| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <CAA93jw6NZc4qP5ArK6tFF1DhP3_Rxsr8VFusU-nSzLHb6tnxjQ@mail.gmail.com> Date: Thu, 10 Nov 2011 15:47:38 +0100 From: Dave Taht <dave.taht@...il.com> To: Johannes Berg <johannes@...solutions.net> Cc: netdev <netdev@...r.kernel.org>, linux-wireless <linux-wireless@...r.kernel.org> Subject: Re: creating netdev queues on the fly? On Thu, Nov 10, 2011 at 2:58 PM, Johannes Berg <johannes@...solutions.net> wrote: > Hi, > > Am I completely crazy? Somewhat. :) Much of your thinking aligns with mine, however my goal was to try and reduce latencies on wireless-n, where we send variable size truckloads of packets to each destination. Solving that one is hard, and requires two levels of active queue management in the packet scheduler layer, and a bit more communication up from the driver itself. We could have a unique 'station identifier' which fits handily into 32 bits as the max allowed range is 0-2008 and map from MAC to that on tx entry. Having that as a flow classifier lets us have per station destination queues easily split up via a std tc filter... and then we have the ability, finally, to manage queue depth on a per station basis. So you end up with four queues, each tied to a hardware queue that then splits things up on a per station basis, fair queues within the queues to each station, and recombines them at the end on a basis bursty enough to aggregate as they exit the radio. I don't mind at all up to 8000 queues, honestly, wasting 99% on mostly unused queue structures via pouring megabytes into useless bufferbloated FIFO only packet buffers seems an acceptible compromise, but I'm easy... As for managing queue depth on a per station basis, some of what has been discussed on "byte queue limits" applies, but given wireless's peculiarites, tsf timestamping on entry to the first qdisc, doing fair queuing inside the per-sta queue (QFQ?), and checking the timestamp on exit from the queue against a sane limit for the queue type would do wonders for overall latencies and network responsiveness. Done right, instead of seeing a single tcp stream capable of inducing multi-second latencies for the next stream, latencies would stay flat up unto the max aggregation depth of different streams on a given sta, subject only the how many other competing stations there are, the net effect of packet loss would be vastly lessened, and world peace, achieved. I dream of 2ms pings and dns lookups, even gaming, under load, on wireless. I do. First steps are getting a station identifier and some useful statistics regarding that stations max (that quantum) packet bundle size, and completion rate, mostly from minstrel... on each packet... a tc classifier that can use it to toss into the tcindex mechanism (if that is what is used), another sane classifier for something like QFQ per station, and a packet 'grouper' that can output correctly sized bursts of packets on a sane basis from the queues in a randomly sane order (not round robin per se', to even out the load it has to start dequeuing groups) How to do all that within tc? Well... I like the idea of throwing out the 32 bitness of tc's calssifiers (mac hashing and ipv6 hashing is not very effective), but I doubt that will fly.. So to fit into the the existing structures the idea of adding the concept of a tc qdisc 'grouper' along with all the other tc filter 'splitters' - that could be multiqueue and multiple hardware queue aware - seems like an answer. Another crazy piece of the idea (courtesy nbd - I'd rather go crazy adding fields to the skb) is to wedge that id and some minstrel statistics and completion rates and the timestamp into each skb's mostly unused 48 byte 'reserved for special uses field... which it has to do under rcu lock anyway. I started coding up time based queue limits the other weekend, actually... some of this has been discussed on the bloat list. > > > johannes > > -- > To unsubscribe from this list: send the line "unsubscribe linux-wireless" in > the body of a message to majordomo@...r.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > -- Dave Täht SKYPE: davetaht US Tel: 1-239-829-5608 FR Tel: 0638645374 http://www.bufferbloat.net -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@...r.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Powered by blists - more mailing lists