xref: /freebsd/sys/dev/cxgbe/t4_netmap.c (revision 3231251399f9e60fc7175b6352eb5291e9296fe8)
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_RDTUN, &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_RDTUN,
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 static int
114 alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int cong)
115 {
116 	int rc, cntxt_id, i;
117 	__be32 v;
118 	struct adapter *sc = vi->pi->adapter;
119 	struct sge_params *sp = &sc->params.sge;
120 	struct netmap_adapter *na = NA(vi->ifp);
121 	struct fw_iq_cmd c;
122 
123 	MPASS(na != NULL);
124 	MPASS(nm_rxq->iq_desc != NULL);
125 	MPASS(nm_rxq->fl_desc != NULL);
126 
127 	bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE);
128 	bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + sp->spg_len);
129 
130 	bzero(&c, sizeof(c));
131 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
132 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
133 	    V_FW_IQ_CMD_VFN(0));
134 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
135 	    FW_LEN16(c));
136 	MPASS(!forwarding_intr_to_fwq(sc));
137 	KASSERT(nm_rxq->intr_idx < sc->intr_count,
138 	    ("%s: invalid direct intr_idx %d", __func__, nm_rxq->intr_idx));
139 	v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
140 	c.type_to_iqandstindex = htobe32(v |
141 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
142 	    V_FW_IQ_CMD_VIID(vi->viid) |
143 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
144 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(vi->pi->tx_chan) |
145 	    F_FW_IQ_CMD_IQGTSMODE |
146 	    V_FW_IQ_CMD_IQINTCNTTHRESH(0) |
147 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
148 	c.iqsize = htobe16(vi->qsize_rxq);
149 	c.iqaddr = htobe64(nm_rxq->iq_ba);
150 	if (cong >= 0) {
151 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN |
152 		    V_FW_IQ_CMD_FL0CNGCHMAP(cong) | F_FW_IQ_CMD_FL0CONGCIF |
153 		    F_FW_IQ_CMD_FL0CONGEN);
154 	}
155 	c.iqns_to_fl0congen |=
156 	    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
157 		F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
158 		(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
159 		(black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0));
160 	c.fl0dcaen_to_fl0cidxfthresh =
161 	    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
162 		X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
163 		V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
164 		X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
165 	c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE);
166 	c.fl0addr = htobe64(nm_rxq->fl_ba);
167 
168 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
169 	if (rc != 0) {
170 		device_printf(sc->dev,
171 		    "failed to create netmap ingress queue: %d\n", rc);
172 		return (rc);
173 	}
174 
175 	nm_rxq->iq_cidx = 0;
176 	MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE);
177 	nm_rxq->iq_gen = F_RSPD_GEN;
178 	nm_rxq->iq_cntxt_id = be16toh(c.iqid);
179 	nm_rxq->iq_abs_id = be16toh(c.physiqid);
180 	cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
181 	if (cntxt_id >= sc->sge.niq) {
182 		panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
183 		    __func__, cntxt_id, sc->sge.niq - 1);
184 	}
185 	sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
186 
187 	nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
188 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
189 	MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
190 	cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
191 	if (cntxt_id >= sc->sge.neq) {
192 		panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
193 		    __func__, cntxt_id, sc->sge.neq - 1);
194 	}
195 	sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
196 
197 	nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) |
198 	    sc->chip_params->sge_fl_db;
199 
200 	if (chip_id(sc) >= CHELSIO_T5 && cong >= 0) {
201 		uint32_t param, val;
202 
203 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
204 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
205 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
206 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
207 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
208 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
209 		if (cong == 0)
210 			val = 1 << 19;
211 		else {
212 			val = 2 << 19;
213 			for (i = 0; i < 4; i++) {
214 				if (cong & (1 << i))
215 					val |= 1 << (i << 2);
216 			}
217 		}
218 
219 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
220 		if (rc != 0) {
221 			/* report error but carry on */
222 			device_printf(sc->dev,
223 			    "failed to set congestion manager context for "
224 			    "ingress queue %d: %d\n", nm_rxq->iq_cntxt_id, rc);
225 		}
226 	}
227 
228 	t4_write_reg(sc, sc->sge_gts_reg,
229 	    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
230 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
231 
232 	return (rc);
233 }
234 
235 static int
236 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
237 {
238 	struct adapter *sc = vi->pi->adapter;
239 	int rc;
240 
241 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
242 	    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
243 	if (rc != 0)
244 		device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
245 		    __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
246 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
247 	return (rc);
248 }
249 
250 static int
251 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
252 {
253 	int rc, cntxt_id;
254 	size_t len;
255 	struct adapter *sc = vi->pi->adapter;
256 	struct netmap_adapter *na = NA(vi->ifp);
257 	struct fw_eq_eth_cmd c;
258 
259 	MPASS(na != NULL);
260 	MPASS(nm_txq->desc != NULL);
261 
262 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
263 	bzero(nm_txq->desc, len);
264 
265 	bzero(&c, sizeof(c));
266 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
267 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
268 	    V_FW_EQ_ETH_CMD_VFN(0));
269 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
270 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
271 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
272 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
273 	c.fetchszm_to_iqid =
274 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
275 		V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
276 		V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
277 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
278 		      V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
279 		      V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
280 	c.eqaddr = htobe64(nm_txq->ba);
281 
282 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
283 	if (rc != 0) {
284 		device_printf(vi->dev,
285 		    "failed to create netmap egress queue: %d\n", rc);
286 		return (rc);
287 	}
288 
289 	nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
290 	cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
291 	if (cntxt_id >= sc->sge.neq)
292 	    panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
293 		cntxt_id, sc->sge.neq - 1);
294 	sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
295 
296 	nm_txq->pidx = nm_txq->cidx = 0;
297 	MPASS(nm_txq->sidx == na->num_tx_desc);
298 	nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
299 
300 	nm_txq->doorbells = sc->doorbells;
301 	if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
302 	    isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
303 	    isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
304 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
305 		uint32_t mask = (1 << s_qpp) - 1;
306 		volatile uint8_t *udb;
307 
308 		udb = sc->udbs_base + UDBS_DB_OFFSET;
309 		udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
310 		nm_txq->udb_qid = nm_txq->cntxt_id & mask;
311 		if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
312 	    		clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
313 		else {
314 			udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
315 			nm_txq->udb_qid = 0;
316 		}
317 		nm_txq->udb = (volatile void *)udb;
318 	}
319 
320 	return (rc);
321 }
322 
323 static int
324 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
325 {
326 	struct adapter *sc = vi->pi->adapter;
327 	int rc;
328 
329 	rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
330 	if (rc != 0)
331 		device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
332 		    nm_txq->cntxt_id, rc);
333 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
334 	return (rc);
335 }
336 
337 static int
338 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
339     struct netmap_adapter *na)
340 {
341 	struct netmap_slot *slot;
342 	struct netmap_kring *kring;
343 	struct sge_nm_rxq *nm_rxq;
344 	struct sge_nm_txq *nm_txq;
345 	int rc, i, j, hwidx, defq, nrssq;
346 	struct hw_buf_info *hwb;
347 
348 	ASSERT_SYNCHRONIZED_OP(sc);
349 
350 	if ((vi->flags & VI_INIT_DONE) == 0 ||
351 	    (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
352 		return (EAGAIN);
353 
354 	hwb = &sc->sge.hw_buf_info[0];
355 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
356 		if (hwb->size == NETMAP_BUF_SIZE(na))
357 			break;
358 	}
359 	if (i >= SGE_FLBUF_SIZES) {
360 		if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
361 		    NETMAP_BUF_SIZE(na));
362 		return (ENXIO);
363 	}
364 	hwidx = i;
365 
366 	/* Must set caps before calling netmap_reset */
367 	nm_set_native_flags(na);
368 
369 	for_each_nm_rxq(vi, i, nm_rxq) {
370 		kring = na->rx_rings[nm_rxq->nid];
371 		if (!nm_kring_pending_on(kring) ||
372 		    nm_rxq->iq_cntxt_id != INVALID_NM_RXQ_CNTXT_ID)
373 			continue;
374 
375 		alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop));
376 		nm_rxq->fl_hwidx = hwidx;
377 		slot = netmap_reset(na, NR_RX, i, 0);
378 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
379 
380 		/* We deal with 8 bufs at a time */
381 		MPASS((na->num_rx_desc & 7) == 0);
382 		MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
383 		for (j = 0; j < nm_rxq->fl_sidx; j++) {
384 			uint64_t ba;
385 
386 			PNMB(na, &slot[j], &ba);
387 			MPASS(ba != 0);
388 			nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
389 		}
390 		j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
391 		MPASS((j & 7) == 0);
392 		j /= 8;	/* driver pidx to hardware pidx */
393 		wmb();
394 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
395 		    nm_rxq->fl_db_val | V_PIDX(j));
396 
397 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_OFF, NM_ON);
398 	}
399 
400 	for_each_nm_txq(vi, i, nm_txq) {
401 		kring = na->tx_rings[nm_txq->nid];
402 		if (!nm_kring_pending_on(kring) ||
403 		    nm_txq->cntxt_id != INVALID_NM_TXQ_CNTXT_ID)
404 			continue;
405 
406 		alloc_nm_txq_hwq(vi, nm_txq);
407 		slot = netmap_reset(na, NR_TX, i, 0);
408 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
409 	}
410 
411 	if (vi->nm_rss == NULL) {
412 		vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
413 		    M_ZERO | M_WAITOK);
414 	}
415 
416 	MPASS(vi->nnmrxq > 0);
417 	if (nm_split_rss == 0 || vi->nnmrxq == 1) {
418 		for (i = 0; i < vi->rss_size;) {
419 			for_each_nm_rxq(vi, j, nm_rxq) {
420 				vi->nm_rss[i++] = nm_rxq->iq_abs_id;
421 				if (i == vi->rss_size)
422 					break;
423 			}
424 		}
425 		defq = vi->nm_rss[0];
426 	} else {
427 		/* We have multiple queues and we want to split the table. */
428 		MPASS(nm_split_rss != 0);
429 		MPASS(vi->nnmrxq > 1);
430 
431 		nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq];
432 		nrssq = vi->nnmrxq;
433 		if (vi->nnmrxq & 1) {
434 			/*
435 			 * Odd number of queues. The first rxq is designated the
436 			 * default queue, the rest are split evenly.
437 			 */
438 			defq = nm_rxq->iq_abs_id;
439 			nm_rxq++;
440 			nrssq--;
441 		} else {
442 			/*
443 			 * Even number of queues split into two halves.  The
444 			 * first rxq in one of the halves is designated the
445 			 * default queue.
446 			 */
447 #if 1
448 			/* First rxq in the first half. */
449 			defq = nm_rxq->iq_abs_id;
450 #else
451 			/* First rxq in the second half. */
452 			defq = nm_rxq[vi->nnmrxq / 2].iq_abs_id;
453 #endif
454 		}
455 
456 		i = 0;
457 		while (i < vi->rss_size / 2) {
458 			for (j = 0; j < nrssq / 2; j++) {
459 				vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
460 				if (i == vi->rss_size / 2)
461 					break;
462 			}
463 		}
464 		while (i < vi->rss_size) {
465 			for (j = nrssq / 2; j < nrssq; j++) {
466 				vi->nm_rss[i++] = nm_rxq[j].iq_abs_id;
467 				if (i == vi->rss_size)
468 					break;
469 			}
470 		}
471 	}
472 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
473 	    vi->nm_rss, vi->rss_size);
474 	if (rc != 0)
475 		if_printf(ifp, "netmap rss_config failed: %d\n", rc);
476 
477 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, defq, 0, 0);
478 	if (rc != 0)
479 		if_printf(ifp, "netmap rss hash/defaultq config failed: %d\n", rc);
480 
481 	return (rc);
482 }
483 
484 static int
485 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
486     struct netmap_adapter *na)
487 {
488 	struct netmap_kring *kring;
489 	int rc, i;
490 	struct sge_nm_txq *nm_txq;
491 	struct sge_nm_rxq *nm_rxq;
492 
493 	ASSERT_SYNCHRONIZED_OP(sc);
494 
495 	if (!nm_netmap_on(na))
496 		return (0);
497 
498 	if ((vi->flags & VI_INIT_DONE) == 0)
499 		return (0);
500 
501 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
502 	    vi->rss, vi->rss_size);
503 	if (rc != 0)
504 		if_printf(ifp, "failed to restore RSS config: %d\n", rc);
505 	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, vi->rss[0], 0, 0);
506 	if (rc != 0)
507 		if_printf(ifp, "failed to restore RSS hash/defaultq: %d\n", rc);
508 	nm_clear_native_flags(na);
509 
510 	for_each_nm_txq(vi, i, nm_txq) {
511 		struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
512 
513 		kring = na->tx_rings[nm_txq->nid];
514 		if (!nm_kring_pending_off(kring) ||
515 		    nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID)
516 			continue;
517 
518 		/* Wait for hw pidx to catch up ... */
519 		while (be16toh(nm_txq->pidx) != spg->pidx)
520 			pause("nmpidx", 1);
521 
522 		/* ... and then for the cidx. */
523 		while (spg->pidx != spg->cidx)
524 			pause("nmcidx", 1);
525 
526 		free_nm_txq_hwq(vi, nm_txq);
527 	}
528 	for_each_nm_rxq(vi, i, nm_rxq) {
529 		kring = na->rx_rings[nm_rxq->nid];
530 		if (!nm_kring_pending_off(kring) ||
531 		    nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID)
532 			continue;
533 
534 		while (!atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_OFF))
535 			pause("nmst", 1);
536 
537 		free_nm_rxq_hwq(vi, nm_rxq);
538 	}
539 
540 	return (rc);
541 }
542 
543 static int
544 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
545 {
546 	struct ifnet *ifp = na->ifp;
547 	struct vi_info *vi = ifp->if_softc;
548 	struct adapter *sc = vi->pi->adapter;
549 	int rc;
550 
551 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
552 	if (rc != 0)
553 		return (rc);
554 	if (on)
555 		rc = cxgbe_netmap_on(sc, vi, ifp, na);
556 	else
557 		rc = cxgbe_netmap_off(sc, vi, ifp, na);
558 	end_synchronized_op(sc, 0);
559 
560 	return (rc);
561 }
562 
563 /* How many packets can a single type1 WR carry in n descriptors */
564 static inline int
565 ndesc_to_npkt(const int n)
566 {
567 
568 	MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
569 
570 	return (n * 2 - 1);
571 }
572 #define MAX_NPKT_IN_TYPE1_WR	(ndesc_to_npkt(SGE_MAX_WR_NDESC))
573 
574 /* Space (in descriptors) needed for a type1 WR that carries n packets */
575 static inline int
576 npkt_to_ndesc(const int n)
577 {
578 
579 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
580 
581 	return ((n + 2) / 2);
582 }
583 
584 /* Space (in 16B units) needed for a type1 WR that carries n packets */
585 static inline int
586 npkt_to_len16(const int n)
587 {
588 
589 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
590 
591 	return (n * 2 + 1);
592 }
593 
594 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
595 
596 static void
597 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
598 {
599 	int n;
600 	u_int db = nm_txq->doorbells;
601 
602 	MPASS(nm_txq->pidx != nm_txq->dbidx);
603 
604 	n = NMIDXDIFF(nm_txq, dbidx);
605 	if (n > 1)
606 		clrbit(&db, DOORBELL_WCWR);
607 	wmb();
608 
609 	switch (ffs(db) - 1) {
610 	case DOORBELL_UDB:
611 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
612 		break;
613 
614 	case DOORBELL_WCWR: {
615 		volatile uint64_t *dst, *src;
616 
617 		/*
618 		 * Queues whose 128B doorbell segment fits in the page do not
619 		 * use relative qid (udb_qid is always 0).  Only queues with
620 		 * doorbell segments can do WCWR.
621 		 */
622 		KASSERT(nm_txq->udb_qid == 0 && n == 1,
623 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
624 		    __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
625 
626 		dst = (volatile void *)((uintptr_t)nm_txq->udb +
627 		    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
628 		src = (void *)&nm_txq->desc[nm_txq->dbidx];
629 		while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
630 			*dst++ = *src++;
631 		wmb();
632 		break;
633 	}
634 
635 	case DOORBELL_UDBWC:
636 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
637 		wmb();
638 		break;
639 
640 	case DOORBELL_KDB:
641 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
642 		    V_QID(nm_txq->cntxt_id) | V_PIDX(n));
643 		break;
644 	}
645 	nm_txq->dbidx = nm_txq->pidx;
646 }
647 
648 /*
649  * Write work requests to send 'npkt' frames and ring the doorbell to send them
650  * on their way.  No need to check for wraparound.
651  */
652 static void
653 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
654     struct netmap_kring *kring, int npkt, int npkt_remaining, int txcsum)
655 {
656 	struct netmap_ring *ring = kring->ring;
657 	struct netmap_slot *slot;
658 	const u_int lim = kring->nkr_num_slots - 1;
659 	struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
660 	uint16_t len;
661 	uint64_t ba;
662 	struct cpl_tx_pkt_core *cpl;
663 	struct ulptx_sgl *usgl;
664 	int i, n;
665 
666 	while (npkt) {
667 		n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
668 		len = 0;
669 
670 		wr = (void *)&nm_txq->desc[nm_txq->pidx];
671 		wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
672 		wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
673 		wr->npkt = n;
674 		wr->r3 = 0;
675 		wr->type = 1;
676 		cpl = (void *)(wr + 1);
677 
678 		for (i = 0; i < n; i++) {
679 			slot = &ring->slot[kring->nr_hwcur];
680 			PNMB(kring->na, slot, &ba);
681 			MPASS(ba != 0);
682 
683 			cpl->ctrl0 = nm_txq->cpl_ctrl0;
684 			cpl->pack = 0;
685 			cpl->len = htobe16(slot->len);
686 			/*
687 			 * netmap(4) says "netmap does not use features such as
688 			 * checksum offloading, TCP segmentation offloading,
689 			 * encryption, VLAN encapsulation/decapsulation, etc."
690 			 *
691 			 * So the ncxl interfaces have tx hardware checksumming
692 			 * disabled by default.  But you can override netmap by
693 			 * enabling IFCAP_TXCSUM on the interface manully.
694 			 */
695 			cpl->ctrl1 = txcsum ? 0 :
696 			    htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
697 
698 			usgl = (void *)(cpl + 1);
699 			usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
700 			    V_ULPTX_NSGE(1));
701 			usgl->len0 = htobe32(slot->len);
702 			usgl->addr0 = htobe64(ba);
703 
704 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
705 			cpl = (void *)(usgl + 1);
706 			MPASS(slot->len + len <= UINT16_MAX);
707 			len += slot->len;
708 			kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
709 		}
710 		wr->plen = htobe16(len);
711 
712 		npkt -= n;
713 		nm_txq->pidx += npkt_to_ndesc(n);
714 		MPASS(nm_txq->pidx <= nm_txq->sidx);
715 		if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
716 			/*
717 			 * This routine doesn't know how to write WRs that wrap
718 			 * around.  Make sure it wasn't asked to.
719 			 */
720 			MPASS(npkt == 0);
721 			nm_txq->pidx = 0;
722 		}
723 
724 		if (npkt == 0 && npkt_remaining == 0) {
725 			/* All done. */
726 			if (lazy_tx_credit_flush == 0) {
727 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
728 				    F_FW_WR_EQUIQ);
729 				nm_txq->equeqidx = nm_txq->pidx;
730 				nm_txq->equiqidx = nm_txq->pidx;
731 			}
732 			ring_nm_txq_db(sc, nm_txq);
733 			return;
734 		}
735 
736 		if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
737 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
738 			    F_FW_WR_EQUIQ);
739 			nm_txq->equeqidx = nm_txq->pidx;
740 			nm_txq->equiqidx = nm_txq->pidx;
741 		} else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
742 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
743 			nm_txq->equeqidx = nm_txq->pidx;
744 		}
745 		if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
746 			ring_nm_txq_db(sc, nm_txq);
747 	}
748 
749 	/* Will get called again. */
750 	MPASS(npkt_remaining);
751 }
752 
753 /* How many contiguous free descriptors starting at pidx */
754 static inline int
755 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
756 {
757 
758 	if (nm_txq->cidx > nm_txq->pidx)
759 		return (nm_txq->cidx - nm_txq->pidx - 1);
760 	else if (nm_txq->cidx > 0)
761 		return (nm_txq->sidx - nm_txq->pidx);
762 	else
763 		return (nm_txq->sidx - nm_txq->pidx - 1);
764 }
765 
766 static int
767 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
768 {
769 	struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
770 	uint16_t hw_cidx = spg->cidx;	/* snapshot */
771 	struct fw_eth_tx_pkts_wr *wr;
772 	int n = 0;
773 
774 	hw_cidx = be16toh(hw_cidx);
775 
776 	while (nm_txq->cidx != hw_cidx) {
777 		wr = (void *)&nm_txq->desc[nm_txq->cidx];
778 
779 		MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)));
780 		MPASS(wr->type == 1);
781 		MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
782 
783 		n += wr->npkt;
784 		nm_txq->cidx += npkt_to_ndesc(wr->npkt);
785 
786 		/*
787 		 * We never sent a WR that wrapped around so the credits coming
788 		 * back, WR by WR, should never cause the cidx to wrap around
789 		 * either.
790 		 */
791 		MPASS(nm_txq->cidx <= nm_txq->sidx);
792 		if (__predict_false(nm_txq->cidx == nm_txq->sidx))
793 			nm_txq->cidx = 0;
794 	}
795 
796 	return (n);
797 }
798 
799 static int
800 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
801 {
802 	struct netmap_adapter *na = kring->na;
803 	struct ifnet *ifp = na->ifp;
804 	struct vi_info *vi = ifp->if_softc;
805 	struct adapter *sc = vi->pi->adapter;
806 	struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
807 	const u_int head = kring->rhead;
808 	u_int reclaimed = 0;
809 	int n, d, npkt_remaining, ndesc_remaining, txcsum;
810 
811 	/*
812 	 * Tx was at kring->nr_hwcur last time around and now we need to advance
813 	 * to kring->rhead.  Note that the driver's pidx moves independent of
814 	 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
815 	 * between descriptors and frames isn't 1:1).
816 	 */
817 
818 	npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
819 	    kring->nkr_num_slots - kring->nr_hwcur + head;
820 	txcsum = ifp->if_capenable & (IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6);
821 	while (npkt_remaining) {
822 		reclaimed += reclaim_nm_tx_desc(nm_txq);
823 		ndesc_remaining = contiguous_ndesc_available(nm_txq);
824 		/* Can't run out of descriptors with packets still remaining */
825 		MPASS(ndesc_remaining > 0);
826 
827 		/* # of desc needed to tx all remaining packets */
828 		d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
829 		if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
830 			d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
831 
832 		if (d <= ndesc_remaining)
833 			n = npkt_remaining;
834 		else {
835 			/* Can't send all, calculate how many can be sent */
836 			n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
837 			    MAX_NPKT_IN_TYPE1_WR;
838 			if (ndesc_remaining % SGE_MAX_WR_NDESC)
839 				n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
840 		}
841 
842 		/* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
843 		npkt_remaining -= n;
844 		cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining, txcsum);
845 	}
846 	MPASS(npkt_remaining == 0);
847 	MPASS(kring->nr_hwcur == head);
848 	MPASS(nm_txq->dbidx == nm_txq->pidx);
849 
850 	/*
851 	 * Second part: reclaim buffers for completed transmissions.
852 	 */
853 	if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
854 		reclaimed += reclaim_nm_tx_desc(nm_txq);
855 		kring->nr_hwtail += reclaimed;
856 		if (kring->nr_hwtail >= kring->nkr_num_slots)
857 			kring->nr_hwtail -= kring->nkr_num_slots;
858 	}
859 
860 	return (0);
861 }
862 
863 static int
864 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
865 {
866 	struct netmap_adapter *na = kring->na;
867 	struct netmap_ring *ring = kring->ring;
868 	struct ifnet *ifp = na->ifp;
869 	struct vi_info *vi = ifp->if_softc;
870 	struct adapter *sc = vi->pi->adapter;
871 	struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
872 	u_int const head = kring->rhead;
873 	u_int n;
874 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
875 
876 	if (black_hole)
877 		return (0);	/* No updates ever. */
878 
879 	if (netmap_no_pendintr || force_update) {
880 		kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
881 		kring->nr_kflags &= ~NKR_PENDINTR;
882 	}
883 
884 	if (nm_rxq->fl_db_saved > 0 && starve_fl == 0) {
885 		wmb();
886 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
887 		    nm_rxq->fl_db_val | V_PIDX(nm_rxq->fl_db_saved));
888 		nm_rxq->fl_db_saved = 0;
889 	}
890 
891 	/* Userspace done with buffers from kring->nr_hwcur to head */
892 	n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
893 	    kring->nkr_num_slots - kring->nr_hwcur + head;
894 	n &= ~7U;
895 	if (n > 0) {
896 		u_int fl_pidx = nm_rxq->fl_pidx;
897 		struct netmap_slot *slot = &ring->slot[fl_pidx];
898 		uint64_t ba;
899 		int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
900 
901 		/*
902 		 * We always deal with 8 buffers at a time.  We must have
903 		 * stopped at an 8B boundary (fl_pidx) last time around and we
904 		 * must have a multiple of 8B buffers to give to the freelist.
905 		 */
906 		MPASS((fl_pidx & 7) == 0);
907 		MPASS((n & 7) == 0);
908 
909 		IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
910 		IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx);
911 
912 		while (n > 0) {
913 			for (i = 0; i < 8; i++, fl_pidx++, slot++) {
914 				PNMB(na, slot, &ba);
915 				MPASS(ba != 0);
916 				nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
917 				slot->flags &= ~NS_BUF_CHANGED;
918 				MPASS(fl_pidx <= nm_rxq->fl_sidx);
919 			}
920 			n -= 8;
921 			if (fl_pidx == nm_rxq->fl_sidx) {
922 				fl_pidx = 0;
923 				slot = &ring->slot[0];
924 			}
925 			if (++dbinc == 8 && n >= 32) {
926 				wmb();
927 				if (starve_fl)
928 					nm_rxq->fl_db_saved += dbinc;
929 				else {
930 					t4_write_reg(sc, sc->sge_kdoorbell_reg,
931 					    nm_rxq->fl_db_val | V_PIDX(dbinc));
932 				}
933 				dbinc = 0;
934 			}
935 		}
936 		MPASS(nm_rxq->fl_pidx == fl_pidx);
937 
938 		if (dbinc > 0) {
939 			wmb();
940 			if (starve_fl)
941 				nm_rxq->fl_db_saved += dbinc;
942 			else {
943 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
944 				    nm_rxq->fl_db_val | V_PIDX(dbinc));
945 			}
946 		}
947 	}
948 
949 	return (0);
950 }
951 
952 void
953 cxgbe_nm_attach(struct vi_info *vi)
954 {
955 	struct port_info *pi;
956 	struct adapter *sc;
957 	struct netmap_adapter na;
958 
959 	MPASS(vi->nnmrxq > 0);
960 	MPASS(vi->ifp != NULL);
961 
962 	pi = vi->pi;
963 	sc = pi->adapter;
964 
965 	bzero(&na, sizeof(na));
966 
967 	na.ifp = vi->ifp;
968 	na.na_flags = NAF_BDG_MAYSLEEP;
969 
970 	/* Netmap doesn't know about the space reserved for the status page. */
971 	na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
972 
973 	/*
974 	 * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
975 	 * num_rx_desc is based on the number of buffers that can be held in the
976 	 * freelist, and not the number of entries in the iq.  (These two are
977 	 * not exactly the same due to the space taken up by the status page).
978 	 */
979 	na.num_rx_desc = rounddown(vi->qsize_rxq, 8);
980 	na.nm_txsync = cxgbe_netmap_txsync;
981 	na.nm_rxsync = cxgbe_netmap_rxsync;
982 	na.nm_register = cxgbe_netmap_reg;
983 	na.num_tx_rings = vi->nnmtxq;
984 	na.num_rx_rings = vi->nnmrxq;
985 	netmap_attach(&na);	/* This adds IFCAP_NETMAP to if_capabilities */
986 }
987 
988 void
989 cxgbe_nm_detach(struct vi_info *vi)
990 {
991 
992 	MPASS(vi->nnmrxq > 0);
993 	MPASS(vi->ifp != NULL);
994 
995 	netmap_detach(vi->ifp);
996 }
997 
998 static inline const void *
999 unwrap_nm_fw6_msg(const struct cpl_fw6_msg *cpl)
1000 {
1001 
1002 	MPASS(cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL);
1003 
1004 	/* data[0] is RSS header */
1005 	return (&cpl->data[1]);
1006 }
1007 
1008 static void
1009 handle_nm_sge_egr_update(struct adapter *sc, struct ifnet *ifp,
1010     const struct cpl_sge_egr_update *egr)
1011 {
1012 	uint32_t oq;
1013 	struct sge_nm_txq *nm_txq;
1014 
1015 	oq = be32toh(egr->opcode_qid);
1016 	MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
1017 	nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
1018 
1019 	netmap_tx_irq(ifp, nm_txq->nid);
1020 }
1021 
1022 void
1023 service_nm_rxq(struct sge_nm_rxq *nm_rxq)
1024 {
1025 	struct vi_info *vi = nm_rxq->vi;
1026 	struct adapter *sc = vi->pi->adapter;
1027 	struct ifnet *ifp = vi->ifp;
1028 	struct netmap_adapter *na = NA(ifp);
1029 	struct netmap_kring *kring = na->rx_rings[nm_rxq->nid];
1030 	struct netmap_ring *ring = kring->ring;
1031 	struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
1032 	const void *cpl;
1033 	uint32_t lq;
1034 	u_int work = 0;
1035 	uint8_t opcode;
1036 	uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
1037 	u_int fl_credits = fl_cidx & 7;
1038 	u_int ndesc = 0;	/* desc processed since last cidx update */
1039 	u_int nframes = 0;	/* frames processed since last netmap wakeup */
1040 
1041 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
1042 
1043 		rmb();
1044 
1045 		lq = be32toh(d->rsp.pldbuflen_qid);
1046 		opcode = d->rss.opcode;
1047 		cpl = &d->cpl[0];
1048 
1049 		switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
1050 		case X_RSPD_TYPE_FLBUF:
1051 
1052 			/* fall through */
1053 
1054 		case X_RSPD_TYPE_CPL:
1055 			MPASS(opcode < NUM_CPL_CMDS);
1056 
1057 			switch (opcode) {
1058 			case CPL_FW4_MSG:
1059 			case CPL_FW6_MSG:
1060 				cpl = unwrap_nm_fw6_msg(cpl);
1061 				/* fall through */
1062 			case CPL_SGE_EGR_UPDATE:
1063 				handle_nm_sge_egr_update(sc, ifp, cpl);
1064 				break;
1065 			case CPL_RX_PKT:
1066 				ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
1067 				    sc->params.sge.fl_pktshift;
1068 				ring->slot[fl_cidx].flags = 0;
1069 				nframes++;
1070 				if (!(lq & F_RSPD_NEWBUF)) {
1071 					MPASS(black_hole == 2);
1072 					break;
1073 				}
1074 				fl_credits++;
1075 				if (__predict_false(++fl_cidx == nm_rxq->fl_sidx))
1076 					fl_cidx = 0;
1077 				break;
1078 			default:
1079 				panic("%s: unexpected opcode 0x%x on nm_rxq %p",
1080 				    __func__, opcode, nm_rxq);
1081 			}
1082 			break;
1083 
1084 		case X_RSPD_TYPE_INTR:
1085 			/* Not equipped to handle forwarded interrupts. */
1086 			panic("%s: netmap queue received interrupt for iq %u\n",
1087 			    __func__, lq);
1088 
1089 		default:
1090 			panic("%s: illegal response type %d on nm_rxq %p",
1091 			    __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
1092 		}
1093 
1094 		d++;
1095 		if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
1096 			nm_rxq->iq_cidx = 0;
1097 			d = &nm_rxq->iq_desc[0];
1098 			nm_rxq->iq_gen ^= F_RSPD_GEN;
1099 		}
1100 
1101 		if (__predict_false(++nframes == rx_nframes) && !black_hole) {
1102 			atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1103 			netmap_rx_irq(ifp, nm_rxq->nid, &work);
1104 			nframes = 0;
1105 		}
1106 
1107 		if (__predict_false(++ndesc == rx_ndesc)) {
1108 			if (black_hole && fl_credits >= 8) {
1109 				fl_credits /= 8;
1110 				IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
1111 				    nm_rxq->fl_sidx);
1112 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
1113 				    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1114 				fl_credits = fl_cidx & 7;
1115 			}
1116 			t4_write_reg(sc, sc->sge_gts_reg,
1117 			    V_CIDXINC(ndesc) |
1118 			    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
1119 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1120 			ndesc = 0;
1121 		}
1122 	}
1123 
1124 	atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1125 	if (black_hole) {
1126 		fl_credits /= 8;
1127 		IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1128 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1129 		    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1130 	} else if (nframes > 0)
1131 		netmap_rx_irq(ifp, nm_rxq->nid, &work);
1132 
1133     	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndesc) |
1134 	    V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1135 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1136 }
1137 #endif
1138