tcp_output.c 92 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612261326142615261626172618261926202621262226232624262526262627262826292630263126322633263426352636263726382639264026412642264326442645264626472648264926502651265226532654265526562657265826592660266126622663266426652666266726682669267026712672267326742675267626772678267926802681268226832684268526862687268826892690269126922693269426952696269726982699270027012702270327042705270627072708270927102711271227132714271527162717271827192720272127222723272427252726272727282729273027312732273327342735273627372738273927402741274227432744274527462747274827492750275127522753275427552756275727582759276027612762276327642765276627672768276927702771277227732774277527762777277827792780278127822783278427852786278727882789279027912792279327942795279627972798279928002801280228032804280528062807280828092810281128122813281428152816281728182819282028212822282328242825282628272828282928302831283228332834283528362837283828392840284128422843284428452846284728482849285028512852285328542855285628572858285928602861286228632864286528662867286828692870287128722873287428752876287728782879288028812882288328842885288628872888288928902891289228932894289528962897289828992900290129022903290429052906290729082909291029112912291329142915291629172918291929202921292229232924292529262927292829292930293129322933293429352936293729382939294029412942294329442945294629472948294929502951295229532954295529562957295829592960296129622963296429652966296729682969297029712972297329742975297629772978297929802981298229832984298529862987298829892990299129922993299429952996299729982999300030013002300330043005300630073008300930103011301230133014301530163017301830193020302130223023302430253026302730283029303030313032303330343035303630373038303930403041304230433044304530463047304830493050305130523053305430553056305730583059306030613062306330643065306630673068306930703071307230733074307530763077307830793080308130823083308430853086308730883089309030913092309330943095309630973098309931003101310231033104310531063107310831093110311131123113311431153116311731183119312031213122312331243125312631273128312931303131313231333134313531363137313831393140314131423143314431453146314731483149315031513152315331543155315631573158315931603161316231633164316531663167316831693170317131723173317431753176317731783179318031813182318331843185318631873188318931903191319231933194319531963197
  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. unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX;
  58. EXPORT_SYMBOL(sysctl_tcp_notsent_lowat);
  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 inet_connection_sock *icsk = inet_csk(sk);
  65. struct tcp_sock *tp = tcp_sk(sk);
  66. unsigned int prior_packets = tp->packets_out;
  67. tcp_advance_send_head(sk, skb);
  68. tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
  69. tp->packets_out += tcp_skb_pcount(skb);
  70. if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
  71. icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
  72. tcp_rearm_rto(sk);
  73. }
  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. const struct dst_entry *dst = __sk_dst_get(sk);
  141. if (sysctl_tcp_slow_start_after_idle &&
  142. (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
  143. tcp_cwnd_restart(sk, __sk_dst_get(sk));
  144. tp->lsndtime = now;
  145. /* If it is a reply for ato after last received
  146. * packet, enter pingpong mode.
  147. */
  148. if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
  149. (!dst || !dst_metric(dst, RTAX_QUICKACK)))
  150. icsk->icsk_ack.pingpong = 1;
  151. }
  152. /* Account for an ACK we sent. */
  153. static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
  154. {
  155. tcp_dec_quickack_mode(sk, pkts);
  156. inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
  157. }
  158. u32 tcp_default_init_rwnd(u32 mss)
  159. {
  160. /* Initial receive window should be twice of TCP_INIT_CWND to
  161. * enable proper sending of new unsent data during fast recovery
  162. * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
  163. * limit when mss is larger than 1460.
  164. */
  165. u32 init_rwnd = TCP_INIT_CWND * 2;
  166. if (mss > 1460)
  167. init_rwnd = max((1460 * init_rwnd) / mss, 2U);
  168. return init_rwnd;
  169. }
  170. /* Determine a window scaling and initial window to offer.
  171. * Based on the assumption that the given amount of space
  172. * will be offered. Store the results in the tp structure.
  173. * NOTE: for smooth operation initial space offering should
  174. * be a multiple of mss if possible. We assume here that mss >= 1.
  175. * This MUST be enforced by all callers.
  176. */
  177. void tcp_select_initial_window(int __space, __u32 mss,
  178. __u32 *rcv_wnd, __u32 *window_clamp,
  179. int wscale_ok, __u8 *rcv_wscale,
  180. __u32 init_rcv_wnd)
  181. {
  182. unsigned int space = (__space < 0 ? 0 : __space);
  183. /* If no clamp set the clamp to the max possible scaled window */
  184. if (*window_clamp == 0)
  185. (*window_clamp) = (65535 << 14);
  186. space = min(*window_clamp, space);
  187. /* Quantize space offering to a multiple of mss if possible. */
  188. if (space > mss)
  189. space = (space / mss) * mss;
  190. /* NOTE: offering an initial window larger than 32767
  191. * will break some buggy TCP stacks. If the admin tells us
  192. * it is likely we could be speaking with such a buggy stack
  193. * we will truncate our initial window offering to 32K-1
  194. * unless the remote has sent us a window scaling option,
  195. * which we interpret as a sign the remote TCP is not
  196. * misinterpreting the window field as a signed quantity.
  197. */
  198. if (sysctl_tcp_workaround_signed_windows)
  199. (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
  200. else
  201. (*rcv_wnd) = space;
  202. (*rcv_wscale) = 0;
  203. if (wscale_ok) {
  204. /* Set window scaling on max possible window
  205. * See RFC1323 for an explanation of the limit to 14
  206. */
  207. space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
  208. space = min_t(u32, space, *window_clamp);
  209. while (space > 65535 && (*rcv_wscale) < 14) {
  210. space >>= 1;
  211. (*rcv_wscale)++;
  212. }
  213. }
  214. if (mss > (1 << *rcv_wscale)) {
  215. if (!init_rcv_wnd) /* Use default unless specified otherwise */
  216. init_rcv_wnd = tcp_default_init_rwnd(mss);
  217. *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
  218. }
  219. /* Set the clamp no higher than max representable value */
  220. (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
  221. }
  222. EXPORT_SYMBOL(tcp_select_initial_window);
  223. /* Chose a new window to advertise, update state in tcp_sock for the
  224. * socket, and return result with RFC1323 scaling applied. The return
  225. * value can be stuffed directly into th->window for an outgoing
  226. * frame.
  227. */
  228. static u16 tcp_select_window(struct sock *sk)
  229. {
  230. struct tcp_sock *tp = tcp_sk(sk);
  231. u32 cur_win = tcp_receive_window(tp);
  232. u32 new_win = __tcp_select_window(sk);
  233. /* Never shrink the offered window */
  234. if (new_win < cur_win) {
  235. /* Danger Will Robinson!
  236. * Don't update rcv_wup/rcv_wnd here or else
  237. * we will not be able to advertise a zero
  238. * window in time. --DaveM
  239. *
  240. * Relax Will Robinson.
  241. */
  242. new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
  243. }
  244. tp->rcv_wnd = new_win;
  245. tp->rcv_wup = tp->rcv_nxt;
  246. /* Make sure we do not exceed the maximum possible
  247. * scaled window.
  248. */
  249. if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
  250. new_win = min(new_win, MAX_TCP_WINDOW);
  251. else
  252. new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
  253. /* RFC1323 scaling applied */
  254. new_win >>= tp->rx_opt.rcv_wscale;
  255. /* If we advertise zero window, disable fast path. */
  256. if (new_win == 0)
  257. tp->pred_flags = 0;
  258. return new_win;
  259. }
  260. /* Packet ECN state for a SYN-ACK */
  261. static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
  262. {
  263. TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
  264. if (!(tp->ecn_flags & TCP_ECN_OK))
  265. TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
  266. }
  267. /* Packet ECN state for a SYN. */
  268. static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
  269. {
  270. struct tcp_sock *tp = tcp_sk(sk);
  271. tp->ecn_flags = 0;
  272. if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
  273. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
  274. tp->ecn_flags = TCP_ECN_OK;
  275. }
  276. }
  277. static __inline__ void
  278. TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
  279. {
  280. if (inet_rsk(req)->ecn_ok)
  281. th->ece = 1;
  282. }
  283. /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
  284. * be sent.
  285. */
  286. static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
  287. int tcp_header_len)
  288. {
  289. struct tcp_sock *tp = tcp_sk(sk);
  290. if (tp->ecn_flags & TCP_ECN_OK) {
  291. /* Not-retransmitted data segment: set ECT and inject CWR. */
  292. if (skb->len != tcp_header_len &&
  293. !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
  294. INET_ECN_xmit(sk);
  295. if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
  296. tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
  297. tcp_hdr(skb)->cwr = 1;
  298. skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
  299. }
  300. } else {
  301. /* ACK or retransmitted segment: clear ECT|CE */
  302. INET_ECN_dontxmit(sk);
  303. }
  304. if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
  305. tcp_hdr(skb)->ece = 1;
  306. }
  307. }
  308. /* Constructs common control bits of non-data skb. If SYN/FIN is present,
  309. * auto increment end seqno.
  310. */
  311. static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
  312. {
  313. skb->ip_summed = CHECKSUM_PARTIAL;
  314. skb->csum = 0;
  315. TCP_SKB_CB(skb)->tcp_flags = flags;
  316. TCP_SKB_CB(skb)->sacked = 0;
  317. skb_shinfo(skb)->gso_segs = 1;
  318. skb_shinfo(skb)->gso_size = 0;
  319. skb_shinfo(skb)->gso_type = 0;
  320. TCP_SKB_CB(skb)->seq = seq;
  321. if (flags & (TCPHDR_SYN | TCPHDR_FIN))
  322. seq++;
  323. TCP_SKB_CB(skb)->end_seq = seq;
  324. }
  325. static inline bool tcp_urg_mode(const struct tcp_sock *tp)
  326. {
  327. return tp->snd_una != tp->snd_up;
  328. }
  329. #define OPTION_SACK_ADVERTISE (1 << 0)
  330. #define OPTION_TS (1 << 1)
  331. #define OPTION_MD5 (1 << 2)
  332. #define OPTION_WSCALE (1 << 3)
  333. #define OPTION_FAST_OPEN_COOKIE (1 << 8)
  334. struct tcp_out_options {
  335. u16 options; /* bit field of OPTION_* */
  336. u16 mss; /* 0 to disable */
  337. u8 ws; /* window scale, 0 to disable */
  338. u8 num_sack_blocks; /* number of SACK blocks to include */
  339. u8 hash_size; /* bytes in hash_location */
  340. __u8 *hash_location; /* temporary pointer, overloaded */
  341. __u32 tsval, tsecr; /* need to include OPTION_TS */
  342. struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
  343. };
  344. /* Write previously computed TCP options to the packet.
  345. *
  346. * Beware: Something in the Internet is very sensitive to the ordering of
  347. * TCP options, we learned this through the hard way, so be careful here.
  348. * Luckily we can at least blame others for their non-compliance but from
  349. * inter-operatibility perspective it seems that we're somewhat stuck with
  350. * the ordering which we have been using if we want to keep working with
  351. * those broken things (not that it currently hurts anybody as there isn't
  352. * particular reason why the ordering would need to be changed).
  353. *
  354. * At least SACK_PERM as the first option is known to lead to a disaster
  355. * (but it may well be that other scenarios fail similarly).
  356. */
  357. static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
  358. struct tcp_out_options *opts)
  359. {
  360. u16 options = opts->options; /* mungable copy */
  361. if (unlikely(OPTION_MD5 & options)) {
  362. *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
  363. (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
  364. /* overload cookie hash location */
  365. opts->hash_location = (__u8 *)ptr;
  366. ptr += 4;
  367. }
  368. if (unlikely(opts->mss)) {
  369. *ptr++ = htonl((TCPOPT_MSS << 24) |
  370. (TCPOLEN_MSS << 16) |
  371. opts->mss);
  372. }
  373. if (likely(OPTION_TS & options)) {
  374. if (unlikely(OPTION_SACK_ADVERTISE & options)) {
  375. *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
  376. (TCPOLEN_SACK_PERM << 16) |
  377. (TCPOPT_TIMESTAMP << 8) |
  378. TCPOLEN_TIMESTAMP);
  379. options &= ~OPTION_SACK_ADVERTISE;
  380. } else {
  381. *ptr++ = htonl((TCPOPT_NOP << 24) |
  382. (TCPOPT_NOP << 16) |
  383. (TCPOPT_TIMESTAMP << 8) |
  384. TCPOLEN_TIMESTAMP);
  385. }
  386. *ptr++ = htonl(opts->tsval);
  387. *ptr++ = htonl(opts->tsecr);
  388. }
  389. if (unlikely(OPTION_SACK_ADVERTISE & options)) {
  390. *ptr++ = htonl((TCPOPT_NOP << 24) |
  391. (TCPOPT_NOP << 16) |
  392. (TCPOPT_SACK_PERM << 8) |
  393. TCPOLEN_SACK_PERM);
  394. }
  395. if (unlikely(OPTION_WSCALE & options)) {
  396. *ptr++ = htonl((TCPOPT_NOP << 24) |
  397. (TCPOPT_WINDOW << 16) |
  398. (TCPOLEN_WINDOW << 8) |
  399. opts->ws);
  400. }
  401. if (unlikely(opts->num_sack_blocks)) {
  402. struct tcp_sack_block *sp = tp->rx_opt.dsack ?
  403. tp->duplicate_sack : tp->selective_acks;
  404. int this_sack;
  405. *ptr++ = htonl((TCPOPT_NOP << 24) |
  406. (TCPOPT_NOP << 16) |
  407. (TCPOPT_SACK << 8) |
  408. (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
  409. TCPOLEN_SACK_PERBLOCK)));
  410. for (this_sack = 0; this_sack < opts->num_sack_blocks;
  411. ++this_sack) {
  412. *ptr++ = htonl(sp[this_sack].start_seq);
  413. *ptr++ = htonl(sp[this_sack].end_seq);
  414. }
  415. tp->rx_opt.dsack = 0;
  416. }
  417. if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
  418. struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
  419. *ptr++ = htonl((TCPOPT_EXP << 24) |
  420. ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
  421. TCPOPT_FASTOPEN_MAGIC);
  422. memcpy(ptr, foc->val, foc->len);
  423. if ((foc->len & 3) == 2) {
  424. u8 *align = ((u8 *)ptr) + foc->len;
  425. align[0] = align[1] = TCPOPT_NOP;
  426. }
  427. ptr += (foc->len + 3) >> 2;
  428. }
  429. }
  430. /* Compute TCP options for SYN packets. This is not the final
  431. * network wire format yet.
  432. */
  433. static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
  434. struct tcp_out_options *opts,
  435. struct tcp_md5sig_key **md5)
  436. {
  437. struct tcp_sock *tp = tcp_sk(sk);
  438. unsigned int remaining = MAX_TCP_OPTION_SPACE;
  439. struct tcp_fastopen_request *fastopen = tp->fastopen_req;
  440. #ifdef CONFIG_TCP_MD5SIG
  441. *md5 = tp->af_specific->md5_lookup(sk, sk);
  442. if (*md5) {
  443. opts->options |= OPTION_MD5;
  444. remaining -= TCPOLEN_MD5SIG_ALIGNED;
  445. }
  446. #else
  447. *md5 = NULL;
  448. #endif
  449. /* We always get an MSS option. The option bytes which will be seen in
  450. * normal data packets should timestamps be used, must be in the MSS
  451. * advertised. But we subtract them from tp->mss_cache so that
  452. * calculations in tcp_sendmsg are simpler etc. So account for this
  453. * fact here if necessary. If we don't do this correctly, as a
  454. * receiver we won't recognize data packets as being full sized when we
  455. * should, and thus we won't abide by the delayed ACK rules correctly.
  456. * SACKs don't matter, we never delay an ACK when we have any of those
  457. * going out. */
  458. opts->mss = tcp_advertise_mss(sk);
  459. remaining -= TCPOLEN_MSS_ALIGNED;
  460. if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
  461. opts->options |= OPTION_TS;
  462. opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
  463. opts->tsecr = tp->rx_opt.ts_recent;
  464. remaining -= TCPOLEN_TSTAMP_ALIGNED;
  465. }
  466. if (likely(sysctl_tcp_window_scaling)) {
  467. opts->ws = tp->rx_opt.rcv_wscale;
  468. opts->options |= OPTION_WSCALE;
  469. remaining -= TCPOLEN_WSCALE_ALIGNED;
  470. }
  471. if (likely(sysctl_tcp_sack)) {
  472. opts->options |= OPTION_SACK_ADVERTISE;
  473. if (unlikely(!(OPTION_TS & opts->options)))
  474. remaining -= TCPOLEN_SACKPERM_ALIGNED;
  475. }
  476. if (fastopen && fastopen->cookie.len >= 0) {
  477. u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
  478. need = (need + 3) & ~3U; /* Align to 32 bits */
  479. if (remaining >= need) {
  480. opts->options |= OPTION_FAST_OPEN_COOKIE;
  481. opts->fastopen_cookie = &fastopen->cookie;
  482. remaining -= need;
  483. tp->syn_fastopen = 1;
  484. }
  485. }
  486. return MAX_TCP_OPTION_SPACE - remaining;
  487. }
  488. /* Set up TCP options for SYN-ACKs. */
  489. static unsigned int tcp_synack_options(struct sock *sk,
  490. struct request_sock *req,
  491. unsigned int mss, struct sk_buff *skb,
  492. struct tcp_out_options *opts,
  493. struct tcp_md5sig_key **md5,
  494. struct tcp_fastopen_cookie *foc)
  495. {
  496. struct inet_request_sock *ireq = inet_rsk(req);
  497. unsigned int remaining = MAX_TCP_OPTION_SPACE;
  498. #ifdef CONFIG_TCP_MD5SIG
  499. *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
  500. if (*md5) {
  501. opts->options |= OPTION_MD5;
  502. remaining -= TCPOLEN_MD5SIG_ALIGNED;
  503. /* We can't fit any SACK blocks in a packet with MD5 + TS
  504. * options. There was discussion about disabling SACK
  505. * rather than TS in order to fit in better with old,
  506. * buggy kernels, but that was deemed to be unnecessary.
  507. */
  508. ireq->tstamp_ok &= !ireq->sack_ok;
  509. }
  510. #else
  511. *md5 = NULL;
  512. #endif
  513. /* We always send an MSS option. */
  514. opts->mss = mss;
  515. remaining -= TCPOLEN_MSS_ALIGNED;
  516. if (likely(ireq->wscale_ok)) {
  517. opts->ws = ireq->rcv_wscale;
  518. opts->options |= OPTION_WSCALE;
  519. remaining -= TCPOLEN_WSCALE_ALIGNED;
  520. }
  521. if (likely(ireq->tstamp_ok)) {
  522. opts->options |= OPTION_TS;
  523. opts->tsval = TCP_SKB_CB(skb)->when;
  524. opts->tsecr = req->ts_recent;
  525. remaining -= TCPOLEN_TSTAMP_ALIGNED;
  526. }
  527. if (likely(ireq->sack_ok)) {
  528. opts->options |= OPTION_SACK_ADVERTISE;
  529. if (unlikely(!ireq->tstamp_ok))
  530. remaining -= TCPOLEN_SACKPERM_ALIGNED;
  531. }
  532. if (foc != NULL) {
  533. u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
  534. need = (need + 3) & ~3U; /* Align to 32 bits */
  535. if (remaining >= need) {
  536. opts->options |= OPTION_FAST_OPEN_COOKIE;
  537. opts->fastopen_cookie = foc;
  538. remaining -= need;
  539. }
  540. }
  541. return MAX_TCP_OPTION_SPACE - remaining;
  542. }
  543. /* Compute TCP options for ESTABLISHED sockets. This is not the
  544. * final wire format yet.
  545. */
  546. static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
  547. struct tcp_out_options *opts,
  548. struct tcp_md5sig_key **md5)
  549. {
  550. struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
  551. struct tcp_sock *tp = tcp_sk(sk);
  552. unsigned int size = 0;
  553. unsigned int eff_sacks;
  554. opts->options = 0;
  555. #ifdef CONFIG_TCP_MD5SIG
  556. *md5 = tp->af_specific->md5_lookup(sk, sk);
  557. if (unlikely(*md5)) {
  558. opts->options |= OPTION_MD5;
  559. size += TCPOLEN_MD5SIG_ALIGNED;
  560. }
  561. #else
  562. *md5 = NULL;
  563. #endif
  564. if (likely(tp->rx_opt.tstamp_ok)) {
  565. opts->options |= OPTION_TS;
  566. opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
  567. opts->tsecr = tp->rx_opt.ts_recent;
  568. size += TCPOLEN_TSTAMP_ALIGNED;
  569. }
  570. eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
  571. if (unlikely(eff_sacks)) {
  572. const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
  573. opts->num_sack_blocks =
  574. min_t(unsigned int, eff_sacks,
  575. (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
  576. TCPOLEN_SACK_PERBLOCK);
  577. size += TCPOLEN_SACK_BASE_ALIGNED +
  578. opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
  579. }
  580. return size;
  581. }
  582. /* TCP SMALL QUEUES (TSQ)
  583. *
  584. * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
  585. * to reduce RTT and bufferbloat.
  586. * We do this using a special skb destructor (tcp_wfree).
  587. *
  588. * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
  589. * needs to be reallocated in a driver.
  590. * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
  591. *
  592. * Since transmit from skb destructor is forbidden, we use a tasklet
  593. * to process all sockets that eventually need to send more skbs.
  594. * We use one tasklet per cpu, with its own queue of sockets.
  595. */
  596. struct tsq_tasklet {
  597. struct tasklet_struct tasklet;
  598. struct list_head head; /* queue of tcp sockets */
  599. };
  600. static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
  601. static void tcp_tsq_handler(struct sock *sk)
  602. {
  603. if ((1 << sk->sk_state) &
  604. (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
  605. TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
  606. tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
  607. }
  608. /*
  609. * One tasklest per cpu tries to send more skbs.
  610. * We run in tasklet context but need to disable irqs when
  611. * transfering tsq->head because tcp_wfree() might
  612. * interrupt us (non NAPI drivers)
  613. */
  614. static void tcp_tasklet_func(unsigned long data)
  615. {
  616. struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
  617. LIST_HEAD(list);
  618. unsigned long flags;
  619. struct list_head *q, *n;
  620. struct tcp_sock *tp;
  621. struct sock *sk;
  622. local_irq_save(flags);
  623. list_splice_init(&tsq->head, &list);
  624. local_irq_restore(flags);
  625. list_for_each_safe(q, n, &list) {
  626. tp = list_entry(q, struct tcp_sock, tsq_node);
  627. list_del(&tp->tsq_node);
  628. sk = (struct sock *)tp;
  629. bh_lock_sock(sk);
  630. if (!sock_owned_by_user(sk)) {
  631. tcp_tsq_handler(sk);
  632. } else {
  633. /* defer the work to tcp_release_cb() */
  634. set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
  635. }
  636. bh_unlock_sock(sk);
  637. clear_bit(TSQ_QUEUED, &tp->tsq_flags);
  638. sk_free(sk);
  639. }
  640. }
  641. #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
  642. (1UL << TCP_WRITE_TIMER_DEFERRED) | \
  643. (1UL << TCP_DELACK_TIMER_DEFERRED) | \
  644. (1UL << TCP_MTU_REDUCED_DEFERRED))
  645. /**
  646. * tcp_release_cb - tcp release_sock() callback
  647. * @sk: socket
  648. *
  649. * called from release_sock() to perform protocol dependent
  650. * actions before socket release.
  651. */
  652. void tcp_release_cb(struct sock *sk)
  653. {
  654. struct tcp_sock *tp = tcp_sk(sk);
  655. unsigned long flags, nflags;
  656. /* perform an atomic operation only if at least one flag is set */
  657. do {
  658. flags = tp->tsq_flags;
  659. if (!(flags & TCP_DEFERRED_ALL))
  660. return;
  661. nflags = flags & ~TCP_DEFERRED_ALL;
  662. } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
  663. if (flags & (1UL << TCP_TSQ_DEFERRED))
  664. tcp_tsq_handler(sk);
  665. if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
  666. tcp_write_timer_handler(sk);
  667. __sock_put(sk);
  668. }
  669. if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
  670. tcp_delack_timer_handler(sk);
  671. __sock_put(sk);
  672. }
  673. if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
  674. sk->sk_prot->mtu_reduced(sk);
  675. __sock_put(sk);
  676. }
  677. }
  678. EXPORT_SYMBOL(tcp_release_cb);
  679. void __init tcp_tasklet_init(void)
  680. {
  681. int i;
  682. for_each_possible_cpu(i) {
  683. struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
  684. INIT_LIST_HEAD(&tsq->head);
  685. tasklet_init(&tsq->tasklet,
  686. tcp_tasklet_func,
  687. (unsigned long)tsq);
  688. }
  689. }
  690. /*
  691. * Write buffer destructor automatically called from kfree_skb.
  692. * We cant xmit new skbs from this context, as we might already
  693. * hold qdisc lock.
  694. */
  695. void tcp_wfree(struct sk_buff *skb)
  696. {
  697. struct sock *sk = skb->sk;
  698. struct tcp_sock *tp = tcp_sk(sk);
  699. if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
  700. !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
  701. unsigned long flags;
  702. struct tsq_tasklet *tsq;
  703. /* Keep a ref on socket.
  704. * This last ref will be released in tcp_tasklet_func()
  705. */
  706. atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
  707. /* queue this socket to tasklet queue */
  708. local_irq_save(flags);
  709. tsq = &__get_cpu_var(tsq_tasklet);
  710. list_add(&tp->tsq_node, &tsq->head);
  711. tasklet_schedule(&tsq->tasklet);
  712. local_irq_restore(flags);
  713. } else {
  714. sock_wfree(skb);
  715. }
  716. }
  717. /* This routine actually transmits TCP packets queued in by
  718. * tcp_do_sendmsg(). This is used by both the initial
  719. * transmission and possible later retransmissions.
  720. * All SKB's seen here are completely headerless. It is our
  721. * job to build the TCP header, and pass the packet down to
  722. * IP so it can do the same plus pass the packet off to the
  723. * device.
  724. *
  725. * We are working here with either a clone of the original
  726. * SKB, or a fresh unique copy made by the retransmit engine.
  727. */
  728. static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
  729. gfp_t gfp_mask)
  730. {
  731. const struct inet_connection_sock *icsk = inet_csk(sk);
  732. struct inet_sock *inet;
  733. struct tcp_sock *tp;
  734. struct tcp_skb_cb *tcb;
  735. struct tcp_out_options opts;
  736. unsigned int tcp_options_size, tcp_header_size;
  737. struct tcp_md5sig_key *md5;
  738. struct tcphdr *th;
  739. int err;
  740. BUG_ON(!skb || !tcp_skb_pcount(skb));
  741. if (clone_it) {
  742. const struct sk_buff *fclone = skb + 1;
  743. /* If congestion control is doing timestamping, we must
  744. * take such a timestamp before we potentially clone/copy.
  745. */
  746. if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
  747. __net_timestamp(skb);
  748. if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
  749. fclone->fclone == SKB_FCLONE_CLONE))
  750. NET_INC_STATS_BH(sock_net(sk),
  751. LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
  752. if (unlikely(skb_cloned(skb)))
  753. skb = pskb_copy(skb, gfp_mask);
  754. else
  755. skb = skb_clone(skb, gfp_mask);
  756. if (unlikely(!skb))
  757. return -ENOBUFS;
  758. }
  759. inet = inet_sk(sk);
  760. tp = tcp_sk(sk);
  761. tcb = TCP_SKB_CB(skb);
  762. memset(&opts, 0, sizeof(opts));
  763. if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
  764. tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
  765. else
  766. tcp_options_size = tcp_established_options(sk, skb, &opts,
  767. &md5);
  768. tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
  769. if (tcp_packets_in_flight(tp) == 0)
  770. tcp_ca_event(sk, CA_EVENT_TX_START);
  771. /* if no packet is in qdisc/device queue, then allow XPS to select
  772. * another queue.
  773. */
  774. skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
  775. skb_push(skb, tcp_header_size);
  776. skb_reset_transport_header(skb);
  777. skb_orphan(skb);
  778. skb->sk = sk;
  779. skb->destructor = tcp_wfree;
  780. atomic_add(skb->truesize, &sk->sk_wmem_alloc);
  781. /* Build TCP header and checksum it. */
  782. th = tcp_hdr(skb);
  783. th->source = inet->inet_sport;
  784. th->dest = inet->inet_dport;
  785. th->seq = htonl(tcb->seq);
  786. th->ack_seq = htonl(tp->rcv_nxt);
  787. *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
  788. tcb->tcp_flags);
  789. if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
  790. /* RFC1323: The window in SYN & SYN/ACK segments
  791. * is never scaled.
  792. */
  793. th->window = htons(min(tp->rcv_wnd, 65535U));
  794. } else {
  795. th->window = htons(tcp_select_window(sk));
  796. }
  797. th->check = 0;
  798. th->urg_ptr = 0;
  799. /* The urg_mode check is necessary during a below snd_una win probe */
  800. if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
  801. if (before(tp->snd_up, tcb->seq + 0x10000)) {
  802. th->urg_ptr = htons(tp->snd_up - tcb->seq);
  803. th->urg = 1;
  804. } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
  805. th->urg_ptr = htons(0xFFFF);
  806. th->urg = 1;
  807. }
  808. }
  809. tcp_options_write((__be32 *)(th + 1), tp, &opts);
  810. if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
  811. TCP_ECN_send(sk, skb, tcp_header_size);
  812. #ifdef CONFIG_TCP_MD5SIG
  813. /* Calculate the MD5 hash, as we have all we need now */
  814. if (md5) {
  815. sk_nocaps_add(sk, NETIF_F_GSO_MASK);
  816. tp->af_specific->calc_md5_hash(opts.hash_location,
  817. md5, sk, NULL, skb);
  818. }
  819. #endif
  820. icsk->icsk_af_ops->send_check(sk, skb);
  821. if (likely(tcb->tcp_flags & TCPHDR_ACK))
  822. tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
  823. if (skb->len != tcp_header_size)
  824. tcp_event_data_sent(tp, sk);
  825. if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
  826. TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
  827. tcp_skb_pcount(skb));
  828. err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
  829. if (likely(err <= 0))
  830. return err;
  831. tcp_enter_cwr(sk, 1);
  832. return net_xmit_eval(err);
  833. }
  834. /* This routine just queues the buffer for sending.
  835. *
  836. * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
  837. * otherwise socket can stall.
  838. */
  839. static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
  840. {
  841. struct tcp_sock *tp = tcp_sk(sk);
  842. /* Advance write_seq and place onto the write_queue. */
  843. tp->write_seq = TCP_SKB_CB(skb)->end_seq;
  844. skb_header_release(skb);
  845. tcp_add_write_queue_tail(sk, skb);
  846. sk->sk_wmem_queued += skb->truesize;
  847. sk_mem_charge(sk, skb->truesize);
  848. }
  849. /* Initialize TSO segments for a packet. */
  850. static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
  851. unsigned int mss_now)
  852. {
  853. if (skb->len <= mss_now || !sk_can_gso(sk) ||
  854. skb->ip_summed == CHECKSUM_NONE) {
  855. /* Avoid the costly divide in the normal
  856. * non-TSO case.
  857. */
  858. skb_shinfo(skb)->gso_segs = 1;
  859. skb_shinfo(skb)->gso_size = 0;
  860. skb_shinfo(skb)->gso_type = 0;
  861. } else {
  862. skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
  863. skb_shinfo(skb)->gso_size = mss_now;
  864. skb_shinfo(skb)->gso_type = sk->sk_gso_type;
  865. }
  866. }
  867. /* When a modification to fackets out becomes necessary, we need to check
  868. * skb is counted to fackets_out or not.
  869. */
  870. static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
  871. int decr)
  872. {
  873. struct tcp_sock *tp = tcp_sk(sk);
  874. if (!tp->sacked_out || tcp_is_reno(tp))
  875. return;
  876. if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
  877. tp->fackets_out -= decr;
  878. }
  879. /* Pcount in the middle of the write queue got changed, we need to do various
  880. * tweaks to fix counters
  881. */
  882. static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
  883. {
  884. struct tcp_sock *tp = tcp_sk(sk);
  885. tp->packets_out -= decr;
  886. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
  887. tp->sacked_out -= decr;
  888. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
  889. tp->retrans_out -= decr;
  890. if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
  891. tp->lost_out -= decr;
  892. /* Reno case is special. Sigh... */
  893. if (tcp_is_reno(tp) && decr > 0)
  894. tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
  895. tcp_adjust_fackets_out(sk, skb, decr);
  896. if (tp->lost_skb_hint &&
  897. before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
  898. (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
  899. tp->lost_cnt_hint -= decr;
  900. tcp_verify_left_out(tp);
  901. }
  902. /* Function to create two new TCP segments. Shrinks the given segment
  903. * to the specified size and appends a new segment with the rest of the
  904. * packet to the list. This won't be called frequently, I hope.
  905. * Remember, these are still headerless SKBs at this point.
  906. */
  907. int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
  908. unsigned int mss_now)
  909. {
  910. struct tcp_sock *tp = tcp_sk(sk);
  911. struct sk_buff *buff;
  912. int nsize, old_factor;
  913. int nlen;
  914. u8 flags;
  915. if (WARN_ON(len > skb->len))
  916. return -EINVAL;
  917. nsize = skb_headlen(skb) - len;
  918. if (nsize < 0)
  919. nsize = 0;
  920. if (skb_cloned(skb) &&
  921. skb_is_nonlinear(skb) &&
  922. pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
  923. return -ENOMEM;
  924. /* Get a new skb... force flag on. */
  925. buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
  926. if (buff == NULL)
  927. return -ENOMEM; /* We'll just try again later. */
  928. sk->sk_wmem_queued += buff->truesize;
  929. sk_mem_charge(sk, buff->truesize);
  930. nlen = skb->len - len - nsize;
  931. buff->truesize += nlen;
  932. skb->truesize -= nlen;
  933. /* Correct the sequence numbers. */
  934. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  935. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  936. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  937. /* PSH and FIN should only be set in the second packet. */
  938. flags = TCP_SKB_CB(skb)->tcp_flags;
  939. TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
  940. TCP_SKB_CB(buff)->tcp_flags = flags;
  941. TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
  942. if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
  943. /* Copy and checksum data tail into the new buffer. */
  944. buff->csum = csum_partial_copy_nocheck(skb->data + len,
  945. skb_put(buff, nsize),
  946. nsize, 0);
  947. skb_trim(skb, len);
  948. skb->csum = csum_block_sub(skb->csum, buff->csum, len);
  949. } else {
  950. skb->ip_summed = CHECKSUM_PARTIAL;
  951. skb_split(skb, buff, len);
  952. }
  953. buff->ip_summed = skb->ip_summed;
  954. /* Looks stupid, but our code really uses when of
  955. * skbs, which it never sent before. --ANK
  956. */
  957. TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
  958. buff->tstamp = skb->tstamp;
  959. old_factor = tcp_skb_pcount(skb);
  960. /* Fix up tso_factor for both original and new SKB. */
  961. tcp_set_skb_tso_segs(sk, skb, mss_now);
  962. tcp_set_skb_tso_segs(sk, buff, mss_now);
  963. /* If this packet has been sent out already, we must
  964. * adjust the various packet counters.
  965. */
  966. if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
  967. int diff = old_factor - tcp_skb_pcount(skb) -
  968. tcp_skb_pcount(buff);
  969. if (diff)
  970. tcp_adjust_pcount(sk, skb, diff);
  971. }
  972. /* Link BUFF into the send queue. */
  973. skb_header_release(buff);
  974. tcp_insert_write_queue_after(skb, buff, sk);
  975. return 0;
  976. }
  977. /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
  978. * eventually). The difference is that pulled data not copied, but
  979. * immediately discarded.
  980. */
  981. static void __pskb_trim_head(struct sk_buff *skb, int len)
  982. {
  983. int i, k, eat;
  984. eat = min_t(int, len, skb_headlen(skb));
  985. if (eat) {
  986. __skb_pull(skb, eat);
  987. len -= eat;
  988. if (!len)
  989. return;
  990. }
  991. eat = len;
  992. k = 0;
  993. for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  994. int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
  995. if (size <= eat) {
  996. skb_frag_unref(skb, i);
  997. eat -= size;
  998. } else {
  999. skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
  1000. if (eat) {
  1001. skb_shinfo(skb)->frags[k].page_offset += eat;
  1002. skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
  1003. eat = 0;
  1004. }
  1005. k++;
  1006. }
  1007. }
  1008. skb_shinfo(skb)->nr_frags = k;
  1009. skb_reset_tail_pointer(skb);
  1010. skb->data_len -= len;
  1011. skb->len = skb->data_len;
  1012. }
  1013. /* Remove acked data from a packet in the transmit queue. */
  1014. int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
  1015. {
  1016. if (skb_unclone(skb, GFP_ATOMIC))
  1017. return -ENOMEM;
  1018. __pskb_trim_head(skb, len);
  1019. TCP_SKB_CB(skb)->seq += len;
  1020. skb->ip_summed = CHECKSUM_PARTIAL;
  1021. skb->truesize -= len;
  1022. sk->sk_wmem_queued -= len;
  1023. sk_mem_uncharge(sk, len);
  1024. sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
  1025. /* Any change of skb->len requires recalculation of tso factor. */
  1026. if (tcp_skb_pcount(skb) > 1)
  1027. tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
  1028. return 0;
  1029. }
  1030. /* Calculate MSS not accounting any TCP options. */
  1031. static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
  1032. {
  1033. const struct tcp_sock *tp = tcp_sk(sk);
  1034. const struct inet_connection_sock *icsk = inet_csk(sk);
  1035. int mss_now;
  1036. /* Calculate base mss without TCP options:
  1037. It is MMS_S - sizeof(tcphdr) of rfc1122
  1038. */
  1039. mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
  1040. /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
  1041. if (icsk->icsk_af_ops->net_frag_header_len) {
  1042. const struct dst_entry *dst = __sk_dst_get(sk);
  1043. if (dst && dst_allfrag(dst))
  1044. mss_now -= icsk->icsk_af_ops->net_frag_header_len;
  1045. }
  1046. /* Clamp it (mss_clamp does not include tcp options) */
  1047. if (mss_now > tp->rx_opt.mss_clamp)
  1048. mss_now = tp->rx_opt.mss_clamp;
  1049. /* Now subtract optional transport overhead */
  1050. mss_now -= icsk->icsk_ext_hdr_len;
  1051. /* Then reserve room for full set of TCP options and 8 bytes of data */
  1052. if (mss_now < 48)
  1053. mss_now = 48;
  1054. return mss_now;
  1055. }
  1056. /* Calculate MSS. Not accounting for SACKs here. */
  1057. int tcp_mtu_to_mss(struct sock *sk, int pmtu)
  1058. {
  1059. /* Subtract TCP options size, not including SACKs */
  1060. return __tcp_mtu_to_mss(sk, pmtu) -
  1061. (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
  1062. }
  1063. /* Inverse of above */
  1064. int tcp_mss_to_mtu(struct sock *sk, int mss)
  1065. {
  1066. const struct tcp_sock *tp = tcp_sk(sk);
  1067. const struct inet_connection_sock *icsk = inet_csk(sk);
  1068. int mtu;
  1069. mtu = mss +
  1070. tp->tcp_header_len +
  1071. icsk->icsk_ext_hdr_len +
  1072. icsk->icsk_af_ops->net_header_len;
  1073. /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
  1074. if (icsk->icsk_af_ops->net_frag_header_len) {
  1075. const struct dst_entry *dst = __sk_dst_get(sk);
  1076. if (dst && dst_allfrag(dst))
  1077. mtu += icsk->icsk_af_ops->net_frag_header_len;
  1078. }
  1079. return mtu;
  1080. }
  1081. /* MTU probing init per socket */
  1082. void tcp_mtup_init(struct sock *sk)
  1083. {
  1084. struct tcp_sock *tp = tcp_sk(sk);
  1085. struct inet_connection_sock *icsk = inet_csk(sk);
  1086. icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
  1087. icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
  1088. icsk->icsk_af_ops->net_header_len;
  1089. icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
  1090. icsk->icsk_mtup.probe_size = 0;
  1091. }
  1092. EXPORT_SYMBOL(tcp_mtup_init);
  1093. /* This function synchronize snd mss to current pmtu/exthdr set.
  1094. tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
  1095. for TCP options, but includes only bare TCP header.
  1096. tp->rx_opt.mss_clamp is mss negotiated at connection setup.
  1097. It is minimum of user_mss and mss received with SYN.
  1098. It also does not include TCP options.
  1099. inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
  1100. tp->mss_cache is current effective sending mss, including
  1101. all tcp options except for SACKs. It is evaluated,
  1102. taking into account current pmtu, but never exceeds
  1103. tp->rx_opt.mss_clamp.
  1104. NOTE1. rfc1122 clearly states that advertised MSS
  1105. DOES NOT include either tcp or ip options.
  1106. NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
  1107. are READ ONLY outside this function. --ANK (980731)
  1108. */
  1109. unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
  1110. {
  1111. struct tcp_sock *tp = tcp_sk(sk);
  1112. struct inet_connection_sock *icsk = inet_csk(sk);
  1113. int mss_now;
  1114. if (icsk->icsk_mtup.search_high > pmtu)
  1115. icsk->icsk_mtup.search_high = pmtu;
  1116. mss_now = tcp_mtu_to_mss(sk, pmtu);
  1117. mss_now = tcp_bound_to_half_wnd(tp, mss_now);
  1118. /* And store cached results */
  1119. icsk->icsk_pmtu_cookie = pmtu;
  1120. if (icsk->icsk_mtup.enabled)
  1121. mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
  1122. tp->mss_cache = mss_now;
  1123. return mss_now;
  1124. }
  1125. EXPORT_SYMBOL(tcp_sync_mss);
  1126. /* Compute the current effective MSS, taking SACKs and IP options,
  1127. * and even PMTU discovery events into account.
  1128. */
  1129. unsigned int tcp_current_mss(struct sock *sk)
  1130. {
  1131. const struct tcp_sock *tp = tcp_sk(sk);
  1132. const struct dst_entry *dst = __sk_dst_get(sk);
  1133. u32 mss_now;
  1134. unsigned int header_len;
  1135. struct tcp_out_options opts;
  1136. struct tcp_md5sig_key *md5;
  1137. mss_now = tp->mss_cache;
  1138. if (dst) {
  1139. u32 mtu = dst_mtu(dst);
  1140. if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
  1141. mss_now = tcp_sync_mss(sk, mtu);
  1142. }
  1143. header_len = tcp_established_options(sk, NULL, &opts, &md5) +
  1144. sizeof(struct tcphdr);
  1145. /* The mss_cache is sized based on tp->tcp_header_len, which assumes
  1146. * some common options. If this is an odd packet (because we have SACK
  1147. * blocks etc) then our calculated header_len will be different, and
  1148. * we have to adjust mss_now correspondingly */
  1149. if (header_len != tp->tcp_header_len) {
  1150. int delta = (int) header_len - tp->tcp_header_len;
  1151. mss_now -= delta;
  1152. }
  1153. return mss_now;
  1154. }
  1155. /* Congestion window validation. (RFC2861) */
  1156. static void tcp_cwnd_validate(struct sock *sk)
  1157. {
  1158. struct tcp_sock *tp = tcp_sk(sk);
  1159. if (tp->packets_out >= tp->snd_cwnd) {
  1160. /* Network is feed fully. */
  1161. tp->snd_cwnd_used = 0;
  1162. tp->snd_cwnd_stamp = tcp_time_stamp;
  1163. } else {
  1164. /* Network starves. */
  1165. if (tp->packets_out > tp->snd_cwnd_used)
  1166. tp->snd_cwnd_used = tp->packets_out;
  1167. if (sysctl_tcp_slow_start_after_idle &&
  1168. (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
  1169. tcp_cwnd_application_limited(sk);
  1170. }
  1171. }
  1172. /* Returns the portion of skb which can be sent right away without
  1173. * introducing MSS oddities to segment boundaries. In rare cases where
  1174. * mss_now != mss_cache, we will request caller to create a small skb
  1175. * per input skb which could be mostly avoided here (if desired).
  1176. *
  1177. * We explicitly want to create a request for splitting write queue tail
  1178. * to a small skb for Nagle purposes while avoiding unnecessary modulos,
  1179. * thus all the complexity (cwnd_len is always MSS multiple which we
  1180. * return whenever allowed by the other factors). Basically we need the
  1181. * modulo only when the receiver window alone is the limiting factor or
  1182. * when we would be allowed to send the split-due-to-Nagle skb fully.
  1183. */
  1184. static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
  1185. unsigned int mss_now, unsigned int max_segs)
  1186. {
  1187. const struct tcp_sock *tp = tcp_sk(sk);
  1188. u32 needed, window, max_len;
  1189. window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  1190. max_len = mss_now * max_segs;
  1191. if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
  1192. return max_len;
  1193. needed = min(skb->len, window);
  1194. if (max_len <= needed)
  1195. return max_len;
  1196. return needed - needed % mss_now;
  1197. }
  1198. /* Can at least one segment of SKB be sent right now, according to the
  1199. * congestion window rules? If so, return how many segments are allowed.
  1200. */
  1201. static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
  1202. const struct sk_buff *skb)
  1203. {
  1204. u32 in_flight, cwnd;
  1205. /* Don't be strict about the congestion window for the final FIN. */
  1206. if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
  1207. tcp_skb_pcount(skb) == 1)
  1208. return 1;
  1209. in_flight = tcp_packets_in_flight(tp);
  1210. cwnd = tp->snd_cwnd;
  1211. if (in_flight < cwnd)
  1212. return (cwnd - in_flight);
  1213. return 0;
  1214. }
  1215. /* Initialize TSO state of a skb.
  1216. * This must be invoked the first time we consider transmitting
  1217. * SKB onto the wire.
  1218. */
  1219. static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
  1220. unsigned int mss_now)
  1221. {
  1222. int tso_segs = tcp_skb_pcount(skb);
  1223. if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
  1224. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1225. tso_segs = tcp_skb_pcount(skb);
  1226. }
  1227. return tso_segs;
  1228. }
  1229. /* Minshall's variant of the Nagle send check. */
  1230. static inline bool tcp_minshall_check(const struct tcp_sock *tp)
  1231. {
  1232. return after(tp->snd_sml, tp->snd_una) &&
  1233. !after(tp->snd_sml, tp->snd_nxt);
  1234. }
  1235. /* Return false, if packet can be sent now without violation Nagle's rules:
  1236. * 1. It is full sized.
  1237. * 2. Or it contains FIN. (already checked by caller)
  1238. * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
  1239. * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
  1240. * With Minshall's modification: all sent small packets are ACKed.
  1241. */
  1242. static inline bool tcp_nagle_check(const struct tcp_sock *tp,
  1243. const struct sk_buff *skb,
  1244. unsigned int mss_now, int nonagle)
  1245. {
  1246. return skb->len < mss_now &&
  1247. ((nonagle & TCP_NAGLE_CORK) ||
  1248. (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
  1249. }
  1250. /* Return true if the Nagle test allows this packet to be
  1251. * sent now.
  1252. */
  1253. static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
  1254. unsigned int cur_mss, int nonagle)
  1255. {
  1256. /* Nagle rule does not apply to frames, which sit in the middle of the
  1257. * write_queue (they have no chances to get new data).
  1258. *
  1259. * This is implemented in the callers, where they modify the 'nonagle'
  1260. * argument based upon the location of SKB in the send queue.
  1261. */
  1262. if (nonagle & TCP_NAGLE_PUSH)
  1263. return true;
  1264. /* Don't use the nagle rule for urgent data (or for the final FIN). */
  1265. if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
  1266. return true;
  1267. if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
  1268. return true;
  1269. return false;
  1270. }
  1271. /* Does at least the first segment of SKB fit into the send window? */
  1272. static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
  1273. const struct sk_buff *skb,
  1274. unsigned int cur_mss)
  1275. {
  1276. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  1277. if (skb->len > cur_mss)
  1278. end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
  1279. return !after(end_seq, tcp_wnd_end(tp));
  1280. }
  1281. /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
  1282. * should be put on the wire right now. If so, it returns the number of
  1283. * packets allowed by the congestion window.
  1284. */
  1285. static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
  1286. unsigned int cur_mss, int nonagle)
  1287. {
  1288. const struct tcp_sock *tp = tcp_sk(sk);
  1289. unsigned int cwnd_quota;
  1290. tcp_init_tso_segs(sk, skb, cur_mss);
  1291. if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
  1292. return 0;
  1293. cwnd_quota = tcp_cwnd_test(tp, skb);
  1294. if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
  1295. cwnd_quota = 0;
  1296. return cwnd_quota;
  1297. }
  1298. /* Test if sending is allowed right now. */
  1299. bool tcp_may_send_now(struct sock *sk)
  1300. {
  1301. const struct tcp_sock *tp = tcp_sk(sk);
  1302. struct sk_buff *skb = tcp_send_head(sk);
  1303. return skb &&
  1304. tcp_snd_test(sk, skb, tcp_current_mss(sk),
  1305. (tcp_skb_is_last(sk, skb) ?
  1306. tp->nonagle : TCP_NAGLE_PUSH));
  1307. }
  1308. /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
  1309. * which is put after SKB on the list. It is very much like
  1310. * tcp_fragment() except that it may make several kinds of assumptions
  1311. * in order to speed up the splitting operation. In particular, we
  1312. * know that all the data is in scatter-gather pages, and that the
  1313. * packet has never been sent out before (and thus is not cloned).
  1314. */
  1315. static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
  1316. unsigned int mss_now, gfp_t gfp)
  1317. {
  1318. struct sk_buff *buff;
  1319. int nlen = skb->len - len;
  1320. u8 flags;
  1321. /* All of a TSO frame must be composed of paged data. */
  1322. if (skb->len != skb->data_len)
  1323. return tcp_fragment(sk, skb, len, mss_now);
  1324. buff = sk_stream_alloc_skb(sk, 0, gfp);
  1325. if (unlikely(buff == NULL))
  1326. return -ENOMEM;
  1327. sk->sk_wmem_queued += buff->truesize;
  1328. sk_mem_charge(sk, buff->truesize);
  1329. buff->truesize += nlen;
  1330. skb->truesize -= nlen;
  1331. /* Correct the sequence numbers. */
  1332. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  1333. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  1334. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  1335. /* PSH and FIN should only be set in the second packet. */
  1336. flags = TCP_SKB_CB(skb)->tcp_flags;
  1337. TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
  1338. TCP_SKB_CB(buff)->tcp_flags = flags;
  1339. /* This packet was never sent out yet, so no SACK bits. */
  1340. TCP_SKB_CB(buff)->sacked = 0;
  1341. buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
  1342. skb_split(skb, buff, len);
  1343. /* Fix up tso_factor for both original and new SKB. */
  1344. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1345. tcp_set_skb_tso_segs(sk, buff, mss_now);
  1346. /* Link BUFF into the send queue. */
  1347. skb_header_release(buff);
  1348. tcp_insert_write_queue_after(skb, buff, sk);
  1349. return 0;
  1350. }
  1351. /* Try to defer sending, if possible, in order to minimize the amount
  1352. * of TSO splitting we do. View it as a kind of TSO Nagle test.
  1353. *
  1354. * This algorithm is from John Heffner.
  1355. */
  1356. static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
  1357. {
  1358. struct tcp_sock *tp = tcp_sk(sk);
  1359. const struct inet_connection_sock *icsk = inet_csk(sk);
  1360. u32 send_win, cong_win, limit, in_flight;
  1361. int win_divisor;
  1362. if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
  1363. goto send_now;
  1364. if (icsk->icsk_ca_state != TCP_CA_Open)
  1365. goto send_now;
  1366. /* Defer for less than two clock ticks. */
  1367. if (tp->tso_deferred &&
  1368. (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
  1369. goto send_now;
  1370. in_flight = tcp_packets_in_flight(tp);
  1371. BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
  1372. send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  1373. /* From in_flight test above, we know that cwnd > in_flight. */
  1374. cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
  1375. limit = min(send_win, cong_win);
  1376. /* If a full-sized TSO skb can be sent, do it. */
  1377. if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
  1378. tp->xmit_size_goal_segs * tp->mss_cache))
  1379. goto send_now;
  1380. /* Middle in queue won't get any more data, full sendable already? */
  1381. if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
  1382. goto send_now;
  1383. win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
  1384. if (win_divisor) {
  1385. u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
  1386. /* If at least some fraction of a window is available,
  1387. * just use it.
  1388. */
  1389. chunk /= win_divisor;
  1390. if (limit >= chunk)
  1391. goto send_now;
  1392. } else {
  1393. /* Different approach, try not to defer past a single
  1394. * ACK. Receiver should ACK every other full sized
  1395. * frame, so if we have space for more than 3 frames
  1396. * then send now.
  1397. */
  1398. if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
  1399. goto send_now;
  1400. }
  1401. /* Ok, it looks like it is advisable to defer.
  1402. * Do not rearm the timer if already set to not break TCP ACK clocking.
  1403. */
  1404. if (!tp->tso_deferred)
  1405. tp->tso_deferred = 1 | (jiffies << 1);
  1406. return true;
  1407. send_now:
  1408. tp->tso_deferred = 0;
  1409. return false;
  1410. }
  1411. /* Create a new MTU probe if we are ready.
  1412. * MTU probe is regularly attempting to increase the path MTU by
  1413. * deliberately sending larger packets. This discovers routing
  1414. * changes resulting in larger path MTUs.
  1415. *
  1416. * Returns 0 if we should wait to probe (no cwnd available),
  1417. * 1 if a probe was sent,
  1418. * -1 otherwise
  1419. */
  1420. static int tcp_mtu_probe(struct sock *sk)
  1421. {
  1422. struct tcp_sock *tp = tcp_sk(sk);
  1423. struct inet_connection_sock *icsk = inet_csk(sk);
  1424. struct sk_buff *skb, *nskb, *next;
  1425. int len;
  1426. int probe_size;
  1427. int size_needed;
  1428. int copy;
  1429. int mss_now;
  1430. /* Not currently probing/verifying,
  1431. * not in recovery,
  1432. * have enough cwnd, and
  1433. * not SACKing (the variable headers throw things off) */
  1434. if (!icsk->icsk_mtup.enabled ||
  1435. icsk->icsk_mtup.probe_size ||
  1436. inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
  1437. tp->snd_cwnd < 11 ||
  1438. tp->rx_opt.num_sacks || tp->rx_opt.dsack)
  1439. return -1;
  1440. /* Very simple search strategy: just double the MSS. */
  1441. mss_now = tcp_current_mss(sk);
  1442. probe_size = 2 * tp->mss_cache;
  1443. size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
  1444. if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
  1445. /* TODO: set timer for probe_converge_event */
  1446. return -1;
  1447. }
  1448. /* Have enough data in the send queue to probe? */
  1449. if (tp->write_seq - tp->snd_nxt < size_needed)
  1450. return -1;
  1451. if (tp->snd_wnd < size_needed)
  1452. return -1;
  1453. if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
  1454. return 0;
  1455. /* Do we need to wait to drain cwnd? With none in flight, don't stall */
  1456. if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
  1457. if (!tcp_packets_in_flight(tp))
  1458. return -1;
  1459. else
  1460. return 0;
  1461. }
  1462. /* We're allowed to probe. Build it now. */
  1463. if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
  1464. return -1;
  1465. sk->sk_wmem_queued += nskb->truesize;
  1466. sk_mem_charge(sk, nskb->truesize);
  1467. skb = tcp_send_head(sk);
  1468. TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
  1469. TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
  1470. TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
  1471. TCP_SKB_CB(nskb)->sacked = 0;
  1472. nskb->csum = 0;
  1473. nskb->ip_summed = skb->ip_summed;
  1474. tcp_insert_write_queue_before(nskb, skb, sk);
  1475. len = 0;
  1476. tcp_for_write_queue_from_safe(skb, next, sk) {
  1477. copy = min_t(int, skb->len, probe_size - len);
  1478. if (nskb->ip_summed)
  1479. skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
  1480. else
  1481. nskb->csum = skb_copy_and_csum_bits(skb, 0,
  1482. skb_put(nskb, copy),
  1483. copy, nskb->csum);
  1484. if (skb->len <= copy) {
  1485. /* We've eaten all the data from this skb.
  1486. * Throw it away. */
  1487. TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
  1488. tcp_unlink_write_queue(skb, sk);
  1489. sk_wmem_free_skb(sk, skb);
  1490. } else {
  1491. TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
  1492. ~(TCPHDR_FIN|TCPHDR_PSH);
  1493. if (!skb_shinfo(skb)->nr_frags) {
  1494. skb_pull(skb, copy);
  1495. if (skb->ip_summed != CHECKSUM_PARTIAL)
  1496. skb->csum = csum_partial(skb->data,
  1497. skb->len, 0);
  1498. } else {
  1499. __pskb_trim_head(skb, copy);
  1500. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1501. }
  1502. TCP_SKB_CB(skb)->seq += copy;
  1503. }
  1504. len += copy;
  1505. if (len >= probe_size)
  1506. break;
  1507. }
  1508. tcp_init_tso_segs(sk, nskb, nskb->len);
  1509. /* We're ready to send. If this fails, the probe will
  1510. * be resegmented into mss-sized pieces by tcp_write_xmit(). */
  1511. TCP_SKB_CB(nskb)->when = tcp_time_stamp;
  1512. if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
  1513. /* Decrement cwnd here because we are sending
  1514. * effectively two packets. */
  1515. tp->snd_cwnd--;
  1516. tcp_event_new_data_sent(sk, nskb);
  1517. icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
  1518. tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
  1519. tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
  1520. return 1;
  1521. }
  1522. return -1;
  1523. }
  1524. /* This routine writes packets to the network. It advances the
  1525. * send_head. This happens as incoming acks open up the remote
  1526. * window for us.
  1527. *
  1528. * LARGESEND note: !tcp_urg_mode is overkill, only frames between
  1529. * snd_up-64k-mss .. snd_up cannot be large. However, taking into
  1530. * account rare use of URG, this is not a big flaw.
  1531. *
  1532. * Send at most one packet when push_one > 0. Temporarily ignore
  1533. * cwnd limit to force at most one packet out when push_one == 2.
  1534. * Returns true, if no segments are in flight and we have queued segments,
  1535. * but cannot send anything now because of SWS or another problem.
  1536. */
  1537. static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  1538. int push_one, gfp_t gfp)
  1539. {
  1540. struct tcp_sock *tp = tcp_sk(sk);
  1541. struct sk_buff *skb;
  1542. unsigned int tso_segs, sent_pkts;
  1543. int cwnd_quota;
  1544. int result;
  1545. sent_pkts = 0;
  1546. if (!push_one) {
  1547. /* Do MTU probing. */
  1548. result = tcp_mtu_probe(sk);
  1549. if (!result) {
  1550. return false;
  1551. } else if (result > 0) {
  1552. sent_pkts = 1;
  1553. }
  1554. }
  1555. while ((skb = tcp_send_head(sk))) {
  1556. unsigned int limit;
  1557. tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
  1558. BUG_ON(!tso_segs);
  1559. if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
  1560. goto repair; /* Skip network transmission */
  1561. cwnd_quota = tcp_cwnd_test(tp, skb);
  1562. if (!cwnd_quota) {
  1563. if (push_one == 2)
  1564. /* Force out a loss probe pkt. */
  1565. cwnd_quota = 1;
  1566. else
  1567. break;
  1568. }
  1569. if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
  1570. break;
  1571. if (tso_segs == 1) {
  1572. if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
  1573. (tcp_skb_is_last(sk, skb) ?
  1574. nonagle : TCP_NAGLE_PUSH))))
  1575. break;
  1576. } else {
  1577. if (!push_one && tcp_tso_should_defer(sk, skb))
  1578. break;
  1579. }
  1580. /* TCP Small Queues :
  1581. * Control number of packets in qdisc/devices to two packets / or ~1 ms.
  1582. * This allows for :
  1583. * - better RTT estimation and ACK scheduling
  1584. * - faster recovery
  1585. * - high rates
  1586. */
  1587. limit = max(skb->truesize, sk->sk_pacing_rate >> 10);
  1588. if (atomic_read(&sk->sk_wmem_alloc) > limit) {
  1589. set_bit(TSQ_THROTTLED, &tp->tsq_flags);
  1590. break;
  1591. }
  1592. limit = mss_now;
  1593. if (tso_segs > 1 && !tcp_urg_mode(tp))
  1594. limit = tcp_mss_split_point(sk, skb, mss_now,
  1595. min_t(unsigned int,
  1596. cwnd_quota,
  1597. sk->sk_gso_max_segs));
  1598. if (skb->len > limit &&
  1599. unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
  1600. break;
  1601. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1602. if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
  1603. break;
  1604. repair:
  1605. /* Advance the send_head. This one is sent out.
  1606. * This call will increment packets_out.
  1607. */
  1608. tcp_event_new_data_sent(sk, skb);
  1609. tcp_minshall_update(tp, mss_now, skb);
  1610. sent_pkts += tcp_skb_pcount(skb);
  1611. if (push_one)
  1612. break;
  1613. }
  1614. if (likely(sent_pkts)) {
  1615. if (tcp_in_cwnd_reduction(sk))
  1616. tp->prr_out += sent_pkts;
  1617. /* Send one loss probe per tail loss episode. */
  1618. if (push_one != 2)
  1619. tcp_schedule_loss_probe(sk);
  1620. tcp_cwnd_validate(sk);
  1621. return false;
  1622. }
  1623. return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
  1624. }
  1625. bool tcp_schedule_loss_probe(struct sock *sk)
  1626. {
  1627. struct inet_connection_sock *icsk = inet_csk(sk);
  1628. struct tcp_sock *tp = tcp_sk(sk);
  1629. u32 timeout, tlp_time_stamp, rto_time_stamp;
  1630. u32 rtt = tp->srtt >> 3;
  1631. if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
  1632. return false;
  1633. /* No consecutive loss probes. */
  1634. if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
  1635. tcp_rearm_rto(sk);
  1636. return false;
  1637. }
  1638. /* Don't do any loss probe on a Fast Open connection before 3WHS
  1639. * finishes.
  1640. */
  1641. if (sk->sk_state == TCP_SYN_RECV)
  1642. return false;
  1643. /* TLP is only scheduled when next timer event is RTO. */
  1644. if (icsk->icsk_pending != ICSK_TIME_RETRANS)
  1645. return false;
  1646. /* Schedule a loss probe in 2*RTT for SACK capable connections
  1647. * in Open state, that are either limited by cwnd or application.
  1648. */
  1649. if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
  1650. !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
  1651. return false;
  1652. if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
  1653. tcp_send_head(sk))
  1654. return false;
  1655. /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
  1656. * for delayed ack when there's one outstanding packet.
  1657. */
  1658. timeout = rtt << 1;
  1659. if (tp->packets_out == 1)
  1660. timeout = max_t(u32, timeout,
  1661. (rtt + (rtt >> 1) + TCP_DELACK_MAX));
  1662. timeout = max_t(u32, timeout, msecs_to_jiffies(10));
  1663. /* If RTO is shorter, just schedule TLP in its place. */
  1664. tlp_time_stamp = tcp_time_stamp + timeout;
  1665. rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
  1666. if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
  1667. s32 delta = rto_time_stamp - tcp_time_stamp;
  1668. if (delta > 0)
  1669. timeout = delta;
  1670. }
  1671. inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
  1672. TCP_RTO_MAX);
  1673. return true;
  1674. }
  1675. /* When probe timeout (PTO) fires, send a new segment if one exists, else
  1676. * retransmit the last segment.
  1677. */
  1678. void tcp_send_loss_probe(struct sock *sk)
  1679. {
  1680. struct tcp_sock *tp = tcp_sk(sk);
  1681. struct sk_buff *skb;
  1682. int pcount;
  1683. int mss = tcp_current_mss(sk);
  1684. int err = -1;
  1685. if (tcp_send_head(sk) != NULL) {
  1686. err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
  1687. goto rearm_timer;
  1688. }
  1689. /* At most one outstanding TLP retransmission. */
  1690. if (tp->tlp_high_seq)
  1691. goto rearm_timer;
  1692. /* Retransmit last segment. */
  1693. skb = tcp_write_queue_tail(sk);
  1694. if (WARN_ON(!skb))
  1695. goto rearm_timer;
  1696. pcount = tcp_skb_pcount(skb);
  1697. if (WARN_ON(!pcount))
  1698. goto rearm_timer;
  1699. if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
  1700. if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
  1701. goto rearm_timer;
  1702. skb = tcp_write_queue_tail(sk);
  1703. }
  1704. if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
  1705. goto rearm_timer;
  1706. /* Probe with zero data doesn't trigger fast recovery. */
  1707. if (skb->len > 0)
  1708. err = __tcp_retransmit_skb(sk, skb);
  1709. /* Record snd_nxt for loss detection. */
  1710. if (likely(!err))
  1711. tp->tlp_high_seq = tp->snd_nxt;
  1712. rearm_timer:
  1713. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  1714. inet_csk(sk)->icsk_rto,
  1715. TCP_RTO_MAX);
  1716. if (likely(!err))
  1717. NET_INC_STATS_BH(sock_net(sk),
  1718. LINUX_MIB_TCPLOSSPROBES);
  1719. return;
  1720. }
  1721. /* Push out any pending frames which were held back due to
  1722. * TCP_CORK or attempt at coalescing tiny packets.
  1723. * The socket must be locked by the caller.
  1724. */
  1725. void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
  1726. int nonagle)
  1727. {
  1728. /* If we are closed, the bytes will have to remain here.
  1729. * In time closedown will finish, we empty the write queue and
  1730. * all will be happy.
  1731. */
  1732. if (unlikely(sk->sk_state == TCP_CLOSE))
  1733. return;
  1734. if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
  1735. sk_gfp_atomic(sk, GFP_ATOMIC)))
  1736. tcp_check_probe_timer(sk);
  1737. }
  1738. /* Send _single_ skb sitting at the send head. This function requires
  1739. * true push pending frames to setup probe timer etc.
  1740. */
  1741. void tcp_push_one(struct sock *sk, unsigned int mss_now)
  1742. {
  1743. struct sk_buff *skb = tcp_send_head(sk);
  1744. BUG_ON(!skb || skb->len < mss_now);
  1745. tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
  1746. }
  1747. /* This function returns the amount that we can raise the
  1748. * usable window based on the following constraints
  1749. *
  1750. * 1. The window can never be shrunk once it is offered (RFC 793)
  1751. * 2. We limit memory per socket
  1752. *
  1753. * RFC 1122:
  1754. * "the suggested [SWS] avoidance algorithm for the receiver is to keep
  1755. * RECV.NEXT + RCV.WIN fixed until:
  1756. * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
  1757. *
  1758. * i.e. don't raise the right edge of the window until you can raise
  1759. * it at least MSS bytes.
  1760. *
  1761. * Unfortunately, the recommended algorithm breaks header prediction,
  1762. * since header prediction assumes th->window stays fixed.
  1763. *
  1764. * Strictly speaking, keeping th->window fixed violates the receiver
  1765. * side SWS prevention criteria. The problem is that under this rule
  1766. * a stream of single byte packets will cause the right side of the
  1767. * window to always advance by a single byte.
  1768. *
  1769. * Of course, if the sender implements sender side SWS prevention
  1770. * then this will not be a problem.
  1771. *
  1772. * BSD seems to make the following compromise:
  1773. *
  1774. * If the free space is less than the 1/4 of the maximum
  1775. * space available and the free space is less than 1/2 mss,
  1776. * then set the window to 0.
  1777. * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
  1778. * Otherwise, just prevent the window from shrinking
  1779. * and from being larger than the largest representable value.
  1780. *
  1781. * This prevents incremental opening of the window in the regime
  1782. * where TCP is limited by the speed of the reader side taking
  1783. * data out of the TCP receive queue. It does nothing about
  1784. * those cases where the window is constrained on the sender side
  1785. * because the pipeline is full.
  1786. *
  1787. * BSD also seems to "accidentally" limit itself to windows that are a
  1788. * multiple of MSS, at least until the free space gets quite small.
  1789. * This would appear to be a side effect of the mbuf implementation.
  1790. * Combining these two algorithms results in the observed behavior
  1791. * of having a fixed window size at almost all times.
  1792. *
  1793. * Below we obtain similar behavior by forcing the offered window to
  1794. * a multiple of the mss when it is feasible to do so.
  1795. *
  1796. * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
  1797. * Regular options like TIMESTAMP are taken into account.
  1798. */
  1799. u32 __tcp_select_window(struct sock *sk)
  1800. {
  1801. struct inet_connection_sock *icsk = inet_csk(sk);
  1802. struct tcp_sock *tp = tcp_sk(sk);
  1803. /* MSS for the peer's data. Previous versions used mss_clamp
  1804. * here. I don't know if the value based on our guesses
  1805. * of peer's MSS is better for the performance. It's more correct
  1806. * but may be worse for the performance because of rcv_mss
  1807. * fluctuations. --SAW 1998/11/1
  1808. */
  1809. int mss = icsk->icsk_ack.rcv_mss;
  1810. int free_space = tcp_space(sk);
  1811. int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
  1812. int window;
  1813. if (mss > full_space)
  1814. mss = full_space;
  1815. if (free_space < (full_space >> 1)) {
  1816. icsk->icsk_ack.quick = 0;
  1817. if (sk_under_memory_pressure(sk))
  1818. tp->rcv_ssthresh = min(tp->rcv_ssthresh,
  1819. 4U * tp->advmss);
  1820. if (free_space < mss)
  1821. return 0;
  1822. }
  1823. if (free_space > tp->rcv_ssthresh)
  1824. free_space = tp->rcv_ssthresh;
  1825. /* Don't do rounding if we are using window scaling, since the
  1826. * scaled window will not line up with the MSS boundary anyway.
  1827. */
  1828. window = tp->rcv_wnd;
  1829. if (tp->rx_opt.rcv_wscale) {
  1830. window = free_space;
  1831. /* Advertise enough space so that it won't get scaled away.
  1832. * Import case: prevent zero window announcement if
  1833. * 1<<rcv_wscale > mss.
  1834. */
  1835. if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
  1836. window = (((window >> tp->rx_opt.rcv_wscale) + 1)
  1837. << tp->rx_opt.rcv_wscale);
  1838. } else {
  1839. /* Get the largest window that is a nice multiple of mss.
  1840. * Window clamp already applied above.
  1841. * If our current window offering is within 1 mss of the
  1842. * free space we just keep it. This prevents the divide
  1843. * and multiply from happening most of the time.
  1844. * We also don't do any window rounding when the free space
  1845. * is too small.
  1846. */
  1847. if (window <= free_space - mss || window > free_space)
  1848. window = (free_space / mss) * mss;
  1849. else if (mss == full_space &&
  1850. free_space > window + (full_space >> 1))
  1851. window = free_space;
  1852. }
  1853. return window;
  1854. }
  1855. /* Collapses two adjacent SKB's during retransmission. */
  1856. static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
  1857. {
  1858. struct tcp_sock *tp = tcp_sk(sk);
  1859. struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
  1860. int skb_size, next_skb_size;
  1861. skb_size = skb->len;
  1862. next_skb_size = next_skb->len;
  1863. BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
  1864. tcp_highest_sack_combine(sk, next_skb, skb);
  1865. tcp_unlink_write_queue(next_skb, sk);
  1866. skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
  1867. next_skb_size);
  1868. if (next_skb->ip_summed == CHECKSUM_PARTIAL)
  1869. skb->ip_summed = CHECKSUM_PARTIAL;
  1870. if (skb->ip_summed != CHECKSUM_PARTIAL)
  1871. skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
  1872. /* Update sequence range on original skb. */
  1873. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
  1874. /* Merge over control information. This moves PSH/FIN etc. over */
  1875. TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
  1876. /* All done, get rid of second SKB and account for it so
  1877. * packet counting does not break.
  1878. */
  1879. TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
  1880. /* changed transmit queue under us so clear hints */
  1881. tcp_clear_retrans_hints_partial(tp);
  1882. if (next_skb == tp->retransmit_skb_hint)
  1883. tp->retransmit_skb_hint = skb;
  1884. tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
  1885. sk_wmem_free_skb(sk, next_skb);
  1886. }
  1887. /* Check if coalescing SKBs is legal. */
  1888. static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
  1889. {
  1890. if (tcp_skb_pcount(skb) > 1)
  1891. return false;
  1892. /* TODO: SACK collapsing could be used to remove this condition */
  1893. if (skb_shinfo(skb)->nr_frags != 0)
  1894. return false;
  1895. if (skb_cloned(skb))
  1896. return false;
  1897. if (skb == tcp_send_head(sk))
  1898. return false;
  1899. /* Some heurestics for collapsing over SACK'd could be invented */
  1900. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
  1901. return false;
  1902. return true;
  1903. }
  1904. /* Collapse packets in the retransmit queue to make to create
  1905. * less packets on the wire. This is only done on retransmission.
  1906. */
  1907. static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
  1908. int space)
  1909. {
  1910. struct tcp_sock *tp = tcp_sk(sk);
  1911. struct sk_buff *skb = to, *tmp;
  1912. bool first = true;
  1913. if (!sysctl_tcp_retrans_collapse)
  1914. return;
  1915. if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
  1916. return;
  1917. tcp_for_write_queue_from_safe(skb, tmp, sk) {
  1918. if (!tcp_can_collapse(sk, skb))
  1919. break;
  1920. space -= skb->len;
  1921. if (first) {
  1922. first = false;
  1923. continue;
  1924. }
  1925. if (space < 0)
  1926. break;
  1927. /* Punt if not enough space exists in the first SKB for
  1928. * the data in the second
  1929. */
  1930. if (skb->len > skb_availroom(to))
  1931. break;
  1932. if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
  1933. break;
  1934. tcp_collapse_retrans(sk, to);
  1935. }
  1936. }
  1937. /* This retransmits one SKB. Policy decisions and retransmit queue
  1938. * state updates are done by the caller. Returns non-zero if an
  1939. * error occurred which prevented the send.
  1940. */
  1941. int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
  1942. {
  1943. struct tcp_sock *tp = tcp_sk(sk);
  1944. struct inet_connection_sock *icsk = inet_csk(sk);
  1945. unsigned int cur_mss;
  1946. /* Inconslusive MTU probe */
  1947. if (icsk->icsk_mtup.probe_size) {
  1948. icsk->icsk_mtup.probe_size = 0;
  1949. }
  1950. /* Do not sent more than we queued. 1/4 is reserved for possible
  1951. * copying overhead: fragmentation, tunneling, mangling etc.
  1952. */
  1953. if (atomic_read(&sk->sk_wmem_alloc) >
  1954. min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
  1955. return -EAGAIN;
  1956. if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
  1957. if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
  1958. BUG();
  1959. if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
  1960. return -ENOMEM;
  1961. }
  1962. if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
  1963. return -EHOSTUNREACH; /* Routing failure or similar. */
  1964. cur_mss = tcp_current_mss(sk);
  1965. /* If receiver has shrunk his window, and skb is out of
  1966. * new window, do not retransmit it. The exception is the
  1967. * case, when window is shrunk to zero. In this case
  1968. * our retransmit serves as a zero window probe.
  1969. */
  1970. if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
  1971. TCP_SKB_CB(skb)->seq != tp->snd_una)
  1972. return -EAGAIN;
  1973. if (skb->len > cur_mss) {
  1974. if (tcp_fragment(sk, skb, cur_mss, cur_mss))
  1975. return -ENOMEM; /* We'll try again later. */
  1976. } else {
  1977. int oldpcount = tcp_skb_pcount(skb);
  1978. if (unlikely(oldpcount > 1)) {
  1979. tcp_init_tso_segs(sk, skb, cur_mss);
  1980. tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
  1981. }
  1982. }
  1983. tcp_retrans_try_collapse(sk, skb, cur_mss);
  1984. /* Some Solaris stacks overoptimize and ignore the FIN on a
  1985. * retransmit when old data is attached. So strip it off
  1986. * since it is cheap to do so and saves bytes on the network.
  1987. */
  1988. if (skb->len > 0 &&
  1989. (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
  1990. tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
  1991. if (!pskb_trim(skb, 0)) {
  1992. /* Reuse, even though it does some unnecessary work */
  1993. tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
  1994. TCP_SKB_CB(skb)->tcp_flags);
  1995. skb->ip_summed = CHECKSUM_NONE;
  1996. }
  1997. }
  1998. /* Make a copy, if the first transmission SKB clone we made
  1999. * is still in somebody's hands, else make a clone.
  2000. */
  2001. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2002. /* make sure skb->data is aligned on arches that require it
  2003. * and check if ack-trimming & collapsing extended the headroom
  2004. * beyond what csum_start can cover.
  2005. */
  2006. if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
  2007. skb_headroom(skb) >= 0xFFFF)) {
  2008. struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
  2009. GFP_ATOMIC);
  2010. return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
  2011. -ENOBUFS;
  2012. } else {
  2013. return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2014. }
  2015. }
  2016. int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
  2017. {
  2018. struct tcp_sock *tp = tcp_sk(sk);
  2019. int err = __tcp_retransmit_skb(sk, skb);
  2020. if (err == 0) {
  2021. /* Update global TCP statistics. */
  2022. TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
  2023. tp->total_retrans++;
  2024. #if FASTRETRANS_DEBUG > 0
  2025. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
  2026. net_dbg_ratelimited("retrans_out leaked\n");
  2027. }
  2028. #endif
  2029. if (!tp->retrans_out)
  2030. tp->lost_retrans_low = tp->snd_nxt;
  2031. TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
  2032. tp->retrans_out += tcp_skb_pcount(skb);
  2033. /* Save stamp of the first retransmit. */
  2034. if (!tp->retrans_stamp)
  2035. tp->retrans_stamp = TCP_SKB_CB(skb)->when;
  2036. tp->undo_retrans += tcp_skb_pcount(skb);
  2037. /* snd_nxt is stored to detect loss of retransmitted segment,
  2038. * see tcp_input.c tcp_sacktag_write_queue().
  2039. */
  2040. TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
  2041. } else {
  2042. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
  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. return;
  2147. NET_INC_STATS_BH(sock_net(sk), mib_idx);
  2148. if (tcp_in_cwnd_reduction(sk))
  2149. tp->prr_out += tcp_skb_pcount(skb);
  2150. if (skb == tcp_write_queue_head(sk))
  2151. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  2152. inet_csk(sk)->icsk_rto,
  2153. TCP_RTO_MAX);
  2154. }
  2155. }
  2156. /* Send a fin. The caller locks the socket for us. This cannot be
  2157. * allowed to fail queueing a FIN frame under any circumstances.
  2158. */
  2159. void tcp_send_fin(struct sock *sk)
  2160. {
  2161. struct tcp_sock *tp = tcp_sk(sk);
  2162. struct sk_buff *skb = tcp_write_queue_tail(sk);
  2163. int mss_now;
  2164. /* Optimization, tack on the FIN if we have a queue of
  2165. * unsent frames. But be careful about outgoing SACKS
  2166. * and IP options.
  2167. */
  2168. mss_now = tcp_current_mss(sk);
  2169. if (tcp_send_head(sk) != NULL) {
  2170. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
  2171. TCP_SKB_CB(skb)->end_seq++;
  2172. tp->write_seq++;
  2173. } else {
  2174. /* Socket is locked, keep trying until memory is available. */
  2175. for (;;) {
  2176. skb = alloc_skb_fclone(MAX_TCP_HEADER,
  2177. sk->sk_allocation);
  2178. if (skb)
  2179. break;
  2180. yield();
  2181. }
  2182. /* Reserve space for headers and prepare control bits. */
  2183. skb_reserve(skb, MAX_TCP_HEADER);
  2184. /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
  2185. tcp_init_nondata_skb(skb, tp->write_seq,
  2186. TCPHDR_ACK | TCPHDR_FIN);
  2187. tcp_queue_skb(sk, skb);
  2188. }
  2189. __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
  2190. }
  2191. /* We get here when a process closes a file descriptor (either due to
  2192. * an explicit close() or as a byproduct of exit()'ing) and there
  2193. * was unread data in the receive queue. This behavior is recommended
  2194. * by RFC 2525, section 2.17. -DaveM
  2195. */
  2196. void tcp_send_active_reset(struct sock *sk, gfp_t priority)
  2197. {
  2198. struct sk_buff *skb;
  2199. /* NOTE: No TCP options attached and we never retransmit this. */
  2200. skb = alloc_skb(MAX_TCP_HEADER, priority);
  2201. if (!skb) {
  2202. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
  2203. return;
  2204. }
  2205. /* Reserve space for headers and prepare control bits. */
  2206. skb_reserve(skb, MAX_TCP_HEADER);
  2207. tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
  2208. TCPHDR_ACK | TCPHDR_RST);
  2209. /* Send it off. */
  2210. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2211. if (tcp_transmit_skb(sk, skb, 0, priority))
  2212. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
  2213. TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
  2214. }
  2215. /* Send a crossed SYN-ACK during socket establishment.
  2216. * WARNING: This routine must only be called when we have already sent
  2217. * a SYN packet that crossed the incoming SYN that caused this routine
  2218. * to get called. If this assumption fails then the initial rcv_wnd
  2219. * and rcv_wscale values will not be correct.
  2220. */
  2221. int tcp_send_synack(struct sock *sk)
  2222. {
  2223. struct sk_buff *skb;
  2224. skb = tcp_write_queue_head(sk);
  2225. if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
  2226. pr_debug("%s: wrong queue state\n", __func__);
  2227. return -EFAULT;
  2228. }
  2229. if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
  2230. if (skb_cloned(skb)) {
  2231. struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
  2232. if (nskb == NULL)
  2233. return -ENOMEM;
  2234. tcp_unlink_write_queue(skb, sk);
  2235. skb_header_release(nskb);
  2236. __tcp_add_write_queue_head(sk, nskb);
  2237. sk_wmem_free_skb(sk, skb);
  2238. sk->sk_wmem_queued += nskb->truesize;
  2239. sk_mem_charge(sk, nskb->truesize);
  2240. skb = nskb;
  2241. }
  2242. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
  2243. TCP_ECN_send_synack(tcp_sk(sk), skb);
  2244. }
  2245. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2246. return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2247. }
  2248. /**
  2249. * tcp_make_synack - Prepare a SYN-ACK.
  2250. * sk: listener socket
  2251. * dst: dst entry attached to the SYNACK
  2252. * req: request_sock pointer
  2253. *
  2254. * Allocate one skb and build a SYNACK packet.
  2255. * @dst is consumed : Caller should not use it again.
  2256. */
  2257. struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
  2258. struct request_sock *req,
  2259. struct tcp_fastopen_cookie *foc)
  2260. {
  2261. struct tcp_out_options opts;
  2262. struct inet_request_sock *ireq = inet_rsk(req);
  2263. struct tcp_sock *tp = tcp_sk(sk);
  2264. struct tcphdr *th;
  2265. struct sk_buff *skb;
  2266. struct tcp_md5sig_key *md5;
  2267. int tcp_header_size;
  2268. int mss;
  2269. skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
  2270. if (unlikely(!skb)) {
  2271. dst_release(dst);
  2272. return NULL;
  2273. }
  2274. /* Reserve space for headers. */
  2275. skb_reserve(skb, MAX_TCP_HEADER);
  2276. skb_dst_set(skb, dst);
  2277. security_skb_owned_by(skb, sk);
  2278. mss = dst_metric_advmss(dst);
  2279. if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
  2280. mss = tp->rx_opt.user_mss;
  2281. if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
  2282. __u8 rcv_wscale;
  2283. /* Set this up on the first call only */
  2284. req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
  2285. /* limit the window selection if the user enforce a smaller rx buffer */
  2286. if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
  2287. (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
  2288. req->window_clamp = tcp_full_space(sk);
  2289. /* tcp_full_space because it is guaranteed to be the first packet */
  2290. tcp_select_initial_window(tcp_full_space(sk),
  2291. mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
  2292. &req->rcv_wnd,
  2293. &req->window_clamp,
  2294. ireq->wscale_ok,
  2295. &rcv_wscale,
  2296. dst_metric(dst, RTAX_INITRWND));
  2297. ireq->rcv_wscale = rcv_wscale;
  2298. }
  2299. memset(&opts, 0, sizeof(opts));
  2300. #ifdef CONFIG_SYN_COOKIES
  2301. if (unlikely(req->cookie_ts))
  2302. TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
  2303. else
  2304. #endif
  2305. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2306. tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
  2307. foc) + sizeof(*th);
  2308. skb_push(skb, tcp_header_size);
  2309. skb_reset_transport_header(skb);
  2310. th = tcp_hdr(skb);
  2311. memset(th, 0, sizeof(struct tcphdr));
  2312. th->syn = 1;
  2313. th->ack = 1;
  2314. TCP_ECN_make_synack(req, th);
  2315. th->source = htons(ireq->ir_num);
  2316. th->dest = ireq->ir_rmt_port;
  2317. /* Setting of flags are superfluous here for callers (and ECE is
  2318. * not even correctly set)
  2319. */
  2320. tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
  2321. TCPHDR_SYN | TCPHDR_ACK);
  2322. th->seq = htonl(TCP_SKB_CB(skb)->seq);
  2323. /* XXX data is queued and acked as is. No buffer/window check */
  2324. th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
  2325. /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
  2326. th->window = htons(min(req->rcv_wnd, 65535U));
  2327. tcp_options_write((__be32 *)(th + 1), tp, &opts);
  2328. th->doff = (tcp_header_size >> 2);
  2329. TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
  2330. #ifdef CONFIG_TCP_MD5SIG
  2331. /* Okay, we have all we need - do the md5 hash if needed */
  2332. if (md5) {
  2333. tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
  2334. md5, NULL, req, skb);
  2335. }
  2336. #endif
  2337. return skb;
  2338. }
  2339. EXPORT_SYMBOL(tcp_make_synack);
  2340. /* Do all connect socket setups that can be done AF independent. */
  2341. void tcp_connect_init(struct sock *sk)
  2342. {
  2343. const struct dst_entry *dst = __sk_dst_get(sk);
  2344. struct tcp_sock *tp = tcp_sk(sk);
  2345. __u8 rcv_wscale;
  2346. /* We'll fix this up when we get a response from the other end.
  2347. * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
  2348. */
  2349. tp->tcp_header_len = sizeof(struct tcphdr) +
  2350. (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
  2351. #ifdef CONFIG_TCP_MD5SIG
  2352. if (tp->af_specific->md5_lookup(sk, sk) != NULL)
  2353. tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
  2354. #endif
  2355. /* If user gave his TCP_MAXSEG, record it to clamp */
  2356. if (tp->rx_opt.user_mss)
  2357. tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
  2358. tp->max_window = 0;
  2359. tcp_mtup_init(sk);
  2360. tcp_sync_mss(sk, dst_mtu(dst));
  2361. if (!tp->window_clamp)
  2362. tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
  2363. tp->advmss = dst_metric_advmss(dst);
  2364. if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
  2365. tp->advmss = tp->rx_opt.user_mss;
  2366. tcp_initialize_rcv_mss(sk);
  2367. /* limit the window selection if the user enforce a smaller rx buffer */
  2368. if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
  2369. (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
  2370. tp->window_clamp = tcp_full_space(sk);
  2371. tcp_select_initial_window(tcp_full_space(sk),
  2372. tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
  2373. &tp->rcv_wnd,
  2374. &tp->window_clamp,
  2375. sysctl_tcp_window_scaling,
  2376. &rcv_wscale,
  2377. dst_metric(dst, RTAX_INITRWND));
  2378. tp->rx_opt.rcv_wscale = rcv_wscale;
  2379. tp->rcv_ssthresh = tp->rcv_wnd;
  2380. sk->sk_err = 0;
  2381. sock_reset_flag(sk, SOCK_DONE);
  2382. tp->snd_wnd = 0;
  2383. tcp_init_wl(tp, 0);
  2384. tp->snd_una = tp->write_seq;
  2385. tp->snd_sml = tp->write_seq;
  2386. tp->snd_up = tp->write_seq;
  2387. tp->snd_nxt = tp->write_seq;
  2388. if (likely(!tp->repair))
  2389. tp->rcv_nxt = 0;
  2390. else
  2391. tp->rcv_tstamp = tcp_time_stamp;
  2392. tp->rcv_wup = tp->rcv_nxt;
  2393. tp->copied_seq = tp->rcv_nxt;
  2394. inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
  2395. inet_csk(sk)->icsk_retransmits = 0;
  2396. tcp_clear_retrans(tp);
  2397. }
  2398. static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
  2399. {
  2400. struct tcp_sock *tp = tcp_sk(sk);
  2401. struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
  2402. tcb->end_seq += skb->len;
  2403. skb_header_release(skb);
  2404. __tcp_add_write_queue_tail(sk, skb);
  2405. sk->sk_wmem_queued += skb->truesize;
  2406. sk_mem_charge(sk, skb->truesize);
  2407. tp->write_seq = tcb->end_seq;
  2408. tp->packets_out += tcp_skb_pcount(skb);
  2409. }
  2410. /* Build and send a SYN with data and (cached) Fast Open cookie. However,
  2411. * queue a data-only packet after the regular SYN, such that regular SYNs
  2412. * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
  2413. * only the SYN sequence, the data are retransmitted in the first ACK.
  2414. * If cookie is not cached or other error occurs, falls back to send a
  2415. * regular SYN with Fast Open cookie request option.
  2416. */
  2417. static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
  2418. {
  2419. struct tcp_sock *tp = tcp_sk(sk);
  2420. struct tcp_fastopen_request *fo = tp->fastopen_req;
  2421. int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
  2422. struct sk_buff *syn_data = NULL, *data;
  2423. unsigned long last_syn_loss = 0;
  2424. tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
  2425. tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
  2426. &syn_loss, &last_syn_loss);
  2427. /* Recurring FO SYN losses: revert to regular handshake temporarily */
  2428. if (syn_loss > 1 &&
  2429. time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
  2430. fo->cookie.len = -1;
  2431. goto fallback;
  2432. }
  2433. if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
  2434. fo->cookie.len = -1;
  2435. else if (fo->cookie.len <= 0)
  2436. goto fallback;
  2437. /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
  2438. * user-MSS. Reserve maximum option space for middleboxes that add
  2439. * private TCP options. The cost is reduced data space in SYN :(
  2440. */
  2441. if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
  2442. tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
  2443. space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
  2444. MAX_TCP_OPTION_SPACE;
  2445. syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
  2446. sk->sk_allocation);
  2447. if (syn_data == NULL)
  2448. goto fallback;
  2449. for (i = 0; i < iovlen && syn_data->len < space; ++i) {
  2450. struct iovec *iov = &fo->data->msg_iov[i];
  2451. unsigned char __user *from = iov->iov_base;
  2452. int len = iov->iov_len;
  2453. if (syn_data->len + len > space)
  2454. len = space - syn_data->len;
  2455. else if (i + 1 == iovlen)
  2456. /* No more data pending in inet_wait_for_connect() */
  2457. fo->data = NULL;
  2458. if (skb_add_data(syn_data, from, len))
  2459. goto fallback;
  2460. }
  2461. /* Queue a data-only packet after the regular SYN for retransmission */
  2462. data = pskb_copy(syn_data, sk->sk_allocation);
  2463. if (data == NULL)
  2464. goto fallback;
  2465. TCP_SKB_CB(data)->seq++;
  2466. TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
  2467. TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
  2468. tcp_connect_queue_skb(sk, data);
  2469. fo->copied = data->len;
  2470. if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
  2471. tp->syn_data = (fo->copied > 0);
  2472. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
  2473. goto done;
  2474. }
  2475. syn_data = NULL;
  2476. fallback:
  2477. /* Send a regular SYN with Fast Open cookie request option */
  2478. if (fo->cookie.len > 0)
  2479. fo->cookie.len = 0;
  2480. err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
  2481. if (err)
  2482. tp->syn_fastopen = 0;
  2483. kfree_skb(syn_data);
  2484. done:
  2485. fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
  2486. return err;
  2487. }
  2488. /* Build a SYN and send it off. */
  2489. int tcp_connect(struct sock *sk)
  2490. {
  2491. struct tcp_sock *tp = tcp_sk(sk);
  2492. struct sk_buff *buff;
  2493. int err;
  2494. tcp_connect_init(sk);
  2495. if (unlikely(tp->repair)) {
  2496. tcp_finish_connect(sk, NULL);
  2497. return 0;
  2498. }
  2499. buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
  2500. if (unlikely(buff == NULL))
  2501. return -ENOBUFS;
  2502. /* Reserve space for headers. */
  2503. skb_reserve(buff, MAX_TCP_HEADER);
  2504. tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
  2505. tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
  2506. tcp_connect_queue_skb(sk, buff);
  2507. TCP_ECN_send_syn(sk, buff);
  2508. /* Send off SYN; include data in Fast Open. */
  2509. err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
  2510. tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
  2511. if (err == -ECONNREFUSED)
  2512. return err;
  2513. /* We change tp->snd_nxt after the tcp_transmit_skb() call
  2514. * in order to make this packet get counted in tcpOutSegs.
  2515. */
  2516. tp->snd_nxt = tp->write_seq;
  2517. tp->pushed_seq = tp->write_seq;
  2518. TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
  2519. /* Timer for repeating the SYN until an answer. */
  2520. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  2521. inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  2522. return 0;
  2523. }
  2524. EXPORT_SYMBOL(tcp_connect);
  2525. /* Send out a delayed ack, the caller does the policy checking
  2526. * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
  2527. * for details.
  2528. */
  2529. void tcp_send_delayed_ack(struct sock *sk)
  2530. {
  2531. struct inet_connection_sock *icsk = inet_csk(sk);
  2532. int ato = icsk->icsk_ack.ato;
  2533. unsigned long timeout;
  2534. if (ato > TCP_DELACK_MIN) {
  2535. const struct tcp_sock *tp = tcp_sk(sk);
  2536. int max_ato = HZ / 2;
  2537. if (icsk->icsk_ack.pingpong ||
  2538. (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
  2539. max_ato = TCP_DELACK_MAX;
  2540. /* Slow path, intersegment interval is "high". */
  2541. /* If some rtt estimate is known, use it to bound delayed ack.
  2542. * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
  2543. * directly.
  2544. */
  2545. if (tp->srtt) {
  2546. int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
  2547. if (rtt < max_ato)
  2548. max_ato = rtt;
  2549. }
  2550. ato = min(ato, max_ato);
  2551. }
  2552. /* Stay within the limit we were given */
  2553. timeout = jiffies + ato;
  2554. /* Use new timeout only if there wasn't a older one earlier. */
  2555. if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
  2556. /* If delack timer was blocked or is about to expire,
  2557. * send ACK now.
  2558. */
  2559. if (icsk->icsk_ack.blocked ||
  2560. time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
  2561. tcp_send_ack(sk);
  2562. return;
  2563. }
  2564. if (!time_before(timeout, icsk->icsk_ack.timeout))
  2565. timeout = icsk->icsk_ack.timeout;
  2566. }
  2567. icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
  2568. icsk->icsk_ack.timeout = timeout;
  2569. sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
  2570. }
  2571. /* This routine sends an ack and also updates the window. */
  2572. void tcp_send_ack(struct sock *sk)
  2573. {
  2574. struct sk_buff *buff;
  2575. /* If we have been reset, we may not send again. */
  2576. if (sk->sk_state == TCP_CLOSE)
  2577. return;
  2578. /* We are not putting this on the write queue, so
  2579. * tcp_transmit_skb() will set the ownership to this
  2580. * sock.
  2581. */
  2582. buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
  2583. if (buff == NULL) {
  2584. inet_csk_schedule_ack(sk);
  2585. inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
  2586. inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
  2587. TCP_DELACK_MAX, TCP_RTO_MAX);
  2588. return;
  2589. }
  2590. /* Reserve space for headers and prepare control bits. */
  2591. skb_reserve(buff, MAX_TCP_HEADER);
  2592. tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
  2593. /* Send it off, this clears delayed acks for us. */
  2594. TCP_SKB_CB(buff)->when = tcp_time_stamp;
  2595. tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
  2596. }
  2597. /* This routine sends a packet with an out of date sequence
  2598. * number. It assumes the other end will try to ack it.
  2599. *
  2600. * Question: what should we make while urgent mode?
  2601. * 4.4BSD forces sending single byte of data. We cannot send
  2602. * out of window data, because we have SND.NXT==SND.MAX...
  2603. *
  2604. * Current solution: to send TWO zero-length segments in urgent mode:
  2605. * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
  2606. * out-of-date with SND.UNA-1 to probe window.
  2607. */
  2608. static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
  2609. {
  2610. struct tcp_sock *tp = tcp_sk(sk);
  2611. struct sk_buff *skb;
  2612. /* We don't queue it, tcp_transmit_skb() sets ownership. */
  2613. skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
  2614. if (skb == NULL)
  2615. return -1;
  2616. /* Reserve space for headers and set control bits. */
  2617. skb_reserve(skb, MAX_TCP_HEADER);
  2618. /* Use a previous sequence. This should cause the other
  2619. * end to send an ack. Don't queue or clone SKB, just
  2620. * send it.
  2621. */
  2622. tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
  2623. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2624. return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
  2625. }
  2626. void tcp_send_window_probe(struct sock *sk)
  2627. {
  2628. if (sk->sk_state == TCP_ESTABLISHED) {
  2629. tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
  2630. tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
  2631. tcp_xmit_probe_skb(sk, 0);
  2632. }
  2633. }
  2634. /* Initiate keepalive or window probe from timer. */
  2635. int tcp_write_wakeup(struct sock *sk)
  2636. {
  2637. struct tcp_sock *tp = tcp_sk(sk);
  2638. struct sk_buff *skb;
  2639. if (sk->sk_state == TCP_CLOSE)
  2640. return -1;
  2641. if ((skb = tcp_send_head(sk)) != NULL &&
  2642. before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
  2643. int err;
  2644. unsigned int mss = tcp_current_mss(sk);
  2645. unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  2646. if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
  2647. tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
  2648. /* We are probing the opening of a window
  2649. * but the window size is != 0
  2650. * must have been a result SWS avoidance ( sender )
  2651. */
  2652. if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
  2653. skb->len > mss) {
  2654. seg_size = min(seg_size, mss);
  2655. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
  2656. if (tcp_fragment(sk, skb, seg_size, mss))
  2657. return -1;
  2658. } else if (!tcp_skb_pcount(skb))
  2659. tcp_set_skb_tso_segs(sk, skb, mss);
  2660. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
  2661. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2662. err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2663. if (!err)
  2664. tcp_event_new_data_sent(sk, skb);
  2665. return err;
  2666. } else {
  2667. if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
  2668. tcp_xmit_probe_skb(sk, 1);
  2669. return tcp_xmit_probe_skb(sk, 0);
  2670. }
  2671. }
  2672. /* A window probe timeout has occurred. If window is not closed send
  2673. * a partial packet else a zero probe.
  2674. */
  2675. void tcp_send_probe0(struct sock *sk)
  2676. {
  2677. struct inet_connection_sock *icsk = inet_csk(sk);
  2678. struct tcp_sock *tp = tcp_sk(sk);
  2679. int err;
  2680. err = tcp_write_wakeup(sk);
  2681. if (tp->packets_out || !tcp_send_head(sk)) {
  2682. /* Cancel probe timer, if it is not required. */
  2683. icsk->icsk_probes_out = 0;
  2684. icsk->icsk_backoff = 0;
  2685. return;
  2686. }
  2687. if (err <= 0) {
  2688. if (icsk->icsk_backoff < sysctl_tcp_retries2)
  2689. icsk->icsk_backoff++;
  2690. icsk->icsk_probes_out++;
  2691. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2692. min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
  2693. TCP_RTO_MAX);
  2694. } else {
  2695. /* If packet was not sent due to local congestion,
  2696. * do not backoff and do not remember icsk_probes_out.
  2697. * Let local senders to fight for local resources.
  2698. *
  2699. * Use accumulated backoff yet.
  2700. */
  2701. if (!icsk->icsk_probes_out)
  2702. icsk->icsk_probes_out = 1;
  2703. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2704. min(icsk->icsk_rto << icsk->icsk_backoff,
  2705. TCP_RESOURCE_PROBE_INTERVAL),
  2706. TCP_RTO_MAX);
  2707. }
  2708. }