xref: /freebsd/sys/dev/cxgbe/t4_netmap.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
1 /*-
2  * Copyright (c) 2014 Chelsio Communications, Inc.
3  * All rights reserved.
4  * Written by: Navdeep Parhar <np@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 
34 #ifdef DEV_NETMAP
35 #include <sys/param.h>
36 #include <sys/bus.h>
37 #include <sys/eventhandler.h>
38 #include <sys/lock.h>
39 #include <sys/mbuf.h>
40 #include <sys/module.h>
41 #include <sys/selinfo.h>
42 #include <sys/socket.h>
43 #include <sys/sockio.h>
44 #include <machine/bus.h>
45 #include <net/ethernet.h>
46 #include <net/if.h>
47 #include <net/if_media.h>
48 #include <net/if_var.h>
49 #include <net/if_clone.h>
50 #include <net/if_types.h>
51 #include <net/netmap.h>
52 #include <dev/netmap/netmap_kern.h>
53 
54 #include "common/common.h"
55 #include "common/t4_regs.h"
56 #include "common/t4_regs_values.h"
57 
58 extern int fl_pad;	/* XXXNM */
59 extern int spg_len;	/* XXXNM */
60 extern int fl_pktshift;	/* XXXNM */
61 
62 SYSCTL_NODE(_hw, OID_AUTO, cxgbe, CTLFLAG_RD, 0, "cxgbe netmap parameters");
63 
64 /*
65  * 0 = normal netmap rx
66  * 1 = black hole
67  * 2 = supermassive black hole (buffer packing enabled)
68  */
69 int black_hole = 0;
70 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_black_hole, CTLFLAG_RDTUN, &black_hole, 0,
71     "Sink incoming packets.");
72 
73 int rx_ndesc = 256;
74 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_rx_ndesc, CTLFLAG_RWTUN,
75     &rx_ndesc, 0, "# of rx descriptors after which the hw cidx is updated.");
76 
77 int holdoff_tmr_idx = 2;
78 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_holdoff_tmr_idx, CTLFLAG_RWTUN,
79     &holdoff_tmr_idx, 0, "Holdoff timer index for netmap rx queues.");
80 
81 /*
82  * Congestion drops.
83  * -1: no congestion feedback (not recommended).
84  *  0: backpressure the channel instead of dropping packets right away.
85  *  1: no backpressure, drop packets for the congested queue immediately.
86  */
87 static int nm_cong_drop = 1;
88 TUNABLE_INT("hw.cxgbe.nm_cong_drop", &nm_cong_drop);
89 
90 /* netmap ifnet routines */
91 static void cxgbe_nm_init(void *);
92 static int cxgbe_nm_ioctl(struct ifnet *, unsigned long, caddr_t);
93 static int cxgbe_nm_transmit(struct ifnet *, struct mbuf *);
94 static void cxgbe_nm_qflush(struct ifnet *);
95 
96 static int cxgbe_nm_init_synchronized(struct vi_info *);
97 static int cxgbe_nm_uninit_synchronized(struct vi_info *);
98 
99 /* T4 netmap VI (ncxgbe) interface */
100 static int ncxgbe_probe(device_t);
101 static int ncxgbe_attach(device_t);
102 static int ncxgbe_detach(device_t);
103 static device_method_t ncxgbe_methods[] = {
104 	DEVMETHOD(device_probe,		ncxgbe_probe),
105 	DEVMETHOD(device_attach,	ncxgbe_attach),
106 	DEVMETHOD(device_detach,	ncxgbe_detach),
107 	{ 0, 0 }
108 };
109 static driver_t ncxgbe_driver = {
110 	"ncxgbe",
111 	ncxgbe_methods,
112 	sizeof(struct vi_info)
113 };
114 
115 /* T5 netmap VI (ncxl) interface */
116 static driver_t ncxl_driver = {
117 	"ncxl",
118 	ncxgbe_methods,
119 	sizeof(struct vi_info)
120 };
121 
122 static void
123 cxgbe_nm_init(void *arg)
124 {
125 	struct vi_info *vi = arg;
126 	struct adapter *sc = vi->pi->adapter;
127 
128 	if (begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nminit") != 0)
129 		return;
130 	cxgbe_nm_init_synchronized(vi);
131 	end_synchronized_op(sc, 0);
132 
133 	return;
134 }
135 
136 static int
137 cxgbe_nm_init_synchronized(struct vi_info *vi)
138 {
139 	struct adapter *sc = vi->pi->adapter;
140 	struct ifnet *ifp = vi->ifp;
141 	int rc = 0;
142 
143 	ASSERT_SYNCHRONIZED_OP(sc);
144 
145 	if (ifp->if_drv_flags & IFF_DRV_RUNNING)
146 		return (0);	/* already running */
147 
148 	if (!(sc->flags & FULL_INIT_DONE) &&
149 	    ((rc = adapter_full_init(sc)) != 0))
150 		return (rc);	/* error message displayed already */
151 
152 	if (!(vi->flags & VI_INIT_DONE) &&
153 	    ((rc = vi_full_init(vi)) != 0))
154 		return (rc);	/* error message displayed already */
155 
156 	rc = update_mac_settings(ifp, XGMAC_ALL);
157 	if (rc)
158 		return (rc);	/* error message displayed already */
159 
160 	ifp->if_drv_flags |= IFF_DRV_RUNNING;
161 	callout_reset(&vi->tick, hz, vi_tick, vi);
162 
163 	return (rc);
164 }
165 
166 static int
167 cxgbe_nm_uninit_synchronized(struct vi_info *vi)
168 {
169 #ifdef INVARIANTS
170 	struct adapter *sc = vi->pi->adapter;
171 #endif
172 	struct ifnet *ifp = vi->ifp;
173 
174 	ASSERT_SYNCHRONIZED_OP(sc);
175 
176 	callout_stop(&vi->tick);
177 	ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
178 
179 	return (0);
180 }
181 
182 static int
183 cxgbe_nm_ioctl(struct ifnet *ifp, unsigned long cmd, caddr_t data)
184 {
185 	int rc = 0, mtu, flags;
186 	struct vi_info *vi = ifp->if_softc;
187 	struct adapter *sc = vi->pi->adapter;
188 	struct ifreq *ifr = (struct ifreq *)data;
189 	uint32_t mask;
190 
191 	MPASS(vi->ifp == ifp);
192 
193 	switch (cmd) {
194 	case SIOCSIFMTU:
195 		mtu = ifr->ifr_mtu;
196 		if ((mtu < ETHERMIN) || (mtu > ETHERMTU_JUMBO))
197 			return (EINVAL);
198 
199 		rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmtu");
200 		if (rc)
201 			return (rc);
202 		ifp->if_mtu = mtu;
203 		if (ifp->if_drv_flags & IFF_DRV_RUNNING)
204 			rc = update_mac_settings(ifp, XGMAC_MTU);
205 		end_synchronized_op(sc, 0);
206 		break;
207 
208 	case SIOCSIFFLAGS:
209 		rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nflg");
210 		if (rc)
211 			return (rc);
212 
213 		if (ifp->if_flags & IFF_UP) {
214 			if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
215 				flags = vi->if_flags;
216 				if ((ifp->if_flags ^ flags) &
217 				    (IFF_PROMISC | IFF_ALLMULTI)) {
218 					rc = update_mac_settings(ifp,
219 					    XGMAC_PROMISC | XGMAC_ALLMULTI);
220 				}
221 			} else
222 				rc = cxgbe_nm_init_synchronized(vi);
223 			vi->if_flags = ifp->if_flags;
224 		} else if (ifp->if_drv_flags & IFF_DRV_RUNNING)
225 			rc = cxgbe_nm_uninit_synchronized(vi);
226 		end_synchronized_op(sc, 0);
227 		break;
228 
229 	case SIOCADDMULTI:
230 	case SIOCDELMULTI: /* these two are called with a mutex held :-( */
231 		rc = begin_synchronized_op(sc, vi, HOLD_LOCK, "t4nmulti");
232 		if (rc)
233 			return (rc);
234 		if (ifp->if_drv_flags & IFF_DRV_RUNNING)
235 			rc = update_mac_settings(ifp, XGMAC_MCADDRS);
236 		end_synchronized_op(sc, LOCK_HELD);
237 		break;
238 
239 	case SIOCSIFCAP:
240 		mask = ifr->ifr_reqcap ^ ifp->if_capenable;
241 		if (mask & IFCAP_TXCSUM) {
242 			ifp->if_capenable ^= IFCAP_TXCSUM;
243 			ifp->if_hwassist ^= (CSUM_TCP | CSUM_UDP | CSUM_IP);
244 		}
245 		if (mask & IFCAP_TXCSUM_IPV6) {
246 			ifp->if_capenable ^= IFCAP_TXCSUM_IPV6;
247 			ifp->if_hwassist ^= (CSUM_UDP_IPV6 | CSUM_TCP_IPV6);
248 		}
249 		if (mask & IFCAP_RXCSUM)
250 			ifp->if_capenable ^= IFCAP_RXCSUM;
251 		if (mask & IFCAP_RXCSUM_IPV6)
252 			ifp->if_capenable ^= IFCAP_RXCSUM_IPV6;
253 		break;
254 
255 	case SIOCSIFMEDIA:
256 	case SIOCGIFMEDIA:
257 		ifmedia_ioctl(ifp, ifr, &vi->media, cmd);
258 		break;
259 
260 	default:
261 		rc = ether_ioctl(ifp, cmd, data);
262 	}
263 
264 	return (rc);
265 }
266 
267 static int
268 cxgbe_nm_transmit(struct ifnet *ifp, struct mbuf *m)
269 {
270 
271 	m_freem(m);
272 	return (0);
273 }
274 
275 static void
276 cxgbe_nm_qflush(struct ifnet *ifp)
277 {
278 
279 	return;
280 }
281 
282 static int
283 alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int cong)
284 {
285 	int rc, cntxt_id, i;
286 	__be32 v;
287 	struct adapter *sc = vi->pi->adapter;
288 	struct netmap_adapter *na = NA(vi->ifp);
289 	struct fw_iq_cmd c;
290 
291 	MPASS(na != NULL);
292 	MPASS(nm_rxq->iq_desc != NULL);
293 	MPASS(nm_rxq->fl_desc != NULL);
294 
295 	bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE);
296 	bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + spg_len);
297 
298 	bzero(&c, sizeof(c));
299 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
300 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
301 	    V_FW_IQ_CMD_VFN(0));
302 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
303 	    FW_LEN16(c));
304 	if (vi->flags & INTR_RXQ) {
305 		KASSERT(nm_rxq->intr_idx < sc->intr_count,
306 		    ("%s: invalid direct intr_idx %d", __func__,
307 		    nm_rxq->intr_idx));
308 		v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
309 	} else {
310 		CXGBE_UNIMPLEMENTED(__func__);	/* XXXNM: needs review */
311 		v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx) |
312 		    F_FW_IQ_CMD_IQANDST;
313 	}
314 	c.type_to_iqandstindex = htobe32(v |
315 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
316 	    V_FW_IQ_CMD_VIID(vi->viid) |
317 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
318 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(vi->pi->tx_chan) |
319 	    F_FW_IQ_CMD_IQGTSMODE |
320 	    V_FW_IQ_CMD_IQINTCNTTHRESH(0) |
321 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
322 	c.iqsize = htobe16(vi->qsize_rxq);
323 	c.iqaddr = htobe64(nm_rxq->iq_ba);
324 	if (cong >= 0) {
325 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN |
326 		    V_FW_IQ_CMD_FL0CNGCHMAP(cong) | F_FW_IQ_CMD_FL0CONGCIF |
327 		    F_FW_IQ_CMD_FL0CONGEN);
328 	}
329 	c.iqns_to_fl0congen |=
330 	    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
331 		F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
332 		(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
333 		(black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0));
334 	c.fl0dcaen_to_fl0cidxfthresh =
335 	    htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_128B) |
336 		V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
337 	c.fl0size = htobe16(na->num_rx_desc / 8 + spg_len / EQ_ESIZE);
338 	c.fl0addr = htobe64(nm_rxq->fl_ba);
339 
340 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
341 	if (rc != 0) {
342 		device_printf(sc->dev,
343 		    "failed to create netmap ingress queue: %d\n", rc);
344 		return (rc);
345 	}
346 
347 	nm_rxq->iq_cidx = 0;
348 	MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - spg_len / IQ_ESIZE);
349 	nm_rxq->iq_gen = F_RSPD_GEN;
350 	nm_rxq->iq_cntxt_id = be16toh(c.iqid);
351 	nm_rxq->iq_abs_id = be16toh(c.physiqid);
352 	cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
353 	if (cntxt_id >= sc->sge.niq) {
354 		panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
355 		    __func__, cntxt_id, sc->sge.niq - 1);
356 	}
357 	sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
358 
359 	nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
360 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
361 	MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
362 	cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
363 	if (cntxt_id >= sc->sge.neq) {
364 		panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
365 		    __func__, cntxt_id, sc->sge.neq - 1);
366 	}
367 	sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
368 
369 	nm_rxq->fl_db_val = F_DBPRIO | V_QID(nm_rxq->fl_cntxt_id) | V_PIDX(0);
370 	if (is_t5(sc))
371 		nm_rxq->fl_db_val |= F_DBTYPE;
372 
373 	if (is_t5(sc) && cong >= 0) {
374 		uint32_t param, val;
375 
376 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
377 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
378 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
379 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
380 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
381 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
382 		if (cong == 0)
383 			val = 1 << 19;
384 		else {
385 			val = 2 << 19;
386 			for (i = 0; i < 4; i++) {
387 				if (cong & (1 << i))
388 					val |= 1 << (i << 2);
389 			}
390 		}
391 
392 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
393 		if (rc != 0) {
394 			/* report error but carry on */
395 			device_printf(sc->dev,
396 			    "failed to set congestion manager context for "
397 			    "ingress queue %d: %d\n", nm_rxq->iq_cntxt_id, rc);
398 		}
399 	}
400 
401 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
402 	    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
403 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
404 
405 	return (rc);
406 }
407 
408 static int
409 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
410 {
411 	struct adapter *sc = vi->pi->adapter;
412 	int rc;
413 
414 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
415 	    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
416 	if (rc != 0)
417 		device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
418 		    __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
419 	return (rc);
420 }
421 
422 static int
423 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
424 {
425 	int rc, cntxt_id;
426 	size_t len;
427 	struct adapter *sc = vi->pi->adapter;
428 	struct netmap_adapter *na = NA(vi->ifp);
429 	struct fw_eq_eth_cmd c;
430 
431 	MPASS(na != NULL);
432 	MPASS(nm_txq->desc != NULL);
433 
434 	len = na->num_tx_desc * EQ_ESIZE + spg_len;
435 	bzero(nm_txq->desc, len);
436 
437 	bzero(&c, sizeof(c));
438 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
439 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
440 	    V_FW_EQ_ETH_CMD_VFN(0));
441 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
442 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
443 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
444 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
445 	c.fetchszm_to_iqid =
446 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
447 		V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
448 		V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
449 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
450 		      V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
451 		      V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
452 	c.eqaddr = htobe64(nm_txq->ba);
453 
454 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
455 	if (rc != 0) {
456 		device_printf(vi->dev,
457 		    "failed to create netmap egress queue: %d\n", rc);
458 		return (rc);
459 	}
460 
461 	nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
462 	cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
463 	if (cntxt_id >= sc->sge.neq)
464 	    panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
465 		cntxt_id, sc->sge.neq - 1);
466 	sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
467 
468 	nm_txq->pidx = nm_txq->cidx = 0;
469 	MPASS(nm_txq->sidx == na->num_tx_desc);
470 	nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
471 
472 	nm_txq->doorbells = sc->doorbells;
473 	if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
474 	    isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
475 	    isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
476 		uint32_t s_qpp = sc->sge.eq_s_qpp;
477 		uint32_t mask = (1 << s_qpp) - 1;
478 		volatile uint8_t *udb;
479 
480 		udb = sc->udbs_base + UDBS_DB_OFFSET;
481 		udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
482 		nm_txq->udb_qid = nm_txq->cntxt_id & mask;
483 		if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
484 	    		clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
485 		else {
486 			udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
487 			nm_txq->udb_qid = 0;
488 		}
489 		nm_txq->udb = (volatile void *)udb;
490 	}
491 
492 	return (rc);
493 }
494 
495 static int
496 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
497 {
498 	struct adapter *sc = vi->pi->adapter;
499 	int rc;
500 
501 	rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
502 	if (rc != 0)
503 		device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
504 		    nm_txq->cntxt_id, rc);
505 	return (rc);
506 }
507 
508 static int
509 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
510     struct netmap_adapter *na)
511 {
512 	struct netmap_slot *slot;
513 	struct sge_nm_rxq *nm_rxq;
514 	struct sge_nm_txq *nm_txq;
515 	int rc, i, j, hwidx;
516 	struct hw_buf_info *hwb;
517 	uint16_t *rss;
518 
519 	ASSERT_SYNCHRONIZED_OP(sc);
520 
521 	if ((vi->flags & VI_INIT_DONE) == 0 ||
522 	    (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
523 		return (EAGAIN);
524 
525 	hwb = &sc->sge.hw_buf_info[0];
526 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
527 		if (hwb->size == NETMAP_BUF_SIZE(na))
528 			break;
529 	}
530 	if (i >= SGE_FLBUF_SIZES) {
531 		if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
532 		    NETMAP_BUF_SIZE(na));
533 		return (ENXIO);
534 	}
535 	hwidx = i;
536 
537 	/* Must set caps before calling netmap_reset */
538 	nm_set_native_flags(na);
539 
540 	for_each_nm_rxq(vi, i, nm_rxq) {
541 		alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop));
542 		nm_rxq->fl_hwidx = hwidx;
543 		slot = netmap_reset(na, NR_RX, i, 0);
544 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
545 
546 		/* We deal with 8 bufs at a time */
547 		MPASS((na->num_rx_desc & 7) == 0);
548 		MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
549 		for (j = 0; j < nm_rxq->fl_sidx; j++) {
550 			uint64_t ba;
551 
552 			PNMB(na, &slot[j], &ba);
553 			MPASS(ba != 0);
554 			nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
555 		}
556 		j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
557 		MPASS((j & 7) == 0);
558 		j /= 8;	/* driver pidx to hardware pidx */
559 		wmb();
560 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
561 		    nm_rxq->fl_db_val | V_PIDX(j));
562 	}
563 
564 	for_each_nm_txq(vi, i, nm_txq) {
565 		alloc_nm_txq_hwq(vi, nm_txq);
566 		slot = netmap_reset(na, NR_TX, i, 0);
567 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
568 	}
569 
570 	rss = malloc(vi->rss_size * sizeof (*rss), M_CXGBE, M_ZERO |
571 	    M_WAITOK);
572 	for (i = 0; i < vi->rss_size;) {
573 		for_each_nm_rxq(vi, j, nm_rxq) {
574 			rss[i++] = nm_rxq->iq_abs_id;
575 			if (i == vi->rss_size)
576 				break;
577 		}
578 	}
579 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
580 	    rss, vi->rss_size);
581 	if (rc != 0)
582 		if_printf(ifp, "netmap rss_config failed: %d\n", rc);
583 	free(rss, M_CXGBE);
584 
585 	rc = -t4_enable_vi(sc, sc->mbox, vi->viid, true, true);
586 	if (rc != 0)
587 		if_printf(ifp, "netmap enable_vi failed: %d\n", rc);
588 
589 	return (rc);
590 }
591 
592 static int
593 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
594     struct netmap_adapter *na)
595 {
596 	int rc, i;
597 	struct sge_nm_txq *nm_txq;
598 	struct sge_nm_rxq *nm_rxq;
599 
600 	ASSERT_SYNCHRONIZED_OP(sc);
601 
602 	if ((vi->flags & VI_INIT_DONE) == 0)
603 		return (0);
604 
605 	rc = -t4_enable_vi(sc, sc->mbox, vi->viid, false, false);
606 	if (rc != 0)
607 		if_printf(ifp, "netmap disable_vi failed: %d\n", rc);
608 	nm_clear_native_flags(na);
609 
610 	for_each_nm_txq(vi, i, nm_txq) {
611 		struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
612 
613 		/* Wait for hw pidx to catch up ... */
614 		while (be16toh(nm_txq->pidx) != spg->pidx)
615 			pause("nmpidx", 1);
616 
617 		/* ... and then for the cidx. */
618 		while (spg->pidx != spg->cidx)
619 			pause("nmcidx", 1);
620 
621 		free_nm_txq_hwq(vi, nm_txq);
622 	}
623 	for_each_nm_rxq(vi, i, nm_rxq) {
624 		free_nm_rxq_hwq(vi, nm_rxq);
625 	}
626 
627 	return (rc);
628 }
629 
630 static int
631 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
632 {
633 	struct ifnet *ifp = na->ifp;
634 	struct vi_info *vi = ifp->if_softc;
635 	struct adapter *sc = vi->pi->adapter;
636 	int rc;
637 
638 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
639 	if (rc != 0)
640 		return (rc);
641 	if (on)
642 		rc = cxgbe_netmap_on(sc, vi, ifp, na);
643 	else
644 		rc = cxgbe_netmap_off(sc, vi, ifp, na);
645 	end_synchronized_op(sc, 0);
646 
647 	return (rc);
648 }
649 
650 /* How many packets can a single type1 WR carry in n descriptors */
651 static inline int
652 ndesc_to_npkt(const int n)
653 {
654 
655 	MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
656 
657 	return (n * 2 - 1);
658 }
659 #define MAX_NPKT_IN_TYPE1_WR	(ndesc_to_npkt(SGE_MAX_WR_NDESC))
660 
661 /* Space (in descriptors) needed for a type1 WR that carries n packets */
662 static inline int
663 npkt_to_ndesc(const int n)
664 {
665 
666 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
667 
668 	return ((n + 2) / 2);
669 }
670 
671 /* Space (in 16B units) needed for a type1 WR that carries n packets */
672 static inline int
673 npkt_to_len16(const int n)
674 {
675 
676 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
677 
678 	return (n * 2 + 1);
679 }
680 
681 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
682 
683 static void
684 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
685 {
686 	int n;
687 	u_int db = nm_txq->doorbells;
688 
689 	MPASS(nm_txq->pidx != nm_txq->dbidx);
690 
691 	n = NMIDXDIFF(nm_txq, dbidx);
692 	if (n > 1)
693 		clrbit(&db, DOORBELL_WCWR);
694 	wmb();
695 
696 	switch (ffs(db) - 1) {
697 	case DOORBELL_UDB:
698 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
699 		break;
700 
701 	case DOORBELL_WCWR: {
702 		volatile uint64_t *dst, *src;
703 
704 		/*
705 		 * Queues whose 128B doorbell segment fits in the page do not
706 		 * use relative qid (udb_qid is always 0).  Only queues with
707 		 * doorbell segments can do WCWR.
708 		 */
709 		KASSERT(nm_txq->udb_qid == 0 && n == 1,
710 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
711 		    __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
712 
713 		dst = (volatile void *)((uintptr_t)nm_txq->udb +
714 		    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
715 		src = (void *)&nm_txq->desc[nm_txq->dbidx];
716 		while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
717 			*dst++ = *src++;
718 		wmb();
719 		break;
720 	}
721 
722 	case DOORBELL_UDBWC:
723 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
724 		wmb();
725 		break;
726 
727 	case DOORBELL_KDB:
728 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
729 		    V_QID(nm_txq->cntxt_id) | V_PIDX(n));
730 		break;
731 	}
732 	nm_txq->dbidx = nm_txq->pidx;
733 }
734 
735 int lazy_tx_credit_flush = 1;
736 
737 /*
738  * Write work requests to send 'npkt' frames and ring the doorbell to send them
739  * on their way.  No need to check for wraparound.
740  */
741 static void
742 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
743     struct netmap_kring *kring, int npkt, int npkt_remaining, int txcsum)
744 {
745 	struct netmap_ring *ring = kring->ring;
746 	struct netmap_slot *slot;
747 	const u_int lim = kring->nkr_num_slots - 1;
748 	struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
749 	uint16_t len;
750 	uint64_t ba;
751 	struct cpl_tx_pkt_core *cpl;
752 	struct ulptx_sgl *usgl;
753 	int i, n;
754 
755 	while (npkt) {
756 		n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
757 		len = 0;
758 
759 		wr = (void *)&nm_txq->desc[nm_txq->pidx];
760 		wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
761 		wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
762 		wr->npkt = n;
763 		wr->r3 = 0;
764 		wr->type = 1;
765 		cpl = (void *)(wr + 1);
766 
767 		for (i = 0; i < n; i++) {
768 			slot = &ring->slot[kring->nr_hwcur];
769 			PNMB(kring->na, slot, &ba);
770 			MPASS(ba != 0);
771 
772 			cpl->ctrl0 = nm_txq->cpl_ctrl0;
773 			cpl->pack = 0;
774 			cpl->len = htobe16(slot->len);
775 			/*
776 			 * netmap(4) says "netmap does not use features such as
777 			 * checksum offloading, TCP segmentation offloading,
778 			 * encryption, VLAN encapsulation/decapsulation, etc."
779 			 *
780 			 * So the ncxl interfaces have tx hardware checksumming
781 			 * disabled by default.  But you can override netmap by
782 			 * enabling IFCAP_TXCSUM on the interface manully.
783 			 */
784 			cpl->ctrl1 = txcsum ? 0 :
785 			    htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
786 
787 			usgl = (void *)(cpl + 1);
788 			usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
789 			    V_ULPTX_NSGE(1));
790 			usgl->len0 = htobe32(slot->len);
791 			usgl->addr0 = htobe64(ba);
792 
793 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
794 			cpl = (void *)(usgl + 1);
795 			MPASS(slot->len + len <= UINT16_MAX);
796 			len += slot->len;
797 			kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
798 		}
799 		wr->plen = htobe16(len);
800 
801 		npkt -= n;
802 		nm_txq->pidx += npkt_to_ndesc(n);
803 		MPASS(nm_txq->pidx <= nm_txq->sidx);
804 		if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
805 			/*
806 			 * This routine doesn't know how to write WRs that wrap
807 			 * around.  Make sure it wasn't asked to.
808 			 */
809 			MPASS(npkt == 0);
810 			nm_txq->pidx = 0;
811 		}
812 
813 		if (npkt == 0 && npkt_remaining == 0) {
814 			/* All done. */
815 			if (lazy_tx_credit_flush == 0) {
816 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
817 				    F_FW_WR_EQUIQ);
818 				nm_txq->equeqidx = nm_txq->pidx;
819 				nm_txq->equiqidx = nm_txq->pidx;
820 			}
821 			ring_nm_txq_db(sc, nm_txq);
822 			return;
823 		}
824 
825 		if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
826 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
827 			    F_FW_WR_EQUIQ);
828 			nm_txq->equeqidx = nm_txq->pidx;
829 			nm_txq->equiqidx = nm_txq->pidx;
830 		} else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
831 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
832 			nm_txq->equeqidx = nm_txq->pidx;
833 		}
834 		if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
835 			ring_nm_txq_db(sc, nm_txq);
836 	}
837 
838 	/* Will get called again. */
839 	MPASS(npkt_remaining);
840 }
841 
842 /* How many contiguous free descriptors starting at pidx */
843 static inline int
844 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
845 {
846 
847 	if (nm_txq->cidx > nm_txq->pidx)
848 		return (nm_txq->cidx - nm_txq->pidx - 1);
849 	else if (nm_txq->cidx > 0)
850 		return (nm_txq->sidx - nm_txq->pidx);
851 	else
852 		return (nm_txq->sidx - nm_txq->pidx - 1);
853 }
854 
855 static int
856 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
857 {
858 	struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
859 	uint16_t hw_cidx = spg->cidx;	/* snapshot */
860 	struct fw_eth_tx_pkts_wr *wr;
861 	int n = 0;
862 
863 	hw_cidx = be16toh(hw_cidx);
864 
865 	while (nm_txq->cidx != hw_cidx) {
866 		wr = (void *)&nm_txq->desc[nm_txq->cidx];
867 
868 		MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)));
869 		MPASS(wr->type == 1);
870 		MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
871 
872 		n += wr->npkt;
873 		nm_txq->cidx += npkt_to_ndesc(wr->npkt);
874 
875 		/*
876 		 * We never sent a WR that wrapped around so the credits coming
877 		 * back, WR by WR, should never cause the cidx to wrap around
878 		 * either.
879 		 */
880 		MPASS(nm_txq->cidx <= nm_txq->sidx);
881 		if (__predict_false(nm_txq->cidx == nm_txq->sidx))
882 			nm_txq->cidx = 0;
883 	}
884 
885 	return (n);
886 }
887 
888 static int
889 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
890 {
891 	struct netmap_adapter *na = kring->na;
892 	struct ifnet *ifp = na->ifp;
893 	struct vi_info *vi = ifp->if_softc;
894 	struct adapter *sc = vi->pi->adapter;
895 	struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_txq + kring->ring_id];
896 	const u_int head = kring->rhead;
897 	u_int reclaimed = 0;
898 	int n, d, npkt_remaining, ndesc_remaining, txcsum;
899 
900 	/*
901 	 * Tx was at kring->nr_hwcur last time around and now we need to advance
902 	 * to kring->rhead.  Note that the driver's pidx moves independent of
903 	 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
904 	 * between descriptors and frames isn't 1:1).
905 	 */
906 
907 	npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
908 	    kring->nkr_num_slots - kring->nr_hwcur + head;
909 	txcsum = ifp->if_capenable & (IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6);
910 	while (npkt_remaining) {
911 		reclaimed += reclaim_nm_tx_desc(nm_txq);
912 		ndesc_remaining = contiguous_ndesc_available(nm_txq);
913 		/* Can't run out of descriptors with packets still remaining */
914 		MPASS(ndesc_remaining > 0);
915 
916 		/* # of desc needed to tx all remaining packets */
917 		d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
918 		if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
919 			d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
920 
921 		if (d <= ndesc_remaining)
922 			n = npkt_remaining;
923 		else {
924 			/* Can't send all, calculate how many can be sent */
925 			n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
926 			    MAX_NPKT_IN_TYPE1_WR;
927 			if (ndesc_remaining % SGE_MAX_WR_NDESC)
928 				n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
929 		}
930 
931 		/* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
932 		npkt_remaining -= n;
933 		cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining, txcsum);
934 	}
935 	MPASS(npkt_remaining == 0);
936 	MPASS(kring->nr_hwcur == head);
937 	MPASS(nm_txq->dbidx == nm_txq->pidx);
938 
939 	/*
940 	 * Second part: reclaim buffers for completed transmissions.
941 	 */
942 	if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
943 		reclaimed += reclaim_nm_tx_desc(nm_txq);
944 		kring->nr_hwtail += reclaimed;
945 		if (kring->nr_hwtail >= kring->nkr_num_slots)
946 			kring->nr_hwtail -= kring->nkr_num_slots;
947 	}
948 
949 	return (0);
950 }
951 
952 static int
953 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
954 {
955 	struct netmap_adapter *na = kring->na;
956 	struct netmap_ring *ring = kring->ring;
957 	struct ifnet *ifp = na->ifp;
958 	struct vi_info *vi = ifp->if_softc;
959 	struct adapter *sc = vi->pi->adapter;
960 	struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_rxq + kring->ring_id];
961 	u_int const head = kring->rhead;
962 	u_int n;
963 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
964 
965 	if (black_hole)
966 		return (0);	/* No updates ever. */
967 
968 	if (netmap_no_pendintr || force_update) {
969 		kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
970 		kring->nr_kflags &= ~NKR_PENDINTR;
971 	}
972 
973 	/* Userspace done with buffers from kring->nr_hwcur to head */
974 	n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
975 	    kring->nkr_num_slots - kring->nr_hwcur + head;
976 	n &= ~7U;
977 	if (n > 0) {
978 		u_int fl_pidx = nm_rxq->fl_pidx;
979 		struct netmap_slot *slot = &ring->slot[fl_pidx];
980 		uint64_t ba;
981 		int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
982 
983 		/*
984 		 * We always deal with 8 buffers at a time.  We must have
985 		 * stopped at an 8B boundary (fl_pidx) last time around and we
986 		 * must have a multiple of 8B buffers to give to the freelist.
987 		 */
988 		MPASS((fl_pidx & 7) == 0);
989 		MPASS((n & 7) == 0);
990 
991 		IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
992 		IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx);
993 
994 		while (n > 0) {
995 			for (i = 0; i < 8; i++, fl_pidx++, slot++) {
996 				PNMB(na, slot, &ba);
997 				MPASS(ba != 0);
998 				nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
999 				slot->flags &= ~NS_BUF_CHANGED;
1000 				MPASS(fl_pidx <= nm_rxq->fl_sidx);
1001 			}
1002 			n -= 8;
1003 			if (fl_pidx == nm_rxq->fl_sidx) {
1004 				fl_pidx = 0;
1005 				slot = &ring->slot[0];
1006 			}
1007 			if (++dbinc == 8 && n >= 32) {
1008 				wmb();
1009 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
1010 				    nm_rxq->fl_db_val | V_PIDX(dbinc));
1011 				dbinc = 0;
1012 			}
1013 		}
1014 		MPASS(nm_rxq->fl_pidx == fl_pidx);
1015 
1016 		if (dbinc > 0) {
1017 			wmb();
1018 			t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
1019 			    nm_rxq->fl_db_val | V_PIDX(dbinc));
1020 		}
1021 	}
1022 
1023 	return (0);
1024 }
1025 
1026 static int
1027 ncxgbe_probe(device_t dev)
1028 {
1029 	char buf[128];
1030 	struct vi_info *vi = device_get_softc(dev);
1031 
1032 	snprintf(buf, sizeof(buf), "port %d netmap vi", vi->pi->port_id);
1033 	device_set_desc_copy(dev, buf);
1034 
1035 	return (BUS_PROBE_DEFAULT);
1036 }
1037 
1038 static int
1039 ncxgbe_attach(device_t dev)
1040 {
1041 	struct vi_info *vi;
1042 	struct port_info *pi;
1043 	struct adapter *sc;
1044 	struct netmap_adapter na;
1045 	struct ifnet *ifp;
1046 	int rc;
1047 
1048 	vi = device_get_softc(dev);
1049 	pi = vi->pi;
1050 	sc = pi->adapter;
1051 
1052 	/*
1053 	 * Allocate a virtual interface exclusively for netmap use.  Give it the
1054 	 * MAC address normally reserved for use by a TOE interface.  (The TOE
1055 	 * driver on FreeBSD doesn't use it).
1056 	 */
1057 	rc = t4_alloc_vi_func(sc, sc->mbox, pi->tx_chan, sc->pf, 0, 1,
1058 	    vi->hw_addr, &vi->rss_size, FW_VI_FUNC_OFLD, 0);
1059 	if (rc < 0) {
1060 		device_printf(dev, "unable to allocate netmap virtual "
1061 		    "interface for port %d: %d\n", pi->port_id, -rc);
1062 		return (-rc);
1063 	}
1064 	vi->viid = rc;
1065 	vi->xact_addr_filt = -1;
1066 	callout_init(&vi->tick, 1);
1067 
1068 	ifp = if_alloc(IFT_ETHER);
1069 	if (ifp == NULL) {
1070 		device_printf(dev, "Cannot allocate netmap ifnet\n");
1071 		return (ENOMEM);
1072 	}
1073 	vi->ifp = ifp;
1074 	ifp->if_softc = vi;
1075 
1076 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
1077 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
1078 
1079 	ifp->if_init = cxgbe_nm_init;
1080 	ifp->if_ioctl = cxgbe_nm_ioctl;
1081 	ifp->if_transmit = cxgbe_nm_transmit;
1082 	ifp->if_qflush = cxgbe_nm_qflush;
1083 	ifp->if_get_counter = cxgbe_get_counter;
1084 
1085 	/*
1086 	 * netmap(4) says "netmap does not use features such as checksum
1087 	 * offloading, TCP segmentation offloading, encryption, VLAN
1088 	 * encapsulation/decapsulation, etc."
1089 	 *
1090 	 * By default we comply with the statement above.  But we do declare the
1091 	 * ifnet capable of L3/L4 checksumming so that a user can override
1092 	 * netmap and have the hardware do the L3/L4 checksums.
1093 	 */
1094 	ifp->if_capabilities = IFCAP_HWCSUM | IFCAP_JUMBO_MTU |
1095 	    IFCAP_HWCSUM_IPV6;
1096 	ifp->if_capenable = 0;
1097 	ifp->if_hwassist = 0;
1098 
1099 	/* vi->media has already been setup by the caller */
1100 
1101 	ether_ifattach(ifp, vi->hw_addr);
1102 
1103 	device_printf(dev, "%d txq, %d rxq (netmap)\n", vi->ntxq, vi->nrxq);
1104 
1105 	vi_sysctls(vi);
1106 
1107 	/*
1108 	 * Register with netmap in the kernel.
1109 	 */
1110 	bzero(&na, sizeof(na));
1111 
1112 	na.ifp = ifp;
1113 	na.na_flags = NAF_BDG_MAYSLEEP;
1114 
1115 	/* Netmap doesn't know about the space reserved for the status page. */
1116 	na.num_tx_desc = vi->qsize_txq - spg_len / EQ_ESIZE;
1117 
1118 	/*
1119 	 * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
1120 	 * num_rx_desc is based on the number of buffers that can be held in the
1121 	 * freelist, and not the number of entries in the iq.  (These two are
1122 	 * not exactly the same due to the space taken up by the status page).
1123 	 */
1124 	na.num_rx_desc = (vi->qsize_rxq / 8) * 8;
1125 	na.nm_txsync = cxgbe_netmap_txsync;
1126 	na.nm_rxsync = cxgbe_netmap_rxsync;
1127 	na.nm_register = cxgbe_netmap_reg;
1128 	na.num_tx_rings = vi->ntxq;
1129 	na.num_rx_rings = vi->nrxq;
1130 	netmap_attach(&na);	/* This adds IFCAP_NETMAP to if_capabilities */
1131 
1132 	return (0);
1133 }
1134 
1135 static int
1136 ncxgbe_detach(device_t dev)
1137 {
1138 	struct vi_info *vi;
1139 	struct adapter *sc;
1140 
1141 	vi = device_get_softc(dev);
1142 	sc = vi->pi->adapter;
1143 
1144 	doom_vi(sc, vi);
1145 
1146 	netmap_detach(vi->ifp);
1147 	ether_ifdetach(vi->ifp);
1148 	cxgbe_nm_uninit_synchronized(vi);
1149 	callout_drain(&vi->tick);
1150 	vi_full_uninit(vi);
1151 	ifmedia_removeall(&vi->media);
1152 	if_free(vi->ifp);
1153 	vi->ifp = NULL;
1154 	t4_free_vi(sc, sc->mbox, sc->pf, 0, vi->viid);
1155 
1156 	end_synchronized_op(sc, 0);
1157 
1158 	return (0);
1159 }
1160 
1161 static void
1162 handle_nm_fw6_msg(struct adapter *sc, struct ifnet *ifp,
1163     const struct cpl_fw6_msg *cpl)
1164 {
1165 	const struct cpl_sge_egr_update *egr;
1166 	uint32_t oq;
1167 	struct sge_nm_txq *nm_txq;
1168 
1169 	if (cpl->type != FW_TYPE_RSSCPL && cpl->type != FW6_TYPE_RSSCPL)
1170 		panic("%s: FW_TYPE 0x%x on nm_rxq.", __func__, cpl->type);
1171 
1172 	/* data[0] is RSS header */
1173 	egr = (const void *)&cpl->data[1];
1174 	oq = be32toh(egr->opcode_qid);
1175 	MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
1176 	nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
1177 
1178 	netmap_tx_irq(ifp, nm_txq->nid);
1179 }
1180 
1181 void
1182 t4_nm_intr(void *arg)
1183 {
1184 	struct sge_nm_rxq *nm_rxq = arg;
1185 	struct vi_info *vi = nm_rxq->vi;
1186 	struct adapter *sc = vi->pi->adapter;
1187 	struct ifnet *ifp = vi->ifp;
1188 	struct netmap_adapter *na = NA(ifp);
1189 	struct netmap_kring *kring = &na->rx_rings[nm_rxq->nid];
1190 	struct netmap_ring *ring = kring->ring;
1191 	struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
1192 	uint32_t lq;
1193 	u_int n = 0, work = 0;
1194 	uint8_t opcode;
1195 	uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
1196 	u_int fl_credits = fl_cidx & 7;
1197 
1198 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
1199 
1200 		rmb();
1201 
1202 		lq = be32toh(d->rsp.pldbuflen_qid);
1203 		opcode = d->rss.opcode;
1204 
1205 		switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
1206 		case X_RSPD_TYPE_FLBUF:
1207 			if (black_hole != 2) {
1208 				/* No buffer packing so new buf every time */
1209 				MPASS(lq & F_RSPD_NEWBUF);
1210 			}
1211 
1212 			/* fall through */
1213 
1214 		case X_RSPD_TYPE_CPL:
1215 			MPASS(opcode < NUM_CPL_CMDS);
1216 
1217 			switch (opcode) {
1218 			case CPL_FW4_MSG:
1219 			case CPL_FW6_MSG:
1220 				handle_nm_fw6_msg(sc, ifp,
1221 				    (const void *)&d->cpl[0]);
1222 				break;
1223 			case CPL_RX_PKT:
1224 				ring->slot[fl_cidx].len = G_RSPD_LEN(lq) - fl_pktshift;
1225 				ring->slot[fl_cidx].flags = kring->nkr_slot_flags;
1226 				fl_cidx += (lq & F_RSPD_NEWBUF) ? 1 : 0;
1227 				fl_credits += (lq & F_RSPD_NEWBUF) ? 1 : 0;
1228 				if (__predict_false(fl_cidx == nm_rxq->fl_sidx))
1229 					fl_cidx = 0;
1230 				break;
1231 			default:
1232 				panic("%s: unexpected opcode 0x%x on nm_rxq %p",
1233 				    __func__, opcode, nm_rxq);
1234 			}
1235 			break;
1236 
1237 		case X_RSPD_TYPE_INTR:
1238 			/* Not equipped to handle forwarded interrupts. */
1239 			panic("%s: netmap queue received interrupt for iq %u\n",
1240 			    __func__, lq);
1241 
1242 		default:
1243 			panic("%s: illegal response type %d on nm_rxq %p",
1244 			    __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
1245 		}
1246 
1247 		d++;
1248 		if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
1249 			nm_rxq->iq_cidx = 0;
1250 			d = &nm_rxq->iq_desc[0];
1251 			nm_rxq->iq_gen ^= F_RSPD_GEN;
1252 		}
1253 
1254 		if (__predict_false(++n == rx_ndesc)) {
1255 			atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1256 			if (black_hole && fl_credits >= 8) {
1257 				fl_credits /= 8;
1258 				IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
1259 				    nm_rxq->fl_sidx);
1260 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
1261 				    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1262 				fl_credits = fl_cidx & 7;
1263 			} else if (!black_hole) {
1264 				netmap_rx_irq(ifp, nm_rxq->nid, &work);
1265 				MPASS(work != 0);
1266 			}
1267 			t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
1268 			    V_CIDXINC(n) | V_INGRESSQID(nm_rxq->iq_cntxt_id) |
1269 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1270 			n = 0;
1271 		}
1272 	}
1273 
1274 	atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1275 	if (black_hole) {
1276 		fl_credits /= 8;
1277 		IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1278 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
1279 		    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1280 	} else
1281 		netmap_rx_irq(ifp, nm_rxq->nid, &work);
1282 
1283 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(n) |
1284 	    V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1285 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1286 }
1287 
1288 static devclass_t ncxgbe_devclass, ncxl_devclass;
1289 
1290 DRIVER_MODULE(ncxgbe, cxgbe, ncxgbe_driver, ncxgbe_devclass, 0, 0);
1291 MODULE_VERSION(ncxgbe, 1);
1292 
1293 DRIVER_MODULE(ncxl, cxl, ncxl_driver, ncxl_devclass, 0, 0);
1294 MODULE_VERSION(ncxl, 1);
1295 #endif
1296