xref: /freebsd/sys/dev/cxgbe/t4_netmap.c (revision 31d62a73c2e6ac0ff413a7a17700ffc7dce254ef)
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 SYSCTL_NODE(_hw, OID_AUTO, cxgbe, CTLFLAG_RD, 0, "cxgbe netmap parameters");
61 
62 /*
63  * 0 = normal netmap rx
64  * 1 = black hole
65  * 2 = supermassive black hole (buffer packing enabled)
66  */
67 int black_hole = 0;
68 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_black_hole, CTLFLAG_RDTUN, &black_hole, 0,
69     "Sink incoming packets.");
70 
71 int rx_ndesc = 256;
72 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_rx_ndesc, CTLFLAG_RWTUN,
73     &rx_ndesc, 0, "# of rx descriptors after which the hw cidx is updated.");
74 
75 int rx_nframes = 64;
76 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_rx_nframes, CTLFLAG_RWTUN,
77     &rx_nframes, 0, "max # of frames received before waking up netmap rx.");
78 
79 int holdoff_tmr_idx = 2;
80 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_holdoff_tmr_idx, CTLFLAG_RWTUN,
81     &holdoff_tmr_idx, 0, "Holdoff timer index for netmap rx queues.");
82 
83 /*
84  * Congestion drops.
85  * -1: no congestion feedback (not recommended).
86  *  0: backpressure the channel instead of dropping packets right away.
87  *  1: no backpressure, drop packets for the congested queue immediately.
88  */
89 static int nm_cong_drop = 1;
90 TUNABLE_INT("hw.cxgbe.nm_cong_drop", &nm_cong_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 static int
105 alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int cong)
106 {
107 	int rc, cntxt_id, i;
108 	__be32 v;
109 	struct adapter *sc = vi->pi->adapter;
110 	struct sge_params *sp = &sc->params.sge;
111 	struct netmap_adapter *na = NA(vi->ifp);
112 	struct fw_iq_cmd c;
113 
114 	MPASS(na != NULL);
115 	MPASS(nm_rxq->iq_desc != NULL);
116 	MPASS(nm_rxq->fl_desc != NULL);
117 
118 	bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE);
119 	bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + sp->spg_len);
120 
121 	bzero(&c, sizeof(c));
122 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
123 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
124 	    V_FW_IQ_CMD_VFN(0));
125 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
126 	    FW_LEN16(c));
127 	MPASS(!forwarding_intr_to_fwq(sc));
128 	KASSERT(nm_rxq->intr_idx < sc->intr_count,
129 	    ("%s: invalid direct intr_idx %d", __func__, nm_rxq->intr_idx));
130 	v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
131 	c.type_to_iqandstindex = htobe32(v |
132 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
133 	    V_FW_IQ_CMD_VIID(vi->viid) |
134 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
135 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(vi->pi->tx_chan) |
136 	    F_FW_IQ_CMD_IQGTSMODE |
137 	    V_FW_IQ_CMD_IQINTCNTTHRESH(0) |
138 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
139 	c.iqsize = htobe16(vi->qsize_rxq);
140 	c.iqaddr = htobe64(nm_rxq->iq_ba);
141 	if (cong >= 0) {
142 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN |
143 		    V_FW_IQ_CMD_FL0CNGCHMAP(cong) | F_FW_IQ_CMD_FL0CONGCIF |
144 		    F_FW_IQ_CMD_FL0CONGEN);
145 	}
146 	c.iqns_to_fl0congen |=
147 	    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
148 		F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
149 		(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
150 		(black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0));
151 	c.fl0dcaen_to_fl0cidxfthresh =
152 	    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
153 		X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
154 		V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
155 		X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
156 	c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE);
157 	c.fl0addr = htobe64(nm_rxq->fl_ba);
158 
159 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
160 	if (rc != 0) {
161 		device_printf(sc->dev,
162 		    "failed to create netmap ingress queue: %d\n", rc);
163 		return (rc);
164 	}
165 
166 	nm_rxq->iq_cidx = 0;
167 	MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE);
168 	nm_rxq->iq_gen = F_RSPD_GEN;
169 	nm_rxq->iq_cntxt_id = be16toh(c.iqid);
170 	nm_rxq->iq_abs_id = be16toh(c.physiqid);
171 	cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
172 	if (cntxt_id >= sc->sge.niq) {
173 		panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
174 		    __func__, cntxt_id, sc->sge.niq - 1);
175 	}
176 	sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
177 
178 	nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
179 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
180 	MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
181 	cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
182 	if (cntxt_id >= sc->sge.neq) {
183 		panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
184 		    __func__, cntxt_id, sc->sge.neq - 1);
185 	}
186 	sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
187 
188 	nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) |
189 	    sc->chip_params->sge_fl_db;
190 
191 	if (chip_id(sc) >= CHELSIO_T5 && cong >= 0) {
192 		uint32_t param, val;
193 
194 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
195 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
196 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
197 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
198 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
199 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
200 		if (cong == 0)
201 			val = 1 << 19;
202 		else {
203 			val = 2 << 19;
204 			for (i = 0; i < 4; i++) {
205 				if (cong & (1 << i))
206 					val |= 1 << (i << 2);
207 			}
208 		}
209 
210 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
211 		if (rc != 0) {
212 			/* report error but carry on */
213 			device_printf(sc->dev,
214 			    "failed to set congestion manager context for "
215 			    "ingress queue %d: %d\n", nm_rxq->iq_cntxt_id, rc);
216 		}
217 	}
218 
219 	t4_write_reg(sc, sc->sge_gts_reg,
220 	    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
221 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
222 
223 	return (rc);
224 }
225 
226 static int
227 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
228 {
229 	struct adapter *sc = vi->pi->adapter;
230 	int rc;
231 
232 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
233 	    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
234 	if (rc != 0)
235 		device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
236 		    __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
237 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
238 	return (rc);
239 }
240 
241 static int
242 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
243 {
244 	int rc, cntxt_id;
245 	size_t len;
246 	struct adapter *sc = vi->pi->adapter;
247 	struct netmap_adapter *na = NA(vi->ifp);
248 	struct fw_eq_eth_cmd c;
249 
250 	MPASS(na != NULL);
251 	MPASS(nm_txq->desc != NULL);
252 
253 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
254 	bzero(nm_txq->desc, len);
255 
256 	bzero(&c, sizeof(c));
257 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
258 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
259 	    V_FW_EQ_ETH_CMD_VFN(0));
260 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
261 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
262 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
263 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
264 	c.fetchszm_to_iqid =
265 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
266 		V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
267 		V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
268 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
269 		      V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
270 		      V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
271 	c.eqaddr = htobe64(nm_txq->ba);
272 
273 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
274 	if (rc != 0) {
275 		device_printf(vi->dev,
276 		    "failed to create netmap egress queue: %d\n", rc);
277 		return (rc);
278 	}
279 
280 	nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
281 	cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
282 	if (cntxt_id >= sc->sge.neq)
283 	    panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
284 		cntxt_id, sc->sge.neq - 1);
285 	sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
286 
287 	nm_txq->pidx = nm_txq->cidx = 0;
288 	MPASS(nm_txq->sidx == na->num_tx_desc);
289 	nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
290 
291 	nm_txq->doorbells = sc->doorbells;
292 	if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
293 	    isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
294 	    isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
295 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
296 		uint32_t mask = (1 << s_qpp) - 1;
297 		volatile uint8_t *udb;
298 
299 		udb = sc->udbs_base + UDBS_DB_OFFSET;
300 		udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
301 		nm_txq->udb_qid = nm_txq->cntxt_id & mask;
302 		if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
303 	    		clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
304 		else {
305 			udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
306 			nm_txq->udb_qid = 0;
307 		}
308 		nm_txq->udb = (volatile void *)udb;
309 	}
310 
311 	return (rc);
312 }
313 
314 static int
315 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
316 {
317 	struct adapter *sc = vi->pi->adapter;
318 	int rc;
319 
320 	rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
321 	if (rc != 0)
322 		device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
323 		    nm_txq->cntxt_id, rc);
324 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
325 	return (rc);
326 }
327 
328 static int
329 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
330     struct netmap_adapter *na)
331 {
332 	struct netmap_slot *slot;
333 	struct netmap_kring *kring;
334 	struct sge_nm_rxq *nm_rxq;
335 	struct sge_nm_txq *nm_txq;
336 	int rc, i, j, hwidx;
337 	struct hw_buf_info *hwb;
338 
339 	ASSERT_SYNCHRONIZED_OP(sc);
340 
341 	if ((vi->flags & VI_INIT_DONE) == 0 ||
342 	    (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
343 		return (EAGAIN);
344 
345 	hwb = &sc->sge.hw_buf_info[0];
346 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
347 		if (hwb->size == NETMAP_BUF_SIZE(na))
348 			break;
349 	}
350 	if (i >= SGE_FLBUF_SIZES) {
351 		if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
352 		    NETMAP_BUF_SIZE(na));
353 		return (ENXIO);
354 	}
355 	hwidx = i;
356 
357 	/* Must set caps before calling netmap_reset */
358 	nm_set_native_flags(na);
359 
360 	for_each_nm_rxq(vi, i, nm_rxq) {
361 		kring = na->rx_rings[nm_rxq->nid];
362 		if (!nm_kring_pending_on(kring) ||
363 		    nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID)
364 			continue;
365 
366 		alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop));
367 		nm_rxq->fl_hwidx = hwidx;
368 		slot = netmap_reset(na, NR_RX, i, 0);
369 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
370 
371 		/* We deal with 8 bufs at a time */
372 		MPASS((na->num_rx_desc & 7) == 0);
373 		MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
374 		for (j = 0; j < nm_rxq->fl_sidx; j++) {
375 			uint64_t ba;
376 
377 			PNMB(na, &slot[j], &ba);
378 			MPASS(ba != 0);
379 			nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
380 		}
381 		j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
382 		MPASS((j & 7) == 0);
383 		j /= 8;	/* driver pidx to hardware pidx */
384 		wmb();
385 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
386 		    nm_rxq->fl_db_val | V_PIDX(j));
387 
388 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_OFF, NM_ON);
389 	}
390 
391 	for_each_nm_txq(vi, i, nm_txq) {
392 		kring = na->tx_rings[nm_txq->nid];
393 		if (!nm_kring_pending_on(kring) ||
394 		    nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID)
395 			continue;
396 
397 		alloc_nm_txq_hwq(vi, nm_txq);
398 		slot = netmap_reset(na, NR_TX, i, 0);
399 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
400 	}
401 
402 	if (vi->nm_rss == NULL) {
403 		vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
404 		    M_ZERO | M_WAITOK);
405 	}
406 	for (i = 0; i < vi->rss_size;) {
407 		for_each_nm_rxq(vi, j, nm_rxq) {
408 			vi->nm_rss[i++] = nm_rxq->iq_abs_id;
409 			if (i == vi->rss_size)
410 				break;
411 		}
412 	}
413 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
414 	    vi->nm_rss, vi->rss_size);
415 	if (rc != 0)
416 		if_printf(ifp, "netmap rss_config failed: %d\n", rc);
417 
418 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen,
419 	    vi->nm_rss[0], 0, 0);
420 	if (rc != 0)
421 		if_printf(ifp, "netmap rss hash/defaultq config failed: %d\n", rc);
422 
423 	return (rc);
424 }
425 
426 static int
427 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
428     struct netmap_adapter *na)
429 {
430 	struct netmap_kring *kring;
431 	int rc, i;
432 	struct sge_nm_txq *nm_txq;
433 	struct sge_nm_rxq *nm_rxq;
434 
435 	ASSERT_SYNCHRONIZED_OP(sc);
436 
437 	if ((vi->flags & VI_INIT_DONE) == 0)
438 		return (0);
439 
440 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
441 	    vi->rss, vi->rss_size);
442 	if (rc != 0)
443 		if_printf(ifp, "failed to restore RSS config: %d\n", rc);
444 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, vi->rss[0], 0, 0);
445 	if (rc != 0)
446 		if_printf(ifp, "failed to restore RSS hash/defaultq: %d\n", rc);
447 	nm_clear_native_flags(na);
448 
449 	for_each_nm_txq(vi, i, nm_txq) {
450 		struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
451 
452 		kring = na->tx_rings[nm_txq->nid];
453 		if (!nm_kring_pending_off(kring) ||
454 		    nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID)
455 			continue;
456 
457 		/* Wait for hw pidx to catch up ... */
458 		while (be16toh(nm_txq->pidx) != spg->pidx)
459 			pause("nmpidx", 1);
460 
461 		/* ... and then for the cidx. */
462 		while (spg->pidx != spg->cidx)
463 			pause("nmcidx", 1);
464 
465 		free_nm_txq_hwq(vi, nm_txq);
466 	}
467 	for_each_nm_rxq(vi, i, nm_rxq) {
468 		kring = na->rx_rings[nm_rxq->nid];
469 		if (!nm_kring_pending_off(kring) ||
470 		    nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID)
471 			continue;
472 
473 		while (!atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_OFF))
474 			pause("nmst", 1);
475 
476 		free_nm_rxq_hwq(vi, nm_rxq);
477 	}
478 
479 	return (rc);
480 }
481 
482 static int
483 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
484 {
485 	struct ifnet *ifp = na->ifp;
486 	struct vi_info *vi = ifp->if_softc;
487 	struct adapter *sc = vi->pi->adapter;
488 	int rc;
489 
490 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
491 	if (rc != 0)
492 		return (rc);
493 	if (on)
494 		rc = cxgbe_netmap_on(sc, vi, ifp, na);
495 	else
496 		rc = cxgbe_netmap_off(sc, vi, ifp, na);
497 	end_synchronized_op(sc, 0);
498 
499 	return (rc);
500 }
501 
502 /* How many packets can a single type1 WR carry in n descriptors */
503 static inline int
504 ndesc_to_npkt(const int n)
505 {
506 
507 	MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
508 
509 	return (n * 2 - 1);
510 }
511 #define MAX_NPKT_IN_TYPE1_WR	(ndesc_to_npkt(SGE_MAX_WR_NDESC))
512 
513 /* Space (in descriptors) needed for a type1 WR that carries n packets */
514 static inline int
515 npkt_to_ndesc(const int n)
516 {
517 
518 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
519 
520 	return ((n + 2) / 2);
521 }
522 
523 /* Space (in 16B units) needed for a type1 WR that carries n packets */
524 static inline int
525 npkt_to_len16(const int n)
526 {
527 
528 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
529 
530 	return (n * 2 + 1);
531 }
532 
533 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
534 
535 static void
536 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
537 {
538 	int n;
539 	u_int db = nm_txq->doorbells;
540 
541 	MPASS(nm_txq->pidx != nm_txq->dbidx);
542 
543 	n = NMIDXDIFF(nm_txq, dbidx);
544 	if (n > 1)
545 		clrbit(&db, DOORBELL_WCWR);
546 	wmb();
547 
548 	switch (ffs(db) - 1) {
549 	case DOORBELL_UDB:
550 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
551 		break;
552 
553 	case DOORBELL_WCWR: {
554 		volatile uint64_t *dst, *src;
555 
556 		/*
557 		 * Queues whose 128B doorbell segment fits in the page do not
558 		 * use relative qid (udb_qid is always 0).  Only queues with
559 		 * doorbell segments can do WCWR.
560 		 */
561 		KASSERT(nm_txq->udb_qid == 0 && n == 1,
562 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
563 		    __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
564 
565 		dst = (volatile void *)((uintptr_t)nm_txq->udb +
566 		    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
567 		src = (void *)&nm_txq->desc[nm_txq->dbidx];
568 		while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
569 			*dst++ = *src++;
570 		wmb();
571 		break;
572 	}
573 
574 	case DOORBELL_UDBWC:
575 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
576 		wmb();
577 		break;
578 
579 	case DOORBELL_KDB:
580 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
581 		    V_QID(nm_txq->cntxt_id) | V_PIDX(n));
582 		break;
583 	}
584 	nm_txq->dbidx = nm_txq->pidx;
585 }
586 
587 /*
588  * Write work requests to send 'npkt' frames and ring the doorbell to send them
589  * on their way.  No need to check for wraparound.
590  */
591 static void
592 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
593     struct netmap_kring *kring, int npkt, int npkt_remaining, int txcsum)
594 {
595 	struct netmap_ring *ring = kring->ring;
596 	struct netmap_slot *slot;
597 	const u_int lim = kring->nkr_num_slots - 1;
598 	struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
599 	uint16_t len;
600 	uint64_t ba;
601 	struct cpl_tx_pkt_core *cpl;
602 	struct ulptx_sgl *usgl;
603 	int i, n;
604 
605 	while (npkt) {
606 		n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
607 		len = 0;
608 
609 		wr = (void *)&nm_txq->desc[nm_txq->pidx];
610 		wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
611 		wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
612 		wr->npkt = n;
613 		wr->r3 = 0;
614 		wr->type = 1;
615 		cpl = (void *)(wr + 1);
616 
617 		for (i = 0; i < n; i++) {
618 			slot = &ring->slot[kring->nr_hwcur];
619 			PNMB(kring->na, slot, &ba);
620 			MPASS(ba != 0);
621 
622 			cpl->ctrl0 = nm_txq->cpl_ctrl0;
623 			cpl->pack = 0;
624 			cpl->len = htobe16(slot->len);
625 			/*
626 			 * netmap(4) says "netmap does not use features such as
627 			 * checksum offloading, TCP segmentation offloading,
628 			 * encryption, VLAN encapsulation/decapsulation, etc."
629 			 *
630 			 * So the ncxl interfaces have tx hardware checksumming
631 			 * disabled by default.  But you can override netmap by
632 			 * enabling IFCAP_TXCSUM on the interface manully.
633 			 */
634 			cpl->ctrl1 = txcsum ? 0 :
635 			    htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
636 
637 			usgl = (void *)(cpl + 1);
638 			usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
639 			    V_ULPTX_NSGE(1));
640 			usgl->len0 = htobe32(slot->len);
641 			usgl->addr0 = htobe64(ba);
642 
643 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
644 			cpl = (void *)(usgl + 1);
645 			MPASS(slot->len + len <= UINT16_MAX);
646 			len += slot->len;
647 			kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
648 		}
649 		wr->plen = htobe16(len);
650 
651 		npkt -= n;
652 		nm_txq->pidx += npkt_to_ndesc(n);
653 		MPASS(nm_txq->pidx <= nm_txq->sidx);
654 		if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
655 			/*
656 			 * This routine doesn't know how to write WRs that wrap
657 			 * around.  Make sure it wasn't asked to.
658 			 */
659 			MPASS(npkt == 0);
660 			nm_txq->pidx = 0;
661 		}
662 
663 		if (npkt == 0 && npkt_remaining == 0) {
664 			/* All done. */
665 			if (lazy_tx_credit_flush == 0) {
666 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
667 				    F_FW_WR_EQUIQ);
668 				nm_txq->equeqidx = nm_txq->pidx;
669 				nm_txq->equiqidx = nm_txq->pidx;
670 			}
671 			ring_nm_txq_db(sc, nm_txq);
672 			return;
673 		}
674 
675 		if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
676 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
677 			    F_FW_WR_EQUIQ);
678 			nm_txq->equeqidx = nm_txq->pidx;
679 			nm_txq->equiqidx = nm_txq->pidx;
680 		} else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
681 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
682 			nm_txq->equeqidx = nm_txq->pidx;
683 		}
684 		if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
685 			ring_nm_txq_db(sc, nm_txq);
686 	}
687 
688 	/* Will get called again. */
689 	MPASS(npkt_remaining);
690 }
691 
692 /* How many contiguous free descriptors starting at pidx */
693 static inline int
694 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
695 {
696 
697 	if (nm_txq->cidx > nm_txq->pidx)
698 		return (nm_txq->cidx - nm_txq->pidx - 1);
699 	else if (nm_txq->cidx > 0)
700 		return (nm_txq->sidx - nm_txq->pidx);
701 	else
702 		return (nm_txq->sidx - nm_txq->pidx - 1);
703 }
704 
705 static int
706 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
707 {
708 	struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
709 	uint16_t hw_cidx = spg->cidx;	/* snapshot */
710 	struct fw_eth_tx_pkts_wr *wr;
711 	int n = 0;
712 
713 	hw_cidx = be16toh(hw_cidx);
714 
715 	while (nm_txq->cidx != hw_cidx) {
716 		wr = (void *)&nm_txq->desc[nm_txq->cidx];
717 
718 		MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)));
719 		MPASS(wr->type == 1);
720 		MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
721 
722 		n += wr->npkt;
723 		nm_txq->cidx += npkt_to_ndesc(wr->npkt);
724 
725 		/*
726 		 * We never sent a WR that wrapped around so the credits coming
727 		 * back, WR by WR, should never cause the cidx to wrap around
728 		 * either.
729 		 */
730 		MPASS(nm_txq->cidx <= nm_txq->sidx);
731 		if (__predict_false(nm_txq->cidx == nm_txq->sidx))
732 			nm_txq->cidx = 0;
733 	}
734 
735 	return (n);
736 }
737 
738 static int
739 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
740 {
741 	struct netmap_adapter *na = kring->na;
742 	struct ifnet *ifp = na->ifp;
743 	struct vi_info *vi = ifp->if_softc;
744 	struct adapter *sc = vi->pi->adapter;
745 	struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
746 	const u_int head = kring->rhead;
747 	u_int reclaimed = 0;
748 	int n, d, npkt_remaining, ndesc_remaining, txcsum;
749 
750 	/*
751 	 * Tx was at kring->nr_hwcur last time around and now we need to advance
752 	 * to kring->rhead.  Note that the driver's pidx moves independent of
753 	 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
754 	 * between descriptors and frames isn't 1:1).
755 	 */
756 
757 	npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
758 	    kring->nkr_num_slots - kring->nr_hwcur + head;
759 	txcsum = ifp->if_capenable & (IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6);
760 	while (npkt_remaining) {
761 		reclaimed += reclaim_nm_tx_desc(nm_txq);
762 		ndesc_remaining = contiguous_ndesc_available(nm_txq);
763 		/* Can't run out of descriptors with packets still remaining */
764 		MPASS(ndesc_remaining > 0);
765 
766 		/* # of desc needed to tx all remaining packets */
767 		d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
768 		if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
769 			d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
770 
771 		if (d <= ndesc_remaining)
772 			n = npkt_remaining;
773 		else {
774 			/* Can't send all, calculate how many can be sent */
775 			n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
776 			    MAX_NPKT_IN_TYPE1_WR;
777 			if (ndesc_remaining % SGE_MAX_WR_NDESC)
778 				n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
779 		}
780 
781 		/* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
782 		npkt_remaining -= n;
783 		cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining, txcsum);
784 	}
785 	MPASS(npkt_remaining == 0);
786 	MPASS(kring->nr_hwcur == head);
787 	MPASS(nm_txq->dbidx == nm_txq->pidx);
788 
789 	/*
790 	 * Second part: reclaim buffers for completed transmissions.
791 	 */
792 	if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
793 		reclaimed += reclaim_nm_tx_desc(nm_txq);
794 		kring->nr_hwtail += reclaimed;
795 		if (kring->nr_hwtail >= kring->nkr_num_slots)
796 			kring->nr_hwtail -= kring->nkr_num_slots;
797 	}
798 
799 	return (0);
800 }
801 
802 static int
803 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
804 {
805 	struct netmap_adapter *na = kring->na;
806 	struct netmap_ring *ring = kring->ring;
807 	struct ifnet *ifp = na->ifp;
808 	struct vi_info *vi = ifp->if_softc;
809 	struct adapter *sc = vi->pi->adapter;
810 	struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
811 	u_int const head = kring->rhead;
812 	u_int n;
813 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
814 
815 	if (black_hole)
816 		return (0);	/* No updates ever. */
817 
818 	if (netmap_no_pendintr || force_update) {
819 		kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
820 		kring->nr_kflags &= ~NKR_PENDINTR;
821 	}
822 
823 	if (nm_rxq->fl_db_saved > 0 && starve_fl == 0) {
824 		wmb();
825 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
826 		    nm_rxq->fl_db_val | V_PIDX(nm_rxq->fl_db_saved));
827 		nm_rxq->fl_db_saved = 0;
828 	}
829 
830 	/* Userspace done with buffers from kring->nr_hwcur to head */
831 	n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
832 	    kring->nkr_num_slots - kring->nr_hwcur + head;
833 	n &= ~7U;
834 	if (n > 0) {
835 		u_int fl_pidx = nm_rxq->fl_pidx;
836 		struct netmap_slot *slot = &ring->slot[fl_pidx];
837 		uint64_t ba;
838 		int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
839 
840 		/*
841 		 * We always deal with 8 buffers at a time.  We must have
842 		 * stopped at an 8B boundary (fl_pidx) last time around and we
843 		 * must have a multiple of 8B buffers to give to the freelist.
844 		 */
845 		MPASS((fl_pidx & 7) == 0);
846 		MPASS((n & 7) == 0);
847 
848 		IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
849 		IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx);
850 
851 		while (n > 0) {
852 			for (i = 0; i < 8; i++, fl_pidx++, slot++) {
853 				PNMB(na, slot, &ba);
854 				MPASS(ba != 0);
855 				nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
856 				slot->flags &= ~NS_BUF_CHANGED;
857 				MPASS(fl_pidx <= nm_rxq->fl_sidx);
858 			}
859 			n -= 8;
860 			if (fl_pidx == nm_rxq->fl_sidx) {
861 				fl_pidx = 0;
862 				slot = &ring->slot[0];
863 			}
864 			if (++dbinc == 8 && n >= 32) {
865 				wmb();
866 				if (starve_fl)
867 					nm_rxq->fl_db_saved += dbinc;
868 				else {
869 					t4_write_reg(sc, sc->sge_kdoorbell_reg,
870 					    nm_rxq->fl_db_val | V_PIDX(dbinc));
871 				}
872 				dbinc = 0;
873 			}
874 		}
875 		MPASS(nm_rxq->fl_pidx == fl_pidx);
876 
877 		if (dbinc > 0) {
878 			wmb();
879 			if (starve_fl)
880 				nm_rxq->fl_db_saved += dbinc;
881 			else {
882 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
883 				    nm_rxq->fl_db_val | V_PIDX(dbinc));
884 			}
885 		}
886 	}
887 
888 	return (0);
889 }
890 
891 void
892 cxgbe_nm_attach(struct vi_info *vi)
893 {
894 	struct port_info *pi;
895 	struct adapter *sc;
896 	struct netmap_adapter na;
897 
898 	MPASS(vi->nnmrxq > 0);
899 	MPASS(vi->ifp != NULL);
900 
901 	pi = vi->pi;
902 	sc = pi->adapter;
903 
904 	bzero(&na, sizeof(na));
905 
906 	na.ifp = vi->ifp;
907 	na.na_flags = NAF_BDG_MAYSLEEP;
908 
909 	/* Netmap doesn't know about the space reserved for the status page. */
910 	na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
911 
912 	/*
913 	 * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
914 	 * num_rx_desc is based on the number of buffers that can be held in the
915 	 * freelist, and not the number of entries in the iq.  (These two are
916 	 * not exactly the same due to the space taken up by the status page).
917 	 */
918 	na.num_rx_desc = rounddown(vi->qsize_rxq, 8);
919 	na.nm_txsync = cxgbe_netmap_txsync;
920 	na.nm_rxsync = cxgbe_netmap_rxsync;
921 	na.nm_register = cxgbe_netmap_reg;
922 	na.num_tx_rings = vi->nnmtxq;
923 	na.num_rx_rings = vi->nnmrxq;
924 	netmap_attach(&na);
925 }
926 
927 void
928 cxgbe_nm_detach(struct vi_info *vi)
929 {
930 
931 	MPASS(vi->nnmrxq > 0);
932 	MPASS(vi->ifp != NULL);
933 
934 	netmap_detach(vi->ifp);
935 }
936 
937 static inline const void *
938 unwrap_nm_fw6_msg(const struct cpl_fw6_msg *cpl)
939 {
940 
941 	MPASS(cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL);
942 
943 	/* data[0] is RSS header */
944 	return (&cpl->data[1]);
945 }
946 
947 static void
948 handle_nm_sge_egr_update(struct adapter *sc, struct ifnet *ifp,
949     const struct cpl_sge_egr_update *egr)
950 {
951 	uint32_t oq;
952 	struct sge_nm_txq *nm_txq;
953 
954 	oq = be32toh(egr->opcode_qid);
955 	MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
956 	nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
957 
958 	netmap_tx_irq(ifp, nm_txq->nid);
959 }
960 
961 void
962 service_nm_rxq(struct sge_nm_rxq *nm_rxq)
963 {
964 	struct vi_info *vi = nm_rxq->vi;
965 	struct adapter *sc = vi->pi->adapter;
966 	struct ifnet *ifp = vi->ifp;
967 	struct netmap_adapter *na = NA(ifp);
968 	struct netmap_kring *kring = na->rx_rings[nm_rxq->nid];
969 	struct netmap_ring *ring = kring->ring;
970 	struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
971 	const void *cpl;
972 	uint32_t lq;
973 	u_int work = 0;
974 	uint8_t opcode;
975 	uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
976 	u_int fl_credits = fl_cidx & 7;
977 	u_int ndesc = 0;	/* desc processed since last cidx update */
978 	u_int nframes = 0;	/* frames processed since last netmap wakeup */
979 
980 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
981 
982 		rmb();
983 
984 		lq = be32toh(d->rsp.pldbuflen_qid);
985 		opcode = d->rss.opcode;
986 		cpl = &d->cpl[0];
987 
988 		switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
989 		case X_RSPD_TYPE_FLBUF:
990 
991 			/* fall through */
992 
993 		case X_RSPD_TYPE_CPL:
994 			MPASS(opcode < NUM_CPL_CMDS);
995 
996 			switch (opcode) {
997 			case CPL_FW4_MSG:
998 			case CPL_FW6_MSG:
999 				cpl = unwrap_nm_fw6_msg(cpl);
1000 				/* fall through */
1001 			case CPL_SGE_EGR_UPDATE:
1002 				handle_nm_sge_egr_update(sc, ifp, cpl);
1003 				break;
1004 			case CPL_RX_PKT:
1005 				ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
1006 				    sc->params.sge.fl_pktshift;
1007 				ring->slot[fl_cidx].flags = 0;
1008 				nframes++;
1009 				if (!(lq & F_RSPD_NEWBUF)) {
1010 					MPASS(black_hole == 2);
1011 					break;
1012 				}
1013 				fl_credits++;
1014 				if (__predict_false(++fl_cidx == nm_rxq->fl_sidx))
1015 					fl_cidx = 0;
1016 				break;
1017 			default:
1018 				panic("%s: unexpected opcode 0x%x on nm_rxq %p",
1019 				    __func__, opcode, nm_rxq);
1020 			}
1021 			break;
1022 
1023 		case X_RSPD_TYPE_INTR:
1024 			/* Not equipped to handle forwarded interrupts. */
1025 			panic("%s: netmap queue received interrupt for iq %u\n",
1026 			    __func__, lq);
1027 
1028 		default:
1029 			panic("%s: illegal response type %d on nm_rxq %p",
1030 			    __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
1031 		}
1032 
1033 		d++;
1034 		if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
1035 			nm_rxq->iq_cidx = 0;
1036 			d = &nm_rxq->iq_desc[0];
1037 			nm_rxq->iq_gen ^= F_RSPD_GEN;
1038 		}
1039 
1040 		if (__predict_false(++nframes == rx_nframes) && !black_hole) {
1041 			atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1042 			netmap_rx_irq(ifp, nm_rxq->nid, &work);
1043 			nframes = 0;
1044 		}
1045 
1046 		if (__predict_false(++ndesc == rx_ndesc)) {
1047 			if (black_hole && fl_credits >= 8) {
1048 				fl_credits /= 8;
1049 				IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
1050 				    nm_rxq->fl_sidx);
1051 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1052 				    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1053 				fl_credits = fl_cidx & 7;
1054 			}
1055 			t4_write_reg(sc, sc->sge_gts_reg,
1056 			    V_CIDXINC(ndesc) |
1057 			    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
1058 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1059 			ndesc = 0;
1060 		}
1061 	}
1062 
1063 	atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1064 	if (black_hole) {
1065 		fl_credits /= 8;
1066 		IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1067 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1068 		    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1069 	} else if (nframes > 0)
1070 		netmap_rx_irq(ifp, nm_rxq->nid, &work);
1071 
1072     	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndesc) |
1073 	    V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1074 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1075 }
1076 #endif
1077