Hello,
----- Original Message -----
From: "Bruce Momjian" <pgman@candle.pha.pa.us>
Sent: September 02, 2002 1:05 AM
> Would someone submit a patch for this?
Attached please find an attempt to fix the volunerability issue below.
Affected files are:
/src/include/libpq/libpq.h
/src/include/libpq/pqformat.h
/src/backend/libpq/pqformat.c
/src/backend/libpq/pqcomm.c
/src/backend/libpq/auth.c
"Briefly" the changes:
Main victims for the change were pq_getstring() and pq_getstr()
(which calls the former) in pqformat.c and pqcomm.c. pq_getstring() is the one reading
until \0 and might possibly render the system run out of memory.
Changing pq_getstring() alone would break a lot code, so I
added a two more functions: pq_getstring_common() and
pq_getstring_bounded(). The former is a big part of what used to be
pq_getstring() and the latter is a copy of the new pq_getstring()
with the string length check. Creating pq_getstring_common()
was suggested by its reuse in pq_getstring() and pq_getstring_bounded()
avoiding code duplication.
Similar changes were done for pq_getstr(). Its common code converting
to MULTIBYTE was placed in pq_getstr_multibyte() and pq_getstr() and
(newly added) pq_getstr_bounded() both call it before returning a result.
WRT above, two places in auth.c were changed to call pq_getstr_bounded()
instead of pq_getstr() on password read. I'm not sure if
there are other places where that might be needed...
Might look ugly for some, but looks like a not-so-bad solution
to me. If I'm completely wrong, I'd like to have some guidance then :)
Please review with care. I'm off to bed.
Thanks,
-s
PS: The patch also fixes a typo in the be-secure.c comment :)
> Tom Lane wrote:
> > Neil Conway <neilc@samurai.com> writes:
> > > (2) The length supplied by the user is completely ignored by
> > > the code, and it simply reads the input until it sees a
> > > NULL terminator (read the comments in the code about 10
> > > lines down.) Therefore, any sanity checking on the length
> > > specified by the user is a waste of time.
> >
> > Agreed; the fact that the protocol requires a length word at all is just
> > a hangover from the past. We can read the length word and forget it.
> >
> > I wonder though if it'd be worthwhile to limit the length of the string
> > that we are willing to read from the client in the second step. We are
> > at this point dealing with an unauthenticated user, so we should be
> > untrusting. And I think Sir Mordred has a point: forcing a backend to
> > allocate a lot of memory can be a form of DoS attack.
> >
> > regards, tom lane
