xref: /freebsd/sys/dev/cxgbe/t4_netmap.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
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, int cong)
280 {
281 	int rc, cntxt_id, i;
282 	__be32 v;
283 	struct adapter *sc = vi->adapter;
284 	struct sge_params *sp = &sc->params.sge;
285 	struct netmap_adapter *na = NA(vi->ifp);
286 	struct fw_iq_cmd c;
287 
288 	MPASS(na != NULL);
289 	MPASS(nm_rxq->iq_desc != NULL);
290 	MPASS(nm_rxq->fl_desc != NULL);
291 
292 	bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE);
293 	bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + sp->spg_len);
294 
295 	bzero(&c, sizeof(c));
296 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
297 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
298 	    V_FW_IQ_CMD_VFN(0));
299 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_IQSTART | FW_LEN16(c));
300 	if (nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID)
301 		c.alloc_to_len16 |= htobe32(F_FW_IQ_CMD_ALLOC);
302 	else {
303 		c.iqid = htobe16(nm_rxq->iq_cntxt_id);
304 		c.fl0id = htobe16(nm_rxq->fl_cntxt_id);
305 		c.fl1id = htobe16(0xffff);
306 		c.physiqid = htobe16(nm_rxq->iq_abs_id);
307 	}
308 	MPASS(!forwarding_intr_to_fwq(sc));
309 	KASSERT(nm_rxq->intr_idx < sc->intr_count,
310 	    ("%s: invalid direct intr_idx %d", __func__, nm_rxq->intr_idx));
311 	v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
312 	c.type_to_iqandstindex = htobe32(v |
313 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
314 	    V_FW_IQ_CMD_VIID(vi->viid) |
315 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
316 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(vi->pi->tx_chan) |
317 	    F_FW_IQ_CMD_IQGTSMODE |
318 	    V_FW_IQ_CMD_IQINTCNTTHRESH(0) |
319 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
320 	c.iqsize = htobe16(vi->qsize_rxq);
321 	c.iqaddr = htobe64(nm_rxq->iq_ba);
322 	if (cong >= 0) {
323 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN |
324 		    V_FW_IQ_CMD_FL0CNGCHMAP(cong) | F_FW_IQ_CMD_FL0CONGCIF |
325 		    F_FW_IQ_CMD_FL0CONGEN);
326 	}
327 	c.iqns_to_fl0congen |=
328 	    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
329 		F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
330 		(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
331 		(black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0));
332 	c.fl0dcaen_to_fl0cidxfthresh =
333 	    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
334 		X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
335 		V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
336 		X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
337 	c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE);
338 	c.fl0addr = htobe64(nm_rxq->fl_ba);
339 
340 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
341 	if (rc != 0) {
342 		device_printf(sc->dev,
343 		    "failed to create netmap ingress queue: %d\n", rc);
344 		return (rc);
345 	}
346 
347 	nm_rxq->iq_cidx = 0;
348 	MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE);
349 	nm_rxq->iq_gen = F_RSPD_GEN;
350 	nm_rxq->iq_cntxt_id = be16toh(c.iqid);
351 	nm_rxq->iq_abs_id = be16toh(c.physiqid);
352 	cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
353 	if (cntxt_id >= sc->sge.iqmap_sz) {
354 		panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
355 		    __func__, cntxt_id, sc->sge.iqmap_sz - 1);
356 	}
357 	sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
358 
359 	nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
360 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
361 	nm_rxq->fl_db_saved = 0;
362 	/* matches the X_FETCHBURSTMAX_512B or X_FETCHBURSTMAX_256B above. */
363 	nm_rxq->fl_db_threshold = chip_id(sc) <= CHELSIO_T5 ? 8 : 4;
364 	MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
365 	cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
366 	if (cntxt_id >= sc->sge.eqmap_sz) {
367 		panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
368 		    __func__, cntxt_id, sc->sge.eqmap_sz - 1);
369 	}
370 	sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
371 
372 	nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) |
373 	    sc->chip_params->sge_fl_db;
374 
375 	if (chip_id(sc) >= CHELSIO_T5 && cong >= 0) {
376 		uint32_t param, val;
377 
378 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
379 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
380 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
381 		if (cong == 0)
382 			val = 1 << 19;
383 		else {
384 			val = 2 << 19;
385 			for (i = 0; i < 4; i++) {
386 				if (cong & (1 << i))
387 					val |= 1 << (i << 2);
388 			}
389 		}
390 
391 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
392 		if (rc != 0) {
393 			/* report error but carry on */
394 			device_printf(sc->dev,
395 			    "failed to set congestion manager context for "
396 			    "ingress queue %d: %d\n", nm_rxq->iq_cntxt_id, rc);
397 		}
398 	}
399 
400 	t4_write_reg(sc, sc->sge_gts_reg,
401 	    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
402 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
403 
404 	return (rc);
405 }
406 
407 static int
408 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
409 {
410 	struct adapter *sc = vi->adapter;
411 	int rc;
412 
413 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
414 	    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
415 	if (rc != 0)
416 		device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
417 		    __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
418 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
419 	return (rc);
420 }
421 
422 static int
423 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
424 {
425 	int rc, cntxt_id;
426 	size_t len;
427 	struct adapter *sc = vi->adapter;
428 	struct netmap_adapter *na = NA(vi->ifp);
429 	struct fw_eq_eth_cmd c;
430 
431 	MPASS(na != NULL);
432 	MPASS(nm_txq->desc != NULL);
433 
434 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
435 	bzero(nm_txq->desc, len);
436 
437 	bzero(&c, sizeof(c));
438 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
439 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
440 	    V_FW_EQ_ETH_CMD_VFN(0));
441 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
442 	if (nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID)
443 		c.alloc_to_len16 |= htobe32(F_FW_EQ_ETH_CMD_ALLOC);
444 	else
445 		c.eqid_pkd = htobe32(V_FW_EQ_ETH_CMD_EQID(nm_txq->cntxt_id));
446 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
447 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
448 	c.fetchszm_to_iqid =
449 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
450 		V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
451 		V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
452 	c.dcaen_to_eqsize =
453 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
454 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
455 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
456 		V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
457 	c.eqaddr = htobe64(nm_txq->ba);
458 
459 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
460 	if (rc != 0) {
461 		device_printf(vi->dev,
462 		    "failed to create netmap egress queue: %d\n", rc);
463 		return (rc);
464 	}
465 
466 	nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
467 	cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
468 	if (cntxt_id >= sc->sge.eqmap_sz)
469 	    panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
470 		cntxt_id, sc->sge.eqmap_sz - 1);
471 	sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
472 
473 	nm_txq->pidx = nm_txq->cidx = 0;
474 	MPASS(nm_txq->sidx == na->num_tx_desc);
475 	nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
476 
477 	nm_txq->doorbells = sc->doorbells;
478 	if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
479 	    isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
480 	    isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
481 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
482 		uint32_t mask = (1 << s_qpp) - 1;
483 		volatile uint8_t *udb;
484 
485 		udb = sc->udbs_base + UDBS_DB_OFFSET;
486 		udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
487 		nm_txq->udb_qid = nm_txq->cntxt_id & mask;
488 		if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
489 	    		clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
490 		else {
491 			udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
492 			nm_txq->udb_qid = 0;
493 		}
494 		nm_txq->udb = (volatile void *)udb;
495 	}
496 
497 	if (sc->params.fw_vers < FW_VERSION32(1, 25, 1, 0)) {
498 		uint32_t param, val;
499 
500 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
501 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) |
502 		    V_FW_PARAMS_PARAM_YZ(nm_txq->cntxt_id);
503 		val = 0xff;
504 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
505 		if (rc != 0) {
506 			device_printf(vi->dev,
507 			    "failed to bind netmap txq %d to class 0xff: %d\n",
508 			    nm_txq->cntxt_id, rc);
509 			rc = 0;
510 		}
511 	}
512 
513 	return (rc);
514 }
515 
516 static int
517 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
518 {
519 	struct adapter *sc = vi->adapter;
520 	int rc;
521 
522 	rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
523 	if (rc != 0)
524 		device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
525 		    nm_txq->cntxt_id, rc);
526 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
527 	return (rc);
528 }
529 
530 static int
531 cxgbe_netmap_simple_rss(struct adapter *sc, struct vi_info *vi,
532     struct ifnet *ifp, struct netmap_adapter *na)
533 {
534 	struct netmap_kring *kring;
535 	struct sge_nm_rxq *nm_rxq;
536 	int rc, i, j, nm_state, defq;
537 	uint16_t *rss;
538 
539 	/*
540 	 * Check if there's at least one active (or about to go active) netmap
541 	 * rx queue.
542 	 */
543 	defq = -1;
544 	for_each_nm_rxq(vi, j, nm_rxq) {
545 		nm_state = atomic_load_int(&nm_rxq->nm_state);
546 		kring = na->rx_rings[nm_rxq->nid];
547 		if ((nm_state != NM_OFF && !nm_kring_pending_off(kring)) ||
548 		    (nm_state == NM_OFF && nm_kring_pending_on(kring))) {
549 			MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
550 			if (defq == -1) {
551 				defq = nm_rxq->iq_abs_id;
552 				break;
553 			}
554 		}
555 	}
556 
557 	if (defq == -1) {
558 		/* No active netmap queues.  Switch back to NIC queues. */
559 		rss = vi->rss;
560 		defq = vi->rss[0];
561 	} else {
562 		for (i = 0; i < vi->rss_size;) {
563 			for_each_nm_rxq(vi, j, nm_rxq) {
564 				nm_state = atomic_load_int(&nm_rxq->nm_state);
565 				kring = na->rx_rings[nm_rxq->nid];
566 				if ((nm_state != NM_OFF &&
567 				    !nm_kring_pending_off(kring)) ||
568 				    (nm_state == NM_OFF &&
569 				    nm_kring_pending_on(kring))) {
570 					MPASS(nm_rxq->iq_cntxt_id !=
571 					    INVALID_NM_RXQ_CNTXT_ID);
572 					vi->nm_rss[i++] = nm_rxq->iq_abs_id;
573 					if (i == vi->rss_size)
574 						break;
575 				}
576 			}
577 		}
578 		rss = vi->nm_rss;
579 	}
580 
581 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size, rss,
582 	    vi->rss_size);
583 	if (rc != 0)
584 		if_printf(ifp, "netmap rss_config failed: %d\n", rc);
585 
586 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
587 	if (rc != 0) {
588 		if_printf(ifp, "netmap defaultq config failed: %d\n", rc);
589 	}
590 
591 	return (rc);
592 }
593 
594 /*
595  * Odd number of rx queues work best for split RSS mode as the first queue can
596  * be dedicated for non-RSS traffic and the rest divided into two equal halves.
597  */
598 static int
599 cxgbe_netmap_split_rss(struct adapter *sc, struct vi_info *vi,
600     struct ifnet *ifp, struct netmap_adapter *na)
601 {
602 	struct netmap_kring *kring;
603 	struct sge_nm_rxq *nm_rxq;
604 	int rc, i, j, nm_state, defq;
605 	int nactive[2] = {0, 0};
606 	int dq[2] = {-1, -1};
607 	bool dq_norss;		/* default queue should not be in RSS table. */
608 
609 	MPASS(nm_split_rss != 0);
610 	MPASS(vi->nnmrxq > 1);
611 
612 	for_each_nm_rxq(vi, i, nm_rxq) {
613 		j = i / ((vi->nnmrxq + 1) / 2);
614 		nm_state = atomic_load_int(&nm_rxq->nm_state);
615 		kring = na->rx_rings[nm_rxq->nid];
616 		if ((nm_state != NM_OFF && !nm_kring_pending_off(kring)) ||
617 		    (nm_state == NM_OFF && nm_kring_pending_on(kring))) {
618 			MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
619 			nactive[j]++;
620 			if (dq[j] == -1) {
621 				dq[j] = nm_rxq->iq_abs_id;
622 				break;
623 			}
624 		}
625 	}
626 
627 	if (nactive[0] == 0 || nactive[1] == 0)
628 		return (cxgbe_netmap_simple_rss(sc, vi, ifp, na));
629 
630 	MPASS(dq[0] != -1 && dq[1] != -1);
631 	if (nactive[0] > nactive[1]) {
632 		defq = dq[0];
633 		dq_norss = true;
634 	} else if (nactive[0] < nactive[1]) {
635 		defq = dq[1];
636 		dq_norss = true;
637 	} else {
638 		defq = dq[0];
639 		dq_norss = false;
640 	}
641 
642 	i = 0;
643 	nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq];
644 	while (i < vi->rss_size / 2) {
645 		for (j = 0; j < (vi->nnmrxq + 1) / 2; 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 / 2)
659 				break;
660 		}
661 	}
662 	while (i < vi->rss_size) {
663 		for (j = (vi->nnmrxq + 1) / 2; j < vi->nnmrxq; j++) {
664 			nm_state = atomic_load_int(&nm_rxq[j].nm_state);
665 			kring = na->rx_rings[nm_rxq[j].nid];
666 			if ((nm_state == NM_OFF &&
667 			    !nm_kring_pending_on(kring)) ||
668 			    (nm_state == NM_ON &&
669 			    nm_kring_pending_off(kring))) {
670 				continue;
671 			}
672 			MPASS(nm_rxq[j].iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
673 			if (dq_norss && defq == nm_rxq[j].iq_abs_id)
674 				continue;
675 			vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
676 			if (i == vi->rss_size)
677 				break;
678 		}
679 	}
680 
681 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
682 	    vi->nm_rss, vi->rss_size);
683 	if (rc != 0)
684 		if_printf(ifp, "netmap split_rss_config failed: %d\n", rc);
685 
686 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
687 	if (rc != 0)
688 		if_printf(ifp, "netmap defaultq config failed: %d\n", rc);
689 
690 	return (rc);
691 }
692 
693 static inline int
694 cxgbe_netmap_rss(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
695     struct netmap_adapter *na)
696 {
697 
698 	if (nm_split_rss == 0 || vi->nnmrxq == 1)
699 		return (cxgbe_netmap_simple_rss(sc, vi, ifp, na));
700 	else
701 		return (cxgbe_netmap_split_rss(sc, vi, ifp, na));
702 }
703 
704 static int
705 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
706     struct netmap_adapter *na)
707 {
708 	struct netmap_slot *slot;
709 	struct netmap_kring *kring;
710 	struct sge_nm_rxq *nm_rxq;
711 	struct sge_nm_txq *nm_txq;
712 	int i, j, hwidx;
713 	struct rx_buf_info *rxb;
714 
715 	ASSERT_SYNCHRONIZED_OP(sc);
716 	MPASS(vi->nnmrxq > 0);
717 	MPASS(vi->nnmtxq > 0);
718 
719 	if ((vi->flags & VI_INIT_DONE) == 0 ||
720 	    (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
721 		if_printf(ifp, "cannot enable netmap operation because "
722 		    "interface is not UP.\n");
723 		return (EAGAIN);
724 	}
725 
726 	rxb = &sc->sge.rx_buf_info[0];
727 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
728 		if (rxb->size1 == NETMAP_BUF_SIZE(na)) {
729 			hwidx = rxb->hwidx1;
730 			break;
731 		}
732 		if (rxb->size2 == NETMAP_BUF_SIZE(na)) {
733 			hwidx = rxb->hwidx2;
734 			break;
735 		}
736 	}
737 	if (i >= SW_ZONE_SIZES) {
738 		if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
739 		    NETMAP_BUF_SIZE(na));
740 		return (ENXIO);
741 	}
742 
743 	/* Must set caps before calling netmap_reset */
744 	nm_set_native_flags(na);
745 
746 	for_each_nm_rxq(vi, i, nm_rxq) {
747 		kring = na->rx_rings[nm_rxq->nid];
748 		if (!nm_kring_pending_on(kring))
749 			continue;
750 
751 		alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop));
752 		nm_rxq->fl_hwidx = hwidx;
753 		slot = netmap_reset(na, NR_RX, i, 0);
754 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
755 
756 		/* We deal with 8 bufs at a time */
757 		MPASS((na->num_rx_desc & 7) == 0);
758 		MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
759 		for (j = 0; j < nm_rxq->fl_sidx; j++) {
760 			uint64_t ba;
761 
762 			PNMB(na, &slot[j], &ba);
763 			MPASS(ba != 0);
764 			nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
765 		}
766 		j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
767 		MPASS((j & 7) == 0);
768 		j /= 8;	/* driver pidx to hardware pidx */
769 		wmb();
770 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
771 		    nm_rxq->fl_db_val | V_PIDX(j));
772 
773 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_OFF, NM_ON);
774 	}
775 
776 	for_each_nm_txq(vi, i, nm_txq) {
777 		kring = na->tx_rings[nm_txq->nid];
778 		if (!nm_kring_pending_on(kring))
779 			continue;
780 
781 		alloc_nm_txq_hwq(vi, nm_txq);
782 		slot = netmap_reset(na, NR_TX, i, 0);
783 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
784 	}
785 
786 	if (vi->nm_rss == NULL) {
787 		vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
788 		    M_ZERO | M_WAITOK);
789 	}
790 
791 	return (cxgbe_netmap_rss(sc, vi, ifp, na));
792 }
793 
794 static int
795 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
796     struct netmap_adapter *na)
797 {
798 	struct netmap_kring *kring;
799 	int rc, i, nm_state, nactive;
800 	struct sge_nm_txq *nm_txq;
801 	struct sge_nm_rxq *nm_rxq;
802 
803 	ASSERT_SYNCHRONIZED_OP(sc);
804 	MPASS(vi->nnmrxq > 0);
805 	MPASS(vi->nnmtxq > 0);
806 
807 	if (!nm_netmap_on(na))
808 		return (0);
809 
810 	if ((vi->flags & VI_INIT_DONE) == 0)
811 		return (0);
812 
813 	/* First remove the queues that are stopping from the RSS table. */
814 	rc = cxgbe_netmap_rss(sc, vi, ifp, na);
815 	if (rc != 0)
816 		return (rc);	/* error message logged already. */
817 
818 	for_each_nm_txq(vi, i, nm_txq) {
819 		kring = na->tx_rings[nm_txq->nid];
820 		if (!nm_kring_pending_off(kring))
821 			continue;
822 		MPASS(nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID);
823 
824 		rc = -t4_eth_eq_stop(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
825 		if (rc != 0) {
826 			device_printf(vi->dev,
827 			    "failed to stop nm_txq[%d]: %d.\n", i, rc);
828 			return (rc);
829 		}
830 
831 		/* XXX: netmap, not the driver, should do this. */
832 		kring->rhead = kring->rcur = kring->nr_hwcur = 0;
833 		kring->rtail = kring->nr_hwtail = kring->nkr_num_slots - 1;
834 	}
835 	nactive = 0;
836 	for_each_nm_rxq(vi, i, nm_rxq) {
837 		nm_state = atomic_load_int(&nm_rxq->nm_state);
838 		kring = na->rx_rings[nm_rxq->nid];
839 		if (nm_state != NM_OFF && !nm_kring_pending_off(kring))
840 			nactive++;
841 		if (!nm_kring_pending_off(kring))
842 			continue;
843 		MPASS(nm_state != NM_OFF);
844 		MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
845 
846 		rc = -t4_iq_stop(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
847 		    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
848 		if (rc != 0) {
849 			device_printf(vi->dev,
850 			    "failed to stop nm_rxq[%d]: %d.\n", i, rc);
851 			return (rc);
852 		}
853 
854 		while (!atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_OFF))
855 			pause("nmst", 1);
856 
857 		/* XXX: netmap, not the driver, should do this. */
858 		kring->rhead = kring->rcur = kring->nr_hwcur = 0;
859 		kring->rtail = kring->nr_hwtail = 0;
860 	}
861 	netmap_krings_mode_commit(na, 0);
862 	if (nactive == 0)
863 		nm_clear_native_flags(na);
864 
865 	return (rc);
866 }
867 
868 static int
869 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
870 {
871 	struct ifnet *ifp = na->ifp;
872 	struct vi_info *vi = ifp->if_softc;
873 	struct adapter *sc = vi->adapter;
874 	int rc;
875 
876 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
877 	if (rc != 0)
878 		return (rc);
879 	if (on)
880 		rc = cxgbe_netmap_on(sc, vi, ifp, na);
881 	else
882 		rc = cxgbe_netmap_off(sc, vi, ifp, na);
883 	end_synchronized_op(sc, 0);
884 
885 	return (rc);
886 }
887 
888 /* How many packets can a single type1 WR carry in n descriptors */
889 static inline int
890 ndesc_to_npkt(const int n)
891 {
892 
893 	MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
894 
895 	return (n * 2 - 1);
896 }
897 #define MAX_NPKT_IN_TYPE1_WR	(ndesc_to_npkt(SGE_MAX_WR_NDESC))
898 
899 /*
900  * Space (in descriptors) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
901  * carries n packets
902  */
903 static inline int
904 npkt_to_ndesc(const int n)
905 {
906 
907 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
908 
909 	return ((n + 2) / 2);
910 }
911 
912 /*
913  * Space (in 16B units) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
914  * carries n packets
915  */
916 static inline int
917 npkt_to_len16(const int n)
918 {
919 
920 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
921 
922 	return (n * 2 + 1);
923 }
924 
925 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
926 
927 static void
928 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
929 {
930 	int n;
931 	u_int db = nm_txq->doorbells;
932 
933 	MPASS(nm_txq->pidx != nm_txq->dbidx);
934 
935 	n = NMIDXDIFF(nm_txq, dbidx);
936 	if (n > 1)
937 		clrbit(&db, DOORBELL_WCWR);
938 	wmb();
939 
940 	switch (ffs(db) - 1) {
941 	case DOORBELL_UDB:
942 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
943 		break;
944 
945 	case DOORBELL_WCWR: {
946 		volatile uint64_t *dst, *src;
947 
948 		/*
949 		 * Queues whose 128B doorbell segment fits in the page do not
950 		 * use relative qid (udb_qid is always 0).  Only queues with
951 		 * doorbell segments can do WCWR.
952 		 */
953 		KASSERT(nm_txq->udb_qid == 0 && n == 1,
954 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
955 		    __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
956 
957 		dst = (volatile void *)((uintptr_t)nm_txq->udb +
958 		    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
959 		src = (void *)&nm_txq->desc[nm_txq->dbidx];
960 		while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
961 			*dst++ = *src++;
962 		wmb();
963 		break;
964 	}
965 
966 	case DOORBELL_UDBWC:
967 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
968 		wmb();
969 		break;
970 
971 	case DOORBELL_KDB:
972 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
973 		    V_QID(nm_txq->cntxt_id) | V_PIDX(n));
974 		break;
975 	}
976 	nm_txq->dbidx = nm_txq->pidx;
977 }
978 
979 /*
980  * Write work requests to send 'npkt' frames and ring the doorbell to send them
981  * on their way.  No need to check for wraparound.
982  */
983 static void
984 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
985     struct netmap_kring *kring, int npkt, int npkt_remaining)
986 {
987 	struct netmap_ring *ring = kring->ring;
988 	struct netmap_slot *slot;
989 	const u_int lim = kring->nkr_num_slots - 1;
990 	struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
991 	uint16_t len;
992 	uint64_t ba;
993 	struct cpl_tx_pkt_core *cpl;
994 	struct ulptx_sgl *usgl;
995 	int i, n;
996 
997 	while (npkt) {
998 		n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
999 		len = 0;
1000 
1001 		wr = (void *)&nm_txq->desc[nm_txq->pidx];
1002 		wr->op_pkd = nm_txq->op_pkd;
1003 		wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
1004 		wr->npkt = n;
1005 		wr->r3 = 0;
1006 		wr->type = 1;
1007 		cpl = (void *)(wr + 1);
1008 
1009 		for (i = 0; i < n; i++) {
1010 			slot = &ring->slot[kring->nr_hwcur];
1011 			PNMB(kring->na, slot, &ba);
1012 			MPASS(ba != 0);
1013 
1014 			cpl->ctrl0 = nm_txq->cpl_ctrl0;
1015 			cpl->pack = 0;
1016 			cpl->len = htobe16(slot->len);
1017 			cpl->ctrl1 = nm_txcsum ? 0 :
1018 			    htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
1019 
1020 			usgl = (void *)(cpl + 1);
1021 			usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1022 			    V_ULPTX_NSGE(1));
1023 			usgl->len0 = htobe32(slot->len);
1024 			usgl->addr0 = htobe64(ba);
1025 
1026 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1027 			cpl = (void *)(usgl + 1);
1028 			MPASS(slot->len + len <= UINT16_MAX);
1029 			len += slot->len;
1030 			kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
1031 		}
1032 		wr->plen = htobe16(len);
1033 
1034 		npkt -= n;
1035 		nm_txq->pidx += npkt_to_ndesc(n);
1036 		MPASS(nm_txq->pidx <= nm_txq->sidx);
1037 		if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
1038 			/*
1039 			 * This routine doesn't know how to write WRs that wrap
1040 			 * around.  Make sure it wasn't asked to.
1041 			 */
1042 			MPASS(npkt == 0);
1043 			nm_txq->pidx = 0;
1044 		}
1045 
1046 		if (npkt == 0 && npkt_remaining == 0) {
1047 			/* All done. */
1048 			if (lazy_tx_credit_flush == 0) {
1049 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
1050 				    F_FW_WR_EQUIQ);
1051 				nm_txq->equeqidx = nm_txq->pidx;
1052 				nm_txq->equiqidx = nm_txq->pidx;
1053 			}
1054 			ring_nm_txq_db(sc, nm_txq);
1055 			return;
1056 		}
1057 
1058 		if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
1059 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
1060 			    F_FW_WR_EQUIQ);
1061 			nm_txq->equeqidx = nm_txq->pidx;
1062 			nm_txq->equiqidx = nm_txq->pidx;
1063 		} else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
1064 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1065 			nm_txq->equeqidx = nm_txq->pidx;
1066 		}
1067 		if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
1068 			ring_nm_txq_db(sc, nm_txq);
1069 	}
1070 
1071 	/* Will get called again. */
1072 	MPASS(npkt_remaining);
1073 }
1074 
1075 /* How many contiguous free descriptors starting at pidx */
1076 static inline int
1077 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
1078 {
1079 
1080 	if (nm_txq->cidx > nm_txq->pidx)
1081 		return (nm_txq->cidx - nm_txq->pidx - 1);
1082 	else if (nm_txq->cidx > 0)
1083 		return (nm_txq->sidx - nm_txq->pidx);
1084 	else
1085 		return (nm_txq->sidx - nm_txq->pidx - 1);
1086 }
1087 
1088 static int
1089 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
1090 {
1091 	struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
1092 	uint16_t hw_cidx = spg->cidx;	/* snapshot */
1093 	struct fw_eth_tx_pkts_wr *wr;
1094 	int n = 0;
1095 
1096 	hw_cidx = be16toh(hw_cidx);
1097 
1098 	while (nm_txq->cidx != hw_cidx) {
1099 		wr = (void *)&nm_txq->desc[nm_txq->cidx];
1100 
1101 		MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)) ||
1102 		    wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR)));
1103 		MPASS(wr->type == 1);
1104 		MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
1105 
1106 		n += wr->npkt;
1107 		nm_txq->cidx += npkt_to_ndesc(wr->npkt);
1108 
1109 		/*
1110 		 * We never sent a WR that wrapped around so the credits coming
1111 		 * back, WR by WR, should never cause the cidx to wrap around
1112 		 * either.
1113 		 */
1114 		MPASS(nm_txq->cidx <= nm_txq->sidx);
1115 		if (__predict_false(nm_txq->cidx == nm_txq->sidx))
1116 			nm_txq->cidx = 0;
1117 	}
1118 
1119 	return (n);
1120 }
1121 
1122 static int
1123 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
1124 {
1125 	struct netmap_adapter *na = kring->na;
1126 	struct ifnet *ifp = na->ifp;
1127 	struct vi_info *vi = ifp->if_softc;
1128 	struct adapter *sc = vi->adapter;
1129 	struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
1130 	const u_int head = kring->rhead;
1131 	u_int reclaimed = 0;
1132 	int n, d, npkt_remaining, ndesc_remaining;
1133 
1134 	/*
1135 	 * Tx was at kring->nr_hwcur last time around and now we need to advance
1136 	 * to kring->rhead.  Note that the driver's pidx moves independent of
1137 	 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
1138 	 * between descriptors and frames isn't 1:1).
1139 	 */
1140 
1141 	npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
1142 	    kring->nkr_num_slots - kring->nr_hwcur + head;
1143 	while (npkt_remaining) {
1144 		reclaimed += reclaim_nm_tx_desc(nm_txq);
1145 		ndesc_remaining = contiguous_ndesc_available(nm_txq);
1146 		/* Can't run out of descriptors with packets still remaining */
1147 		MPASS(ndesc_remaining > 0);
1148 
1149 		/* # of desc needed to tx all remaining packets */
1150 		d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
1151 		if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
1152 			d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
1153 
1154 		if (d <= ndesc_remaining)
1155 			n = npkt_remaining;
1156 		else {
1157 			/* Can't send all, calculate how many can be sent */
1158 			n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
1159 			    MAX_NPKT_IN_TYPE1_WR;
1160 			if (ndesc_remaining % SGE_MAX_WR_NDESC)
1161 				n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
1162 		}
1163 
1164 		/* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
1165 		npkt_remaining -= n;
1166 		cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining);
1167 	}
1168 	MPASS(npkt_remaining == 0);
1169 	MPASS(kring->nr_hwcur == head);
1170 	MPASS(nm_txq->dbidx == nm_txq->pidx);
1171 
1172 	/*
1173 	 * Second part: reclaim buffers for completed transmissions.
1174 	 */
1175 	if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
1176 		reclaimed += reclaim_nm_tx_desc(nm_txq);
1177 		kring->nr_hwtail += reclaimed;
1178 		if (kring->nr_hwtail >= kring->nkr_num_slots)
1179 			kring->nr_hwtail -= kring->nkr_num_slots;
1180 	}
1181 
1182 	return (0);
1183 }
1184 
1185 static int
1186 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
1187 {
1188 	struct netmap_adapter *na = kring->na;
1189 	struct netmap_ring *ring = kring->ring;
1190 	struct ifnet *ifp = na->ifp;
1191 	struct vi_info *vi = ifp->if_softc;
1192 	struct adapter *sc = vi->adapter;
1193 	struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
1194 	u_int const head = kring->rhead;
1195 	u_int n;
1196 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1197 
1198 	if (black_hole)
1199 		return (0);	/* No updates ever. */
1200 
1201 	if (netmap_no_pendintr || force_update) {
1202 		kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
1203 		kring->nr_kflags &= ~NKR_PENDINTR;
1204 	}
1205 
1206 	if (nm_rxq->fl_db_saved > 0 && starve_fl == 0) {
1207 		wmb();
1208 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1209 		    nm_rxq->fl_db_val | V_PIDX(nm_rxq->fl_db_saved));
1210 		nm_rxq->fl_db_saved = 0;
1211 	}
1212 
1213 	/* Userspace done with buffers from kring->nr_hwcur to head */
1214 	n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
1215 	    kring->nkr_num_slots - kring->nr_hwcur + head;
1216 	n &= ~7U;
1217 	if (n > 0) {
1218 		u_int fl_pidx = nm_rxq->fl_pidx;
1219 		struct netmap_slot *slot = &ring->slot[fl_pidx];
1220 		uint64_t ba;
1221 		int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
1222 
1223 		/*
1224 		 * We always deal with 8 buffers at a time.  We must have
1225 		 * stopped at an 8B boundary (fl_pidx) last time around and we
1226 		 * must have a multiple of 8B buffers to give to the freelist.
1227 		 */
1228 		MPASS((fl_pidx & 7) == 0);
1229 		MPASS((n & 7) == 0);
1230 
1231 		IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
1232 		IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx2);
1233 
1234 		while (n > 0) {
1235 			for (i = 0; i < 8; i++, fl_pidx++, slot++) {
1236 				PNMB(na, slot, &ba);
1237 				MPASS(ba != 0);
1238 				nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
1239 				slot->flags &= ~NS_BUF_CHANGED;
1240 				MPASS(fl_pidx <= nm_rxq->fl_sidx2);
1241 			}
1242 			n -= 8;
1243 			if (fl_pidx == nm_rxq->fl_sidx2) {
1244 				fl_pidx = 0;
1245 				slot = &ring->slot[0];
1246 			}
1247 			if (++dbinc == nm_rxq->fl_db_threshold) {
1248 				wmb();
1249 				if (starve_fl)
1250 					nm_rxq->fl_db_saved += dbinc;
1251 				else {
1252 					t4_write_reg(sc, sc->sge_kdoorbell_reg,
1253 					    nm_rxq->fl_db_val | V_PIDX(dbinc));
1254 				}
1255 				dbinc = 0;
1256 			}
1257 		}
1258 		MPASS(nm_rxq->fl_pidx == fl_pidx);
1259 
1260 		if (dbinc > 0) {
1261 			wmb();
1262 			if (starve_fl)
1263 				nm_rxq->fl_db_saved += dbinc;
1264 			else {
1265 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1266 				    nm_rxq->fl_db_val | V_PIDX(dbinc));
1267 			}
1268 		}
1269 	}
1270 
1271 	return (0);
1272 }
1273 
1274 void
1275 cxgbe_nm_attach(struct vi_info *vi)
1276 {
1277 	struct port_info *pi;
1278 	struct adapter *sc;
1279 	struct netmap_adapter na;
1280 
1281 	MPASS(vi->nnmrxq > 0);
1282 	MPASS(vi->ifp != NULL);
1283 
1284 	pi = vi->pi;
1285 	sc = pi->adapter;
1286 
1287 	bzero(&na, sizeof(na));
1288 
1289 	na.ifp = vi->ifp;
1290 	na.na_flags = NAF_BDG_MAYSLEEP;
1291 
1292 	/* Netmap doesn't know about the space reserved for the status page. */
1293 	na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
1294 
1295 	/*
1296 	 * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
1297 	 * num_rx_desc is based on the number of buffers that can be held in the
1298 	 * freelist, and not the number of entries in the iq.  (These two are
1299 	 * not exactly the same due to the space taken up by the status page).
1300 	 */
1301 	na.num_rx_desc = rounddown(vi->qsize_rxq, 8);
1302 	na.nm_txsync = cxgbe_netmap_txsync;
1303 	na.nm_rxsync = cxgbe_netmap_rxsync;
1304 	na.nm_register = cxgbe_netmap_reg;
1305 	na.num_tx_rings = vi->nnmtxq;
1306 	na.num_rx_rings = vi->nnmrxq;
1307 	na.rx_buf_maxsize = MAX_MTU;
1308 	netmap_attach(&na);	/* This adds IFCAP_NETMAP to if_capabilities */
1309 }
1310 
1311 void
1312 cxgbe_nm_detach(struct vi_info *vi)
1313 {
1314 
1315 	MPASS(vi->nnmrxq > 0);
1316 	MPASS(vi->ifp != NULL);
1317 
1318 	netmap_detach(vi->ifp);
1319 }
1320 
1321 static inline const void *
1322 unwrap_nm_fw6_msg(const struct cpl_fw6_msg *cpl)
1323 {
1324 
1325 	MPASS(cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL);
1326 
1327 	/* data[0] is RSS header */
1328 	return (&cpl->data[1]);
1329 }
1330 
1331 static void
1332 handle_nm_sge_egr_update(struct adapter *sc, struct ifnet *ifp,
1333     const struct cpl_sge_egr_update *egr)
1334 {
1335 	uint32_t oq;
1336 	struct sge_nm_txq *nm_txq;
1337 
1338 	oq = be32toh(egr->opcode_qid);
1339 	MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
1340 	nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
1341 
1342 	netmap_tx_irq(ifp, nm_txq->nid);
1343 }
1344 
1345 void
1346 service_nm_rxq(struct sge_nm_rxq *nm_rxq)
1347 {
1348 	struct vi_info *vi = nm_rxq->vi;
1349 	struct adapter *sc = vi->adapter;
1350 	struct ifnet *ifp = vi->ifp;
1351 	struct netmap_adapter *na = NA(ifp);
1352 	struct netmap_kring *kring = na->rx_rings[nm_rxq->nid];
1353 	struct netmap_ring *ring = kring->ring;
1354 	struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
1355 	const void *cpl;
1356 	uint32_t lq;
1357 	u_int work = 0;
1358 	uint8_t opcode;
1359 	uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
1360 	u_int fl_credits = fl_cidx & 7;
1361 	u_int ndesc = 0;	/* desc processed since last cidx update */
1362 	u_int nframes = 0;	/* frames processed since last netmap wakeup */
1363 
1364 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
1365 
1366 		rmb();
1367 
1368 		lq = be32toh(d->rsp.pldbuflen_qid);
1369 		opcode = d->rss.opcode;
1370 		cpl = &d->cpl[0];
1371 
1372 		switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
1373 		case X_RSPD_TYPE_FLBUF:
1374 
1375 			/* fall through */
1376 
1377 		case X_RSPD_TYPE_CPL:
1378 			MPASS(opcode < NUM_CPL_CMDS);
1379 
1380 			switch (opcode) {
1381 			case CPL_FW4_MSG:
1382 			case CPL_FW6_MSG:
1383 				cpl = unwrap_nm_fw6_msg(cpl);
1384 				/* fall through */
1385 			case CPL_SGE_EGR_UPDATE:
1386 				handle_nm_sge_egr_update(sc, ifp, cpl);
1387 				break;
1388 			case CPL_RX_PKT:
1389 				ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
1390 				    sc->params.sge.fl_pktshift;
1391 				ring->slot[fl_cidx].flags = 0;
1392 				nframes++;
1393 				if (!(lq & F_RSPD_NEWBUF)) {
1394 					MPASS(black_hole == 2);
1395 					break;
1396 				}
1397 				fl_credits++;
1398 				if (__predict_false(++fl_cidx == nm_rxq->fl_sidx))
1399 					fl_cidx = 0;
1400 				break;
1401 			default:
1402 				panic("%s: unexpected opcode 0x%x on nm_rxq %p",
1403 				    __func__, opcode, nm_rxq);
1404 			}
1405 			break;
1406 
1407 		case X_RSPD_TYPE_INTR:
1408 			/* Not equipped to handle forwarded interrupts. */
1409 			panic("%s: netmap queue received interrupt for iq %u\n",
1410 			    __func__, lq);
1411 
1412 		default:
1413 			panic("%s: illegal response type %d on nm_rxq %p",
1414 			    __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
1415 		}
1416 
1417 		d++;
1418 		if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
1419 			nm_rxq->iq_cidx = 0;
1420 			d = &nm_rxq->iq_desc[0];
1421 			nm_rxq->iq_gen ^= F_RSPD_GEN;
1422 		}
1423 
1424 		if (__predict_false(++nframes == rx_nframes) && !black_hole) {
1425 			atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1426 			netmap_rx_irq(ifp, nm_rxq->nid, &work);
1427 			nframes = 0;
1428 		}
1429 
1430 		if (__predict_false(++ndesc == rx_ndesc)) {
1431 			if (black_hole && fl_credits >= 8) {
1432 				fl_credits /= 8;
1433 				IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
1434 				    nm_rxq->fl_sidx);
1435 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1436 				    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1437 				fl_credits = fl_cidx & 7;
1438 			}
1439 			t4_write_reg(sc, sc->sge_gts_reg,
1440 			    V_CIDXINC(ndesc) |
1441 			    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
1442 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1443 			ndesc = 0;
1444 		}
1445 	}
1446 
1447 	atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1448 	if (black_hole) {
1449 		fl_credits /= 8;
1450 		IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1451 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1452 		    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1453 	} else if (nframes > 0)
1454 		netmap_rx_irq(ifp, nm_rxq->nid, &work);
1455 
1456     	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndesc) |
1457 	    V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1458 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1459 }
1460 #endif
1461