1/*- 2 * Copyright (c) 2012 Chelsio Communications, Inc. 3 * All rights reserved. 4 * Written by: Navdeep Parhar <np@FreeBSD.org> 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28#include <sys/cdefs.h> 29__FBSDID("$FreeBSD$"); 30 31#include "opt_inet.h" 32 33#ifdef TCP_OFFLOAD 34#include <sys/param.h> 35#include <sys/types.h> 36#include <sys/kernel.h> 37#include <sys/ktr.h> 38#include <sys/module.h> 39#include <sys/protosw.h> 40#include <sys/domain.h> 41#include <sys/socket.h> 42#include <sys/socketvar.h> 43#include <sys/sglist.h> 44#include <netinet/in.h> 45#include <netinet/in_pcb.h> 46#include <netinet/ip.h> 47#include <netinet/tcp_var.h> 48#define TCPSTATES 49#include <netinet/tcp_fsm.h> 50#include <netinet/tcp_seq.h> 51#include <netinet/toecore.h> 52 53#include "common/common.h" 54#include "common/t4_msg.h" 55#include "common/t4_regs.h" 56#include "common/t4_tcb.h" 57#include "tom/t4_tom_l2t.h" 58#include "tom/t4_tom.h" 59 60VNET_DECLARE(int, tcp_do_autosndbuf); 61#define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) 62VNET_DECLARE(int, tcp_autosndbuf_inc); 63#define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) 64VNET_DECLARE(int, tcp_autosndbuf_max); 65#define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) 66VNET_DECLARE(int, tcp_do_autorcvbuf); 67#define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf) 68VNET_DECLARE(int, tcp_autorcvbuf_inc); 69#define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc) 70VNET_DECLARE(int, tcp_autorcvbuf_max); 71#define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max) 72 73void 74send_flowc_wr(struct toepcb *toep, struct flowc_tx_params *ftxp) 75{ 76 struct wrqe *wr; 77 struct fw_flowc_wr *flowc; 78 unsigned int nparams = ftxp ? 8 : 6, flowclen; 79 struct port_info *pi = toep->port; 80 struct adapter *sc = pi->adapter; 81 unsigned int pfvf = G_FW_VIID_PFN(pi->viid) << S_FW_VIID_PFN; 82 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; 83 84 KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT), 85 ("%s: flowc for tid %u sent already", __func__, toep->tid)); 86 87 CTR2(KTR_CXGBE, "%s: tid %u", __func__, toep->tid); 88 89 flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval); 90 91 wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq); 92 if (wr == NULL) { 93 /* XXX */ 94 panic("%s: allocation failure.", __func__); 95 } 96 flowc = wrtod(wr); 97 memset(flowc, 0, wr->wr_len); 98 99 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | 100 V_FW_FLOWC_WR_NPARAMS(nparams)); 101 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) | 102 V_FW_WR_FLOWID(toep->tid)); 103 104 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN; 105 flowc->mnemval[0].val = htobe32(pfvf); 106 flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH; 107 flowc->mnemval[1].val = htobe32(pi->tx_chan); 108 flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT; 109 flowc->mnemval[2].val = htobe32(pi->tx_chan); 110 flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID; 111 flowc->mnemval[3].val = htobe32(toep->ofld_rxq->iq.abs_id); 112 if (ftxp) { 113 uint32_t sndbuf = min(ftxp->snd_space, sc->tt.sndbuf); 114 115 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDNXT; 116 flowc->mnemval[4].val = htobe32(ftxp->snd_nxt); 117 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_RCVNXT; 118 flowc->mnemval[5].val = htobe32(ftxp->rcv_nxt); 119 flowc->mnemval[6].mnemonic = FW_FLOWC_MNEM_SNDBUF; 120 flowc->mnemval[6].val = htobe32(sndbuf); 121 flowc->mnemval[7].mnemonic = FW_FLOWC_MNEM_MSS; 122 flowc->mnemval[7].val = htobe32(ftxp->mss); 123 } else { 124 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF; 125 flowc->mnemval[4].val = htobe32(512); 126 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS; 127 flowc->mnemval[5].val = htobe32(512); 128 } 129 130 txsd->tx_credits = howmany(flowclen, 16); 131 txsd->plen = 0; 132 KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0, 133 ("%s: not enough credits (%d)", __func__, toep->tx_credits)); 134 toep->tx_credits -= txsd->tx_credits; 135 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) 136 toep->txsd_pidx = 0; 137 toep->txsd_avail--; 138 139 toep->flags |= TPF_FLOWC_WR_SENT; 140 t4_wrq_tx(sc, wr); 141} 142 143void 144send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt) 145{ 146 struct wrqe *wr; 147 struct cpl_abort_req *req; 148 int tid = toep->tid; 149 struct inpcb *inp = toep->inp; 150 struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */ 151 152 INP_WLOCK_ASSERT(inp); 153 154 CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s", 155 __func__, toep->tid, 156 inp->inp_flags & INP_DROPPED ? "inp dropped" : 157 tcpstates[tp->t_state], 158 toep->flags, inp->inp_flags, 159 toep->flags & TPF_ABORT_SHUTDOWN ? 160 " (abort already in progress)" : ""); 161 162 if (toep->flags & TPF_ABORT_SHUTDOWN) 163 return; /* abort already in progress */ 164 165 toep->flags |= TPF_ABORT_SHUTDOWN; 166 167 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 168 ("%s: flowc_wr not sent for tid %d.", __func__, tid)); 169 170 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq); 171 if (wr == NULL) { 172 /* XXX */ 173 panic("%s: allocation failure.", __func__); 174 } 175 req = wrtod(wr); 176 177 INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid); 178 if (inp->inp_flags & INP_DROPPED) 179 req->rsvd0 = htobe32(snd_nxt); 180 else 181 req->rsvd0 = htobe32(tp->snd_nxt); 182 req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT); 183 req->cmd = CPL_ABORT_SEND_RST; 184 185 /* 186 * XXX: What's the correct way to tell that the inp hasn't been detached 187 * from its socket? Should I even be flushing the snd buffer here? 188 */ 189 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { 190 struct socket *so = inp->inp_socket; 191 192 if (so != NULL) /* because I'm not sure. See comment above */ 193 sbflush(&so->so_snd); 194 } 195 196 t4_l2t_send(sc, wr, toep->l2te); 197} 198 199/* 200 * Called when a connection is established to translate the TCP options 201 * reported by HW to FreeBSD's native format. 202 */ 203static void 204assign_rxopt(struct tcpcb *tp, unsigned int opt) 205{ 206 struct toepcb *toep = tp->t_toe; 207 struct adapter *sc = td_adapter(toep->td); 208 209 INP_LOCK_ASSERT(tp->t_inpcb); 210 211 tp->t_maxseg = tp->t_maxopd = sc->params.mtus[G_TCPOPT_MSS(opt)] - 40; 212 213 if (G_TCPOPT_TSTAMP(opt)) { 214 tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */ 215 tp->ts_recent = 0; /* hmmm */ 216 tp->ts_recent_age = tcp_ts_getticks(); 217 tp->t_maxseg -= TCPOLEN_TSTAMP_APPA; 218 } 219 220 if (G_TCPOPT_SACK(opt)) 221 tp->t_flags |= TF_SACK_PERMIT; /* should already be set */ 222 else 223 tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */ 224 225 if (G_TCPOPT_WSCALE_OK(opt)) 226 tp->t_flags |= TF_RCVD_SCALE; 227 228 /* Doing window scaling? */ 229 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 230 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 231 tp->rcv_scale = tp->request_r_scale; 232 tp->snd_scale = G_TCPOPT_SND_WSCALE(opt); 233 } 234} 235 236/* 237 * Completes some final bits of initialization for just established connections 238 * and changes their state to TCPS_ESTABLISHED. 239 * 240 * The ISNs are from after the exchange of SYNs. i.e., the true ISN + 1. 241 */ 242void 243make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn, 244 uint16_t opt) 245{ 246 struct inpcb *inp = toep->inp; 247 struct socket *so = inp->inp_socket; 248 struct tcpcb *tp = intotcpcb(inp); 249 long bufsize; 250 uint32_t iss = be32toh(snd_isn) - 1; /* true ISS */ 251 uint32_t irs = be32toh(rcv_isn) - 1; /* true IRS */ 252 uint16_t tcpopt = be16toh(opt); 253 struct flowc_tx_params ftxp; 254 255 INP_WLOCK_ASSERT(inp); 256 KASSERT(tp->t_state == TCPS_SYN_SENT || 257 tp->t_state == TCPS_SYN_RECEIVED, 258 ("%s: TCP state %s", __func__, tcpstates[tp->t_state])); 259 260 CTR4(KTR_CXGBE, "%s: tid %d, toep %p, inp %p", 261 __func__, toep->tid, toep, inp); 262 263 tp->t_state = TCPS_ESTABLISHED; 264 tp->t_starttime = ticks; 265 TCPSTAT_INC(tcps_connects); 266 267 tp->irs = irs; 268 tcp_rcvseqinit(tp); 269 tp->rcv_wnd = toep->rx_credits << 10; 270 tp->rcv_adv += tp->rcv_wnd; 271 tp->last_ack_sent = tp->rcv_nxt; 272 273 /* 274 * If we were unable to send all rx credits via opt0, save the remainder 275 * in rx_credits so that they can be handed over with the next credit 276 * update. 277 */ 278 SOCKBUF_LOCK(&so->so_rcv); 279 bufsize = select_rcv_wnd(so); 280 SOCKBUF_UNLOCK(&so->so_rcv); 281 toep->rx_credits = bufsize - tp->rcv_wnd; 282 283 tp->iss = iss; 284 tcp_sendseqinit(tp); 285 tp->snd_una = iss + 1; 286 tp->snd_nxt = iss + 1; 287 tp->snd_max = iss + 1; 288 289 assign_rxopt(tp, tcpopt); 290 291 SOCKBUF_LOCK(&so->so_snd); 292 if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf) 293 bufsize = V_tcp_autosndbuf_max; 294 else 295 bufsize = sbspace(&so->so_snd); 296 SOCKBUF_UNLOCK(&so->so_snd); 297 298 ftxp.snd_nxt = tp->snd_nxt; 299 ftxp.rcv_nxt = tp->rcv_nxt; 300 ftxp.snd_space = bufsize; 301 ftxp.mss = tp->t_maxseg; 302 send_flowc_wr(toep, &ftxp); 303 304 soisconnected(so); 305} 306 307static int 308send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits) 309{ 310 struct wrqe *wr; 311 struct cpl_rx_data_ack *req; 312 uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1); 313 314 KASSERT(credits >= 0, ("%s: %d credits", __func__, credits)); 315 316 wr = alloc_wrqe(sizeof(*req), toep->ctrlq); 317 if (wr == NULL) 318 return (0); 319 req = wrtod(wr); 320 321 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid); 322 req->credit_dack = htobe32(dack | V_RX_CREDITS(credits)); 323 324 t4_wrq_tx(sc, wr); 325 return (credits); 326} 327 328void 329t4_rcvd(struct toedev *tod, struct tcpcb *tp) 330{ 331 struct adapter *sc = tod->tod_softc; 332 struct inpcb *inp = tp->t_inpcb; 333 struct socket *so = inp->inp_socket; 334 struct sockbuf *sb = &so->so_rcv; 335 struct toepcb *toep = tp->t_toe; 336 int credits; 337 338 INP_WLOCK_ASSERT(inp); 339 340 SOCKBUF_LOCK(sb); 341 KASSERT(toep->sb_cc >= sb->sb_cc, 342 ("%s: sb %p has more data (%d) than last time (%d).", 343 __func__, sb, sb->sb_cc, toep->sb_cc)); 344 toep->rx_credits += toep->sb_cc - sb->sb_cc; 345 toep->sb_cc = sb->sb_cc; 346 credits = toep->rx_credits; 347 SOCKBUF_UNLOCK(sb); 348 349 if (credits > 0 && 350 (credits + 16384 >= tp->rcv_wnd || credits >= 15 * 1024)) { 351 352 credits = send_rx_credits(sc, toep, credits); 353 SOCKBUF_LOCK(sb); 354 toep->rx_credits -= credits; 355 SOCKBUF_UNLOCK(sb); 356 tp->rcv_wnd += credits; 357 tp->rcv_adv += credits; 358 } 359} 360 361/* 362 * Close a connection by sending a CPL_CLOSE_CON_REQ message. 363 */ 364static int 365close_conn(struct adapter *sc, struct toepcb *toep) 366{ 367 struct wrqe *wr; 368 struct cpl_close_con_req *req; 369 unsigned int tid = toep->tid; 370 371 CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid, 372 toep->flags & TPF_FIN_SENT ? ", IGNORED" : ""); 373 374 if (toep->flags & TPF_FIN_SENT) 375 return (0); 376 377 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 378 ("%s: flowc_wr not sent for tid %u.", __func__, tid)); 379 380 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq); 381 if (wr == NULL) { 382 /* XXX */ 383 panic("%s: allocation failure.", __func__); 384 } 385 req = wrtod(wr); 386 387 req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) | 388 V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr))); 389 req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) | 390 V_FW_WR_FLOWID(tid)); 391 req->wr.wr_lo = cpu_to_be64(0); 392 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid)); 393 req->rsvd = 0; 394 395 toep->flags |= TPF_FIN_SENT; 396 toep->flags &= ~TPF_SEND_FIN; 397 t4_l2t_send(sc, wr, toep->l2te); 398 399 return (0); 400} 401 402#define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16) 403#define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16)) 404 405/* Maximum amount of immediate data we could stuff in a WR */ 406static inline int 407max_imm_payload(int tx_credits) 408{ 409 const int n = 2; /* Use only up to 2 desc for imm. data WR */ 410 411 KASSERT(tx_credits >= 0 && 412 tx_credits <= MAX_OFLD_TX_CREDITS, 413 ("%s: %d credits", __func__, tx_credits)); 414 415 if (tx_credits < MIN_OFLD_TX_CREDITS) 416 return (0); 417 418 if (tx_credits >= (n * EQ_ESIZE) / 16) 419 return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr)); 420 else 421 return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr)); 422} 423 424/* Maximum number of SGL entries we could stuff in a WR */ 425static inline int 426max_dsgl_nsegs(int tx_credits) 427{ 428 int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */ 429 int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS; 430 431 KASSERT(tx_credits >= 0 && 432 tx_credits <= MAX_OFLD_TX_CREDITS, 433 ("%s: %d credits", __func__, tx_credits)); 434 435 if (tx_credits < MIN_OFLD_TX_CREDITS) 436 return (0); 437 438 nseg += 2 * (sge_pair_credits * 16 / 24); 439 if ((sge_pair_credits * 16) % 24 == 16) 440 nseg++; 441 442 return (nseg); 443} 444 445static inline void 446write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen, 447 unsigned int plen, uint8_t credits, int shove) 448{ 449 struct fw_ofld_tx_data_wr *txwr = dst; 450 451 txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) | 452 V_FW_WR_IMMDLEN(immdlen)); 453 txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) | 454 V_FW_WR_LEN16(credits)); 455 txwr->lsodisable_to_proxy = 456 htobe32(V_FW_OFLD_TX_DATA_WR_ULPMODE(toep->ulp_mode) | 457 V_FW_OFLD_TX_DATA_WR_URGENT(0) | /* XXX */ 458 V_FW_OFLD_TX_DATA_WR_SHOVE(shove)); 459 txwr->plen = htobe32(plen); 460} 461 462/* 463 * Generate a DSGL from a starting mbuf. The total number of segments and the 464 * maximum segments in any one mbuf are provided. 465 */ 466static void 467write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n) 468{ 469 struct mbuf *m; 470 struct ulptx_sgl *usgl = dst; 471 int i, j, rc; 472 struct sglist sg; 473 struct sglist_seg segs[n]; 474 475 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__)); 476 477 sglist_init(&sg, n, segs); 478 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | 479 V_ULPTX_NSGE(nsegs)); 480 481 i = -1; 482 for (m = start; m != stop; m = m->m_next) { 483 rc = sglist_append(&sg, mtod(m, void *), m->m_len); 484 if (__predict_false(rc != 0)) 485 panic("%s: sglist_append %d", __func__, rc); 486 487 for (j = 0; j < sg.sg_nseg; i++, j++) { 488 if (i < 0) { 489 usgl->len0 = htobe32(segs[j].ss_len); 490 usgl->addr0 = htobe64(segs[j].ss_paddr); 491 } else { 492 usgl->sge[i / 2].len[i & 1] = 493 htobe32(segs[j].ss_len); 494 usgl->sge[i / 2].addr[i & 1] = 495 htobe64(segs[j].ss_paddr); 496 } 497#ifdef INVARIANTS 498 nsegs--; 499#endif 500 } 501 sglist_reset(&sg); 502 } 503 if (i & 1) 504 usgl->sge[i / 2].len[1] = htobe32(0); 505 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p", 506 __func__, nsegs, start, stop)); 507} 508 509/* 510 * Max number of SGL entries an offload tx work request can have. This is 41 511 * (1 + 40) for a full 512B work request. 512 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40) 513 */ 514#define OFLD_SGL_LEN (41) 515 516/* 517 * Send data and/or a FIN to the peer. 518 * 519 * The socket's so_snd buffer consists of a stream of data starting with sb_mb 520 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that 521 * was transmitted. 522 * 523 * drop indicates the number of bytes that should be dropped from the head of 524 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating 525 * contention on the send buffer lock (before this change it used to do 526 * sowwakeup and then t4_push_frames right after that when recovering from tx 527 * stalls). When drop is set this function MUST drop the bytes and wake up any 528 * writers. 529 */ 530static void 531t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop) 532{ 533 struct mbuf *sndptr, *m, *sb_sndptr; 534 struct fw_ofld_tx_data_wr *txwr; 535 struct wrqe *wr; 536 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf; 537 struct inpcb *inp = toep->inp; 538 struct tcpcb *tp = intotcpcb(inp); 539 struct socket *so = inp->inp_socket; 540 struct sockbuf *sb = &so->so_snd; 541 int tx_credits, shove, compl, space, sowwakeup; 542 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; 543 544 INP_WLOCK_ASSERT(inp); 545 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 546 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); 547 548 if (__predict_false(toep->ulp_mode != ULP_MODE_NONE && 549 toep->ulp_mode != ULP_MODE_TCPDDP)) 550 CXGBE_UNIMPLEMENTED("ulp_mode"); 551 552 /* 553 * This function doesn't resume by itself. Someone else must clear the 554 * flag and call this function. 555 */ 556 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { 557 KASSERT(drop == 0, 558 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); 559 return; 560 } 561 562 do { 563 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); 564 max_imm = max_imm_payload(tx_credits); 565 max_nsegs = max_dsgl_nsegs(tx_credits); 566 567 SOCKBUF_LOCK(sb); 568 sowwakeup = drop; 569 if (drop) { 570 sbdrop_locked(sb, drop); 571 drop = 0; 572 } 573 sb_sndptr = sb->sb_sndptr; 574 sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb; 575 plen = 0; 576 nsegs = 0; 577 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */ 578 for (m = sndptr; m != NULL; m = m->m_next) { 579 int n = sglist_count(mtod(m, void *), m->m_len); 580 581 nsegs += n; 582 plen += m->m_len; 583 584 /* This mbuf sent us _over_ the nsegs limit, back out */ 585 if (plen > max_imm && nsegs > max_nsegs) { 586 nsegs -= n; 587 plen -= m->m_len; 588 if (plen == 0) { 589 /* Too few credits */ 590 toep->flags |= TPF_TX_SUSPENDED; 591 if (sowwakeup) 592 sowwakeup_locked(so); 593 else 594 SOCKBUF_UNLOCK(sb); 595 SOCKBUF_UNLOCK_ASSERT(sb); 596 return; 597 } 598 break; 599 } 600 601 if (max_nsegs_1mbuf < n) 602 max_nsegs_1mbuf = n; 603 sb_sndptr = m; /* new sb->sb_sndptr if all goes well */ 604 605 /* This mbuf put us right at the max_nsegs limit */ 606 if (plen > max_imm && nsegs == max_nsegs) { 607 m = m->m_next; 608 break; 609 } 610 } 611 612 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME); 613 space = sbspace(sb); 614 615 if (space <= sb->sb_hiwat * 3 / 8 && 616 toep->plen_nocompl + plen >= sb->sb_hiwat / 4) 617 compl = 1; 618 else 619 compl = 0; 620 621 if (sb->sb_flags & SB_AUTOSIZE && 622 V_tcp_do_autosndbuf && 623 sb->sb_hiwat < V_tcp_autosndbuf_max && 624 space < sb->sb_hiwat / 8) { 625 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc, 626 V_tcp_autosndbuf_max); 627 628 if (!sbreserve_locked(sb, newsize, so, NULL)) 629 sb->sb_flags &= ~SB_AUTOSIZE; 630 else 631 sowwakeup = 1; /* room available */ 632 } 633 if (sowwakeup) 634 sowwakeup_locked(so); 635 else 636 SOCKBUF_UNLOCK(sb); 637 SOCKBUF_UNLOCK_ASSERT(sb); 638 639 /* nothing to send */ 640 if (plen == 0) { 641 KASSERT(m == NULL, 642 ("%s: nothing to send, but m != NULL", __func__)); 643 break; 644 } 645 646 if (__predict_false(toep->flags & TPF_FIN_SENT)) 647 panic("%s: excess tx.", __func__); 648 649 if (plen <= max_imm) { 650 651 /* Immediate data tx */ 652 653 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16), 654 toep->ofld_txq); 655 if (wr == NULL) { 656 /* XXX: how will we recover from this? */ 657 toep->flags |= TPF_TX_SUSPENDED; 658 return; 659 } 660 txwr = wrtod(wr); 661 credits = howmany(wr->wr_len, 16); 662 write_tx_wr(txwr, toep, plen, plen, credits, shove); 663 m_copydata(sndptr, 0, plen, (void *)(txwr + 1)); 664 nsegs = 0; 665 } else { 666 int wr_len; 667 668 /* DSGL tx */ 669 670 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) + 671 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; 672 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq); 673 if (wr == NULL) { 674 /* XXX: how will we recover from this? */ 675 toep->flags |= TPF_TX_SUSPENDED; 676 return; 677 } 678 txwr = wrtod(wr); 679 credits = howmany(wr_len, 16); 680 write_tx_wr(txwr, toep, 0, plen, credits, shove); 681 write_tx_sgl(txwr + 1, sndptr, m, nsegs, 682 max_nsegs_1mbuf); 683 if (wr_len & 0xf) { 684 uint64_t *pad = (uint64_t *) 685 ((uintptr_t)txwr + wr_len); 686 *pad = 0; 687 } 688 } 689 690 KASSERT(toep->tx_credits >= credits, 691 ("%s: not enough credits", __func__)); 692 693 toep->tx_credits -= credits; 694 toep->tx_nocompl += credits; 695 toep->plen_nocompl += plen; 696 if (toep->tx_credits <= toep->tx_total * 3 / 8 && 697 toep->tx_nocompl >= toep->tx_total / 4) 698 compl = 1; 699 700 if (compl) { 701 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL); 702 toep->tx_nocompl = 0; 703 toep->plen_nocompl = 0; 704 } 705 706 tp->snd_nxt += plen; 707 tp->snd_max += plen; 708 709 SOCKBUF_LOCK(sb); 710 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__)); 711 sb->sb_sndptr = sb_sndptr; 712 SOCKBUF_UNLOCK(sb); 713 714 toep->flags |= TPF_TX_DATA_SENT; 715 if (toep->tx_credits < MIN_OFLD_TX_CREDITS) 716 toep->flags |= TPF_TX_SUSPENDED; 717 718 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); 719 txsd->plen = plen; 720 txsd->tx_credits = credits; 721 txsd++; 722 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { 723 toep->txsd_pidx = 0; 724 txsd = &toep->txsd[0]; 725 } 726 toep->txsd_avail--; 727 728 t4_l2t_send(sc, wr, toep->l2te); 729 } while (m != NULL); 730 731 /* Send a FIN if requested, but only if there's no more data to send */ 732 if (m == NULL && toep->flags & TPF_SEND_FIN) 733 close_conn(sc, toep); 734} 735 736int 737t4_tod_output(struct toedev *tod, struct tcpcb *tp) 738{ 739 struct adapter *sc = tod->tod_softc; 740#ifdef INVARIANTS 741 struct inpcb *inp = tp->t_inpcb; 742#endif 743 struct toepcb *toep = tp->t_toe; 744 745 INP_WLOCK_ASSERT(inp); 746 KASSERT((inp->inp_flags & INP_DROPPED) == 0, 747 ("%s: inp %p dropped.", __func__, inp)); 748 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 749 750 t4_push_frames(sc, toep, 0); 751 752 return (0); 753} 754 755int 756t4_send_fin(struct toedev *tod, struct tcpcb *tp) 757{ 758 struct adapter *sc = tod->tod_softc; 759#ifdef INVARIANTS 760 struct inpcb *inp = tp->t_inpcb; 761#endif 762 struct toepcb *toep = tp->t_toe; 763 764 INP_WLOCK_ASSERT(inp); 765 KASSERT((inp->inp_flags & INP_DROPPED) == 0, 766 ("%s: inp %p dropped.", __func__, inp)); 767 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 768 769 toep->flags |= TPF_SEND_FIN; 770 if (tp->t_state >= TCPS_ESTABLISHED) 771 t4_push_frames(sc, toep, 0); 772 773 return (0); 774} 775 776int 777t4_send_rst(struct toedev *tod, struct tcpcb *tp) 778{ 779 struct adapter *sc = tod->tod_softc; 780#if defined(INVARIANTS) 781 struct inpcb *inp = tp->t_inpcb; 782#endif 783 struct toepcb *toep = tp->t_toe; 784 785 INP_WLOCK_ASSERT(inp); 786 KASSERT((inp->inp_flags & INP_DROPPED) == 0, 787 ("%s: inp %p dropped.", __func__, inp)); 788 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 789 790 /* hmmmm */ 791 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 792 ("%s: flowc for tid %u [%s] not sent already", 793 __func__, toep->tid, tcpstates[tp->t_state])); 794 795 send_reset(sc, toep, 0); 796 return (0); 797} 798 799/* 800 * Peer has sent us a FIN. 801 */ 802static int 803do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 804{ 805 struct adapter *sc = iq->adapter; 806 const struct cpl_peer_close *cpl = (const void *)(rss + 1); 807 unsigned int tid = GET_TID(cpl); 808 struct toepcb *toep = lookup_tid(sc, tid); 809 struct inpcb *inp = toep->inp; 810 struct tcpcb *tp = NULL; 811 struct socket *so; 812 struct sockbuf *sb; 813#ifdef INVARIANTS 814 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 815#endif 816 817 KASSERT(opcode == CPL_PEER_CLOSE, 818 ("%s: unexpected opcode 0x%x", __func__, opcode)); 819 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 820 821 if (__predict_false(toep->flags & TPF_SYNQE)) { 822#ifdef INVARIANTS 823 struct synq_entry *synqe = (void *)toep; 824 825 INP_WLOCK(synqe->lctx->inp); 826 if (synqe->flags & TPF_SYNQE_HAS_L2TE) { 827 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, 828 ("%s: listen socket closed but tid %u not aborted.", 829 __func__, tid)); 830 } else { 831 /* 832 * do_pass_accept_req is still running and will 833 * eventually take care of this tid. 834 */ 835 } 836 INP_WUNLOCK(synqe->lctx->inp); 837#endif 838 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, 839 toep, toep->flags); 840 return (0); 841 } 842 843 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 844 845 INP_INFO_WLOCK(&V_tcbinfo); 846 INP_WLOCK(inp); 847 tp = intotcpcb(inp); 848 849 CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__, 850 tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp); 851 852 if (toep->flags & TPF_ABORT_SHUTDOWN) 853 goto done; 854 855 tp->rcv_nxt++; /* FIN */ 856 857 so = inp->inp_socket; 858 sb = &so->so_rcv; 859 SOCKBUF_LOCK(sb); 860 if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) { 861 m = get_ddp_mbuf(be32toh(cpl->rcv_nxt) - tp->rcv_nxt); 862 tp->rcv_nxt = be32toh(cpl->rcv_nxt); 863 toep->ddp_flags &= ~(DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE); 864 865 KASSERT(toep->sb_cc >= sb->sb_cc, 866 ("%s: sb %p has more data (%d) than last time (%d).", 867 __func__, sb, sb->sb_cc, toep->sb_cc)); 868 toep->rx_credits += toep->sb_cc - sb->sb_cc; 869#ifdef USE_DDP_RX_FLOW_CONTROL 870 toep->rx_credits -= m->m_len; /* adjust for F_RX_FC_DDP */ 871#endif 872 sbappendstream_locked(sb, m); 873 toep->sb_cc = sb->sb_cc; 874 } 875 socantrcvmore_locked(so); /* unlocks the sockbuf */ 876 877 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt), 878 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt, 879 be32toh(cpl->rcv_nxt))); 880 881 switch (tp->t_state) { 882 case TCPS_SYN_RECEIVED: 883 tp->t_starttime = ticks; 884 /* FALLTHROUGH */ 885 886 case TCPS_ESTABLISHED: 887 tp->t_state = TCPS_CLOSE_WAIT; 888 break; 889 890 case TCPS_FIN_WAIT_1: 891 tp->t_state = TCPS_CLOSING; 892 break; 893 894 case TCPS_FIN_WAIT_2: 895 tcp_twstart(tp); 896 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ 897 INP_INFO_WUNLOCK(&V_tcbinfo); 898 899 INP_WLOCK(inp); 900 final_cpl_received(toep); 901 return (0); 902 903 default: 904 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n", 905 __func__, tid, tp->t_state); 906 } 907done: 908 INP_WUNLOCK(inp); 909 INP_INFO_WUNLOCK(&V_tcbinfo); 910 return (0); 911} 912 913/* 914 * Peer has ACK'd our FIN. 915 */ 916static int 917do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss, 918 struct mbuf *m) 919{ 920 struct adapter *sc = iq->adapter; 921 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1); 922 unsigned int tid = GET_TID(cpl); 923 struct toepcb *toep = lookup_tid(sc, tid); 924 struct inpcb *inp = toep->inp; 925 struct tcpcb *tp = NULL; 926 struct socket *so = NULL; 927#ifdef INVARIANTS 928 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 929#endif 930 931 KASSERT(opcode == CPL_CLOSE_CON_RPL, 932 ("%s: unexpected opcode 0x%x", __func__, opcode)); 933 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 934 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 935 936 INP_INFO_WLOCK(&V_tcbinfo); 937 INP_WLOCK(inp); 938 tp = intotcpcb(inp); 939 940 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x", 941 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags); 942 943 if (toep->flags & TPF_ABORT_SHUTDOWN) 944 goto done; 945 946 so = inp->inp_socket; 947 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */ 948 949 switch (tp->t_state) { 950 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */ 951 tcp_twstart(tp); 952release: 953 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ 954 INP_INFO_WUNLOCK(&V_tcbinfo); 955 956 INP_WLOCK(inp); 957 final_cpl_received(toep); /* no more CPLs expected */ 958 959 return (0); 960 case TCPS_LAST_ACK: 961 if (tcp_close(tp)) 962 INP_WUNLOCK(inp); 963 goto release; 964 965 case TCPS_FIN_WAIT_1: 966 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 967 soisdisconnected(so); 968 tp->t_state = TCPS_FIN_WAIT_2; 969 break; 970 971 default: 972 log(LOG_ERR, 973 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n", 974 __func__, tid, tcpstates[tp->t_state]); 975 } 976done: 977 INP_WUNLOCK(inp); 978 INP_INFO_WUNLOCK(&V_tcbinfo); 979 return (0); 980} 981 982void 983send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid, 984 int rst_status) 985{ 986 struct wrqe *wr; 987 struct cpl_abort_rpl *cpl; 988 989 wr = alloc_wrqe(sizeof(*cpl), ofld_txq); 990 if (wr == NULL) { 991 /* XXX */ 992 panic("%s: allocation failure.", __func__); 993 } 994 cpl = wrtod(wr); 995 996 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid); 997 cpl->cmd = rst_status; 998 999 t4_wrq_tx(sc, wr); 1000} 1001 1002static int 1003abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason) 1004{ 1005 switch (abort_reason) { 1006 case CPL_ERR_BAD_SYN: 1007 case CPL_ERR_CONN_RESET: 1008 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET); 1009 case CPL_ERR_XMIT_TIMEDOUT: 1010 case CPL_ERR_PERSIST_TIMEDOUT: 1011 case CPL_ERR_FINWAIT2_TIMEDOUT: 1012 case CPL_ERR_KEEPALIVE_TIMEDOUT: 1013 return (ETIMEDOUT); 1014 default: 1015 return (EIO); 1016 } 1017} 1018 1019/* 1020 * TCP RST from the peer, timeout, or some other such critical error. 1021 */ 1022static int 1023do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1024{ 1025 struct adapter *sc = iq->adapter; 1026 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1); 1027 unsigned int tid = GET_TID(cpl); 1028 struct toepcb *toep = lookup_tid(sc, tid); 1029 struct sge_wrq *ofld_txq = toep->ofld_txq; 1030 struct inpcb *inp; 1031 struct tcpcb *tp; 1032#ifdef INVARIANTS 1033 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1034#endif 1035 1036 KASSERT(opcode == CPL_ABORT_REQ_RSS, 1037 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1038 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1039 1040 if (toep->flags & TPF_SYNQE) 1041 return (do_abort_req_synqe(iq, rss, m)); 1042 1043 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1044 1045 if (negative_advice(cpl->status)) { 1046 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)", 1047 __func__, cpl->status, tid, toep->flags); 1048 return (0); /* Ignore negative advice */ 1049 } 1050 1051 inp = toep->inp; 1052 INP_INFO_WLOCK(&V_tcbinfo); /* for tcp_close */ 1053 INP_WLOCK(inp); 1054 1055 tp = intotcpcb(inp); 1056 1057 CTR6(KTR_CXGBE, 1058 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d", 1059 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, 1060 inp->inp_flags, cpl->status); 1061 1062 /* 1063 * If we'd initiated an abort earlier the reply to it is responsible for 1064 * cleaning up resources. Otherwise we tear everything down right here 1065 * right now. We owe the T4 a CPL_ABORT_RPL no matter what. 1066 */ 1067 if (toep->flags & TPF_ABORT_SHUTDOWN) { 1068 INP_WUNLOCK(inp); 1069 goto done; 1070 } 1071 toep->flags |= TPF_ABORT_SHUTDOWN; 1072 1073 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { 1074 struct socket *so = inp->inp_socket; 1075 1076 if (so != NULL) 1077 so_error_set(so, abort_status_to_errno(tp, 1078 cpl->status)); 1079 tp = tcp_close(tp); 1080 if (tp == NULL) 1081 INP_WLOCK(inp); /* re-acquire */ 1082 } 1083 1084 final_cpl_received(toep); 1085done: 1086 INP_INFO_WUNLOCK(&V_tcbinfo); 1087 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST); 1088 return (0); 1089} 1090 1091/* 1092 * Reply to the CPL_ABORT_REQ (send_reset) 1093 */ 1094static int 1095do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1096{ 1097 struct adapter *sc = iq->adapter; 1098 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1); 1099 unsigned int tid = GET_TID(cpl); 1100 struct toepcb *toep = lookup_tid(sc, tid); 1101 struct inpcb *inp = toep->inp; 1102#ifdef INVARIANTS 1103 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1104#endif 1105 1106 KASSERT(opcode == CPL_ABORT_RPL_RSS, 1107 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1108 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1109 1110 if (toep->flags & TPF_SYNQE) 1111 return (do_abort_rpl_synqe(iq, rss, m)); 1112 1113 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1114 1115 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d", 1116 __func__, tid, toep, inp, cpl->status); 1117 1118 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, 1119 ("%s: wasn't expecting abort reply", __func__)); 1120 1121 INP_WLOCK(inp); 1122 final_cpl_received(toep); 1123 1124 return (0); 1125} 1126 1127static int 1128do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1129{ 1130 struct adapter *sc = iq->adapter; 1131 const struct cpl_rx_data *cpl = mtod(m, const void *); 1132 unsigned int tid = GET_TID(cpl); 1133 struct toepcb *toep = lookup_tid(sc, tid); 1134 struct inpcb *inp = toep->inp; 1135 struct tcpcb *tp; 1136 struct socket *so; 1137 struct sockbuf *sb; 1138 int len; 1139 uint32_t ddp_placed = 0; 1140 1141 if (__predict_false(toep->flags & TPF_SYNQE)) { 1142#ifdef INVARIANTS 1143 struct synq_entry *synqe = (void *)toep; 1144 1145 INP_WLOCK(synqe->lctx->inp); 1146 if (synqe->flags & TPF_SYNQE_HAS_L2TE) { 1147 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, 1148 ("%s: listen socket closed but tid %u not aborted.", 1149 __func__, tid)); 1150 } else { 1151 /* 1152 * do_pass_accept_req is still running and will 1153 * eventually take care of this tid. 1154 */ 1155 } 1156 INP_WUNLOCK(synqe->lctx->inp); 1157#endif 1158 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, 1159 toep, toep->flags); 1160 m_freem(m); 1161 return (0); 1162 } 1163 1164 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1165 1166 /* strip off CPL header */ 1167 m_adj(m, sizeof(*cpl)); 1168 len = m->m_pkthdr.len; 1169 1170 INP_WLOCK(inp); 1171 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) { 1172 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x", 1173 __func__, tid, len, inp->inp_flags); 1174 INP_WUNLOCK(inp); 1175 m_freem(m); 1176 return (0); 1177 } 1178 1179 tp = intotcpcb(inp); 1180 1181 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq))) 1182 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt; 1183 1184 tp->rcv_nxt += len; 1185 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__)); 1186 tp->rcv_wnd -= len; 1187 tp->t_rcvtime = ticks; 1188 1189 so = inp_inpcbtosocket(inp); 1190 sb = &so->so_rcv; 1191 SOCKBUF_LOCK(sb); 1192 1193 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) { 1194 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)", 1195 __func__, tid, len); 1196 m_freem(m); 1197 SOCKBUF_UNLOCK(sb); 1198 INP_WUNLOCK(inp); 1199 1200 INP_INFO_WLOCK(&V_tcbinfo); 1201 INP_WLOCK(inp); 1202 tp = tcp_drop(tp, ECONNRESET); 1203 if (tp) 1204 INP_WUNLOCK(inp); 1205 INP_INFO_WUNLOCK(&V_tcbinfo); 1206 1207 return (0); 1208 } 1209 1210 /* receive buffer autosize */ 1211 if (sb->sb_flags & SB_AUTOSIZE && 1212 V_tcp_do_autorcvbuf && 1213 sb->sb_hiwat < V_tcp_autorcvbuf_max && 1214 len > (sbspace(sb) / 8 * 7)) { 1215 unsigned int hiwat = sb->sb_hiwat; 1216 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc, 1217 V_tcp_autorcvbuf_max); 1218 1219 if (!sbreserve_locked(sb, newsize, so, NULL)) 1220 sb->sb_flags &= ~SB_AUTOSIZE; 1221 else 1222 toep->rx_credits += newsize - hiwat; 1223 } 1224 1225 if (toep->ulp_mode == ULP_MODE_TCPDDP) { 1226 int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off; 1227 1228 if (changed) { 1229 if (toep->ddp_flags & DDP_SC_REQ) 1230 toep->ddp_flags ^= DDP_ON | DDP_SC_REQ; 1231 else { 1232 KASSERT(cpl->ddp_off == 1, 1233 ("%s: DDP switched on by itself.", 1234 __func__)); 1235 1236 /* Fell out of DDP mode */ 1237 toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE | 1238 DDP_BUF1_ACTIVE); 1239 1240 if (ddp_placed) 1241 insert_ddp_data(toep, ddp_placed); 1242 } 1243 } 1244 1245 if ((toep->ddp_flags & DDP_OK) == 0 && 1246 time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) { 1247 toep->ddp_score = DDP_LOW_SCORE; 1248 toep->ddp_flags |= DDP_OK; 1249 CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u", 1250 __func__, tid, time_uptime); 1251 } 1252 1253 if (toep->ddp_flags & DDP_ON) { 1254 1255 /* 1256 * CPL_RX_DATA with DDP on can only be an indicate. Ask 1257 * soreceive to post a buffer or disable DDP. The 1258 * payload that arrived in this indicate is appended to 1259 * the socket buffer as usual. 1260 */ 1261 1262#if 0 1263 CTR5(KTR_CXGBE, 1264 "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)", 1265 __func__, tid, toep->flags, be32toh(cpl->seq), len); 1266#endif 1267 sb->sb_flags |= SB_DDP_INDICATE; 1268 } else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK && 1269 tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) { 1270 1271 /* 1272 * DDP allowed but isn't on (and a request to switch it 1273 * on isn't pending either), and conditions are ripe for 1274 * it to work. Switch it on. 1275 */ 1276 1277 enable_ddp(sc, toep); 1278 } 1279 } 1280 1281 KASSERT(toep->sb_cc >= sb->sb_cc, 1282 ("%s: sb %p has more data (%d) than last time (%d).", 1283 __func__, sb, sb->sb_cc, toep->sb_cc)); 1284 toep->rx_credits += toep->sb_cc - sb->sb_cc; 1285 sbappendstream_locked(sb, m); 1286 toep->sb_cc = sb->sb_cc; 1287 sorwakeup_locked(so); 1288 SOCKBUF_UNLOCK_ASSERT(sb); 1289 1290 INP_WUNLOCK(inp); 1291 return (0); 1292} 1293 1294#define S_CPL_FW4_ACK_OPCODE 24 1295#define M_CPL_FW4_ACK_OPCODE 0xff 1296#define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE) 1297#define G_CPL_FW4_ACK_OPCODE(x) \ 1298 (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE) 1299 1300#define S_CPL_FW4_ACK_FLOWID 0 1301#define M_CPL_FW4_ACK_FLOWID 0xffffff 1302#define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID) 1303#define G_CPL_FW4_ACK_FLOWID(x) \ 1304 (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID) 1305 1306#define S_CPL_FW4_ACK_CR 24 1307#define M_CPL_FW4_ACK_CR 0xff 1308#define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR) 1309#define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR) 1310 1311#define S_CPL_FW4_ACK_SEQVAL 0 1312#define M_CPL_FW4_ACK_SEQVAL 0x1 1313#define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL) 1314#define G_CPL_FW4_ACK_SEQVAL(x) \ 1315 (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL) 1316#define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U) 1317 1318static int 1319do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1320{ 1321 struct adapter *sc = iq->adapter; 1322 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1); 1323 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl))); 1324 struct toepcb *toep = lookup_tid(sc, tid); 1325 struct inpcb *inp; 1326 struct tcpcb *tp; 1327 struct socket *so; 1328 uint8_t credits = cpl->credits; 1329 struct ofld_tx_sdesc *txsd; 1330 int plen; 1331#ifdef INVARIANTS 1332 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl))); 1333#endif 1334 1335 /* 1336 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and 1337 * now this comes back carrying the credits for the flowc. 1338 */ 1339 if (__predict_false(toep->flags & TPF_SYNQE)) { 1340 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, 1341 ("%s: credits for a synq entry %p", __func__, toep)); 1342 return (0); 1343 } 1344 1345 inp = toep->inp; 1346 1347 KASSERT(opcode == CPL_FW4_ACK, 1348 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1349 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1350 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1351 1352 INP_WLOCK(inp); 1353 1354 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) { 1355 INP_WUNLOCK(inp); 1356 return (0); 1357 } 1358 1359 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0, 1360 ("%s: inp_flags 0x%x", __func__, inp->inp_flags)); 1361 1362 tp = intotcpcb(inp); 1363 1364 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) { 1365 tcp_seq snd_una = be32toh(cpl->snd_una); 1366 1367#ifdef INVARIANTS 1368 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) { 1369 log(LOG_ERR, 1370 "%s: unexpected seq# %x for TID %u, snd_una %x\n", 1371 __func__, snd_una, toep->tid, tp->snd_una); 1372 } 1373#endif 1374 1375 if (tp->snd_una != snd_una) { 1376 tp->snd_una = snd_una; 1377 tp->ts_recent_age = tcp_ts_getticks(); 1378 } 1379 } 1380 1381 so = inp->inp_socket; 1382 txsd = &toep->txsd[toep->txsd_cidx]; 1383 plen = 0; 1384 while (credits) { 1385 KASSERT(credits >= txsd->tx_credits, 1386 ("%s: too many (or partial) credits", __func__)); 1387 credits -= txsd->tx_credits; 1388 toep->tx_credits += txsd->tx_credits; 1389 plen += txsd->plen; 1390 txsd++; 1391 toep->txsd_avail++; 1392 KASSERT(toep->txsd_avail <= toep->txsd_total, 1393 ("%s: txsd avail > total", __func__)); 1394 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) { 1395 txsd = &toep->txsd[0]; 1396 toep->txsd_cidx = 0; 1397 } 1398 } 1399 1400 if (toep->tx_credits == toep->tx_total) { 1401 toep->tx_nocompl = 0; 1402 toep->plen_nocompl = 0; 1403 } 1404 1405 if (toep->flags & TPF_TX_SUSPENDED && 1406 toep->tx_credits >= toep->tx_total / 4) { 1407 toep->flags &= ~TPF_TX_SUSPENDED; 1408 t4_push_frames(sc, toep, plen); 1409 } else if (plen > 0) { 1410 struct sockbuf *sb = &so->so_snd; 1411 1412 SOCKBUF_LOCK(sb); 1413 sbdrop_locked(sb, plen); 1414 sowwakeup_locked(so); 1415 SOCKBUF_UNLOCK_ASSERT(sb); 1416 } 1417 1418 INP_WUNLOCK(inp); 1419 1420 return (0); 1421} 1422 1423static int 1424do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1425{ 1426 struct adapter *sc = iq->adapter; 1427 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1); 1428 unsigned int tid = GET_TID(cpl); 1429#ifdef INVARIANTS 1430 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1431#endif 1432 1433 KASSERT(opcode == CPL_SET_TCB_RPL, 1434 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1435 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1436 1437 if (is_ftid(sc, tid)) 1438 return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */ 1439 1440 CXGBE_UNIMPLEMENTED(__func__); 1441} 1442 1443void 1444t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl, 1445 uint16_t word, uint64_t mask, uint64_t val) 1446{ 1447 struct wrqe *wr; 1448 struct cpl_set_tcb_field *req; 1449 1450 wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq); 1451 if (wr == NULL) { 1452 /* XXX */ 1453 panic("%s: allocation failure.", __func__); 1454 } 1455 req = wrtod(wr); 1456 1457 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid); 1458 req->reply_ctrl = htobe16(V_NO_REPLY(1) | 1459 V_QUEUENO(toep->ofld_rxq->iq.abs_id)); 1460 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0)); 1461 req->mask = htobe64(mask); 1462 req->val = htobe64(val); 1463 1464 t4_wrq_tx(sc, wr); 1465} 1466 1467void 1468t4_init_cpl_io_handlers(struct adapter *sc) 1469{ 1470 1471 t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close); 1472 t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl); 1473 t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req); 1474 t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl); 1475 t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data); 1476 t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack); 1477 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl); 1478} 1479 1480void 1481t4_uninit_cpl_io_handlers(struct adapter *sc) 1482{ 1483 1484 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl); 1485} 1486#endif 1487