Linux Network Traffic ControlImplementation OverviewWord下载.docx

Linux Network Traffic ControlImplementation OverviewWord下载.docx

《Linux Network Traffic ControlImplementation OverviewWord下载.docx》由会员分享，可在线阅读，更多相关《Linux Network Traffic ControlImplementation OverviewWord下载.docx（19页珍藏版）》请在冰豆网上搜索。

intserv"

group[2],andwillserveasthebasisforsupportingthemorerecentworkof"

diffserv"

[3].Seealso[4]forfurtherdetailsonhowintservanddiffservarerelated.ThisdocumentillustratestheunderlyingarchitectureanddescribeshownewtrafficcontrolfunctionscanbeaddedtotheLinuxkernel.Thekernelversionweusedis2.2.6.

Figure1showsroughlyhowthekernelprocessesdatareceivedfromthenetwork,andhowitgeneratesnewdatatobesentonthenetwork:

incomingpacketsareexaminedandtheneitherdirectlyforwardedtothenetwork（e.g.onadifferentinterface,ifthemachineisactingasarouterorabridge）,ortheyarepasseduptohigherlayersintheprotocolstack（e.g.toatransportprotocollikeUDPorTCP）forfurtherprocessing.Thosehigherlayersmayalsogeneratedataontheirownandhandittothelowerlayersfortaskslikeencapsulation,routing,andeventuallytransmission.

"

Forwarding"

includestheselectionoftheoutputinterface,theselectionofthenexthop,encapsulation,etc.Onceallthisisdone,packetsarequeuedontherespectiveoutputinterface.Thisisthepointwheretrafficcontrolcomesintoplay.Trafficcontrolcan,amongotherthings,decideifpacketsarequeuedoriftheyaredropped（e.g.ifthequeuehasreachedsomelengthlimit,orifthetrafficexceedssomeratelimit）,itcandecideinwhichorderpacketsaresent（e.g.togiveprioritytocertainflows）,itcandelaythesendingofpackets（e.g.tolimittherateofoutboundtraffic）,etc.

Oncetrafficcontrolhasreleasedapacketforsending,thedevicedriverpicksitupandemitsitonthenetwork.

Sections2to4giveanoverviewandexplainsometerminology.Sections5to8describetheelementsoftrafficcontrolintheLinuxkernelinmoredetail.Section9describesaqueuingdisciplinethathasbeenimplementedbytheauthor.

2Overview

ThetrafficcontrolcodeintheLinuxkernelconsistsofthefollowingmajorconceptualcomponents:

●queuingdisciplines

●classes（withinaqueuingdiscipline）

●filters

●policing

Figure1:

Processingofnetworkdata.

Eachnetworkdevicehasaqueuingdisciplineassociatedwithit,whichcontrolshowpacketsenqueuedonthatdevicearetreated.Averysimplequeuingdisciplinemayjustconsistofasinglequeue,whereallpacketsarestoredintheorderinwhichtheyhavebeenenqueued,andwhichisemptiedasfastastherespectivedevicecansend.Seefigure2forsuchaqueuingdisciplinewithoutexternallyvisibleinternalstructure.

Figure2:

Asimplequeuingdisciplinewithoutclasses.

Moreelaboratequeuingdisciplinesmayusefilterstodistinguishamongdifferentclassesofpacketsandprocesseachclassinaspecificway,e.g.bygivingoneclasspriorityoverotherclasses.

Figure3showsanexampleofsuchaqueuingdiscipline.Notethatmultiplefiltersmaymaptothesameclass.

Queuingdisciplinesandclassesareintimatelytiedtogether:

thepresenceofclassesandtheirsemanticsarefundamentalpropertiesofthequeuingdiscipline.Incontrasttothat,filterscanbecombinedarbitrarilywithqueuingdisciplinesandclassesaslongasthequeuingdisciplinehasclassesatall.Butflexibilitydoesn'

tendyet-classesnormallydon'

ttakecareofstoringtheirpacketsthemselves,buttheyuseanotherqueuingdisciplinetotakecareofthat.Thatqueuingdisciplinecanbearbitrarilychosenfromthesetofavailablequeuingdisciplines,anditmaywellhaveclasses,whichinturnusequeuingdisciplines,etc.

Figure4showsanexampleofsuchastack:

first,thereisaqueuingdisciplinewithtwodelaypriorities.Packetswhichareselectedbythefiltergotothehigh-priorityclass,whileallotherpacketsgotothelow-priorityclass.Whenevertherearepacketsinthehigh-priorityqueue,theyaresentbeforepacketsinthelow-priorityqueue（e.g.thesch_prioqueuingdisciplineworksthisway）.Inordertopreventhigh-prioritytrafficfromstarvinglow-prioritytraffic,weuseatokenbucketfilter（TBF）,whichenforcesarateofatmost1Mbps.Finally,thequeuingoflow-prioritypacketsisdonebyaFIFOqueuingdiscipline.Notethattherearebetterwaystoaccomplishwhatwe'

vedonehere,e.g.byusingclass-basedqueuing（CBQ）[5].

Packetsareenqueuedasfollows:

whentheenqueuedfunctionofaqueuingdisciplineiscalled,itrunsonefilteraftertheotheruntiloneofthemindicatesamatch.Itthenqueuesthepacketforthecorrespondingclass,whichusuallymeanstoinvoketheenqueuedfunctionofthequeuingdiscipline"

owned"

bythatclass.Packetswhichdonotmatchanyofthefiltersaretypicallyattributedtosomedefaultclass.

Typically,eachclass"

owns"

onequeue,butitisinprinciplealsopossiblethatseveralclassessharethesamequeueoreventhatasinglequeueisusedbyallclassesoftherespectivequeuingdiscipline.Notehoweverthatpacketsdonotcarryanyexplicitindicationofwhichclasstheywereattributedto.Queuingdisciplinesthatchangeper-classinformationwhendequeuingpackets（e.g.CBQ）maythereforenotworkproperlyifthe"

inner"

queuesareshared,unlesstheyareableeithertorepeattheclassificationortopasstheclassificationresultfromenqueuedtodequeuebysomeothermeans.

Usuallywhenenqueuingpackets,thecorrespondingflow（s）canbepoliced,e.g.bydiscardingpacketswhichexceedacertainrate.

Wewillnottrytointroducenewterminologytodistinguishamongalgorithms,theirimplementations,andinstancesofsuchelements,butratherusethetermsqueuingdiscipline,class,andfilterthrough-outmostofthisdocument,torefertoallthreelevelsofabstractionatthesametime.

3Resources

Linuxtrafficcontrolisspreadoveracomparablylargenumberoffiles.Notethatallpathnamesarerelativetothebasedirectoryoftherespectivecomponent,Filtere.g.fortheLinuxkernelthisis/usr/src/linux/,forthetcprogramiproute2/tc/.

Figure3:

Asimplequeuingdisciplinewithmultipleclasses.

Figure4:

Combinationofpriority,TBF,andFIFOqueuingdisciplines.

Tcisauser-spaceprogramusedtomanipulateindividualtrafficcontrolelements.Itssourceisinthefileiproute2-version.tar.gz,whichcanbeobtainedfromftp:

//linux.wauug.org/pub/net/ip-routing/.

Thekernelcoderesidesmainlyinthedirectorynet/sched/.Theinterfacesbetweenkerneltrafficcontrolelementsanduserspaceprogramsusingthemaredeclaredininclude/linux/pkt_cls.handinclude/linux/pkt_sched.h.Declarationsusedonlyinsidethekernelandthedefinitionsofsomeinlinefunctionscanbefoundininclude/net/pkt_cls.handinclude/net/pkt_sched.h.

Thertnetlinkmechanismusedforcommunicationbetweentrafficcontrolelementsinuser-spaceandinthekernelisimplementedinnet/core/rtnetlink.candinclude/linux/rtnetlink.h.rtnetlinkisbasedonnetlink,whichcanbefoundinnet/netlink/andinclude/linux/netlink.h.

Thekernelsourcecanbeobtainedfromtheusualwell-knownplaces,e.g.fromftp:

//ftp.kernel.org/pub/linux/kernel/v2.2/.

Theexampleinsection9isincludedintheATMonLinuxdistribution,whichcanbedownloadedfromhttp:

//icawww1.epfl.ch/linux-atm/dist.html.

TheDifferentiatedServicesonLinuxproject（http:

//icawww1.epfl.ch/linux-diffserv/）hasproducedfurtherexamplesforextensionsofLinuxtrafficcontrolandtheiruse.

4Terminology

Unfortunately,theterminologyusedtodescribetrafficcontrolelementsisfarfromconsistentinliterature,andtherearesomevariationsevenwithinLinuxtrafficcontrol.Thepurposeofthissectionistohelptoputthingsintocontext.

Figure5showsthearchitecturalmodelsandtheterminologyusedintheIETFgroups"

[6]and"

[7,8],andhowelementsofLinuxtrafficcontrolarerelatedtothem.Notethatclassesplayanambivalentrole,becausetheydeterminethefinaloutcomeofaclassificationandtheycanalsobepartofthemechanismthatimplementsacertainqueuingorschedulingbehavior.

Table1summarizesthekeywordsusedatthetccommandline,thefilenamesusedinthekernel（innet/sched/）,andthefilenamesusedinthesourceoftc.

Figure5:

RelationofelementsoftheintservanddiffservarchitecturetotrafficcontrolintheLinuxkernel.

Table1:

Keywordsandfilenamesusedfortrafficcontrolelements.

5Queuingdisciplines

Eachqueuingdisciplineprovidesthefollowingsetoffunctionstocontrolitsoperation（seestructQdisc_opsininclude/net/pkt_sched.h）:

enqueueenqueuesapacketwiththequeuingdiscipline.Ifthequeuingdisciplinehasclasses,theenqueuedfunctionfirstselectsaclassandtheninvokestheenqueuefunctionofthecorrespondingqueuingdisciplineforfurtherenqueuing.

dequeuereturnsthenextpacketeligibleforsending.Ifthequeuingdisciplinehasnopacketstosend（e.g.becausethequeueisemptyorbecausethey'

renotscheduledtobesentyet）,dequeuereturnsNULL.

requeueputsapacketbackintothequeueafterdequeuingitwithdequeue.Thisdiffersfromenqueueinthatthepacketshouldbequeuedate