xref: /freebsd/sys/dev/cxgbe/t4_netmap.c (revision 46e1e307)
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 rx_buf_info *rxb;
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 	rxb = &sc->sge.rx_buf_info[0];
357 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
358 		if (rxb->size1 == NETMAP_BUF_SIZE(na)) {
359 			hwidx = rxb->hwidx1;
360 			break;
361 		}
362 		if (rxb->size2 == NETMAP_BUF_SIZE(na)) {
363 			hwidx = rxb->hwidx2;
364 			break;
365 		}
366 	}
367 	if (i >= SW_ZONE_SIZES) {
368 		if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
369 		    NETMAP_BUF_SIZE(na));
370 		return (ENXIO);
371 	}
372 
373 	/* Must set caps before calling netmap_reset */
374 	nm_set_native_flags(na);
375 
376 	for_each_nm_rxq(vi, i, nm_rxq) {
377 		kring = na->rx_rings[nm_rxq->nid];
378 		if (!nm_kring_pending_on(kring) ||
379 		    nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID)
380 			continue;
381 
382 		alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop));
383 		nm_rxq->fl_hwidx = hwidx;
384 		slot = netmap_reset(na, NR_RX, i, 0);
385 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
386 
387 		/* We deal with 8 bufs at a time */
388 		MPASS((na->num_rx_desc & 7) == 0);
389 		MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
390 		for (j = 0; j < nm_rxq->fl_sidx; j++) {
391 			uint64_t ba;
392 
393 			PNMB(na, &slot[j], &ba);
394 			MPASS(ba != 0);
395 			nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
396 		}
397 		j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
398 		MPASS((j & 7) == 0);
399 		j /= 8;	/* driver pidx to hardware pidx */
400 		wmb();
401 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
402 		    nm_rxq->fl_db_val | V_PIDX(j));
403 
404 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_OFF, NM_ON);
405 	}
406 
407 	for_each_nm_txq(vi, i, nm_txq) {
408 		kring = na->tx_rings[nm_txq->nid];
409 		if (!nm_kring_pending_on(kring) ||
410 		    nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID)
411 			continue;
412 
413 		alloc_nm_txq_hwq(vi, nm_txq);
414 		slot = netmap_reset(na, NR_TX, i, 0);
415 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
416 	}
417 
418 	if (vi->nm_rss == NULL) {
419 		vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
420 		    M_ZERO | M_WAITOK);
421 	}
422 
423 	MPASS(vi->nnmrxq > 0);
424 	if (nm_split_rss == 0 || vi->nnmrxq == 1) {
425 		for (i = 0; i < vi->rss_size;) {
426 			for_each_nm_rxq(vi, j, nm_rxq) {
427 				vi->nm_rss[i++] = nm_rxq->iq_abs_id;
428 				if (i == vi->rss_size)
429 					break;
430 			}
431 		}
432 		defq = vi->nm_rss[0];
433 	} else {
434 		/* We have multiple queues and we want to split the table. */
435 		MPASS(nm_split_rss != 0);
436 		MPASS(vi->nnmrxq > 1);
437 
438 		nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq];
439 		nrssq = vi->nnmrxq;
440 		if (vi->nnmrxq & 1) {
441 			/*
442 			 * Odd number of queues. The first rxq is designated the
443 			 * default queue, the rest are split evenly.
444 			 */
445 			defq = nm_rxq->iq_abs_id;
446 			nm_rxq++;
447 			nrssq--;
448 		} else {
449 			/*
450 			 * Even number of queues split into two halves.  The
451 			 * first rxq in one of the halves is designated the
452 			 * default queue.
453 			 */
454 #if 1
455 			/* First rxq in the first half. */
456 			defq = nm_rxq->iq_abs_id;
457 #else
458 			/* First rxq in the second half. */
459 			defq = nm_rxq[vi->nnmrxq / 2].iq_abs_id;
460 #endif
461 		}
462 
463 		i = 0;
464 		while (i < vi->rss_size / 2) {
465 			for (j = 0; j < nrssq / 2; j++) {
466 				vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
467 				if (i == vi->rss_size / 2)
468 					break;
469 			}
470 		}
471 		while (i < vi->rss_size) {
472 			for (j = nrssq / 2; j < nrssq; j++) {
473 				vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
474 				if (i == vi->rss_size)
475 					break;
476 			}
477 		}
478 	}
479 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
480 	    vi->nm_rss, vi->rss_size);
481 	if (rc != 0)
482 		if_printf(ifp, "netmap rss_config failed: %d\n", rc);
483 
484 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
485 	if (rc != 0)
486 		if_printf(ifp, "netmap rss hash/defaultq config failed: %d\n", rc);
487 
488 	return (rc);
489 }
490 
491 static int
492 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
493     struct netmap_adapter *na)
494 {
495 	struct netmap_kring *kring;
496 	int rc, i;
497 	struct sge_nm_txq *nm_txq;
498 	struct sge_nm_rxq *nm_rxq;
499 
500 	ASSERT_SYNCHRONIZED_OP(sc);
501 
502 	if (!nm_netmap_on(na))
503 		return (0);
504 
505 	if ((vi->flags & VI_INIT_DONE) == 0)
506 		return (0);
507 
508 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
509 	    vi->rss, vi->rss_size);
510 	if (rc != 0)
511 		if_printf(ifp, "failed to restore RSS config: %d\n", rc);
512 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, vi->rss[0], 0, 0);
513 	if (rc != 0)
514 		if_printf(ifp, "failed to restore RSS hash/defaultq: %d\n", rc);
515 	nm_clear_native_flags(na);
516 
517 	for_each_nm_txq(vi, i, nm_txq) {
518 		struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
519 
520 		kring = na->tx_rings[nm_txq->nid];
521 		if (!nm_kring_pending_off(kring) ||
522 		    nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID)
523 			continue;
524 
525 		/* Wait for hw pidx to catch up ... */
526 		while (be16toh(nm_txq->pidx) != spg->pidx)
527 			pause("nmpidx", 1);
528 
529 		/* ... and then for the cidx. */
530 		while (spg->pidx != spg->cidx)
531 			pause("nmcidx", 1);
532 
533 		free_nm_txq_hwq(vi, nm_txq);
534 	}
535 	for_each_nm_rxq(vi, i, nm_rxq) {
536 		kring = na->rx_rings[nm_rxq->nid];
537 		if (!nm_kring_pending_off(kring) ||
538 		    nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID)
539 			continue;
540 
541 		while (!atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_OFF))
542 			pause("nmst", 1);
543 
544 		free_nm_rxq_hwq(vi, nm_rxq);
545 	}
546 
547 	return (rc);
548 }
549 
550 static int
551 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
552 {
553 	struct ifnet *ifp = na->ifp;
554 	struct vi_info *vi = ifp->if_softc;
555 	struct adapter *sc = vi->pi->adapter;
556 	int rc;
557 
558 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
559 	if (rc != 0)
560 		return (rc);
561 	if (on)
562 		rc = cxgbe_netmap_on(sc, vi, ifp, na);
563 	else
564 		rc = cxgbe_netmap_off(sc, vi, ifp, na);
565 	end_synchronized_op(sc, 0);
566 
567 	return (rc);
568 }
569 
570 /* How many packets can a single type1 WR carry in n descriptors */
571 static inline int
572 ndesc_to_npkt(const int n)
573 {
574 
575 	MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
576 
577 	return (n * 2 - 1);
578 }
579 #define MAX_NPKT_IN_TYPE1_WR	(ndesc_to_npkt(SGE_MAX_WR_NDESC))
580 
581 /*
582  * Space (in descriptors) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
583  * carries n packets
584  */
585 static inline int
586 npkt_to_ndesc(const int n)
587 {
588 
589 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
590 
591 	return ((n + 2) / 2);
592 }
593 
594 /*
595  * Space (in 16B units) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
596  * carries n packets
597  */
598 static inline int
599 npkt_to_len16(const int n)
600 {
601 
602 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
603 
604 	return (n * 2 + 1);
605 }
606 
607 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
608 
609 static void
610 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
611 {
612 	int n;
613 	u_int db = nm_txq->doorbells;
614 
615 	MPASS(nm_txq->pidx != nm_txq->dbidx);
616 
617 	n = NMIDXDIFF(nm_txq, dbidx);
618 	if (n > 1)
619 		clrbit(&db, DOORBELL_WCWR);
620 	wmb();
621 
622 	switch (ffs(db) - 1) {
623 	case DOORBELL_UDB:
624 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
625 		break;
626 
627 	case DOORBELL_WCWR: {
628 		volatile uint64_t *dst, *src;
629 
630 		/*
631 		 * Queues whose 128B doorbell segment fits in the page do not
632 		 * use relative qid (udb_qid is always 0).  Only queues with
633 		 * doorbell segments can do WCWR.
634 		 */
635 		KASSERT(nm_txq->udb_qid == 0 && n == 1,
636 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
637 		    __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
638 
639 		dst = (volatile void *)((uintptr_t)nm_txq->udb +
640 		    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
641 		src = (void *)&nm_txq->desc[nm_txq->dbidx];
642 		while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
643 			*dst++ = *src++;
644 		wmb();
645 		break;
646 	}
647 
648 	case DOORBELL_UDBWC:
649 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
650 		wmb();
651 		break;
652 
653 	case DOORBELL_KDB:
654 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
655 		    V_QID(nm_txq->cntxt_id) | V_PIDX(n));
656 		break;
657 	}
658 	nm_txq->dbidx = nm_txq->pidx;
659 }
660 
661 /*
662  * Write work requests to send 'npkt' frames and ring the doorbell to send them
663  * on their way.  No need to check for wraparound.
664  */
665 static void
666 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
667     struct netmap_kring *kring, int npkt, int npkt_remaining)
668 {
669 	struct netmap_ring *ring = kring->ring;
670 	struct netmap_slot *slot;
671 	const u_int lim = kring->nkr_num_slots - 1;
672 	struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
673 	uint16_t len;
674 	uint64_t ba;
675 	struct cpl_tx_pkt_core *cpl;
676 	struct ulptx_sgl *usgl;
677 	int i, n;
678 
679 	while (npkt) {
680 		n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
681 		len = 0;
682 
683 		wr = (void *)&nm_txq->desc[nm_txq->pidx];
684 		wr->op_pkd = nm_txq->op_pkd;
685 		wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
686 		wr->npkt = n;
687 		wr->r3 = 0;
688 		wr->type = 1;
689 		cpl = (void *)(wr + 1);
690 
691 		for (i = 0; i < n; i++) {
692 			slot = &ring->slot[kring->nr_hwcur];
693 			PNMB(kring->na, slot, &ba);
694 			MPASS(ba != 0);
695 
696 			cpl->ctrl0 = nm_txq->cpl_ctrl0;
697 			cpl->pack = 0;
698 			cpl->len = htobe16(slot->len);
699 			/*
700 			 * netmap(4) says "netmap does not use features such as
701 			 * checksum offloading, TCP segmentation offloading,
702 			 * encryption, VLAN encapsulation/decapsulation, etc."
703 			 */
704 			cpl->ctrl1 = htobe64(F_TXPKT_IPCSUM_DIS |
705 			    F_TXPKT_L4CSUM_DIS);
706 
707 			usgl = (void *)(cpl + 1);
708 			usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
709 			    V_ULPTX_NSGE(1));
710 			usgl->len0 = htobe32(slot->len);
711 			usgl->addr0 = htobe64(ba);
712 
713 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
714 			cpl = (void *)(usgl + 1);
715 			MPASS(slot->len + len <= UINT16_MAX);
716 			len += slot->len;
717 			kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
718 		}
719 		wr->plen = htobe16(len);
720 
721 		npkt -= n;
722 		nm_txq->pidx += npkt_to_ndesc(n);
723 		MPASS(nm_txq->pidx <= nm_txq->sidx);
724 		if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
725 			/*
726 			 * This routine doesn't know how to write WRs that wrap
727 			 * around.  Make sure it wasn't asked to.
728 			 */
729 			MPASS(npkt == 0);
730 			nm_txq->pidx = 0;
731 		}
732 
733 		if (npkt == 0 && npkt_remaining == 0) {
734 			/* All done. */
735 			if (lazy_tx_credit_flush == 0) {
736 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
737 				    F_FW_WR_EQUIQ);
738 				nm_txq->equeqidx = nm_txq->pidx;
739 				nm_txq->equiqidx = nm_txq->pidx;
740 			}
741 			ring_nm_txq_db(sc, nm_txq);
742 			return;
743 		}
744 
745 		if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
746 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
747 			    F_FW_WR_EQUIQ);
748 			nm_txq->equeqidx = nm_txq->pidx;
749 			nm_txq->equiqidx = nm_txq->pidx;
750 		} else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
751 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
752 			nm_txq->equeqidx = nm_txq->pidx;
753 		}
754 		if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
755 			ring_nm_txq_db(sc, nm_txq);
756 	}
757 
758 	/* Will get called again. */
759 	MPASS(npkt_remaining);
760 }
761 
762 /* How many contiguous free descriptors starting at pidx */
763 static inline int
764 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
765 {
766 
767 	if (nm_txq->cidx > nm_txq->pidx)
768 		return (nm_txq->cidx - nm_txq->pidx - 1);
769 	else if (nm_txq->cidx > 0)
770 		return (nm_txq->sidx - nm_txq->pidx);
771 	else
772 		return (nm_txq->sidx - nm_txq->pidx - 1);
773 }
774 
775 static int
776 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
777 {
778 	struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
779 	uint16_t hw_cidx = spg->cidx;	/* snapshot */
780 	struct fw_eth_tx_pkts_wr *wr;
781 	int n = 0;
782 
783 	hw_cidx = be16toh(hw_cidx);
784 
785 	while (nm_txq->cidx != hw_cidx) {
786 		wr = (void *)&nm_txq->desc[nm_txq->cidx];
787 
788 		MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)) ||
789 		    wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR)));
790 		MPASS(wr->type == 1);
791 		MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
792 
793 		n += wr->npkt;
794 		nm_txq->cidx += npkt_to_ndesc(wr->npkt);
795 
796 		/*
797 		 * We never sent a WR that wrapped around so the credits coming
798 		 * back, WR by WR, should never cause the cidx to wrap around
799 		 * either.
800 		 */
801 		MPASS(nm_txq->cidx <= nm_txq->sidx);
802 		if (__predict_false(nm_txq->cidx == nm_txq->sidx))
803 			nm_txq->cidx = 0;
804 	}
805 
806 	return (n);
807 }
808 
809 static int
810 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
811 {
812 	struct netmap_adapter *na = kring->na;
813 	struct ifnet *ifp = na->ifp;
814 	struct vi_info *vi = ifp->if_softc;
815 	struct adapter *sc = vi->pi->adapter;
816 	struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
817 	const u_int head = kring->rhead;
818 	u_int reclaimed = 0;
819 	int n, d, npkt_remaining, ndesc_remaining;
820 
821 	/*
822 	 * Tx was at kring->nr_hwcur last time around and now we need to advance
823 	 * to kring->rhead.  Note that the driver's pidx moves independent of
824 	 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
825 	 * between descriptors and frames isn't 1:1).
826 	 */
827 
828 	npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
829 	    kring->nkr_num_slots - kring->nr_hwcur + head;
830 	while (npkt_remaining) {
831 		reclaimed += reclaim_nm_tx_desc(nm_txq);
832 		ndesc_remaining = contiguous_ndesc_available(nm_txq);
833 		/* Can't run out of descriptors with packets still remaining */
834 		MPASS(ndesc_remaining > 0);
835 
836 		/* # of desc needed to tx all remaining packets */
837 		d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
838 		if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
839 			d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
840 
841 		if (d <= ndesc_remaining)
842 			n = npkt_remaining;
843 		else {
844 			/* Can't send all, calculate how many can be sent */
845 			n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
846 			    MAX_NPKT_IN_TYPE1_WR;
847 			if (ndesc_remaining % SGE_MAX_WR_NDESC)
848 				n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
849 		}
850 
851 		/* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
852 		npkt_remaining -= n;
853 		cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining);
854 	}
855 	MPASS(npkt_remaining == 0);
856 	MPASS(kring->nr_hwcur == head);
857 	MPASS(nm_txq->dbidx == nm_txq->pidx);
858 
859 	/*
860 	 * Second part: reclaim buffers for completed transmissions.
861 	 */
862 	if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
863 		reclaimed += reclaim_nm_tx_desc(nm_txq);
864 		kring->nr_hwtail += reclaimed;
865 		if (kring->nr_hwtail >= kring->nkr_num_slots)
866 			kring->nr_hwtail -= kring->nkr_num_slots;
867 	}
868 
869 	return (0);
870 }
871 
872 static int
873 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
874 {
875 	struct netmap_adapter *na = kring->na;
876 	struct netmap_ring *ring = kring->ring;
877 	struct ifnet *ifp = na->ifp;
878 	struct vi_info *vi = ifp->if_softc;
879 	struct adapter *sc = vi->pi->adapter;
880 	struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
881 	u_int const head = kring->rhead;
882 	u_int n;
883 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
884 
885 	if (black_hole)
886 		return (0);	/* No updates ever. */
887 
888 	if (netmap_no_pendintr || force_update) {
889 		kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
890 		kring->nr_kflags &= ~NKR_PENDINTR;
891 	}
892 
893 	if (nm_rxq->fl_db_saved > 0 && starve_fl == 0) {
894 		wmb();
895 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
896 		    nm_rxq->fl_db_val | V_PIDX(nm_rxq->fl_db_saved));
897 		nm_rxq->fl_db_saved = 0;
898 	}
899 
900 	/* Userspace done with buffers from kring->nr_hwcur to head */
901 	n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
902 	    kring->nkr_num_slots - kring->nr_hwcur + head;
903 	n &= ~7U;
904 	if (n > 0) {
905 		u_int fl_pidx = nm_rxq->fl_pidx;
906 		struct netmap_slot *slot = &ring->slot[fl_pidx];
907 		uint64_t ba;
908 		int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
909 
910 		/*
911 		 * We always deal with 8 buffers at a time.  We must have
912 		 * stopped at an 8B boundary (fl_pidx) last time around and we
913 		 * must have a multiple of 8B buffers to give to the freelist.
914 		 */
915 		MPASS((fl_pidx & 7) == 0);
916 		MPASS((n & 7) == 0);
917 
918 		IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
919 		IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx);
920 
921 		while (n > 0) {
922 			for (i = 0; i < 8; i++, fl_pidx++, slot++) {
923 				PNMB(na, slot, &ba);
924 				MPASS(ba != 0);
925 				nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
926 				slot->flags &= ~NS_BUF_CHANGED;
927 				MPASS(fl_pidx <= nm_rxq->fl_sidx);
928 			}
929 			n -= 8;
930 			if (fl_pidx == nm_rxq->fl_sidx) {
931 				fl_pidx = 0;
932 				slot = &ring->slot[0];
933 			}
934 			if (++dbinc == 8 && n >= 32) {
935 				wmb();
936 				if (starve_fl)
937 					nm_rxq->fl_db_saved += dbinc;
938 				else {
939 					t4_write_reg(sc, sc->sge_kdoorbell_reg,
940 					    nm_rxq->fl_db_val | V_PIDX(dbinc));
941 				}
942 				dbinc = 0;
943 			}
944 		}
945 		MPASS(nm_rxq->fl_pidx == fl_pidx);
946 
947 		if (dbinc > 0) {
948 			wmb();
949 			if (starve_fl)
950 				nm_rxq->fl_db_saved += dbinc;
951 			else {
952 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
953 				    nm_rxq->fl_db_val | V_PIDX(dbinc));
954 			}
955 		}
956 	}
957 
958 	return (0);
959 }
960 
961 void
962 cxgbe_nm_attach(struct vi_info *vi)
963 {
964 	struct port_info *pi;
965 	struct adapter *sc;
966 	struct netmap_adapter na;
967 
968 	MPASS(vi->nnmrxq > 0);
969 	MPASS(vi->ifp != NULL);
970 
971 	pi = vi->pi;
972 	sc = pi->adapter;
973 
974 	bzero(&na, sizeof(na));
975 
976 	na.ifp = vi->ifp;
977 	na.na_flags = NAF_BDG_MAYSLEEP;
978 
979 	/* Netmap doesn't know about the space reserved for the status page. */
980 	na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
981 
982 	/*
983 	 * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
984 	 * num_rx_desc is based on the number of buffers that can be held in the
985 	 * freelist, and not the number of entries in the iq.  (These two are
986 	 * not exactly the same due to the space taken up by the status page).
987 	 */
988 	na.num_rx_desc = rounddown(vi->qsize_rxq, 8);
989 	na.nm_txsync = cxgbe_netmap_txsync;
990 	na.nm_rxsync = cxgbe_netmap_rxsync;
991 	na.nm_register = cxgbe_netmap_reg;
992 	na.num_tx_rings = vi->nnmtxq;
993 	na.num_rx_rings = vi->nnmrxq;
994 	netmap_attach(&na);	/* This adds IFCAP_NETMAP to if_capabilities */
995 }
996 
997 void
998 cxgbe_nm_detach(struct vi_info *vi)
999 {
1000 
1001 	MPASS(vi->nnmrxq > 0);
1002 	MPASS(vi->ifp != NULL);
1003 
1004 	netmap_detach(vi->ifp);
1005 }
1006 
1007 static inline const void *
1008 unwrap_nm_fw6_msg(const struct cpl_fw6_msg *cpl)
1009 {
1010 
1011 	MPASS(cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL);
1012 
1013 	/* data[0] is RSS header */
1014 	return (&cpl->data[1]);
1015 }
1016 
1017 static void
1018 handle_nm_sge_egr_update(struct adapter *sc, struct ifnet *ifp,
1019     const struct cpl_sge_egr_update *egr)
1020 {
1021 	uint32_t oq;
1022 	struct sge_nm_txq *nm_txq;
1023 
1024 	oq = be32toh(egr->opcode_qid);
1025 	MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
1026 	nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
1027 
1028 	netmap_tx_irq(ifp, nm_txq->nid);
1029 }
1030 
1031 void
1032 service_nm_rxq(struct sge_nm_rxq *nm_rxq)
1033 {
1034 	struct vi_info *vi = nm_rxq->vi;
1035 	struct adapter *sc = vi->pi->adapter;
1036 	struct ifnet *ifp = vi->ifp;
1037 	struct netmap_adapter *na = NA(ifp);
1038 	struct netmap_kring *kring = na->rx_rings[nm_rxq->nid];
1039 	struct netmap_ring *ring = kring->ring;
1040 	struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
1041 	const void *cpl;
1042 	uint32_t lq;
1043 	u_int work = 0;
1044 	uint8_t opcode;
1045 	uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
1046 	u_int fl_credits = fl_cidx & 7;
1047 	u_int ndesc = 0;	/* desc processed since last cidx update */
1048 	u_int nframes = 0;	/* frames processed since last netmap wakeup */
1049 
1050 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
1051 
1052 		rmb();
1053 
1054 		lq = be32toh(d->rsp.pldbuflen_qid);
1055 		opcode = d->rss.opcode;
1056 		cpl = &d->cpl[0];
1057 
1058 		switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
1059 		case X_RSPD_TYPE_FLBUF:
1060 
1061 			/* fall through */
1062 
1063 		case X_RSPD_TYPE_CPL:
1064 			MPASS(opcode < NUM_CPL_CMDS);
1065 
1066 			switch (opcode) {
1067 			case CPL_FW4_MSG:
1068 			case CPL_FW6_MSG:
1069 				cpl = unwrap_nm_fw6_msg(cpl);
1070 				/* fall through */
1071 			case CPL_SGE_EGR_UPDATE:
1072 				handle_nm_sge_egr_update(sc, ifp, cpl);
1073 				break;
1074 			case CPL_RX_PKT:
1075 				ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
1076 				    sc->params.sge.fl_pktshift;
1077 				ring->slot[fl_cidx].flags = 0;
1078 				nframes++;
1079 				if (!(lq & F_RSPD_NEWBUF)) {
1080 					MPASS(black_hole == 2);
1081 					break;
1082 				}
1083 				fl_credits++;
1084 				if (__predict_false(++fl_cidx == nm_rxq->fl_sidx))
1085 					fl_cidx = 0;
1086 				break;
1087 			default:
1088 				panic("%s: unexpected opcode 0x%x on nm_rxq %p",
1089 				    __func__, opcode, nm_rxq);
1090 			}
1091 			break;
1092 
1093 		case X_RSPD_TYPE_INTR:
1094 			/* Not equipped to handle forwarded interrupts. */
1095 			panic("%s: netmap queue received interrupt for iq %u\n",
1096 			    __func__, lq);
1097 
1098 		default:
1099 			panic("%s: illegal response type %d on nm_rxq %p",
1100 			    __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
1101 		}
1102 
1103 		d++;
1104 		if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
1105 			nm_rxq->iq_cidx = 0;
1106 			d = &nm_rxq->iq_desc[0];
1107 			nm_rxq->iq_gen ^= F_RSPD_GEN;
1108 		}
1109 
1110 		if (__predict_false(++nframes == rx_nframes) && !black_hole) {
1111 			atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1112 			netmap_rx_irq(ifp, nm_rxq->nid, &work);
1113 			nframes = 0;
1114 		}
1115 
1116 		if (__predict_false(++ndesc == rx_ndesc)) {
1117 			if (black_hole && fl_credits >= 8) {
1118 				fl_credits /= 8;
1119 				IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
1120 				    nm_rxq->fl_sidx);
1121 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1122 				    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1123 				fl_credits = fl_cidx & 7;
1124 			}
1125 			t4_write_reg(sc, sc->sge_gts_reg,
1126 			    V_CIDXINC(ndesc) |
1127 			    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
1128 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1129 			ndesc = 0;
1130 		}
1131 	}
1132 
1133 	atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1134 	if (black_hole) {
1135 		fl_credits /= 8;
1136 		IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1137 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1138 		    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1139 	} else if (nframes > 0)
1140 		netmap_rx_irq(ifp, nm_rxq->nid, &work);
1141 
1142     	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndesc) |
1143 	    V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1144 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1145 }
1146 #endif
1147