xref: /freebsd/sys/dev/cxgbe/t4_netmap.c (revision 3fc36ee018bb836bd1796067cf4ef8683f166ebc)
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 SYSCTL_NODE(_hw, OID_AUTO, cxgbe, CTLFLAG_RD, 0, "cxgbe netmap parameters");
61 
62 /*
63  * 0 = normal netmap rx
64  * 1 = black hole
65  * 2 = supermassive black hole (buffer packing enabled)
66  */
67 int black_hole = 0;
68 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_black_hole, CTLFLAG_RDTUN, &black_hole, 0,
69     "Sink incoming packets.");
70 
71 int rx_ndesc = 256;
72 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_rx_ndesc, CTLFLAG_RWTUN,
73     &rx_ndesc, 0, "# of rx descriptors after which the hw cidx is updated.");
74 
75 int holdoff_tmr_idx = 2;
76 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_holdoff_tmr_idx, CTLFLAG_RWTUN,
77     &holdoff_tmr_idx, 0, "Holdoff timer index for netmap rx queues.");
78 
79 /*
80  * Congestion drops.
81  * -1: no congestion feedback (not recommended).
82  *  0: backpressure the channel instead of dropping packets right away.
83  *  1: no backpressure, drop packets for the congested queue immediately.
84  */
85 static int nm_cong_drop = 1;
86 TUNABLE_INT("hw.cxgbe.nm_cong_drop", &nm_cong_drop);
87 
88 static int
89 alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int cong)
90 {
91 	int rc, cntxt_id, i;
92 	__be32 v;
93 	struct adapter *sc = vi->pi->adapter;
94 	struct sge_params *sp = &sc->params.sge;
95 	struct netmap_adapter *na = NA(vi->ifp);
96 	struct fw_iq_cmd c;
97 
98 	MPASS(na != NULL);
99 	MPASS(nm_rxq->iq_desc != NULL);
100 	MPASS(nm_rxq->fl_desc != NULL);
101 
102 	bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE);
103 	bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + sp->spg_len);
104 
105 	bzero(&c, sizeof(c));
106 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
107 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
108 	    V_FW_IQ_CMD_VFN(0));
109 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
110 	    FW_LEN16(c));
111 	if (vi->flags & INTR_RXQ) {
112 		KASSERT(nm_rxq->intr_idx < sc->intr_count,
113 		    ("%s: invalid direct intr_idx %d", __func__,
114 		    nm_rxq->intr_idx));
115 		v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
116 	} else {
117 		CXGBE_UNIMPLEMENTED(__func__);	/* XXXNM: needs review */
118 		v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx) |
119 		    F_FW_IQ_CMD_IQANDST;
120 	}
121 	c.type_to_iqandstindex = htobe32(v |
122 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
123 	    V_FW_IQ_CMD_VIID(vi->viid) |
124 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
125 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(vi->pi->tx_chan) |
126 	    F_FW_IQ_CMD_IQGTSMODE |
127 	    V_FW_IQ_CMD_IQINTCNTTHRESH(0) |
128 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
129 	c.iqsize = htobe16(vi->qsize_rxq);
130 	c.iqaddr = htobe64(nm_rxq->iq_ba);
131 	if (cong >= 0) {
132 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN |
133 		    V_FW_IQ_CMD_FL0CNGCHMAP(cong) | F_FW_IQ_CMD_FL0CONGCIF |
134 		    F_FW_IQ_CMD_FL0CONGEN);
135 	}
136 	c.iqns_to_fl0congen |=
137 	    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
138 		F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
139 		(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
140 		(black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0));
141 	c.fl0dcaen_to_fl0cidxfthresh =
142 	    htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_128B) |
143 		V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
144 	c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE);
145 	c.fl0addr = htobe64(nm_rxq->fl_ba);
146 
147 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
148 	if (rc != 0) {
149 		device_printf(sc->dev,
150 		    "failed to create netmap ingress queue: %d\n", rc);
151 		return (rc);
152 	}
153 
154 	nm_rxq->iq_cidx = 0;
155 	MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE);
156 	nm_rxq->iq_gen = F_RSPD_GEN;
157 	nm_rxq->iq_cntxt_id = be16toh(c.iqid);
158 	nm_rxq->iq_abs_id = be16toh(c.physiqid);
159 	cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
160 	if (cntxt_id >= sc->sge.niq) {
161 		panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
162 		    __func__, cntxt_id, sc->sge.niq - 1);
163 	}
164 	sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
165 
166 	nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
167 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
168 	MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
169 	cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
170 	if (cntxt_id >= sc->sge.neq) {
171 		panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
172 		    __func__, cntxt_id, sc->sge.neq - 1);
173 	}
174 	sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
175 
176 	nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) |
177 	    sc->chip_params->sge_fl_db;
178 
179 	if (is_t5(sc) && cong >= 0) {
180 		uint32_t param, val;
181 
182 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
183 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
184 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
185 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
186 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
187 		    V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
188 		if (cong == 0)
189 			val = 1 << 19;
190 		else {
191 			val = 2 << 19;
192 			for (i = 0; i < 4; i++) {
193 				if (cong & (1 << i))
194 					val |= 1 << (i << 2);
195 			}
196 		}
197 
198 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
199 		if (rc != 0) {
200 			/* report error but carry on */
201 			device_printf(sc->dev,
202 			    "failed to set congestion manager context for "
203 			    "ingress queue %d: %d\n", nm_rxq->iq_cntxt_id, rc);
204 		}
205 	}
206 
207 	t4_write_reg(sc, sc->sge_gts_reg,
208 	    V_INGRESSQID(nm_rxq->iq_cntxt_id) |
209 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
210 
211 	return (rc);
212 }
213 
214 static int
215 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
216 {
217 	struct adapter *sc = vi->pi->adapter;
218 	int rc;
219 
220 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
221 	    nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
222 	if (rc != 0)
223 		device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
224 		    __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
225 	return (rc);
226 }
227 
228 static int
229 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
230 {
231 	int rc, cntxt_id;
232 	size_t len;
233 	struct adapter *sc = vi->pi->adapter;
234 	struct netmap_adapter *na = NA(vi->ifp);
235 	struct fw_eq_eth_cmd c;
236 
237 	MPASS(na != NULL);
238 	MPASS(nm_txq->desc != NULL);
239 
240 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
241 	bzero(nm_txq->desc, len);
242 
243 	bzero(&c, sizeof(c));
244 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
245 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
246 	    V_FW_EQ_ETH_CMD_VFN(0));
247 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
248 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
249 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
250 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
251 	c.fetchszm_to_iqid =
252 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
253 		V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
254 		V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
255 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
256 		      V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
257 		      V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
258 	c.eqaddr = htobe64(nm_txq->ba);
259 
260 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
261 	if (rc != 0) {
262 		device_printf(vi->dev,
263 		    "failed to create netmap egress queue: %d\n", rc);
264 		return (rc);
265 	}
266 
267 	nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
268 	cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
269 	if (cntxt_id >= sc->sge.neq)
270 	    panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
271 		cntxt_id, sc->sge.neq - 1);
272 	sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
273 
274 	nm_txq->pidx = nm_txq->cidx = 0;
275 	MPASS(nm_txq->sidx == na->num_tx_desc);
276 	nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
277 
278 	nm_txq->doorbells = sc->doorbells;
279 	if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
280 	    isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
281 	    isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
282 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
283 		uint32_t mask = (1 << s_qpp) - 1;
284 		volatile uint8_t *udb;
285 
286 		udb = sc->udbs_base + UDBS_DB_OFFSET;
287 		udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
288 		nm_txq->udb_qid = nm_txq->cntxt_id & mask;
289 		if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
290 	    		clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
291 		else {
292 			udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
293 			nm_txq->udb_qid = 0;
294 		}
295 		nm_txq->udb = (volatile void *)udb;
296 	}
297 
298 	return (rc);
299 }
300 
301 static int
302 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
303 {
304 	struct adapter *sc = vi->pi->adapter;
305 	int rc;
306 
307 	rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
308 	if (rc != 0)
309 		device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
310 		    nm_txq->cntxt_id, rc);
311 	return (rc);
312 }
313 
314 static int
315 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
316     struct netmap_adapter *na)
317 {
318 	struct netmap_slot *slot;
319 	struct sge_nm_rxq *nm_rxq;
320 	struct sge_nm_txq *nm_txq;
321 	int rc, i, j, hwidx;
322 	struct hw_buf_info *hwb;
323 
324 	ASSERT_SYNCHRONIZED_OP(sc);
325 
326 	if ((vi->flags & VI_INIT_DONE) == 0 ||
327 	    (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
328 		return (EAGAIN);
329 
330 	hwb = &sc->sge.hw_buf_info[0];
331 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
332 		if (hwb->size == NETMAP_BUF_SIZE(na))
333 			break;
334 	}
335 	if (i >= SGE_FLBUF_SIZES) {
336 		if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
337 		    NETMAP_BUF_SIZE(na));
338 		return (ENXIO);
339 	}
340 	hwidx = i;
341 
342 	/* Must set caps before calling netmap_reset */
343 	nm_set_native_flags(na);
344 
345 	for_each_nm_rxq(vi, i, nm_rxq) {
346 		struct irq *irq = &sc->irq[vi->first_intr + i];
347 
348 		alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop));
349 		nm_rxq->fl_hwidx = hwidx;
350 		slot = netmap_reset(na, NR_RX, i, 0);
351 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
352 
353 		/* We deal with 8 bufs at a time */
354 		MPASS((na->num_rx_desc & 7) == 0);
355 		MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
356 		for (j = 0; j < nm_rxq->fl_sidx; j++) {
357 			uint64_t ba;
358 
359 			PNMB(na, &slot[j], &ba);
360 			MPASS(ba != 0);
361 			nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
362 		}
363 		j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
364 		MPASS((j & 7) == 0);
365 		j /= 8;	/* driver pidx to hardware pidx */
366 		wmb();
367 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
368 		    nm_rxq->fl_db_val | V_PIDX(j));
369 
370 		atomic_cmpset_int(&irq->nm_state, NM_OFF, NM_ON);
371 	}
372 
373 	for_each_nm_txq(vi, i, nm_txq) {
374 		alloc_nm_txq_hwq(vi, nm_txq);
375 		slot = netmap_reset(na, NR_TX, i, 0);
376 		MPASS(slot != NULL);	/* XXXNM: error check, not assert */
377 	}
378 
379 	if (vi->nm_rss == NULL) {
380 		vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
381 		    M_ZERO | M_WAITOK);
382 	}
383 	for (i = 0; i < vi->rss_size;) {
384 		for_each_nm_rxq(vi, j, nm_rxq) {
385 			vi->nm_rss[i++] = nm_rxq->iq_abs_id;
386 			if (i == vi->rss_size)
387 				break;
388 		}
389 	}
390 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
391 	    vi->nm_rss, vi->rss_size);
392 	if (rc != 0)
393 		if_printf(ifp, "netmap rss_config failed: %d\n", rc);
394 
395 	return (rc);
396 }
397 
398 static int
399 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
400     struct netmap_adapter *na)
401 {
402 	int rc, i;
403 	struct sge_nm_txq *nm_txq;
404 	struct sge_nm_rxq *nm_rxq;
405 
406 	ASSERT_SYNCHRONIZED_OP(sc);
407 
408 	if ((vi->flags & VI_INIT_DONE) == 0)
409 		return (0);
410 
411 	rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
412 	    vi->rss, vi->rss_size);
413 	if (rc != 0)
414 		if_printf(ifp, "failed to restore RSS config: %d\n", rc);
415 	nm_clear_native_flags(na);
416 
417 	for_each_nm_txq(vi, i, nm_txq) {
418 		struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
419 
420 		/* Wait for hw pidx to catch up ... */
421 		while (be16toh(nm_txq->pidx) != spg->pidx)
422 			pause("nmpidx", 1);
423 
424 		/* ... and then for the cidx. */
425 		while (spg->pidx != spg->cidx)
426 			pause("nmcidx", 1);
427 
428 		free_nm_txq_hwq(vi, nm_txq);
429 	}
430 	for_each_nm_rxq(vi, i, nm_rxq) {
431 		struct irq *irq = &sc->irq[vi->first_intr + i];
432 
433 		while (!atomic_cmpset_int(&irq->nm_state, NM_ON, NM_OFF))
434 			pause("nmst", 1);
435 
436 		free_nm_rxq_hwq(vi, nm_rxq);
437 	}
438 
439 	return (rc);
440 }
441 
442 static int
443 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
444 {
445 	struct ifnet *ifp = na->ifp;
446 	struct vi_info *vi = ifp->if_softc;
447 	struct adapter *sc = vi->pi->adapter;
448 	int rc;
449 
450 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
451 	if (rc != 0)
452 		return (rc);
453 	if (on)
454 		rc = cxgbe_netmap_on(sc, vi, ifp, na);
455 	else
456 		rc = cxgbe_netmap_off(sc, vi, ifp, na);
457 	end_synchronized_op(sc, 0);
458 
459 	return (rc);
460 }
461 
462 /* How many packets can a single type1 WR carry in n descriptors */
463 static inline int
464 ndesc_to_npkt(const int n)
465 {
466 
467 	MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
468 
469 	return (n * 2 - 1);
470 }
471 #define MAX_NPKT_IN_TYPE1_WR	(ndesc_to_npkt(SGE_MAX_WR_NDESC))
472 
473 /* Space (in descriptors) needed for a type1 WR that carries n packets */
474 static inline int
475 npkt_to_ndesc(const int n)
476 {
477 
478 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
479 
480 	return ((n + 2) / 2);
481 }
482 
483 /* Space (in 16B units) needed for a type1 WR that carries n packets */
484 static inline int
485 npkt_to_len16(const int n)
486 {
487 
488 	MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
489 
490 	return (n * 2 + 1);
491 }
492 
493 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
494 
495 static void
496 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
497 {
498 	int n;
499 	u_int db = nm_txq->doorbells;
500 
501 	MPASS(nm_txq->pidx != nm_txq->dbidx);
502 
503 	n = NMIDXDIFF(nm_txq, dbidx);
504 	if (n > 1)
505 		clrbit(&db, DOORBELL_WCWR);
506 	wmb();
507 
508 	switch (ffs(db) - 1) {
509 	case DOORBELL_UDB:
510 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
511 		break;
512 
513 	case DOORBELL_WCWR: {
514 		volatile uint64_t *dst, *src;
515 
516 		/*
517 		 * Queues whose 128B doorbell segment fits in the page do not
518 		 * use relative qid (udb_qid is always 0).  Only queues with
519 		 * doorbell segments can do WCWR.
520 		 */
521 		KASSERT(nm_txq->udb_qid == 0 && n == 1,
522 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
523 		    __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
524 
525 		dst = (volatile void *)((uintptr_t)nm_txq->udb +
526 		    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
527 		src = (void *)&nm_txq->desc[nm_txq->dbidx];
528 		while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
529 			*dst++ = *src++;
530 		wmb();
531 		break;
532 	}
533 
534 	case DOORBELL_UDBWC:
535 		*nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
536 		wmb();
537 		break;
538 
539 	case DOORBELL_KDB:
540 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
541 		    V_QID(nm_txq->cntxt_id) | V_PIDX(n));
542 		break;
543 	}
544 	nm_txq->dbidx = nm_txq->pidx;
545 }
546 
547 int lazy_tx_credit_flush = 1;
548 
549 /*
550  * Write work requests to send 'npkt' frames and ring the doorbell to send them
551  * on their way.  No need to check for wraparound.
552  */
553 static void
554 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
555     struct netmap_kring *kring, int npkt, int npkt_remaining, int txcsum)
556 {
557 	struct netmap_ring *ring = kring->ring;
558 	struct netmap_slot *slot;
559 	const u_int lim = kring->nkr_num_slots - 1;
560 	struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
561 	uint16_t len;
562 	uint64_t ba;
563 	struct cpl_tx_pkt_core *cpl;
564 	struct ulptx_sgl *usgl;
565 	int i, n;
566 
567 	while (npkt) {
568 		n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
569 		len = 0;
570 
571 		wr = (void *)&nm_txq->desc[nm_txq->pidx];
572 		wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
573 		wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
574 		wr->npkt = n;
575 		wr->r3 = 0;
576 		wr->type = 1;
577 		cpl = (void *)(wr + 1);
578 
579 		for (i = 0; i < n; i++) {
580 			slot = &ring->slot[kring->nr_hwcur];
581 			PNMB(kring->na, slot, &ba);
582 			MPASS(ba != 0);
583 
584 			cpl->ctrl0 = nm_txq->cpl_ctrl0;
585 			cpl->pack = 0;
586 			cpl->len = htobe16(slot->len);
587 			/*
588 			 * netmap(4) says "netmap does not use features such as
589 			 * checksum offloading, TCP segmentation offloading,
590 			 * encryption, VLAN encapsulation/decapsulation, etc."
591 			 *
592 			 * So the ncxl interfaces have tx hardware checksumming
593 			 * disabled by default.  But you can override netmap by
594 			 * enabling IFCAP_TXCSUM on the interface manully.
595 			 */
596 			cpl->ctrl1 = txcsum ? 0 :
597 			    htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
598 
599 			usgl = (void *)(cpl + 1);
600 			usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
601 			    V_ULPTX_NSGE(1));
602 			usgl->len0 = htobe32(slot->len);
603 			usgl->addr0 = htobe64(ba);
604 
605 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
606 			cpl = (void *)(usgl + 1);
607 			MPASS(slot->len + len <= UINT16_MAX);
608 			len += slot->len;
609 			kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
610 		}
611 		wr->plen = htobe16(len);
612 
613 		npkt -= n;
614 		nm_txq->pidx += npkt_to_ndesc(n);
615 		MPASS(nm_txq->pidx <= nm_txq->sidx);
616 		if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
617 			/*
618 			 * This routine doesn't know how to write WRs that wrap
619 			 * around.  Make sure it wasn't asked to.
620 			 */
621 			MPASS(npkt == 0);
622 			nm_txq->pidx = 0;
623 		}
624 
625 		if (npkt == 0 && npkt_remaining == 0) {
626 			/* All done. */
627 			if (lazy_tx_credit_flush == 0) {
628 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
629 				    F_FW_WR_EQUIQ);
630 				nm_txq->equeqidx = nm_txq->pidx;
631 				nm_txq->equiqidx = nm_txq->pidx;
632 			}
633 			ring_nm_txq_db(sc, nm_txq);
634 			return;
635 		}
636 
637 		if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
638 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
639 			    F_FW_WR_EQUIQ);
640 			nm_txq->equeqidx = nm_txq->pidx;
641 			nm_txq->equiqidx = nm_txq->pidx;
642 		} else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
643 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
644 			nm_txq->equeqidx = nm_txq->pidx;
645 		}
646 		if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
647 			ring_nm_txq_db(sc, nm_txq);
648 	}
649 
650 	/* Will get called again. */
651 	MPASS(npkt_remaining);
652 }
653 
654 /* How many contiguous free descriptors starting at pidx */
655 static inline int
656 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
657 {
658 
659 	if (nm_txq->cidx > nm_txq->pidx)
660 		return (nm_txq->cidx - nm_txq->pidx - 1);
661 	else if (nm_txq->cidx > 0)
662 		return (nm_txq->sidx - nm_txq->pidx);
663 	else
664 		return (nm_txq->sidx - nm_txq->pidx - 1);
665 }
666 
667 static int
668 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
669 {
670 	struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
671 	uint16_t hw_cidx = spg->cidx;	/* snapshot */
672 	struct fw_eth_tx_pkts_wr *wr;
673 	int n = 0;
674 
675 	hw_cidx = be16toh(hw_cidx);
676 
677 	while (nm_txq->cidx != hw_cidx) {
678 		wr = (void *)&nm_txq->desc[nm_txq->cidx];
679 
680 		MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)));
681 		MPASS(wr->type == 1);
682 		MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
683 
684 		n += wr->npkt;
685 		nm_txq->cidx += npkt_to_ndesc(wr->npkt);
686 
687 		/*
688 		 * We never sent a WR that wrapped around so the credits coming
689 		 * back, WR by WR, should never cause the cidx to wrap around
690 		 * either.
691 		 */
692 		MPASS(nm_txq->cidx <= nm_txq->sidx);
693 		if (__predict_false(nm_txq->cidx == nm_txq->sidx))
694 			nm_txq->cidx = 0;
695 	}
696 
697 	return (n);
698 }
699 
700 static int
701 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
702 {
703 	struct netmap_adapter *na = kring->na;
704 	struct ifnet *ifp = na->ifp;
705 	struct vi_info *vi = ifp->if_softc;
706 	struct adapter *sc = vi->pi->adapter;
707 	struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
708 	const u_int head = kring->rhead;
709 	u_int reclaimed = 0;
710 	int n, d, npkt_remaining, ndesc_remaining, txcsum;
711 
712 	/*
713 	 * Tx was at kring->nr_hwcur last time around and now we need to advance
714 	 * to kring->rhead.  Note that the driver's pidx moves independent of
715 	 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
716 	 * between descriptors and frames isn't 1:1).
717 	 */
718 
719 	npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
720 	    kring->nkr_num_slots - kring->nr_hwcur + head;
721 	txcsum = ifp->if_capenable & (IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6);
722 	while (npkt_remaining) {
723 		reclaimed += reclaim_nm_tx_desc(nm_txq);
724 		ndesc_remaining = contiguous_ndesc_available(nm_txq);
725 		/* Can't run out of descriptors with packets still remaining */
726 		MPASS(ndesc_remaining > 0);
727 
728 		/* # of desc needed to tx all remaining packets */
729 		d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
730 		if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
731 			d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
732 
733 		if (d <= ndesc_remaining)
734 			n = npkt_remaining;
735 		else {
736 			/* Can't send all, calculate how many can be sent */
737 			n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
738 			    MAX_NPKT_IN_TYPE1_WR;
739 			if (ndesc_remaining % SGE_MAX_WR_NDESC)
740 				n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
741 		}
742 
743 		/* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
744 		npkt_remaining -= n;
745 		cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining, txcsum);
746 	}
747 	MPASS(npkt_remaining == 0);
748 	MPASS(kring->nr_hwcur == head);
749 	MPASS(nm_txq->dbidx == nm_txq->pidx);
750 
751 	/*
752 	 * Second part: reclaim buffers for completed transmissions.
753 	 */
754 	if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
755 		reclaimed += reclaim_nm_tx_desc(nm_txq);
756 		kring->nr_hwtail += reclaimed;
757 		if (kring->nr_hwtail >= kring->nkr_num_slots)
758 			kring->nr_hwtail -= kring->nkr_num_slots;
759 	}
760 
761 	return (0);
762 }
763 
764 static int
765 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
766 {
767 	struct netmap_adapter *na = kring->na;
768 	struct netmap_ring *ring = kring->ring;
769 	struct ifnet *ifp = na->ifp;
770 	struct vi_info *vi = ifp->if_softc;
771 	struct adapter *sc = vi->pi->adapter;
772 	struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
773 	u_int const head = kring->rhead;
774 	u_int n;
775 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
776 
777 	if (black_hole)
778 		return (0);	/* No updates ever. */
779 
780 	if (netmap_no_pendintr || force_update) {
781 		kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
782 		kring->nr_kflags &= ~NKR_PENDINTR;
783 	}
784 
785 	/* Userspace done with buffers from kring->nr_hwcur to head */
786 	n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
787 	    kring->nkr_num_slots - kring->nr_hwcur + head;
788 	n &= ~7U;
789 	if (n > 0) {
790 		u_int fl_pidx = nm_rxq->fl_pidx;
791 		struct netmap_slot *slot = &ring->slot[fl_pidx];
792 		uint64_t ba;
793 		int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
794 
795 		/*
796 		 * We always deal with 8 buffers at a time.  We must have
797 		 * stopped at an 8B boundary (fl_pidx) last time around and we
798 		 * must have a multiple of 8B buffers to give to the freelist.
799 		 */
800 		MPASS((fl_pidx & 7) == 0);
801 		MPASS((n & 7) == 0);
802 
803 		IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
804 		IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx);
805 
806 		while (n > 0) {
807 			for (i = 0; i < 8; i++, fl_pidx++, slot++) {
808 				PNMB(na, slot, &ba);
809 				MPASS(ba != 0);
810 				nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
811 				slot->flags &= ~NS_BUF_CHANGED;
812 				MPASS(fl_pidx <= nm_rxq->fl_sidx);
813 			}
814 			n -= 8;
815 			if (fl_pidx == nm_rxq->fl_sidx) {
816 				fl_pidx = 0;
817 				slot = &ring->slot[0];
818 			}
819 			if (++dbinc == 8 && n >= 32) {
820 				wmb();
821 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
822 				    nm_rxq->fl_db_val | V_PIDX(dbinc));
823 				dbinc = 0;
824 			}
825 		}
826 		MPASS(nm_rxq->fl_pidx == fl_pidx);
827 
828 		if (dbinc > 0) {
829 			wmb();
830 			t4_write_reg(sc, sc->sge_kdoorbell_reg,
831 			    nm_rxq->fl_db_val | V_PIDX(dbinc));
832 		}
833 	}
834 
835 	return (0);
836 }
837 
838 void
839 cxgbe_nm_attach(struct vi_info *vi)
840 {
841 	struct port_info *pi;
842 	struct adapter *sc;
843 	struct netmap_adapter na;
844 
845 	MPASS(vi->nnmrxq > 0);
846 	MPASS(vi->ifp != NULL);
847 
848 	pi = vi->pi;
849 	sc = pi->adapter;
850 
851 	bzero(&na, sizeof(na));
852 
853 	na.ifp = vi->ifp;
854 	na.na_flags = NAF_BDG_MAYSLEEP;
855 
856 	/* Netmap doesn't know about the space reserved for the status page. */
857 	na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
858 
859 	/*
860 	 * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
861 	 * num_rx_desc is based on the number of buffers that can be held in the
862 	 * freelist, and not the number of entries in the iq.  (These two are
863 	 * not exactly the same due to the space taken up by the status page).
864 	 */
865 	na.num_rx_desc = rounddown(vi->qsize_rxq, 8);
866 	na.nm_txsync = cxgbe_netmap_txsync;
867 	na.nm_rxsync = cxgbe_netmap_rxsync;
868 	na.nm_register = cxgbe_netmap_reg;
869 	na.num_tx_rings = vi->nnmtxq;
870 	na.num_rx_rings = vi->nnmrxq;
871 	netmap_attach(&na);	/* This adds IFCAP_NETMAP to if_capabilities */
872 }
873 
874 void
875 cxgbe_nm_detach(struct vi_info *vi)
876 {
877 
878 	MPASS(vi->nnmrxq > 0);
879 	MPASS(vi->ifp != NULL);
880 
881 	netmap_detach(vi->ifp);
882 }
883 
884 static void
885 handle_nm_fw6_msg(struct adapter *sc, struct ifnet *ifp,
886     const struct cpl_fw6_msg *cpl)
887 {
888 	const struct cpl_sge_egr_update *egr;
889 	uint32_t oq;
890 	struct sge_nm_txq *nm_txq;
891 
892 	if (cpl->type != FW_TYPE_RSSCPL && cpl->type != FW6_TYPE_RSSCPL)
893 		panic("%s: FW_TYPE 0x%x on nm_rxq.", __func__, cpl->type);
894 
895 	/* data[0] is RSS header */
896 	egr = (const void *)&cpl->data[1];
897 	oq = be32toh(egr->opcode_qid);
898 	MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
899 	nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
900 
901 	netmap_tx_irq(ifp, nm_txq->nid);
902 }
903 
904 void
905 t4_nm_intr(void *arg)
906 {
907 	struct sge_nm_rxq *nm_rxq = arg;
908 	struct vi_info *vi = nm_rxq->vi;
909 	struct adapter *sc = vi->pi->adapter;
910 	struct ifnet *ifp = vi->ifp;
911 	struct netmap_adapter *na = NA(ifp);
912 	struct netmap_kring *kring = &na->rx_rings[nm_rxq->nid];
913 	struct netmap_ring *ring = kring->ring;
914 	struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
915 	uint32_t lq;
916 	u_int n = 0, work = 0;
917 	uint8_t opcode;
918 	uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
919 	u_int fl_credits = fl_cidx & 7;
920 
921 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
922 
923 		rmb();
924 
925 		lq = be32toh(d->rsp.pldbuflen_qid);
926 		opcode = d->rss.opcode;
927 
928 		switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
929 		case X_RSPD_TYPE_FLBUF:
930 			if (black_hole != 2) {
931 				/* No buffer packing so new buf every time */
932 				MPASS(lq & F_RSPD_NEWBUF);
933 			}
934 
935 			/* fall through */
936 
937 		case X_RSPD_TYPE_CPL:
938 			MPASS(opcode < NUM_CPL_CMDS);
939 
940 			switch (opcode) {
941 			case CPL_FW4_MSG:
942 			case CPL_FW6_MSG:
943 				handle_nm_fw6_msg(sc, ifp,
944 				    (const void *)&d->cpl[0]);
945 				break;
946 			case CPL_RX_PKT:
947 				ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
948 				    sc->params.sge.fl_pktshift;
949 				ring->slot[fl_cidx].flags = kring->nkr_slot_flags;
950 				fl_cidx += (lq & F_RSPD_NEWBUF) ? 1 : 0;
951 				fl_credits += (lq & F_RSPD_NEWBUF) ? 1 : 0;
952 				if (__predict_false(fl_cidx == nm_rxq->fl_sidx))
953 					fl_cidx = 0;
954 				break;
955 			default:
956 				panic("%s: unexpected opcode 0x%x on nm_rxq %p",
957 				    __func__, opcode, nm_rxq);
958 			}
959 			break;
960 
961 		case X_RSPD_TYPE_INTR:
962 			/* Not equipped to handle forwarded interrupts. */
963 			panic("%s: netmap queue received interrupt for iq %u\n",
964 			    __func__, lq);
965 
966 		default:
967 			panic("%s: illegal response type %d on nm_rxq %p",
968 			    __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
969 		}
970 
971 		d++;
972 		if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
973 			nm_rxq->iq_cidx = 0;
974 			d = &nm_rxq->iq_desc[0];
975 			nm_rxq->iq_gen ^= F_RSPD_GEN;
976 		}
977 
978 		if (__predict_false(++n == rx_ndesc)) {
979 			atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
980 			if (black_hole && fl_credits >= 8) {
981 				fl_credits /= 8;
982 				IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
983 				    nm_rxq->fl_sidx);
984 				t4_write_reg(sc, sc->sge_kdoorbell_reg,
985 				    nm_rxq->fl_db_val | V_PIDX(fl_credits));
986 				fl_credits = fl_cidx & 7;
987 			} else if (!black_hole) {
988 				netmap_rx_irq(ifp, nm_rxq->nid, &work);
989 				MPASS(work != 0);
990 			}
991 			t4_write_reg(sc, sc->sge_gts_reg,
992 			    V_CIDXINC(n) | V_INGRESSQID(nm_rxq->iq_cntxt_id) |
993 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
994 			n = 0;
995 		}
996 	}
997 
998 	atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
999 	if (black_hole) {
1000 		fl_credits /= 8;
1001 		IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1002 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
1003 		    nm_rxq->fl_db_val | V_PIDX(fl_credits));
1004 	} else
1005 		netmap_rx_irq(ifp, nm_rxq->nid, &work);
1006 
1007 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(n) |
1008 	    V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1009 	    V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1010 }
1011 #endif
1012