1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (c) 2012 Chelsio Communications, Inc.
5 * All rights reserved.
6 * Written by: Navdeep Parhar <np@FreeBSD.org>
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
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.
16 *
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
27 * SUCH DAMAGE.
28 */
29
30 #include <sys/cdefs.h>
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_kern_tls.h"
34 #include "opt_ratelimit.h"
35
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/ktr.h>
41 #include <sys/lock.h>
42 #include <sys/limits.h>
43 #include <sys/module.h>
44 #include <sys/protosw.h>
45 #include <sys/domain.h>
46 #include <sys/refcount.h>
47 #include <sys/rmlock.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/taskqueue.h>
52 #include <net/if.h>
53 #include <net/if_var.h>
54 #include <net/if_types.h>
55 #include <net/if_vlan_var.h>
56 #include <netinet/in.h>
57 #include <netinet/in_pcb.h>
58 #include <netinet/in_var.h>
59 #include <netinet/ip.h>
60 #include <netinet/ip6.h>
61 #include <netinet6/scope6_var.h>
62 #define TCPSTATES
63 #include <netinet/tcp_fsm.h>
64 #include <netinet/tcp_seq.h>
65 #include <netinet/tcp_timer.h>
66 #include <netinet/tcp_var.h>
67 #include <netinet/toecore.h>
68 #include <netinet/cc/cc.h>
69
70 #ifdef TCP_OFFLOAD
71 #include "common/common.h"
72 #include "common/t4_msg.h"
73 #include "common/t4_regs.h"
74 #include "common/t4_regs_values.h"
75 #include "common/t4_tcb.h"
76 #include "t4_clip.h"
77 #include "tom/t4_tom_l2t.h"
78 #include "tom/t4_tom.h"
79 #include "tom/t4_tls.h"
80
81 static struct protosw toe_protosw;
82 static struct protosw toe6_protosw;
83
84 /* Module ops */
85 static int t4_tom_mod_load(void);
86 static int t4_tom_mod_unload(void);
87 static int t4_tom_modevent(module_t, int, void *);
88
89 /* ULD ops and helpers */
90 static int t4_tom_activate(struct adapter *);
91 static int t4_tom_deactivate(struct adapter *);
92 static int t4_tom_stop(struct adapter *);
93 static int t4_tom_restart(struct adapter *);
94
95 static struct uld_info tom_uld_info = {
96 .uld_activate = t4_tom_activate,
97 .uld_deactivate = t4_tom_deactivate,
98 .uld_stop = t4_tom_stop,
99 .uld_restart = t4_tom_restart,
100 };
101
102 static void release_offload_resources(struct toepcb *);
103 static void done_with_toepcb(struct toepcb *);
104 static int alloc_tid_tabs(struct adapter *);
105 static void free_tid_tabs(struct adapter *);
106 static void free_tom_data(struct adapter *, struct tom_data *);
107 static void reclaim_wr_resources(void *, int);
108 static void cleanup_stranded_tids(void *, int);
109
110 struct toepcb *
alloc_toepcb(struct vi_info * vi,int flags)111 alloc_toepcb(struct vi_info *vi, int flags)
112 {
113 struct port_info *pi = vi->pi;
114 struct adapter *sc = pi->adapter;
115 struct toepcb *toep;
116 int tx_credits, txsd_total, len;
117
118 /*
119 * The firmware counts tx work request credits in units of 16 bytes
120 * each. Reserve room for an ABORT_REQ so the driver never has to worry
121 * about tx credits if it wants to abort a connection.
122 */
123 tx_credits = sc->params.ofldq_wr_cred;
124 tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
125
126 /*
127 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
128 * immediate payload, and firmware counts tx work request credits in
129 * units of 16 byte. Calculate the maximum work requests possible.
130 */
131 txsd_total = tx_credits /
132 howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16);
133
134 len = offsetof(struct toepcb, txsd) +
135 txsd_total * sizeof(struct ofld_tx_sdesc);
136
137 toep = malloc(len, M_CXGBE, M_ZERO | flags);
138 if (toep == NULL)
139 return (NULL);
140
141 refcount_init(&toep->refcount, 1);
142 toep->td = sc->tom_softc;
143 toep->incarnation = sc->incarnation;
144 toep->vi = vi;
145 toep->tid = -1;
146 toep->tx_total = tx_credits;
147 toep->tx_credits = tx_credits;
148 mbufq_init(&toep->ulp_pduq, INT_MAX);
149 mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX);
150 toep->txsd_total = txsd_total;
151 toep->txsd_avail = txsd_total;
152 toep->txsd_pidx = 0;
153 toep->txsd_cidx = 0;
154 aiotx_init_toep(toep);
155
156 return (toep);
157 }
158
159 /*
160 * Initialize a toepcb after its params have been filled out.
161 */
162 int
init_toepcb(struct vi_info * vi,struct toepcb * toep)163 init_toepcb(struct vi_info *vi, struct toepcb *toep)
164 {
165 struct conn_params *cp = &toep->params;
166 struct port_info *pi = vi->pi;
167 struct adapter *sc = pi->adapter;
168 struct tx_cl_rl_params *tc;
169
170 if (cp->tc_idx >= 0 && cp->tc_idx < sc->params.nsched_cls) {
171 tc = &pi->sched_params->cl_rl[cp->tc_idx];
172 mtx_lock(&sc->tc_lock);
173 if (tc->state != CS_HW_CONFIGURED) {
174 CH_ERR(vi, "tid %d cannot be bound to traffic class %d "
175 "because it is not configured (its state is %d)\n",
176 toep->tid, cp->tc_idx, tc->state);
177 cp->tc_idx = -1;
178 } else {
179 tc->refcount++;
180 }
181 mtx_unlock(&sc->tc_lock);
182 }
183 toep->ofld_txq = &sc->sge.ofld_txq[cp->txq_idx];
184 toep->ofld_rxq = &sc->sge.ofld_rxq[cp->rxq_idx];
185 toep->ctrlq = &sc->sge.ctrlq[cp->ctrlq_idx];
186
187 tls_init_toep(toep);
188 MPASS(ulp_mode(toep) != ULP_MODE_TCPDDP);
189
190 toep->flags |= TPF_INITIALIZED;
191
192 return (0);
193 }
194
195 struct toepcb *
hold_toepcb(struct toepcb * toep)196 hold_toepcb(struct toepcb *toep)
197 {
198
199 refcount_acquire(&toep->refcount);
200 return (toep);
201 }
202
203 void
free_toepcb(struct toepcb * toep)204 free_toepcb(struct toepcb *toep)
205 {
206
207 if (refcount_release(&toep->refcount) == 0)
208 return;
209
210 KASSERT(!(toep->flags & TPF_ATTACHED),
211 ("%s: attached to an inpcb", __func__));
212 KASSERT(!(toep->flags & TPF_CPL_PENDING),
213 ("%s: CPL pending", __func__));
214
215 if (toep->flags & TPF_INITIALIZED) {
216 if (ulp_mode(toep) == ULP_MODE_TCPDDP)
217 ddp_uninit_toep(toep);
218 tls_uninit_toep(toep);
219 }
220 free(toep, M_CXGBE);
221 }
222
223 /*
224 * Set up the socket for TCP offload.
225 */
226 void
offload_socket(struct socket * so,struct toepcb * toep)227 offload_socket(struct socket *so, struct toepcb *toep)
228 {
229 struct tom_data *td = toep->td;
230 struct inpcb *inp = sotoinpcb(so);
231 struct tcpcb *tp = intotcpcb(inp);
232 struct sockbuf *sb;
233
234 INP_WLOCK_ASSERT(inp);
235
236 /* Update socket */
237 sb = &so->so_snd;
238 SOCKBUF_LOCK(sb);
239 sb->sb_flags |= SB_NOCOALESCE;
240 SOCKBUF_UNLOCK(sb);
241 sb = &so->so_rcv;
242 SOCKBUF_LOCK(sb);
243 sb->sb_flags |= SB_NOCOALESCE;
244 if (inp->inp_vflag & INP_IPV6)
245 so->so_proto = &toe6_protosw;
246 else
247 so->so_proto = &toe_protosw;
248 SOCKBUF_UNLOCK(sb);
249
250 /* Update TCP PCB */
251 tp->tod = &td->tod;
252 tp->t_toe = toep;
253 tp->t_flags |= TF_TOE;
254
255 /* Install an extra hold on inp */
256 toep->inp = inp;
257 toep->flags |= TPF_ATTACHED;
258 in_pcbref(inp);
259 }
260
261 void
restore_so_proto(struct socket * so,bool v6)262 restore_so_proto(struct socket *so, bool v6)
263 {
264 if (v6)
265 so->so_proto = &tcp6_protosw;
266 else
267 so->so_proto = &tcp_protosw;
268 }
269
270 /* This is _not_ the normal way to "unoffload" a socket. */
271 void
undo_offload_socket(struct socket * so)272 undo_offload_socket(struct socket *so)
273 {
274 struct inpcb *inp = sotoinpcb(so);
275 struct tcpcb *tp = intotcpcb(inp);
276 struct toepcb *toep = tp->t_toe;
277 struct sockbuf *sb;
278
279 INP_WLOCK_ASSERT(inp);
280
281 sb = &so->so_snd;
282 SOCKBUF_LOCK(sb);
283 sb->sb_flags &= ~SB_NOCOALESCE;
284 SOCKBUF_UNLOCK(sb);
285 sb = &so->so_rcv;
286 SOCKBUF_LOCK(sb);
287 sb->sb_flags &= ~SB_NOCOALESCE;
288 restore_so_proto(so, inp->inp_vflag & INP_IPV6);
289 SOCKBUF_UNLOCK(sb);
290
291 tp->tod = NULL;
292 tp->t_toe = NULL;
293 tp->t_flags &= ~TF_TOE;
294
295 toep->inp = NULL;
296 toep->flags &= ~TPF_ATTACHED;
297 if (in_pcbrele_wlocked(inp))
298 panic("%s: inp freed.", __func__);
299 }
300
301 static void
release_offload_resources(struct toepcb * toep)302 release_offload_resources(struct toepcb *toep)
303 {
304 struct tom_data *td = toep->td;
305 struct adapter *sc = td_adapter(td);
306 int tid = toep->tid;
307
308 KASSERT(!(toep->flags & TPF_CPL_PENDING),
309 ("%s: %p has CPL pending.", __func__, toep));
310
311 CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
312 __func__, toep, tid, toep->l2te, toep->ce);
313
314 if (toep->l2te) {
315 t4_l2t_release(toep->l2te);
316 toep->l2te = NULL;
317 }
318 if (tid >= 0) {
319 remove_tid(sc, tid, toep->ce ? 2 : 1);
320 release_tid(sc, tid, toep->ctrlq);
321 toep->tid = -1;
322 mtx_lock(&td->toep_list_lock);
323 if (toep->flags & TPF_IN_TOEP_LIST) {
324 toep->flags &= ~TPF_IN_TOEP_LIST;
325 TAILQ_REMOVE(&td->toep_list, toep, link);
326 }
327 mtx_unlock(&td->toep_list_lock);
328 }
329 if (toep->ce) {
330 t4_release_clip_entry(sc, toep->ce);
331 toep->ce = NULL;
332 }
333 if (toep->params.tc_idx != -1)
334 t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->params.tc_idx);
335 }
336
337 /*
338 * Both the driver and kernel are done with the toepcb.
339 */
340 static void
done_with_toepcb(struct toepcb * toep)341 done_with_toepcb(struct toepcb *toep)
342 {
343 KASSERT(!(toep->flags & TPF_CPL_PENDING),
344 ("%s: %p has CPL pending.", __func__, toep));
345 KASSERT(!(toep->flags & TPF_ATTACHED),
346 ("%s: %p is still attached.", __func__, toep));
347
348 CTR(KTR_CXGBE, "%s: toep %p (0x%x)", __func__, toep, toep->flags);
349
350 /*
351 * These queues should have been emptied at approximately the same time
352 * that a normal connection's socket's so_snd would have been purged or
353 * drained. Do _not_ clean up here.
354 */
355 MPASS(mbufq_empty(&toep->ulp_pduq));
356 MPASS(mbufq_empty(&toep->ulp_pdu_reclaimq));
357 #ifdef INVARIANTS
358 if (ulp_mode(toep) == ULP_MODE_TCPDDP)
359 ddp_assert_empty(toep);
360 #endif
361 MPASS(TAILQ_EMPTY(&toep->aiotx_jobq));
362 MPASS(toep->tid == -1);
363 MPASS(toep->l2te == NULL);
364 MPASS(toep->ce == NULL);
365 MPASS((toep->flags & TPF_IN_TOEP_LIST) == 0);
366
367 free_toepcb(toep);
368 }
369
370 /*
371 * The kernel is done with the TCP PCB and this is our opportunity to unhook the
372 * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no
373 * pending CPL) then it is time to release all resources tied to the toepcb.
374 *
375 * Also gets called when an offloaded active open fails and the TOM wants the
376 * kernel to take the TCP PCB back.
377 */
378 void
t4_pcb_detach(struct toedev * tod __unused,struct tcpcb * tp)379 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
380 {
381 #if defined(KTR) || defined(INVARIANTS)
382 struct inpcb *inp = tptoinpcb(tp);
383 #endif
384 struct toepcb *toep = tp->t_toe;
385
386 INP_WLOCK_ASSERT(inp);
387
388 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
389 KASSERT(toep->flags & TPF_ATTACHED,
390 ("%s: not attached", __func__));
391
392 #ifdef KTR
393 if (tp->t_state == TCPS_SYN_SENT) {
394 CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
395 __func__, toep->tid, toep, toep->flags, inp,
396 inp->inp_flags);
397 } else {
398 CTR6(KTR_CXGBE,
399 "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
400 toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
401 inp->inp_flags);
402 }
403 #endif
404
405 tp->tod = NULL;
406 tp->t_toe = NULL;
407 tp->t_flags &= ~TF_TOE;
408 toep->flags &= ~TPF_ATTACHED;
409
410 if (!(toep->flags & TPF_CPL_PENDING))
411 done_with_toepcb(toep);
412 }
413
414 /*
415 * setsockopt handler.
416 */
417 static void
t4_ctloutput(struct toedev * tod,struct tcpcb * tp,int dir,int name)418 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
419 {
420 struct adapter *sc = tod->tod_softc;
421 struct toepcb *toep = tp->t_toe;
422
423 if (dir == SOPT_GET)
424 return;
425
426 CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
427
428 switch (name) {
429 case TCP_NODELAY:
430 if (tp->t_state != TCPS_ESTABLISHED)
431 break;
432 toep->params.nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
433 t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
434 V_TF_NAGLE(1), V_TF_NAGLE(toep->params.nagle), 0, 0);
435 break;
436 default:
437 break;
438 }
439 }
440
441 static inline uint64_t
get_tcb_tflags(const uint64_t * tcb)442 get_tcb_tflags(const uint64_t *tcb)
443 {
444
445 return ((be64toh(tcb[14]) << 32) | (be64toh(tcb[15]) >> 32));
446 }
447
448 static inline uint32_t
get_tcb_field(const uint64_t * tcb,u_int word,uint32_t mask,u_int shift)449 get_tcb_field(const uint64_t *tcb, u_int word, uint32_t mask, u_int shift)
450 {
451 #define LAST_WORD ((TCB_SIZE / 4) - 1)
452 uint64_t t1, t2;
453 int flit_idx;
454
455 MPASS(mask != 0);
456 MPASS(word <= LAST_WORD);
457 MPASS(shift < 32);
458
459 flit_idx = (LAST_WORD - word) / 2;
460 if (word & 0x1)
461 shift += 32;
462 t1 = be64toh(tcb[flit_idx]) >> shift;
463 t2 = 0;
464 if (fls(mask) > 64 - shift) {
465 /*
466 * Will spill over into the next logical flit, which is the flit
467 * before this one. The flit_idx before this one must be valid.
468 */
469 MPASS(flit_idx > 0);
470 t2 = be64toh(tcb[flit_idx - 1]) << (64 - shift);
471 }
472 return ((t2 | t1) & mask);
473 #undef LAST_WORD
474 }
475 #define GET_TCB_FIELD(tcb, F) \
476 get_tcb_field(tcb, W_TCB_##F, M_TCB_##F, S_TCB_##F)
477
478 /*
479 * Issues a CPL_GET_TCB to read the entire TCB for the tid.
480 */
481 static int
send_get_tcb(struct adapter * sc,u_int tid)482 send_get_tcb(struct adapter *sc, u_int tid)
483 {
484 struct cpl_get_tcb *cpl;
485 struct wrq_cookie cookie;
486
487 MPASS(tid >= sc->tids.tid_base);
488 MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
489
490 cpl = start_wrq_wr(&sc->sge.ctrlq[0], howmany(sizeof(*cpl), 16),
491 &cookie);
492 if (__predict_false(cpl == NULL))
493 return (ENOMEM);
494 bzero(cpl, sizeof(*cpl));
495 INIT_TP_WR(cpl, tid);
496 OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_GET_TCB, tid));
497 if (chip_id(sc) >= CHELSIO_T7) {
498 cpl->reply_ctrl =
499 htobe16(V_T7_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id) |
500 V_T7_REPLY_CHAN(0) | V_NO_REPLY(0));
501 } else {
502 cpl->reply_ctrl =
503 htobe16(V_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id) |
504 V_REPLY_CHAN(0) | V_NO_REPLY(0));
505 }
506 cpl->cookie = 0xff;
507 commit_wrq_wr(&sc->sge.ctrlq[0], cpl, &cookie);
508
509 return (0);
510 }
511
512 static struct tcb_histent *
alloc_tcb_histent(struct adapter * sc,u_int tid,int flags)513 alloc_tcb_histent(struct adapter *sc, u_int tid, int flags)
514 {
515 struct tcb_histent *te;
516
517 MPASS(flags == M_NOWAIT || flags == M_WAITOK);
518
519 te = malloc(sizeof(*te), M_CXGBE, M_ZERO | flags);
520 if (te == NULL)
521 return (NULL);
522 mtx_init(&te->te_lock, "TCB entry", NULL, MTX_DEF);
523 callout_init_mtx(&te->te_callout, &te->te_lock, 0);
524 te->te_adapter = sc;
525 te->te_tid = tid;
526
527 return (te);
528 }
529
530 static void
free_tcb_histent(struct tcb_histent * te)531 free_tcb_histent(struct tcb_histent *te)
532 {
533
534 mtx_destroy(&te->te_lock);
535 free(te, M_CXGBE);
536 }
537
538 /*
539 * Start tracking the tid in the TCB history.
540 */
541 int
add_tid_to_history(struct adapter * sc,u_int tid)542 add_tid_to_history(struct adapter *sc, u_int tid)
543 {
544 struct tcb_histent *te = NULL;
545 struct tom_data *td = sc->tom_softc;
546 int rc;
547
548 MPASS(tid >= sc->tids.tid_base);
549 MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
550
551 if (td->tcb_history == NULL)
552 return (ENXIO);
553
554 rw_wlock(&td->tcb_history_lock);
555 if (td->tcb_history[tid] != NULL) {
556 rc = EEXIST;
557 goto done;
558 }
559 te = alloc_tcb_histent(sc, tid, M_NOWAIT);
560 if (te == NULL) {
561 rc = ENOMEM;
562 goto done;
563 }
564 mtx_lock(&te->te_lock);
565 rc = send_get_tcb(sc, tid);
566 if (rc == 0) {
567 te->te_flags |= TE_RPL_PENDING;
568 td->tcb_history[tid] = te;
569 } else {
570 free(te, M_CXGBE);
571 }
572 mtx_unlock(&te->te_lock);
573 done:
574 rw_wunlock(&td->tcb_history_lock);
575 return (rc);
576 }
577
578 static void
remove_tcb_histent(struct tcb_histent * te)579 remove_tcb_histent(struct tcb_histent *te)
580 {
581 struct adapter *sc = te->te_adapter;
582 struct tom_data *td = sc->tom_softc;
583
584 rw_assert(&td->tcb_history_lock, RA_WLOCKED);
585 mtx_assert(&te->te_lock, MA_OWNED);
586 MPASS(td->tcb_history[te->te_tid] == te);
587
588 td->tcb_history[te->te_tid] = NULL;
589 free_tcb_histent(te);
590 rw_wunlock(&td->tcb_history_lock);
591 }
592
593 static inline struct tcb_histent *
lookup_tcb_histent(struct adapter * sc,u_int tid,bool addrem)594 lookup_tcb_histent(struct adapter *sc, u_int tid, bool addrem)
595 {
596 struct tcb_histent *te;
597 struct tom_data *td = sc->tom_softc;
598
599 MPASS(tid >= sc->tids.tid_base);
600 MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
601
602 if (td->tcb_history == NULL)
603 return (NULL);
604
605 if (addrem)
606 rw_wlock(&td->tcb_history_lock);
607 else
608 rw_rlock(&td->tcb_history_lock);
609 te = td->tcb_history[tid];
610 if (te != NULL) {
611 mtx_lock(&te->te_lock);
612 return (te); /* with both locks held */
613 }
614 if (addrem)
615 rw_wunlock(&td->tcb_history_lock);
616 else
617 rw_runlock(&td->tcb_history_lock);
618
619 return (te);
620 }
621
622 static inline void
release_tcb_histent(struct tcb_histent * te)623 release_tcb_histent(struct tcb_histent *te)
624 {
625 struct adapter *sc = te->te_adapter;
626 struct tom_data *td = sc->tom_softc;
627
628 mtx_assert(&te->te_lock, MA_OWNED);
629 mtx_unlock(&te->te_lock);
630 rw_assert(&td->tcb_history_lock, RA_RLOCKED);
631 rw_runlock(&td->tcb_history_lock);
632 }
633
634 static void
request_tcb(void * arg)635 request_tcb(void *arg)
636 {
637 struct tcb_histent *te = arg;
638
639 mtx_assert(&te->te_lock, MA_OWNED);
640
641 /* Noone else is supposed to update the histent. */
642 MPASS(!(te->te_flags & TE_RPL_PENDING));
643 if (send_get_tcb(te->te_adapter, te->te_tid) == 0)
644 te->te_flags |= TE_RPL_PENDING;
645 else
646 callout_schedule(&te->te_callout, hz / 100);
647 }
648
649 static void
update_tcb_histent(struct tcb_histent * te,const uint64_t * tcb)650 update_tcb_histent(struct tcb_histent *te, const uint64_t *tcb)
651 {
652 struct tom_data *td = te->te_adapter->tom_softc;
653 uint64_t tflags = get_tcb_tflags(tcb);
654 uint8_t sample = 0;
655
656 if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != GET_TCB_FIELD(tcb, SND_UNA_RAW)) {
657 if (GET_TCB_FIELD(tcb, T_RXTSHIFT) != 0)
658 sample |= TS_RTO;
659 if (GET_TCB_FIELD(tcb, T_DUPACKS) != 0)
660 sample |= TS_DUPACKS;
661 if (GET_TCB_FIELD(tcb, T_DUPACKS) >= td->dupack_threshold)
662 sample |= TS_FASTREXMT;
663 }
664
665 if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != 0) {
666 uint32_t snd_wnd;
667
668 sample |= TS_SND_BACKLOGGED; /* for whatever reason. */
669
670 snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
671 if (tflags & V_TF_RECV_SCALE(1))
672 snd_wnd <<= GET_TCB_FIELD(tcb, RCV_SCALE);
673 if (GET_TCB_FIELD(tcb, SND_CWND) < snd_wnd)
674 sample |= TS_CWND_LIMITED; /* maybe due to CWND */
675 }
676
677 if (tflags & V_TF_CCTRL_ECN(1)) {
678
679 /*
680 * CE marker on incoming IP hdr, echoing ECE back in the TCP
681 * hdr. Indicates congestion somewhere on the way from the peer
682 * to this node.
683 */
684 if (tflags & V_TF_CCTRL_ECE(1))
685 sample |= TS_ECN_ECE;
686
687 /*
688 * ECE seen and CWR sent (or about to be sent). Might indicate
689 * congestion on the way to the peer. This node is reducing its
690 * congestion window in response.
691 */
692 if (tflags & (V_TF_CCTRL_CWR(1) | V_TF_CCTRL_RFR(1)))
693 sample |= TS_ECN_CWR;
694 }
695
696 te->te_sample[te->te_pidx] = sample;
697 if (++te->te_pidx == nitems(te->te_sample))
698 te->te_pidx = 0;
699 memcpy(te->te_tcb, tcb, TCB_SIZE);
700 te->te_flags |= TE_ACTIVE;
701 }
702
703 static int
do_get_tcb_rpl(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)704 do_get_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
705 {
706 struct adapter *sc = iq->adapter;
707 const struct cpl_get_tcb_rpl *cpl = mtod(m, const void *);
708 const uint64_t *tcb = (const uint64_t *)(const void *)(cpl + 1);
709 struct tcb_histent *te;
710 const u_int tid = GET_TID(cpl);
711 bool remove;
712
713 remove = GET_TCB_FIELD(tcb, T_STATE) == TCPS_CLOSED;
714 te = lookup_tcb_histent(sc, tid, remove);
715 if (te == NULL) {
716 /* Not in the history. Who issued the GET_TCB for this? */
717 device_printf(sc->dev, "tcb %u: flags 0x%016jx, state %u, "
718 "srtt %u, sscale %u, rscale %u, cookie 0x%x\n", tid,
719 (uintmax_t)get_tcb_tflags(tcb), GET_TCB_FIELD(tcb, T_STATE),
720 GET_TCB_FIELD(tcb, T_SRTT), GET_TCB_FIELD(tcb, SND_SCALE),
721 GET_TCB_FIELD(tcb, RCV_SCALE), cpl->cookie);
722 goto done;
723 }
724
725 MPASS(te->te_flags & TE_RPL_PENDING);
726 te->te_flags &= ~TE_RPL_PENDING;
727 if (remove) {
728 remove_tcb_histent(te);
729 } else {
730 update_tcb_histent(te, tcb);
731 callout_reset(&te->te_callout, hz / 10, request_tcb, te);
732 release_tcb_histent(te);
733 }
734 done:
735 m_freem(m);
736 return (0);
737 }
738
739 static void
fill_tcp_info_from_tcb(struct adapter * sc,uint64_t * tcb,struct tcp_info * ti)740 fill_tcp_info_from_tcb(struct adapter *sc, uint64_t *tcb, struct tcp_info *ti)
741 {
742 uint32_t v;
743
744 ti->tcpi_state = GET_TCB_FIELD(tcb, T_STATE);
745
746 v = GET_TCB_FIELD(tcb, T_SRTT);
747 ti->tcpi_rtt = tcp_ticks_to_us(sc, v);
748
749 v = GET_TCB_FIELD(tcb, T_RTTVAR);
750 ti->tcpi_rttvar = tcp_ticks_to_us(sc, v);
751
752 ti->tcpi_snd_ssthresh = GET_TCB_FIELD(tcb, SND_SSTHRESH);
753 ti->tcpi_snd_cwnd = GET_TCB_FIELD(tcb, SND_CWND);
754 ti->tcpi_rcv_nxt = GET_TCB_FIELD(tcb, RCV_NXT);
755 ti->tcpi_rcv_adv = GET_TCB_FIELD(tcb, RCV_ADV);
756 ti->tcpi_dupacks = GET_TCB_FIELD(tcb, T_DUPACKS);
757
758 v = GET_TCB_FIELD(tcb, TX_MAX);
759 ti->tcpi_snd_nxt = v - GET_TCB_FIELD(tcb, SND_NXT_RAW);
760 ti->tcpi_snd_una = v - GET_TCB_FIELD(tcb, SND_UNA_RAW);
761 ti->tcpi_snd_max = v - GET_TCB_FIELD(tcb, SND_MAX_RAW);
762
763 /* Receive window being advertised by us. */
764 ti->tcpi_rcv_wscale = GET_TCB_FIELD(tcb, SND_SCALE); /* Yes, SND. */
765 ti->tcpi_rcv_space = GET_TCB_FIELD(tcb, RCV_WND);
766
767 /* Send window */
768 ti->tcpi_snd_wscale = GET_TCB_FIELD(tcb, RCV_SCALE); /* Yes, RCV. */
769 ti->tcpi_snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
770 if (get_tcb_tflags(tcb) & V_TF_RECV_SCALE(1))
771 ti->tcpi_snd_wnd <<= ti->tcpi_snd_wscale;
772 else
773 ti->tcpi_snd_wscale = 0;
774
775 }
776
777 static void
fill_tcp_info_from_history(struct adapter * sc,struct tcb_histent * te,struct tcp_info * ti)778 fill_tcp_info_from_history(struct adapter *sc, struct tcb_histent *te,
779 struct tcp_info *ti)
780 {
781
782 fill_tcp_info_from_tcb(sc, te->te_tcb, ti);
783 }
784
785 /*
786 * Reads the TCB for the given tid using a memory window and copies it to 'buf'
787 * in the same format as CPL_GET_TCB_RPL.
788 */
789 static void
read_tcb_using_memwin(struct adapter * sc,u_int tid,uint64_t * buf)790 read_tcb_using_memwin(struct adapter *sc, u_int tid, uint64_t *buf)
791 {
792 int i, j, k, rc;
793 uint32_t addr;
794 u_char *tcb, tmp;
795
796 MPASS(tid >= sc->tids.tid_base);
797 MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
798
799 addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + tid * TCB_SIZE;
800 rc = read_via_memwin(sc, 2, addr, (uint32_t *)buf, TCB_SIZE);
801 if (rc != 0)
802 return;
803
804 tcb = (u_char *)buf;
805 for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) {
806 for (k = 0; k < 16; k++) {
807 tmp = tcb[i + k];
808 tcb[i + k] = tcb[j + k];
809 tcb[j + k] = tmp;
810 }
811 }
812 }
813
814 static void
fill_tcp_info(struct adapter * sc,u_int tid,struct tcp_info * ti)815 fill_tcp_info(struct adapter *sc, u_int tid, struct tcp_info *ti)
816 {
817 uint64_t tcb[TCB_SIZE / sizeof(uint64_t)];
818 struct tcb_histent *te;
819
820 ti->tcpi_toe_tid = tid;
821 te = lookup_tcb_histent(sc, tid, false);
822 if (te != NULL) {
823 fill_tcp_info_from_history(sc, te, ti);
824 release_tcb_histent(te);
825 } else {
826 if (!(sc->debug_flags & DF_DISABLE_TCB_CACHE)) {
827 /* XXX: tell firmware to flush TCB cache. */
828 }
829 read_tcb_using_memwin(sc, tid, tcb);
830 fill_tcp_info_from_tcb(sc, tcb, ti);
831 }
832 }
833
834 /*
835 * Called by the kernel to allow the TOE driver to "refine" values filled up in
836 * the tcp_info for an offloaded connection.
837 */
838 static void
t4_tcp_info(struct toedev * tod,const struct tcpcb * tp,struct tcp_info * ti)839 t4_tcp_info(struct toedev *tod, const struct tcpcb *tp, struct tcp_info *ti)
840 {
841 struct adapter *sc = tod->tod_softc;
842 struct toepcb *toep = tp->t_toe;
843
844 INP_LOCK_ASSERT(tptoinpcb(tp));
845 MPASS(ti != NULL);
846
847 fill_tcp_info(sc, toep->tid, ti);
848 }
849
850 #ifdef KERN_TLS
851 static int
t4_alloc_tls_session(struct toedev * tod,struct tcpcb * tp,struct ktls_session * tls,int direction)852 t4_alloc_tls_session(struct toedev *tod, struct tcpcb *tp,
853 struct ktls_session *tls, int direction)
854 {
855 struct toepcb *toep = tp->t_toe;
856
857 INP_WLOCK_ASSERT(tptoinpcb(tp));
858 MPASS(tls != NULL);
859
860 return (tls_alloc_ktls(toep, tls, direction));
861 }
862 #endif
863
864 static void
send_mss_flowc_wr(struct adapter * sc,struct toepcb * toep)865 send_mss_flowc_wr(struct adapter *sc, struct toepcb *toep)
866 {
867 struct wrq_cookie cookie;
868 struct fw_flowc_wr *flowc;
869 struct ofld_tx_sdesc *txsd;
870 const int flowclen = sizeof(*flowc) + sizeof(struct fw_flowc_mnemval);
871 const int flowclen16 = howmany(flowclen, 16);
872
873 if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0) {
874 CH_ERR(sc, "%s: tid %u out of tx credits (%d, %d).\n", __func__,
875 toep->tid, toep->tx_credits, toep->txsd_avail);
876 return;
877 }
878
879 flowc = start_wrq_wr(&toep->ofld_txq->wrq, flowclen16, &cookie);
880 if (__predict_false(flowc == NULL)) {
881 CH_ERR(sc, "ENOMEM in %s for tid %u.\n", __func__, toep->tid);
882 return;
883 }
884 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
885 V_FW_FLOWC_WR_NPARAMS(1));
886 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
887 V_FW_WR_FLOWID(toep->tid));
888 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_MSS;
889 flowc->mnemval[0].val = htobe32(toep->params.emss);
890
891 txsd = &toep->txsd[toep->txsd_pidx];
892 _Static_assert(flowclen16 <= MAX_OFLD_TX_SDESC_CREDITS,
893 "MAX_OFLD_TX_SDESC_CREDITS too small");
894 txsd->tx_credits = flowclen16;
895 txsd->plen = 0;
896 toep->tx_credits -= txsd->tx_credits;
897 if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
898 toep->txsd_pidx = 0;
899 toep->txsd_avail--;
900 commit_wrq_wr(&toep->ofld_txq->wrq, flowc, &cookie);
901 }
902
903 static void
t4_pmtu_update(struct toedev * tod,struct tcpcb * tp,tcp_seq seq,int mtu)904 t4_pmtu_update(struct toedev *tod, struct tcpcb *tp, tcp_seq seq, int mtu)
905 {
906 struct work_request_hdr *wrh;
907 struct ulp_txpkt *ulpmc;
908 int idx, len;
909 struct wrq_cookie cookie;
910 struct inpcb *inp = tptoinpcb(tp);
911 struct toepcb *toep = tp->t_toe;
912 struct adapter *sc = td_adapter(toep->td);
913 unsigned short *mtus = &sc->params.mtus[0];
914
915 INP_WLOCK_ASSERT(inp);
916 MPASS(mtu > 0); /* kernel is supposed to provide something usable. */
917
918 /* tp->snd_una and snd_max are in host byte order too. */
919 seq = be32toh(seq);
920
921 CTR6(KTR_CXGBE, "%s: tid %d, seq 0x%08x, mtu %u, mtu_idx %u (%d)",
922 __func__, toep->tid, seq, mtu, toep->params.mtu_idx,
923 mtus[toep->params.mtu_idx]);
924
925 if (ulp_mode(toep) == ULP_MODE_NONE && /* XXX: Read TCB otherwise? */
926 (SEQ_LT(seq, tp->snd_una) || SEQ_GEQ(seq, tp->snd_max))) {
927 CTR5(KTR_CXGBE,
928 "%s: tid %d, seq 0x%08x not in range [0x%08x, 0x%08x).",
929 __func__, toep->tid, seq, tp->snd_una, tp->snd_max);
930 return;
931 }
932
933 /* Find the best mtu_idx for the suggested MTU. */
934 for (idx = 0; idx < NMTUS - 1 && mtus[idx + 1] <= mtu; idx++)
935 continue;
936 if (idx >= toep->params.mtu_idx)
937 return; /* Never increase the PMTU (just like the kernel). */
938
939 /*
940 * We'll send a compound work request with 2 SET_TCB_FIELDs -- the first
941 * one updates the mtu_idx and the second one triggers a retransmit.
942 */
943 len = sizeof(*wrh) + 2 * roundup2(LEN__SET_TCB_FIELD_ULP, 16);
944 wrh = start_wrq_wr(toep->ctrlq, howmany(len, 16), &cookie);
945 if (wrh == NULL) {
946 CH_ERR(sc, "failed to change mtu_idx of tid %d (%u -> %u).\n",
947 toep->tid, toep->params.mtu_idx, idx);
948 return;
949 }
950 INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */
951 ulpmc = (struct ulp_txpkt *)(wrh + 1);
952 ulpmc = mk_set_tcb_field_ulp(sc, ulpmc, toep->tid, W_TCB_T_MAXSEG,
953 V_TCB_T_MAXSEG(M_TCB_T_MAXSEG), V_TCB_T_MAXSEG(idx));
954 ulpmc = mk_set_tcb_field_ulp(sc, ulpmc, toep->tid, W_TCB_TIMESTAMP,
955 V_TCB_TIMESTAMP(0x7FFFFULL << 11), 0);
956 commit_wrq_wr(toep->ctrlq, wrh, &cookie);
957
958 /* Update the software toepcb and tcpcb. */
959 toep->params.mtu_idx = idx;
960 tp->t_maxseg = mtus[toep->params.mtu_idx];
961 if (inp->inp_inc.inc_flags & INC_ISIPV6)
962 tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
963 else
964 tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
965 toep->params.emss = tp->t_maxseg;
966 if (tp->t_flags & TF_RCVD_TSTMP)
967 toep->params.emss -= TCPOLEN_TSTAMP_APPA;
968
969 /* Update the firmware flowc. */
970 send_mss_flowc_wr(sc, toep);
971
972 /* Update the MTU in the kernel's hostcache. */
973 if (sc->tt.update_hc_on_pmtu_change != 0) {
974 struct in_conninfo inc = {0};
975
976 inc.inc_fibnum = inp->inp_inc.inc_fibnum;
977 if (inp->inp_inc.inc_flags & INC_ISIPV6) {
978 inc.inc_flags |= INC_ISIPV6;
979 inc.inc6_faddr = inp->inp_inc.inc6_faddr;
980 } else {
981 inc.inc_faddr = inp->inp_inc.inc_faddr;
982 }
983 tcp_hc_updatemtu(&inc, mtu);
984 }
985
986 CTR6(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u), t_maxseg %u, emss %u",
987 __func__, toep->tid, toep->params.mtu_idx,
988 mtus[toep->params.mtu_idx], tp->t_maxseg, toep->params.emss);
989 }
990
991 /*
992 * The TOE driver will not receive any more CPLs for the tid associated with the
993 * toepcb; release the hold on the inpcb.
994 */
995 void
final_cpl_received(struct toepcb * toep)996 final_cpl_received(struct toepcb *toep)
997 {
998 struct inpcb *inp = toep->inp;
999 bool need_wakeup;
1000
1001 KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
1002 INP_WLOCK_ASSERT(inp);
1003 KASSERT(toep->flags & TPF_CPL_PENDING,
1004 ("%s: CPL not pending already?", __func__));
1005
1006 CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
1007 __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
1008
1009 if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1010 release_ddp_resources(toep);
1011 toep->inp = NULL;
1012 need_wakeup = (toep->flags & TPF_WAITING_FOR_FINAL) != 0;
1013 toep->flags &= ~(TPF_CPL_PENDING | TPF_WAITING_FOR_FINAL);
1014 mbufq_drain(&toep->ulp_pduq);
1015 mbufq_drain(&toep->ulp_pdu_reclaimq);
1016 release_offload_resources(toep);
1017 if (!(toep->flags & TPF_ATTACHED))
1018 done_with_toepcb(toep);
1019
1020 if (!in_pcbrele_wlocked(inp))
1021 INP_WUNLOCK(inp);
1022
1023 if (need_wakeup) {
1024 struct mtx *lock = mtx_pool_find(mtxpool_sleep, toep);
1025
1026 mtx_lock(lock);
1027 wakeup(toep);
1028 mtx_unlock(lock);
1029 }
1030 }
1031
1032 void
insert_tid(struct adapter * sc,int tid,void * ctx,int ntids)1033 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
1034 {
1035 struct tid_info *t = &sc->tids;
1036
1037 MPASS(tid >= t->tid_base);
1038 MPASS(tid - t->tid_base < t->ntids);
1039
1040 t->tid_tab[tid - t->tid_base] = ctx;
1041 atomic_add_int(&t->tids_in_use, ntids);
1042 }
1043
1044 void *
lookup_tid(struct adapter * sc,int tid)1045 lookup_tid(struct adapter *sc, int tid)
1046 {
1047 struct tid_info *t = &sc->tids;
1048
1049 return (t->tid_tab[tid - t->tid_base]);
1050 }
1051
1052 void
update_tid(struct adapter * sc,int tid,void * ctx)1053 update_tid(struct adapter *sc, int tid, void *ctx)
1054 {
1055 struct tid_info *t = &sc->tids;
1056
1057 t->tid_tab[tid - t->tid_base] = ctx;
1058 }
1059
1060 void
remove_tid(struct adapter * sc,int tid,int ntids)1061 remove_tid(struct adapter *sc, int tid, int ntids)
1062 {
1063 struct tid_info *t = &sc->tids;
1064
1065 t->tid_tab[tid - t->tid_base] = NULL;
1066 atomic_subtract_int(&t->tids_in_use, ntids);
1067 }
1068
1069 /*
1070 * What mtu_idx to use, given a 4-tuple. Note that both s->mss and tcp_mssopt
1071 * have the MSS that we should advertise in our SYN. Advertised MSS doesn't
1072 * account for any TCP options so the effective MSS (only payload, no headers or
1073 * options) could be different.
1074 */
1075 static int
find_best_mtu_idx(struct adapter * sc,struct in_conninfo * inc,struct offload_settings * s)1076 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc,
1077 struct offload_settings *s)
1078 {
1079 unsigned short *mtus = &sc->params.mtus[0];
1080 int i, mss, mtu;
1081
1082 MPASS(inc != NULL);
1083
1084 mss = s->mss > 0 ? s->mss : tcp_mssopt(inc);
1085 if (inc->inc_flags & INC_ISIPV6)
1086 mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1087 else
1088 mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr);
1089
1090 for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++)
1091 continue;
1092
1093 return (i);
1094 }
1095
1096 /*
1097 * Determine the receive window size for a socket.
1098 */
1099 u_long
select_rcv_wnd(struct socket * so)1100 select_rcv_wnd(struct socket *so)
1101 {
1102 unsigned long wnd;
1103
1104 SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1105
1106 wnd = sbspace(&so->so_rcv);
1107 if (wnd < MIN_RCV_WND)
1108 wnd = MIN_RCV_WND;
1109
1110 return min(wnd, MAX_RCV_WND);
1111 }
1112
1113 int
select_rcv_wscale(void)1114 select_rcv_wscale(void)
1115 {
1116 int wscale = 0;
1117 unsigned long space = sb_max;
1118
1119 if (space > MAX_RCV_WND)
1120 space = MAX_RCV_WND;
1121
1122 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
1123 wscale++;
1124
1125 return (wscale);
1126 }
1127
1128 __be64
calc_options0(struct vi_info * vi,struct conn_params * cp)1129 calc_options0(struct vi_info *vi, struct conn_params *cp)
1130 {
1131 uint64_t opt0 = 0;
1132
1133 opt0 |= F_TCAM_BYPASS;
1134
1135 MPASS(cp->wscale >= 0 && cp->wscale <= M_WND_SCALE);
1136 opt0 |= V_WND_SCALE(cp->wscale);
1137
1138 MPASS(cp->mtu_idx >= 0 && cp->mtu_idx < NMTUS);
1139 opt0 |= V_MSS_IDX(cp->mtu_idx);
1140
1141 MPASS(cp->ulp_mode >= 0 && cp->ulp_mode <= M_ULP_MODE);
1142 opt0 |= V_ULP_MODE(cp->ulp_mode);
1143
1144 MPASS(cp->opt0_bufsize >= 0 && cp->opt0_bufsize <= M_RCV_BUFSIZ);
1145 opt0 |= V_RCV_BUFSIZ(cp->opt0_bufsize);
1146
1147 MPASS(cp->l2t_idx >= 0 && cp->l2t_idx < vi->adapter->vres.l2t.size);
1148 opt0 |= V_L2T_IDX(cp->l2t_idx);
1149
1150 opt0 |= V_SMAC_SEL(vi->smt_idx);
1151 opt0 |= V_TX_CHAN(vi->pi->tx_chan);
1152
1153 MPASS(cp->keepalive == 0 || cp->keepalive == 1);
1154 opt0 |= V_KEEP_ALIVE(cp->keepalive);
1155
1156 MPASS(cp->nagle == 0 || cp->nagle == 1);
1157 opt0 |= V_NAGLE(cp->nagle);
1158
1159 return (htobe64(opt0));
1160 }
1161
1162 __be32
calc_options2(struct vi_info * vi,struct conn_params * cp)1163 calc_options2(struct vi_info *vi, struct conn_params *cp)
1164 {
1165 uint32_t opt2 = 0;
1166 struct port_info *pi = vi->pi;
1167 struct adapter *sc = pi->adapter;
1168
1169 /*
1170 * rx flow control, rx coalesce, congestion control, and tx pace are all
1171 * explicitly set by the driver. On T5+ the ISS is also set by the
1172 * driver to the value picked by the kernel.
1173 */
1174 if (is_t4(sc)) {
1175 opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID;
1176 opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID;
1177 } else {
1178 opt2 |= F_T5_OPT_2_VALID; /* all 4 valid */
1179 opt2 |= F_T5_ISS; /* ISS provided in CPL */
1180 }
1181
1182 MPASS(cp->sack == 0 || cp->sack == 1);
1183 opt2 |= V_SACK_EN(cp->sack);
1184
1185 MPASS(cp->tstamp == 0 || cp->tstamp == 1);
1186 opt2 |= V_TSTAMPS_EN(cp->tstamp);
1187
1188 if (cp->wscale > 0)
1189 opt2 |= F_WND_SCALE_EN;
1190
1191 MPASS(cp->ecn == 0 || cp->ecn == 1);
1192 opt2 |= V_CCTRL_ECN(cp->ecn);
1193
1194 opt2 |= V_TX_QUEUE(TX_MODQ(pi->tx_chan));
1195 opt2 |= V_PACE(0);
1196 opt2 |= F_RSS_QUEUE_VALID;
1197 opt2 |= V_RSS_QUEUE(sc->sge.ofld_rxq[cp->rxq_idx].iq.abs_id);
1198 if (chip_id(sc) <= CHELSIO_T6) {
1199 MPASS(pi->rx_chan == 0 || pi->rx_chan == 1);
1200 opt2 |= V_RX_CHANNEL(pi->rx_chan);
1201 }
1202
1203 MPASS(cp->cong_algo >= 0 && cp->cong_algo <= M_CONG_CNTRL);
1204 opt2 |= V_CONG_CNTRL(cp->cong_algo);
1205
1206 MPASS(cp->rx_coalesce == 0 || cp->rx_coalesce == 1);
1207 if (cp->rx_coalesce == 1)
1208 opt2 |= V_RX_COALESCE(M_RX_COALESCE);
1209
1210 opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0);
1211 MPASS(cp->ulp_mode != ULP_MODE_TCPDDP);
1212
1213 return (htobe32(opt2));
1214 }
1215
1216 uint64_t
select_ntuple(struct vi_info * vi,struct l2t_entry * e)1217 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
1218 {
1219 struct adapter *sc = vi->adapter;
1220 struct tp_params *tp = &sc->params.tp;
1221 uint64_t ntuple = 0;
1222
1223 /*
1224 * Initialize each of the fields which we care about which are present
1225 * in the Compressed Filter Tuple.
1226 */
1227 if (tp->vlan_shift >= 0 && EVL_VLANOFTAG(e->vlan) != CPL_L2T_VLAN_NONE)
1228 ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
1229
1230 if (tp->port_shift >= 0)
1231 ntuple |= (uint64_t)e->hw_port << tp->port_shift;
1232
1233 if (tp->protocol_shift >= 0)
1234 ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
1235
1236 if (tp->vnic_shift >= 0 && tp->vnic_mode == FW_VNIC_MODE_PF_VF) {
1237 ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vi->vin) |
1238 V_FT_VNID_ID_PF(sc->pf) | V_FT_VNID_ID_VLD(vi->vfvld)) <<
1239 tp->vnic_shift;
1240 }
1241
1242 return (ntuple);
1243 }
1244
1245 /*
1246 * Initialize various connection parameters.
1247 */
1248 void
init_conn_params(struct vi_info * vi,struct offload_settings * s,struct in_conninfo * inc,struct socket * so,const struct tcp_options * tcpopt,int16_t l2t_idx,struct conn_params * cp)1249 init_conn_params(struct vi_info *vi , struct offload_settings *s,
1250 struct in_conninfo *inc, struct socket *so,
1251 const struct tcp_options *tcpopt, int16_t l2t_idx, struct conn_params *cp)
1252 {
1253 struct port_info *pi = vi->pi;
1254 struct adapter *sc = pi->adapter;
1255 struct tom_tunables *tt = &sc->tt;
1256 struct inpcb *inp = sotoinpcb(so);
1257 struct tcpcb *tp = intotcpcb(inp);
1258 u_long wnd;
1259 u_int q_idx;
1260
1261 MPASS(s->offload != 0);
1262
1263 /* Congestion control algorithm */
1264 if (s->cong_algo >= 0)
1265 cp->cong_algo = s->cong_algo & M_CONG_CNTRL;
1266 else if (sc->tt.cong_algorithm >= 0)
1267 cp->cong_algo = tt->cong_algorithm & M_CONG_CNTRL;
1268 else {
1269 struct cc_algo *cc = CC_ALGO(tp);
1270
1271 if (strcasecmp(cc->name, "reno") == 0)
1272 cp->cong_algo = CONG_ALG_RENO;
1273 else if (strcasecmp(cc->name, "tahoe") == 0)
1274 cp->cong_algo = CONG_ALG_TAHOE;
1275 if (strcasecmp(cc->name, "newreno") == 0)
1276 cp->cong_algo = CONG_ALG_NEWRENO;
1277 if (strcasecmp(cc->name, "highspeed") == 0)
1278 cp->cong_algo = CONG_ALG_HIGHSPEED;
1279 else {
1280 /*
1281 * Use newreno in case the algorithm selected by the
1282 * host stack is not supported by the hardware.
1283 */
1284 cp->cong_algo = CONG_ALG_NEWRENO;
1285 }
1286 }
1287
1288 /* Tx traffic scheduling class. */
1289 if (s->sched_class >= 0 && s->sched_class < sc->params.nsched_cls)
1290 cp->tc_idx = s->sched_class;
1291 else
1292 cp->tc_idx = -1;
1293
1294 /* Nagle's algorithm. */
1295 if (s->nagle >= 0)
1296 cp->nagle = s->nagle > 0 ? 1 : 0;
1297 else
1298 cp->nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
1299
1300 /* TCP Keepalive. */
1301 if (V_tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE)
1302 cp->keepalive = 1;
1303 else
1304 cp->keepalive = 0;
1305
1306 /* Optimization that's specific to T5 @ 40G. */
1307 if (tt->tx_align >= 0)
1308 cp->tx_align = tt->tx_align > 0 ? 1 : 0;
1309 else if (chip_id(sc) == CHELSIO_T5 &&
1310 (port_top_speed(pi) > 10 || sc->params.nports > 2))
1311 cp->tx_align = 1;
1312 else
1313 cp->tx_align = 0;
1314
1315 /* ULP mode. */
1316 cp->ulp_mode = ULP_MODE_NONE;
1317
1318 /* Rx coalescing. */
1319 if (s->rx_coalesce >= 0)
1320 cp->rx_coalesce = s->rx_coalesce > 0 ? 1 : 0;
1321 else if (tt->rx_coalesce >= 0)
1322 cp->rx_coalesce = tt->rx_coalesce > 0 ? 1 : 0;
1323 else
1324 cp->rx_coalesce = 1; /* default */
1325
1326 /*
1327 * Index in the PMTU table. This controls the MSS that we announce in
1328 * our SYN initially, but after ESTABLISHED it controls the MSS that we
1329 * use to send data.
1330 */
1331 cp->mtu_idx = find_best_mtu_idx(sc, inc, s);
1332
1333 /* Control queue. */
1334 cp->ctrlq_idx = vi->pi->port_id;
1335
1336 /* Tx queue for this connection. */
1337 if (s->txq == QUEUE_RANDOM)
1338 q_idx = arc4random();
1339 else if (s->txq == QUEUE_ROUNDROBIN)
1340 q_idx = atomic_fetchadd_int(&vi->txq_rr, 1);
1341 else
1342 q_idx = s->txq;
1343 cp->txq_idx = vi->first_ofld_txq + q_idx % vi->nofldtxq;
1344
1345 /* Rx queue for this connection. */
1346 if (s->rxq == QUEUE_RANDOM)
1347 q_idx = arc4random();
1348 else if (s->rxq == QUEUE_ROUNDROBIN)
1349 q_idx = atomic_fetchadd_int(&vi->rxq_rr, 1);
1350 else
1351 q_idx = s->rxq;
1352 cp->rxq_idx = vi->first_ofld_rxq + q_idx % vi->nofldrxq;
1353
1354 if (SOLISTENING(so)) {
1355 /* Passive open */
1356 MPASS(tcpopt != NULL);
1357
1358 /* TCP timestamp option */
1359 if (tcpopt->tstamp &&
1360 (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323)))
1361 cp->tstamp = 1;
1362 else
1363 cp->tstamp = 0;
1364
1365 /* SACK */
1366 if (tcpopt->sack &&
1367 (s->sack > 0 || (s->sack < 0 && V_tcp_do_sack)))
1368 cp->sack = 1;
1369 else
1370 cp->sack = 0;
1371
1372 /* Receive window scaling. */
1373 if (tcpopt->wsf > 0 && tcpopt->wsf < 15 && V_tcp_do_rfc1323)
1374 cp->wscale = select_rcv_wscale();
1375 else
1376 cp->wscale = 0;
1377
1378 /* ECN */
1379 if (tcpopt->ecn && /* XXX: review. */
1380 (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn)))
1381 cp->ecn = 1;
1382 else
1383 cp->ecn = 0;
1384
1385 wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND);
1386 cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1387
1388 if (tt->sndbuf > 0)
1389 cp->sndbuf = tt->sndbuf;
1390 else if (so->sol_sbsnd_flags & SB_AUTOSIZE &&
1391 V_tcp_do_autosndbuf)
1392 cp->sndbuf = 256 * 1024;
1393 else
1394 cp->sndbuf = so->sol_sbsnd_hiwat;
1395 } else {
1396 /* Active open */
1397
1398 /* TCP timestamp option */
1399 if (s->tstamp > 0 ||
1400 (s->tstamp < 0 && (tp->t_flags & TF_REQ_TSTMP)))
1401 cp->tstamp = 1;
1402 else
1403 cp->tstamp = 0;
1404
1405 /* SACK */
1406 if (s->sack > 0 ||
1407 (s->sack < 0 && (tp->t_flags & TF_SACK_PERMIT)))
1408 cp->sack = 1;
1409 else
1410 cp->sack = 0;
1411
1412 /* Receive window scaling */
1413 if (tp->t_flags & TF_REQ_SCALE)
1414 cp->wscale = select_rcv_wscale();
1415 else
1416 cp->wscale = 0;
1417
1418 /* ECN */
1419 if (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn == 1))
1420 cp->ecn = 1;
1421 else
1422 cp->ecn = 0;
1423
1424 SOCKBUF_LOCK(&so->so_rcv);
1425 wnd = max(select_rcv_wnd(so), MIN_RCV_WND);
1426 SOCKBUF_UNLOCK(&so->so_rcv);
1427 cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1428
1429 if (tt->sndbuf > 0)
1430 cp->sndbuf = tt->sndbuf;
1431 else {
1432 SOCKBUF_LOCK(&so->so_snd);
1433 if (so->so_snd.sb_flags & SB_AUTOSIZE &&
1434 V_tcp_do_autosndbuf)
1435 cp->sndbuf = 256 * 1024;
1436 else
1437 cp->sndbuf = so->so_snd.sb_hiwat;
1438 SOCKBUF_UNLOCK(&so->so_snd);
1439 }
1440 }
1441
1442 cp->l2t_idx = l2t_idx;
1443
1444 /* This will be initialized on ESTABLISHED. */
1445 cp->emss = 0;
1446 }
1447
1448 void
update_tid_qid_sel(struct vi_info * vi,struct conn_params * cp,int tid)1449 update_tid_qid_sel(struct vi_info *vi, struct conn_params *cp, int tid)
1450 {
1451 struct adapter *sc = vi->adapter;
1452 const int mask = sc->params.tid_qid_sel_mask;
1453 struct sge_ofld_txq *ofld_txq = &sc->sge.ofld_txq[cp->txq_idx];
1454 uint32_t ngroup;
1455 int g, nqpg;
1456
1457 cp->ctrlq_idx = ofld_txq_group(tid, mask);
1458 CTR(KTR_CXGBE, "tid %u is on core %u", tid, cp->ctrlq_idx);
1459 if ((ofld_txq->wrq.eq.cntxt_id & mask) == (tid & mask))
1460 return;
1461
1462 ngroup = 1 << bitcount32(mask);
1463 MPASS(vi->nofldtxq % ngroup == 0);
1464 g = ofld_txq_group(tid, mask);
1465 nqpg = vi->nofldtxq / ngroup;
1466 cp->txq_idx = vi->first_ofld_txq + g * nqpg + arc4random() % nqpg;
1467 #ifdef INVARIANTS
1468 MPASS(cp->txq_idx < vi->first_ofld_txq + vi->nofldtxq);
1469 ofld_txq = &sc->sge.ofld_txq[cp->txq_idx];
1470 MPASS((ofld_txq->wrq.eq.cntxt_id & mask) == (tid & mask));
1471 #endif
1472 }
1473
1474 int
negative_advice(int status)1475 negative_advice(int status)
1476 {
1477
1478 return (status == CPL_ERR_RTX_NEG_ADVICE ||
1479 status == CPL_ERR_PERSIST_NEG_ADVICE ||
1480 status == CPL_ERR_KEEPALV_NEG_ADVICE);
1481 }
1482
1483 static int
alloc_tid_tab(struct adapter * sc)1484 alloc_tid_tab(struct adapter *sc)
1485 {
1486 struct tid_info *t = &sc->tids;
1487
1488 MPASS(t->ntids > 0);
1489 MPASS(t->tid_tab == NULL);
1490
1491 t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE,
1492 M_ZERO | M_NOWAIT);
1493 if (t->tid_tab == NULL)
1494 return (ENOMEM);
1495 atomic_store_rel_int(&t->tids_in_use, 0);
1496
1497 return (0);
1498 }
1499
1500 static void
free_tid_tab(struct adapter * sc)1501 free_tid_tab(struct adapter *sc)
1502 {
1503 struct tid_info *t = &sc->tids;
1504
1505 KASSERT(t->tids_in_use == 0,
1506 ("%s: %d tids still in use.", __func__, t->tids_in_use));
1507
1508 free(t->tid_tab, M_CXGBE);
1509 t->tid_tab = NULL;
1510 }
1511
1512 static void
free_tid_tabs(struct adapter * sc)1513 free_tid_tabs(struct adapter *sc)
1514 {
1515 free_tid_tab(sc);
1516 free_stid_tab(sc);
1517 }
1518
1519 static int
alloc_tid_tabs(struct adapter * sc)1520 alloc_tid_tabs(struct adapter *sc)
1521 {
1522 int rc;
1523
1524 rc = alloc_tid_tab(sc);
1525 if (rc != 0)
1526 goto failed;
1527
1528 rc = alloc_stid_tab(sc);
1529 if (rc != 0)
1530 goto failed;
1531
1532 return (0);
1533 failed:
1534 free_tid_tabs(sc);
1535 return (rc);
1536 }
1537
1538 static inline void
alloc_tcb_history(struct adapter * sc,struct tom_data * td)1539 alloc_tcb_history(struct adapter *sc, struct tom_data *td)
1540 {
1541
1542 if (sc->tids.ntids == 0 || sc->tids.ntids > 1024)
1543 return;
1544 rw_init(&td->tcb_history_lock, "TCB history");
1545 td->tcb_history = malloc(sc->tids.ntids * sizeof(*td->tcb_history),
1546 M_CXGBE, M_ZERO | M_NOWAIT);
1547 td->dupack_threshold = G_DUPACKTHRESH(t4_read_reg(sc, A_TP_PARA_REG0));
1548 }
1549
1550 static inline void
free_tcb_history(struct adapter * sc,struct tom_data * td)1551 free_tcb_history(struct adapter *sc, struct tom_data *td)
1552 {
1553 #ifdef INVARIANTS
1554 int i;
1555
1556 if (td->tcb_history != NULL) {
1557 for (i = 0; i < sc->tids.ntids; i++) {
1558 MPASS(td->tcb_history[i] == NULL);
1559 }
1560 }
1561 #endif
1562 free(td->tcb_history, M_CXGBE);
1563 if (rw_initialized(&td->tcb_history_lock))
1564 rw_destroy(&td->tcb_history_lock);
1565 }
1566
1567 static void
free_tom_data(struct adapter * sc,struct tom_data * td)1568 free_tom_data(struct adapter *sc, struct tom_data *td)
1569 {
1570
1571 ASSERT_SYNCHRONIZED_OP(sc);
1572
1573 KASSERT(TAILQ_EMPTY(&td->toep_list),
1574 ("%s: TOE PCB list is not empty.", __func__));
1575 KASSERT(td->lctx_count == 0,
1576 ("%s: lctx hash table is not empty.", __func__));
1577
1578 t4_free_ppod_region(&td->pr);
1579
1580 if (td->listen_mask != 0)
1581 hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
1582
1583 if (mtx_initialized(&td->unsent_wr_lock))
1584 mtx_destroy(&td->unsent_wr_lock);
1585 if (mtx_initialized(&td->lctx_hash_lock))
1586 mtx_destroy(&td->lctx_hash_lock);
1587 if (mtx_initialized(&td->toep_list_lock))
1588 mtx_destroy(&td->toep_list_lock);
1589
1590 free_tcb_history(sc, td);
1591 free_tid_tabs(sc);
1592 free(td, M_CXGBE);
1593 }
1594
1595 static char *
prepare_pkt(int open_type,uint16_t vtag,struct inpcb * inp,int * pktlen,int * buflen)1596 prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen,
1597 int *buflen)
1598 {
1599 char *pkt;
1600 struct tcphdr *th;
1601 int ipv6, len;
1602 const int maxlen =
1603 max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) +
1604 max(sizeof(struct ip), sizeof(struct ip6_hdr)) +
1605 sizeof(struct tcphdr);
1606
1607 MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN);
1608
1609 pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT);
1610 if (pkt == NULL)
1611 return (NULL);
1612
1613 ipv6 = inp->inp_vflag & INP_IPV6;
1614 len = 0;
1615
1616 if (EVL_VLANOFTAG(vtag) == 0xfff) {
1617 struct ether_header *eh = (void *)pkt;
1618
1619 if (ipv6)
1620 eh->ether_type = htons(ETHERTYPE_IPV6);
1621 else
1622 eh->ether_type = htons(ETHERTYPE_IP);
1623
1624 len += sizeof(*eh);
1625 } else {
1626 struct ether_vlan_header *evh = (void *)pkt;
1627
1628 evh->evl_encap_proto = htons(ETHERTYPE_VLAN);
1629 evh->evl_tag = htons(vtag);
1630 if (ipv6)
1631 evh->evl_proto = htons(ETHERTYPE_IPV6);
1632 else
1633 evh->evl_proto = htons(ETHERTYPE_IP);
1634
1635 len += sizeof(*evh);
1636 }
1637
1638 if (ipv6) {
1639 struct ip6_hdr *ip6 = (void *)&pkt[len];
1640
1641 ip6->ip6_vfc = IPV6_VERSION;
1642 ip6->ip6_plen = htons(sizeof(struct tcphdr));
1643 ip6->ip6_nxt = IPPROTO_TCP;
1644 if (open_type == OPEN_TYPE_ACTIVE) {
1645 ip6->ip6_src = inp->in6p_laddr;
1646 ip6->ip6_dst = inp->in6p_faddr;
1647 } else if (open_type == OPEN_TYPE_LISTEN) {
1648 ip6->ip6_src = inp->in6p_laddr;
1649 ip6->ip6_dst = ip6->ip6_src;
1650 }
1651
1652 len += sizeof(*ip6);
1653 } else {
1654 struct ip *ip = (void *)&pkt[len];
1655
1656 ip->ip_v = IPVERSION;
1657 ip->ip_hl = sizeof(*ip) >> 2;
1658 ip->ip_tos = inp->inp_ip_tos;
1659 ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr));
1660 ip->ip_ttl = inp->inp_ip_ttl;
1661 ip->ip_p = IPPROTO_TCP;
1662 if (open_type == OPEN_TYPE_ACTIVE) {
1663 ip->ip_src = inp->inp_laddr;
1664 ip->ip_dst = inp->inp_faddr;
1665 } else if (open_type == OPEN_TYPE_LISTEN) {
1666 ip->ip_src = inp->inp_laddr;
1667 ip->ip_dst = ip->ip_src;
1668 }
1669
1670 len += sizeof(*ip);
1671 }
1672
1673 th = (void *)&pkt[len];
1674 if (open_type == OPEN_TYPE_ACTIVE) {
1675 th->th_sport = inp->inp_lport; /* network byte order already */
1676 th->th_dport = inp->inp_fport; /* ditto */
1677 } else if (open_type == OPEN_TYPE_LISTEN) {
1678 th->th_sport = inp->inp_lport; /* network byte order already */
1679 th->th_dport = th->th_sport;
1680 }
1681 len += sizeof(th);
1682
1683 *pktlen = *buflen = len;
1684 return (pkt);
1685 }
1686
1687 const struct offload_settings *
lookup_offload_policy(struct adapter * sc,int open_type,struct mbuf * m,uint16_t vtag,struct inpcb * inp)1688 lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m,
1689 uint16_t vtag, struct inpcb *inp)
1690 {
1691 const struct t4_offload_policy *op;
1692 char *pkt;
1693 struct offload_rule *r;
1694 int i, matched, pktlen, buflen;
1695 static const struct offload_settings allow_offloading_settings = {
1696 .offload = 1,
1697 .rx_coalesce = -1,
1698 .cong_algo = -1,
1699 .sched_class = -1,
1700 .tstamp = -1,
1701 .sack = -1,
1702 .nagle = -1,
1703 .ecn = -1,
1704 .ddp = -1,
1705 .tls = -1,
1706 .txq = QUEUE_RANDOM,
1707 .rxq = QUEUE_RANDOM,
1708 .mss = -1,
1709 };
1710 static const struct offload_settings disallow_offloading_settings = {
1711 .offload = 0,
1712 /* rest is irrelevant when offload is off. */
1713 };
1714
1715 rw_assert(&sc->policy_lock, RA_LOCKED);
1716
1717 /*
1718 * If there's no Connection Offloading Policy attached to the device
1719 * then we need to return a default static policy. If
1720 * "cop_managed_offloading" is true, then we need to disallow
1721 * offloading until a COP is attached to the device. Otherwise we
1722 * allow offloading ...
1723 */
1724 op = sc->policy;
1725 if (op == NULL) {
1726 if (sc->tt.cop_managed_offloading)
1727 return (&disallow_offloading_settings);
1728 else
1729 return (&allow_offloading_settings);
1730 }
1731
1732 switch (open_type) {
1733 case OPEN_TYPE_ACTIVE:
1734 case OPEN_TYPE_LISTEN:
1735 pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen);
1736 break;
1737 case OPEN_TYPE_PASSIVE:
1738 MPASS(m != NULL);
1739 pkt = mtod(m, char *);
1740 MPASS(*pkt == CPL_PASS_ACCEPT_REQ);
1741 pkt += sizeof(struct cpl_pass_accept_req);
1742 pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req);
1743 buflen = m->m_len - sizeof(struct cpl_pass_accept_req);
1744 break;
1745 default:
1746 MPASS(0);
1747 return (&disallow_offloading_settings);
1748 }
1749
1750 if (pkt == NULL || pktlen == 0 || buflen == 0)
1751 return (&disallow_offloading_settings);
1752
1753 matched = 0;
1754 r = &op->rule[0];
1755 for (i = 0; i < op->nrules; i++, r++) {
1756 if (r->open_type != open_type &&
1757 r->open_type != OPEN_TYPE_DONTCARE) {
1758 continue;
1759 }
1760 matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen);
1761 if (matched)
1762 break;
1763 }
1764
1765 if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN)
1766 free(pkt, M_CXGBE);
1767
1768 return (matched ? &r->settings : &disallow_offloading_settings);
1769 }
1770
1771 static void
reclaim_wr_resources(void * arg,int count)1772 reclaim_wr_resources(void *arg, int count)
1773 {
1774 struct tom_data *td = arg;
1775 STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
1776 struct cpl_act_open_req *cpl;
1777 u_int opcode, atid, tid;
1778 struct wrqe *wr;
1779 struct adapter *sc = td_adapter(td);
1780
1781 mtx_lock(&td->unsent_wr_lock);
1782 STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
1783 mtx_unlock(&td->unsent_wr_lock);
1784
1785 while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
1786 STAILQ_REMOVE_HEAD(&twr_list, link);
1787
1788 cpl = wrtod(wr);
1789 opcode = GET_OPCODE(cpl);
1790
1791 switch (opcode) {
1792 case CPL_ACT_OPEN_REQ:
1793 case CPL_ACT_OPEN_REQ6:
1794 atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
1795 CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1796 act_open_failure_cleanup(sc, lookup_atid(sc, atid),
1797 EHOSTUNREACH);
1798 free(wr, M_CXGBE);
1799 break;
1800 case CPL_PASS_ACCEPT_RPL:
1801 tid = GET_TID(cpl);
1802 CTR2(KTR_CXGBE, "%s: tid %u ", __func__, tid);
1803 synack_failure_cleanup(sc, lookup_tid(sc, tid));
1804 free(wr, M_CXGBE);
1805 break;
1806 default:
1807 log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1808 "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1809 /* WR not freed here; go look at it with a debugger. */
1810 }
1811 }
1812 }
1813
1814 /*
1815 * Based on do_abort_req. We treat an abrupt hardware stop as a connection
1816 * abort from the hardware.
1817 */
1818 static void
live_tid_failure_cleanup(struct adapter * sc,struct toepcb * toep,u_int status)1819 live_tid_failure_cleanup(struct adapter *sc, struct toepcb *toep, u_int status)
1820 {
1821 struct inpcb *inp;
1822 struct tcpcb *tp;
1823 struct epoch_tracker et;
1824
1825 MPASS(!(toep->flags & TPF_SYNQE));
1826
1827 inp = toep->inp;
1828 CURVNET_SET(toep->vnet);
1829 NET_EPOCH_ENTER(et); /* for tcp_close */
1830 INP_WLOCK(inp);
1831 tp = intotcpcb(inp);
1832 toep->flags |= TPF_ABORT_SHUTDOWN;
1833 if ((inp->inp_flags & INP_DROPPED) == 0) {
1834 struct socket *so = inp->inp_socket;
1835
1836 if (so != NULL)
1837 so_error_set(so, status);
1838 tp = tcp_close(tp);
1839 if (tp == NULL)
1840 INP_WLOCK(inp); /* re-acquire */
1841 }
1842 final_cpl_received(toep);
1843 NET_EPOCH_EXIT(et);
1844 CURVNET_RESTORE();
1845 }
1846
1847 static void
cleanup_stranded_tids(void * arg,int count)1848 cleanup_stranded_tids(void *arg, int count)
1849 {
1850 TAILQ_HEAD(, toepcb) tlist = TAILQ_HEAD_INITIALIZER(tlist);
1851 TAILQ_HEAD(, synq_entry) slist = TAILQ_HEAD_INITIALIZER(slist);
1852 struct tom_data *td = arg;
1853 struct adapter *sc = td_adapter(td);
1854 struct toepcb *toep;
1855 struct synq_entry *synqe;
1856
1857 /* Clean up synq entries. */
1858 mtx_lock(&td->toep_list_lock);
1859 TAILQ_SWAP(&td->stranded_synqe, &slist, synq_entry, link);
1860 mtx_unlock(&td->toep_list_lock);
1861 while ((synqe = TAILQ_FIRST(&slist)) != NULL) {
1862 TAILQ_REMOVE(&slist, synqe, link);
1863 MPASS(synqe->tid >= 0); /* stale, was kept around for debug */
1864 synqe->tid = -1;
1865 synack_failure_cleanup(sc, synqe);
1866 }
1867
1868 /* Clean up in-flight active opens. */
1869 mtx_lock(&td->toep_list_lock);
1870 TAILQ_SWAP(&td->stranded_atids, &tlist, toepcb, link);
1871 mtx_unlock(&td->toep_list_lock);
1872 while ((toep = TAILQ_FIRST(&tlist)) != NULL) {
1873 TAILQ_REMOVE(&tlist, toep, link);
1874 MPASS(toep->tid >= 0); /* stale, was kept around for debug */
1875 toep->tid = -1;
1876 act_open_failure_cleanup(sc, toep, EHOSTUNREACH);
1877 }
1878
1879 /* Clean up live connections. */
1880 mtx_lock(&td->toep_list_lock);
1881 TAILQ_SWAP(&td->stranded_tids, &tlist, toepcb, link);
1882 mtx_unlock(&td->toep_list_lock);
1883 while ((toep = TAILQ_FIRST(&tlist)) != NULL) {
1884 TAILQ_REMOVE(&tlist, toep, link);
1885 MPASS(toep->tid >= 0); /* stale, was kept around for debug */
1886 toep->tid = -1;
1887 live_tid_failure_cleanup(sc, toep, ECONNABORTED);
1888 }
1889 }
1890
1891 /*
1892 * Ground control to Major TOM
1893 * Commencing countdown, engines on
1894 */
1895 static int
t4_tom_activate(struct adapter * sc)1896 t4_tom_activate(struct adapter *sc)
1897 {
1898 struct tom_data *td;
1899 struct toedev *tod;
1900 struct vi_info *vi;
1901 int i, rc, v;
1902
1903 ASSERT_SYNCHRONIZED_OP(sc);
1904
1905 /* per-adapter softc for TOM */
1906 td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1907 if (td == NULL)
1908 return (ENOMEM);
1909
1910 /* List of TOE PCBs and associated lock */
1911 mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1912 TAILQ_INIT(&td->toep_list);
1913 TAILQ_INIT(&td->synqe_list);
1914 TAILQ_INIT(&td->stranded_atids);
1915 TAILQ_INIT(&td->stranded_tids);
1916 TASK_INIT(&td->cleanup_stranded_tids, 0, cleanup_stranded_tids, td);
1917
1918 /* Listen context */
1919 mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1920 td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1921 &td->listen_mask, HASH_NOWAIT);
1922
1923 /* List of WRs for which L2 resolution failed */
1924 mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1925 STAILQ_INIT(&td->unsent_wr_list);
1926 TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1927
1928 /* TID tables */
1929 rc = alloc_tid_tabs(sc);
1930 if (rc != 0)
1931 goto done;
1932
1933 rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1934 t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1935 if (rc != 0)
1936 goto done;
1937 t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1938 V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1939
1940 alloc_tcb_history(sc, td);
1941
1942 /* toedev ops */
1943 tod = &td->tod;
1944 init_toedev(tod);
1945 tod->tod_softc = sc;
1946 tod->tod_connect = t4_connect;
1947 tod->tod_listen_start = t4_listen_start;
1948 tod->tod_listen_stop = t4_listen_stop;
1949 tod->tod_rcvd = t4_rcvd;
1950 tod->tod_output = t4_tod_output;
1951 tod->tod_send_rst = t4_send_rst;
1952 tod->tod_send_fin = t4_send_fin;
1953 tod->tod_pcb_detach = t4_pcb_detach;
1954 tod->tod_l2_update = t4_l2_update;
1955 tod->tod_syncache_added = t4_syncache_added;
1956 tod->tod_syncache_removed = t4_syncache_removed;
1957 tod->tod_syncache_respond = t4_syncache_respond;
1958 tod->tod_offload_socket = t4_offload_socket;
1959 tod->tod_ctloutput = t4_ctloutput;
1960 tod->tod_tcp_info = t4_tcp_info;
1961 #ifdef KERN_TLS
1962 tod->tod_alloc_tls_session = t4_alloc_tls_session;
1963 #endif
1964 tod->tod_pmtu_update = t4_pmtu_update;
1965
1966 for_each_port(sc, i) {
1967 for_each_vi(sc->port[i], v, vi) {
1968 SETTOEDEV(vi->ifp, &td->tod);
1969 }
1970 }
1971
1972 sc->tom_softc = td;
1973 register_toedev(sc->tom_softc);
1974
1975 done:
1976 if (rc != 0)
1977 free_tom_data(sc, td);
1978 return (rc);
1979 }
1980
1981 static int
t4_tom_deactivate(struct adapter * sc)1982 t4_tom_deactivate(struct adapter *sc)
1983 {
1984 int rc = 0, i, v;
1985 struct tom_data *td = sc->tom_softc;
1986 struct vi_info *vi;
1987
1988 ASSERT_SYNCHRONIZED_OP(sc);
1989
1990 if (td == NULL)
1991 return (0); /* XXX. KASSERT? */
1992
1993 /* These ULDs rely on the TOE. */
1994 if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI) ||
1995 uld_active(sc, ULD_NVME))
1996 return (EBUSY);
1997
1998 if (sc->offload_map != 0) {
1999 for_each_port(sc, i) {
2000 for_each_vi(sc->port[i], v, vi) {
2001 toe_capability(vi, false);
2002 if_setcapenablebit(vi->ifp, 0, IFCAP_TOE);
2003 SETTOEDEV(vi->ifp, NULL);
2004 }
2005 }
2006 MPASS(sc->offload_map == 0);
2007 }
2008
2009 mtx_lock(&td->toep_list_lock);
2010 if (!TAILQ_EMPTY(&td->toep_list))
2011 rc = EBUSY;
2012 MPASS(TAILQ_EMPTY(&td->synqe_list));
2013 MPASS(TAILQ_EMPTY(&td->stranded_tids));
2014 mtx_unlock(&td->toep_list_lock);
2015
2016 mtx_lock(&td->lctx_hash_lock);
2017 if (td->lctx_count > 0)
2018 rc = EBUSY;
2019 mtx_unlock(&td->lctx_hash_lock);
2020
2021 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
2022 taskqueue_drain(taskqueue_thread, &td->cleanup_stranded_tids);
2023 mtx_lock(&td->unsent_wr_lock);
2024 if (!STAILQ_EMPTY(&td->unsent_wr_list))
2025 rc = EBUSY;
2026 mtx_unlock(&td->unsent_wr_lock);
2027
2028 if (rc == 0) {
2029 unregister_toedev(sc->tom_softc);
2030 free_tom_data(sc, td);
2031 sc->tom_softc = NULL;
2032 }
2033
2034 return (rc);
2035 }
2036
2037 static void
stop_atids(struct adapter * sc)2038 stop_atids(struct adapter *sc)
2039 {
2040 struct tom_data *td = sc->tom_softc;
2041 struct tid_info *t = &sc->tids;
2042 struct toepcb *toep;
2043 int atid;
2044
2045 /*
2046 * Hashfilters and T6-KTLS are the only other users of atids but they're
2047 * both mutually exclusive with TOE. That means t4_tom owns all the
2048 * atids in the table.
2049 */
2050 MPASS(!is_hashfilter(sc));
2051 if (is_t6(sc))
2052 MPASS(!(sc->flags & KERN_TLS_ON));
2053
2054 /* New atids are not being allocated. */
2055 #ifdef INVARIANTS
2056 mtx_lock(&t->atid_lock);
2057 MPASS(t->atid_alloc_stopped == true);
2058 mtx_unlock(&t->atid_lock);
2059 #endif
2060
2061 /*
2062 * In-use atids fall in one of these two categories:
2063 * a) Those waiting for L2 resolution before being submitted to
2064 * hardware.
2065 * b) Those that have been submitted to hardware and are awaiting
2066 * replies that will never arrive because the LLD is stopped.
2067 */
2068 for (atid = 0; atid < t->natids; atid++) {
2069 toep = lookup_atid(sc, atid);
2070 if ((uintptr_t)toep >= (uintptr_t)&t->atid_tab[0] &&
2071 (uintptr_t)toep < (uintptr_t)&t->atid_tab[t->natids])
2072 continue;
2073 if (__predict_false(toep == NULL))
2074 continue;
2075 MPASS(toep->tid == atid);
2076 MPASS(toep->incarnation == sc->incarnation);
2077 /*
2078 * Take the atid out of the lookup table. toep->tid is stale
2079 * after this but useful for debug.
2080 */
2081 CTR(KTR_CXGBE, "%s: atid %d@%d STRANDED, removed from table",
2082 __func__, atid, toep->incarnation);
2083 free_atid(sc, toep->tid);
2084 #if 0
2085 toep->tid = -1;
2086 #endif
2087 mtx_lock(&td->toep_list_lock);
2088 toep->flags &= ~TPF_IN_TOEP_LIST;
2089 TAILQ_REMOVE(&td->toep_list, toep, link);
2090 TAILQ_INSERT_TAIL(&td->stranded_atids, toep, link);
2091 mtx_unlock(&td->toep_list_lock);
2092 }
2093 MPASS(atomic_load_int(&t->atids_in_use) == 0);
2094 }
2095
2096 static void
stop_tids(struct adapter * sc)2097 stop_tids(struct adapter *sc)
2098 {
2099 struct tom_data *td = sc->tom_softc;
2100 struct toepcb *toep;
2101 #ifdef INVARIANTS
2102 struct tid_info *t = &sc->tids;
2103 #endif
2104
2105 /*
2106 * The LLD's offload queues are stopped so do_act_establish and
2107 * do_pass_accept_req cannot run and insert tids in parallel with this
2108 * thread. stop_stid_tab has also run and removed the synq entries'
2109 * tids from the table. The only tids in the table are for connections
2110 * at or beyond ESTABLISHED that are still waiting for the final CPL.
2111 */
2112 mtx_lock(&td->toep_list_lock);
2113 TAILQ_FOREACH(toep, &td->toep_list, link) {
2114 MPASS(sc->incarnation == toep->incarnation);
2115 MPASS(toep->tid >= 0);
2116 MPASS(toep == lookup_tid(sc, toep->tid));
2117 /* Remove tid from the lookup table immediately. */
2118 CTR(KTR_CXGBE, "%s: tid %d@%d STRANDED, removed from table",
2119 __func__, toep->tid, toep->incarnation);
2120 remove_tid(sc, toep->tid, toep->ce ? 2 : 1);
2121 #if 0
2122 /* toep->tid is stale now but left alone for debug. */
2123 toep->tid = -1;
2124 #endif
2125 /* All toep in this list will get bulk moved to stranded_tids */
2126 toep->flags &= ~TPF_IN_TOEP_LIST;
2127 }
2128 MPASS(TAILQ_EMPTY(&td->stranded_tids));
2129 TAILQ_CONCAT(&td->stranded_tids, &td->toep_list, link);
2130 MPASS(TAILQ_EMPTY(&td->toep_list));
2131 mtx_unlock(&td->toep_list_lock);
2132
2133 MPASS(atomic_load_int(&t->tids_in_use) == 0);
2134 }
2135
2136 /*
2137 * L2T is stable because
2138 * 1. stop_lld stopped all new allocations.
2139 * 2. stop_lld also stopped the tx wrq so nothing is enqueueing new WRs to the
2140 * queue or to l2t_entry->wr_list.
2141 * 3. t4_l2t_update is ignoring all L2 updates.
2142 */
2143 static void
stop_tom_l2t(struct adapter * sc)2144 stop_tom_l2t(struct adapter *sc)
2145 {
2146 struct l2t_data *d = sc->l2t;
2147 struct tom_data *td = sc->tom_softc;
2148 struct l2t_entry *e;
2149 struct wrqe *wr;
2150 int i;
2151
2152 /*
2153 * This task cannot be enqueued because L2 state changes are not being
2154 * processed. But if it's already scheduled or running then we need to
2155 * wait for it to cleanup the atids in the unsent_wr_list.
2156 */
2157 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
2158 MPASS(STAILQ_EMPTY(&td->unsent_wr_list));
2159
2160 for (i = 0; i < d->l2t_size; i++) {
2161 e = &d->l2tab[i];
2162 mtx_lock(&e->lock);
2163 if (e->state == L2T_STATE_VALID || e->state == L2T_STATE_STALE)
2164 e->state = L2T_STATE_RESOLVING;
2165 /*
2166 * stop_atids is going to clean up _all_ atids in use, including
2167 * these that were pending L2 resolution. Just discard the WRs.
2168 */
2169 while ((wr = STAILQ_FIRST(&e->wr_list)) != NULL) {
2170 STAILQ_REMOVE_HEAD(&e->wr_list, link);
2171 free(wr, M_CXGBE);
2172 }
2173 mtx_unlock(&e->lock);
2174 }
2175 }
2176
2177 static int
t4_tom_stop(struct adapter * sc)2178 t4_tom_stop(struct adapter *sc)
2179 {
2180 struct tid_info *t = &sc->tids;
2181 struct tom_data *td = sc->tom_softc;
2182
2183 ASSERT_SYNCHRONIZED_OP(sc);
2184
2185 stop_tom_l2t(sc);
2186 if (atomic_load_int(&t->atids_in_use) > 0)
2187 stop_atids(sc);
2188 if (atomic_load_int(&t->stids_in_use) > 0)
2189 stop_stid_tab(sc);
2190 if (atomic_load_int(&t->tids_in_use) > 0)
2191 stop_tids(sc);
2192 taskqueue_enqueue(taskqueue_thread, &td->cleanup_stranded_tids);
2193
2194 /*
2195 * L2T and atid_tab are restarted before t4_tom_restart so this assert
2196 * is not valid in t4_tom_restart. This is the next best place for it.
2197 */
2198 MPASS(STAILQ_EMPTY(&td->unsent_wr_list));
2199
2200 return (0);
2201 }
2202
2203 static int
t4_tom_restart(struct adapter * sc)2204 t4_tom_restart(struct adapter *sc)
2205 {
2206 ASSERT_SYNCHRONIZED_OP(sc);
2207
2208 restart_stid_tab(sc);
2209
2210 return (0);
2211 }
2212
2213 static int
t4_ctloutput_tom(struct socket * so,struct sockopt * sopt)2214 t4_ctloutput_tom(struct socket *so, struct sockopt *sopt)
2215 {
2216 struct tcpcb *tp = sototcpcb(so);
2217 struct toepcb *toep = tp->t_toe;
2218 int error, optval;
2219
2220 if (sopt->sopt_level == IPPROTO_TCP && sopt->sopt_name == TCP_USE_DDP) {
2221 if (sopt->sopt_dir != SOPT_SET)
2222 return (EOPNOTSUPP);
2223
2224 if (sopt->sopt_td != NULL) {
2225 /* Only settable by the kernel. */
2226 return (EPERM);
2227 }
2228
2229 error = sooptcopyin(sopt, &optval, sizeof(optval),
2230 sizeof(optval));
2231 if (error != 0)
2232 return (error);
2233
2234 if (optval != 0)
2235 return (t4_enable_ddp_rcv(so, toep));
2236 else
2237 return (EOPNOTSUPP);
2238 }
2239 return (tcp_ctloutput(so, sopt));
2240 }
2241
2242 static int
t4_aio_queue_tom(struct socket * so,struct kaiocb * job)2243 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
2244 {
2245 struct tcpcb *tp = sototcpcb(so);
2246 struct toepcb *toep = tp->t_toe;
2247 int error;
2248
2249 /*
2250 * No lock is needed as TOE sockets never change between
2251 * active and passive.
2252 */
2253 if (SOLISTENING(so))
2254 return (EINVAL);
2255
2256 if (ulp_mode(toep) == ULP_MODE_TCPDDP ||
2257 ulp_mode(toep) == ULP_MODE_NONE) {
2258 error = t4_aio_queue_ddp(so, job);
2259 if (error == 0)
2260 return (0);
2261 else if (error != EOPNOTSUPP)
2262 return (soaio_queue_generic(so, job));
2263 }
2264
2265 if (t4_aio_queue_aiotx(so, job) != 0)
2266 return (soaio_queue_generic(so, job));
2267 else
2268 return (0);
2269 }
2270
2271 /*
2272 * Request/response structure used to find out the adapter offloading
2273 * a socket.
2274 */
2275 struct find_offload_adapter_data {
2276 struct socket *so;
2277 struct adapter *sc; /* result */
2278 };
2279
2280 static void
find_offload_adapter_cb(struct adapter * sc,void * arg)2281 find_offload_adapter_cb(struct adapter *sc, void *arg)
2282 {
2283 struct find_offload_adapter_data *fa = arg;
2284 struct socket *so = fa->so;
2285 struct tom_data *td = sc->tom_softc;
2286 struct tcpcb *tp;
2287 struct inpcb *inp;
2288
2289 /* Non-TCP were filtered out earlier. */
2290 MPASS(so->so_proto->pr_protocol == IPPROTO_TCP);
2291
2292 if (fa->sc != NULL)
2293 return; /* Found already. */
2294
2295 if (td == NULL)
2296 return; /* TOE not enabled on this adapter. */
2297
2298 inp = sotoinpcb(so);
2299 INP_WLOCK(inp);
2300 if ((inp->inp_flags & INP_DROPPED) == 0) {
2301 tp = intotcpcb(inp);
2302 if (tp->t_flags & TF_TOE && tp->tod == &td->tod)
2303 fa->sc = sc; /* Found. */
2304 }
2305 INP_WUNLOCK(inp);
2306 }
2307
2308 struct adapter *
find_offload_adapter(struct socket * so)2309 find_offload_adapter(struct socket *so)
2310 {
2311 struct find_offload_adapter_data fa;
2312
2313 fa.sc = NULL;
2314 fa.so = so;
2315 t4_iterate(find_offload_adapter_cb, &fa);
2316 return (fa.sc);
2317 }
2318
2319 void
send_txdataplen_max_flowc_wr(struct adapter * sc,struct toepcb * toep,int maxlen)2320 send_txdataplen_max_flowc_wr(struct adapter *sc, struct toepcb *toep,
2321 int maxlen)
2322 {
2323 struct wrqe *wr;
2324 struct fw_flowc_wr *flowc;
2325 const u_int nparams = 1;
2326 u_int flowclen;
2327 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
2328
2329 CTR(KTR_CXGBE, "%s: tid %u maxlen=%d", __func__, toep->tid, maxlen);
2330
2331 flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
2332
2333 wr = alloc_wrqe(roundup2(flowclen, 16), &toep->ofld_txq->wrq);
2334 if (wr == NULL) {
2335 /* XXX */
2336 panic("%s: allocation failure.", __func__);
2337 }
2338 flowc = wrtod(wr);
2339 memset(flowc, 0, wr->wr_len);
2340
2341 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
2342 V_FW_FLOWC_WR_NPARAMS(nparams));
2343 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
2344 V_FW_WR_FLOWID(toep->tid));
2345
2346 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_TXDATAPLEN_MAX;
2347 flowc->mnemval[0].val = htobe32(maxlen);
2348
2349 KASSERT(howmany(flowclen, 16) <= MAX_OFLD_TX_SDESC_CREDITS,
2350 ("%s: tx_credits %u too large", __func__, howmany(flowclen, 16)));
2351 txsd->tx_credits = howmany(flowclen, 16);
2352 txsd->plen = 0;
2353 KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
2354 ("%s: not enough credits (%d)", __func__, toep->tx_credits));
2355 toep->tx_credits -= txsd->tx_credits;
2356 if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
2357 toep->txsd_pidx = 0;
2358 toep->txsd_avail--;
2359
2360 t4_wrq_tx(sc, wr);
2361 }
2362
2363 static int
t4_tom_mod_load(void)2364 t4_tom_mod_load(void)
2365 {
2366 /* CPL handlers */
2367 t4_register_cpl_handler(CPL_GET_TCB_RPL, do_get_tcb_rpl);
2368 t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2,
2369 CPL_COOKIE_TOM);
2370 t4_init_connect_cpl_handlers();
2371 t4_init_listen_cpl_handlers();
2372 t4_init_cpl_io_handlers();
2373
2374 t4_ddp_mod_load();
2375 t4_tls_mod_load();
2376
2377 bcopy(&tcp_protosw, &toe_protosw, sizeof(toe_protosw));
2378 toe_protosw.pr_ctloutput = t4_ctloutput_tom;
2379 toe_protosw.pr_aio_queue = t4_aio_queue_tom;
2380
2381 bcopy(&tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
2382 toe6_protosw.pr_ctloutput = t4_ctloutput_tom;
2383 toe6_protosw.pr_aio_queue = t4_aio_queue_tom;
2384
2385 return (t4_register_uld(&tom_uld_info, ULD_TOM));
2386 }
2387
2388 static void
tom_uninit(struct adapter * sc,void * arg)2389 tom_uninit(struct adapter *sc, void *arg)
2390 {
2391 bool *ok_to_unload = arg;
2392
2393 if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
2394 return;
2395
2396 /* Try to free resources (works only if no port has IFCAP_TOE) */
2397 if (uld_active(sc, ULD_TOM) && t4_deactivate_uld(sc, ULD_TOM) != 0)
2398 *ok_to_unload = false;
2399
2400 end_synchronized_op(sc, 0);
2401 }
2402
2403 static int
t4_tom_mod_unload(void)2404 t4_tom_mod_unload(void)
2405 {
2406 bool ok_to_unload = true;
2407
2408 t4_iterate(tom_uninit, &ok_to_unload);
2409 if (!ok_to_unload)
2410 return (EBUSY);
2411
2412 if (t4_unregister_uld(&tom_uld_info, ULD_TOM) == EBUSY)
2413 return (EBUSY);
2414
2415 t4_tls_mod_unload();
2416 t4_ddp_mod_unload();
2417
2418 t4_uninit_connect_cpl_handlers();
2419 t4_uninit_listen_cpl_handlers();
2420 t4_uninit_cpl_io_handlers();
2421 t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM);
2422 t4_register_cpl_handler(CPL_GET_TCB_RPL, NULL);
2423
2424 return (0);
2425 }
2426 #endif /* TCP_OFFLOAD */
2427
2428 static int
t4_tom_modevent(module_t mod,int cmd,void * arg)2429 t4_tom_modevent(module_t mod, int cmd, void *arg)
2430 {
2431 int rc = 0;
2432
2433 #ifdef TCP_OFFLOAD
2434 switch (cmd) {
2435 case MOD_LOAD:
2436 rc = t4_tom_mod_load();
2437 break;
2438
2439 case MOD_UNLOAD:
2440 rc = t4_tom_mod_unload();
2441 break;
2442
2443 default:
2444 rc = EINVAL;
2445 }
2446 #else
2447 printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
2448 rc = EOPNOTSUPP;
2449 #endif
2450 return (rc);
2451 }
2452
2453 static moduledata_t t4_tom_moddata= {
2454 "t4_tom",
2455 t4_tom_modevent,
2456 0
2457 };
2458
2459 MODULE_VERSION(t4_tom, 1);
2460 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
2461 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
2462 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
2463