On Thu, Jan 30, 2025 at 01:20:40PM -0500, Laine Stump wrote: > On 1/30/25 10:48 AM, Andrea Bolognani wrote: > > Debian 12 doesn't come with a new enough libvirt version anyway, but > > FYI a few months back I switched the default backend in Debian to > > nftables (matching Fedora) only to walk back the decision after > > getting several reports of it breaking software that's just too > > popular to ignore. See [1] for more details. > > > > I don't expect that Debian will be able to move off the iptables > > backend any time soon, at least when it comes to the default. > > Changing the backend on a per-system basis is of course totally > > possible, as long as you understand the caveats. > > > > > > [1] https://bugs.debian.org/1090355 > > Sigh. > > In the days of iptables, there were 3 main tables (filter, nat, and mangle) > and everybody's rules went into those same 3 tables. Within that single > table, if a packet reached a REJECT or DROP rule before an ACCEPT rule (or > the end of the table) then the packet would be dropped, but if it reached an > ACCEPT rule first, then it would never see the REJECT rule, and be accepted. > > But with nftables, there are an infinite number of "base tables", all > traffic is evaluated against *all* tables *all the way* to either > accept/reject in *all* cases, and it must get to the end of *every single > table* without encountering a reject rule in order for the traffic to be > accepted - there is no "early exit on accept" that skips all the rest of the > tables if the traffic is accepted by one table. [...] > There is no generic way to fix this problem. libvirt can't possible find > every possible firewall system and add rules to the table of every single > one that passes traffic from libvirt guests. I guess the best we can > theoretically do is make a list of "supported firewall enemies" and add in > extra stuff just like we currently do with firewalld - 1) attempt to > autodetect if that enemy package is installed and active and then 2) add > whatever rules are necessary to the enemy package's table (or the "filter" > table if the enemy is still using iptables) in order to get our traffic > through) > > I'm not sure where I'm going with this, other than to say that, on a bad > day, trying to win this seems like a losing battle - let's say we add > something to counteract docker's rules, and ufw's rules (and later when > those packages add nftables support then we'll need to detect whether each > is using iptables or nftables, and add counteracting rules for those as > well); then what's next? Sorry, it's only Thursday and I'm already feeling > defeated :-/ If things really work the way you describe them, it sounds like an unsolvable problem indeed. Any scenario in which all possible components need to be aware of each other obviously doesn't scale. Have the nftables maintainers expressed their opinion about this? Surely they would have considered how to make filtering work without forcing extremely tight coupling. > So what do you consider libvirt could do to make it acceptable to have > nftables as the default backend on debian? Automatically add rules for the > current state of what docker and ufw do? Or is there some other slightly > more obscure package that we also need to compensate for before nftables > backend is acceptable as default? (seriously, let's declare an enumerated > list and then (hopefully, time permitting) take action on it. I would love > to completely eliminate the iptables backend if I possibly could, and that > certainly can't happen if some systems still have it as the default. I don't know if we can just draft a list. Docker and ufw would certainly need to work, but there might very well exist some other fairly popular package that I'm simply not aware of, for which being broken out of the box is not considered acceptable. That's exactly why tight coupling is undesirable. I'll note that the nwfilter driver not having an nftables backend is another, if secondary, reason to stick with iptables by default. The main goal for most people is to create a deployment that's completely free of the legacy userspace, and if some other driver is going to drag it in anyway, a big part of the benefit is immediately lost... -- Andrea Bolognani / Red Hat / Virtualization
