xref: /freebsd/sys/dev/cxgbe/t4_netmap.c (revision aa1a8ff2d6dbc51ef058f46f3db5a8bb77967145)
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 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 
32 #ifdef DEV_NETMAP
33 #include <sys/param.h>
34 #include <sys/bus.h>
35 #include <sys/eventhandler.h>
36 #include <sys/lock.h>
37 #include <sys/mbuf.h>
38 #include <sys/module.h>
39 #include <sys/selinfo.h>
40 #include <sys/socket.h>
41 #include <sys/sockio.h>
42 #include <machine/bus.h>
43 #include <net/ethernet.h>
44 #include <net/if.h>
45 #include <net/if_media.h>
46 #include <net/if_var.h>
47 #include <net/if_clone.h>
48 #include <net/if_types.h>
49 #include <net/netmap.h>
50 #include <dev/netmap/netmap_kern.h>
51 
52 #include "common/common.h"
53 #include "common/t4_regs.h"
54 #include "common/t4_regs_values.h"
55 
56 extern int fl_pad;	/* XXXNM */
57 
58 /*
59  * 0 = normal netmap rx
60  * 1 = black hole
61  * 2 = supermassive black hole (buffer packing enabled)
62  */
63 int black_hole = 0;
64 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_black_hole, CTLFLAG_RWTUN, &black_hole, 0,
65     "Sink incoming packets.");
66 
67 int rx_ndesc = 256;
68 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_rx_ndesc, CTLFLAG_RWTUN,
69     &rx_ndesc, 0, "# of rx descriptors after which the hw cidx is updated.");
70 
71 int rx_nframes = 64;
72 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_rx_nframes, CTLFLAG_RWTUN,
73     &rx_nframes, 0, "max # of frames received before waking up netmap rx.");
74 
75 int holdoff_tmr_idx = 2;
76 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_holdoff_tmr_idx, CTLFLAG_RWTUN,
77     &holdoff_tmr_idx, 0, "Holdoff timer index for netmap rx queues.");
78 
79 /*
80  * Congestion drops.
81  * -1: no congestion feedback (not recommended).
82  *  0: backpressure the channel instead of dropping packets right away.
83  *  1: no backpressure, drop packets for the congested queue immediately.
84  */
85 static int nm_cong_drop = 1;
86 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_cong_drop, CTLFLAG_RWTUN,
87     &nm_cong_drop, 0,
88     "Congestion control for netmap rx queues (0 = backpressure, 1 = drop");
89 
90 int starve_fl = 0;
91 SYSCTL_INT(_hw_cxgbe, OID_AUTO, starve_fl, CTLFLAG_RWTUN,
92     &starve_fl, 0, "Don't ring fl db for netmap rx queues.");
93 
94 /*
95  * Try to process tx credits in bulk.  This may cause a delay in the return of
96  * tx credits and is suitable for bursty or non-stop tx only.
97  */
98 int lazy_tx_credit_flush = 1;
99 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lazy_tx_credit_flush, CTLFLAG_RWTUN,
100     &lazy_tx_credit_flush, 0, "lazy credit flush for netmap tx queues.");
101 
102 /*
103  * Split the netmap rx queues into two groups that populate separate halves of
104  * the RSS indirection table.  This allows filters with hashmask to steer to a
105  * particular group of queues.
106  */
107 static int nm_split_rss = 0;
108 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_split_rss, CTLFLAG_RWTUN,
109     &nm_split_rss, 0, "Split the netmap rx queues into two groups.");
110 
111 /*
112  * netmap(4) says "netmap does not use features such as checksum offloading, TCP
113  * segmentation offloading, encryption, VLAN encapsulation/decapsulation, etc."
114  * but this knob can be used to get the hardware to checksum all tx traffic
115  * anyway.
116  */
117 static int nm_txcsum = 0;
118 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_txcsum, CTLFLAG_RWTUN,
119     &nm_txcsum, 0, "Enable transmit checksum offloading.");
120 
121 static int free_nm_rxq_hwq(struct vi_info *, struct sge_nm_rxq *);
122 static int free_nm_txq_hwq(struct vi_info *, struct sge_nm_txq *);
123 
124 int
125 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
126     int idx)
127 {
128 	int rc;
129 	struct sysctl_oid *oid;
130 	struct sysctl_oid_list *children;
131 	struct sysctl_ctx_list *ctx;
132 	char name[16];
133 	size_t len;
134 	struct adapter *sc = vi->adapter;
135 	struct netmap_adapter *na = NA(vi->ifp);
136 
137 	MPASS(na != NULL);
138 
139 	len = vi->qsize_rxq * IQ_ESIZE;
140 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
141 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
142 	if (rc != 0)
143 		return (rc);
144 
145 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
146 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
147 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
148 	if (rc != 0)
149 		return (rc);
150 
151 	nm_rxq->vi = vi;
152 	nm_rxq->nid = idx;
153 	nm_rxq->iq_cidx = 0;
154 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
155 	nm_rxq->iq_gen = F_RSPD_GEN;
156 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
157 	nm_rxq->fl_sidx = na->num_rx_desc;
158 	nm_rxq->fl_sidx2 = nm_rxq->fl_sidx;	/* copy for rxsync cacheline */
159 	nm_rxq->intr_idx = intr_idx;
160 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
161 
162 	ctx = &vi->ctx;
163 	children = SYSCTL_CHILDREN(vi->nm_rxq_oid);
164 
165 	snprintf(name, sizeof(name), "%d", idx);
166 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name,
167 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queue");
168 	children = SYSCTL_CHILDREN(oid);
169 
170 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
171 	    &nm_rxq->iq_abs_id, 0, "absolute id of the queue");
172 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
173 	    &nm_rxq->iq_cntxt_id, 0, "SGE context id of the queue");
174 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
175 	    &nm_rxq->iq_cidx, 0, "consumer index");
176 
177 	children = SYSCTL_CHILDREN(oid);
178 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl",
179 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist");
180 	children = SYSCTL_CHILDREN(oid);
181 
182 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
183 	    &nm_rxq->fl_cntxt_id, 0, "SGE context id of the freelist");
184 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
185 	    &nm_rxq->fl_cidx, 0, "consumer index");
186 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
187 	    &nm_rxq->fl_pidx, 0, "producer index");
188 
189 	return (rc);
190 }
191 
192 int
193 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
194 {
195 	struct adapter *sc = vi->adapter;
196 
197 	if (!(vi->flags & VI_INIT_DONE))
198 		return (0);
199 
200 	if (nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID)
201 		free_nm_rxq_hwq(vi, nm_rxq);
202 	MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
203 
204 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
205 	    nm_rxq->iq_desc);
206 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
207 	    nm_rxq->fl_desc);
208 
209 	return (0);
210 }
211 
212 int
213 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx)
214 {
215 	int rc;
216 	size_t len;
217 	struct port_info *pi = vi->pi;
218 	struct adapter *sc = pi->adapter;
219 	struct netmap_adapter *na = NA(vi->ifp);
220 	char name[16];
221 	struct sysctl_oid *oid;
222 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(vi->nm_txq_oid);
223 
224 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
225 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
226 	    &nm_txq->ba, (void **)&nm_txq->desc);
227 	if (rc)
228 		return (rc);
229 
230 	nm_txq->pidx = nm_txq->cidx = 0;
231 	nm_txq->sidx = na->num_tx_desc;
232 	nm_txq->nid = idx;
233 	nm_txq->iqidx = iqidx;
234 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
235 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
236 	    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
237 	if (sc->params.fw_vers >= FW_VERSION32(1, 24, 11, 0))
238 		nm_txq->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR));
239 	else
240 		nm_txq->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
241 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
242 
243 	snprintf(name, sizeof(name), "%d", idx);
244 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name,
245 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "netmap tx queue");
246 	children = SYSCTL_CHILDREN(oid);
247 
248 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
249 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
250 	SYSCTL_ADD_U16(&vi->ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
251 	    &nm_txq->cidx, 0, "consumer index");
252 	SYSCTL_ADD_U16(&vi->ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
253 	    &nm_txq->pidx, 0, "producer index");
254 
255 	return (rc);
256 }
257 
258 int
259 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
260 {
261 	struct adapter *sc = vi->adapter;
262 
263 	if (!(vi->flags & VI_INIT_DONE))
264 		return (0);
265 
266 	if (nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID)
267 		free_nm_txq_hwq(vi, nm_txq);
268 	MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
269 
270 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
271 	    nm_txq->desc);
272 
273 	return (0);
274 }
275 
276 static int
277 alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
278 {
279 	int rc, cntxt_id;
280 	__be32 v;
281 	struct adapter *sc = vi->adapter;
282 	struct port_info *pi = vi->pi;
283 	struct sge_params *sp = &sc->params.sge;
284 	struct netmap_adapter *na = NA(vi->ifp);
285 	struct fw_iq_cmd c;
286 	const int cong_drop = nm_cong_drop;
287 	const int cong_map = pi->rx_e_chan_map;
288 
289 	MPASS(na != NULL);
290 	MPASS(nm_rxq->iq_desc != NULL);
291 	MPASS(nm_rxq->fl_desc != NULL);
292 
293 	bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE);
294 	bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + sp->spg_len);
295 
296 	bzero(&c, sizeof(c));
297 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
298 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
299 	    V_FW_IQ_CMD_VFN(0));
300 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_IQSTART | FW_LEN16(c));
301 	if (nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID)
302 		c.alloc_to_len16 |= htobe32(F_FW_IQ_CMD_ALLOC);
303 	else {
304 		c.iqid = htobe16(nm_rxq->iq_cntxt_id);
305 		c.fl0id = htobe16(nm_rxq->fl_cntxt_id);
306 		c.fl1id = htobe16(0xffff);
307 		c.physiqid = htobe16(nm_rxq->iq_abs_id);
308 	}
309 	MPASS(!forwarding_intr_to_fwq(sc));
310 	KASSERT(nm_rxq->intr_idx < sc->intr_count,
311 	    ("%s: invalid direct intr_idx %d", __func__, nm_rxq->intr_idx));
312 	v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
313 	c.type_to_iqandstindex = htobe32(v |
314 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
315 	    V_FW_IQ_CMD_VIID(vi->viid) |
316 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
317 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
318 	    F_FW_IQ_CMD_IQGTSMODE |
319 	    V_FW_IQ_CMD_IQINTCNTTHRESH(0) |
320 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
321 	c.iqsize = htobe16(vi->qsize_rxq);
322 	c.iqaddr = htobe64(nm_rxq->iq_ba);
323 	if (cong_drop != -1) {
324 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN |
325 		    V_FW_IQ_CMD_FL0CNGCHMAP(cong_map) | F_FW_IQ_CMD_FL0CONGCIF |
326 		    F_FW_IQ_CMD_FL0CONGEN);
327 	}
328 	c.iqns_to_fl0congen |=
329 	    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
330 		V_FW_IQ_CMD_IQTYPE(FW_IQ_IQTYPE_NIC) |
331 		F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
332 		(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
333 		(black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0));
334 	c.fl0dcaen_to_fl0cidxfthresh =
335 	    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
336 		X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
337 		V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
338 		X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
339 	c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE);
340 	c.fl0addr = htobe64(nm_rxq->fl_ba);
341 
342 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
343 	if (rc != 0) {
344 		device_printf(sc->dev,
345 		    "failed to create netmap ingress queue: %d\n", rc);
346 		return (rc);
347 	}
348 
349 	nm_rxq->iq_cidx = 0;
350 	MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE);
351 	nm_rxq->iq_gen = F_RSPD_GEN;
352 	nm_rxq->iq_cntxt_id = be16toh(c.iqid);
353 	nm_rxq->iq_abs_id = be16toh(c.physiqid);
354 	cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
355 	if (cntxt_id >= sc->sge.iqmap_sz) {
356 		panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
357 		    __func__, cntxt_id, sc->sge.iqmap_sz - 1);
358 	}
359 	sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
360 
361 	nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
362 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
363 	nm_rxq->fl_db_saved = 0;
364 	/* matches the X_FETCHBURSTMAX_512B or X_FETCHBURSTMAX_256B above. */
365 	nm_rxq->fl_db_threshold = chip_id(sc) <= CHELSIO_T5 ? 8 : 4;
366 	MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
367 	cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
368 	if (cntxt_id >= sc->sge.eqmap_sz) {
369 		panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
370 		    __func__, cntxt_id, sc->sge.eqmap_sz - 1);
371 	}
372 	sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
373 
374 	nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) |
375 	    sc->chip_params->sge_fl_db;
376 
377 	if (chip_id(sc) >= CHELSIO_T5 && cong_drop != -1) {
378 		t4_sge_set_conm_context(sc, nm_rxq->iq_cntxt_id, cong_drop,
379 		    cong_map);
380 	}
381 
382 	t4_write_reg(sc, sc->sge_gts_reg,
383 	    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
384 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
385 
386 	return (rc);
387 }
388 
389 static int
390 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
391 {
392 	struct adapter *sc = vi->adapter;
393 	int rc;
394 
395 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
396 	    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
397 	if (rc != 0)
398 		device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
399 		    __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
400 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
401 	return (rc);
402 }
403 
404 static int
405 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
406 {
407 	int rc, cntxt_id;
408 	size_t len;
409 	struct adapter *sc = vi->adapter;
410 	struct netmap_adapter *na = NA(vi->ifp);
411 	struct fw_eq_eth_cmd c;
412 
413 	MPASS(na != NULL);
414 	MPASS(nm_txq->desc != NULL);
415 
416 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
417 	bzero(nm_txq->desc, len);
418 
419 	bzero(&c, sizeof(c));
420 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
421 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
422 	    V_FW_EQ_ETH_CMD_VFN(0));
423 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
424 	if (nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID)
425 		c.alloc_to_len16 |= htobe32(F_FW_EQ_ETH_CMD_ALLOC);
426 	else
427 		c.eqid_pkd = htobe32(V_FW_EQ_ETH_CMD_EQID(nm_txq->cntxt_id));
428 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
429 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
430 	c.fetchszm_to_iqid =
431 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
432 		V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
433 		V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
434 	c.dcaen_to_eqsize =
435 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
436 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
437 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
438 		V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
439 	c.eqaddr = htobe64(nm_txq->ba);
440 
441 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
442 	if (rc != 0) {
443 		device_printf(vi->dev,
444 		    "failed to create netmap egress queue: %d\n", rc);
445 		return (rc);
446 	}
447 
448 	nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
449 	cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
450 	if (cntxt_id >= sc->sge.eqmap_sz)
451 	    panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
452 		cntxt_id, sc->sge.eqmap_sz - 1);
453 	sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
454 
455 	nm_txq->pidx = nm_txq->cidx = 0;
456 	MPASS(nm_txq->sidx == na->num_tx_desc);
457 	nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
458 
459 	nm_txq->doorbells = sc->doorbells;
460 	if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
461 	    isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
462 	    isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
463 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
464 		uint32_t mask = (1 << s_qpp) - 1;
465 		volatile uint8_t *udb;
466 
467 		udb = sc->udbs_base + UDBS_DB_OFFSET;
468 		udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
469 		nm_txq->udb_qid = nm_txq->cntxt_id & mask;
470 		if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
471 			clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
472 		else {
473 			udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
474 			nm_txq->udb_qid = 0;
475 		}
476 		nm_txq->udb = (volatile void *)udb;
477 	}
478 
479 	if (sc->params.fw_vers < FW_VERSION32(1, 25, 1, 0)) {
480 		uint32_t param, val;
481 
482 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
483 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) |
484 		    V_FW_PARAMS_PARAM_YZ(nm_txq->cntxt_id);
485 		val = 0xff;
486 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
487 		if (rc != 0) {
488 			device_printf(vi->dev,
489 			    "failed to bind netmap txq %d to class 0xff: %d\n",
490 			    nm_txq->cntxt_id, rc);
491 			rc = 0;
492 		}
493 	}
494 
495 	return (rc);
496 }
497 
498 static int
499 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
500 {
501 	struct adapter *sc = vi->adapter;
502 	int rc;
503 
504 	rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
505 	if (rc != 0)
506 		device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
507 		    nm_txq->cntxt_id, rc);
508 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
509 	return (rc);
510 }
511 
512 static int
513 cxgbe_netmap_simple_rss(struct adapter *sc, struct vi_info *vi,
514     if_t ifp, struct netmap_adapter *na)
515 {
516 	struct netmap_kring *kring;
517 	struct sge_nm_rxq *nm_rxq;
518 	int rc, i, j, nm_state, defq;
519 	uint16_t *rss;
520 
521 	/*
522 	 * Check if there's at least one active (or about to go active) netmap
523 	 * rx queue.
524 	 */
525 	defq = -1;
526 	for_each_nm_rxq(vi, j, nm_rxq) {
527 		nm_state = atomic_load_int(&nm_rxq->nm_state);
528 		kring = na->rx_rings[nm_rxq->nid];
529 		if ((nm_state != NM_OFF && !nm_kring_pending_off(kring)) ||
530 		    (nm_state == NM_OFF && nm_kring_pending_on(kring))) {
531 			MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
532 			if (defq == -1) {
533 				defq = nm_rxq->iq_abs_id;
534 				break;
535 			}
536 		}
537 	}
538 
539 	if (defq == -1) {
540 		/* No active netmap queues.  Switch back to NIC queues. */
541 		rss = vi->rss;
542 		defq = vi->rss[0];
543 	} else {
544 		for (i = 0; i < vi->rss_size;) {
545 			for_each_nm_rxq(vi, j, nm_rxq) {
546 				nm_state = atomic_load_int(&nm_rxq->nm_state);
547 				kring = na->rx_rings[nm_rxq->nid];
548 				if ((nm_state != NM_OFF &&
549 				    !nm_kring_pending_off(kring)) ||
550 				    (nm_state == NM_OFF &&
551 				    nm_kring_pending_on(kring))) {
552 					MPASS(nm_rxq->iq_cntxt_id !=
553 					    INVALID_NM_RXQ_CNTXT_ID);
554 					vi->nm_rss[i++] = nm_rxq->iq_abs_id;
555 					if (i == vi->rss_size)
556 						break;
557 				}
558 			}
559 		}
560 		rss = vi->nm_rss;
561 	}
562 
563 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size, rss,
564 	    vi->rss_size);
565 	if (rc != 0)
566 		if_printf(ifp, "netmap rss_config failed: %d\n", rc);
567 
568 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
569 	if (rc != 0) {
570 		if_printf(ifp, "netmap defaultq config failed: %d\n", rc);
571 	}
572 
573 	return (rc);
574 }
575 
576 /*
577  * Odd number of rx queues work best for split RSS mode as the first queue can
578  * be dedicated for non-RSS traffic and the rest divided into two equal halves.
579  */
580 static int
581 cxgbe_netmap_split_rss(struct adapter *sc, struct vi_info *vi,
582     if_t ifp, struct netmap_adapter *na)
583 {
584 	struct netmap_kring *kring;
585 	struct sge_nm_rxq *nm_rxq;
586 	int rc, i, j, nm_state, defq;
587 	int nactive[2] = {0, 0};
588 	int dq[2] = {-1, -1};
589 	bool dq_norss;		/* default queue should not be in RSS table. */
590 
591 	MPASS(nm_split_rss != 0);
592 	MPASS(vi->nnmrxq > 1);
593 
594 	for_each_nm_rxq(vi, i, nm_rxq) {
595 		j = i / ((vi->nnmrxq + 1) / 2);
596 		nm_state = atomic_load_int(&nm_rxq->nm_state);
597 		kring = na->rx_rings[nm_rxq->nid];
598 		if ((nm_state != NM_OFF && !nm_kring_pending_off(kring)) ||
599 		    (nm_state == NM_OFF && nm_kring_pending_on(kring))) {
600 			MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
601 			nactive[j]++;
602 			if (dq[j] == -1) {
603 				dq[j] = nm_rxq->iq_abs_id;
604 				break;
605 			}
606 		}
607 	}
608 
609 	if (nactive[0] == 0 || nactive[1] == 0)
610 		return (cxgbe_netmap_simple_rss(sc, vi, ifp, na));
611 
612 	MPASS(dq[0] != -1 && dq[1] != -1);
613 	if (nactive[0] > nactive[1]) {
614 		defq = dq[0];
615 		dq_norss = true;
616 	} else if (nactive[0] < nactive[1]) {
617 		defq = dq[1];
618 		dq_norss = true;
619 	} else {
620 		defq = dq[0];
621 		dq_norss = false;
622 	}
623 
624 	i = 0;
625 	nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq];
626 	while (i < vi->rss_size / 2) {
627 		for (j = 0; j < (vi->nnmrxq + 1) / 2; j++) {
628 			nm_state = atomic_load_int(&nm_rxq[j].nm_state);
629 			kring = na->rx_rings[nm_rxq[j].nid];
630 			if ((nm_state == NM_OFF &&
631 			    !nm_kring_pending_on(kring)) ||
632 			    (nm_state == NM_ON &&
633 			    nm_kring_pending_off(kring))) {
634 				continue;
635 			}
636 			MPASS(nm_rxq[j].iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
637 			if (dq_norss && defq == nm_rxq[j].iq_abs_id)
638 				continue;
639 			vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
640 			if (i == vi->rss_size / 2)
641 				break;
642 		}
643 	}
644 	while (i < vi->rss_size) {
645 		for (j = (vi->nnmrxq + 1) / 2; j < vi->nnmrxq; j++) {
646 			nm_state = atomic_load_int(&nm_rxq[j].nm_state);
647 			kring = na->rx_rings[nm_rxq[j].nid];
648 			if ((nm_state == NM_OFF &&
649 			    !nm_kring_pending_on(kring)) ||
650 			    (nm_state == NM_ON &&
651 			    nm_kring_pending_off(kring))) {
652 				continue;
653 			}
654 			MPASS(nm_rxq[j].iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
655 			if (dq_norss && defq == nm_rxq[j].iq_abs_id)
656 				continue;
657 			vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
658 			if (i == vi->rss_size)
659 				break;
660 		}
661 	}
662 
663 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
664 	    vi->nm_rss, vi->rss_size);
665 	if (rc != 0)
666 		if_printf(ifp, "netmap split_rss_config failed: %d\n", rc);
667 
668 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
669 	if (rc != 0)
670 		if_printf(ifp, "netmap defaultq config failed: %d\n", rc);
671 
672 	return (rc);
673 }
674 
675 static inline int
676 cxgbe_netmap_rss(struct adapter *sc, struct vi_info *vi, if_t ifp,
677     struct netmap_adapter *na)
678 {
679 
680 	if (nm_split_rss == 0 || vi->nnmrxq == 1)
681 		return (cxgbe_netmap_simple_rss(sc, vi, ifp, na));
682 	else
683 		return (cxgbe_netmap_split_rss(sc, vi, ifp, na));
684 }
685 
686 static int
687 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, if_t ifp,
688     struct netmap_adapter *na)
689 {
690 	struct netmap_slot *slot;
691 	struct netmap_kring *kring;
692 	struct sge_nm_rxq *nm_rxq;
693 	struct sge_nm_txq *nm_txq;
694 	int i, j, hwidx;
695 	struct rx_buf_info *rxb;
696 
697 	ASSERT_SYNCHRONIZED_OP(sc);
698 	MPASS(vi->nnmrxq > 0);
699 	MPASS(vi->nnmtxq > 0);
700 
701 	if ((vi->flags & VI_INIT_DONE) == 0 ||
702 	    (if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0) {
703 		if_printf(ifp, "cannot enable netmap operation because "
704 		    "interface is not UP.\n");
705 		return (EAGAIN);
706 	}
707 
708 	rxb = &sc->sge.rx_buf_info[0];
709 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
710 		if (rxb->size1 == NETMAP_BUF_SIZE(na)) {
711 			hwidx = rxb->hwidx1;
712 			break;
713 		}
714 		if (rxb->size2 == NETMAP_BUF_SIZE(na)) {
715 			hwidx = rxb->hwidx2;
716 			break;
717 		}
718 	}
719 	if (i >= SW_ZONE_SIZES) {
720 		if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
721 		    NETMAP_BUF_SIZE(na));
722 		return (ENXIO);
723 	}
724 
725 	/* Must set caps before calling netmap_reset */
726 	nm_set_native_flags(na);
727 
728 	for_each_nm_rxq(vi, i, nm_rxq) {
729 		kring = na->rx_rings[nm_rxq->nid];
730 		if (!nm_kring_pending_on(kring))
731 			continue;
732 
733 		alloc_nm_rxq_hwq(vi, nm_rxq);
734 		nm_rxq->fl_hwidx = hwidx;
735 		slot = netmap_reset(na, NR_RX, i, 0);
736 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
737 
738 		/* We deal with 8 bufs at a time */
739 		MPASS((na->num_rx_desc & 7) == 0);
740 		MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
741 		for (j = 0; j < nm_rxq->fl_sidx; j++) {
742 			uint64_t ba;
743 
744 			PNMB(na, &slot[j], &ba);
745 			MPASS(ba != 0);
746 			nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
747 		}
748 		j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
749 		MPASS((j & 7) == 0);
750 		j /= 8;	/* driver pidx to hardware pidx */
751 		wmb();
752 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
753 		    nm_rxq->fl_db_val | V_PIDX(j));
754 
755 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_OFF, NM_ON);
756 	}
757 
758 	for_each_nm_txq(vi, i, nm_txq) {
759 		kring = na->tx_rings[nm_txq->nid];
760 		if (!nm_kring_pending_on(kring))
761 			continue;
762 
763 		alloc_nm_txq_hwq(vi, nm_txq);
764 		slot = netmap_reset(na, NR_TX, i, 0);
765 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
766 	}
767 
768 	if (vi->nm_rss == NULL) {
769 		vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
770 		    M_ZERO | M_WAITOK);
771 	}
772 
773 	return (cxgbe_netmap_rss(sc, vi, ifp, na));
774 }
775 
776 static int
777 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, if_t ifp,
778     struct netmap_adapter *na)
779 {
780 	struct netmap_kring *kring;
781 	int rc, i, nm_state, nactive;
782 	struct sge_nm_txq *nm_txq;
783 	struct sge_nm_rxq *nm_rxq;
784 
785 	ASSERT_SYNCHRONIZED_OP(sc);
786 	MPASS(vi->nnmrxq > 0);
787 	MPASS(vi->nnmtxq > 0);
788 
789 	if (!nm_netmap_on(na))
790 		return (0);
791 
792 	if ((vi->flags & VI_INIT_DONE) == 0)
793 		return (0);
794 
795 	/* First remove the queues that are stopping from the RSS table. */
796 	rc = cxgbe_netmap_rss(sc, vi, ifp, na);
797 	if (rc != 0)
798 		return (rc);	/* error message logged already. */
799 
800 	for_each_nm_txq(vi, i, nm_txq) {
801 		kring = na->tx_rings[nm_txq->nid];
802 		if (!nm_kring_pending_off(kring))
803 			continue;
804 		MPASS(nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID);
805 
806 		rc = -t4_eth_eq_stop(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
807 		if (rc != 0) {
808 			device_printf(vi->dev,
809 			    "failed to stop nm_txq[%d]: %d.\n", i, rc);
810 			return (rc);
811 		}
812 
813 		/* XXX: netmap, not the driver, should do this. */
814 		kring->rhead = kring->rcur = kring->nr_hwcur = 0;
815 		kring->rtail = kring->nr_hwtail = kring->nkr_num_slots - 1;
816 	}
817 	nactive = 0;
818 	for_each_nm_rxq(vi, i, nm_rxq) {
819 		nm_state = atomic_load_int(&nm_rxq->nm_state);
820 		kring = na->rx_rings[nm_rxq->nid];
821 		if (nm_state != NM_OFF && !nm_kring_pending_off(kring))
822 			nactive++;
823 		if (!nm_kring_pending_off(kring))
824 			continue;
825 		MPASS(nm_state != NM_OFF);
826 		MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
827 
828 		rc = -t4_iq_stop(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
829 		    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
830 		if (rc != 0) {
831 			device_printf(vi->dev,
832 			    "failed to stop nm_rxq[%d]: %d.\n", i, rc);
833 			return (rc);
834 		}
835 
836 		while (!atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_OFF))
837 			pause("nmst", 1);
838 
839 		/* XXX: netmap, not the driver, should do this. */
840 		kring->rhead = kring->rcur = kring->nr_hwcur = 0;
841 		kring->rtail = kring->nr_hwtail = 0;
842 	}
843 	netmap_krings_mode_commit(na, 0);
844 	if (nactive == 0)
845 		nm_clear_native_flags(na);
846 
847 	return (rc);
848 }
849 
850 static int
851 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
852 {
853 	if_t ifp = na->ifp;
854 	struct vi_info *vi = if_getsoftc(ifp);
855 	struct adapter *sc = vi->adapter;
856 	int rc;
857 
858 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
859 	if (rc != 0)
860 		return (rc);
861 	if (on)
862 		rc = cxgbe_netmap_on(sc, vi, ifp, na);
863 	else
864 		rc = cxgbe_netmap_off(sc, vi, ifp, na);
865 	end_synchronized_op(sc, 0);
866 
867 	return (rc);
868 }
869 
870 /* How many packets can a single type1 WR carry in n descriptors */
871 static inline int
872 ndesc_to_npkt(const int n)
873 {
874 
875 	MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
876 
877 	return (n * 2 - 1);
878 }
879 #define MAX_NPKT_IN_TYPE1_WR	(ndesc_to_npkt(SGE_MAX_WR_NDESC))
880 
881 /*
882  * Space (in descriptors) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
883  * carries n packets
884  */
885 static inline int
886 npkt_to_ndesc(const int n)
887 {
888 
889 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
890 
891 	return ((n + 2) / 2);
892 }
893 
894 /*
895  * Space (in 16B units) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
896  * carries n packets
897  */
898 static inline int
899 npkt_to_len16(const int n)
900 {
901 
902 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
903 
904 	return (n * 2 + 1);
905 }
906 
907 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
908 
909 static void
910 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
911 {
912 	int n;
913 	u_int db = nm_txq->doorbells;
914 
915 	MPASS(nm_txq->pidx != nm_txq->dbidx);
916 
917 	n = NMIDXDIFF(nm_txq, dbidx);
918 	if (n > 1)
919 		clrbit(&db, DOORBELL_WCWR);
920 	wmb();
921 
922 	switch (ffs(db) - 1) {
923 	case DOORBELL_UDB:
924 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
925 		break;
926 
927 	case DOORBELL_WCWR: {
928 		volatile uint64_t *dst, *src;
929 
930 		/*
931 		 * Queues whose 128B doorbell segment fits in the page do not
932 		 * use relative qid (udb_qid is always 0).  Only queues with
933 		 * doorbell segments can do WCWR.
934 		 */
935 		KASSERT(nm_txq->udb_qid == 0 && n == 1,
936 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
937 		    __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
938 
939 		dst = (volatile void *)((uintptr_t)nm_txq->udb +
940 		    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
941 		src = (void *)&nm_txq->desc[nm_txq->dbidx];
942 		while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
943 			*dst++ = *src++;
944 		wmb();
945 		break;
946 	}
947 
948 	case DOORBELL_UDBWC:
949 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
950 		wmb();
951 		break;
952 
953 	case DOORBELL_KDB:
954 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
955 		    V_QID(nm_txq->cntxt_id) | V_PIDX(n));
956 		break;
957 	}
958 	nm_txq->dbidx = nm_txq->pidx;
959 }
960 
961 /*
962  * Write work requests to send 'npkt' frames and ring the doorbell to send them
963  * on their way.  No need to check for wraparound.
964  */
965 static void
966 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
967     struct netmap_kring *kring, int npkt, int npkt_remaining)
968 {
969 	struct netmap_ring *ring = kring->ring;
970 	struct netmap_slot *slot;
971 	const u_int lim = kring->nkr_num_slots - 1;
972 	struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
973 	uint16_t len;
974 	uint64_t ba;
975 	struct cpl_tx_pkt_core *cpl;
976 	struct ulptx_sgl *usgl;
977 	int i, n;
978 
979 	while (npkt) {
980 		n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
981 		len = 0;
982 
983 		wr = (void *)&nm_txq->desc[nm_txq->pidx];
984 		wr->op_pkd = nm_txq->op_pkd;
985 		wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
986 		wr->npkt = n;
987 		wr->r3 = 0;
988 		wr->type = 1;
989 		cpl = (void *)(wr + 1);
990 
991 		for (i = 0; i < n; i++) {
992 			slot = &ring->slot[kring->nr_hwcur];
993 			PNMB(kring->na, slot, &ba);
994 			MPASS(ba != 0);
995 
996 			cpl->ctrl0 = nm_txq->cpl_ctrl0;
997 			cpl->pack = 0;
998 			cpl->len = htobe16(slot->len);
999 			cpl->ctrl1 = nm_txcsum ? 0 :
1000 			    htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
1001 
1002 			usgl = (void *)(cpl + 1);
1003 			usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1004 			    V_ULPTX_NSGE(1));
1005 			usgl->len0 = htobe32(slot->len);
1006 			usgl->addr0 = htobe64(ba + nm_get_offset(kring, slot));
1007 
1008 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1009 			cpl = (void *)(usgl + 1);
1010 			MPASS(slot->len + len <= UINT16_MAX);
1011 			len += slot->len;
1012 			kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
1013 		}
1014 		wr->plen = htobe16(len);
1015 
1016 		npkt -= n;
1017 		nm_txq->pidx += npkt_to_ndesc(n);
1018 		MPASS(nm_txq->pidx <= nm_txq->sidx);
1019 		if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
1020 			/*
1021 			 * This routine doesn't know how to write WRs that wrap
1022 			 * around.  Make sure it wasn't asked to.
1023 			 */
1024 			MPASS(npkt == 0);
1025 			nm_txq->pidx = 0;
1026 		}
1027 
1028 		if (npkt == 0 && npkt_remaining == 0) {
1029 			/* All done. */
1030 			if (lazy_tx_credit_flush == 0) {
1031 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
1032 				    F_FW_WR_EQUIQ);
1033 				nm_txq->equeqidx = nm_txq->pidx;
1034 				nm_txq->equiqidx = nm_txq->pidx;
1035 			}
1036 			ring_nm_txq_db(sc, nm_txq);
1037 			return;
1038 		}
1039 
1040 		if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
1041 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
1042 			    F_FW_WR_EQUIQ);
1043 			nm_txq->equeqidx = nm_txq->pidx;
1044 			nm_txq->equiqidx = nm_txq->pidx;
1045 		} else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
1046 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1047 			nm_txq->equeqidx = nm_txq->pidx;
1048 		}
1049 		if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
1050 			ring_nm_txq_db(sc, nm_txq);
1051 	}
1052 
1053 	/* Will get called again. */
1054 	MPASS(npkt_remaining);
1055 }
1056 
1057 /* How many contiguous free descriptors starting at pidx */
1058 static inline int
1059 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
1060 {
1061 
1062 	if (nm_txq->cidx > nm_txq->pidx)
1063 		return (nm_txq->cidx - nm_txq->pidx - 1);
1064 	else if (nm_txq->cidx > 0)
1065 		return (nm_txq->sidx - nm_txq->pidx);
1066 	else
1067 		return (nm_txq->sidx - nm_txq->pidx - 1);
1068 }
1069 
1070 static int
1071 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
1072 {
1073 	struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
1074 	uint16_t hw_cidx = spg->cidx;	/* snapshot */
1075 	struct fw_eth_tx_pkts_wr *wr;
1076 	int n = 0;
1077 
1078 	hw_cidx = be16toh(hw_cidx);
1079 
1080 	while (nm_txq->cidx != hw_cidx) {
1081 		wr = (void *)&nm_txq->desc[nm_txq->cidx];
1082 
1083 		MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)) ||
1084 		    wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR)));
1085 		MPASS(wr->type == 1);
1086 		MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
1087 
1088 		n += wr->npkt;
1089 		nm_txq->cidx += npkt_to_ndesc(wr->npkt);
1090 
1091 		/*
1092 		 * We never sent a WR that wrapped around so the credits coming
1093 		 * back, WR by WR, should never cause the cidx to wrap around
1094 		 * either.
1095 		 */
1096 		MPASS(nm_txq->cidx <= nm_txq->sidx);
1097 		if (__predict_false(nm_txq->cidx == nm_txq->sidx))
1098 			nm_txq->cidx = 0;
1099 	}
1100 
1101 	return (n);
1102 }
1103 
1104 static int
1105 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
1106 {
1107 	struct netmap_adapter *na = kring->na;
1108 	if_t ifp = na->ifp;
1109 	struct vi_info *vi = if_getsoftc(ifp);
1110 	struct adapter *sc = vi->adapter;
1111 	struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
1112 	const u_int head = kring->rhead;
1113 	u_int reclaimed = 0;
1114 	int n, d, npkt_remaining, ndesc_remaining;
1115 
1116 	/*
1117 	 * Tx was at kring->nr_hwcur last time around and now we need to advance
1118 	 * to kring->rhead.  Note that the driver's pidx moves independent of
1119 	 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
1120 	 * between descriptors and frames isn't 1:1).
1121 	 */
1122 
1123 	npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
1124 	    kring->nkr_num_slots - kring->nr_hwcur + head;
1125 	while (npkt_remaining) {
1126 		reclaimed += reclaim_nm_tx_desc(nm_txq);
1127 		ndesc_remaining = contiguous_ndesc_available(nm_txq);
1128 		/* Can't run out of descriptors with packets still remaining */
1129 		MPASS(ndesc_remaining > 0);
1130 
1131 		/* # of desc needed to tx all remaining packets */
1132 		d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
1133 		if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
1134 			d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
1135 
1136 		if (d <= ndesc_remaining)
1137 			n = npkt_remaining;
1138 		else {
1139 			/* Can't send all, calculate how many can be sent */
1140 			n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
1141 			    MAX_NPKT_IN_TYPE1_WR;
1142 			if (ndesc_remaining % SGE_MAX_WR_NDESC)
1143 				n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
1144 		}
1145 
1146 		/* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
1147 		npkt_remaining -= n;
1148 		cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining);
1149 	}
1150 	MPASS(npkt_remaining == 0);
1151 	MPASS(kring->nr_hwcur == head);
1152 	MPASS(nm_txq->dbidx == nm_txq->pidx);
1153 
1154 	/*
1155 	 * Second part: reclaim buffers for completed transmissions.
1156 	 */
1157 	if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
1158 		reclaimed += reclaim_nm_tx_desc(nm_txq);
1159 		kring->nr_hwtail += reclaimed;
1160 		if (kring->nr_hwtail >= kring->nkr_num_slots)
1161 			kring->nr_hwtail -= kring->nkr_num_slots;
1162 	}
1163 
1164 	return (0);
1165 }
1166 
1167 static int
1168 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
1169 {
1170 	struct netmap_adapter *na = kring->na;
1171 	struct netmap_ring *ring = kring->ring;
1172 	if_t ifp = na->ifp;
1173 	struct vi_info *vi = if_getsoftc(ifp);
1174 	struct adapter *sc = vi->adapter;
1175 	struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
1176 	u_int const head = kring->rhead;
1177 	u_int n;
1178 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1179 
1180 	if (black_hole)
1181 		return (0);	/* No updates ever. */
1182 
1183 	if (netmap_no_pendintr || force_update) {
1184 		kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
1185 		kring->nr_kflags &= ~NKR_PENDINTR;
1186 	}
1187 
1188 	if (nm_rxq->fl_db_saved > 0 && starve_fl == 0) {
1189 		wmb();
1190 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1191 		    nm_rxq->fl_db_val | V_PIDX(nm_rxq->fl_db_saved));
1192 		nm_rxq->fl_db_saved = 0;
1193 	}
1194 
1195 	/* Userspace done with buffers from kring->nr_hwcur to head */
1196 	n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
1197 	    kring->nkr_num_slots - kring->nr_hwcur + head;
1198 	n &= ~7U;
1199 	if (n > 0) {
1200 		u_int fl_pidx = nm_rxq->fl_pidx;
1201 		struct netmap_slot *slot = &ring->slot[fl_pidx];
1202 		uint64_t ba;
1203 		int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
1204 
1205 		/*
1206 		 * We always deal with 8 buffers at a time.  We must have
1207 		 * stopped at an 8B boundary (fl_pidx) last time around and we
1208 		 * must have a multiple of 8B buffers to give to the freelist.
1209 		 */
1210 		MPASS((fl_pidx & 7) == 0);
1211 		MPASS((n & 7) == 0);
1212 
1213 		IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
1214 		IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx2);
1215 
1216 		while (n > 0) {
1217 			for (i = 0; i < 8; i++, fl_pidx++, slot++) {
1218 				PNMB(na, slot, &ba);
1219 				MPASS(ba != 0);
1220 				nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
1221 				slot->flags &= ~NS_BUF_CHANGED;
1222 				MPASS(fl_pidx <= nm_rxq->fl_sidx2);
1223 			}
1224 			n -= 8;
1225 			if (fl_pidx == nm_rxq->fl_sidx2) {
1226 				fl_pidx = 0;
1227 				slot = &ring->slot[0];
1228 			}
1229 			if (++dbinc == nm_rxq->fl_db_threshold) {
1230 				wmb();
1231 				if (starve_fl)
1232 					nm_rxq->fl_db_saved += dbinc;
1233 				else {
1234 					t4_write_reg(sc, sc->sge_kdoorbell_reg,
1235 					    nm_rxq->fl_db_val | V_PIDX(dbinc));
1236 				}
1237 				dbinc = 0;
1238 			}
1239 		}
1240 		MPASS(nm_rxq->fl_pidx == fl_pidx);
1241 
1242 		if (dbinc > 0) {
1243 			wmb();
1244 			if (starve_fl)
1245 				nm_rxq->fl_db_saved += dbinc;
1246 			else {
1247 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1248 				    nm_rxq->fl_db_val | V_PIDX(dbinc));
1249 			}
1250 		}
1251 	}
1252 
1253 	return (0);
1254 }
1255 
1256 void
1257 cxgbe_nm_attach(struct vi_info *vi)
1258 {
1259 	struct port_info *pi;
1260 	struct adapter *sc;
1261 	struct netmap_adapter na;
1262 
1263 	MPASS(vi->nnmrxq > 0);
1264 	MPASS(vi->ifp != NULL);
1265 
1266 	pi = vi->pi;
1267 	sc = pi->adapter;
1268 
1269 	bzero(&na, sizeof(na));
1270 
1271 	na.ifp = vi->ifp;
1272 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_OFFSETS;
1273 
1274 	/* Netmap doesn't know about the space reserved for the status page. */
1275 	na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
1276 
1277 	/*
1278 	 * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
1279 	 * num_rx_desc is based on the number of buffers that can be held in the
1280 	 * freelist, and not the number of entries in the iq.  (These two are
1281 	 * not exactly the same due to the space taken up by the status page).
1282 	 */
1283 	na.num_rx_desc = rounddown(vi->qsize_rxq, 8);
1284 	na.nm_txsync = cxgbe_netmap_txsync;
1285 	na.nm_rxsync = cxgbe_netmap_rxsync;
1286 	na.nm_register = cxgbe_netmap_reg;
1287 	na.num_tx_rings = vi->nnmtxq;
1288 	na.num_rx_rings = vi->nnmrxq;
1289 	na.rx_buf_maxsize = MAX_MTU + sc->params.sge.fl_pktshift;
1290 	netmap_attach(&na);	/* This adds IFCAP_NETMAP to if_capabilities */
1291 }
1292 
1293 void
1294 cxgbe_nm_detach(struct vi_info *vi)
1295 {
1296 
1297 	MPASS(vi->nnmrxq > 0);
1298 	MPASS(vi->ifp != NULL);
1299 
1300 	netmap_detach(vi->ifp);
1301 }
1302 
1303 static inline const void *
1304 unwrap_nm_fw6_msg(const struct cpl_fw6_msg *cpl)
1305 {
1306 
1307 	MPASS(cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL);
1308 
1309 	/* data[0] is RSS header */
1310 	return (&cpl->data[1]);
1311 }
1312 
1313 static void
1314 handle_nm_sge_egr_update(struct adapter *sc, if_t ifp,
1315     const struct cpl_sge_egr_update *egr)
1316 {
1317 	uint32_t oq;
1318 	struct sge_nm_txq *nm_txq;
1319 
1320 	oq = be32toh(egr->opcode_qid);
1321 	MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
1322 	nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
1323 
1324 	netmap_tx_irq(ifp, nm_txq->nid);
1325 }
1326 
1327 void
1328 service_nm_rxq(struct sge_nm_rxq *nm_rxq)
1329 {
1330 	struct vi_info *vi = nm_rxq->vi;
1331 	struct adapter *sc = vi->adapter;
1332 	if_t ifp = vi->ifp;
1333 	struct netmap_adapter *na = NA(ifp);
1334 	struct netmap_kring *kring = na->rx_rings[nm_rxq->nid];
1335 	struct netmap_ring *ring = kring->ring;
1336 	struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
1337 	const void *cpl;
1338 	uint32_t lq;
1339 	u_int work = 0;
1340 	uint8_t opcode;
1341 	uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
1342 	u_int fl_credits = fl_cidx & 7;
1343 	u_int ndesc = 0;	/* desc processed since last cidx update */
1344 	u_int nframes = 0;	/* frames processed since last netmap wakeup */
1345 
1346 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
1347 
1348 		rmb();
1349 
1350 		lq = be32toh(d->rsp.pldbuflen_qid);
1351 		opcode = d->rss.opcode;
1352 		cpl = &d->cpl[0];
1353 
1354 		switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
1355 		case X_RSPD_TYPE_FLBUF:
1356 
1357 			/* fall through */
1358 
1359 		case X_RSPD_TYPE_CPL:
1360 			MPASS(opcode < NUM_CPL_CMDS);
1361 
1362 			switch (opcode) {
1363 			case CPL_FW4_MSG:
1364 			case CPL_FW6_MSG:
1365 				cpl = unwrap_nm_fw6_msg(cpl);
1366 				/* fall through */
1367 			case CPL_SGE_EGR_UPDATE:
1368 				handle_nm_sge_egr_update(sc, ifp, cpl);
1369 				break;
1370 			case CPL_RX_PKT:
1371 				/*
1372 				 * Note that the application must have netmap
1373 				 * offsets (NETMAP_REQ_OPT_OFFSETS) enabled on
1374 				 * the ring or its rx will not work correctly
1375 				 * when fl_pktshift > 0.
1376 				 */
1377 				nm_write_offset(kring, &ring->slot[fl_cidx],
1378 				    sc->params.sge.fl_pktshift);
1379 				ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
1380 				    sc->params.sge.fl_pktshift;
1381 				ring->slot[fl_cidx].flags = 0;
1382 				nframes++;
1383 				if (!(lq & F_RSPD_NEWBUF)) {
1384 					MPASS(black_hole == 2);
1385 					break;
1386 				}
1387 				fl_credits++;
1388 				if (__predict_false(++fl_cidx == nm_rxq->fl_sidx))
1389 					fl_cidx = 0;
1390 				break;
1391 			default:
1392 				panic("%s: unexpected opcode 0x%x on nm_rxq %p",
1393 				    __func__, opcode, nm_rxq);
1394 			}
1395 			break;
1396 
1397 		case X_RSPD_TYPE_INTR:
1398 			/* Not equipped to handle forwarded interrupts. */
1399 			panic("%s: netmap queue received interrupt for iq %u\n",
1400 			    __func__, lq);
1401 
1402 		default:
1403 			panic("%s: illegal response type %d on nm_rxq %p",
1404 			    __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
1405 		}
1406 
1407 		d++;
1408 		if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
1409 			nm_rxq->iq_cidx = 0;
1410 			d = &nm_rxq->iq_desc[0];
1411 			nm_rxq->iq_gen ^= F_RSPD_GEN;
1412 		}
1413 
1414 		if (__predict_false(++nframes == rx_nframes) && !black_hole) {
1415 			atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1416 			netmap_rx_irq(ifp, nm_rxq->nid, &work);
1417 			nframes = 0;
1418 		}
1419 
1420 		if (__predict_false(++ndesc == rx_ndesc)) {
1421 			if (black_hole && fl_credits >= 8) {
1422 				fl_credits /= 8;
1423 				IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
1424 				    nm_rxq->fl_sidx);
1425 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1426 				    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1427 				fl_credits = fl_cidx & 7;
1428 			}
1429 			t4_write_reg(sc, sc->sge_gts_reg,
1430 			    V_CIDXINC(ndesc) |
1431 			    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
1432 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1433 			ndesc = 0;
1434 		}
1435 	}
1436 
1437 	atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1438 	if (black_hole) {
1439 		fl_credits /= 8;
1440 		IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1441 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1442 		    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1443 	} else if (nframes > 0)
1444 		netmap_rx_irq(ifp, nm_rxq->nid, &work);
1445 
1446 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndesc) |
1447 	    V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1448 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1449 }
1450 #endif
1451