tcp_output.c 108 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611261226132614261526162617261826192620262126222623262426252626262726282629263026312632263326342635263626372638263926402641264226432644264526462647264826492650265126522653265426552656265726582659266026612662266326642665266626672668266926702671267226732674267526762677267826792680268126822683268426852686268726882689269026912692269326942695269626972698269927002701270227032704270527062707270827092710271127122713271427152716271727182719272027212722272327242725272627272728272927302731273227332734273527362737273827392740274127422743274427452746274727482749275027512752275327542755275627572758275927602761276227632764276527662767276827692770277127722773277427752776277727782779278027812782278327842785278627872788278927902791279227932794279527962797279827992800280128022803280428052806280728082809281028112812281328142815281628172818281928202821282228232824282528262827282828292830283128322833283428352836283728382839284028412842284328442845284628472848284928502851285228532854285528562857285828592860286128622863286428652866286728682869287028712872287328742875287628772878287928802881288228832884288528862887288828892890289128922893289428952896289728982899290029012902290329042905290629072908290929102911291229132914291529162917291829192920292129222923292429252926292729282929293029312932293329342935293629372938293929402941294229432944294529462947294829492950295129522953295429552956295729582959296029612962296329642965296629672968296929702971297229732974297529762977297829792980298129822983298429852986298729882989299029912992299329942995299629972998299930003001300230033004300530063007300830093010301130123013301430153016301730183019302030213022302330243025302630273028302930303031303230333034303530363037303830393040304130423043304430453046304730483049305030513052305330543055305630573058305930603061306230633064306530663067306830693070307130723073307430753076307730783079308030813082308330843085308630873088308930903091309230933094309530963097309830993100310131023103310431053106310731083109311031113112311331143115311631173118311931203121312231233124312531263127312831293130313131323133313431353136313731383139314031413142314331443145314631473148314931503151315231533154315531563157315831593160316131623163316431653166316731683169317031713172317331743175317631773178317931803181318231833184318531863187318831893190319131923193319431953196319731983199320032013202320332043205320632073208320932103211321232133214321532163217321832193220322132223223322432253226322732283229323032313232323332343235323632373238323932403241324232433244324532463247324832493250325132523253325432553256325732583259326032613262326332643265326632673268326932703271327232733274327532763277327832793280328132823283328432853286328732883289329032913292329332943295329632973298329933003301330233033304330533063307330833093310331133123313331433153316331733183319332033213322332333243325332633273328332933303331333233333334333533363337333833393340334133423343334433453346334733483349335033513352335333543355335633573358335933603361336233633364336533663367336833693370337133723373337433753376337733783379338033813382338333843385338633873388338933903391339233933394339533963397339833993400340134023403340434053406340734083409341034113412341334143415341634173418341934203421342234233424342534263427342834293430343134323433343434353436343734383439344034413442344334443445344634473448344934503451345234533454345534563457345834593460346134623463346434653466346734683469347034713472347334743475347634773478347934803481348234833484348534863487348834893490349134923493349434953496349734983499350035013502350335043505350635073508350935103511351235133514351535163517351835193520352135223523352435253526352735283529353035313532353335343535353635373538353935403541354235433544354535463547354835493550355135523553355435553556355735583559356035613562356335643565356635673568356935703571357235733574357535763577357835793580358135823583358435853586358735883589359035913592359335943595359635973598359936003601360236033604360536063607360836093610361136123613361436153616361736183619362036213622362336243625362636273628362936303631363236333634363536363637363836393640364136423643364436453646364736483649365036513652365336543655365636573658365936603661366236633664366536663667366836693670367136723673367436753676367736783679368036813682368336843685368636873688368936903691369236933694369536963697369836993700370137023703370437053706370737083709371037113712371337143715371637173718371937203721372237233724372537263727372837293730373137323733373437353736373737383739374037413742374337443745374637473748374937503751375237533754375537563757375837593760
  1. /*
  2. * INET An implementation of the TCP/IP protocol suite for the LINUX
  3. * operating system. INET is implemented using the BSD Socket
  4. * interface as the means of communication with the user level.
  5. *
  6. * Implementation of the Transmission Control Protocol(TCP).
  7. *
  8. * Authors: Ross Biro
  9. * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10. * Mark Evans, <evansmp@uhura.aston.ac.uk>
  11. * Corey Minyard <wf-rch!minyard@relay.EU.net>
  12. * Florian La Roche, <flla@stud.uni-sb.de>
  13. * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14. * Linus Torvalds, <torvalds@cs.helsinki.fi>
  15. * Alan Cox, <gw4pts@gw4pts.ampr.org>
  16. * Matthew Dillon, <dillon@apollo.west.oic.com>
  17. * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18. * Jorge Cwik, <jorge@laser.satlink.net>
  19. */
  20. /*
  21. * Changes: Pedro Roque : Retransmit queue handled by TCP.
  22. * : Fragmentation on mtu decrease
  23. * : Segment collapse on retransmit
  24. * : AF independence
  25. *
  26. * Linus Torvalds : send_delayed_ack
  27. * David S. Miller : Charge memory using the right skb
  28. * during syn/ack processing.
  29. * David S. Miller : Output engine completely rewritten.
  30. * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
  31. * Cacophonix Gaul : draft-minshall-nagle-01
  32. * J Hadi Salim : ECN support
  33. *
  34. */
  35. #define pr_fmt(fmt) "TCP: " fmt
  36. #include <net/tcp.h>
  37. #include <linux/compiler.h>
  38. #include <linux/gfp.h>
  39. #include <linux/module.h>
  40. #include <linux/static_key.h>
  41. #include <trace/events/tcp.h>
  42. static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  43. int push_one, gfp_t gfp);
  44. /* Account for new data that has been sent to the network. */
  45. static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  46. {
  47. struct inet_connection_sock *icsk = inet_csk(sk);
  48. struct tcp_sock *tp = tcp_sk(sk);
  49. unsigned int prior_packets = tp->packets_out;
  50. tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
  51. __skb_unlink(skb, &sk->sk_write_queue);
  52. tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
  53. tp->packets_out += tcp_skb_pcount(skb);
  54. if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
  55. tcp_rearm_rto(sk);
  56. NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
  57. tcp_skb_pcount(skb));
  58. }
  59. /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
  60. * window scaling factor due to loss of precision.
  61. * If window has been shrunk, what should we make? It is not clear at all.
  62. * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  63. * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  64. * invalid. OK, let's make this for now:
  65. */
  66. static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  67. {
  68. const struct tcp_sock *tp = tcp_sk(sk);
  69. if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
  70. (tp->rx_opt.wscale_ok &&
  71. ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
  72. return tp->snd_nxt;
  73. else
  74. return tcp_wnd_end(tp);
  75. }
  76. /* Calculate mss to advertise in SYN segment.
  77. * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
  78. *
  79. * 1. It is independent of path mtu.
  80. * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
  81. * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
  82. * attached devices, because some buggy hosts are confused by
  83. * large MSS.
  84. * 4. We do not make 3, we advertise MSS, calculated from first
  85. * hop device mtu, but allow to raise it to ip_rt_min_advmss.
  86. * This may be overridden via information stored in routing table.
  87. * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
  88. * probably even Jumbo".
  89. */
  90. static __u16 tcp_advertise_mss(struct sock *sk)
  91. {
  92. struct tcp_sock *tp = tcp_sk(sk);
  93. const struct dst_entry *dst = __sk_dst_get(sk);
  94. int mss = tp->advmss;
  95. if (dst) {
  96. unsigned int metric = dst_metric_advmss(dst);
  97. if (metric < mss) {
  98. mss = metric;
  99. tp->advmss = mss;
  100. }
  101. }
  102. return (__u16)mss;
  103. }
  104. /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
  105. * This is the first part of cwnd validation mechanism.
  106. */
  107. void tcp_cwnd_restart(struct sock *sk, s32 delta)
  108. {
  109. struct tcp_sock *tp = tcp_sk(sk);
  110. u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
  111. u32 cwnd = tp->snd_cwnd;
  112. tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
  113. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  114. restart_cwnd = min(restart_cwnd, cwnd);
  115. while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
  116. cwnd >>= 1;
  117. tp->snd_cwnd = max(cwnd, restart_cwnd);
  118. tp->snd_cwnd_stamp = tcp_jiffies32;
  119. tp->snd_cwnd_used = 0;
  120. }
  121. /* Congestion state accounting after a packet has been sent. */
  122. static void tcp_event_data_sent(struct tcp_sock *tp,
  123. struct sock *sk)
  124. {
  125. struct inet_connection_sock *icsk = inet_csk(sk);
  126. const u32 now = tcp_jiffies32;
  127. if (tcp_packets_in_flight(tp) == 0)
  128. tcp_ca_event(sk, CA_EVENT_TX_START);
  129. tp->lsndtime = now;
  130. /* If it is a reply for ato after last received
  131. * packet, enter pingpong mode.
  132. */
  133. if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
  134. icsk->icsk_ack.pingpong = 1;
  135. }
  136. /* Account for an ACK we sent. */
  137. static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
  138. u32 rcv_nxt)
  139. {
  140. struct tcp_sock *tp = tcp_sk(sk);
  141. if (unlikely(tp->compressed_ack)) {
  142. NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
  143. tp->compressed_ack);
  144. tp->compressed_ack = 0;
  145. if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
  146. __sock_put(sk);
  147. }
  148. if (unlikely(rcv_nxt != tp->rcv_nxt))
  149. return; /* Special ACK sent by DCTCP to reflect ECN */
  150. tcp_dec_quickack_mode(sk, pkts);
  151. inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
  152. }
  153. u32 tcp_default_init_rwnd(u32 mss)
  154. {
  155. /* Initial receive window should be twice of TCP_INIT_CWND to
  156. * enable proper sending of new unsent data during fast recovery
  157. * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
  158. * limit when mss is larger than 1460.
  159. */
  160. u32 init_rwnd = TCP_INIT_CWND * 2;
  161. if (mss > 1460)
  162. init_rwnd = max((1460 * init_rwnd) / mss, 2U);
  163. return init_rwnd;
  164. }
  165. /* Determine a window scaling and initial window to offer.
  166. * Based on the assumption that the given amount of space
  167. * will be offered. Store the results in the tp structure.
  168. * NOTE: for smooth operation initial space offering should
  169. * be a multiple of mss if possible. We assume here that mss >= 1.
  170. * This MUST be enforced by all callers.
  171. */
  172. void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
  173. __u32 *rcv_wnd, __u32 *window_clamp,
  174. int wscale_ok, __u8 *rcv_wscale,
  175. __u32 init_rcv_wnd)
  176. {
  177. unsigned int space = (__space < 0 ? 0 : __space);
  178. /* If no clamp set the clamp to the max possible scaled window */
  179. if (*window_clamp == 0)
  180. (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
  181. space = min(*window_clamp, space);
  182. /* Quantize space offering to a multiple of mss if possible. */
  183. if (space > mss)
  184. space = rounddown(space, mss);
  185. /* NOTE: offering an initial window larger than 32767
  186. * will break some buggy TCP stacks. If the admin tells us
  187. * it is likely we could be speaking with such a buggy stack
  188. * we will truncate our initial window offering to 32K-1
  189. * unless the remote has sent us a window scaling option,
  190. * which we interpret as a sign the remote TCP is not
  191. * misinterpreting the window field as a signed quantity.
  192. */
  193. if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
  194. (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
  195. else
  196. (*rcv_wnd) = space;
  197. (*rcv_wscale) = 0;
  198. if (wscale_ok) {
  199. /* Set window scaling on max possible window */
  200. space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
  201. space = max_t(u32, space, sysctl_rmem_max);
  202. space = min_t(u32, space, *window_clamp);
  203. while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
  204. space >>= 1;
  205. (*rcv_wscale)++;
  206. }
  207. }
  208. if (!init_rcv_wnd) /* Use default unless specified otherwise */
  209. init_rcv_wnd = tcp_default_init_rwnd(mss);
  210. *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
  211. /* Set the clamp no higher than max representable value */
  212. (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
  213. }
  214. EXPORT_SYMBOL(tcp_select_initial_window);
  215. /* Chose a new window to advertise, update state in tcp_sock for the
  216. * socket, and return result with RFC1323 scaling applied. The return
  217. * value can be stuffed directly into th->window for an outgoing
  218. * frame.
  219. */
  220. static u16 tcp_select_window(struct sock *sk)
  221. {
  222. struct tcp_sock *tp = tcp_sk(sk);
  223. u32 old_win = tp->rcv_wnd;
  224. u32 cur_win = tcp_receive_window(tp);
  225. u32 new_win = __tcp_select_window(sk);
  226. /* Never shrink the offered window */
  227. if (new_win < cur_win) {
  228. /* Danger Will Robinson!
  229. * Don't update rcv_wup/rcv_wnd here or else
  230. * we will not be able to advertise a zero
  231. * window in time. --DaveM
  232. *
  233. * Relax Will Robinson.
  234. */
  235. if (new_win == 0)
  236. NET_INC_STATS(sock_net(sk),
  237. LINUX_MIB_TCPWANTZEROWINDOWADV);
  238. new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
  239. }
  240. tp->rcv_wnd = new_win;
  241. tp->rcv_wup = tp->rcv_nxt;
  242. /* Make sure we do not exceed the maximum possible
  243. * scaled window.
  244. */
  245. if (!tp->rx_opt.rcv_wscale &&
  246. sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
  247. new_win = min(new_win, MAX_TCP_WINDOW);
  248. else
  249. new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
  250. /* RFC1323 scaling applied */
  251. new_win >>= tp->rx_opt.rcv_wscale;
  252. /* If we advertise zero window, disable fast path. */
  253. if (new_win == 0) {
  254. tp->pred_flags = 0;
  255. if (old_win)
  256. NET_INC_STATS(sock_net(sk),
  257. LINUX_MIB_TCPTOZEROWINDOWADV);
  258. } else if (old_win == 0) {
  259. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
  260. }
  261. return new_win;
  262. }
  263. /* Packet ECN state for a SYN-ACK */
  264. static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
  265. {
  266. const struct tcp_sock *tp = tcp_sk(sk);
  267. TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
  268. if (!(tp->ecn_flags & TCP_ECN_OK))
  269. TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
  270. else if (tcp_ca_needs_ecn(sk) ||
  271. tcp_bpf_ca_needs_ecn(sk))
  272. INET_ECN_xmit(sk);
  273. }
  274. /* Packet ECN state for a SYN. */
  275. static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
  276. {
  277. struct tcp_sock *tp = tcp_sk(sk);
  278. bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
  279. bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
  280. tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
  281. if (!use_ecn) {
  282. const struct dst_entry *dst = __sk_dst_get(sk);
  283. if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
  284. use_ecn = true;
  285. }
  286. tp->ecn_flags = 0;
  287. if (use_ecn) {
  288. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
  289. tp->ecn_flags = TCP_ECN_OK;
  290. if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
  291. INET_ECN_xmit(sk);
  292. }
  293. }
  294. static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
  295. {
  296. if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
  297. /* tp->ecn_flags are cleared at a later point in time when
  298. * SYN ACK is ultimatively being received.
  299. */
  300. TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
  301. }
  302. static void
  303. tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
  304. {
  305. if (inet_rsk(req)->ecn_ok)
  306. th->ece = 1;
  307. }
  308. /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
  309. * be sent.
  310. */
  311. static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
  312. struct tcphdr *th, int tcp_header_len)
  313. {
  314. struct tcp_sock *tp = tcp_sk(sk);
  315. if (tp->ecn_flags & TCP_ECN_OK) {
  316. /* Not-retransmitted data segment: set ECT and inject CWR. */
  317. if (skb->len != tcp_header_len &&
  318. !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
  319. INET_ECN_xmit(sk);
  320. if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
  321. tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
  322. th->cwr = 1;
  323. skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
  324. }
  325. } else if (!tcp_ca_needs_ecn(sk)) {
  326. /* ACK or retransmitted segment: clear ECT|CE */
  327. INET_ECN_dontxmit(sk);
  328. }
  329. if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
  330. th->ece = 1;
  331. }
  332. }
  333. /* Constructs common control bits of non-data skb. If SYN/FIN is present,
  334. * auto increment end seqno.
  335. */
  336. static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
  337. {
  338. skb->ip_summed = CHECKSUM_PARTIAL;
  339. TCP_SKB_CB(skb)->tcp_flags = flags;
  340. TCP_SKB_CB(skb)->sacked = 0;
  341. tcp_skb_pcount_set(skb, 1);
  342. TCP_SKB_CB(skb)->seq = seq;
  343. if (flags & (TCPHDR_SYN | TCPHDR_FIN))
  344. seq++;
  345. TCP_SKB_CB(skb)->end_seq = seq;
  346. }
  347. static inline bool tcp_urg_mode(const struct tcp_sock *tp)
  348. {
  349. return tp->snd_una != tp->snd_up;
  350. }
  351. #define OPTION_SACK_ADVERTISE (1 << 0)
  352. #define OPTION_TS (1 << 1)
  353. #define OPTION_MD5 (1 << 2)
  354. #define OPTION_WSCALE (1 << 3)
  355. #define OPTION_FAST_OPEN_COOKIE (1 << 8)
  356. #define OPTION_SMC (1 << 9)
  357. static void smc_options_write(__be32 *ptr, u16 *options)
  358. {
  359. #if IS_ENABLED(CONFIG_SMC)
  360. if (static_branch_unlikely(&tcp_have_smc)) {
  361. if (unlikely(OPTION_SMC & *options)) {
  362. *ptr++ = htonl((TCPOPT_NOP << 24) |
  363. (TCPOPT_NOP << 16) |
  364. (TCPOPT_EXP << 8) |
  365. (TCPOLEN_EXP_SMC_BASE));
  366. *ptr++ = htonl(TCPOPT_SMC_MAGIC);
  367. }
  368. }
  369. #endif
  370. }
  371. struct tcp_out_options {
  372. u16 options; /* bit field of OPTION_* */
  373. u16 mss; /* 0 to disable */
  374. u8 ws; /* window scale, 0 to disable */
  375. u8 num_sack_blocks; /* number of SACK blocks to include */
  376. u8 hash_size; /* bytes in hash_location */
  377. __u8 *hash_location; /* temporary pointer, overloaded */
  378. __u32 tsval, tsecr; /* need to include OPTION_TS */
  379. struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
  380. };
  381. /* Write previously computed TCP options to the packet.
  382. *
  383. * Beware: Something in the Internet is very sensitive to the ordering of
  384. * TCP options, we learned this through the hard way, so be careful here.
  385. * Luckily we can at least blame others for their non-compliance but from
  386. * inter-operability perspective it seems that we're somewhat stuck with
  387. * the ordering which we have been using if we want to keep working with
  388. * those broken things (not that it currently hurts anybody as there isn't
  389. * particular reason why the ordering would need to be changed).
  390. *
  391. * At least SACK_PERM as the first option is known to lead to a disaster
  392. * (but it may well be that other scenarios fail similarly).
  393. */
  394. static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
  395. struct tcp_out_options *opts)
  396. {
  397. u16 options = opts->options; /* mungable copy */
  398. if (unlikely(OPTION_MD5 & options)) {
  399. *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
  400. (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
  401. /* overload cookie hash location */
  402. opts->hash_location = (__u8 *)ptr;
  403. ptr += 4;
  404. }
  405. if (unlikely(opts->mss)) {
  406. *ptr++ = htonl((TCPOPT_MSS << 24) |
  407. (TCPOLEN_MSS << 16) |
  408. opts->mss);
  409. }
  410. if (likely(OPTION_TS & options)) {
  411. if (unlikely(OPTION_SACK_ADVERTISE & options)) {
  412. *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
  413. (TCPOLEN_SACK_PERM << 16) |
  414. (TCPOPT_TIMESTAMP << 8) |
  415. TCPOLEN_TIMESTAMP);
  416. options &= ~OPTION_SACK_ADVERTISE;
  417. } else {
  418. *ptr++ = htonl((TCPOPT_NOP << 24) |
  419. (TCPOPT_NOP << 16) |
  420. (TCPOPT_TIMESTAMP << 8) |
  421. TCPOLEN_TIMESTAMP);
  422. }
  423. *ptr++ = htonl(opts->tsval);
  424. *ptr++ = htonl(opts->tsecr);
  425. }
  426. if (unlikely(OPTION_SACK_ADVERTISE & options)) {
  427. *ptr++ = htonl((TCPOPT_NOP << 24) |
  428. (TCPOPT_NOP << 16) |
  429. (TCPOPT_SACK_PERM << 8) |
  430. TCPOLEN_SACK_PERM);
  431. }
  432. if (unlikely(OPTION_WSCALE & options)) {
  433. *ptr++ = htonl((TCPOPT_NOP << 24) |
  434. (TCPOPT_WINDOW << 16) |
  435. (TCPOLEN_WINDOW << 8) |
  436. opts->ws);
  437. }
  438. if (unlikely(opts->num_sack_blocks)) {
  439. struct tcp_sack_block *sp = tp->rx_opt.dsack ?
  440. tp->duplicate_sack : tp->selective_acks;
  441. int this_sack;
  442. *ptr++ = htonl((TCPOPT_NOP << 24) |
  443. (TCPOPT_NOP << 16) |
  444. (TCPOPT_SACK << 8) |
  445. (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
  446. TCPOLEN_SACK_PERBLOCK)));
  447. for (this_sack = 0; this_sack < opts->num_sack_blocks;
  448. ++this_sack) {
  449. *ptr++ = htonl(sp[this_sack].start_seq);
  450. *ptr++ = htonl(sp[this_sack].end_seq);
  451. }
  452. tp->rx_opt.dsack = 0;
  453. }
  454. if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
  455. struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
  456. u8 *p = (u8 *)ptr;
  457. u32 len; /* Fast Open option length */
  458. if (foc->exp) {
  459. len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
  460. *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
  461. TCPOPT_FASTOPEN_MAGIC);
  462. p += TCPOLEN_EXP_FASTOPEN_BASE;
  463. } else {
  464. len = TCPOLEN_FASTOPEN_BASE + foc->len;
  465. *p++ = TCPOPT_FASTOPEN;
  466. *p++ = len;
  467. }
  468. memcpy(p, foc->val, foc->len);
  469. if ((len & 3) == 2) {
  470. p[foc->len] = TCPOPT_NOP;
  471. p[foc->len + 1] = TCPOPT_NOP;
  472. }
  473. ptr += (len + 3) >> 2;
  474. }
  475. smc_options_write(ptr, &options);
  476. }
  477. static void smc_set_option(const struct tcp_sock *tp,
  478. struct tcp_out_options *opts,
  479. unsigned int *remaining)
  480. {
  481. #if IS_ENABLED(CONFIG_SMC)
  482. if (static_branch_unlikely(&tcp_have_smc)) {
  483. if (tp->syn_smc) {
  484. if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
  485. opts->options |= OPTION_SMC;
  486. *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
  487. }
  488. }
  489. }
  490. #endif
  491. }
  492. static void smc_set_option_cond(const struct tcp_sock *tp,
  493. const struct inet_request_sock *ireq,
  494. struct tcp_out_options *opts,
  495. unsigned int *remaining)
  496. {
  497. #if IS_ENABLED(CONFIG_SMC)
  498. if (static_branch_unlikely(&tcp_have_smc)) {
  499. if (tp->syn_smc && ireq->smc_ok) {
  500. if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
  501. opts->options |= OPTION_SMC;
  502. *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
  503. }
  504. }
  505. }
  506. #endif
  507. }
  508. /* Compute TCP options for SYN packets. This is not the final
  509. * network wire format yet.
  510. */
  511. static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
  512. struct tcp_out_options *opts,
  513. struct tcp_md5sig_key **md5)
  514. {
  515. struct tcp_sock *tp = tcp_sk(sk);
  516. unsigned int remaining = MAX_TCP_OPTION_SPACE;
  517. struct tcp_fastopen_request *fastopen = tp->fastopen_req;
  518. *md5 = NULL;
  519. #ifdef CONFIG_TCP_MD5SIG
  520. if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
  521. *md5 = tp->af_specific->md5_lookup(sk, sk);
  522. if (*md5) {
  523. opts->options |= OPTION_MD5;
  524. remaining -= TCPOLEN_MD5SIG_ALIGNED;
  525. }
  526. }
  527. #endif
  528. /* We always get an MSS option. The option bytes which will be seen in
  529. * normal data packets should timestamps be used, must be in the MSS
  530. * advertised. But we subtract them from tp->mss_cache so that
  531. * calculations in tcp_sendmsg are simpler etc. So account for this
  532. * fact here if necessary. If we don't do this correctly, as a
  533. * receiver we won't recognize data packets as being full sized when we
  534. * should, and thus we won't abide by the delayed ACK rules correctly.
  535. * SACKs don't matter, we never delay an ACK when we have any of those
  536. * going out. */
  537. opts->mss = tcp_advertise_mss(sk);
  538. remaining -= TCPOLEN_MSS_ALIGNED;
  539. if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
  540. opts->options |= OPTION_TS;
  541. opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
  542. opts->tsecr = tp->rx_opt.ts_recent;
  543. remaining -= TCPOLEN_TSTAMP_ALIGNED;
  544. }
  545. if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
  546. opts->ws = tp->rx_opt.rcv_wscale;
  547. opts->options |= OPTION_WSCALE;
  548. remaining -= TCPOLEN_WSCALE_ALIGNED;
  549. }
  550. if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
  551. opts->options |= OPTION_SACK_ADVERTISE;
  552. if (unlikely(!(OPTION_TS & opts->options)))
  553. remaining -= TCPOLEN_SACKPERM_ALIGNED;
  554. }
  555. if (fastopen && fastopen->cookie.len >= 0) {
  556. u32 need = fastopen->cookie.len;
  557. need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
  558. TCPOLEN_FASTOPEN_BASE;
  559. need = (need + 3) & ~3U; /* Align to 32 bits */
  560. if (remaining >= need) {
  561. opts->options |= OPTION_FAST_OPEN_COOKIE;
  562. opts->fastopen_cookie = &fastopen->cookie;
  563. remaining -= need;
  564. tp->syn_fastopen = 1;
  565. tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
  566. }
  567. }
  568. smc_set_option(tp, opts, &remaining);
  569. return MAX_TCP_OPTION_SPACE - remaining;
  570. }
  571. /* Set up TCP options for SYN-ACKs. */
  572. static unsigned int tcp_synack_options(const struct sock *sk,
  573. struct request_sock *req,
  574. unsigned int mss, struct sk_buff *skb,
  575. struct tcp_out_options *opts,
  576. const struct tcp_md5sig_key *md5,
  577. struct tcp_fastopen_cookie *foc)
  578. {
  579. struct inet_request_sock *ireq = inet_rsk(req);
  580. unsigned int remaining = MAX_TCP_OPTION_SPACE;
  581. #ifdef CONFIG_TCP_MD5SIG
  582. if (md5) {
  583. opts->options |= OPTION_MD5;
  584. remaining -= TCPOLEN_MD5SIG_ALIGNED;
  585. /* We can't fit any SACK blocks in a packet with MD5 + TS
  586. * options. There was discussion about disabling SACK
  587. * rather than TS in order to fit in better with old,
  588. * buggy kernels, but that was deemed to be unnecessary.
  589. */
  590. ireq->tstamp_ok &= !ireq->sack_ok;
  591. }
  592. #endif
  593. /* We always send an MSS option. */
  594. opts->mss = mss;
  595. remaining -= TCPOLEN_MSS_ALIGNED;
  596. if (likely(ireq->wscale_ok)) {
  597. opts->ws = ireq->rcv_wscale;
  598. opts->options |= OPTION_WSCALE;
  599. remaining -= TCPOLEN_WSCALE_ALIGNED;
  600. }
  601. if (likely(ireq->tstamp_ok)) {
  602. opts->options |= OPTION_TS;
  603. opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
  604. opts->tsecr = req->ts_recent;
  605. remaining -= TCPOLEN_TSTAMP_ALIGNED;
  606. }
  607. if (likely(ireq->sack_ok)) {
  608. opts->options |= OPTION_SACK_ADVERTISE;
  609. if (unlikely(!ireq->tstamp_ok))
  610. remaining -= TCPOLEN_SACKPERM_ALIGNED;
  611. }
  612. if (foc != NULL && foc->len >= 0) {
  613. u32 need = foc->len;
  614. need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
  615. TCPOLEN_FASTOPEN_BASE;
  616. need = (need + 3) & ~3U; /* Align to 32 bits */
  617. if (remaining >= need) {
  618. opts->options |= OPTION_FAST_OPEN_COOKIE;
  619. opts->fastopen_cookie = foc;
  620. remaining -= need;
  621. }
  622. }
  623. smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
  624. return MAX_TCP_OPTION_SPACE - remaining;
  625. }
  626. /* Compute TCP options for ESTABLISHED sockets. This is not the
  627. * final wire format yet.
  628. */
  629. static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
  630. struct tcp_out_options *opts,
  631. struct tcp_md5sig_key **md5)
  632. {
  633. struct tcp_sock *tp = tcp_sk(sk);
  634. unsigned int size = 0;
  635. unsigned int eff_sacks;
  636. opts->options = 0;
  637. *md5 = NULL;
  638. #ifdef CONFIG_TCP_MD5SIG
  639. if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
  640. *md5 = tp->af_specific->md5_lookup(sk, sk);
  641. if (*md5) {
  642. opts->options |= OPTION_MD5;
  643. size += TCPOLEN_MD5SIG_ALIGNED;
  644. }
  645. }
  646. #endif
  647. if (likely(tp->rx_opt.tstamp_ok)) {
  648. opts->options |= OPTION_TS;
  649. opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
  650. opts->tsecr = tp->rx_opt.ts_recent;
  651. size += TCPOLEN_TSTAMP_ALIGNED;
  652. }
  653. eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
  654. if (unlikely(eff_sacks)) {
  655. const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
  656. opts->num_sack_blocks =
  657. min_t(unsigned int, eff_sacks,
  658. (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
  659. TCPOLEN_SACK_PERBLOCK);
  660. size += TCPOLEN_SACK_BASE_ALIGNED +
  661. opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
  662. }
  663. return size;
  664. }
  665. /* TCP SMALL QUEUES (TSQ)
  666. *
  667. * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
  668. * to reduce RTT and bufferbloat.
  669. * We do this using a special skb destructor (tcp_wfree).
  670. *
  671. * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
  672. * needs to be reallocated in a driver.
  673. * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
  674. *
  675. * Since transmit from skb destructor is forbidden, we use a tasklet
  676. * to process all sockets that eventually need to send more skbs.
  677. * We use one tasklet per cpu, with its own queue of sockets.
  678. */
  679. struct tsq_tasklet {
  680. struct tasklet_struct tasklet;
  681. struct list_head head; /* queue of tcp sockets */
  682. };
  683. static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
  684. static void tcp_tsq_write(struct sock *sk)
  685. {
  686. if ((1 << sk->sk_state) &
  687. (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
  688. TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
  689. struct tcp_sock *tp = tcp_sk(sk);
  690. if (tp->lost_out > tp->retrans_out &&
  691. tp->snd_cwnd > tcp_packets_in_flight(tp)) {
  692. tcp_mstamp_refresh(tp);
  693. tcp_xmit_retransmit_queue(sk);
  694. }
  695. tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
  696. 0, GFP_ATOMIC);
  697. }
  698. }
  699. static void tcp_tsq_handler(struct sock *sk)
  700. {
  701. bh_lock_sock(sk);
  702. if (!sock_owned_by_user(sk))
  703. tcp_tsq_write(sk);
  704. else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
  705. sock_hold(sk);
  706. bh_unlock_sock(sk);
  707. }
  708. /*
  709. * One tasklet per cpu tries to send more skbs.
  710. * We run in tasklet context but need to disable irqs when
  711. * transferring tsq->head because tcp_wfree() might
  712. * interrupt us (non NAPI drivers)
  713. */
  714. static void tcp_tasklet_func(unsigned long data)
  715. {
  716. struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
  717. LIST_HEAD(list);
  718. unsigned long flags;
  719. struct list_head *q, *n;
  720. struct tcp_sock *tp;
  721. struct sock *sk;
  722. local_irq_save(flags);
  723. list_splice_init(&tsq->head, &list);
  724. local_irq_restore(flags);
  725. list_for_each_safe(q, n, &list) {
  726. tp = list_entry(q, struct tcp_sock, tsq_node);
  727. list_del(&tp->tsq_node);
  728. sk = (struct sock *)tp;
  729. smp_mb__before_atomic();
  730. clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
  731. tcp_tsq_handler(sk);
  732. sk_free(sk);
  733. }
  734. }
  735. #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
  736. TCPF_WRITE_TIMER_DEFERRED | \
  737. TCPF_DELACK_TIMER_DEFERRED | \
  738. TCPF_MTU_REDUCED_DEFERRED)
  739. /**
  740. * tcp_release_cb - tcp release_sock() callback
  741. * @sk: socket
  742. *
  743. * called from release_sock() to perform protocol dependent
  744. * actions before socket release.
  745. */
  746. void tcp_release_cb(struct sock *sk)
  747. {
  748. unsigned long flags, nflags;
  749. /* perform an atomic operation only if at least one flag is set */
  750. do {
  751. flags = sk->sk_tsq_flags;
  752. if (!(flags & TCP_DEFERRED_ALL))
  753. return;
  754. nflags = flags & ~TCP_DEFERRED_ALL;
  755. } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
  756. if (flags & TCPF_TSQ_DEFERRED) {
  757. tcp_tsq_write(sk);
  758. __sock_put(sk);
  759. }
  760. /* Here begins the tricky part :
  761. * We are called from release_sock() with :
  762. * 1) BH disabled
  763. * 2) sk_lock.slock spinlock held
  764. * 3) socket owned by us (sk->sk_lock.owned == 1)
  765. *
  766. * But following code is meant to be called from BH handlers,
  767. * so we should keep BH disabled, but early release socket ownership
  768. */
  769. sock_release_ownership(sk);
  770. if (flags & TCPF_WRITE_TIMER_DEFERRED) {
  771. tcp_write_timer_handler(sk);
  772. __sock_put(sk);
  773. }
  774. if (flags & TCPF_DELACK_TIMER_DEFERRED) {
  775. tcp_delack_timer_handler(sk);
  776. __sock_put(sk);
  777. }
  778. if (flags & TCPF_MTU_REDUCED_DEFERRED) {
  779. inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
  780. __sock_put(sk);
  781. }
  782. }
  783. EXPORT_SYMBOL(tcp_release_cb);
  784. void __init tcp_tasklet_init(void)
  785. {
  786. int i;
  787. for_each_possible_cpu(i) {
  788. struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
  789. INIT_LIST_HEAD(&tsq->head);
  790. tasklet_init(&tsq->tasklet,
  791. tcp_tasklet_func,
  792. (unsigned long)tsq);
  793. }
  794. }
  795. /*
  796. * Write buffer destructor automatically called from kfree_skb.
  797. * We can't xmit new skbs from this context, as we might already
  798. * hold qdisc lock.
  799. */
  800. void tcp_wfree(struct sk_buff *skb)
  801. {
  802. struct sock *sk = skb->sk;
  803. struct tcp_sock *tp = tcp_sk(sk);
  804. unsigned long flags, nval, oval;
  805. /* Keep one reference on sk_wmem_alloc.
  806. * Will be released by sk_free() from here or tcp_tasklet_func()
  807. */
  808. WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
  809. /* If this softirq is serviced by ksoftirqd, we are likely under stress.
  810. * Wait until our queues (qdisc + devices) are drained.
  811. * This gives :
  812. * - less callbacks to tcp_write_xmit(), reducing stress (batches)
  813. * - chance for incoming ACK (processed by another cpu maybe)
  814. * to migrate this flow (skb->ooo_okay will be eventually set)
  815. */
  816. if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
  817. goto out;
  818. for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
  819. struct tsq_tasklet *tsq;
  820. bool empty;
  821. if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
  822. goto out;
  823. nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
  824. nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
  825. if (nval != oval)
  826. continue;
  827. /* queue this socket to tasklet queue */
  828. local_irq_save(flags);
  829. tsq = this_cpu_ptr(&tsq_tasklet);
  830. empty = list_empty(&tsq->head);
  831. list_add(&tp->tsq_node, &tsq->head);
  832. if (empty)
  833. tasklet_schedule(&tsq->tasklet);
  834. local_irq_restore(flags);
  835. return;
  836. }
  837. out:
  838. sk_free(sk);
  839. }
  840. /* Note: Called under soft irq.
  841. * We can call TCP stack right away, unless socket is owned by user.
  842. */
  843. enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
  844. {
  845. struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
  846. struct sock *sk = (struct sock *)tp;
  847. tcp_tsq_handler(sk);
  848. sock_put(sk);
  849. return HRTIMER_NORESTART;
  850. }
  851. static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
  852. {
  853. u64 len_ns;
  854. u32 rate;
  855. if (!tcp_needs_internal_pacing(sk))
  856. return;
  857. rate = sk->sk_pacing_rate;
  858. if (!rate || rate == ~0U)
  859. return;
  860. len_ns = (u64)skb->len * NSEC_PER_SEC;
  861. do_div(len_ns, rate);
  862. hrtimer_start(&tcp_sk(sk)->pacing_timer,
  863. ktime_add_ns(ktime_get(), len_ns),
  864. HRTIMER_MODE_ABS_PINNED_SOFT);
  865. sock_hold(sk);
  866. }
  867. static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb)
  868. {
  869. skb->skb_mstamp = tp->tcp_mstamp;
  870. list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
  871. }
  872. /* This routine actually transmits TCP packets queued in by
  873. * tcp_do_sendmsg(). This is used by both the initial
  874. * transmission and possible later retransmissions.
  875. * All SKB's seen here are completely headerless. It is our
  876. * job to build the TCP header, and pass the packet down to
  877. * IP so it can do the same plus pass the packet off to the
  878. * device.
  879. *
  880. * We are working here with either a clone of the original
  881. * SKB, or a fresh unique copy made by the retransmit engine.
  882. */
  883. static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
  884. int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
  885. {
  886. const struct inet_connection_sock *icsk = inet_csk(sk);
  887. struct inet_sock *inet;
  888. struct tcp_sock *tp;
  889. struct tcp_skb_cb *tcb;
  890. struct tcp_out_options opts;
  891. unsigned int tcp_options_size, tcp_header_size;
  892. struct sk_buff *oskb = NULL;
  893. struct tcp_md5sig_key *md5;
  894. struct tcphdr *th;
  895. int err;
  896. BUG_ON(!skb || !tcp_skb_pcount(skb));
  897. tp = tcp_sk(sk);
  898. if (clone_it) {
  899. TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
  900. - tp->snd_una;
  901. oskb = skb;
  902. tcp_skb_tsorted_save(oskb) {
  903. if (unlikely(skb_cloned(oskb)))
  904. skb = pskb_copy(oskb, gfp_mask);
  905. else
  906. skb = skb_clone(oskb, gfp_mask);
  907. } tcp_skb_tsorted_restore(oskb);
  908. if (unlikely(!skb))
  909. return -ENOBUFS;
  910. }
  911. skb->skb_mstamp = tp->tcp_mstamp;
  912. inet = inet_sk(sk);
  913. tcb = TCP_SKB_CB(skb);
  914. memset(&opts, 0, sizeof(opts));
  915. if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
  916. tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
  917. else
  918. tcp_options_size = tcp_established_options(sk, skb, &opts,
  919. &md5);
  920. tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
  921. /* if no packet is in qdisc/device queue, then allow XPS to select
  922. * another queue. We can be called from tcp_tsq_handler()
  923. * which holds one reference to sk.
  924. *
  925. * TODO: Ideally, in-flight pure ACK packets should not matter here.
  926. * One way to get this would be to set skb->truesize = 2 on them.
  927. */
  928. skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
  929. /* If we had to use memory reserve to allocate this skb,
  930. * this might cause drops if packet is looped back :
  931. * Other socket might not have SOCK_MEMALLOC.
  932. * Packets not looped back do not care about pfmemalloc.
  933. */
  934. skb->pfmemalloc = 0;
  935. skb_push(skb, tcp_header_size);
  936. skb_reset_transport_header(skb);
  937. skb_orphan(skb);
  938. skb->sk = sk;
  939. skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
  940. skb_set_hash_from_sk(skb, sk);
  941. refcount_add(skb->truesize, &sk->sk_wmem_alloc);
  942. skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
  943. /* Build TCP header and checksum it. */
  944. th = (struct tcphdr *)skb->data;
  945. th->source = inet->inet_sport;
  946. th->dest = inet->inet_dport;
  947. th->seq = htonl(tcb->seq);
  948. th->ack_seq = htonl(rcv_nxt);
  949. *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
  950. tcb->tcp_flags);
  951. th->check = 0;
  952. th->urg_ptr = 0;
  953. /* The urg_mode check is necessary during a below snd_una win probe */
  954. if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
  955. if (before(tp->snd_up, tcb->seq + 0x10000)) {
  956. th->urg_ptr = htons(tp->snd_up - tcb->seq);
  957. th->urg = 1;
  958. } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
  959. th->urg_ptr = htons(0xFFFF);
  960. th->urg = 1;
  961. }
  962. }
  963. tcp_options_write((__be32 *)(th + 1), tp, &opts);
  964. skb_shinfo(skb)->gso_type = sk->sk_gso_type;
  965. if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
  966. th->window = htons(tcp_select_window(sk));
  967. tcp_ecn_send(sk, skb, th, tcp_header_size);
  968. } else {
  969. /* RFC1323: The window in SYN & SYN/ACK segments
  970. * is never scaled.
  971. */
  972. th->window = htons(min(tp->rcv_wnd, 65535U));
  973. }
  974. #ifdef CONFIG_TCP_MD5SIG
  975. /* Calculate the MD5 hash, as we have all we need now */
  976. if (md5) {
  977. sk_nocaps_add(sk, NETIF_F_GSO_MASK);
  978. tp->af_specific->calc_md5_hash(opts.hash_location,
  979. md5, sk, skb);
  980. }
  981. #endif
  982. icsk->icsk_af_ops->send_check(sk, skb);
  983. if (likely(tcb->tcp_flags & TCPHDR_ACK))
  984. tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
  985. if (skb->len != tcp_header_size) {
  986. tcp_event_data_sent(tp, sk);
  987. tp->data_segs_out += tcp_skb_pcount(skb);
  988. tp->bytes_sent += skb->len - tcp_header_size;
  989. tcp_internal_pacing(sk, skb);
  990. }
  991. if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
  992. TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
  993. tcp_skb_pcount(skb));
  994. tp->segs_out += tcp_skb_pcount(skb);
  995. /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
  996. skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
  997. skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
  998. /* Our usage of tstamp should remain private */
  999. skb->tstamp = 0;
  1000. /* Cleanup our debris for IP stacks */
  1001. memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
  1002. sizeof(struct inet6_skb_parm)));
  1003. err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
  1004. if (unlikely(err > 0)) {
  1005. tcp_enter_cwr(sk);
  1006. err = net_xmit_eval(err);
  1007. }
  1008. if (!err && oskb) {
  1009. tcp_update_skb_after_send(tp, oskb);
  1010. tcp_rate_skb_sent(sk, oskb);
  1011. }
  1012. return err;
  1013. }
  1014. static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
  1015. gfp_t gfp_mask)
  1016. {
  1017. return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
  1018. tcp_sk(sk)->rcv_nxt);
  1019. }
  1020. /* This routine just queues the buffer for sending.
  1021. *
  1022. * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
  1023. * otherwise socket can stall.
  1024. */
  1025. static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
  1026. {
  1027. struct tcp_sock *tp = tcp_sk(sk);
  1028. /* Advance write_seq and place onto the write_queue. */
  1029. tp->write_seq = TCP_SKB_CB(skb)->end_seq;
  1030. __skb_header_release(skb);
  1031. tcp_add_write_queue_tail(sk, skb);
  1032. sk->sk_wmem_queued += skb->truesize;
  1033. sk_mem_charge(sk, skb->truesize);
  1034. }
  1035. /* Initialize TSO segments for a packet. */
  1036. static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
  1037. {
  1038. if (skb->len <= mss_now) {
  1039. /* Avoid the costly divide in the normal
  1040. * non-TSO case.
  1041. */
  1042. tcp_skb_pcount_set(skb, 1);
  1043. TCP_SKB_CB(skb)->tcp_gso_size = 0;
  1044. } else {
  1045. tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
  1046. TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
  1047. }
  1048. }
  1049. /* Pcount in the middle of the write queue got changed, we need to do various
  1050. * tweaks to fix counters
  1051. */
  1052. static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
  1053. {
  1054. struct tcp_sock *tp = tcp_sk(sk);
  1055. tp->packets_out -= decr;
  1056. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
  1057. tp->sacked_out -= decr;
  1058. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
  1059. tp->retrans_out -= decr;
  1060. if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
  1061. tp->lost_out -= decr;
  1062. /* Reno case is special. Sigh... */
  1063. if (tcp_is_reno(tp) && decr > 0)
  1064. tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
  1065. if (tp->lost_skb_hint &&
  1066. before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
  1067. (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
  1068. tp->lost_cnt_hint -= decr;
  1069. tcp_verify_left_out(tp);
  1070. }
  1071. static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
  1072. {
  1073. return TCP_SKB_CB(skb)->txstamp_ack ||
  1074. (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
  1075. }
  1076. static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
  1077. {
  1078. struct skb_shared_info *shinfo = skb_shinfo(skb);
  1079. if (unlikely(tcp_has_tx_tstamp(skb)) &&
  1080. !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
  1081. struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
  1082. u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
  1083. shinfo->tx_flags &= ~tsflags;
  1084. shinfo2->tx_flags |= tsflags;
  1085. swap(shinfo->tskey, shinfo2->tskey);
  1086. TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
  1087. TCP_SKB_CB(skb)->txstamp_ack = 0;
  1088. }
  1089. }
  1090. static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
  1091. {
  1092. TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
  1093. TCP_SKB_CB(skb)->eor = 0;
  1094. }
  1095. /* Insert buff after skb on the write or rtx queue of sk. */
  1096. static void tcp_insert_write_queue_after(struct sk_buff *skb,
  1097. struct sk_buff *buff,
  1098. struct sock *sk,
  1099. enum tcp_queue tcp_queue)
  1100. {
  1101. if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
  1102. __skb_queue_after(&sk->sk_write_queue, skb, buff);
  1103. else
  1104. tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
  1105. }
  1106. /* Function to create two new TCP segments. Shrinks the given segment
  1107. * to the specified size and appends a new segment with the rest of the
  1108. * packet to the list. This won't be called frequently, I hope.
  1109. * Remember, these are still headerless SKBs at this point.
  1110. */
  1111. int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
  1112. struct sk_buff *skb, u32 len,
  1113. unsigned int mss_now, gfp_t gfp)
  1114. {
  1115. struct tcp_sock *tp = tcp_sk(sk);
  1116. struct sk_buff *buff;
  1117. int nsize, old_factor;
  1118. int nlen;
  1119. u8 flags;
  1120. if (WARN_ON(len > skb->len))
  1121. return -EINVAL;
  1122. nsize = skb_headlen(skb) - len;
  1123. if (nsize < 0)
  1124. nsize = 0;
  1125. if (skb_unclone(skb, gfp))
  1126. return -ENOMEM;
  1127. /* Get a new skb... force flag on. */
  1128. buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
  1129. if (!buff)
  1130. return -ENOMEM; /* We'll just try again later. */
  1131. sk->sk_wmem_queued += buff->truesize;
  1132. sk_mem_charge(sk, buff->truesize);
  1133. nlen = skb->len - len - nsize;
  1134. buff->truesize += nlen;
  1135. skb->truesize -= nlen;
  1136. /* Correct the sequence numbers. */
  1137. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  1138. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  1139. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  1140. /* PSH and FIN should only be set in the second packet. */
  1141. flags = TCP_SKB_CB(skb)->tcp_flags;
  1142. TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
  1143. TCP_SKB_CB(buff)->tcp_flags = flags;
  1144. TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
  1145. tcp_skb_fragment_eor(skb, buff);
  1146. skb_split(skb, buff, len);
  1147. buff->ip_summed = CHECKSUM_PARTIAL;
  1148. buff->tstamp = skb->tstamp;
  1149. tcp_fragment_tstamp(skb, buff);
  1150. old_factor = tcp_skb_pcount(skb);
  1151. /* Fix up tso_factor for both original and new SKB. */
  1152. tcp_set_skb_tso_segs(skb, mss_now);
  1153. tcp_set_skb_tso_segs(buff, mss_now);
  1154. /* Update delivered info for the new segment */
  1155. TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
  1156. /* If this packet has been sent out already, we must
  1157. * adjust the various packet counters.
  1158. */
  1159. if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
  1160. int diff = old_factor - tcp_skb_pcount(skb) -
  1161. tcp_skb_pcount(buff);
  1162. if (diff)
  1163. tcp_adjust_pcount(sk, skb, diff);
  1164. }
  1165. /* Link BUFF into the send queue. */
  1166. __skb_header_release(buff);
  1167. tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
  1168. if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
  1169. list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
  1170. return 0;
  1171. }
  1172. /* This is similar to __pskb_pull_tail(). The difference is that pulled
  1173. * data is not copied, but immediately discarded.
  1174. */
  1175. static int __pskb_trim_head(struct sk_buff *skb, int len)
  1176. {
  1177. struct skb_shared_info *shinfo;
  1178. int i, k, eat;
  1179. eat = min_t(int, len, skb_headlen(skb));
  1180. if (eat) {
  1181. __skb_pull(skb, eat);
  1182. len -= eat;
  1183. if (!len)
  1184. return 0;
  1185. }
  1186. eat = len;
  1187. k = 0;
  1188. shinfo = skb_shinfo(skb);
  1189. for (i = 0; i < shinfo->nr_frags; i++) {
  1190. int size = skb_frag_size(&shinfo->frags[i]);
  1191. if (size <= eat) {
  1192. skb_frag_unref(skb, i);
  1193. eat -= size;
  1194. } else {
  1195. shinfo->frags[k] = shinfo->frags[i];
  1196. if (eat) {
  1197. shinfo->frags[k].page_offset += eat;
  1198. skb_frag_size_sub(&shinfo->frags[k], eat);
  1199. eat = 0;
  1200. }
  1201. k++;
  1202. }
  1203. }
  1204. shinfo->nr_frags = k;
  1205. skb->data_len -= len;
  1206. skb->len = skb->data_len;
  1207. return len;
  1208. }
  1209. /* Remove acked data from a packet in the transmit queue. */
  1210. int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
  1211. {
  1212. u32 delta_truesize;
  1213. if (skb_unclone(skb, GFP_ATOMIC))
  1214. return -ENOMEM;
  1215. delta_truesize = __pskb_trim_head(skb, len);
  1216. TCP_SKB_CB(skb)->seq += len;
  1217. skb->ip_summed = CHECKSUM_PARTIAL;
  1218. if (delta_truesize) {
  1219. skb->truesize -= delta_truesize;
  1220. sk->sk_wmem_queued -= delta_truesize;
  1221. sk_mem_uncharge(sk, delta_truesize);
  1222. sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
  1223. }
  1224. /* Any change of skb->len requires recalculation of tso factor. */
  1225. if (tcp_skb_pcount(skb) > 1)
  1226. tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
  1227. return 0;
  1228. }
  1229. /* Calculate MSS not accounting any TCP options. */
  1230. static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
  1231. {
  1232. const struct tcp_sock *tp = tcp_sk(sk);
  1233. const struct inet_connection_sock *icsk = inet_csk(sk);
  1234. int mss_now;
  1235. /* Calculate base mss without TCP options:
  1236. It is MMS_S - sizeof(tcphdr) of rfc1122
  1237. */
  1238. mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
  1239. /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
  1240. if (icsk->icsk_af_ops->net_frag_header_len) {
  1241. const struct dst_entry *dst = __sk_dst_get(sk);
  1242. if (dst && dst_allfrag(dst))
  1243. mss_now -= icsk->icsk_af_ops->net_frag_header_len;
  1244. }
  1245. /* Clamp it (mss_clamp does not include tcp options) */
  1246. if (mss_now > tp->rx_opt.mss_clamp)
  1247. mss_now = tp->rx_opt.mss_clamp;
  1248. /* Now subtract optional transport overhead */
  1249. mss_now -= icsk->icsk_ext_hdr_len;
  1250. /* Then reserve room for full set of TCP options and 8 bytes of data */
  1251. if (mss_now < 48)
  1252. mss_now = 48;
  1253. return mss_now;
  1254. }
  1255. /* Calculate MSS. Not accounting for SACKs here. */
  1256. int tcp_mtu_to_mss(struct sock *sk, int pmtu)
  1257. {
  1258. /* Subtract TCP options size, not including SACKs */
  1259. return __tcp_mtu_to_mss(sk, pmtu) -
  1260. (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
  1261. }
  1262. /* Inverse of above */
  1263. int tcp_mss_to_mtu(struct sock *sk, int mss)
  1264. {
  1265. const struct tcp_sock *tp = tcp_sk(sk);
  1266. const struct inet_connection_sock *icsk = inet_csk(sk);
  1267. int mtu;
  1268. mtu = mss +
  1269. tp->tcp_header_len +
  1270. icsk->icsk_ext_hdr_len +
  1271. icsk->icsk_af_ops->net_header_len;
  1272. /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
  1273. if (icsk->icsk_af_ops->net_frag_header_len) {
  1274. const struct dst_entry *dst = __sk_dst_get(sk);
  1275. if (dst && dst_allfrag(dst))
  1276. mtu += icsk->icsk_af_ops->net_frag_header_len;
  1277. }
  1278. return mtu;
  1279. }
  1280. EXPORT_SYMBOL(tcp_mss_to_mtu);
  1281. /* MTU probing init per socket */
  1282. void tcp_mtup_init(struct sock *sk)
  1283. {
  1284. struct tcp_sock *tp = tcp_sk(sk);
  1285. struct inet_connection_sock *icsk = inet_csk(sk);
  1286. struct net *net = sock_net(sk);
  1287. icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
  1288. icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
  1289. icsk->icsk_af_ops->net_header_len;
  1290. icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
  1291. icsk->icsk_mtup.probe_size = 0;
  1292. if (icsk->icsk_mtup.enabled)
  1293. icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
  1294. }
  1295. EXPORT_SYMBOL(tcp_mtup_init);
  1296. /* This function synchronize snd mss to current pmtu/exthdr set.
  1297. tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
  1298. for TCP options, but includes only bare TCP header.
  1299. tp->rx_opt.mss_clamp is mss negotiated at connection setup.
  1300. It is minimum of user_mss and mss received with SYN.
  1301. It also does not include TCP options.
  1302. inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
  1303. tp->mss_cache is current effective sending mss, including
  1304. all tcp options except for SACKs. It is evaluated,
  1305. taking into account current pmtu, but never exceeds
  1306. tp->rx_opt.mss_clamp.
  1307. NOTE1. rfc1122 clearly states that advertised MSS
  1308. DOES NOT include either tcp or ip options.
  1309. NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
  1310. are READ ONLY outside this function. --ANK (980731)
  1311. */
  1312. unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
  1313. {
  1314. struct tcp_sock *tp = tcp_sk(sk);
  1315. struct inet_connection_sock *icsk = inet_csk(sk);
  1316. int mss_now;
  1317. if (icsk->icsk_mtup.search_high > pmtu)
  1318. icsk->icsk_mtup.search_high = pmtu;
  1319. mss_now = tcp_mtu_to_mss(sk, pmtu);
  1320. mss_now = tcp_bound_to_half_wnd(tp, mss_now);
  1321. /* And store cached results */
  1322. icsk->icsk_pmtu_cookie = pmtu;
  1323. if (icsk->icsk_mtup.enabled)
  1324. mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
  1325. tp->mss_cache = mss_now;
  1326. return mss_now;
  1327. }
  1328. EXPORT_SYMBOL(tcp_sync_mss);
  1329. /* Compute the current effective MSS, taking SACKs and IP options,
  1330. * and even PMTU discovery events into account.
  1331. */
  1332. unsigned int tcp_current_mss(struct sock *sk)
  1333. {
  1334. const struct tcp_sock *tp = tcp_sk(sk);
  1335. const struct dst_entry *dst = __sk_dst_get(sk);
  1336. u32 mss_now;
  1337. unsigned int header_len;
  1338. struct tcp_out_options opts;
  1339. struct tcp_md5sig_key *md5;
  1340. mss_now = tp->mss_cache;
  1341. if (dst) {
  1342. u32 mtu = dst_mtu(dst);
  1343. if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
  1344. mss_now = tcp_sync_mss(sk, mtu);
  1345. }
  1346. header_len = tcp_established_options(sk, NULL, &opts, &md5) +
  1347. sizeof(struct tcphdr);
  1348. /* The mss_cache is sized based on tp->tcp_header_len, which assumes
  1349. * some common options. If this is an odd packet (because we have SACK
  1350. * blocks etc) then our calculated header_len will be different, and
  1351. * we have to adjust mss_now correspondingly */
  1352. if (header_len != tp->tcp_header_len) {
  1353. int delta = (int) header_len - tp->tcp_header_len;
  1354. mss_now -= delta;
  1355. }
  1356. return mss_now;
  1357. }
  1358. /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
  1359. * As additional protections, we do not touch cwnd in retransmission phases,
  1360. * and if application hit its sndbuf limit recently.
  1361. */
  1362. static void tcp_cwnd_application_limited(struct sock *sk)
  1363. {
  1364. struct tcp_sock *tp = tcp_sk(sk);
  1365. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
  1366. sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
  1367. /* Limited by application or receiver window. */
  1368. u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
  1369. u32 win_used = max(tp->snd_cwnd_used, init_win);
  1370. if (win_used < tp->snd_cwnd) {
  1371. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  1372. tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
  1373. }
  1374. tp->snd_cwnd_used = 0;
  1375. }
  1376. tp->snd_cwnd_stamp = tcp_jiffies32;
  1377. }
  1378. static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
  1379. {
  1380. const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
  1381. struct tcp_sock *tp = tcp_sk(sk);
  1382. /* Track the maximum number of outstanding packets in each
  1383. * window, and remember whether we were cwnd-limited then.
  1384. */
  1385. if (!before(tp->snd_una, tp->max_packets_seq) ||
  1386. tp->packets_out > tp->max_packets_out) {
  1387. tp->max_packets_out = tp->packets_out;
  1388. tp->max_packets_seq = tp->snd_nxt;
  1389. tp->is_cwnd_limited = is_cwnd_limited;
  1390. }
  1391. if (tcp_is_cwnd_limited(sk)) {
  1392. /* Network is feed fully. */
  1393. tp->snd_cwnd_used = 0;
  1394. tp->snd_cwnd_stamp = tcp_jiffies32;
  1395. } else {
  1396. /* Network starves. */
  1397. if (tp->packets_out > tp->snd_cwnd_used)
  1398. tp->snd_cwnd_used = tp->packets_out;
  1399. if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
  1400. (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
  1401. !ca_ops->cong_control)
  1402. tcp_cwnd_application_limited(sk);
  1403. /* The following conditions together indicate the starvation
  1404. * is caused by insufficient sender buffer:
  1405. * 1) just sent some data (see tcp_write_xmit)
  1406. * 2) not cwnd limited (this else condition)
  1407. * 3) no more data to send (tcp_write_queue_empty())
  1408. * 4) application is hitting buffer limit (SOCK_NOSPACE)
  1409. */
  1410. if (tcp_write_queue_empty(sk) && sk->sk_socket &&
  1411. test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
  1412. (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
  1413. tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
  1414. }
  1415. }
  1416. /* Minshall's variant of the Nagle send check. */
  1417. static bool tcp_minshall_check(const struct tcp_sock *tp)
  1418. {
  1419. return after(tp->snd_sml, tp->snd_una) &&
  1420. !after(tp->snd_sml, tp->snd_nxt);
  1421. }
  1422. /* Update snd_sml if this skb is under mss
  1423. * Note that a TSO packet might end with a sub-mss segment
  1424. * The test is really :
  1425. * if ((skb->len % mss) != 0)
  1426. * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
  1427. * But we can avoid doing the divide again given we already have
  1428. * skb_pcount = skb->len / mss_now
  1429. */
  1430. static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
  1431. const struct sk_buff *skb)
  1432. {
  1433. if (skb->len < tcp_skb_pcount(skb) * mss_now)
  1434. tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
  1435. }
  1436. /* Return false, if packet can be sent now without violation Nagle's rules:
  1437. * 1. It is full sized. (provided by caller in %partial bool)
  1438. * 2. Or it contains FIN. (already checked by caller)
  1439. * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
  1440. * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
  1441. * With Minshall's modification: all sent small packets are ACKed.
  1442. */
  1443. static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
  1444. int nonagle)
  1445. {
  1446. return partial &&
  1447. ((nonagle & TCP_NAGLE_CORK) ||
  1448. (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
  1449. }
  1450. /* Return how many segs we'd like on a TSO packet,
  1451. * to send one TSO packet per ms
  1452. */
  1453. static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
  1454. int min_tso_segs)
  1455. {
  1456. u32 bytes, segs;
  1457. bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
  1458. sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
  1459. /* Goal is to send at least one packet per ms,
  1460. * not one big TSO packet every 100 ms.
  1461. * This preserves ACK clocking and is consistent
  1462. * with tcp_tso_should_defer() heuristic.
  1463. */
  1464. segs = max_t(u32, bytes / mss_now, min_tso_segs);
  1465. return segs;
  1466. }
  1467. /* Return the number of segments we want in the skb we are transmitting.
  1468. * See if congestion control module wants to decide; otherwise, autosize.
  1469. */
  1470. static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
  1471. {
  1472. const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
  1473. u32 min_tso, tso_segs;
  1474. min_tso = ca_ops->min_tso_segs ?
  1475. ca_ops->min_tso_segs(sk) :
  1476. sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
  1477. tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
  1478. return min_t(u32, tso_segs, sk->sk_gso_max_segs);
  1479. }
  1480. /* Returns the portion of skb which can be sent right away */
  1481. static unsigned int tcp_mss_split_point(const struct sock *sk,
  1482. const struct sk_buff *skb,
  1483. unsigned int mss_now,
  1484. unsigned int max_segs,
  1485. int nonagle)
  1486. {
  1487. const struct tcp_sock *tp = tcp_sk(sk);
  1488. u32 partial, needed, window, max_len;
  1489. window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  1490. max_len = mss_now * max_segs;
  1491. if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
  1492. return max_len;
  1493. needed = min(skb->len, window);
  1494. if (max_len <= needed)
  1495. return max_len;
  1496. partial = needed % mss_now;
  1497. /* If last segment is not a full MSS, check if Nagle rules allow us
  1498. * to include this last segment in this skb.
  1499. * Otherwise, we'll split the skb at last MSS boundary
  1500. */
  1501. if (tcp_nagle_check(partial != 0, tp, nonagle))
  1502. return needed - partial;
  1503. return needed;
  1504. }
  1505. /* Can at least one segment of SKB be sent right now, according to the
  1506. * congestion window rules? If so, return how many segments are allowed.
  1507. */
  1508. static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
  1509. const struct sk_buff *skb)
  1510. {
  1511. u32 in_flight, cwnd, halfcwnd;
  1512. /* Don't be strict about the congestion window for the final FIN. */
  1513. if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
  1514. tcp_skb_pcount(skb) == 1)
  1515. return 1;
  1516. in_flight = tcp_packets_in_flight(tp);
  1517. cwnd = tp->snd_cwnd;
  1518. if (in_flight >= cwnd)
  1519. return 0;
  1520. /* For better scheduling, ensure we have at least
  1521. * 2 GSO packets in flight.
  1522. */
  1523. halfcwnd = max(cwnd >> 1, 1U);
  1524. return min(halfcwnd, cwnd - in_flight);
  1525. }
  1526. /* Initialize TSO state of a skb.
  1527. * This must be invoked the first time we consider transmitting
  1528. * SKB onto the wire.
  1529. */
  1530. static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
  1531. {
  1532. int tso_segs = tcp_skb_pcount(skb);
  1533. if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
  1534. tcp_set_skb_tso_segs(skb, mss_now);
  1535. tso_segs = tcp_skb_pcount(skb);
  1536. }
  1537. return tso_segs;
  1538. }
  1539. /* Return true if the Nagle test allows this packet to be
  1540. * sent now.
  1541. */
  1542. static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
  1543. unsigned int cur_mss, int nonagle)
  1544. {
  1545. /* Nagle rule does not apply to frames, which sit in the middle of the
  1546. * write_queue (they have no chances to get new data).
  1547. *
  1548. * This is implemented in the callers, where they modify the 'nonagle'
  1549. * argument based upon the location of SKB in the send queue.
  1550. */
  1551. if (nonagle & TCP_NAGLE_PUSH)
  1552. return true;
  1553. /* Don't use the nagle rule for urgent data (or for the final FIN). */
  1554. if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
  1555. return true;
  1556. if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
  1557. return true;
  1558. return false;
  1559. }
  1560. /* Does at least the first segment of SKB fit into the send window? */
  1561. static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
  1562. const struct sk_buff *skb,
  1563. unsigned int cur_mss)
  1564. {
  1565. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  1566. if (skb->len > cur_mss)
  1567. end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
  1568. return !after(end_seq, tcp_wnd_end(tp));
  1569. }
  1570. /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
  1571. * which is put after SKB on the list. It is very much like
  1572. * tcp_fragment() except that it may make several kinds of assumptions
  1573. * in order to speed up the splitting operation. In particular, we
  1574. * know that all the data is in scatter-gather pages, and that the
  1575. * packet has never been sent out before (and thus is not cloned).
  1576. */
  1577. static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
  1578. struct sk_buff *skb, unsigned int len,
  1579. unsigned int mss_now, gfp_t gfp)
  1580. {
  1581. struct sk_buff *buff;
  1582. int nlen = skb->len - len;
  1583. u8 flags;
  1584. /* All of a TSO frame must be composed of paged data. */
  1585. if (skb->len != skb->data_len)
  1586. return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
  1587. buff = sk_stream_alloc_skb(sk, 0, gfp, true);
  1588. if (unlikely(!buff))
  1589. return -ENOMEM;
  1590. sk->sk_wmem_queued += buff->truesize;
  1591. sk_mem_charge(sk, buff->truesize);
  1592. buff->truesize += nlen;
  1593. skb->truesize -= nlen;
  1594. /* Correct the sequence numbers. */
  1595. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  1596. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  1597. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  1598. /* PSH and FIN should only be set in the second packet. */
  1599. flags = TCP_SKB_CB(skb)->tcp_flags;
  1600. TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
  1601. TCP_SKB_CB(buff)->tcp_flags = flags;
  1602. /* This packet was never sent out yet, so no SACK bits. */
  1603. TCP_SKB_CB(buff)->sacked = 0;
  1604. tcp_skb_fragment_eor(skb, buff);
  1605. buff->ip_summed = CHECKSUM_PARTIAL;
  1606. skb_split(skb, buff, len);
  1607. tcp_fragment_tstamp(skb, buff);
  1608. /* Fix up tso_factor for both original and new SKB. */
  1609. tcp_set_skb_tso_segs(skb, mss_now);
  1610. tcp_set_skb_tso_segs(buff, mss_now);
  1611. /* Link BUFF into the send queue. */
  1612. __skb_header_release(buff);
  1613. tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
  1614. return 0;
  1615. }
  1616. /* Try to defer sending, if possible, in order to minimize the amount
  1617. * of TSO splitting we do. View it as a kind of TSO Nagle test.
  1618. *
  1619. * This algorithm is from John Heffner.
  1620. */
  1621. static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
  1622. bool *is_cwnd_limited, u32 max_segs)
  1623. {
  1624. const struct inet_connection_sock *icsk = inet_csk(sk);
  1625. u32 age, send_win, cong_win, limit, in_flight;
  1626. struct tcp_sock *tp = tcp_sk(sk);
  1627. struct sk_buff *head;
  1628. int win_divisor;
  1629. if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
  1630. goto send_now;
  1631. if (icsk->icsk_ca_state >= TCP_CA_Recovery)
  1632. goto send_now;
  1633. /* Avoid bursty behavior by allowing defer
  1634. * only if the last write was recent.
  1635. */
  1636. if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
  1637. goto send_now;
  1638. in_flight = tcp_packets_in_flight(tp);
  1639. BUG_ON(tcp_skb_pcount(skb) <= 1);
  1640. BUG_ON(tp->snd_cwnd <= in_flight);
  1641. send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  1642. /* From in_flight test above, we know that cwnd > in_flight. */
  1643. cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
  1644. limit = min(send_win, cong_win);
  1645. /* If a full-sized TSO skb can be sent, do it. */
  1646. if (limit >= max_segs * tp->mss_cache)
  1647. goto send_now;
  1648. /* Middle in queue won't get any more data, full sendable already? */
  1649. if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
  1650. goto send_now;
  1651. win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
  1652. if (win_divisor) {
  1653. u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
  1654. /* If at least some fraction of a window is available,
  1655. * just use it.
  1656. */
  1657. chunk /= win_divisor;
  1658. if (limit >= chunk)
  1659. goto send_now;
  1660. } else {
  1661. /* Different approach, try not to defer past a single
  1662. * ACK. Receiver should ACK every other full sized
  1663. * frame, so if we have space for more than 3 frames
  1664. * then send now.
  1665. */
  1666. if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
  1667. goto send_now;
  1668. }
  1669. /* TODO : use tsorted_sent_queue ? */
  1670. head = tcp_rtx_queue_head(sk);
  1671. if (!head)
  1672. goto send_now;
  1673. age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
  1674. /* If next ACK is likely to come too late (half srtt), do not defer */
  1675. if (age < (tp->srtt_us >> 4))
  1676. goto send_now;
  1677. /* Ok, it looks like it is advisable to defer. */
  1678. if (cong_win < send_win && cong_win <= skb->len)
  1679. *is_cwnd_limited = true;
  1680. return true;
  1681. send_now:
  1682. return false;
  1683. }
  1684. static inline void tcp_mtu_check_reprobe(struct sock *sk)
  1685. {
  1686. struct inet_connection_sock *icsk = inet_csk(sk);
  1687. struct tcp_sock *tp = tcp_sk(sk);
  1688. struct net *net = sock_net(sk);
  1689. u32 interval;
  1690. s32 delta;
  1691. interval = net->ipv4.sysctl_tcp_probe_interval;
  1692. delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
  1693. if (unlikely(delta >= interval * HZ)) {
  1694. int mss = tcp_current_mss(sk);
  1695. /* Update current search range */
  1696. icsk->icsk_mtup.probe_size = 0;
  1697. icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
  1698. sizeof(struct tcphdr) +
  1699. icsk->icsk_af_ops->net_header_len;
  1700. icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
  1701. /* Update probe time stamp */
  1702. icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
  1703. }
  1704. }
  1705. static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
  1706. {
  1707. struct sk_buff *skb, *next;
  1708. skb = tcp_send_head(sk);
  1709. tcp_for_write_queue_from_safe(skb, next, sk) {
  1710. if (len <= skb->len)
  1711. break;
  1712. if (unlikely(TCP_SKB_CB(skb)->eor))
  1713. return false;
  1714. len -= skb->len;
  1715. }
  1716. return true;
  1717. }
  1718. /* Create a new MTU probe if we are ready.
  1719. * MTU probe is regularly attempting to increase the path MTU by
  1720. * deliberately sending larger packets. This discovers routing
  1721. * changes resulting in larger path MTUs.
  1722. *
  1723. * Returns 0 if we should wait to probe (no cwnd available),
  1724. * 1 if a probe was sent,
  1725. * -1 otherwise
  1726. */
  1727. static int tcp_mtu_probe(struct sock *sk)
  1728. {
  1729. struct inet_connection_sock *icsk = inet_csk(sk);
  1730. struct tcp_sock *tp = tcp_sk(sk);
  1731. struct sk_buff *skb, *nskb, *next;
  1732. struct net *net = sock_net(sk);
  1733. int probe_size;
  1734. int size_needed;
  1735. int copy, len;
  1736. int mss_now;
  1737. int interval;
  1738. /* Not currently probing/verifying,
  1739. * not in recovery,
  1740. * have enough cwnd, and
  1741. * not SACKing (the variable headers throw things off)
  1742. */
  1743. if (likely(!icsk->icsk_mtup.enabled ||
  1744. icsk->icsk_mtup.probe_size ||
  1745. inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
  1746. tp->snd_cwnd < 11 ||
  1747. tp->rx_opt.num_sacks || tp->rx_opt.dsack))
  1748. return -1;
  1749. /* Use binary search for probe_size between tcp_mss_base,
  1750. * and current mss_clamp. if (search_high - search_low)
  1751. * smaller than a threshold, backoff from probing.
  1752. */
  1753. mss_now = tcp_current_mss(sk);
  1754. probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
  1755. icsk->icsk_mtup.search_low) >> 1);
  1756. size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
  1757. interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
  1758. /* When misfortune happens, we are reprobing actively,
  1759. * and then reprobe timer has expired. We stick with current
  1760. * probing process by not resetting search range to its orignal.
  1761. */
  1762. if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
  1763. interval < net->ipv4.sysctl_tcp_probe_threshold) {
  1764. /* Check whether enough time has elaplased for
  1765. * another round of probing.
  1766. */
  1767. tcp_mtu_check_reprobe(sk);
  1768. return -1;
  1769. }
  1770. /* Have enough data in the send queue to probe? */
  1771. if (tp->write_seq - tp->snd_nxt < size_needed)
  1772. return -1;
  1773. if (tp->snd_wnd < size_needed)
  1774. return -1;
  1775. if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
  1776. return 0;
  1777. /* Do we need to wait to drain cwnd? With none in flight, don't stall */
  1778. if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
  1779. if (!tcp_packets_in_flight(tp))
  1780. return -1;
  1781. else
  1782. return 0;
  1783. }
  1784. if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
  1785. return -1;
  1786. /* We're allowed to probe. Build it now. */
  1787. nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
  1788. if (!nskb)
  1789. return -1;
  1790. sk->sk_wmem_queued += nskb->truesize;
  1791. sk_mem_charge(sk, nskb->truesize);
  1792. skb = tcp_send_head(sk);
  1793. TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
  1794. TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
  1795. TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
  1796. TCP_SKB_CB(nskb)->sacked = 0;
  1797. nskb->csum = 0;
  1798. nskb->ip_summed = CHECKSUM_PARTIAL;
  1799. tcp_insert_write_queue_before(nskb, skb, sk);
  1800. tcp_highest_sack_replace(sk, skb, nskb);
  1801. len = 0;
  1802. tcp_for_write_queue_from_safe(skb, next, sk) {
  1803. copy = min_t(int, skb->len, probe_size - len);
  1804. skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
  1805. if (skb->len <= copy) {
  1806. /* We've eaten all the data from this skb.
  1807. * Throw it away. */
  1808. TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
  1809. /* If this is the last SKB we copy and eor is set
  1810. * we need to propagate it to the new skb.
  1811. */
  1812. TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
  1813. tcp_unlink_write_queue(skb, sk);
  1814. sk_wmem_free_skb(sk, skb);
  1815. } else {
  1816. TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
  1817. ~(TCPHDR_FIN|TCPHDR_PSH);
  1818. if (!skb_shinfo(skb)->nr_frags) {
  1819. skb_pull(skb, copy);
  1820. } else {
  1821. __pskb_trim_head(skb, copy);
  1822. tcp_set_skb_tso_segs(skb, mss_now);
  1823. }
  1824. TCP_SKB_CB(skb)->seq += copy;
  1825. }
  1826. len += copy;
  1827. if (len >= probe_size)
  1828. break;
  1829. }
  1830. tcp_init_tso_segs(nskb, nskb->len);
  1831. /* We're ready to send. If this fails, the probe will
  1832. * be resegmented into mss-sized pieces by tcp_write_xmit().
  1833. */
  1834. if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
  1835. /* Decrement cwnd here because we are sending
  1836. * effectively two packets. */
  1837. tp->snd_cwnd--;
  1838. tcp_event_new_data_sent(sk, nskb);
  1839. icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
  1840. tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
  1841. tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
  1842. return 1;
  1843. }
  1844. return -1;
  1845. }
  1846. static bool tcp_pacing_check(const struct sock *sk)
  1847. {
  1848. return tcp_needs_internal_pacing(sk) &&
  1849. hrtimer_is_queued(&tcp_sk(sk)->pacing_timer);
  1850. }
  1851. /* TCP Small Queues :
  1852. * Control number of packets in qdisc/devices to two packets / or ~1 ms.
  1853. * (These limits are doubled for retransmits)
  1854. * This allows for :
  1855. * - better RTT estimation and ACK scheduling
  1856. * - faster recovery
  1857. * - high rates
  1858. * Alas, some drivers / subsystems require a fair amount
  1859. * of queued bytes to ensure line rate.
  1860. * One example is wifi aggregation (802.11 AMPDU)
  1861. */
  1862. static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
  1863. unsigned int factor)
  1864. {
  1865. unsigned int limit;
  1866. limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
  1867. limit = min_t(u32, limit,
  1868. sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
  1869. limit <<= factor;
  1870. if (refcount_read(&sk->sk_wmem_alloc) > limit) {
  1871. /* Always send skb if rtx queue is empty.
  1872. * No need to wait for TX completion to call us back,
  1873. * after softirq/tasklet schedule.
  1874. * This helps when TX completions are delayed too much.
  1875. */
  1876. if (tcp_rtx_queue_empty(sk))
  1877. return false;
  1878. set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
  1879. /* It is possible TX completion already happened
  1880. * before we set TSQ_THROTTLED, so we must
  1881. * test again the condition.
  1882. */
  1883. smp_mb__after_atomic();
  1884. if (refcount_read(&sk->sk_wmem_alloc) > limit)
  1885. return true;
  1886. }
  1887. return false;
  1888. }
  1889. static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
  1890. {
  1891. const u32 now = tcp_jiffies32;
  1892. enum tcp_chrono old = tp->chrono_type;
  1893. if (old > TCP_CHRONO_UNSPEC)
  1894. tp->chrono_stat[old - 1] += now - tp->chrono_start;
  1895. tp->chrono_start = now;
  1896. tp->chrono_type = new;
  1897. }
  1898. void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
  1899. {
  1900. struct tcp_sock *tp = tcp_sk(sk);
  1901. /* If there are multiple conditions worthy of tracking in a
  1902. * chronograph then the highest priority enum takes precedence
  1903. * over the other conditions. So that if something "more interesting"
  1904. * starts happening, stop the previous chrono and start a new one.
  1905. */
  1906. if (type > tp->chrono_type)
  1907. tcp_chrono_set(tp, type);
  1908. }
  1909. void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
  1910. {
  1911. struct tcp_sock *tp = tcp_sk(sk);
  1912. /* There are multiple conditions worthy of tracking in a
  1913. * chronograph, so that the highest priority enum takes
  1914. * precedence over the other conditions (see tcp_chrono_start).
  1915. * If a condition stops, we only stop chrono tracking if
  1916. * it's the "most interesting" or current chrono we are
  1917. * tracking and starts busy chrono if we have pending data.
  1918. */
  1919. if (tcp_rtx_and_write_queues_empty(sk))
  1920. tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
  1921. else if (type == tp->chrono_type)
  1922. tcp_chrono_set(tp, TCP_CHRONO_BUSY);
  1923. }
  1924. /* This routine writes packets to the network. It advances the
  1925. * send_head. This happens as incoming acks open up the remote
  1926. * window for us.
  1927. *
  1928. * LARGESEND note: !tcp_urg_mode is overkill, only frames between
  1929. * snd_up-64k-mss .. snd_up cannot be large. However, taking into
  1930. * account rare use of URG, this is not a big flaw.
  1931. *
  1932. * Send at most one packet when push_one > 0. Temporarily ignore
  1933. * cwnd limit to force at most one packet out when push_one == 2.
  1934. * Returns true, if no segments are in flight and we have queued segments,
  1935. * but cannot send anything now because of SWS or another problem.
  1936. */
  1937. static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  1938. int push_one, gfp_t gfp)
  1939. {
  1940. struct tcp_sock *tp = tcp_sk(sk);
  1941. struct sk_buff *skb;
  1942. unsigned int tso_segs, sent_pkts;
  1943. int cwnd_quota;
  1944. int result;
  1945. bool is_cwnd_limited = false, is_rwnd_limited = false;
  1946. u32 max_segs;
  1947. sent_pkts = 0;
  1948. tcp_mstamp_refresh(tp);
  1949. if (!push_one) {
  1950. /* Do MTU probing. */
  1951. result = tcp_mtu_probe(sk);
  1952. if (!result) {
  1953. return false;
  1954. } else if (result > 0) {
  1955. sent_pkts = 1;
  1956. }
  1957. }
  1958. max_segs = tcp_tso_segs(sk, mss_now);
  1959. while ((skb = tcp_send_head(sk))) {
  1960. unsigned int limit;
  1961. if (tcp_pacing_check(sk))
  1962. break;
  1963. tso_segs = tcp_init_tso_segs(skb, mss_now);
  1964. BUG_ON(!tso_segs);
  1965. if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
  1966. /* "skb_mstamp" is used as a start point for the retransmit timer */
  1967. tcp_update_skb_after_send(tp, skb);
  1968. goto repair; /* Skip network transmission */
  1969. }
  1970. cwnd_quota = tcp_cwnd_test(tp, skb);
  1971. if (!cwnd_quota) {
  1972. if (push_one == 2)
  1973. /* Force out a loss probe pkt. */
  1974. cwnd_quota = 1;
  1975. else
  1976. break;
  1977. }
  1978. if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
  1979. is_rwnd_limited = true;
  1980. break;
  1981. }
  1982. if (tso_segs == 1) {
  1983. if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
  1984. (tcp_skb_is_last(sk, skb) ?
  1985. nonagle : TCP_NAGLE_PUSH))))
  1986. break;
  1987. } else {
  1988. if (!push_one &&
  1989. tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
  1990. max_segs))
  1991. break;
  1992. }
  1993. limit = mss_now;
  1994. if (tso_segs > 1 && !tcp_urg_mode(tp))
  1995. limit = tcp_mss_split_point(sk, skb, mss_now,
  1996. min_t(unsigned int,
  1997. cwnd_quota,
  1998. max_segs),
  1999. nonagle);
  2000. if (skb->len > limit &&
  2001. unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
  2002. skb, limit, mss_now, gfp)))
  2003. break;
  2004. if (tcp_small_queue_check(sk, skb, 0))
  2005. break;
  2006. if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
  2007. break;
  2008. repair:
  2009. /* Advance the send_head. This one is sent out.
  2010. * This call will increment packets_out.
  2011. */
  2012. tcp_event_new_data_sent(sk, skb);
  2013. tcp_minshall_update(tp, mss_now, skb);
  2014. sent_pkts += tcp_skb_pcount(skb);
  2015. if (push_one)
  2016. break;
  2017. }
  2018. if (is_rwnd_limited)
  2019. tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
  2020. else
  2021. tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
  2022. if (likely(sent_pkts)) {
  2023. if (tcp_in_cwnd_reduction(sk))
  2024. tp->prr_out += sent_pkts;
  2025. /* Send one loss probe per tail loss episode. */
  2026. if (push_one != 2)
  2027. tcp_schedule_loss_probe(sk, false);
  2028. is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
  2029. tcp_cwnd_validate(sk, is_cwnd_limited);
  2030. return false;
  2031. }
  2032. return !tp->packets_out && !tcp_write_queue_empty(sk);
  2033. }
  2034. bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
  2035. {
  2036. struct inet_connection_sock *icsk = inet_csk(sk);
  2037. struct tcp_sock *tp = tcp_sk(sk);
  2038. u32 timeout, rto_delta_us;
  2039. int early_retrans;
  2040. /* Don't do any loss probe on a Fast Open connection before 3WHS
  2041. * finishes.
  2042. */
  2043. if (tp->fastopen_rsk)
  2044. return false;
  2045. early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
  2046. /* Schedule a loss probe in 2*RTT for SACK capable connections
  2047. * not in loss recovery, that are either limited by cwnd or application.
  2048. */
  2049. if ((early_retrans != 3 && early_retrans != 4) ||
  2050. !tp->packets_out || !tcp_is_sack(tp) ||
  2051. (icsk->icsk_ca_state != TCP_CA_Open &&
  2052. icsk->icsk_ca_state != TCP_CA_CWR))
  2053. return false;
  2054. /* Probe timeout is 2*rtt. Add minimum RTO to account
  2055. * for delayed ack when there's one outstanding packet. If no RTT
  2056. * sample is available then probe after TCP_TIMEOUT_INIT.
  2057. */
  2058. if (tp->srtt_us) {
  2059. timeout = usecs_to_jiffies(tp->srtt_us >> 2);
  2060. if (tp->packets_out == 1)
  2061. timeout += TCP_RTO_MIN;
  2062. else
  2063. timeout += TCP_TIMEOUT_MIN;
  2064. } else {
  2065. timeout = TCP_TIMEOUT_INIT;
  2066. }
  2067. /* If the RTO formula yields an earlier time, then use that time. */
  2068. rto_delta_us = advancing_rto ?
  2069. jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
  2070. tcp_rto_delta_us(sk); /* How far in future is RTO? */
  2071. if (rto_delta_us > 0)
  2072. timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
  2073. inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
  2074. TCP_RTO_MAX);
  2075. return true;
  2076. }
  2077. /* Thanks to skb fast clones, we can detect if a prior transmit of
  2078. * a packet is still in a qdisc or driver queue.
  2079. * In this case, there is very little point doing a retransmit !
  2080. */
  2081. static bool skb_still_in_host_queue(const struct sock *sk,
  2082. const struct sk_buff *skb)
  2083. {
  2084. if (unlikely(skb_fclone_busy(sk, skb))) {
  2085. NET_INC_STATS(sock_net(sk),
  2086. LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
  2087. return true;
  2088. }
  2089. return false;
  2090. }
  2091. /* When probe timeout (PTO) fires, try send a new segment if possible, else
  2092. * retransmit the last segment.
  2093. */
  2094. void tcp_send_loss_probe(struct sock *sk)
  2095. {
  2096. struct tcp_sock *tp = tcp_sk(sk);
  2097. struct sk_buff *skb;
  2098. int pcount;
  2099. int mss = tcp_current_mss(sk);
  2100. skb = tcp_send_head(sk);
  2101. if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
  2102. pcount = tp->packets_out;
  2103. tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
  2104. if (tp->packets_out > pcount)
  2105. goto probe_sent;
  2106. goto rearm_timer;
  2107. }
  2108. skb = skb_rb_last(&sk->tcp_rtx_queue);
  2109. /* At most one outstanding TLP retransmission. */
  2110. if (tp->tlp_high_seq)
  2111. goto rearm_timer;
  2112. /* Retransmit last segment. */
  2113. if (WARN_ON(!skb))
  2114. goto rearm_timer;
  2115. if (skb_still_in_host_queue(sk, skb))
  2116. goto rearm_timer;
  2117. pcount = tcp_skb_pcount(skb);
  2118. if (WARN_ON(!pcount))
  2119. goto rearm_timer;
  2120. if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
  2121. if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
  2122. (pcount - 1) * mss, mss,
  2123. GFP_ATOMIC)))
  2124. goto rearm_timer;
  2125. skb = skb_rb_next(skb);
  2126. }
  2127. if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
  2128. goto rearm_timer;
  2129. if (__tcp_retransmit_skb(sk, skb, 1))
  2130. goto rearm_timer;
  2131. /* Record snd_nxt for loss detection. */
  2132. tp->tlp_high_seq = tp->snd_nxt;
  2133. probe_sent:
  2134. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
  2135. /* Reset s.t. tcp_rearm_rto will restart timer from now */
  2136. inet_csk(sk)->icsk_pending = 0;
  2137. rearm_timer:
  2138. tcp_rearm_rto(sk);
  2139. }
  2140. /* Push out any pending frames which were held back due to
  2141. * TCP_CORK or attempt at coalescing tiny packets.
  2142. * The socket must be locked by the caller.
  2143. */
  2144. void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
  2145. int nonagle)
  2146. {
  2147. /* If we are closed, the bytes will have to remain here.
  2148. * In time closedown will finish, we empty the write queue and
  2149. * all will be happy.
  2150. */
  2151. if (unlikely(sk->sk_state == TCP_CLOSE))
  2152. return;
  2153. if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
  2154. sk_gfp_mask(sk, GFP_ATOMIC)))
  2155. tcp_check_probe_timer(sk);
  2156. }
  2157. /* Send _single_ skb sitting at the send head. This function requires
  2158. * true push pending frames to setup probe timer etc.
  2159. */
  2160. void tcp_push_one(struct sock *sk, unsigned int mss_now)
  2161. {
  2162. struct sk_buff *skb = tcp_send_head(sk);
  2163. BUG_ON(!skb || skb->len < mss_now);
  2164. tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
  2165. }
  2166. /* This function returns the amount that we can raise the
  2167. * usable window based on the following constraints
  2168. *
  2169. * 1. The window can never be shrunk once it is offered (RFC 793)
  2170. * 2. We limit memory per socket
  2171. *
  2172. * RFC 1122:
  2173. * "the suggested [SWS] avoidance algorithm for the receiver is to keep
  2174. * RECV.NEXT + RCV.WIN fixed until:
  2175. * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
  2176. *
  2177. * i.e. don't raise the right edge of the window until you can raise
  2178. * it at least MSS bytes.
  2179. *
  2180. * Unfortunately, the recommended algorithm breaks header prediction,
  2181. * since header prediction assumes th->window stays fixed.
  2182. *
  2183. * Strictly speaking, keeping th->window fixed violates the receiver
  2184. * side SWS prevention criteria. The problem is that under this rule
  2185. * a stream of single byte packets will cause the right side of the
  2186. * window to always advance by a single byte.
  2187. *
  2188. * Of course, if the sender implements sender side SWS prevention
  2189. * then this will not be a problem.
  2190. *
  2191. * BSD seems to make the following compromise:
  2192. *
  2193. * If the free space is less than the 1/4 of the maximum
  2194. * space available and the free space is less than 1/2 mss,
  2195. * then set the window to 0.
  2196. * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
  2197. * Otherwise, just prevent the window from shrinking
  2198. * and from being larger than the largest representable value.
  2199. *
  2200. * This prevents incremental opening of the window in the regime
  2201. * where TCP is limited by the speed of the reader side taking
  2202. * data out of the TCP receive queue. It does nothing about
  2203. * those cases where the window is constrained on the sender side
  2204. * because the pipeline is full.
  2205. *
  2206. * BSD also seems to "accidentally" limit itself to windows that are a
  2207. * multiple of MSS, at least until the free space gets quite small.
  2208. * This would appear to be a side effect of the mbuf implementation.
  2209. * Combining these two algorithms results in the observed behavior
  2210. * of having a fixed window size at almost all times.
  2211. *
  2212. * Below we obtain similar behavior by forcing the offered window to
  2213. * a multiple of the mss when it is feasible to do so.
  2214. *
  2215. * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
  2216. * Regular options like TIMESTAMP are taken into account.
  2217. */
  2218. u32 __tcp_select_window(struct sock *sk)
  2219. {
  2220. struct inet_connection_sock *icsk = inet_csk(sk);
  2221. struct tcp_sock *tp = tcp_sk(sk);
  2222. /* MSS for the peer's data. Previous versions used mss_clamp
  2223. * here. I don't know if the value based on our guesses
  2224. * of peer's MSS is better for the performance. It's more correct
  2225. * but may be worse for the performance because of rcv_mss
  2226. * fluctuations. --SAW 1998/11/1
  2227. */
  2228. int mss = icsk->icsk_ack.rcv_mss;
  2229. int free_space = tcp_space(sk);
  2230. int allowed_space = tcp_full_space(sk);
  2231. int full_space = min_t(int, tp->window_clamp, allowed_space);
  2232. int window;
  2233. if (unlikely(mss > full_space)) {
  2234. mss = full_space;
  2235. if (mss <= 0)
  2236. return 0;
  2237. }
  2238. if (free_space < (full_space >> 1)) {
  2239. icsk->icsk_ack.quick = 0;
  2240. if (tcp_under_memory_pressure(sk))
  2241. tp->rcv_ssthresh = min(tp->rcv_ssthresh,
  2242. 4U * tp->advmss);
  2243. /* free_space might become our new window, make sure we don't
  2244. * increase it due to wscale.
  2245. */
  2246. free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
  2247. /* if free space is less than mss estimate, or is below 1/16th
  2248. * of the maximum allowed, try to move to zero-window, else
  2249. * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
  2250. * new incoming data is dropped due to memory limits.
  2251. * With large window, mss test triggers way too late in order
  2252. * to announce zero window in time before rmem limit kicks in.
  2253. */
  2254. if (free_space < (allowed_space >> 4) || free_space < mss)
  2255. return 0;
  2256. }
  2257. if (free_space > tp->rcv_ssthresh)
  2258. free_space = tp->rcv_ssthresh;
  2259. /* Don't do rounding if we are using window scaling, since the
  2260. * scaled window will not line up with the MSS boundary anyway.
  2261. */
  2262. if (tp->rx_opt.rcv_wscale) {
  2263. window = free_space;
  2264. /* Advertise enough space so that it won't get scaled away.
  2265. * Import case: prevent zero window announcement if
  2266. * 1<<rcv_wscale > mss.
  2267. */
  2268. window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
  2269. } else {
  2270. window = tp->rcv_wnd;
  2271. /* Get the largest window that is a nice multiple of mss.
  2272. * Window clamp already applied above.
  2273. * If our current window offering is within 1 mss of the
  2274. * free space we just keep it. This prevents the divide
  2275. * and multiply from happening most of the time.
  2276. * We also don't do any window rounding when the free space
  2277. * is too small.
  2278. */
  2279. if (window <= free_space - mss || window > free_space)
  2280. window = rounddown(free_space, mss);
  2281. else if (mss == full_space &&
  2282. free_space > window + (full_space >> 1))
  2283. window = free_space;
  2284. }
  2285. return window;
  2286. }
  2287. void tcp_skb_collapse_tstamp(struct sk_buff *skb,
  2288. const struct sk_buff *next_skb)
  2289. {
  2290. if (unlikely(tcp_has_tx_tstamp(next_skb))) {
  2291. const struct skb_shared_info *next_shinfo =
  2292. skb_shinfo(next_skb);
  2293. struct skb_shared_info *shinfo = skb_shinfo(skb);
  2294. shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
  2295. shinfo->tskey = next_shinfo->tskey;
  2296. TCP_SKB_CB(skb)->txstamp_ack |=
  2297. TCP_SKB_CB(next_skb)->txstamp_ack;
  2298. }
  2299. }
  2300. /* Collapses two adjacent SKB's during retransmission. */
  2301. static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
  2302. {
  2303. struct tcp_sock *tp = tcp_sk(sk);
  2304. struct sk_buff *next_skb = skb_rb_next(skb);
  2305. int skb_size, next_skb_size;
  2306. skb_size = skb->len;
  2307. next_skb_size = next_skb->len;
  2308. BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
  2309. if (next_skb_size) {
  2310. if (next_skb_size <= skb_availroom(skb))
  2311. skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
  2312. next_skb_size);
  2313. else if (!skb_shift(skb, next_skb, next_skb_size))
  2314. return false;
  2315. }
  2316. tcp_highest_sack_replace(sk, next_skb, skb);
  2317. /* Update sequence range on original skb. */
  2318. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
  2319. /* Merge over control information. This moves PSH/FIN etc. over */
  2320. TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
  2321. /* All done, get rid of second SKB and account for it so
  2322. * packet counting does not break.
  2323. */
  2324. TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
  2325. TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
  2326. /* changed transmit queue under us so clear hints */
  2327. tcp_clear_retrans_hints_partial(tp);
  2328. if (next_skb == tp->retransmit_skb_hint)
  2329. tp->retransmit_skb_hint = skb;
  2330. tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
  2331. tcp_skb_collapse_tstamp(skb, next_skb);
  2332. tcp_rtx_queue_unlink_and_free(next_skb, sk);
  2333. return true;
  2334. }
  2335. /* Check if coalescing SKBs is legal. */
  2336. static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
  2337. {
  2338. if (tcp_skb_pcount(skb) > 1)
  2339. return false;
  2340. if (skb_cloned(skb))
  2341. return false;
  2342. /* Some heuristics for collapsing over SACK'd could be invented */
  2343. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
  2344. return false;
  2345. return true;
  2346. }
  2347. /* Collapse packets in the retransmit queue to make to create
  2348. * less packets on the wire. This is only done on retransmission.
  2349. */
  2350. static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
  2351. int space)
  2352. {
  2353. struct tcp_sock *tp = tcp_sk(sk);
  2354. struct sk_buff *skb = to, *tmp;
  2355. bool first = true;
  2356. if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
  2357. return;
  2358. if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
  2359. return;
  2360. skb_rbtree_walk_from_safe(skb, tmp) {
  2361. if (!tcp_can_collapse(sk, skb))
  2362. break;
  2363. if (!tcp_skb_can_collapse_to(to))
  2364. break;
  2365. space -= skb->len;
  2366. if (first) {
  2367. first = false;
  2368. continue;
  2369. }
  2370. if (space < 0)
  2371. break;
  2372. if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
  2373. break;
  2374. if (!tcp_collapse_retrans(sk, to))
  2375. break;
  2376. }
  2377. }
  2378. /* This retransmits one SKB. Policy decisions and retransmit queue
  2379. * state updates are done by the caller. Returns non-zero if an
  2380. * error occurred which prevented the send.
  2381. */
  2382. int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
  2383. {
  2384. struct inet_connection_sock *icsk = inet_csk(sk);
  2385. struct tcp_sock *tp = tcp_sk(sk);
  2386. unsigned int cur_mss;
  2387. int diff, len, err;
  2388. /* Inconclusive MTU probe */
  2389. if (icsk->icsk_mtup.probe_size)
  2390. icsk->icsk_mtup.probe_size = 0;
  2391. /* Do not sent more than we queued. 1/4 is reserved for possible
  2392. * copying overhead: fragmentation, tunneling, mangling etc.
  2393. */
  2394. if (refcount_read(&sk->sk_wmem_alloc) >
  2395. min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
  2396. sk->sk_sndbuf))
  2397. return -EAGAIN;
  2398. if (skb_still_in_host_queue(sk, skb))
  2399. return -EBUSY;
  2400. if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
  2401. if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
  2402. WARN_ON_ONCE(1);
  2403. return -EINVAL;
  2404. }
  2405. if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
  2406. return -ENOMEM;
  2407. }
  2408. if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
  2409. return -EHOSTUNREACH; /* Routing failure or similar. */
  2410. cur_mss = tcp_current_mss(sk);
  2411. /* If receiver has shrunk his window, and skb is out of
  2412. * new window, do not retransmit it. The exception is the
  2413. * case, when window is shrunk to zero. In this case
  2414. * our retransmit serves as a zero window probe.
  2415. */
  2416. if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
  2417. TCP_SKB_CB(skb)->seq != tp->snd_una)
  2418. return -EAGAIN;
  2419. len = cur_mss * segs;
  2420. if (skb->len > len) {
  2421. if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
  2422. cur_mss, GFP_ATOMIC))
  2423. return -ENOMEM; /* We'll try again later. */
  2424. } else {
  2425. if (skb_unclone(skb, GFP_ATOMIC))
  2426. return -ENOMEM;
  2427. diff = tcp_skb_pcount(skb);
  2428. tcp_set_skb_tso_segs(skb, cur_mss);
  2429. diff -= tcp_skb_pcount(skb);
  2430. if (diff)
  2431. tcp_adjust_pcount(sk, skb, diff);
  2432. if (skb->len < cur_mss)
  2433. tcp_retrans_try_collapse(sk, skb, cur_mss);
  2434. }
  2435. /* RFC3168, section 6.1.1.1. ECN fallback */
  2436. if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
  2437. tcp_ecn_clear_syn(sk, skb);
  2438. /* Update global and local TCP statistics. */
  2439. segs = tcp_skb_pcount(skb);
  2440. TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
  2441. if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
  2442. __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
  2443. tp->total_retrans += segs;
  2444. tp->bytes_retrans += skb->len;
  2445. /* make sure skb->data is aligned on arches that require it
  2446. * and check if ack-trimming & collapsing extended the headroom
  2447. * beyond what csum_start can cover.
  2448. */
  2449. if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
  2450. skb_headroom(skb) >= 0xFFFF)) {
  2451. struct sk_buff *nskb;
  2452. tcp_skb_tsorted_save(skb) {
  2453. nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
  2454. err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
  2455. -ENOBUFS;
  2456. } tcp_skb_tsorted_restore(skb);
  2457. if (!err) {
  2458. tcp_update_skb_after_send(tp, skb);
  2459. tcp_rate_skb_sent(sk, skb);
  2460. }
  2461. } else {
  2462. err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2463. }
  2464. if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
  2465. tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
  2466. TCP_SKB_CB(skb)->seq, segs, err);
  2467. if (likely(!err)) {
  2468. TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
  2469. trace_tcp_retransmit_skb(sk, skb);
  2470. } else if (err != -EBUSY) {
  2471. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
  2472. }
  2473. return err;
  2474. }
  2475. int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
  2476. {
  2477. struct tcp_sock *tp = tcp_sk(sk);
  2478. int err = __tcp_retransmit_skb(sk, skb, segs);
  2479. if (err == 0) {
  2480. #if FASTRETRANS_DEBUG > 0
  2481. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
  2482. net_dbg_ratelimited("retrans_out leaked\n");
  2483. }
  2484. #endif
  2485. TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
  2486. tp->retrans_out += tcp_skb_pcount(skb);
  2487. /* Save stamp of the first retransmit. */
  2488. if (!tp->retrans_stamp)
  2489. tp->retrans_stamp = tcp_skb_timestamp(skb);
  2490. }
  2491. if (tp->undo_retrans < 0)
  2492. tp->undo_retrans = 0;
  2493. tp->undo_retrans += tcp_skb_pcount(skb);
  2494. return err;
  2495. }
  2496. /* This gets called after a retransmit timeout, and the initially
  2497. * retransmitted data is acknowledged. It tries to continue
  2498. * resending the rest of the retransmit queue, until either
  2499. * we've sent it all or the congestion window limit is reached.
  2500. */
  2501. void tcp_xmit_retransmit_queue(struct sock *sk)
  2502. {
  2503. const struct inet_connection_sock *icsk = inet_csk(sk);
  2504. struct sk_buff *skb, *rtx_head, *hole = NULL;
  2505. struct tcp_sock *tp = tcp_sk(sk);
  2506. u32 max_segs;
  2507. int mib_idx;
  2508. if (!tp->packets_out)
  2509. return;
  2510. rtx_head = tcp_rtx_queue_head(sk);
  2511. skb = tp->retransmit_skb_hint ?: rtx_head;
  2512. max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
  2513. skb_rbtree_walk_from(skb) {
  2514. __u8 sacked;
  2515. int segs;
  2516. if (tcp_pacing_check(sk))
  2517. break;
  2518. /* we could do better than to assign each time */
  2519. if (!hole)
  2520. tp->retransmit_skb_hint = skb;
  2521. segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
  2522. if (segs <= 0)
  2523. return;
  2524. sacked = TCP_SKB_CB(skb)->sacked;
  2525. /* In case tcp_shift_skb_data() have aggregated large skbs,
  2526. * we need to make sure not sending too bigs TSO packets
  2527. */
  2528. segs = min_t(int, segs, max_segs);
  2529. if (tp->retrans_out >= tp->lost_out) {
  2530. break;
  2531. } else if (!(sacked & TCPCB_LOST)) {
  2532. if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
  2533. hole = skb;
  2534. continue;
  2535. } else {
  2536. if (icsk->icsk_ca_state != TCP_CA_Loss)
  2537. mib_idx = LINUX_MIB_TCPFASTRETRANS;
  2538. else
  2539. mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
  2540. }
  2541. if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
  2542. continue;
  2543. if (tcp_small_queue_check(sk, skb, 1))
  2544. return;
  2545. if (tcp_retransmit_skb(sk, skb, segs))
  2546. return;
  2547. NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
  2548. if (tcp_in_cwnd_reduction(sk))
  2549. tp->prr_out += tcp_skb_pcount(skb);
  2550. if (skb == rtx_head &&
  2551. icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
  2552. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  2553. inet_csk(sk)->icsk_rto,
  2554. TCP_RTO_MAX);
  2555. }
  2556. }
  2557. /* We allow to exceed memory limits for FIN packets to expedite
  2558. * connection tear down and (memory) recovery.
  2559. * Otherwise tcp_send_fin() could be tempted to either delay FIN
  2560. * or even be forced to close flow without any FIN.
  2561. * In general, we want to allow one skb per socket to avoid hangs
  2562. * with edge trigger epoll()
  2563. */
  2564. void sk_forced_mem_schedule(struct sock *sk, int size)
  2565. {
  2566. int amt;
  2567. if (size <= sk->sk_forward_alloc)
  2568. return;
  2569. amt = sk_mem_pages(size);
  2570. sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
  2571. sk_memory_allocated_add(sk, amt);
  2572. if (mem_cgroup_sockets_enabled && sk->sk_memcg)
  2573. mem_cgroup_charge_skmem(sk->sk_memcg, amt);
  2574. }
  2575. /* Send a FIN. The caller locks the socket for us.
  2576. * We should try to send a FIN packet really hard, but eventually give up.
  2577. */
  2578. void tcp_send_fin(struct sock *sk)
  2579. {
  2580. struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
  2581. struct tcp_sock *tp = tcp_sk(sk);
  2582. /* Optimization, tack on the FIN if we have one skb in write queue and
  2583. * this skb was not yet sent, or we are under memory pressure.
  2584. * Note: in the latter case, FIN packet will be sent after a timeout,
  2585. * as TCP stack thinks it has already been transmitted.
  2586. */
  2587. if (!tskb && tcp_under_memory_pressure(sk))
  2588. tskb = skb_rb_last(&sk->tcp_rtx_queue);
  2589. if (tskb) {
  2590. coalesce:
  2591. TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
  2592. TCP_SKB_CB(tskb)->end_seq++;
  2593. tp->write_seq++;
  2594. if (tcp_write_queue_empty(sk)) {
  2595. /* This means tskb was already sent.
  2596. * Pretend we included the FIN on previous transmit.
  2597. * We need to set tp->snd_nxt to the value it would have
  2598. * if FIN had been sent. This is because retransmit path
  2599. * does not change tp->snd_nxt.
  2600. */
  2601. tp->snd_nxt++;
  2602. return;
  2603. }
  2604. } else {
  2605. skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
  2606. if (unlikely(!skb)) {
  2607. if (tskb)
  2608. goto coalesce;
  2609. return;
  2610. }
  2611. INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
  2612. skb_reserve(skb, MAX_TCP_HEADER);
  2613. sk_forced_mem_schedule(sk, skb->truesize);
  2614. /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
  2615. tcp_init_nondata_skb(skb, tp->write_seq,
  2616. TCPHDR_ACK | TCPHDR_FIN);
  2617. tcp_queue_skb(sk, skb);
  2618. }
  2619. __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
  2620. }
  2621. /* We get here when a process closes a file descriptor (either due to
  2622. * an explicit close() or as a byproduct of exit()'ing) and there
  2623. * was unread data in the receive queue. This behavior is recommended
  2624. * by RFC 2525, section 2.17. -DaveM
  2625. */
  2626. void tcp_send_active_reset(struct sock *sk, gfp_t priority)
  2627. {
  2628. struct sk_buff *skb;
  2629. TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
  2630. /* NOTE: No TCP options attached and we never retransmit this. */
  2631. skb = alloc_skb(MAX_TCP_HEADER, priority);
  2632. if (!skb) {
  2633. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
  2634. return;
  2635. }
  2636. /* Reserve space for headers and prepare control bits. */
  2637. skb_reserve(skb, MAX_TCP_HEADER);
  2638. tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
  2639. TCPHDR_ACK | TCPHDR_RST);
  2640. tcp_mstamp_refresh(tcp_sk(sk));
  2641. /* Send it off. */
  2642. if (tcp_transmit_skb(sk, skb, 0, priority))
  2643. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
  2644. /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
  2645. * skb here is different to the troublesome skb, so use NULL
  2646. */
  2647. trace_tcp_send_reset(sk, NULL);
  2648. }
  2649. /* Send a crossed SYN-ACK during socket establishment.
  2650. * WARNING: This routine must only be called when we have already sent
  2651. * a SYN packet that crossed the incoming SYN that caused this routine
  2652. * to get called. If this assumption fails then the initial rcv_wnd
  2653. * and rcv_wscale values will not be correct.
  2654. */
  2655. int tcp_send_synack(struct sock *sk)
  2656. {
  2657. struct sk_buff *skb;
  2658. skb = tcp_rtx_queue_head(sk);
  2659. if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
  2660. pr_err("%s: wrong queue state\n", __func__);
  2661. return -EFAULT;
  2662. }
  2663. if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
  2664. if (skb_cloned(skb)) {
  2665. struct sk_buff *nskb;
  2666. tcp_skb_tsorted_save(skb) {
  2667. nskb = skb_copy(skb, GFP_ATOMIC);
  2668. } tcp_skb_tsorted_restore(skb);
  2669. if (!nskb)
  2670. return -ENOMEM;
  2671. INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
  2672. tcp_rtx_queue_unlink_and_free(skb, sk);
  2673. __skb_header_release(nskb);
  2674. tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
  2675. sk->sk_wmem_queued += nskb->truesize;
  2676. sk_mem_charge(sk, nskb->truesize);
  2677. skb = nskb;
  2678. }
  2679. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
  2680. tcp_ecn_send_synack(sk, skb);
  2681. }
  2682. return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2683. }
  2684. /**
  2685. * tcp_make_synack - Prepare a SYN-ACK.
  2686. * sk: listener socket
  2687. * dst: dst entry attached to the SYNACK
  2688. * req: request_sock pointer
  2689. *
  2690. * Allocate one skb and build a SYNACK packet.
  2691. * @dst is consumed : Caller should not use it again.
  2692. */
  2693. struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
  2694. struct request_sock *req,
  2695. struct tcp_fastopen_cookie *foc,
  2696. enum tcp_synack_type synack_type)
  2697. {
  2698. struct inet_request_sock *ireq = inet_rsk(req);
  2699. const struct tcp_sock *tp = tcp_sk(sk);
  2700. struct tcp_md5sig_key *md5 = NULL;
  2701. struct tcp_out_options opts;
  2702. struct sk_buff *skb;
  2703. int tcp_header_size;
  2704. struct tcphdr *th;
  2705. int mss;
  2706. skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
  2707. if (unlikely(!skb)) {
  2708. dst_release(dst);
  2709. return NULL;
  2710. }
  2711. /* Reserve space for headers. */
  2712. skb_reserve(skb, MAX_TCP_HEADER);
  2713. switch (synack_type) {
  2714. case TCP_SYNACK_NORMAL:
  2715. skb_set_owner_w(skb, req_to_sk(req));
  2716. break;
  2717. case TCP_SYNACK_COOKIE:
  2718. /* Under synflood, we do not attach skb to a socket,
  2719. * to avoid false sharing.
  2720. */
  2721. break;
  2722. case TCP_SYNACK_FASTOPEN:
  2723. /* sk is a const pointer, because we want to express multiple
  2724. * cpu might call us concurrently.
  2725. * sk->sk_wmem_alloc in an atomic, we can promote to rw.
  2726. */
  2727. skb_set_owner_w(skb, (struct sock *)sk);
  2728. break;
  2729. }
  2730. skb_dst_set(skb, dst);
  2731. mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
  2732. memset(&opts, 0, sizeof(opts));
  2733. #ifdef CONFIG_SYN_COOKIES
  2734. if (unlikely(req->cookie_ts))
  2735. skb->skb_mstamp = cookie_init_timestamp(req);
  2736. else
  2737. #endif
  2738. skb->skb_mstamp = tcp_clock_us();
  2739. #ifdef CONFIG_TCP_MD5SIG
  2740. rcu_read_lock();
  2741. md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
  2742. #endif
  2743. skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
  2744. tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
  2745. foc) + sizeof(*th);
  2746. skb_push(skb, tcp_header_size);
  2747. skb_reset_transport_header(skb);
  2748. th = (struct tcphdr *)skb->data;
  2749. memset(th, 0, sizeof(struct tcphdr));
  2750. th->syn = 1;
  2751. th->ack = 1;
  2752. tcp_ecn_make_synack(req, th);
  2753. th->source = htons(ireq->ir_num);
  2754. th->dest = ireq->ir_rmt_port;
  2755. skb->mark = ireq->ir_mark;
  2756. skb->ip_summed = CHECKSUM_PARTIAL;
  2757. th->seq = htonl(tcp_rsk(req)->snt_isn);
  2758. /* XXX data is queued and acked as is. No buffer/window check */
  2759. th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
  2760. /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
  2761. th->window = htons(min(req->rsk_rcv_wnd, 65535U));
  2762. tcp_options_write((__be32 *)(th + 1), NULL, &opts);
  2763. th->doff = (tcp_header_size >> 2);
  2764. __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
  2765. #ifdef CONFIG_TCP_MD5SIG
  2766. /* Okay, we have all we need - do the md5 hash if needed */
  2767. if (md5)
  2768. tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
  2769. md5, req_to_sk(req), skb);
  2770. rcu_read_unlock();
  2771. #endif
  2772. /* Do not fool tcpdump (if any), clean our debris */
  2773. skb->tstamp = 0;
  2774. return skb;
  2775. }
  2776. EXPORT_SYMBOL(tcp_make_synack);
  2777. static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
  2778. {
  2779. struct inet_connection_sock *icsk = inet_csk(sk);
  2780. const struct tcp_congestion_ops *ca;
  2781. u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
  2782. if (ca_key == TCP_CA_UNSPEC)
  2783. return;
  2784. rcu_read_lock();
  2785. ca = tcp_ca_find_key(ca_key);
  2786. if (likely(ca && try_module_get(ca->owner))) {
  2787. module_put(icsk->icsk_ca_ops->owner);
  2788. icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
  2789. icsk->icsk_ca_ops = ca;
  2790. }
  2791. rcu_read_unlock();
  2792. }
  2793. /* Do all connect socket setups that can be done AF independent. */
  2794. static void tcp_connect_init(struct sock *sk)
  2795. {
  2796. const struct dst_entry *dst = __sk_dst_get(sk);
  2797. struct tcp_sock *tp = tcp_sk(sk);
  2798. __u8 rcv_wscale;
  2799. u32 rcv_wnd;
  2800. /* We'll fix this up when we get a response from the other end.
  2801. * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
  2802. */
  2803. tp->tcp_header_len = sizeof(struct tcphdr);
  2804. if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
  2805. tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
  2806. #ifdef CONFIG_TCP_MD5SIG
  2807. if (tp->af_specific->md5_lookup(sk, sk))
  2808. tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
  2809. #endif
  2810. /* If user gave his TCP_MAXSEG, record it to clamp */
  2811. if (tp->rx_opt.user_mss)
  2812. tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
  2813. tp->max_window = 0;
  2814. tcp_mtup_init(sk);
  2815. tcp_sync_mss(sk, dst_mtu(dst));
  2816. tcp_ca_dst_init(sk, dst);
  2817. if (!tp->window_clamp)
  2818. tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
  2819. tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
  2820. tcp_initialize_rcv_mss(sk);
  2821. /* limit the window selection if the user enforce a smaller rx buffer */
  2822. if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
  2823. (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
  2824. tp->window_clamp = tcp_full_space(sk);
  2825. rcv_wnd = tcp_rwnd_init_bpf(sk);
  2826. if (rcv_wnd == 0)
  2827. rcv_wnd = dst_metric(dst, RTAX_INITRWND);
  2828. tcp_select_initial_window(sk, tcp_full_space(sk),
  2829. tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
  2830. &tp->rcv_wnd,
  2831. &tp->window_clamp,
  2832. sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
  2833. &rcv_wscale,
  2834. rcv_wnd);
  2835. tp->rx_opt.rcv_wscale = rcv_wscale;
  2836. tp->rcv_ssthresh = tp->rcv_wnd;
  2837. sk->sk_err = 0;
  2838. sock_reset_flag(sk, SOCK_DONE);
  2839. tp->snd_wnd = 0;
  2840. tcp_init_wl(tp, 0);
  2841. tcp_write_queue_purge(sk);
  2842. tp->snd_una = tp->write_seq;
  2843. tp->snd_sml = tp->write_seq;
  2844. tp->snd_up = tp->write_seq;
  2845. tp->snd_nxt = tp->write_seq;
  2846. if (likely(!tp->repair))
  2847. tp->rcv_nxt = 0;
  2848. else
  2849. tp->rcv_tstamp = tcp_jiffies32;
  2850. tp->rcv_wup = tp->rcv_nxt;
  2851. tp->copied_seq = tp->rcv_nxt;
  2852. inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
  2853. inet_csk(sk)->icsk_retransmits = 0;
  2854. tcp_clear_retrans(tp);
  2855. }
  2856. static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
  2857. {
  2858. struct tcp_sock *tp = tcp_sk(sk);
  2859. struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
  2860. tcb->end_seq += skb->len;
  2861. __skb_header_release(skb);
  2862. sk->sk_wmem_queued += skb->truesize;
  2863. sk_mem_charge(sk, skb->truesize);
  2864. tp->write_seq = tcb->end_seq;
  2865. tp->packets_out += tcp_skb_pcount(skb);
  2866. }
  2867. /* Build and send a SYN with data and (cached) Fast Open cookie. However,
  2868. * queue a data-only packet after the regular SYN, such that regular SYNs
  2869. * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
  2870. * only the SYN sequence, the data are retransmitted in the first ACK.
  2871. * If cookie is not cached or other error occurs, falls back to send a
  2872. * regular SYN with Fast Open cookie request option.
  2873. */
  2874. static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
  2875. {
  2876. struct tcp_sock *tp = tcp_sk(sk);
  2877. struct tcp_fastopen_request *fo = tp->fastopen_req;
  2878. int space, err = 0;
  2879. struct sk_buff *syn_data;
  2880. tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
  2881. if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
  2882. goto fallback;
  2883. /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
  2884. * user-MSS. Reserve maximum option space for middleboxes that add
  2885. * private TCP options. The cost is reduced data space in SYN :(
  2886. */
  2887. tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
  2888. space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
  2889. MAX_TCP_OPTION_SPACE;
  2890. space = min_t(size_t, space, fo->size);
  2891. /* limit to order-0 allocations */
  2892. space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
  2893. syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
  2894. if (!syn_data)
  2895. goto fallback;
  2896. syn_data->ip_summed = CHECKSUM_PARTIAL;
  2897. memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
  2898. if (space) {
  2899. int copied = copy_from_iter(skb_put(syn_data, space), space,
  2900. &fo->data->msg_iter);
  2901. if (unlikely(!copied)) {
  2902. tcp_skb_tsorted_anchor_cleanup(syn_data);
  2903. kfree_skb(syn_data);
  2904. goto fallback;
  2905. }
  2906. if (copied != space) {
  2907. skb_trim(syn_data, copied);
  2908. space = copied;
  2909. }
  2910. }
  2911. /* No more data pending in inet_wait_for_connect() */
  2912. if (space == fo->size)
  2913. fo->data = NULL;
  2914. fo->copied = space;
  2915. tcp_connect_queue_skb(sk, syn_data);
  2916. if (syn_data->len)
  2917. tcp_chrono_start(sk, TCP_CHRONO_BUSY);
  2918. err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
  2919. syn->skb_mstamp = syn_data->skb_mstamp;
  2920. /* Now full SYN+DATA was cloned and sent (or not),
  2921. * remove the SYN from the original skb (syn_data)
  2922. * we keep in write queue in case of a retransmit, as we
  2923. * also have the SYN packet (with no data) in the same queue.
  2924. */
  2925. TCP_SKB_CB(syn_data)->seq++;
  2926. TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
  2927. if (!err) {
  2928. tp->syn_data = (fo->copied > 0);
  2929. tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
  2930. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
  2931. goto done;
  2932. }
  2933. /* data was not sent, put it in write_queue */
  2934. __skb_queue_tail(&sk->sk_write_queue, syn_data);
  2935. tp->packets_out -= tcp_skb_pcount(syn_data);
  2936. fallback:
  2937. /* Send a regular SYN with Fast Open cookie request option */
  2938. if (fo->cookie.len > 0)
  2939. fo->cookie.len = 0;
  2940. err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
  2941. if (err)
  2942. tp->syn_fastopen = 0;
  2943. done:
  2944. fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
  2945. return err;
  2946. }
  2947. /* Build a SYN and send it off. */
  2948. int tcp_connect(struct sock *sk)
  2949. {
  2950. struct tcp_sock *tp = tcp_sk(sk);
  2951. struct sk_buff *buff;
  2952. int err;
  2953. tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
  2954. if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
  2955. return -EHOSTUNREACH; /* Routing failure or similar. */
  2956. tcp_connect_init(sk);
  2957. if (unlikely(tp->repair)) {
  2958. tcp_finish_connect(sk, NULL);
  2959. return 0;
  2960. }
  2961. buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
  2962. if (unlikely(!buff))
  2963. return -ENOBUFS;
  2964. tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
  2965. tcp_mstamp_refresh(tp);
  2966. tp->retrans_stamp = tcp_time_stamp(tp);
  2967. tcp_connect_queue_skb(sk, buff);
  2968. tcp_ecn_send_syn(sk, buff);
  2969. tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
  2970. /* Send off SYN; include data in Fast Open. */
  2971. err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
  2972. tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
  2973. if (err == -ECONNREFUSED)
  2974. return err;
  2975. /* We change tp->snd_nxt after the tcp_transmit_skb() call
  2976. * in order to make this packet get counted in tcpOutSegs.
  2977. */
  2978. tp->snd_nxt = tp->write_seq;
  2979. tp->pushed_seq = tp->write_seq;
  2980. buff = tcp_send_head(sk);
  2981. if (unlikely(buff)) {
  2982. tp->snd_nxt = TCP_SKB_CB(buff)->seq;
  2983. tp->pushed_seq = TCP_SKB_CB(buff)->seq;
  2984. }
  2985. TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
  2986. /* Timer for repeating the SYN until an answer. */
  2987. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  2988. inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  2989. return 0;
  2990. }
  2991. EXPORT_SYMBOL(tcp_connect);
  2992. /* Send out a delayed ack, the caller does the policy checking
  2993. * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
  2994. * for details.
  2995. */
  2996. void tcp_send_delayed_ack(struct sock *sk)
  2997. {
  2998. struct inet_connection_sock *icsk = inet_csk(sk);
  2999. int ato = icsk->icsk_ack.ato;
  3000. unsigned long timeout;
  3001. if (ato > TCP_DELACK_MIN) {
  3002. const struct tcp_sock *tp = tcp_sk(sk);
  3003. int max_ato = HZ / 2;
  3004. if (icsk->icsk_ack.pingpong ||
  3005. (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
  3006. max_ato = TCP_DELACK_MAX;
  3007. /* Slow path, intersegment interval is "high". */
  3008. /* If some rtt estimate is known, use it to bound delayed ack.
  3009. * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
  3010. * directly.
  3011. */
  3012. if (tp->srtt_us) {
  3013. int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
  3014. TCP_DELACK_MIN);
  3015. if (rtt < max_ato)
  3016. max_ato = rtt;
  3017. }
  3018. ato = min(ato, max_ato);
  3019. }
  3020. /* Stay within the limit we were given */
  3021. timeout = jiffies + ato;
  3022. /* Use new timeout only if there wasn't a older one earlier. */
  3023. if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
  3024. /* If delack timer was blocked or is about to expire,
  3025. * send ACK now.
  3026. */
  3027. if (icsk->icsk_ack.blocked ||
  3028. time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
  3029. tcp_send_ack(sk);
  3030. return;
  3031. }
  3032. if (!time_before(timeout, icsk->icsk_ack.timeout))
  3033. timeout = icsk->icsk_ack.timeout;
  3034. }
  3035. icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
  3036. icsk->icsk_ack.timeout = timeout;
  3037. sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
  3038. }
  3039. /* This routine sends an ack and also updates the window. */
  3040. void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
  3041. {
  3042. struct sk_buff *buff;
  3043. /* If we have been reset, we may not send again. */
  3044. if (sk->sk_state == TCP_CLOSE)
  3045. return;
  3046. /* We are not putting this on the write queue, so
  3047. * tcp_transmit_skb() will set the ownership to this
  3048. * sock.
  3049. */
  3050. buff = alloc_skb(MAX_TCP_HEADER,
  3051. sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
  3052. if (unlikely(!buff)) {
  3053. inet_csk_schedule_ack(sk);
  3054. inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
  3055. inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
  3056. TCP_DELACK_MAX, TCP_RTO_MAX);
  3057. return;
  3058. }
  3059. /* Reserve space for headers and prepare control bits. */
  3060. skb_reserve(buff, MAX_TCP_HEADER);
  3061. tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
  3062. /* We do not want pure acks influencing TCP Small Queues or fq/pacing
  3063. * too much.
  3064. * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
  3065. */
  3066. skb_set_tcp_pure_ack(buff);
  3067. /* Send it off, this clears delayed acks for us. */
  3068. __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
  3069. }
  3070. EXPORT_SYMBOL_GPL(__tcp_send_ack);
  3071. void tcp_send_ack(struct sock *sk)
  3072. {
  3073. __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
  3074. }
  3075. /* This routine sends a packet with an out of date sequence
  3076. * number. It assumes the other end will try to ack it.
  3077. *
  3078. * Question: what should we make while urgent mode?
  3079. * 4.4BSD forces sending single byte of data. We cannot send
  3080. * out of window data, because we have SND.NXT==SND.MAX...
  3081. *
  3082. * Current solution: to send TWO zero-length segments in urgent mode:
  3083. * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
  3084. * out-of-date with SND.UNA-1 to probe window.
  3085. */
  3086. static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
  3087. {
  3088. struct tcp_sock *tp = tcp_sk(sk);
  3089. struct sk_buff *skb;
  3090. /* We don't queue it, tcp_transmit_skb() sets ownership. */
  3091. skb = alloc_skb(MAX_TCP_HEADER,
  3092. sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
  3093. if (!skb)
  3094. return -1;
  3095. /* Reserve space for headers and set control bits. */
  3096. skb_reserve(skb, MAX_TCP_HEADER);
  3097. /* Use a previous sequence. This should cause the other
  3098. * end to send an ack. Don't queue or clone SKB, just
  3099. * send it.
  3100. */
  3101. tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
  3102. NET_INC_STATS(sock_net(sk), mib);
  3103. return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
  3104. }
  3105. /* Called from setsockopt( ... TCP_REPAIR ) */
  3106. void tcp_send_window_probe(struct sock *sk)
  3107. {
  3108. if (sk->sk_state == TCP_ESTABLISHED) {
  3109. tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
  3110. tcp_mstamp_refresh(tcp_sk(sk));
  3111. tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
  3112. }
  3113. }
  3114. /* Initiate keepalive or window probe from timer. */
  3115. int tcp_write_wakeup(struct sock *sk, int mib)
  3116. {
  3117. struct tcp_sock *tp = tcp_sk(sk);
  3118. struct sk_buff *skb;
  3119. if (sk->sk_state == TCP_CLOSE)
  3120. return -1;
  3121. skb = tcp_send_head(sk);
  3122. if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
  3123. int err;
  3124. unsigned int mss = tcp_current_mss(sk);
  3125. unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  3126. if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
  3127. tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
  3128. /* We are probing the opening of a window
  3129. * but the window size is != 0
  3130. * must have been a result SWS avoidance ( sender )
  3131. */
  3132. if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
  3133. skb->len > mss) {
  3134. seg_size = min(seg_size, mss);
  3135. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
  3136. if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
  3137. skb, seg_size, mss, GFP_ATOMIC))
  3138. return -1;
  3139. } else if (!tcp_skb_pcount(skb))
  3140. tcp_set_skb_tso_segs(skb, mss);
  3141. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
  3142. err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  3143. if (!err)
  3144. tcp_event_new_data_sent(sk, skb);
  3145. return err;
  3146. } else {
  3147. if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
  3148. tcp_xmit_probe_skb(sk, 1, mib);
  3149. return tcp_xmit_probe_skb(sk, 0, mib);
  3150. }
  3151. }
  3152. /* A window probe timeout has occurred. If window is not closed send
  3153. * a partial packet else a zero probe.
  3154. */
  3155. void tcp_send_probe0(struct sock *sk)
  3156. {
  3157. struct inet_connection_sock *icsk = inet_csk(sk);
  3158. struct tcp_sock *tp = tcp_sk(sk);
  3159. struct net *net = sock_net(sk);
  3160. unsigned long probe_max;
  3161. int err;
  3162. err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
  3163. if (tp->packets_out || tcp_write_queue_empty(sk)) {
  3164. /* Cancel probe timer, if it is not required. */
  3165. icsk->icsk_probes_out = 0;
  3166. icsk->icsk_backoff = 0;
  3167. return;
  3168. }
  3169. if (err <= 0) {
  3170. if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
  3171. icsk->icsk_backoff++;
  3172. icsk->icsk_probes_out++;
  3173. probe_max = TCP_RTO_MAX;
  3174. } else {
  3175. /* If packet was not sent due to local congestion,
  3176. * do not backoff and do not remember icsk_probes_out.
  3177. * Let local senders to fight for local resources.
  3178. *
  3179. * Use accumulated backoff yet.
  3180. */
  3181. if (!icsk->icsk_probes_out)
  3182. icsk->icsk_probes_out = 1;
  3183. probe_max = TCP_RESOURCE_PROBE_INTERVAL;
  3184. }
  3185. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  3186. tcp_probe0_when(sk, probe_max),
  3187. TCP_RTO_MAX);
  3188. }
  3189. int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
  3190. {
  3191. const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
  3192. struct flowi fl;
  3193. int res;
  3194. tcp_rsk(req)->txhash = net_tx_rndhash();
  3195. res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
  3196. if (!res) {
  3197. __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
  3198. __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
  3199. if (unlikely(tcp_passive_fastopen(sk)))
  3200. tcp_sk(sk)->total_retrans++;
  3201. trace_tcp_retransmit_synack(sk, req);
  3202. }
  3203. return res;
  3204. }
  3205. EXPORT_SYMBOL(tcp_rtx_synack);