xref: /freebsd/sys/dev/cxgbe/cxgbei/cxgbei.c (revision 6be3386466ab79a84b48429ae66244f21526d3df)
1 /*-
2  * Copyright (c) 2012 Chelsio Communications, Inc.
3  * All rights reserved.
4  *
5  * Chelsio T5xx iSCSI driver
6  *
7  * Written by: Sreenivasa Honnur <shonnur@chelsio.com>
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include "opt_inet.h"
35 #include "opt_inet6.h"
36 
37 #include <sys/types.h>
38 #include <sys/param.h>
39 #include <sys/kernel.h>
40 #include <sys/ktr.h>
41 #include <sys/module.h>
42 #include <sys/systm.h>
43 
44 #ifdef TCP_OFFLOAD
45 #include <sys/errno.h>
46 #include <sys/kthread.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/mbuf.h>
51 #include <sys/lock.h>
52 #include <sys/mutex.h>
53 #include <sys/condvar.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/in_pcb.h>
57 #include <netinet/toecore.h>
58 #include <netinet/tcp_var.h>
59 #include <netinet/tcp_fsm.h>
60 
61 #include <cam/scsi/scsi_all.h>
62 #include <cam/scsi/scsi_da.h>
63 #include <cam/ctl/ctl_io.h>
64 #include <cam/ctl/ctl.h>
65 #include <cam/ctl/ctl_backend.h>
66 #include <cam/ctl/ctl_error.h>
67 #include <cam/ctl/ctl_frontend.h>
68 #include <cam/ctl/ctl_debug.h>
69 #include <cam/ctl/ctl_ha.h>
70 #include <cam/ctl/ctl_ioctl.h>
71 
72 #include <dev/iscsi/icl.h>
73 #include <dev/iscsi/iscsi_proto.h>
74 #include <dev/iscsi/iscsi_ioctl.h>
75 #include <dev/iscsi/iscsi.h>
76 #include <cam/ctl/ctl_frontend_iscsi.h>
77 
78 #include <cam/cam.h>
79 #include <cam/cam_ccb.h>
80 #include <cam/cam_xpt.h>
81 #include <cam/cam_debug.h>
82 #include <cam/cam_sim.h>
83 #include <cam/cam_xpt_sim.h>
84 #include <cam/cam_xpt_periph.h>
85 #include <cam/cam_periph.h>
86 #include <cam/cam_compat.h>
87 #include <cam/scsi/scsi_message.h>
88 
89 #include "common/common.h"
90 #include "common/t4_msg.h"
91 #include "common/t4_regs.h"     /* for PCIE_MEM_ACCESS */
92 #include "tom/t4_tom.h"
93 #include "cxgbei.h"
94 
95 static int worker_thread_count;
96 static struct cxgbei_worker_thread_softc *cwt_softc;
97 static struct proc *cxgbei_proc;
98 
99 static void
100 free_ci_counters(struct cxgbei_data *ci)
101 {
102 
103 #define FREE_CI_COUNTER(x) do { \
104 	if (ci->x != NULL) { \
105 		counter_u64_free(ci->x); \
106 		ci->x = NULL; \
107 	} \
108 } while (0)
109 
110 	FREE_CI_COUNTER(ddp_setup_ok);
111 	FREE_CI_COUNTER(ddp_setup_error);
112 	FREE_CI_COUNTER(ddp_bytes);
113 	FREE_CI_COUNTER(ddp_pdus);
114 	FREE_CI_COUNTER(fl_bytes);
115 	FREE_CI_COUNTER(fl_pdus);
116 #undef FREE_CI_COUNTER
117 }
118 
119 static int
120 alloc_ci_counters(struct cxgbei_data *ci)
121 {
122 
123 #define ALLOC_CI_COUNTER(x) do { \
124 	ci->x = counter_u64_alloc(M_WAITOK); \
125 	if (ci->x == NULL) \
126 		goto fail; \
127 } while (0)
128 
129 	ALLOC_CI_COUNTER(ddp_setup_ok);
130 	ALLOC_CI_COUNTER(ddp_setup_error);
131 	ALLOC_CI_COUNTER(ddp_bytes);
132 	ALLOC_CI_COUNTER(ddp_pdus);
133 	ALLOC_CI_COUNTER(fl_bytes);
134 	ALLOC_CI_COUNTER(fl_pdus);
135 #undef ALLOC_CI_COUNTER
136 
137 	return (0);
138 fail:
139 	free_ci_counters(ci);
140 	return (ENOMEM);
141 }
142 
143 static void
144 read_pdu_limits(struct adapter *sc, uint32_t *max_tx_pdu_len,
145     uint32_t *max_rx_pdu_len)
146 {
147 	uint32_t tx_len, rx_len, r, v;
148 
149 	rx_len = t4_read_reg(sc, A_TP_PMM_RX_PAGE_SIZE);
150 	tx_len = t4_read_reg(sc, A_TP_PMM_TX_PAGE_SIZE);
151 
152 	r = t4_read_reg(sc, A_TP_PARA_REG2);
153 	rx_len = min(rx_len, G_MAXRXDATA(r));
154 	tx_len = min(tx_len, G_MAXRXDATA(r));
155 
156 	r = t4_read_reg(sc, A_TP_PARA_REG7);
157 	v = min(G_PMMAXXFERLEN0(r), G_PMMAXXFERLEN1(r));
158 	rx_len = min(rx_len, v);
159 	tx_len = min(tx_len, v);
160 
161 	/* Remove after FW_FLOWC_MNEM_TXDATAPLEN_MAX fix in firmware. */
162 	tx_len = min(tx_len, 3 * 4096);
163 
164 	*max_tx_pdu_len = rounddown2(tx_len, 512);
165 	*max_rx_pdu_len = rounddown2(rx_len, 512);
166 }
167 
168 /*
169  * Initialize the software state of the iSCSI ULP driver.
170  *
171  * ENXIO means firmware didn't set up something that it was supposed to.
172  */
173 static int
174 cxgbei_init(struct adapter *sc, struct cxgbei_data *ci)
175 {
176 	struct sysctl_oid *oid;
177 	struct sysctl_oid_list *children;
178 	struct ppod_region *pr;
179 	uint32_t r;
180 	int rc;
181 
182 	MPASS(sc->vres.iscsi.size > 0);
183 	MPASS(ci != NULL);
184 
185 	rc = alloc_ci_counters(ci);
186 	if (rc != 0)
187 		return (rc);
188 
189 	read_pdu_limits(sc, &ci->max_tx_pdu_len, &ci->max_rx_pdu_len);
190 
191 	pr = &ci->pr;
192 	r = t4_read_reg(sc, A_ULP_RX_ISCSI_PSZ);
193 	rc = t4_init_ppod_region(pr, &sc->vres.iscsi, r, "iSCSI page pods");
194 	if (rc != 0) {
195 		device_printf(sc->dev,
196 		    "%s: failed to initialize the iSCSI page pod region: %u.\n",
197 		    __func__, rc);
198 		free_ci_counters(ci);
199 		return (rc);
200 	}
201 
202 	r = t4_read_reg(sc, A_ULP_RX_ISCSI_TAGMASK);
203 	r &= V_ISCSITAGMASK(M_ISCSITAGMASK);
204 	if (r != pr->pr_tag_mask) {
205 		/*
206 		 * Recent firmwares are supposed to set up the iSCSI tagmask
207 		 * but we'll do it ourselves it the computed value doesn't match
208 		 * what's in the register.
209 		 */
210 		device_printf(sc->dev,
211 		    "tagmask 0x%08x does not match computed mask 0x%08x.\n", r,
212 		    pr->pr_tag_mask);
213 		t4_set_reg_field(sc, A_ULP_RX_ISCSI_TAGMASK,
214 		    V_ISCSITAGMASK(M_ISCSITAGMASK), pr->pr_tag_mask);
215 	}
216 
217 	sysctl_ctx_init(&ci->ctx);
218 	oid = device_get_sysctl_tree(sc->dev);	/* dev.t5nex.X */
219 	children = SYSCTL_CHILDREN(oid);
220 
221 	oid = SYSCTL_ADD_NODE(&ci->ctx, children, OID_AUTO, "iscsi",
222 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "iSCSI ULP statistics");
223 	children = SYSCTL_CHILDREN(oid);
224 
225 	SYSCTL_ADD_COUNTER_U64(&ci->ctx, children, OID_AUTO, "ddp_setup_ok",
226 	    CTLFLAG_RD, &ci->ddp_setup_ok,
227 	    "# of times DDP buffer was setup successfully.");
228 
229 	SYSCTL_ADD_COUNTER_U64(&ci->ctx, children, OID_AUTO, "ddp_setup_error",
230 	    CTLFLAG_RD, &ci->ddp_setup_error,
231 	    "# of times DDP buffer setup failed.");
232 
233 	SYSCTL_ADD_COUNTER_U64(&ci->ctx, children, OID_AUTO, "ddp_bytes",
234 	    CTLFLAG_RD, &ci->ddp_bytes, "# of bytes placed directly");
235 
236 	SYSCTL_ADD_COUNTER_U64(&ci->ctx, children, OID_AUTO, "ddp_pdus",
237 	    CTLFLAG_RD, &ci->ddp_pdus, "# of PDUs with data placed directly.");
238 
239 	SYSCTL_ADD_COUNTER_U64(&ci->ctx, children, OID_AUTO, "fl_bytes",
240 	    CTLFLAG_RD, &ci->fl_bytes, "# of data bytes delivered in freelist");
241 
242 	SYSCTL_ADD_COUNTER_U64(&ci->ctx, children, OID_AUTO, "fl_pdus",
243 	    CTLFLAG_RD, &ci->fl_pdus,
244 	    "# of PDUs with data delivered in freelist");
245 
246 	ci->ddp_threshold = 2048;
247 	SYSCTL_ADD_UINT(&ci->ctx, children, OID_AUTO, "ddp_threshold",
248 	    CTLFLAG_RW, &ci->ddp_threshold, 0, "Rx zero copy threshold");
249 
250 	return (0);
251 }
252 
253 static int
254 do_rx_iscsi_hdr(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
255 {
256 	struct adapter *sc = iq->adapter;
257 	struct cpl_iscsi_hdr *cpl = mtod(m, struct cpl_iscsi_hdr *);
258 	u_int tid = GET_TID(cpl);
259 	struct toepcb *toep = lookup_tid(sc, tid);
260 	struct icl_pdu *ip;
261 	struct icl_cxgbei_pdu *icp;
262 	uint16_t len_ddp = be16toh(cpl->pdu_len_ddp);
263 	uint16_t len = be16toh(cpl->len);
264 
265 	M_ASSERTPKTHDR(m);
266 	MPASS(m->m_pkthdr.len == len + sizeof(*cpl));
267 
268 	ip = icl_cxgbei_new_pdu(M_NOWAIT);
269 	if (ip == NULL)
270 		CXGBE_UNIMPLEMENTED("PDU allocation failure");
271 	m_copydata(m, sizeof(*cpl), ISCSI_BHS_SIZE, (caddr_t)ip->ip_bhs);
272 	ip->ip_data_len = G_ISCSI_PDU_LEN(len_ddp) - len;
273 	icp = ip_to_icp(ip);
274 	icp->icp_seq = ntohl(cpl->seq);
275 	icp->icp_flags = ICPF_RX_HDR;
276 
277 	/* This is the start of a new PDU.  There should be no old state. */
278 	MPASS(toep->ulpcb2 == NULL);
279 	toep->ulpcb2 = icp;
280 
281 #if 0
282 	CTR5(KTR_CXGBE, "%s: tid %u, cpl->len %u, pdu_len_ddp 0x%04x, icp %p",
283 	    __func__, tid, len, len_ddp, icp);
284 #endif
285 
286 	m_freem(m);
287 	return (0);
288 }
289 
290 static int
291 do_rx_iscsi_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
292 {
293 	struct adapter *sc = iq->adapter;
294 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
295 	struct cpl_iscsi_data *cpl =  mtod(m, struct cpl_iscsi_data *);
296 	u_int tid = GET_TID(cpl);
297 	struct toepcb *toep = lookup_tid(sc, tid);
298 	struct icl_cxgbei_pdu *icp = toep->ulpcb2;
299 
300 	M_ASSERTPKTHDR(m);
301 	MPASS(m->m_pkthdr.len == be16toh(cpl->len) + sizeof(*cpl));
302 
303 	/* Must already have received the header (but not the data). */
304 	MPASS(icp != NULL);
305 	MPASS(icp->icp_flags == ICPF_RX_HDR);
306 	MPASS(icp->ip.ip_data_mbuf == NULL);
307 
308 
309 	m_adj(m, sizeof(*cpl));
310 	MPASS(icp->ip.ip_data_len == m->m_pkthdr.len);
311 
312 	icp->icp_flags |= ICPF_RX_FLBUF;
313 	icp->ip.ip_data_mbuf = m;
314 	counter_u64_add(ci->fl_pdus, 1);
315 	counter_u64_add(ci->fl_bytes, m->m_pkthdr.len);
316 
317 #if 0
318 	CTR3(KTR_CXGBE, "%s: tid %u, cpl->len %u", __func__, tid,
319 	    be16toh(cpl->len));
320 #endif
321 
322 	return (0);
323 }
324 
325 static int
326 do_rx_iscsi_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
327 {
328 	struct adapter *sc = iq->adapter;
329 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
330 	const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
331 	u_int tid = GET_TID(cpl);
332 	struct toepcb *toep = lookup_tid(sc, tid);
333 	struct inpcb *inp = toep->inp;
334 	struct socket *so;
335 	struct sockbuf *sb;
336 	struct tcpcb *tp;
337 	struct icl_cxgbei_conn *icc;
338 	struct icl_conn *ic;
339 	struct icl_cxgbei_pdu *icp = toep->ulpcb2;
340 	struct icl_pdu *ip;
341 	u_int pdu_len, val;
342 	struct epoch_tracker et;
343 
344 	MPASS(m == NULL);
345 
346 	/* Must already be assembling a PDU. */
347 	MPASS(icp != NULL);
348 	MPASS(icp->icp_flags & ICPF_RX_HDR);	/* Data is optional. */
349 	MPASS((icp->icp_flags & ICPF_RX_STATUS) == 0);
350 
351 	pdu_len = be16toh(cpl->len);	/* includes everything. */
352 	val = be32toh(cpl->ddpvld);
353 
354 #if 0
355 	CTR5(KTR_CXGBE,
356 	    "%s: tid %u, cpl->len %u, ddpvld 0x%08x, icp_flags 0x%08x",
357 	    __func__, tid, pdu_len, val, icp->icp_flags);
358 #endif
359 
360 	icp->icp_flags |= ICPF_RX_STATUS;
361 	ip = &icp->ip;
362 	if (val & F_DDP_PADDING_ERR)
363 		icp->icp_flags |= ICPF_PAD_ERR;
364 	if (val & F_DDP_HDRCRC_ERR)
365 		icp->icp_flags |= ICPF_HCRC_ERR;
366 	if (val & F_DDP_DATACRC_ERR)
367 		icp->icp_flags |= ICPF_DCRC_ERR;
368 	if (val & F_DDP_PDU && ip->ip_data_mbuf == NULL) {
369 		MPASS((icp->icp_flags & ICPF_RX_FLBUF) == 0);
370 		MPASS(ip->ip_data_len > 0);
371 		icp->icp_flags |= ICPF_RX_DDP;
372 		counter_u64_add(ci->ddp_pdus, 1);
373 		counter_u64_add(ci->ddp_bytes, ip->ip_data_len);
374 	}
375 
376 	INP_WLOCK(inp);
377 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
378 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
379 		    __func__, tid, pdu_len, inp->inp_flags);
380 		INP_WUNLOCK(inp);
381 		icl_cxgbei_conn_pdu_free(NULL, ip);
382 #ifdef INVARIANTS
383 		toep->ulpcb2 = NULL;
384 #endif
385 		return (0);
386 	}
387 
388 	tp = intotcpcb(inp);
389 	MPASS(icp->icp_seq == tp->rcv_nxt);
390 	MPASS(tp->rcv_wnd >= pdu_len);
391 	tp->rcv_nxt += pdu_len;
392 	tp->rcv_wnd -= pdu_len;
393 	tp->t_rcvtime = ticks;
394 
395 	/* update rx credits */
396 	t4_rcvd(&toep->td->tod, tp);	/* XXX: sc->tom_softc.tod */
397 
398 	so = inp->inp_socket;
399 	sb = &so->so_rcv;
400 	SOCKBUF_LOCK(sb);
401 
402 	icc = toep->ulpcb;
403 	if (__predict_false(icc == NULL || sb->sb_state & SBS_CANTRCVMORE)) {
404 		CTR5(KTR_CXGBE,
405 		    "%s: tid %u, excess rx (%d bytes), icc %p, sb_state 0x%x",
406 		    __func__, tid, pdu_len, icc, sb->sb_state);
407 		SOCKBUF_UNLOCK(sb);
408 		INP_WUNLOCK(inp);
409 
410 		CURVNET_SET(so->so_vnet);
411 		NET_EPOCH_ENTER(et);
412 		INP_WLOCK(inp);
413 		tp = tcp_drop(tp, ECONNRESET);
414 		if (tp)
415 			INP_WUNLOCK(inp);
416 		NET_EPOCH_EXIT(et);
417 		CURVNET_RESTORE();
418 
419 		icl_cxgbei_conn_pdu_free(NULL, ip);
420 #ifdef INVARIANTS
421 		toep->ulpcb2 = NULL;
422 #endif
423 		return (0);
424 	}
425 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
426 	ic = &icc->ic;
427 	icl_cxgbei_new_pdu_set_conn(ip, ic);
428 
429 	MPASS(m == NULL); /* was unused, we'll use it now. */
430 	m = sbcut_locked(sb, sbused(sb)); /* XXXNP: toep->sb_cc accounting? */
431 	if (__predict_false(m != NULL)) {
432 		int len = m_length(m, NULL);
433 
434 		/*
435 		 * PDUs were received before the tid transitioned to ULP mode.
436 		 * Convert them to icl_cxgbei_pdus and send them to ICL before
437 		 * the PDU in icp/ip.
438 		 */
439 		CTR3(KTR_CXGBE, "%s: tid %u, %u bytes in so_rcv", __func__, tid,
440 		    len);
441 
442 		/* XXXNP: needs to be rewritten. */
443 		if (len == sizeof(struct iscsi_bhs) || len == 4 + sizeof(struct
444 		    iscsi_bhs)) {
445 			struct icl_cxgbei_pdu *icp0;
446 			struct icl_pdu *ip0;
447 
448 			ip0 = icl_cxgbei_new_pdu(M_NOWAIT);
449 			if (ip0 == NULL)
450 				CXGBE_UNIMPLEMENTED("PDU allocation failure");
451 			icl_cxgbei_new_pdu_set_conn(ip0, ic);
452 			icp0 = ip_to_icp(ip0);
453 			icp0->icp_seq = 0; /* XXX */
454 			icp0->icp_flags = ICPF_RX_HDR | ICPF_RX_STATUS;
455 			m_copydata(m, 0, sizeof(struct iscsi_bhs), (void *)ip0->ip_bhs);
456 			STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip0, ip_next);
457 		}
458 		m_freem(m);
459 	}
460 
461 	STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip, ip_next);
462 	if ((icc->rx_flags & RXF_ACTIVE) == 0) {
463 		struct cxgbei_worker_thread_softc *cwt = &cwt_softc[icc->cwt];
464 
465 		mtx_lock(&cwt->cwt_lock);
466 		icc->rx_flags |= RXF_ACTIVE;
467 		TAILQ_INSERT_TAIL(&cwt->rx_head, icc, rx_link);
468 		if (cwt->cwt_state == CWT_SLEEPING) {
469 			cwt->cwt_state = CWT_RUNNING;
470 			cv_signal(&cwt->cwt_cv);
471 		}
472 		mtx_unlock(&cwt->cwt_lock);
473 	}
474 	SOCKBUF_UNLOCK(sb);
475 	INP_WUNLOCK(inp);
476 
477 #ifdef INVARIANTS
478 	toep->ulpcb2 = NULL;
479 #endif
480 
481 	return (0);
482 }
483 
484 static int
485 cxgbei_activate(struct adapter *sc)
486 {
487 	struct cxgbei_data *ci;
488 	int rc;
489 
490 	ASSERT_SYNCHRONIZED_OP(sc);
491 
492 	if (uld_active(sc, ULD_ISCSI)) {
493 		KASSERT(0, ("%s: iSCSI offload already enabled on adapter %p",
494 		    __func__, sc));
495 		return (0);
496 	}
497 
498 	if (sc->iscsicaps == 0 || sc->vres.iscsi.size == 0) {
499 		device_printf(sc->dev,
500 		    "not iSCSI offload capable, or capability disabled.\n");
501 		return (ENOSYS);
502 	}
503 
504 	/* per-adapter softc for iSCSI */
505 	ci = malloc(sizeof(*ci), M_CXGBE, M_ZERO | M_WAITOK);
506 	if (ci == NULL)
507 		return (ENOMEM);
508 
509 	rc = cxgbei_init(sc, ci);
510 	if (rc != 0) {
511 		free(ci, M_CXGBE);
512 		return (rc);
513 	}
514 
515 	sc->iscsi_ulp_softc = ci;
516 
517 	return (0);
518 }
519 
520 static int
521 cxgbei_deactivate(struct adapter *sc)
522 {
523 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
524 
525 	ASSERT_SYNCHRONIZED_OP(sc);
526 
527 	if (ci != NULL) {
528 		sysctl_ctx_free(&ci->ctx);
529 		t4_free_ppod_region(&ci->pr);
530 		free_ci_counters(ci);
531 		free(ci, M_CXGBE);
532 		sc->iscsi_ulp_softc = NULL;
533 	}
534 
535 	return (0);
536 }
537 
538 static void
539 cxgbei_activate_all(struct adapter *sc, void *arg __unused)
540 {
541 
542 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4isact") != 0)
543 		return;
544 
545 	/* Activate iSCSI if any port on this adapter has IFCAP_TOE enabled. */
546 	if (sc->offload_map && !uld_active(sc, ULD_ISCSI))
547 		(void) t4_activate_uld(sc, ULD_ISCSI);
548 
549 	end_synchronized_op(sc, 0);
550 }
551 
552 static void
553 cxgbei_deactivate_all(struct adapter *sc, void *arg __unused)
554 {
555 
556 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4isdea") != 0)
557 		return;
558 
559 	if (uld_active(sc, ULD_ISCSI))
560 	    (void) t4_deactivate_uld(sc, ULD_ISCSI);
561 
562 	end_synchronized_op(sc, 0);
563 }
564 
565 static struct uld_info cxgbei_uld_info = {
566 	.uld_id = ULD_ISCSI,
567 	.activate = cxgbei_activate,
568 	.deactivate = cxgbei_deactivate,
569 };
570 
571 static void
572 cwt_main(void *arg)
573 {
574 	struct cxgbei_worker_thread_softc *cwt = arg;
575 	struct icl_cxgbei_conn *icc = NULL;
576 	struct icl_conn *ic;
577 	struct icl_pdu *ip;
578 	struct sockbuf *sb;
579 	STAILQ_HEAD(, icl_pdu) rx_pdus = STAILQ_HEAD_INITIALIZER(rx_pdus);
580 
581 	MPASS(cwt != NULL);
582 
583 	mtx_lock(&cwt->cwt_lock);
584 	MPASS(cwt->cwt_state == 0);
585 	cwt->cwt_state = CWT_RUNNING;
586 	cv_signal(&cwt->cwt_cv);
587 
588 	while (__predict_true(cwt->cwt_state != CWT_STOP)) {
589 		cwt->cwt_state = CWT_RUNNING;
590 		while ((icc = TAILQ_FIRST(&cwt->rx_head)) != NULL) {
591 			TAILQ_REMOVE(&cwt->rx_head, icc, rx_link);
592 			mtx_unlock(&cwt->cwt_lock);
593 
594 			ic = &icc->ic;
595 			sb = &ic->ic_socket->so_rcv;
596 
597 			SOCKBUF_LOCK(sb);
598 			MPASS(icc->rx_flags & RXF_ACTIVE);
599 			if (__predict_true(!(sb->sb_state & SBS_CANTRCVMORE))) {
600 				MPASS(STAILQ_EMPTY(&rx_pdus));
601 				STAILQ_SWAP(&icc->rcvd_pdus, &rx_pdus, icl_pdu);
602 				SOCKBUF_UNLOCK(sb);
603 
604 				/* Hand over PDUs to ICL. */
605 				while ((ip = STAILQ_FIRST(&rx_pdus)) != NULL) {
606 					STAILQ_REMOVE_HEAD(&rx_pdus, ip_next);
607 					ic->ic_receive(ip);
608 				}
609 
610 				SOCKBUF_LOCK(sb);
611 				MPASS(STAILQ_EMPTY(&rx_pdus));
612 			}
613 			MPASS(icc->rx_flags & RXF_ACTIVE);
614 			if (STAILQ_EMPTY(&icc->rcvd_pdus) ||
615 			    __predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
616 				icc->rx_flags &= ~RXF_ACTIVE;
617 			} else {
618 				/*
619 				 * More PDUs were received while we were busy
620 				 * handing over the previous batch to ICL.
621 				 * Re-add this connection to the end of the
622 				 * queue.
623 				 */
624 				mtx_lock(&cwt->cwt_lock);
625 				TAILQ_INSERT_TAIL(&cwt->rx_head, icc,
626 				    rx_link);
627 				mtx_unlock(&cwt->cwt_lock);
628 			}
629 			SOCKBUF_UNLOCK(sb);
630 
631 			mtx_lock(&cwt->cwt_lock);
632 		}
633 
634 		/* Inner loop doesn't check for CWT_STOP, do that first. */
635 		if (__predict_false(cwt->cwt_state == CWT_STOP))
636 			break;
637 		cwt->cwt_state = CWT_SLEEPING;
638 		cv_wait(&cwt->cwt_cv, &cwt->cwt_lock);
639 	}
640 
641 	MPASS(TAILQ_FIRST(&cwt->rx_head) == NULL);
642 	mtx_assert(&cwt->cwt_lock, MA_OWNED);
643 	cwt->cwt_state = CWT_STOPPED;
644 	cv_signal(&cwt->cwt_cv);
645 	mtx_unlock(&cwt->cwt_lock);
646 	kthread_exit();
647 }
648 
649 static int
650 start_worker_threads(void)
651 {
652 	int i, rc;
653 	struct cxgbei_worker_thread_softc *cwt;
654 
655 	worker_thread_count = min(mp_ncpus, 32);
656 	cwt_softc = malloc(worker_thread_count * sizeof(*cwt), M_CXGBE,
657 	    M_WAITOK | M_ZERO);
658 
659 	MPASS(cxgbei_proc == NULL);
660 	for (i = 0, cwt = &cwt_softc[0]; i < worker_thread_count; i++, cwt++) {
661 		mtx_init(&cwt->cwt_lock, "cwt lock", NULL, MTX_DEF);
662 		cv_init(&cwt->cwt_cv, "cwt cv");
663 		TAILQ_INIT(&cwt->rx_head);
664 		rc = kproc_kthread_add(cwt_main, cwt, &cxgbei_proc, NULL, 0, 0,
665 		    "cxgbei", "%d", i);
666 		if (rc != 0) {
667 			printf("cxgbei: failed to start thread #%d/%d (%d)\n",
668 			    i + 1, worker_thread_count, rc);
669 			mtx_destroy(&cwt->cwt_lock);
670 			cv_destroy(&cwt->cwt_cv);
671 			bzero(cwt, sizeof(*cwt));
672 			if (i == 0) {
673 				free(cwt_softc, M_CXGBE);
674 				worker_thread_count = 0;
675 
676 				return (rc);
677 			}
678 
679 			/* Not fatal, carry on with fewer threads. */
680 			worker_thread_count = i;
681 			rc = 0;
682 			break;
683 		}
684 
685 		/* Wait for thread to start before moving on to the next one. */
686 		mtx_lock(&cwt->cwt_lock);
687 		while (cwt->cwt_state == 0)
688 			cv_wait(&cwt->cwt_cv, &cwt->cwt_lock);
689 		mtx_unlock(&cwt->cwt_lock);
690 	}
691 
692 	MPASS(cwt_softc != NULL);
693 	MPASS(worker_thread_count > 0);
694 	return (0);
695 }
696 
697 static void
698 stop_worker_threads(void)
699 {
700 	int i;
701 	struct cxgbei_worker_thread_softc *cwt = &cwt_softc[0];
702 
703 	MPASS(worker_thread_count >= 0);
704 
705 	for (i = 0, cwt = &cwt_softc[0]; i < worker_thread_count; i++, cwt++) {
706 		mtx_lock(&cwt->cwt_lock);
707 		MPASS(cwt->cwt_state == CWT_RUNNING ||
708 		    cwt->cwt_state == CWT_SLEEPING);
709 		cwt->cwt_state = CWT_STOP;
710 		cv_signal(&cwt->cwt_cv);
711 		do {
712 			cv_wait(&cwt->cwt_cv, &cwt->cwt_lock);
713 		} while (cwt->cwt_state != CWT_STOPPED);
714 		mtx_unlock(&cwt->cwt_lock);
715 		mtx_destroy(&cwt->cwt_lock);
716 		cv_destroy(&cwt->cwt_cv);
717 	}
718 	free(cwt_softc, M_CXGBE);
719 }
720 
721 /* Select a worker thread for a connection. */
722 u_int
723 cxgbei_select_worker_thread(struct icl_cxgbei_conn *icc)
724 {
725 	struct adapter *sc = icc->sc;
726 	struct toepcb *toep = icc->toep;
727 	u_int i, n;
728 
729 	n = worker_thread_count / sc->sge.nofldrxq;
730 	if (n > 0)
731 		i = toep->vi->pi->port_id * n + arc4random() % n;
732 	else
733 		i = arc4random() % worker_thread_count;
734 
735 	CTR3(KTR_CXGBE, "%s: tid %u, cwt %u", __func__, toep->tid, i);
736 
737 	return (i);
738 }
739 
740 static int
741 cxgbei_mod_load(void)
742 {
743 	int rc;
744 
745 	t4_register_cpl_handler(CPL_ISCSI_HDR, do_rx_iscsi_hdr);
746 	t4_register_cpl_handler(CPL_ISCSI_DATA, do_rx_iscsi_data);
747 	t4_register_cpl_handler(CPL_RX_ISCSI_DDP, do_rx_iscsi_ddp);
748 
749 	rc = start_worker_threads();
750 	if (rc != 0)
751 		return (rc);
752 
753 	rc = t4_register_uld(&cxgbei_uld_info);
754 	if (rc != 0) {
755 		stop_worker_threads();
756 		return (rc);
757 	}
758 
759 	t4_iterate(cxgbei_activate_all, NULL);
760 
761 	return (rc);
762 }
763 
764 static int
765 cxgbei_mod_unload(void)
766 {
767 
768 	t4_iterate(cxgbei_deactivate_all, NULL);
769 
770 	if (t4_unregister_uld(&cxgbei_uld_info) == EBUSY)
771 		return (EBUSY);
772 
773 	stop_worker_threads();
774 
775 	t4_register_cpl_handler(CPL_ISCSI_HDR, NULL);
776 	t4_register_cpl_handler(CPL_ISCSI_DATA, NULL);
777 	t4_register_cpl_handler(CPL_RX_ISCSI_DDP, NULL);
778 
779 	return (0);
780 }
781 #endif
782 
783 static int
784 cxgbei_modevent(module_t mod, int cmd, void *arg)
785 {
786 	int rc = 0;
787 
788 #ifdef TCP_OFFLOAD
789 	switch (cmd) {
790 	case MOD_LOAD:
791 		rc = cxgbei_mod_load();
792 		if (rc == 0)
793 			rc = icl_cxgbei_mod_load();
794 		break;
795 
796 	case MOD_UNLOAD:
797 		rc = icl_cxgbei_mod_unload();
798 		if (rc == 0)
799 			rc = cxgbei_mod_unload();
800 		break;
801 
802 	default:
803 		rc = EINVAL;
804 	}
805 #else
806 	printf("cxgbei: compiled without TCP_OFFLOAD support.\n");
807 	rc = EOPNOTSUPP;
808 #endif
809 
810 	return (rc);
811 }
812 
813 static moduledata_t cxgbei_mod = {
814 	"cxgbei",
815 	cxgbei_modevent,
816 	NULL,
817 };
818 
819 MODULE_VERSION(cxgbei, 1);
820 DECLARE_MODULE(cxgbei, cxgbei_mod, SI_SUB_EXEC, SI_ORDER_ANY);
821 MODULE_DEPEND(cxgbei, t4_tom, 1, 1, 1);
822 MODULE_DEPEND(cxgbei, cxgbe, 1, 1, 1);
823 MODULE_DEPEND(cxgbei, icl, 1, 1, 1);
824