Re: [RFC] Faster libvirtd restart with nwfilter rules, one more time

On 3/20/20 5:25 AM, nshirokovskiy wrote: > Hi, all. >                                                                                  Some time ago I posted RFC [1] concerning an issue of unresponsive > libvird during restart if there is large number of VMs that have network > filters on their interfaces. It was identified that in most cases we > don't need actually to reinstall network filter rules on daemon restart. > Thus I proposed patches [2] that check whether we need to reapply rules > or not. >                                                                                  The first version has a drawback that daemon won't reapply rules if > someone mangled them between daemon stop and start (and this can be done > just by restarting firewalld). The second one is just ugly :) >                                                                                  Around that time Florian Westphal in a letter off the mailing list > suggested to use {iptables|ebtables}-restore to apply rules in one > binary call. These binaries has --noflush option so that we won't reset > current state of tables.  We also need one more -L call for > iptables/ebtables to query current filter state to be able to construct > input for restore binaries. So are you considering doing something with this idea? At the end of our discussion, both libvirt and firewalld people agreed that we're gaining nothing from setting our rules via firewalld passthrough, and we would be potentially gaining *a lot* by setting them in batch mode with "iptables-restore -n". Perhaps we could just add a new firewall backend (in util/virfirewall.c) that checked for the presence of iptables-restore (and ip6tables-restore and ebtables-restore), and if they are found it would use a backend that just put all the rules for each layer together in a temporary file and send them to *-restore (the internals would need to be reorganized a bit, so that args like -w, -l, and -n could be added in during virFirewallApply (if necessary) rather than when initially adding rules). Ooh! I just tried it, and iptables-restore also accepts (and acts on) lines with "-D" to delete rules! So we could do everything in a single go - intermixing -D and -A rules in the same file (to minimize the time when the firewall would be incorrect while still taking advantage of the efficiency of doing everything in a batch).