On Friday 06 September 2013 20:00:19 Jozsef Kadlecsik wrote: > Hi Oliver, > > On Fri, 6 Sep 2013, Oliver wrote: > > Alright, I took a look at the changes - I'll wait for the extension > > cleanup then send out a rebased version of hash:net,net since that's > > presumably going to be the difference between having to declare struct > > combinations for the various extensions versus not having to do it. I'm > > very intrigued to see what you can come up with that's shorter than my > > rework :P > > There's a technical problem with your refactoring patchset (it doesn't > prevent unaligned access to extension members), and a design issue: the > data about the extensions (offsets) should belong to the core. There's > still code duplication between the bitmap and hash types in it when even > the list type is missing. The solution means an overall rearrangig and > that's what I'm working on. Sounds good. I was working under the expectation that explicitly setting the alignment of extension structs wasn't that big of a deal, but I agree that it's better if it doesn't need to be manually taken care of and moving more stuff into the core would definitely improve things. > > > On a related note - I was thinking about the impact of recursively > > walking the CIDRs and it did occur to me that this operation is very > > parallelisable - however, I found that when I ran ipset -T on the set, > > the in-kernel code is running in an interrupt context (somehow, I > > couldn't see what made it like that), which means it's not possible to > > spin up kthreads (unless there's something that I don't know about) - > > when testing of an ipset is called via xt_set, is that also running in > > an interrupt context? if not, it could be parallelised quite nicely, > > which I anticipate would give some gains on multicore systems, and > > especially for IPv6, one would think. > > How could the walking of the CIDRs be parallellized, when the lookup must > follow the priority of the networks? (With the exception of checking > multiple same sized networks.) Well, pretty easily... as the most basic (albeit not necessarily most resource efficient) example: you spin up a kthread for every CIDR size and then check the return value of each kthread you spun up in order that you started them - the first result you get to that has a non-zero return value is obviously the one you require and you're done. Naturally the time saving here is coming from the hash table searches for every given CIDR size being done in parallel; the thread controller code is still going to iterate, but it's just checking a return code. I already took a stab at this, but it's currently impossible because the moment kthread_create_on_node() is called and goes into sched_setscheduler_nocheck(), it triggers BUG_ON(in_interrupt()) - I poked through all the base code but I don't see exactly how or why we're inside an interrupt context when responding to ipset userspace commands. Kind Regards, Oliver. -- To unsubscribe from this list: send the line "unsubscribe netfilter-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
