tcp_output.c 74 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612
  1. /*
  2. * INET An implementation of the TCP/IP protocol suite for the LINUX
  3. * operating system. INET is implemented using the BSD Socket
  4. * interface as the means of communication with the user level.
  5. *
  6. * Implementation of the Transmission Control Protocol(TCP).
  7. *
  8. * Authors: Ross Biro
  9. * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10. * Mark Evans, <evansmp@uhura.aston.ac.uk>
  11. * Corey Minyard <wf-rch!minyard@relay.EU.net>
  12. * Florian La Roche, <flla@stud.uni-sb.de>
  13. * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14. * Linus Torvalds, <torvalds@cs.helsinki.fi>
  15. * Alan Cox, <gw4pts@gw4pts.ampr.org>
  16. * Matthew Dillon, <dillon@apollo.west.oic.com>
  17. * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18. * Jorge Cwik, <jorge@laser.satlink.net>
  19. */
  20. /*
  21. * Changes: Pedro Roque : Retransmit queue handled by TCP.
  22. * : Fragmentation on mtu decrease
  23. * : Segment collapse on retransmit
  24. * : AF independence
  25. *
  26. * Linus Torvalds : send_delayed_ack
  27. * David S. Miller : Charge memory using the right skb
  28. * during syn/ack processing.
  29. * David S. Miller : Output engine completely rewritten.
  30. * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
  31. * Cacophonix Gaul : draft-minshall-nagle-01
  32. * J Hadi Salim : ECN support
  33. *
  34. */
  35. #include <net/tcp.h>
  36. #include <linux/compiler.h>
  37. #include <linux/module.h>
  38. /* People can turn this off for buggy TCP's found in printers etc. */
  39. int sysctl_tcp_retrans_collapse __read_mostly = 1;
  40. /* People can turn this on to work with those rare, broken TCPs that
  41. * interpret the window field as a signed quantity.
  42. */
  43. int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
  44. /* This limits the percentage of the congestion window which we
  45. * will allow a single TSO frame to consume. Building TSO frames
  46. * which are too large can cause TCP streams to be bursty.
  47. */
  48. int sysctl_tcp_tso_win_divisor __read_mostly = 3;
  49. int sysctl_tcp_mtu_probing __read_mostly = 0;
  50. int sysctl_tcp_base_mss __read_mostly = 512;
  51. /* By default, RFC2861 behavior. */
  52. int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
  53. static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  54. {
  55. struct tcp_sock *tp = tcp_sk(sk);
  56. unsigned int prior_packets = tp->packets_out;
  57. tcp_advance_send_head(sk, skb);
  58. tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
  59. /* Don't override Nagle indefinately with F-RTO */
  60. if (tp->frto_counter == 2)
  61. tp->frto_counter = 3;
  62. tp->packets_out += tcp_skb_pcount(skb);
  63. if (!prior_packets)
  64. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  65. inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  66. }
  67. /* SND.NXT, if window was not shrunk.
  68. * If window has been shrunk, what should we make? It is not clear at all.
  69. * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  70. * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  71. * invalid. OK, let's make this for now:
  72. */
  73. static inline __u32 tcp_acceptable_seq(struct sock *sk)
  74. {
  75. struct tcp_sock *tp = tcp_sk(sk);
  76. if (!before(tcp_wnd_end(tp), tp->snd_nxt))
  77. return tp->snd_nxt;
  78. else
  79. return tcp_wnd_end(tp);
  80. }
  81. /* Calculate mss to advertise in SYN segment.
  82. * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
  83. *
  84. * 1. It is independent of path mtu.
  85. * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
  86. * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
  87. * attached devices, because some buggy hosts are confused by
  88. * large MSS.
  89. * 4. We do not make 3, we advertise MSS, calculated from first
  90. * hop device mtu, but allow to raise it to ip_rt_min_advmss.
  91. * This may be overridden via information stored in routing table.
  92. * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
  93. * probably even Jumbo".
  94. */
  95. static __u16 tcp_advertise_mss(struct sock *sk)
  96. {
  97. struct tcp_sock *tp = tcp_sk(sk);
  98. struct dst_entry *dst = __sk_dst_get(sk);
  99. int mss = tp->advmss;
  100. if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
  101. mss = dst_metric(dst, RTAX_ADVMSS);
  102. tp->advmss = mss;
  103. }
  104. return (__u16)mss;
  105. }
  106. /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
  107. * This is the first part of cwnd validation mechanism. */
  108. static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
  109. {
  110. struct tcp_sock *tp = tcp_sk(sk);
  111. s32 delta = tcp_time_stamp - tp->lsndtime;
  112. u32 restart_cwnd = tcp_init_cwnd(tp, dst);
  113. u32 cwnd = tp->snd_cwnd;
  114. tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
  115. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  116. restart_cwnd = min(restart_cwnd, cwnd);
  117. while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
  118. cwnd >>= 1;
  119. tp->snd_cwnd = max(cwnd, restart_cwnd);
  120. tp->snd_cwnd_stamp = tcp_time_stamp;
  121. tp->snd_cwnd_used = 0;
  122. }
  123. static void tcp_event_data_sent(struct tcp_sock *tp,
  124. struct sk_buff *skb, struct sock *sk)
  125. {
  126. struct inet_connection_sock *icsk = inet_csk(sk);
  127. const u32 now = tcp_time_stamp;
  128. if (sysctl_tcp_slow_start_after_idle &&
  129. (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
  130. tcp_cwnd_restart(sk, __sk_dst_get(sk));
  131. tp->lsndtime = now;
  132. /* If it is a reply for ato after last received
  133. * packet, enter pingpong mode.
  134. */
  135. if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
  136. icsk->icsk_ack.pingpong = 1;
  137. }
  138. static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
  139. {
  140. tcp_dec_quickack_mode(sk, pkts);
  141. inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
  142. }
  143. /* Determine a window scaling and initial window to offer.
  144. * Based on the assumption that the given amount of space
  145. * will be offered. Store the results in the tp structure.
  146. * NOTE: for smooth operation initial space offering should
  147. * be a multiple of mss if possible. We assume here that mss >= 1.
  148. * This MUST be enforced by all callers.
  149. */
  150. void tcp_select_initial_window(int __space, __u32 mss,
  151. __u32 *rcv_wnd, __u32 *window_clamp,
  152. int wscale_ok, __u8 *rcv_wscale)
  153. {
  154. unsigned int space = (__space < 0 ? 0 : __space);
  155. /* If no clamp set the clamp to the max possible scaled window */
  156. if (*window_clamp == 0)
  157. (*window_clamp) = (65535 << 14);
  158. space = min(*window_clamp, space);
  159. /* Quantize space offering to a multiple of mss if possible. */
  160. if (space > mss)
  161. space = (space / mss) * mss;
  162. /* NOTE: offering an initial window larger than 32767
  163. * will break some buggy TCP stacks. If the admin tells us
  164. * it is likely we could be speaking with such a buggy stack
  165. * we will truncate our initial window offering to 32K-1
  166. * unless the remote has sent us a window scaling option,
  167. * which we interpret as a sign the remote TCP is not
  168. * misinterpreting the window field as a signed quantity.
  169. */
  170. if (sysctl_tcp_workaround_signed_windows)
  171. (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
  172. else
  173. (*rcv_wnd) = space;
  174. (*rcv_wscale) = 0;
  175. if (wscale_ok) {
  176. /* Set window scaling on max possible window
  177. * See RFC1323 for an explanation of the limit to 14
  178. */
  179. space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
  180. space = min_t(u32, space, *window_clamp);
  181. while (space > 65535 && (*rcv_wscale) < 14) {
  182. space >>= 1;
  183. (*rcv_wscale)++;
  184. }
  185. }
  186. /* Set initial window to value enough for senders,
  187. * following RFC2414. Senders, not following this RFC,
  188. * will be satisfied with 2.
  189. */
  190. if (mss > (1 << *rcv_wscale)) {
  191. int init_cwnd = 4;
  192. if (mss > 1460 * 3)
  193. init_cwnd = 2;
  194. else if (mss > 1460)
  195. init_cwnd = 3;
  196. if (*rcv_wnd > init_cwnd * mss)
  197. *rcv_wnd = init_cwnd * mss;
  198. }
  199. /* Set the clamp no higher than max representable value */
  200. (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
  201. }
  202. /* Chose a new window to advertise, update state in tcp_sock for the
  203. * socket, and return result with RFC1323 scaling applied. The return
  204. * value can be stuffed directly into th->window for an outgoing
  205. * frame.
  206. */
  207. static u16 tcp_select_window(struct sock *sk)
  208. {
  209. struct tcp_sock *tp = tcp_sk(sk);
  210. u32 cur_win = tcp_receive_window(tp);
  211. u32 new_win = __tcp_select_window(sk);
  212. /* Never shrink the offered window */
  213. if (new_win < cur_win) {
  214. /* Danger Will Robinson!
  215. * Don't update rcv_wup/rcv_wnd here or else
  216. * we will not be able to advertise a zero
  217. * window in time. --DaveM
  218. *
  219. * Relax Will Robinson.
  220. */
  221. new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
  222. }
  223. tp->rcv_wnd = new_win;
  224. tp->rcv_wup = tp->rcv_nxt;
  225. /* Make sure we do not exceed the maximum possible
  226. * scaled window.
  227. */
  228. if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
  229. new_win = min(new_win, MAX_TCP_WINDOW);
  230. else
  231. new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
  232. /* RFC1323 scaling applied */
  233. new_win >>= tp->rx_opt.rcv_wscale;
  234. /* If we advertise zero window, disable fast path. */
  235. if (new_win == 0)
  236. tp->pred_flags = 0;
  237. return new_win;
  238. }
  239. static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
  240. {
  241. TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
  242. if (!(tp->ecn_flags & TCP_ECN_OK))
  243. TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
  244. }
  245. static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
  246. {
  247. struct tcp_sock *tp = tcp_sk(sk);
  248. tp->ecn_flags = 0;
  249. if (sysctl_tcp_ecn) {
  250. TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR;
  251. tp->ecn_flags = TCP_ECN_OK;
  252. }
  253. }
  254. static __inline__ void
  255. TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
  256. {
  257. if (inet_rsk(req)->ecn_ok)
  258. th->ece = 1;
  259. }
  260. static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
  261. int tcp_header_len)
  262. {
  263. struct tcp_sock *tp = tcp_sk(sk);
  264. if (tp->ecn_flags & TCP_ECN_OK) {
  265. /* Not-retransmitted data segment: set ECT and inject CWR. */
  266. if (skb->len != tcp_header_len &&
  267. !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
  268. INET_ECN_xmit(sk);
  269. if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
  270. tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
  271. tcp_hdr(skb)->cwr = 1;
  272. skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
  273. }
  274. } else {
  275. /* ACK or retransmitted segment: clear ECT|CE */
  276. INET_ECN_dontxmit(sk);
  277. }
  278. if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
  279. tcp_hdr(skb)->ece = 1;
  280. }
  281. }
  282. /* Constructs common control bits of non-data skb. If SYN/FIN is present,
  283. * auto increment end seqno.
  284. */
  285. static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
  286. {
  287. skb->csum = 0;
  288. TCP_SKB_CB(skb)->flags = flags;
  289. TCP_SKB_CB(skb)->sacked = 0;
  290. skb_shinfo(skb)->gso_segs = 1;
  291. skb_shinfo(skb)->gso_size = 0;
  292. skb_shinfo(skb)->gso_type = 0;
  293. TCP_SKB_CB(skb)->seq = seq;
  294. if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN))
  295. seq++;
  296. TCP_SKB_CB(skb)->end_seq = seq;
  297. }
  298. static inline int tcp_urg_mode(const struct tcp_sock *tp)
  299. {
  300. return tp->snd_una != tp->snd_up;
  301. }
  302. #define OPTION_SACK_ADVERTISE (1 << 0)
  303. #define OPTION_TS (1 << 1)
  304. #define OPTION_MD5 (1 << 2)
  305. struct tcp_out_options {
  306. u8 options; /* bit field of OPTION_* */
  307. u8 ws; /* window scale, 0 to disable */
  308. u8 num_sack_blocks; /* number of SACK blocks to include */
  309. u16 mss; /* 0 to disable */
  310. __u32 tsval, tsecr; /* need to include OPTION_TS */
  311. };
  312. /* Beware: Something in the Internet is very sensitive to the ordering of
  313. * TCP options, we learned this through the hard way, so be careful here.
  314. * Luckily we can at least blame others for their non-compliance but from
  315. * inter-operatibility perspective it seems that we're somewhat stuck with
  316. * the ordering which we have been using if we want to keep working with
  317. * those broken things (not that it currently hurts anybody as there isn't
  318. * particular reason why the ordering would need to be changed).
  319. *
  320. * At least SACK_PERM as the first option is known to lead to a disaster
  321. * (but it may well be that other scenarios fail similarly).
  322. */
  323. static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
  324. const struct tcp_out_options *opts,
  325. __u8 **md5_hash) {
  326. if (unlikely(OPTION_MD5 & opts->options)) {
  327. *ptr++ = htonl((TCPOPT_NOP << 24) |
  328. (TCPOPT_NOP << 16) |
  329. (TCPOPT_MD5SIG << 8) |
  330. TCPOLEN_MD5SIG);
  331. *md5_hash = (__u8 *)ptr;
  332. ptr += 4;
  333. } else {
  334. *md5_hash = NULL;
  335. }
  336. if (unlikely(opts->mss)) {
  337. *ptr++ = htonl((TCPOPT_MSS << 24) |
  338. (TCPOLEN_MSS << 16) |
  339. opts->mss);
  340. }
  341. if (likely(OPTION_TS & opts->options)) {
  342. if (unlikely(OPTION_SACK_ADVERTISE & opts->options)) {
  343. *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
  344. (TCPOLEN_SACK_PERM << 16) |
  345. (TCPOPT_TIMESTAMP << 8) |
  346. TCPOLEN_TIMESTAMP);
  347. } else {
  348. *ptr++ = htonl((TCPOPT_NOP << 24) |
  349. (TCPOPT_NOP << 16) |
  350. (TCPOPT_TIMESTAMP << 8) |
  351. TCPOLEN_TIMESTAMP);
  352. }
  353. *ptr++ = htonl(opts->tsval);
  354. *ptr++ = htonl(opts->tsecr);
  355. }
  356. if (unlikely(OPTION_SACK_ADVERTISE & opts->options &&
  357. !(OPTION_TS & opts->options))) {
  358. *ptr++ = htonl((TCPOPT_NOP << 24) |
  359. (TCPOPT_NOP << 16) |
  360. (TCPOPT_SACK_PERM << 8) |
  361. TCPOLEN_SACK_PERM);
  362. }
  363. if (unlikely(opts->ws)) {
  364. *ptr++ = htonl((TCPOPT_NOP << 24) |
  365. (TCPOPT_WINDOW << 16) |
  366. (TCPOLEN_WINDOW << 8) |
  367. opts->ws);
  368. }
  369. if (unlikely(opts->num_sack_blocks)) {
  370. struct tcp_sack_block *sp = tp->rx_opt.dsack ?
  371. tp->duplicate_sack : tp->selective_acks;
  372. int this_sack;
  373. *ptr++ = htonl((TCPOPT_NOP << 24) |
  374. (TCPOPT_NOP << 16) |
  375. (TCPOPT_SACK << 8) |
  376. (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
  377. TCPOLEN_SACK_PERBLOCK)));
  378. for (this_sack = 0; this_sack < opts->num_sack_blocks;
  379. ++this_sack) {
  380. *ptr++ = htonl(sp[this_sack].start_seq);
  381. *ptr++ = htonl(sp[this_sack].end_seq);
  382. }
  383. if (tp->rx_opt.dsack) {
  384. tp->rx_opt.dsack = 0;
  385. tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks;
  386. }
  387. }
  388. }
  389. static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
  390. struct tcp_out_options *opts,
  391. struct tcp_md5sig_key **md5) {
  392. struct tcp_sock *tp = tcp_sk(sk);
  393. unsigned size = 0;
  394. #ifdef CONFIG_TCP_MD5SIG
  395. *md5 = tp->af_specific->md5_lookup(sk, sk);
  396. if (*md5) {
  397. opts->options |= OPTION_MD5;
  398. size += TCPOLEN_MD5SIG_ALIGNED;
  399. }
  400. #else
  401. *md5 = NULL;
  402. #endif
  403. /* We always get an MSS option. The option bytes which will be seen in
  404. * normal data packets should timestamps be used, must be in the MSS
  405. * advertised. But we subtract them from tp->mss_cache so that
  406. * calculations in tcp_sendmsg are simpler etc. So account for this
  407. * fact here if necessary. If we don't do this correctly, as a
  408. * receiver we won't recognize data packets as being full sized when we
  409. * should, and thus we won't abide by the delayed ACK rules correctly.
  410. * SACKs don't matter, we never delay an ACK when we have any of those
  411. * going out. */
  412. opts->mss = tcp_advertise_mss(sk);
  413. size += TCPOLEN_MSS_ALIGNED;
  414. if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
  415. opts->options |= OPTION_TS;
  416. opts->tsval = TCP_SKB_CB(skb)->when;
  417. opts->tsecr = tp->rx_opt.ts_recent;
  418. size += TCPOLEN_TSTAMP_ALIGNED;
  419. }
  420. if (likely(sysctl_tcp_window_scaling)) {
  421. opts->ws = tp->rx_opt.rcv_wscale;
  422. if (likely(opts->ws))
  423. size += TCPOLEN_WSCALE_ALIGNED;
  424. }
  425. if (likely(sysctl_tcp_sack)) {
  426. opts->options |= OPTION_SACK_ADVERTISE;
  427. if (unlikely(!(OPTION_TS & opts->options)))
  428. size += TCPOLEN_SACKPERM_ALIGNED;
  429. }
  430. return size;
  431. }
  432. static unsigned tcp_synack_options(struct sock *sk,
  433. struct request_sock *req,
  434. unsigned mss, struct sk_buff *skb,
  435. struct tcp_out_options *opts,
  436. struct tcp_md5sig_key **md5) {
  437. unsigned size = 0;
  438. struct inet_request_sock *ireq = inet_rsk(req);
  439. char doing_ts;
  440. #ifdef CONFIG_TCP_MD5SIG
  441. *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
  442. if (*md5) {
  443. opts->options |= OPTION_MD5;
  444. size += TCPOLEN_MD5SIG_ALIGNED;
  445. }
  446. #else
  447. *md5 = NULL;
  448. #endif
  449. /* we can't fit any SACK blocks in a packet with MD5 + TS
  450. options. There was discussion about disabling SACK rather than TS in
  451. order to fit in better with old, buggy kernels, but that was deemed
  452. to be unnecessary. */
  453. doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok);
  454. opts->mss = mss;
  455. size += TCPOLEN_MSS_ALIGNED;
  456. if (likely(ireq->wscale_ok)) {
  457. opts->ws = ireq->rcv_wscale;
  458. if (likely(opts->ws))
  459. size += TCPOLEN_WSCALE_ALIGNED;
  460. }
  461. if (likely(doing_ts)) {
  462. opts->options |= OPTION_TS;
  463. opts->tsval = TCP_SKB_CB(skb)->when;
  464. opts->tsecr = req->ts_recent;
  465. size += TCPOLEN_TSTAMP_ALIGNED;
  466. }
  467. if (likely(ireq->sack_ok)) {
  468. opts->options |= OPTION_SACK_ADVERTISE;
  469. if (unlikely(!doing_ts))
  470. size += TCPOLEN_SACKPERM_ALIGNED;
  471. }
  472. return size;
  473. }
  474. static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
  475. struct tcp_out_options *opts,
  476. struct tcp_md5sig_key **md5) {
  477. struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
  478. struct tcp_sock *tp = tcp_sk(sk);
  479. unsigned size = 0;
  480. #ifdef CONFIG_TCP_MD5SIG
  481. *md5 = tp->af_specific->md5_lookup(sk, sk);
  482. if (unlikely(*md5)) {
  483. opts->options |= OPTION_MD5;
  484. size += TCPOLEN_MD5SIG_ALIGNED;
  485. }
  486. #else
  487. *md5 = NULL;
  488. #endif
  489. if (likely(tp->rx_opt.tstamp_ok)) {
  490. opts->options |= OPTION_TS;
  491. opts->tsval = tcb ? tcb->when : 0;
  492. opts->tsecr = tp->rx_opt.ts_recent;
  493. size += TCPOLEN_TSTAMP_ALIGNED;
  494. }
  495. if (unlikely(tp->rx_opt.eff_sacks)) {
  496. const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
  497. opts->num_sack_blocks =
  498. min_t(unsigned, tp->rx_opt.eff_sacks,
  499. (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
  500. TCPOLEN_SACK_PERBLOCK);
  501. size += TCPOLEN_SACK_BASE_ALIGNED +
  502. opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
  503. }
  504. return size;
  505. }
  506. /* This routine actually transmits TCP packets queued in by
  507. * tcp_do_sendmsg(). This is used by both the initial
  508. * transmission and possible later retransmissions.
  509. * All SKB's seen here are completely headerless. It is our
  510. * job to build the TCP header, and pass the packet down to
  511. * IP so it can do the same plus pass the packet off to the
  512. * device.
  513. *
  514. * We are working here with either a clone of the original
  515. * SKB, or a fresh unique copy made by the retransmit engine.
  516. */
  517. static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
  518. gfp_t gfp_mask)
  519. {
  520. const struct inet_connection_sock *icsk = inet_csk(sk);
  521. struct inet_sock *inet;
  522. struct tcp_sock *tp;
  523. struct tcp_skb_cb *tcb;
  524. struct tcp_out_options opts;
  525. unsigned tcp_options_size, tcp_header_size;
  526. struct tcp_md5sig_key *md5;
  527. __u8 *md5_hash_location;
  528. struct tcphdr *th;
  529. int err;
  530. BUG_ON(!skb || !tcp_skb_pcount(skb));
  531. /* If congestion control is doing timestamping, we must
  532. * take such a timestamp before we potentially clone/copy.
  533. */
  534. if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
  535. __net_timestamp(skb);
  536. if (likely(clone_it)) {
  537. if (unlikely(skb_cloned(skb)))
  538. skb = pskb_copy(skb, gfp_mask);
  539. else
  540. skb = skb_clone(skb, gfp_mask);
  541. if (unlikely(!skb))
  542. return -ENOBUFS;
  543. }
  544. inet = inet_sk(sk);
  545. tp = tcp_sk(sk);
  546. tcb = TCP_SKB_CB(skb);
  547. memset(&opts, 0, sizeof(opts));
  548. if (unlikely(tcb->flags & TCPCB_FLAG_SYN))
  549. tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
  550. else
  551. tcp_options_size = tcp_established_options(sk, skb, &opts,
  552. &md5);
  553. tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
  554. if (tcp_packets_in_flight(tp) == 0)
  555. tcp_ca_event(sk, CA_EVENT_TX_START);
  556. skb_push(skb, tcp_header_size);
  557. skb_reset_transport_header(skb);
  558. skb_set_owner_w(skb, sk);
  559. /* Build TCP header and checksum it. */
  560. th = tcp_hdr(skb);
  561. th->source = inet->sport;
  562. th->dest = inet->dport;
  563. th->seq = htonl(tcb->seq);
  564. th->ack_seq = htonl(tp->rcv_nxt);
  565. *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
  566. tcb->flags);
  567. if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
  568. /* RFC1323: The window in SYN & SYN/ACK segments
  569. * is never scaled.
  570. */
  571. th->window = htons(min(tp->rcv_wnd, 65535U));
  572. } else {
  573. th->window = htons(tcp_select_window(sk));
  574. }
  575. th->check = 0;
  576. th->urg_ptr = 0;
  577. /* The urg_mode check is necessary during a below snd_una win probe */
  578. if (unlikely(tcp_urg_mode(tp) &&
  579. between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) {
  580. th->urg_ptr = htons(tp->snd_up - tcb->seq);
  581. th->urg = 1;
  582. }
  583. tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
  584. if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0))
  585. TCP_ECN_send(sk, skb, tcp_header_size);
  586. #ifdef CONFIG_TCP_MD5SIG
  587. /* Calculate the MD5 hash, as we have all we need now */
  588. if (md5) {
  589. sk->sk_route_caps &= ~NETIF_F_GSO_MASK;
  590. tp->af_specific->calc_md5_hash(md5_hash_location,
  591. md5, sk, NULL, skb);
  592. }
  593. #endif
  594. icsk->icsk_af_ops->send_check(sk, skb->len, skb);
  595. if (likely(tcb->flags & TCPCB_FLAG_ACK))
  596. tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
  597. if (skb->len != tcp_header_size)
  598. tcp_event_data_sent(tp, skb, sk);
  599. if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
  600. TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
  601. err = icsk->icsk_af_ops->queue_xmit(skb, 0);
  602. if (likely(err <= 0))
  603. return err;
  604. tcp_enter_cwr(sk, 1);
  605. return net_xmit_eval(err);
  606. }
  607. /* This routine just queue's the buffer
  608. *
  609. * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
  610. * otherwise socket can stall.
  611. */
  612. static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
  613. {
  614. struct tcp_sock *tp = tcp_sk(sk);
  615. /* Advance write_seq and place onto the write_queue. */
  616. tp->write_seq = TCP_SKB_CB(skb)->end_seq;
  617. skb_header_release(skb);
  618. tcp_add_write_queue_tail(sk, skb);
  619. sk->sk_wmem_queued += skb->truesize;
  620. sk_mem_charge(sk, skb->truesize);
  621. }
  622. static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
  623. unsigned int mss_now)
  624. {
  625. if (skb->len <= mss_now || !sk_can_gso(sk)) {
  626. /* Avoid the costly divide in the normal
  627. * non-TSO case.
  628. */
  629. skb_shinfo(skb)->gso_segs = 1;
  630. skb_shinfo(skb)->gso_size = 0;
  631. skb_shinfo(skb)->gso_type = 0;
  632. } else {
  633. skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
  634. skb_shinfo(skb)->gso_size = mss_now;
  635. skb_shinfo(skb)->gso_type = sk->sk_gso_type;
  636. }
  637. }
  638. /* When a modification to fackets out becomes necessary, we need to check
  639. * skb is counted to fackets_out or not.
  640. */
  641. static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
  642. int decr)
  643. {
  644. struct tcp_sock *tp = tcp_sk(sk);
  645. if (!tp->sacked_out || tcp_is_reno(tp))
  646. return;
  647. if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
  648. tp->fackets_out -= decr;
  649. }
  650. /* Function to create two new TCP segments. Shrinks the given segment
  651. * to the specified size and appends a new segment with the rest of the
  652. * packet to the list. This won't be called frequently, I hope.
  653. * Remember, these are still headerless SKBs at this point.
  654. */
  655. int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
  656. unsigned int mss_now)
  657. {
  658. struct tcp_sock *tp = tcp_sk(sk);
  659. struct sk_buff *buff;
  660. int nsize, old_factor;
  661. int nlen;
  662. u16 flags;
  663. BUG_ON(len > skb->len);
  664. tcp_clear_retrans_hints_partial(tp);
  665. nsize = skb_headlen(skb) - len;
  666. if (nsize < 0)
  667. nsize = 0;
  668. if (skb_cloned(skb) &&
  669. skb_is_nonlinear(skb) &&
  670. pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
  671. return -ENOMEM;
  672. /* Get a new skb... force flag on. */
  673. buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
  674. if (buff == NULL)
  675. return -ENOMEM; /* We'll just try again later. */
  676. sk->sk_wmem_queued += buff->truesize;
  677. sk_mem_charge(sk, buff->truesize);
  678. nlen = skb->len - len - nsize;
  679. buff->truesize += nlen;
  680. skb->truesize -= nlen;
  681. /* Correct the sequence numbers. */
  682. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  683. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  684. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  685. /* PSH and FIN should only be set in the second packet. */
  686. flags = TCP_SKB_CB(skb)->flags;
  687. TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
  688. TCP_SKB_CB(buff)->flags = flags;
  689. TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
  690. if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
  691. /* Copy and checksum data tail into the new buffer. */
  692. buff->csum = csum_partial_copy_nocheck(skb->data + len,
  693. skb_put(buff, nsize),
  694. nsize, 0);
  695. skb_trim(skb, len);
  696. skb->csum = csum_block_sub(skb->csum, buff->csum, len);
  697. } else {
  698. skb->ip_summed = CHECKSUM_PARTIAL;
  699. skb_split(skb, buff, len);
  700. }
  701. buff->ip_summed = skb->ip_summed;
  702. /* Looks stupid, but our code really uses when of
  703. * skbs, which it never sent before. --ANK
  704. */
  705. TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
  706. buff->tstamp = skb->tstamp;
  707. old_factor = tcp_skb_pcount(skb);
  708. /* Fix up tso_factor for both original and new SKB. */
  709. tcp_set_skb_tso_segs(sk, skb, mss_now);
  710. tcp_set_skb_tso_segs(sk, buff, mss_now);
  711. /* If this packet has been sent out already, we must
  712. * adjust the various packet counters.
  713. */
  714. if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
  715. int diff = old_factor - tcp_skb_pcount(skb) -
  716. tcp_skb_pcount(buff);
  717. tp->packets_out -= diff;
  718. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
  719. tp->sacked_out -= diff;
  720. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
  721. tp->retrans_out -= diff;
  722. if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
  723. tp->lost_out -= diff;
  724. /* Adjust Reno SACK estimate. */
  725. if (tcp_is_reno(tp) && diff > 0) {
  726. tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
  727. tcp_verify_left_out(tp);
  728. }
  729. tcp_adjust_fackets_out(sk, skb, diff);
  730. }
  731. /* Link BUFF into the send queue. */
  732. skb_header_release(buff);
  733. tcp_insert_write_queue_after(skb, buff, sk);
  734. return 0;
  735. }
  736. /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
  737. * eventually). The difference is that pulled data not copied, but
  738. * immediately discarded.
  739. */
  740. static void __pskb_trim_head(struct sk_buff *skb, int len)
  741. {
  742. int i, k, eat;
  743. eat = len;
  744. k = 0;
  745. for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  746. if (skb_shinfo(skb)->frags[i].size <= eat) {
  747. put_page(skb_shinfo(skb)->frags[i].page);
  748. eat -= skb_shinfo(skb)->frags[i].size;
  749. } else {
  750. skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
  751. if (eat) {
  752. skb_shinfo(skb)->frags[k].page_offset += eat;
  753. skb_shinfo(skb)->frags[k].size -= eat;
  754. eat = 0;
  755. }
  756. k++;
  757. }
  758. }
  759. skb_shinfo(skb)->nr_frags = k;
  760. skb_reset_tail_pointer(skb);
  761. skb->data_len -= len;
  762. skb->len = skb->data_len;
  763. }
  764. int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
  765. {
  766. if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
  767. return -ENOMEM;
  768. /* If len == headlen, we avoid __skb_pull to preserve alignment. */
  769. if (unlikely(len < skb_headlen(skb)))
  770. __skb_pull(skb, len);
  771. else
  772. __pskb_trim_head(skb, len - skb_headlen(skb));
  773. TCP_SKB_CB(skb)->seq += len;
  774. skb->ip_summed = CHECKSUM_PARTIAL;
  775. skb->truesize -= len;
  776. sk->sk_wmem_queued -= len;
  777. sk_mem_uncharge(sk, len);
  778. sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
  779. /* Any change of skb->len requires recalculation of tso
  780. * factor and mss.
  781. */
  782. if (tcp_skb_pcount(skb) > 1)
  783. tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
  784. return 0;
  785. }
  786. /* Not accounting for SACKs here. */
  787. int tcp_mtu_to_mss(struct sock *sk, int pmtu)
  788. {
  789. struct tcp_sock *tp = tcp_sk(sk);
  790. struct inet_connection_sock *icsk = inet_csk(sk);
  791. int mss_now;
  792. /* Calculate base mss without TCP options:
  793. It is MMS_S - sizeof(tcphdr) of rfc1122
  794. */
  795. mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
  796. /* Clamp it (mss_clamp does not include tcp options) */
  797. if (mss_now > tp->rx_opt.mss_clamp)
  798. mss_now = tp->rx_opt.mss_clamp;
  799. /* Now subtract optional transport overhead */
  800. mss_now -= icsk->icsk_ext_hdr_len;
  801. /* Then reserve room for full set of TCP options and 8 bytes of data */
  802. if (mss_now < 48)
  803. mss_now = 48;
  804. /* Now subtract TCP options size, not including SACKs */
  805. mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
  806. return mss_now;
  807. }
  808. /* Inverse of above */
  809. int tcp_mss_to_mtu(struct sock *sk, int mss)
  810. {
  811. struct tcp_sock *tp = tcp_sk(sk);
  812. struct inet_connection_sock *icsk = inet_csk(sk);
  813. int mtu;
  814. mtu = mss +
  815. tp->tcp_header_len +
  816. icsk->icsk_ext_hdr_len +
  817. icsk->icsk_af_ops->net_header_len;
  818. return mtu;
  819. }
  820. void tcp_mtup_init(struct sock *sk)
  821. {
  822. struct tcp_sock *tp = tcp_sk(sk);
  823. struct inet_connection_sock *icsk = inet_csk(sk);
  824. icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
  825. icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
  826. icsk->icsk_af_ops->net_header_len;
  827. icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
  828. icsk->icsk_mtup.probe_size = 0;
  829. }
  830. /* Bound MSS / TSO packet size with the half of the window */
  831. static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
  832. {
  833. if (tp->max_window && pktsize > (tp->max_window >> 1))
  834. return max(tp->max_window >> 1, 68U - tp->tcp_header_len);
  835. else
  836. return pktsize;
  837. }
  838. /* This function synchronize snd mss to current pmtu/exthdr set.
  839. tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
  840. for TCP options, but includes only bare TCP header.
  841. tp->rx_opt.mss_clamp is mss negotiated at connection setup.
  842. It is minimum of user_mss and mss received with SYN.
  843. It also does not include TCP options.
  844. inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
  845. tp->mss_cache is current effective sending mss, including
  846. all tcp options except for SACKs. It is evaluated,
  847. taking into account current pmtu, but never exceeds
  848. tp->rx_opt.mss_clamp.
  849. NOTE1. rfc1122 clearly states that advertised MSS
  850. DOES NOT include either tcp or ip options.
  851. NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
  852. are READ ONLY outside this function. --ANK (980731)
  853. */
  854. unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
  855. {
  856. struct tcp_sock *tp = tcp_sk(sk);
  857. struct inet_connection_sock *icsk = inet_csk(sk);
  858. int mss_now;
  859. if (icsk->icsk_mtup.search_high > pmtu)
  860. icsk->icsk_mtup.search_high = pmtu;
  861. mss_now = tcp_mtu_to_mss(sk, pmtu);
  862. mss_now = tcp_bound_to_half_wnd(tp, mss_now);
  863. /* And store cached results */
  864. icsk->icsk_pmtu_cookie = pmtu;
  865. if (icsk->icsk_mtup.enabled)
  866. mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
  867. tp->mss_cache = mss_now;
  868. return mss_now;
  869. }
  870. /* Compute the current effective MSS, taking SACKs and IP options,
  871. * and even PMTU discovery events into account.
  872. *
  873. * LARGESEND note: !tcp_urg_mode is overkill, only frames up to snd_up
  874. * cannot be large. However, taking into account rare use of URG, this
  875. * is not a big flaw.
  876. */
  877. unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
  878. {
  879. struct tcp_sock *tp = tcp_sk(sk);
  880. struct dst_entry *dst = __sk_dst_get(sk);
  881. u32 mss_now;
  882. u16 xmit_size_goal;
  883. int doing_tso = 0;
  884. unsigned header_len;
  885. struct tcp_out_options opts;
  886. struct tcp_md5sig_key *md5;
  887. mss_now = tp->mss_cache;
  888. if (large_allowed && sk_can_gso(sk) && !tcp_urg_mode(tp))
  889. doing_tso = 1;
  890. if (dst) {
  891. u32 mtu = dst_mtu(dst);
  892. if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
  893. mss_now = tcp_sync_mss(sk, mtu);
  894. }
  895. header_len = tcp_established_options(sk, NULL, &opts, &md5) +
  896. sizeof(struct tcphdr);
  897. /* The mss_cache is sized based on tp->tcp_header_len, which assumes
  898. * some common options. If this is an odd packet (because we have SACK
  899. * blocks etc) then our calculated header_len will be different, and
  900. * we have to adjust mss_now correspondingly */
  901. if (header_len != tp->tcp_header_len) {
  902. int delta = (int) header_len - tp->tcp_header_len;
  903. mss_now -= delta;
  904. }
  905. xmit_size_goal = mss_now;
  906. if (doing_tso) {
  907. xmit_size_goal = ((sk->sk_gso_max_size - 1) -
  908. inet_csk(sk)->icsk_af_ops->net_header_len -
  909. inet_csk(sk)->icsk_ext_hdr_len -
  910. tp->tcp_header_len);
  911. xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
  912. xmit_size_goal -= (xmit_size_goal % mss_now);
  913. }
  914. tp->xmit_size_goal = xmit_size_goal;
  915. return mss_now;
  916. }
  917. /* Congestion window validation. (RFC2861) */
  918. static void tcp_cwnd_validate(struct sock *sk)
  919. {
  920. struct tcp_sock *tp = tcp_sk(sk);
  921. if (tp->packets_out >= tp->snd_cwnd) {
  922. /* Network is feed fully. */
  923. tp->snd_cwnd_used = 0;
  924. tp->snd_cwnd_stamp = tcp_time_stamp;
  925. } else {
  926. /* Network starves. */
  927. if (tp->packets_out > tp->snd_cwnd_used)
  928. tp->snd_cwnd_used = tp->packets_out;
  929. if (sysctl_tcp_slow_start_after_idle &&
  930. (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
  931. tcp_cwnd_application_limited(sk);
  932. }
  933. }
  934. /* Returns the portion of skb which can be sent right away without
  935. * introducing MSS oddities to segment boundaries. In rare cases where
  936. * mss_now != mss_cache, we will request caller to create a small skb
  937. * per input skb which could be mostly avoided here (if desired).
  938. *
  939. * We explicitly want to create a request for splitting write queue tail
  940. * to a small skb for Nagle purposes while avoiding unnecessary modulos,
  941. * thus all the complexity (cwnd_len is always MSS multiple which we
  942. * return whenever allowed by the other factors). Basically we need the
  943. * modulo only when the receiver window alone is the limiting factor or
  944. * when we would be allowed to send the split-due-to-Nagle skb fully.
  945. */
  946. static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
  947. unsigned int mss_now, unsigned int cwnd)
  948. {
  949. struct tcp_sock *tp = tcp_sk(sk);
  950. u32 needed, window, cwnd_len;
  951. window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  952. cwnd_len = mss_now * cwnd;
  953. if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
  954. return cwnd_len;
  955. needed = min(skb->len, window);
  956. if (cwnd_len <= needed)
  957. return cwnd_len;
  958. return needed - needed % mss_now;
  959. }
  960. /* Can at least one segment of SKB be sent right now, according to the
  961. * congestion window rules? If so, return how many segments are allowed.
  962. */
  963. static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
  964. struct sk_buff *skb)
  965. {
  966. u32 in_flight, cwnd;
  967. /* Don't be strict about the congestion window for the final FIN. */
  968. if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
  969. tcp_skb_pcount(skb) == 1)
  970. return 1;
  971. in_flight = tcp_packets_in_flight(tp);
  972. cwnd = tp->snd_cwnd;
  973. if (in_flight < cwnd)
  974. return (cwnd - in_flight);
  975. return 0;
  976. }
  977. /* This must be invoked the first time we consider transmitting
  978. * SKB onto the wire.
  979. */
  980. static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
  981. unsigned int mss_now)
  982. {
  983. int tso_segs = tcp_skb_pcount(skb);
  984. if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
  985. tcp_set_skb_tso_segs(sk, skb, mss_now);
  986. tso_segs = tcp_skb_pcount(skb);
  987. }
  988. return tso_segs;
  989. }
  990. static inline int tcp_minshall_check(const struct tcp_sock *tp)
  991. {
  992. return after(tp->snd_sml, tp->snd_una) &&
  993. !after(tp->snd_sml, tp->snd_nxt);
  994. }
  995. /* Return 0, if packet can be sent now without violation Nagle's rules:
  996. * 1. It is full sized.
  997. * 2. Or it contains FIN. (already checked by caller)
  998. * 3. Or TCP_NODELAY was set.
  999. * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
  1000. * With Minshall's modification: all sent small packets are ACKed.
  1001. */
  1002. static inline int tcp_nagle_check(const struct tcp_sock *tp,
  1003. const struct sk_buff *skb,
  1004. unsigned mss_now, int nonagle)
  1005. {
  1006. return (skb->len < mss_now &&
  1007. ((nonagle & TCP_NAGLE_CORK) ||
  1008. (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
  1009. }
  1010. /* Return non-zero if the Nagle test allows this packet to be
  1011. * sent now.
  1012. */
  1013. static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
  1014. unsigned int cur_mss, int nonagle)
  1015. {
  1016. /* Nagle rule does not apply to frames, which sit in the middle of the
  1017. * write_queue (they have no chances to get new data).
  1018. *
  1019. * This is implemented in the callers, where they modify the 'nonagle'
  1020. * argument based upon the location of SKB in the send queue.
  1021. */
  1022. if (nonagle & TCP_NAGLE_PUSH)
  1023. return 1;
  1024. /* Don't use the nagle rule for urgent data (or for the final FIN).
  1025. * Nagle can be ignored during F-RTO too (see RFC4138).
  1026. */
  1027. if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
  1028. (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
  1029. return 1;
  1030. if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
  1031. return 1;
  1032. return 0;
  1033. }
  1034. /* Does at least the first segment of SKB fit into the send window? */
  1035. static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
  1036. unsigned int cur_mss)
  1037. {
  1038. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  1039. if (skb->len > cur_mss)
  1040. end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
  1041. return !after(end_seq, tcp_wnd_end(tp));
  1042. }
  1043. /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
  1044. * should be put on the wire right now. If so, it returns the number of
  1045. * packets allowed by the congestion window.
  1046. */
  1047. static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
  1048. unsigned int cur_mss, int nonagle)
  1049. {
  1050. struct tcp_sock *tp = tcp_sk(sk);
  1051. unsigned int cwnd_quota;
  1052. tcp_init_tso_segs(sk, skb, cur_mss);
  1053. if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
  1054. return 0;
  1055. cwnd_quota = tcp_cwnd_test(tp, skb);
  1056. if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
  1057. cwnd_quota = 0;
  1058. return cwnd_quota;
  1059. }
  1060. int tcp_may_send_now(struct sock *sk)
  1061. {
  1062. struct tcp_sock *tp = tcp_sk(sk);
  1063. struct sk_buff *skb = tcp_send_head(sk);
  1064. return (skb &&
  1065. tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
  1066. (tcp_skb_is_last(sk, skb) ?
  1067. tp->nonagle : TCP_NAGLE_PUSH)));
  1068. }
  1069. /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
  1070. * which is put after SKB on the list. It is very much like
  1071. * tcp_fragment() except that it may make several kinds of assumptions
  1072. * in order to speed up the splitting operation. In particular, we
  1073. * know that all the data is in scatter-gather pages, and that the
  1074. * packet has never been sent out before (and thus is not cloned).
  1075. */
  1076. static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
  1077. unsigned int mss_now)
  1078. {
  1079. struct sk_buff *buff;
  1080. int nlen = skb->len - len;
  1081. u16 flags;
  1082. /* All of a TSO frame must be composed of paged data. */
  1083. if (skb->len != skb->data_len)
  1084. return tcp_fragment(sk, skb, len, mss_now);
  1085. buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
  1086. if (unlikely(buff == NULL))
  1087. return -ENOMEM;
  1088. sk->sk_wmem_queued += buff->truesize;
  1089. sk_mem_charge(sk, buff->truesize);
  1090. buff->truesize += nlen;
  1091. skb->truesize -= nlen;
  1092. /* Correct the sequence numbers. */
  1093. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  1094. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  1095. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  1096. /* PSH and FIN should only be set in the second packet. */
  1097. flags = TCP_SKB_CB(skb)->flags;
  1098. TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
  1099. TCP_SKB_CB(buff)->flags = flags;
  1100. /* This packet was never sent out yet, so no SACK bits. */
  1101. TCP_SKB_CB(buff)->sacked = 0;
  1102. buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
  1103. skb_split(skb, buff, len);
  1104. /* Fix up tso_factor for both original and new SKB. */
  1105. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1106. tcp_set_skb_tso_segs(sk, buff, mss_now);
  1107. /* Link BUFF into the send queue. */
  1108. skb_header_release(buff);
  1109. tcp_insert_write_queue_after(skb, buff, sk);
  1110. return 0;
  1111. }
  1112. /* Try to defer sending, if possible, in order to minimize the amount
  1113. * of TSO splitting we do. View it as a kind of TSO Nagle test.
  1114. *
  1115. * This algorithm is from John Heffner.
  1116. */
  1117. static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
  1118. {
  1119. struct tcp_sock *tp = tcp_sk(sk);
  1120. const struct inet_connection_sock *icsk = inet_csk(sk);
  1121. u32 send_win, cong_win, limit, in_flight;
  1122. if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
  1123. goto send_now;
  1124. if (icsk->icsk_ca_state != TCP_CA_Open)
  1125. goto send_now;
  1126. /* Defer for less than two clock ticks. */
  1127. if (tp->tso_deferred &&
  1128. ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
  1129. goto send_now;
  1130. in_flight = tcp_packets_in_flight(tp);
  1131. BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
  1132. send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  1133. /* From in_flight test above, we know that cwnd > in_flight. */
  1134. cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
  1135. limit = min(send_win, cong_win);
  1136. /* If a full-sized TSO skb can be sent, do it. */
  1137. if (limit >= sk->sk_gso_max_size)
  1138. goto send_now;
  1139. if (sysctl_tcp_tso_win_divisor) {
  1140. u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
  1141. /* If at least some fraction of a window is available,
  1142. * just use it.
  1143. */
  1144. chunk /= sysctl_tcp_tso_win_divisor;
  1145. if (limit >= chunk)
  1146. goto send_now;
  1147. } else {
  1148. /* Different approach, try not to defer past a single
  1149. * ACK. Receiver should ACK every other full sized
  1150. * frame, so if we have space for more than 3 frames
  1151. * then send now.
  1152. */
  1153. if (limit > tcp_max_burst(tp) * tp->mss_cache)
  1154. goto send_now;
  1155. }
  1156. /* Ok, it looks like it is advisable to defer. */
  1157. tp->tso_deferred = 1 | (jiffies << 1);
  1158. return 1;
  1159. send_now:
  1160. tp->tso_deferred = 0;
  1161. return 0;
  1162. }
  1163. /* Create a new MTU probe if we are ready.
  1164. * Returns 0 if we should wait to probe (no cwnd available),
  1165. * 1 if a probe was sent,
  1166. * -1 otherwise
  1167. */
  1168. static int tcp_mtu_probe(struct sock *sk)
  1169. {
  1170. struct tcp_sock *tp = tcp_sk(sk);
  1171. struct inet_connection_sock *icsk = inet_csk(sk);
  1172. struct sk_buff *skb, *nskb, *next;
  1173. int len;
  1174. int probe_size;
  1175. int size_needed;
  1176. int copy;
  1177. int mss_now;
  1178. /* Not currently probing/verifying,
  1179. * not in recovery,
  1180. * have enough cwnd, and
  1181. * not SACKing (the variable headers throw things off) */
  1182. if (!icsk->icsk_mtup.enabled ||
  1183. icsk->icsk_mtup.probe_size ||
  1184. inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
  1185. tp->snd_cwnd < 11 ||
  1186. tp->rx_opt.eff_sacks)
  1187. return -1;
  1188. /* Very simple search strategy: just double the MSS. */
  1189. mss_now = tcp_current_mss(sk, 0);
  1190. probe_size = 2 * tp->mss_cache;
  1191. size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
  1192. if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
  1193. /* TODO: set timer for probe_converge_event */
  1194. return -1;
  1195. }
  1196. /* Have enough data in the send queue to probe? */
  1197. if (tp->write_seq - tp->snd_nxt < size_needed)
  1198. return -1;
  1199. if (tp->snd_wnd < size_needed)
  1200. return -1;
  1201. if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
  1202. return 0;
  1203. /* Do we need to wait to drain cwnd? With none in flight, don't stall */
  1204. if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
  1205. if (!tcp_packets_in_flight(tp))
  1206. return -1;
  1207. else
  1208. return 0;
  1209. }
  1210. /* We're allowed to probe. Build it now. */
  1211. if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
  1212. return -1;
  1213. sk->sk_wmem_queued += nskb->truesize;
  1214. sk_mem_charge(sk, nskb->truesize);
  1215. skb = tcp_send_head(sk);
  1216. TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
  1217. TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
  1218. TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
  1219. TCP_SKB_CB(nskb)->sacked = 0;
  1220. nskb->csum = 0;
  1221. nskb->ip_summed = skb->ip_summed;
  1222. tcp_insert_write_queue_before(nskb, skb, sk);
  1223. len = 0;
  1224. tcp_for_write_queue_from_safe(skb, next, sk) {
  1225. copy = min_t(int, skb->len, probe_size - len);
  1226. if (nskb->ip_summed)
  1227. skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
  1228. else
  1229. nskb->csum = skb_copy_and_csum_bits(skb, 0,
  1230. skb_put(nskb, copy),
  1231. copy, nskb->csum);
  1232. if (skb->len <= copy) {
  1233. /* We've eaten all the data from this skb.
  1234. * Throw it away. */
  1235. TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
  1236. tcp_unlink_write_queue(skb, sk);
  1237. sk_wmem_free_skb(sk, skb);
  1238. } else {
  1239. TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
  1240. ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
  1241. if (!skb_shinfo(skb)->nr_frags) {
  1242. skb_pull(skb, copy);
  1243. if (skb->ip_summed != CHECKSUM_PARTIAL)
  1244. skb->csum = csum_partial(skb->data,
  1245. skb->len, 0);
  1246. } else {
  1247. __pskb_trim_head(skb, copy);
  1248. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1249. }
  1250. TCP_SKB_CB(skb)->seq += copy;
  1251. }
  1252. len += copy;
  1253. if (len >= probe_size)
  1254. break;
  1255. }
  1256. tcp_init_tso_segs(sk, nskb, nskb->len);
  1257. /* We're ready to send. If this fails, the probe will
  1258. * be resegmented into mss-sized pieces by tcp_write_xmit(). */
  1259. TCP_SKB_CB(nskb)->when = tcp_time_stamp;
  1260. if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
  1261. /* Decrement cwnd here because we are sending
  1262. * effectively two packets. */
  1263. tp->snd_cwnd--;
  1264. tcp_event_new_data_sent(sk, nskb);
  1265. icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
  1266. tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
  1267. tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
  1268. return 1;
  1269. }
  1270. return -1;
  1271. }
  1272. /* This routine writes packets to the network. It advances the
  1273. * send_head. This happens as incoming acks open up the remote
  1274. * window for us.
  1275. *
  1276. * Returns 1, if no segments are in flight and we have queued segments, but
  1277. * cannot send anything now because of SWS or another problem.
  1278. */
  1279. static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
  1280. {
  1281. struct tcp_sock *tp = tcp_sk(sk);
  1282. struct sk_buff *skb;
  1283. unsigned int tso_segs, sent_pkts;
  1284. int cwnd_quota;
  1285. int result;
  1286. /* If we are closed, the bytes will have to remain here.
  1287. * In time closedown will finish, we empty the write queue and all
  1288. * will be happy.
  1289. */
  1290. if (unlikely(sk->sk_state == TCP_CLOSE))
  1291. return 0;
  1292. sent_pkts = 0;
  1293. /* Do MTU probing. */
  1294. if ((result = tcp_mtu_probe(sk)) == 0) {
  1295. return 0;
  1296. } else if (result > 0) {
  1297. sent_pkts = 1;
  1298. }
  1299. while ((skb = tcp_send_head(sk))) {
  1300. unsigned int limit;
  1301. tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
  1302. BUG_ON(!tso_segs);
  1303. cwnd_quota = tcp_cwnd_test(tp, skb);
  1304. if (!cwnd_quota)
  1305. break;
  1306. if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
  1307. break;
  1308. if (tso_segs == 1) {
  1309. if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
  1310. (tcp_skb_is_last(sk, skb) ?
  1311. nonagle : TCP_NAGLE_PUSH))))
  1312. break;
  1313. } else {
  1314. if (tcp_tso_should_defer(sk, skb))
  1315. break;
  1316. }
  1317. limit = mss_now;
  1318. if (tso_segs > 1)
  1319. limit = tcp_mss_split_point(sk, skb, mss_now,
  1320. cwnd_quota);
  1321. if (skb->len > limit &&
  1322. unlikely(tso_fragment(sk, skb, limit, mss_now)))
  1323. break;
  1324. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1325. if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
  1326. break;
  1327. /* Advance the send_head. This one is sent out.
  1328. * This call will increment packets_out.
  1329. */
  1330. tcp_event_new_data_sent(sk, skb);
  1331. tcp_minshall_update(tp, mss_now, skb);
  1332. sent_pkts++;
  1333. }
  1334. if (likely(sent_pkts)) {
  1335. tcp_cwnd_validate(sk);
  1336. return 0;
  1337. }
  1338. return !tp->packets_out && tcp_send_head(sk);
  1339. }
  1340. /* Push out any pending frames which were held back due to
  1341. * TCP_CORK or attempt at coalescing tiny packets.
  1342. * The socket must be locked by the caller.
  1343. */
  1344. void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
  1345. int nonagle)
  1346. {
  1347. struct sk_buff *skb = tcp_send_head(sk);
  1348. if (skb) {
  1349. if (tcp_write_xmit(sk, cur_mss, nonagle))
  1350. tcp_check_probe_timer(sk);
  1351. }
  1352. }
  1353. /* Send _single_ skb sitting at the send head. This function requires
  1354. * true push pending frames to setup probe timer etc.
  1355. */
  1356. void tcp_push_one(struct sock *sk, unsigned int mss_now)
  1357. {
  1358. struct sk_buff *skb = tcp_send_head(sk);
  1359. unsigned int tso_segs, cwnd_quota;
  1360. BUG_ON(!skb || skb->len < mss_now);
  1361. tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
  1362. cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
  1363. if (likely(cwnd_quota)) {
  1364. unsigned int limit;
  1365. BUG_ON(!tso_segs);
  1366. limit = mss_now;
  1367. if (tso_segs > 1)
  1368. limit = tcp_mss_split_point(sk, skb, mss_now,
  1369. cwnd_quota);
  1370. if (skb->len > limit &&
  1371. unlikely(tso_fragment(sk, skb, limit, mss_now)))
  1372. return;
  1373. /* Send it out now. */
  1374. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1375. if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
  1376. tcp_event_new_data_sent(sk, skb);
  1377. tcp_cwnd_validate(sk);
  1378. return;
  1379. }
  1380. }
  1381. }
  1382. /* This function returns the amount that we can raise the
  1383. * usable window based on the following constraints
  1384. *
  1385. * 1. The window can never be shrunk once it is offered (RFC 793)
  1386. * 2. We limit memory per socket
  1387. *
  1388. * RFC 1122:
  1389. * "the suggested [SWS] avoidance algorithm for the receiver is to keep
  1390. * RECV.NEXT + RCV.WIN fixed until:
  1391. * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
  1392. *
  1393. * i.e. don't raise the right edge of the window until you can raise
  1394. * it at least MSS bytes.
  1395. *
  1396. * Unfortunately, the recommended algorithm breaks header prediction,
  1397. * since header prediction assumes th->window stays fixed.
  1398. *
  1399. * Strictly speaking, keeping th->window fixed violates the receiver
  1400. * side SWS prevention criteria. The problem is that under this rule
  1401. * a stream of single byte packets will cause the right side of the
  1402. * window to always advance by a single byte.
  1403. *
  1404. * Of course, if the sender implements sender side SWS prevention
  1405. * then this will not be a problem.
  1406. *
  1407. * BSD seems to make the following compromise:
  1408. *
  1409. * If the free space is less than the 1/4 of the maximum
  1410. * space available and the free space is less than 1/2 mss,
  1411. * then set the window to 0.
  1412. * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
  1413. * Otherwise, just prevent the window from shrinking
  1414. * and from being larger than the largest representable value.
  1415. *
  1416. * This prevents incremental opening of the window in the regime
  1417. * where TCP is limited by the speed of the reader side taking
  1418. * data out of the TCP receive queue. It does nothing about
  1419. * those cases where the window is constrained on the sender side
  1420. * because the pipeline is full.
  1421. *
  1422. * BSD also seems to "accidentally" limit itself to windows that are a
  1423. * multiple of MSS, at least until the free space gets quite small.
  1424. * This would appear to be a side effect of the mbuf implementation.
  1425. * Combining these two algorithms results in the observed behavior
  1426. * of having a fixed window size at almost all times.
  1427. *
  1428. * Below we obtain similar behavior by forcing the offered window to
  1429. * a multiple of the mss when it is feasible to do so.
  1430. *
  1431. * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
  1432. * Regular options like TIMESTAMP are taken into account.
  1433. */
  1434. u32 __tcp_select_window(struct sock *sk)
  1435. {
  1436. struct inet_connection_sock *icsk = inet_csk(sk);
  1437. struct tcp_sock *tp = tcp_sk(sk);
  1438. /* MSS for the peer's data. Previous versions used mss_clamp
  1439. * here. I don't know if the value based on our guesses
  1440. * of peer's MSS is better for the performance. It's more correct
  1441. * but may be worse for the performance because of rcv_mss
  1442. * fluctuations. --SAW 1998/11/1
  1443. */
  1444. int mss = icsk->icsk_ack.rcv_mss;
  1445. int free_space = tcp_space(sk);
  1446. int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
  1447. int window;
  1448. if (mss > full_space)
  1449. mss = full_space;
  1450. if (free_space < (full_space >> 1)) {
  1451. icsk->icsk_ack.quick = 0;
  1452. if (tcp_memory_pressure)
  1453. tp->rcv_ssthresh = min(tp->rcv_ssthresh,
  1454. 4U * tp->advmss);
  1455. if (free_space < mss)
  1456. return 0;
  1457. }
  1458. if (free_space > tp->rcv_ssthresh)
  1459. free_space = tp->rcv_ssthresh;
  1460. /* Don't do rounding if we are using window scaling, since the
  1461. * scaled window will not line up with the MSS boundary anyway.
  1462. */
  1463. window = tp->rcv_wnd;
  1464. if (tp->rx_opt.rcv_wscale) {
  1465. window = free_space;
  1466. /* Advertise enough space so that it won't get scaled away.
  1467. * Import case: prevent zero window announcement if
  1468. * 1<<rcv_wscale > mss.
  1469. */
  1470. if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
  1471. window = (((window >> tp->rx_opt.rcv_wscale) + 1)
  1472. << tp->rx_opt.rcv_wscale);
  1473. } else {
  1474. /* Get the largest window that is a nice multiple of mss.
  1475. * Window clamp already applied above.
  1476. * If our current window offering is within 1 mss of the
  1477. * free space we just keep it. This prevents the divide
  1478. * and multiply from happening most of the time.
  1479. * We also don't do any window rounding when the free space
  1480. * is too small.
  1481. */
  1482. if (window <= free_space - mss || window > free_space)
  1483. window = (free_space / mss) * mss;
  1484. else if (mss == full_space &&
  1485. free_space > window + (full_space >> 1))
  1486. window = free_space;
  1487. }
  1488. return window;
  1489. }
  1490. /* Collapses two adjacent SKB's during retransmission. */
  1491. static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
  1492. {
  1493. struct tcp_sock *tp = tcp_sk(sk);
  1494. struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
  1495. int skb_size, next_skb_size;
  1496. u16 flags;
  1497. skb_size = skb->len;
  1498. next_skb_size = next_skb->len;
  1499. flags = TCP_SKB_CB(skb)->flags;
  1500. BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
  1501. tcp_highest_sack_combine(sk, next_skb, skb);
  1502. tcp_unlink_write_queue(next_skb, sk);
  1503. skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
  1504. next_skb_size);
  1505. if (next_skb->ip_summed == CHECKSUM_PARTIAL)
  1506. skb->ip_summed = CHECKSUM_PARTIAL;
  1507. if (skb->ip_summed != CHECKSUM_PARTIAL)
  1508. skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
  1509. /* Update sequence range on original skb. */
  1510. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
  1511. /* Merge over control information. */
  1512. flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
  1513. TCP_SKB_CB(skb)->flags = flags;
  1514. /* All done, get rid of second SKB and account for it so
  1515. * packet counting does not break.
  1516. */
  1517. TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
  1518. if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
  1519. tp->retrans_out -= tcp_skb_pcount(next_skb);
  1520. if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
  1521. tp->lost_out -= tcp_skb_pcount(next_skb);
  1522. /* Reno case is special. Sigh... */
  1523. if (tcp_is_reno(tp) && tp->sacked_out)
  1524. tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
  1525. tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
  1526. tp->packets_out -= tcp_skb_pcount(next_skb);
  1527. /* changed transmit queue under us so clear hints */
  1528. tcp_clear_retrans_hints_partial(tp);
  1529. if (next_skb == tp->retransmit_skb_hint)
  1530. tp->retransmit_skb_hint = skb;
  1531. sk_wmem_free_skb(sk, next_skb);
  1532. }
  1533. static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
  1534. {
  1535. if (tcp_skb_pcount(skb) > 1)
  1536. return 0;
  1537. /* TODO: SACK collapsing could be used to remove this condition */
  1538. if (skb_shinfo(skb)->nr_frags != 0)
  1539. return 0;
  1540. if (skb_cloned(skb))
  1541. return 0;
  1542. if (skb == tcp_send_head(sk))
  1543. return 0;
  1544. /* Some heurestics for collapsing over SACK'd could be invented */
  1545. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
  1546. return 0;
  1547. return 1;
  1548. }
  1549. static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
  1550. int space)
  1551. {
  1552. struct tcp_sock *tp = tcp_sk(sk);
  1553. struct sk_buff *skb = to, *tmp;
  1554. int first = 1;
  1555. if (!sysctl_tcp_retrans_collapse)
  1556. return;
  1557. if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)
  1558. return;
  1559. tcp_for_write_queue_from_safe(skb, tmp, sk) {
  1560. if (!tcp_can_collapse(sk, skb))
  1561. break;
  1562. space -= skb->len;
  1563. if (first) {
  1564. first = 0;
  1565. continue;
  1566. }
  1567. if (space < 0)
  1568. break;
  1569. /* Punt if not enough space exists in the first SKB for
  1570. * the data in the second
  1571. */
  1572. if (skb->len > skb_tailroom(to))
  1573. break;
  1574. if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
  1575. break;
  1576. tcp_collapse_retrans(sk, to);
  1577. }
  1578. }
  1579. /* This retransmits one SKB. Policy decisions and retransmit queue
  1580. * state updates are done by the caller. Returns non-zero if an
  1581. * error occurred which prevented the send.
  1582. */
  1583. int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
  1584. {
  1585. struct tcp_sock *tp = tcp_sk(sk);
  1586. struct inet_connection_sock *icsk = inet_csk(sk);
  1587. unsigned int cur_mss;
  1588. int err;
  1589. /* Inconslusive MTU probe */
  1590. if (icsk->icsk_mtup.probe_size) {
  1591. icsk->icsk_mtup.probe_size = 0;
  1592. }
  1593. /* Do not sent more than we queued. 1/4 is reserved for possible
  1594. * copying overhead: fragmentation, tunneling, mangling etc.
  1595. */
  1596. if (atomic_read(&sk->sk_wmem_alloc) >
  1597. min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
  1598. return -EAGAIN;
  1599. if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
  1600. if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
  1601. BUG();
  1602. if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
  1603. return -ENOMEM;
  1604. }
  1605. if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
  1606. return -EHOSTUNREACH; /* Routing failure or similar. */
  1607. cur_mss = tcp_current_mss(sk, 0);
  1608. /* If receiver has shrunk his window, and skb is out of
  1609. * new window, do not retransmit it. The exception is the
  1610. * case, when window is shrunk to zero. In this case
  1611. * our retransmit serves as a zero window probe.
  1612. */
  1613. if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
  1614. && TCP_SKB_CB(skb)->seq != tp->snd_una)
  1615. return -EAGAIN;
  1616. if (skb->len > cur_mss) {
  1617. if (tcp_fragment(sk, skb, cur_mss, cur_mss))
  1618. return -ENOMEM; /* We'll try again later. */
  1619. }
  1620. tcp_retrans_try_collapse(sk, skb, cur_mss);
  1621. /* Some Solaris stacks overoptimize and ignore the FIN on a
  1622. * retransmit when old data is attached. So strip it off
  1623. * since it is cheap to do so and saves bytes on the network.
  1624. */
  1625. if (skb->len > 0 &&
  1626. (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
  1627. tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
  1628. if (!pskb_trim(skb, 0)) {
  1629. /* Reuse, even though it does some unnecessary work */
  1630. tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
  1631. TCP_SKB_CB(skb)->flags);
  1632. skb->ip_summed = CHECKSUM_NONE;
  1633. }
  1634. }
  1635. /* Make a copy, if the first transmission SKB clone we made
  1636. * is still in somebody's hands, else make a clone.
  1637. */
  1638. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1639. err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  1640. if (err == 0) {
  1641. /* Update global TCP statistics. */
  1642. TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
  1643. tp->total_retrans++;
  1644. #if FASTRETRANS_DEBUG > 0
  1645. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
  1646. if (net_ratelimit())
  1647. printk(KERN_DEBUG "retrans_out leaked.\n");
  1648. }
  1649. #endif
  1650. if (!tp->retrans_out)
  1651. tp->lost_retrans_low = tp->snd_nxt;
  1652. TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
  1653. tp->retrans_out += tcp_skb_pcount(skb);
  1654. /* Save stamp of the first retransmit. */
  1655. if (!tp->retrans_stamp)
  1656. tp->retrans_stamp = TCP_SKB_CB(skb)->when;
  1657. tp->undo_retrans++;
  1658. /* snd_nxt is stored to detect loss of retransmitted segment,
  1659. * see tcp_input.c tcp_sacktag_write_queue().
  1660. */
  1661. TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
  1662. }
  1663. return err;
  1664. }
  1665. static int tcp_can_forward_retransmit(struct sock *sk)
  1666. {
  1667. const struct inet_connection_sock *icsk = inet_csk(sk);
  1668. struct tcp_sock *tp = tcp_sk(sk);
  1669. /* Forward retransmissions are possible only during Recovery. */
  1670. if (icsk->icsk_ca_state != TCP_CA_Recovery)
  1671. return 0;
  1672. /* No forward retransmissions in Reno are possible. */
  1673. if (tcp_is_reno(tp))
  1674. return 0;
  1675. /* Yeah, we have to make difficult choice between forward transmission
  1676. * and retransmission... Both ways have their merits...
  1677. *
  1678. * For now we do not retransmit anything, while we have some new
  1679. * segments to send. In the other cases, follow rule 3 for
  1680. * NextSeg() specified in RFC3517.
  1681. */
  1682. if (tcp_may_send_now(sk))
  1683. return 0;
  1684. return 1;
  1685. }
  1686. /* This gets called after a retransmit timeout, and the initially
  1687. * retransmitted data is acknowledged. It tries to continue
  1688. * resending the rest of the retransmit queue, until either
  1689. * we've sent it all or the congestion window limit is reached.
  1690. * If doing SACK, the first ACK which comes back for a timeout
  1691. * based retransmit packet might feed us FACK information again.
  1692. * If so, we use it to avoid unnecessarily retransmissions.
  1693. */
  1694. void tcp_xmit_retransmit_queue(struct sock *sk)
  1695. {
  1696. const struct inet_connection_sock *icsk = inet_csk(sk);
  1697. struct tcp_sock *tp = tcp_sk(sk);
  1698. struct sk_buff *skb;
  1699. struct sk_buff *hole = NULL;
  1700. u32 last_lost;
  1701. int mib_idx;
  1702. int fwd_rexmitting = 0;
  1703. if (!tp->lost_out)
  1704. tp->retransmit_high = tp->snd_una;
  1705. if (tp->retransmit_skb_hint) {
  1706. skb = tp->retransmit_skb_hint;
  1707. last_lost = TCP_SKB_CB(skb)->end_seq;
  1708. if (after(last_lost, tp->retransmit_high))
  1709. last_lost = tp->retransmit_high;
  1710. } else {
  1711. skb = tcp_write_queue_head(sk);
  1712. last_lost = tp->snd_una;
  1713. }
  1714. /* First pass: retransmit lost packets. */
  1715. tcp_for_write_queue_from(skb, sk) {
  1716. __u8 sacked = TCP_SKB_CB(skb)->sacked;
  1717. if (skb == tcp_send_head(sk))
  1718. break;
  1719. /* we could do better than to assign each time */
  1720. if (hole == NULL)
  1721. tp->retransmit_skb_hint = skb;
  1722. /* Assume this retransmit will generate
  1723. * only one packet for congestion window
  1724. * calculation purposes. This works because
  1725. * tcp_retransmit_skb() will chop up the
  1726. * packet to be MSS sized and all the
  1727. * packet counting works out.
  1728. */
  1729. if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
  1730. return;
  1731. if (fwd_rexmitting) {
  1732. begin_fwd:
  1733. if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
  1734. break;
  1735. mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
  1736. } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
  1737. tp->retransmit_high = last_lost;
  1738. if (!tcp_can_forward_retransmit(sk))
  1739. break;
  1740. /* Backtrack if necessary to non-L'ed skb */
  1741. if (hole != NULL) {
  1742. skb = hole;
  1743. hole = NULL;
  1744. }
  1745. fwd_rexmitting = 1;
  1746. goto begin_fwd;
  1747. } else if (!(sacked & TCPCB_LOST)) {
  1748. if (hole == NULL && !(sacked & TCPCB_SACKED_RETRANS))
  1749. hole = skb;
  1750. continue;
  1751. } else {
  1752. last_lost = TCP_SKB_CB(skb)->end_seq;
  1753. if (icsk->icsk_ca_state != TCP_CA_Loss)
  1754. mib_idx = LINUX_MIB_TCPFASTRETRANS;
  1755. else
  1756. mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
  1757. }
  1758. if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
  1759. continue;
  1760. if (tcp_retransmit_skb(sk, skb))
  1761. return;
  1762. NET_INC_STATS_BH(sock_net(sk), mib_idx);
  1763. if (skb == tcp_write_queue_head(sk))
  1764. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  1765. inet_csk(sk)->icsk_rto,
  1766. TCP_RTO_MAX);
  1767. }
  1768. }
  1769. /* Send a fin. The caller locks the socket for us. This cannot be
  1770. * allowed to fail queueing a FIN frame under any circumstances.
  1771. */
  1772. void tcp_send_fin(struct sock *sk)
  1773. {
  1774. struct tcp_sock *tp = tcp_sk(sk);
  1775. struct sk_buff *skb = tcp_write_queue_tail(sk);
  1776. int mss_now;
  1777. /* Optimization, tack on the FIN if we have a queue of
  1778. * unsent frames. But be careful about outgoing SACKS
  1779. * and IP options.
  1780. */
  1781. mss_now = tcp_current_mss(sk, 1);
  1782. if (tcp_send_head(sk) != NULL) {
  1783. TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
  1784. TCP_SKB_CB(skb)->end_seq++;
  1785. tp->write_seq++;
  1786. } else {
  1787. /* Socket is locked, keep trying until memory is available. */
  1788. for (;;) {
  1789. skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
  1790. if (skb)
  1791. break;
  1792. yield();
  1793. }
  1794. /* Reserve space for headers and prepare control bits. */
  1795. skb_reserve(skb, MAX_TCP_HEADER);
  1796. /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
  1797. tcp_init_nondata_skb(skb, tp->write_seq,
  1798. TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
  1799. tcp_queue_skb(sk, skb);
  1800. }
  1801. __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
  1802. }
  1803. /* We get here when a process closes a file descriptor (either due to
  1804. * an explicit close() or as a byproduct of exit()'ing) and there
  1805. * was unread data in the receive queue. This behavior is recommended
  1806. * by RFC 2525, section 2.17. -DaveM
  1807. */
  1808. void tcp_send_active_reset(struct sock *sk, gfp_t priority)
  1809. {
  1810. struct sk_buff *skb;
  1811. /* NOTE: No TCP options attached and we never retransmit this. */
  1812. skb = alloc_skb(MAX_TCP_HEADER, priority);
  1813. if (!skb) {
  1814. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
  1815. return;
  1816. }
  1817. /* Reserve space for headers and prepare control bits. */
  1818. skb_reserve(skb, MAX_TCP_HEADER);
  1819. tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
  1820. TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
  1821. /* Send it off. */
  1822. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1823. if (tcp_transmit_skb(sk, skb, 0, priority))
  1824. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
  1825. TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
  1826. }
  1827. /* WARNING: This routine must only be called when we have already sent
  1828. * a SYN packet that crossed the incoming SYN that caused this routine
  1829. * to get called. If this assumption fails then the initial rcv_wnd
  1830. * and rcv_wscale values will not be correct.
  1831. */
  1832. int tcp_send_synack(struct sock *sk)
  1833. {
  1834. struct sk_buff *skb;
  1835. skb = tcp_write_queue_head(sk);
  1836. if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
  1837. printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
  1838. return -EFAULT;
  1839. }
  1840. if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
  1841. if (skb_cloned(skb)) {
  1842. struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
  1843. if (nskb == NULL)
  1844. return -ENOMEM;
  1845. tcp_unlink_write_queue(skb, sk);
  1846. skb_header_release(nskb);
  1847. __tcp_add_write_queue_head(sk, nskb);
  1848. sk_wmem_free_skb(sk, skb);
  1849. sk->sk_wmem_queued += nskb->truesize;
  1850. sk_mem_charge(sk, nskb->truesize);
  1851. skb = nskb;
  1852. }
  1853. TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
  1854. TCP_ECN_send_synack(tcp_sk(sk), skb);
  1855. }
  1856. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1857. return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  1858. }
  1859. /*
  1860. * Prepare a SYN-ACK.
  1861. */
  1862. struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
  1863. struct request_sock *req)
  1864. {
  1865. struct inet_request_sock *ireq = inet_rsk(req);
  1866. struct tcp_sock *tp = tcp_sk(sk);
  1867. struct tcphdr *th;
  1868. int tcp_header_size;
  1869. struct tcp_out_options opts;
  1870. struct sk_buff *skb;
  1871. struct tcp_md5sig_key *md5;
  1872. __u8 *md5_hash_location;
  1873. int mss;
  1874. skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
  1875. if (skb == NULL)
  1876. return NULL;
  1877. /* Reserve space for headers. */
  1878. skb_reserve(skb, MAX_TCP_HEADER);
  1879. skb->dst = dst_clone(dst);
  1880. mss = dst_metric(dst, RTAX_ADVMSS);
  1881. if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
  1882. mss = tp->rx_opt.user_mss;
  1883. if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
  1884. __u8 rcv_wscale;
  1885. /* Set this up on the first call only */
  1886. req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
  1887. /* tcp_full_space because it is guaranteed to be the first packet */
  1888. tcp_select_initial_window(tcp_full_space(sk),
  1889. mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
  1890. &req->rcv_wnd,
  1891. &req->window_clamp,
  1892. ireq->wscale_ok,
  1893. &rcv_wscale);
  1894. ireq->rcv_wscale = rcv_wscale;
  1895. }
  1896. memset(&opts, 0, sizeof(opts));
  1897. #ifdef CONFIG_SYN_COOKIES
  1898. if (unlikely(req->cookie_ts))
  1899. TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
  1900. else
  1901. #endif
  1902. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1903. tcp_header_size = tcp_synack_options(sk, req, mss,
  1904. skb, &opts, &md5) +
  1905. sizeof(struct tcphdr);
  1906. skb_push(skb, tcp_header_size);
  1907. skb_reset_transport_header(skb);
  1908. th = tcp_hdr(skb);
  1909. memset(th, 0, sizeof(struct tcphdr));
  1910. th->syn = 1;
  1911. th->ack = 1;
  1912. TCP_ECN_make_synack(req, th);
  1913. th->source = ireq->loc_port;
  1914. th->dest = ireq->rmt_port;
  1915. /* Setting of flags are superfluous here for callers (and ECE is
  1916. * not even correctly set)
  1917. */
  1918. tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
  1919. TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
  1920. th->seq = htonl(TCP_SKB_CB(skb)->seq);
  1921. th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
  1922. /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
  1923. th->window = htons(min(req->rcv_wnd, 65535U));
  1924. tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
  1925. th->doff = (tcp_header_size >> 2);
  1926. TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
  1927. #ifdef CONFIG_TCP_MD5SIG
  1928. /* Okay, we have all we need - do the md5 hash if needed */
  1929. if (md5) {
  1930. tp->af_specific->calc_md5_hash(md5_hash_location,
  1931. md5, NULL, req, skb);
  1932. }
  1933. #endif
  1934. return skb;
  1935. }
  1936. /*
  1937. * Do all connect socket setups that can be done AF independent.
  1938. */
  1939. static void tcp_connect_init(struct sock *sk)
  1940. {
  1941. struct dst_entry *dst = __sk_dst_get(sk);
  1942. struct tcp_sock *tp = tcp_sk(sk);
  1943. __u8 rcv_wscale;
  1944. /* We'll fix this up when we get a response from the other end.
  1945. * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
  1946. */
  1947. tp->tcp_header_len = sizeof(struct tcphdr) +
  1948. (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
  1949. #ifdef CONFIG_TCP_MD5SIG
  1950. if (tp->af_specific->md5_lookup(sk, sk) != NULL)
  1951. tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
  1952. #endif
  1953. /* If user gave his TCP_MAXSEG, record it to clamp */
  1954. if (tp->rx_opt.user_mss)
  1955. tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
  1956. tp->max_window = 0;
  1957. tcp_mtup_init(sk);
  1958. tcp_sync_mss(sk, dst_mtu(dst));
  1959. if (!tp->window_clamp)
  1960. tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
  1961. tp->advmss = dst_metric(dst, RTAX_ADVMSS);
  1962. if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
  1963. tp->advmss = tp->rx_opt.user_mss;
  1964. tcp_initialize_rcv_mss(sk);
  1965. tcp_select_initial_window(tcp_full_space(sk),
  1966. tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
  1967. &tp->rcv_wnd,
  1968. &tp->window_clamp,
  1969. sysctl_tcp_window_scaling,
  1970. &rcv_wscale);
  1971. tp->rx_opt.rcv_wscale = rcv_wscale;
  1972. tp->rcv_ssthresh = tp->rcv_wnd;
  1973. sk->sk_err = 0;
  1974. sock_reset_flag(sk, SOCK_DONE);
  1975. tp->snd_wnd = 0;
  1976. tcp_init_wl(tp, tp->write_seq, 0);
  1977. tp->snd_una = tp->write_seq;
  1978. tp->snd_sml = tp->write_seq;
  1979. tp->snd_up = tp->write_seq;
  1980. tp->rcv_nxt = 0;
  1981. tp->rcv_wup = 0;
  1982. tp->copied_seq = 0;
  1983. inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
  1984. inet_csk(sk)->icsk_retransmits = 0;
  1985. tcp_clear_retrans(tp);
  1986. }
  1987. /*
  1988. * Build a SYN and send it off.
  1989. */
  1990. int tcp_connect(struct sock *sk)
  1991. {
  1992. struct tcp_sock *tp = tcp_sk(sk);
  1993. struct sk_buff *buff;
  1994. tcp_connect_init(sk);
  1995. buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
  1996. if (unlikely(buff == NULL))
  1997. return -ENOBUFS;
  1998. /* Reserve space for headers. */
  1999. skb_reserve(buff, MAX_TCP_HEADER);
  2000. tp->snd_nxt = tp->write_seq;
  2001. tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
  2002. TCP_ECN_send_syn(sk, buff);
  2003. /* Send it off. */
  2004. TCP_SKB_CB(buff)->when = tcp_time_stamp;
  2005. tp->retrans_stamp = TCP_SKB_CB(buff)->when;
  2006. skb_header_release(buff);
  2007. __tcp_add_write_queue_tail(sk, buff);
  2008. sk->sk_wmem_queued += buff->truesize;
  2009. sk_mem_charge(sk, buff->truesize);
  2010. tp->packets_out += tcp_skb_pcount(buff);
  2011. tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
  2012. /* We change tp->snd_nxt after the tcp_transmit_skb() call
  2013. * in order to make this packet get counted in tcpOutSegs.
  2014. */
  2015. tp->snd_nxt = tp->write_seq;
  2016. tp->pushed_seq = tp->write_seq;
  2017. TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
  2018. /* Timer for repeating the SYN until an answer. */
  2019. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  2020. inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  2021. return 0;
  2022. }
  2023. /* Send out a delayed ack, the caller does the policy checking
  2024. * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
  2025. * for details.
  2026. */
  2027. void tcp_send_delayed_ack(struct sock *sk)
  2028. {
  2029. struct inet_connection_sock *icsk = inet_csk(sk);
  2030. int ato = icsk->icsk_ack.ato;
  2031. unsigned long timeout;
  2032. if (ato > TCP_DELACK_MIN) {
  2033. const struct tcp_sock *tp = tcp_sk(sk);
  2034. int max_ato = HZ / 2;
  2035. if (icsk->icsk_ack.pingpong ||
  2036. (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
  2037. max_ato = TCP_DELACK_MAX;
  2038. /* Slow path, intersegment interval is "high". */
  2039. /* If some rtt estimate is known, use it to bound delayed ack.
  2040. * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
  2041. * directly.
  2042. */
  2043. if (tp->srtt) {
  2044. int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
  2045. if (rtt < max_ato)
  2046. max_ato = rtt;
  2047. }
  2048. ato = min(ato, max_ato);
  2049. }
  2050. /* Stay within the limit we were given */
  2051. timeout = jiffies + ato;
  2052. /* Use new timeout only if there wasn't a older one earlier. */
  2053. if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
  2054. /* If delack timer was blocked or is about to expire,
  2055. * send ACK now.
  2056. */
  2057. if (icsk->icsk_ack.blocked ||
  2058. time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
  2059. tcp_send_ack(sk);
  2060. return;
  2061. }
  2062. if (!time_before(timeout, icsk->icsk_ack.timeout))
  2063. timeout = icsk->icsk_ack.timeout;
  2064. }
  2065. icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
  2066. icsk->icsk_ack.timeout = timeout;
  2067. sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
  2068. }
  2069. /* This routine sends an ack and also updates the window. */
  2070. void tcp_send_ack(struct sock *sk)
  2071. {
  2072. struct sk_buff *buff;
  2073. /* If we have been reset, we may not send again. */
  2074. if (sk->sk_state == TCP_CLOSE)
  2075. return;
  2076. /* We are not putting this on the write queue, so
  2077. * tcp_transmit_skb() will set the ownership to this
  2078. * sock.
  2079. */
  2080. buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
  2081. if (buff == NULL) {
  2082. inet_csk_schedule_ack(sk);
  2083. inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
  2084. inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
  2085. TCP_DELACK_MAX, TCP_RTO_MAX);
  2086. return;
  2087. }
  2088. /* Reserve space for headers and prepare control bits. */
  2089. skb_reserve(buff, MAX_TCP_HEADER);
  2090. tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
  2091. /* Send it off, this clears delayed acks for us. */
  2092. TCP_SKB_CB(buff)->when = tcp_time_stamp;
  2093. tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
  2094. }
  2095. /* This routine sends a packet with an out of date sequence
  2096. * number. It assumes the other end will try to ack it.
  2097. *
  2098. * Question: what should we make while urgent mode?
  2099. * 4.4BSD forces sending single byte of data. We cannot send
  2100. * out of window data, because we have SND.NXT==SND.MAX...
  2101. *
  2102. * Current solution: to send TWO zero-length segments in urgent mode:
  2103. * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
  2104. * out-of-date with SND.UNA-1 to probe window.
  2105. */
  2106. static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
  2107. {
  2108. struct tcp_sock *tp = tcp_sk(sk);
  2109. struct sk_buff *skb;
  2110. /* We don't queue it, tcp_transmit_skb() sets ownership. */
  2111. skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
  2112. if (skb == NULL)
  2113. return -1;
  2114. /* Reserve space for headers and set control bits. */
  2115. skb_reserve(skb, MAX_TCP_HEADER);
  2116. /* Use a previous sequence. This should cause the other
  2117. * end to send an ack. Don't queue or clone SKB, just
  2118. * send it.
  2119. */
  2120. tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
  2121. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2122. return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
  2123. }
  2124. int tcp_write_wakeup(struct sock *sk)
  2125. {
  2126. struct tcp_sock *tp = tcp_sk(sk);
  2127. struct sk_buff *skb;
  2128. if (sk->sk_state == TCP_CLOSE)
  2129. return -1;
  2130. if ((skb = tcp_send_head(sk)) != NULL &&
  2131. before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
  2132. int err;
  2133. unsigned int mss = tcp_current_mss(sk, 0);
  2134. unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  2135. if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
  2136. tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
  2137. /* We are probing the opening of a window
  2138. * but the window size is != 0
  2139. * must have been a result SWS avoidance ( sender )
  2140. */
  2141. if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
  2142. skb->len > mss) {
  2143. seg_size = min(seg_size, mss);
  2144. TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
  2145. if (tcp_fragment(sk, skb, seg_size, mss))
  2146. return -1;
  2147. } else if (!tcp_skb_pcount(skb))
  2148. tcp_set_skb_tso_segs(sk, skb, mss);
  2149. TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
  2150. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2151. err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2152. if (!err)
  2153. tcp_event_new_data_sent(sk, skb);
  2154. return err;
  2155. } else {
  2156. if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
  2157. tcp_xmit_probe_skb(sk, 1);
  2158. return tcp_xmit_probe_skb(sk, 0);
  2159. }
  2160. }
  2161. /* A window probe timeout has occurred. If window is not closed send
  2162. * a partial packet else a zero probe.
  2163. */
  2164. void tcp_send_probe0(struct sock *sk)
  2165. {
  2166. struct inet_connection_sock *icsk = inet_csk(sk);
  2167. struct tcp_sock *tp = tcp_sk(sk);
  2168. int err;
  2169. err = tcp_write_wakeup(sk);
  2170. if (tp->packets_out || !tcp_send_head(sk)) {
  2171. /* Cancel probe timer, if it is not required. */
  2172. icsk->icsk_probes_out = 0;
  2173. icsk->icsk_backoff = 0;
  2174. return;
  2175. }
  2176. if (err <= 0) {
  2177. if (icsk->icsk_backoff < sysctl_tcp_retries2)
  2178. icsk->icsk_backoff++;
  2179. icsk->icsk_probes_out++;
  2180. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2181. min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
  2182. TCP_RTO_MAX);
  2183. } else {
  2184. /* If packet was not sent due to local congestion,
  2185. * do not backoff and do not remember icsk_probes_out.
  2186. * Let local senders to fight for local resources.
  2187. *
  2188. * Use accumulated backoff yet.
  2189. */
  2190. if (!icsk->icsk_probes_out)
  2191. icsk->icsk_probes_out = 1;
  2192. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2193. min(icsk->icsk_rto << icsk->icsk_backoff,
  2194. TCP_RESOURCE_PROBE_INTERVAL),
  2195. TCP_RTO_MAX);
  2196. }
  2197. }
  2198. EXPORT_SYMBOL(tcp_select_initial_window);
  2199. EXPORT_SYMBOL(tcp_connect);
  2200. EXPORT_SYMBOL(tcp_make_synack);
  2201. EXPORT_SYMBOL(tcp_simple_retransmit);
  2202. EXPORT_SYMBOL(tcp_sync_mss);
  2203. EXPORT_SYMBOL(tcp_mtup_init);