Index | Thread | Search

From:
Stuart Henderson <stu@spacehopper.org>
Subject:
pf_lb port range retries [Re: CVS: cvs.openbsd.org: src]
To:
Alexandr Nedvedicky <sashan@cvs.openbsd.org>
Cc:
tech <tech@openbsd.org>
Date:
Mon, 9 Nov 2026 11:13:20 +0000

Download raw body.

Thread
On 2026/11/09 03:56, Alexandr Nedvedicky wrote:
> CVSROOT:	/cvs
> Module name:	src
> Changes by:	sashan@cvs.openbsd.org	2026/11/09 03:56:59
> 
> Modified files:
> 	sys/net        : pf_lb.c 
> 
> Log message:
> pf(4): depleted source port pool for NAT starves other packets
> 
> The function pf_get_sport() in current selects the source port
> number for network address translation (a.k.a. NAT). It basically
> chooses a random source port and uses pf_find_state_all()
> to check if the chosen port number is already in use for a particular NAT
> address. If the port is already in use, pf_get_sport() selects
> another port number and retries. This process continues as long
> as the chosen port number collides with an existing address:port tuple.
> The problem is that all this happens under the protection of
> PF_LOCK(). This effectively stops pf(4) rule evaluation
> for other packets that enter the firewall and have no state.
> 
> The fix introduces a limit on retries. The limit is set to 128.
> If no free port number is found within 128 attempts, the function
> gives up.

I think we need to watch for issues with this, it seems like it will
be bad on systems with a lot of machines behind a nat (though it is
true that they will already be locking a lot..)

There is a DPFPRINTF for port allocation failure for every new state,
but it's LOG_NOTICE which is not enabled by default (and I think it
ends up pretty noisy if you do enable it). I don't think it's
included in any pfctl -si counter.

>           The packet is dropped and PF_LOCK() is released so other
> packets waiting for a rule check can be inspected by pf_test().

This isn't correct. The nat rule is not matched but ruleset traversal
continues. If the ruleset allows it to pass (common where the nat-to
is on a match rule), the packet is sent untranslated unless there's
a subsequent translate rule which matches and has a free port.

If people bump into this, increasing the nat port range should help
(at least to an extent). Default range is 50001:65535, defined in
pfctl_parser.h PF_NAT_PROXY_PORT_LOW / PF_NAT_PROXY_PORT_HIGH, but
it can be set per-rule. If overriding, try to avoid starving the
net.inet.ip.porthi* range to prevent affecting connections from the
machine itself too badly. Something like 30000:60000 would likely
be a better choice than e.g. 40000:65535.