xref: /freebsd/sys/dev/cxgbe/tom/t4_tom.c (revision d940bfec8c329dd82d8d54efebd81c8aa420503b)
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 #include "opt_inet6.h"
33 
34 #include <sys/param.h>
35 #include <sys/types.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/ktr.h>
39 #include <sys/module.h>
40 #include <sys/protosw.h>
41 #include <sys/domain.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
44 #include <sys/taskqueue.h>
45 #include <net/if.h>
46 #include <netinet/in.h>
47 #include <netinet/in_pcb.h>
48 #include <netinet/in_var.h>
49 #include <netinet/ip.h>
50 #include <netinet/ip6.h>
51 #include <netinet/tcp_var.h>
52 #include <netinet6/scope6_var.h>
53 #define TCPSTATES
54 #include <netinet/tcp_fsm.h>
55 #include <netinet/toecore.h>
56 
57 #ifdef TCP_OFFLOAD
58 #include "common/common.h"
59 #include "common/t4_msg.h"
60 #include "common/t4_regs.h"
61 #include "common/t4_regs_values.h"
62 #include "common/t4_tcb.h"
63 #include "tom/t4_tom_l2t.h"
64 #include "tom/t4_tom.h"
65 
66 static struct protosw ddp_protosw;
67 static struct pr_usrreqs ddp_usrreqs;
68 
69 static struct protosw ddp6_protosw;
70 static struct pr_usrreqs ddp6_usrreqs;
71 
72 /* Module ops */
73 static int t4_tom_mod_load(void);
74 static int t4_tom_mod_unload(void);
75 static int t4_tom_modevent(module_t, int, void *);
76 
77 /* ULD ops and helpers */
78 static int t4_tom_activate(struct adapter *);
79 static int t4_tom_deactivate(struct adapter *);
80 
81 static struct uld_info tom_uld_info = {
82 	.uld_id = ULD_TOM,
83 	.activate = t4_tom_activate,
84 	.deactivate = t4_tom_deactivate,
85 };
86 
87 static void queue_tid_release(struct adapter *, int);
88 static void release_offload_resources(struct toepcb *);
89 static int alloc_tid_tabs(struct tid_info *);
90 static void free_tid_tabs(struct tid_info *);
91 static int add_lip(struct adapter *, struct in6_addr *);
92 static int delete_lip(struct adapter *, struct in6_addr *);
93 static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *);
94 static void init_clip_table(struct adapter *, struct tom_data *);
95 static void update_clip(struct adapter *, void *);
96 static void t4_clip_task(void *, int);
97 static void update_clip_table(struct adapter *, struct tom_data *);
98 static void destroy_clip_table(struct adapter *, struct tom_data *);
99 static void free_tom_data(struct adapter *, struct tom_data *);
100 
101 static int in6_ifaddr_gen;
102 static eventhandler_tag ifaddr_evhandler;
103 static struct timeout_task clip_task;
104 
105 struct toepcb *
106 alloc_toepcb(struct port_info *pi, int txqid, int rxqid, int flags)
107 {
108 	struct adapter *sc = pi->adapter;
109 	struct toepcb *toep;
110 	int tx_credits, txsd_total, len;
111 
112 	/*
113 	 * The firmware counts tx work request credits in units of 16 bytes
114 	 * each.  Reserve room for an ABORT_REQ so the driver never has to worry
115 	 * about tx credits if it wants to abort a connection.
116 	 */
117 	tx_credits = sc->params.ofldq_wr_cred;
118 	tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
119 
120 	/*
121 	 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
122 	 * immediate payload, and firmware counts tx work request credits in
123 	 * units of 16 byte.  Calculate the maximum work requests possible.
124 	 */
125 	txsd_total = tx_credits /
126 	    howmany((sizeof(struct fw_ofld_tx_data_wr) + 1), 16);
127 
128 	if (txqid < 0)
129 		txqid = (arc4random() % pi->nofldtxq) + pi->first_ofld_txq;
130 	KASSERT(txqid >= pi->first_ofld_txq &&
131 	    txqid < pi->first_ofld_txq + pi->nofldtxq,
132 	    ("%s: txqid %d for port %p (first %d, n %d)", __func__, txqid, pi,
133 		pi->first_ofld_txq, pi->nofldtxq));
134 
135 	if (rxqid < 0)
136 		rxqid = (arc4random() % pi->nofldrxq) + pi->first_ofld_rxq;
137 	KASSERT(rxqid >= pi->first_ofld_rxq &&
138 	    rxqid < pi->first_ofld_rxq + pi->nofldrxq,
139 	    ("%s: rxqid %d for port %p (first %d, n %d)", __func__, rxqid, pi,
140 		pi->first_ofld_rxq, pi->nofldrxq));
141 
142 	len = offsetof(struct toepcb, txsd) +
143 	    txsd_total * sizeof(struct ofld_tx_sdesc);
144 
145 	toep = malloc(len, M_CXGBE, M_ZERO | flags);
146 	if (toep == NULL)
147 		return (NULL);
148 
149 	toep->td = sc->tom_softc;
150 	toep->port = pi;
151 	toep->tx_credits = tx_credits;
152 	toep->ofld_txq = &sc->sge.ofld_txq[txqid];
153 	toep->ofld_rxq = &sc->sge.ofld_rxq[rxqid];
154 	toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
155 	toep->txsd_total = txsd_total;
156 	toep->txsd_avail = txsd_total;
157 	toep->txsd_pidx = 0;
158 	toep->txsd_cidx = 0;
159 
160 	return (toep);
161 }
162 
163 void
164 free_toepcb(struct toepcb *toep)
165 {
166 
167 	KASSERT(!(toep->flags & TPF_ATTACHED),
168 	    ("%s: attached to an inpcb", __func__));
169 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
170 	    ("%s: CPL pending", __func__));
171 
172 	free(toep, M_CXGBE);
173 }
174 
175 /*
176  * Set up the socket for TCP offload.
177  */
178 void
179 offload_socket(struct socket *so, struct toepcb *toep)
180 {
181 	struct tom_data *td = toep->td;
182 	struct inpcb *inp = sotoinpcb(so);
183 	struct tcpcb *tp = intotcpcb(inp);
184 	struct sockbuf *sb;
185 
186 	INP_WLOCK_ASSERT(inp);
187 
188 	/* Update socket */
189 	sb = &so->so_snd;
190 	SOCKBUF_LOCK(sb);
191 	sb->sb_flags |= SB_NOCOALESCE;
192 	SOCKBUF_UNLOCK(sb);
193 	sb = &so->so_rcv;
194 	SOCKBUF_LOCK(sb);
195 	sb->sb_flags |= SB_NOCOALESCE;
196 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
197 		if (inp->inp_vflag & INP_IPV6)
198 			so->so_proto = &ddp6_protosw;
199 		else
200 			so->so_proto = &ddp_protosw;
201 	}
202 	SOCKBUF_UNLOCK(sb);
203 
204 	/* Update TCP PCB */
205 	tp->tod = &td->tod;
206 	tp->t_toe = toep;
207 	tp->t_flags |= TF_TOE;
208 
209 	/* Install an extra hold on inp */
210 	toep->inp = inp;
211 	toep->flags |= TPF_ATTACHED;
212 	in_pcbref(inp);
213 
214 	/* Add the TOE PCB to the active list */
215 	mtx_lock(&td->toep_list_lock);
216 	TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
217 	mtx_unlock(&td->toep_list_lock);
218 }
219 
220 /* This is _not_ the normal way to "unoffload" a socket. */
221 void
222 undo_offload_socket(struct socket *so)
223 {
224 	struct inpcb *inp = sotoinpcb(so);
225 	struct tcpcb *tp = intotcpcb(inp);
226 	struct toepcb *toep = tp->t_toe;
227 	struct tom_data *td = toep->td;
228 	struct sockbuf *sb;
229 
230 	INP_WLOCK_ASSERT(inp);
231 
232 	sb = &so->so_snd;
233 	SOCKBUF_LOCK(sb);
234 	sb->sb_flags &= ~SB_NOCOALESCE;
235 	SOCKBUF_UNLOCK(sb);
236 	sb = &so->so_rcv;
237 	SOCKBUF_LOCK(sb);
238 	sb->sb_flags &= ~SB_NOCOALESCE;
239 	SOCKBUF_UNLOCK(sb);
240 
241 	tp->tod = NULL;
242 	tp->t_toe = NULL;
243 	tp->t_flags &= ~TF_TOE;
244 
245 	toep->inp = NULL;
246 	toep->flags &= ~TPF_ATTACHED;
247 	if (in_pcbrele_wlocked(inp))
248 		panic("%s: inp freed.", __func__);
249 
250 	mtx_lock(&td->toep_list_lock);
251 	TAILQ_REMOVE(&td->toep_list, toep, link);
252 	mtx_unlock(&td->toep_list_lock);
253 }
254 
255 static void
256 release_offload_resources(struct toepcb *toep)
257 {
258 	struct tom_data *td = toep->td;
259 	struct adapter *sc = td_adapter(td);
260 	int tid = toep->tid;
261 
262 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
263 	    ("%s: %p has CPL pending.", __func__, toep));
264 	KASSERT(!(toep->flags & TPF_ATTACHED),
265 	    ("%s: %p is still attached.", __func__, toep));
266 
267 	CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
268 	    __func__, toep, tid, toep->l2te, toep->ce);
269 
270 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
271 		release_ddp_resources(toep);
272 
273 	if (toep->l2te)
274 		t4_l2t_release(toep->l2te);
275 
276 	if (tid >= 0) {
277 		remove_tid(sc, tid);
278 		release_tid(sc, tid, toep->ctrlq);
279 	}
280 
281 	if (toep->ce)
282 		release_lip(td, toep->ce);
283 
284 	mtx_lock(&td->toep_list_lock);
285 	TAILQ_REMOVE(&td->toep_list, toep, link);
286 	mtx_unlock(&td->toep_list_lock);
287 
288 	free_toepcb(toep);
289 }
290 
291 /*
292  * The kernel is done with the TCP PCB and this is our opportunity to unhook the
293  * toepcb hanging off of it.  If the TOE driver is also done with the toepcb (no
294  * pending CPL) then it is time to release all resources tied to the toepcb.
295  *
296  * Also gets called when an offloaded active open fails and the TOM wants the
297  * kernel to take the TCP PCB back.
298  */
299 static void
300 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
301 {
302 #if defined(KTR) || defined(INVARIANTS)
303 	struct inpcb *inp = tp->t_inpcb;
304 #endif
305 	struct toepcb *toep = tp->t_toe;
306 
307 	INP_WLOCK_ASSERT(inp);
308 
309 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
310 	KASSERT(toep->flags & TPF_ATTACHED,
311 	    ("%s: not attached", __func__));
312 
313 #ifdef KTR
314 	if (tp->t_state == TCPS_SYN_SENT) {
315 		CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
316 		    __func__, toep->tid, toep, toep->flags, inp,
317 		    inp->inp_flags);
318 	} else {
319 		CTR6(KTR_CXGBE,
320 		    "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
321 		    toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
322 		    inp->inp_flags);
323 	}
324 #endif
325 
326 	tp->t_toe = NULL;
327 	tp->t_flags &= ~TF_TOE;
328 	toep->flags &= ~TPF_ATTACHED;
329 
330 	if (!(toep->flags & TPF_CPL_PENDING))
331 		release_offload_resources(toep);
332 }
333 
334 /*
335  * setsockopt handler.
336  */
337 static void
338 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
339 {
340 	struct adapter *sc = tod->tod_softc;
341 	struct toepcb *toep = tp->t_toe;
342 
343 	if (dir == SOPT_GET)
344 		return;
345 
346 	CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
347 
348 	switch (name) {
349 	case TCP_NODELAY:
350 		t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS, V_TF_NAGLE(1),
351 		    V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1));
352 		break;
353 	default:
354 		break;
355 	}
356 }
357 
358 /*
359  * The TOE driver will not receive any more CPLs for the tid associated with the
360  * toepcb; release the hold on the inpcb.
361  */
362 void
363 final_cpl_received(struct toepcb *toep)
364 {
365 	struct inpcb *inp = toep->inp;
366 
367 	KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
368 	INP_WLOCK_ASSERT(inp);
369 	KASSERT(toep->flags & TPF_CPL_PENDING,
370 	    ("%s: CPL not pending already?", __func__));
371 
372 	CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
373 	    __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
374 
375 	toep->inp = NULL;
376 	toep->flags &= ~TPF_CPL_PENDING;
377 
378 	if (!(toep->flags & TPF_ATTACHED))
379 		release_offload_resources(toep);
380 
381 	if (!in_pcbrele_wlocked(inp))
382 		INP_WUNLOCK(inp);
383 }
384 
385 void
386 insert_tid(struct adapter *sc, int tid, void *ctx)
387 {
388 	struct tid_info *t = &sc->tids;
389 
390 	t->tid_tab[tid] = ctx;
391 	atomic_add_int(&t->tids_in_use, 1);
392 }
393 
394 void *
395 lookup_tid(struct adapter *sc, int tid)
396 {
397 	struct tid_info *t = &sc->tids;
398 
399 	return (t->tid_tab[tid]);
400 }
401 
402 void
403 update_tid(struct adapter *sc, int tid, void *ctx)
404 {
405 	struct tid_info *t = &sc->tids;
406 
407 	t->tid_tab[tid] = ctx;
408 }
409 
410 void
411 remove_tid(struct adapter *sc, int tid)
412 {
413 	struct tid_info *t = &sc->tids;
414 
415 	t->tid_tab[tid] = NULL;
416 	atomic_subtract_int(&t->tids_in_use, 1);
417 }
418 
419 void
420 release_tid(struct adapter *sc, int tid, struct sge_wrq *ctrlq)
421 {
422 	struct wrqe *wr;
423 	struct cpl_tid_release *req;
424 
425 	wr = alloc_wrqe(sizeof(*req), ctrlq);
426 	if (wr == NULL) {
427 		queue_tid_release(sc, tid);	/* defer */
428 		return;
429 	}
430 	req = wrtod(wr);
431 
432 	INIT_TP_WR_MIT_CPL(req, CPL_TID_RELEASE, tid);
433 
434 	t4_wrq_tx(sc, wr);
435 }
436 
437 static void
438 queue_tid_release(struct adapter *sc, int tid)
439 {
440 
441 	CXGBE_UNIMPLEMENTED("deferred tid release");
442 }
443 
444 /*
445  * What mtu_idx to use, given a 4-tuple and/or an MSS cap
446  */
447 int
448 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, int pmss)
449 {
450 	unsigned short *mtus = &sc->params.mtus[0];
451 	int i, mss, n;
452 
453 	KASSERT(inc != NULL || pmss > 0,
454 	    ("%s: at least one of inc/pmss must be specified", __func__));
455 
456 	mss = inc ? tcp_mssopt(inc) : pmss;
457 	if (pmss > 0 && mss > pmss)
458 		mss = pmss;
459 
460 	if (inc->inc_flags & INC_ISIPV6)
461 		n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
462 	else
463 		n = sizeof(struct ip) + sizeof(struct tcphdr);
464 
465 	for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mss + n; i++)
466 		continue;
467 
468 	return (i);
469 }
470 
471 /*
472  * Determine the receive window size for a socket.
473  */
474 u_long
475 select_rcv_wnd(struct socket *so)
476 {
477 	unsigned long wnd;
478 
479 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
480 
481 	wnd = sbspace(&so->so_rcv);
482 	if (wnd < MIN_RCV_WND)
483 		wnd = MIN_RCV_WND;
484 
485 	return min(wnd, MAX_RCV_WND);
486 }
487 
488 int
489 select_rcv_wscale(void)
490 {
491 	int wscale = 0;
492 	unsigned long space = sb_max;
493 
494 	if (space > MAX_RCV_WND)
495 		space = MAX_RCV_WND;
496 
497 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
498 		wscale++;
499 
500 	return (wscale);
501 }
502 
503 extern int always_keepalive;
504 #define VIID_SMACIDX(v)	(((unsigned int)(v) & 0x7f) << 1)
505 
506 /*
507  * socket so could be a listening socket too.
508  */
509 uint64_t
510 calc_opt0(struct socket *so, struct port_info *pi, struct l2t_entry *e,
511     int mtu_idx, int rscale, int rx_credits, int ulp_mode)
512 {
513 	uint64_t opt0;
514 
515 	KASSERT(rx_credits <= M_RCV_BUFSIZ,
516 	    ("%s: rcv_bufsiz too high", __func__));
517 
518 	opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) |
519 	    V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits);
520 
521 	if (so != NULL) {
522 		struct inpcb *inp = sotoinpcb(so);
523 		struct tcpcb *tp = intotcpcb(inp);
524 		int keepalive = always_keepalive ||
525 		    so_options_get(so) & SO_KEEPALIVE;
526 
527 		opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0);
528 		opt0 |= V_KEEP_ALIVE(keepalive != 0);
529 	}
530 
531 	if (e != NULL)
532 		opt0 |= V_L2T_IDX(e->idx);
533 
534 	if (pi != NULL) {
535 		opt0 |= V_SMAC_SEL(VIID_SMACIDX(pi->viid));
536 		opt0 |= V_TX_CHAN(pi->tx_chan);
537 	}
538 
539 	return htobe64(opt0);
540 }
541 
542 uint64_t
543 select_ntuple(struct port_info *pi, struct l2t_entry *e)
544 {
545 	struct adapter *sc = pi->adapter;
546 	struct tp_params *tp = &sc->params.tp;
547 	uint16_t viid = pi->viid;
548 	uint64_t ntuple = 0;
549 
550 	/*
551 	 * Initialize each of the fields which we care about which are present
552 	 * in the Compressed Filter Tuple.
553 	 */
554 	if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE)
555 		ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
556 
557 	if (tp->port_shift >= 0)
558 		ntuple |= (uint64_t)e->lport << tp->port_shift;
559 
560 	if (tp->protocol_shift >= 0)
561 		ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
562 
563 	if (tp->vnic_shift >= 0) {
564 		uint32_t vf = G_FW_VIID_VIN(viid);
565 		uint32_t pf = G_FW_VIID_PFN(viid);
566 		uint32_t vld = G_FW_VIID_VIVLD(viid);
567 
568 		ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) |
569 		    V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift;
570 	}
571 
572 	if (is_t4(sc))
573 		return (htobe32((uint32_t)ntuple));
574 	else
575 		return (htobe64(V_FILTER_TUPLE(ntuple)));
576 }
577 
578 void
579 set_tcpddp_ulp_mode(struct toepcb *toep)
580 {
581 
582 	toep->ulp_mode = ULP_MODE_TCPDDP;
583 	toep->ddp_flags = DDP_OK;
584 	toep->ddp_score = DDP_LOW_SCORE;
585 }
586 
587 int
588 negative_advice(int status)
589 {
590 
591 	return (status == CPL_ERR_RTX_NEG_ADVICE ||
592 	    status == CPL_ERR_PERSIST_NEG_ADVICE ||
593 	    status == CPL_ERR_KEEPALV_NEG_ADVICE);
594 }
595 
596 static int
597 alloc_tid_tabs(struct tid_info *t)
598 {
599 	size_t size;
600 	unsigned int i;
601 
602 	size = t->ntids * sizeof(*t->tid_tab) +
603 	    t->natids * sizeof(*t->atid_tab) +
604 	    t->nstids * sizeof(*t->stid_tab);
605 
606 	t->tid_tab = malloc(size, M_CXGBE, M_ZERO | M_NOWAIT);
607 	if (t->tid_tab == NULL)
608 		return (ENOMEM);
609 
610 	mtx_init(&t->atid_lock, "atid lock", NULL, MTX_DEF);
611 	t->atid_tab = (union aopen_entry *)&t->tid_tab[t->ntids];
612 	t->afree = t->atid_tab;
613 	t->atids_in_use = 0;
614 	for (i = 1; i < t->natids; i++)
615 		t->atid_tab[i - 1].next = &t->atid_tab[i];
616 	t->atid_tab[t->natids - 1].next = NULL;
617 
618 	mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
619 	t->stid_tab = (struct listen_ctx **)&t->atid_tab[t->natids];
620 	t->stids_in_use = 0;
621 	TAILQ_INIT(&t->stids);
622 	t->nstids_free_head = t->nstids;
623 
624 	atomic_store_rel_int(&t->tids_in_use, 0);
625 
626 	return (0);
627 }
628 
629 static void
630 free_tid_tabs(struct tid_info *t)
631 {
632 	KASSERT(t->tids_in_use == 0,
633 	    ("%s: %d tids still in use.", __func__, t->tids_in_use));
634 	KASSERT(t->atids_in_use == 0,
635 	    ("%s: %d atids still in use.", __func__, t->atids_in_use));
636 	KASSERT(t->stids_in_use == 0,
637 	    ("%s: %d tids still in use.", __func__, t->stids_in_use));
638 
639 	free(t->tid_tab, M_CXGBE);
640 	t->tid_tab = NULL;
641 
642 	if (mtx_initialized(&t->atid_lock))
643 		mtx_destroy(&t->atid_lock);
644 	if (mtx_initialized(&t->stid_lock))
645 		mtx_destroy(&t->stid_lock);
646 }
647 
648 static int
649 add_lip(struct adapter *sc, struct in6_addr *lip)
650 {
651         struct fw_clip_cmd c;
652 
653 	ASSERT_SYNCHRONIZED_OP(sc);
654 	/* mtx_assert(&td->clip_table_lock, MA_OWNED); */
655 
656         memset(&c, 0, sizeof(c));
657 	c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
658 	    F_FW_CMD_WRITE);
659         c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c));
660         c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
661         c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
662 
663 	return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
664 }
665 
666 static int
667 delete_lip(struct adapter *sc, struct in6_addr *lip)
668 {
669 	struct fw_clip_cmd c;
670 
671 	ASSERT_SYNCHRONIZED_OP(sc);
672 	/* mtx_assert(&td->clip_table_lock, MA_OWNED); */
673 
674 	memset(&c, 0, sizeof(c));
675 	c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
676 	    F_FW_CMD_READ);
677         c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c));
678         c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
679         c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
680 
681 	return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
682 }
683 
684 static struct clip_entry *
685 search_lip(struct tom_data *td, struct in6_addr *lip)
686 {
687 	struct clip_entry *ce;
688 
689 	mtx_assert(&td->clip_table_lock, MA_OWNED);
690 
691 	TAILQ_FOREACH(ce, &td->clip_table, link) {
692 		if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
693 			return (ce);
694 	}
695 
696 	return (NULL);
697 }
698 
699 struct clip_entry *
700 hold_lip(struct tom_data *td, struct in6_addr *lip)
701 {
702 	struct clip_entry *ce;
703 
704 	mtx_lock(&td->clip_table_lock);
705 	ce = search_lip(td, lip);
706 	if (ce != NULL)
707 		ce->refcount++;
708 	mtx_unlock(&td->clip_table_lock);
709 
710 	return (ce);
711 }
712 
713 void
714 release_lip(struct tom_data *td, struct clip_entry *ce)
715 {
716 
717 	mtx_lock(&td->clip_table_lock);
718 	KASSERT(search_lip(td, &ce->lip) == ce,
719 	    ("%s: CLIP entry %p p not in CLIP table.", __func__, ce));
720 	KASSERT(ce->refcount > 0,
721 	    ("%s: CLIP entry %p has refcount 0", __func__, ce));
722 	--ce->refcount;
723 	mtx_unlock(&td->clip_table_lock);
724 }
725 
726 static void
727 init_clip_table(struct adapter *sc, struct tom_data *td)
728 {
729 
730 	ASSERT_SYNCHRONIZED_OP(sc);
731 
732 	mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF);
733 	TAILQ_INIT(&td->clip_table);
734 	td->clip_gen = -1;
735 
736 	update_clip_table(sc, td);
737 }
738 
739 static void
740 update_clip(struct adapter *sc, void *arg __unused)
741 {
742 
743 	if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc"))
744 		return;
745 
746 	if (sc->flags & TOM_INIT_DONE)
747 		update_clip_table(sc, sc->tom_softc);
748 
749 	end_synchronized_op(sc, LOCK_HELD);
750 }
751 
752 static void
753 t4_clip_task(void *arg, int count)
754 {
755 
756 	t4_iterate(update_clip, NULL);
757 }
758 
759 static void
760 update_clip_table(struct adapter *sc, struct tom_data *td)
761 {
762 	struct in6_ifaddr *ia;
763 	struct in6_addr *lip, tlip;
764 	struct clip_head stale;
765 	struct clip_entry *ce, *ce_temp;
766 	int rc, gen = atomic_load_acq_int(&in6_ifaddr_gen);
767 
768 	ASSERT_SYNCHRONIZED_OP(sc);
769 
770 	IN6_IFADDR_RLOCK();
771 	mtx_lock(&td->clip_table_lock);
772 
773 	if (gen == td->clip_gen)
774 		goto done;
775 
776 	TAILQ_INIT(&stale);
777 	TAILQ_CONCAT(&stale, &td->clip_table, link);
778 
779 	TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) {
780 		lip = &ia->ia_addr.sin6_addr;
781 
782 		KASSERT(!IN6_IS_ADDR_MULTICAST(lip),
783 		    ("%s: mcast address in in6_ifaddr list", __func__));
784 
785 		if (IN6_IS_ADDR_LOOPBACK(lip))
786 			continue;
787 		if (IN6_IS_SCOPE_EMBED(lip)) {
788 			/* Remove the embedded scope */
789 			tlip = *lip;
790 			lip = &tlip;
791 			in6_clearscope(lip);
792 		}
793 		/*
794 		 * XXX: how to weed out the link local address for the loopback
795 		 * interface?  It's fe80::1 usually (always?).
796 		 */
797 
798 		/*
799 		 * If it's in the main list then we already know it's not stale.
800 		 */
801 		TAILQ_FOREACH(ce, &td->clip_table, link) {
802 			if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
803 				goto next;
804 		}
805 
806 		/*
807 		 * If it's in the stale list we should move it to the main list.
808 		 */
809 		TAILQ_FOREACH(ce, &stale, link) {
810 			if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) {
811 				TAILQ_REMOVE(&stale, ce, link);
812 				TAILQ_INSERT_TAIL(&td->clip_table, ce, link);
813 				goto next;
814 			}
815 		}
816 
817 		/* A new IP6 address; add it to the CLIP table */
818 		ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT);
819 		memcpy(&ce->lip, lip, sizeof(ce->lip));
820 		ce->refcount = 0;
821 		rc = add_lip(sc, lip);
822 		if (rc == 0)
823 			TAILQ_INSERT_TAIL(&td->clip_table, ce, link);
824 		else {
825 			char ip[INET6_ADDRSTRLEN];
826 
827 			inet_ntop(AF_INET6, &ce->lip, &ip[0], sizeof(ip));
828 			log(LOG_ERR, "%s: could not add %s (%d)\n",
829 			    __func__, ip, rc);
830 			free(ce, M_CXGBE);
831 		}
832 next:
833 		continue;
834 	}
835 
836 	/*
837 	 * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are
838 	 * no longer referenced by the driver.
839 	 */
840 	TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) {
841 		if (ce->refcount == 0) {
842 			rc = delete_lip(sc, &ce->lip);
843 			if (rc == 0) {
844 				TAILQ_REMOVE(&stale, ce, link);
845 				free(ce, M_CXGBE);
846 			} else {
847 				char ip[INET6_ADDRSTRLEN];
848 
849 				inet_ntop(AF_INET6, &ce->lip, &ip[0],
850 				    sizeof(ip));
851 				log(LOG_ERR, "%s: could not delete %s (%d)\n",
852 				    __func__, ip, rc);
853 			}
854 		}
855 	}
856 	/* The ones that are still referenced need to stay in the CLIP table */
857 	TAILQ_CONCAT(&td->clip_table, &stale, link);
858 
859 	td->clip_gen = gen;
860 done:
861 	mtx_unlock(&td->clip_table_lock);
862 	IN6_IFADDR_RUNLOCK();
863 }
864 
865 static void
866 destroy_clip_table(struct adapter *sc, struct tom_data *td)
867 {
868 	struct clip_entry *ce, *ce_temp;
869 
870 	if (mtx_initialized(&td->clip_table_lock)) {
871 		mtx_lock(&td->clip_table_lock);
872 		TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) {
873 			KASSERT(ce->refcount == 0,
874 			    ("%s: CLIP entry %p still in use (%d)", __func__,
875 			    ce, ce->refcount));
876 			TAILQ_REMOVE(&td->clip_table, ce, link);
877 			delete_lip(sc, &ce->lip);
878 			free(ce, M_CXGBE);
879 		}
880 		mtx_unlock(&td->clip_table_lock);
881 		mtx_destroy(&td->clip_table_lock);
882 	}
883 }
884 
885 static void
886 free_tom_data(struct adapter *sc, struct tom_data *td)
887 {
888 
889 	ASSERT_SYNCHRONIZED_OP(sc);
890 
891 	KASSERT(TAILQ_EMPTY(&td->toep_list),
892 	    ("%s: TOE PCB list is not empty.", __func__));
893 	KASSERT(td->lctx_count == 0,
894 	    ("%s: lctx hash table is not empty.", __func__));
895 
896 	t4_uninit_l2t_cpl_handlers(sc);
897 	t4_uninit_cpl_io_handlers(sc);
898 	t4_uninit_ddp(sc, td);
899 	destroy_clip_table(sc, td);
900 
901 	if (td->listen_mask != 0)
902 		hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
903 
904 	if (mtx_initialized(&td->lctx_hash_lock))
905 		mtx_destroy(&td->lctx_hash_lock);
906 	if (mtx_initialized(&td->toep_list_lock))
907 		mtx_destroy(&td->toep_list_lock);
908 
909 	free_tid_tabs(&sc->tids);
910 	free(td, M_CXGBE);
911 }
912 
913 /*
914  * Ground control to Major TOM
915  * Commencing countdown, engines on
916  */
917 static int
918 t4_tom_activate(struct adapter *sc)
919 {
920 	struct tom_data *td;
921 	struct toedev *tod;
922 	int i, rc;
923 
924 	ASSERT_SYNCHRONIZED_OP(sc);
925 
926 	/* per-adapter softc for TOM */
927 	td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
928 	if (td == NULL)
929 		return (ENOMEM);
930 
931 	/* List of TOE PCBs and associated lock */
932 	mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
933 	TAILQ_INIT(&td->toep_list);
934 
935 	/* Listen context */
936 	mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
937 	td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
938 	    &td->listen_mask, HASH_NOWAIT);
939 
940 	/* TID tables */
941 	rc = alloc_tid_tabs(&sc->tids);
942 	if (rc != 0)
943 		goto done;
944 
945 	/* DDP page pods and CPL handlers */
946 	t4_init_ddp(sc, td);
947 
948 	/* CLIP table for IPv6 offload */
949 	init_clip_table(sc, td);
950 
951 	/* CPL handlers */
952 	t4_init_connect_cpl_handlers(sc);
953 	t4_init_l2t_cpl_handlers(sc);
954 	t4_init_listen_cpl_handlers(sc);
955 	t4_init_cpl_io_handlers(sc);
956 
957 	/* toedev ops */
958 	tod = &td->tod;
959 	init_toedev(tod);
960 	tod->tod_softc = sc;
961 	tod->tod_connect = t4_connect;
962 	tod->tod_listen_start = t4_listen_start;
963 	tod->tod_listen_stop = t4_listen_stop;
964 	tod->tod_rcvd = t4_rcvd;
965 	tod->tod_output = t4_tod_output;
966 	tod->tod_send_rst = t4_send_rst;
967 	tod->tod_send_fin = t4_send_fin;
968 	tod->tod_pcb_detach = t4_pcb_detach;
969 	tod->tod_l2_update = t4_l2_update;
970 	tod->tod_syncache_added = t4_syncache_added;
971 	tod->tod_syncache_removed = t4_syncache_removed;
972 	tod->tod_syncache_respond = t4_syncache_respond;
973 	tod->tod_offload_socket = t4_offload_socket;
974 	tod->tod_ctloutput = t4_ctloutput;
975 
976 	for_each_port(sc, i)
977 		TOEDEV(sc->port[i]->ifp) = &td->tod;
978 
979 	sc->tom_softc = td;
980 	sc->flags |= TOM_INIT_DONE;
981 	register_toedev(sc->tom_softc);
982 
983 done:
984 	if (rc != 0)
985 		free_tom_data(sc, td);
986 	return (rc);
987 }
988 
989 static int
990 t4_tom_deactivate(struct adapter *sc)
991 {
992 	int rc = 0;
993 	struct tom_data *td = sc->tom_softc;
994 
995 	ASSERT_SYNCHRONIZED_OP(sc);
996 
997 	if (td == NULL)
998 		return (0);	/* XXX. KASSERT? */
999 
1000 	if (sc->offload_map != 0)
1001 		return (EBUSY);	/* at least one port has IFCAP_TOE enabled */
1002 
1003 	mtx_lock(&td->toep_list_lock);
1004 	if (!TAILQ_EMPTY(&td->toep_list))
1005 		rc = EBUSY;
1006 	mtx_unlock(&td->toep_list_lock);
1007 
1008 	mtx_lock(&td->lctx_hash_lock);
1009 	if (td->lctx_count > 0)
1010 		rc = EBUSY;
1011 	mtx_unlock(&td->lctx_hash_lock);
1012 
1013 	if (rc == 0) {
1014 		unregister_toedev(sc->tom_softc);
1015 		free_tom_data(sc, td);
1016 		sc->tom_softc = NULL;
1017 		sc->flags &= ~TOM_INIT_DONE;
1018 	}
1019 
1020 	return (rc);
1021 }
1022 
1023 static void
1024 t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp)
1025 {
1026 
1027 	atomic_add_rel_int(&in6_ifaddr_gen, 1);
1028 	taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4);
1029 }
1030 
1031 static int
1032 t4_tom_mod_load(void)
1033 {
1034 	int rc;
1035 	struct protosw *tcp_protosw, *tcp6_protosw;
1036 
1037 	tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM);
1038 	if (tcp_protosw == NULL)
1039 		return (ENOPROTOOPT);
1040 	bcopy(tcp_protosw, &ddp_protosw, sizeof(ddp_protosw));
1041 	bcopy(tcp_protosw->pr_usrreqs, &ddp_usrreqs, sizeof(ddp_usrreqs));
1042 	ddp_usrreqs.pru_soreceive = t4_soreceive_ddp;
1043 	ddp_protosw.pr_usrreqs = &ddp_usrreqs;
1044 
1045 	tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM);
1046 	if (tcp6_protosw == NULL)
1047 		return (ENOPROTOOPT);
1048 	bcopy(tcp6_protosw, &ddp6_protosw, sizeof(ddp6_protosw));
1049 	bcopy(tcp6_protosw->pr_usrreqs, &ddp6_usrreqs, sizeof(ddp6_usrreqs));
1050 	ddp6_usrreqs.pru_soreceive = t4_soreceive_ddp;
1051 	ddp6_protosw.pr_usrreqs = &ddp6_usrreqs;
1052 
1053 	TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL);
1054 	ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event,
1055 	    t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY);
1056 
1057 	rc = t4_register_uld(&tom_uld_info);
1058 	if (rc != 0)
1059 		t4_tom_mod_unload();
1060 
1061 	return (rc);
1062 }
1063 
1064 static void
1065 tom_uninit(struct adapter *sc, void *arg __unused)
1066 {
1067 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
1068 		return;
1069 
1070 	/* Try to free resources (works only if no port has IFCAP_TOE) */
1071 	if (sc->flags & TOM_INIT_DONE)
1072 		t4_deactivate_uld(sc, ULD_TOM);
1073 
1074 	end_synchronized_op(sc, 0);
1075 }
1076 
1077 static int
1078 t4_tom_mod_unload(void)
1079 {
1080 	t4_iterate(tom_uninit, NULL);
1081 
1082 	if (t4_unregister_uld(&tom_uld_info) == EBUSY)
1083 		return (EBUSY);
1084 
1085 	if (ifaddr_evhandler) {
1086 		EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler);
1087 		taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL);
1088 	}
1089 
1090 	return (0);
1091 }
1092 #endif	/* TCP_OFFLOAD */
1093 
1094 static int
1095 t4_tom_modevent(module_t mod, int cmd, void *arg)
1096 {
1097 	int rc = 0;
1098 
1099 #ifdef TCP_OFFLOAD
1100 	switch (cmd) {
1101 	case MOD_LOAD:
1102 		rc = t4_tom_mod_load();
1103 		break;
1104 
1105 	case MOD_UNLOAD:
1106 		rc = t4_tom_mod_unload();
1107 		break;
1108 
1109 	default:
1110 		rc = EINVAL;
1111 	}
1112 #else
1113 	printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
1114 	rc = EOPNOTSUPP;
1115 #endif
1116 	return (rc);
1117 }
1118 
1119 static moduledata_t t4_tom_moddata= {
1120 	"t4_tom",
1121 	t4_tom_modevent,
1122 	0
1123 };
1124 
1125 MODULE_VERSION(t4_tom, 1);
1126 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
1127 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
1128 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
1129