xref: /freebsd/sys/dev/cxgbe/t4_netmap.c (revision d30a1689f5b37e78ea189232a8b94a7011dc0dc8)
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 free_nm_rxq_hwq(struct vi_info *, struct sge_nm_rxq *);
124 static int free_nm_txq_hwq(struct vi_info *, struct sge_nm_txq *);
125 
126 int
127 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
128     int idx)
129 {
130 	int rc;
131 	struct sysctl_oid *oid;
132 	struct sysctl_oid_list *children;
133 	struct sysctl_ctx_list *ctx;
134 	char name[16];
135 	size_t len;
136 	struct adapter *sc = vi->adapter;
137 	struct netmap_adapter *na = NA(vi->ifp);
138 
139 	MPASS(na != NULL);
140 
141 	len = vi->qsize_rxq * IQ_ESIZE;
142 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
143 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
144 	if (rc != 0)
145 		return (rc);
146 
147 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
148 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
149 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
150 	if (rc != 0)
151 		return (rc);
152 
153 	nm_rxq->vi = vi;
154 	nm_rxq->nid = idx;
155 	nm_rxq->iq_cidx = 0;
156 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
157 	nm_rxq->iq_gen = F_RSPD_GEN;
158 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
159 	nm_rxq->fl_sidx = na->num_rx_desc;
160 	nm_rxq->fl_sidx2 = nm_rxq->fl_sidx;	/* copy for rxsync cacheline */
161 	nm_rxq->intr_idx = intr_idx;
162 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
163 
164 	ctx = &vi->ctx;
165 	children = SYSCTL_CHILDREN(vi->nm_rxq_oid);
166 
167 	snprintf(name, sizeof(name), "%d", idx);
168 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name,
169 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queue");
170 	children = SYSCTL_CHILDREN(oid);
171 
172 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
173 	    &nm_rxq->iq_abs_id, 0, "absolute id of the queue");
174 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
175 	    &nm_rxq->iq_cntxt_id, 0, "SGE context id of the queue");
176 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
177 	    &nm_rxq->iq_cidx, 0, "consumer index");
178 
179 	children = SYSCTL_CHILDREN(oid);
180 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl",
181 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist");
182 	children = SYSCTL_CHILDREN(oid);
183 
184 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
185 	    &nm_rxq->fl_cntxt_id, 0, "SGE context id of the freelist");
186 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
187 	    &nm_rxq->fl_cidx, 0, "consumer index");
188 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
189 	    &nm_rxq->fl_pidx, 0, "producer index");
190 
191 	return (rc);
192 }
193 
194 int
195 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
196 {
197 	struct adapter *sc = vi->adapter;
198 
199 	if (!(vi->flags & VI_INIT_DONE))
200 		return (0);
201 
202 	if (nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID)
203 		free_nm_rxq_hwq(vi, nm_rxq);
204 	MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
205 
206 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
207 	    nm_rxq->iq_desc);
208 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
209 	    nm_rxq->fl_desc);
210 
211 	return (0);
212 }
213 
214 int
215 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx)
216 {
217 	int rc;
218 	size_t len;
219 	struct port_info *pi = vi->pi;
220 	struct adapter *sc = pi->adapter;
221 	struct netmap_adapter *na = NA(vi->ifp);
222 	char name[16];
223 	struct sysctl_oid *oid;
224 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(vi->nm_txq_oid);
225 
226 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
227 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
228 	    &nm_txq->ba, (void **)&nm_txq->desc);
229 	if (rc)
230 		return (rc);
231 
232 	nm_txq->pidx = nm_txq->cidx = 0;
233 	nm_txq->sidx = na->num_tx_desc;
234 	nm_txq->nid = idx;
235 	nm_txq->iqidx = iqidx;
236 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
237 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
238 	    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
239 	if (sc->params.fw_vers >= FW_VERSION32(1, 24, 11, 0))
240 		nm_txq->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR));
241 	else
242 		nm_txq->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
243 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
244 
245 	snprintf(name, sizeof(name), "%d", idx);
246 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name,
247 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "netmap tx queue");
248 	children = SYSCTL_CHILDREN(oid);
249 
250 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
251 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
252 	SYSCTL_ADD_U16(&vi->ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
253 	    &nm_txq->cidx, 0, "consumer index");
254 	SYSCTL_ADD_U16(&vi->ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
255 	    &nm_txq->pidx, 0, "producer index");
256 
257 	return (rc);
258 }
259 
260 int
261 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
262 {
263 	struct adapter *sc = vi->adapter;
264 
265 	if (!(vi->flags & VI_INIT_DONE))
266 		return (0);
267 
268 	if (nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID)
269 		free_nm_txq_hwq(vi, nm_txq);
270 	MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
271 
272 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
273 	    nm_txq->desc);
274 
275 	return (0);
276 }
277 
278 static int
279 alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
280 {
281 	int rc, cntxt_id;
282 	__be32 v;
283 	struct adapter *sc = vi->adapter;
284 	struct port_info *pi = vi->pi;
285 	struct sge_params *sp = &sc->params.sge;
286 	struct netmap_adapter *na = NA(vi->ifp);
287 	struct fw_iq_cmd c;
288 	const int cong_drop = nm_cong_drop;
289 	const int cong_map = pi->rx_e_chan_map;
290 
291 	MPASS(na != NULL);
292 	MPASS(nm_rxq->iq_desc != NULL);
293 	MPASS(nm_rxq->fl_desc != NULL);
294 
295 	bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE);
296 	bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + sp->spg_len);
297 
298 	bzero(&c, sizeof(c));
299 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
300 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
301 	    V_FW_IQ_CMD_VFN(0));
302 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_IQSTART | FW_LEN16(c));
303 	if (nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID)
304 		c.alloc_to_len16 |= htobe32(F_FW_IQ_CMD_ALLOC);
305 	else {
306 		c.iqid = htobe16(nm_rxq->iq_cntxt_id);
307 		c.fl0id = htobe16(nm_rxq->fl_cntxt_id);
308 		c.fl1id = htobe16(0xffff);
309 		c.physiqid = htobe16(nm_rxq->iq_abs_id);
310 	}
311 	MPASS(!forwarding_intr_to_fwq(sc));
312 	KASSERT(nm_rxq->intr_idx < sc->intr_count,
313 	    ("%s: invalid direct intr_idx %d", __func__, nm_rxq->intr_idx));
314 	v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
315 	c.type_to_iqandstindex = htobe32(v |
316 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
317 	    V_FW_IQ_CMD_VIID(vi->viid) |
318 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
319 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
320 	    F_FW_IQ_CMD_IQGTSMODE |
321 	    V_FW_IQ_CMD_IQINTCNTTHRESH(0) |
322 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
323 	c.iqsize = htobe16(vi->qsize_rxq);
324 	c.iqaddr = htobe64(nm_rxq->iq_ba);
325 	if (cong_drop != -1) {
326 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN |
327 		    V_FW_IQ_CMD_FL0CNGCHMAP(cong_map) | F_FW_IQ_CMD_FL0CONGCIF |
328 		    F_FW_IQ_CMD_FL0CONGEN);
329 	}
330 	c.iqns_to_fl0congen |=
331 	    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
332 		V_FW_IQ_CMD_IQTYPE(FW_IQ_IQTYPE_NIC) |
333 		F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
334 		(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
335 		(black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0));
336 	c.fl0dcaen_to_fl0cidxfthresh =
337 	    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
338 		X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
339 		V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
340 		X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
341 	c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE);
342 	c.fl0addr = htobe64(nm_rxq->fl_ba);
343 
344 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
345 	if (rc != 0) {
346 		device_printf(sc->dev,
347 		    "failed to create netmap ingress queue: %d\n", rc);
348 		return (rc);
349 	}
350 
351 	nm_rxq->iq_cidx = 0;
352 	MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE);
353 	nm_rxq->iq_gen = F_RSPD_GEN;
354 	nm_rxq->iq_cntxt_id = be16toh(c.iqid);
355 	nm_rxq->iq_abs_id = be16toh(c.physiqid);
356 	cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
357 	if (cntxt_id >= sc->sge.iqmap_sz) {
358 		panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
359 		    __func__, cntxt_id, sc->sge.iqmap_sz - 1);
360 	}
361 	sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
362 
363 	nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
364 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
365 	nm_rxq->fl_db_saved = 0;
366 	/* matches the X_FETCHBURSTMAX_512B or X_FETCHBURSTMAX_256B above. */
367 	nm_rxq->fl_db_threshold = chip_id(sc) <= CHELSIO_T5 ? 8 : 4;
368 	MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
369 	cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
370 	if (cntxt_id >= sc->sge.eqmap_sz) {
371 		panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
372 		    __func__, cntxt_id, sc->sge.eqmap_sz - 1);
373 	}
374 	sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
375 
376 	nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) |
377 	    sc->chip_params->sge_fl_db;
378 
379 	if (chip_id(sc) >= CHELSIO_T5 && cong_drop != -1) {
380 		t4_sge_set_conm_context(sc, nm_rxq->iq_cntxt_id, cong_drop,
381 		    cong_map);
382 	}
383 
384 	t4_write_reg(sc, sc->sge_gts_reg,
385 	    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
386 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
387 
388 	return (rc);
389 }
390 
391 static int
392 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
393 {
394 	struct adapter *sc = vi->adapter;
395 	int rc;
396 
397 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
398 	    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
399 	if (rc != 0)
400 		device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
401 		    __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
402 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
403 	return (rc);
404 }
405 
406 static int
407 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
408 {
409 	int rc, cntxt_id;
410 	size_t len;
411 	struct adapter *sc = vi->adapter;
412 	struct netmap_adapter *na = NA(vi->ifp);
413 	struct fw_eq_eth_cmd c;
414 
415 	MPASS(na != NULL);
416 	MPASS(nm_txq->desc != NULL);
417 
418 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
419 	bzero(nm_txq->desc, len);
420 
421 	bzero(&c, sizeof(c));
422 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
423 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
424 	    V_FW_EQ_ETH_CMD_VFN(0));
425 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
426 	if (nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID)
427 		c.alloc_to_len16 |= htobe32(F_FW_EQ_ETH_CMD_ALLOC);
428 	else
429 		c.eqid_pkd = htobe32(V_FW_EQ_ETH_CMD_EQID(nm_txq->cntxt_id));
430 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
431 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
432 	c.fetchszm_to_iqid =
433 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
434 		V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
435 		V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
436 	c.dcaen_to_eqsize =
437 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
438 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
439 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
440 		V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
441 	c.eqaddr = htobe64(nm_txq->ba);
442 
443 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
444 	if (rc != 0) {
445 		device_printf(vi->dev,
446 		    "failed to create netmap egress queue: %d\n", rc);
447 		return (rc);
448 	}
449 
450 	nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
451 	cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
452 	if (cntxt_id >= sc->sge.eqmap_sz)
453 	    panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
454 		cntxt_id, sc->sge.eqmap_sz - 1);
455 	sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
456 
457 	nm_txq->pidx = nm_txq->cidx = 0;
458 	MPASS(nm_txq->sidx == na->num_tx_desc);
459 	nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
460 
461 	nm_txq->doorbells = sc->doorbells;
462 	if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
463 	    isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
464 	    isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
465 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
466 		uint32_t mask = (1 << s_qpp) - 1;
467 		volatile uint8_t *udb;
468 
469 		udb = sc->udbs_base + UDBS_DB_OFFSET;
470 		udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
471 		nm_txq->udb_qid = nm_txq->cntxt_id & mask;
472 		if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
473 	    		clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
474 		else {
475 			udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
476 			nm_txq->udb_qid = 0;
477 		}
478 		nm_txq->udb = (volatile void *)udb;
479 	}
480 
481 	if (sc->params.fw_vers < FW_VERSION32(1, 25, 1, 0)) {
482 		uint32_t param, val;
483 
484 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
485 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) |
486 		    V_FW_PARAMS_PARAM_YZ(nm_txq->cntxt_id);
487 		val = 0xff;
488 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
489 		if (rc != 0) {
490 			device_printf(vi->dev,
491 			    "failed to bind netmap txq %d to class 0xff: %d\n",
492 			    nm_txq->cntxt_id, rc);
493 			rc = 0;
494 		}
495 	}
496 
497 	return (rc);
498 }
499 
500 static int
501 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
502 {
503 	struct adapter *sc = vi->adapter;
504 	int rc;
505 
506 	rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
507 	if (rc != 0)
508 		device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
509 		    nm_txq->cntxt_id, rc);
510 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
511 	return (rc);
512 }
513 
514 static int
515 cxgbe_netmap_simple_rss(struct adapter *sc, struct vi_info *vi,
516     struct ifnet *ifp, struct netmap_adapter *na)
517 {
518 	struct netmap_kring *kring;
519 	struct sge_nm_rxq *nm_rxq;
520 	int rc, i, j, nm_state, defq;
521 	uint16_t *rss;
522 
523 	/*
524 	 * Check if there's at least one active (or about to go active) netmap
525 	 * rx queue.
526 	 */
527 	defq = -1;
528 	for_each_nm_rxq(vi, j, nm_rxq) {
529 		nm_state = atomic_load_int(&nm_rxq->nm_state);
530 		kring = na->rx_rings[nm_rxq->nid];
531 		if ((nm_state != NM_OFF && !nm_kring_pending_off(kring)) ||
532 		    (nm_state == NM_OFF && nm_kring_pending_on(kring))) {
533 			MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
534 			if (defq == -1) {
535 				defq = nm_rxq->iq_abs_id;
536 				break;
537 			}
538 		}
539 	}
540 
541 	if (defq == -1) {
542 		/* No active netmap queues.  Switch back to NIC queues. */
543 		rss = vi->rss;
544 		defq = vi->rss[0];
545 	} else {
546 		for (i = 0; i < vi->rss_size;) {
547 			for_each_nm_rxq(vi, j, nm_rxq) {
548 				nm_state = atomic_load_int(&nm_rxq->nm_state);
549 				kring = na->rx_rings[nm_rxq->nid];
550 				if ((nm_state != NM_OFF &&
551 				    !nm_kring_pending_off(kring)) ||
552 				    (nm_state == NM_OFF &&
553 				    nm_kring_pending_on(kring))) {
554 					MPASS(nm_rxq->iq_cntxt_id !=
555 					    INVALID_NM_RXQ_CNTXT_ID);
556 					vi->nm_rss[i++] = nm_rxq->iq_abs_id;
557 					if (i == vi->rss_size)
558 						break;
559 				}
560 			}
561 		}
562 		rss = vi->nm_rss;
563 	}
564 
565 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size, rss,
566 	    vi->rss_size);
567 	if (rc != 0)
568 		if_printf(ifp, "netmap rss_config failed: %d\n", rc);
569 
570 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
571 	if (rc != 0) {
572 		if_printf(ifp, "netmap defaultq config failed: %d\n", rc);
573 	}
574 
575 	return (rc);
576 }
577 
578 /*
579  * Odd number of rx queues work best for split RSS mode as the first queue can
580  * be dedicated for non-RSS traffic and the rest divided into two equal halves.
581  */
582 static int
583 cxgbe_netmap_split_rss(struct adapter *sc, struct vi_info *vi,
584     struct ifnet *ifp, struct netmap_adapter *na)
585 {
586 	struct netmap_kring *kring;
587 	struct sge_nm_rxq *nm_rxq;
588 	int rc, i, j, nm_state, defq;
589 	int nactive[2] = {0, 0};
590 	int dq[2] = {-1, -1};
591 	bool dq_norss;		/* default queue should not be in RSS table. */
592 
593 	MPASS(nm_split_rss != 0);
594 	MPASS(vi->nnmrxq > 1);
595 
596 	for_each_nm_rxq(vi, i, nm_rxq) {
597 		j = i / ((vi->nnmrxq + 1) / 2);
598 		nm_state = atomic_load_int(&nm_rxq->nm_state);
599 		kring = na->rx_rings[nm_rxq->nid];
600 		if ((nm_state != NM_OFF && !nm_kring_pending_off(kring)) ||
601 		    (nm_state == NM_OFF && nm_kring_pending_on(kring))) {
602 			MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
603 			nactive[j]++;
604 			if (dq[j] == -1) {
605 				dq[j] = nm_rxq->iq_abs_id;
606 				break;
607 			}
608 		}
609 	}
610 
611 	if (nactive[0] == 0 || nactive[1] == 0)
612 		return (cxgbe_netmap_simple_rss(sc, vi, ifp, na));
613 
614 	MPASS(dq[0] != -1 && dq[1] != -1);
615 	if (nactive[0] > nactive[1]) {
616 		defq = dq[0];
617 		dq_norss = true;
618 	} else if (nactive[0] < nactive[1]) {
619 		defq = dq[1];
620 		dq_norss = true;
621 	} else {
622 		defq = dq[0];
623 		dq_norss = false;
624 	}
625 
626 	i = 0;
627 	nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq];
628 	while (i < vi->rss_size / 2) {
629 		for (j = 0; j < (vi->nnmrxq + 1) / 2; j++) {
630 			nm_state = atomic_load_int(&nm_rxq[j].nm_state);
631 			kring = na->rx_rings[nm_rxq[j].nid];
632 			if ((nm_state == NM_OFF &&
633 			    !nm_kring_pending_on(kring)) ||
634 			    (nm_state == NM_ON &&
635 			    nm_kring_pending_off(kring))) {
636 				continue;
637 			}
638 			MPASS(nm_rxq[j].iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
639 			if (dq_norss && defq == nm_rxq[j].iq_abs_id)
640 				continue;
641 			vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
642 			if (i == vi->rss_size / 2)
643 				break;
644 		}
645 	}
646 	while (i < vi->rss_size) {
647 		for (j = (vi->nnmrxq + 1) / 2; j < vi->nnmrxq; j++) {
648 			nm_state = atomic_load_int(&nm_rxq[j].nm_state);
649 			kring = na->rx_rings[nm_rxq[j].nid];
650 			if ((nm_state == NM_OFF &&
651 			    !nm_kring_pending_on(kring)) ||
652 			    (nm_state == NM_ON &&
653 			    nm_kring_pending_off(kring))) {
654 				continue;
655 			}
656 			MPASS(nm_rxq[j].iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
657 			if (dq_norss && defq == nm_rxq[j].iq_abs_id)
658 				continue;
659 			vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
660 			if (i == vi->rss_size)
661 				break;
662 		}
663 	}
664 
665 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
666 	    vi->nm_rss, vi->rss_size);
667 	if (rc != 0)
668 		if_printf(ifp, "netmap split_rss_config failed: %d\n", rc);
669 
670 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
671 	if (rc != 0)
672 		if_printf(ifp, "netmap defaultq config failed: %d\n", rc);
673 
674 	return (rc);
675 }
676 
677 static inline int
678 cxgbe_netmap_rss(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
679     struct netmap_adapter *na)
680 {
681 
682 	if (nm_split_rss == 0 || vi->nnmrxq == 1)
683 		return (cxgbe_netmap_simple_rss(sc, vi, ifp, na));
684 	else
685 		return (cxgbe_netmap_split_rss(sc, vi, ifp, na));
686 }
687 
688 static int
689 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
690     struct netmap_adapter *na)
691 {
692 	struct netmap_slot *slot;
693 	struct netmap_kring *kring;
694 	struct sge_nm_rxq *nm_rxq;
695 	struct sge_nm_txq *nm_txq;
696 	int i, j, hwidx;
697 	struct rx_buf_info *rxb;
698 
699 	ASSERT_SYNCHRONIZED_OP(sc);
700 	MPASS(vi->nnmrxq > 0);
701 	MPASS(vi->nnmtxq > 0);
702 
703 	if ((vi->flags & VI_INIT_DONE) == 0 ||
704 	    (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
705 		if_printf(ifp, "cannot enable netmap operation because "
706 		    "interface is not UP.\n");
707 		return (EAGAIN);
708 	}
709 
710 	rxb = &sc->sge.rx_buf_info[0];
711 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
712 		if (rxb->size1 == NETMAP_BUF_SIZE(na)) {
713 			hwidx = rxb->hwidx1;
714 			break;
715 		}
716 		if (rxb->size2 == NETMAP_BUF_SIZE(na)) {
717 			hwidx = rxb->hwidx2;
718 			break;
719 		}
720 	}
721 	if (i >= SW_ZONE_SIZES) {
722 		if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
723 		    NETMAP_BUF_SIZE(na));
724 		return (ENXIO);
725 	}
726 
727 	/* Must set caps before calling netmap_reset */
728 	nm_set_native_flags(na);
729 
730 	for_each_nm_rxq(vi, i, nm_rxq) {
731 		kring = na->rx_rings[nm_rxq->nid];
732 		if (!nm_kring_pending_on(kring))
733 			continue;
734 
735 		alloc_nm_rxq_hwq(vi, nm_rxq);
736 		nm_rxq->fl_hwidx = hwidx;
737 		slot = netmap_reset(na, NR_RX, i, 0);
738 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
739 
740 		/* We deal with 8 bufs at a time */
741 		MPASS((na->num_rx_desc & 7) == 0);
742 		MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
743 		for (j = 0; j < nm_rxq->fl_sidx; j++) {
744 			uint64_t ba;
745 
746 			PNMB(na, &slot[j], &ba);
747 			MPASS(ba != 0);
748 			nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
749 		}
750 		j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
751 		MPASS((j & 7) == 0);
752 		j /= 8;	/* driver pidx to hardware pidx */
753 		wmb();
754 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
755 		    nm_rxq->fl_db_val | V_PIDX(j));
756 
757 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_OFF, NM_ON);
758 	}
759 
760 	for_each_nm_txq(vi, i, nm_txq) {
761 		kring = na->tx_rings[nm_txq->nid];
762 		if (!nm_kring_pending_on(kring))
763 			continue;
764 
765 		alloc_nm_txq_hwq(vi, nm_txq);
766 		slot = netmap_reset(na, NR_TX, i, 0);
767 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
768 	}
769 
770 	if (vi->nm_rss == NULL) {
771 		vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
772 		    M_ZERO | M_WAITOK);
773 	}
774 
775 	return (cxgbe_netmap_rss(sc, vi, ifp, na));
776 }
777 
778 static int
779 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
780     struct netmap_adapter *na)
781 {
782 	struct netmap_kring *kring;
783 	int rc, i, nm_state, nactive;
784 	struct sge_nm_txq *nm_txq;
785 	struct sge_nm_rxq *nm_rxq;
786 
787 	ASSERT_SYNCHRONIZED_OP(sc);
788 	MPASS(vi->nnmrxq > 0);
789 	MPASS(vi->nnmtxq > 0);
790 
791 	if (!nm_netmap_on(na))
792 		return (0);
793 
794 	if ((vi->flags & VI_INIT_DONE) == 0)
795 		return (0);
796 
797 	/* First remove the queues that are stopping from the RSS table. */
798 	rc = cxgbe_netmap_rss(sc, vi, ifp, na);
799 	if (rc != 0)
800 		return (rc);	/* error message logged already. */
801 
802 	for_each_nm_txq(vi, i, nm_txq) {
803 		kring = na->tx_rings[nm_txq->nid];
804 		if (!nm_kring_pending_off(kring))
805 			continue;
806 		MPASS(nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID);
807 
808 		rc = -t4_eth_eq_stop(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
809 		if (rc != 0) {
810 			device_printf(vi->dev,
811 			    "failed to stop nm_txq[%d]: %d.\n", i, rc);
812 			return (rc);
813 		}
814 
815 		/* XXX: netmap, not the driver, should do this. */
816 		kring->rhead = kring->rcur = kring->nr_hwcur = 0;
817 		kring->rtail = kring->nr_hwtail = kring->nkr_num_slots - 1;
818 	}
819 	nactive = 0;
820 	for_each_nm_rxq(vi, i, nm_rxq) {
821 		nm_state = atomic_load_int(&nm_rxq->nm_state);
822 		kring = na->rx_rings[nm_rxq->nid];
823 		if (nm_state != NM_OFF && !nm_kring_pending_off(kring))
824 			nactive++;
825 		if (!nm_kring_pending_off(kring))
826 			continue;
827 		MPASS(nm_state != NM_OFF);
828 		MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
829 
830 		rc = -t4_iq_stop(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
831 		    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
832 		if (rc != 0) {
833 			device_printf(vi->dev,
834 			    "failed to stop nm_rxq[%d]: %d.\n", i, rc);
835 			return (rc);
836 		}
837 
838 		while (!atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_OFF))
839 			pause("nmst", 1);
840 
841 		/* XXX: netmap, not the driver, should do this. */
842 		kring->rhead = kring->rcur = kring->nr_hwcur = 0;
843 		kring->rtail = kring->nr_hwtail = 0;
844 	}
845 	netmap_krings_mode_commit(na, 0);
846 	if (nactive == 0)
847 		nm_clear_native_flags(na);
848 
849 	return (rc);
850 }
851 
852 static int
853 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
854 {
855 	struct ifnet *ifp = na->ifp;
856 	struct vi_info *vi = ifp->if_softc;
857 	struct adapter *sc = vi->adapter;
858 	int rc;
859 
860 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
861 	if (rc != 0)
862 		return (rc);
863 	if (on)
864 		rc = cxgbe_netmap_on(sc, vi, ifp, na);
865 	else
866 		rc = cxgbe_netmap_off(sc, vi, ifp, na);
867 	end_synchronized_op(sc, 0);
868 
869 	return (rc);
870 }
871 
872 /* How many packets can a single type1 WR carry in n descriptors */
873 static inline int
874 ndesc_to_npkt(const int n)
875 {
876 
877 	MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
878 
879 	return (n * 2 - 1);
880 }
881 #define MAX_NPKT_IN_TYPE1_WR	(ndesc_to_npkt(SGE_MAX_WR_NDESC))
882 
883 /*
884  * Space (in descriptors) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
885  * carries n packets
886  */
887 static inline int
888 npkt_to_ndesc(const int n)
889 {
890 
891 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
892 
893 	return ((n + 2) / 2);
894 }
895 
896 /*
897  * Space (in 16B units) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
898  * carries n packets
899  */
900 static inline int
901 npkt_to_len16(const int n)
902 {
903 
904 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
905 
906 	return (n * 2 + 1);
907 }
908 
909 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
910 
911 static void
912 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
913 {
914 	int n;
915 	u_int db = nm_txq->doorbells;
916 
917 	MPASS(nm_txq->pidx != nm_txq->dbidx);
918 
919 	n = NMIDXDIFF(nm_txq, dbidx);
920 	if (n > 1)
921 		clrbit(&db, DOORBELL_WCWR);
922 	wmb();
923 
924 	switch (ffs(db) - 1) {
925 	case DOORBELL_UDB:
926 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
927 		break;
928 
929 	case DOORBELL_WCWR: {
930 		volatile uint64_t *dst, *src;
931 
932 		/*
933 		 * Queues whose 128B doorbell segment fits in the page do not
934 		 * use relative qid (udb_qid is always 0).  Only queues with
935 		 * doorbell segments can do WCWR.
936 		 */
937 		KASSERT(nm_txq->udb_qid == 0 && n == 1,
938 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
939 		    __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
940 
941 		dst = (volatile void *)((uintptr_t)nm_txq->udb +
942 		    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
943 		src = (void *)&nm_txq->desc[nm_txq->dbidx];
944 		while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
945 			*dst++ = *src++;
946 		wmb();
947 		break;
948 	}
949 
950 	case DOORBELL_UDBWC:
951 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
952 		wmb();
953 		break;
954 
955 	case DOORBELL_KDB:
956 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
957 		    V_QID(nm_txq->cntxt_id) | V_PIDX(n));
958 		break;
959 	}
960 	nm_txq->dbidx = nm_txq->pidx;
961 }
962 
963 /*
964  * Write work requests to send 'npkt' frames and ring the doorbell to send them
965  * on their way.  No need to check for wraparound.
966  */
967 static void
968 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
969     struct netmap_kring *kring, int npkt, int npkt_remaining)
970 {
971 	struct netmap_ring *ring = kring->ring;
972 	struct netmap_slot *slot;
973 	const u_int lim = kring->nkr_num_slots - 1;
974 	struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
975 	uint16_t len;
976 	uint64_t ba;
977 	struct cpl_tx_pkt_core *cpl;
978 	struct ulptx_sgl *usgl;
979 	int i, n;
980 
981 	while (npkt) {
982 		n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
983 		len = 0;
984 
985 		wr = (void *)&nm_txq->desc[nm_txq->pidx];
986 		wr->op_pkd = nm_txq->op_pkd;
987 		wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
988 		wr->npkt = n;
989 		wr->r3 = 0;
990 		wr->type = 1;
991 		cpl = (void *)(wr + 1);
992 
993 		for (i = 0; i < n; i++) {
994 			slot = &ring->slot[kring->nr_hwcur];
995 			PNMB(kring->na, slot, &ba);
996 			MPASS(ba != 0);
997 
998 			cpl->ctrl0 = nm_txq->cpl_ctrl0;
999 			cpl->pack = 0;
1000 			cpl->len = htobe16(slot->len);
1001 			cpl->ctrl1 = nm_txcsum ? 0 :
1002 			    htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
1003 
1004 			usgl = (void *)(cpl + 1);
1005 			usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1006 			    V_ULPTX_NSGE(1));
1007 			usgl->len0 = htobe32(slot->len);
1008 			usgl->addr0 = htobe64(ba);
1009 
1010 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1011 			cpl = (void *)(usgl + 1);
1012 			MPASS(slot->len + len <= UINT16_MAX);
1013 			len += slot->len;
1014 			kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
1015 		}
1016 		wr->plen = htobe16(len);
1017 
1018 		npkt -= n;
1019 		nm_txq->pidx += npkt_to_ndesc(n);
1020 		MPASS(nm_txq->pidx <= nm_txq->sidx);
1021 		if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
1022 			/*
1023 			 * This routine doesn't know how to write WRs that wrap
1024 			 * around.  Make sure it wasn't asked to.
1025 			 */
1026 			MPASS(npkt == 0);
1027 			nm_txq->pidx = 0;
1028 		}
1029 
1030 		if (npkt == 0 && npkt_remaining == 0) {
1031 			/* All done. */
1032 			if (lazy_tx_credit_flush == 0) {
1033 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
1034 				    F_FW_WR_EQUIQ);
1035 				nm_txq->equeqidx = nm_txq->pidx;
1036 				nm_txq->equiqidx = nm_txq->pidx;
1037 			}
1038 			ring_nm_txq_db(sc, nm_txq);
1039 			return;
1040 		}
1041 
1042 		if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
1043 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
1044 			    F_FW_WR_EQUIQ);
1045 			nm_txq->equeqidx = nm_txq->pidx;
1046 			nm_txq->equiqidx = nm_txq->pidx;
1047 		} else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
1048 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1049 			nm_txq->equeqidx = nm_txq->pidx;
1050 		}
1051 		if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
1052 			ring_nm_txq_db(sc, nm_txq);
1053 	}
1054 
1055 	/* Will get called again. */
1056 	MPASS(npkt_remaining);
1057 }
1058 
1059 /* How many contiguous free descriptors starting at pidx */
1060 static inline int
1061 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
1062 {
1063 
1064 	if (nm_txq->cidx > nm_txq->pidx)
1065 		return (nm_txq->cidx - nm_txq->pidx - 1);
1066 	else if (nm_txq->cidx > 0)
1067 		return (nm_txq->sidx - nm_txq->pidx);
1068 	else
1069 		return (nm_txq->sidx - nm_txq->pidx - 1);
1070 }
1071 
1072 static int
1073 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
1074 {
1075 	struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
1076 	uint16_t hw_cidx = spg->cidx;	/* snapshot */
1077 	struct fw_eth_tx_pkts_wr *wr;
1078 	int n = 0;
1079 
1080 	hw_cidx = be16toh(hw_cidx);
1081 
1082 	while (nm_txq->cidx != hw_cidx) {
1083 		wr = (void *)&nm_txq->desc[nm_txq->cidx];
1084 
1085 		MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)) ||
1086 		    wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR)));
1087 		MPASS(wr->type == 1);
1088 		MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
1089 
1090 		n += wr->npkt;
1091 		nm_txq->cidx += npkt_to_ndesc(wr->npkt);
1092 
1093 		/*
1094 		 * We never sent a WR that wrapped around so the credits coming
1095 		 * back, WR by WR, should never cause the cidx to wrap around
1096 		 * either.
1097 		 */
1098 		MPASS(nm_txq->cidx <= nm_txq->sidx);
1099 		if (__predict_false(nm_txq->cidx == nm_txq->sidx))
1100 			nm_txq->cidx = 0;
1101 	}
1102 
1103 	return (n);
1104 }
1105 
1106 static int
1107 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
1108 {
1109 	struct netmap_adapter *na = kring->na;
1110 	struct ifnet *ifp = na->ifp;
1111 	struct vi_info *vi = ifp->if_softc;
1112 	struct adapter *sc = vi->adapter;
1113 	struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
1114 	const u_int head = kring->rhead;
1115 	u_int reclaimed = 0;
1116 	int n, d, npkt_remaining, ndesc_remaining;
1117 
1118 	/*
1119 	 * Tx was at kring->nr_hwcur last time around and now we need to advance
1120 	 * to kring->rhead.  Note that the driver's pidx moves independent of
1121 	 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
1122 	 * between descriptors and frames isn't 1:1).
1123 	 */
1124 
1125 	npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
1126 	    kring->nkr_num_slots - kring->nr_hwcur + head;
1127 	while (npkt_remaining) {
1128 		reclaimed += reclaim_nm_tx_desc(nm_txq);
1129 		ndesc_remaining = contiguous_ndesc_available(nm_txq);
1130 		/* Can't run out of descriptors with packets still remaining */
1131 		MPASS(ndesc_remaining > 0);
1132 
1133 		/* # of desc needed to tx all remaining packets */
1134 		d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
1135 		if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
1136 			d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
1137 
1138 		if (d <= ndesc_remaining)
1139 			n = npkt_remaining;
1140 		else {
1141 			/* Can't send all, calculate how many can be sent */
1142 			n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
1143 			    MAX_NPKT_IN_TYPE1_WR;
1144 			if (ndesc_remaining % SGE_MAX_WR_NDESC)
1145 				n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
1146 		}
1147 
1148 		/* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
1149 		npkt_remaining -= n;
1150 		cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining);
1151 	}
1152 	MPASS(npkt_remaining == 0);
1153 	MPASS(kring->nr_hwcur == head);
1154 	MPASS(nm_txq->dbidx == nm_txq->pidx);
1155 
1156 	/*
1157 	 * Second part: reclaim buffers for completed transmissions.
1158 	 */
1159 	if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
1160 		reclaimed += reclaim_nm_tx_desc(nm_txq);
1161 		kring->nr_hwtail += reclaimed;
1162 		if (kring->nr_hwtail >= kring->nkr_num_slots)
1163 			kring->nr_hwtail -= kring->nkr_num_slots;
1164 	}
1165 
1166 	return (0);
1167 }
1168 
1169 static int
1170 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
1171 {
1172 	struct netmap_adapter *na = kring->na;
1173 	struct netmap_ring *ring = kring->ring;
1174 	struct ifnet *ifp = na->ifp;
1175 	struct vi_info *vi = ifp->if_softc;
1176 	struct adapter *sc = vi->adapter;
1177 	struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
1178 	u_int const head = kring->rhead;
1179 	u_int n;
1180 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1181 
1182 	if (black_hole)
1183 		return (0);	/* No updates ever. */
1184 
1185 	if (netmap_no_pendintr || force_update) {
1186 		kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
1187 		kring->nr_kflags &= ~NKR_PENDINTR;
1188 	}
1189 
1190 	if (nm_rxq->fl_db_saved > 0 && starve_fl == 0) {
1191 		wmb();
1192 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1193 		    nm_rxq->fl_db_val | V_PIDX(nm_rxq->fl_db_saved));
1194 		nm_rxq->fl_db_saved = 0;
1195 	}
1196 
1197 	/* Userspace done with buffers from kring->nr_hwcur to head */
1198 	n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
1199 	    kring->nkr_num_slots - kring->nr_hwcur + head;
1200 	n &= ~7U;
1201 	if (n > 0) {
1202 		u_int fl_pidx = nm_rxq->fl_pidx;
1203 		struct netmap_slot *slot = &ring->slot[fl_pidx];
1204 		uint64_t ba;
1205 		int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
1206 
1207 		/*
1208 		 * We always deal with 8 buffers at a time.  We must have
1209 		 * stopped at an 8B boundary (fl_pidx) last time around and we
1210 		 * must have a multiple of 8B buffers to give to the freelist.
1211 		 */
1212 		MPASS((fl_pidx & 7) == 0);
1213 		MPASS((n & 7) == 0);
1214 
1215 		IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
1216 		IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx2);
1217 
1218 		while (n > 0) {
1219 			for (i = 0; i < 8; i++, fl_pidx++, slot++) {
1220 				PNMB(na, slot, &ba);
1221 				MPASS(ba != 0);
1222 				nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
1223 				slot->flags &= ~NS_BUF_CHANGED;
1224 				MPASS(fl_pidx <= nm_rxq->fl_sidx2);
1225 			}
1226 			n -= 8;
1227 			if (fl_pidx == nm_rxq->fl_sidx2) {
1228 				fl_pidx = 0;
1229 				slot = &ring->slot[0];
1230 			}
1231 			if (++dbinc == nm_rxq->fl_db_threshold) {
1232 				wmb();
1233 				if (starve_fl)
1234 					nm_rxq->fl_db_saved += dbinc;
1235 				else {
1236 					t4_write_reg(sc, sc->sge_kdoorbell_reg,
1237 					    nm_rxq->fl_db_val | V_PIDX(dbinc));
1238 				}
1239 				dbinc = 0;
1240 			}
1241 		}
1242 		MPASS(nm_rxq->fl_pidx == fl_pidx);
1243 
1244 		if (dbinc > 0) {
1245 			wmb();
1246 			if (starve_fl)
1247 				nm_rxq->fl_db_saved += dbinc;
1248 			else {
1249 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1250 				    nm_rxq->fl_db_val | V_PIDX(dbinc));
1251 			}
1252 		}
1253 	}
1254 
1255 	return (0);
1256 }
1257 
1258 void
1259 cxgbe_nm_attach(struct vi_info *vi)
1260 {
1261 	struct port_info *pi;
1262 	struct adapter *sc;
1263 	struct netmap_adapter na;
1264 
1265 	MPASS(vi->nnmrxq > 0);
1266 	MPASS(vi->ifp != NULL);
1267 
1268 	pi = vi->pi;
1269 	sc = pi->adapter;
1270 
1271 	bzero(&na, sizeof(na));
1272 
1273 	na.ifp = vi->ifp;
1274 	na.na_flags = NAF_BDG_MAYSLEEP;
1275 
1276 	/* Netmap doesn't know about the space reserved for the status page. */
1277 	na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
1278 
1279 	/*
1280 	 * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
1281 	 * num_rx_desc is based on the number of buffers that can be held in the
1282 	 * freelist, and not the number of entries in the iq.  (These two are
1283 	 * not exactly the same due to the space taken up by the status page).
1284 	 */
1285 	na.num_rx_desc = rounddown(vi->qsize_rxq, 8);
1286 	na.nm_txsync = cxgbe_netmap_txsync;
1287 	na.nm_rxsync = cxgbe_netmap_rxsync;
1288 	na.nm_register = cxgbe_netmap_reg;
1289 	na.num_tx_rings = vi->nnmtxq;
1290 	na.num_rx_rings = vi->nnmrxq;
1291 	na.rx_buf_maxsize = MAX_MTU;
1292 	netmap_attach(&na);	/* This adds IFCAP_NETMAP to if_capabilities */
1293 }
1294 
1295 void
1296 cxgbe_nm_detach(struct vi_info *vi)
1297 {
1298 
1299 	MPASS(vi->nnmrxq > 0);
1300 	MPASS(vi->ifp != NULL);
1301 
1302 	netmap_detach(vi->ifp);
1303 }
1304 
1305 static inline const void *
1306 unwrap_nm_fw6_msg(const struct cpl_fw6_msg *cpl)
1307 {
1308 
1309 	MPASS(cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL);
1310 
1311 	/* data[0] is RSS header */
1312 	return (&cpl->data[1]);
1313 }
1314 
1315 static void
1316 handle_nm_sge_egr_update(struct adapter *sc, struct ifnet *ifp,
1317     const struct cpl_sge_egr_update *egr)
1318 {
1319 	uint32_t oq;
1320 	struct sge_nm_txq *nm_txq;
1321 
1322 	oq = be32toh(egr->opcode_qid);
1323 	MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
1324 	nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
1325 
1326 	netmap_tx_irq(ifp, nm_txq->nid);
1327 }
1328 
1329 void
1330 service_nm_rxq(struct sge_nm_rxq *nm_rxq)
1331 {
1332 	struct vi_info *vi = nm_rxq->vi;
1333 	struct adapter *sc = vi->adapter;
1334 	struct ifnet *ifp = vi->ifp;
1335 	struct netmap_adapter *na = NA(ifp);
1336 	struct netmap_kring *kring = na->rx_rings[nm_rxq->nid];
1337 	struct netmap_ring *ring = kring->ring;
1338 	struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
1339 	const void *cpl;
1340 	uint32_t lq;
1341 	u_int work = 0;
1342 	uint8_t opcode;
1343 	uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
1344 	u_int fl_credits = fl_cidx & 7;
1345 	u_int ndesc = 0;	/* desc processed since last cidx update */
1346 	u_int nframes = 0;	/* frames processed since last netmap wakeup */
1347 
1348 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
1349 
1350 		rmb();
1351 
1352 		lq = be32toh(d->rsp.pldbuflen_qid);
1353 		opcode = d->rss.opcode;
1354 		cpl = &d->cpl[0];
1355 
1356 		switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
1357 		case X_RSPD_TYPE_FLBUF:
1358 
1359 			/* fall through */
1360 
1361 		case X_RSPD_TYPE_CPL:
1362 			MPASS(opcode < NUM_CPL_CMDS);
1363 
1364 			switch (opcode) {
1365 			case CPL_FW4_MSG:
1366 			case CPL_FW6_MSG:
1367 				cpl = unwrap_nm_fw6_msg(cpl);
1368 				/* fall through */
1369 			case CPL_SGE_EGR_UPDATE:
1370 				handle_nm_sge_egr_update(sc, ifp, cpl);
1371 				break;
1372 			case CPL_RX_PKT:
1373 				ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
1374 				    sc->params.sge.fl_pktshift;
1375 				ring->slot[fl_cidx].flags = 0;
1376 				nframes++;
1377 				if (!(lq & F_RSPD_NEWBUF)) {
1378 					MPASS(black_hole == 2);
1379 					break;
1380 				}
1381 				fl_credits++;
1382 				if (__predict_false(++fl_cidx == nm_rxq->fl_sidx))
1383 					fl_cidx = 0;
1384 				break;
1385 			default:
1386 				panic("%s: unexpected opcode 0x%x on nm_rxq %p",
1387 				    __func__, opcode, nm_rxq);
1388 			}
1389 			break;
1390 
1391 		case X_RSPD_TYPE_INTR:
1392 			/* Not equipped to handle forwarded interrupts. */
1393 			panic("%s: netmap queue received interrupt for iq %u\n",
1394 			    __func__, lq);
1395 
1396 		default:
1397 			panic("%s: illegal response type %d on nm_rxq %p",
1398 			    __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
1399 		}
1400 
1401 		d++;
1402 		if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
1403 			nm_rxq->iq_cidx = 0;
1404 			d = &nm_rxq->iq_desc[0];
1405 			nm_rxq->iq_gen ^= F_RSPD_GEN;
1406 		}
1407 
1408 		if (__predict_false(++nframes == rx_nframes) && !black_hole) {
1409 			atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1410 			netmap_rx_irq(ifp, nm_rxq->nid, &work);
1411 			nframes = 0;
1412 		}
1413 
1414 		if (__predict_false(++ndesc == rx_ndesc)) {
1415 			if (black_hole && fl_credits >= 8) {
1416 				fl_credits /= 8;
1417 				IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
1418 				    nm_rxq->fl_sidx);
1419 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1420 				    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1421 				fl_credits = fl_cidx & 7;
1422 			}
1423 			t4_write_reg(sc, sc->sge_gts_reg,
1424 			    V_CIDXINC(ndesc) |
1425 			    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
1426 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1427 			ndesc = 0;
1428 		}
1429 	}
1430 
1431 	atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1432 	if (black_hole) {
1433 		fl_credits /= 8;
1434 		IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1435 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1436 		    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1437 	} else if (nframes > 0)
1438 		netmap_rx_irq(ifp, nm_rxq->nid, &work);
1439 
1440     	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndesc) |
1441 	    V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1442 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1443 }
1444 #endif
1445