Re: [PATCH iptables 4/4] nft: keep old cache around until batch is refreshed in case of ERESTART

[Phil Sutter <phil@nwl.cc>]

On Mon, May 20, 2019 at 02:06:20PM +0200, Florian Westphal wrote:
> Phil Sutter <phil@nwl.cc> wrote:
> > On Sun, May 19, 2019 at 06:45:08PM +0200, Phil Sutter wrote:
> > [...]
> > > The only way to make the above work is by keeping the original cache
> > > copy around until mnl_batch_talk has finally succeeded or failed with
> > > something else than ERESTART.
> > 
> > How about a completely different approach:
> > 
> > If memory serves right (and from reading the related Red Hat ticket),
> > the actual problem we're trying to solve is that iptables-nft-restore
> > creates NFT_MSG_DELTABLE only if that table exists already at the point
> > of parsing but another client might create it in the mean time before
> > committing.
> > 
> > My idea for solving this was to unconditionally create NFT_MSG_NEWTABLE
> > followed by NFT_MSG_DELTABLE - in case the table exists, the first one
> > is a noop; in case the table doesn't exist, the second one won't provoke
> > an error message from kernel space.
> 
> Does that work even work?
> new table x
> del table x
> add rule to x // table was deleted?
> 
> Or are you talking about a new/del/new sequence?

Oh, yes of course. Existing iptables-nft-restore does:

1) delete table x if exists
2) add table x
3) add table x content

My idea is to:

+ 0) add table x
- 1) delete table x if exists
+ 1) delete table x
  2) add table x
  3) add table x content

> If it works, ok/fine, but it seems ugly.

Did you consider rule insert with index in your batch replay logic? I
did when duplicating it for nft, but decided it's not worth it and
people using 'add rule ... index IDX' have been warned already anyway.

> > Since NFT_MSG_DELTABLE removes the table recursively, we don't need to
> > care about any content added by the other client.
> 
> Yes, we can't do this in the --noflush case though.

My state from our last talk about it was "--noflush users are screwed
anyway". :)

> > Or is this about a generic solution for all commands not just
> > iptables-nft-restore (without --noflush)?
> 
> Its for ipt-nft-restore, including --noflush.
> 
> It would be good though to also speed up 'iptables-nft -A' later on
> by eliding cache completely.

The problem is that we can't unconditionally create NFT_MSG_NEWCHAIN
because that will reset the chain policy if non-default. So we need at
least a cache containing tables and chains, even for that simple rule
append command.

Cheers, Phil