xref: /freebsd/sys/dev/cxgbe/cxgbei/cxgbei.c (revision e92ffd9b626833ebdbf2742c8ffddc6cd94b963e)
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 read_pdu_limits(struct adapter *sc, uint32_t *max_tx_data_len,
101     uint32_t *max_rx_data_len, struct ppod_region *pr)
102 {
103 	uint32_t tx_len, rx_len, r, v;
104 
105 	rx_len = t4_read_reg(sc, A_TP_PMM_RX_PAGE_SIZE);
106 	tx_len = t4_read_reg(sc, A_TP_PMM_TX_PAGE_SIZE);
107 
108 	r = t4_read_reg(sc, A_TP_PARA_REG2);
109 	rx_len = min(rx_len, G_MAXRXDATA(r));
110 	tx_len = min(tx_len, G_MAXRXDATA(r));
111 
112 	r = t4_read_reg(sc, A_TP_PARA_REG7);
113 	v = min(G_PMMAXXFERLEN0(r), G_PMMAXXFERLEN1(r));
114 	rx_len = min(rx_len, v);
115 	tx_len = min(tx_len, v);
116 
117 	/*
118 	 * AHS is not supported by the kernel so we'll not account for
119 	 * it either in our PDU len -> data segment len conversions.
120 	 */
121 	rx_len -= ISCSI_BHS_SIZE + ISCSI_HEADER_DIGEST_SIZE +
122 	    ISCSI_DATA_DIGEST_SIZE;
123 	tx_len -= ISCSI_BHS_SIZE + ISCSI_HEADER_DIGEST_SIZE +
124 	    ISCSI_DATA_DIGEST_SIZE;
125 
126 	/*
127 	 * DDP can place only 4 pages for a single PDU.  A single
128 	 * request might use larger pages than the smallest page size,
129 	 * but that cannot be guaranteed.  Assume the smallest DDP
130 	 * page size for this limit.
131 	 */
132 	rx_len = min(rx_len, 4 * (1U << pr->pr_page_shift[0]));
133 
134 	if (chip_id(sc) == CHELSIO_T5) {
135 		tx_len = min(tx_len, 15360);
136 
137 		rx_len = rounddown2(rx_len, 512);
138 		tx_len = rounddown2(tx_len, 512);
139 	}
140 
141 	*max_tx_data_len = tx_len;
142 	*max_rx_data_len = rx_len;
143 }
144 
145 /*
146  * Initialize the software state of the iSCSI ULP driver.
147  *
148  * ENXIO means firmware didn't set up something that it was supposed to.
149  */
150 static int
151 cxgbei_init(struct adapter *sc, struct cxgbei_data *ci)
152 {
153 	struct sysctl_oid *oid;
154 	struct sysctl_oid_list *children;
155 	struct ppod_region *pr;
156 	uint32_t r;
157 	int rc;
158 
159 	MPASS(sc->vres.iscsi.size > 0);
160 	MPASS(ci != NULL);
161 
162 	pr = &ci->pr;
163 	r = t4_read_reg(sc, A_ULP_RX_ISCSI_PSZ);
164 	rc = t4_init_ppod_region(pr, &sc->vres.iscsi, r, "iSCSI page pods");
165 	if (rc != 0) {
166 		device_printf(sc->dev,
167 		    "%s: failed to initialize the iSCSI page pod region: %u.\n",
168 		    __func__, rc);
169 		return (rc);
170 	}
171 
172 	r = t4_read_reg(sc, A_ULP_RX_ISCSI_TAGMASK);
173 	r &= V_ISCSITAGMASK(M_ISCSITAGMASK);
174 	if (r != pr->pr_tag_mask) {
175 		/*
176 		 * Recent firmwares are supposed to set up the iSCSI tagmask
177 		 * but we'll do it ourselves it the computed value doesn't match
178 		 * what's in the register.
179 		 */
180 		device_printf(sc->dev,
181 		    "tagmask 0x%08x does not match computed mask 0x%08x.\n", r,
182 		    pr->pr_tag_mask);
183 		t4_set_reg_field(sc, A_ULP_RX_ISCSI_TAGMASK,
184 		    V_ISCSITAGMASK(M_ISCSITAGMASK), pr->pr_tag_mask);
185 	}
186 
187 	read_pdu_limits(sc, &ci->max_tx_data_len, &ci->max_rx_data_len, pr);
188 
189 	sysctl_ctx_init(&ci->ctx);
190 	oid = device_get_sysctl_tree(sc->dev);	/* dev.t5nex.X */
191 	children = SYSCTL_CHILDREN(oid);
192 
193 	oid = SYSCTL_ADD_NODE(&ci->ctx, children, OID_AUTO, "iscsi",
194 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "iSCSI ULP settings");
195 	children = SYSCTL_CHILDREN(oid);
196 
197 	ci->ddp_threshold = 2048;
198 	SYSCTL_ADD_UINT(&ci->ctx, children, OID_AUTO, "ddp_threshold",
199 	    CTLFLAG_RW, &ci->ddp_threshold, 0, "Rx zero copy threshold");
200 
201 	SYSCTL_ADD_UINT(&ci->ctx, children, OID_AUTO, "max_rx_data_len",
202 	    CTLFLAG_RD, &ci->max_rx_data_len, 0,
203 	    "Maximum receive data segment length");
204 	SYSCTL_ADD_UINT(&ci->ctx, children, OID_AUTO, "max_tx_data_len",
205 	    CTLFLAG_RD, &ci->max_tx_data_len, 0,
206 	    "Maximum transmit data segment length");
207 
208 	return (0);
209 }
210 
211 static int
212 do_rx_iscsi_hdr(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
213 {
214 	struct adapter *sc = iq->adapter;
215 	struct cpl_iscsi_hdr *cpl = mtod(m, struct cpl_iscsi_hdr *);
216 	u_int tid = GET_TID(cpl);
217 	struct toepcb *toep = lookup_tid(sc, tid);
218 	struct icl_pdu *ip;
219 	struct icl_cxgbei_pdu *icp;
220 	uint16_t len_ddp = be16toh(cpl->pdu_len_ddp);
221 	uint16_t len = be16toh(cpl->len);
222 
223 	M_ASSERTPKTHDR(m);
224 	MPASS(m->m_pkthdr.len == len + sizeof(*cpl));
225 
226 	ip = icl_cxgbei_new_pdu(M_NOWAIT);
227 	if (ip == NULL)
228 		CXGBE_UNIMPLEMENTED("PDU allocation failure");
229 	m_copydata(m, sizeof(*cpl), ISCSI_BHS_SIZE, (caddr_t)ip->ip_bhs);
230 	ip->ip_data_len = G_ISCSI_PDU_LEN(len_ddp) - len;
231 	icp = ip_to_icp(ip);
232 	icp->icp_seq = ntohl(cpl->seq);
233 	icp->icp_flags = ICPF_RX_HDR;
234 
235 	/* This is the start of a new PDU.  There should be no old state. */
236 	MPASS(toep->ulpcb2 == NULL);
237 	toep->ulpcb2 = icp;
238 
239 #if 0
240 	CTR5(KTR_CXGBE, "%s: tid %u, cpl->len %u, pdu_len_ddp 0x%04x, icp %p",
241 	    __func__, tid, len, len_ddp, icp);
242 #endif
243 
244 	m_freem(m);
245 	return (0);
246 }
247 
248 static int
249 do_rx_iscsi_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
250 {
251 	struct adapter *sc = iq->adapter;
252 	struct cpl_iscsi_data *cpl =  mtod(m, struct cpl_iscsi_data *);
253 	u_int tid = GET_TID(cpl);
254 	struct toepcb *toep = lookup_tid(sc, tid);
255 	struct icl_cxgbei_pdu *icp = toep->ulpcb2;
256 	struct icl_pdu *ip;
257 
258 	M_ASSERTPKTHDR(m);
259 	MPASS(m->m_pkthdr.len == be16toh(cpl->len) + sizeof(*cpl));
260 
261 	if (icp == NULL) {
262 		/*
263 		 * T6 completion enabled, start of a new pdu. Header
264 		 * will come in completion CPL.
265 		 */
266 	        ip = icl_cxgbei_new_pdu(M_NOWAIT);
267 	        if (ip == NULL)
268 			CXGBE_UNIMPLEMENTED("PDU allocation failure");
269 		icp = ip_to_icp(ip);
270 	} else {
271 		/* T5 mode, header is already received. */
272 		MPASS(icp->icp_flags == ICPF_RX_HDR);
273 		MPASS(icp->ip.ip_data_mbuf == NULL);
274 		MPASS(icp->ip.ip_data_len == m->m_pkthdr.len - sizeof(*cpl));
275 	}
276 
277 	/* Trim the cpl header from mbuf. */
278 	m_adj(m, sizeof(*cpl));
279 
280 	icp->icp_flags |= ICPF_RX_FLBUF;
281 	icp->ip.ip_data_mbuf = m;
282 	toep->ofld_rxq->rx_iscsi_fl_pdus++;
283 	toep->ofld_rxq->rx_iscsi_fl_octets += m->m_pkthdr.len;
284 
285 	/*
286 	 * For T6, save the icp for further processing in the
287 	 * completion handler.
288 	 */
289 	if (icp->icp_flags == ICPF_RX_FLBUF) {
290 		MPASS(toep->ulpcb2 == NULL);
291 		toep->ulpcb2 = icp;
292 	}
293 
294 #if 0
295 	CTR4(KTR_CXGBE, "%s: tid %u, cpl->len %u, icp %p", __func__, tid,
296 	    be16toh(cpl->len), icp);
297 #endif
298 
299 	return (0);
300 }
301 
302 static int
303 do_rx_iscsi_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
304 {
305 	struct adapter *sc = iq->adapter;
306 	const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
307 	u_int tid = GET_TID(cpl);
308 	struct toepcb *toep = lookup_tid(sc, tid);
309 	struct inpcb *inp = toep->inp;
310 	struct socket *so;
311 	struct sockbuf *sb;
312 	struct tcpcb *tp;
313 	struct icl_cxgbei_conn *icc;
314 	struct icl_conn *ic;
315 	struct icl_cxgbei_pdu *icp = toep->ulpcb2;
316 	struct icl_pdu *ip;
317 	u_int pdu_len, val;
318 	struct epoch_tracker et;
319 
320 	MPASS(m == NULL);
321 
322 	/* Must already be assembling a PDU. */
323 	MPASS(icp != NULL);
324 	MPASS(icp->icp_flags & ICPF_RX_HDR);	/* Data is optional. */
325 	MPASS((icp->icp_flags & ICPF_RX_STATUS) == 0);
326 
327 	pdu_len = be16toh(cpl->len);	/* includes everything. */
328 	val = be32toh(cpl->ddpvld);
329 
330 #if 0
331 	CTR5(KTR_CXGBE,
332 	    "%s: tid %u, cpl->len %u, ddpvld 0x%08x, icp_flags 0x%08x",
333 	    __func__, tid, pdu_len, val, icp->icp_flags);
334 #endif
335 
336 	icp->icp_flags |= ICPF_RX_STATUS;
337 	ip = &icp->ip;
338 	if (val & F_DDP_PADDING_ERR) {
339 		ICL_WARN("received PDU 0x%02x with invalid padding",
340 		    ip->ip_bhs->bhs_opcode);
341 		toep->ofld_rxq->rx_iscsi_padding_errors++;
342 	}
343 	if (val & F_DDP_HDRCRC_ERR) {
344 		ICL_WARN("received PDU 0x%02x with invalid header digest",
345 		    ip->ip_bhs->bhs_opcode);
346 		toep->ofld_rxq->rx_iscsi_header_digest_errors++;
347 	}
348 	if (val & F_DDP_DATACRC_ERR) {
349 		ICL_WARN("received PDU 0x%02x with invalid data digest",
350 		    ip->ip_bhs->bhs_opcode);
351 		toep->ofld_rxq->rx_iscsi_data_digest_errors++;
352 	}
353 	if (val & F_DDP_PDU && ip->ip_data_mbuf == NULL) {
354 		MPASS((icp->icp_flags & ICPF_RX_FLBUF) == 0);
355 		MPASS(ip->ip_data_len > 0);
356 		icp->icp_flags |= ICPF_RX_DDP;
357 		toep->ofld_rxq->rx_iscsi_ddp_pdus++;
358 		toep->ofld_rxq->rx_iscsi_ddp_octets += ip->ip_data_len;
359 	}
360 
361 	INP_WLOCK(inp);
362 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
363 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
364 		    __func__, tid, pdu_len, inp->inp_flags);
365 		INP_WUNLOCK(inp);
366 		icl_cxgbei_conn_pdu_free(NULL, ip);
367 		toep->ulpcb2 = NULL;
368 		return (0);
369 	}
370 
371 	/*
372 	 * T6+ does not report data PDUs received via DDP without F
373 	 * set.  This can result in gaps in the TCP sequence space.
374 	 */
375 	tp = intotcpcb(inp);
376 	MPASS(chip_id(sc) >= CHELSIO_T6 || icp->icp_seq == tp->rcv_nxt);
377 	tp->rcv_nxt = icp->icp_seq + pdu_len;
378 	tp->t_rcvtime = ticks;
379 
380 	/*
381 	 * Don't update the window size or return credits since RX
382 	 * flow control is disabled.
383 	 */
384 
385 	so = inp->inp_socket;
386 	sb = &so->so_rcv;
387 	SOCKBUF_LOCK(sb);
388 
389 	icc = toep->ulpcb;
390 	if (__predict_false(icc == NULL || sb->sb_state & SBS_CANTRCVMORE)) {
391 		CTR5(KTR_CXGBE,
392 		    "%s: tid %u, excess rx (%d bytes), icc %p, sb_state 0x%x",
393 		    __func__, tid, pdu_len, icc, sb->sb_state);
394 		SOCKBUF_UNLOCK(sb);
395 		INP_WUNLOCK(inp);
396 
397 		CURVNET_SET(so->so_vnet);
398 		NET_EPOCH_ENTER(et);
399 		INP_WLOCK(inp);
400 		tp = tcp_drop(tp, ECONNRESET);
401 		if (tp)
402 			INP_WUNLOCK(inp);
403 		NET_EPOCH_EXIT(et);
404 		CURVNET_RESTORE();
405 
406 		icl_cxgbei_conn_pdu_free(NULL, ip);
407 		toep->ulpcb2 = NULL;
408 		return (0);
409 	}
410 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
411 	ic = &icc->ic;
412 	if ((val & (F_DDP_PADDING_ERR | F_DDP_HDRCRC_ERR |
413 	    F_DDP_DATACRC_ERR)) != 0) {
414 		SOCKBUF_UNLOCK(sb);
415 		INP_WUNLOCK(inp);
416 
417 		icl_cxgbei_conn_pdu_free(NULL, ip);
418 		toep->ulpcb2 = NULL;
419 		ic->ic_error(ic);
420 		return (0);
421 	}
422 	icl_cxgbei_new_pdu_set_conn(ip, ic);
423 
424 	MPASS(m == NULL); /* was unused, we'll use it now. */
425 	m = sbcut_locked(sb, sbused(sb)); /* XXXNP: toep->sb_cc accounting? */
426 	if (__predict_false(m != NULL)) {
427 		int len = m_length(m, NULL);
428 
429 		/*
430 		 * PDUs were received before the tid transitioned to ULP mode.
431 		 * Convert them to icl_cxgbei_pdus and send them to ICL before
432 		 * the PDU in icp/ip.
433 		 */
434 		CTR3(KTR_CXGBE, "%s: tid %u, %u bytes in so_rcv", __func__, tid,
435 		    len);
436 
437 		/* XXXNP: needs to be rewritten. */
438 		if (len == sizeof(struct iscsi_bhs) || len == 4 + sizeof(struct
439 		    iscsi_bhs)) {
440 			struct icl_cxgbei_pdu *icp0;
441 			struct icl_pdu *ip0;
442 
443 			ip0 = icl_cxgbei_new_pdu(M_NOWAIT);
444 			if (ip0 == NULL)
445 				CXGBE_UNIMPLEMENTED("PDU allocation failure");
446 			icl_cxgbei_new_pdu_set_conn(ip0, ic);
447 			icp0 = ip_to_icp(ip0);
448 			icp0->icp_seq = 0; /* XXX */
449 			icp0->icp_flags = ICPF_RX_HDR | ICPF_RX_STATUS;
450 			m_copydata(m, 0, sizeof(struct iscsi_bhs), (void *)ip0->ip_bhs);
451 			STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip0, ip_next);
452 		}
453 		m_freem(m);
454 	}
455 
456 	STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip, ip_next);
457 	if ((icc->rx_flags & RXF_ACTIVE) == 0) {
458 		struct cxgbei_worker_thread_softc *cwt = &cwt_softc[icc->cwt];
459 
460 		mtx_lock(&cwt->cwt_lock);
461 		icc->rx_flags |= RXF_ACTIVE;
462 		TAILQ_INSERT_TAIL(&cwt->rx_head, icc, rx_link);
463 		if (cwt->cwt_state == CWT_SLEEPING) {
464 			cwt->cwt_state = CWT_RUNNING;
465 			cv_signal(&cwt->cwt_cv);
466 		}
467 		mtx_unlock(&cwt->cwt_lock);
468 	}
469 	SOCKBUF_UNLOCK(sb);
470 	INP_WUNLOCK(inp);
471 
472 	toep->ulpcb2 = NULL;
473 
474 	return (0);
475 }
476 
477 static int
478 do_rx_iscsi_cmp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
479 {
480 	struct epoch_tracker et;
481 	struct adapter *sc = iq->adapter;
482 	struct cpl_rx_iscsi_cmp *cpl = mtod(m, struct cpl_rx_iscsi_cmp *);
483 	u_int tid = GET_TID(cpl);
484 	struct toepcb *toep = lookup_tid(sc, tid);
485 	struct icl_cxgbei_pdu *icp = toep->ulpcb2;
486 	struct icl_pdu *ip;
487 	struct cxgbei_cmp *cmp;
488 	struct inpcb *inp = toep->inp;
489 #ifdef INVARIANTS
490 	uint16_t len = be16toh(cpl->len);
491 	u_int data_digest_len;
492 #endif
493 	struct socket *so;
494 	struct sockbuf *sb;
495 	struct tcpcb *tp;
496 	struct icl_cxgbei_conn *icc;
497 	struct icl_conn *ic;
498 	struct iscsi_bhs_data_out *bhsdo;
499 	u_int val = be32toh(cpl->ddpvld);
500 	u_int npdus, pdu_len;
501 	uint32_t prev_seg_len;
502 
503 	M_ASSERTPKTHDR(m);
504 	MPASS(m->m_pkthdr.len == len + sizeof(*cpl));
505 
506 	if ((val & F_DDP_PDU) == 0) {
507 		MPASS(icp != NULL);
508 		MPASS((icp->icp_flags & ICPF_RX_STATUS) == 0);
509 		ip = &icp->ip;
510 	}
511 
512 	if (icp == NULL) {
513 		/* T6 completion enabled, start of a new PDU. */
514 		ip = icl_cxgbei_new_pdu(M_NOWAIT);
515 		if (ip == NULL)
516 			CXGBE_UNIMPLEMENTED("PDU allocation failure");
517 		icp = ip_to_icp(ip);
518 	}
519 	pdu_len = G_ISCSI_PDU_LEN(be16toh(cpl->pdu_len_ddp));
520 
521 #if 0
522 	CTR5(KTR_CXGBE,
523 	    "%s: tid %u, cpl->len %u, ddpvld 0x%08x, icp %p",
524 	    __func__, tid, pdu_len, val, icp);
525 #endif
526 
527 	/* Copy header */
528 	m_copydata(m, sizeof(*cpl), ISCSI_BHS_SIZE, (caddr_t)ip->ip_bhs);
529 	bhsdo = (struct iscsi_bhs_data_out *)ip->ip_bhs;
530 	ip->ip_data_len = bhsdo->bhsdo_data_segment_len[0] << 16 |
531 	    bhsdo->bhsdo_data_segment_len[1] << 8 |
532 	    bhsdo->bhsdo_data_segment_len[2];
533 	icp->icp_seq = ntohl(cpl->seq);
534 	icp->icp_flags |= ICPF_RX_HDR;
535 	icp->icp_flags |= ICPF_RX_STATUS;
536 
537 	if (val & F_DDP_PADDING_ERR) {
538 		ICL_WARN("received PDU 0x%02x with invalid padding",
539 		    ip->ip_bhs->bhs_opcode);
540 		toep->ofld_rxq->rx_iscsi_padding_errors++;
541 	}
542 	if (val & F_DDP_HDRCRC_ERR) {
543 		ICL_WARN("received PDU 0x%02x with invalid header digest",
544 		    ip->ip_bhs->bhs_opcode);
545 		toep->ofld_rxq->rx_iscsi_header_digest_errors++;
546 	}
547 	if (val & F_DDP_DATACRC_ERR) {
548 		ICL_WARN("received PDU 0x%02x with invalid data digest",
549 		    ip->ip_bhs->bhs_opcode);
550 		toep->ofld_rxq->rx_iscsi_data_digest_errors++;
551 	}
552 
553 	INP_WLOCK(inp);
554 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
555 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
556 		    __func__, tid, pdu_len, inp->inp_flags);
557 		INP_WUNLOCK(inp);
558 		icl_cxgbei_conn_pdu_free(NULL, ip);
559 		toep->ulpcb2 = NULL;
560 		m_freem(m);
561 		return (0);
562 	}
563 
564 	tp = intotcpcb(inp);
565 
566 	/*
567 	 * If icc is NULL, the connection is being closed in
568 	 * icl_cxgbei_conn_close(), just drop this data.
569 	 */
570 	icc = toep->ulpcb;
571 	if (__predict_false(icc == NULL)) {
572 		CTR4(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes), icc %p",
573 		    __func__, tid, pdu_len, icc);
574 
575 		/*
576 		 * Update rcv_nxt so the sequence number of the FIN
577 		 * doesn't appear wrong.
578 		 */
579 		tp->rcv_nxt = icp->icp_seq + pdu_len;
580 		tp->t_rcvtime = ticks;
581 		INP_WUNLOCK(inp);
582 
583 		icl_cxgbei_conn_pdu_free(NULL, ip);
584 		toep->ulpcb2 = NULL;
585 		m_freem(m);
586 		return (0);
587 	}
588 
589 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
590 	ic = &icc->ic;
591 	if ((val & (F_DDP_PADDING_ERR | F_DDP_HDRCRC_ERR |
592 	    F_DDP_DATACRC_ERR)) != 0) {
593 		INP_WUNLOCK(inp);
594 
595 		icl_cxgbei_conn_pdu_free(NULL, ip);
596 		toep->ulpcb2 = NULL;
597 		m_freem(m);
598 		ic->ic_error(ic);
599 		return (0);
600 	}
601 
602 #ifdef INVARIANTS
603 	data_digest_len = (icc->ulp_submode & ULP_CRC_DATA) ?
604 	    ISCSI_DATA_DIGEST_SIZE : 0;
605 	MPASS(roundup2(ip->ip_data_len, 4) == pdu_len - len - data_digest_len);
606 #endif
607 
608 	if (val & F_DDP_PDU && ip->ip_data_mbuf == NULL) {
609 		MPASS((icp->icp_flags & ICPF_RX_FLBUF) == 0);
610 		MPASS(ip->ip_data_len > 0);
611 		icp->icp_flags |= ICPF_RX_DDP;
612 		bhsdo = (struct iscsi_bhs_data_out *)ip->ip_bhs;
613 
614 		switch (ip->ip_bhs->bhs_opcode & ~ISCSI_BHS_OPCODE_IMMEDIATE) {
615 		case ISCSI_BHS_OPCODE_SCSI_DATA_IN:
616 			cmp = cxgbei_find_cmp(icc,
617 			    be32toh(bhsdo->bhsdo_initiator_task_tag));
618 			break;
619 		case ISCSI_BHS_OPCODE_SCSI_DATA_OUT:
620 			cmp = cxgbei_find_cmp(icc,
621 			    be32toh(bhsdo->bhsdo_target_transfer_tag));
622 			break;
623 		default:
624 			__assert_unreachable();
625 		}
626 		MPASS(cmp != NULL);
627 
628 		/*
629 		 * The difference between the end of the last burst
630 		 * and the offset of the last PDU in this burst is
631 		 * the additional data received via DDP.
632 		 */
633 		prev_seg_len = be32toh(bhsdo->bhsdo_buffer_offset) -
634 		    cmp->next_buffer_offset;
635 
636 		if (prev_seg_len != 0) {
637 			uint32_t orig_datasn;
638 
639 			/*
640 			 * Return a "large" PDU representing the burst
641 			 * of PDUs.  Adjust the offset and length of
642 			 * this PDU to represent the entire burst.
643 			 */
644 			ip->ip_data_len += prev_seg_len;
645 			bhsdo->bhsdo_data_segment_len[2] = ip->ip_data_len;
646 			bhsdo->bhsdo_data_segment_len[1] = ip->ip_data_len >> 8;
647 			bhsdo->bhsdo_data_segment_len[0] = ip->ip_data_len >> 16;
648 			bhsdo->bhsdo_buffer_offset =
649 			    htobe32(cmp->next_buffer_offset);
650 
651 			orig_datasn = htobe32(bhsdo->bhsdo_datasn);
652 			npdus = orig_datasn - cmp->last_datasn;
653 			bhsdo->bhsdo_datasn = htobe32(cmp->last_datasn + 1);
654 			cmp->last_datasn = orig_datasn;
655 			ip->ip_additional_pdus = npdus - 1;
656 		} else {
657 			MPASS(htobe32(bhsdo->bhsdo_datasn) ==
658 			    cmp->last_datasn + 1);
659 			npdus = 1;
660 			cmp->last_datasn = htobe32(bhsdo->bhsdo_datasn);
661 		}
662 
663 		cmp->next_buffer_offset += ip->ip_data_len;
664 		toep->ofld_rxq->rx_iscsi_ddp_pdus += npdus;
665 		toep->ofld_rxq->rx_iscsi_ddp_octets += ip->ip_data_len;
666 	} else {
667 		MPASS(icp->icp_flags & (ICPF_RX_FLBUF));
668 		MPASS(ip->ip_data_len == ip->ip_data_mbuf->m_pkthdr.len);
669 	}
670 
671 	tp->rcv_nxt = icp->icp_seq + pdu_len;
672 	tp->t_rcvtime = ticks;
673 
674 	/*
675 	 * Don't update the window size or return credits since RX
676 	 * flow control is disabled.
677 	 */
678 
679 	so = inp->inp_socket;
680 	sb = &so->so_rcv;
681 	SOCKBUF_LOCK(sb);
682 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
683 		CTR5(KTR_CXGBE,
684 		    "%s: tid %u, excess rx (%d bytes), icc %p, sb_state 0x%x",
685 		    __func__, tid, pdu_len, icc, sb->sb_state);
686 		SOCKBUF_UNLOCK(sb);
687 		INP_WUNLOCK(inp);
688 
689 		CURVNET_SET(so->so_vnet);
690 		NET_EPOCH_ENTER(et);
691 		INP_WLOCK(inp);
692 		tp = tcp_drop(tp, ECONNRESET);
693 		if (tp != NULL)
694 			INP_WUNLOCK(inp);
695 		NET_EPOCH_EXIT(et);
696 		CURVNET_RESTORE();
697 
698 		icl_cxgbei_conn_pdu_free(NULL, ip);
699 		toep->ulpcb2 = NULL;
700 		m_freem(m);
701 		return (0);
702 	}
703 	icl_cxgbei_new_pdu_set_conn(ip, ic);
704 
705 	/* Enqueue the PDU to the received pdus queue. */
706 	STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip, ip_next);
707 	if ((icc->rx_flags & RXF_ACTIVE) == 0) {
708 		struct cxgbei_worker_thread_softc *cwt = &cwt_softc[icc->cwt];
709 
710 		mtx_lock(&cwt->cwt_lock);
711 		icc->rx_flags |= RXF_ACTIVE;
712 		TAILQ_INSERT_TAIL(&cwt->rx_head, icc, rx_link);
713 		if (cwt->cwt_state == CWT_SLEEPING) {
714 			cwt->cwt_state = CWT_RUNNING;
715 			cv_signal(&cwt->cwt_cv);
716 		}
717 		mtx_unlock(&cwt->cwt_lock);
718 	}
719 	SOCKBUF_UNLOCK(sb);
720 	INP_WUNLOCK(inp);
721 
722 	toep->ulpcb2 = NULL;
723 	m_freem(m);
724 
725 	return (0);
726 }
727 
728 static int
729 cxgbei_activate(struct adapter *sc)
730 {
731 	struct cxgbei_data *ci;
732 	int rc;
733 
734 	ASSERT_SYNCHRONIZED_OP(sc);
735 
736 	if (uld_active(sc, ULD_ISCSI)) {
737 		KASSERT(0, ("%s: iSCSI offload already enabled on adapter %p",
738 		    __func__, sc));
739 		return (0);
740 	}
741 
742 	if (sc->iscsicaps == 0 || sc->vres.iscsi.size == 0) {
743 		device_printf(sc->dev,
744 		    "not iSCSI offload capable, or capability disabled.\n");
745 		return (ENOSYS);
746 	}
747 
748 	/* per-adapter softc for iSCSI */
749 	ci = malloc(sizeof(*ci), M_CXGBE, M_ZERO | M_WAITOK);
750 	if (ci == NULL)
751 		return (ENOMEM);
752 
753 	rc = cxgbei_init(sc, ci);
754 	if (rc != 0) {
755 		free(ci, M_CXGBE);
756 		return (rc);
757 	}
758 
759 	sc->iscsi_ulp_softc = ci;
760 
761 	return (0);
762 }
763 
764 static int
765 cxgbei_deactivate(struct adapter *sc)
766 {
767 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
768 
769 	ASSERT_SYNCHRONIZED_OP(sc);
770 
771 	if (ci != NULL) {
772 		sysctl_ctx_free(&ci->ctx);
773 		t4_free_ppod_region(&ci->pr);
774 		free(ci, M_CXGBE);
775 		sc->iscsi_ulp_softc = NULL;
776 	}
777 
778 	return (0);
779 }
780 
781 static void
782 cxgbei_activate_all(struct adapter *sc, void *arg __unused)
783 {
784 
785 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4isact") != 0)
786 		return;
787 
788 	/* Activate iSCSI if any port on this adapter has IFCAP_TOE enabled. */
789 	if (sc->offload_map && !uld_active(sc, ULD_ISCSI))
790 		(void) t4_activate_uld(sc, ULD_ISCSI);
791 
792 	end_synchronized_op(sc, 0);
793 }
794 
795 static void
796 cxgbei_deactivate_all(struct adapter *sc, void *arg __unused)
797 {
798 
799 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4isdea") != 0)
800 		return;
801 
802 	if (uld_active(sc, ULD_ISCSI))
803 	    (void) t4_deactivate_uld(sc, ULD_ISCSI);
804 
805 	end_synchronized_op(sc, 0);
806 }
807 
808 static struct uld_info cxgbei_uld_info = {
809 	.uld_id = ULD_ISCSI,
810 	.activate = cxgbei_activate,
811 	.deactivate = cxgbei_deactivate,
812 };
813 
814 static void
815 cwt_main(void *arg)
816 {
817 	struct cxgbei_worker_thread_softc *cwt = arg;
818 	struct icl_cxgbei_conn *icc = NULL;
819 	struct icl_conn *ic;
820 	struct icl_pdu *ip;
821 	struct sockbuf *sb;
822 	STAILQ_HEAD(, icl_pdu) rx_pdus = STAILQ_HEAD_INITIALIZER(rx_pdus);
823 
824 	MPASS(cwt != NULL);
825 
826 	mtx_lock(&cwt->cwt_lock);
827 	MPASS(cwt->cwt_state == 0);
828 	cwt->cwt_state = CWT_RUNNING;
829 	cv_signal(&cwt->cwt_cv);
830 
831 	while (__predict_true(cwt->cwt_state != CWT_STOP)) {
832 		cwt->cwt_state = CWT_RUNNING;
833 		while ((icc = TAILQ_FIRST(&cwt->rx_head)) != NULL) {
834 			TAILQ_REMOVE(&cwt->rx_head, icc, rx_link);
835 			mtx_unlock(&cwt->cwt_lock);
836 
837 			ic = &icc->ic;
838 			sb = &ic->ic_socket->so_rcv;
839 
840 			SOCKBUF_LOCK(sb);
841 			MPASS(icc->rx_flags & RXF_ACTIVE);
842 			if (__predict_true(!(sb->sb_state & SBS_CANTRCVMORE))) {
843 				MPASS(STAILQ_EMPTY(&rx_pdus));
844 				STAILQ_SWAP(&icc->rcvd_pdus, &rx_pdus, icl_pdu);
845 				SOCKBUF_UNLOCK(sb);
846 
847 				/* Hand over PDUs to ICL. */
848 				while ((ip = STAILQ_FIRST(&rx_pdus)) != NULL) {
849 					STAILQ_REMOVE_HEAD(&rx_pdus, ip_next);
850 					ic->ic_receive(ip);
851 				}
852 
853 				SOCKBUF_LOCK(sb);
854 				MPASS(STAILQ_EMPTY(&rx_pdus));
855 			}
856 			MPASS(icc->rx_flags & RXF_ACTIVE);
857 			if (STAILQ_EMPTY(&icc->rcvd_pdus) ||
858 			    __predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
859 				icc->rx_flags &= ~RXF_ACTIVE;
860 			} else {
861 				/*
862 				 * More PDUs were received while we were busy
863 				 * handing over the previous batch to ICL.
864 				 * Re-add this connection to the end of the
865 				 * queue.
866 				 */
867 				mtx_lock(&cwt->cwt_lock);
868 				TAILQ_INSERT_TAIL(&cwt->rx_head, icc,
869 				    rx_link);
870 				mtx_unlock(&cwt->cwt_lock);
871 			}
872 			SOCKBUF_UNLOCK(sb);
873 
874 			mtx_lock(&cwt->cwt_lock);
875 		}
876 
877 		/* Inner loop doesn't check for CWT_STOP, do that first. */
878 		if (__predict_false(cwt->cwt_state == CWT_STOP))
879 			break;
880 		cwt->cwt_state = CWT_SLEEPING;
881 		cv_wait(&cwt->cwt_cv, &cwt->cwt_lock);
882 	}
883 
884 	MPASS(TAILQ_FIRST(&cwt->rx_head) == NULL);
885 	mtx_assert(&cwt->cwt_lock, MA_OWNED);
886 	cwt->cwt_state = CWT_STOPPED;
887 	cv_signal(&cwt->cwt_cv);
888 	mtx_unlock(&cwt->cwt_lock);
889 	kthread_exit();
890 }
891 
892 static int
893 start_worker_threads(void)
894 {
895 	int i, rc;
896 	struct cxgbei_worker_thread_softc *cwt;
897 
898 	worker_thread_count = min(mp_ncpus, 32);
899 	cwt_softc = malloc(worker_thread_count * sizeof(*cwt), M_CXGBE,
900 	    M_WAITOK | M_ZERO);
901 
902 	MPASS(cxgbei_proc == NULL);
903 	for (i = 0, cwt = &cwt_softc[0]; i < worker_thread_count; i++, cwt++) {
904 		mtx_init(&cwt->cwt_lock, "cwt lock", NULL, MTX_DEF);
905 		cv_init(&cwt->cwt_cv, "cwt cv");
906 		TAILQ_INIT(&cwt->rx_head);
907 		rc = kproc_kthread_add(cwt_main, cwt, &cxgbei_proc, NULL, 0, 0,
908 		    "cxgbei", "%d", i);
909 		if (rc != 0) {
910 			printf("cxgbei: failed to start thread #%d/%d (%d)\n",
911 			    i + 1, worker_thread_count, rc);
912 			mtx_destroy(&cwt->cwt_lock);
913 			cv_destroy(&cwt->cwt_cv);
914 			bzero(cwt, sizeof(*cwt));
915 			if (i == 0) {
916 				free(cwt_softc, M_CXGBE);
917 				worker_thread_count = 0;
918 
919 				return (rc);
920 			}
921 
922 			/* Not fatal, carry on with fewer threads. */
923 			worker_thread_count = i;
924 			rc = 0;
925 			break;
926 		}
927 
928 		/* Wait for thread to start before moving on to the next one. */
929 		mtx_lock(&cwt->cwt_lock);
930 		while (cwt->cwt_state == 0)
931 			cv_wait(&cwt->cwt_cv, &cwt->cwt_lock);
932 		mtx_unlock(&cwt->cwt_lock);
933 	}
934 
935 	MPASS(cwt_softc != NULL);
936 	MPASS(worker_thread_count > 0);
937 	return (0);
938 }
939 
940 static void
941 stop_worker_threads(void)
942 {
943 	int i;
944 	struct cxgbei_worker_thread_softc *cwt = &cwt_softc[0];
945 
946 	MPASS(worker_thread_count >= 0);
947 
948 	for (i = 0, cwt = &cwt_softc[0]; i < worker_thread_count; i++, cwt++) {
949 		mtx_lock(&cwt->cwt_lock);
950 		MPASS(cwt->cwt_state == CWT_RUNNING ||
951 		    cwt->cwt_state == CWT_SLEEPING);
952 		cwt->cwt_state = CWT_STOP;
953 		cv_signal(&cwt->cwt_cv);
954 		do {
955 			cv_wait(&cwt->cwt_cv, &cwt->cwt_lock);
956 		} while (cwt->cwt_state != CWT_STOPPED);
957 		mtx_unlock(&cwt->cwt_lock);
958 		mtx_destroy(&cwt->cwt_lock);
959 		cv_destroy(&cwt->cwt_cv);
960 	}
961 	free(cwt_softc, M_CXGBE);
962 }
963 
964 /* Select a worker thread for a connection. */
965 u_int
966 cxgbei_select_worker_thread(struct icl_cxgbei_conn *icc)
967 {
968 	struct adapter *sc = icc->sc;
969 	struct toepcb *toep = icc->toep;
970 	u_int i, n;
971 
972 	n = worker_thread_count / sc->sge.nofldrxq;
973 	if (n > 0)
974 		i = toep->vi->pi->port_id * n + arc4random() % n;
975 	else
976 		i = arc4random() % worker_thread_count;
977 
978 	CTR3(KTR_CXGBE, "%s: tid %u, cwt %u", __func__, toep->tid, i);
979 
980 	return (i);
981 }
982 
983 static int
984 cxgbei_mod_load(void)
985 {
986 	int rc;
987 
988 	t4_register_cpl_handler(CPL_ISCSI_HDR, do_rx_iscsi_hdr);
989 	t4_register_cpl_handler(CPL_ISCSI_DATA, do_rx_iscsi_data);
990 	t4_register_cpl_handler(CPL_RX_ISCSI_DDP, do_rx_iscsi_ddp);
991 	t4_register_cpl_handler(CPL_RX_ISCSI_CMP, do_rx_iscsi_cmp);
992 
993 	rc = start_worker_threads();
994 	if (rc != 0)
995 		return (rc);
996 
997 	rc = t4_register_uld(&cxgbei_uld_info);
998 	if (rc != 0) {
999 		stop_worker_threads();
1000 		return (rc);
1001 	}
1002 
1003 	t4_iterate(cxgbei_activate_all, NULL);
1004 
1005 	return (rc);
1006 }
1007 
1008 static int
1009 cxgbei_mod_unload(void)
1010 {
1011 
1012 	t4_iterate(cxgbei_deactivate_all, NULL);
1013 
1014 	if (t4_unregister_uld(&cxgbei_uld_info) == EBUSY)
1015 		return (EBUSY);
1016 
1017 	stop_worker_threads();
1018 
1019 	t4_register_cpl_handler(CPL_ISCSI_HDR, NULL);
1020 	t4_register_cpl_handler(CPL_ISCSI_DATA, NULL);
1021 	t4_register_cpl_handler(CPL_RX_ISCSI_DDP, NULL);
1022 	t4_register_cpl_handler(CPL_RX_ISCSI_CMP, NULL);
1023 
1024 	return (0);
1025 }
1026 #endif
1027 
1028 static int
1029 cxgbei_modevent(module_t mod, int cmd, void *arg)
1030 {
1031 	int rc = 0;
1032 
1033 #ifdef TCP_OFFLOAD
1034 	switch (cmd) {
1035 	case MOD_LOAD:
1036 		rc = cxgbei_mod_load();
1037 		if (rc == 0)
1038 			rc = icl_cxgbei_mod_load();
1039 		break;
1040 
1041 	case MOD_UNLOAD:
1042 		rc = icl_cxgbei_mod_unload();
1043 		if (rc == 0)
1044 			rc = cxgbei_mod_unload();
1045 		break;
1046 
1047 	default:
1048 		rc = EINVAL;
1049 	}
1050 #else
1051 	printf("cxgbei: compiled without TCP_OFFLOAD support.\n");
1052 	rc = EOPNOTSUPP;
1053 #endif
1054 
1055 	return (rc);
1056 }
1057 
1058 static moduledata_t cxgbei_mod = {
1059 	"cxgbei",
1060 	cxgbei_modevent,
1061 	NULL,
1062 };
1063 
1064 MODULE_VERSION(cxgbei, 1);
1065 DECLARE_MODULE(cxgbei, cxgbei_mod, SI_SUB_EXEC, SI_ORDER_ANY);
1066 MODULE_DEPEND(cxgbei, t4_tom, 1, 1, 1);
1067 MODULE_DEPEND(cxgbei, cxgbe, 1, 1, 1);
1068 MODULE_DEPEND(cxgbei, icl, 1, 1, 1);
1069