2 * Copyright (c) 2012 Chelsio Communications, Inc.
4 * Written by: Navdeep Parhar <np@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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.
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
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
32 #include "opt_inet6.h"
33 #include "opt_ratelimit.h"
35 #include <sys/param.h>
36 #include <sys/types.h>
37 #include <sys/systm.h>
38 #include <sys/kernel.h>
41 #include <sys/limits.h>
42 #include <sys/module.h>
43 #include <sys/protosw.h>
44 #include <sys/domain.h>
45 #include <sys/refcount.h>
46 #include <sys/rmlock.h>
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/taskqueue.h>
51 #include <net/if_var.h>
52 #include <netinet/in.h>
53 #include <netinet/in_pcb.h>
54 #include <netinet/in_var.h>
55 #include <netinet/ip.h>
56 #include <netinet/ip6.h>
57 #include <netinet6/scope6_var.h>
59 #include <netinet/tcp_fsm.h>
60 #include <netinet/tcp_var.h>
61 #include <netinet/toecore.h>
64 #include "common/common.h"
65 #include "common/t4_msg.h"
66 #include "common/t4_regs.h"
67 #include "common/t4_regs_values.h"
68 #include "common/t4_tcb.h"
69 #include "tom/t4_tom_l2t.h"
70 #include "tom/t4_tom.h"
72 static struct protosw toe_protosw;
73 static struct pr_usrreqs toe_usrreqs;
75 static struct protosw toe6_protosw;
76 static struct pr_usrreqs toe6_usrreqs;
79 static int t4_tom_mod_load(void);
80 static int t4_tom_mod_unload(void);
81 static int t4_tom_modevent(module_t, int, void *);
83 /* ULD ops and helpers */
84 static int t4_tom_activate(struct adapter *);
85 static int t4_tom_deactivate(struct adapter *);
87 static struct uld_info tom_uld_info = {
89 .activate = t4_tom_activate,
90 .deactivate = t4_tom_deactivate,
93 static void queue_tid_release(struct adapter *, int);
94 static void release_offload_resources(struct toepcb *);
95 static int alloc_tid_tabs(struct tid_info *);
96 static void free_tid_tabs(struct tid_info *);
97 static int add_lip(struct adapter *, struct in6_addr *);
98 static int delete_lip(struct adapter *, struct in6_addr *);
99 static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *);
100 static void init_clip_table(struct adapter *, struct tom_data *);
101 static void update_clip(struct adapter *, void *);
102 static void t4_clip_task(void *, int);
103 static void update_clip_table(struct adapter *, struct tom_data *);
104 static void destroy_clip_table(struct adapter *, struct tom_data *);
105 static void free_tom_data(struct adapter *, struct tom_data *);
106 static void reclaim_wr_resources(void *, int);
108 static int in6_ifaddr_gen;
109 static eventhandler_tag ifaddr_evhandler;
110 static struct timeout_task clip_task;
113 alloc_toepcb(struct vi_info *vi, int txqid, int rxqid, int flags)
115 struct port_info *pi = vi->pi;
116 struct adapter *sc = pi->adapter;
118 int tx_credits, txsd_total, len;
121 * The firmware counts tx work request credits in units of 16 bytes
122 * each. Reserve room for an ABORT_REQ so the driver never has to worry
123 * about tx credits if it wants to abort a connection.
125 tx_credits = sc->params.ofldq_wr_cred;
126 tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
129 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
130 * immediate payload, and firmware counts tx work request credits in
131 * units of 16 byte. Calculate the maximum work requests possible.
133 txsd_total = tx_credits /
134 howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16);
137 txqid = (arc4random() % vi->nofldtxq) + vi->first_ofld_txq;
138 KASSERT(txqid >= vi->first_ofld_txq &&
139 txqid < vi->first_ofld_txq + vi->nofldtxq,
140 ("%s: txqid %d for vi %p (first %d, n %d)", __func__, txqid, vi,
141 vi->first_ofld_txq, vi->nofldtxq));
144 rxqid = (arc4random() % vi->nofldrxq) + vi->first_ofld_rxq;
145 KASSERT(rxqid >= vi->first_ofld_rxq &&
146 rxqid < vi->first_ofld_rxq + vi->nofldrxq,
147 ("%s: rxqid %d for vi %p (first %d, n %d)", __func__, rxqid, vi,
148 vi->first_ofld_rxq, vi->nofldrxq));
150 len = offsetof(struct toepcb, txsd) +
151 txsd_total * sizeof(struct ofld_tx_sdesc);
153 toep = malloc(len, M_CXGBE, M_ZERO | flags);
157 refcount_init(&toep->refcount, 1);
158 toep->td = sc->tom_softc;
161 toep->tx_total = tx_credits;
162 toep->tx_credits = tx_credits;
163 toep->ofld_txq = &sc->sge.ofld_txq[txqid];
164 toep->ofld_rxq = &sc->sge.ofld_rxq[rxqid];
165 toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
166 mbufq_init(&toep->ulp_pduq, INT_MAX);
167 mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX);
168 toep->txsd_total = txsd_total;
169 toep->txsd_avail = txsd_total;
172 aiotx_init_toep(toep);
179 hold_toepcb(struct toepcb *toep)
182 refcount_acquire(&toep->refcount);
187 free_toepcb(struct toepcb *toep)
190 if (refcount_release(&toep->refcount) == 0)
193 KASSERT(!(toep->flags & TPF_ATTACHED),
194 ("%s: attached to an inpcb", __func__));
195 KASSERT(!(toep->flags & TPF_CPL_PENDING),
196 ("%s: CPL pending", __func__));
198 ddp_uninit_toep(toep);
203 * Set up the socket for TCP offload.
206 offload_socket(struct socket *so, struct toepcb *toep)
208 struct tom_data *td = toep->td;
209 struct inpcb *inp = sotoinpcb(so);
210 struct tcpcb *tp = intotcpcb(inp);
213 INP_WLOCK_ASSERT(inp);
218 sb->sb_flags |= SB_NOCOALESCE;
222 sb->sb_flags |= SB_NOCOALESCE;
223 if (inp->inp_vflag & INP_IPV6)
224 so->so_proto = &toe6_protosw;
226 so->so_proto = &toe_protosw;
232 tp->t_flags |= TF_TOE;
234 /* Install an extra hold on inp */
236 toep->flags |= TPF_ATTACHED;
239 /* Add the TOE PCB to the active list */
240 mtx_lock(&td->toep_list_lock);
241 TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
242 mtx_unlock(&td->toep_list_lock);
245 /* This is _not_ the normal way to "unoffload" a socket. */
247 undo_offload_socket(struct socket *so)
249 struct inpcb *inp = sotoinpcb(so);
250 struct tcpcb *tp = intotcpcb(inp);
251 struct toepcb *toep = tp->t_toe;
252 struct tom_data *td = toep->td;
255 INP_WLOCK_ASSERT(inp);
259 sb->sb_flags &= ~SB_NOCOALESCE;
263 sb->sb_flags &= ~SB_NOCOALESCE;
268 tp->t_flags &= ~TF_TOE;
271 toep->flags &= ~TPF_ATTACHED;
272 if (in_pcbrele_wlocked(inp))
273 panic("%s: inp freed.", __func__);
275 mtx_lock(&td->toep_list_lock);
276 TAILQ_REMOVE(&td->toep_list, toep, link);
277 mtx_unlock(&td->toep_list_lock);
281 release_offload_resources(struct toepcb *toep)
283 struct tom_data *td = toep->td;
284 struct adapter *sc = td_adapter(td);
287 KASSERT(!(toep->flags & TPF_CPL_PENDING),
288 ("%s: %p has CPL pending.", __func__, toep));
289 KASSERT(!(toep->flags & TPF_ATTACHED),
290 ("%s: %p is still attached.", __func__, toep));
292 CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
293 __func__, toep, tid, toep->l2te, toep->ce);
296 * These queues should have been emptied at approximately the same time
297 * that a normal connection's socket's so_snd would have been purged or
298 * drained. Do _not_ clean up here.
300 MPASS(mbufq_len(&toep->ulp_pduq) == 0);
301 MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0);
303 ddp_assert_empty(toep);
307 t4_l2t_release(toep->l2te);
310 remove_tid(sc, tid, toep->ce ? 2 : 1);
311 release_tid(sc, tid, toep->ctrlq);
315 release_lip(td, toep->ce);
318 if (toep->tc_idx != -1)
319 t4_release_cl_rl_kbps(sc, toep->vi->pi->port_id, toep->tc_idx);
321 mtx_lock(&td->toep_list_lock);
322 TAILQ_REMOVE(&td->toep_list, toep, link);
323 mtx_unlock(&td->toep_list_lock);
329 * The kernel is done with the TCP PCB and this is our opportunity to unhook the
330 * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no
331 * pending CPL) then it is time to release all resources tied to the toepcb.
333 * Also gets called when an offloaded active open fails and the TOM wants the
334 * kernel to take the TCP PCB back.
337 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
339 #if defined(KTR) || defined(INVARIANTS)
340 struct inpcb *inp = tp->t_inpcb;
342 struct toepcb *toep = tp->t_toe;
344 INP_WLOCK_ASSERT(inp);
346 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
347 KASSERT(toep->flags & TPF_ATTACHED,
348 ("%s: not attached", __func__));
351 if (tp->t_state == TCPS_SYN_SENT) {
352 CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
353 __func__, toep->tid, toep, toep->flags, inp,
357 "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
358 toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
364 tp->t_flags &= ~TF_TOE;
365 toep->flags &= ~TPF_ATTACHED;
367 if (!(toep->flags & TPF_CPL_PENDING))
368 release_offload_resources(toep);
372 * setsockopt handler.
375 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
377 struct adapter *sc = tod->tod_softc;
378 struct toepcb *toep = tp->t_toe;
383 CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
387 if (tp->t_state != TCPS_ESTABLISHED)
389 t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_T_FLAGS,
390 V_TF_NAGLE(1), V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1),
391 0, 0, toep->ofld_rxq->iq.abs_id);
399 * The TOE driver will not receive any more CPLs for the tid associated with the
400 * toepcb; release the hold on the inpcb.
403 final_cpl_received(struct toepcb *toep)
405 struct inpcb *inp = toep->inp;
407 KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
408 INP_WLOCK_ASSERT(inp);
409 KASSERT(toep->flags & TPF_CPL_PENDING,
410 ("%s: CPL not pending already?", __func__));
412 CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
413 __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
415 if (toep->ulp_mode == ULP_MODE_TCPDDP)
416 release_ddp_resources(toep);
418 toep->flags &= ~TPF_CPL_PENDING;
419 mbufq_drain(&toep->ulp_pdu_reclaimq);
421 if (!(toep->flags & TPF_ATTACHED))
422 release_offload_resources(toep);
424 if (!in_pcbrele_wlocked(inp))
429 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
431 struct tid_info *t = &sc->tids;
433 t->tid_tab[tid] = ctx;
434 atomic_add_int(&t->tids_in_use, ntids);
438 lookup_tid(struct adapter *sc, int tid)
440 struct tid_info *t = &sc->tids;
442 return (t->tid_tab[tid]);
446 update_tid(struct adapter *sc, int tid, void *ctx)
448 struct tid_info *t = &sc->tids;
450 t->tid_tab[tid] = ctx;
454 remove_tid(struct adapter *sc, int tid, int ntids)
456 struct tid_info *t = &sc->tids;
458 t->tid_tab[tid] = NULL;
459 atomic_subtract_int(&t->tids_in_use, ntids);
463 release_tid(struct adapter *sc, int tid, struct sge_wrq *ctrlq)
466 struct cpl_tid_release *req;
468 wr = alloc_wrqe(sizeof(*req), ctrlq);
470 queue_tid_release(sc, tid); /* defer */
475 INIT_TP_WR_MIT_CPL(req, CPL_TID_RELEASE, tid);
481 queue_tid_release(struct adapter *sc, int tid)
484 CXGBE_UNIMPLEMENTED("deferred tid release");
488 * What mtu_idx to use, given a 4-tuple and/or an MSS cap
491 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, int pmss)
493 unsigned short *mtus = &sc->params.mtus[0];
496 KASSERT(inc != NULL || pmss > 0,
497 ("%s: at least one of inc/pmss must be specified", __func__));
499 mss = inc ? tcp_mssopt(inc) : pmss;
500 if (pmss > 0 && mss > pmss)
503 if (inc->inc_flags & INC_ISIPV6)
504 n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
506 n = sizeof(struct ip) + sizeof(struct tcphdr);
508 for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mss + n; i++)
515 * Determine the receive window size for a socket.
518 select_rcv_wnd(struct socket *so)
522 SOCKBUF_LOCK_ASSERT(&so->so_rcv);
524 wnd = sbspace(&so->so_rcv);
525 if (wnd < MIN_RCV_WND)
528 return min(wnd, MAX_RCV_WND);
532 select_rcv_wscale(void)
535 unsigned long space = sb_max;
537 if (space > MAX_RCV_WND)
540 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
546 extern int always_keepalive;
549 * socket so could be a listening socket too.
552 calc_opt0(struct socket *so, struct vi_info *vi, struct l2t_entry *e,
553 int mtu_idx, int rscale, int rx_credits, int ulp_mode)
557 KASSERT(rx_credits <= M_RCV_BUFSIZ,
558 ("%s: rcv_bufsiz too high", __func__));
560 opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) |
561 V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits);
564 struct inpcb *inp = sotoinpcb(so);
565 struct tcpcb *tp = intotcpcb(inp);
566 int keepalive = always_keepalive ||
567 so_options_get(so) & SO_KEEPALIVE;
569 opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0);
570 opt0 |= V_KEEP_ALIVE(keepalive != 0);
574 opt0 |= V_L2T_IDX(e->idx);
577 opt0 |= V_SMAC_SEL(vi->smt_idx);
578 opt0 |= V_TX_CHAN(vi->pi->tx_chan);
581 return htobe64(opt0);
585 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
587 struct adapter *sc = vi->pi->adapter;
588 struct tp_params *tp = &sc->params.tp;
589 uint16_t viid = vi->viid;
593 * Initialize each of the fields which we care about which are present
594 * in the Compressed Filter Tuple.
596 if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE)
597 ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
599 if (tp->port_shift >= 0)
600 ntuple |= (uint64_t)e->lport << tp->port_shift;
602 if (tp->protocol_shift >= 0)
603 ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
605 if (tp->vnic_shift >= 0) {
606 uint32_t vf = G_FW_VIID_VIN(viid);
607 uint32_t pf = G_FW_VIID_PFN(viid);
608 uint32_t vld = G_FW_VIID_VIVLD(viid);
610 ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) |
611 V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift;
615 return (htobe32((uint32_t)ntuple));
617 return (htobe64(V_FILTER_TUPLE(ntuple)));
621 set_tcpddp_ulp_mode(struct toepcb *toep)
624 toep->ulp_mode = ULP_MODE_TCPDDP;
625 toep->ddp_flags = DDP_OK;
629 negative_advice(int status)
632 return (status == CPL_ERR_RTX_NEG_ADVICE ||
633 status == CPL_ERR_PERSIST_NEG_ADVICE ||
634 status == CPL_ERR_KEEPALV_NEG_ADVICE);
638 alloc_tid_tabs(struct tid_info *t)
643 size = t->ntids * sizeof(*t->tid_tab) +
644 t->natids * sizeof(*t->atid_tab) +
645 t->nstids * sizeof(*t->stid_tab);
647 t->tid_tab = malloc(size, M_CXGBE, M_ZERO | M_NOWAIT);
648 if (t->tid_tab == NULL)
651 mtx_init(&t->atid_lock, "atid lock", NULL, MTX_DEF);
652 t->atid_tab = (union aopen_entry *)&t->tid_tab[t->ntids];
653 t->afree = t->atid_tab;
655 for (i = 1; i < t->natids; i++)
656 t->atid_tab[i - 1].next = &t->atid_tab[i];
657 t->atid_tab[t->natids - 1].next = NULL;
659 mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
660 t->stid_tab = (struct listen_ctx **)&t->atid_tab[t->natids];
662 TAILQ_INIT(&t->stids);
663 t->nstids_free_head = t->nstids;
665 atomic_store_rel_int(&t->tids_in_use, 0);
671 free_tid_tabs(struct tid_info *t)
673 KASSERT(t->tids_in_use == 0,
674 ("%s: %d tids still in use.", __func__, t->tids_in_use));
675 KASSERT(t->atids_in_use == 0,
676 ("%s: %d atids still in use.", __func__, t->atids_in_use));
677 KASSERT(t->stids_in_use == 0,
678 ("%s: %d tids still in use.", __func__, t->stids_in_use));
680 free(t->tid_tab, M_CXGBE);
683 if (mtx_initialized(&t->atid_lock))
684 mtx_destroy(&t->atid_lock);
685 if (mtx_initialized(&t->stid_lock))
686 mtx_destroy(&t->stid_lock);
690 add_lip(struct adapter *sc, struct in6_addr *lip)
692 struct fw_clip_cmd c;
694 ASSERT_SYNCHRONIZED_OP(sc);
695 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */
697 memset(&c, 0, sizeof(c));
698 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
700 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c));
701 c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
702 c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
704 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
708 delete_lip(struct adapter *sc, struct in6_addr *lip)
710 struct fw_clip_cmd c;
712 ASSERT_SYNCHRONIZED_OP(sc);
713 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */
715 memset(&c, 0, sizeof(c));
716 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
718 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c));
719 c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
720 c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
722 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
725 static struct clip_entry *
726 search_lip(struct tom_data *td, struct in6_addr *lip)
728 struct clip_entry *ce;
730 mtx_assert(&td->clip_table_lock, MA_OWNED);
732 TAILQ_FOREACH(ce, &td->clip_table, link) {
733 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
741 hold_lip(struct tom_data *td, struct in6_addr *lip, struct clip_entry *ce)
744 mtx_lock(&td->clip_table_lock);
746 ce = search_lip(td, lip);
749 mtx_unlock(&td->clip_table_lock);
755 release_lip(struct tom_data *td, struct clip_entry *ce)
758 mtx_lock(&td->clip_table_lock);
759 KASSERT(search_lip(td, &ce->lip) == ce,
760 ("%s: CLIP entry %p p not in CLIP table.", __func__, ce));
761 KASSERT(ce->refcount > 0,
762 ("%s: CLIP entry %p has refcount 0", __func__, ce));
764 mtx_unlock(&td->clip_table_lock);
768 init_clip_table(struct adapter *sc, struct tom_data *td)
771 ASSERT_SYNCHRONIZED_OP(sc);
773 mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF);
774 TAILQ_INIT(&td->clip_table);
777 update_clip_table(sc, td);
781 update_clip(struct adapter *sc, void *arg __unused)
784 if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc"))
787 if (uld_active(sc, ULD_TOM))
788 update_clip_table(sc, sc->tom_softc);
790 end_synchronized_op(sc, LOCK_HELD);
794 t4_clip_task(void *arg, int count)
797 t4_iterate(update_clip, NULL);
801 update_clip_table(struct adapter *sc, struct tom_data *td)
803 struct rm_priotracker in6_ifa_tracker;
804 struct in6_ifaddr *ia;
805 struct in6_addr *lip, tlip;
806 struct clip_head stale;
807 struct clip_entry *ce, *ce_temp;
812 ASSERT_SYNCHRONIZED_OP(sc);
814 IN6_IFADDR_RLOCK(&in6_ifa_tracker);
815 mtx_lock(&td->clip_table_lock);
817 gen = atomic_load_acq_int(&in6_ifaddr_gen);
818 if (gen == td->clip_gen)
822 TAILQ_CONCAT(&stale, &td->clip_table, link);
825 * last_vnet optimizes the common cases where all if_vnet = NULL (no
826 * VIMAGE) or all if_vnet = vnet0.
828 last_vnet = (uintptr_t)(-1);
830 for_each_vi(sc->port[i], j, vi) {
831 if (last_vnet == (uintptr_t)vi->ifp->if_vnet)
834 /* XXX: races with if_vmove */
835 CURVNET_SET(vi->ifp->if_vnet);
836 TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) {
837 lip = &ia->ia_addr.sin6_addr;
839 KASSERT(!IN6_IS_ADDR_MULTICAST(lip),
840 ("%s: mcast address in in6_ifaddr list", __func__));
842 if (IN6_IS_ADDR_LOOPBACK(lip))
844 if (IN6_IS_SCOPE_EMBED(lip)) {
845 /* Remove the embedded scope */
851 * XXX: how to weed out the link local address for the
852 * loopback interface? It's fe80::1 usually (always?).
856 * If it's in the main list then we already know it's
859 TAILQ_FOREACH(ce, &td->clip_table, link) {
860 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
865 * If it's in the stale list we should move it to the
868 TAILQ_FOREACH(ce, &stale, link) {
869 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) {
870 TAILQ_REMOVE(&stale, ce, link);
871 TAILQ_INSERT_TAIL(&td->clip_table, ce,
877 /* A new IP6 address; add it to the CLIP table */
878 ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT);
879 memcpy(&ce->lip, lip, sizeof(ce->lip));
881 rc = add_lip(sc, lip);
883 TAILQ_INSERT_TAIL(&td->clip_table, ce, link);
885 char ip[INET6_ADDRSTRLEN];
887 inet_ntop(AF_INET6, &ce->lip, &ip[0],
889 log(LOG_ERR, "%s: could not add %s (%d)\n",
897 last_vnet = (uintptr_t)vi->ifp->if_vnet;
901 * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are
902 * no longer referenced by the driver.
904 TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) {
905 if (ce->refcount == 0) {
906 rc = delete_lip(sc, &ce->lip);
908 TAILQ_REMOVE(&stale, ce, link);
911 char ip[INET6_ADDRSTRLEN];
913 inet_ntop(AF_INET6, &ce->lip, &ip[0],
915 log(LOG_ERR, "%s: could not delete %s (%d)\n",
920 /* The ones that are still referenced need to stay in the CLIP table */
921 TAILQ_CONCAT(&td->clip_table, &stale, link);
925 mtx_unlock(&td->clip_table_lock);
926 IN6_IFADDR_RUNLOCK(&in6_ifa_tracker);
930 destroy_clip_table(struct adapter *sc, struct tom_data *td)
932 struct clip_entry *ce, *ce_temp;
934 if (mtx_initialized(&td->clip_table_lock)) {
935 mtx_lock(&td->clip_table_lock);
936 TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) {
937 KASSERT(ce->refcount == 0,
938 ("%s: CLIP entry %p still in use (%d)", __func__,
940 TAILQ_REMOVE(&td->clip_table, ce, link);
941 delete_lip(sc, &ce->lip);
944 mtx_unlock(&td->clip_table_lock);
945 mtx_destroy(&td->clip_table_lock);
950 free_tom_data(struct adapter *sc, struct tom_data *td)
953 ASSERT_SYNCHRONIZED_OP(sc);
955 KASSERT(TAILQ_EMPTY(&td->toep_list),
956 ("%s: TOE PCB list is not empty.", __func__));
957 KASSERT(td->lctx_count == 0,
958 ("%s: lctx hash table is not empty.", __func__));
960 t4_free_ppod_region(&td->pr);
961 destroy_clip_table(sc, td);
963 if (td->listen_mask != 0)
964 hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
966 if (mtx_initialized(&td->unsent_wr_lock))
967 mtx_destroy(&td->unsent_wr_lock);
968 if (mtx_initialized(&td->lctx_hash_lock))
969 mtx_destroy(&td->lctx_hash_lock);
970 if (mtx_initialized(&td->toep_list_lock))
971 mtx_destroy(&td->toep_list_lock);
973 free_tid_tabs(&sc->tids);
978 reclaim_wr_resources(void *arg, int count)
980 struct tom_data *td = arg;
981 STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
982 struct cpl_act_open_req *cpl;
987 mtx_lock(&td->unsent_wr_lock);
988 STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
989 mtx_unlock(&td->unsent_wr_lock);
991 while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
992 STAILQ_REMOVE_HEAD(&twr_list, link);
995 opcode = GET_OPCODE(cpl);
998 case CPL_ACT_OPEN_REQ:
999 case CPL_ACT_OPEN_REQ6:
1000 atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
1001 sc = td_adapter(td);
1003 CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1004 act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
1008 log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1009 "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1010 /* WR not freed here; go look at it with a debugger. */
1016 * Ground control to Major TOM
1017 * Commencing countdown, engines on
1020 t4_tom_activate(struct adapter *sc)
1022 struct tom_data *td;
1025 struct sge_ofld_rxq *ofld_rxq;
1028 ASSERT_SYNCHRONIZED_OP(sc);
1030 /* per-adapter softc for TOM */
1031 td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1035 /* List of TOE PCBs and associated lock */
1036 mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1037 TAILQ_INIT(&td->toep_list);
1039 /* Listen context */
1040 mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1041 td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1042 &td->listen_mask, HASH_NOWAIT);
1044 /* List of WRs for which L2 resolution failed */
1045 mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1046 STAILQ_INIT(&td->unsent_wr_list);
1047 TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1050 rc = alloc_tid_tabs(&sc->tids);
1054 rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1055 t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1058 t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1059 V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1061 /* CLIP table for IPv6 offload */
1062 init_clip_table(sc, td);
1067 tod->tod_softc = sc;
1068 tod->tod_connect = t4_connect;
1069 tod->tod_listen_start = t4_listen_start;
1070 tod->tod_listen_stop = t4_listen_stop;
1071 tod->tod_rcvd = t4_rcvd;
1072 tod->tod_output = t4_tod_output;
1073 tod->tod_send_rst = t4_send_rst;
1074 tod->tod_send_fin = t4_send_fin;
1075 tod->tod_pcb_detach = t4_pcb_detach;
1076 tod->tod_l2_update = t4_l2_update;
1077 tod->tod_syncache_added = t4_syncache_added;
1078 tod->tod_syncache_removed = t4_syncache_removed;
1079 tod->tod_syncache_respond = t4_syncache_respond;
1080 tod->tod_offload_socket = t4_offload_socket;
1081 tod->tod_ctloutput = t4_ctloutput;
1083 for_each_port(sc, i) {
1084 for_each_vi(sc->port[i], v, vi) {
1085 TOEDEV(vi->ifp) = &td->tod;
1086 for_each_ofld_rxq(vi, j, ofld_rxq) {
1087 ofld_rxq->iq.set_tcb_rpl = do_set_tcb_rpl;
1088 ofld_rxq->iq.l2t_write_rpl = do_l2t_write_rpl2;
1094 register_toedev(sc->tom_softc);
1098 free_tom_data(sc, td);
1103 t4_tom_deactivate(struct adapter *sc)
1106 struct tom_data *td = sc->tom_softc;
1108 ASSERT_SYNCHRONIZED_OP(sc);
1111 return (0); /* XXX. KASSERT? */
1113 if (sc->offload_map != 0)
1114 return (EBUSY); /* at least one port has IFCAP_TOE enabled */
1116 if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1117 return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */
1119 mtx_lock(&td->toep_list_lock);
1120 if (!TAILQ_EMPTY(&td->toep_list))
1122 mtx_unlock(&td->toep_list_lock);
1124 mtx_lock(&td->lctx_hash_lock);
1125 if (td->lctx_count > 0)
1127 mtx_unlock(&td->lctx_hash_lock);
1129 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1130 mtx_lock(&td->unsent_wr_lock);
1131 if (!STAILQ_EMPTY(&td->unsent_wr_list))
1133 mtx_unlock(&td->unsent_wr_lock);
1136 unregister_toedev(sc->tom_softc);
1137 free_tom_data(sc, td);
1138 sc->tom_softc = NULL;
1145 t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp)
1148 atomic_add_rel_int(&in6_ifaddr_gen, 1);
1149 taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4);
1153 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
1155 struct tcpcb *tp = so_sototcpcb(so);
1156 struct toepcb *toep = tp->t_toe;
1159 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1160 error = t4_aio_queue_ddp(so, job);
1161 if (error != EOPNOTSUPP)
1165 return (t4_aio_queue_aiotx(so, job));
1169 t4_tom_mod_load(void)
1172 struct protosw *tcp_protosw, *tcp6_protosw;
1175 t4_init_connect_cpl_handlers();
1176 t4_init_listen_cpl_handlers();
1177 t4_init_cpl_io_handlers();
1179 rc = t4_ddp_mod_load();
1183 tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM);
1184 if (tcp_protosw == NULL)
1185 return (ENOPROTOOPT);
1186 bcopy(tcp_protosw, &toe_protosw, sizeof(toe_protosw));
1187 bcopy(tcp_protosw->pr_usrreqs, &toe_usrreqs, sizeof(toe_usrreqs));
1188 toe_usrreqs.pru_aio_queue = t4_aio_queue_tom;
1189 toe_protosw.pr_usrreqs = &toe_usrreqs;
1191 tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM);
1192 if (tcp6_protosw == NULL)
1193 return (ENOPROTOOPT);
1194 bcopy(tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
1195 bcopy(tcp6_protosw->pr_usrreqs, &toe6_usrreqs, sizeof(toe6_usrreqs));
1196 toe6_usrreqs.pru_aio_queue = t4_aio_queue_tom;
1197 toe6_protosw.pr_usrreqs = &toe6_usrreqs;
1199 TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL);
1200 ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event,
1201 t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY);
1203 rc = t4_register_uld(&tom_uld_info);
1205 t4_tom_mod_unload();
1211 tom_uninit(struct adapter *sc, void *arg __unused)
1213 if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
1216 /* Try to free resources (works only if no port has IFCAP_TOE) */
1217 if (uld_active(sc, ULD_TOM))
1218 t4_deactivate_uld(sc, ULD_TOM);
1220 end_synchronized_op(sc, 0);
1224 t4_tom_mod_unload(void)
1226 t4_iterate(tom_uninit, NULL);
1228 if (t4_unregister_uld(&tom_uld_info) == EBUSY)
1231 if (ifaddr_evhandler) {
1232 EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler);
1233 taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL);
1236 t4_ddp_mod_unload();
1238 t4_uninit_connect_cpl_handlers();
1239 t4_uninit_listen_cpl_handlers();
1240 t4_uninit_cpl_io_handlers();
1244 #endif /* TCP_OFFLOAD */
1247 t4_tom_modevent(module_t mod, int cmd, void *arg)
1254 rc = t4_tom_mod_load();
1258 rc = t4_tom_mod_unload();
1265 printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
1271 static moduledata_t t4_tom_moddata= {
1277 MODULE_VERSION(t4_tom, 1);
1278 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
1279 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
1280 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);