tcp_output.c 83 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715271627172718271927202721272227232724272527262727272827292730273127322733273427352736273727382739274027412742274327442745274627472748274927502751275227532754275527562757275827592760276127622763276427652766276727682769277027712772277327742775277627772778277927802781278227832784278527862787278827892790279127922793279427952796279727982799280028012802280328042805280628072808280928102811281228132814281528162817281828192820282128222823282428252826282728282829283028312832283328342835283628372838283928402841284228432844284528462847284828492850285128522853285428552856285728582859286028612862286328642865286628672868286928702871287228732874287528762877287828792880288128822883288428852886288728882889289028912892289328942895289628972898289929002901
  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/gfp.h>
  38. #include <linux/module.h>
  39. /* People can turn this off for buggy TCP's found in printers etc. */
  40. int sysctl_tcp_retrans_collapse __read_mostly = 1;
  41. /* People can turn this on to work with those rare, broken TCPs that
  42. * interpret the window field as a signed quantity.
  43. */
  44. int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
  45. /* This limits the percentage of the congestion window which we
  46. * will allow a single TSO frame to consume. Building TSO frames
  47. * which are too large can cause TCP streams to be bursty.
  48. */
  49. int sysctl_tcp_tso_win_divisor __read_mostly = 3;
  50. int sysctl_tcp_mtu_probing __read_mostly = 0;
  51. int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
  52. int sysctl_tcp_min_snd_mss __read_mostly = TCP_MIN_SND_MSS;
  53. /* By default, RFC2861 behavior. */
  54. int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
  55. int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
  56. EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
  57. /* Account for new data that has been sent to the network. */
  58. static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
  59. {
  60. struct tcp_sock *tp = tcp_sk(sk);
  61. unsigned int prior_packets = tp->packets_out;
  62. tcp_advance_send_head(sk, skb);
  63. tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
  64. /* Don't override Nagle indefinitely with F-RTO */
  65. if (tp->frto_counter == 2)
  66. tp->frto_counter = 3;
  67. tp->packets_out += tcp_skb_pcount(skb);
  68. if (!prior_packets)
  69. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  70. inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  71. }
  72. /* SND.NXT, if window was not shrunk.
  73. * If window has been shrunk, what should we make? It is not clear at all.
  74. * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  75. * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  76. * invalid. OK, let's make this for now:
  77. */
  78. static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  79. {
  80. const struct tcp_sock *tp = tcp_sk(sk);
  81. if (!before(tcp_wnd_end(tp), tp->snd_nxt))
  82. return tp->snd_nxt;
  83. else
  84. return tcp_wnd_end(tp);
  85. }
  86. /* Calculate mss to advertise in SYN segment.
  87. * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
  88. *
  89. * 1. It is independent of path mtu.
  90. * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
  91. * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
  92. * attached devices, because some buggy hosts are confused by
  93. * large MSS.
  94. * 4. We do not make 3, we advertise MSS, calculated from first
  95. * hop device mtu, but allow to raise it to ip_rt_min_advmss.
  96. * This may be overridden via information stored in routing table.
  97. * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
  98. * probably even Jumbo".
  99. */
  100. static __u16 tcp_advertise_mss(struct sock *sk)
  101. {
  102. struct tcp_sock *tp = tcp_sk(sk);
  103. const struct dst_entry *dst = __sk_dst_get(sk);
  104. int mss = tp->advmss;
  105. if (dst) {
  106. unsigned int metric = dst_metric_advmss(dst);
  107. if (metric < mss) {
  108. mss = metric;
  109. tp->advmss = mss;
  110. }
  111. }
  112. return (__u16)mss;
  113. }
  114. /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
  115. * This is the first part of cwnd validation mechanism. */
  116. static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
  117. {
  118. struct tcp_sock *tp = tcp_sk(sk);
  119. s32 delta = tcp_time_stamp - tp->lsndtime;
  120. u32 restart_cwnd = tcp_init_cwnd(tp, dst);
  121. u32 cwnd = tp->snd_cwnd;
  122. tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
  123. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  124. restart_cwnd = min(restart_cwnd, cwnd);
  125. while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
  126. cwnd >>= 1;
  127. tp->snd_cwnd = max(cwnd, restart_cwnd);
  128. tp->snd_cwnd_stamp = tcp_time_stamp;
  129. tp->snd_cwnd_used = 0;
  130. }
  131. /* Congestion state accounting after a packet has been sent. */
  132. static void tcp_event_data_sent(struct tcp_sock *tp,
  133. struct sock *sk)
  134. {
  135. struct inet_connection_sock *icsk = inet_csk(sk);
  136. const u32 now = tcp_time_stamp;
  137. if (sysctl_tcp_slow_start_after_idle &&
  138. (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
  139. tcp_cwnd_restart(sk, __sk_dst_get(sk));
  140. tp->lsndtime = now;
  141. /* If it is a reply for ato after last received
  142. * packet, enter pingpong mode.
  143. */
  144. if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
  145. icsk->icsk_ack.pingpong = 1;
  146. }
  147. /* Account for an ACK we sent. */
  148. static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
  149. {
  150. tcp_dec_quickack_mode(sk, pkts);
  151. inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
  152. }
  153. /* Determine a window scaling and initial window to offer.
  154. * Based on the assumption that the given amount of space
  155. * will be offered. Store the results in the tp structure.
  156. * NOTE: for smooth operation initial space offering should
  157. * be a multiple of mss if possible. We assume here that mss >= 1.
  158. * This MUST be enforced by all callers.
  159. */
  160. void tcp_select_initial_window(int __space, __u32 mss,
  161. __u32 *rcv_wnd, __u32 *window_clamp,
  162. int wscale_ok, __u8 *rcv_wscale,
  163. __u32 init_rcv_wnd)
  164. {
  165. unsigned int space = (__space < 0 ? 0 : __space);
  166. /* If no clamp set the clamp to the max possible scaled window */
  167. if (*window_clamp == 0)
  168. (*window_clamp) = (65535 << 14);
  169. space = min(*window_clamp, space);
  170. /* Quantize space offering to a multiple of mss if possible. */
  171. if (space > mss)
  172. space = (space / mss) * mss;
  173. /* NOTE: offering an initial window larger than 32767
  174. * will break some buggy TCP stacks. If the admin tells us
  175. * it is likely we could be speaking with such a buggy stack
  176. * we will truncate our initial window offering to 32K-1
  177. * unless the remote has sent us a window scaling option,
  178. * which we interpret as a sign the remote TCP is not
  179. * misinterpreting the window field as a signed quantity.
  180. */
  181. if (sysctl_tcp_workaround_signed_windows)
  182. (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
  183. else
  184. (*rcv_wnd) = space;
  185. (*rcv_wscale) = 0;
  186. if (wscale_ok) {
  187. /* Set window scaling on max possible window
  188. * See RFC1323 for an explanation of the limit to 14
  189. */
  190. space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
  191. space = min_t(u32, space, *window_clamp);
  192. while (space > 65535 && (*rcv_wscale) < 14) {
  193. space >>= 1;
  194. (*rcv_wscale)++;
  195. }
  196. }
  197. /* Set initial window to a value enough for senders starting with
  198. * initial congestion window of sysctl_tcp_default_init_rwnd. Place
  199. * a limit on the initial window when mss is larger than 1460.
  200. */
  201. if (mss > (1 << *rcv_wscale)) {
  202. int init_cwnd = sysctl_tcp_default_init_rwnd;
  203. if (mss > 1460)
  204. init_cwnd = max_t(u32, (1460 * init_cwnd) / mss, 2);
  205. /* when initializing use the value from init_rcv_wnd
  206. * rather than the default from above
  207. */
  208. if (init_rcv_wnd)
  209. *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
  210. else
  211. *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
  212. }
  213. /* Set the clamp no higher than max representable value */
  214. (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
  215. }
  216. EXPORT_SYMBOL(tcp_select_initial_window);
  217. /* Chose a new window to advertise, update state in tcp_sock for the
  218. * socket, and return result with RFC1323 scaling applied. The return
  219. * value can be stuffed directly into th->window for an outgoing
  220. * frame.
  221. */
  222. static u16 tcp_select_window(struct sock *sk)
  223. {
  224. struct tcp_sock *tp = tcp_sk(sk);
  225. u32 cur_win = tcp_receive_window(tp);
  226. u32 new_win = __tcp_select_window(sk);
  227. /* Never shrink the offered window */
  228. if (new_win < cur_win) {
  229. /* Danger Will Robinson!
  230. * Don't update rcv_wup/rcv_wnd here or else
  231. * we will not be able to advertise a zero
  232. * window in time. --DaveM
  233. *
  234. * Relax Will Robinson.
  235. */
  236. new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
  237. }
  238. tp->rcv_wnd = new_win;
  239. tp->rcv_wup = tp->rcv_nxt;
  240. /* Make sure we do not exceed the maximum possible
  241. * scaled window.
  242. */
  243. if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
  244. new_win = min(new_win, MAX_TCP_WINDOW);
  245. else
  246. new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
  247. /* RFC1323 scaling applied */
  248. new_win >>= tp->rx_opt.rcv_wscale;
  249. /* If we advertise zero window, disable fast path. */
  250. if (new_win == 0)
  251. tp->pred_flags = 0;
  252. return new_win;
  253. }
  254. /* Packet ECN state for a SYN-ACK */
  255. static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
  256. {
  257. TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
  258. if (!(tp->ecn_flags & TCP_ECN_OK))
  259. TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
  260. }
  261. /* Packet ECN state for a SYN. */
  262. static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
  263. {
  264. struct tcp_sock *tp = tcp_sk(sk);
  265. tp->ecn_flags = 0;
  266. if (sysctl_tcp_ecn == 1) {
  267. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
  268. tp->ecn_flags = TCP_ECN_OK;
  269. }
  270. }
  271. static __inline__ void
  272. TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
  273. {
  274. if (inet_rsk(req)->ecn_ok)
  275. th->ece = 1;
  276. }
  277. /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
  278. * be sent.
  279. */
  280. static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
  281. int tcp_header_len)
  282. {
  283. struct tcp_sock *tp = tcp_sk(sk);
  284. if (tp->ecn_flags & TCP_ECN_OK) {
  285. /* Not-retransmitted data segment: set ECT and inject CWR. */
  286. if (skb->len != tcp_header_len &&
  287. !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
  288. INET_ECN_xmit(sk);
  289. if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
  290. tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
  291. tcp_hdr(skb)->cwr = 1;
  292. skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
  293. }
  294. } else {
  295. /* ACK or retransmitted segment: clear ECT|CE */
  296. INET_ECN_dontxmit(sk);
  297. }
  298. if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
  299. tcp_hdr(skb)->ece = 1;
  300. }
  301. }
  302. /* Constructs common control bits of non-data skb. If SYN/FIN is present,
  303. * auto increment end seqno.
  304. */
  305. static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
  306. {
  307. skb->ip_summed = CHECKSUM_PARTIAL;
  308. skb->csum = 0;
  309. TCP_SKB_CB(skb)->tcp_flags = flags;
  310. TCP_SKB_CB(skb)->sacked = 0;
  311. skb_shinfo(skb)->gso_segs = 1;
  312. skb_shinfo(skb)->gso_size = 0;
  313. skb_shinfo(skb)->gso_type = 0;
  314. TCP_SKB_CB(skb)->seq = seq;
  315. if (flags & (TCPHDR_SYN | TCPHDR_FIN))
  316. seq++;
  317. TCP_SKB_CB(skb)->end_seq = seq;
  318. }
  319. static inline int tcp_urg_mode(const struct tcp_sock *tp)
  320. {
  321. return tp->snd_una != tp->snd_up;
  322. }
  323. #define OPTION_SACK_ADVERTISE (1 << 0)
  324. #define OPTION_TS (1 << 1)
  325. #define OPTION_MD5 (1 << 2)
  326. #define OPTION_WSCALE (1 << 3)
  327. #define OPTION_COOKIE_EXTENSION (1 << 4)
  328. struct tcp_out_options {
  329. u8 options; /* bit field of OPTION_* */
  330. u8 ws; /* window scale, 0 to disable */
  331. u8 num_sack_blocks; /* number of SACK blocks to include */
  332. u8 hash_size; /* bytes in hash_location */
  333. u16 mss; /* 0 to disable */
  334. __u32 tsval, tsecr; /* need to include OPTION_TS */
  335. __u8 *hash_location; /* temporary pointer, overloaded */
  336. };
  337. /* The sysctl int routines are generic, so check consistency here.
  338. */
  339. static u8 tcp_cookie_size_check(u8 desired)
  340. {
  341. int cookie_size;
  342. if (desired > 0)
  343. /* previously specified */
  344. return desired;
  345. cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
  346. if (cookie_size <= 0)
  347. /* no default specified */
  348. return 0;
  349. if (cookie_size <= TCP_COOKIE_MIN)
  350. /* value too small, specify minimum */
  351. return TCP_COOKIE_MIN;
  352. if (cookie_size >= TCP_COOKIE_MAX)
  353. /* value too large, specify maximum */
  354. return TCP_COOKIE_MAX;
  355. if (cookie_size & 1)
  356. /* 8-bit multiple, illegal, fix it */
  357. cookie_size++;
  358. return (u8)cookie_size;
  359. }
  360. /* Write previously computed TCP options to the packet.
  361. *
  362. * Beware: Something in the Internet is very sensitive to the ordering of
  363. * TCP options, we learned this through the hard way, so be careful here.
  364. * Luckily we can at least blame others for their non-compliance but from
  365. * inter-operatibility perspective it seems that we're somewhat stuck with
  366. * the ordering which we have been using if we want to keep working with
  367. * those broken things (not that it currently hurts anybody as there isn't
  368. * particular reason why the ordering would need to be changed).
  369. *
  370. * At least SACK_PERM as the first option is known to lead to a disaster
  371. * (but it may well be that other scenarios fail similarly).
  372. */
  373. static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
  374. struct tcp_out_options *opts)
  375. {
  376. u8 options = opts->options; /* mungable copy */
  377. /* Having both authentication and cookies for security is redundant,
  378. * and there's certainly not enough room. Instead, the cookie-less
  379. * extension variant is proposed.
  380. *
  381. * Consider the pessimal case with authentication. The options
  382. * could look like:
  383. * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
  384. */
  385. if (unlikely(OPTION_MD5 & options)) {
  386. if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
  387. *ptr++ = htonl((TCPOPT_COOKIE << 24) |
  388. (TCPOLEN_COOKIE_BASE << 16) |
  389. (TCPOPT_MD5SIG << 8) |
  390. TCPOLEN_MD5SIG);
  391. } else {
  392. *ptr++ = htonl((TCPOPT_NOP << 24) |
  393. (TCPOPT_NOP << 16) |
  394. (TCPOPT_MD5SIG << 8) |
  395. TCPOLEN_MD5SIG);
  396. }
  397. options &= ~OPTION_COOKIE_EXTENSION;
  398. /* overload cookie hash location */
  399. opts->hash_location = (__u8 *)ptr;
  400. ptr += 4;
  401. }
  402. if (unlikely(opts->mss)) {
  403. *ptr++ = htonl((TCPOPT_MSS << 24) |
  404. (TCPOLEN_MSS << 16) |
  405. opts->mss);
  406. }
  407. if (likely(OPTION_TS & options)) {
  408. if (unlikely(OPTION_SACK_ADVERTISE & options)) {
  409. *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
  410. (TCPOLEN_SACK_PERM << 16) |
  411. (TCPOPT_TIMESTAMP << 8) |
  412. TCPOLEN_TIMESTAMP);
  413. options &= ~OPTION_SACK_ADVERTISE;
  414. } else {
  415. *ptr++ = htonl((TCPOPT_NOP << 24) |
  416. (TCPOPT_NOP << 16) |
  417. (TCPOPT_TIMESTAMP << 8) |
  418. TCPOLEN_TIMESTAMP);
  419. }
  420. *ptr++ = htonl(opts->tsval);
  421. *ptr++ = htonl(opts->tsecr);
  422. }
  423. /* Specification requires after timestamp, so do it now.
  424. *
  425. * Consider the pessimal case without authentication. The options
  426. * could look like:
  427. * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
  428. */
  429. if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
  430. __u8 *cookie_copy = opts->hash_location;
  431. u8 cookie_size = opts->hash_size;
  432. /* 8-bit multiple handled in tcp_cookie_size_check() above,
  433. * and elsewhere.
  434. */
  435. if (0x2 & cookie_size) {
  436. __u8 *p = (__u8 *)ptr;
  437. /* 16-bit multiple */
  438. *p++ = TCPOPT_COOKIE;
  439. *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
  440. *p++ = *cookie_copy++;
  441. *p++ = *cookie_copy++;
  442. ptr++;
  443. cookie_size -= 2;
  444. } else {
  445. /* 32-bit multiple */
  446. *ptr++ = htonl(((TCPOPT_NOP << 24) |
  447. (TCPOPT_NOP << 16) |
  448. (TCPOPT_COOKIE << 8) |
  449. TCPOLEN_COOKIE_BASE) +
  450. cookie_size);
  451. }
  452. if (cookie_size > 0) {
  453. memcpy(ptr, cookie_copy, cookie_size);
  454. ptr += (cookie_size / 4);
  455. }
  456. }
  457. if (unlikely(OPTION_SACK_ADVERTISE & options)) {
  458. *ptr++ = htonl((TCPOPT_NOP << 24) |
  459. (TCPOPT_NOP << 16) |
  460. (TCPOPT_SACK_PERM << 8) |
  461. TCPOLEN_SACK_PERM);
  462. }
  463. if (unlikely(OPTION_WSCALE & options)) {
  464. *ptr++ = htonl((TCPOPT_NOP << 24) |
  465. (TCPOPT_WINDOW << 16) |
  466. (TCPOLEN_WINDOW << 8) |
  467. opts->ws);
  468. }
  469. if (unlikely(opts->num_sack_blocks)) {
  470. struct tcp_sack_block *sp = tp->rx_opt.dsack ?
  471. tp->duplicate_sack : tp->selective_acks;
  472. int this_sack;
  473. *ptr++ = htonl((TCPOPT_NOP << 24) |
  474. (TCPOPT_NOP << 16) |
  475. (TCPOPT_SACK << 8) |
  476. (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
  477. TCPOLEN_SACK_PERBLOCK)));
  478. for (this_sack = 0; this_sack < opts->num_sack_blocks;
  479. ++this_sack) {
  480. *ptr++ = htonl(sp[this_sack].start_seq);
  481. *ptr++ = htonl(sp[this_sack].end_seq);
  482. }
  483. tp->rx_opt.dsack = 0;
  484. }
  485. }
  486. /* Compute TCP options for SYN packets. This is not the final
  487. * network wire format yet.
  488. */
  489. static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
  490. struct tcp_out_options *opts,
  491. struct tcp_md5sig_key **md5)
  492. {
  493. struct tcp_sock *tp = tcp_sk(sk);
  494. struct tcp_cookie_values *cvp = tp->cookie_values;
  495. unsigned remaining = MAX_TCP_OPTION_SPACE;
  496. u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
  497. tcp_cookie_size_check(cvp->cookie_desired) :
  498. 0;
  499. #ifdef CONFIG_TCP_MD5SIG
  500. *md5 = tp->af_specific->md5_lookup(sk, sk);
  501. if (*md5) {
  502. opts->options |= OPTION_MD5;
  503. remaining -= TCPOLEN_MD5SIG_ALIGNED;
  504. }
  505. #else
  506. *md5 = NULL;
  507. #endif
  508. /* We always get an MSS option. The option bytes which will be seen in
  509. * normal data packets should timestamps be used, must be in the MSS
  510. * advertised. But we subtract them from tp->mss_cache so that
  511. * calculations in tcp_sendmsg are simpler etc. So account for this
  512. * fact here if necessary. If we don't do this correctly, as a
  513. * receiver we won't recognize data packets as being full sized when we
  514. * should, and thus we won't abide by the delayed ACK rules correctly.
  515. * SACKs don't matter, we never delay an ACK when we have any of those
  516. * going out. */
  517. opts->mss = tcp_advertise_mss(sk);
  518. remaining -= TCPOLEN_MSS_ALIGNED;
  519. if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
  520. opts->options |= OPTION_TS;
  521. opts->tsval = TCP_SKB_CB(skb)->when;
  522. opts->tsecr = tp->rx_opt.ts_recent;
  523. remaining -= TCPOLEN_TSTAMP_ALIGNED;
  524. }
  525. if (likely(sysctl_tcp_window_scaling)) {
  526. opts->ws = tp->rx_opt.rcv_wscale;
  527. opts->options |= OPTION_WSCALE;
  528. remaining -= TCPOLEN_WSCALE_ALIGNED;
  529. }
  530. if (likely(sysctl_tcp_sack)) {
  531. opts->options |= OPTION_SACK_ADVERTISE;
  532. if (unlikely(!(OPTION_TS & opts->options)))
  533. remaining -= TCPOLEN_SACKPERM_ALIGNED;
  534. }
  535. /* Note that timestamps are required by the specification.
  536. *
  537. * Odd numbers of bytes are prohibited by the specification, ensuring
  538. * that the cookie is 16-bit aligned, and the resulting cookie pair is
  539. * 32-bit aligned.
  540. */
  541. if (*md5 == NULL &&
  542. (OPTION_TS & opts->options) &&
  543. cookie_size > 0) {
  544. int need = TCPOLEN_COOKIE_BASE + cookie_size;
  545. if (0x2 & need) {
  546. /* 32-bit multiple */
  547. need += 2; /* NOPs */
  548. if (need > remaining) {
  549. /* try shrinking cookie to fit */
  550. cookie_size -= 2;
  551. need -= 4;
  552. }
  553. }
  554. while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
  555. cookie_size -= 4;
  556. need -= 4;
  557. }
  558. if (TCP_COOKIE_MIN <= cookie_size) {
  559. opts->options |= OPTION_COOKIE_EXTENSION;
  560. opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
  561. opts->hash_size = cookie_size;
  562. /* Remember for future incarnations. */
  563. cvp->cookie_desired = cookie_size;
  564. if (cvp->cookie_desired != cvp->cookie_pair_size) {
  565. /* Currently use random bytes as a nonce,
  566. * assuming these are completely unpredictable
  567. * by hostile users of the same system.
  568. */
  569. get_random_bytes(&cvp->cookie_pair[0],
  570. cookie_size);
  571. cvp->cookie_pair_size = cookie_size;
  572. }
  573. remaining -= need;
  574. }
  575. }
  576. return MAX_TCP_OPTION_SPACE - remaining;
  577. }
  578. /* Set up TCP options for SYN-ACKs. */
  579. static unsigned tcp_synack_options(struct sock *sk,
  580. struct request_sock *req,
  581. unsigned mss, struct sk_buff *skb,
  582. struct tcp_out_options *opts,
  583. struct tcp_md5sig_key **md5,
  584. struct tcp_extend_values *xvp)
  585. {
  586. struct inet_request_sock *ireq = inet_rsk(req);
  587. unsigned remaining = MAX_TCP_OPTION_SPACE;
  588. u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
  589. xvp->cookie_plus :
  590. 0;
  591. #ifdef CONFIG_TCP_MD5SIG
  592. *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
  593. if (*md5) {
  594. opts->options |= OPTION_MD5;
  595. remaining -= TCPOLEN_MD5SIG_ALIGNED;
  596. /* We can't fit any SACK blocks in a packet with MD5 + TS
  597. * options. There was discussion about disabling SACK
  598. * rather than TS in order to fit in better with old,
  599. * buggy kernels, but that was deemed to be unnecessary.
  600. */
  601. ireq->tstamp_ok &= !ireq->sack_ok;
  602. }
  603. #else
  604. *md5 = NULL;
  605. #endif
  606. /* We always send an MSS option. */
  607. opts->mss = mss;
  608. remaining -= TCPOLEN_MSS_ALIGNED;
  609. if (likely(ireq->wscale_ok)) {
  610. opts->ws = ireq->rcv_wscale;
  611. opts->options |= OPTION_WSCALE;
  612. remaining -= TCPOLEN_WSCALE_ALIGNED;
  613. }
  614. if (likely(ireq->tstamp_ok)) {
  615. opts->options |= OPTION_TS;
  616. opts->tsval = TCP_SKB_CB(skb)->when;
  617. opts->tsecr = req->ts_recent;
  618. remaining -= TCPOLEN_TSTAMP_ALIGNED;
  619. }
  620. if (likely(ireq->sack_ok)) {
  621. opts->options |= OPTION_SACK_ADVERTISE;
  622. if (unlikely(!ireq->tstamp_ok))
  623. remaining -= TCPOLEN_SACKPERM_ALIGNED;
  624. }
  625. /* Similar rationale to tcp_syn_options() applies here, too.
  626. * If the <SYN> options fit, the same options should fit now!
  627. */
  628. if (*md5 == NULL &&
  629. ireq->tstamp_ok &&
  630. cookie_plus > TCPOLEN_COOKIE_BASE) {
  631. int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
  632. if (0x2 & need) {
  633. /* 32-bit multiple */
  634. need += 2; /* NOPs */
  635. }
  636. if (need <= remaining) {
  637. opts->options |= OPTION_COOKIE_EXTENSION;
  638. opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
  639. remaining -= need;
  640. } else {
  641. /* There's no error return, so flag it. */
  642. xvp->cookie_out_never = 1; /* true */
  643. opts->hash_size = 0;
  644. }
  645. }
  646. return MAX_TCP_OPTION_SPACE - remaining;
  647. }
  648. /* Compute TCP options for ESTABLISHED sockets. This is not the
  649. * final wire format yet.
  650. */
  651. static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
  652. struct tcp_out_options *opts,
  653. struct tcp_md5sig_key **md5)
  654. {
  655. struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
  656. struct tcp_sock *tp = tcp_sk(sk);
  657. unsigned size = 0;
  658. unsigned int eff_sacks;
  659. #ifdef CONFIG_TCP_MD5SIG
  660. *md5 = tp->af_specific->md5_lookup(sk, sk);
  661. if (unlikely(*md5)) {
  662. opts->options |= OPTION_MD5;
  663. size += TCPOLEN_MD5SIG_ALIGNED;
  664. }
  665. #else
  666. *md5 = NULL;
  667. #endif
  668. if (likely(tp->rx_opt.tstamp_ok)) {
  669. opts->options |= OPTION_TS;
  670. opts->tsval = tcb ? tcb->when : 0;
  671. opts->tsecr = tp->rx_opt.ts_recent;
  672. size += TCPOLEN_TSTAMP_ALIGNED;
  673. }
  674. eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
  675. if (unlikely(eff_sacks)) {
  676. const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
  677. opts->num_sack_blocks =
  678. min_t(unsigned, eff_sacks,
  679. (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
  680. TCPOLEN_SACK_PERBLOCK);
  681. size += TCPOLEN_SACK_BASE_ALIGNED +
  682. opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
  683. }
  684. return size;
  685. }
  686. /* This routine actually transmits TCP packets queued in by
  687. * tcp_do_sendmsg(). This is used by both the initial
  688. * transmission and possible later retransmissions.
  689. * All SKB's seen here are completely headerless. It is our
  690. * job to build the TCP header, and pass the packet down to
  691. * IP so it can do the same plus pass the packet off to the
  692. * device.
  693. *
  694. * We are working here with either a clone of the original
  695. * SKB, or a fresh unique copy made by the retransmit engine.
  696. */
  697. static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
  698. gfp_t gfp_mask)
  699. {
  700. const struct inet_connection_sock *icsk = inet_csk(sk);
  701. struct inet_sock *inet;
  702. struct tcp_sock *tp;
  703. struct tcp_skb_cb *tcb;
  704. struct tcp_out_options opts;
  705. unsigned tcp_options_size, tcp_header_size;
  706. struct tcp_md5sig_key *md5;
  707. struct tcphdr *th;
  708. int err;
  709. BUG_ON(!skb || !tcp_skb_pcount(skb));
  710. /* If congestion control is doing timestamping, we must
  711. * take such a timestamp before we potentially clone/copy.
  712. */
  713. if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
  714. __net_timestamp(skb);
  715. if (likely(clone_it)) {
  716. if (unlikely(skb_cloned(skb)))
  717. skb = pskb_copy(skb, gfp_mask);
  718. else
  719. skb = skb_clone(skb, gfp_mask);
  720. if (unlikely(!skb))
  721. return -ENOBUFS;
  722. }
  723. inet = inet_sk(sk);
  724. tp = tcp_sk(sk);
  725. tcb = TCP_SKB_CB(skb);
  726. memset(&opts, 0, sizeof(opts));
  727. if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
  728. tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
  729. else
  730. tcp_options_size = tcp_established_options(sk, skb, &opts,
  731. &md5);
  732. tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
  733. if (tcp_packets_in_flight(tp) == 0)
  734. tcp_ca_event(sk, CA_EVENT_TX_START);
  735. /* if no packet is in qdisc/device queue, then allow XPS to select
  736. * another queue.
  737. */
  738. skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
  739. skb_push(skb, tcp_header_size);
  740. skb_reset_transport_header(skb);
  741. skb_set_owner_w(skb, sk);
  742. /* Build TCP header and checksum it. */
  743. th = tcp_hdr(skb);
  744. th->source = inet->inet_sport;
  745. th->dest = inet->inet_dport;
  746. th->seq = htonl(tcb->seq);
  747. th->ack_seq = htonl(tp->rcv_nxt);
  748. *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
  749. tcb->tcp_flags);
  750. if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
  751. /* RFC1323: The window in SYN & SYN/ACK segments
  752. * is never scaled.
  753. */
  754. th->window = htons(min(tp->rcv_wnd, 65535U));
  755. } else {
  756. th->window = htons(tcp_select_window(sk));
  757. }
  758. th->check = 0;
  759. th->urg_ptr = 0;
  760. /* The urg_mode check is necessary during a below snd_una win probe */
  761. if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
  762. if (before(tp->snd_up, tcb->seq + 0x10000)) {
  763. th->urg_ptr = htons(tp->snd_up - tcb->seq);
  764. th->urg = 1;
  765. } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
  766. th->urg_ptr = htons(0xFFFF);
  767. th->urg = 1;
  768. }
  769. }
  770. tcp_options_write((__be32 *)(th + 1), tp, &opts);
  771. if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
  772. TCP_ECN_send(sk, skb, tcp_header_size);
  773. #ifdef CONFIG_TCP_MD5SIG
  774. /* Calculate the MD5 hash, as we have all we need now */
  775. if (md5) {
  776. sk_nocaps_add(sk, NETIF_F_GSO_MASK);
  777. tp->af_specific->calc_md5_hash(opts.hash_location,
  778. md5, sk, NULL, skb);
  779. }
  780. #endif
  781. icsk->icsk_af_ops->send_check(sk, skb);
  782. if (likely(tcb->tcp_flags & TCPHDR_ACK))
  783. tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
  784. if (skb->len != tcp_header_size)
  785. tcp_event_data_sent(tp, sk);
  786. if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
  787. TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
  788. tcp_skb_pcount(skb));
  789. err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
  790. if (likely(err <= 0))
  791. return err;
  792. tcp_enter_cwr(sk, 1);
  793. return net_xmit_eval(err);
  794. }
  795. /* This routine just queues the buffer for sending.
  796. *
  797. * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
  798. * otherwise socket can stall.
  799. */
  800. static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
  801. {
  802. struct tcp_sock *tp = tcp_sk(sk);
  803. /* Advance write_seq and place onto the write_queue. */
  804. tp->write_seq = TCP_SKB_CB(skb)->end_seq;
  805. skb_header_release(skb);
  806. tcp_add_write_queue_tail(sk, skb);
  807. sk->sk_wmem_queued += skb->truesize;
  808. sk_mem_charge(sk, skb->truesize);
  809. }
  810. /* Initialize TSO segments for a packet. */
  811. static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
  812. unsigned int mss_now)
  813. {
  814. /* Make sure we own this skb before messing gso_size/gso_segs */
  815. WARN_ON_ONCE(skb_cloned(skb));
  816. if (skb->len <= mss_now || !sk_can_gso(sk) ||
  817. skb->ip_summed == CHECKSUM_NONE) {
  818. /* Avoid the costly divide in the normal
  819. * non-TSO case.
  820. */
  821. skb_shinfo(skb)->gso_segs = 1;
  822. skb_shinfo(skb)->gso_size = 0;
  823. skb_shinfo(skb)->gso_type = 0;
  824. } else {
  825. skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
  826. skb_shinfo(skb)->gso_size = mss_now;
  827. skb_shinfo(skb)->gso_type = sk->sk_gso_type;
  828. }
  829. }
  830. /* When a modification to fackets out becomes necessary, we need to check
  831. * skb is counted to fackets_out or not.
  832. */
  833. static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
  834. int decr)
  835. {
  836. struct tcp_sock *tp = tcp_sk(sk);
  837. if (!tp->sacked_out || tcp_is_reno(tp))
  838. return;
  839. if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
  840. tp->fackets_out -= decr;
  841. }
  842. /* Pcount in the middle of the write queue got changed, we need to do various
  843. * tweaks to fix counters
  844. */
  845. static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
  846. {
  847. struct tcp_sock *tp = tcp_sk(sk);
  848. tp->packets_out -= decr;
  849. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
  850. tp->sacked_out -= decr;
  851. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
  852. tp->retrans_out -= decr;
  853. if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
  854. tp->lost_out -= decr;
  855. /* Reno case is special. Sigh... */
  856. if (tcp_is_reno(tp) && decr > 0)
  857. tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
  858. tcp_adjust_fackets_out(sk, skb, decr);
  859. if (tp->lost_skb_hint &&
  860. before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
  861. (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
  862. tp->lost_cnt_hint -= decr;
  863. tcp_verify_left_out(tp);
  864. }
  865. /* Function to create two new TCP segments. Shrinks the given segment
  866. * to the specified size and appends a new segment with the rest of the
  867. * packet to the list. This won't be called frequently, I hope.
  868. * Remember, these are still headerless SKBs at this point.
  869. */
  870. int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
  871. unsigned int mss_now)
  872. {
  873. struct tcp_sock *tp = tcp_sk(sk);
  874. struct sk_buff *buff;
  875. int nsize, old_factor;
  876. int nlen;
  877. u8 flags;
  878. if (WARN_ON(len > skb->len))
  879. return -EINVAL;
  880. nsize = skb_headlen(skb) - len;
  881. if (nsize < 0)
  882. nsize = 0;
  883. if (unlikely((sk->sk_wmem_queued >> 1) > sk->sk_sndbuf + 0x20000)) {
  884. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
  885. return -ENOMEM;
  886. }
  887. if (skb_unclone(skb, GFP_ATOMIC))
  888. return -ENOMEM;
  889. /* Get a new skb... force flag on. */
  890. buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
  891. if (buff == NULL)
  892. return -ENOMEM; /* We'll just try again later. */
  893. sk->sk_wmem_queued += buff->truesize;
  894. sk_mem_charge(sk, buff->truesize);
  895. nlen = skb->len - len - nsize;
  896. buff->truesize += nlen;
  897. skb->truesize -= nlen;
  898. /* Correct the sequence numbers. */
  899. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  900. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  901. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  902. /* PSH and FIN should only be set in the second packet. */
  903. flags = TCP_SKB_CB(skb)->tcp_flags;
  904. TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
  905. TCP_SKB_CB(buff)->tcp_flags = flags;
  906. TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
  907. if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
  908. /* Copy and checksum data tail into the new buffer. */
  909. buff->csum = csum_partial_copy_nocheck(skb->data + len,
  910. skb_put(buff, nsize),
  911. nsize, 0);
  912. skb_trim(skb, len);
  913. skb->csum = csum_block_sub(skb->csum, buff->csum, len);
  914. } else {
  915. skb->ip_summed = CHECKSUM_PARTIAL;
  916. skb_split(skb, buff, len);
  917. }
  918. buff->ip_summed = skb->ip_summed;
  919. /* Looks stupid, but our code really uses when of
  920. * skbs, which it never sent before. --ANK
  921. */
  922. TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
  923. buff->tstamp = skb->tstamp;
  924. old_factor = tcp_skb_pcount(skb);
  925. /* Fix up tso_factor for both original and new SKB. */
  926. tcp_set_skb_tso_segs(sk, skb, mss_now);
  927. tcp_set_skb_tso_segs(sk, buff, mss_now);
  928. /* If this packet has been sent out already, we must
  929. * adjust the various packet counters.
  930. */
  931. if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
  932. int diff = old_factor - tcp_skb_pcount(skb) -
  933. tcp_skb_pcount(buff);
  934. if (diff)
  935. tcp_adjust_pcount(sk, skb, diff);
  936. }
  937. /* Link BUFF into the send queue. */
  938. skb_header_release(buff);
  939. tcp_insert_write_queue_after(skb, buff, sk);
  940. return 0;
  941. }
  942. /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
  943. * eventually). The difference is that pulled data not copied, but
  944. * immediately discarded.
  945. */
  946. static void __pskb_trim_head(struct sk_buff *skb, int len)
  947. {
  948. int i, k, eat;
  949. eat = min_t(int, len, skb_headlen(skb));
  950. if (eat) {
  951. __skb_pull(skb, eat);
  952. len -= eat;
  953. if (!len)
  954. return;
  955. }
  956. eat = len;
  957. k = 0;
  958. for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  959. int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
  960. if (size <= eat) {
  961. skb_frag_unref(skb, i);
  962. eat -= size;
  963. } else {
  964. skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
  965. if (eat) {
  966. skb_shinfo(skb)->frags[k].page_offset += eat;
  967. skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
  968. eat = 0;
  969. }
  970. k++;
  971. }
  972. }
  973. skb_shinfo(skb)->nr_frags = k;
  974. skb_reset_tail_pointer(skb);
  975. skb->data_len -= len;
  976. skb->len = skb->data_len;
  977. }
  978. /* Remove acked data from a packet in the transmit queue. */
  979. int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
  980. {
  981. if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
  982. return -ENOMEM;
  983. __pskb_trim_head(skb, len);
  984. TCP_SKB_CB(skb)->seq += len;
  985. skb->ip_summed = CHECKSUM_PARTIAL;
  986. skb->truesize -= len;
  987. sk->sk_wmem_queued -= len;
  988. sk_mem_uncharge(sk, len);
  989. sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
  990. /* Any change of skb->len requires recalculation of tso factor. */
  991. if (tcp_skb_pcount(skb) > 1)
  992. tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
  993. return 0;
  994. }
  995. /* Calculate MSS. Not accounting for SACKs here. */
  996. int tcp_mtu_to_mss(const struct sock *sk, int pmtu)
  997. {
  998. const struct tcp_sock *tp = tcp_sk(sk);
  999. const struct inet_connection_sock *icsk = inet_csk(sk);
  1000. int mss_now;
  1001. /* Calculate base mss without TCP options:
  1002. It is MMS_S - sizeof(tcphdr) of rfc1122
  1003. */
  1004. mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
  1005. /* Clamp it (mss_clamp does not include tcp options) */
  1006. if (mss_now > tp->rx_opt.mss_clamp)
  1007. mss_now = tp->rx_opt.mss_clamp;
  1008. /* Now subtract optional transport overhead */
  1009. mss_now -= icsk->icsk_ext_hdr_len;
  1010. /* Then reserve room for full set of TCP options and 8 bytes of data */
  1011. mss_now = max(mss_now, sysctl_tcp_min_snd_mss);
  1012. /* Now subtract TCP options size, not including SACKs */
  1013. mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
  1014. return mss_now;
  1015. }
  1016. /* Inverse of above */
  1017. int tcp_mss_to_mtu(const struct sock *sk, int mss)
  1018. {
  1019. const struct tcp_sock *tp = tcp_sk(sk);
  1020. const struct inet_connection_sock *icsk = inet_csk(sk);
  1021. int mtu;
  1022. mtu = mss +
  1023. tp->tcp_header_len +
  1024. icsk->icsk_ext_hdr_len +
  1025. icsk->icsk_af_ops->net_header_len;
  1026. return mtu;
  1027. }
  1028. /* MTU probing init per socket */
  1029. void tcp_mtup_init(struct sock *sk)
  1030. {
  1031. struct tcp_sock *tp = tcp_sk(sk);
  1032. struct inet_connection_sock *icsk = inet_csk(sk);
  1033. icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
  1034. icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
  1035. icsk->icsk_af_ops->net_header_len;
  1036. icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
  1037. icsk->icsk_mtup.probe_size = 0;
  1038. }
  1039. EXPORT_SYMBOL(tcp_mtup_init);
  1040. /* This function synchronize snd mss to current pmtu/exthdr set.
  1041. tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
  1042. for TCP options, but includes only bare TCP header.
  1043. tp->rx_opt.mss_clamp is mss negotiated at connection setup.
  1044. It is minimum of user_mss and mss received with SYN.
  1045. It also does not include TCP options.
  1046. inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
  1047. tp->mss_cache is current effective sending mss, including
  1048. all tcp options except for SACKs. It is evaluated,
  1049. taking into account current pmtu, but never exceeds
  1050. tp->rx_opt.mss_clamp.
  1051. NOTE1. rfc1122 clearly states that advertised MSS
  1052. DOES NOT include either tcp or ip options.
  1053. NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
  1054. are READ ONLY outside this function. --ANK (980731)
  1055. */
  1056. unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
  1057. {
  1058. struct tcp_sock *tp = tcp_sk(sk);
  1059. struct inet_connection_sock *icsk = inet_csk(sk);
  1060. int mss_now;
  1061. if (icsk->icsk_mtup.search_high > pmtu)
  1062. icsk->icsk_mtup.search_high = pmtu;
  1063. mss_now = tcp_mtu_to_mss(sk, pmtu);
  1064. mss_now = tcp_bound_to_half_wnd(tp, mss_now);
  1065. /* And store cached results */
  1066. icsk->icsk_pmtu_cookie = pmtu;
  1067. if (icsk->icsk_mtup.enabled)
  1068. mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
  1069. tp->mss_cache = mss_now;
  1070. return mss_now;
  1071. }
  1072. EXPORT_SYMBOL(tcp_sync_mss);
  1073. /* Compute the current effective MSS, taking SACKs and IP options,
  1074. * and even PMTU discovery events into account.
  1075. */
  1076. unsigned int tcp_current_mss(struct sock *sk)
  1077. {
  1078. const struct tcp_sock *tp = tcp_sk(sk);
  1079. const struct dst_entry *dst = __sk_dst_get(sk);
  1080. u32 mss_now;
  1081. unsigned header_len;
  1082. struct tcp_out_options opts;
  1083. struct tcp_md5sig_key *md5;
  1084. mss_now = tp->mss_cache;
  1085. if (dst) {
  1086. u32 mtu = dst_mtu(dst);
  1087. if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
  1088. mss_now = tcp_sync_mss(sk, mtu);
  1089. }
  1090. header_len = tcp_established_options(sk, NULL, &opts, &md5) +
  1091. sizeof(struct tcphdr);
  1092. /* The mss_cache is sized based on tp->tcp_header_len, which assumes
  1093. * some common options. If this is an odd packet (because we have SACK
  1094. * blocks etc) then our calculated header_len will be different, and
  1095. * we have to adjust mss_now correspondingly */
  1096. if (header_len != tp->tcp_header_len) {
  1097. int delta = (int) header_len - tp->tcp_header_len;
  1098. mss_now -= delta;
  1099. }
  1100. return mss_now;
  1101. }
  1102. /* Congestion window validation. (RFC2861) */
  1103. static void tcp_cwnd_validate(struct sock *sk)
  1104. {
  1105. struct tcp_sock *tp = tcp_sk(sk);
  1106. if (tp->packets_out >= tp->snd_cwnd) {
  1107. /* Network is feed fully. */
  1108. tp->snd_cwnd_used = 0;
  1109. tp->snd_cwnd_stamp = tcp_time_stamp;
  1110. } else {
  1111. /* Network starves. */
  1112. if (tp->packets_out > tp->snd_cwnd_used)
  1113. tp->snd_cwnd_used = tp->packets_out;
  1114. if (sysctl_tcp_slow_start_after_idle &&
  1115. (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
  1116. tcp_cwnd_application_limited(sk);
  1117. }
  1118. }
  1119. /* Returns the portion of skb which can be sent right away without
  1120. * introducing MSS oddities to segment boundaries. In rare cases where
  1121. * mss_now != mss_cache, we will request caller to create a small skb
  1122. * per input skb which could be mostly avoided here (if desired).
  1123. *
  1124. * We explicitly want to create a request for splitting write queue tail
  1125. * to a small skb for Nagle purposes while avoiding unnecessary modulos,
  1126. * thus all the complexity (cwnd_len is always MSS multiple which we
  1127. * return whenever allowed by the other factors). Basically we need the
  1128. * modulo only when the receiver window alone is the limiting factor or
  1129. * when we would be allowed to send the split-due-to-Nagle skb fully.
  1130. */
  1131. static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
  1132. unsigned int mss_now, unsigned int max_segs)
  1133. {
  1134. const struct tcp_sock *tp = tcp_sk(sk);
  1135. u32 needed, window, max_len;
  1136. window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  1137. max_len = mss_now * max_segs;
  1138. if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
  1139. return max_len;
  1140. needed = min(skb->len, window);
  1141. if (max_len <= needed)
  1142. return max_len;
  1143. return needed - needed % mss_now;
  1144. }
  1145. /* Can at least one segment of SKB be sent right now, according to the
  1146. * congestion window rules? If so, return how many segments are allowed.
  1147. */
  1148. static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
  1149. const struct sk_buff *skb)
  1150. {
  1151. u32 in_flight, cwnd;
  1152. /* Don't be strict about the congestion window for the final FIN. */
  1153. if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
  1154. tcp_skb_pcount(skb) == 1)
  1155. return 1;
  1156. in_flight = tcp_packets_in_flight(tp);
  1157. cwnd = tp->snd_cwnd;
  1158. if (in_flight < cwnd)
  1159. return (cwnd - in_flight);
  1160. return 0;
  1161. }
  1162. /* Initialize TSO state of a skb.
  1163. * This must be invoked the first time we consider transmitting
  1164. * SKB onto the wire.
  1165. */
  1166. static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
  1167. unsigned int mss_now)
  1168. {
  1169. int tso_segs = tcp_skb_pcount(skb);
  1170. if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
  1171. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1172. tso_segs = tcp_skb_pcount(skb);
  1173. }
  1174. return tso_segs;
  1175. }
  1176. /* Minshall's variant of the Nagle send check. */
  1177. static inline int tcp_minshall_check(const struct tcp_sock *tp)
  1178. {
  1179. return after(tp->snd_sml, tp->snd_una) &&
  1180. !after(tp->snd_sml, tp->snd_nxt);
  1181. }
  1182. /* Return 0, if packet can be sent now without violation Nagle's rules:
  1183. * 1. It is full sized.
  1184. * 2. Or it contains FIN. (already checked by caller)
  1185. * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
  1186. * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
  1187. * With Minshall's modification: all sent small packets are ACKed.
  1188. */
  1189. static inline int tcp_nagle_check(const struct tcp_sock *tp,
  1190. const struct sk_buff *skb,
  1191. unsigned mss_now, int nonagle)
  1192. {
  1193. return skb->len < mss_now &&
  1194. ((nonagle & TCP_NAGLE_CORK) ||
  1195. (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
  1196. }
  1197. /* Return non-zero if the Nagle test allows this packet to be
  1198. * sent now.
  1199. */
  1200. static inline int tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
  1201. unsigned int cur_mss, int nonagle)
  1202. {
  1203. /* Nagle rule does not apply to frames, which sit in the middle of the
  1204. * write_queue (they have no chances to get new data).
  1205. *
  1206. * This is implemented in the callers, where they modify the 'nonagle'
  1207. * argument based upon the location of SKB in the send queue.
  1208. */
  1209. if (nonagle & TCP_NAGLE_PUSH)
  1210. return 1;
  1211. /* Don't use the nagle rule for urgent data (or for the final FIN).
  1212. * Nagle can be ignored during F-RTO too (see RFC4138).
  1213. */
  1214. if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
  1215. (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
  1216. return 1;
  1217. if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
  1218. return 1;
  1219. return 0;
  1220. }
  1221. /* Does at least the first segment of SKB fit into the send window? */
  1222. static inline int tcp_snd_wnd_test(const struct tcp_sock *tp, const struct sk_buff *skb,
  1223. unsigned int cur_mss)
  1224. {
  1225. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  1226. if (skb->len > cur_mss)
  1227. end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
  1228. return !after(end_seq, tcp_wnd_end(tp));
  1229. }
  1230. /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
  1231. * should be put on the wire right now. If so, it returns the number of
  1232. * packets allowed by the congestion window.
  1233. */
  1234. static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
  1235. unsigned int cur_mss, int nonagle)
  1236. {
  1237. const struct tcp_sock *tp = tcp_sk(sk);
  1238. unsigned int cwnd_quota;
  1239. tcp_init_tso_segs(sk, skb, cur_mss);
  1240. if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
  1241. return 0;
  1242. cwnd_quota = tcp_cwnd_test(tp, skb);
  1243. if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
  1244. cwnd_quota = 0;
  1245. return cwnd_quota;
  1246. }
  1247. /* Test if sending is allowed right now. */
  1248. int tcp_may_send_now(struct sock *sk)
  1249. {
  1250. const struct tcp_sock *tp = tcp_sk(sk);
  1251. struct sk_buff *skb = tcp_send_head(sk);
  1252. return skb &&
  1253. tcp_snd_test(sk, skb, tcp_current_mss(sk),
  1254. (tcp_skb_is_last(sk, skb) ?
  1255. tp->nonagle : TCP_NAGLE_PUSH));
  1256. }
  1257. /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
  1258. * which is put after SKB on the list. It is very much like
  1259. * tcp_fragment() except that it may make several kinds of assumptions
  1260. * in order to speed up the splitting operation. In particular, we
  1261. * know that all the data is in scatter-gather pages, and that the
  1262. * packet has never been sent out before (and thus is not cloned).
  1263. */
  1264. static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
  1265. unsigned int mss_now, gfp_t gfp)
  1266. {
  1267. struct sk_buff *buff;
  1268. int nlen = skb->len - len;
  1269. u8 flags;
  1270. /* All of a TSO frame must be composed of paged data. */
  1271. if (skb->len != skb->data_len)
  1272. return tcp_fragment(sk, skb, len, mss_now);
  1273. buff = sk_stream_alloc_skb(sk, 0, gfp);
  1274. if (unlikely(buff == NULL))
  1275. return -ENOMEM;
  1276. sk->sk_wmem_queued += buff->truesize;
  1277. sk_mem_charge(sk, buff->truesize);
  1278. buff->truesize += nlen;
  1279. skb->truesize -= nlen;
  1280. /* Correct the sequence numbers. */
  1281. TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
  1282. TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
  1283. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
  1284. /* PSH and FIN should only be set in the second packet. */
  1285. flags = TCP_SKB_CB(skb)->tcp_flags;
  1286. TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
  1287. TCP_SKB_CB(buff)->tcp_flags = flags;
  1288. /* This packet was never sent out yet, so no SACK bits. */
  1289. TCP_SKB_CB(buff)->sacked = 0;
  1290. buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
  1291. skb_split(skb, buff, len);
  1292. /* Fix up tso_factor for both original and new SKB. */
  1293. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1294. tcp_set_skb_tso_segs(sk, buff, mss_now);
  1295. /* Link BUFF into the send queue. */
  1296. skb_header_release(buff);
  1297. tcp_insert_write_queue_after(skb, buff, sk);
  1298. return 0;
  1299. }
  1300. /* Try to defer sending, if possible, in order to minimize the amount
  1301. * of TSO splitting we do. View it as a kind of TSO Nagle test.
  1302. *
  1303. * This algorithm is from John Heffner.
  1304. */
  1305. static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
  1306. {
  1307. struct tcp_sock *tp = tcp_sk(sk);
  1308. const struct inet_connection_sock *icsk = inet_csk(sk);
  1309. u32 send_win, cong_win, limit, in_flight;
  1310. int win_divisor;
  1311. if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
  1312. goto send_now;
  1313. if (icsk->icsk_ca_state != TCP_CA_Open)
  1314. goto send_now;
  1315. /* Defer for less than two clock ticks. */
  1316. if (tp->tso_deferred &&
  1317. (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
  1318. goto send_now;
  1319. in_flight = tcp_packets_in_flight(tp);
  1320. BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
  1321. send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  1322. /* From in_flight test above, we know that cwnd > in_flight. */
  1323. cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
  1324. limit = min(send_win, cong_win);
  1325. /* If a full-sized TSO skb can be sent, do it. */
  1326. if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
  1327. sk->sk_gso_max_segs * tp->mss_cache))
  1328. goto send_now;
  1329. /* Middle in queue won't get any more data, full sendable already? */
  1330. if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
  1331. goto send_now;
  1332. win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
  1333. if (win_divisor) {
  1334. u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
  1335. /* If at least some fraction of a window is available,
  1336. * just use it.
  1337. */
  1338. chunk /= win_divisor;
  1339. if (limit >= chunk)
  1340. goto send_now;
  1341. } else {
  1342. /* Different approach, try not to defer past a single
  1343. * ACK. Receiver should ACK every other full sized
  1344. * frame, so if we have space for more than 3 frames
  1345. * then send now.
  1346. */
  1347. if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
  1348. goto send_now;
  1349. }
  1350. /* Ok, it looks like it is advisable to defer.
  1351. * Do not rearm the timer if already set to not break TCP ACK clocking.
  1352. */
  1353. if (!tp->tso_deferred)
  1354. tp->tso_deferred = 1 | (jiffies << 1);
  1355. return 1;
  1356. send_now:
  1357. tp->tso_deferred = 0;
  1358. return 0;
  1359. }
  1360. /* Create a new MTU probe if we are ready.
  1361. * MTU probe is regularly attempting to increase the path MTU by
  1362. * deliberately sending larger packets. This discovers routing
  1363. * changes resulting in larger path MTUs.
  1364. *
  1365. * Returns 0 if we should wait to probe (no cwnd available),
  1366. * 1 if a probe was sent,
  1367. * -1 otherwise
  1368. */
  1369. static int tcp_mtu_probe(struct sock *sk)
  1370. {
  1371. struct tcp_sock *tp = tcp_sk(sk);
  1372. struct inet_connection_sock *icsk = inet_csk(sk);
  1373. struct sk_buff *skb, *nskb, *next;
  1374. int len;
  1375. int probe_size;
  1376. int size_needed;
  1377. int copy;
  1378. int mss_now;
  1379. /* Not currently probing/verifying,
  1380. * not in recovery,
  1381. * have enough cwnd, and
  1382. * not SACKing (the variable headers throw things off) */
  1383. if (!icsk->icsk_mtup.enabled ||
  1384. icsk->icsk_mtup.probe_size ||
  1385. inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
  1386. tp->snd_cwnd < 11 ||
  1387. tp->rx_opt.num_sacks || tp->rx_opt.dsack)
  1388. return -1;
  1389. /* Very simple search strategy: just double the MSS. */
  1390. mss_now = tcp_current_mss(sk);
  1391. probe_size = 2 * tp->mss_cache;
  1392. size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
  1393. if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
  1394. /* TODO: set timer for probe_converge_event */
  1395. return -1;
  1396. }
  1397. /* Have enough data in the send queue to probe? */
  1398. if (tp->write_seq - tp->snd_nxt < size_needed)
  1399. return -1;
  1400. if (tp->snd_wnd < size_needed)
  1401. return -1;
  1402. if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
  1403. return 0;
  1404. /* Do we need to wait to drain cwnd? With none in flight, don't stall */
  1405. if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
  1406. if (!tcp_packets_in_flight(tp))
  1407. return -1;
  1408. else
  1409. return 0;
  1410. }
  1411. /* We're allowed to probe. Build it now. */
  1412. if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
  1413. return -1;
  1414. sk->sk_wmem_queued += nskb->truesize;
  1415. sk_mem_charge(sk, nskb->truesize);
  1416. skb = tcp_send_head(sk);
  1417. TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
  1418. TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
  1419. TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
  1420. TCP_SKB_CB(nskb)->sacked = 0;
  1421. nskb->csum = 0;
  1422. nskb->ip_summed = skb->ip_summed;
  1423. tcp_insert_write_queue_before(nskb, skb, sk);
  1424. len = 0;
  1425. tcp_for_write_queue_from_safe(skb, next, sk) {
  1426. copy = min_t(int, skb->len, probe_size - len);
  1427. if (nskb->ip_summed)
  1428. skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
  1429. else
  1430. nskb->csum = skb_copy_and_csum_bits(skb, 0,
  1431. skb_put(nskb, copy),
  1432. copy, nskb->csum);
  1433. if (skb->len <= copy) {
  1434. /* We've eaten all the data from this skb.
  1435. * Throw it away. */
  1436. TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
  1437. tcp_unlink_write_queue(skb, sk);
  1438. sk_wmem_free_skb(sk, skb);
  1439. } else {
  1440. TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
  1441. ~(TCPHDR_FIN|TCPHDR_PSH);
  1442. if (!skb_shinfo(skb)->nr_frags) {
  1443. skb_pull(skb, copy);
  1444. if (skb->ip_summed != CHECKSUM_PARTIAL)
  1445. skb->csum = csum_partial(skb->data,
  1446. skb->len, 0);
  1447. } else {
  1448. __pskb_trim_head(skb, copy);
  1449. tcp_set_skb_tso_segs(sk, skb, mss_now);
  1450. }
  1451. TCP_SKB_CB(skb)->seq += copy;
  1452. }
  1453. len += copy;
  1454. if (len >= probe_size)
  1455. break;
  1456. }
  1457. tcp_init_tso_segs(sk, nskb, nskb->len);
  1458. /* We're ready to send. If this fails, the probe will
  1459. * be resegmented into mss-sized pieces by tcp_write_xmit(). */
  1460. TCP_SKB_CB(nskb)->when = tcp_time_stamp;
  1461. if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
  1462. /* Decrement cwnd here because we are sending
  1463. * effectively two packets. */
  1464. tp->snd_cwnd--;
  1465. tcp_event_new_data_sent(sk, nskb);
  1466. icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
  1467. tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
  1468. tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
  1469. return 1;
  1470. }
  1471. return -1;
  1472. }
  1473. /* This routine writes packets to the network. It advances the
  1474. * send_head. This happens as incoming acks open up the remote
  1475. * window for us.
  1476. *
  1477. * LARGESEND note: !tcp_urg_mode is overkill, only frames between
  1478. * snd_up-64k-mss .. snd_up cannot be large. However, taking into
  1479. * account rare use of URG, this is not a big flaw.
  1480. *
  1481. * Returns 1, if no segments are in flight and we have queued segments, but
  1482. * cannot send anything now because of SWS or another problem.
  1483. */
  1484. static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  1485. int push_one, gfp_t gfp)
  1486. {
  1487. struct tcp_sock *tp = tcp_sk(sk);
  1488. struct sk_buff *skb;
  1489. unsigned int tso_segs, sent_pkts;
  1490. int cwnd_quota;
  1491. int result;
  1492. sent_pkts = 0;
  1493. if (!push_one) {
  1494. /* Do MTU probing. */
  1495. result = tcp_mtu_probe(sk);
  1496. if (!result) {
  1497. return 0;
  1498. } else if (result > 0) {
  1499. sent_pkts = 1;
  1500. }
  1501. }
  1502. while ((skb = tcp_send_head(sk))) {
  1503. unsigned int limit;
  1504. tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
  1505. BUG_ON(!tso_segs);
  1506. cwnd_quota = tcp_cwnd_test(tp, skb);
  1507. if (!cwnd_quota)
  1508. break;
  1509. if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
  1510. break;
  1511. if (tso_segs == 1) {
  1512. if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
  1513. (tcp_skb_is_last(sk, skb) ?
  1514. nonagle : TCP_NAGLE_PUSH))))
  1515. break;
  1516. } else {
  1517. if (!push_one && tcp_tso_should_defer(sk, skb))
  1518. break;
  1519. }
  1520. limit = mss_now;
  1521. if (tso_segs > 1 && !tcp_urg_mode(tp))
  1522. limit = tcp_mss_split_point(sk, skb, mss_now,
  1523. min_t(unsigned int,
  1524. cwnd_quota,
  1525. sk->sk_gso_max_segs));
  1526. if (skb->len > limit &&
  1527. unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
  1528. break;
  1529. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1530. if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
  1531. break;
  1532. /* Advance the send_head. This one is sent out.
  1533. * This call will increment packets_out.
  1534. */
  1535. tcp_event_new_data_sent(sk, skb);
  1536. tcp_minshall_update(tp, mss_now, skb);
  1537. sent_pkts += tcp_skb_pcount(skb);
  1538. if (push_one)
  1539. break;
  1540. }
  1541. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
  1542. tp->prr_out += sent_pkts;
  1543. if (likely(sent_pkts)) {
  1544. tcp_cwnd_validate(sk);
  1545. return 0;
  1546. }
  1547. return !tp->packets_out && tcp_send_head(sk);
  1548. }
  1549. /* Push out any pending frames which were held back due to
  1550. * TCP_CORK or attempt at coalescing tiny packets.
  1551. * The socket must be locked by the caller.
  1552. */
  1553. void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
  1554. int nonagle)
  1555. {
  1556. /* If we are closed, the bytes will have to remain here.
  1557. * In time closedown will finish, we empty the write queue and
  1558. * all will be happy.
  1559. */
  1560. if (unlikely(sk->sk_state == TCP_CLOSE))
  1561. return;
  1562. if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
  1563. tcp_check_probe_timer(sk);
  1564. }
  1565. /* Send _single_ skb sitting at the send head. This function requires
  1566. * true push pending frames to setup probe timer etc.
  1567. */
  1568. void tcp_push_one(struct sock *sk, unsigned int mss_now)
  1569. {
  1570. struct sk_buff *skb = tcp_send_head(sk);
  1571. BUG_ON(!skb || skb->len < mss_now);
  1572. tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
  1573. }
  1574. /* This function returns the amount that we can raise the
  1575. * usable window based on the following constraints
  1576. *
  1577. * 1. The window can never be shrunk once it is offered (RFC 793)
  1578. * 2. We limit memory per socket
  1579. *
  1580. * RFC 1122:
  1581. * "the suggested [SWS] avoidance algorithm for the receiver is to keep
  1582. * RECV.NEXT + RCV.WIN fixed until:
  1583. * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
  1584. *
  1585. * i.e. don't raise the right edge of the window until you can raise
  1586. * it at least MSS bytes.
  1587. *
  1588. * Unfortunately, the recommended algorithm breaks header prediction,
  1589. * since header prediction assumes th->window stays fixed.
  1590. *
  1591. * Strictly speaking, keeping th->window fixed violates the receiver
  1592. * side SWS prevention criteria. The problem is that under this rule
  1593. * a stream of single byte packets will cause the right side of the
  1594. * window to always advance by a single byte.
  1595. *
  1596. * Of course, if the sender implements sender side SWS prevention
  1597. * then this will not be a problem.
  1598. *
  1599. * BSD seems to make the following compromise:
  1600. *
  1601. * If the free space is less than the 1/4 of the maximum
  1602. * space available and the free space is less than 1/2 mss,
  1603. * then set the window to 0.
  1604. * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
  1605. * Otherwise, just prevent the window from shrinking
  1606. * and from being larger than the largest representable value.
  1607. *
  1608. * This prevents incremental opening of the window in the regime
  1609. * where TCP is limited by the speed of the reader side taking
  1610. * data out of the TCP receive queue. It does nothing about
  1611. * those cases where the window is constrained on the sender side
  1612. * because the pipeline is full.
  1613. *
  1614. * BSD also seems to "accidentally" limit itself to windows that are a
  1615. * multiple of MSS, at least until the free space gets quite small.
  1616. * This would appear to be a side effect of the mbuf implementation.
  1617. * Combining these two algorithms results in the observed behavior
  1618. * of having a fixed window size at almost all times.
  1619. *
  1620. * Below we obtain similar behavior by forcing the offered window to
  1621. * a multiple of the mss when it is feasible to do so.
  1622. *
  1623. * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
  1624. * Regular options like TIMESTAMP are taken into account.
  1625. */
  1626. u32 __tcp_select_window(struct sock *sk)
  1627. {
  1628. struct inet_connection_sock *icsk = inet_csk(sk);
  1629. struct tcp_sock *tp = tcp_sk(sk);
  1630. /* MSS for the peer's data. Previous versions used mss_clamp
  1631. * here. I don't know if the value based on our guesses
  1632. * of peer's MSS is better for the performance. It's more correct
  1633. * but may be worse for the performance because of rcv_mss
  1634. * fluctuations. --SAW 1998/11/1
  1635. */
  1636. int mss = icsk->icsk_ack.rcv_mss;
  1637. int free_space = tcp_space(sk);
  1638. int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
  1639. int window;
  1640. if (mss > full_space)
  1641. mss = full_space;
  1642. if (free_space < (full_space >> 1)) {
  1643. icsk->icsk_ack.quick = 0;
  1644. if (sk_under_memory_pressure(sk))
  1645. tp->rcv_ssthresh = min(tp->rcv_ssthresh,
  1646. 4U * tp->advmss);
  1647. if (free_space < mss)
  1648. return 0;
  1649. }
  1650. if (free_space > tp->rcv_ssthresh)
  1651. free_space = tp->rcv_ssthresh;
  1652. /* Don't do rounding if we are using window scaling, since the
  1653. * scaled window will not line up with the MSS boundary anyway.
  1654. */
  1655. window = tp->rcv_wnd;
  1656. if (tp->rx_opt.rcv_wscale) {
  1657. window = free_space;
  1658. /* Advertise enough space so that it won't get scaled away.
  1659. * Import case: prevent zero window announcement if
  1660. * 1<<rcv_wscale > mss.
  1661. */
  1662. if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
  1663. window = (((window >> tp->rx_opt.rcv_wscale) + 1)
  1664. << tp->rx_opt.rcv_wscale);
  1665. } else {
  1666. /* Get the largest window that is a nice multiple of mss.
  1667. * Window clamp already applied above.
  1668. * If our current window offering is within 1 mss of the
  1669. * free space we just keep it. This prevents the divide
  1670. * and multiply from happening most of the time.
  1671. * We also don't do any window rounding when the free space
  1672. * is too small.
  1673. */
  1674. if (window <= free_space - mss || window > free_space)
  1675. window = (free_space / mss) * mss;
  1676. else if (mss == full_space &&
  1677. free_space > window + (full_space >> 1))
  1678. window = free_space;
  1679. }
  1680. return window;
  1681. }
  1682. /* Collapses two adjacent SKB's during retransmission. */
  1683. static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
  1684. {
  1685. struct tcp_sock *tp = tcp_sk(sk);
  1686. struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
  1687. int skb_size, next_skb_size;
  1688. skb_size = skb->len;
  1689. next_skb_size = next_skb->len;
  1690. BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
  1691. tcp_highest_sack_combine(sk, next_skb, skb);
  1692. tcp_unlink_write_queue(next_skb, sk);
  1693. skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
  1694. next_skb_size);
  1695. if (next_skb->ip_summed == CHECKSUM_PARTIAL)
  1696. skb->ip_summed = CHECKSUM_PARTIAL;
  1697. if (skb->ip_summed != CHECKSUM_PARTIAL)
  1698. skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
  1699. /* Update sequence range on original skb. */
  1700. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
  1701. /* Merge over control information. This moves PSH/FIN etc. over */
  1702. TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
  1703. /* All done, get rid of second SKB and account for it so
  1704. * packet counting does not break.
  1705. */
  1706. TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
  1707. /* changed transmit queue under us so clear hints */
  1708. tcp_clear_retrans_hints_partial(tp);
  1709. if (next_skb == tp->retransmit_skb_hint)
  1710. tp->retransmit_skb_hint = skb;
  1711. tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
  1712. sk_wmem_free_skb(sk, next_skb);
  1713. }
  1714. /* Check if coalescing SKBs is legal. */
  1715. static int tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
  1716. {
  1717. if (tcp_skb_pcount(skb) > 1)
  1718. return 0;
  1719. /* TODO: SACK collapsing could be used to remove this condition */
  1720. if (skb_shinfo(skb)->nr_frags != 0)
  1721. return 0;
  1722. if (skb_cloned(skb))
  1723. return 0;
  1724. if (skb == tcp_send_head(sk))
  1725. return 0;
  1726. /* Some heurestics for collapsing over SACK'd could be invented */
  1727. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
  1728. return 0;
  1729. return 1;
  1730. }
  1731. /* Collapse packets in the retransmit queue to make to create
  1732. * less packets on the wire. This is only done on retransmission.
  1733. */
  1734. static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
  1735. int space)
  1736. {
  1737. struct tcp_sock *tp = tcp_sk(sk);
  1738. struct sk_buff *skb = to, *tmp;
  1739. int first = 1;
  1740. if (!sysctl_tcp_retrans_collapse)
  1741. return;
  1742. if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
  1743. return;
  1744. tcp_for_write_queue_from_safe(skb, tmp, sk) {
  1745. if (!tcp_can_collapse(sk, skb))
  1746. break;
  1747. space -= skb->len;
  1748. if (first) {
  1749. first = 0;
  1750. continue;
  1751. }
  1752. if (space < 0)
  1753. break;
  1754. /* Punt if not enough space exists in the first SKB for
  1755. * the data in the second
  1756. */
  1757. if (skb->len > skb_availroom(to))
  1758. break;
  1759. if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
  1760. break;
  1761. tcp_collapse_retrans(sk, to);
  1762. }
  1763. }
  1764. /* This retransmits one SKB. Policy decisions and retransmit queue
  1765. * state updates are done by the caller. Returns non-zero if an
  1766. * error occurred which prevented the send.
  1767. */
  1768. int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
  1769. {
  1770. struct tcp_sock *tp = tcp_sk(sk);
  1771. struct inet_connection_sock *icsk = inet_csk(sk);
  1772. unsigned int cur_mss;
  1773. int err;
  1774. /* Inconslusive MTU probe */
  1775. if (icsk->icsk_mtup.probe_size) {
  1776. icsk->icsk_mtup.probe_size = 0;
  1777. }
  1778. /* Do not sent more than we queued. 1/4 is reserved for possible
  1779. * copying overhead: fragmentation, tunneling, mangling etc.
  1780. */
  1781. if (atomic_read(&sk->sk_wmem_alloc) >
  1782. min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
  1783. return -EAGAIN;
  1784. if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
  1785. if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
  1786. WARN_ON_ONCE(1);
  1787. return -EINVAL;
  1788. }
  1789. if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
  1790. return -ENOMEM;
  1791. }
  1792. if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
  1793. return -EHOSTUNREACH; /* Routing failure or similar. */
  1794. cur_mss = tcp_current_mss(sk);
  1795. /* If receiver has shrunk his window, and skb is out of
  1796. * new window, do not retransmit it. The exception is the
  1797. * case, when window is shrunk to zero. In this case
  1798. * our retransmit serves as a zero window probe.
  1799. */
  1800. if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
  1801. TCP_SKB_CB(skb)->seq != tp->snd_una)
  1802. return -EAGAIN;
  1803. if (skb->len > cur_mss) {
  1804. if (tcp_fragment(sk, skb, cur_mss, cur_mss))
  1805. return -ENOMEM; /* We'll try again later. */
  1806. } else {
  1807. int oldpcount = tcp_skb_pcount(skb);
  1808. if (unlikely(oldpcount > 1)) {
  1809. if (skb_unclone(skb, GFP_ATOMIC))
  1810. return -ENOMEM;
  1811. tcp_init_tso_segs(sk, skb, cur_mss);
  1812. tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
  1813. }
  1814. }
  1815. tcp_retrans_try_collapse(sk, skb, cur_mss);
  1816. /* Some Solaris stacks overoptimize and ignore the FIN on a
  1817. * retransmit when old data is attached. So strip it off
  1818. * since it is cheap to do so and saves bytes on the network.
  1819. */
  1820. if (skb->len > 0 &&
  1821. (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
  1822. tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
  1823. if (!pskb_trim(skb, 0)) {
  1824. /* Reuse, even though it does some unnecessary work */
  1825. tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
  1826. TCP_SKB_CB(skb)->tcp_flags);
  1827. skb->ip_summed = CHECKSUM_NONE;
  1828. }
  1829. }
  1830. /* Make a copy, if the first transmission SKB clone we made
  1831. * is still in somebody's hands, else make a clone.
  1832. */
  1833. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  1834. /* make sure skb->data is aligned on arches that require it
  1835. * and check if ack-trimming & collapsing extended the headroom
  1836. * beyond what csum_start can cover.
  1837. */
  1838. if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
  1839. skb_headroom(skb) >= 0xFFFF)) {
  1840. struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
  1841. GFP_ATOMIC);
  1842. err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
  1843. -ENOBUFS;
  1844. } else {
  1845. err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  1846. }
  1847. if (err == 0) {
  1848. /* Update global TCP statistics. */
  1849. TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
  1850. tp->total_retrans++;
  1851. #if FASTRETRANS_DEBUG > 0
  1852. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
  1853. if (net_ratelimit())
  1854. printk(KERN_DEBUG "retrans_out leaked.\n");
  1855. }
  1856. #endif
  1857. if (!tp->retrans_out)
  1858. tp->lost_retrans_low = tp->snd_nxt;
  1859. TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
  1860. tp->retrans_out += tcp_skb_pcount(skb);
  1861. /* Save stamp of the first retransmit. */
  1862. if (!tp->retrans_stamp)
  1863. tp->retrans_stamp = TCP_SKB_CB(skb)->when;
  1864. /* snd_nxt is stored to detect loss of retransmitted segment,
  1865. * see tcp_input.c tcp_sacktag_write_queue().
  1866. */
  1867. TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
  1868. }
  1869. if (tp->undo_retrans < 0)
  1870. tp->undo_retrans = 0;
  1871. tp->undo_retrans += tcp_skb_pcount(skb);
  1872. return err;
  1873. }
  1874. /* Check if we forward retransmits are possible in the current
  1875. * window/congestion state.
  1876. */
  1877. static int tcp_can_forward_retransmit(struct sock *sk)
  1878. {
  1879. const struct inet_connection_sock *icsk = inet_csk(sk);
  1880. const struct tcp_sock *tp = tcp_sk(sk);
  1881. /* Forward retransmissions are possible only during Recovery. */
  1882. if (icsk->icsk_ca_state != TCP_CA_Recovery)
  1883. return 0;
  1884. /* No forward retransmissions in Reno are possible. */
  1885. if (tcp_is_reno(tp))
  1886. return 0;
  1887. /* Yeah, we have to make difficult choice between forward transmission
  1888. * and retransmission... Both ways have their merits...
  1889. *
  1890. * For now we do not retransmit anything, while we have some new
  1891. * segments to send. In the other cases, follow rule 3 for
  1892. * NextSeg() specified in RFC3517.
  1893. */
  1894. if (tcp_may_send_now(sk))
  1895. return 0;
  1896. return 1;
  1897. }
  1898. /* This gets called after a retransmit timeout, and the initially
  1899. * retransmitted data is acknowledged. It tries to continue
  1900. * resending the rest of the retransmit queue, until either
  1901. * we've sent it all or the congestion window limit is reached.
  1902. * If doing SACK, the first ACK which comes back for a timeout
  1903. * based retransmit packet might feed us FACK information again.
  1904. * If so, we use it to avoid unnecessarily retransmissions.
  1905. */
  1906. void tcp_xmit_retransmit_queue(struct sock *sk)
  1907. {
  1908. const struct inet_connection_sock *icsk = inet_csk(sk);
  1909. struct tcp_sock *tp = tcp_sk(sk);
  1910. struct sk_buff *skb;
  1911. struct sk_buff *hole = NULL;
  1912. u32 last_lost;
  1913. int mib_idx;
  1914. int fwd_rexmitting = 0;
  1915. if (!tp->packets_out)
  1916. return;
  1917. if (!tp->lost_out)
  1918. tp->retransmit_high = tp->snd_una;
  1919. if (tp->retransmit_skb_hint) {
  1920. skb = tp->retransmit_skb_hint;
  1921. last_lost = TCP_SKB_CB(skb)->end_seq;
  1922. if (after(last_lost, tp->retransmit_high))
  1923. last_lost = tp->retransmit_high;
  1924. } else {
  1925. skb = tcp_write_queue_head(sk);
  1926. last_lost = tp->snd_una;
  1927. }
  1928. tcp_for_write_queue_from(skb, sk) {
  1929. __u8 sacked = TCP_SKB_CB(skb)->sacked;
  1930. if (skb == tcp_send_head(sk))
  1931. break;
  1932. /* we could do better than to assign each time */
  1933. if (hole == NULL)
  1934. tp->retransmit_skb_hint = skb;
  1935. /* Assume this retransmit will generate
  1936. * only one packet for congestion window
  1937. * calculation purposes. This works because
  1938. * tcp_retransmit_skb() will chop up the
  1939. * packet to be MSS sized and all the
  1940. * packet counting works out.
  1941. */
  1942. if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
  1943. return;
  1944. if (fwd_rexmitting) {
  1945. begin_fwd:
  1946. if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
  1947. break;
  1948. mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
  1949. } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
  1950. tp->retransmit_high = last_lost;
  1951. if (!tcp_can_forward_retransmit(sk))
  1952. break;
  1953. /* Backtrack if necessary to non-L'ed skb */
  1954. if (hole != NULL) {
  1955. skb = hole;
  1956. hole = NULL;
  1957. }
  1958. fwd_rexmitting = 1;
  1959. goto begin_fwd;
  1960. } else if (!(sacked & TCPCB_LOST)) {
  1961. if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
  1962. hole = skb;
  1963. continue;
  1964. } else {
  1965. last_lost = TCP_SKB_CB(skb)->end_seq;
  1966. if (icsk->icsk_ca_state != TCP_CA_Loss)
  1967. mib_idx = LINUX_MIB_TCPFASTRETRANS;
  1968. else
  1969. mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
  1970. }
  1971. if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
  1972. continue;
  1973. if (tcp_retransmit_skb(sk, skb)) {
  1974. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
  1975. return;
  1976. }
  1977. NET_INC_STATS_BH(sock_net(sk), mib_idx);
  1978. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
  1979. tp->prr_out += tcp_skb_pcount(skb);
  1980. if (skb == tcp_write_queue_head(sk))
  1981. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  1982. inet_csk(sk)->icsk_rto,
  1983. TCP_RTO_MAX);
  1984. }
  1985. }
  1986. /* Send a fin. The caller locks the socket for us. This cannot be
  1987. * allowed to fail queueing a FIN frame under any circumstances.
  1988. */
  1989. void tcp_send_fin(struct sock *sk)
  1990. {
  1991. struct tcp_sock *tp = tcp_sk(sk);
  1992. struct sk_buff *skb = tcp_write_queue_tail(sk);
  1993. int mss_now;
  1994. /* Optimization, tack on the FIN if we have a queue of
  1995. * unsent frames. But be careful about outgoing SACKS
  1996. * and IP options.
  1997. */
  1998. mss_now = tcp_current_mss(sk);
  1999. if (tcp_send_head(sk) != NULL) {
  2000. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
  2001. TCP_SKB_CB(skb)->end_seq++;
  2002. tp->write_seq++;
  2003. } else {
  2004. /* Socket is locked, keep trying until memory is available. */
  2005. for (;;) {
  2006. skb = alloc_skb_fclone(MAX_TCP_HEADER,
  2007. sk->sk_allocation);
  2008. if (skb)
  2009. break;
  2010. yield();
  2011. }
  2012. /* Reserve space for headers and prepare control bits. */
  2013. skb_reserve(skb, MAX_TCP_HEADER);
  2014. /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
  2015. tcp_init_nondata_skb(skb, tp->write_seq,
  2016. TCPHDR_ACK | TCPHDR_FIN);
  2017. tcp_queue_skb(sk, skb);
  2018. }
  2019. __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
  2020. }
  2021. /* We get here when a process closes a file descriptor (either due to
  2022. * an explicit close() or as a byproduct of exit()'ing) and there
  2023. * was unread data in the receive queue. This behavior is recommended
  2024. * by RFC 2525, section 2.17. -DaveM
  2025. */
  2026. void tcp_send_active_reset(struct sock *sk, gfp_t priority)
  2027. {
  2028. struct sk_buff *skb;
  2029. /* NOTE: No TCP options attached and we never retransmit this. */
  2030. skb = alloc_skb(MAX_TCP_HEADER, priority);
  2031. if (!skb) {
  2032. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
  2033. return;
  2034. }
  2035. /* Reserve space for headers and prepare control bits. */
  2036. skb_reserve(skb, MAX_TCP_HEADER);
  2037. tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
  2038. TCPHDR_ACK | TCPHDR_RST);
  2039. /* Send it off. */
  2040. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2041. if (tcp_transmit_skb(sk, skb, 0, priority))
  2042. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
  2043. TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
  2044. }
  2045. /* Send a crossed SYN-ACK during socket establishment.
  2046. * WARNING: This routine must only be called when we have already sent
  2047. * a SYN packet that crossed the incoming SYN that caused this routine
  2048. * to get called. If this assumption fails then the initial rcv_wnd
  2049. * and rcv_wscale values will not be correct.
  2050. */
  2051. int tcp_send_synack(struct sock *sk)
  2052. {
  2053. struct sk_buff *skb;
  2054. skb = tcp_write_queue_head(sk);
  2055. if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
  2056. printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
  2057. return -EFAULT;
  2058. }
  2059. if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
  2060. if (skb_cloned(skb)) {
  2061. struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
  2062. if (nskb == NULL)
  2063. return -ENOMEM;
  2064. tcp_unlink_write_queue(skb, sk);
  2065. skb_header_release(nskb);
  2066. __tcp_add_write_queue_head(sk, nskb);
  2067. sk_wmem_free_skb(sk, skb);
  2068. sk->sk_wmem_queued += nskb->truesize;
  2069. sk_mem_charge(sk, nskb->truesize);
  2070. skb = nskb;
  2071. }
  2072. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
  2073. TCP_ECN_send_synack(tcp_sk(sk), skb);
  2074. }
  2075. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2076. return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2077. }
  2078. /* Prepare a SYN-ACK. */
  2079. struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
  2080. struct request_sock *req,
  2081. struct request_values *rvp)
  2082. {
  2083. struct tcp_out_options opts;
  2084. struct tcp_extend_values *xvp = tcp_xv(rvp);
  2085. struct inet_request_sock *ireq = inet_rsk(req);
  2086. struct tcp_sock *tp = tcp_sk(sk);
  2087. const struct tcp_cookie_values *cvp = tp->cookie_values;
  2088. struct tcphdr *th;
  2089. struct sk_buff *skb;
  2090. struct tcp_md5sig_key *md5;
  2091. int tcp_header_size;
  2092. int mss;
  2093. int s_data_desired = 0;
  2094. if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
  2095. s_data_desired = cvp->s_data_desired;
  2096. skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
  2097. if (skb == NULL)
  2098. return NULL;
  2099. /* Reserve space for headers. */
  2100. skb_reserve(skb, MAX_TCP_HEADER);
  2101. skb_dst_set(skb, dst_clone(dst));
  2102. mss = dst_metric_advmss(dst);
  2103. if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
  2104. mss = tp->rx_opt.user_mss;
  2105. if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
  2106. __u8 rcv_wscale;
  2107. /* Set this up on the first call only */
  2108. req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
  2109. /* limit the window selection if the user enforce a smaller rx buffer */
  2110. if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
  2111. (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
  2112. req->window_clamp = tcp_full_space(sk);
  2113. /* tcp_full_space because it is guaranteed to be the first packet */
  2114. tcp_select_initial_window(tcp_full_space(sk),
  2115. mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
  2116. &req->rcv_wnd,
  2117. &req->window_clamp,
  2118. ireq->wscale_ok,
  2119. &rcv_wscale,
  2120. dst_metric(dst, RTAX_INITRWND));
  2121. ireq->rcv_wscale = rcv_wscale;
  2122. }
  2123. memset(&opts, 0, sizeof(opts));
  2124. #ifdef CONFIG_SYN_COOKIES
  2125. if (unlikely(req->cookie_ts))
  2126. TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
  2127. else
  2128. #endif
  2129. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2130. tcp_header_size = tcp_synack_options(sk, req, mss,
  2131. skb, &opts, &md5, xvp)
  2132. + sizeof(*th);
  2133. skb_push(skb, tcp_header_size);
  2134. skb_reset_transport_header(skb);
  2135. th = tcp_hdr(skb);
  2136. memset(th, 0, sizeof(struct tcphdr));
  2137. th->syn = 1;
  2138. th->ack = 1;
  2139. TCP_ECN_make_synack(req, th);
  2140. th->source = ireq->loc_port;
  2141. th->dest = ireq->rmt_port;
  2142. /* Setting of flags are superfluous here for callers (and ECE is
  2143. * not even correctly set)
  2144. */
  2145. tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
  2146. TCPHDR_SYN | TCPHDR_ACK);
  2147. if (OPTION_COOKIE_EXTENSION & opts.options) {
  2148. if (s_data_desired) {
  2149. u8 *buf = skb_put(skb, s_data_desired);
  2150. /* copy data directly from the listening socket. */
  2151. memcpy(buf, cvp->s_data_payload, s_data_desired);
  2152. TCP_SKB_CB(skb)->end_seq += s_data_desired;
  2153. }
  2154. if (opts.hash_size > 0) {
  2155. __u32 workspace[SHA_WORKSPACE_WORDS];
  2156. u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
  2157. u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
  2158. /* Secret recipe depends on the Timestamp, (future)
  2159. * Sequence and Acknowledgment Numbers, Initiator
  2160. * Cookie, and others handled by IP variant caller.
  2161. */
  2162. *tail-- ^= opts.tsval;
  2163. *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
  2164. *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
  2165. /* recommended */
  2166. *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
  2167. *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
  2168. sha_transform((__u32 *)&xvp->cookie_bakery[0],
  2169. (char *)mess,
  2170. &workspace[0]);
  2171. opts.hash_location =
  2172. (__u8 *)&xvp->cookie_bakery[0];
  2173. }
  2174. }
  2175. th->seq = htonl(TCP_SKB_CB(skb)->seq);
  2176. th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
  2177. /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
  2178. th->window = htons(min(req->rcv_wnd, 65535U));
  2179. tcp_options_write((__be32 *)(th + 1), tp, &opts);
  2180. th->doff = (tcp_header_size >> 2);
  2181. TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
  2182. #ifdef CONFIG_TCP_MD5SIG
  2183. /* Okay, we have all we need - do the md5 hash if needed */
  2184. if (md5) {
  2185. tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
  2186. md5, NULL, req, skb);
  2187. }
  2188. #endif
  2189. return skb;
  2190. }
  2191. EXPORT_SYMBOL(tcp_make_synack);
  2192. /* Do all connect socket setups that can be done AF independent. */
  2193. static void tcp_connect_init(struct sock *sk)
  2194. {
  2195. const struct dst_entry *dst = __sk_dst_get(sk);
  2196. struct tcp_sock *tp = tcp_sk(sk);
  2197. __u8 rcv_wscale;
  2198. /* We'll fix this up when we get a response from the other end.
  2199. * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
  2200. */
  2201. tp->tcp_header_len = sizeof(struct tcphdr) +
  2202. (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
  2203. #ifdef CONFIG_TCP_MD5SIG
  2204. if (tp->af_specific->md5_lookup(sk, sk) != NULL)
  2205. tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
  2206. #endif
  2207. /* If user gave his TCP_MAXSEG, record it to clamp */
  2208. if (tp->rx_opt.user_mss)
  2209. tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
  2210. tp->max_window = 0;
  2211. tcp_mtup_init(sk);
  2212. tcp_sync_mss(sk, dst_mtu(dst));
  2213. if (!tp->window_clamp)
  2214. tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
  2215. tp->advmss = dst_metric_advmss(dst);
  2216. if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
  2217. tp->advmss = tp->rx_opt.user_mss;
  2218. tcp_initialize_rcv_mss(sk);
  2219. /* limit the window selection if the user enforce a smaller rx buffer */
  2220. if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
  2221. (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
  2222. tp->window_clamp = tcp_full_space(sk);
  2223. tcp_select_initial_window(tcp_full_space(sk),
  2224. tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
  2225. &tp->rcv_wnd,
  2226. &tp->window_clamp,
  2227. sysctl_tcp_window_scaling,
  2228. &rcv_wscale,
  2229. dst_metric(dst, RTAX_INITRWND));
  2230. tp->rx_opt.rcv_wscale = rcv_wscale;
  2231. tp->rcv_ssthresh = tp->rcv_wnd;
  2232. sk->sk_err = 0;
  2233. sock_reset_flag(sk, SOCK_DONE);
  2234. tp->snd_wnd = 0;
  2235. tcp_init_wl(tp, 0);
  2236. tcp_write_queue_purge(sk);
  2237. tp->snd_una = tp->write_seq;
  2238. tp->snd_sml = tp->write_seq;
  2239. tp->snd_up = tp->write_seq;
  2240. tp->rcv_nxt = 0;
  2241. tp->rcv_wup = 0;
  2242. tp->copied_seq = 0;
  2243. inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
  2244. inet_csk(sk)->icsk_retransmits = 0;
  2245. tcp_clear_retrans(tp);
  2246. }
  2247. /* Build a SYN and send it off. */
  2248. int tcp_connect(struct sock *sk)
  2249. {
  2250. struct tcp_sock *tp = tcp_sk(sk);
  2251. struct sk_buff *buff;
  2252. int err;
  2253. tcp_connect_init(sk);
  2254. buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
  2255. if (unlikely(buff == NULL))
  2256. return -ENOBUFS;
  2257. /* Reserve space for headers. */
  2258. skb_reserve(buff, MAX_TCP_HEADER);
  2259. tp->snd_nxt = tp->write_seq;
  2260. tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
  2261. TCP_ECN_send_syn(sk, buff);
  2262. /* Send it off. */
  2263. TCP_SKB_CB(buff)->when = tcp_time_stamp;
  2264. tp->retrans_stamp = TCP_SKB_CB(buff)->when;
  2265. skb_header_release(buff);
  2266. __tcp_add_write_queue_tail(sk, buff);
  2267. sk->sk_wmem_queued += buff->truesize;
  2268. sk_mem_charge(sk, buff->truesize);
  2269. tp->packets_out += tcp_skb_pcount(buff);
  2270. err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
  2271. if (err == -ECONNREFUSED)
  2272. return err;
  2273. /* We change tp->snd_nxt after the tcp_transmit_skb() call
  2274. * in order to make this packet get counted in tcpOutSegs.
  2275. */
  2276. tp->snd_nxt = tp->write_seq;
  2277. tp->pushed_seq = tp->write_seq;
  2278. TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
  2279. /* Timer for repeating the SYN until an answer. */
  2280. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  2281. inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  2282. return 0;
  2283. }
  2284. EXPORT_SYMBOL(tcp_connect);
  2285. /* Send out a delayed ack, the caller does the policy checking
  2286. * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
  2287. * for details.
  2288. */
  2289. void tcp_send_delayed_ack(struct sock *sk)
  2290. {
  2291. struct inet_connection_sock *icsk = inet_csk(sk);
  2292. int ato = icsk->icsk_ack.ato;
  2293. unsigned long timeout;
  2294. if (ato > TCP_DELACK_MIN) {
  2295. const struct tcp_sock *tp = tcp_sk(sk);
  2296. int max_ato = HZ / 2;
  2297. if (icsk->icsk_ack.pingpong ||
  2298. (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
  2299. max_ato = TCP_DELACK_MAX;
  2300. /* Slow path, intersegment interval is "high". */
  2301. /* If some rtt estimate is known, use it to bound delayed ack.
  2302. * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
  2303. * directly.
  2304. */
  2305. if (tp->srtt) {
  2306. int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
  2307. if (rtt < max_ato)
  2308. max_ato = rtt;
  2309. }
  2310. ato = min(ato, max_ato);
  2311. }
  2312. /* Stay within the limit we were given */
  2313. timeout = jiffies + ato;
  2314. /* Use new timeout only if there wasn't a older one earlier. */
  2315. if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
  2316. /* If delack timer was blocked or is about to expire,
  2317. * send ACK now.
  2318. */
  2319. if (icsk->icsk_ack.blocked ||
  2320. time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
  2321. tcp_send_ack(sk);
  2322. return;
  2323. }
  2324. if (!time_before(timeout, icsk->icsk_ack.timeout))
  2325. timeout = icsk->icsk_ack.timeout;
  2326. }
  2327. icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
  2328. icsk->icsk_ack.timeout = timeout;
  2329. sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
  2330. }
  2331. /* This routine sends an ack and also updates the window. */
  2332. void tcp_send_ack(struct sock *sk)
  2333. {
  2334. struct sk_buff *buff;
  2335. /* If we have been reset, we may not send again. */
  2336. if (sk->sk_state == TCP_CLOSE)
  2337. return;
  2338. /* We are not putting this on the write queue, so
  2339. * tcp_transmit_skb() will set the ownership to this
  2340. * sock.
  2341. */
  2342. buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
  2343. if (buff == NULL) {
  2344. inet_csk_schedule_ack(sk);
  2345. inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
  2346. inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
  2347. TCP_DELACK_MAX, TCP_RTO_MAX);
  2348. return;
  2349. }
  2350. /* Reserve space for headers and prepare control bits. */
  2351. skb_reserve(buff, MAX_TCP_HEADER);
  2352. tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
  2353. /* Send it off, this clears delayed acks for us. */
  2354. TCP_SKB_CB(buff)->when = tcp_time_stamp;
  2355. tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
  2356. }
  2357. /* This routine sends a packet with an out of date sequence
  2358. * number. It assumes the other end will try to ack it.
  2359. *
  2360. * Question: what should we make while urgent mode?
  2361. * 4.4BSD forces sending single byte of data. We cannot send
  2362. * out of window data, because we have SND.NXT==SND.MAX...
  2363. *
  2364. * Current solution: to send TWO zero-length segments in urgent mode:
  2365. * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
  2366. * out-of-date with SND.UNA-1 to probe window.
  2367. */
  2368. static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
  2369. {
  2370. struct tcp_sock *tp = tcp_sk(sk);
  2371. struct sk_buff *skb;
  2372. /* We don't queue it, tcp_transmit_skb() sets ownership. */
  2373. skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
  2374. if (skb == NULL)
  2375. return -1;
  2376. /* Reserve space for headers and set control bits. */
  2377. skb_reserve(skb, MAX_TCP_HEADER);
  2378. /* Use a previous sequence. This should cause the other
  2379. * end to send an ack. Don't queue or clone SKB, just
  2380. * send it.
  2381. */
  2382. tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
  2383. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2384. return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
  2385. }
  2386. /* Initiate keepalive or window probe from timer. */
  2387. int tcp_write_wakeup(struct sock *sk)
  2388. {
  2389. struct tcp_sock *tp = tcp_sk(sk);
  2390. struct sk_buff *skb;
  2391. if (sk->sk_state == TCP_CLOSE)
  2392. return -1;
  2393. if ((skb = tcp_send_head(sk)) != NULL &&
  2394. before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
  2395. int err;
  2396. unsigned int mss = tcp_current_mss(sk);
  2397. unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
  2398. if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
  2399. tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
  2400. /* We are probing the opening of a window
  2401. * but the window size is != 0
  2402. * must have been a result SWS avoidance ( sender )
  2403. */
  2404. if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
  2405. skb->len > mss) {
  2406. seg_size = min(seg_size, mss);
  2407. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
  2408. if (tcp_fragment(sk, skb, seg_size, mss))
  2409. return -1;
  2410. } else if (!tcp_skb_pcount(skb))
  2411. tcp_set_skb_tso_segs(sk, skb, mss);
  2412. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
  2413. TCP_SKB_CB(skb)->when = tcp_time_stamp;
  2414. err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
  2415. if (!err)
  2416. tcp_event_new_data_sent(sk, skb);
  2417. return err;
  2418. } else {
  2419. if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
  2420. tcp_xmit_probe_skb(sk, 1);
  2421. return tcp_xmit_probe_skb(sk, 0);
  2422. }
  2423. }
  2424. /* A window probe timeout has occurred. If window is not closed send
  2425. * a partial packet else a zero probe.
  2426. */
  2427. void tcp_send_probe0(struct sock *sk)
  2428. {
  2429. struct inet_connection_sock *icsk = inet_csk(sk);
  2430. struct tcp_sock *tp = tcp_sk(sk);
  2431. int err;
  2432. err = tcp_write_wakeup(sk);
  2433. if (tp->packets_out || !tcp_send_head(sk)) {
  2434. /* Cancel probe timer, if it is not required. */
  2435. icsk->icsk_probes_out = 0;
  2436. icsk->icsk_backoff = 0;
  2437. return;
  2438. }
  2439. if (err <= 0) {
  2440. if (icsk->icsk_backoff < sysctl_tcp_retries2)
  2441. icsk->icsk_backoff++;
  2442. icsk->icsk_probes_out++;
  2443. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2444. min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
  2445. TCP_RTO_MAX);
  2446. } else {
  2447. /* If packet was not sent due to local congestion,
  2448. * do not backoff and do not remember icsk_probes_out.
  2449. * Let local senders to fight for local resources.
  2450. *
  2451. * Use accumulated backoff yet.
  2452. */
  2453. if (!icsk->icsk_probes_out)
  2454. icsk->icsk_probes_out = 1;
  2455. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2456. min(icsk->icsk_rto << icsk->icsk_backoff,
  2457. TCP_RESOURCE_PROBE_INTERVAL),
  2458. TCP_RTO_MAX);
  2459. }
  2460. }