[Heikki Linnakangas <hlinnaka@iki.fi>]

On 05/25/2017 06:36 PM, Michael Paquier wrote:
> On Thu, May 25, 2017 at 10:52 AM, Michael Paquier
> <michael.paquier@gmail.com> wrote:
>> Actually, I don't think that we are completely done here. Using the
>> patch of upthread to enforce a failure on SASLInitialResponse, I see
>> that connecting without SSL causes the following error:
>> psql: FATAL:  password authentication failed for user "mpaquier"
>> But connecting with SSL returns that:
>> psql: duplicate SASL authentication request
>>
>> I have not looked at that in details yet, but it seems to me that we
>> should not take pg_SASL_init() twice in the scram authentication code
>> path in libpq for a single attempt.
>
> Gotcha. This happens because of sslmode=prefer, on which
> pqDropConnection is used to clean up the connection state. So it seems
> to me that the correct fix is to move the cleanup of sasl_state to
> pqDropConnection() instead of closePGconn(). Once I do so the failures
> are correct, showing to the user two FATAL errors because of the two
> attempts:
> psql: FATAL:  password authentication failed for user "mpaquier"
> FATAL:  password authentication failed for user "mpaquier"

Hmm, the SASL state cleanup is done pretty much the same way that 
GSS/SSPI cleanup is. Don't we have a similar problem with GSS?

I tested that, and I got an error from glibc, complaining about a 
double-free:

> *** Error in `/home/heikki/pgsql.master/bin/psql': double free or corruption (out): 0x000056157d13be00 ***
> ======= Backtrace: =========
> /lib/x86_64-linux-gnu/libc.so.6(+0x70bcb)[0x7f6a460b7bcb]
> /lib/x86_64-linux-gnu/libc.so.6(+0x76f96)[0x7f6a460bdf96]
> /lib/x86_64-linux-gnu/libc.so.6(+0x7778e)[0x7f6a460be78e]
> /home/heikki/pgsql.master/lib/libpq.so.5(+0xa1dc)[0x7f6a46d651dc]
> /home/heikki/pgsql.master/lib/libpq.so.5(+0xa4b3)[0x7f6a46d654b3]
> /home/heikki/pgsql.master/lib/libpq.so.5(+0xacb9)[0x7f6a46d65cb9]
> /home/heikki/pgsql.master/lib/libpq.so.5(PQconnectPoll+0x14e7)[0x7f6a46d6ae81]
> /home/heikki/pgsql.master/lib/libpq.so.5(+0xe895)[0x7f6a46d69895]
> /home/heikki/pgsql.master/lib/libpq.so.5(PQconnectdbParams+0x4f)[0x7f6a46d675b9]
> /home/heikki/pgsql.master/bin/psql(+0x3daa9)[0x56157ccafaa9]
> /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf1)[0x7f6a460672b1]
> /home/heikki/pgsql.master/bin/psql(+0x1163a)[0x56157cc8363a]

I bisected that; the culprit was commit 61bf96cab0, where I refactored 
the libpq authentication code in preparation for SCRAM. The logic around 
that free() was always a bit wonky, but the refactoring made it outright 
broken. Attached is a patch for that, see commit message for details. 
(Review of that would be welcome.)

So, after fixing that, back to the original question; don't we have a 
similar "duplicate authentication request" problem with GSS? Yes, turns 
out that we do, even on stable branches:

psql "sslmode=prefer dbname=postgres hostaddr=127.0.0.1 
krbsrvname=postgres  host=localhost user=krbtestuser"
psql: duplicate GSS authentication request

To fix, I suppose we can do what you did for SASL in your patch, and 
move the cleanup of conn->gctx from closePGconn to pgDropConnection. And 
I presume we need to do the same for the SSPI state too, but I don't 
have a Windows set up to test that at the moment.

- Heikki

-- 
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers

Attachments:

0001-Fix-double-free-bug-in-GSS-authentication.patch (invalid/octet-stream, 1.9K) download
From c75b69e777a52520f9688a8926b9d6986c88c8cc Mon Sep 17 00:00:00 2001
From: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Date: Thu, 1 Jun 2017 13:26:06 +0300
Subject: [PATCH 1/1] Fix double-free bug in GSS authentication.

The logic to free the buffer after the gss_init_sec_context() call was
always a bit wonky. Because gss_init_sec_context() sets the GSS context
variable, conn->gctx, we would in fact always attempt to free the buffer.
That only works, because previously conn->ginbuf.value was initialized
to NULL, and free(NULL) is a no-op. Commit 61bf96cab0 refactored things so
that the GSS input token buffer is allocated locally in pg_GSS_continue,
and not held in the PGconn object. After that, the now-local ginbuf.value
variable isn't initialized when it's not used, so we pass a bogus pointer
to free().

To fix, only try to free the input buffer if we allocated it. That was
the intention, certainly after the refactoring, and probably even before
that.
---
 src/interfaces/libpq/fe-auth.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/src/interfaces/libpq/fe-auth.c b/src/interfaces/libpq/fe-auth.c
index f4397afc64..f30fb04809 100644
--- a/src/interfaces/libpq/fe-auth.c
+++ b/src/interfaces/libpq/fe-auth.c
@@ -133,6 +133,11 @@ pg_GSS_continue(PGconn *conn, int payloadlen)
 			return STATUS_ERROR;
 		}
 	}
+	else
+	{
+		ginbuf.length = 0;
+		ginbuf.value = NULL;
+	}
 
 	maj_stat = gss_init_sec_context(&min_stat,
 									GSS_C_NO_CREDENTIAL,
@@ -142,13 +147,13 @@ pg_GSS_continue(PGconn *conn, int payloadlen)
 									GSS_C_MUTUAL_FLAG,
 									0,
 									GSS_C_NO_CHANNEL_BINDINGS,
-				(conn->gctx == GSS_C_NO_CONTEXT) ? GSS_C_NO_BUFFER : &ginbuf,
+						  (ginbuf.value == NULL) ? GSS_C_NO_BUFFER : &ginbuf,
 									NULL,
 									&goutbuf,
 									NULL,
 									NULL);
 
-	if (conn->gctx != GSS_C_NO_CONTEXT)
+	if (ginbuf.value)
 		free(ginbuf.value);
 
 	if (goutbuf.length != 0)
-- 
2.11.0