compile, link and run but I've not got a good testbed
for kerberos 4.

The Kerberos 4 stuff works fine with real Kerberos 4 libraries.

While Kerberos 5 authentication and authorization is nice, I'd like to
investigate the possibility of adding encryption as well.

Absolutely. This should be specified in the pg_hba.conf file, so that
you could demand Kerberos authentication plus encryption for sensitive
data. When not demanded by pg_hba.conf, it should be a client option.

-tih
--
Popularity is the hallmark of mediocrity. --Niles Crane, "Frasier"

On 20 May 1998, Tom Ivar Helbekkmo wrote:

While Kerberos 5 authentication and authorization is nice, I'd like to
investigate the possibility of adding encryption as well.

Absolutely. This should be specified in the pg_hba.conf file, so that
you could demand Kerberos authentication plus encryption for sensitive
data. When not demanded by pg_hba.conf, it should be a client option.

I read through the SSL patch and am convinced that we need a little more
coherent arrangment of interface methods. Allowing direct manipulation of
the file descriptors is really going to make adding stuff like this (SSL,
Kerb5 encryption etc) next to impossible.

Take a look at Apache 1.2 vx. 1.3 for an idea of what I'm talking about.

Also, allowing writes of single characters is bad; you incur a context
switch each write. The client and server should be writing things into
largish buffers and writing those instead of doing small writes.

The existence of the following scare me...

pqPutShort(int integer, FILE *f)
pqPutLong(int integer, FILE *f)
pqGetShort(int *result, FILE *f)
pqGetLong(int *result, FILE *f)
pqGetNBytes(char *s, size_t len, FILE *f)
pqPutNBytes(const char *s, size_t len, FILE *f)
pqGetString(char *s, size_t len, FILE *f)
pqPutString(const char *s, FILE *f)
pqGetByte(FILE *f)
pqPutByte(int c, FILE *f)

(from src/backend/libpq/pqcomprim.c)

A select based I/O buffering system would seem to be in order here...

I'd like to see these routines passing around a connection information
struct that contains the file handle and other connection options as well.

I'll not bother beating on this anymore as I'm unlikely to cover anything
that has not already been covered. Regardless, this issue needs some
critical analysis before any code is changed.

Failing to address this issue really raises the cost of adding stuff like
SSL and Kerberos5 encryption.

Take a look at src/main/buff.c and src/include/buff.h in Apache 1.3 at how
they use their 'struct buff_struct' for some interesting examples.

/*
Matthew N. Dodd | A memory retaining a love you had for life
winter@jurai.net | As cruel as it seems nothing ever seems to
http://www.jurai.net/~winter | go right - FLA M 3.1:53
*/

At 11:02 AM -0700 5/20/98, Matthew N. Dodd wrote:

Also, allowing writes of single characters is bad; you incur a context
switch each write. The client and server should be writing things into
largish buffers and writing those instead of doing small writes.

The existence of the following scare me...

pqPutShort(int integer, FILE *f)

.
.
.

Can't these be defined as macros the way get/put stuff is done in stdio.h?

Signature failed Preliminary Design Review.
Feasibility of a new signature is currently being evaluated.
h.b.hotz@jpl.nasa.gov, or hbhotz@oxy.edu

On Wed, 20 May 1998, Henry B. Hotz wrote:

Can't these be defined as macros the way get/put stuff is done in stdio.h?

Which macros?

Looking at stdio.h and the FILE struct/typedef; I wonder if its possible
to override the _read and _write function pointers and sub in our own
depending on when encryption scheme is in use.

/*
Matthew N. Dodd | A memory retaining a love you had for life
winter@jurai.net | As cruel as it seems nothing ever seems to
http://www.jurai.net/~winter | go right - FLA M 3.1:53
*/

"Matthew N. Dodd" <winter@jurai.net> writes:

Also, allowing writes of single characters is bad; you incur a context
switch each write. The client and server should be writing things into
largish buffers and writing those instead of doing small writes.

The existence of the following scare me...
pqPutShort(int integer, FILE *f)
pqPutLong(int integer, FILE *f)
[etc]

Look again. Those functions use <stdio.h>, which provides buffering.
They don't need to do it themselves.

It might be good to put a layer underneath these functions to allow
insertion of encryption or something like that, but efficiency is not
a valid argument for doing it.

On the client side, in the recent libpq rewrite I took out usage of
stdio and did my own buffering instead, but that was just so that
I could control when and how the client would block for input.
I don't think it bought any speedup.

regards, tom lane

At 11:46 AM -0700 5/20/98, Matthew N. Dodd wrote:

On Wed, 20 May 1998, Henry B. Hotz wrote:

Can't these be defined as macros the way get/put stuff is done in stdio.h?

Which macros?

I haven't actually looked, but I think it's pretty standard for
getchar/putchar to just do I/O from some local-to-the-program buffers.
Only when they overflow does it become a real system/library call, but they
look like function calls to the C program.

It's also true that you can play games with incremental linking and symbol
table stripping to insert your own wrapper on a system routine, but I would
not recommend that. It's much too likely to create portability problems.

Signature failed Preliminary Design Review.
Feasibility of a new signature is currently being evaluated.
h.b.hotz@jpl.nasa.gov, or hbhotz@oxy.edu

Originally the JDBC driver had these small writes, because the libpq code
had them. This had the downside that there was a lot of small packets being
sent to the backend, slowing things up a lot.

Anyhow, I've been supplied with a patch that does this buffering, and while
testing this, it does improve things.

I haven't posted it to the patches list yet as another part of the driver is
having some problems at the moment.

This doesn't help with SSL or KB5 though, as Java doesn't have them (yet).

--
Peter T Mount, peter@maidstone.gov.uk, peter@retep.org.uk
JDBC FAQ: http://www.retep.org.uk/postgres
Please note that this is from my works email. If you reply, please cc my
home address.

-----Original Message-----
From: owner-pgsql-hackers@hub.org [mailto:owner-pgsql-hackers@hub.org]On
Behalf Of Matthew N. Dodd
Sent: Wednesday, May 20, 1998 7:22 PM
To: Tom Ivar Helbekkmo
Cc: Bruce Momjian; pgsql-hackers@postgreSQL.org
Subject: Re: [HACKERS] Kerberos 5 breakage.

On 20 May 1998, Tom Ivar Helbekkmo wrote:

While Kerberos 5 authentication and authorization is nice, I'd like to
investigate the possibility of adding encryption as well.

Absolutely. This should be specified in the pg_hba.conf file, so that
you could demand Kerberos authentication plus encryption for sensitive
data. When not demanded by pg_hba.conf, it should be a client option.

I read through the SSL patch and am convinced that we need a little more
coherent arrangment of interface methods. Allowing direct manipulation of
the file descriptors is really going to make adding stuff like this (SSL,
Kerb5 encryption etc) next to impossible.

Take a look at Apache 1.2 vx. 1.3 for an idea of what I'm talking about.

Also, allowing writes of single characters is bad; you incur a context
switch each write. The client and server should be writing things into
largish buffers and writing those instead of doing small writes.

The existence of the following scare me...

pqPutShort(int integer, FILE *f)
pqPutLong(int integer, FILE *f)
pqGetShort(int *result, FILE *f)
pqGetLong(int *result, FILE *f)
pqGetNBytes(char *s, size_t len, FILE *f)
pqPutNBytes(const char *s, size_t len, FILE *f)
pqGetString(char *s, size_t len, FILE *f)
pqPutString(const char *s, FILE *f)
pqGetByte(FILE *f)
pqPutByte(int c, FILE *f)

(from src/backend/libpq/pqcomprim.c)

A select based I/O buffering system would seem to be in order here...

I'd like to see these routines passing around a connection information
struct that contains the file handle and other connection options as well.

I'll not bother beating on this anymore as I'm unlikely to cover anything
that has not already been covered. Regardless, this issue needs some
critical analysis before any code is changed.

Failing to address this issue really raises the cost of adding stuff like
SSL and Kerberos5 encryption.

Take a look at src/main/buff.c and src/include/buff.h in Apache 1.3 at how
they use their 'struct buff_struct' for some interesting examples.

/*
Matthew N. Dodd | A memory retaining a love you had for life
winter@jurai.net | As cruel as it seems nothing ever seems to
http://www.jurai.net/~winter | go right - FLA M 3.1:53
*/