tcp_output.c 93 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612261326142615261626172618261926202621262226232624262526262627262826292630263126322633263426352636263726382639264026412642264326442645264626472648264926502651265226532654265526562657265826592660266126622663266426652666266726682669267026712672267326742675267626772678267926802681268226832684268526862687268826892690269126922693269426952696269726982699270027012702270327042705270627072708270927102711271227132714271527162717271827192720272127222723272427252726272727282729273027312732273327342735273627372738273927402741274227432744274527462747274827492750275127522753275427552756275727582759276027612762276327642765276627672768276927702771277227732774277527762777277827792780278127822783278427852786278727882789279027912792279327942795279627972798279928002801280228032804280528062807280828092810281128122813281428152816281728182819282028212822282328242825282628272828282928302831283228332834283528362837283828392840284128422843284428452846284728482849285028512852285328542855285628572858285928602861286228632864286528662867286828692870287128722873287428752876287728782879288028812882288328842885288628872888288928902891289228932894289528962897289828992900290129022903290429052906290729082909291029112912291329142915291629172918291929202921292229232924292529262927292829292930293129322933293429352936293729382939294029412942294329442945294629472948294929502951295229532954295529562957295829592960296129622963296429652966296729682969297029712972297329742975297629772978297929802981298229832984298529862987298829892990299129922993299429952996299729982999300030013002300330043005300630073008300930103011301230133014301530163017301830193020302130223023302430253026302730283029303030313032303330343035303630373038303930403041304230433044304530463047304830493050305130523053305430553056305730583059306030613062306330643065306630673068306930703071307230733074307530763077307830793080308130823083308430853086308730883089309030913092309330943095309630973098309931003101310231033104310531063107310831093110311131123113311431153116311731183119312031213122312331243125312631273128312931303131313231333134313531363137313831393140314131423143314431453146314731483149315031513152315331543155315631573158315931603161316231633164316531663167316831693170317131723173317431753176317731783179318031813182318331843185318631873188318931903191319231933194319531963197319831993200320132023203320432053206320732083209321032113212321332143215321632173218
  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. /* People can turn this off for buggy TCP's found in printers etc. */
  41. int sysctl_tcp_retrans_collapse __read_mostly = 1;
  42. /* People can turn this on to work with those rare, broken TCPs that
  43. * interpret the window field as a signed quantity.
  44. */
  45. int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
  46. /* Default TSQ limit of two TSO segments */
  47. int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
  48. /* This limits the percentage of the congestion window which we
  49. * will allow a single TSO frame to consume. Building TSO frames
  50. * which are too large can cause TCP streams to be bursty.
  51. */
  52. int sysctl_tcp_tso_win_divisor __read_mostly = 3;
  53. int sysctl_tcp_mtu_probing __read_mostly = 0;
  54. int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
  55. /* By default, RFC2861 behavior. */
  56. int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
  57. int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
  58. EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
  59. static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  60. int push_one, gfp_t gfp);
  61. /* Account for new data that has been sent to the network. */
  62. static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
  63. {
  64. struct tcp_sock *tp = tcp_sk(sk);
  65. unsigned int prior_packets = tp->packets_out;
  66. tcp_advance_send_head(sk, skb);
  67. tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
  68. /* Don't override Nagle indefinitely with F-RTO */
  69. if (tp->frto_counter == 2)
  70. tp->frto_counter = 3;
  71. tp->packets_out += tcp_skb_pcount(skb);
  72. if (!prior_packets || tp->early_retrans_delayed)
  73. tcp_rearm_rto(sk);
  74. }
  75. /* SND.NXT, if window was not shrunk.
  76. * If window has been shrunk, what should we make? It is not clear at all.
  77. * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  78. * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  79. * invalid. OK, let's make this for now:
  80. */
  81. static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  82. {
  83. const struct tcp_sock *tp = tcp_sk(sk);
  84. if (!before(tcp_wnd_end(tp), tp->snd_nxt))
  85. return tp->snd_nxt;
  86. else
  87. return tcp_wnd_end(tp);
  88. }
  89. /* Calculate mss to advertise in SYN segment.
  90. * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
  91. *
  92. * 1. It is independent of path mtu.
  93. * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
  94. * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
  95. * attached devices, because some buggy hosts are confused by
  96. * large MSS.
  97. * 4. We do not make 3, we advertise MSS, calculated from first
  98. * hop device mtu, but allow to raise it to ip_rt_min_advmss.
  99. * This may be overridden via information stored in routing table.
  100. * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
  101. * probably even Jumbo".
  102. */
  103. static __u16 tcp_advertise_mss(struct sock *sk)
  104. {
  105. struct tcp_sock *tp = tcp_sk(sk);
  106. const struct dst_entry *dst = __sk_dst_get(sk);
  107. int mss = tp->advmss;
  108. if (dst) {
  109. unsigned int metric = dst_metric_advmss(dst);
  110. if (metric < mss) {
  111. mss = metric;
  112. tp->advmss = mss;
  113. }
  114. }
  115. return (__u16)mss;
  116. }
  117. /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
  118. * This is the first part of cwnd validation mechanism. */
  119. static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
  120. {
  121. struct tcp_sock *tp = tcp_sk(sk);
  122. s32 delta = tcp_time_stamp - tp->lsndtime;
  123. u32 restart_cwnd = tcp_init_cwnd(tp, dst);
  124. u32 cwnd = tp->snd_cwnd;
  125. tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
  126. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  127. restart_cwnd = min(restart_cwnd, cwnd);
  128. while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
  129. cwnd >>= 1;
  130. tp->snd_cwnd = max(cwnd, restart_cwnd);
  131. tp->snd_cwnd_stamp = tcp_time_stamp;
  132. tp->snd_cwnd_used = 0;
  133. }
  134. /* Congestion state accounting after a packet has been sent. */
  135. static void tcp_event_data_sent(struct tcp_sock *tp,
  136. struct sock *sk)
  137. {
  138. struct inet_connection_sock *icsk = inet_csk(sk);
  139. const u32 now = tcp_time_stamp;
  140. if (sysctl_tcp_slow_start_after_idle &&
  141. (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
  142. tcp_cwnd_restart(sk, __sk_dst_get(sk));
  143. tp->lsndtime = now;
  144. /* If it is a reply for ato after last received
  145. * packet, enter pingpong mode.
  146. */
  147. if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
  148. icsk->icsk_ack.pingpong = 1;
  149. }
  150. /* Account for an ACK we sent. */
  151. static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
  152. {
  153. tcp_dec_quickack_mode(sk, pkts);
  154. inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
  155. }
  156. /* Determine a window scaling and initial window to offer.
  157. * Based on the assumption that the given amount of space
  158. * will be offered. Store the results in the tp structure.
  159. * NOTE: for smooth operation initial space offering should
  160. * be a multiple of mss if possible. We assume here that mss >= 1.
  161. * This MUST be enforced by all callers.
  162. */
  163. void tcp_select_initial_window(int __space, __u32 mss,
  164. __u32 *rcv_wnd, __u32 *window_clamp,
  165. int wscale_ok, __u8 *rcv_wscale,
  166. __u32 init_rcv_wnd)
  167. {
  168. unsigned int space = (__space < 0 ? 0 : __space);
  169. /* If no clamp set the clamp to the max possible scaled window */
  170. if (*window_clamp == 0)
  171. (*window_clamp) = (65535 << 14);
  172. space = min(*window_clamp, space);
  173. /* Quantize space offering to a multiple of mss if possible. */
  174. if (space > mss)
  175. space = (space / mss) * mss;
  176. /* NOTE: offering an initial window larger than 32767
  177. * will break some buggy TCP stacks. If the admin tells us
  178. * it is likely we could be speaking with such a buggy stack
  179. * we will truncate our initial window offering to 32K-1
  180. * unless the remote has sent us a window scaling option,
  181. * which we interpret as a sign the remote TCP is not
  182. * misinterpreting the window field as a signed quantity.
  183. */
  184. if (sysctl_tcp_workaround_signed_windows)
  185. (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
  186. else
  187. (*rcv_wnd) = space;
  188. (*rcv_wscale) = 0;
  189. if (wscale_ok) {
  190. /* Set window scaling on max possible window
  191. * See RFC1323 for an explanation of the limit to 14
  192. */
  193. space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
  194. space = min_t(u32, space, *window_clamp);
  195. while (space > 65535 && (*rcv_wscale) < 14) {
  196. space >>= 1;
  197. (*rcv_wscale)++;
  198. }
  199. }
  200. /* Set initial window to a value enough for senders starting with
  201. * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
  202. * a limit on the initial window when mss is larger than 1460.
  203. */
  204. if (mss > (1 << *rcv_wscale)) {
  205. int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
  206. if (mss > 1460)
  207. init_cwnd =
  208. max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
  209. /* when initializing use the value from init_rcv_wnd
  210. * rather than the default from above
  211. */
  212. if (init_rcv_wnd)
  213. *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
  214. else
  215. *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
  216. }
  217. /* Set the clamp no higher than max representable value */
  218. (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
  219. }
  220. EXPORT_SYMBOL(tcp_select_initial_window);
  221. /* Chose a new window to advertise, update state in tcp_sock for the
  222. * socket, and return result with RFC1323 scaling applied. The return
  223. * value can be stuffed directly into th->window for an outgoing
  224. * frame.
  225. */
  226. static u16 tcp_select_window(struct sock *sk)
  227. {
  228. struct tcp_sock *tp = tcp_sk(sk);
  229. u32 cur_win = tcp_receive_window(tp);
  230. u32 new_win = __tcp_select_window(sk);
  231. /* Never shrink the offered window */
  232. if (new_win < cur_win) {
  233. /* Danger Will Robinson!
  234. * Don't update rcv_wup/rcv_wnd here or else
  235. * we will not be able to advertise a zero
  236. * window in time. --DaveM
  237. *
  238. * Relax Will Robinson.
  239. */
  240. new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
  241. }
  242. tp->rcv_wnd = new_win;
  243. tp->rcv_wup = tp->rcv_nxt;
  244. /* Make sure we do not exceed the maximum possible
  245. * scaled window.
  246. */
  247. if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
  248. new_win = min(new_win, MAX_TCP_WINDOW);
  249. else
  250. new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
  251. /* RFC1323 scaling applied */
  252. new_win >>= tp->rx_opt.rcv_wscale;
  253. /* If we advertise zero window, disable fast path. */
  254. if (new_win == 0)
  255. tp->pred_flags = 0;
  256. return new_win;
  257. }
  258. /* Packet ECN state for a SYN-ACK */
  259. static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
  260. {
  261. TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
  262. if (!(tp->ecn_flags & TCP_ECN_OK))
  263. TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
  264. }
  265. /* Packet ECN state for a SYN. */
  266. static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
  267. {
  268. struct tcp_sock *tp = tcp_sk(sk);
  269. tp->ecn_flags = 0;
  270. if (sysctl_tcp_ecn == 1) {
  271. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
  272. tp->ecn_flags = TCP_ECN_OK;
  273. }
  274. }
  275. static __inline__ void
  276. TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
  277. {
  278. if (inet_rsk(req)->ecn_ok)
  279. th->ece = 1;
  280. }
  281. /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
  282. * be sent.
  283. */
  284. static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
  285. int tcp_header_len)
  286. {
  287. struct tcp_sock *tp = tcp_sk(sk);
  288. if (tp->ecn_flags & TCP_ECN_OK) {
  289. /* Not-retransmitted data segment: set ECT and inject CWR. */
  290. if (skb->len != tcp_header_len &&
  291. !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
  292. INET_ECN_xmit(sk);
  293. if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
  294. tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
  295. tcp_hdr(skb)->cwr = 1;
  296. skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
  297. }
  298. } else {
  299. /* ACK or retransmitted segment: clear ECT|CE */
  300. INET_ECN_dontxmit(sk);
  301. }
  302. if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
  303. tcp_hdr(skb)->ece = 1;
  304. }
  305. }
  306. /* Constructs common control bits of non-data skb. If SYN/FIN is present,
  307. * auto increment end seqno.
  308. */
  309. static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
  310. {
  311. skb->ip_summed = CHECKSUM_PARTIAL;
  312. skb->csum = 0;
  313. TCP_SKB_CB(skb)->tcp_flags = flags;
  314. TCP_SKB_CB(skb)->sacked = 0;
  315. skb_shinfo(skb)->gso_segs = 1;
  316. skb_shinfo(skb)->gso_size = 0;
  317. skb_shinfo(skb)->gso_type = 0;
  318. TCP_SKB_CB(skb)->seq = seq;
  319. if (flags & (TCPHDR_SYN | TCPHDR_FIN))
  320. seq++;
  321. TCP_SKB_CB(skb)->end_seq = seq;
  322. }
  323. static inline bool tcp_urg_mode(const struct tcp_sock *tp)
  324. {
  325. return tp->snd_una != tp->snd_up;
  326. }
  327. #define OPTION_SACK_ADVERTISE (1 << 0)
  328. #define OPTION_TS (1 << 1)
  329. #define OPTION_MD5 (1 << 2)
  330. #define OPTION_WSCALE (1 << 3)
  331. #define OPTION_COOKIE_EXTENSION (1 << 4)
  332. #define OPTION_FAST_OPEN_COOKIE (1 << 8)
  333. struct tcp_out_options {
  334. u16 options; /* bit field of OPTION_* */
  335. u16 mss; /* 0 to disable */
  336. u8 ws; /* window scale, 0 to disable */
  337. u8 num_sack_blocks; /* number of SACK blocks to include */
  338. u8 hash_size; /* bytes in hash_location */
  339. __u8 *hash_location; /* temporary pointer, overloaded */
  340. __u32 tsval, tsecr; /* need to include OPTION_TS */
  341. struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
  342. };
  343. /* The sysctl int routines are generic, so check consistency here.
  344. */
  345. static u8 tcp_cookie_size_check(u8 desired)
  346. {
  347. int cookie_size;
  348. if (desired > 0)
  349. /* previously specified */
  350. return desired;
  351. cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
  352. if (cookie_size <= 0)
  353. /* no default specified */
  354. return 0;
  355. if (cookie_size <= TCP_COOKIE_MIN)
  356. /* value too small, specify minimum */
  357. return TCP_COOKIE_MIN;
  358. if (cookie_size >= TCP_COOKIE_MAX)
  359. /* value too large, specify maximum */
  360. return TCP_COOKIE_MAX;
  361. if (cookie_size & 1)
  362. /* 8-bit multiple, illegal, fix it */
  363. cookie_size++;
  364. return (u8)cookie_size;
  365. }
  366. /* Write previously computed TCP options to the packet.
  367. *
  368. * Beware: Something in the Internet is very sensitive to the ordering of
  369. * TCP options, we learned this through the hard way, so be careful here.
  370. * Luckily we can at least blame others for their non-compliance but from
  371. * inter-operatibility perspective it seems that we're somewhat stuck with
  372. * the ordering which we have been using if we want to keep working with
  373. * those broken things (not that it currently hurts anybody as there isn't
  374. * particular reason why the ordering would need to be changed).
  375. *
  376. * At least SACK_PERM as the first option is known to lead to a disaster
  377. * (but it may well be that other scenarios fail similarly).
  378. */
  379. static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
  380. struct tcp_out_options *opts)
  381. {
  382. u16 options = opts->options; /* mungable copy */
  383. /* Having both authentication and cookies for security is redundant,
  384. * and there's certainly not enough room. Instead, the cookie-less
  385. * extension variant is proposed.
  386. *
  387. * Consider the pessimal case with authentication. The options
  388. * could look like:
  389. * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
  390. */
  391. if (unlikely(OPTION_MD5 & options)) {
  392. if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
  393. *ptr++ = htonl((TCPOPT_COOKIE << 24) |
  394. (TCPOLEN_COOKIE_BASE << 16) |
  395. (TCPOPT_MD5SIG << 8) |
  396. TCPOLEN_MD5SIG);
  397. } else {
  398. *ptr++ = htonl((TCPOPT_NOP << 24) |
  399. (TCPOPT_NOP << 16) |
  400. (TCPOPT_MD5SIG << 8) |
  401. TCPOLEN_MD5SIG);
  402. }
  403. options &= ~OPTION_COOKIE_EXTENSION;
  404. /* overload cookie hash location */
  405. opts->hash_location = (__u8 *)ptr;
  406. ptr += 4;
  407. }
  408. if (unlikely(opts->mss)) {
  409. *ptr++ = htonl((TCPOPT_MSS << 24) |
  410. (TCPOLEN_MSS << 16) |
  411. opts->mss);
  412. }
  413. if (likely(OPTION_TS & options)) {
  414. if (unlikely(OPTION_SACK_ADVERTISE & options)) {
  415. *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
  416. (TCPOLEN_SACK_PERM << 16) |
  417. (TCPOPT_TIMESTAMP << 8) |
  418. TCPOLEN_TIMESTAMP);
  419. options &= ~OPTION_SACK_ADVERTISE;
  420. } else {
  421. *ptr++ = htonl((TCPOPT_NOP << 24) |
  422. (TCPOPT_NOP << 16) |
  423. (TCPOPT_TIMESTAMP << 8) |
  424. TCPOLEN_TIMESTAMP);
  425. }
  426. *ptr++ = htonl(opts->tsval);
  427. *ptr++ = htonl(opts->tsecr);
  428. }
  429. /* Specification requires after timestamp, so do it now.
  430. *
  431. * Consider the pessimal case without authentication. The options
  432. * could look like:
  433. * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
  434. */
  435. if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
  436. __u8 *cookie_copy = opts->hash_location;
  437. u8 cookie_size = opts->hash_size;
  438. /* 8-bit multiple handled in tcp_cookie_size_check() above,
  439. * and elsewhere.
  440. */
  441. if (0x2 & cookie_size) {
  442. __u8 *p = (__u8 *)ptr;
  443. /* 16-bit multiple */
  444. *p++ = TCPOPT_COOKIE;
  445. *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
  446. *p++ = *cookie_copy++;
  447. *p++ = *cookie_copy++;
  448. ptr++;
  449. cookie_size -= 2;
  450. } else {
  451. /* 32-bit multiple */
  452. *ptr++ = htonl(((TCPOPT_NOP << 24) |
  453. (TCPOPT_NOP << 16) |
  454. (TCPOPT_COOKIE << 8) |
  455. TCPOLEN_COOKIE_BASE) +
  456. cookie_size);
  457. }
  458. if (cookie_size > 0) {
  459. memcpy(ptr, cookie_copy, cookie_size);
  460. ptr += (cookie_size / 4);
  461. }
  462. }
  463. if (unlikely(OPTION_SACK_ADVERTISE & options)) {
  464. *ptr++ = htonl((TCPOPT_NOP << 24) |
  465. (TCPOPT_NOP << 16) |
  466. (TCPOPT_SACK_PERM << 8) |
  467. TCPOLEN_SACK_PERM);
  468. }
  469. if (unlikely(OPTION_WSCALE & options)) {
  470. *ptr++ = htonl((TCPOPT_NOP << 24) |
  471. (TCPOPT_WINDOW << 16) |
  472. (TCPOLEN_WINDOW << 8) |
  473. opts->ws);
  474. }
  475. if (unlikely(opts->num_sack_blocks)) {
  476. struct tcp_sack_block *sp = tp->rx_opt.dsack ?
  477. tp->duplicate_sack : tp->selective_acks;
  478. int this_sack;
  479. *ptr++ = htonl((TCPOPT_NOP << 24) |
  480. (TCPOPT_NOP << 16) |
  481. (TCPOPT_SACK << 8) |
  482. (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
  483. TCPOLEN_SACK_PERBLOCK)));
  484. for (this_sack = 0; this_sack < opts->num_sack_blocks;
  485. ++this_sack) {
  486. *ptr++ = htonl(sp[this_sack].start_seq);
  487. *ptr++ = htonl(sp[this_sack].end_seq);
  488. }
  489. tp->rx_opt.dsack = 0;
  490. }
  491. if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
  492. struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
  493. *ptr++ = htonl((TCPOPT_EXP << 24) |
  494. ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
  495. TCPOPT_FASTOPEN_MAGIC);
  496. memcpy(ptr, foc->val, foc->len);
  497. if ((foc->len & 3) == 2) {
  498. u8 *align = ((u8 *)ptr) + foc->len;
  499. align[0] = align[1] = TCPOPT_NOP;
  500. }
  501. ptr += (foc->len + 3) >> 2;
  502. }
  503. }
  504. /* Compute TCP options for SYN packets. This is not the final
  505. * network wire format yet.
  506. */
  507. static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
  508. struct tcp_out_options *opts,
  509. struct tcp_md5sig_key **md5)
  510. {
  511. struct tcp_sock *tp = tcp_sk(sk);
  512. struct tcp_cookie_values *cvp = tp->cookie_values;
  513. unsigned int remaining = MAX_TCP_OPTION_SPACE;
  514. u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
  515. tcp_cookie_size_check(cvp->cookie_desired) :
  516. 0;
  517. struct tcp_fastopen_request *fastopen = tp->fastopen_req;
  518. #ifdef CONFIG_TCP_MD5SIG
  519. *md5 = tp->af_specific->md5_lookup(sk, sk);
  520. if (*md5) {
  521. opts->options |= OPTION_MD5;
  522. remaining -= TCPOLEN_MD5SIG_ALIGNED;
  523. }
  524. #else
  525. *md5 = NULL;
  526. #endif
  527. /* We always get an MSS option. The option bytes which will be seen in
  528. * normal data packets should timestamps be used, must be in the MSS
  529. * advertised. But we subtract them from tp->mss_cache so that
  530. * calculations in tcp_sendmsg are simpler etc. So account for this
  531. * fact here if necessary. If we don't do this correctly, as a
  532. * receiver we won't recognize data packets as being full sized when we
  533. * should, and thus we won't abide by the delayed ACK rules correctly.
  534. * SACKs don't matter, we never delay an ACK when we have any of those
  535. * going out. */
  536. opts->mss = tcp_advertise_mss(sk);
  537. remaining -= TCPOLEN_MSS_ALIGNED;
  538. if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
  539. opts->options |= OPTION_TS;
  540. opts->tsval = TCP_SKB_CB(skb)->when;
  541. opts->tsecr = tp->rx_opt.ts_recent;
  542. remaining -= TCPOLEN_TSTAMP_ALIGNED;
  543. }
  544. if (likely(sysctl_tcp_window_scaling)) {
  545. opts->ws = tp->rx_opt.rcv_wscale;
  546. opts->options |= OPTION_WSCALE;
  547. remaining -= TCPOLEN_WSCALE_ALIGNED;
  548. }
  549. if (likely(sysctl_tcp_sack)) {
  550. opts->options |= OPTION_SACK_ADVERTISE;
  551. if (unlikely(!(OPTION_TS & opts->options)))
  552. remaining -= TCPOLEN_SACKPERM_ALIGNED;
  553. }
  554. if (fastopen && fastopen->cookie.len >= 0) {
  555. u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
  556. need = (need + 3) & ~3U; /* Align to 32 bits */
  557. if (remaining >= need) {
  558. opts->options |= OPTION_FAST_OPEN_COOKIE;
  559. opts->fastopen_cookie = &fastopen->cookie;
  560. remaining -= need;
  561. tp->syn_fastopen = 1;
  562. }
  563. }
  564. /* Note that timestamps are required by the specification.
  565. *
  566. * Odd numbers of bytes are prohibited by the specification, ensuring
  567. * that the cookie is 16-bit aligned, and the resulting cookie pair is
  568. * 32-bit aligned.
  569. */
  570. if (*md5 == NULL &&
  571. (OPTION_TS & opts->options) &&
  572. cookie_size > 0) {
  573. int need = TCPOLEN_COOKIE_BASE + cookie_size;
  574. if (0x2 & need) {
  575. /* 32-bit multiple */
  576. need += 2; /* NOPs */
  577. if (need > remaining) {
  578. /* try shrinking cookie to fit */
  579. cookie_size -= 2;
  580. need -= 4;
  581. }
  582. }
  583. while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
  584. cookie_size -= 4;
  585. need -= 4;
  586. }
  587. if (TCP_COOKIE_MIN <= cookie_size) {
  588. opts->options |= OPTION_COOKIE_EXTENSION;
  589. opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
  590. opts->hash_size = cookie_size;
  591. /* Remember for future incarnations. */
  592. cvp->cookie_desired = cookie_size;
  593. if (cvp->cookie_desired != cvp->cookie_pair_size) {
  594. /* Currently use random bytes as a nonce,
  595. * assuming these are completely unpredictable
  596. * by hostile users of the same system.
  597. */
  598. get_random_bytes(&cvp->cookie_pair[0],
  599. cookie_size);
  600. cvp->cookie_pair_size = cookie_size;
  601. }
  602. remaining -= need;
  603. }
  604. }
  605. return MAX_TCP_OPTION_SPACE - remaining;
  606. }
  607. /* Set up TCP options for SYN-ACKs. */
  608. static unsigned int tcp_synack_options(struct sock *sk,
  609. struct request_sock *req,
  610. unsigned int mss, struct sk_buff *skb,
  611. struct tcp_out_options *opts,
  612. struct tcp_md5sig_key **md5,
  613. struct tcp_extend_values *xvp)
  614. {
  615. struct inet_request_sock *ireq = inet_rsk(req);
  616. unsigned int remaining = MAX_TCP_OPTION_SPACE;
  617. u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
  618. xvp->cookie_plus :
  619. 0;
  620. #ifdef CONFIG_TCP_MD5SIG
  621. *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
  622. if (*md5) {
  623. opts->options |= OPTION_MD5;
  624. remaining -= TCPOLEN_MD5SIG_ALIGNED;
  625. /* We can't fit any SACK blocks in a packet with MD5 + TS
  626. * options. There was discussion about disabling SACK
  627. * rather than TS in order to fit in better with old,
  628. * buggy kernels, but that was deemed to be unnecessary.
  629. */
  630. ireq->tstamp_ok &= !ireq->sack_ok;
  631. }
  632. #else
  633. *md5 = NULL;
  634. #endif
  635. /* We always send an MSS option. */
  636. opts->mss = mss;
  637. remaining -= TCPOLEN_MSS_ALIGNED;
  638. if (likely(ireq->wscale_ok)) {
  639. opts->ws = ireq->rcv_wscale;
  640. opts->options |= OPTION_WSCALE;
  641. remaining -= TCPOLEN_WSCALE_ALIGNED;
  642. }
  643. if (likely(ireq->tstamp_ok)) {
  644. opts->options |= OPTION_TS;
  645. opts->tsval = TCP_SKB_CB(skb)->when;
  646. opts->tsecr = req->ts_recent;
  647. remaining -= TCPOLEN_TSTAMP_ALIGNED;
  648. }
  649. if (likely(ireq->sack_ok)) {
  650. opts->options |= OPTION_SACK_ADVERTISE;
  651. if (unlikely(!ireq->tstamp_ok))
  652. remaining -= TCPOLEN_SACKPERM_ALIGNED;
  653. }
  654. /* Similar rationale to tcp_syn_options() applies here, too.
  655. * If the <SYN> options fit, the same options should fit now!
  656. */
  657. if (*md5 == NULL &&
  658. ireq->tstamp_ok &&
  659. cookie_plus > TCPOLEN_COOKIE_BASE) {
  660. int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
  661. if (0x2 & need) {
  662. /* 32-bit multiple */
  663. need += 2; /* NOPs */
  664. }
  665. if (need <= remaining) {
  666. opts->options |= OPTION_COOKIE_EXTENSION;
  667. opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
  668. remaining -= need;
  669. } else {
  670. /* There's no error return, so flag it. */
  671. xvp->cookie_out_never = 1; /* true */
  672. opts->hash_size = 0;
  673. }
  674. }
  675. return MAX_TCP_OPTION_SPACE - remaining;
  676. }
  677. /* Compute TCP options for ESTABLISHED sockets. This is not the
  678. * final wire format yet.
  679. */
  680. static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
  681. struct tcp_out_options *opts,
  682. struct tcp_md5sig_key **md5)
  683. {
  684. struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
  685. struct tcp_sock *tp = tcp_sk(sk);
  686. unsigned int size = 0;
  687. unsigned int eff_sacks;
  688. #ifdef CONFIG_TCP_MD5SIG
  689. *md5 = tp->af_specific->md5_lookup(sk, sk);
  690. if (unlikely(*md5)) {
  691. opts->options |= OPTION_MD5;
  692. size += TCPOLEN_MD5SIG_ALIGNED;
  693. }
  694. #else
  695. *md5 = NULL;
  696. #endif
  697. if (likely(tp->rx_opt.tstamp_ok)) {
  698. opts->options |= OPTION_TS;
  699. opts->tsval = tcb ? tcb->when : 0;
  700. opts->tsecr = tp->rx_opt.ts_recent;
  701. size += TCPOLEN_TSTAMP_ALIGNED;
  702. }
  703. eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
  704. if (unlikely(eff_sacks)) {
  705. const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
  706. opts->num_sack_blocks =
  707. min_t(unsigned int, eff_sacks,
  708. (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
  709. TCPOLEN_SACK_PERBLOCK);
  710. size += TCPOLEN_SACK_BASE_ALIGNED +
  711. opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
  712. }
  713. return size;
  714. }
  715. /* TCP SMALL QUEUES (TSQ)
  716. *
  717. * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
  718. * to reduce RTT and bufferbloat.
  719. * We do this using a special skb destructor (tcp_wfree).
  720. *
  721. * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
  722. * needs to be reallocated in a driver.
  723. * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
  724. *
  725. * Since transmit from skb destructor is forbidden, we use a tasklet
  726. * to process all sockets that eventually need to send more skbs.
  727. * We use one tasklet per cpu, with its own queue of sockets.
  728. */
  729. struct tsq_tasklet {
  730. struct tasklet_struct tasklet;
  731. struct list_head head; /* queue of tcp sockets */
  732. };
  733. static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
  734. static void tcp_tsq_handler(struct sock *sk)
  735. {
  736. if ((1 << sk->sk_state) &
  737. (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
  738. TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
  739. tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
  740. }
  741. /*
  742. * One tasklest per cpu tries to send more skbs.
  743. * We run in tasklet context but need to disable irqs when
  744. * transfering tsq->head because tcp_wfree() might
  745. * interrupt us (non NAPI drivers)
  746. */
  747. static void tcp_tasklet_func(unsigned long data)
  748. {
  749. struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
  750. LIST_HEAD(list);
  751. unsigned long flags;
  752. struct list_head *q, *n;
  753. struct tcp_sock *tp;
  754. struct sock *sk;
  755. local_irq_save(flags);
  756. list_splice_init(&tsq->head, &list);
  757. local_irq_restore(flags);
  758. list_for_each_safe(q, n, &list) {
  759. tp = list_entry(q, struct tcp_sock, tsq_node);
  760. list_del(&tp->tsq_node);
  761. sk = (struct sock *)tp;
  762. bh_lock_sock(sk);
  763. if (!sock_owned_by_user(sk)) {
  764. tcp_tsq_handler(sk);
  765. } else {
  766. /* defer the work to tcp_release_cb() */
  767. set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
  768. }
  769. bh_unlock_sock(sk);
  770. clear_bit(TSQ_QUEUED, &tp->tsq_flags);
  771. sk_free(sk);
  772. }
  773. }
  774. #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
  775. (1UL << TCP_WRITE_TIMER_DEFERRED) | \
  776. (1UL << TCP_DELACK_TIMER_DEFERRED) | \
  777. (1UL << TCP_MTU_REDUCED_DEFERRED))
  778. /**
  779. * tcp_release_cb - tcp release_sock() callback
  780. * @sk: socket
  781. *
  782. * called from release_sock() to perform protocol dependent
  783. * actions before socket release.
  784. */
  785. void tcp_release_cb(struct sock *sk)
  786. {
  787. struct tcp_sock *tp = tcp_sk(sk);
  788. unsigned long flags, nflags;
  789. /* perform an atomic operation only if at least one flag is set */
  790. do {
  791. flags = tp->tsq_flags;
  792. if (!(flags & TCP_DEFERRED_ALL))
  793. return;
  794. nflags = flags & ~TCP_DEFERRED_ALL;
  795. } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
  796. if (flags & (1UL << TCP_TSQ_DEFERRED))
  797. tcp_tsq_handler(sk);
  798. if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
  799. tcp_write_timer_handler(sk);
  800. __sock_put(sk);
  801. }
  802. if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
  803. tcp_delack_timer_handler(sk);
  804. __sock_put(sk);
  805. }
  806. if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
  807. sk->sk_prot->mtu_reduced(sk);
  808. __sock_put(sk);
  809. }
  810. }
  811. EXPORT_SYMBOL(tcp_release_cb);
  812. void __init tcp_tasklet_init(void)
  813. {
  814. int i;
  815. for_each_possible_cpu(i) {
  816. struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
  817. INIT_LIST_HEAD(&tsq->head);
  818. tasklet_init(&tsq->tasklet,
  819. tcp_tasklet_func,
  820. (unsigned long)tsq);
  821. }
  822. }
  823. /*
  824. * Write buffer destructor automatically called from kfree_skb.
  825. * We cant xmit new skbs from this context, as we might already
  826. * hold qdisc lock.
  827. */
  828. static void tcp_wfree(struct sk_buff *skb)
  829. {
  830. struct sock *sk = skb->sk;
  831. struct tcp_sock *tp = tcp_sk(sk);
  832. if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
  833. !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
  834. unsigned long flags;
  835. struct tsq_tasklet *tsq;
  836. /* Keep a ref on socket.
  837. * This last ref will be released in tcp_tasklet_func()
  838. */
  839. atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
  840. /* queue this socket to tasklet queue */
  841. local_irq_save(flags);
  842. tsq = &__get_cpu_var(tsq_tasklet);
  843. list_add(&tp->tsq_node, &tsq->head);
  844. tasklet_schedule(&tsq->tasklet);
  845. local_irq_restore(flags);
  846. } else {
  847. sock_wfree(skb);
  848. }
  849. }
  850. /* This routine actually transmits TCP packets queued in by
  851. * tcp_do_sendmsg(). This is used by both the initial
  852. * transmission and possible later retransmissions.
  853. * All SKB's seen here are completely headerless. It is our
  854. * job to build the TCP header, and pass the packet down to
  855. * IP so it can do the same plus pass the packet off to the
  856. * device.
  857. *
  858. * We are working here with either a clone of the original
  859. * SKB, or a fresh unique copy made by the retransmit engine.
  860. */
  861. static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
  862. gfp_t gfp_mask)
  863. {
  864. const struct inet_connection_sock *icsk = inet_csk(sk);
  865. struct inet_sock *inet;
  866. struct tcp_sock *tp;
  867. struct tcp_skb_cb *tcb;
  868. struct tcp_out_options opts;
  869. unsigned int tcp_options_size, tcp_header_size;
  870. struct tcp_md5sig_key *md5;
  871. struct tcphdr *th;
  872. int err;
  873. BUG_ON(!skb || !tcp_skb_pcount(skb));
  874. /* If congestion control is doing timestamping, we must
  875. * take such a timestamp before we potentially clone/copy.
  876. */
  877. if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
  878. __net_timestamp(skb);
  879. if (likely(clone_it)) {
  880. if (unlikely(skb_cloned(skb)))
  881. skb = pskb_copy(skb, gfp_mask);
  882. else
  883. skb = skb_clone(skb, gfp_mask);
  884. if (unlikely(!skb))
  885. return -ENOBUFS;
  886. }
  887. inet = inet_sk(sk);
  888. tp = tcp_sk(sk);
  889. tcb = TCP_SKB_CB(skb);
  890. memset(&opts, 0, sizeof(opts));
  891. if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
  892. tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
  893. else
  894. tcp_options_size = tcp_established_options(sk, skb, &opts,
  895. &md5);
  896. tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
  897. if (tcp_packets_in_flight(tp) == 0) {
  898. tcp_ca_event(sk, CA_EVENT_TX_START);
  899. skb->ooo_okay = 1;
  900. } else
  901. skb->ooo_okay = 0;
  902. skb_push(skb, tcp_header_size);
  903. skb_reset_transport_header(skb);
  904. skb_orphan(skb);
  905. skb->sk = sk;
  906. skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
  907. tcp_wfree : sock_wfree;
  908. atomic_add(skb->truesize, &sk->sk_wmem_alloc);
  909. /* Build TCP header and checksum it. */
  910. th = tcp_hdr(skb);
  911. th->source = inet->inet_sport;
  912. th->dest = inet->inet_dport;
  913. th->seq = htonl(tcb->seq);
  914. th->ack_seq = htonl(tp->rcv_nxt);
  915. *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
  916. tcb->tcp_flags);
  917. if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
  918. /* RFC1323: The window in SYN & SYN/ACK segments
  919. * is never scaled.
  920. */
  921. th->window = htons(min(tp->rcv_wnd, 65535U));
  922. } else {
  923. th->window = htons(tcp_select_window(sk));
  924. }
  925. th->check = 0;
  926. th->urg_ptr = 0;
  927. /* The urg_mode check is necessary during a below snd_una win probe */
  928. if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
  929. if (before(tp->snd_up, tcb->seq + 0x10000)) {
  930. th->urg_ptr = htons(tp->snd_up - tcb->seq);
  931. th->urg = 1;
  932. } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
  933. th->urg_ptr = htons(0xFFFF);
  934. th->urg = 1;
  935. }
  936. }
  937. tcp_options_write((__be32 *)(th + 1), tp, &opts);
  938. if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
  939. TCP_ECN_send(sk, skb, tcp_header_size);
  940. #ifdef CONFIG_TCP_MD5SIG
  941. /* Calculate the MD5 hash, as we have all we need now */
  942. if (md5) {
  943. sk_nocaps_add(sk, NETIF_F_GSO_MASK);
  944. tp->af_specific->calc_md5_hash(opts.hash_location,
  945. md5, sk, NULL, skb);
  946. }
  947. #endif
  948. icsk->icsk_af_ops->send_check(sk, skb);
  949. if (likely(tcb->tcp_flags & TCPHDR_ACK))
  950. tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
  951. if (skb->len != tcp_header_size)
  952. tcp_event_data_sent(tp, sk);
  953. if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
  954. TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
  955. tcp_skb_pcount(skb));
  956. err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
  957. if (likely(err <= 0))
  958. return err;
  959. tcp_enter_cwr(sk, 1);
  960. return net_xmit_eval(err);
  961. }
  962. /* This routine just queues the buffer for sending.
  963. *
  964. * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
  965. * otherwise socket can stall.
  966. */
  967. static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
  968. {
  969. struct tcp_sock *tp = tcp_sk(sk);
  970. /* Advance write_seq and place onto the write_queue. */
  971. tp->write_seq = TCP_SKB_CB(skb)->end_seq;
  972. skb_header_release(skb);
  973. tcp_add_write_queue_tail(sk, skb);
  974. sk->sk_wmem_queued += skb->truesize;
  975. sk_mem_charge(sk, skb->truesize);
  976. }
  977. /* Initialize TSO segments for a packet. */
  978. static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
  979. unsigned int mss_now)
  980. {
  981. if (skb->len <= mss_now || !sk_can_gso(sk) ||
  982. skb->ip_summed == CHECKSUM_NONE) {
  983. /* Avoid the costly divide in the normal
  984. * non-TSO case.
  985. */
  986. skb_shinfo(skb)->gso_segs = 1;
  987. skb_shinfo(skb)->gso_size = 0;
  988. skb_shinfo(skb)->gso_type = 0;
  989. } else {
  990. skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
  991. skb_shinfo(skb)->gso_size = mss_now;
  992. skb_shinfo(skb)->gso_type = sk->sk_gso_type;
  993. }
  994. }
  995. /* When a modification to fackets out becomes necessary, we need to check
  996. * skb is counted to fackets_out or not.
  997. */
  998. static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
  999. int decr)
  1000. {
  1001. struct tcp_sock *tp = tcp_sk(sk);
  1002. if (!tp->sacked_out || tcp_is_reno(tp))
  1003. return;
  1004. if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
  1005. tp->fackets_out -= decr;
  1006. }
  1007. /* Pcount in the middle of the write queue got changed, we need to do various
  1008. * tweaks to fix counters
  1009. */
  1010. static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
  1011. {
  1012. struct tcp_sock *tp = tcp_sk(sk);
  1013. tp->packets_out -= decr;
  1014. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
  1015. tp->sacked_out -= decr;
  1016. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
  1017. tp->retrans_out -= decr;
  1018. if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
  1019. tp->lost_out -= decr;
  1020. /* Reno case is special. Sigh... */
  1021. if (tcp_is_reno(tp) && decr > 0)
  1022. tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
  1023. tcp_adjust_fackets_out(sk, skb, decr);
  1024. if (tp->lost_skb_hint &&
  1025. before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
  1026. (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
  1027. tp->lost_cnt_hint -= decr;
  1028. tcp_verify_left_out(tp);
  1029. }
  1030. /* Function to create two new TCP segments. Shrinks the given segment
  1031. * to the specified size and appends a new segment with the rest of the
  1032. * packet to the list. This won't be called frequently, I hope.
  1033. * Remember, these are still headerless SKBs at this point.
  1034. */
  1035. int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
  1036. unsigned int mss_now)
  1037. {
  1038. struct tcp_sock *tp = tcp_sk(sk);
  1039. struct sk_buff *buff;
  1040. int nsize, old_factor;
  1041. int nlen;
  1042. u8 flags;
  1043. if (WARN_ON(len > skb->len))
  1044. return -EINVAL;
  1045. nsize = skb_headlen(skb) - len;
  1046. if (nsize < 0)
  1047. nsize = 0;
  1048. if (skb_cloned(skb) &&
  1049. skb_is_nonlinear(skb) &&
  1050. pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
  1051. return -ENOMEM;
  1052. /* Get a new skb... force flag on. */
  1053. buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
  1054. if (buff == NULL)
  1055. return -ENOMEM; /* We'll just try again later. */
  1056. sk->sk_wmem_queued += buff->truesize;
  1057. sk_mem_charge(sk, buff->truesize);
  1058. nlen = skb->len - len - nsize;
  1059. buff->truesize += nlen;
  1060. skb->truesize -= nlen;
  1061. /* Correct the sequence numbers. */
  1062. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  1063. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  1064. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  1065. /* PSH and FIN should only be set in the second packet. */
  1066. flags = TCP_SKB_CB(skb)->tcp_flags;
  1067. TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
  1068. TCP_SKB_CB(buff)->tcp_flags = flags;
  1069. TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
  1070. if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
  1071. /* Copy and checksum data tail into the new buffer. */
  1072. buff->csum = csum_partial_copy_nocheck(skb->data + len,
  1073. skb_put(buff, nsize),
  1074. nsize, 0);
  1075. skb_trim(skb, len);
  1076. skb->csum = csum_block_sub(skb->csum, buff->csum, len);
  1077. } else {
  1078. skb->ip_summed = CHECKSUM_PARTIAL;
  1079. skb_split(skb, buff, len);
  1080. }
  1081. buff->ip_summed = skb->ip_summed;
  1082. /* Looks stupid, but our code really uses when of
  1083. * skbs, which it never sent before. --ANK
  1084. */
  1085. TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
  1086. buff->tstamp = skb->tstamp;
  1087. old_factor = tcp_skb_pcount(skb);
  1088. /* Fix up tso_factor for both original and new SKB. */
  1089. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1090. tcp_set_skb_tso_segs(sk, buff, mss_now);
  1091. /* If this packet has been sent out already, we must
  1092. * adjust the various packet counters.
  1093. */
  1094. if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
  1095. int diff = old_factor - tcp_skb_pcount(skb) -
  1096. tcp_skb_pcount(buff);
  1097. if (diff)
  1098. tcp_adjust_pcount(sk, skb, diff);
  1099. }
  1100. /* Link BUFF into the send queue. */
  1101. skb_header_release(buff);
  1102. tcp_insert_write_queue_after(skb, buff, sk);
  1103. return 0;
  1104. }
  1105. /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
  1106. * eventually). The difference is that pulled data not copied, but
  1107. * immediately discarded.
  1108. */
  1109. static void __pskb_trim_head(struct sk_buff *skb, int len)
  1110. {
  1111. int i, k, eat;
  1112. eat = min_t(int, len, skb_headlen(skb));
  1113. if (eat) {
  1114. __skb_pull(skb, eat);
  1115. skb->avail_size -= eat;
  1116. len -= eat;
  1117. if (!len)
  1118. return;
  1119. }
  1120. eat = len;
  1121. k = 0;
  1122. for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  1123. int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
  1124. if (size <= eat) {
  1125. skb_frag_unref(skb, i);
  1126. eat -= size;
  1127. } else {
  1128. skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
  1129. if (eat) {
  1130. skb_shinfo(skb)->frags[k].page_offset += eat;
  1131. skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
  1132. eat = 0;
  1133. }
  1134. k++;
  1135. }
  1136. }
  1137. skb_shinfo(skb)->nr_frags = k;
  1138. skb_reset_tail_pointer(skb);
  1139. skb->data_len -= len;
  1140. skb->len = skb->data_len;
  1141. }
  1142. /* Remove acked data from a packet in the transmit queue. */
  1143. int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
  1144. {
  1145. if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
  1146. return -ENOMEM;
  1147. __pskb_trim_head(skb, len);
  1148. TCP_SKB_CB(skb)->seq += len;
  1149. skb->ip_summed = CHECKSUM_PARTIAL;
  1150. skb->truesize -= len;
  1151. sk->sk_wmem_queued -= len;
  1152. sk_mem_uncharge(sk, len);
  1153. sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
  1154. /* Any change of skb->len requires recalculation of tso factor. */
  1155. if (tcp_skb_pcount(skb) > 1)
  1156. tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
  1157. return 0;
  1158. }
  1159. /* Calculate MSS. Not accounting for SACKs here. */
  1160. int tcp_mtu_to_mss(struct sock *sk, int pmtu)
  1161. {
  1162. const struct tcp_sock *tp = tcp_sk(sk);
  1163. const struct inet_connection_sock *icsk = inet_csk(sk);
  1164. int mss_now;
  1165. /* Calculate base mss without TCP options:
  1166. It is MMS_S - sizeof(tcphdr) of rfc1122
  1167. */
  1168. mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
  1169. /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
  1170. if (icsk->icsk_af_ops->net_frag_header_len) {
  1171. const struct dst_entry *dst = __sk_dst_get(sk);
  1172. if (dst && dst_allfrag(dst))
  1173. mss_now -= icsk->icsk_af_ops->net_frag_header_len;
  1174. }
  1175. /* Clamp it (mss_clamp does not include tcp options) */
  1176. if (mss_now > tp->rx_opt.mss_clamp)
  1177. mss_now = tp->rx_opt.mss_clamp;
  1178. /* Now subtract optional transport overhead */
  1179. mss_now -= icsk->icsk_ext_hdr_len;
  1180. /* Then reserve room for full set of TCP options and 8 bytes of data */
  1181. if (mss_now < 48)
  1182. mss_now = 48;
  1183. /* Now subtract TCP options size, not including SACKs */
  1184. mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
  1185. return mss_now;
  1186. }
  1187. /* Inverse of above */
  1188. int tcp_mss_to_mtu(struct sock *sk, int mss)
  1189. {
  1190. const struct tcp_sock *tp = tcp_sk(sk);
  1191. const struct inet_connection_sock *icsk = inet_csk(sk);
  1192. int mtu;
  1193. mtu = mss +
  1194. tp->tcp_header_len +
  1195. icsk->icsk_ext_hdr_len +
  1196. icsk->icsk_af_ops->net_header_len;
  1197. /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
  1198. if (icsk->icsk_af_ops->net_frag_header_len) {
  1199. const struct dst_entry *dst = __sk_dst_get(sk);
  1200. if (dst && dst_allfrag(dst))
  1201. mtu += icsk->icsk_af_ops->net_frag_header_len;
  1202. }
  1203. return mtu;
  1204. }
  1205. /* MTU probing init per socket */
  1206. void tcp_mtup_init(struct sock *sk)
  1207. {
  1208. struct tcp_sock *tp = tcp_sk(sk);
  1209. struct inet_connection_sock *icsk = inet_csk(sk);
  1210. icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
  1211. icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
  1212. icsk->icsk_af_ops->net_header_len;
  1213. icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
  1214. icsk->icsk_mtup.probe_size = 0;
  1215. }
  1216. EXPORT_SYMBOL(tcp_mtup_init);
  1217. /* This function synchronize snd mss to current pmtu/exthdr set.
  1218. tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
  1219. for TCP options, but includes only bare TCP header.
  1220. tp->rx_opt.mss_clamp is mss negotiated at connection setup.
  1221. It is minimum of user_mss and mss received with SYN.
  1222. It also does not include TCP options.
  1223. inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
  1224. tp->mss_cache is current effective sending mss, including
  1225. all tcp options except for SACKs. It is evaluated,
  1226. taking into account current pmtu, but never exceeds
  1227. tp->rx_opt.mss_clamp.
  1228. NOTE1. rfc1122 clearly states that advertised MSS
  1229. DOES NOT include either tcp or ip options.
  1230. NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
  1231. are READ ONLY outside this function. --ANK (980731)
  1232. */
  1233. unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
  1234. {
  1235. struct tcp_sock *tp = tcp_sk(sk);
  1236. struct inet_connection_sock *icsk = inet_csk(sk);
  1237. int mss_now;
  1238. if (icsk->icsk_mtup.search_high > pmtu)
  1239. icsk->icsk_mtup.search_high = pmtu;
  1240. mss_now = tcp_mtu_to_mss(sk, pmtu);
  1241. mss_now = tcp_bound_to_half_wnd(tp, mss_now);
  1242. /* And store cached results */
  1243. icsk->icsk_pmtu_cookie = pmtu;
  1244. if (icsk->icsk_mtup.enabled)
  1245. mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
  1246. tp->mss_cache = mss_now;
  1247. return mss_now;
  1248. }
  1249. EXPORT_SYMBOL(tcp_sync_mss);
  1250. /* Compute the current effective MSS, taking SACKs and IP options,
  1251. * and even PMTU discovery events into account.
  1252. */
  1253. unsigned int tcp_current_mss(struct sock *sk)
  1254. {
  1255. const struct tcp_sock *tp = tcp_sk(sk);
  1256. const struct dst_entry *dst = __sk_dst_get(sk);
  1257. u32 mss_now;
  1258. unsigned int header_len;
  1259. struct tcp_out_options opts;
  1260. struct tcp_md5sig_key *md5;
  1261. mss_now = tp->mss_cache;
  1262. if (dst) {
  1263. u32 mtu = dst_mtu(dst);
  1264. if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
  1265. mss_now = tcp_sync_mss(sk, mtu);
  1266. }
  1267. header_len = tcp_established_options(sk, NULL, &opts, &md5) +
  1268. sizeof(struct tcphdr);
  1269. /* The mss_cache is sized based on tp->tcp_header_len, which assumes
  1270. * some common options. If this is an odd packet (because we have SACK
  1271. * blocks etc) then our calculated header_len will be different, and
  1272. * we have to adjust mss_now correspondingly */
  1273. if (header_len != tp->tcp_header_len) {
  1274. int delta = (int) header_len - tp->tcp_header_len;
  1275. mss_now -= delta;
  1276. }
  1277. return mss_now;
  1278. }
  1279. /* Congestion window validation. (RFC2861) */
  1280. static void tcp_cwnd_validate(struct sock *sk)
  1281. {
  1282. struct tcp_sock *tp = tcp_sk(sk);
  1283. if (tp->packets_out >= tp->snd_cwnd) {
  1284. /* Network is feed fully. */
  1285. tp->snd_cwnd_used = 0;
  1286. tp->snd_cwnd_stamp = tcp_time_stamp;
  1287. } else {
  1288. /* Network starves. */
  1289. if (tp->packets_out > tp->snd_cwnd_used)
  1290. tp->snd_cwnd_used = tp->packets_out;
  1291. if (sysctl_tcp_slow_start_after_idle &&
  1292. (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
  1293. tcp_cwnd_application_limited(sk);
  1294. }
  1295. }
  1296. /* Returns the portion of skb which can be sent right away without
  1297. * introducing MSS oddities to segment boundaries. In rare cases where
  1298. * mss_now != mss_cache, we will request caller to create a small skb
  1299. * per input skb which could be mostly avoided here (if desired).
  1300. *
  1301. * We explicitly want to create a request for splitting write queue tail
  1302. * to a small skb for Nagle purposes while avoiding unnecessary modulos,
  1303. * thus all the complexity (cwnd_len is always MSS multiple which we
  1304. * return whenever allowed by the other factors). Basically we need the
  1305. * modulo only when the receiver window alone is the limiting factor or
  1306. * when we would be allowed to send the split-due-to-Nagle skb fully.
  1307. */
  1308. static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
  1309. unsigned int mss_now, unsigned int max_segs)
  1310. {
  1311. const struct tcp_sock *tp = tcp_sk(sk);
  1312. u32 needed, window, max_len;
  1313. window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  1314. max_len = mss_now * max_segs;
  1315. if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
  1316. return max_len;
  1317. needed = min(skb->len, window);
  1318. if (max_len <= needed)
  1319. return max_len;
  1320. return needed - needed % mss_now;
  1321. }
  1322. /* Can at least one segment of SKB be sent right now, according to the
  1323. * congestion window rules? If so, return how many segments are allowed.
  1324. */
  1325. static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
  1326. const struct sk_buff *skb)
  1327. {
  1328. u32 in_flight, cwnd;
  1329. /* Don't be strict about the congestion window for the final FIN. */
  1330. if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
  1331. tcp_skb_pcount(skb) == 1)
  1332. return 1;
  1333. in_flight = tcp_packets_in_flight(tp);
  1334. cwnd = tp->snd_cwnd;
  1335. if (in_flight < cwnd)
  1336. return (cwnd - in_flight);
  1337. return 0;
  1338. }
  1339. /* Initialize TSO state of a skb.
  1340. * This must be invoked the first time we consider transmitting
  1341. * SKB onto the wire.
  1342. */
  1343. static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
  1344. unsigned int mss_now)
  1345. {
  1346. int tso_segs = tcp_skb_pcount(skb);
  1347. if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
  1348. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1349. tso_segs = tcp_skb_pcount(skb);
  1350. }
  1351. return tso_segs;
  1352. }
  1353. /* Minshall's variant of the Nagle send check. */
  1354. static inline bool tcp_minshall_check(const struct tcp_sock *tp)
  1355. {
  1356. return after(tp->snd_sml, tp->snd_una) &&
  1357. !after(tp->snd_sml, tp->snd_nxt);
  1358. }
  1359. /* Return false, if packet can be sent now without violation Nagle's rules:
  1360. * 1. It is full sized.
  1361. * 2. Or it contains FIN. (already checked by caller)
  1362. * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
  1363. * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
  1364. * With Minshall's modification: all sent small packets are ACKed.
  1365. */
  1366. static inline bool tcp_nagle_check(const struct tcp_sock *tp,
  1367. const struct sk_buff *skb,
  1368. unsigned int mss_now, int nonagle)
  1369. {
  1370. return skb->len < mss_now &&
  1371. ((nonagle & TCP_NAGLE_CORK) ||
  1372. (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
  1373. }
  1374. /* Return true if the Nagle test allows this packet to be
  1375. * sent now.
  1376. */
  1377. static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
  1378. unsigned int cur_mss, int nonagle)
  1379. {
  1380. /* Nagle rule does not apply to frames, which sit in the middle of the
  1381. * write_queue (they have no chances to get new data).
  1382. *
  1383. * This is implemented in the callers, where they modify the 'nonagle'
  1384. * argument based upon the location of SKB in the send queue.
  1385. */
  1386. if (nonagle & TCP_NAGLE_PUSH)
  1387. return true;
  1388. /* Don't use the nagle rule for urgent data (or for the final FIN).
  1389. * Nagle can be ignored during F-RTO too (see RFC4138).
  1390. */
  1391. if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
  1392. (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
  1393. return true;
  1394. if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
  1395. return true;
  1396. return false;
  1397. }
  1398. /* Does at least the first segment of SKB fit into the send window? */
  1399. static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
  1400. const struct sk_buff *skb,
  1401. unsigned int cur_mss)
  1402. {
  1403. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  1404. if (skb->len > cur_mss)
  1405. end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
  1406. return !after(end_seq, tcp_wnd_end(tp));
  1407. }
  1408. /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
  1409. * should be put on the wire right now. If so, it returns the number of
  1410. * packets allowed by the congestion window.
  1411. */
  1412. static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
  1413. unsigned int cur_mss, int nonagle)
  1414. {
  1415. const struct tcp_sock *tp = tcp_sk(sk);
  1416. unsigned int cwnd_quota;
  1417. tcp_init_tso_segs(sk, skb, cur_mss);
  1418. if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
  1419. return 0;
  1420. cwnd_quota = tcp_cwnd_test(tp, skb);
  1421. if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
  1422. cwnd_quota = 0;
  1423. return cwnd_quota;
  1424. }
  1425. /* Test if sending is allowed right now. */
  1426. bool tcp_may_send_now(struct sock *sk)
  1427. {
  1428. const struct tcp_sock *tp = tcp_sk(sk);
  1429. struct sk_buff *skb = tcp_send_head(sk);
  1430. return skb &&
  1431. tcp_snd_test(sk, skb, tcp_current_mss(sk),
  1432. (tcp_skb_is_last(sk, skb) ?
  1433. tp->nonagle : TCP_NAGLE_PUSH));
  1434. }
  1435. /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
  1436. * which is put after SKB on the list. It is very much like
  1437. * tcp_fragment() except that it may make several kinds of assumptions
  1438. * in order to speed up the splitting operation. In particular, we
  1439. * know that all the data is in scatter-gather pages, and that the
  1440. * packet has never been sent out before (and thus is not cloned).
  1441. */
  1442. static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
  1443. unsigned int mss_now, gfp_t gfp)
  1444. {
  1445. struct sk_buff *buff;
  1446. int nlen = skb->len - len;
  1447. u8 flags;
  1448. /* All of a TSO frame must be composed of paged data. */
  1449. if (skb->len != skb->data_len)
  1450. return tcp_fragment(sk, skb, len, mss_now);
  1451. buff = sk_stream_alloc_skb(sk, 0, gfp);
  1452. if (unlikely(buff == NULL))
  1453. return -ENOMEM;
  1454. sk->sk_wmem_queued += buff->truesize;
  1455. sk_mem_charge(sk, buff->truesize);
  1456. buff->truesize += nlen;
  1457. skb->truesize -= nlen;
  1458. /* Correct the sequence numbers. */
  1459. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  1460. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  1461. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  1462. /* PSH and FIN should only be set in the second packet. */
  1463. flags = TCP_SKB_CB(skb)->tcp_flags;
  1464. TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
  1465. TCP_SKB_CB(buff)->tcp_flags = flags;
  1466. /* This packet was never sent out yet, so no SACK bits. */
  1467. TCP_SKB_CB(buff)->sacked = 0;
  1468. buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
  1469. skb_split(skb, buff, len);
  1470. /* Fix up tso_factor for both original and new SKB. */
  1471. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1472. tcp_set_skb_tso_segs(sk, buff, mss_now);
  1473. /* Link BUFF into the send queue. */
  1474. skb_header_release(buff);
  1475. tcp_insert_write_queue_after(skb, buff, sk);
  1476. return 0;
  1477. }
  1478. /* Try to defer sending, if possible, in order to minimize the amount
  1479. * of TSO splitting we do. View it as a kind of TSO Nagle test.
  1480. *
  1481. * This algorithm is from John Heffner.
  1482. */
  1483. static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
  1484. {
  1485. struct tcp_sock *tp = tcp_sk(sk);
  1486. const struct inet_connection_sock *icsk = inet_csk(sk);
  1487. u32 send_win, cong_win, limit, in_flight;
  1488. int win_divisor;
  1489. if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
  1490. goto send_now;
  1491. if (icsk->icsk_ca_state != TCP_CA_Open)
  1492. goto send_now;
  1493. /* Defer for less than two clock ticks. */
  1494. if (tp->tso_deferred &&
  1495. (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
  1496. goto send_now;
  1497. in_flight = tcp_packets_in_flight(tp);
  1498. BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
  1499. send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  1500. /* From in_flight test above, we know that cwnd > in_flight. */
  1501. cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
  1502. limit = min(send_win, cong_win);
  1503. /* If a full-sized TSO skb can be sent, do it. */
  1504. if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
  1505. sk->sk_gso_max_segs * tp->mss_cache))
  1506. goto send_now;
  1507. /* Middle in queue won't get any more data, full sendable already? */
  1508. if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
  1509. goto send_now;
  1510. win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
  1511. if (win_divisor) {
  1512. u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
  1513. /* If at least some fraction of a window is available,
  1514. * just use it.
  1515. */
  1516. chunk /= win_divisor;
  1517. if (limit >= chunk)
  1518. goto send_now;
  1519. } else {
  1520. /* Different approach, try not to defer past a single
  1521. * ACK. Receiver should ACK every other full sized
  1522. * frame, so if we have space for more than 3 frames
  1523. * then send now.
  1524. */
  1525. if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
  1526. goto send_now;
  1527. }
  1528. /* Ok, it looks like it is advisable to defer. */
  1529. tp->tso_deferred = 1 | (jiffies << 1);
  1530. return true;
  1531. send_now:
  1532. tp->tso_deferred = 0;
  1533. return false;
  1534. }
  1535. /* Create a new MTU probe if we are ready.
  1536. * MTU probe is regularly attempting to increase the path MTU by
  1537. * deliberately sending larger packets. This discovers routing
  1538. * changes resulting in larger path MTUs.
  1539. *
  1540. * Returns 0 if we should wait to probe (no cwnd available),
  1541. * 1 if a probe was sent,
  1542. * -1 otherwise
  1543. */
  1544. static int tcp_mtu_probe(struct sock *sk)
  1545. {
  1546. struct tcp_sock *tp = tcp_sk(sk);
  1547. struct inet_connection_sock *icsk = inet_csk(sk);
  1548. struct sk_buff *skb, *nskb, *next;
  1549. int len;
  1550. int probe_size;
  1551. int size_needed;
  1552. int copy;
  1553. int mss_now;
  1554. /* Not currently probing/verifying,
  1555. * not in recovery,
  1556. * have enough cwnd, and
  1557. * not SACKing (the variable headers throw things off) */
  1558. if (!icsk->icsk_mtup.enabled ||
  1559. icsk->icsk_mtup.probe_size ||
  1560. inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
  1561. tp->snd_cwnd < 11 ||
  1562. tp->rx_opt.num_sacks || tp->rx_opt.dsack)
  1563. return -1;
  1564. /* Very simple search strategy: just double the MSS. */
  1565. mss_now = tcp_current_mss(sk);
  1566. probe_size = 2 * tp->mss_cache;
  1567. size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
  1568. if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
  1569. /* TODO: set timer for probe_converge_event */
  1570. return -1;
  1571. }
  1572. /* Have enough data in the send queue to probe? */
  1573. if (tp->write_seq - tp->snd_nxt < size_needed)
  1574. return -1;
  1575. if (tp->snd_wnd < size_needed)
  1576. return -1;
  1577. if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
  1578. return 0;
  1579. /* Do we need to wait to drain cwnd? With none in flight, don't stall */
  1580. if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
  1581. if (!tcp_packets_in_flight(tp))
  1582. return -1;
  1583. else
  1584. return 0;
  1585. }
  1586. /* We're allowed to probe. Build it now. */
  1587. if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
  1588. return -1;
  1589. sk->sk_wmem_queued += nskb->truesize;
  1590. sk_mem_charge(sk, nskb->truesize);
  1591. skb = tcp_send_head(sk);
  1592. TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
  1593. TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
  1594. TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
  1595. TCP_SKB_CB(nskb)->sacked = 0;
  1596. nskb->csum = 0;
  1597. nskb->ip_summed = skb->ip_summed;
  1598. tcp_insert_write_queue_before(nskb, skb, sk);
  1599. len = 0;
  1600. tcp_for_write_queue_from_safe(skb, next, sk) {
  1601. copy = min_t(int, skb->len, probe_size - len);
  1602. if (nskb->ip_summed)
  1603. skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
  1604. else
  1605. nskb->csum = skb_copy_and_csum_bits(skb, 0,
  1606. skb_put(nskb, copy),
  1607. copy, nskb->csum);
  1608. if (skb->len <= copy) {
  1609. /* We've eaten all the data from this skb.
  1610. * Throw it away. */
  1611. TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
  1612. tcp_unlink_write_queue(skb, sk);
  1613. sk_wmem_free_skb(sk, skb);
  1614. } else {
  1615. TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
  1616. ~(TCPHDR_FIN|TCPHDR_PSH);
  1617. if (!skb_shinfo(skb)->nr_frags) {
  1618. skb_pull(skb, copy);
  1619. if (skb->ip_summed != CHECKSUM_PARTIAL)
  1620. skb->csum = csum_partial(skb->data,
  1621. skb->len, 0);
  1622. } else {
  1623. __pskb_trim_head(skb, copy);
  1624. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1625. }
  1626. TCP_SKB_CB(skb)->seq += copy;
  1627. }
  1628. len += copy;
  1629. if (len >= probe_size)
  1630. break;
  1631. }
  1632. tcp_init_tso_segs(sk, nskb, nskb->len);
  1633. /* We're ready to send. If this fails, the probe will
  1634. * be resegmented into mss-sized pieces by tcp_write_xmit(). */
  1635. TCP_SKB_CB(nskb)->when = tcp_time_stamp;
  1636. if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
  1637. /* Decrement cwnd here because we are sending
  1638. * effectively two packets. */
  1639. tp->snd_cwnd--;
  1640. tcp_event_new_data_sent(sk, nskb);
  1641. icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
  1642. tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
  1643. tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
  1644. return 1;
  1645. }
  1646. return -1;
  1647. }
  1648. /* This routine writes packets to the network. It advances the
  1649. * send_head. This happens as incoming acks open up the remote
  1650. * window for us.
  1651. *
  1652. * LARGESEND note: !tcp_urg_mode is overkill, only frames between
  1653. * snd_up-64k-mss .. snd_up cannot be large. However, taking into
  1654. * account rare use of URG, this is not a big flaw.
  1655. *
  1656. * Returns true, if no segments are in flight and we have queued segments,
  1657. * but cannot send anything now because of SWS or another problem.
  1658. */
  1659. static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  1660. int push_one, gfp_t gfp)
  1661. {
  1662. struct tcp_sock *tp = tcp_sk(sk);
  1663. struct sk_buff *skb;
  1664. unsigned int tso_segs, sent_pkts;
  1665. int cwnd_quota;
  1666. int result;
  1667. sent_pkts = 0;
  1668. if (!push_one) {
  1669. /* Do MTU probing. */
  1670. result = tcp_mtu_probe(sk);
  1671. if (!result) {
  1672. return false;
  1673. } else if (result > 0) {
  1674. sent_pkts = 1;
  1675. }
  1676. }
  1677. while ((skb = tcp_send_head(sk))) {
  1678. unsigned int limit;
  1679. tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
  1680. BUG_ON(!tso_segs);
  1681. cwnd_quota = tcp_cwnd_test(tp, skb);
  1682. if (!cwnd_quota)
  1683. break;
  1684. if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
  1685. break;
  1686. if (tso_segs == 1) {
  1687. if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
  1688. (tcp_skb_is_last(sk, skb) ?
  1689. nonagle : TCP_NAGLE_PUSH))))
  1690. break;
  1691. } else {
  1692. if (!push_one && tcp_tso_should_defer(sk, skb))
  1693. break;
  1694. }
  1695. /* TSQ : sk_wmem_alloc accounts skb truesize,
  1696. * including skb overhead. But thats OK.
  1697. */
  1698. if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
  1699. set_bit(TSQ_THROTTLED, &tp->tsq_flags);
  1700. break;
  1701. }
  1702. limit = mss_now;
  1703. if (tso_segs > 1 && !tcp_urg_mode(tp))
  1704. limit = tcp_mss_split_point(sk, skb, mss_now,
  1705. min_t(unsigned int,
  1706. cwnd_quota,
  1707. sk->sk_gso_max_segs));
  1708. if (skb->len > limit &&
  1709. unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
  1710. break;
  1711. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1712. if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
  1713. break;
  1714. /* Advance the send_head. This one is sent out.
  1715. * This call will increment packets_out.
  1716. */
  1717. tcp_event_new_data_sent(sk, skb);
  1718. tcp_minshall_update(tp, mss_now, skb);
  1719. sent_pkts += tcp_skb_pcount(skb);
  1720. if (push_one)
  1721. break;
  1722. }
  1723. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
  1724. tp->prr_out += sent_pkts;
  1725. if (likely(sent_pkts)) {
  1726. tcp_cwnd_validate(sk);
  1727. return false;
  1728. }
  1729. return !tp->packets_out && tcp_send_head(sk);
  1730. }
  1731. /* Push out any pending frames which were held back due to
  1732. * TCP_CORK or attempt at coalescing tiny packets.
  1733. * The socket must be locked by the caller.
  1734. */
  1735. void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
  1736. int nonagle)
  1737. {
  1738. /* If we are closed, the bytes will have to remain here.
  1739. * In time closedown will finish, we empty the write queue and
  1740. * all will be happy.
  1741. */
  1742. if (unlikely(sk->sk_state == TCP_CLOSE))
  1743. return;
  1744. if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
  1745. sk_gfp_atomic(sk, GFP_ATOMIC)))
  1746. tcp_check_probe_timer(sk);
  1747. }
  1748. /* Send _single_ skb sitting at the send head. This function requires
  1749. * true push pending frames to setup probe timer etc.
  1750. */
  1751. void tcp_push_one(struct sock *sk, unsigned int mss_now)
  1752. {
  1753. struct sk_buff *skb = tcp_send_head(sk);
  1754. BUG_ON(!skb || skb->len < mss_now);
  1755. tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
  1756. }
  1757. /* This function returns the amount that we can raise the
  1758. * usable window based on the following constraints
  1759. *
  1760. * 1. The window can never be shrunk once it is offered (RFC 793)
  1761. * 2. We limit memory per socket
  1762. *
  1763. * RFC 1122:
  1764. * "the suggested [SWS] avoidance algorithm for the receiver is to keep
  1765. * RECV.NEXT + RCV.WIN fixed until:
  1766. * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
  1767. *
  1768. * i.e. don't raise the right edge of the window until you can raise
  1769. * it at least MSS bytes.
  1770. *
  1771. * Unfortunately, the recommended algorithm breaks header prediction,
  1772. * since header prediction assumes th->window stays fixed.
  1773. *
  1774. * Strictly speaking, keeping th->window fixed violates the receiver
  1775. * side SWS prevention criteria. The problem is that under this rule
  1776. * a stream of single byte packets will cause the right side of the
  1777. * window to always advance by a single byte.
  1778. *
  1779. * Of course, if the sender implements sender side SWS prevention
  1780. * then this will not be a problem.
  1781. *
  1782. * BSD seems to make the following compromise:
  1783. *
  1784. * If the free space is less than the 1/4 of the maximum
  1785. * space available and the free space is less than 1/2 mss,
  1786. * then set the window to 0.
  1787. * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
  1788. * Otherwise, just prevent the window from shrinking
  1789. * and from being larger than the largest representable value.
  1790. *
  1791. * This prevents incremental opening of the window in the regime
  1792. * where TCP is limited by the speed of the reader side taking
  1793. * data out of the TCP receive queue. It does nothing about
  1794. * those cases where the window is constrained on the sender side
  1795. * because the pipeline is full.
  1796. *
  1797. * BSD also seems to "accidentally" limit itself to windows that are a
  1798. * multiple of MSS, at least until the free space gets quite small.
  1799. * This would appear to be a side effect of the mbuf implementation.
  1800. * Combining these two algorithms results in the observed behavior
  1801. * of having a fixed window size at almost all times.
  1802. *
  1803. * Below we obtain similar behavior by forcing the offered window to
  1804. * a multiple of the mss when it is feasible to do so.
  1805. *
  1806. * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
  1807. * Regular options like TIMESTAMP are taken into account.
  1808. */
  1809. u32 __tcp_select_window(struct sock *sk)
  1810. {
  1811. struct inet_connection_sock *icsk = inet_csk(sk);
  1812. struct tcp_sock *tp = tcp_sk(sk);
  1813. /* MSS for the peer's data. Previous versions used mss_clamp
  1814. * here. I don't know if the value based on our guesses
  1815. * of peer's MSS is better for the performance. It's more correct
  1816. * but may be worse for the performance because of rcv_mss
  1817. * fluctuations. --SAW 1998/11/1
  1818. */
  1819. int mss = icsk->icsk_ack.rcv_mss;
  1820. int free_space = tcp_space(sk);
  1821. int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
  1822. int window;
  1823. if (mss > full_space)
  1824. mss = full_space;
  1825. if (free_space < (full_space >> 1)) {
  1826. icsk->icsk_ack.quick = 0;
  1827. if (sk_under_memory_pressure(sk))
  1828. tp->rcv_ssthresh = min(tp->rcv_ssthresh,
  1829. 4U * tp->advmss);
  1830. if (free_space < mss)
  1831. return 0;
  1832. }
  1833. if (free_space > tp->rcv_ssthresh)
  1834. free_space = tp->rcv_ssthresh;
  1835. /* Don't do rounding if we are using window scaling, since the
  1836. * scaled window will not line up with the MSS boundary anyway.
  1837. */
  1838. window = tp->rcv_wnd;
  1839. if (tp->rx_opt.rcv_wscale) {
  1840. window = free_space;
  1841. /* Advertise enough space so that it won't get scaled away.
  1842. * Import case: prevent zero window announcement if
  1843. * 1<<rcv_wscale > mss.
  1844. */
  1845. if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
  1846. window = (((window >> tp->rx_opt.rcv_wscale) + 1)
  1847. << tp->rx_opt.rcv_wscale);
  1848. } else {
  1849. /* Get the largest window that is a nice multiple of mss.
  1850. * Window clamp already applied above.
  1851. * If our current window offering is within 1 mss of the
  1852. * free space we just keep it. This prevents the divide
  1853. * and multiply from happening most of the time.
  1854. * We also don't do any window rounding when the free space
  1855. * is too small.
  1856. */
  1857. if (window <= free_space - mss || window > free_space)
  1858. window = (free_space / mss) * mss;
  1859. else if (mss == full_space &&
  1860. free_space > window + (full_space >> 1))
  1861. window = free_space;
  1862. }
  1863. return window;
  1864. }
  1865. /* Collapses two adjacent SKB's during retransmission. */
  1866. static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
  1867. {
  1868. struct tcp_sock *tp = tcp_sk(sk);
  1869. struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
  1870. int skb_size, next_skb_size;
  1871. skb_size = skb->len;
  1872. next_skb_size = next_skb->len;
  1873. BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
  1874. tcp_highest_sack_combine(sk, next_skb, skb);
  1875. tcp_unlink_write_queue(next_skb, sk);
  1876. skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
  1877. next_skb_size);
  1878. if (next_skb->ip_summed == CHECKSUM_PARTIAL)
  1879. skb->ip_summed = CHECKSUM_PARTIAL;
  1880. if (skb->ip_summed != CHECKSUM_PARTIAL)
  1881. skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
  1882. /* Update sequence range on original skb. */
  1883. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
  1884. /* Merge over control information. This moves PSH/FIN etc. over */
  1885. TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
  1886. /* All done, get rid of second SKB and account for it so
  1887. * packet counting does not break.
  1888. */
  1889. TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
  1890. /* changed transmit queue under us so clear hints */
  1891. tcp_clear_retrans_hints_partial(tp);
  1892. if (next_skb == tp->retransmit_skb_hint)
  1893. tp->retransmit_skb_hint = skb;
  1894. tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
  1895. sk_wmem_free_skb(sk, next_skb);
  1896. }
  1897. /* Check if coalescing SKBs is legal. */
  1898. static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
  1899. {
  1900. if (tcp_skb_pcount(skb) > 1)
  1901. return false;
  1902. /* TODO: SACK collapsing could be used to remove this condition */
  1903. if (skb_shinfo(skb)->nr_frags != 0)
  1904. return false;
  1905. if (skb_cloned(skb))
  1906. return false;
  1907. if (skb == tcp_send_head(sk))
  1908. return false;
  1909. /* Some heurestics for collapsing over SACK'd could be invented */
  1910. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
  1911. return false;
  1912. return true;
  1913. }
  1914. /* Collapse packets in the retransmit queue to make to create
  1915. * less packets on the wire. This is only done on retransmission.
  1916. */
  1917. static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
  1918. int space)
  1919. {
  1920. struct tcp_sock *tp = tcp_sk(sk);
  1921. struct sk_buff *skb = to, *tmp;
  1922. bool first = true;
  1923. if (!sysctl_tcp_retrans_collapse)
  1924. return;
  1925. if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
  1926. return;
  1927. tcp_for_write_queue_from_safe(skb, tmp, sk) {
  1928. if (!tcp_can_collapse(sk, skb))
  1929. break;
  1930. space -= skb->len;
  1931. if (first) {
  1932. first = false;
  1933. continue;
  1934. }
  1935. if (space < 0)
  1936. break;
  1937. /* Punt if not enough space exists in the first SKB for
  1938. * the data in the second
  1939. */
  1940. if (skb->len > skb_availroom(to))
  1941. break;
  1942. if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
  1943. break;
  1944. tcp_collapse_retrans(sk, to);
  1945. }
  1946. }
  1947. /* This retransmits one SKB. Policy decisions and retransmit queue
  1948. * state updates are done by the caller. Returns non-zero if an
  1949. * error occurred which prevented the send.
  1950. */
  1951. int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
  1952. {
  1953. struct tcp_sock *tp = tcp_sk(sk);
  1954. struct inet_connection_sock *icsk = inet_csk(sk);
  1955. unsigned int cur_mss;
  1956. int err;
  1957. /* Inconslusive MTU probe */
  1958. if (icsk->icsk_mtup.probe_size) {
  1959. icsk->icsk_mtup.probe_size = 0;
  1960. }
  1961. /* Do not sent more than we queued. 1/4 is reserved for possible
  1962. * copying overhead: fragmentation, tunneling, mangling etc.
  1963. */
  1964. if (atomic_read(&sk->sk_wmem_alloc) >
  1965. min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
  1966. return -EAGAIN;
  1967. if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
  1968. if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
  1969. BUG();
  1970. if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
  1971. return -ENOMEM;
  1972. }
  1973. if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
  1974. return -EHOSTUNREACH; /* Routing failure or similar. */
  1975. cur_mss = tcp_current_mss(sk);
  1976. /* If receiver has shrunk his window, and skb is out of
  1977. * new window, do not retransmit it. The exception is the
  1978. * case, when window is shrunk to zero. In this case
  1979. * our retransmit serves as a zero window probe.
  1980. */
  1981. if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
  1982. TCP_SKB_CB(skb)->seq != tp->snd_una)
  1983. return -EAGAIN;
  1984. if (skb->len > cur_mss) {
  1985. if (tcp_fragment(sk, skb, cur_mss, cur_mss))
  1986. return -ENOMEM; /* We'll try again later. */
  1987. } else {
  1988. int oldpcount = tcp_skb_pcount(skb);
  1989. if (unlikely(oldpcount > 1)) {
  1990. tcp_init_tso_segs(sk, skb, cur_mss);
  1991. tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
  1992. }
  1993. }
  1994. tcp_retrans_try_collapse(sk, skb, cur_mss);
  1995. /* Some Solaris stacks overoptimize and ignore the FIN on a
  1996. * retransmit when old data is attached. So strip it off
  1997. * since it is cheap to do so and saves bytes on the network.
  1998. */
  1999. if (skb->len > 0 &&
  2000. (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
  2001. tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
  2002. if (!pskb_trim(skb, 0)) {
  2003. /* Reuse, even though it does some unnecessary work */
  2004. tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
  2005. TCP_SKB_CB(skb)->tcp_flags);
  2006. skb->ip_summed = CHECKSUM_NONE;
  2007. }
  2008. }
  2009. /* Make a copy, if the first transmission SKB clone we made
  2010. * is still in somebody's hands, else make a clone.
  2011. */
  2012. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2013. /* make sure skb->data is aligned on arches that require it */
  2014. if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) {
  2015. struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
  2016. GFP_ATOMIC);
  2017. err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
  2018. -ENOBUFS;
  2019. } else {
  2020. err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2021. }
  2022. if (err == 0) {
  2023. /* Update global TCP statistics. */
  2024. TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
  2025. tp->total_retrans++;
  2026. #if FASTRETRANS_DEBUG > 0
  2027. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
  2028. net_dbg_ratelimited("retrans_out leaked\n");
  2029. }
  2030. #endif
  2031. if (!tp->retrans_out)
  2032. tp->lost_retrans_low = tp->snd_nxt;
  2033. TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
  2034. tp->retrans_out += tcp_skb_pcount(skb);
  2035. /* Save stamp of the first retransmit. */
  2036. if (!tp->retrans_stamp)
  2037. tp->retrans_stamp = TCP_SKB_CB(skb)->when;
  2038. tp->undo_retrans += tcp_skb_pcount(skb);
  2039. /* snd_nxt is stored to detect loss of retransmitted segment,
  2040. * see tcp_input.c tcp_sacktag_write_queue().
  2041. */
  2042. TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
  2043. }
  2044. return err;
  2045. }
  2046. /* Check if we forward retransmits are possible in the current
  2047. * window/congestion state.
  2048. */
  2049. static bool tcp_can_forward_retransmit(struct sock *sk)
  2050. {
  2051. const struct inet_connection_sock *icsk = inet_csk(sk);
  2052. const struct tcp_sock *tp = tcp_sk(sk);
  2053. /* Forward retransmissions are possible only during Recovery. */
  2054. if (icsk->icsk_ca_state != TCP_CA_Recovery)
  2055. return false;
  2056. /* No forward retransmissions in Reno are possible. */
  2057. if (tcp_is_reno(tp))
  2058. return false;
  2059. /* Yeah, we have to make difficult choice between forward transmission
  2060. * and retransmission... Both ways have their merits...
  2061. *
  2062. * For now we do not retransmit anything, while we have some new
  2063. * segments to send. In the other cases, follow rule 3 for
  2064. * NextSeg() specified in RFC3517.
  2065. */
  2066. if (tcp_may_send_now(sk))
  2067. return false;
  2068. return true;
  2069. }
  2070. /* This gets called after a retransmit timeout, and the initially
  2071. * retransmitted data is acknowledged. It tries to continue
  2072. * resending the rest of the retransmit queue, until either
  2073. * we've sent it all or the congestion window limit is reached.
  2074. * If doing SACK, the first ACK which comes back for a timeout
  2075. * based retransmit packet might feed us FACK information again.
  2076. * If so, we use it to avoid unnecessarily retransmissions.
  2077. */
  2078. void tcp_xmit_retransmit_queue(struct sock *sk)
  2079. {
  2080. const struct inet_connection_sock *icsk = inet_csk(sk);
  2081. struct tcp_sock *tp = tcp_sk(sk);
  2082. struct sk_buff *skb;
  2083. struct sk_buff *hole = NULL;
  2084. u32 last_lost;
  2085. int mib_idx;
  2086. int fwd_rexmitting = 0;
  2087. if (!tp->packets_out)
  2088. return;
  2089. if (!tp->lost_out)
  2090. tp->retransmit_high = tp->snd_una;
  2091. if (tp->retransmit_skb_hint) {
  2092. skb = tp->retransmit_skb_hint;
  2093. last_lost = TCP_SKB_CB(skb)->end_seq;
  2094. if (after(last_lost, tp->retransmit_high))
  2095. last_lost = tp->retransmit_high;
  2096. } else {
  2097. skb = tcp_write_queue_head(sk);
  2098. last_lost = tp->snd_una;
  2099. }
  2100. tcp_for_write_queue_from(skb, sk) {
  2101. __u8 sacked = TCP_SKB_CB(skb)->sacked;
  2102. if (skb == tcp_send_head(sk))
  2103. break;
  2104. /* we could do better than to assign each time */
  2105. if (hole == NULL)
  2106. tp->retransmit_skb_hint = skb;
  2107. /* Assume this retransmit will generate
  2108. * only one packet for congestion window
  2109. * calculation purposes. This works because
  2110. * tcp_retransmit_skb() will chop up the
  2111. * packet to be MSS sized and all the
  2112. * packet counting works out.
  2113. */
  2114. if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
  2115. return;
  2116. if (fwd_rexmitting) {
  2117. begin_fwd:
  2118. if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
  2119. break;
  2120. mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
  2121. } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
  2122. tp->retransmit_high = last_lost;
  2123. if (!tcp_can_forward_retransmit(sk))
  2124. break;
  2125. /* Backtrack if necessary to non-L'ed skb */
  2126. if (hole != NULL) {
  2127. skb = hole;
  2128. hole = NULL;
  2129. }
  2130. fwd_rexmitting = 1;
  2131. goto begin_fwd;
  2132. } else if (!(sacked & TCPCB_LOST)) {
  2133. if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
  2134. hole = skb;
  2135. continue;
  2136. } else {
  2137. last_lost = TCP_SKB_CB(skb)->end_seq;
  2138. if (icsk->icsk_ca_state != TCP_CA_Loss)
  2139. mib_idx = LINUX_MIB_TCPFASTRETRANS;
  2140. else
  2141. mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
  2142. }
  2143. if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
  2144. continue;
  2145. if (tcp_retransmit_skb(sk, skb)) {
  2146. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
  2147. return;
  2148. }
  2149. NET_INC_STATS_BH(sock_net(sk), mib_idx);
  2150. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
  2151. tp->prr_out += tcp_skb_pcount(skb);
  2152. if (skb == tcp_write_queue_head(sk))
  2153. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  2154. inet_csk(sk)->icsk_rto,
  2155. TCP_RTO_MAX);
  2156. }
  2157. }
  2158. /* Send a fin. The caller locks the socket for us. This cannot be
  2159. * allowed to fail queueing a FIN frame under any circumstances.
  2160. */
  2161. void tcp_send_fin(struct sock *sk)
  2162. {
  2163. struct tcp_sock *tp = tcp_sk(sk);
  2164. struct sk_buff *skb = tcp_write_queue_tail(sk);
  2165. int mss_now;
  2166. /* Optimization, tack on the FIN if we have a queue of
  2167. * unsent frames. But be careful about outgoing SACKS
  2168. * and IP options.
  2169. */
  2170. mss_now = tcp_current_mss(sk);
  2171. if (tcp_send_head(sk) != NULL) {
  2172. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
  2173. TCP_SKB_CB(skb)->end_seq++;
  2174. tp->write_seq++;
  2175. } else {
  2176. /* Socket is locked, keep trying until memory is available. */
  2177. for (;;) {
  2178. skb = alloc_skb_fclone(MAX_TCP_HEADER,
  2179. sk->sk_allocation);
  2180. if (skb)
  2181. break;
  2182. yield();
  2183. }
  2184. /* Reserve space for headers and prepare control bits. */
  2185. skb_reserve(skb, MAX_TCP_HEADER);
  2186. /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
  2187. tcp_init_nondata_skb(skb, tp->write_seq,
  2188. TCPHDR_ACK | TCPHDR_FIN);
  2189. tcp_queue_skb(sk, skb);
  2190. }
  2191. __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
  2192. }
  2193. /* We get here when a process closes a file descriptor (either due to
  2194. * an explicit close() or as a byproduct of exit()'ing) and there
  2195. * was unread data in the receive queue. This behavior is recommended
  2196. * by RFC 2525, section 2.17. -DaveM
  2197. */
  2198. void tcp_send_active_reset(struct sock *sk, gfp_t priority)
  2199. {
  2200. struct sk_buff *skb;
  2201. /* NOTE: No TCP options attached and we never retransmit this. */
  2202. skb = alloc_skb(MAX_TCP_HEADER, priority);
  2203. if (!skb) {
  2204. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
  2205. return;
  2206. }
  2207. /* Reserve space for headers and prepare control bits. */
  2208. skb_reserve(skb, MAX_TCP_HEADER);
  2209. tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
  2210. TCPHDR_ACK | TCPHDR_RST);
  2211. /* Send it off. */
  2212. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2213. if (tcp_transmit_skb(sk, skb, 0, priority))
  2214. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
  2215. TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
  2216. }
  2217. /* Send a crossed SYN-ACK during socket establishment.
  2218. * WARNING: This routine must only be called when we have already sent
  2219. * a SYN packet that crossed the incoming SYN that caused this routine
  2220. * to get called. If this assumption fails then the initial rcv_wnd
  2221. * and rcv_wscale values will not be correct.
  2222. */
  2223. int tcp_send_synack(struct sock *sk)
  2224. {
  2225. struct sk_buff *skb;
  2226. skb = tcp_write_queue_head(sk);
  2227. if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
  2228. pr_debug("%s: wrong queue state\n", __func__);
  2229. return -EFAULT;
  2230. }
  2231. if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
  2232. if (skb_cloned(skb)) {
  2233. struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
  2234. if (nskb == NULL)
  2235. return -ENOMEM;
  2236. tcp_unlink_write_queue(skb, sk);
  2237. skb_header_release(nskb);
  2238. __tcp_add_write_queue_head(sk, nskb);
  2239. sk_wmem_free_skb(sk, skb);
  2240. sk->sk_wmem_queued += nskb->truesize;
  2241. sk_mem_charge(sk, nskb->truesize);
  2242. skb = nskb;
  2243. }
  2244. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
  2245. TCP_ECN_send_synack(tcp_sk(sk), skb);
  2246. }
  2247. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2248. return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2249. }
  2250. /**
  2251. * tcp_make_synack - Prepare a SYN-ACK.
  2252. * sk: listener socket
  2253. * dst: dst entry attached to the SYNACK
  2254. * req: request_sock pointer
  2255. * rvp: request_values pointer
  2256. *
  2257. * Allocate one skb and build a SYNACK packet.
  2258. * @dst is consumed : Caller should not use it again.
  2259. */
  2260. struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
  2261. struct request_sock *req,
  2262. struct request_values *rvp)
  2263. {
  2264. struct tcp_out_options opts;
  2265. struct tcp_extend_values *xvp = tcp_xv(rvp);
  2266. struct inet_request_sock *ireq = inet_rsk(req);
  2267. struct tcp_sock *tp = tcp_sk(sk);
  2268. const struct tcp_cookie_values *cvp = tp->cookie_values;
  2269. struct tcphdr *th;
  2270. struct sk_buff *skb;
  2271. struct tcp_md5sig_key *md5;
  2272. int tcp_header_size;
  2273. int mss;
  2274. int s_data_desired = 0;
  2275. if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
  2276. s_data_desired = cvp->s_data_desired;
  2277. skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired,
  2278. sk_gfp_atomic(sk, GFP_ATOMIC));
  2279. if (unlikely(!skb)) {
  2280. dst_release(dst);
  2281. return NULL;
  2282. }
  2283. /* Reserve space for headers. */
  2284. skb_reserve(skb, MAX_TCP_HEADER);
  2285. skb_dst_set(skb, dst);
  2286. mss = dst_metric_advmss(dst);
  2287. if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
  2288. mss = tp->rx_opt.user_mss;
  2289. if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
  2290. __u8 rcv_wscale;
  2291. /* Set this up on the first call only */
  2292. req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
  2293. /* limit the window selection if the user enforce a smaller rx buffer */
  2294. if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
  2295. (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
  2296. req->window_clamp = tcp_full_space(sk);
  2297. /* tcp_full_space because it is guaranteed to be the first packet */
  2298. tcp_select_initial_window(tcp_full_space(sk),
  2299. mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
  2300. &req->rcv_wnd,
  2301. &req->window_clamp,
  2302. ireq->wscale_ok,
  2303. &rcv_wscale,
  2304. dst_metric(dst, RTAX_INITRWND));
  2305. ireq->rcv_wscale = rcv_wscale;
  2306. }
  2307. memset(&opts, 0, sizeof(opts));
  2308. #ifdef CONFIG_SYN_COOKIES
  2309. if (unlikely(req->cookie_ts))
  2310. TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
  2311. else
  2312. #endif
  2313. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2314. tcp_header_size = tcp_synack_options(sk, req, mss,
  2315. skb, &opts, &md5, xvp)
  2316. + sizeof(*th);
  2317. skb_push(skb, tcp_header_size);
  2318. skb_reset_transport_header(skb);
  2319. th = tcp_hdr(skb);
  2320. memset(th, 0, sizeof(struct tcphdr));
  2321. th->syn = 1;
  2322. th->ack = 1;
  2323. TCP_ECN_make_synack(req, th);
  2324. th->source = ireq->loc_port;
  2325. th->dest = ireq->rmt_port;
  2326. /* Setting of flags are superfluous here for callers (and ECE is
  2327. * not even correctly set)
  2328. */
  2329. tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
  2330. TCPHDR_SYN | TCPHDR_ACK);
  2331. if (OPTION_COOKIE_EXTENSION & opts.options) {
  2332. if (s_data_desired) {
  2333. u8 *buf = skb_put(skb, s_data_desired);
  2334. /* copy data directly from the listening socket. */
  2335. memcpy(buf, cvp->s_data_payload, s_data_desired);
  2336. TCP_SKB_CB(skb)->end_seq += s_data_desired;
  2337. }
  2338. if (opts.hash_size > 0) {
  2339. __u32 workspace[SHA_WORKSPACE_WORDS];
  2340. u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
  2341. u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
  2342. /* Secret recipe depends on the Timestamp, (future)
  2343. * Sequence and Acknowledgment Numbers, Initiator
  2344. * Cookie, and others handled by IP variant caller.
  2345. */
  2346. *tail-- ^= opts.tsval;
  2347. *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
  2348. *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
  2349. /* recommended */
  2350. *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
  2351. *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
  2352. sha_transform((__u32 *)&xvp->cookie_bakery[0],
  2353. (char *)mess,
  2354. &workspace[0]);
  2355. opts.hash_location =
  2356. (__u8 *)&xvp->cookie_bakery[0];
  2357. }
  2358. }
  2359. th->seq = htonl(TCP_SKB_CB(skb)->seq);
  2360. th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
  2361. /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
  2362. th->window = htons(min(req->rcv_wnd, 65535U));
  2363. tcp_options_write((__be32 *)(th + 1), tp, &opts);
  2364. th->doff = (tcp_header_size >> 2);
  2365. TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
  2366. #ifdef CONFIG_TCP_MD5SIG
  2367. /* Okay, we have all we need - do the md5 hash if needed */
  2368. if (md5) {
  2369. tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
  2370. md5, NULL, req, skb);
  2371. }
  2372. #endif
  2373. return skb;
  2374. }
  2375. EXPORT_SYMBOL(tcp_make_synack);
  2376. /* Do all connect socket setups that can be done AF independent. */
  2377. void tcp_connect_init(struct sock *sk)
  2378. {
  2379. const struct dst_entry *dst = __sk_dst_get(sk);
  2380. struct tcp_sock *tp = tcp_sk(sk);
  2381. __u8 rcv_wscale;
  2382. /* We'll fix this up when we get a response from the other end.
  2383. * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
  2384. */
  2385. tp->tcp_header_len = sizeof(struct tcphdr) +
  2386. (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
  2387. #ifdef CONFIG_TCP_MD5SIG
  2388. if (tp->af_specific->md5_lookup(sk, sk) != NULL)
  2389. tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
  2390. #endif
  2391. /* If user gave his TCP_MAXSEG, record it to clamp */
  2392. if (tp->rx_opt.user_mss)
  2393. tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
  2394. tp->max_window = 0;
  2395. tcp_mtup_init(sk);
  2396. tcp_sync_mss(sk, dst_mtu(dst));
  2397. if (!tp->window_clamp)
  2398. tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
  2399. tp->advmss = dst_metric_advmss(dst);
  2400. if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
  2401. tp->advmss = tp->rx_opt.user_mss;
  2402. tcp_initialize_rcv_mss(sk);
  2403. /* limit the window selection if the user enforce a smaller rx buffer */
  2404. if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
  2405. (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
  2406. tp->window_clamp = tcp_full_space(sk);
  2407. tcp_select_initial_window(tcp_full_space(sk),
  2408. tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
  2409. &tp->rcv_wnd,
  2410. &tp->window_clamp,
  2411. sysctl_tcp_window_scaling,
  2412. &rcv_wscale,
  2413. dst_metric(dst, RTAX_INITRWND));
  2414. tp->rx_opt.rcv_wscale = rcv_wscale;
  2415. tp->rcv_ssthresh = tp->rcv_wnd;
  2416. sk->sk_err = 0;
  2417. sock_reset_flag(sk, SOCK_DONE);
  2418. tp->snd_wnd = 0;
  2419. tcp_init_wl(tp, 0);
  2420. tp->snd_una = tp->write_seq;
  2421. tp->snd_sml = tp->write_seq;
  2422. tp->snd_up = tp->write_seq;
  2423. tp->snd_nxt = tp->write_seq;
  2424. if (likely(!tp->repair))
  2425. tp->rcv_nxt = 0;
  2426. tp->rcv_wup = tp->rcv_nxt;
  2427. tp->copied_seq = tp->rcv_nxt;
  2428. inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
  2429. inet_csk(sk)->icsk_retransmits = 0;
  2430. tcp_clear_retrans(tp);
  2431. }
  2432. static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
  2433. {
  2434. struct tcp_sock *tp = tcp_sk(sk);
  2435. struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
  2436. tcb->end_seq += skb->len;
  2437. skb_header_release(skb);
  2438. __tcp_add_write_queue_tail(sk, skb);
  2439. sk->sk_wmem_queued += skb->truesize;
  2440. sk_mem_charge(sk, skb->truesize);
  2441. tp->write_seq = tcb->end_seq;
  2442. tp->packets_out += tcp_skb_pcount(skb);
  2443. }
  2444. /* Build and send a SYN with data and (cached) Fast Open cookie. However,
  2445. * queue a data-only packet after the regular SYN, such that regular SYNs
  2446. * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
  2447. * only the SYN sequence, the data are retransmitted in the first ACK.
  2448. * If cookie is not cached or other error occurs, falls back to send a
  2449. * regular SYN with Fast Open cookie request option.
  2450. */
  2451. static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
  2452. {
  2453. struct tcp_sock *tp = tcp_sk(sk);
  2454. struct tcp_fastopen_request *fo = tp->fastopen_req;
  2455. int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
  2456. struct sk_buff *syn_data = NULL, *data;
  2457. unsigned long last_syn_loss = 0;
  2458. tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
  2459. tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
  2460. &syn_loss, &last_syn_loss);
  2461. /* Recurring FO SYN losses: revert to regular handshake temporarily */
  2462. if (syn_loss > 1 &&
  2463. time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
  2464. fo->cookie.len = -1;
  2465. goto fallback;
  2466. }
  2467. if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
  2468. fo->cookie.len = -1;
  2469. else if (fo->cookie.len <= 0)
  2470. goto fallback;
  2471. /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
  2472. * user-MSS. Reserve maximum option space for middleboxes that add
  2473. * private TCP options. The cost is reduced data space in SYN :(
  2474. */
  2475. if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
  2476. tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
  2477. space = tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
  2478. MAX_TCP_OPTION_SPACE;
  2479. syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
  2480. sk->sk_allocation);
  2481. if (syn_data == NULL)
  2482. goto fallback;
  2483. for (i = 0; i < iovlen && syn_data->len < space; ++i) {
  2484. struct iovec *iov = &fo->data->msg_iov[i];
  2485. unsigned char __user *from = iov->iov_base;
  2486. int len = iov->iov_len;
  2487. if (syn_data->len + len > space)
  2488. len = space - syn_data->len;
  2489. else if (i + 1 == iovlen)
  2490. /* No more data pending in inet_wait_for_connect() */
  2491. fo->data = NULL;
  2492. if (skb_add_data(syn_data, from, len))
  2493. goto fallback;
  2494. }
  2495. /* Queue a data-only packet after the regular SYN for retransmission */
  2496. data = pskb_copy(syn_data, sk->sk_allocation);
  2497. if (data == NULL)
  2498. goto fallback;
  2499. TCP_SKB_CB(data)->seq++;
  2500. TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
  2501. TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
  2502. tcp_connect_queue_skb(sk, data);
  2503. fo->copied = data->len;
  2504. if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
  2505. tp->syn_data = (fo->copied > 0);
  2506. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
  2507. goto done;
  2508. }
  2509. syn_data = NULL;
  2510. fallback:
  2511. /* Send a regular SYN with Fast Open cookie request option */
  2512. if (fo->cookie.len > 0)
  2513. fo->cookie.len = 0;
  2514. err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
  2515. if (err)
  2516. tp->syn_fastopen = 0;
  2517. kfree_skb(syn_data);
  2518. done:
  2519. fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
  2520. return err;
  2521. }
  2522. /* Build a SYN and send it off. */
  2523. int tcp_connect(struct sock *sk)
  2524. {
  2525. struct tcp_sock *tp = tcp_sk(sk);
  2526. struct sk_buff *buff;
  2527. int err;
  2528. tcp_connect_init(sk);
  2529. buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
  2530. if (unlikely(buff == NULL))
  2531. return -ENOBUFS;
  2532. /* Reserve space for headers. */
  2533. skb_reserve(buff, MAX_TCP_HEADER);
  2534. tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
  2535. tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
  2536. tcp_connect_queue_skb(sk, buff);
  2537. TCP_ECN_send_syn(sk, buff);
  2538. /* Send off SYN; include data in Fast Open. */
  2539. err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
  2540. tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
  2541. if (err == -ECONNREFUSED)
  2542. return err;
  2543. /* We change tp->snd_nxt after the tcp_transmit_skb() call
  2544. * in order to make this packet get counted in tcpOutSegs.
  2545. */
  2546. tp->snd_nxt = tp->write_seq;
  2547. tp->pushed_seq = tp->write_seq;
  2548. TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
  2549. /* Timer for repeating the SYN until an answer. */
  2550. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  2551. inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  2552. return 0;
  2553. }
  2554. EXPORT_SYMBOL(tcp_connect);
  2555. /* Send out a delayed ack, the caller does the policy checking
  2556. * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
  2557. * for details.
  2558. */
  2559. void tcp_send_delayed_ack(struct sock *sk)
  2560. {
  2561. struct inet_connection_sock *icsk = inet_csk(sk);
  2562. int ato = icsk->icsk_ack.ato;
  2563. unsigned long timeout;
  2564. if (ato > TCP_DELACK_MIN) {
  2565. const struct tcp_sock *tp = tcp_sk(sk);
  2566. int max_ato = HZ / 2;
  2567. if (icsk->icsk_ack.pingpong ||
  2568. (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
  2569. max_ato = TCP_DELACK_MAX;
  2570. /* Slow path, intersegment interval is "high". */
  2571. /* If some rtt estimate is known, use it to bound delayed ack.
  2572. * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
  2573. * directly.
  2574. */
  2575. if (tp->srtt) {
  2576. int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
  2577. if (rtt < max_ato)
  2578. max_ato = rtt;
  2579. }
  2580. ato = min(ato, max_ato);
  2581. }
  2582. /* Stay within the limit we were given */
  2583. timeout = jiffies + ato;
  2584. /* Use new timeout only if there wasn't a older one earlier. */
  2585. if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
  2586. /* If delack timer was blocked or is about to expire,
  2587. * send ACK now.
  2588. */
  2589. if (icsk->icsk_ack.blocked ||
  2590. time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
  2591. tcp_send_ack(sk);
  2592. return;
  2593. }
  2594. if (!time_before(timeout, icsk->icsk_ack.timeout))
  2595. timeout = icsk->icsk_ack.timeout;
  2596. }
  2597. icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
  2598. icsk->icsk_ack.timeout = timeout;
  2599. sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
  2600. }
  2601. /* This routine sends an ack and also updates the window. */
  2602. void tcp_send_ack(struct sock *sk)
  2603. {
  2604. struct sk_buff *buff;
  2605. /* If we have been reset, we may not send again. */
  2606. if (sk->sk_state == TCP_CLOSE)
  2607. return;
  2608. /* We are not putting this on the write queue, so
  2609. * tcp_transmit_skb() will set the ownership to this
  2610. * sock.
  2611. */
  2612. buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
  2613. if (buff == NULL) {
  2614. inet_csk_schedule_ack(sk);
  2615. inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
  2616. inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
  2617. TCP_DELACK_MAX, TCP_RTO_MAX);
  2618. return;
  2619. }
  2620. /* Reserve space for headers and prepare control bits. */
  2621. skb_reserve(buff, MAX_TCP_HEADER);
  2622. tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
  2623. /* Send it off, this clears delayed acks for us. */
  2624. TCP_SKB_CB(buff)->when = tcp_time_stamp;
  2625. tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
  2626. }
  2627. /* This routine sends a packet with an out of date sequence
  2628. * number. It assumes the other end will try to ack it.
  2629. *
  2630. * Question: what should we make while urgent mode?
  2631. * 4.4BSD forces sending single byte of data. We cannot send
  2632. * out of window data, because we have SND.NXT==SND.MAX...
  2633. *
  2634. * Current solution: to send TWO zero-length segments in urgent mode:
  2635. * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
  2636. * out-of-date with SND.UNA-1 to probe window.
  2637. */
  2638. static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
  2639. {
  2640. struct tcp_sock *tp = tcp_sk(sk);
  2641. struct sk_buff *skb;
  2642. /* We don't queue it, tcp_transmit_skb() sets ownership. */
  2643. skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
  2644. if (skb == NULL)
  2645. return -1;
  2646. /* Reserve space for headers and set control bits. */
  2647. skb_reserve(skb, MAX_TCP_HEADER);
  2648. /* Use a previous sequence. This should cause the other
  2649. * end to send an ack. Don't queue or clone SKB, just
  2650. * send it.
  2651. */
  2652. tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
  2653. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2654. return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
  2655. }
  2656. void tcp_send_window_probe(struct sock *sk)
  2657. {
  2658. if (sk->sk_state == TCP_ESTABLISHED) {
  2659. tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
  2660. tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
  2661. tcp_xmit_probe_skb(sk, 0);
  2662. }
  2663. }
  2664. /* Initiate keepalive or window probe from timer. */
  2665. int tcp_write_wakeup(struct sock *sk)
  2666. {
  2667. struct tcp_sock *tp = tcp_sk(sk);
  2668. struct sk_buff *skb;
  2669. if (sk->sk_state == TCP_CLOSE)
  2670. return -1;
  2671. if ((skb = tcp_send_head(sk)) != NULL &&
  2672. before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
  2673. int err;
  2674. unsigned int mss = tcp_current_mss(sk);
  2675. unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  2676. if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
  2677. tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
  2678. /* We are probing the opening of a window
  2679. * but the window size is != 0
  2680. * must have been a result SWS avoidance ( sender )
  2681. */
  2682. if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
  2683. skb->len > mss) {
  2684. seg_size = min(seg_size, mss);
  2685. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
  2686. if (tcp_fragment(sk, skb, seg_size, mss))
  2687. return -1;
  2688. } else if (!tcp_skb_pcount(skb))
  2689. tcp_set_skb_tso_segs(sk, skb, mss);
  2690. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
  2691. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2692. err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2693. if (!err)
  2694. tcp_event_new_data_sent(sk, skb);
  2695. return err;
  2696. } else {
  2697. if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
  2698. tcp_xmit_probe_skb(sk, 1);
  2699. return tcp_xmit_probe_skb(sk, 0);
  2700. }
  2701. }
  2702. /* A window probe timeout has occurred. If window is not closed send
  2703. * a partial packet else a zero probe.
  2704. */
  2705. void tcp_send_probe0(struct sock *sk)
  2706. {
  2707. struct inet_connection_sock *icsk = inet_csk(sk);
  2708. struct tcp_sock *tp = tcp_sk(sk);
  2709. int err;
  2710. err = tcp_write_wakeup(sk);
  2711. if (tp->packets_out || !tcp_send_head(sk)) {
  2712. /* Cancel probe timer, if it is not required. */
  2713. icsk->icsk_probes_out = 0;
  2714. icsk->icsk_backoff = 0;
  2715. return;
  2716. }
  2717. if (err <= 0) {
  2718. if (icsk->icsk_backoff < sysctl_tcp_retries2)
  2719. icsk->icsk_backoff++;
  2720. icsk->icsk_probes_out++;
  2721. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2722. min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
  2723. TCP_RTO_MAX);
  2724. } else {
  2725. /* If packet was not sent due to local congestion,
  2726. * do not backoff and do not remember icsk_probes_out.
  2727. * Let local senders to fight for local resources.
  2728. *
  2729. * Use accumulated backoff yet.
  2730. */
  2731. if (!icsk->icsk_probes_out)
  2732. icsk->icsk_probes_out = 1;
  2733. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2734. min(icsk->icsk_rto << icsk->icsk_backoff,
  2735. TCP_RESOURCE_PROBE_INTERVAL),
  2736. TCP_RTO_MAX);
  2737. }
  2738. }