xref: /freebsd/sys/dev/cxgbe/t4_netmap.c (revision f2d48b5e2c3b45850585e4d7aee324fe148afbf2)
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, struct sysctl_oid *oid)
129 {
130 	int rc;
131 	struct sysctl_oid_list *children;
132 	struct sysctl_ctx_list *ctx;
133 	char name[16];
134 	size_t len;
135 	struct adapter *sc = vi->adapter;
136 	struct netmap_adapter *na = NA(vi->ifp);
137 
138 	MPASS(na != NULL);
139 
140 	len = vi->qsize_rxq * IQ_ESIZE;
141 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
142 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
143 	if (rc != 0)
144 		return (rc);
145 
146 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
147 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
148 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
149 	if (rc != 0)
150 		return (rc);
151 
152 	nm_rxq->vi = vi;
153 	nm_rxq->nid = idx;
154 	nm_rxq->iq_cidx = 0;
155 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
156 	nm_rxq->iq_gen = F_RSPD_GEN;
157 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
158 	nm_rxq->fl_sidx = na->num_rx_desc;
159 	nm_rxq->fl_sidx2 = nm_rxq->fl_sidx;	/* copy for rxsync cacheline */
160 	nm_rxq->intr_idx = intr_idx;
161 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
162 
163 	ctx = &vi->ctx;
164 	children = SYSCTL_CHILDREN(oid);
165 
166 	snprintf(name, sizeof(name), "%d", idx);
167 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name,
168 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queue");
169 	children = SYSCTL_CHILDREN(oid);
170 
171 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
172 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &nm_rxq->iq_abs_id,
173 	    0, sysctl_uint16, "I", "absolute id of the queue");
174 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
175 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &nm_rxq->iq_cntxt_id,
176 	    0, sysctl_uint16, "I", "SGE context id of the queue");
177 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
178 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &nm_rxq->iq_cidx, 0,
179 	    sysctl_uint16, "I", "consumer index");
180 
181 	children = SYSCTL_CHILDREN(oid);
182 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl",
183 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist");
184 	children = SYSCTL_CHILDREN(oid);
185 
186 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
187 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &nm_rxq->fl_cntxt_id,
188 	    0, sysctl_uint16, "I", "SGE context id of the freelist");
189 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
190 	    &nm_rxq->fl_cidx, 0, "consumer index");
191 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
192 	    &nm_rxq->fl_pidx, 0, "producer index");
193 
194 	return (rc);
195 }
196 
197 int
198 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
199 {
200 	struct adapter *sc = vi->adapter;
201 
202 	if (!(vi->flags & VI_INIT_DONE))
203 		return (0);
204 
205 	if (nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID)
206 		free_nm_rxq_hwq(vi, nm_rxq);
207 	MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
208 
209 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
210 	    nm_rxq->iq_desc);
211 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
212 	    nm_rxq->fl_desc);
213 
214 	return (0);
215 }
216 
217 int
218 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
219     struct sysctl_oid *oid)
220 {
221 	int rc;
222 	size_t len;
223 	struct port_info *pi = vi->pi;
224 	struct adapter *sc = pi->adapter;
225 	struct netmap_adapter *na = NA(vi->ifp);
226 	char name[16];
227 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
228 
229 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
230 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
231 	    &nm_txq->ba, (void **)&nm_txq->desc);
232 	if (rc)
233 		return (rc);
234 
235 	nm_txq->pidx = nm_txq->cidx = 0;
236 	nm_txq->sidx = na->num_tx_desc;
237 	nm_txq->nid = idx;
238 	nm_txq->iqidx = iqidx;
239 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
240 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
241 	    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
242 	if (sc->params.fw_vers >= FW_VERSION32(1, 24, 11, 0))
243 		nm_txq->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR));
244 	else
245 		nm_txq->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
246 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
247 
248 	snprintf(name, sizeof(name), "%d", idx);
249 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name,
250 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "netmap tx queue");
251 	children = SYSCTL_CHILDREN(oid);
252 
253 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
254 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
255 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
256 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &nm_txq->cidx, 0,
257 	    sysctl_uint16, "I", "consumer index");
258 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
259 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &nm_txq->pidx, 0,
260 	    sysctl_uint16, "I", "producer index");
261 
262 	return (rc);
263 }
264 
265 int
266 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
267 {
268 	struct adapter *sc = vi->adapter;
269 
270 	if (!(vi->flags & VI_INIT_DONE))
271 		return (0);
272 
273 	if (nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID)
274 		free_nm_txq_hwq(vi, nm_txq);
275 	MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
276 
277 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
278 	    nm_txq->desc);
279 
280 	return (0);
281 }
282 
283 static int
284 alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int cong)
285 {
286 	int rc, cntxt_id, i;
287 	__be32 v;
288 	struct adapter *sc = vi->adapter;
289 	struct sge_params *sp = &sc->params.sge;
290 	struct netmap_adapter *na = NA(vi->ifp);
291 	struct fw_iq_cmd c;
292 
293 	MPASS(na != NULL);
294 	MPASS(nm_rxq->iq_desc != NULL);
295 	MPASS(nm_rxq->fl_desc != NULL);
296 
297 	bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE);
298 	bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + sp->spg_len);
299 
300 	bzero(&c, sizeof(c));
301 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
302 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
303 	    V_FW_IQ_CMD_VFN(0));
304 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_IQSTART | FW_LEN16(c));
305 	if (nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID)
306 		c.alloc_to_len16 |= htobe32(F_FW_IQ_CMD_ALLOC);
307 	else {
308 		c.iqid = htobe16(nm_rxq->iq_cntxt_id);
309 		c.fl0id = htobe16(nm_rxq->fl_cntxt_id);
310 		c.fl1id = htobe16(0xffff);
311 		c.physiqid = htobe16(nm_rxq->iq_abs_id);
312 	}
313 	MPASS(!forwarding_intr_to_fwq(sc));
314 	KASSERT(nm_rxq->intr_idx < sc->intr_count,
315 	    ("%s: invalid direct intr_idx %d", __func__, nm_rxq->intr_idx));
316 	v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
317 	c.type_to_iqandstindex = htobe32(v |
318 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
319 	    V_FW_IQ_CMD_VIID(vi->viid) |
320 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
321 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(vi->pi->tx_chan) |
322 	    F_FW_IQ_CMD_IQGTSMODE |
323 	    V_FW_IQ_CMD_IQINTCNTTHRESH(0) |
324 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
325 	c.iqsize = htobe16(vi->qsize_rxq);
326 	c.iqaddr = htobe64(nm_rxq->iq_ba);
327 	if (cong >= 0) {
328 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN |
329 		    V_FW_IQ_CMD_FL0CNGCHMAP(cong) | F_FW_IQ_CMD_FL0CONGCIF |
330 		    F_FW_IQ_CMD_FL0CONGEN);
331 	}
332 	c.iqns_to_fl0congen |=
333 	    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
334 		F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
335 		(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
336 		(black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0));
337 	c.fl0dcaen_to_fl0cidxfthresh =
338 	    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
339 		X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
340 		V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
341 		X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
342 	c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE);
343 	c.fl0addr = htobe64(nm_rxq->fl_ba);
344 
345 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
346 	if (rc != 0) {
347 		device_printf(sc->dev,
348 		    "failed to create netmap ingress queue: %d\n", rc);
349 		return (rc);
350 	}
351 
352 	nm_rxq->iq_cidx = 0;
353 	MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE);
354 	nm_rxq->iq_gen = F_RSPD_GEN;
355 	nm_rxq->iq_cntxt_id = be16toh(c.iqid);
356 	nm_rxq->iq_abs_id = be16toh(c.physiqid);
357 	cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
358 	if (cntxt_id >= sc->sge.iqmap_sz) {
359 		panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
360 		    __func__, cntxt_id, sc->sge.iqmap_sz - 1);
361 	}
362 	sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
363 
364 	nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
365 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
366 	nm_rxq->fl_db_saved = 0;
367 	/* matches the X_FETCHBURSTMAX_512B or X_FETCHBURSTMAX_256B above. */
368 	nm_rxq->fl_db_threshold = chip_id(sc) <= CHELSIO_T5 ? 8 : 4;
369 	MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
370 	cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
371 	if (cntxt_id >= sc->sge.eqmap_sz) {
372 		panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
373 		    __func__, cntxt_id, sc->sge.eqmap_sz - 1);
374 	}
375 	sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
376 
377 	nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) |
378 	    sc->chip_params->sge_fl_db;
379 
380 	if (chip_id(sc) >= CHELSIO_T5 && cong >= 0) {
381 		uint32_t param, val;
382 
383 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
384 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
385 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
386 		if (cong == 0)
387 			val = 1 << 19;
388 		else {
389 			val = 2 << 19;
390 			for (i = 0; i < 4; i++) {
391 				if (cong & (1 << i))
392 					val |= 1 << (i << 2);
393 			}
394 		}
395 
396 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
397 		if (rc != 0) {
398 			/* report error but carry on */
399 			device_printf(sc->dev,
400 			    "failed to set congestion manager context for "
401 			    "ingress queue %d: %d\n", nm_rxq->iq_cntxt_id, rc);
402 		}
403 	}
404 
405 	t4_write_reg(sc, sc->sge_gts_reg,
406 	    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
407 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
408 
409 	return (rc);
410 }
411 
412 static int
413 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
414 {
415 	struct adapter *sc = vi->adapter;
416 	int rc;
417 
418 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
419 	    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
420 	if (rc != 0)
421 		device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
422 		    __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
423 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
424 	return (rc);
425 }
426 
427 static int
428 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
429 {
430 	int rc, cntxt_id;
431 	size_t len;
432 	struct adapter *sc = vi->adapter;
433 	struct netmap_adapter *na = NA(vi->ifp);
434 	struct fw_eq_eth_cmd c;
435 
436 	MPASS(na != NULL);
437 	MPASS(nm_txq->desc != NULL);
438 
439 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
440 	bzero(nm_txq->desc, len);
441 
442 	bzero(&c, sizeof(c));
443 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
444 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
445 	    V_FW_EQ_ETH_CMD_VFN(0));
446 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
447 	if (nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID)
448 		c.alloc_to_len16 |= htobe32(F_FW_EQ_ETH_CMD_ALLOC);
449 	else
450 		c.eqid_pkd = htobe32(V_FW_EQ_ETH_CMD_EQID(nm_txq->cntxt_id));
451 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
452 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
453 	c.fetchszm_to_iqid =
454 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
455 		V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
456 		V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
457 	c.dcaen_to_eqsize =
458 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
459 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
460 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
461 		V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
462 	c.eqaddr = htobe64(nm_txq->ba);
463 
464 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
465 	if (rc != 0) {
466 		device_printf(vi->dev,
467 		    "failed to create netmap egress queue: %d\n", rc);
468 		return (rc);
469 	}
470 
471 	nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
472 	cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
473 	if (cntxt_id >= sc->sge.eqmap_sz)
474 	    panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
475 		cntxt_id, sc->sge.eqmap_sz - 1);
476 	sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
477 
478 	nm_txq->pidx = nm_txq->cidx = 0;
479 	MPASS(nm_txq->sidx == na->num_tx_desc);
480 	nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
481 
482 	nm_txq->doorbells = sc->doorbells;
483 	if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
484 	    isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
485 	    isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
486 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
487 		uint32_t mask = (1 << s_qpp) - 1;
488 		volatile uint8_t *udb;
489 
490 		udb = sc->udbs_base + UDBS_DB_OFFSET;
491 		udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
492 		nm_txq->udb_qid = nm_txq->cntxt_id & mask;
493 		if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
494 	    		clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
495 		else {
496 			udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
497 			nm_txq->udb_qid = 0;
498 		}
499 		nm_txq->udb = (volatile void *)udb;
500 	}
501 
502 	if (sc->params.fw_vers < FW_VERSION32(1, 25, 1, 0)) {
503 		uint32_t param, val;
504 
505 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
506 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) |
507 		    V_FW_PARAMS_PARAM_YZ(nm_txq->cntxt_id);
508 		val = 0xff;
509 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
510 		if (rc != 0) {
511 			device_printf(vi->dev,
512 			    "failed to bind netmap txq %d to class 0xff: %d\n",
513 			    nm_txq->cntxt_id, rc);
514 			rc = 0;
515 		}
516 	}
517 
518 	return (rc);
519 }
520 
521 static int
522 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
523 {
524 	struct adapter *sc = vi->adapter;
525 	int rc;
526 
527 	rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
528 	if (rc != 0)
529 		device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
530 		    nm_txq->cntxt_id, rc);
531 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
532 	return (rc);
533 }
534 
535 static int
536 cxgbe_netmap_simple_rss(struct adapter *sc, struct vi_info *vi,
537     struct ifnet *ifp, struct netmap_adapter *na)
538 {
539 	struct netmap_kring *kring;
540 	struct sge_nm_rxq *nm_rxq;
541 	int rc, i, j, nm_state, defq;
542 	uint16_t *rss;
543 
544 	/*
545 	 * Check if there's at least one active (or about to go active) netmap
546 	 * rx queue.
547 	 */
548 	defq = -1;
549 	for_each_nm_rxq(vi, j, nm_rxq) {
550 		nm_state = atomic_load_int(&nm_rxq->nm_state);
551 		kring = na->rx_rings[nm_rxq->nid];
552 		if ((nm_state != NM_OFF && !nm_kring_pending_off(kring)) ||
553 		    (nm_state == NM_OFF && nm_kring_pending_on(kring))) {
554 			MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
555 			if (defq == -1) {
556 				defq = nm_rxq->iq_abs_id;
557 				break;
558 			}
559 		}
560 	}
561 
562 	if (defq == -1) {
563 		/* No active netmap queues.  Switch back to NIC queues. */
564 		rss = vi->rss;
565 		defq = vi->rss[0];
566 	} else {
567 		for (i = 0; i < vi->rss_size;) {
568 			for_each_nm_rxq(vi, j, nm_rxq) {
569 				nm_state = atomic_load_int(&nm_rxq->nm_state);
570 				kring = na->rx_rings[nm_rxq->nid];
571 				if ((nm_state != NM_OFF &&
572 				    !nm_kring_pending_off(kring)) ||
573 				    (nm_state == NM_OFF &&
574 				    nm_kring_pending_on(kring))) {
575 					MPASS(nm_rxq->iq_cntxt_id !=
576 					    INVALID_NM_RXQ_CNTXT_ID);
577 					vi->nm_rss[i++] = nm_rxq->iq_abs_id;
578 					if (i == vi->rss_size)
579 						break;
580 				}
581 			}
582 		}
583 		rss = vi->nm_rss;
584 	}
585 
586 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size, rss,
587 	    vi->rss_size);
588 	if (rc != 0)
589 		if_printf(ifp, "netmap rss_config failed: %d\n", rc);
590 
591 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
592 	if (rc != 0) {
593 		if_printf(ifp, "netmap defaultq config failed: %d\n", rc);
594 	}
595 
596 	return (rc);
597 }
598 
599 /*
600  * Odd number of rx queues work best for split RSS mode as the first queue can
601  * be dedicated for non-RSS traffic and the rest divided into two equal halves.
602  */
603 static int
604 cxgbe_netmap_split_rss(struct adapter *sc, struct vi_info *vi,
605     struct ifnet *ifp, struct netmap_adapter *na)
606 {
607 	struct netmap_kring *kring;
608 	struct sge_nm_rxq *nm_rxq;
609 	int rc, i, j, nm_state, defq;
610 	int nactive[2] = {0, 0};
611 	int dq[2] = {-1, -1};
612 	bool dq_norss;		/* default queue should not be in RSS table. */
613 
614 	MPASS(nm_split_rss != 0);
615 	MPASS(vi->nnmrxq > 1);
616 
617 	for_each_nm_rxq(vi, i, nm_rxq) {
618 		j = i / ((vi->nnmrxq + 1) / 2);
619 		nm_state = atomic_load_int(&nm_rxq->nm_state);
620 		kring = na->rx_rings[nm_rxq->nid];
621 		if ((nm_state != NM_OFF && !nm_kring_pending_off(kring)) ||
622 		    (nm_state == NM_OFF && nm_kring_pending_on(kring))) {
623 			MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
624 			nactive[j]++;
625 			if (dq[j] == -1) {
626 				dq[j] = nm_rxq->iq_abs_id;
627 				break;
628 			}
629 		}
630 	}
631 
632 	if (nactive[0] == 0 || nactive[1] == 0)
633 		return (cxgbe_netmap_simple_rss(sc, vi, ifp, na));
634 
635 	MPASS(dq[0] != -1 && dq[1] != -1);
636 	if (nactive[0] > nactive[1]) {
637 		defq = dq[0];
638 		dq_norss = true;
639 	} else if (nactive[0] < nactive[1]) {
640 		defq = dq[1];
641 		dq_norss = true;
642 	} else {
643 		defq = dq[0];
644 		dq_norss = false;
645 	}
646 
647 	i = 0;
648 	nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq];
649 	while (i < vi->rss_size / 2) {
650 		for (j = 0; j < (vi->nnmrxq + 1) / 2; j++) {
651 			nm_state = atomic_load_int(&nm_rxq[j].nm_state);
652 			kring = na->rx_rings[nm_rxq[j].nid];
653 			if ((nm_state == NM_OFF &&
654 			    !nm_kring_pending_on(kring)) ||
655 			    (nm_state == NM_ON &&
656 			    nm_kring_pending_off(kring))) {
657 				continue;
658 			}
659 			MPASS(nm_rxq[j].iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
660 			if (dq_norss && defq == nm_rxq[j].iq_abs_id)
661 				continue;
662 			vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
663 			if (i == vi->rss_size / 2)
664 				break;
665 		}
666 	}
667 	while (i < vi->rss_size) {
668 		for (j = (vi->nnmrxq + 1) / 2; j < vi->nnmrxq; j++) {
669 			nm_state = atomic_load_int(&nm_rxq[j].nm_state);
670 			kring = na->rx_rings[nm_rxq[j].nid];
671 			if ((nm_state == NM_OFF &&
672 			    !nm_kring_pending_on(kring)) ||
673 			    (nm_state == NM_ON &&
674 			    nm_kring_pending_off(kring))) {
675 				continue;
676 			}
677 			MPASS(nm_rxq[j].iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
678 			if (dq_norss && defq == nm_rxq[j].iq_abs_id)
679 				continue;
680 			vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
681 			if (i == vi->rss_size)
682 				break;
683 		}
684 	}
685 
686 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
687 	    vi->nm_rss, vi->rss_size);
688 	if (rc != 0)
689 		if_printf(ifp, "netmap split_rss_config failed: %d\n", rc);
690 
691 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
692 	if (rc != 0)
693 		if_printf(ifp, "netmap defaultq config failed: %d\n", rc);
694 
695 	return (rc);
696 }
697 
698 static inline int
699 cxgbe_netmap_rss(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
700     struct netmap_adapter *na)
701 {
702 
703 	if (nm_split_rss == 0 || vi->nnmrxq == 1)
704 		return (cxgbe_netmap_simple_rss(sc, vi, ifp, na));
705 	else
706 		return (cxgbe_netmap_split_rss(sc, vi, ifp, na));
707 }
708 
709 static int
710 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
711     struct netmap_adapter *na)
712 {
713 	struct netmap_slot *slot;
714 	struct netmap_kring *kring;
715 	struct sge_nm_rxq *nm_rxq;
716 	struct sge_nm_txq *nm_txq;
717 	int i, j, hwidx;
718 	struct rx_buf_info *rxb;
719 
720 	ASSERT_SYNCHRONIZED_OP(sc);
721 	MPASS(vi->nnmrxq > 0);
722 	MPASS(vi->nnmtxq > 0);
723 
724 	if ((vi->flags & VI_INIT_DONE) == 0 ||
725 	    (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
726 		if_printf(ifp, "cannot enable netmap operation because "
727 		    "interface is not UP.\n");
728 		return (EAGAIN);
729 	}
730 
731 	rxb = &sc->sge.rx_buf_info[0];
732 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
733 		if (rxb->size1 == NETMAP_BUF_SIZE(na)) {
734 			hwidx = rxb->hwidx1;
735 			break;
736 		}
737 		if (rxb->size2 == NETMAP_BUF_SIZE(na)) {
738 			hwidx = rxb->hwidx2;
739 			break;
740 		}
741 	}
742 	if (i >= SW_ZONE_SIZES) {
743 		if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
744 		    NETMAP_BUF_SIZE(na));
745 		return (ENXIO);
746 	}
747 
748 	/* Must set caps before calling netmap_reset */
749 	nm_set_native_flags(na);
750 
751 	for_each_nm_rxq(vi, i, nm_rxq) {
752 		kring = na->rx_rings[nm_rxq->nid];
753 		if (!nm_kring_pending_on(kring))
754 			continue;
755 
756 		alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop));
757 		nm_rxq->fl_hwidx = hwidx;
758 		slot = netmap_reset(na, NR_RX, i, 0);
759 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
760 
761 		/* We deal with 8 bufs at a time */
762 		MPASS((na->num_rx_desc & 7) == 0);
763 		MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
764 		for (j = 0; j < nm_rxq->fl_sidx; j++) {
765 			uint64_t ba;
766 
767 			PNMB(na, &slot[j], &ba);
768 			MPASS(ba != 0);
769 			nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
770 		}
771 		j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
772 		MPASS((j & 7) == 0);
773 		j /= 8;	/* driver pidx to hardware pidx */
774 		wmb();
775 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
776 		    nm_rxq->fl_db_val | V_PIDX(j));
777 
778 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_OFF, NM_ON);
779 	}
780 
781 	for_each_nm_txq(vi, i, nm_txq) {
782 		kring = na->tx_rings[nm_txq->nid];
783 		if (!nm_kring_pending_on(kring))
784 			continue;
785 
786 		alloc_nm_txq_hwq(vi, nm_txq);
787 		slot = netmap_reset(na, NR_TX, i, 0);
788 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
789 	}
790 
791 	if (vi->nm_rss == NULL) {
792 		vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
793 		    M_ZERO | M_WAITOK);
794 	}
795 
796 	return (cxgbe_netmap_rss(sc, vi, ifp, na));
797 }
798 
799 static int
800 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
801     struct netmap_adapter *na)
802 {
803 	struct netmap_kring *kring;
804 	int rc, i, nm_state, nactive;
805 	struct sge_nm_txq *nm_txq;
806 	struct sge_nm_rxq *nm_rxq;
807 
808 	ASSERT_SYNCHRONIZED_OP(sc);
809 	MPASS(vi->nnmrxq > 0);
810 	MPASS(vi->nnmtxq > 0);
811 
812 	if (!nm_netmap_on(na))
813 		return (0);
814 
815 	if ((vi->flags & VI_INIT_DONE) == 0)
816 		return (0);
817 
818 	/* First remove the queues that are stopping from the RSS table. */
819 	rc = cxgbe_netmap_rss(sc, vi, ifp, na);
820 	if (rc != 0)
821 		return (rc);	/* error message logged already. */
822 
823 	for_each_nm_txq(vi, i, nm_txq) {
824 		kring = na->tx_rings[nm_txq->nid];
825 		if (!nm_kring_pending_off(kring))
826 			continue;
827 		MPASS(nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID);
828 
829 		rc = -t4_eth_eq_stop(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
830 		if (rc != 0) {
831 			device_printf(vi->dev,
832 			    "failed to stop nm_txq[%d]: %d.\n", i, rc);
833 			return (rc);
834 		}
835 
836 		/* XXX: netmap, not the driver, should do this. */
837 		kring->rhead = kring->rcur = kring->nr_hwcur = 0;
838 		kring->rtail = kring->nr_hwtail = kring->nkr_num_slots - 1;
839 	}
840 	nactive = 0;
841 	for_each_nm_rxq(vi, i, nm_rxq) {
842 		nm_state = atomic_load_int(&nm_rxq->nm_state);
843 		kring = na->rx_rings[nm_rxq->nid];
844 		if (nm_state != NM_OFF && !nm_kring_pending_off(kring))
845 			nactive++;
846 		if (!nm_kring_pending_off(kring))
847 			continue;
848 		MPASS(nm_state != NM_OFF);
849 		MPASS(nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID);
850 
851 		rc = -t4_iq_stop(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
852 		    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
853 		if (rc != 0) {
854 			device_printf(vi->dev,
855 			    "failed to stop nm_rxq[%d]: %d.\n", i, rc);
856 			return (rc);
857 		}
858 
859 		while (!atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_OFF))
860 			pause("nmst", 1);
861 
862 		/* XXX: netmap, not the driver, should do this. */
863 		kring->rhead = kring->rcur = kring->nr_hwcur = 0;
864 		kring->rtail = kring->nr_hwtail = 0;
865 	}
866 	netmap_krings_mode_commit(na, 0);
867 	if (nactive == 0)
868 		nm_clear_native_flags(na);
869 
870 	return (rc);
871 }
872 
873 static int
874 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
875 {
876 	struct ifnet *ifp = na->ifp;
877 	struct vi_info *vi = ifp->if_softc;
878 	struct adapter *sc = vi->adapter;
879 	int rc;
880 
881 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
882 	if (rc != 0)
883 		return (rc);
884 	if (on)
885 		rc = cxgbe_netmap_on(sc, vi, ifp, na);
886 	else
887 		rc = cxgbe_netmap_off(sc, vi, ifp, na);
888 	end_synchronized_op(sc, 0);
889 
890 	return (rc);
891 }
892 
893 /* How many packets can a single type1 WR carry in n descriptors */
894 static inline int
895 ndesc_to_npkt(const int n)
896 {
897 
898 	MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
899 
900 	return (n * 2 - 1);
901 }
902 #define MAX_NPKT_IN_TYPE1_WR	(ndesc_to_npkt(SGE_MAX_WR_NDESC))
903 
904 /*
905  * Space (in descriptors) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
906  * carries n packets
907  */
908 static inline int
909 npkt_to_ndesc(const int n)
910 {
911 
912 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
913 
914 	return ((n + 2) / 2);
915 }
916 
917 /*
918  * Space (in 16B units) needed for a type1 WR (TX_PKTS or TX_PKTS2) that
919  * carries n packets
920  */
921 static inline int
922 npkt_to_len16(const int n)
923 {
924 
925 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
926 
927 	return (n * 2 + 1);
928 }
929 
930 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
931 
932 static void
933 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
934 {
935 	int n;
936 	u_int db = nm_txq->doorbells;
937 
938 	MPASS(nm_txq->pidx != nm_txq->dbidx);
939 
940 	n = NMIDXDIFF(nm_txq, dbidx);
941 	if (n > 1)
942 		clrbit(&db, DOORBELL_WCWR);
943 	wmb();
944 
945 	switch (ffs(db) - 1) {
946 	case DOORBELL_UDB:
947 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
948 		break;
949 
950 	case DOORBELL_WCWR: {
951 		volatile uint64_t *dst, *src;
952 
953 		/*
954 		 * Queues whose 128B doorbell segment fits in the page do not
955 		 * use relative qid (udb_qid is always 0).  Only queues with
956 		 * doorbell segments can do WCWR.
957 		 */
958 		KASSERT(nm_txq->udb_qid == 0 && n == 1,
959 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
960 		    __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
961 
962 		dst = (volatile void *)((uintptr_t)nm_txq->udb +
963 		    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
964 		src = (void *)&nm_txq->desc[nm_txq->dbidx];
965 		while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
966 			*dst++ = *src++;
967 		wmb();
968 		break;
969 	}
970 
971 	case DOORBELL_UDBWC:
972 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
973 		wmb();
974 		break;
975 
976 	case DOORBELL_KDB:
977 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
978 		    V_QID(nm_txq->cntxt_id) | V_PIDX(n));
979 		break;
980 	}
981 	nm_txq->dbidx = nm_txq->pidx;
982 }
983 
984 /*
985  * Write work requests to send 'npkt' frames and ring the doorbell to send them
986  * on their way.  No need to check for wraparound.
987  */
988 static void
989 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
990     struct netmap_kring *kring, int npkt, int npkt_remaining)
991 {
992 	struct netmap_ring *ring = kring->ring;
993 	struct netmap_slot *slot;
994 	const u_int lim = kring->nkr_num_slots - 1;
995 	struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
996 	uint16_t len;
997 	uint64_t ba;
998 	struct cpl_tx_pkt_core *cpl;
999 	struct ulptx_sgl *usgl;
1000 	int i, n;
1001 
1002 	while (npkt) {
1003 		n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
1004 		len = 0;
1005 
1006 		wr = (void *)&nm_txq->desc[nm_txq->pidx];
1007 		wr->op_pkd = nm_txq->op_pkd;
1008 		wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
1009 		wr->npkt = n;
1010 		wr->r3 = 0;
1011 		wr->type = 1;
1012 		cpl = (void *)(wr + 1);
1013 
1014 		for (i = 0; i < n; i++) {
1015 			slot = &ring->slot[kring->nr_hwcur];
1016 			PNMB(kring->na, slot, &ba);
1017 			MPASS(ba != 0);
1018 
1019 			cpl->ctrl0 = nm_txq->cpl_ctrl0;
1020 			cpl->pack = 0;
1021 			cpl->len = htobe16(slot->len);
1022 			cpl->ctrl1 = nm_txcsum ? 0 :
1023 			    htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
1024 
1025 			usgl = (void *)(cpl + 1);
1026 			usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1027 			    V_ULPTX_NSGE(1));
1028 			usgl->len0 = htobe32(slot->len);
1029 			usgl->addr0 = htobe64(ba);
1030 
1031 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1032 			cpl = (void *)(usgl + 1);
1033 			MPASS(slot->len + len <= UINT16_MAX);
1034 			len += slot->len;
1035 			kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
1036 		}
1037 		wr->plen = htobe16(len);
1038 
1039 		npkt -= n;
1040 		nm_txq->pidx += npkt_to_ndesc(n);
1041 		MPASS(nm_txq->pidx <= nm_txq->sidx);
1042 		if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
1043 			/*
1044 			 * This routine doesn't know how to write WRs that wrap
1045 			 * around.  Make sure it wasn't asked to.
1046 			 */
1047 			MPASS(npkt == 0);
1048 			nm_txq->pidx = 0;
1049 		}
1050 
1051 		if (npkt == 0 && npkt_remaining == 0) {
1052 			/* All done. */
1053 			if (lazy_tx_credit_flush == 0) {
1054 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
1055 				    F_FW_WR_EQUIQ);
1056 				nm_txq->equeqidx = nm_txq->pidx;
1057 				nm_txq->equiqidx = nm_txq->pidx;
1058 			}
1059 			ring_nm_txq_db(sc, nm_txq);
1060 			return;
1061 		}
1062 
1063 		if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
1064 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
1065 			    F_FW_WR_EQUIQ);
1066 			nm_txq->equeqidx = nm_txq->pidx;
1067 			nm_txq->equiqidx = nm_txq->pidx;
1068 		} else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
1069 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1070 			nm_txq->equeqidx = nm_txq->pidx;
1071 		}
1072 		if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
1073 			ring_nm_txq_db(sc, nm_txq);
1074 	}
1075 
1076 	/* Will get called again. */
1077 	MPASS(npkt_remaining);
1078 }
1079 
1080 /* How many contiguous free descriptors starting at pidx */
1081 static inline int
1082 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
1083 {
1084 
1085 	if (nm_txq->cidx > nm_txq->pidx)
1086 		return (nm_txq->cidx - nm_txq->pidx - 1);
1087 	else if (nm_txq->cidx > 0)
1088 		return (nm_txq->sidx - nm_txq->pidx);
1089 	else
1090 		return (nm_txq->sidx - nm_txq->pidx - 1);
1091 }
1092 
1093 static int
1094 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
1095 {
1096 	struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
1097 	uint16_t hw_cidx = spg->cidx;	/* snapshot */
1098 	struct fw_eth_tx_pkts_wr *wr;
1099 	int n = 0;
1100 
1101 	hw_cidx = be16toh(hw_cidx);
1102 
1103 	while (nm_txq->cidx != hw_cidx) {
1104 		wr = (void *)&nm_txq->desc[nm_txq->cidx];
1105 
1106 		MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)) ||
1107 		    wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR)));
1108 		MPASS(wr->type == 1);
1109 		MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
1110 
1111 		n += wr->npkt;
1112 		nm_txq->cidx += npkt_to_ndesc(wr->npkt);
1113 
1114 		/*
1115 		 * We never sent a WR that wrapped around so the credits coming
1116 		 * back, WR by WR, should never cause the cidx to wrap around
1117 		 * either.
1118 		 */
1119 		MPASS(nm_txq->cidx <= nm_txq->sidx);
1120 		if (__predict_false(nm_txq->cidx == nm_txq->sidx))
1121 			nm_txq->cidx = 0;
1122 	}
1123 
1124 	return (n);
1125 }
1126 
1127 static int
1128 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
1129 {
1130 	struct netmap_adapter *na = kring->na;
1131 	struct ifnet *ifp = na->ifp;
1132 	struct vi_info *vi = ifp->if_softc;
1133 	struct adapter *sc = vi->adapter;
1134 	struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
1135 	const u_int head = kring->rhead;
1136 	u_int reclaimed = 0;
1137 	int n, d, npkt_remaining, ndesc_remaining;
1138 
1139 	/*
1140 	 * Tx was at kring->nr_hwcur last time around and now we need to advance
1141 	 * to kring->rhead.  Note that the driver's pidx moves independent of
1142 	 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
1143 	 * between descriptors and frames isn't 1:1).
1144 	 */
1145 
1146 	npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
1147 	    kring->nkr_num_slots - kring->nr_hwcur + head;
1148 	while (npkt_remaining) {
1149 		reclaimed += reclaim_nm_tx_desc(nm_txq);
1150 		ndesc_remaining = contiguous_ndesc_available(nm_txq);
1151 		/* Can't run out of descriptors with packets still remaining */
1152 		MPASS(ndesc_remaining > 0);
1153 
1154 		/* # of desc needed to tx all remaining packets */
1155 		d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
1156 		if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
1157 			d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
1158 
1159 		if (d <= ndesc_remaining)
1160 			n = npkt_remaining;
1161 		else {
1162 			/* Can't send all, calculate how many can be sent */
1163 			n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
1164 			    MAX_NPKT_IN_TYPE1_WR;
1165 			if (ndesc_remaining % SGE_MAX_WR_NDESC)
1166 				n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
1167 		}
1168 
1169 		/* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
1170 		npkt_remaining -= n;
1171 		cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining);
1172 	}
1173 	MPASS(npkt_remaining == 0);
1174 	MPASS(kring->nr_hwcur == head);
1175 	MPASS(nm_txq->dbidx == nm_txq->pidx);
1176 
1177 	/*
1178 	 * Second part: reclaim buffers for completed transmissions.
1179 	 */
1180 	if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
1181 		reclaimed += reclaim_nm_tx_desc(nm_txq);
1182 		kring->nr_hwtail += reclaimed;
1183 		if (kring->nr_hwtail >= kring->nkr_num_slots)
1184 			kring->nr_hwtail -= kring->nkr_num_slots;
1185 	}
1186 
1187 	return (0);
1188 }
1189 
1190 static int
1191 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
1192 {
1193 	struct netmap_adapter *na = kring->na;
1194 	struct netmap_ring *ring = kring->ring;
1195 	struct ifnet *ifp = na->ifp;
1196 	struct vi_info *vi = ifp->if_softc;
1197 	struct adapter *sc = vi->adapter;
1198 	struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
1199 	u_int const head = kring->rhead;
1200 	u_int n;
1201 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1202 
1203 	if (black_hole)
1204 		return (0);	/* No updates ever. */
1205 
1206 	if (netmap_no_pendintr || force_update) {
1207 		kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
1208 		kring->nr_kflags &= ~NKR_PENDINTR;
1209 	}
1210 
1211 	if (nm_rxq->fl_db_saved > 0 && starve_fl == 0) {
1212 		wmb();
1213 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1214 		    nm_rxq->fl_db_val | V_PIDX(nm_rxq->fl_db_saved));
1215 		nm_rxq->fl_db_saved = 0;
1216 	}
1217 
1218 	/* Userspace done with buffers from kring->nr_hwcur to head */
1219 	n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
1220 	    kring->nkr_num_slots - kring->nr_hwcur + head;
1221 	n &= ~7U;
1222 	if (n > 0) {
1223 		u_int fl_pidx = nm_rxq->fl_pidx;
1224 		struct netmap_slot *slot = &ring->slot[fl_pidx];
1225 		uint64_t ba;
1226 		int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
1227 
1228 		/*
1229 		 * We always deal with 8 buffers at a time.  We must have
1230 		 * stopped at an 8B boundary (fl_pidx) last time around and we
1231 		 * must have a multiple of 8B buffers to give to the freelist.
1232 		 */
1233 		MPASS((fl_pidx & 7) == 0);
1234 		MPASS((n & 7) == 0);
1235 
1236 		IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
1237 		IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx2);
1238 
1239 		while (n > 0) {
1240 			for (i = 0; i < 8; i++, fl_pidx++, slot++) {
1241 				PNMB(na, slot, &ba);
1242 				MPASS(ba != 0);
1243 				nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
1244 				slot->flags &= ~NS_BUF_CHANGED;
1245 				MPASS(fl_pidx <= nm_rxq->fl_sidx2);
1246 			}
1247 			n -= 8;
1248 			if (fl_pidx == nm_rxq->fl_sidx2) {
1249 				fl_pidx = 0;
1250 				slot = &ring->slot[0];
1251 			}
1252 			if (++dbinc == nm_rxq->fl_db_threshold) {
1253 				wmb();
1254 				if (starve_fl)
1255 					nm_rxq->fl_db_saved += dbinc;
1256 				else {
1257 					t4_write_reg(sc, sc->sge_kdoorbell_reg,
1258 					    nm_rxq->fl_db_val | V_PIDX(dbinc));
1259 				}
1260 				dbinc = 0;
1261 			}
1262 		}
1263 		MPASS(nm_rxq->fl_pidx == fl_pidx);
1264 
1265 		if (dbinc > 0) {
1266 			wmb();
1267 			if (starve_fl)
1268 				nm_rxq->fl_db_saved += dbinc;
1269 			else {
1270 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1271 				    nm_rxq->fl_db_val | V_PIDX(dbinc));
1272 			}
1273 		}
1274 	}
1275 
1276 	return (0);
1277 }
1278 
1279 void
1280 cxgbe_nm_attach(struct vi_info *vi)
1281 {
1282 	struct port_info *pi;
1283 	struct adapter *sc;
1284 	struct netmap_adapter na;
1285 
1286 	MPASS(vi->nnmrxq > 0);
1287 	MPASS(vi->ifp != NULL);
1288 
1289 	pi = vi->pi;
1290 	sc = pi->adapter;
1291 
1292 	bzero(&na, sizeof(na));
1293 
1294 	na.ifp = vi->ifp;
1295 	na.na_flags = NAF_BDG_MAYSLEEP;
1296 
1297 	/* Netmap doesn't know about the space reserved for the status page. */
1298 	na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
1299 
1300 	/*
1301 	 * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
1302 	 * num_rx_desc is based on the number of buffers that can be held in the
1303 	 * freelist, and not the number of entries in the iq.  (These two are
1304 	 * not exactly the same due to the space taken up by the status page).
1305 	 */
1306 	na.num_rx_desc = rounddown(vi->qsize_rxq, 8);
1307 	na.nm_txsync = cxgbe_netmap_txsync;
1308 	na.nm_rxsync = cxgbe_netmap_rxsync;
1309 	na.nm_register = cxgbe_netmap_reg;
1310 	na.num_tx_rings = vi->nnmtxq;
1311 	na.num_rx_rings = vi->nnmrxq;
1312 	na.rx_buf_maxsize = MAX_MTU;
1313 	netmap_attach(&na);	/* This adds IFCAP_NETMAP to if_capabilities */
1314 }
1315 
1316 void
1317 cxgbe_nm_detach(struct vi_info *vi)
1318 {
1319 
1320 	MPASS(vi->nnmrxq > 0);
1321 	MPASS(vi->ifp != NULL);
1322 
1323 	netmap_detach(vi->ifp);
1324 }
1325 
1326 static inline const void *
1327 unwrap_nm_fw6_msg(const struct cpl_fw6_msg *cpl)
1328 {
1329 
1330 	MPASS(cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL);
1331 
1332 	/* data[0] is RSS header */
1333 	return (&cpl->data[1]);
1334 }
1335 
1336 static void
1337 handle_nm_sge_egr_update(struct adapter *sc, struct ifnet *ifp,
1338     const struct cpl_sge_egr_update *egr)
1339 {
1340 	uint32_t oq;
1341 	struct sge_nm_txq *nm_txq;
1342 
1343 	oq = be32toh(egr->opcode_qid);
1344 	MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
1345 	nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
1346 
1347 	netmap_tx_irq(ifp, nm_txq->nid);
1348 }
1349 
1350 void
1351 service_nm_rxq(struct sge_nm_rxq *nm_rxq)
1352 {
1353 	struct vi_info *vi = nm_rxq->vi;
1354 	struct adapter *sc = vi->adapter;
1355 	struct ifnet *ifp = vi->ifp;
1356 	struct netmap_adapter *na = NA(ifp);
1357 	struct netmap_kring *kring = na->rx_rings[nm_rxq->nid];
1358 	struct netmap_ring *ring = kring->ring;
1359 	struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
1360 	const void *cpl;
1361 	uint32_t lq;
1362 	u_int work = 0;
1363 	uint8_t opcode;
1364 	uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
1365 	u_int fl_credits = fl_cidx & 7;
1366 	u_int ndesc = 0;	/* desc processed since last cidx update */
1367 	u_int nframes = 0;	/* frames processed since last netmap wakeup */
1368 
1369 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
1370 
1371 		rmb();
1372 
1373 		lq = be32toh(d->rsp.pldbuflen_qid);
1374 		opcode = d->rss.opcode;
1375 		cpl = &d->cpl[0];
1376 
1377 		switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
1378 		case X_RSPD_TYPE_FLBUF:
1379 
1380 			/* fall through */
1381 
1382 		case X_RSPD_TYPE_CPL:
1383 			MPASS(opcode < NUM_CPL_CMDS);
1384 
1385 			switch (opcode) {
1386 			case CPL_FW4_MSG:
1387 			case CPL_FW6_MSG:
1388 				cpl = unwrap_nm_fw6_msg(cpl);
1389 				/* fall through */
1390 			case CPL_SGE_EGR_UPDATE:
1391 				handle_nm_sge_egr_update(sc, ifp, cpl);
1392 				break;
1393 			case CPL_RX_PKT:
1394 				ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
1395 				    sc->params.sge.fl_pktshift;
1396 				ring->slot[fl_cidx].flags = 0;
1397 				nframes++;
1398 				if (!(lq & F_RSPD_NEWBUF)) {
1399 					MPASS(black_hole == 2);
1400 					break;
1401 				}
1402 				fl_credits++;
1403 				if (__predict_false(++fl_cidx == nm_rxq->fl_sidx))
1404 					fl_cidx = 0;
1405 				break;
1406 			default:
1407 				panic("%s: unexpected opcode 0x%x on nm_rxq %p",
1408 				    __func__, opcode, nm_rxq);
1409 			}
1410 			break;
1411 
1412 		case X_RSPD_TYPE_INTR:
1413 			/* Not equipped to handle forwarded interrupts. */
1414 			panic("%s: netmap queue received interrupt for iq %u\n",
1415 			    __func__, lq);
1416 
1417 		default:
1418 			panic("%s: illegal response type %d on nm_rxq %p",
1419 			    __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
1420 		}
1421 
1422 		d++;
1423 		if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
1424 			nm_rxq->iq_cidx = 0;
1425 			d = &nm_rxq->iq_desc[0];
1426 			nm_rxq->iq_gen ^= F_RSPD_GEN;
1427 		}
1428 
1429 		if (__predict_false(++nframes == rx_nframes) && !black_hole) {
1430 			atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1431 			netmap_rx_irq(ifp, nm_rxq->nid, &work);
1432 			nframes = 0;
1433 		}
1434 
1435 		if (__predict_false(++ndesc == rx_ndesc)) {
1436 			if (black_hole && fl_credits >= 8) {
1437 				fl_credits /= 8;
1438 				IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
1439 				    nm_rxq->fl_sidx);
1440 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1441 				    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1442 				fl_credits = fl_cidx & 7;
1443 			}
1444 			t4_write_reg(sc, sc->sge_gts_reg,
1445 			    V_CIDXINC(ndesc) |
1446 			    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
1447 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1448 			ndesc = 0;
1449 		}
1450 	}
1451 
1452 	atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1453 	if (black_hole) {
1454 		fl_credits /= 8;
1455 		IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1456 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1457 		    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1458 	} else if (nframes > 0)
1459 		netmap_rx_irq(ifp, nm_rxq->nid, &work);
1460 
1461     	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndesc) |
1462 	    V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1463 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1464 }
1465 #endif
1466