xref: /freebsd/sys/dev/cxgbe/t4_netmap.c (revision 45951bf0a941ff2dcb072dd55a71bd8a6dfe890d)
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 
60 /*
61  * 0 = normal netmap rx
62  * 1 = black hole
63  * 2 = supermassive black hole (buffer packing enabled)
64  */
65 int black_hole = 0;
66 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_black_hole, CTLFLAG_RDTUN, &black_hole, 0,
67     "Sink incoming packets.");
68 
69 int rx_ndesc = 256;
70 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_rx_ndesc, CTLFLAG_RWTUN,
71     &rx_ndesc, 0, "# of rx descriptors after which the hw cidx is updated.");
72 
73 int rx_nframes = 64;
74 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_rx_nframes, CTLFLAG_RWTUN,
75     &rx_nframes, 0, "max # of frames received before waking up netmap rx.");
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 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_cong_drop, CTLFLAG_RDTUN,
89     &nm_cong_drop, 0,
90     "Congestion control for netmap rx queues (0 = backpressure, 1 = drop");
91 
92 int starve_fl = 0;
93 SYSCTL_INT(_hw_cxgbe, OID_AUTO, starve_fl, CTLFLAG_RWTUN,
94     &starve_fl, 0, "Don't ring fl db for netmap rx queues.");
95 
96 /*
97  * Try to process tx credits in bulk.  This may cause a delay in the return of
98  * tx credits and is suitable for bursty or non-stop tx only.
99  */
100 int lazy_tx_credit_flush = 1;
101 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lazy_tx_credit_flush, CTLFLAG_RWTUN,
102     &lazy_tx_credit_flush, 0, "lazy credit flush for netmap tx queues.");
103 
104 /*
105  * Split the netmap rx queues into two groups that populate separate halves of
106  * the RSS indirection table.  This allows filters with hashmask to steer to a
107  * particular group of queues.
108  */
109 static int nm_split_rss = 0;
110 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_split_rss, CTLFLAG_RWTUN,
111     &nm_split_rss, 0, "Split the netmap rx queues into two groups.");
112 
113 static int
114 alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int cong)
115 {
116 	int rc, cntxt_id, i;
117 	__be32 v;
118 	struct adapter *sc = vi->pi->adapter;
119 	struct sge_params *sp = &sc->params.sge;
120 	struct netmap_adapter *na = NA(vi->ifp);
121 	struct fw_iq_cmd c;
122 
123 	MPASS(na != NULL);
124 	MPASS(nm_rxq->iq_desc != NULL);
125 	MPASS(nm_rxq->fl_desc != NULL);
126 
127 	bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE);
128 	bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + sp->spg_len);
129 
130 	bzero(&c, sizeof(c));
131 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
132 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
133 	    V_FW_IQ_CMD_VFN(0));
134 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
135 	    FW_LEN16(c));
136 	MPASS(!forwarding_intr_to_fwq(sc));
137 	KASSERT(nm_rxq->intr_idx < sc->intr_count,
138 	    ("%s: invalid direct intr_idx %d", __func__, nm_rxq->intr_idx));
139 	v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
140 	c.type_to_iqandstindex = htobe32(v |
141 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
142 	    V_FW_IQ_CMD_VIID(vi->viid) |
143 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
144 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(vi->pi->tx_chan) |
145 	    F_FW_IQ_CMD_IQGTSMODE |
146 	    V_FW_IQ_CMD_IQINTCNTTHRESH(0) |
147 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
148 	c.iqsize = htobe16(vi->qsize_rxq);
149 	c.iqaddr = htobe64(nm_rxq->iq_ba);
150 	if (cong >= 0) {
151 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN |
152 		    V_FW_IQ_CMD_FL0CNGCHMAP(cong) | F_FW_IQ_CMD_FL0CONGCIF |
153 		    F_FW_IQ_CMD_FL0CONGEN);
154 	}
155 	c.iqns_to_fl0congen |=
156 	    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
157 		F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
158 		(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
159 		(black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0));
160 	c.fl0dcaen_to_fl0cidxfthresh =
161 	    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
162 		X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
163 		V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
164 		X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
165 	c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE);
166 	c.fl0addr = htobe64(nm_rxq->fl_ba);
167 
168 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
169 	if (rc != 0) {
170 		device_printf(sc->dev,
171 		    "failed to create netmap ingress queue: %d\n", rc);
172 		return (rc);
173 	}
174 
175 	nm_rxq->iq_cidx = 0;
176 	MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE);
177 	nm_rxq->iq_gen = F_RSPD_GEN;
178 	nm_rxq->iq_cntxt_id = be16toh(c.iqid);
179 	nm_rxq->iq_abs_id = be16toh(c.physiqid);
180 	cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
181 	if (cntxt_id >= sc->sge.niq) {
182 		panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
183 		    __func__, cntxt_id, sc->sge.niq - 1);
184 	}
185 	sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
186 
187 	nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
188 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
189 	MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
190 	cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
191 	if (cntxt_id >= sc->sge.neq) {
192 		panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
193 		    __func__, cntxt_id, sc->sge.neq - 1);
194 	}
195 	sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
196 
197 	nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) |
198 	    sc->chip_params->sge_fl_db;
199 
200 	if (chip_id(sc) >= CHELSIO_T5 && cong >= 0) {
201 		uint32_t param, val;
202 
203 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
204 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
205 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
206 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
207 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
208 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
209 		if (cong == 0)
210 			val = 1 << 19;
211 		else {
212 			val = 2 << 19;
213 			for (i = 0; i < 4; i++) {
214 				if (cong & (1 << i))
215 					val |= 1 << (i << 2);
216 			}
217 		}
218 
219 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
220 		if (rc != 0) {
221 			/* report error but carry on */
222 			device_printf(sc->dev,
223 			    "failed to set congestion manager context for "
224 			    "ingress queue %d: %d\n", nm_rxq->iq_cntxt_id, rc);
225 		}
226 	}
227 
228 	t4_write_reg(sc, sc->sge_gts_reg,
229 	    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
230 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
231 
232 	return (rc);
233 }
234 
235 static int
236 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
237 {
238 	struct adapter *sc = vi->pi->adapter;
239 	int rc;
240 
241 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
242 	    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
243 	if (rc != 0)
244 		device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
245 		    __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
246 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
247 	return (rc);
248 }
249 
250 static int
251 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
252 {
253 	int rc, cntxt_id;
254 	size_t len;
255 	struct adapter *sc = vi->pi->adapter;
256 	struct netmap_adapter *na = NA(vi->ifp);
257 	struct fw_eq_eth_cmd c;
258 
259 	MPASS(na != NULL);
260 	MPASS(nm_txq->desc != NULL);
261 
262 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
263 	bzero(nm_txq->desc, len);
264 
265 	bzero(&c, sizeof(c));
266 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
267 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
268 	    V_FW_EQ_ETH_CMD_VFN(0));
269 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
270 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
271 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
272 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
273 	c.fetchszm_to_iqid =
274 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
275 		V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
276 		V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
277 	c.dcaen_to_eqsize =
278 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
279 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
280 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
281 		V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
282 	c.eqaddr = htobe64(nm_txq->ba);
283 
284 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
285 	if (rc != 0) {
286 		device_printf(vi->dev,
287 		    "failed to create netmap egress queue: %d\n", rc);
288 		return (rc);
289 	}
290 
291 	nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
292 	cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
293 	if (cntxt_id >= sc->sge.neq)
294 	    panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
295 		cntxt_id, sc->sge.neq - 1);
296 	sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
297 
298 	nm_txq->pidx = nm_txq->cidx = 0;
299 	MPASS(nm_txq->sidx == na->num_tx_desc);
300 	nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
301 
302 	nm_txq->doorbells = sc->doorbells;
303 	if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
304 	    isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
305 	    isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
306 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
307 		uint32_t mask = (1 << s_qpp) - 1;
308 		volatile uint8_t *udb;
309 
310 		udb = sc->udbs_base + UDBS_DB_OFFSET;
311 		udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
312 		nm_txq->udb_qid = nm_txq->cntxt_id & mask;
313 		if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
314 	    		clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
315 		else {
316 			udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
317 			nm_txq->udb_qid = 0;
318 		}
319 		nm_txq->udb = (volatile void *)udb;
320 	}
321 
322 	return (rc);
323 }
324 
325 static int
326 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
327 {
328 	struct adapter *sc = vi->pi->adapter;
329 	int rc;
330 
331 	rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
332 	if (rc != 0)
333 		device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
334 		    nm_txq->cntxt_id, rc);
335 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
336 	return (rc);
337 }
338 
339 static int
340 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
341     struct netmap_adapter *na)
342 {
343 	struct netmap_slot *slot;
344 	struct netmap_kring *kring;
345 	struct sge_nm_rxq *nm_rxq;
346 	struct sge_nm_txq *nm_txq;
347 	int rc, i, j, hwidx, defq, nrssq;
348 	struct hw_buf_info *hwb;
349 
350 	ASSERT_SYNCHRONIZED_OP(sc);
351 
352 	if ((vi->flags & VI_INIT_DONE) == 0 ||
353 	    (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
354 		return (EAGAIN);
355 
356 	hwb = &sc->sge.hw_buf_info[0];
357 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
358 		if (hwb->size == NETMAP_BUF_SIZE(na))
359 			break;
360 	}
361 	if (i >= SGE_FLBUF_SIZES) {
362 		if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
363 		    NETMAP_BUF_SIZE(na));
364 		return (ENXIO);
365 	}
366 	hwidx = i;
367 
368 	/* Must set caps before calling netmap_reset */
369 	nm_set_native_flags(na);
370 
371 	for_each_nm_rxq(vi, i, nm_rxq) {
372 		kring = na->rx_rings[nm_rxq->nid];
373 		if (!nm_kring_pending_on(kring) ||
374 		    nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID)
375 			continue;
376 
377 		alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop));
378 		nm_rxq->fl_hwidx = hwidx;
379 		slot = netmap_reset(na, NR_RX, i, 0);
380 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
381 
382 		/* We deal with 8 bufs at a time */
383 		MPASS((na->num_rx_desc & 7) == 0);
384 		MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
385 		for (j = 0; j < nm_rxq->fl_sidx; j++) {
386 			uint64_t ba;
387 
388 			PNMB(na, &slot[j], &ba);
389 			MPASS(ba != 0);
390 			nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
391 		}
392 		j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
393 		MPASS((j & 7) == 0);
394 		j /= 8;	/* driver pidx to hardware pidx */
395 		wmb();
396 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
397 		    nm_rxq->fl_db_val | V_PIDX(j));
398 
399 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_OFF, NM_ON);
400 	}
401 
402 	for_each_nm_txq(vi, i, nm_txq) {
403 		kring = na->tx_rings[nm_txq->nid];
404 		if (!nm_kring_pending_on(kring) ||
405 		    nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID)
406 			continue;
407 
408 		alloc_nm_txq_hwq(vi, nm_txq);
409 		slot = netmap_reset(na, NR_TX, i, 0);
410 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
411 	}
412 
413 	if (vi->nm_rss == NULL) {
414 		vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
415 		    M_ZERO | M_WAITOK);
416 	}
417 
418 	MPASS(vi->nnmrxq > 0);
419 	if (nm_split_rss == 0 || vi->nnmrxq == 1) {
420 		for (i = 0; i < vi->rss_size;) {
421 			for_each_nm_rxq(vi, j, nm_rxq) {
422 				vi->nm_rss[i++] = nm_rxq->iq_abs_id;
423 				if (i == vi->rss_size)
424 					break;
425 			}
426 		}
427 		defq = vi->nm_rss[0];
428 	} else {
429 		/* We have multiple queues and we want to split the table. */
430 		MPASS(nm_split_rss != 0);
431 		MPASS(vi->nnmrxq > 1);
432 
433 		nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq];
434 		nrssq = vi->nnmrxq;
435 		if (vi->nnmrxq & 1) {
436 			/*
437 			 * Odd number of queues. The first rxq is designated the
438 			 * default queue, the rest are split evenly.
439 			 */
440 			defq = nm_rxq->iq_abs_id;
441 			nm_rxq++;
442 			nrssq--;
443 		} else {
444 			/*
445 			 * Even number of queues split into two halves.  The
446 			 * first rxq in one of the halves is designated the
447 			 * default queue.
448 			 */
449 #if 1
450 			/* First rxq in the first half. */
451 			defq = nm_rxq->iq_abs_id;
452 #else
453 			/* First rxq in the second half. */
454 			defq = nm_rxq[vi->nnmrxq / 2].iq_abs_id;
455 #endif
456 		}
457 
458 		i = 0;
459 		while (i < vi->rss_size / 2) {
460 			for (j = 0; j < nrssq / 2; j++) {
461 				vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
462 				if (i == vi->rss_size / 2)
463 					break;
464 			}
465 		}
466 		while (i < vi->rss_size) {
467 			for (j = nrssq / 2; j < nrssq; j++) {
468 				vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
469 				if (i == vi->rss_size)
470 					break;
471 			}
472 		}
473 	}
474 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
475 	    vi->nm_rss, vi->rss_size);
476 	if (rc != 0)
477 		if_printf(ifp, "netmap rss_config failed: %d\n", rc);
478 
479 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
480 	if (rc != 0)
481 		if_printf(ifp, "netmap rss hash/defaultq config failed: %d\n", rc);
482 
483 	return (rc);
484 }
485 
486 static int
487 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
488     struct netmap_adapter *na)
489 {
490 	struct netmap_kring *kring;
491 	int rc, i;
492 	struct sge_nm_txq *nm_txq;
493 	struct sge_nm_rxq *nm_rxq;
494 
495 	ASSERT_SYNCHRONIZED_OP(sc);
496 
497 	if (!nm_netmap_on(na))
498 		return (0);
499 
500 	if ((vi->flags & VI_INIT_DONE) == 0)
501 		return (0);
502 
503 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
504 	    vi->rss, vi->rss_size);
505 	if (rc != 0)
506 		if_printf(ifp, "failed to restore RSS config: %d\n", rc);
507 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, vi->rss[0], 0, 0);
508 	if (rc != 0)
509 		if_printf(ifp, "failed to restore RSS hash/defaultq: %d\n", rc);
510 	nm_clear_native_flags(na);
511 
512 	for_each_nm_txq(vi, i, nm_txq) {
513 		struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
514 
515 		kring = na->tx_rings[nm_txq->nid];
516 		if (!nm_kring_pending_off(kring) ||
517 		    nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID)
518 			continue;
519 
520 		/* Wait for hw pidx to catch up ... */
521 		while (be16toh(nm_txq->pidx) != spg->pidx)
522 			pause("nmpidx", 1);
523 
524 		/* ... and then for the cidx. */
525 		while (spg->pidx != spg->cidx)
526 			pause("nmcidx", 1);
527 
528 		free_nm_txq_hwq(vi, nm_txq);
529 	}
530 	for_each_nm_rxq(vi, i, nm_rxq) {
531 		kring = na->rx_rings[nm_rxq->nid];
532 		if (!nm_kring_pending_off(kring) ||
533 		    nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID)
534 			continue;
535 
536 		while (!atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_OFF))
537 			pause("nmst", 1);
538 
539 		free_nm_rxq_hwq(vi, nm_rxq);
540 	}
541 
542 	return (rc);
543 }
544 
545 static int
546 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
547 {
548 	struct ifnet *ifp = na->ifp;
549 	struct vi_info *vi = ifp->if_softc;
550 	struct adapter *sc = vi->pi->adapter;
551 	int rc;
552 
553 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
554 	if (rc != 0)
555 		return (rc);
556 	if (on)
557 		rc = cxgbe_netmap_on(sc, vi, ifp, na);
558 	else
559 		rc = cxgbe_netmap_off(sc, vi, ifp, na);
560 	end_synchronized_op(sc, 0);
561 
562 	return (rc);
563 }
564 
565 /* How many packets can a single type1 WR carry in n descriptors */
566 static inline int
567 ndesc_to_npkt(const int n)
568 {
569 
570 	MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
571 
572 	return (n * 2 - 1);
573 }
574 #define MAX_NPKT_IN_TYPE1_WR	(ndesc_to_npkt(SGE_MAX_WR_NDESC))
575 
576 /*
577  * Space (in descriptors) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
578  * carries n packets
579  */
580 static inline int
581 npkt_to_ndesc(const int n)
582 {
583 
584 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
585 
586 	return ((n + 2) / 2);
587 }
588 
589 /*
590  * Space (in 16B units) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
591  * carries n packets
592  */
593 static inline int
594 npkt_to_len16(const int n)
595 {
596 
597 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
598 
599 	return (n * 2 + 1);
600 }
601 
602 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
603 
604 static void
605 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
606 {
607 	int n;
608 	u_int db = nm_txq->doorbells;
609 
610 	MPASS(nm_txq->pidx != nm_txq->dbidx);
611 
612 	n = NMIDXDIFF(nm_txq, dbidx);
613 	if (n > 1)
614 		clrbit(&db, DOORBELL_WCWR);
615 	wmb();
616 
617 	switch (ffs(db) - 1) {
618 	case DOORBELL_UDB:
619 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
620 		break;
621 
622 	case DOORBELL_WCWR: {
623 		volatile uint64_t *dst, *src;
624 
625 		/*
626 		 * Queues whose 128B doorbell segment fits in the page do not
627 		 * use relative qid (udb_qid is always 0).  Only queues with
628 		 * doorbell segments can do WCWR.
629 		 */
630 		KASSERT(nm_txq->udb_qid == 0 && n == 1,
631 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
632 		    __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
633 
634 		dst = (volatile void *)((uintptr_t)nm_txq->udb +
635 		    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
636 		src = (void *)&nm_txq->desc[nm_txq->dbidx];
637 		while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
638 			*dst++ = *src++;
639 		wmb();
640 		break;
641 	}
642 
643 	case DOORBELL_UDBWC:
644 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
645 		wmb();
646 		break;
647 
648 	case DOORBELL_KDB:
649 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
650 		    V_QID(nm_txq->cntxt_id) | V_PIDX(n));
651 		break;
652 	}
653 	nm_txq->dbidx = nm_txq->pidx;
654 }
655 
656 /*
657  * Write work requests to send 'npkt' frames and ring the doorbell to send them
658  * on their way.  No need to check for wraparound.
659  */
660 static void
661 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
662     struct netmap_kring *kring, int npkt, int npkt_remaining)
663 {
664 	struct netmap_ring *ring = kring->ring;
665 	struct netmap_slot *slot;
666 	const u_int lim = kring->nkr_num_slots - 1;
667 	struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
668 	uint16_t len;
669 	uint64_t ba;
670 	struct cpl_tx_pkt_core *cpl;
671 	struct ulptx_sgl *usgl;
672 	int i, n;
673 
674 	while (npkt) {
675 		n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
676 		len = 0;
677 
678 		wr = (void *)&nm_txq->desc[nm_txq->pidx];
679 		wr->op_pkd = nm_txq->op_pkd;
680 		wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
681 		wr->npkt = n;
682 		wr->r3 = 0;
683 		wr->type = 1;
684 		cpl = (void *)(wr + 1);
685 
686 		for (i = 0; i < n; i++) {
687 			slot = &ring->slot[kring->nr_hwcur];
688 			PNMB(kring->na, slot, &ba);
689 			MPASS(ba != 0);
690 
691 			cpl->ctrl0 = nm_txq->cpl_ctrl0;
692 			cpl->pack = 0;
693 			cpl->len = htobe16(slot->len);
694 			/*
695 			 * netmap(4) says "netmap does not use features such as
696 			 * checksum offloading, TCP segmentation offloading,
697 			 * encryption, VLAN encapsulation/decapsulation, etc."
698 			 */
699 			cpl->ctrl1 = htobe64(F_TXPKT_IPCSUM_DIS |
700 			    F_TXPKT_L4CSUM_DIS);
701 
702 			usgl = (void *)(cpl + 1);
703 			usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
704 			    V_ULPTX_NSGE(1));
705 			usgl->len0 = htobe32(slot->len);
706 			usgl->addr0 = htobe64(ba);
707 
708 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
709 			cpl = (void *)(usgl + 1);
710 			MPASS(slot->len + len <= UINT16_MAX);
711 			len += slot->len;
712 			kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
713 		}
714 		wr->plen = htobe16(len);
715 
716 		npkt -= n;
717 		nm_txq->pidx += npkt_to_ndesc(n);
718 		MPASS(nm_txq->pidx <= nm_txq->sidx);
719 		if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
720 			/*
721 			 * This routine doesn't know how to write WRs that wrap
722 			 * around.  Make sure it wasn't asked to.
723 			 */
724 			MPASS(npkt == 0);
725 			nm_txq->pidx = 0;
726 		}
727 
728 		if (npkt == 0 && npkt_remaining == 0) {
729 			/* All done. */
730 			if (lazy_tx_credit_flush == 0) {
731 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
732 				    F_FW_WR_EQUIQ);
733 				nm_txq->equeqidx = nm_txq->pidx;
734 				nm_txq->equiqidx = nm_txq->pidx;
735 			}
736 			ring_nm_txq_db(sc, nm_txq);
737 			return;
738 		}
739 
740 		if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
741 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
742 			    F_FW_WR_EQUIQ);
743 			nm_txq->equeqidx = nm_txq->pidx;
744 			nm_txq->equiqidx = nm_txq->pidx;
745 		} else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
746 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
747 			nm_txq->equeqidx = nm_txq->pidx;
748 		}
749 		if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
750 			ring_nm_txq_db(sc, nm_txq);
751 	}
752 
753 	/* Will get called again. */
754 	MPASS(npkt_remaining);
755 }
756 
757 /* How many contiguous free descriptors starting at pidx */
758 static inline int
759 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
760 {
761 
762 	if (nm_txq->cidx > nm_txq->pidx)
763 		return (nm_txq->cidx - nm_txq->pidx - 1);
764 	else if (nm_txq->cidx > 0)
765 		return (nm_txq->sidx - nm_txq->pidx);
766 	else
767 		return (nm_txq->sidx - nm_txq->pidx - 1);
768 }
769 
770 static int
771 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
772 {
773 	struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
774 	uint16_t hw_cidx = spg->cidx;	/* snapshot */
775 	struct fw_eth_tx_pkts_wr *wr;
776 	int n = 0;
777 
778 	hw_cidx = be16toh(hw_cidx);
779 
780 	while (nm_txq->cidx != hw_cidx) {
781 		wr = (void *)&nm_txq->desc[nm_txq->cidx];
782 
783 		MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)) ||
784 		    wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR)));
785 		MPASS(wr->type == 1);
786 		MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
787 
788 		n += wr->npkt;
789 		nm_txq->cidx += npkt_to_ndesc(wr->npkt);
790 
791 		/*
792 		 * We never sent a WR that wrapped around so the credits coming
793 		 * back, WR by WR, should never cause the cidx to wrap around
794 		 * either.
795 		 */
796 		MPASS(nm_txq->cidx <= nm_txq->sidx);
797 		if (__predict_false(nm_txq->cidx == nm_txq->sidx))
798 			nm_txq->cidx = 0;
799 	}
800 
801 	return (n);
802 }
803 
804 static int
805 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
806 {
807 	struct netmap_adapter *na = kring->na;
808 	struct ifnet *ifp = na->ifp;
809 	struct vi_info *vi = ifp->if_softc;
810 	struct adapter *sc = vi->pi->adapter;
811 	struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
812 	const u_int head = kring->rhead;
813 	u_int reclaimed = 0;
814 	int n, d, npkt_remaining, ndesc_remaining;
815 
816 	/*
817 	 * Tx was at kring->nr_hwcur last time around and now we need to advance
818 	 * to kring->rhead.  Note that the driver's pidx moves independent of
819 	 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
820 	 * between descriptors and frames isn't 1:1).
821 	 */
822 
823 	npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
824 	    kring->nkr_num_slots - kring->nr_hwcur + head;
825 	while (npkt_remaining) {
826 		reclaimed += reclaim_nm_tx_desc(nm_txq);
827 		ndesc_remaining = contiguous_ndesc_available(nm_txq);
828 		/* Can't run out of descriptors with packets still remaining */
829 		MPASS(ndesc_remaining > 0);
830 
831 		/* # of desc needed to tx all remaining packets */
832 		d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
833 		if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
834 			d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
835 
836 		if (d <= ndesc_remaining)
837 			n = npkt_remaining;
838 		else {
839 			/* Can't send all, calculate how many can be sent */
840 			n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
841 			    MAX_NPKT_IN_TYPE1_WR;
842 			if (ndesc_remaining % SGE_MAX_WR_NDESC)
843 				n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
844 		}
845 
846 		/* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
847 		npkt_remaining -= n;
848 		cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining);
849 	}
850 	MPASS(npkt_remaining == 0);
851 	MPASS(kring->nr_hwcur == head);
852 	MPASS(nm_txq->dbidx == nm_txq->pidx);
853 
854 	/*
855 	 * Second part: reclaim buffers for completed transmissions.
856 	 */
857 	if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
858 		reclaimed += reclaim_nm_tx_desc(nm_txq);
859 		kring->nr_hwtail += reclaimed;
860 		if (kring->nr_hwtail >= kring->nkr_num_slots)
861 			kring->nr_hwtail -= kring->nkr_num_slots;
862 	}
863 
864 	return (0);
865 }
866 
867 static int
868 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
869 {
870 	struct netmap_adapter *na = kring->na;
871 	struct netmap_ring *ring = kring->ring;
872 	struct ifnet *ifp = na->ifp;
873 	struct vi_info *vi = ifp->if_softc;
874 	struct adapter *sc = vi->pi->adapter;
875 	struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
876 	u_int const head = kring->rhead;
877 	u_int n;
878 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
879 
880 	if (black_hole)
881 		return (0);	/* No updates ever. */
882 
883 	if (netmap_no_pendintr || force_update) {
884 		kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
885 		kring->nr_kflags &= ~NKR_PENDINTR;
886 	}
887 
888 	if (nm_rxq->fl_db_saved > 0 && starve_fl == 0) {
889 		wmb();
890 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
891 		    nm_rxq->fl_db_val | V_PIDX(nm_rxq->fl_db_saved));
892 		nm_rxq->fl_db_saved = 0;
893 	}
894 
895 	/* Userspace done with buffers from kring->nr_hwcur to head */
896 	n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
897 	    kring->nkr_num_slots - kring->nr_hwcur + head;
898 	n &= ~7U;
899 	if (n > 0) {
900 		u_int fl_pidx = nm_rxq->fl_pidx;
901 		struct netmap_slot *slot = &ring->slot[fl_pidx];
902 		uint64_t ba;
903 		int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
904 
905 		/*
906 		 * We always deal with 8 buffers at a time.  We must have
907 		 * stopped at an 8B boundary (fl_pidx) last time around and we
908 		 * must have a multiple of 8B buffers to give to the freelist.
909 		 */
910 		MPASS((fl_pidx & 7) == 0);
911 		MPASS((n & 7) == 0);
912 
913 		IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
914 		IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx);
915 
916 		while (n > 0) {
917 			for (i = 0; i < 8; i++, fl_pidx++, slot++) {
918 				PNMB(na, slot, &ba);
919 				MPASS(ba != 0);
920 				nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
921 				slot->flags &= ~NS_BUF_CHANGED;
922 				MPASS(fl_pidx <= nm_rxq->fl_sidx);
923 			}
924 			n -= 8;
925 			if (fl_pidx == nm_rxq->fl_sidx) {
926 				fl_pidx = 0;
927 				slot = &ring->slot[0];
928 			}
929 			if (++dbinc == 8 && n >= 32) {
930 				wmb();
931 				if (starve_fl)
932 					nm_rxq->fl_db_saved += dbinc;
933 				else {
934 					t4_write_reg(sc, sc->sge_kdoorbell_reg,
935 					    nm_rxq->fl_db_val | V_PIDX(dbinc));
936 				}
937 				dbinc = 0;
938 			}
939 		}
940 		MPASS(nm_rxq->fl_pidx == fl_pidx);
941 
942 		if (dbinc > 0) {
943 			wmb();
944 			if (starve_fl)
945 				nm_rxq->fl_db_saved += dbinc;
946 			else {
947 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
948 				    nm_rxq->fl_db_val | V_PIDX(dbinc));
949 			}
950 		}
951 	}
952 
953 	return (0);
954 }
955 
956 void
957 cxgbe_nm_attach(struct vi_info *vi)
958 {
959 	struct port_info *pi;
960 	struct adapter *sc;
961 	struct netmap_adapter na;
962 
963 	MPASS(vi->nnmrxq > 0);
964 	MPASS(vi->ifp != NULL);
965 
966 	pi = vi->pi;
967 	sc = pi->adapter;
968 
969 	bzero(&na, sizeof(na));
970 
971 	na.ifp = vi->ifp;
972 	na.na_flags = NAF_BDG_MAYSLEEP;
973 
974 	/* Netmap doesn't know about the space reserved for the status page. */
975 	na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
976 
977 	/*
978 	 * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
979 	 * num_rx_desc is based on the number of buffers that can be held in the
980 	 * freelist, and not the number of entries in the iq.  (These two are
981 	 * not exactly the same due to the space taken up by the status page).
982 	 */
983 	na.num_rx_desc = rounddown(vi->qsize_rxq, 8);
984 	na.nm_txsync = cxgbe_netmap_txsync;
985 	na.nm_rxsync = cxgbe_netmap_rxsync;
986 	na.nm_register = cxgbe_netmap_reg;
987 	na.num_tx_rings = vi->nnmtxq;
988 	na.num_rx_rings = vi->nnmrxq;
989 	netmap_attach(&na);	/* This adds IFCAP_NETMAP to if_capabilities */
990 }
991 
992 void
993 cxgbe_nm_detach(struct vi_info *vi)
994 {
995 
996 	MPASS(vi->nnmrxq > 0);
997 	MPASS(vi->ifp != NULL);
998 
999 	netmap_detach(vi->ifp);
1000 }
1001 
1002 static inline const void *
1003 unwrap_nm_fw6_msg(const struct cpl_fw6_msg *cpl)
1004 {
1005 
1006 	MPASS(cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL);
1007 
1008 	/* data[0] is RSS header */
1009 	return (&cpl->data[1]);
1010 }
1011 
1012 static void
1013 handle_nm_sge_egr_update(struct adapter *sc, struct ifnet *ifp,
1014     const struct cpl_sge_egr_update *egr)
1015 {
1016 	uint32_t oq;
1017 	struct sge_nm_txq *nm_txq;
1018 
1019 	oq = be32toh(egr->opcode_qid);
1020 	MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
1021 	nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
1022 
1023 	netmap_tx_irq(ifp, nm_txq->nid);
1024 }
1025 
1026 void
1027 service_nm_rxq(struct sge_nm_rxq *nm_rxq)
1028 {
1029 	struct vi_info *vi = nm_rxq->vi;
1030 	struct adapter *sc = vi->pi->adapter;
1031 	struct ifnet *ifp = vi->ifp;
1032 	struct netmap_adapter *na = NA(ifp);
1033 	struct netmap_kring *kring = na->rx_rings[nm_rxq->nid];
1034 	struct netmap_ring *ring = kring->ring;
1035 	struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
1036 	const void *cpl;
1037 	uint32_t lq;
1038 	u_int work = 0;
1039 	uint8_t opcode;
1040 	uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
1041 	u_int fl_credits = fl_cidx & 7;
1042 	u_int ndesc = 0;	/* desc processed since last cidx update */
1043 	u_int nframes = 0;	/* frames processed since last netmap wakeup */
1044 
1045 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
1046 
1047 		rmb();
1048 
1049 		lq = be32toh(d->rsp.pldbuflen_qid);
1050 		opcode = d->rss.opcode;
1051 		cpl = &d->cpl[0];
1052 
1053 		switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
1054 		case X_RSPD_TYPE_FLBUF:
1055 
1056 			/* fall through */
1057 
1058 		case X_RSPD_TYPE_CPL:
1059 			MPASS(opcode < NUM_CPL_CMDS);
1060 
1061 			switch (opcode) {
1062 			case CPL_FW4_MSG:
1063 			case CPL_FW6_MSG:
1064 				cpl = unwrap_nm_fw6_msg(cpl);
1065 				/* fall through */
1066 			case CPL_SGE_EGR_UPDATE:
1067 				handle_nm_sge_egr_update(sc, ifp, cpl);
1068 				break;
1069 			case CPL_RX_PKT:
1070 				ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
1071 				    sc->params.sge.fl_pktshift;
1072 				ring->slot[fl_cidx].flags = 0;
1073 				nframes++;
1074 				if (!(lq & F_RSPD_NEWBUF)) {
1075 					MPASS(black_hole == 2);
1076 					break;
1077 				}
1078 				fl_credits++;
1079 				if (__predict_false(++fl_cidx == nm_rxq->fl_sidx))
1080 					fl_cidx = 0;
1081 				break;
1082 			default:
1083 				panic("%s: unexpected opcode 0x%x on nm_rxq %p",
1084 				    __func__, opcode, nm_rxq);
1085 			}
1086 			break;
1087 
1088 		case X_RSPD_TYPE_INTR:
1089 			/* Not equipped to handle forwarded interrupts. */
1090 			panic("%s: netmap queue received interrupt for iq %u\n",
1091 			    __func__, lq);
1092 
1093 		default:
1094 			panic("%s: illegal response type %d on nm_rxq %p",
1095 			    __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
1096 		}
1097 
1098 		d++;
1099 		if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
1100 			nm_rxq->iq_cidx = 0;
1101 			d = &nm_rxq->iq_desc[0];
1102 			nm_rxq->iq_gen ^= F_RSPD_GEN;
1103 		}
1104 
1105 		if (__predict_false(++nframes == rx_nframes) && !black_hole) {
1106 			atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1107 			netmap_rx_irq(ifp, nm_rxq->nid, &work);
1108 			nframes = 0;
1109 		}
1110 
1111 		if (__predict_false(++ndesc == rx_ndesc)) {
1112 			if (black_hole && fl_credits >= 8) {
1113 				fl_credits /= 8;
1114 				IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
1115 				    nm_rxq->fl_sidx);
1116 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1117 				    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1118 				fl_credits = fl_cidx & 7;
1119 			}
1120 			t4_write_reg(sc, sc->sge_gts_reg,
1121 			    V_CIDXINC(ndesc) |
1122 			    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
1123 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1124 			ndesc = 0;
1125 		}
1126 	}
1127 
1128 	atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1129 	if (black_hole) {
1130 		fl_credits /= 8;
1131 		IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1132 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1133 		    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1134 	} else if (nframes > 0)
1135 		netmap_rx_irq(ifp, nm_rxq->nid, &work);
1136 
1137     	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndesc) |
1138 	    V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1139 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1140 }
1141 #endif
1142