tcp_output.c 74 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585
  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. * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
  9. *
  10. * Authors: Ross Biro
  11. * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  12. * Mark Evans, <evansmp@uhura.aston.ac.uk>
  13. * Corey Minyard <wf-rch!minyard@relay.EU.net>
  14. * Florian La Roche, <flla@stud.uni-sb.de>
  15. * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  16. * Linus Torvalds, <torvalds@cs.helsinki.fi>
  17. * Alan Cox, <gw4pts@gw4pts.ampr.org>
  18. * Matthew Dillon, <dillon@apollo.west.oic.com>
  19. * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  20. * Jorge Cwik, <jorge@laser.satlink.net>
  21. */
  22. /*
  23. * Changes: Pedro Roque : Retransmit queue handled by TCP.
  24. * : Fragmentation on mtu decrease
  25. * : Segment collapse on retransmit
  26. * : AF independence
  27. *
  28. * Linus Torvalds : send_delayed_ack
  29. * David S. Miller : Charge memory using the right skb
  30. * during syn/ack processing.
  31. * David S. Miller : Output engine completely rewritten.
  32. * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
  33. * Cacophonix Gaul : draft-minshall-nagle-01
  34. * J Hadi Salim : ECN support
  35. *
  36. */
  37. #include <net/tcp.h>
  38. #include <linux/compiler.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. /* This limits the percentage of the congestion window which we
  47. * will allow a single TSO frame to consume. Building TSO frames
  48. * which are too large can cause TCP streams to be bursty.
  49. */
  50. int sysctl_tcp_tso_win_divisor __read_mostly = 3;
  51. int sysctl_tcp_mtu_probing __read_mostly = 0;
  52. int sysctl_tcp_base_mss __read_mostly = 512;
  53. /* By default, RFC2861 behavior. */
  54. int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
  55. static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  56. {
  57. struct tcp_sock *tp = tcp_sk(sk);
  58. unsigned int prior_packets = tp->packets_out;
  59. tcp_advance_send_head(sk, skb);
  60. tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
  61. /* Don't override Nagle indefinately with F-RTO */
  62. if (tp->frto_counter == 2)
  63. tp->frto_counter = 3;
  64. tp->packets_out += tcp_skb_pcount(skb);
  65. if (!prior_packets)
  66. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  67. inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  68. }
  69. /* SND.NXT, if window was not shrunk.
  70. * If window has been shrunk, what should we make? It is not clear at all.
  71. * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  72. * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  73. * invalid. OK, let's make this for now:
  74. */
  75. static inline __u32 tcp_acceptable_seq(struct sock *sk)
  76. {
  77. struct tcp_sock *tp = tcp_sk(sk);
  78. if (!before(tcp_wnd_end(tp), tp->snd_nxt))
  79. return tp->snd_nxt;
  80. else
  81. return tcp_wnd_end(tp);
  82. }
  83. /* Calculate mss to advertise in SYN segment.
  84. * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
  85. *
  86. * 1. It is independent of path mtu.
  87. * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
  88. * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
  89. * attached devices, because some buggy hosts are confused by
  90. * large MSS.
  91. * 4. We do not make 3, we advertise MSS, calculated from first
  92. * hop device mtu, but allow to raise it to ip_rt_min_advmss.
  93. * This may be overridden via information stored in routing table.
  94. * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
  95. * probably even Jumbo".
  96. */
  97. static __u16 tcp_advertise_mss(struct sock *sk)
  98. {
  99. struct tcp_sock *tp = tcp_sk(sk);
  100. struct dst_entry *dst = __sk_dst_get(sk);
  101. int mss = tp->advmss;
  102. if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
  103. mss = dst_metric(dst, RTAX_ADVMSS);
  104. tp->advmss = mss;
  105. }
  106. return (__u16)mss;
  107. }
  108. /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
  109. * This is the first part of cwnd validation mechanism. */
  110. static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
  111. {
  112. struct tcp_sock *tp = tcp_sk(sk);
  113. s32 delta = tcp_time_stamp - tp->lsndtime;
  114. u32 restart_cwnd = tcp_init_cwnd(tp, dst);
  115. u32 cwnd = tp->snd_cwnd;
  116. tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
  117. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  118. restart_cwnd = min(restart_cwnd, cwnd);
  119. while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
  120. cwnd >>= 1;
  121. tp->snd_cwnd = max(cwnd, restart_cwnd);
  122. tp->snd_cwnd_stamp = tcp_time_stamp;
  123. tp->snd_cwnd_used = 0;
  124. }
  125. static void tcp_event_data_sent(struct tcp_sock *tp,
  126. struct sk_buff *skb, struct sock *sk)
  127. {
  128. struct inet_connection_sock *icsk = inet_csk(sk);
  129. const u32 now = tcp_time_stamp;
  130. if (sysctl_tcp_slow_start_after_idle &&
  131. (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
  132. tcp_cwnd_restart(sk, __sk_dst_get(sk));
  133. tp->lsndtime = now;
  134. /* If it is a reply for ato after last received
  135. * packet, enter pingpong mode.
  136. */
  137. if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
  138. icsk->icsk_ack.pingpong = 1;
  139. }
  140. static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
  141. {
  142. tcp_dec_quickack_mode(sk, pkts);
  143. inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
  144. }
  145. /* Determine a window scaling and initial window to offer.
  146. * Based on the assumption that the given amount of space
  147. * will be offered. Store the results in the tp structure.
  148. * NOTE: for smooth operation initial space offering should
  149. * be a multiple of mss if possible. We assume here that mss >= 1.
  150. * This MUST be enforced by all callers.
  151. */
  152. void tcp_select_initial_window(int __space, __u32 mss,
  153. __u32 *rcv_wnd, __u32 *window_clamp,
  154. int wscale_ok, __u8 *rcv_wscale)
  155. {
  156. unsigned int space = (__space < 0 ? 0 : __space);
  157. /* If no clamp set the clamp to the max possible scaled window */
  158. if (*window_clamp == 0)
  159. (*window_clamp) = (65535 << 14);
  160. space = min(*window_clamp, space);
  161. /* Quantize space offering to a multiple of mss if possible. */
  162. if (space > mss)
  163. space = (space / mss) * mss;
  164. /* NOTE: offering an initial window larger than 32767
  165. * will break some buggy TCP stacks. If the admin tells us
  166. * it is likely we could be speaking with such a buggy stack
  167. * we will truncate our initial window offering to 32K-1
  168. * unless the remote has sent us a window scaling option,
  169. * which we interpret as a sign the remote TCP is not
  170. * misinterpreting the window field as a signed quantity.
  171. */
  172. if (sysctl_tcp_workaround_signed_windows)
  173. (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
  174. else
  175. (*rcv_wnd) = space;
  176. (*rcv_wscale) = 0;
  177. if (wscale_ok) {
  178. /* Set window scaling on max possible window
  179. * See RFC1323 for an explanation of the limit to 14
  180. */
  181. space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
  182. space = min_t(u32, space, *window_clamp);
  183. while (space > 65535 && (*rcv_wscale) < 14) {
  184. space >>= 1;
  185. (*rcv_wscale)++;
  186. }
  187. }
  188. /* Set initial window to value enough for senders,
  189. * following RFC2414. Senders, not following this RFC,
  190. * will be satisfied with 2.
  191. */
  192. if (mss > (1<<*rcv_wscale)) {
  193. int init_cwnd = 4;
  194. if (mss > 1460*3)
  195. init_cwnd = 2;
  196. else if (mss > 1460)
  197. init_cwnd = 3;
  198. if (*rcv_wnd > init_cwnd*mss)
  199. *rcv_wnd = init_cwnd*mss;
  200. }
  201. /* Set the clamp no higher than max representable value */
  202. (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
  203. }
  204. /* Chose a new window to advertise, update state in tcp_sock for the
  205. * socket, and return result with RFC1323 scaling applied. The return
  206. * value can be stuffed directly into th->window for an outgoing
  207. * frame.
  208. */
  209. static u16 tcp_select_window(struct sock *sk)
  210. {
  211. struct tcp_sock *tp = tcp_sk(sk);
  212. u32 cur_win = tcp_receive_window(tp);
  213. u32 new_win = __tcp_select_window(sk);
  214. /* Never shrink the offered window */
  215. if (new_win < cur_win) {
  216. /* Danger Will Robinson!
  217. * Don't update rcv_wup/rcv_wnd here or else
  218. * we will not be able to advertise a zero
  219. * window in time. --DaveM
  220. *
  221. * Relax Will Robinson.
  222. */
  223. new_win = cur_win;
  224. }
  225. tp->rcv_wnd = new_win;
  226. tp->rcv_wup = tp->rcv_nxt;
  227. /* Make sure we do not exceed the maximum possible
  228. * scaled window.
  229. */
  230. if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
  231. new_win = min(new_win, MAX_TCP_WINDOW);
  232. else
  233. new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
  234. /* RFC1323 scaling applied */
  235. new_win >>= tp->rx_opt.rcv_wscale;
  236. /* If we advertise zero window, disable fast path. */
  237. if (new_win == 0)
  238. tp->pred_flags = 0;
  239. return new_win;
  240. }
  241. static inline void TCP_ECN_send_synack(struct tcp_sock *tp,
  242. struct sk_buff *skb)
  243. {
  244. TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
  245. if (!(tp->ecn_flags&TCP_ECN_OK))
  246. TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
  247. }
  248. static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
  249. {
  250. struct tcp_sock *tp = tcp_sk(sk);
  251. tp->ecn_flags = 0;
  252. if (sysctl_tcp_ecn) {
  253. TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE|TCPCB_FLAG_CWR;
  254. tp->ecn_flags = TCP_ECN_OK;
  255. }
  256. }
  257. static __inline__ void
  258. TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
  259. {
  260. if (inet_rsk(req)->ecn_ok)
  261. th->ece = 1;
  262. }
  263. static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
  264. int tcp_header_len)
  265. {
  266. struct tcp_sock *tp = tcp_sk(sk);
  267. if (tp->ecn_flags & TCP_ECN_OK) {
  268. /* Not-retransmitted data segment: set ECT and inject CWR. */
  269. if (skb->len != tcp_header_len &&
  270. !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
  271. INET_ECN_xmit(sk);
  272. if (tp->ecn_flags&TCP_ECN_QUEUE_CWR) {
  273. tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
  274. tcp_hdr(skb)->cwr = 1;
  275. skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
  276. }
  277. } else {
  278. /* ACK or retransmitted segment: clear ECT|CE */
  279. INET_ECN_dontxmit(sk);
  280. }
  281. if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
  282. tcp_hdr(skb)->ece = 1;
  283. }
  284. }
  285. static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
  286. __u32 tstamp, __u8 **md5_hash)
  287. {
  288. if (tp->rx_opt.tstamp_ok) {
  289. *ptr++ = htonl((TCPOPT_NOP << 24) |
  290. (TCPOPT_NOP << 16) |
  291. (TCPOPT_TIMESTAMP << 8) |
  292. TCPOLEN_TIMESTAMP);
  293. *ptr++ = htonl(tstamp);
  294. *ptr++ = htonl(tp->rx_opt.ts_recent);
  295. }
  296. if (tp->rx_opt.eff_sacks) {
  297. struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
  298. int this_sack;
  299. *ptr++ = htonl((TCPOPT_NOP << 24) |
  300. (TCPOPT_NOP << 16) |
  301. (TCPOPT_SACK << 8) |
  302. (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
  303. TCPOLEN_SACK_PERBLOCK)));
  304. for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
  305. *ptr++ = htonl(sp[this_sack].start_seq);
  306. *ptr++ = htonl(sp[this_sack].end_seq);
  307. }
  308. if (tp->rx_opt.dsack) {
  309. tp->rx_opt.dsack = 0;
  310. tp->rx_opt.eff_sacks--;
  311. }
  312. }
  313. #ifdef CONFIG_TCP_MD5SIG
  314. if (md5_hash) {
  315. *ptr++ = htonl((TCPOPT_NOP << 24) |
  316. (TCPOPT_NOP << 16) |
  317. (TCPOPT_MD5SIG << 8) |
  318. TCPOLEN_MD5SIG);
  319. *md5_hash = (__u8 *)ptr;
  320. }
  321. #endif
  322. }
  323. /* Construct a tcp options header for a SYN or SYN_ACK packet.
  324. * If this is every changed make sure to change the definition of
  325. * MAX_SYN_SIZE to match the new maximum number of options that you
  326. * can generate.
  327. *
  328. * Note - that with the RFC2385 TCP option, we make room for the
  329. * 16 byte MD5 hash. This will be filled in later, so the pointer for the
  330. * location to be filled is passed back up.
  331. */
  332. static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
  333. int offer_wscale, int wscale, __u32 tstamp,
  334. __u32 ts_recent, __u8 **md5_hash)
  335. {
  336. /* We always get an MSS option.
  337. * The option bytes which will be seen in normal data
  338. * packets should timestamps be used, must be in the MSS
  339. * advertised. But we subtract them from tp->mss_cache so
  340. * that calculations in tcp_sendmsg are simpler etc.
  341. * So account for this fact here if necessary. If we
  342. * don't do this correctly, as a receiver we won't
  343. * recognize data packets as being full sized when we
  344. * should, and thus we won't abide by the delayed ACK
  345. * rules correctly.
  346. * SACKs don't matter, we never delay an ACK when we
  347. * have any of those going out.
  348. */
  349. *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
  350. if (ts) {
  351. if (sack)
  352. *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
  353. (TCPOLEN_SACK_PERM << 16) |
  354. (TCPOPT_TIMESTAMP << 8) |
  355. TCPOLEN_TIMESTAMP);
  356. else
  357. *ptr++ = htonl((TCPOPT_NOP << 24) |
  358. (TCPOPT_NOP << 16) |
  359. (TCPOPT_TIMESTAMP << 8) |
  360. TCPOLEN_TIMESTAMP);
  361. *ptr++ = htonl(tstamp); /* TSVAL */
  362. *ptr++ = htonl(ts_recent); /* TSECR */
  363. } else if (sack)
  364. *ptr++ = htonl((TCPOPT_NOP << 24) |
  365. (TCPOPT_NOP << 16) |
  366. (TCPOPT_SACK_PERM << 8) |
  367. TCPOLEN_SACK_PERM);
  368. if (offer_wscale)
  369. *ptr++ = htonl((TCPOPT_NOP << 24) |
  370. (TCPOPT_WINDOW << 16) |
  371. (TCPOLEN_WINDOW << 8) |
  372. (wscale));
  373. #ifdef CONFIG_TCP_MD5SIG
  374. /*
  375. * If MD5 is enabled, then we set the option, and include the size
  376. * (always 18). The actual MD5 hash is added just before the
  377. * packet is sent.
  378. */
  379. if (md5_hash) {
  380. *ptr++ = htonl((TCPOPT_NOP << 24) |
  381. (TCPOPT_NOP << 16) |
  382. (TCPOPT_MD5SIG << 8) |
  383. TCPOLEN_MD5SIG);
  384. *md5_hash = (__u8 *) ptr;
  385. }
  386. #endif
  387. }
  388. /* This routine actually transmits TCP packets queued in by
  389. * tcp_do_sendmsg(). This is used by both the initial
  390. * transmission and possible later retransmissions.
  391. * All SKB's seen here are completely headerless. It is our
  392. * job to build the TCP header, and pass the packet down to
  393. * IP so it can do the same plus pass the packet off to the
  394. * device.
  395. *
  396. * We are working here with either a clone of the original
  397. * SKB, or a fresh unique copy made by the retransmit engine.
  398. */
  399. static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
  400. {
  401. const struct inet_connection_sock *icsk = inet_csk(sk);
  402. struct inet_sock *inet;
  403. struct tcp_sock *tp;
  404. struct tcp_skb_cb *tcb;
  405. int tcp_header_size;
  406. #ifdef CONFIG_TCP_MD5SIG
  407. struct tcp_md5sig_key *md5;
  408. __u8 *md5_hash_location;
  409. #endif
  410. struct tcphdr *th;
  411. int sysctl_flags;
  412. int err;
  413. BUG_ON(!skb || !tcp_skb_pcount(skb));
  414. /* If congestion control is doing timestamping, we must
  415. * take such a timestamp before we potentially clone/copy.
  416. */
  417. if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
  418. __net_timestamp(skb);
  419. if (likely(clone_it)) {
  420. if (unlikely(skb_cloned(skb)))
  421. skb = pskb_copy(skb, gfp_mask);
  422. else
  423. skb = skb_clone(skb, gfp_mask);
  424. if (unlikely(!skb))
  425. return -ENOBUFS;
  426. }
  427. inet = inet_sk(sk);
  428. tp = tcp_sk(sk);
  429. tcb = TCP_SKB_CB(skb);
  430. tcp_header_size = tp->tcp_header_len;
  431. #define SYSCTL_FLAG_TSTAMPS 0x1
  432. #define SYSCTL_FLAG_WSCALE 0x2
  433. #define SYSCTL_FLAG_SACK 0x4
  434. sysctl_flags = 0;
  435. if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
  436. tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
  437. if (sysctl_tcp_timestamps) {
  438. tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
  439. sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
  440. }
  441. if (sysctl_tcp_window_scaling) {
  442. tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
  443. sysctl_flags |= SYSCTL_FLAG_WSCALE;
  444. }
  445. if (sysctl_tcp_sack) {
  446. sysctl_flags |= SYSCTL_FLAG_SACK;
  447. if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
  448. tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
  449. }
  450. } else if (unlikely(tp->rx_opt.eff_sacks)) {
  451. /* A SACK is 2 pad bytes, a 2 byte header, plus
  452. * 2 32-bit sequence numbers for each SACK block.
  453. */
  454. tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
  455. (tp->rx_opt.eff_sacks *
  456. TCPOLEN_SACK_PERBLOCK));
  457. }
  458. if (tcp_packets_in_flight(tp) == 0)
  459. tcp_ca_event(sk, CA_EVENT_TX_START);
  460. #ifdef CONFIG_TCP_MD5SIG
  461. /*
  462. * Are we doing MD5 on this segment? If so - make
  463. * room for it.
  464. */
  465. md5 = tp->af_specific->md5_lookup(sk, sk);
  466. if (md5)
  467. tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
  468. #endif
  469. skb_push(skb, tcp_header_size);
  470. skb_reset_transport_header(skb);
  471. skb_set_owner_w(skb, sk);
  472. /* Build TCP header and checksum it. */
  473. th = tcp_hdr(skb);
  474. th->source = inet->sport;
  475. th->dest = inet->dport;
  476. th->seq = htonl(tcb->seq);
  477. th->ack_seq = htonl(tp->rcv_nxt);
  478. *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
  479. tcb->flags);
  480. if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
  481. /* RFC1323: The window in SYN & SYN/ACK segments
  482. * is never scaled.
  483. */
  484. th->window = htons(min(tp->rcv_wnd, 65535U));
  485. } else {
  486. th->window = htons(tcp_select_window(sk));
  487. }
  488. th->check = 0;
  489. th->urg_ptr = 0;
  490. if (unlikely(tp->urg_mode &&
  491. between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
  492. th->urg_ptr = htons(tp->snd_up-tcb->seq);
  493. th->urg = 1;
  494. }
  495. if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
  496. tcp_syn_build_options((__be32 *)(th + 1),
  497. tcp_advertise_mss(sk),
  498. (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
  499. (sysctl_flags & SYSCTL_FLAG_SACK),
  500. (sysctl_flags & SYSCTL_FLAG_WSCALE),
  501. tp->rx_opt.rcv_wscale,
  502. tcb->when,
  503. tp->rx_opt.ts_recent,
  504. #ifdef CONFIG_TCP_MD5SIG
  505. md5 ? &md5_hash_location :
  506. #endif
  507. NULL);
  508. } else {
  509. tcp_build_and_update_options((__be32 *)(th + 1),
  510. tp, tcb->when,
  511. #ifdef CONFIG_TCP_MD5SIG
  512. md5 ? &md5_hash_location :
  513. #endif
  514. NULL);
  515. TCP_ECN_send(sk, skb, tcp_header_size);
  516. }
  517. #ifdef CONFIG_TCP_MD5SIG
  518. /* Calculate the MD5 hash, as we have all we need now */
  519. if (md5) {
  520. tp->af_specific->calc_md5_hash(md5_hash_location,
  521. md5,
  522. sk, NULL, NULL,
  523. tcp_hdr(skb),
  524. sk->sk_protocol,
  525. skb->len);
  526. }
  527. #endif
  528. icsk->icsk_af_ops->send_check(sk, skb->len, skb);
  529. if (likely(tcb->flags & TCPCB_FLAG_ACK))
  530. tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
  531. if (skb->len != tcp_header_size)
  532. tcp_event_data_sent(tp, skb, sk);
  533. if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
  534. TCP_INC_STATS(TCP_MIB_OUTSEGS);
  535. err = icsk->icsk_af_ops->queue_xmit(skb, 0);
  536. if (likely(err <= 0))
  537. return err;
  538. tcp_enter_cwr(sk, 1);
  539. return net_xmit_eval(err);
  540. #undef SYSCTL_FLAG_TSTAMPS
  541. #undef SYSCTL_FLAG_WSCALE
  542. #undef SYSCTL_FLAG_SACK
  543. }
  544. /* This routine just queue's the buffer
  545. *
  546. * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
  547. * otherwise socket can stall.
  548. */
  549. static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
  550. {
  551. struct tcp_sock *tp = tcp_sk(sk);
  552. /* Advance write_seq and place onto the write_queue. */
  553. tp->write_seq = TCP_SKB_CB(skb)->end_seq;
  554. skb_header_release(skb);
  555. tcp_add_write_queue_tail(sk, skb);
  556. sk->sk_wmem_queued += skb->truesize;
  557. sk_mem_charge(sk, skb->truesize);
  558. }
  559. static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
  560. {
  561. if (skb->len <= mss_now || !sk_can_gso(sk)) {
  562. /* Avoid the costly divide in the normal
  563. * non-TSO case.
  564. */
  565. skb_shinfo(skb)->gso_segs = 1;
  566. skb_shinfo(skb)->gso_size = 0;
  567. skb_shinfo(skb)->gso_type = 0;
  568. } else {
  569. skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
  570. skb_shinfo(skb)->gso_size = mss_now;
  571. skb_shinfo(skb)->gso_type = sk->sk_gso_type;
  572. }
  573. }
  574. /* When a modification to fackets out becomes necessary, we need to check
  575. * skb is counted to fackets_out or not.
  576. */
  577. static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
  578. int decr)
  579. {
  580. struct tcp_sock *tp = tcp_sk(sk);
  581. if (!tp->sacked_out || tcp_is_reno(tp))
  582. return;
  583. if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
  584. tp->fackets_out -= decr;
  585. }
  586. /* Function to create two new TCP segments. Shrinks the given segment
  587. * to the specified size and appends a new segment with the rest of the
  588. * packet to the list. This won't be called frequently, I hope.
  589. * Remember, these are still headerless SKBs at this point.
  590. */
  591. int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
  592. {
  593. struct tcp_sock *tp = tcp_sk(sk);
  594. struct sk_buff *buff;
  595. int nsize, old_factor;
  596. int nlen;
  597. u16 flags;
  598. BUG_ON(len > skb->len);
  599. tcp_clear_retrans_hints_partial(tp);
  600. nsize = skb_headlen(skb) - len;
  601. if (nsize < 0)
  602. nsize = 0;
  603. if (skb_cloned(skb) &&
  604. skb_is_nonlinear(skb) &&
  605. pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
  606. return -ENOMEM;
  607. /* Get a new skb... force flag on. */
  608. buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
  609. if (buff == NULL)
  610. return -ENOMEM; /* We'll just try again later. */
  611. sk->sk_wmem_queued += buff->truesize;
  612. sk_mem_charge(sk, buff->truesize);
  613. nlen = skb->len - len - nsize;
  614. buff->truesize += nlen;
  615. skb->truesize -= nlen;
  616. /* Correct the sequence numbers. */
  617. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  618. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  619. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  620. /* PSH and FIN should only be set in the second packet. */
  621. flags = TCP_SKB_CB(skb)->flags;
  622. TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
  623. TCP_SKB_CB(buff)->flags = flags;
  624. TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
  625. if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
  626. /* Copy and checksum data tail into the new buffer. */
  627. buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
  628. nsize, 0);
  629. skb_trim(skb, len);
  630. skb->csum = csum_block_sub(skb->csum, buff->csum, len);
  631. } else {
  632. skb->ip_summed = CHECKSUM_PARTIAL;
  633. skb_split(skb, buff, len);
  634. }
  635. buff->ip_summed = skb->ip_summed;
  636. /* Looks stupid, but our code really uses when of
  637. * skbs, which it never sent before. --ANK
  638. */
  639. TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
  640. buff->tstamp = skb->tstamp;
  641. old_factor = tcp_skb_pcount(skb);
  642. /* Fix up tso_factor for both original and new SKB. */
  643. tcp_set_skb_tso_segs(sk, skb, mss_now);
  644. tcp_set_skb_tso_segs(sk, buff, mss_now);
  645. /* If this packet has been sent out already, we must
  646. * adjust the various packet counters.
  647. */
  648. if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
  649. int diff = old_factor - tcp_skb_pcount(skb) -
  650. tcp_skb_pcount(buff);
  651. tp->packets_out -= diff;
  652. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
  653. tp->sacked_out -= diff;
  654. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
  655. tp->retrans_out -= diff;
  656. if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
  657. tp->lost_out -= diff;
  658. /* Adjust Reno SACK estimate. */
  659. if (tcp_is_reno(tp) && diff > 0) {
  660. tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
  661. tcp_verify_left_out(tp);
  662. }
  663. tcp_adjust_fackets_out(sk, skb, diff);
  664. }
  665. /* Link BUFF into the send queue. */
  666. skb_header_release(buff);
  667. tcp_insert_write_queue_after(skb, buff, sk);
  668. return 0;
  669. }
  670. /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
  671. * eventually). The difference is that pulled data not copied, but
  672. * immediately discarded.
  673. */
  674. static void __pskb_trim_head(struct sk_buff *skb, int len)
  675. {
  676. int i, k, eat;
  677. eat = len;
  678. k = 0;
  679. for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
  680. if (skb_shinfo(skb)->frags[i].size <= eat) {
  681. put_page(skb_shinfo(skb)->frags[i].page);
  682. eat -= skb_shinfo(skb)->frags[i].size;
  683. } else {
  684. skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
  685. if (eat) {
  686. skb_shinfo(skb)->frags[k].page_offset += eat;
  687. skb_shinfo(skb)->frags[k].size -= eat;
  688. eat = 0;
  689. }
  690. k++;
  691. }
  692. }
  693. skb_shinfo(skb)->nr_frags = k;
  694. skb_reset_tail_pointer(skb);
  695. skb->data_len -= len;
  696. skb->len = skb->data_len;
  697. }
  698. int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
  699. {
  700. if (skb_cloned(skb) &&
  701. pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
  702. return -ENOMEM;
  703. /* If len == headlen, we avoid __skb_pull to preserve alignment. */
  704. if (unlikely(len < skb_headlen(skb)))
  705. __skb_pull(skb, len);
  706. else
  707. __pskb_trim_head(skb, len - skb_headlen(skb));
  708. TCP_SKB_CB(skb)->seq += len;
  709. skb->ip_summed = CHECKSUM_PARTIAL;
  710. skb->truesize -= len;
  711. sk->sk_wmem_queued -= len;
  712. sk_mem_uncharge(sk, len);
  713. sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
  714. /* Any change of skb->len requires recalculation of tso
  715. * factor and mss.
  716. */
  717. if (tcp_skb_pcount(skb) > 1)
  718. tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
  719. return 0;
  720. }
  721. /* Not accounting for SACKs here. */
  722. int tcp_mtu_to_mss(struct sock *sk, int pmtu)
  723. {
  724. struct tcp_sock *tp = tcp_sk(sk);
  725. struct inet_connection_sock *icsk = inet_csk(sk);
  726. int mss_now;
  727. /* Calculate base mss without TCP options:
  728. It is MMS_S - sizeof(tcphdr) of rfc1122
  729. */
  730. mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
  731. /* Clamp it (mss_clamp does not include tcp options) */
  732. if (mss_now > tp->rx_opt.mss_clamp)
  733. mss_now = tp->rx_opt.mss_clamp;
  734. /* Now subtract optional transport overhead */
  735. mss_now -= icsk->icsk_ext_hdr_len;
  736. /* Then reserve room for full set of TCP options and 8 bytes of data */
  737. if (mss_now < 48)
  738. mss_now = 48;
  739. /* Now subtract TCP options size, not including SACKs */
  740. mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
  741. return mss_now;
  742. }
  743. /* Inverse of above */
  744. int tcp_mss_to_mtu(struct sock *sk, int mss)
  745. {
  746. struct tcp_sock *tp = tcp_sk(sk);
  747. struct inet_connection_sock *icsk = inet_csk(sk);
  748. int mtu;
  749. mtu = mss +
  750. tp->tcp_header_len +
  751. icsk->icsk_ext_hdr_len +
  752. icsk->icsk_af_ops->net_header_len;
  753. return mtu;
  754. }
  755. void tcp_mtup_init(struct sock *sk)
  756. {
  757. struct tcp_sock *tp = tcp_sk(sk);
  758. struct inet_connection_sock *icsk = inet_csk(sk);
  759. icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
  760. icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
  761. icsk->icsk_af_ops->net_header_len;
  762. icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
  763. icsk->icsk_mtup.probe_size = 0;
  764. }
  765. /* This function synchronize snd mss to current pmtu/exthdr set.
  766. tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
  767. for TCP options, but includes only bare TCP header.
  768. tp->rx_opt.mss_clamp is mss negotiated at connection setup.
  769. It is minimum of user_mss and mss received with SYN.
  770. It also does not include TCP options.
  771. inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
  772. tp->mss_cache is current effective sending mss, including
  773. all tcp options except for SACKs. It is evaluated,
  774. taking into account current pmtu, but never exceeds
  775. tp->rx_opt.mss_clamp.
  776. NOTE1. rfc1122 clearly states that advertised MSS
  777. DOES NOT include either tcp or ip options.
  778. NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
  779. are READ ONLY outside this function. --ANK (980731)
  780. */
  781. unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
  782. {
  783. struct tcp_sock *tp = tcp_sk(sk);
  784. struct inet_connection_sock *icsk = inet_csk(sk);
  785. int mss_now;
  786. if (icsk->icsk_mtup.search_high > pmtu)
  787. icsk->icsk_mtup.search_high = pmtu;
  788. mss_now = tcp_mtu_to_mss(sk, pmtu);
  789. /* Bound mss with half of window */
  790. if (tp->max_window && mss_now > (tp->max_window>>1))
  791. mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
  792. /* And store cached results */
  793. icsk->icsk_pmtu_cookie = pmtu;
  794. if (icsk->icsk_mtup.enabled)
  795. mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
  796. tp->mss_cache = mss_now;
  797. return mss_now;
  798. }
  799. /* Compute the current effective MSS, taking SACKs and IP options,
  800. * and even PMTU discovery events into account.
  801. *
  802. * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
  803. * cannot be large. However, taking into account rare use of URG, this
  804. * is not a big flaw.
  805. */
  806. unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
  807. {
  808. struct tcp_sock *tp = tcp_sk(sk);
  809. struct dst_entry *dst = __sk_dst_get(sk);
  810. u32 mss_now;
  811. u16 xmit_size_goal;
  812. int doing_tso = 0;
  813. mss_now = tp->mss_cache;
  814. if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
  815. doing_tso = 1;
  816. if (dst) {
  817. u32 mtu = dst_mtu(dst);
  818. if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
  819. mss_now = tcp_sync_mss(sk, mtu);
  820. }
  821. if (tp->rx_opt.eff_sacks)
  822. mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
  823. (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
  824. #ifdef CONFIG_TCP_MD5SIG
  825. if (tp->af_specific->md5_lookup(sk, sk))
  826. mss_now -= TCPOLEN_MD5SIG_ALIGNED;
  827. #endif
  828. xmit_size_goal = mss_now;
  829. if (doing_tso) {
  830. xmit_size_goal = (65535 -
  831. inet_csk(sk)->icsk_af_ops->net_header_len -
  832. inet_csk(sk)->icsk_ext_hdr_len -
  833. tp->tcp_header_len);
  834. if (tp->max_window &&
  835. (xmit_size_goal > (tp->max_window >> 1)))
  836. xmit_size_goal = max((tp->max_window >> 1),
  837. 68U - tp->tcp_header_len);
  838. xmit_size_goal -= (xmit_size_goal % mss_now);
  839. }
  840. tp->xmit_size_goal = xmit_size_goal;
  841. return mss_now;
  842. }
  843. /* Congestion window validation. (RFC2861) */
  844. static void tcp_cwnd_validate(struct sock *sk)
  845. {
  846. struct tcp_sock *tp = tcp_sk(sk);
  847. __u32 packets_out = tp->packets_out;
  848. if (packets_out >= tp->snd_cwnd) {
  849. /* Network is feed fully. */
  850. tp->snd_cwnd_used = 0;
  851. tp->snd_cwnd_stamp = tcp_time_stamp;
  852. } else {
  853. /* Network starves. */
  854. if (tp->packets_out > tp->snd_cwnd_used)
  855. tp->snd_cwnd_used = tp->packets_out;
  856. if (sysctl_tcp_slow_start_after_idle &&
  857. (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
  858. tcp_cwnd_application_limited(sk);
  859. }
  860. }
  861. /* Returns the portion of skb which can be sent right away without
  862. * introducing MSS oddities to segment boundaries. In rare cases where
  863. * mss_now != mss_cache, we will request caller to create a small skb
  864. * per input skb which could be mostly avoided here (if desired).
  865. */
  866. static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
  867. unsigned int mss_now,
  868. unsigned int cwnd)
  869. {
  870. struct tcp_sock *tp = tcp_sk(sk);
  871. u32 needed, window, cwnd_len;
  872. window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  873. cwnd_len = mss_now * cwnd;
  874. if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
  875. return cwnd_len;
  876. if (skb == tcp_write_queue_tail(sk) && cwnd_len <= skb->len)
  877. return cwnd_len;
  878. needed = min(skb->len, window);
  879. return needed - needed % mss_now;
  880. }
  881. /* Can at least one segment of SKB be sent right now, according to the
  882. * congestion window rules? If so, return how many segments are allowed.
  883. */
  884. static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
  885. {
  886. u32 in_flight, cwnd;
  887. /* Don't be strict about the congestion window for the final FIN. */
  888. if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
  889. tcp_skb_pcount(skb) == 1)
  890. return 1;
  891. in_flight = tcp_packets_in_flight(tp);
  892. cwnd = tp->snd_cwnd;
  893. if (in_flight < cwnd)
  894. return (cwnd - in_flight);
  895. return 0;
  896. }
  897. /* This must be invoked the first time we consider transmitting
  898. * SKB onto the wire.
  899. */
  900. static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
  901. {
  902. int tso_segs = tcp_skb_pcount(skb);
  903. if (!tso_segs ||
  904. (tso_segs > 1 &&
  905. tcp_skb_mss(skb) != mss_now)) {
  906. tcp_set_skb_tso_segs(sk, skb, mss_now);
  907. tso_segs = tcp_skb_pcount(skb);
  908. }
  909. return tso_segs;
  910. }
  911. static inline int tcp_minshall_check(const struct tcp_sock *tp)
  912. {
  913. return after(tp->snd_sml,tp->snd_una) &&
  914. !after(tp->snd_sml, tp->snd_nxt);
  915. }
  916. /* Return 0, if packet can be sent now without violation Nagle's rules:
  917. * 1. It is full sized.
  918. * 2. Or it contains FIN. (already checked by caller)
  919. * 3. Or TCP_NODELAY was set.
  920. * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
  921. * With Minshall's modification: all sent small packets are ACKed.
  922. */
  923. static inline int tcp_nagle_check(const struct tcp_sock *tp,
  924. const struct sk_buff *skb,
  925. unsigned mss_now, int nonagle)
  926. {
  927. return (skb->len < mss_now &&
  928. ((nonagle&TCP_NAGLE_CORK) ||
  929. (!nonagle &&
  930. tp->packets_out &&
  931. tcp_minshall_check(tp))));
  932. }
  933. /* Return non-zero if the Nagle test allows this packet to be
  934. * sent now.
  935. */
  936. static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
  937. unsigned int cur_mss, int nonagle)
  938. {
  939. /* Nagle rule does not apply to frames, which sit in the middle of the
  940. * write_queue (they have no chances to get new data).
  941. *
  942. * This is implemented in the callers, where they modify the 'nonagle'
  943. * argument based upon the location of SKB in the send queue.
  944. */
  945. if (nonagle & TCP_NAGLE_PUSH)
  946. return 1;
  947. /* Don't use the nagle rule for urgent data (or for the final FIN).
  948. * Nagle can be ignored during F-RTO too (see RFC4138).
  949. */
  950. if (tp->urg_mode || (tp->frto_counter == 2) ||
  951. (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
  952. return 1;
  953. if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
  954. return 1;
  955. return 0;
  956. }
  957. /* Does at least the first segment of SKB fit into the send window? */
  958. static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
  959. {
  960. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  961. if (skb->len > cur_mss)
  962. end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
  963. return !after(end_seq, tcp_wnd_end(tp));
  964. }
  965. /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
  966. * should be put on the wire right now. If so, it returns the number of
  967. * packets allowed by the congestion window.
  968. */
  969. static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
  970. unsigned int cur_mss, int nonagle)
  971. {
  972. struct tcp_sock *tp = tcp_sk(sk);
  973. unsigned int cwnd_quota;
  974. tcp_init_tso_segs(sk, skb, cur_mss);
  975. if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
  976. return 0;
  977. cwnd_quota = tcp_cwnd_test(tp, skb);
  978. if (cwnd_quota &&
  979. !tcp_snd_wnd_test(tp, skb, cur_mss))
  980. cwnd_quota = 0;
  981. return cwnd_quota;
  982. }
  983. int tcp_may_send_now(struct sock *sk)
  984. {
  985. struct tcp_sock *tp = tcp_sk(sk);
  986. struct sk_buff *skb = tcp_send_head(sk);
  987. return (skb &&
  988. tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
  989. (tcp_skb_is_last(sk, skb) ?
  990. tp->nonagle : TCP_NAGLE_PUSH)));
  991. }
  992. /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
  993. * which is put after SKB on the list. It is very much like
  994. * tcp_fragment() except that it may make several kinds of assumptions
  995. * in order to speed up the splitting operation. In particular, we
  996. * know that all the data is in scatter-gather pages, and that the
  997. * packet has never been sent out before (and thus is not cloned).
  998. */
  999. static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
  1000. {
  1001. struct sk_buff *buff;
  1002. int nlen = skb->len - len;
  1003. u16 flags;
  1004. /* All of a TSO frame must be composed of paged data. */
  1005. if (skb->len != skb->data_len)
  1006. return tcp_fragment(sk, skb, len, mss_now);
  1007. buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
  1008. if (unlikely(buff == NULL))
  1009. return -ENOMEM;
  1010. sk->sk_wmem_queued += buff->truesize;
  1011. sk_mem_charge(sk, buff->truesize);
  1012. buff->truesize += nlen;
  1013. skb->truesize -= nlen;
  1014. /* Correct the sequence numbers. */
  1015. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  1016. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  1017. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  1018. /* PSH and FIN should only be set in the second packet. */
  1019. flags = TCP_SKB_CB(skb)->flags;
  1020. TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
  1021. TCP_SKB_CB(buff)->flags = flags;
  1022. /* This packet was never sent out yet, so no SACK bits. */
  1023. TCP_SKB_CB(buff)->sacked = 0;
  1024. buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
  1025. skb_split(skb, buff, len);
  1026. /* Fix up tso_factor for both original and new SKB. */
  1027. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1028. tcp_set_skb_tso_segs(sk, buff, mss_now);
  1029. /* Link BUFF into the send queue. */
  1030. skb_header_release(buff);
  1031. tcp_insert_write_queue_after(skb, buff, sk);
  1032. return 0;
  1033. }
  1034. /* Try to defer sending, if possible, in order to minimize the amount
  1035. * of TSO splitting we do. View it as a kind of TSO Nagle test.
  1036. *
  1037. * This algorithm is from John Heffner.
  1038. */
  1039. static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
  1040. {
  1041. struct tcp_sock *tp = tcp_sk(sk);
  1042. const struct inet_connection_sock *icsk = inet_csk(sk);
  1043. u32 send_win, cong_win, limit, in_flight;
  1044. if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
  1045. goto send_now;
  1046. if (icsk->icsk_ca_state != TCP_CA_Open)
  1047. goto send_now;
  1048. /* Defer for less than two clock ticks. */
  1049. if (tp->tso_deferred &&
  1050. ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
  1051. goto send_now;
  1052. in_flight = tcp_packets_in_flight(tp);
  1053. BUG_ON(tcp_skb_pcount(skb) <= 1 ||
  1054. (tp->snd_cwnd <= in_flight));
  1055. send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  1056. /* From in_flight test above, we know that cwnd > in_flight. */
  1057. cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
  1058. limit = min(send_win, cong_win);
  1059. /* If a full-sized TSO skb can be sent, do it. */
  1060. if (limit >= 65536)
  1061. goto send_now;
  1062. if (sysctl_tcp_tso_win_divisor) {
  1063. u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
  1064. /* If at least some fraction of a window is available,
  1065. * just use it.
  1066. */
  1067. chunk /= sysctl_tcp_tso_win_divisor;
  1068. if (limit >= chunk)
  1069. goto send_now;
  1070. } else {
  1071. /* Different approach, try not to defer past a single
  1072. * ACK. Receiver should ACK every other full sized
  1073. * frame, so if we have space for more than 3 frames
  1074. * then send now.
  1075. */
  1076. if (limit > tcp_max_burst(tp) * tp->mss_cache)
  1077. goto send_now;
  1078. }
  1079. /* Ok, it looks like it is advisable to defer. */
  1080. tp->tso_deferred = 1 | (jiffies<<1);
  1081. return 1;
  1082. send_now:
  1083. tp->tso_deferred = 0;
  1084. return 0;
  1085. }
  1086. /* Create a new MTU probe if we are ready.
  1087. * Returns 0 if we should wait to probe (no cwnd available),
  1088. * 1 if a probe was sent,
  1089. * -1 otherwise */
  1090. static int tcp_mtu_probe(struct sock *sk)
  1091. {
  1092. struct tcp_sock *tp = tcp_sk(sk);
  1093. struct inet_connection_sock *icsk = inet_csk(sk);
  1094. struct sk_buff *skb, *nskb, *next;
  1095. int len;
  1096. int probe_size;
  1097. int size_needed;
  1098. int copy;
  1099. int mss_now;
  1100. /* Not currently probing/verifying,
  1101. * not in recovery,
  1102. * have enough cwnd, and
  1103. * not SACKing (the variable headers throw things off) */
  1104. if (!icsk->icsk_mtup.enabled ||
  1105. icsk->icsk_mtup.probe_size ||
  1106. inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
  1107. tp->snd_cwnd < 11 ||
  1108. tp->rx_opt.eff_sacks)
  1109. return -1;
  1110. /* Very simple search strategy: just double the MSS. */
  1111. mss_now = tcp_current_mss(sk, 0);
  1112. probe_size = 2*tp->mss_cache;
  1113. size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
  1114. if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
  1115. /* TODO: set timer for probe_converge_event */
  1116. return -1;
  1117. }
  1118. /* Have enough data in the send queue to probe? */
  1119. if (tp->write_seq - tp->snd_nxt < size_needed)
  1120. return -1;
  1121. if (tp->snd_wnd < size_needed)
  1122. return -1;
  1123. if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
  1124. return 0;
  1125. /* Do we need to wait to drain cwnd? With none in flight, don't stall */
  1126. if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
  1127. if (!tcp_packets_in_flight(tp))
  1128. return -1;
  1129. else
  1130. return 0;
  1131. }
  1132. /* We're allowed to probe. Build it now. */
  1133. if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
  1134. return -1;
  1135. sk->sk_wmem_queued += nskb->truesize;
  1136. sk_mem_charge(sk, nskb->truesize);
  1137. skb = tcp_send_head(sk);
  1138. TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
  1139. TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
  1140. TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
  1141. TCP_SKB_CB(nskb)->sacked = 0;
  1142. nskb->csum = 0;
  1143. nskb->ip_summed = skb->ip_summed;
  1144. tcp_insert_write_queue_before(nskb, skb, sk);
  1145. len = 0;
  1146. tcp_for_write_queue_from_safe(skb, next, sk) {
  1147. copy = min_t(int, skb->len, probe_size - len);
  1148. if (nskb->ip_summed)
  1149. skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
  1150. else
  1151. nskb->csum = skb_copy_and_csum_bits(skb, 0,
  1152. skb_put(nskb, copy), copy, nskb->csum);
  1153. if (skb->len <= copy) {
  1154. /* We've eaten all the data from this skb.
  1155. * Throw it away. */
  1156. TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
  1157. tcp_unlink_write_queue(skb, sk);
  1158. sk_wmem_free_skb(sk, skb);
  1159. } else {
  1160. TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
  1161. ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
  1162. if (!skb_shinfo(skb)->nr_frags) {
  1163. skb_pull(skb, copy);
  1164. if (skb->ip_summed != CHECKSUM_PARTIAL)
  1165. skb->csum = csum_partial(skb->data, skb->len, 0);
  1166. } else {
  1167. __pskb_trim_head(skb, copy);
  1168. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1169. }
  1170. TCP_SKB_CB(skb)->seq += copy;
  1171. }
  1172. len += copy;
  1173. if (len >= probe_size)
  1174. break;
  1175. }
  1176. tcp_init_tso_segs(sk, nskb, nskb->len);
  1177. /* We're ready to send. If this fails, the probe will
  1178. * be resegmented into mss-sized pieces by tcp_write_xmit(). */
  1179. TCP_SKB_CB(nskb)->when = tcp_time_stamp;
  1180. if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
  1181. /* Decrement cwnd here because we are sending
  1182. * effectively two packets. */
  1183. tp->snd_cwnd--;
  1184. tcp_event_new_data_sent(sk, nskb);
  1185. icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
  1186. tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
  1187. tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
  1188. return 1;
  1189. }
  1190. return -1;
  1191. }
  1192. /* This routine writes packets to the network. It advances the
  1193. * send_head. This happens as incoming acks open up the remote
  1194. * window for us.
  1195. *
  1196. * Returns 1, if no segments are in flight and we have queued segments, but
  1197. * cannot send anything now because of SWS or another problem.
  1198. */
  1199. static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
  1200. {
  1201. struct tcp_sock *tp = tcp_sk(sk);
  1202. struct sk_buff *skb;
  1203. unsigned int tso_segs, sent_pkts;
  1204. int cwnd_quota;
  1205. int result;
  1206. /* If we are closed, the bytes will have to remain here.
  1207. * In time closedown will finish, we empty the write queue and all
  1208. * will be happy.
  1209. */
  1210. if (unlikely(sk->sk_state == TCP_CLOSE))
  1211. return 0;
  1212. sent_pkts = 0;
  1213. /* Do MTU probing. */
  1214. if ((result = tcp_mtu_probe(sk)) == 0) {
  1215. return 0;
  1216. } else if (result > 0) {
  1217. sent_pkts = 1;
  1218. }
  1219. while ((skb = tcp_send_head(sk))) {
  1220. unsigned int limit;
  1221. tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
  1222. BUG_ON(!tso_segs);
  1223. cwnd_quota = tcp_cwnd_test(tp, skb);
  1224. if (!cwnd_quota)
  1225. break;
  1226. if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
  1227. break;
  1228. if (tso_segs == 1) {
  1229. if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
  1230. (tcp_skb_is_last(sk, skb) ?
  1231. nonagle : TCP_NAGLE_PUSH))))
  1232. break;
  1233. } else {
  1234. if (tcp_tso_should_defer(sk, skb))
  1235. break;
  1236. }
  1237. limit = mss_now;
  1238. if (tso_segs > 1)
  1239. limit = tcp_mss_split_point(sk, skb, mss_now,
  1240. cwnd_quota);
  1241. if (skb->len > limit &&
  1242. unlikely(tso_fragment(sk, skb, limit, mss_now)))
  1243. break;
  1244. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1245. if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
  1246. break;
  1247. /* Advance the send_head. This one is sent out.
  1248. * This call will increment packets_out.
  1249. */
  1250. tcp_event_new_data_sent(sk, skb);
  1251. tcp_minshall_update(tp, mss_now, skb);
  1252. sent_pkts++;
  1253. }
  1254. if (likely(sent_pkts)) {
  1255. tcp_cwnd_validate(sk);
  1256. return 0;
  1257. }
  1258. return !tp->packets_out && tcp_send_head(sk);
  1259. }
  1260. /* Push out any pending frames which were held back due to
  1261. * TCP_CORK or attempt at coalescing tiny packets.
  1262. * The socket must be locked by the caller.
  1263. */
  1264. void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
  1265. int nonagle)
  1266. {
  1267. struct sk_buff *skb = tcp_send_head(sk);
  1268. if (skb) {
  1269. if (tcp_write_xmit(sk, cur_mss, nonagle))
  1270. tcp_check_probe_timer(sk);
  1271. }
  1272. }
  1273. /* Send _single_ skb sitting at the send head. This function requires
  1274. * true push pending frames to setup probe timer etc.
  1275. */
  1276. void tcp_push_one(struct sock *sk, unsigned int mss_now)
  1277. {
  1278. struct sk_buff *skb = tcp_send_head(sk);
  1279. unsigned int tso_segs, cwnd_quota;
  1280. BUG_ON(!skb || skb->len < mss_now);
  1281. tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
  1282. cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
  1283. if (likely(cwnd_quota)) {
  1284. unsigned int limit;
  1285. BUG_ON(!tso_segs);
  1286. limit = mss_now;
  1287. if (tso_segs > 1)
  1288. limit = tcp_mss_split_point(sk, skb, mss_now,
  1289. cwnd_quota);
  1290. if (skb->len > limit &&
  1291. unlikely(tso_fragment(sk, skb, limit, mss_now)))
  1292. return;
  1293. /* Send it out now. */
  1294. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1295. if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
  1296. tcp_event_new_data_sent(sk, skb);
  1297. tcp_cwnd_validate(sk);
  1298. return;
  1299. }
  1300. }
  1301. }
  1302. /* This function returns the amount that we can raise the
  1303. * usable window based on the following constraints
  1304. *
  1305. * 1. The window can never be shrunk once it is offered (RFC 793)
  1306. * 2. We limit memory per socket
  1307. *
  1308. * RFC 1122:
  1309. * "the suggested [SWS] avoidance algorithm for the receiver is to keep
  1310. * RECV.NEXT + RCV.WIN fixed until:
  1311. * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
  1312. *
  1313. * i.e. don't raise the right edge of the window until you can raise
  1314. * it at least MSS bytes.
  1315. *
  1316. * Unfortunately, the recommended algorithm breaks header prediction,
  1317. * since header prediction assumes th->window stays fixed.
  1318. *
  1319. * Strictly speaking, keeping th->window fixed violates the receiver
  1320. * side SWS prevention criteria. The problem is that under this rule
  1321. * a stream of single byte packets will cause the right side of the
  1322. * window to always advance by a single byte.
  1323. *
  1324. * Of course, if the sender implements sender side SWS prevention
  1325. * then this will not be a problem.
  1326. *
  1327. * BSD seems to make the following compromise:
  1328. *
  1329. * If the free space is less than the 1/4 of the maximum
  1330. * space available and the free space is less than 1/2 mss,
  1331. * then set the window to 0.
  1332. * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
  1333. * Otherwise, just prevent the window from shrinking
  1334. * and from being larger than the largest representable value.
  1335. *
  1336. * This prevents incremental opening of the window in the regime
  1337. * where TCP is limited by the speed of the reader side taking
  1338. * data out of the TCP receive queue. It does nothing about
  1339. * those cases where the window is constrained on the sender side
  1340. * because the pipeline is full.
  1341. *
  1342. * BSD also seems to "accidentally" limit itself to windows that are a
  1343. * multiple of MSS, at least until the free space gets quite small.
  1344. * This would appear to be a side effect of the mbuf implementation.
  1345. * Combining these two algorithms results in the observed behavior
  1346. * of having a fixed window size at almost all times.
  1347. *
  1348. * Below we obtain similar behavior by forcing the offered window to
  1349. * a multiple of the mss when it is feasible to do so.
  1350. *
  1351. * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
  1352. * Regular options like TIMESTAMP are taken into account.
  1353. */
  1354. u32 __tcp_select_window(struct sock *sk)
  1355. {
  1356. struct inet_connection_sock *icsk = inet_csk(sk);
  1357. struct tcp_sock *tp = tcp_sk(sk);
  1358. /* MSS for the peer's data. Previous versions used mss_clamp
  1359. * here. I don't know if the value based on our guesses
  1360. * of peer's MSS is better for the performance. It's more correct
  1361. * but may be worse for the performance because of rcv_mss
  1362. * fluctuations. --SAW 1998/11/1
  1363. */
  1364. int mss = icsk->icsk_ack.rcv_mss;
  1365. int free_space = tcp_space(sk);
  1366. int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
  1367. int window;
  1368. if (mss > full_space)
  1369. mss = full_space;
  1370. if (free_space < (full_space >> 1)) {
  1371. icsk->icsk_ack.quick = 0;
  1372. if (tcp_memory_pressure)
  1373. tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
  1374. if (free_space < mss)
  1375. return 0;
  1376. }
  1377. if (free_space > tp->rcv_ssthresh)
  1378. free_space = tp->rcv_ssthresh;
  1379. /* Don't do rounding if we are using window scaling, since the
  1380. * scaled window will not line up with the MSS boundary anyway.
  1381. */
  1382. window = tp->rcv_wnd;
  1383. if (tp->rx_opt.rcv_wscale) {
  1384. window = free_space;
  1385. /* Advertise enough space so that it won't get scaled away.
  1386. * Import case: prevent zero window announcement if
  1387. * 1<<rcv_wscale > mss.
  1388. */
  1389. if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
  1390. window = (((window >> tp->rx_opt.rcv_wscale) + 1)
  1391. << tp->rx_opt.rcv_wscale);
  1392. } else {
  1393. /* Get the largest window that is a nice multiple of mss.
  1394. * Window clamp already applied above.
  1395. * If our current window offering is within 1 mss of the
  1396. * free space we just keep it. This prevents the divide
  1397. * and multiply from happening most of the time.
  1398. * We also don't do any window rounding when the free space
  1399. * is too small.
  1400. */
  1401. if (window <= free_space - mss || window > free_space)
  1402. window = (free_space/mss)*mss;
  1403. else if (mss == full_space &&
  1404. free_space > window + (full_space >> 1))
  1405. window = free_space;
  1406. }
  1407. return window;
  1408. }
  1409. /* Attempt to collapse two adjacent SKB's during retransmission. */
  1410. static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
  1411. {
  1412. struct tcp_sock *tp = tcp_sk(sk);
  1413. struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
  1414. int skb_size, next_skb_size;
  1415. u16 flags;
  1416. /* The first test we must make is that neither of these two
  1417. * SKB's are still referenced by someone else.
  1418. */
  1419. if (skb_cloned(skb) || skb_cloned(next_skb))
  1420. return;
  1421. skb_size = skb->len;
  1422. next_skb_size = next_skb->len;
  1423. flags = TCP_SKB_CB(skb)->flags;
  1424. /* Also punt if next skb has been SACK'd. */
  1425. if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
  1426. return;
  1427. /* Next skb is out of window. */
  1428. if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp)))
  1429. return;
  1430. /* Punt if not enough space exists in the first SKB for
  1431. * the data in the second, or the total combined payload
  1432. * would exceed the MSS.
  1433. */
  1434. if ((next_skb_size > skb_tailroom(skb)) ||
  1435. ((skb_size + next_skb_size) > mss_now))
  1436. return;
  1437. BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
  1438. tcp_highest_sack_combine(sk, next_skb, skb);
  1439. /* Ok. We will be able to collapse the packet. */
  1440. tcp_unlink_write_queue(next_skb, sk);
  1441. skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
  1442. next_skb_size);
  1443. if (next_skb->ip_summed == CHECKSUM_PARTIAL)
  1444. skb->ip_summed = CHECKSUM_PARTIAL;
  1445. if (skb->ip_summed != CHECKSUM_PARTIAL)
  1446. skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
  1447. /* Update sequence range on original skb. */
  1448. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
  1449. /* Merge over control information. */
  1450. flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
  1451. TCP_SKB_CB(skb)->flags = flags;
  1452. /* All done, get rid of second SKB and account for it so
  1453. * packet counting does not break.
  1454. */
  1455. TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
  1456. if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
  1457. tp->retrans_out -= tcp_skb_pcount(next_skb);
  1458. if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
  1459. tp->lost_out -= tcp_skb_pcount(next_skb);
  1460. /* Reno case is special. Sigh... */
  1461. if (tcp_is_reno(tp) && tp->sacked_out)
  1462. tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
  1463. tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
  1464. tp->packets_out -= tcp_skb_pcount(next_skb);
  1465. /* changed transmit queue under us so clear hints */
  1466. tcp_clear_retrans_hints_partial(tp);
  1467. sk_wmem_free_skb(sk, next_skb);
  1468. }
  1469. /* Do a simple retransmit without using the backoff mechanisms in
  1470. * tcp_timer. This is used for path mtu discovery.
  1471. * The socket is already locked here.
  1472. */
  1473. void tcp_simple_retransmit(struct sock *sk)
  1474. {
  1475. const struct inet_connection_sock *icsk = inet_csk(sk);
  1476. struct tcp_sock *tp = tcp_sk(sk);
  1477. struct sk_buff *skb;
  1478. unsigned int mss = tcp_current_mss(sk, 0);
  1479. int lost = 0;
  1480. tcp_for_write_queue(skb, sk) {
  1481. if (skb == tcp_send_head(sk))
  1482. break;
  1483. if (skb->len > mss &&
  1484. !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
  1485. if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
  1486. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1487. tp->retrans_out -= tcp_skb_pcount(skb);
  1488. }
  1489. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
  1490. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1491. tp->lost_out += tcp_skb_pcount(skb);
  1492. lost = 1;
  1493. }
  1494. }
  1495. }
  1496. tcp_clear_all_retrans_hints(tp);
  1497. if (!lost)
  1498. return;
  1499. tcp_verify_left_out(tp);
  1500. /* Don't muck with the congestion window here.
  1501. * Reason is that we do not increase amount of _data_
  1502. * in network, but units changed and effective
  1503. * cwnd/ssthresh really reduced now.
  1504. */
  1505. if (icsk->icsk_ca_state != TCP_CA_Loss) {
  1506. tp->high_seq = tp->snd_nxt;
  1507. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  1508. tp->prior_ssthresh = 0;
  1509. tp->undo_marker = 0;
  1510. tcp_set_ca_state(sk, TCP_CA_Loss);
  1511. }
  1512. tcp_xmit_retransmit_queue(sk);
  1513. }
  1514. /* This retransmits one SKB. Policy decisions and retransmit queue
  1515. * state updates are done by the caller. Returns non-zero if an
  1516. * error occurred which prevented the send.
  1517. */
  1518. int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
  1519. {
  1520. struct tcp_sock *tp = tcp_sk(sk);
  1521. struct inet_connection_sock *icsk = inet_csk(sk);
  1522. unsigned int cur_mss = tcp_current_mss(sk, 0);
  1523. int err;
  1524. /* Inconslusive MTU probe */
  1525. if (icsk->icsk_mtup.probe_size) {
  1526. icsk->icsk_mtup.probe_size = 0;
  1527. }
  1528. /* Do not sent more than we queued. 1/4 is reserved for possible
  1529. * copying overhead: fragmentation, tunneling, mangling etc.
  1530. */
  1531. if (atomic_read(&sk->sk_wmem_alloc) >
  1532. min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
  1533. return -EAGAIN;
  1534. if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
  1535. if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
  1536. BUG();
  1537. if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
  1538. return -ENOMEM;
  1539. }
  1540. /* If receiver has shrunk his window, and skb is out of
  1541. * new window, do not retransmit it. The exception is the
  1542. * case, when window is shrunk to zero. In this case
  1543. * our retransmit serves as a zero window probe.
  1544. */
  1545. if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
  1546. && TCP_SKB_CB(skb)->seq != tp->snd_una)
  1547. return -EAGAIN;
  1548. if (skb->len > cur_mss) {
  1549. if (tcp_fragment(sk, skb, cur_mss, cur_mss))
  1550. return -ENOMEM; /* We'll try again later. */
  1551. }
  1552. /* Collapse two adjacent packets if worthwhile and we can. */
  1553. if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
  1554. (skb->len < (cur_mss >> 1)) &&
  1555. (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
  1556. (!tcp_skb_is_last(sk, skb)) &&
  1557. (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
  1558. (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
  1559. (sysctl_tcp_retrans_collapse != 0))
  1560. tcp_retrans_try_collapse(sk, skb, cur_mss);
  1561. if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
  1562. return -EHOSTUNREACH; /* Routing failure or similar. */
  1563. /* Some Solaris stacks overoptimize and ignore the FIN on a
  1564. * retransmit when old data is attached. So strip it off
  1565. * since it is cheap to do so and saves bytes on the network.
  1566. */
  1567. if (skb->len > 0 &&
  1568. (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
  1569. tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
  1570. if (!pskb_trim(skb, 0)) {
  1571. TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
  1572. skb_shinfo(skb)->gso_segs = 1;
  1573. skb_shinfo(skb)->gso_size = 0;
  1574. skb_shinfo(skb)->gso_type = 0;
  1575. skb->ip_summed = CHECKSUM_NONE;
  1576. skb->csum = 0;
  1577. }
  1578. }
  1579. /* Make a copy, if the first transmission SKB clone we made
  1580. * is still in somebody's hands, else make a clone.
  1581. */
  1582. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1583. err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  1584. if (err == 0) {
  1585. /* Update global TCP statistics. */
  1586. TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
  1587. tp->total_retrans++;
  1588. #if FASTRETRANS_DEBUG > 0
  1589. if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
  1590. if (net_ratelimit())
  1591. printk(KERN_DEBUG "retrans_out leaked.\n");
  1592. }
  1593. #endif
  1594. if (!tp->retrans_out)
  1595. tp->lost_retrans_low = tp->snd_nxt;
  1596. TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
  1597. tp->retrans_out += tcp_skb_pcount(skb);
  1598. /* Save stamp of the first retransmit. */
  1599. if (!tp->retrans_stamp)
  1600. tp->retrans_stamp = TCP_SKB_CB(skb)->when;
  1601. tp->undo_retrans++;
  1602. /* snd_nxt is stored to detect loss of retransmitted segment,
  1603. * see tcp_input.c tcp_sacktag_write_queue().
  1604. */
  1605. TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
  1606. }
  1607. return err;
  1608. }
  1609. /* This gets called after a retransmit timeout, and the initially
  1610. * retransmitted data is acknowledged. It tries to continue
  1611. * resending the rest of the retransmit queue, until either
  1612. * we've sent it all or the congestion window limit is reached.
  1613. * If doing SACK, the first ACK which comes back for a timeout
  1614. * based retransmit packet might feed us FACK information again.
  1615. * If so, we use it to avoid unnecessarily retransmissions.
  1616. */
  1617. void tcp_xmit_retransmit_queue(struct sock *sk)
  1618. {
  1619. const struct inet_connection_sock *icsk = inet_csk(sk);
  1620. struct tcp_sock *tp = tcp_sk(sk);
  1621. struct sk_buff *skb;
  1622. int packet_cnt;
  1623. if (tp->retransmit_skb_hint) {
  1624. skb = tp->retransmit_skb_hint;
  1625. packet_cnt = tp->retransmit_cnt_hint;
  1626. }else{
  1627. skb = tcp_write_queue_head(sk);
  1628. packet_cnt = 0;
  1629. }
  1630. /* First pass: retransmit lost packets. */
  1631. if (tp->lost_out) {
  1632. tcp_for_write_queue_from(skb, sk) {
  1633. __u8 sacked = TCP_SKB_CB(skb)->sacked;
  1634. if (skb == tcp_send_head(sk))
  1635. break;
  1636. /* we could do better than to assign each time */
  1637. tp->retransmit_skb_hint = skb;
  1638. tp->retransmit_cnt_hint = packet_cnt;
  1639. /* Assume this retransmit will generate
  1640. * only one packet for congestion window
  1641. * calculation purposes. This works because
  1642. * tcp_retransmit_skb() will chop up the
  1643. * packet to be MSS sized and all the
  1644. * packet counting works out.
  1645. */
  1646. if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
  1647. return;
  1648. if (sacked & TCPCB_LOST) {
  1649. if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
  1650. if (tcp_retransmit_skb(sk, skb)) {
  1651. tp->retransmit_skb_hint = NULL;
  1652. return;
  1653. }
  1654. if (icsk->icsk_ca_state != TCP_CA_Loss)
  1655. NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
  1656. else
  1657. NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
  1658. if (skb == tcp_write_queue_head(sk))
  1659. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  1660. inet_csk(sk)->icsk_rto,
  1661. TCP_RTO_MAX);
  1662. }
  1663. packet_cnt += tcp_skb_pcount(skb);
  1664. if (packet_cnt >= tp->lost_out)
  1665. break;
  1666. }
  1667. }
  1668. }
  1669. /* OK, demanded retransmission is finished. */
  1670. /* Forward retransmissions are possible only during Recovery. */
  1671. if (icsk->icsk_ca_state != TCP_CA_Recovery)
  1672. return;
  1673. /* No forward retransmissions in Reno are possible. */
  1674. if (tcp_is_reno(tp))
  1675. return;
  1676. /* Yeah, we have to make difficult choice between forward transmission
  1677. * and retransmission... Both ways have their merits...
  1678. *
  1679. * For now we do not retransmit anything, while we have some new
  1680. * segments to send. In the other cases, follow rule 3 for
  1681. * NextSeg() specified in RFC3517.
  1682. */
  1683. if (tcp_may_send_now(sk))
  1684. return;
  1685. /* If nothing is SACKed, highest_sack in the loop won't be valid */
  1686. if (!tp->sacked_out)
  1687. return;
  1688. if (tp->forward_skb_hint)
  1689. skb = tp->forward_skb_hint;
  1690. else
  1691. skb = tcp_write_queue_head(sk);
  1692. tcp_for_write_queue_from(skb, sk) {
  1693. if (skb == tcp_send_head(sk))
  1694. break;
  1695. tp->forward_skb_hint = skb;
  1696. if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
  1697. break;
  1698. if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
  1699. break;
  1700. if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
  1701. continue;
  1702. /* Ok, retransmit it. */
  1703. if (tcp_retransmit_skb(sk, skb)) {
  1704. tp->forward_skb_hint = NULL;
  1705. break;
  1706. }
  1707. if (skb == tcp_write_queue_head(sk))
  1708. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  1709. inet_csk(sk)->icsk_rto,
  1710. TCP_RTO_MAX);
  1711. NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
  1712. }
  1713. }
  1714. /* Send a fin. The caller locks the socket for us. This cannot be
  1715. * allowed to fail queueing a FIN frame under any circumstances.
  1716. */
  1717. void tcp_send_fin(struct sock *sk)
  1718. {
  1719. struct tcp_sock *tp = tcp_sk(sk);
  1720. struct sk_buff *skb = tcp_write_queue_tail(sk);
  1721. int mss_now;
  1722. /* Optimization, tack on the FIN if we have a queue of
  1723. * unsent frames. But be careful about outgoing SACKS
  1724. * and IP options.
  1725. */
  1726. mss_now = tcp_current_mss(sk, 1);
  1727. if (tcp_send_head(sk) != NULL) {
  1728. TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
  1729. TCP_SKB_CB(skb)->end_seq++;
  1730. tp->write_seq++;
  1731. } else {
  1732. /* Socket is locked, keep trying until memory is available. */
  1733. for (;;) {
  1734. skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
  1735. if (skb)
  1736. break;
  1737. yield();
  1738. }
  1739. /* Reserve space for headers and prepare control bits. */
  1740. skb_reserve(skb, MAX_TCP_HEADER);
  1741. skb->csum = 0;
  1742. TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
  1743. TCP_SKB_CB(skb)->sacked = 0;
  1744. skb_shinfo(skb)->gso_segs = 1;
  1745. skb_shinfo(skb)->gso_size = 0;
  1746. skb_shinfo(skb)->gso_type = 0;
  1747. /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
  1748. TCP_SKB_CB(skb)->seq = tp->write_seq;
  1749. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
  1750. tcp_queue_skb(sk, skb);
  1751. }
  1752. __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
  1753. }
  1754. /* We get here when a process closes a file descriptor (either due to
  1755. * an explicit close() or as a byproduct of exit()'ing) and there
  1756. * was unread data in the receive queue. This behavior is recommended
  1757. * by RFC 2525, section 2.17. -DaveM
  1758. */
  1759. void tcp_send_active_reset(struct sock *sk, gfp_t priority)
  1760. {
  1761. struct sk_buff *skb;
  1762. /* NOTE: No TCP options attached and we never retransmit this. */
  1763. skb = alloc_skb(MAX_TCP_HEADER, priority);
  1764. if (!skb) {
  1765. NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
  1766. return;
  1767. }
  1768. /* Reserve space for headers and prepare control bits. */
  1769. skb_reserve(skb, MAX_TCP_HEADER);
  1770. skb->csum = 0;
  1771. TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
  1772. TCP_SKB_CB(skb)->sacked = 0;
  1773. skb_shinfo(skb)->gso_segs = 1;
  1774. skb_shinfo(skb)->gso_size = 0;
  1775. skb_shinfo(skb)->gso_type = 0;
  1776. /* Send it off. */
  1777. TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk);
  1778. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
  1779. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1780. if (tcp_transmit_skb(sk, skb, 0, priority))
  1781. NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
  1782. }
  1783. /* WARNING: This routine must only be called when we have already sent
  1784. * a SYN packet that crossed the incoming SYN that caused this routine
  1785. * to get called. If this assumption fails then the initial rcv_wnd
  1786. * and rcv_wscale values will not be correct.
  1787. */
  1788. int tcp_send_synack(struct sock *sk)
  1789. {
  1790. struct sk_buff* skb;
  1791. skb = tcp_write_queue_head(sk);
  1792. if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
  1793. printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
  1794. return -EFAULT;
  1795. }
  1796. if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
  1797. if (skb_cloned(skb)) {
  1798. struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
  1799. if (nskb == NULL)
  1800. return -ENOMEM;
  1801. tcp_unlink_write_queue(skb, sk);
  1802. skb_header_release(nskb);
  1803. __tcp_add_write_queue_head(sk, nskb);
  1804. sk_wmem_free_skb(sk, skb);
  1805. sk->sk_wmem_queued += nskb->truesize;
  1806. sk_mem_charge(sk, nskb->truesize);
  1807. skb = nskb;
  1808. }
  1809. TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
  1810. TCP_ECN_send_synack(tcp_sk(sk), skb);
  1811. }
  1812. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1813. return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  1814. }
  1815. /*
  1816. * Prepare a SYN-ACK.
  1817. */
  1818. struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
  1819. struct request_sock *req)
  1820. {
  1821. struct inet_request_sock *ireq = inet_rsk(req);
  1822. struct tcp_sock *tp = tcp_sk(sk);
  1823. struct tcphdr *th;
  1824. int tcp_header_size;
  1825. struct sk_buff *skb;
  1826. #ifdef CONFIG_TCP_MD5SIG
  1827. struct tcp_md5sig_key *md5;
  1828. __u8 *md5_hash_location;
  1829. #endif
  1830. skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
  1831. if (skb == NULL)
  1832. return NULL;
  1833. /* Reserve space for headers. */
  1834. skb_reserve(skb, MAX_TCP_HEADER);
  1835. skb->dst = dst_clone(dst);
  1836. tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
  1837. (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
  1838. (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
  1839. /* SACK_PERM is in the place of NOP NOP of TS */
  1840. ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
  1841. #ifdef CONFIG_TCP_MD5SIG
  1842. /* Are we doing MD5 on this segment? If so - make room for it */
  1843. md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
  1844. if (md5)
  1845. tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
  1846. #endif
  1847. skb_push(skb, tcp_header_size);
  1848. skb_reset_transport_header(skb);
  1849. th = tcp_hdr(skb);
  1850. memset(th, 0, sizeof(struct tcphdr));
  1851. th->syn = 1;
  1852. th->ack = 1;
  1853. TCP_ECN_make_synack(req, th);
  1854. th->source = inet_sk(sk)->sport;
  1855. th->dest = ireq->rmt_port;
  1856. TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
  1857. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
  1858. TCP_SKB_CB(skb)->sacked = 0;
  1859. skb_shinfo(skb)->gso_segs = 1;
  1860. skb_shinfo(skb)->gso_size = 0;
  1861. skb_shinfo(skb)->gso_type = 0;
  1862. th->seq = htonl(TCP_SKB_CB(skb)->seq);
  1863. th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
  1864. if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
  1865. __u8 rcv_wscale;
  1866. /* Set this up on the first call only */
  1867. req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
  1868. /* tcp_full_space because it is guaranteed to be the first packet */
  1869. tcp_select_initial_window(tcp_full_space(sk),
  1870. dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
  1871. &req->rcv_wnd,
  1872. &req->window_clamp,
  1873. ireq->wscale_ok,
  1874. &rcv_wscale);
  1875. ireq->rcv_wscale = rcv_wscale;
  1876. }
  1877. /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
  1878. th->window = htons(min(req->rcv_wnd, 65535U));
  1879. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1880. tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
  1881. ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
  1882. TCP_SKB_CB(skb)->when,
  1883. req->ts_recent,
  1884. (
  1885. #ifdef CONFIG_TCP_MD5SIG
  1886. md5 ? &md5_hash_location :
  1887. #endif
  1888. NULL)
  1889. );
  1890. skb->csum = 0;
  1891. th->doff = (tcp_header_size >> 2);
  1892. TCP_INC_STATS(TCP_MIB_OUTSEGS);
  1893. #ifdef CONFIG_TCP_MD5SIG
  1894. /* Okay, we have all we need - do the md5 hash if needed */
  1895. if (md5) {
  1896. tp->af_specific->calc_md5_hash(md5_hash_location,
  1897. md5,
  1898. NULL, dst, req,
  1899. tcp_hdr(skb), sk->sk_protocol,
  1900. skb->len);
  1901. }
  1902. #endif
  1903. return skb;
  1904. }
  1905. /*
  1906. * Do all connect socket setups that can be done AF independent.
  1907. */
  1908. static void tcp_connect_init(struct sock *sk)
  1909. {
  1910. struct dst_entry *dst = __sk_dst_get(sk);
  1911. struct tcp_sock *tp = tcp_sk(sk);
  1912. __u8 rcv_wscale;
  1913. /* We'll fix this up when we get a response from the other end.
  1914. * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
  1915. */
  1916. tp->tcp_header_len = sizeof(struct tcphdr) +
  1917. (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
  1918. #ifdef CONFIG_TCP_MD5SIG
  1919. if (tp->af_specific->md5_lookup(sk, sk) != NULL)
  1920. tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
  1921. #endif
  1922. /* If user gave his TCP_MAXSEG, record it to clamp */
  1923. if (tp->rx_opt.user_mss)
  1924. tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
  1925. tp->max_window = 0;
  1926. tcp_mtup_init(sk);
  1927. tcp_sync_mss(sk, dst_mtu(dst));
  1928. if (!tp->window_clamp)
  1929. tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
  1930. tp->advmss = dst_metric(dst, RTAX_ADVMSS);
  1931. tcp_initialize_rcv_mss(sk);
  1932. tcp_select_initial_window(tcp_full_space(sk),
  1933. tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
  1934. &tp->rcv_wnd,
  1935. &tp->window_clamp,
  1936. sysctl_tcp_window_scaling,
  1937. &rcv_wscale);
  1938. tp->rx_opt.rcv_wscale = rcv_wscale;
  1939. tp->rcv_ssthresh = tp->rcv_wnd;
  1940. sk->sk_err = 0;
  1941. sock_reset_flag(sk, SOCK_DONE);
  1942. tp->snd_wnd = 0;
  1943. tcp_init_wl(tp, tp->write_seq, 0);
  1944. tp->snd_una = tp->write_seq;
  1945. tp->snd_sml = tp->write_seq;
  1946. tp->rcv_nxt = 0;
  1947. tp->rcv_wup = 0;
  1948. tp->copied_seq = 0;
  1949. inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
  1950. inet_csk(sk)->icsk_retransmits = 0;
  1951. tcp_clear_retrans(tp);
  1952. }
  1953. /*
  1954. * Build a SYN and send it off.
  1955. */
  1956. int tcp_connect(struct sock *sk)
  1957. {
  1958. struct tcp_sock *tp = tcp_sk(sk);
  1959. struct sk_buff *buff;
  1960. tcp_connect_init(sk);
  1961. buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
  1962. if (unlikely(buff == NULL))
  1963. return -ENOBUFS;
  1964. /* Reserve space for headers. */
  1965. skb_reserve(buff, MAX_TCP_HEADER);
  1966. TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
  1967. TCP_ECN_send_syn(sk, buff);
  1968. TCP_SKB_CB(buff)->sacked = 0;
  1969. skb_shinfo(buff)->gso_segs = 1;
  1970. skb_shinfo(buff)->gso_size = 0;
  1971. skb_shinfo(buff)->gso_type = 0;
  1972. buff->csum = 0;
  1973. tp->snd_nxt = tp->write_seq;
  1974. TCP_SKB_CB(buff)->seq = tp->write_seq++;
  1975. TCP_SKB_CB(buff)->end_seq = tp->write_seq;
  1976. /* Send it off. */
  1977. TCP_SKB_CB(buff)->when = tcp_time_stamp;
  1978. tp->retrans_stamp = TCP_SKB_CB(buff)->when;
  1979. skb_header_release(buff);
  1980. __tcp_add_write_queue_tail(sk, buff);
  1981. sk->sk_wmem_queued += buff->truesize;
  1982. sk_mem_charge(sk, buff->truesize);
  1983. tp->packets_out += tcp_skb_pcount(buff);
  1984. tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
  1985. /* We change tp->snd_nxt after the tcp_transmit_skb() call
  1986. * in order to make this packet get counted in tcpOutSegs.
  1987. */
  1988. tp->snd_nxt = tp->write_seq;
  1989. tp->pushed_seq = tp->write_seq;
  1990. TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
  1991. /* Timer for repeating the SYN until an answer. */
  1992. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  1993. inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  1994. return 0;
  1995. }
  1996. /* Send out a delayed ack, the caller does the policy checking
  1997. * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
  1998. * for details.
  1999. */
  2000. void tcp_send_delayed_ack(struct sock *sk)
  2001. {
  2002. struct inet_connection_sock *icsk = inet_csk(sk);
  2003. int ato = icsk->icsk_ack.ato;
  2004. unsigned long timeout;
  2005. if (ato > TCP_DELACK_MIN) {
  2006. const struct tcp_sock *tp = tcp_sk(sk);
  2007. int max_ato = HZ/2;
  2008. if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
  2009. max_ato = TCP_DELACK_MAX;
  2010. /* Slow path, intersegment interval is "high". */
  2011. /* If some rtt estimate is known, use it to bound delayed ack.
  2012. * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
  2013. * directly.
  2014. */
  2015. if (tp->srtt) {
  2016. int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
  2017. if (rtt < max_ato)
  2018. max_ato = rtt;
  2019. }
  2020. ato = min(ato, max_ato);
  2021. }
  2022. /* Stay within the limit we were given */
  2023. timeout = jiffies + ato;
  2024. /* Use new timeout only if there wasn't a older one earlier. */
  2025. if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
  2026. /* If delack timer was blocked or is about to expire,
  2027. * send ACK now.
  2028. */
  2029. if (icsk->icsk_ack.blocked ||
  2030. time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
  2031. tcp_send_ack(sk);
  2032. return;
  2033. }
  2034. if (!time_before(timeout, icsk->icsk_ack.timeout))
  2035. timeout = icsk->icsk_ack.timeout;
  2036. }
  2037. icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
  2038. icsk->icsk_ack.timeout = timeout;
  2039. sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
  2040. }
  2041. /* This routine sends an ack and also updates the window. */
  2042. void tcp_send_ack(struct sock *sk)
  2043. {
  2044. struct sk_buff *buff;
  2045. /* If we have been reset, we may not send again. */
  2046. if (sk->sk_state == TCP_CLOSE)
  2047. return;
  2048. /* We are not putting this on the write queue, so
  2049. * tcp_transmit_skb() will set the ownership to this
  2050. * sock.
  2051. */
  2052. buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
  2053. if (buff == NULL) {
  2054. inet_csk_schedule_ack(sk);
  2055. inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
  2056. inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
  2057. TCP_DELACK_MAX, TCP_RTO_MAX);
  2058. return;
  2059. }
  2060. /* Reserve space for headers and prepare control bits. */
  2061. skb_reserve(buff, MAX_TCP_HEADER);
  2062. buff->csum = 0;
  2063. TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
  2064. TCP_SKB_CB(buff)->sacked = 0;
  2065. skb_shinfo(buff)->gso_segs = 1;
  2066. skb_shinfo(buff)->gso_size = 0;
  2067. skb_shinfo(buff)->gso_type = 0;
  2068. /* Send it off, this clears delayed acks for us. */
  2069. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk);
  2070. TCP_SKB_CB(buff)->when = tcp_time_stamp;
  2071. tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
  2072. }
  2073. /* This routine sends a packet with an out of date sequence
  2074. * number. It assumes the other end will try to ack it.
  2075. *
  2076. * Question: what should we make while urgent mode?
  2077. * 4.4BSD forces sending single byte of data. We cannot send
  2078. * out of window data, because we have SND.NXT==SND.MAX...
  2079. *
  2080. * Current solution: to send TWO zero-length segments in urgent mode:
  2081. * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
  2082. * out-of-date with SND.UNA-1 to probe window.
  2083. */
  2084. static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
  2085. {
  2086. struct tcp_sock *tp = tcp_sk(sk);
  2087. struct sk_buff *skb;
  2088. /* We don't queue it, tcp_transmit_skb() sets ownership. */
  2089. skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
  2090. if (skb == NULL)
  2091. return -1;
  2092. /* Reserve space for headers and set control bits. */
  2093. skb_reserve(skb, MAX_TCP_HEADER);
  2094. skb->csum = 0;
  2095. TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
  2096. TCP_SKB_CB(skb)->sacked = 0;
  2097. skb_shinfo(skb)->gso_segs = 1;
  2098. skb_shinfo(skb)->gso_size = 0;
  2099. skb_shinfo(skb)->gso_type = 0;
  2100. /* Use a previous sequence. This should cause the other
  2101. * end to send an ack. Don't queue or clone SKB, just
  2102. * send it.
  2103. */
  2104. TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
  2105. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
  2106. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2107. return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
  2108. }
  2109. int tcp_write_wakeup(struct sock *sk)
  2110. {
  2111. struct tcp_sock *tp = tcp_sk(sk);
  2112. struct sk_buff *skb;
  2113. if (sk->sk_state == TCP_CLOSE)
  2114. return -1;
  2115. if ((skb = tcp_send_head(sk)) != NULL &&
  2116. before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
  2117. int err;
  2118. unsigned int mss = tcp_current_mss(sk, 0);
  2119. unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  2120. if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
  2121. tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
  2122. /* We are probing the opening of a window
  2123. * but the window size is != 0
  2124. * must have been a result SWS avoidance ( sender )
  2125. */
  2126. if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
  2127. skb->len > mss) {
  2128. seg_size = min(seg_size, mss);
  2129. TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
  2130. if (tcp_fragment(sk, skb, seg_size, mss))
  2131. return -1;
  2132. } else if (!tcp_skb_pcount(skb))
  2133. tcp_set_skb_tso_segs(sk, skb, mss);
  2134. TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
  2135. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2136. err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2137. if (!err)
  2138. tcp_event_new_data_sent(sk, skb);
  2139. return err;
  2140. } else {
  2141. if (tp->urg_mode &&
  2142. between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
  2143. tcp_xmit_probe_skb(sk, 1);
  2144. return tcp_xmit_probe_skb(sk, 0);
  2145. }
  2146. }
  2147. /* A window probe timeout has occurred. If window is not closed send
  2148. * a partial packet else a zero probe.
  2149. */
  2150. void tcp_send_probe0(struct sock *sk)
  2151. {
  2152. struct inet_connection_sock *icsk = inet_csk(sk);
  2153. struct tcp_sock *tp = tcp_sk(sk);
  2154. int err;
  2155. err = tcp_write_wakeup(sk);
  2156. if (tp->packets_out || !tcp_send_head(sk)) {
  2157. /* Cancel probe timer, if it is not required. */
  2158. icsk->icsk_probes_out = 0;
  2159. icsk->icsk_backoff = 0;
  2160. return;
  2161. }
  2162. if (err <= 0) {
  2163. if (icsk->icsk_backoff < sysctl_tcp_retries2)
  2164. icsk->icsk_backoff++;
  2165. icsk->icsk_probes_out++;
  2166. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2167. min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
  2168. TCP_RTO_MAX);
  2169. } else {
  2170. /* If packet was not sent due to local congestion,
  2171. * do not backoff and do not remember icsk_probes_out.
  2172. * Let local senders to fight for local resources.
  2173. *
  2174. * Use accumulated backoff yet.
  2175. */
  2176. if (!icsk->icsk_probes_out)
  2177. icsk->icsk_probes_out = 1;
  2178. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2179. min(icsk->icsk_rto << icsk->icsk_backoff,
  2180. TCP_RESOURCE_PROBE_INTERVAL),
  2181. TCP_RTO_MAX);
  2182. }
  2183. }
  2184. EXPORT_SYMBOL(tcp_connect);
  2185. EXPORT_SYMBOL(tcp_make_synack);
  2186. EXPORT_SYMBOL(tcp_simple_retransmit);
  2187. EXPORT_SYMBOL(tcp_sync_mss);
  2188. EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
  2189. EXPORT_SYMBOL(tcp_mtup_init);