Skip site navigation (1) Skip section navigation (2)

Re: [HACKERS] Re: Cancel key now ready

From: Bruce Momjian <maillist(at)candle(dot)pha(dot)pa(dot)us>
To: tgl(at)sss(dot)pgh(dot)pa(dot)us (Tom Lane)
Cc: hackers(at)postgreSQL(dot)org
Subject: Re: [HACKERS] Re: Cancel key now ready
Date: 1998-06-08 21:50:59
Message-ID: 199806082150.RAA15151@candle.pha.pa.us (view raw or flat)
Thread:
Lists: pgsql-hackers
> 
> Bruce Momjian <maillist(at)candle(dot)pha(dot)pa(dot)us> writes:
> > I thought about this.  I can force a re-seeding of random in the
> > backend on first use.
> 
> No you can't; you might make PostmasterRandom behave that way, but
> that doesn't stop an installed function from just calling random()
> directly.  You really need to wipe out the state saved by the random
> library function.

You can't just call random directly.  You have to call an install
function with a pg_proc entry for it to work.  If we set that one up to
initialize itself, it should work.

> 
> > Could re-seed on every
> > startup, but again, could be an expensive function.
> 
> srandom() is generally not much more than a store into a
> static variable.  If there's anything expensive about this,
> it'd be the gettimeofday() call to produce the new seed value.

But we don't call that for every backend startup, just twice for the
life of the postmaster, once for postmaster startup, and once for the
startup of the first backend.  What I don't want is to profile the
backend and find that random/srandom() is showing up as significant.

> 
> > If they have access the backend address space, they can see the entire
> > postmaster backend structure at time of fork(), so seeing the seed is
> > meanless.
> 
> That's a good point --- in particular, they could trace the postmaster
> backend-process list to find out everyone else's cancel keys.  This
> sort of thing is one of the disadvantages of not using an exec().
> 
> What do you think of freeing that process list as part of backend startup?

Again, being able to connect to the backend, and accessing its address
space are two separate privs.  Only the postgres user can do such
access.

> 
> > Basically, for any user who is installing their own functions
> > or stuff is already able to do more severe damage than just cancel. 
> > They can write directly into the database.
> 
> That's certainly true ... but last week we were trying to make the
> cancel protocol proof against someone with the ability to spy on
> TCP packets in transit (which is not that easy) and now you seem
> to be unworried about attacks that only require loading a function
> into one of the backends.  I'd prefer to do whatever we can easily
> do to defend against that.

Only the postgres super-user can load functions.  This is something we
have protection against.  Someone snooping the wire may not even have
permissions to access the database.

> p
> >> * fflush before forking, to eliminate double-buffering problems
> >> between postmaster and backends.
> 
> > Can you elaborate on what this fixes?
> 
> I have not seen any failure cases, if that's what you mean; but I
> haven't yet done anything with the new no-exec code.  The risk is
> that if any data is waiting in a postmaster stdio output buffer,
> it will eventually get written twice, once by the postmaster and
> once by the backend.  You want to flush it out before forking
> to ensure that doesn't happen.  This wasn't an issue before with
> the exec-based code, because the child process' copy of the postmaster's
> stdio buffers got thrown away when the exec() occurred.  With no
> exec, the unflushed buffers are still there and still valid as far
> as the stdio library in the child knows.

Yes.  Excellent point.

> 
> > The code is similar to taking a random() and doing:
> > 	rand % 10
> > 	(rand / 10) % 10
> > which for a random of 123456 returns 6 and 5.  In the postmaster case
> > the values are 62 and not 10, but the concept is the same.  No reason to
> > call random() twice.  May be an expensive function on some platforms.
> 
> It's not that expensive (you were doing it twice before, with no visible
> problem).  I'm concerned that the new way exposes more info about the
> current state of the postmaster's random sequence.  For that matter,
> I'm probably going to want to change the computation of the cancel key
> later on --- the code I just sent in was only
> 	MyCancelKey = PostmasterRandom();
> but I think it would be better to synthesize the cancel key from
> fragments of a couple of random values.  This code will do to get the
> protocol working but I don't think it's cryptographically secure.

Again, XOR'ing with the seed should do what we need.

-- 
Bruce Momjian                          |  830 Blythe Avenue
maillist(at)candle(dot)pha(dot)pa(dot)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)

In response to

pgsql-hackers by date

Next:From: Bruce MomjianDate: 1998-06-08 21:55:22
Subject: Re: [HACKERS] Re: Cancel key now ready
Previous:From: Egon SchmidDate: 1998-06-08 20:33:24
Subject: Re: [HACKERS] Re: Cancel key now ready

Privacy Policy | About PostgreSQL
Copyright © 1996-2014 The PostgreSQL Global Development Group