2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2012 Chelsio Communications, Inc.
6 * Written by: Navdeep Parhar <np@FreeBSD.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
34 #include "opt_inet6.h"
35 #include "opt_ratelimit.h"
37 #include <sys/param.h>
38 #include <sys/types.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
43 #include <sys/limits.h>
44 #include <sys/module.h>
45 #include <sys/protosw.h>
46 #include <sys/domain.h>
47 #include <sys/refcount.h>
48 #include <sys/rmlock.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/taskqueue.h>
53 #include <net/if_var.h>
54 #include <netinet/in.h>
55 #include <netinet/in_pcb.h>
56 #include <netinet/in_var.h>
57 #include <netinet/ip.h>
58 #include <netinet/ip6.h>
59 #include <netinet6/scope6_var.h>
61 #include <netinet/tcp_fsm.h>
62 #include <netinet/tcp_var.h>
63 #include <netinet/toecore.h>
66 #include "common/common.h"
67 #include "common/t4_msg.h"
68 #include "common/t4_regs.h"
69 #include "common/t4_regs_values.h"
70 #include "common/t4_tcb.h"
71 #include "tom/t4_tom_l2t.h"
72 #include "tom/t4_tom.h"
74 static struct protosw toe_protosw;
75 static struct pr_usrreqs toe_usrreqs;
77 static struct protosw toe6_protosw;
78 static struct pr_usrreqs toe6_usrreqs;
81 static int t4_tom_mod_load(void);
82 static int t4_tom_mod_unload(void);
83 static int t4_tom_modevent(module_t, int, void *);
85 /* ULD ops and helpers */
86 static int t4_tom_activate(struct adapter *);
87 static int t4_tom_deactivate(struct adapter *);
89 static struct uld_info tom_uld_info = {
91 .activate = t4_tom_activate,
92 .deactivate = t4_tom_deactivate,
95 static void queue_tid_release(struct adapter *, int);
96 static void release_offload_resources(struct toepcb *);
97 static int alloc_tid_tabs(struct tid_info *);
98 static void free_tid_tabs(struct tid_info *);
99 static int add_lip(struct adapter *, struct in6_addr *);
100 static int delete_lip(struct adapter *, struct in6_addr *);
101 static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *);
102 static void init_clip_table(struct adapter *, struct tom_data *);
103 static void update_clip(struct adapter *, void *);
104 static void t4_clip_task(void *, int);
105 static void update_clip_table(struct adapter *, struct tom_data *);
106 static void destroy_clip_table(struct adapter *, struct tom_data *);
107 static void free_tom_data(struct adapter *, struct tom_data *);
108 static void reclaim_wr_resources(void *, int);
110 static int in6_ifaddr_gen;
111 static eventhandler_tag ifaddr_evhandler;
112 static struct timeout_task clip_task;
115 alloc_toepcb(struct vi_info *vi, int txqid, int rxqid, int flags)
117 struct port_info *pi = vi->pi;
118 struct adapter *sc = pi->adapter;
120 int tx_credits, txsd_total, len;
123 * The firmware counts tx work request credits in units of 16 bytes
124 * each. Reserve room for an ABORT_REQ so the driver never has to worry
125 * about tx credits if it wants to abort a connection.
127 tx_credits = sc->params.ofldq_wr_cred;
128 tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
131 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
132 * immediate payload, and firmware counts tx work request credits in
133 * units of 16 byte. Calculate the maximum work requests possible.
135 txsd_total = tx_credits /
136 howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16);
139 txqid = (arc4random() % vi->nofldtxq) + vi->first_ofld_txq;
140 KASSERT(txqid >= vi->first_ofld_txq &&
141 txqid < vi->first_ofld_txq + vi->nofldtxq,
142 ("%s: txqid %d for vi %p (first %d, n %d)", __func__, txqid, vi,
143 vi->first_ofld_txq, vi->nofldtxq));
146 rxqid = (arc4random() % vi->nofldrxq) + vi->first_ofld_rxq;
147 KASSERT(rxqid >= vi->first_ofld_rxq &&
148 rxqid < vi->first_ofld_rxq + vi->nofldrxq,
149 ("%s: rxqid %d for vi %p (first %d, n %d)", __func__, rxqid, vi,
150 vi->first_ofld_rxq, vi->nofldrxq));
152 len = offsetof(struct toepcb, txsd) +
153 txsd_total * sizeof(struct ofld_tx_sdesc);
155 toep = malloc(len, M_CXGBE, M_ZERO | flags);
159 refcount_init(&toep->refcount, 1);
160 toep->td = sc->tom_softc;
163 toep->tx_total = tx_credits;
164 toep->tx_credits = tx_credits;
165 toep->ofld_txq = &sc->sge.ofld_txq[txqid];
166 toep->ofld_rxq = &sc->sge.ofld_rxq[rxqid];
167 toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
168 mbufq_init(&toep->ulp_pduq, INT_MAX);
169 mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX);
170 toep->txsd_total = txsd_total;
171 toep->txsd_avail = txsd_total;
174 aiotx_init_toep(toep);
181 hold_toepcb(struct toepcb *toep)
184 refcount_acquire(&toep->refcount);
189 free_toepcb(struct toepcb *toep)
192 if (refcount_release(&toep->refcount) == 0)
195 KASSERT(!(toep->flags & TPF_ATTACHED),
196 ("%s: attached to an inpcb", __func__));
197 KASSERT(!(toep->flags & TPF_CPL_PENDING),
198 ("%s: CPL pending", __func__));
200 ddp_uninit_toep(toep);
205 * Set up the socket for TCP offload.
208 offload_socket(struct socket *so, struct toepcb *toep)
210 struct tom_data *td = toep->td;
211 struct inpcb *inp = sotoinpcb(so);
212 struct tcpcb *tp = intotcpcb(inp);
215 INP_WLOCK_ASSERT(inp);
220 sb->sb_flags |= SB_NOCOALESCE;
224 sb->sb_flags |= SB_NOCOALESCE;
225 if (inp->inp_vflag & INP_IPV6)
226 so->so_proto = &toe6_protosw;
228 so->so_proto = &toe_protosw;
234 tp->t_flags |= TF_TOE;
236 /* Install an extra hold on inp */
238 toep->flags |= TPF_ATTACHED;
241 /* Add the TOE PCB to the active list */
242 mtx_lock(&td->toep_list_lock);
243 TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
244 mtx_unlock(&td->toep_list_lock);
247 /* This is _not_ the normal way to "unoffload" a socket. */
249 undo_offload_socket(struct socket *so)
251 struct inpcb *inp = sotoinpcb(so);
252 struct tcpcb *tp = intotcpcb(inp);
253 struct toepcb *toep = tp->t_toe;
254 struct tom_data *td = toep->td;
257 INP_WLOCK_ASSERT(inp);
261 sb->sb_flags &= ~SB_NOCOALESCE;
265 sb->sb_flags &= ~SB_NOCOALESCE;
270 tp->t_flags &= ~TF_TOE;
273 toep->flags &= ~TPF_ATTACHED;
274 if (in_pcbrele_wlocked(inp))
275 panic("%s: inp freed.", __func__);
277 mtx_lock(&td->toep_list_lock);
278 TAILQ_REMOVE(&td->toep_list, toep, link);
279 mtx_unlock(&td->toep_list_lock);
283 release_offload_resources(struct toepcb *toep)
285 struct tom_data *td = toep->td;
286 struct adapter *sc = td_adapter(td);
289 KASSERT(!(toep->flags & TPF_CPL_PENDING),
290 ("%s: %p has CPL pending.", __func__, toep));
291 KASSERT(!(toep->flags & TPF_ATTACHED),
292 ("%s: %p is still attached.", __func__, toep));
294 CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
295 __func__, toep, tid, toep->l2te, toep->ce);
298 * These queues should have been emptied at approximately the same time
299 * that a normal connection's socket's so_snd would have been purged or
300 * drained. Do _not_ clean up here.
302 MPASS(mbufq_len(&toep->ulp_pduq) == 0);
303 MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0);
305 ddp_assert_empty(toep);
309 t4_l2t_release(toep->l2te);
312 remove_tid(sc, tid, toep->ce ? 2 : 1);
313 release_tid(sc, tid, toep->ctrlq);
317 release_lip(td, toep->ce);
320 if (toep->tc_idx != -1)
321 t4_release_cl_rl_kbps(sc, toep->vi->pi->port_id, toep->tc_idx);
323 mtx_lock(&td->toep_list_lock);
324 TAILQ_REMOVE(&td->toep_list, toep, link);
325 mtx_unlock(&td->toep_list_lock);
331 * The kernel is done with the TCP PCB and this is our opportunity to unhook the
332 * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no
333 * pending CPL) then it is time to release all resources tied to the toepcb.
335 * Also gets called when an offloaded active open fails and the TOM wants the
336 * kernel to take the TCP PCB back.
339 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
341 #if defined(KTR) || defined(INVARIANTS)
342 struct inpcb *inp = tp->t_inpcb;
344 struct toepcb *toep = tp->t_toe;
346 INP_WLOCK_ASSERT(inp);
348 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
349 KASSERT(toep->flags & TPF_ATTACHED,
350 ("%s: not attached", __func__));
353 if (tp->t_state == TCPS_SYN_SENT) {
354 CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
355 __func__, toep->tid, toep, toep->flags, inp,
359 "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
360 toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
366 tp->t_flags &= ~TF_TOE;
367 toep->flags &= ~TPF_ATTACHED;
369 if (!(toep->flags & TPF_CPL_PENDING))
370 release_offload_resources(toep);
374 * setsockopt handler.
377 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
379 struct adapter *sc = tod->tod_softc;
380 struct toepcb *toep = tp->t_toe;
385 CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
389 if (tp->t_state != TCPS_ESTABLISHED)
391 t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_T_FLAGS,
392 V_TF_NAGLE(1), V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1),
393 0, 0, toep->ofld_rxq->iq.abs_id);
401 * The TOE driver will not receive any more CPLs for the tid associated with the
402 * toepcb; release the hold on the inpcb.
405 final_cpl_received(struct toepcb *toep)
407 struct inpcb *inp = toep->inp;
409 KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
410 INP_WLOCK_ASSERT(inp);
411 KASSERT(toep->flags & TPF_CPL_PENDING,
412 ("%s: CPL not pending already?", __func__));
414 CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
415 __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
417 if (toep->ulp_mode == ULP_MODE_TCPDDP)
418 release_ddp_resources(toep);
420 toep->flags &= ~TPF_CPL_PENDING;
421 mbufq_drain(&toep->ulp_pdu_reclaimq);
423 if (!(toep->flags & TPF_ATTACHED))
424 release_offload_resources(toep);
426 if (!in_pcbrele_wlocked(inp))
431 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
433 struct tid_info *t = &sc->tids;
435 t->tid_tab[tid] = ctx;
436 atomic_add_int(&t->tids_in_use, ntids);
440 lookup_tid(struct adapter *sc, int tid)
442 struct tid_info *t = &sc->tids;
444 return (t->tid_tab[tid]);
448 update_tid(struct adapter *sc, int tid, void *ctx)
450 struct tid_info *t = &sc->tids;
452 t->tid_tab[tid] = ctx;
456 remove_tid(struct adapter *sc, int tid, int ntids)
458 struct tid_info *t = &sc->tids;
460 t->tid_tab[tid] = NULL;
461 atomic_subtract_int(&t->tids_in_use, ntids);
465 release_tid(struct adapter *sc, int tid, struct sge_wrq *ctrlq)
468 struct cpl_tid_release *req;
470 wr = alloc_wrqe(sizeof(*req), ctrlq);
472 queue_tid_release(sc, tid); /* defer */
477 INIT_TP_WR_MIT_CPL(req, CPL_TID_RELEASE, tid);
483 queue_tid_release(struct adapter *sc, int tid)
486 CXGBE_UNIMPLEMENTED("deferred tid release");
490 * What mtu_idx to use, given a 4-tuple and/or an MSS cap
493 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, int pmss)
495 unsigned short *mtus = &sc->params.mtus[0];
498 KASSERT(inc != NULL || pmss > 0,
499 ("%s: at least one of inc/pmss must be specified", __func__));
501 mss = inc ? tcp_mssopt(inc) : pmss;
502 if (pmss > 0 && mss > pmss)
505 if (inc->inc_flags & INC_ISIPV6)
506 n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
508 n = sizeof(struct ip) + sizeof(struct tcphdr);
510 for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mss + n; i++)
517 * Determine the receive window size for a socket.
520 select_rcv_wnd(struct socket *so)
524 SOCKBUF_LOCK_ASSERT(&so->so_rcv);
526 wnd = sbspace(&so->so_rcv);
527 if (wnd < MIN_RCV_WND)
530 return min(wnd, MAX_RCV_WND);
534 select_rcv_wscale(void)
537 unsigned long space = sb_max;
539 if (space > MAX_RCV_WND)
542 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
548 extern int always_keepalive;
551 * socket so could be a listening socket too.
554 calc_opt0(struct socket *so, struct vi_info *vi, struct l2t_entry *e,
555 int mtu_idx, int rscale, int rx_credits, int ulp_mode)
559 KASSERT(rx_credits <= M_RCV_BUFSIZ,
560 ("%s: rcv_bufsiz too high", __func__));
562 opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) |
563 V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits);
566 struct inpcb *inp = sotoinpcb(so);
567 struct tcpcb *tp = intotcpcb(inp);
568 int keepalive = always_keepalive ||
569 so_options_get(so) & SO_KEEPALIVE;
571 opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0);
572 opt0 |= V_KEEP_ALIVE(keepalive != 0);
576 opt0 |= V_L2T_IDX(e->idx);
579 opt0 |= V_SMAC_SEL(vi->smt_idx);
580 opt0 |= V_TX_CHAN(vi->pi->tx_chan);
583 return htobe64(opt0);
587 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
589 struct adapter *sc = vi->pi->adapter;
590 struct tp_params *tp = &sc->params.tp;
591 uint16_t viid = vi->viid;
595 * Initialize each of the fields which we care about which are present
596 * in the Compressed Filter Tuple.
598 if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE)
599 ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
601 if (tp->port_shift >= 0)
602 ntuple |= (uint64_t)e->lport << tp->port_shift;
604 if (tp->protocol_shift >= 0)
605 ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
607 if (tp->vnic_shift >= 0) {
608 uint32_t vf = G_FW_VIID_VIN(viid);
609 uint32_t pf = G_FW_VIID_PFN(viid);
610 uint32_t vld = G_FW_VIID_VIVLD(viid);
612 ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) |
613 V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift;
617 return (htobe32((uint32_t)ntuple));
619 return (htobe64(V_FILTER_TUPLE(ntuple)));
623 set_tcpddp_ulp_mode(struct toepcb *toep)
626 toep->ulp_mode = ULP_MODE_TCPDDP;
627 toep->ddp_flags = DDP_OK;
631 negative_advice(int status)
634 return (status == CPL_ERR_RTX_NEG_ADVICE ||
635 status == CPL_ERR_PERSIST_NEG_ADVICE ||
636 status == CPL_ERR_KEEPALV_NEG_ADVICE);
640 alloc_tid_tabs(struct tid_info *t)
645 size = t->ntids * sizeof(*t->tid_tab) +
646 t->natids * sizeof(*t->atid_tab) +
647 t->nstids * sizeof(*t->stid_tab);
649 t->tid_tab = malloc(size, M_CXGBE, M_ZERO | M_NOWAIT);
650 if (t->tid_tab == NULL)
653 mtx_init(&t->atid_lock, "atid lock", NULL, MTX_DEF);
654 t->atid_tab = (union aopen_entry *)&t->tid_tab[t->ntids];
655 t->afree = t->atid_tab;
657 for (i = 1; i < t->natids; i++)
658 t->atid_tab[i - 1].next = &t->atid_tab[i];
659 t->atid_tab[t->natids - 1].next = NULL;
661 mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
662 t->stid_tab = (struct listen_ctx **)&t->atid_tab[t->natids];
664 TAILQ_INIT(&t->stids);
665 t->nstids_free_head = t->nstids;
667 atomic_store_rel_int(&t->tids_in_use, 0);
673 free_tid_tabs(struct tid_info *t)
675 KASSERT(t->tids_in_use == 0,
676 ("%s: %d tids still in use.", __func__, t->tids_in_use));
677 KASSERT(t->atids_in_use == 0,
678 ("%s: %d atids still in use.", __func__, t->atids_in_use));
679 KASSERT(t->stids_in_use == 0,
680 ("%s: %d tids still in use.", __func__, t->stids_in_use));
682 free(t->tid_tab, M_CXGBE);
685 if (mtx_initialized(&t->atid_lock))
686 mtx_destroy(&t->atid_lock);
687 if (mtx_initialized(&t->stid_lock))
688 mtx_destroy(&t->stid_lock);
692 add_lip(struct adapter *sc, struct in6_addr *lip)
694 struct fw_clip_cmd c;
696 ASSERT_SYNCHRONIZED_OP(sc);
697 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */
699 memset(&c, 0, sizeof(c));
700 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
702 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c));
703 c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
704 c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
706 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
710 delete_lip(struct adapter *sc, struct in6_addr *lip)
712 struct fw_clip_cmd c;
714 ASSERT_SYNCHRONIZED_OP(sc);
715 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */
717 memset(&c, 0, sizeof(c));
718 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
720 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c));
721 c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
722 c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
724 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
727 static struct clip_entry *
728 search_lip(struct tom_data *td, struct in6_addr *lip)
730 struct clip_entry *ce;
732 mtx_assert(&td->clip_table_lock, MA_OWNED);
734 TAILQ_FOREACH(ce, &td->clip_table, link) {
735 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
743 hold_lip(struct tom_data *td, struct in6_addr *lip, struct clip_entry *ce)
746 mtx_lock(&td->clip_table_lock);
748 ce = search_lip(td, lip);
751 mtx_unlock(&td->clip_table_lock);
757 release_lip(struct tom_data *td, struct clip_entry *ce)
760 mtx_lock(&td->clip_table_lock);
761 KASSERT(search_lip(td, &ce->lip) == ce,
762 ("%s: CLIP entry %p p not in CLIP table.", __func__, ce));
763 KASSERT(ce->refcount > 0,
764 ("%s: CLIP entry %p has refcount 0", __func__, ce));
766 mtx_unlock(&td->clip_table_lock);
770 init_clip_table(struct adapter *sc, struct tom_data *td)
773 ASSERT_SYNCHRONIZED_OP(sc);
775 mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF);
776 TAILQ_INIT(&td->clip_table);
779 update_clip_table(sc, td);
783 update_clip(struct adapter *sc, void *arg __unused)
786 if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc"))
789 if (uld_active(sc, ULD_TOM))
790 update_clip_table(sc, sc->tom_softc);
792 end_synchronized_op(sc, LOCK_HELD);
796 t4_clip_task(void *arg, int count)
799 t4_iterate(update_clip, NULL);
803 update_clip_table(struct adapter *sc, struct tom_data *td)
805 struct rm_priotracker in6_ifa_tracker;
806 struct in6_ifaddr *ia;
807 struct in6_addr *lip, tlip;
808 struct clip_head stale;
809 struct clip_entry *ce, *ce_temp;
814 ASSERT_SYNCHRONIZED_OP(sc);
816 IN6_IFADDR_RLOCK(&in6_ifa_tracker);
817 mtx_lock(&td->clip_table_lock);
819 gen = atomic_load_acq_int(&in6_ifaddr_gen);
820 if (gen == td->clip_gen)
824 TAILQ_CONCAT(&stale, &td->clip_table, link);
827 * last_vnet optimizes the common cases where all if_vnet = NULL (no
828 * VIMAGE) or all if_vnet = vnet0.
830 last_vnet = (uintptr_t)(-1);
832 for_each_vi(sc->port[i], j, vi) {
833 if (last_vnet == (uintptr_t)vi->ifp->if_vnet)
836 /* XXX: races with if_vmove */
837 CURVNET_SET(vi->ifp->if_vnet);
838 TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) {
839 lip = &ia->ia_addr.sin6_addr;
841 KASSERT(!IN6_IS_ADDR_MULTICAST(lip),
842 ("%s: mcast address in in6_ifaddr list", __func__));
844 if (IN6_IS_ADDR_LOOPBACK(lip))
846 if (IN6_IS_SCOPE_EMBED(lip)) {
847 /* Remove the embedded scope */
853 * XXX: how to weed out the link local address for the
854 * loopback interface? It's fe80::1 usually (always?).
858 * If it's in the main list then we already know it's
861 TAILQ_FOREACH(ce, &td->clip_table, link) {
862 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
867 * If it's in the stale list we should move it to the
870 TAILQ_FOREACH(ce, &stale, link) {
871 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) {
872 TAILQ_REMOVE(&stale, ce, link);
873 TAILQ_INSERT_TAIL(&td->clip_table, ce,
879 /* A new IP6 address; add it to the CLIP table */
880 ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT);
881 memcpy(&ce->lip, lip, sizeof(ce->lip));
883 rc = add_lip(sc, lip);
885 TAILQ_INSERT_TAIL(&td->clip_table, ce, link);
887 char ip[INET6_ADDRSTRLEN];
889 inet_ntop(AF_INET6, &ce->lip, &ip[0],
891 log(LOG_ERR, "%s: could not add %s (%d)\n",
899 last_vnet = (uintptr_t)vi->ifp->if_vnet;
903 * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are
904 * no longer referenced by the driver.
906 TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) {
907 if (ce->refcount == 0) {
908 rc = delete_lip(sc, &ce->lip);
910 TAILQ_REMOVE(&stale, ce, link);
913 char ip[INET6_ADDRSTRLEN];
915 inet_ntop(AF_INET6, &ce->lip, &ip[0],
917 log(LOG_ERR, "%s: could not delete %s (%d)\n",
922 /* The ones that are still referenced need to stay in the CLIP table */
923 TAILQ_CONCAT(&td->clip_table, &stale, link);
927 mtx_unlock(&td->clip_table_lock);
928 IN6_IFADDR_RUNLOCK(&in6_ifa_tracker);
932 destroy_clip_table(struct adapter *sc, struct tom_data *td)
934 struct clip_entry *ce, *ce_temp;
936 if (mtx_initialized(&td->clip_table_lock)) {
937 mtx_lock(&td->clip_table_lock);
938 TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) {
939 KASSERT(ce->refcount == 0,
940 ("%s: CLIP entry %p still in use (%d)", __func__,
942 TAILQ_REMOVE(&td->clip_table, ce, link);
943 delete_lip(sc, &ce->lip);
946 mtx_unlock(&td->clip_table_lock);
947 mtx_destroy(&td->clip_table_lock);
952 free_tom_data(struct adapter *sc, struct tom_data *td)
955 ASSERT_SYNCHRONIZED_OP(sc);
957 KASSERT(TAILQ_EMPTY(&td->toep_list),
958 ("%s: TOE PCB list is not empty.", __func__));
959 KASSERT(td->lctx_count == 0,
960 ("%s: lctx hash table is not empty.", __func__));
962 t4_free_ppod_region(&td->pr);
963 destroy_clip_table(sc, td);
965 if (td->listen_mask != 0)
966 hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
968 if (mtx_initialized(&td->unsent_wr_lock))
969 mtx_destroy(&td->unsent_wr_lock);
970 if (mtx_initialized(&td->lctx_hash_lock))
971 mtx_destroy(&td->lctx_hash_lock);
972 if (mtx_initialized(&td->toep_list_lock))
973 mtx_destroy(&td->toep_list_lock);
975 free_tid_tabs(&sc->tids);
980 reclaim_wr_resources(void *arg, int count)
982 struct tom_data *td = arg;
983 STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
984 struct cpl_act_open_req *cpl;
989 mtx_lock(&td->unsent_wr_lock);
990 STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
991 mtx_unlock(&td->unsent_wr_lock);
993 while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
994 STAILQ_REMOVE_HEAD(&twr_list, link);
997 opcode = GET_OPCODE(cpl);
1000 case CPL_ACT_OPEN_REQ:
1001 case CPL_ACT_OPEN_REQ6:
1002 atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
1003 sc = td_adapter(td);
1005 CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1006 act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
1010 log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1011 "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1012 /* WR not freed here; go look at it with a debugger. */
1018 * Ground control to Major TOM
1019 * Commencing countdown, engines on
1022 t4_tom_activate(struct adapter *sc)
1024 struct tom_data *td;
1027 struct sge_ofld_rxq *ofld_rxq;
1030 ASSERT_SYNCHRONIZED_OP(sc);
1032 /* per-adapter softc for TOM */
1033 td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1037 /* List of TOE PCBs and associated lock */
1038 mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1039 TAILQ_INIT(&td->toep_list);
1041 /* Listen context */
1042 mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1043 td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1044 &td->listen_mask, HASH_NOWAIT);
1046 /* List of WRs for which L2 resolution failed */
1047 mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1048 STAILQ_INIT(&td->unsent_wr_list);
1049 TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1052 rc = alloc_tid_tabs(&sc->tids);
1056 rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1057 t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1060 t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1061 V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1063 /* CLIP table for IPv6 offload */
1064 init_clip_table(sc, td);
1069 tod->tod_softc = sc;
1070 tod->tod_connect = t4_connect;
1071 tod->tod_listen_start = t4_listen_start;
1072 tod->tod_listen_stop = t4_listen_stop;
1073 tod->tod_rcvd = t4_rcvd;
1074 tod->tod_output = t4_tod_output;
1075 tod->tod_send_rst = t4_send_rst;
1076 tod->tod_send_fin = t4_send_fin;
1077 tod->tod_pcb_detach = t4_pcb_detach;
1078 tod->tod_l2_update = t4_l2_update;
1079 tod->tod_syncache_added = t4_syncache_added;
1080 tod->tod_syncache_removed = t4_syncache_removed;
1081 tod->tod_syncache_respond = t4_syncache_respond;
1082 tod->tod_offload_socket = t4_offload_socket;
1083 tod->tod_ctloutput = t4_ctloutput;
1085 for_each_port(sc, i) {
1086 for_each_vi(sc->port[i], v, vi) {
1087 TOEDEV(vi->ifp) = &td->tod;
1088 for_each_ofld_rxq(vi, j, ofld_rxq) {
1089 ofld_rxq->iq.set_tcb_rpl = do_set_tcb_rpl;
1090 ofld_rxq->iq.l2t_write_rpl = do_l2t_write_rpl2;
1096 register_toedev(sc->tom_softc);
1100 free_tom_data(sc, td);
1105 t4_tom_deactivate(struct adapter *sc)
1108 struct tom_data *td = sc->tom_softc;
1110 ASSERT_SYNCHRONIZED_OP(sc);
1113 return (0); /* XXX. KASSERT? */
1115 if (sc->offload_map != 0)
1116 return (EBUSY); /* at least one port has IFCAP_TOE enabled */
1118 if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1119 return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */
1121 mtx_lock(&td->toep_list_lock);
1122 if (!TAILQ_EMPTY(&td->toep_list))
1124 mtx_unlock(&td->toep_list_lock);
1126 mtx_lock(&td->lctx_hash_lock);
1127 if (td->lctx_count > 0)
1129 mtx_unlock(&td->lctx_hash_lock);
1131 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1132 mtx_lock(&td->unsent_wr_lock);
1133 if (!STAILQ_EMPTY(&td->unsent_wr_list))
1135 mtx_unlock(&td->unsent_wr_lock);
1138 unregister_toedev(sc->tom_softc);
1139 free_tom_data(sc, td);
1140 sc->tom_softc = NULL;
1147 t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp)
1150 atomic_add_rel_int(&in6_ifaddr_gen, 1);
1151 taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4);
1155 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
1157 struct tcpcb *tp = so_sototcpcb(so);
1158 struct toepcb *toep = tp->t_toe;
1161 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1162 error = t4_aio_queue_ddp(so, job);
1163 if (error != EOPNOTSUPP)
1167 return (t4_aio_queue_aiotx(so, job));
1171 t4_tom_mod_load(void)
1174 struct protosw *tcp_protosw, *tcp6_protosw;
1177 t4_init_connect_cpl_handlers();
1178 t4_init_listen_cpl_handlers();
1179 t4_init_cpl_io_handlers();
1181 rc = t4_ddp_mod_load();
1185 tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM);
1186 if (tcp_protosw == NULL)
1187 return (ENOPROTOOPT);
1188 bcopy(tcp_protosw, &toe_protosw, sizeof(toe_protosw));
1189 bcopy(tcp_protosw->pr_usrreqs, &toe_usrreqs, sizeof(toe_usrreqs));
1190 toe_usrreqs.pru_aio_queue = t4_aio_queue_tom;
1191 toe_protosw.pr_usrreqs = &toe_usrreqs;
1193 tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM);
1194 if (tcp6_protosw == NULL)
1195 return (ENOPROTOOPT);
1196 bcopy(tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
1197 bcopy(tcp6_protosw->pr_usrreqs, &toe6_usrreqs, sizeof(toe6_usrreqs));
1198 toe6_usrreqs.pru_aio_queue = t4_aio_queue_tom;
1199 toe6_protosw.pr_usrreqs = &toe6_usrreqs;
1201 TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL);
1202 ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event,
1203 t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY);
1205 rc = t4_register_uld(&tom_uld_info);
1207 t4_tom_mod_unload();
1213 tom_uninit(struct adapter *sc, void *arg __unused)
1215 if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
1218 /* Try to free resources (works only if no port has IFCAP_TOE) */
1219 if (uld_active(sc, ULD_TOM))
1220 t4_deactivate_uld(sc, ULD_TOM);
1222 end_synchronized_op(sc, 0);
1226 t4_tom_mod_unload(void)
1228 t4_iterate(tom_uninit, NULL);
1230 if (t4_unregister_uld(&tom_uld_info) == EBUSY)
1233 if (ifaddr_evhandler) {
1234 EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler);
1235 taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL);
1238 t4_ddp_mod_unload();
1240 t4_uninit_connect_cpl_handlers();
1241 t4_uninit_listen_cpl_handlers();
1242 t4_uninit_cpl_io_handlers();
1246 #endif /* TCP_OFFLOAD */
1249 t4_tom_modevent(module_t mod, int cmd, void *arg)
1256 rc = t4_tom_mod_load();
1260 rc = t4_tom_mod_unload();
1267 printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
1273 static moduledata_t t4_tom_moddata= {
1279 MODULE_VERSION(t4_tom, 1);
1280 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
1281 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
1282 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);