xref: /freebsd/sys/dev/cxgbe/t4_netmap.c (revision 2ed3236082a4473c1da8f72c1ebc071a7b54321f)
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_RWTUN, &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_RWTUN,
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 /*
114  * netmap(4) says "netmap does not use features such as checksum offloading, TCP
115  * segmentation offloading, encryption, VLAN encapsulation/decapsulation, etc."
116  * but this knob can be used to get the hardware to checksum all tx traffic
117  * anyway.
118  */
119 static int nm_txcsum = 0;
120 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_txcsum, CTLFLAG_RWTUN,
121     &nm_txcsum, 0, "Enable transmit checksum offloading.");
122 
123 static int
124 alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int cong)
125 {
126 	int rc, cntxt_id, i;
127 	__be32 v;
128 	struct adapter *sc = vi->adapter;
129 	struct sge_params *sp = &sc->params.sge;
130 	struct netmap_adapter *na = NA(vi->ifp);
131 	struct fw_iq_cmd c;
132 
133 	MPASS(na != NULL);
134 	MPASS(nm_rxq->iq_desc != NULL);
135 	MPASS(nm_rxq->fl_desc != NULL);
136 
137 	bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE);
138 	bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + sp->spg_len);
139 
140 	bzero(&c, sizeof(c));
141 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
142 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
143 	    V_FW_IQ_CMD_VFN(0));
144 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
145 	    FW_LEN16(c));
146 	MPASS(!forwarding_intr_to_fwq(sc));
147 	KASSERT(nm_rxq->intr_idx < sc->intr_count,
148 	    ("%s: invalid direct intr_idx %d", __func__, nm_rxq->intr_idx));
149 	v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
150 	c.type_to_iqandstindex = htobe32(v |
151 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
152 	    V_FW_IQ_CMD_VIID(vi->viid) |
153 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
154 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(vi->pi->tx_chan) |
155 	    F_FW_IQ_CMD_IQGTSMODE |
156 	    V_FW_IQ_CMD_IQINTCNTTHRESH(0) |
157 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
158 	c.iqsize = htobe16(vi->qsize_rxq);
159 	c.iqaddr = htobe64(nm_rxq->iq_ba);
160 	if (cong >= 0) {
161 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN |
162 		    V_FW_IQ_CMD_FL0CNGCHMAP(cong) | F_FW_IQ_CMD_FL0CONGCIF |
163 		    F_FW_IQ_CMD_FL0CONGEN);
164 	}
165 	c.iqns_to_fl0congen |=
166 	    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
167 		F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
168 		(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
169 		(black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0));
170 	c.fl0dcaen_to_fl0cidxfthresh =
171 	    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
172 		X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
173 		V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
174 		X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
175 	c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE);
176 	c.fl0addr = htobe64(nm_rxq->fl_ba);
177 
178 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
179 	if (rc != 0) {
180 		device_printf(sc->dev,
181 		    "failed to create netmap ingress queue: %d\n", rc);
182 		return (rc);
183 	}
184 
185 	nm_rxq->iq_cidx = 0;
186 	MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE);
187 	nm_rxq->iq_gen = F_RSPD_GEN;
188 	nm_rxq->iq_cntxt_id = be16toh(c.iqid);
189 	nm_rxq->iq_abs_id = be16toh(c.physiqid);
190 	cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
191 	if (cntxt_id >= sc->sge.niq) {
192 		panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
193 		    __func__, cntxt_id, sc->sge.niq - 1);
194 	}
195 	sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
196 
197 	nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
198 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
199 	nm_rxq->fl_db_saved = 0;
200 	/* matches the X_FETCHBURSTMAX_512B or X_FETCHBURSTMAX_256B above. */
201 	nm_rxq->fl_db_threshold = chip_id(sc) <= CHELSIO_T5 ? 8 : 4;
202 	MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
203 	cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
204 	if (cntxt_id >= sc->sge.neq) {
205 		panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
206 		    __func__, cntxt_id, sc->sge.neq - 1);
207 	}
208 	sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
209 
210 	nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) |
211 	    sc->chip_params->sge_fl_db;
212 
213 	if (chip_id(sc) >= CHELSIO_T5 && cong >= 0) {
214 		uint32_t param, val;
215 
216 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
217 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
218 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
219 		if (cong == 0)
220 			val = 1 << 19;
221 		else {
222 			val = 2 << 19;
223 			for (i = 0; i < 4; i++) {
224 				if (cong & (1 << i))
225 					val |= 1 << (i << 2);
226 			}
227 		}
228 
229 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
230 		if (rc != 0) {
231 			/* report error but carry on */
232 			device_printf(sc->dev,
233 			    "failed to set congestion manager context for "
234 			    "ingress queue %d: %d\n", nm_rxq->iq_cntxt_id, rc);
235 		}
236 	}
237 
238 	t4_write_reg(sc, sc->sge_gts_reg,
239 	    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
240 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
241 
242 	return (rc);
243 }
244 
245 static int
246 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
247 {
248 	struct adapter *sc = vi->adapter;
249 	int rc;
250 
251 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
252 	    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
253 	if (rc != 0)
254 		device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
255 		    __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
256 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
257 	return (rc);
258 }
259 
260 static int
261 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
262 {
263 	int rc, cntxt_id;
264 	size_t len;
265 	struct adapter *sc = vi->adapter;
266 	struct netmap_adapter *na = NA(vi->ifp);
267 	struct fw_eq_eth_cmd c;
268 
269 	MPASS(na != NULL);
270 	MPASS(nm_txq->desc != NULL);
271 
272 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
273 	bzero(nm_txq->desc, len);
274 
275 	bzero(&c, sizeof(c));
276 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
277 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
278 	    V_FW_EQ_ETH_CMD_VFN(0));
279 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
280 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
281 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
282 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
283 	c.fetchszm_to_iqid =
284 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
285 		V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
286 		V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
287 	c.dcaen_to_eqsize =
288 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
289 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
290 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
291 		V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
292 	c.eqaddr = htobe64(nm_txq->ba);
293 
294 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
295 	if (rc != 0) {
296 		device_printf(vi->dev,
297 		    "failed to create netmap egress queue: %d\n", rc);
298 		return (rc);
299 	}
300 
301 	nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
302 	cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
303 	if (cntxt_id >= sc->sge.neq)
304 	    panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
305 		cntxt_id, sc->sge.neq - 1);
306 	sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
307 
308 	nm_txq->pidx = nm_txq->cidx = 0;
309 	MPASS(nm_txq->sidx == na->num_tx_desc);
310 	nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
311 
312 	nm_txq->doorbells = sc->doorbells;
313 	if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
314 	    isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
315 	    isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
316 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
317 		uint32_t mask = (1 << s_qpp) - 1;
318 		volatile uint8_t *udb;
319 
320 		udb = sc->udbs_base + UDBS_DB_OFFSET;
321 		udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
322 		nm_txq->udb_qid = nm_txq->cntxt_id & mask;
323 		if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
324 	    		clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
325 		else {
326 			udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
327 			nm_txq->udb_qid = 0;
328 		}
329 		nm_txq->udb = (volatile void *)udb;
330 	}
331 
332 	if (sc->params.fw_vers < FW_VERSION32(1, 25, 1, 0)) {
333 		uint32_t param, val;
334 
335 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
336 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) |
337 		    V_FW_PARAMS_PARAM_YZ(nm_txq->cntxt_id);
338 		val = 0xff;
339 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
340 		if (rc != 0) {
341 			device_printf(vi->dev,
342 			    "failed to bind netmap txq %d to class 0xff: %d\n",
343 			    nm_txq->cntxt_id, rc);
344 			rc = 0;
345 		}
346 	}
347 
348 	return (rc);
349 }
350 
351 static int
352 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
353 {
354 	struct adapter *sc = vi->adapter;
355 	int rc;
356 
357 	rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
358 	if (rc != 0)
359 		device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
360 		    nm_txq->cntxt_id, rc);
361 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
362 	return (rc);
363 }
364 
365 static int
366 cxgbe_netmap_simple_rss(struct adapter *sc, struct vi_info *vi,
367     struct ifnet *ifp, struct netmap_adapter *na)
368 {
369 	struct netmap_kring *kring;
370 	struct sge_nm_rxq *nm_rxq;
371 	int rc, i, j, nm_state, defq;
372 	uint16_t *rss;
373 
374 	/*
375 	 * Check if there's at least one active (or about to go active) netmap
376 	 * rx queue.
377 	 */
378 	defq = -1;
379 	for_each_nm_rxq(vi, j, nm_rxq) {
380 		nm_state = atomic_load_int(&nm_rxq->nm_state);
381 		kring = na->rx_rings[nm_rxq->nid];
382 		if ((nm_state != NM_OFF && !nm_kring_pending_off(kring)) ||
383 		    (nm_state == NM_OFF && nm_kring_pending_on(kring))) {
384 			MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
385 			if (defq == -1) {
386 				defq = nm_rxq->iq_abs_id;
387 				break;
388 			}
389 		}
390 	}
391 
392 	if (defq == -1) {
393 		/* No active netmap queues.  Switch back to NIC queues. */
394 		rss = vi->rss;
395 		defq = vi->rss[0];
396 	} else {
397 		for (i = 0; i < vi->rss_size;) {
398 			for_each_nm_rxq(vi, j, nm_rxq) {
399 				nm_state = atomic_load_int(&nm_rxq->nm_state);
400 				kring = na->rx_rings[nm_rxq->nid];
401 				if ((nm_state != NM_OFF &&
402 				    !nm_kring_pending_off(kring)) ||
403 				    (nm_state == NM_OFF &&
404 				    nm_kring_pending_on(kring))) {
405 					MPASS(nm_rxq->iq_cntxt_id !=
406 					    INVALID_NM_RXQ_CNTXT_ID);
407 					vi->nm_rss[i++] = nm_rxq->iq_abs_id;
408 					if (i == vi->rss_size)
409 						break;
410 				}
411 			}
412 		}
413 		rss = vi->nm_rss;
414 	}
415 
416 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size, rss,
417 	    vi->rss_size);
418 	if (rc != 0)
419 		if_printf(ifp, "netmap rss_config failed: %d\n", rc);
420 
421 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
422 	if (rc != 0) {
423 		if_printf(ifp, "netmap defaultq config failed: %d\n", rc);
424 	}
425 
426 	return (rc);
427 }
428 
429 /*
430  * Odd number of rx queues work best for split RSS mode as the first queue can
431  * be dedicated for non-RSS traffic and the rest divided into two equal halves.
432  */
433 static int
434 cxgbe_netmap_split_rss(struct adapter *sc, struct vi_info *vi,
435     struct ifnet *ifp, struct netmap_adapter *na)
436 {
437 	struct netmap_kring *kring;
438 	struct sge_nm_rxq *nm_rxq;
439 	int rc, i, j, nm_state, defq;
440 	int nactive[2] = {0, 0};
441 	int dq[2] = {-1, -1};
442 	bool dq_norss;		/* default queue should not be in RSS table. */
443 
444 	MPASS(nm_split_rss != 0);
445 	MPASS(vi->nnmrxq > 1);
446 
447 	for_each_nm_rxq(vi, i, nm_rxq) {
448 		j = i / ((vi->nnmrxq + 1) / 2);
449 		nm_state = atomic_load_int(&nm_rxq->nm_state);
450 		kring = na->rx_rings[nm_rxq->nid];
451 		if ((nm_state != NM_OFF && !nm_kring_pending_off(kring)) ||
452 		    (nm_state == NM_OFF && nm_kring_pending_on(kring))) {
453 			MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
454 			nactive[j]++;
455 			if (dq[j] == -1) {
456 				dq[j] = nm_rxq->iq_abs_id;
457 				break;
458 			}
459 		}
460 	}
461 
462 	if (nactive[0] == 0 || nactive[1] == 0)
463 		return (cxgbe_netmap_simple_rss(sc, vi, ifp, na));
464 
465 	MPASS(dq[0] != -1 && dq[1] != -1);
466 	if (nactive[0] > nactive[1]) {
467 		defq = dq[0];
468 		dq_norss = true;
469 	} else if (nactive[0] < nactive[1]) {
470 		defq = dq[1];
471 		dq_norss = true;
472 	} else {
473 		defq = dq[0];
474 		dq_norss = false;
475 	}
476 
477 	i = 0;
478 	nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq];
479 	while (i < vi->rss_size / 2) {
480 		for (j = 0; j < (vi->nnmrxq + 1) / 2; j++) {
481 			nm_state = atomic_load_int(&nm_rxq[j].nm_state);
482 			kring = na->rx_rings[nm_rxq[j].nid];
483 			if ((nm_state == NM_OFF &&
484 			    !nm_kring_pending_on(kring)) ||
485 			    (nm_state == NM_ON &&
486 			    nm_kring_pending_off(kring))) {
487 				continue;
488 			}
489 			MPASS(nm_rxq[j].iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
490 			if (dq_norss && defq == nm_rxq[j].iq_abs_id)
491 				continue;
492 			vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
493 			if (i == vi->rss_size / 2)
494 				break;
495 		}
496 	}
497 	while (i < vi->rss_size) {
498 		for (j = (vi->nnmrxq + 1) / 2; j < vi->nnmrxq; j++) {
499 			nm_state = atomic_load_int(&nm_rxq[j].nm_state);
500 			kring = na->rx_rings[nm_rxq[j].nid];
501 			if ((nm_state == NM_OFF &&
502 			    !nm_kring_pending_on(kring)) ||
503 			    (nm_state == NM_ON &&
504 			    nm_kring_pending_off(kring))) {
505 				continue;
506 			}
507 			MPASS(nm_rxq[j].iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
508 			if (dq_norss && defq == nm_rxq[j].iq_abs_id)
509 				continue;
510 			vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
511 			if (i == vi->rss_size)
512 				break;
513 		}
514 	}
515 
516 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
517 	    vi->nm_rss, vi->rss_size);
518 	if (rc != 0)
519 		if_printf(ifp, "netmap split_rss_config failed: %d\n", rc);
520 
521 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
522 	if (rc != 0)
523 		if_printf(ifp, "netmap defaultq config failed: %d\n", rc);
524 
525 	return (rc);
526 }
527 
528 static inline int
529 cxgbe_netmap_rss(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
530     struct netmap_adapter *na)
531 {
532 
533 	if (nm_split_rss == 0 || vi->nnmrxq == 1)
534 		return (cxgbe_netmap_simple_rss(sc, vi, ifp, na));
535 	else
536 		return (cxgbe_netmap_split_rss(sc, vi, ifp, na));
537 }
538 
539 static int
540 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
541     struct netmap_adapter *na)
542 {
543 	struct netmap_slot *slot;
544 	struct netmap_kring *kring;
545 	struct sge_nm_rxq *nm_rxq;
546 	struct sge_nm_txq *nm_txq;
547 	int i, j, hwidx;
548 	struct rx_buf_info *rxb;
549 
550 	ASSERT_SYNCHRONIZED_OP(sc);
551 	MPASS(vi->nnmrxq > 0);
552 	MPASS(vi->nnmtxq > 0);
553 
554 	if ((vi->flags & VI_INIT_DONE) == 0 ||
555 	    (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
556 		if_printf(ifp, "cannot enable netmap operation because "
557 		    "interface is not UP.\n");
558 		return (EAGAIN);
559 	}
560 
561 	rxb = &sc->sge.rx_buf_info[0];
562 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
563 		if (rxb->size1 == NETMAP_BUF_SIZE(na)) {
564 			hwidx = rxb->hwidx1;
565 			break;
566 		}
567 		if (rxb->size2 == NETMAP_BUF_SIZE(na)) {
568 			hwidx = rxb->hwidx2;
569 			break;
570 		}
571 	}
572 	if (i >= SW_ZONE_SIZES) {
573 		if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
574 		    NETMAP_BUF_SIZE(na));
575 		return (ENXIO);
576 	}
577 
578 	/* Must set caps before calling netmap_reset */
579 	nm_set_native_flags(na);
580 
581 	for_each_nm_rxq(vi, i, nm_rxq) {
582 		kring = na->rx_rings[nm_rxq->nid];
583 		if (!nm_kring_pending_on(kring) ||
584 		    nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID)
585 			continue;
586 
587 		alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop));
588 		nm_rxq->fl_hwidx = hwidx;
589 		slot = netmap_reset(na, NR_RX, i, 0);
590 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
591 
592 		/* We deal with 8 bufs at a time */
593 		MPASS((na->num_rx_desc & 7) == 0);
594 		MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
595 		for (j = 0; j < nm_rxq->fl_sidx; j++) {
596 			uint64_t ba;
597 
598 			PNMB(na, &slot[j], &ba);
599 			MPASS(ba != 0);
600 			nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
601 		}
602 		j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
603 		MPASS((j & 7) == 0);
604 		j /= 8;	/* driver pidx to hardware pidx */
605 		wmb();
606 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
607 		    nm_rxq->fl_db_val | V_PIDX(j));
608 
609 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_OFF, NM_ON);
610 	}
611 
612 	for_each_nm_txq(vi, i, nm_txq) {
613 		kring = na->tx_rings[nm_txq->nid];
614 		if (!nm_kring_pending_on(kring) ||
615 		    nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID)
616 			continue;
617 
618 		alloc_nm_txq_hwq(vi, nm_txq);
619 		slot = netmap_reset(na, NR_TX, i, 0);
620 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
621 	}
622 
623 	if (vi->nm_rss == NULL) {
624 		vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
625 		    M_ZERO | M_WAITOK);
626 	}
627 
628 	return (cxgbe_netmap_rss(sc, vi, ifp, na));
629 }
630 
631 static int
632 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
633     struct netmap_adapter *na)
634 {
635 	struct netmap_kring *kring;
636 	int rc, i, nm_state, nactive;
637 	struct sge_nm_txq *nm_txq;
638 	struct sge_nm_rxq *nm_rxq;
639 
640 	ASSERT_SYNCHRONIZED_OP(sc);
641 	MPASS(vi->nnmrxq > 0);
642 	MPASS(vi->nnmtxq > 0);
643 
644 	if (!nm_netmap_on(na))
645 		return (0);
646 
647 	if ((vi->flags & VI_INIT_DONE) == 0)
648 		return (0);
649 
650 	/* First remove the queues that are stopping from the RSS table. */
651 	rc = cxgbe_netmap_rss(sc, vi, ifp, na);
652 	if (rc != 0)
653 		return (rc);	/* error message logged already. */
654 
655 	for_each_nm_txq(vi, i, nm_txq) {
656 		struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
657 
658 		kring = na->tx_rings[nm_txq->nid];
659 		if (!nm_kring_pending_off(kring) ||
660 		    nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID)
661 			continue;
662 
663 		/* Wait for hw pidx to catch up ... */
664 		while (be16toh(nm_txq->pidx) != spg->pidx)
665 			pause("nmpidx", 1);
666 
667 		/* ... and then for the cidx. */
668 		while (spg->pidx != spg->cidx)
669 			pause("nmcidx", 1);
670 
671 		free_nm_txq_hwq(vi, nm_txq);
672 
673 		/* XXX: netmap, not the driver, should do this. */
674 		kring->rhead = kring->rcur = kring->nr_hwcur = 0;
675 		kring->rtail = kring->nr_hwtail = kring->nkr_num_slots - 1;
676 	}
677 	nactive = 0;
678 	for_each_nm_rxq(vi, i, nm_rxq) {
679 		nm_state = atomic_load_int(&nm_rxq->nm_state);
680 		kring = na->rx_rings[nm_rxq->nid];
681 		if (nm_state != NM_OFF && !nm_kring_pending_off(kring))
682 			nactive++;
683 		if (nm_state == NM_OFF || !nm_kring_pending_off(kring))
684 			continue;
685 
686 		MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
687 		while (!atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_OFF))
688 			pause("nmst", 1);
689 
690 		free_nm_rxq_hwq(vi, nm_rxq);
691 
692 		/* XXX: netmap, not the driver, should do this. */
693 		kring->rhead = kring->rcur = kring->nr_hwcur = 0;
694 		kring->rtail = kring->nr_hwtail = 0;
695 	}
696 	netmap_krings_mode_commit(na, 0);
697 	if (nactive == 0)
698 		nm_clear_native_flags(na);
699 
700 	return (rc);
701 }
702 
703 static int
704 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
705 {
706 	struct ifnet *ifp = na->ifp;
707 	struct vi_info *vi = ifp->if_softc;
708 	struct adapter *sc = vi->adapter;
709 	int rc;
710 
711 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
712 	if (rc != 0)
713 		return (rc);
714 	if (on)
715 		rc = cxgbe_netmap_on(sc, vi, ifp, na);
716 	else
717 		rc = cxgbe_netmap_off(sc, vi, ifp, na);
718 	end_synchronized_op(sc, 0);
719 
720 	return (rc);
721 }
722 
723 /* How many packets can a single type1 WR carry in n descriptors */
724 static inline int
725 ndesc_to_npkt(const int n)
726 {
727 
728 	MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
729 
730 	return (n * 2 - 1);
731 }
732 #define MAX_NPKT_IN_TYPE1_WR	(ndesc_to_npkt(SGE_MAX_WR_NDESC))
733 
734 /*
735  * Space (in descriptors) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
736  * carries n packets
737  */
738 static inline int
739 npkt_to_ndesc(const int n)
740 {
741 
742 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
743 
744 	return ((n + 2) / 2);
745 }
746 
747 /*
748  * Space (in 16B units) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
749  * carries n packets
750  */
751 static inline int
752 npkt_to_len16(const int n)
753 {
754 
755 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
756 
757 	return (n * 2 + 1);
758 }
759 
760 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
761 
762 static void
763 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
764 {
765 	int n;
766 	u_int db = nm_txq->doorbells;
767 
768 	MPASS(nm_txq->pidx != nm_txq->dbidx);
769 
770 	n = NMIDXDIFF(nm_txq, dbidx);
771 	if (n > 1)
772 		clrbit(&db, DOORBELL_WCWR);
773 	wmb();
774 
775 	switch (ffs(db) - 1) {
776 	case DOORBELL_UDB:
777 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
778 		break;
779 
780 	case DOORBELL_WCWR: {
781 		volatile uint64_t *dst, *src;
782 
783 		/*
784 		 * Queues whose 128B doorbell segment fits in the page do not
785 		 * use relative qid (udb_qid is always 0).  Only queues with
786 		 * doorbell segments can do WCWR.
787 		 */
788 		KASSERT(nm_txq->udb_qid == 0 && n == 1,
789 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
790 		    __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
791 
792 		dst = (volatile void *)((uintptr_t)nm_txq->udb +
793 		    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
794 		src = (void *)&nm_txq->desc[nm_txq->dbidx];
795 		while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
796 			*dst++ = *src++;
797 		wmb();
798 		break;
799 	}
800 
801 	case DOORBELL_UDBWC:
802 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
803 		wmb();
804 		break;
805 
806 	case DOORBELL_KDB:
807 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
808 		    V_QID(nm_txq->cntxt_id) | V_PIDX(n));
809 		break;
810 	}
811 	nm_txq->dbidx = nm_txq->pidx;
812 }
813 
814 /*
815  * Write work requests to send 'npkt' frames and ring the doorbell to send them
816  * on their way.  No need to check for wraparound.
817  */
818 static void
819 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
820     struct netmap_kring *kring, int npkt, int npkt_remaining)
821 {
822 	struct netmap_ring *ring = kring->ring;
823 	struct netmap_slot *slot;
824 	const u_int lim = kring->nkr_num_slots - 1;
825 	struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
826 	uint16_t len;
827 	uint64_t ba;
828 	struct cpl_tx_pkt_core *cpl;
829 	struct ulptx_sgl *usgl;
830 	int i, n;
831 
832 	while (npkt) {
833 		n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
834 		len = 0;
835 
836 		wr = (void *)&nm_txq->desc[nm_txq->pidx];
837 		wr->op_pkd = nm_txq->op_pkd;
838 		wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
839 		wr->npkt = n;
840 		wr->r3 = 0;
841 		wr->type = 1;
842 		cpl = (void *)(wr + 1);
843 
844 		for (i = 0; i < n; i++) {
845 			slot = &ring->slot[kring->nr_hwcur];
846 			PNMB(kring->na, slot, &ba);
847 			MPASS(ba != 0);
848 
849 			cpl->ctrl0 = nm_txq->cpl_ctrl0;
850 			cpl->pack = 0;
851 			cpl->len = htobe16(slot->len);
852 			cpl->ctrl1 = nm_txcsum ? 0 :
853 			    htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
854 
855 			usgl = (void *)(cpl + 1);
856 			usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
857 			    V_ULPTX_NSGE(1));
858 			usgl->len0 = htobe32(slot->len);
859 			usgl->addr0 = htobe64(ba);
860 
861 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
862 			cpl = (void *)(usgl + 1);
863 			MPASS(slot->len + len <= UINT16_MAX);
864 			len += slot->len;
865 			kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
866 		}
867 		wr->plen = htobe16(len);
868 
869 		npkt -= n;
870 		nm_txq->pidx += npkt_to_ndesc(n);
871 		MPASS(nm_txq->pidx <= nm_txq->sidx);
872 		if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
873 			/*
874 			 * This routine doesn't know how to write WRs that wrap
875 			 * around.  Make sure it wasn't asked to.
876 			 */
877 			MPASS(npkt == 0);
878 			nm_txq->pidx = 0;
879 		}
880 
881 		if (npkt == 0 && npkt_remaining == 0) {
882 			/* All done. */
883 			if (lazy_tx_credit_flush == 0) {
884 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
885 				    F_FW_WR_EQUIQ);
886 				nm_txq->equeqidx = nm_txq->pidx;
887 				nm_txq->equiqidx = nm_txq->pidx;
888 			}
889 			ring_nm_txq_db(sc, nm_txq);
890 			return;
891 		}
892 
893 		if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
894 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
895 			    F_FW_WR_EQUIQ);
896 			nm_txq->equeqidx = nm_txq->pidx;
897 			nm_txq->equiqidx = nm_txq->pidx;
898 		} else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
899 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
900 			nm_txq->equeqidx = nm_txq->pidx;
901 		}
902 		if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
903 			ring_nm_txq_db(sc, nm_txq);
904 	}
905 
906 	/* Will get called again. */
907 	MPASS(npkt_remaining);
908 }
909 
910 /* How many contiguous free descriptors starting at pidx */
911 static inline int
912 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
913 {
914 
915 	if (nm_txq->cidx > nm_txq->pidx)
916 		return (nm_txq->cidx - nm_txq->pidx - 1);
917 	else if (nm_txq->cidx > 0)
918 		return (nm_txq->sidx - nm_txq->pidx);
919 	else
920 		return (nm_txq->sidx - nm_txq->pidx - 1);
921 }
922 
923 static int
924 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
925 {
926 	struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
927 	uint16_t hw_cidx = spg->cidx;	/* snapshot */
928 	struct fw_eth_tx_pkts_wr *wr;
929 	int n = 0;
930 
931 	hw_cidx = be16toh(hw_cidx);
932 
933 	while (nm_txq->cidx != hw_cidx) {
934 		wr = (void *)&nm_txq->desc[nm_txq->cidx];
935 
936 		MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)) ||
937 		    wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR)));
938 		MPASS(wr->type == 1);
939 		MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
940 
941 		n += wr->npkt;
942 		nm_txq->cidx += npkt_to_ndesc(wr->npkt);
943 
944 		/*
945 		 * We never sent a WR that wrapped around so the credits coming
946 		 * back, WR by WR, should never cause the cidx to wrap around
947 		 * either.
948 		 */
949 		MPASS(nm_txq->cidx <= nm_txq->sidx);
950 		if (__predict_false(nm_txq->cidx == nm_txq->sidx))
951 			nm_txq->cidx = 0;
952 	}
953 
954 	return (n);
955 }
956 
957 static int
958 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
959 {
960 	struct netmap_adapter *na = kring->na;
961 	struct ifnet *ifp = na->ifp;
962 	struct vi_info *vi = ifp->if_softc;
963 	struct adapter *sc = vi->adapter;
964 	struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
965 	const u_int head = kring->rhead;
966 	u_int reclaimed = 0;
967 	int n, d, npkt_remaining, ndesc_remaining;
968 
969 	/*
970 	 * Tx was at kring->nr_hwcur last time around and now we need to advance
971 	 * to kring->rhead.  Note that the driver's pidx moves independent of
972 	 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
973 	 * between descriptors and frames isn't 1:1).
974 	 */
975 
976 	npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
977 	    kring->nkr_num_slots - kring->nr_hwcur + head;
978 	while (npkt_remaining) {
979 		reclaimed += reclaim_nm_tx_desc(nm_txq);
980 		ndesc_remaining = contiguous_ndesc_available(nm_txq);
981 		/* Can't run out of descriptors with packets still remaining */
982 		MPASS(ndesc_remaining > 0);
983 
984 		/* # of desc needed to tx all remaining packets */
985 		d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
986 		if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
987 			d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
988 
989 		if (d <= ndesc_remaining)
990 			n = npkt_remaining;
991 		else {
992 			/* Can't send all, calculate how many can be sent */
993 			n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
994 			    MAX_NPKT_IN_TYPE1_WR;
995 			if (ndesc_remaining % SGE_MAX_WR_NDESC)
996 				n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
997 		}
998 
999 		/* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
1000 		npkt_remaining -= n;
1001 		cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining);
1002 	}
1003 	MPASS(npkt_remaining == 0);
1004 	MPASS(kring->nr_hwcur == head);
1005 	MPASS(nm_txq->dbidx == nm_txq->pidx);
1006 
1007 	/*
1008 	 * Second part: reclaim buffers for completed transmissions.
1009 	 */
1010 	if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
1011 		reclaimed += reclaim_nm_tx_desc(nm_txq);
1012 		kring->nr_hwtail += reclaimed;
1013 		if (kring->nr_hwtail >= kring->nkr_num_slots)
1014 			kring->nr_hwtail -= kring->nkr_num_slots;
1015 	}
1016 
1017 	return (0);
1018 }
1019 
1020 static int
1021 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
1022 {
1023 	struct netmap_adapter *na = kring->na;
1024 	struct netmap_ring *ring = kring->ring;
1025 	struct ifnet *ifp = na->ifp;
1026 	struct vi_info *vi = ifp->if_softc;
1027 	struct adapter *sc = vi->adapter;
1028 	struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
1029 	u_int const head = kring->rhead;
1030 	u_int n;
1031 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1032 
1033 	if (black_hole)
1034 		return (0);	/* No updates ever. */
1035 
1036 	if (netmap_no_pendintr || force_update) {
1037 		kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
1038 		kring->nr_kflags &= ~NKR_PENDINTR;
1039 	}
1040 
1041 	if (nm_rxq->fl_db_saved > 0 && starve_fl == 0) {
1042 		wmb();
1043 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1044 		    nm_rxq->fl_db_val | V_PIDX(nm_rxq->fl_db_saved));
1045 		nm_rxq->fl_db_saved = 0;
1046 	}
1047 
1048 	/* Userspace done with buffers from kring->nr_hwcur to head */
1049 	n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
1050 	    kring->nkr_num_slots - kring->nr_hwcur + head;
1051 	n &= ~7U;
1052 	if (n > 0) {
1053 		u_int fl_pidx = nm_rxq->fl_pidx;
1054 		struct netmap_slot *slot = &ring->slot[fl_pidx];
1055 		uint64_t ba;
1056 		int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
1057 
1058 		/*
1059 		 * We always deal with 8 buffers at a time.  We must have
1060 		 * stopped at an 8B boundary (fl_pidx) last time around and we
1061 		 * must have a multiple of 8B buffers to give to the freelist.
1062 		 */
1063 		MPASS((fl_pidx & 7) == 0);
1064 		MPASS((n & 7) == 0);
1065 
1066 		IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
1067 		IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx2);
1068 
1069 		while (n > 0) {
1070 			for (i = 0; i < 8; i++, fl_pidx++, slot++) {
1071 				PNMB(na, slot, &ba);
1072 				MPASS(ba != 0);
1073 				nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
1074 				slot->flags &= ~NS_BUF_CHANGED;
1075 				MPASS(fl_pidx <= nm_rxq->fl_sidx2);
1076 			}
1077 			n -= 8;
1078 			if (fl_pidx == nm_rxq->fl_sidx2) {
1079 				fl_pidx = 0;
1080 				slot = &ring->slot[0];
1081 			}
1082 			if (++dbinc == nm_rxq->fl_db_threshold) {
1083 				wmb();
1084 				if (starve_fl)
1085 					nm_rxq->fl_db_saved += dbinc;
1086 				else {
1087 					t4_write_reg(sc, sc->sge_kdoorbell_reg,
1088 					    nm_rxq->fl_db_val | V_PIDX(dbinc));
1089 				}
1090 				dbinc = 0;
1091 			}
1092 		}
1093 		MPASS(nm_rxq->fl_pidx == fl_pidx);
1094 
1095 		if (dbinc > 0) {
1096 			wmb();
1097 			if (starve_fl)
1098 				nm_rxq->fl_db_saved += dbinc;
1099 			else {
1100 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1101 				    nm_rxq->fl_db_val | V_PIDX(dbinc));
1102 			}
1103 		}
1104 	}
1105 
1106 	return (0);
1107 }
1108 
1109 void
1110 cxgbe_nm_attach(struct vi_info *vi)
1111 {
1112 	struct port_info *pi;
1113 	struct adapter *sc;
1114 	struct netmap_adapter na;
1115 
1116 	MPASS(vi->nnmrxq > 0);
1117 	MPASS(vi->ifp != NULL);
1118 
1119 	pi = vi->pi;
1120 	sc = pi->adapter;
1121 
1122 	bzero(&na, sizeof(na));
1123 
1124 	na.ifp = vi->ifp;
1125 	na.na_flags = NAF_BDG_MAYSLEEP;
1126 
1127 	/* Netmap doesn't know about the space reserved for the status page. */
1128 	na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
1129 
1130 	/*
1131 	 * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
1132 	 * num_rx_desc is based on the number of buffers that can be held in the
1133 	 * freelist, and not the number of entries in the iq.  (These two are
1134 	 * not exactly the same due to the space taken up by the status page).
1135 	 */
1136 	na.num_rx_desc = rounddown(vi->qsize_rxq, 8);
1137 	na.nm_txsync = cxgbe_netmap_txsync;
1138 	na.nm_rxsync = cxgbe_netmap_rxsync;
1139 	na.nm_register = cxgbe_netmap_reg;
1140 	na.num_tx_rings = vi->nnmtxq;
1141 	na.num_rx_rings = vi->nnmrxq;
1142 	na.rx_buf_maxsize = MAX_MTU;
1143 	netmap_attach(&na);	/* This adds IFCAP_NETMAP to if_capabilities */
1144 }
1145 
1146 void
1147 cxgbe_nm_detach(struct vi_info *vi)
1148 {
1149 
1150 	MPASS(vi->nnmrxq > 0);
1151 	MPASS(vi->ifp != NULL);
1152 
1153 	netmap_detach(vi->ifp);
1154 }
1155 
1156 static inline const void *
1157 unwrap_nm_fw6_msg(const struct cpl_fw6_msg *cpl)
1158 {
1159 
1160 	MPASS(cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL);
1161 
1162 	/* data[0] is RSS header */
1163 	return (&cpl->data[1]);
1164 }
1165 
1166 static void
1167 handle_nm_sge_egr_update(struct adapter *sc, struct ifnet *ifp,
1168     const struct cpl_sge_egr_update *egr)
1169 {
1170 	uint32_t oq;
1171 	struct sge_nm_txq *nm_txq;
1172 
1173 	oq = be32toh(egr->opcode_qid);
1174 	MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
1175 	nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
1176 
1177 	netmap_tx_irq(ifp, nm_txq->nid);
1178 }
1179 
1180 void
1181 service_nm_rxq(struct sge_nm_rxq *nm_rxq)
1182 {
1183 	struct vi_info *vi = nm_rxq->vi;
1184 	struct adapter *sc = vi->adapter;
1185 	struct ifnet *ifp = vi->ifp;
1186 	struct netmap_adapter *na = NA(ifp);
1187 	struct netmap_kring *kring = na->rx_rings[nm_rxq->nid];
1188 	struct netmap_ring *ring = kring->ring;
1189 	struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
1190 	const void *cpl;
1191 	uint32_t lq;
1192 	u_int work = 0;
1193 	uint8_t opcode;
1194 	uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
1195 	u_int fl_credits = fl_cidx & 7;
1196 	u_int ndesc = 0;	/* desc processed since last cidx update */
1197 	u_int nframes = 0;	/* frames processed since last netmap wakeup */
1198 
1199 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
1200 
1201 		rmb();
1202 
1203 		lq = be32toh(d->rsp.pldbuflen_qid);
1204 		opcode = d->rss.opcode;
1205 		cpl = &d->cpl[0];
1206 
1207 		switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
1208 		case X_RSPD_TYPE_FLBUF:
1209 
1210 			/* fall through */
1211 
1212 		case X_RSPD_TYPE_CPL:
1213 			MPASS(opcode < NUM_CPL_CMDS);
1214 
1215 			switch (opcode) {
1216 			case CPL_FW4_MSG:
1217 			case CPL_FW6_MSG:
1218 				cpl = unwrap_nm_fw6_msg(cpl);
1219 				/* fall through */
1220 			case CPL_SGE_EGR_UPDATE:
1221 				handle_nm_sge_egr_update(sc, ifp, cpl);
1222 				break;
1223 			case CPL_RX_PKT:
1224 				ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
1225 				    sc->params.sge.fl_pktshift;
1226 				ring->slot[fl_cidx].flags = 0;
1227 				nframes++;
1228 				if (!(lq & F_RSPD_NEWBUF)) {
1229 					MPASS(black_hole == 2);
1230 					break;
1231 				}
1232 				fl_credits++;
1233 				if (__predict_false(++fl_cidx == nm_rxq->fl_sidx))
1234 					fl_cidx = 0;
1235 				break;
1236 			default:
1237 				panic("%s: unexpected opcode 0x%x on nm_rxq %p",
1238 				    __func__, opcode, nm_rxq);
1239 			}
1240 			break;
1241 
1242 		case X_RSPD_TYPE_INTR:
1243 			/* Not equipped to handle forwarded interrupts. */
1244 			panic("%s: netmap queue received interrupt for iq %u\n",
1245 			    __func__, lq);
1246 
1247 		default:
1248 			panic("%s: illegal response type %d on nm_rxq %p",
1249 			    __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
1250 		}
1251 
1252 		d++;
1253 		if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
1254 			nm_rxq->iq_cidx = 0;
1255 			d = &nm_rxq->iq_desc[0];
1256 			nm_rxq->iq_gen ^= F_RSPD_GEN;
1257 		}
1258 
1259 		if (__predict_false(++nframes == rx_nframes) && !black_hole) {
1260 			atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1261 			netmap_rx_irq(ifp, nm_rxq->nid, &work);
1262 			nframes = 0;
1263 		}
1264 
1265 		if (__predict_false(++ndesc == rx_ndesc)) {
1266 			if (black_hole && fl_credits >= 8) {
1267 				fl_credits /= 8;
1268 				IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
1269 				    nm_rxq->fl_sidx);
1270 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1271 				    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1272 				fl_credits = fl_cidx & 7;
1273 			}
1274 			t4_write_reg(sc, sc->sge_gts_reg,
1275 			    V_CIDXINC(ndesc) |
1276 			    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
1277 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1278 			ndesc = 0;
1279 		}
1280 	}
1281 
1282 	atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1283 	if (black_hole) {
1284 		fl_credits /= 8;
1285 		IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1286 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1287 		    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1288 	} else if (nframes > 0)
1289 		netmap_rx_irq(ifp, nm_rxq->nid, &work);
1290 
1291     	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndesc) |
1292 	    V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1293 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1294 }
1295 #endif
1296