xref: /freebsd/sys/dev/mana/mana_en.c (revision d5b0e70f7e04d971691517ce1304d86a1e367e2e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2021 Microsoft Corp.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  *
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
23  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
24  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/bus.h>
36 #include <sys/kernel.h>
37 #include <sys/kthread.h>
38 #include <sys/malloc.h>
39 #include <sys/mbuf.h>
40 #include <sys/smp.h>
41 #include <sys/socket.h>
42 #include <sys/sockio.h>
43 #include <sys/time.h>
44 #include <sys/eventhandler.h>
45 
46 #include <machine/bus.h>
47 #include <machine/resource.h>
48 #include <machine/in_cksum.h>
49 
50 #include <net/if.h>
51 #include <net/if_var.h>
52 #include <net/if_types.h>
53 #include <net/if_vlan_var.h>
54 #ifdef RSS
55 #include <net/rss_config.h>
56 #endif
57 
58 #include <netinet/in_systm.h>
59 #include <netinet/in.h>
60 #include <netinet/if_ether.h>
61 #include <netinet/ip.h>
62 #include <netinet/ip6.h>
63 #include <netinet/tcp.h>
64 #include <netinet/udp.h>
65 
66 #include "mana.h"
67 #include "mana_sysctl.h"
68 
69 static int mana_up(struct mana_port_context *apc);
70 static int mana_down(struct mana_port_context *apc);
71 
72 static void
73 mana_rss_key_fill(void *k, size_t size)
74 {
75 	static bool rss_key_generated = false;
76 	static uint8_t rss_key[MANA_HASH_KEY_SIZE];
77 
78 	KASSERT(size <= MANA_HASH_KEY_SIZE,
79 	    ("Request more buytes than MANA RSS key can hold"));
80 
81 	if (!rss_key_generated) {
82 		arc4random_buf(rss_key, MANA_HASH_KEY_SIZE);
83 		rss_key_generated = true;
84 	}
85 	memcpy(k, rss_key, size);
86 }
87 
88 static int
89 mana_ifmedia_change(struct ifnet *ifp __unused)
90 {
91 	return EOPNOTSUPP;
92 }
93 
94 static void
95 mana_ifmedia_status(struct ifnet *ifp, struct ifmediareq *ifmr)
96 {
97 	struct mana_port_context *apc = if_getsoftc(ifp);
98 
99 	if (!apc) {
100 		if_printf(ifp, "Port not available\n");
101 		return;
102 	}
103 
104 	MANA_APC_LOCK_LOCK(apc);
105 
106 	ifmr->ifm_status = IFM_AVALID;
107 	ifmr->ifm_active = IFM_ETHER;
108 
109 	if (!apc->port_is_up) {
110 		MANA_APC_LOCK_UNLOCK(apc);
111 		mana_dbg(NULL, "Port %u link is down\n", apc->port_idx);
112 		return;
113 	}
114 
115 	ifmr->ifm_status |= IFM_ACTIVE;
116 	ifmr->ifm_active |= IFM_100G_DR | IFM_FDX;
117 
118 	MANA_APC_LOCK_UNLOCK(apc);
119 }
120 
121 static uint64_t
122 mana_get_counter(struct ifnet *ifp, ift_counter cnt)
123 {
124 	struct mana_port_context *apc = if_getsoftc(ifp);
125 	struct mana_port_stats *stats = &apc->port_stats;
126 
127 	switch (cnt) {
128 	case IFCOUNTER_IPACKETS:
129 		return (counter_u64_fetch(stats->rx_packets));
130 	case IFCOUNTER_OPACKETS:
131 		return (counter_u64_fetch(stats->tx_packets));
132 	case IFCOUNTER_IBYTES:
133 		return (counter_u64_fetch(stats->rx_bytes));
134 	case IFCOUNTER_OBYTES:
135 		return (counter_u64_fetch(stats->tx_bytes));
136 	case IFCOUNTER_IQDROPS:
137 		return (counter_u64_fetch(stats->rx_drops));
138 	case IFCOUNTER_OQDROPS:
139 		return (counter_u64_fetch(stats->tx_drops));
140 	default:
141 		return (if_get_counter_default(ifp, cnt));
142 	}
143 }
144 
145 static void
146 mana_qflush(struct ifnet *ifp)
147 {
148 	if_qflush(ifp);
149 }
150 
151 int
152 mana_restart(struct mana_port_context *apc)
153 {
154 	int rc = 0;
155 
156 	MANA_APC_LOCK_LOCK(apc);
157 	if (apc->port_is_up)
158 		 mana_down(apc);
159 
160 	rc = mana_up(apc);
161 	MANA_APC_LOCK_UNLOCK(apc);
162 
163 	return (rc);
164 }
165 
166 static int
167 mana_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
168 {
169 	struct mana_port_context *apc = if_getsoftc(ifp);
170 	struct ifrsskey *ifrk;
171 	struct ifrsshash *ifrh;
172 	struct ifreq *ifr;
173 	uint16_t new_mtu;
174 	int rc = 0;
175 
176 	switch (command) {
177 	case SIOCSIFMTU:
178 		ifr = (struct ifreq *)data;
179 		new_mtu = ifr->ifr_mtu;
180 		if (ifp->if_mtu == new_mtu)
181 			break;
182 		if ((new_mtu + 18 > MAX_FRAME_SIZE) ||
183 		    (new_mtu + 18 < MIN_FRAME_SIZE)) {
184 			if_printf(ifp, "Invalid MTU. new_mtu: %d, "
185 			    "max allowed: %d, min allowed: %d\n",
186 			    new_mtu, MAX_FRAME_SIZE - 18, MIN_FRAME_SIZE - 18);
187 			return EINVAL;
188 		}
189 		MANA_APC_LOCK_LOCK(apc);
190 		if (apc->port_is_up)
191 			mana_down(apc);
192 
193 		apc->frame_size = new_mtu + 18;
194 		if_setmtu(ifp, new_mtu);
195 		mana_dbg(NULL, "Set MTU to %d\n", new_mtu);
196 
197 		rc = mana_up(apc);
198 		MANA_APC_LOCK_UNLOCK(apc);
199 		break;
200 
201 	case SIOCSIFFLAGS:
202 		if (ifp->if_flags & IFF_UP) {
203 			if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
204 				MANA_APC_LOCK_LOCK(apc);
205 				if (!apc->port_is_up)
206 					rc = mana_up(apc);
207 				MANA_APC_LOCK_UNLOCK(apc);
208 			}
209 		} else {
210 			if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
211 				MANA_APC_LOCK_LOCK(apc);
212 				if (apc->port_is_up)
213 					mana_down(apc);
214 				MANA_APC_LOCK_UNLOCK(apc);
215 			}
216 		}
217 		break;
218 
219 	case SIOCSIFMEDIA:
220 	case SIOCGIFMEDIA:
221 	case SIOCGIFXMEDIA:
222 		ifr = (struct ifreq *)data;
223 		rc = ifmedia_ioctl(ifp, ifr, &apc->media, command);
224 		break;
225 
226 	case SIOCGIFRSSKEY:
227 		ifrk = (struct ifrsskey *)data;
228 		ifrk->ifrk_func = RSS_FUNC_TOEPLITZ;
229 		ifrk->ifrk_keylen = MANA_HASH_KEY_SIZE;
230 		memcpy(ifrk->ifrk_key, apc->hashkey, MANA_HASH_KEY_SIZE);
231 		break;
232 
233 	case SIOCGIFRSSHASH:
234 		ifrh = (struct ifrsshash *)data;
235 		ifrh->ifrh_func = RSS_FUNC_TOEPLITZ;
236 		ifrh->ifrh_types =
237 		    RSS_TYPE_TCP_IPV4 |
238 		    RSS_TYPE_UDP_IPV4 |
239 		    RSS_TYPE_TCP_IPV6 |
240 		    RSS_TYPE_UDP_IPV6;
241 		break;
242 
243 	default:
244 		rc = ether_ioctl(ifp, command, data);
245 		break;
246 	}
247 
248 	return (rc);
249 }
250 
251 static inline void
252 mana_alloc_counters(counter_u64_t *begin, int size)
253 {
254 	counter_u64_t *end = (counter_u64_t *)((char *)begin + size);
255 
256 	for (; begin < end; ++begin)
257 		*begin = counter_u64_alloc(M_WAITOK);
258 }
259 
260 static inline void
261 mana_free_counters(counter_u64_t *begin, int size)
262 {
263 	counter_u64_t *end = (counter_u64_t *)((char *)begin + size);
264 
265 	for (; begin < end; ++begin)
266 		counter_u64_free(*begin);
267 }
268 
269 static bool
270 mana_can_tx(struct gdma_queue *wq)
271 {
272 	return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE;
273 }
274 
275 static inline int
276 mana_tx_map_mbuf(struct mana_port_context *apc,
277     struct mana_send_buf_info *tx_info,
278     struct mbuf **m_head, struct mana_tx_package *tp,
279     struct mana_stats *tx_stats)
280 {
281 	struct gdma_dev *gd = apc->ac->gdma_dev;
282 	bus_dma_segment_t segs[MAX_MBUF_FRAGS];
283 	struct mbuf *m = *m_head;
284 	int err, nsegs, i;
285 
286 	err = bus_dmamap_load_mbuf_sg(apc->tx_buf_tag, tx_info->dma_map,
287 	    m, segs, &nsegs, BUS_DMA_NOWAIT);
288 	if (err == EFBIG) {
289 		struct mbuf *m_new;
290 
291 		counter_u64_add(tx_stats->collapse, 1);
292 		m_new = m_collapse(m, M_NOWAIT, MAX_MBUF_FRAGS);
293 		if (unlikely(m_new == NULL)) {
294 			counter_u64_add(tx_stats->collapse_err, 1);
295 			return ENOBUFS;
296 		} else {
297 			*m_head = m = m_new;
298 		}
299 
300 		mana_warn(NULL,
301 		    "Too many segs in orig mbuf, m_collapse called\n");
302 
303 		err = bus_dmamap_load_mbuf_sg(apc->tx_buf_tag,
304 		    tx_info->dma_map, m, segs, &nsegs, BUS_DMA_NOWAIT);
305 	}
306 	if (!err) {
307 		for (i = 0; i < nsegs; i++) {
308 			tp->wqe_req.sgl[i].address = segs[i].ds_addr;
309 			tp->wqe_req.sgl[i].mem_key = gd->gpa_mkey;
310 			tp->wqe_req.sgl[i].size = segs[i].ds_len;
311 		}
312 		tp->wqe_req.num_sge = nsegs;
313 
314 		tx_info->mbuf = *m_head;
315 
316 		bus_dmamap_sync(apc->tx_buf_tag, tx_info->dma_map,
317 		    BUS_DMASYNC_PREWRITE);
318 	}
319 
320 	return err;
321 }
322 
323 static inline void
324 mana_tx_unmap_mbuf(struct mana_port_context *apc,
325     struct mana_send_buf_info *tx_info)
326 {
327 	bus_dmamap_sync(apc->tx_buf_tag, tx_info->dma_map,
328 	    BUS_DMASYNC_POSTWRITE);
329 	bus_dmamap_unload(apc->tx_buf_tag, tx_info->dma_map);
330 	if (tx_info->mbuf) {
331 		m_freem(tx_info->mbuf);
332 		tx_info->mbuf = NULL;
333 	}
334 }
335 
336 static inline int
337 mana_load_rx_mbuf(struct mana_port_context *apc, struct mana_rxq *rxq,
338     struct mana_recv_buf_oob *rx_oob, bool alloc_mbuf)
339 {
340 	bus_dma_segment_t segs[1];
341 	struct mbuf *mbuf;
342 	int nsegs, err;
343 	uint32_t mlen;
344 
345 	if (alloc_mbuf) {
346 		mbuf = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, rxq->datasize);
347 		if (unlikely(mbuf == NULL)) {
348 			mbuf = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
349 			if (unlikely(mbuf == NULL)) {
350 				return ENOMEM;
351 			}
352 			mlen = MCLBYTES;
353 		} else {
354 			mlen = rxq->datasize;
355 		}
356 
357 		mbuf->m_pkthdr.len = mbuf->m_len = mlen;
358 	} else {
359 		if (rx_oob->mbuf) {
360 			mbuf = rx_oob->mbuf;
361 			mlen = rx_oob->mbuf->m_pkthdr.len;
362 		} else {
363 			return ENOMEM;
364 		}
365 	}
366 
367 	err = bus_dmamap_load_mbuf_sg(apc->rx_buf_tag, rx_oob->dma_map,
368 	    mbuf, segs, &nsegs, BUS_DMA_NOWAIT);
369 
370 	if (unlikely((err != 0) || (nsegs != 1))) {
371 		mana_warn(NULL, "Failed to map mbuf, error: %d, "
372 		    "nsegs: %d\n", err, nsegs);
373 		counter_u64_add(rxq->stats.dma_mapping_err, 1);
374 		goto error;
375 	}
376 
377 	bus_dmamap_sync(apc->rx_buf_tag, rx_oob->dma_map,
378 	    BUS_DMASYNC_PREREAD);
379 
380 	rx_oob->mbuf = mbuf;
381 	rx_oob->num_sge = 1;
382 	rx_oob->sgl[0].address = segs[0].ds_addr;
383 	rx_oob->sgl[0].size = mlen;
384 	rx_oob->sgl[0].mem_key = apc->ac->gdma_dev->gpa_mkey;
385 
386 	return 0;
387 
388 error:
389 	m_freem(mbuf);
390 	return EFAULT;
391 }
392 
393 static inline void
394 mana_unload_rx_mbuf(struct mana_port_context *apc, struct mana_rxq *rxq,
395     struct mana_recv_buf_oob *rx_oob, bool free_mbuf)
396 {
397 	bus_dmamap_sync(apc->rx_buf_tag, rx_oob->dma_map,
398 	    BUS_DMASYNC_POSTREAD);
399 	bus_dmamap_unload(apc->rx_buf_tag, rx_oob->dma_map);
400 
401 	if (free_mbuf && rx_oob->mbuf) {
402 		m_freem(rx_oob->mbuf);
403 		rx_oob->mbuf = NULL;
404 	}
405 }
406 
407 
408 /* Use couple mbuf PH_loc spaces for l3 and l4 protocal type */
409 #define MANA_L3_PROTO(_mbuf)	((_mbuf)->m_pkthdr.PH_loc.sixteen[0])
410 #define MANA_L4_PROTO(_mbuf)	((_mbuf)->m_pkthdr.PH_loc.sixteen[1])
411 
412 #define MANA_TXQ_FULL	(IFF_DRV_RUNNING | IFF_DRV_OACTIVE)
413 
414 static void
415 mana_xmit(struct mana_txq *txq)
416 {
417 	enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT;
418 	struct mana_send_buf_info *tx_info;
419 	struct ifnet *ndev = txq->ndev;
420 	struct mbuf *mbuf;
421 	struct mana_port_context *apc = if_getsoftc(ndev);
422 	struct mana_port_stats *port_stats = &apc->port_stats;
423 	struct gdma_dev *gd = apc->ac->gdma_dev;
424 	uint64_t packets, bytes;
425 	uint16_t next_to_use;
426 	struct mana_tx_package pkg = {};
427 	struct mana_stats *tx_stats;
428 	struct gdma_queue *gdma_sq;
429 	struct mana_cq *cq;
430 	int err, len;
431 
432 	gdma_sq = txq->gdma_sq;
433 	cq = &apc->tx_qp[txq->idx].tx_cq;
434 	tx_stats = &txq->stats;
435 
436 	packets = 0;
437 	bytes = 0;
438 	next_to_use = txq->next_to_use;
439 
440 	while ((mbuf = drbr_peek(ndev, txq->txq_br)) != NULL) {
441 		if (!apc->port_is_up ||
442 		    (if_getdrvflags(ndev) & MANA_TXQ_FULL) != IFF_DRV_RUNNING) {
443 			drbr_putback(ndev, txq->txq_br, mbuf);
444 			break;
445 		}
446 
447 		if (!mana_can_tx(gdma_sq)) {
448 			/* SQ is full. Set the IFF_DRV_OACTIVE flag */
449 			if_setdrvflagbits(apc->ndev, IFF_DRV_OACTIVE, 0);
450 			counter_u64_add(tx_stats->stop, 1);
451 			uint64_t stops = counter_u64_fetch(tx_stats->stop);
452 			uint64_t wakeups = counter_u64_fetch(tx_stats->wakeup);
453 #define MANA_TXQ_STOP_THRESHOLD		50
454 			if (stops > MANA_TXQ_STOP_THRESHOLD && wakeups > 0 &&
455 			    stops > wakeups && txq->alt_txq_idx == txq->idx) {
456 				txq->alt_txq_idx =
457 				    (txq->idx + (stops / wakeups))
458 				    % apc->num_queues;
459 				counter_u64_add(tx_stats->alt_chg, 1);
460 			}
461 
462 			drbr_putback(ndev, txq->txq_br, mbuf);
463 
464 			taskqueue_enqueue(cq->cleanup_tq, &cq->cleanup_task);
465 			break;
466 		}
467 
468 		tx_info = &txq->tx_buf_info[next_to_use];
469 
470 		memset(&pkg, 0, sizeof(struct mana_tx_package));
471 		pkg.wqe_req.sgl = pkg.sgl_array;
472 
473 		err = mana_tx_map_mbuf(apc, tx_info, &mbuf, &pkg, tx_stats);
474 		if (unlikely(err)) {
475 			mana_dbg(NULL,
476 			    "Failed to map tx mbuf, err %d\n", err);
477 
478 			counter_u64_add(tx_stats->dma_mapping_err, 1);
479 
480 			/* The mbuf is still there. Free it */
481 			m_freem(mbuf);
482 			/* Advance the drbr queue */
483 			drbr_advance(ndev, txq->txq_br);
484 			continue;
485 		}
486 
487 		pkg.tx_oob.s_oob.vcq_num = cq->gdma_id;
488 		pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame;
489 
490 		if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) {
491 			pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset;
492 			pkt_fmt = MANA_LONG_PKT_FMT;
493 		} else {
494 			pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset;
495 		}
496 
497 		pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt;
498 
499 		if (pkt_fmt == MANA_SHORT_PKT_FMT)
500 			pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob);
501 		else
502 			pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob);
503 
504 		pkg.wqe_req.inline_oob_data = &pkg.tx_oob;
505 		pkg.wqe_req.flags = 0;
506 		pkg.wqe_req.client_data_unit = 0;
507 
508 		if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) {
509 			if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IP)
510 				pkg.tx_oob.s_oob.is_outer_ipv4 = 1;
511 			else
512 				pkg.tx_oob.s_oob.is_outer_ipv6 = 1;
513 
514 			pkg.tx_oob.s_oob.comp_iphdr_csum = 1;
515 			pkg.tx_oob.s_oob.comp_tcp_csum = 1;
516 			pkg.tx_oob.s_oob.trans_off = mbuf->m_pkthdr.l3hlen;
517 
518 			pkg.wqe_req.client_data_unit = mbuf->m_pkthdr.tso_segsz;
519 			pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0;
520 		} else if (mbuf->m_pkthdr.csum_flags &
521 		    (CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP6_UDP | CSUM_IP6_TCP)) {
522 			if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IP) {
523 				pkg.tx_oob.s_oob.is_outer_ipv4 = 1;
524 				pkg.tx_oob.s_oob.comp_iphdr_csum = 1;
525 			} else {
526 				pkg.tx_oob.s_oob.is_outer_ipv6 = 1;
527 			}
528 
529 			if (MANA_L4_PROTO(mbuf) == IPPROTO_TCP) {
530 				pkg.tx_oob.s_oob.comp_tcp_csum = 1;
531 				pkg.tx_oob.s_oob.trans_off =
532 				    mbuf->m_pkthdr.l3hlen;
533 			} else {
534 				pkg.tx_oob.s_oob.comp_udp_csum = 1;
535 			}
536 		} else if (mbuf->m_pkthdr.csum_flags & CSUM_IP) {
537 			pkg.tx_oob.s_oob.is_outer_ipv4 = 1;
538 			pkg.tx_oob.s_oob.comp_iphdr_csum = 1;
539 		} else {
540 			if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IP)
541 				pkg.tx_oob.s_oob.is_outer_ipv4 = 1;
542 			else if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IPV6)
543 				pkg.tx_oob.s_oob.is_outer_ipv6 = 1;
544 		}
545 
546 		len = mbuf->m_pkthdr.len;
547 
548 		err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req,
549 		    (struct gdma_posted_wqe_info *)&tx_info->wqe_inf);
550 		if (unlikely(err)) {
551 			/* Should not happen */
552 			if_printf(ndev, "Failed to post TX OOB: %d\n", err);
553 
554 			mana_tx_unmap_mbuf(apc, tx_info);
555 
556 			drbr_advance(ndev, txq->txq_br);
557 			continue;
558 		}
559 
560 		next_to_use =
561 		    (next_to_use + 1) % MAX_SEND_BUFFERS_PER_QUEUE;
562 
563 		(void)atomic_inc_return(&txq->pending_sends);
564 
565 		drbr_advance(ndev, txq->txq_br);
566 
567 		mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq);
568 
569 		packets++;
570 		bytes += len;
571 	}
572 
573 	counter_enter();
574 	counter_u64_add_protected(tx_stats->packets, packets);
575 	counter_u64_add_protected(port_stats->tx_packets, packets);
576 	counter_u64_add_protected(tx_stats->bytes, bytes);
577 	counter_u64_add_protected(port_stats->tx_bytes, bytes);
578 	counter_exit();
579 
580 	txq->next_to_use = next_to_use;
581 }
582 
583 static void
584 mana_xmit_taskfunc(void *arg, int pending)
585 {
586 	struct mana_txq *txq = (struct mana_txq *)arg;
587 	struct ifnet *ndev = txq->ndev;
588 	struct mana_port_context *apc = if_getsoftc(ndev);
589 
590 	while (!drbr_empty(ndev, txq->txq_br) && apc->port_is_up &&
591 	    (if_getdrvflags(ndev) & MANA_TXQ_FULL) == IFF_DRV_RUNNING) {
592 		mtx_lock(&txq->txq_mtx);
593 		mana_xmit(txq);
594 		mtx_unlock(&txq->txq_mtx);
595 	}
596 }
597 
598 #define PULLUP_HDR(m, len)				\
599 do {							\
600 	if (unlikely((m)->m_len < (len))) {		\
601 		(m) = m_pullup((m), (len));		\
602 		if ((m) == NULL)			\
603 			return (NULL);			\
604 	}						\
605 } while (0)
606 
607 /*
608  * If this function failed, the mbuf would be freed.
609  */
610 static inline struct mbuf *
611 mana_tso_fixup(struct mbuf *mbuf)
612 {
613 	struct ether_vlan_header *eh = mtod(mbuf, struct ether_vlan_header *);
614 	struct tcphdr *th;
615 	uint16_t etype;
616 	int ehlen;
617 
618 	if (eh->evl_encap_proto == ntohs(ETHERTYPE_VLAN)) {
619 		etype = ntohs(eh->evl_proto);
620 		ehlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
621 	} else {
622 		etype = ntohs(eh->evl_encap_proto);
623 		ehlen = ETHER_HDR_LEN;
624 	}
625 
626 	if (etype == ETHERTYPE_IP) {
627 		struct ip *ip;
628 		int iphlen;
629 
630 		PULLUP_HDR(mbuf, ehlen + sizeof(*ip));
631 		ip = mtodo(mbuf, ehlen);
632 		iphlen = ip->ip_hl << 2;
633 		mbuf->m_pkthdr.l3hlen = ehlen + iphlen;
634 
635 		PULLUP_HDR(mbuf, ehlen + iphlen + sizeof(*th));
636 		th = mtodo(mbuf, ehlen + iphlen);
637 
638 		ip->ip_len = 0;
639 		ip->ip_sum = 0;
640 		th->th_sum = in_pseudo(ip->ip_src.s_addr,
641 		    ip->ip_dst.s_addr, htons(IPPROTO_TCP));
642 	} else if (etype == ETHERTYPE_IPV6) {
643 		struct ip6_hdr *ip6;
644 
645 		PULLUP_HDR(mbuf, ehlen + sizeof(*ip6) + sizeof(*th));
646 		ip6 = mtodo(mbuf, ehlen);
647 		if (ip6->ip6_nxt != IPPROTO_TCP) {
648 			/* Realy something wrong, just return */
649 			mana_dbg(NULL, "TSO mbuf not TCP, freed.\n");
650 			m_freem(mbuf);
651 			return NULL;
652 		}
653 		mbuf->m_pkthdr.l3hlen = ehlen + sizeof(*ip6);
654 
655 		th = mtodo(mbuf, ehlen + sizeof(*ip6));
656 
657 		ip6->ip6_plen = 0;
658 		th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
659 	} else {
660 		/* CSUM_TSO is set but not IP protocol. */
661 		mana_warn(NULL, "TSO mbuf not right, freed.\n");
662 		m_freem(mbuf);
663 		return NULL;
664 	}
665 
666 	MANA_L3_PROTO(mbuf) = etype;
667 
668 	return (mbuf);
669 }
670 
671 /*
672  * If this function failed, the mbuf would be freed.
673  */
674 static inline struct mbuf *
675 mana_mbuf_csum_check(struct mbuf *mbuf)
676 {
677 	struct ether_vlan_header *eh = mtod(mbuf, struct ether_vlan_header *);
678 	struct mbuf *mbuf_next;
679 	uint16_t etype;
680 	int offset;
681 	int ehlen;
682 
683 	if (eh->evl_encap_proto == ntohs(ETHERTYPE_VLAN)) {
684 		etype = ntohs(eh->evl_proto);
685 		ehlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
686 	} else {
687 		etype = ntohs(eh->evl_encap_proto);
688 		ehlen = ETHER_HDR_LEN;
689 	}
690 
691 	mbuf_next = m_getptr(mbuf, ehlen, &offset);
692 
693 	MANA_L4_PROTO(mbuf) = 0;
694 	if (etype == ETHERTYPE_IP) {
695 		const struct ip *ip;
696 		int iphlen;
697 
698 		ip = (struct ip *)(mtodo(mbuf_next, offset));
699 		iphlen = ip->ip_hl << 2;
700 		mbuf->m_pkthdr.l3hlen = ehlen + iphlen;
701 
702 		MANA_L4_PROTO(mbuf) = ip->ip_p;
703 	} else if (etype == ETHERTYPE_IPV6) {
704 		const struct ip6_hdr *ip6;
705 
706 		ip6 = (struct ip6_hdr *)(mtodo(mbuf_next, offset));
707 		mbuf->m_pkthdr.l3hlen = ehlen + sizeof(*ip6);
708 
709 		MANA_L4_PROTO(mbuf) = ip6->ip6_nxt;
710 	} else {
711 		MANA_L4_PROTO(mbuf) = 0;
712 	}
713 
714 	MANA_L3_PROTO(mbuf) = etype;
715 
716 	return (mbuf);
717 }
718 
719 static int
720 mana_start_xmit(struct ifnet *ifp, struct mbuf *m)
721 {
722 	struct mana_port_context *apc = if_getsoftc(ifp);
723 	struct mana_txq *txq;
724 	int is_drbr_empty;
725 	uint16_t txq_id;
726 	int err;
727 
728 	if (unlikely((!apc->port_is_up) ||
729 	    (if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0))
730 		return ENODEV;
731 
732 	if (m->m_pkthdr.csum_flags & CSUM_TSO) {
733 		m = mana_tso_fixup(m);
734 		if (unlikely(m == NULL)) {
735 			counter_enter();
736 			counter_u64_add_protected(apc->port_stats.tx_drops, 1);
737 			counter_exit();
738 			return EIO;
739 		}
740 	} else {
741 		m = mana_mbuf_csum_check(m);
742 		if (unlikely(m == NULL)) {
743 			counter_enter();
744 			counter_u64_add_protected(apc->port_stats.tx_drops, 1);
745 			counter_exit();
746 			return EIO;
747 		}
748 	}
749 
750 	if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
751 		uint32_t hash = m->m_pkthdr.flowid;
752 		txq_id = apc->indir_table[(hash) & MANA_INDIRECT_TABLE_MASK] %
753 		    apc->num_queues;
754 	} else {
755 		txq_id = m->m_pkthdr.flowid % apc->num_queues;
756 	}
757 
758 	if (apc->enable_tx_altq)
759 		txq_id = apc->tx_qp[txq_id].txq.alt_txq_idx;
760 
761 	txq = &apc->tx_qp[txq_id].txq;
762 
763 	is_drbr_empty = drbr_empty(ifp, txq->txq_br);
764 	err = drbr_enqueue(ifp, txq->txq_br, m);
765 	if (unlikely(err)) {
766 		mana_warn(NULL, "txq %u failed to enqueue: %d\n",
767 		    txq_id, err);
768 		taskqueue_enqueue(txq->enqueue_tq, &txq->enqueue_task);
769 		return err;
770 	}
771 
772 	if (is_drbr_empty && mtx_trylock(&txq->txq_mtx)) {
773 		mana_xmit(txq);
774 		mtx_unlock(&txq->txq_mtx);
775 	} else {
776 		taskqueue_enqueue(txq->enqueue_tq, &txq->enqueue_task);
777 	}
778 
779 	return 0;
780 }
781 
782 static void
783 mana_cleanup_port_context(struct mana_port_context *apc)
784 {
785 	bus_dma_tag_destroy(apc->tx_buf_tag);
786 	bus_dma_tag_destroy(apc->rx_buf_tag);
787 	apc->rx_buf_tag = NULL;
788 
789 	free(apc->rxqs, M_DEVBUF);
790 	apc->rxqs = NULL;
791 
792 	mana_free_counters((counter_u64_t *)&apc->port_stats,
793 	    sizeof(struct mana_port_stats));
794 }
795 
796 static int
797 mana_init_port_context(struct mana_port_context *apc)
798 {
799 	device_t dev = apc->ac->gdma_dev->gdma_context->dev;
800 	uint32_t tso_maxsize;
801 	int err;
802 
803 	tso_maxsize = MAX_MBUF_FRAGS * MANA_TSO_MAXSEG_SZ -
804 	    (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
805 
806 	/* Create DMA tag for tx bufs */
807 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
808 	    1, 0,			/* alignment, boundary	*/
809 	    BUS_SPACE_MAXADDR,		/* lowaddr		*/
810 	    BUS_SPACE_MAXADDR,		/* highaddr		*/
811 	    NULL, NULL,			/* filter, filterarg	*/
812 	    tso_maxsize,		/* maxsize		*/
813 	    MAX_MBUF_FRAGS,		/* nsegments		*/
814 	    tso_maxsize,		/* maxsegsize		*/
815 	    0,				/* flags		*/
816 	    NULL, NULL,			/* lockfunc, lockfuncarg*/
817 	    &apc->tx_buf_tag);
818 	if (unlikely(err)) {
819 		device_printf(dev, "Feiled to create TX DMA tag\n");
820 		return err;
821 	}
822 
823 	/* Create DMA tag for rx bufs */
824 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
825 	    64, 0,			/* alignment, boundary	*/
826 	    BUS_SPACE_MAXADDR,		/* lowaddr		*/
827 	    BUS_SPACE_MAXADDR,		/* highaddr		*/
828 	    NULL, NULL,			/* filter, filterarg	*/
829 	    MJUMPAGESIZE,		/* maxsize		*/
830 	    1,				/* nsegments		*/
831 	    MJUMPAGESIZE,		/* maxsegsize		*/
832 	    0,				/* flags		*/
833 	    NULL, NULL,			/* lockfunc, lockfuncarg*/
834 	    &apc->rx_buf_tag);
835 	if (unlikely(err)) {
836 		device_printf(dev, "Feiled to create RX DMA tag\n");
837 		return err;
838 	}
839 
840 	apc->rxqs = mallocarray(apc->num_queues, sizeof(struct mana_rxq *),
841 	    M_DEVBUF, M_WAITOK | M_ZERO);
842 
843 	if (!apc->rxqs) {
844 		bus_dma_tag_destroy(apc->tx_buf_tag);
845 		bus_dma_tag_destroy(apc->rx_buf_tag);
846 		apc->rx_buf_tag = NULL;
847 		return ENOMEM;
848 	}
849 
850 	return 0;
851 }
852 
853 static int
854 mana_send_request(struct mana_context *ac, void *in_buf,
855     uint32_t in_len, void *out_buf, uint32_t out_len)
856 {
857 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
858 	struct gdma_resp_hdr *resp = out_buf;
859 	struct gdma_req_hdr *req = in_buf;
860 	device_t dev = gc->dev;
861 	static atomic_t activity_id;
862 	int err;
863 
864 	req->dev_id = gc->mana.dev_id;
865 	req->activity_id = atomic_inc_return(&activity_id);
866 
867 	mana_dbg(NULL, "activity_id  = %u\n", activity_id);
868 
869 	err = mana_gd_send_request(gc, in_len, in_buf, out_len,
870 	    out_buf);
871 	if (err || resp->status) {
872 		device_printf(dev, "Failed to send mana message: %d, 0x%x\n",
873 			err, resp->status);
874 		return err ? err : EPROTO;
875 	}
876 
877 	if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 ||
878 	    req->activity_id != resp->activity_id) {
879 		device_printf(dev,
880 		    "Unexpected mana message response: %x,%x,%x,%x\n",
881 		    req->dev_id.as_uint32, resp->dev_id.as_uint32,
882 		    req->activity_id, resp->activity_id);
883 		return EPROTO;
884 	}
885 
886 	return 0;
887 }
888 
889 static int
890 mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr,
891     const enum mana_command_code expected_code,
892     const uint32_t min_size)
893 {
894 	if (resp_hdr->response.msg_type != expected_code)
895 		return EPROTO;
896 
897 	if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1)
898 		return EPROTO;
899 
900 	if (resp_hdr->response.msg_size < min_size)
901 		return EPROTO;
902 
903 	return 0;
904 }
905 
906 static int
907 mana_query_device_cfg(struct mana_context *ac, uint32_t proto_major_ver,
908     uint32_t proto_minor_ver, uint32_t proto_micro_ver,
909     uint16_t *max_num_vports)
910 {
911 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
912 	struct mana_query_device_cfg_resp resp = {};
913 	struct mana_query_device_cfg_req req = {};
914 	device_t dev = gc->dev;
915 	int err = 0;
916 
917 	mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG,
918 	    sizeof(req), sizeof(resp));
919 	req.proto_major_ver = proto_major_ver;
920 	req.proto_minor_ver = proto_minor_ver;
921 	req.proto_micro_ver = proto_micro_ver;
922 
923 	err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp));
924 	if (err) {
925 		device_printf(dev, "Failed to query config: %d", err);
926 		return err;
927 	}
928 
929 	err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG,
930 	    sizeof(resp));
931 	if (err || resp.hdr.status) {
932 		device_printf(dev, "Invalid query result: %d, 0x%x\n", err,
933 		    resp.hdr.status);
934 		if (!err)
935 			err = EPROTO;
936 		return err;
937 	}
938 
939 	*max_num_vports = resp.max_num_vports;
940 
941 	mana_dbg(NULL, "mana max_num_vports from device = %d\n",
942 	    *max_num_vports);
943 
944 	return 0;
945 }
946 
947 static int
948 mana_query_vport_cfg(struct mana_port_context *apc, uint32_t vport_index,
949     uint32_t *max_sq, uint32_t *max_rq, uint32_t *num_indir_entry)
950 {
951 	struct mana_query_vport_cfg_resp resp = {};
952 	struct mana_query_vport_cfg_req req = {};
953 	int err;
954 
955 	mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG,
956 	    sizeof(req), sizeof(resp));
957 
958 	req.vport_index = vport_index;
959 
960 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
961 	    sizeof(resp));
962 	if (err)
963 		return err;
964 
965 	err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG,
966 	    sizeof(resp));
967 	if (err)
968 		return err;
969 
970 	if (resp.hdr.status)
971 		return EPROTO;
972 
973 	*max_sq = resp.max_num_sq;
974 	*max_rq = resp.max_num_rq;
975 	*num_indir_entry = resp.num_indirection_ent;
976 
977 	apc->port_handle = resp.vport;
978 	memcpy(apc->mac_addr, resp.mac_addr, ETHER_ADDR_LEN);
979 
980 	return 0;
981 }
982 
983 void
984 mana_uncfg_vport(struct mana_port_context *apc)
985 {
986 	MANA_APC_LOCK_LOCK(apc);
987 	apc->vport_use_count--;
988 	if (apc->vport_use_count < 0) {
989 		mana_err(NULL,
990 		    "WARNING: vport_use_count less than 0: %u\n",
991 		    apc->vport_use_count);
992 	}
993 	MANA_APC_LOCK_UNLOCK(apc);
994 }
995 
996 int
997 mana_cfg_vport(struct mana_port_context *apc, uint32_t protection_dom_id,
998     uint32_t doorbell_pg_id)
999 {
1000 	struct mana_config_vport_resp resp = {};
1001 	struct mana_config_vport_req req = {};
1002 	int err;
1003 
1004 	/* This function is used to program the Ethernet port in the hardware
1005 	 * table. It can be called from the Ethernet driver or the RDMA driver.
1006 	 *
1007 	 * For Ethernet usage, the hardware supports only one active user on a
1008 	 * physical port. The driver checks on the port usage before programming
1009 	 * the hardware when creating the RAW QP (RDMA driver) or exposing the
1010 	 * device to kernel NET layer (Ethernet driver).
1011 	 *
1012 	 * Because the RDMA driver doesn't know in advance which QP type the
1013 	 * user will create, it exposes the device with all its ports. The user
1014 	 * may not be able to create RAW QP on a port if this port is already
1015 	 * in used by the Ethernet driver from the kernel.
1016 	 *
1017 	 * This physical port limitation only applies to the RAW QP. For RC QP,
1018 	 * the hardware doesn't have this limitation. The user can create RC
1019 	 * QPs on a physical port up to the hardware limits independent of the
1020 	 * Ethernet usage on the same port.
1021 	 */
1022 	MANA_APC_LOCK_LOCK(apc);
1023 	if (apc->vport_use_count > 0) {
1024 		MANA_APC_LOCK_UNLOCK(apc);
1025 		return EBUSY;
1026 	}
1027 	apc->vport_use_count++;
1028 	MANA_APC_LOCK_UNLOCK(apc);
1029 
1030 	mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX,
1031 	    sizeof(req), sizeof(resp));
1032 	req.vport = apc->port_handle;
1033 	req.pdid = protection_dom_id;
1034 	req.doorbell_pageid = doorbell_pg_id;
1035 
1036 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1037 	    sizeof(resp));
1038 	if (err) {
1039 		if_printf(apc->ndev, "Failed to configure vPort: %d\n", err);
1040 		goto out;
1041 	}
1042 
1043 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX,
1044 	    sizeof(resp));
1045 	if (err || resp.hdr.status) {
1046 		if_printf(apc->ndev, "Failed to configure vPort: %d, 0x%x\n",
1047 		    err, resp.hdr.status);
1048 		if (!err)
1049 			err = EPROTO;
1050 
1051 		goto out;
1052 	}
1053 
1054 	apc->tx_shortform_allowed = resp.short_form_allowed;
1055 	apc->tx_vp_offset = resp.tx_vport_offset;
1056 
1057 	if_printf(apc->ndev, "Configured vPort %ju PD %u DB %u\n",
1058 	    apc->port_handle, protection_dom_id, doorbell_pg_id);
1059 
1060 out:
1061 	if (err)
1062 		mana_uncfg_vport(apc);
1063 
1064 	return err;
1065 }
1066 
1067 static int
1068 mana_cfg_vport_steering(struct mana_port_context *apc,
1069     enum TRI_STATE rx,
1070     bool update_default_rxobj, bool update_key,
1071     bool update_tab)
1072 {
1073 	uint16_t num_entries = MANA_INDIRECT_TABLE_SIZE;
1074 	struct mana_cfg_rx_steer_req *req = NULL;
1075 	struct mana_cfg_rx_steer_resp resp = {};
1076 	struct ifnet *ndev = apc->ndev;
1077 	mana_handle_t *req_indir_tab;
1078 	uint32_t req_buf_size;
1079 	int err;
1080 
1081 	req_buf_size = sizeof(*req) + sizeof(mana_handle_t) * num_entries;
1082 	req = malloc(req_buf_size, M_DEVBUF, M_WAITOK | M_ZERO);
1083 	if (!req)
1084 		return ENOMEM;
1085 
1086 	mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
1087 	    sizeof(resp));
1088 
1089 	req->vport = apc->port_handle;
1090 	req->num_indir_entries = num_entries;
1091 	req->indir_tab_offset = sizeof(*req);
1092 	req->rx_enable = rx;
1093 	req->rss_enable = apc->rss_state;
1094 	req->update_default_rxobj = update_default_rxobj;
1095 	req->update_hashkey = update_key;
1096 	req->update_indir_tab = update_tab;
1097 	req->default_rxobj = apc->default_rxobj;
1098 
1099 	if (update_key)
1100 		memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE);
1101 
1102 	if (update_tab) {
1103 		req_indir_tab = (mana_handle_t *)(req + 1);
1104 		memcpy(req_indir_tab, apc->rxobj_table,
1105 		       req->num_indir_entries * sizeof(mana_handle_t));
1106 	}
1107 
1108 	err = mana_send_request(apc->ac, req, req_buf_size, &resp,
1109 	    sizeof(resp));
1110 	if (err) {
1111 		if_printf(ndev, "Failed to configure vPort RX: %d\n", err);
1112 		goto out;
1113 	}
1114 
1115 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX,
1116 	    sizeof(resp));
1117 	if (err) {
1118 		if_printf(ndev, "vPort RX configuration failed: %d\n", err);
1119 		goto out;
1120 	}
1121 
1122 	if (resp.hdr.status) {
1123 		if_printf(ndev, "vPort RX configuration failed: 0x%x\n",
1124 		    resp.hdr.status);
1125 		err = EPROTO;
1126 	}
1127 
1128 	if_printf(ndev, "Configured steering vPort %ju entries %u\n",
1129 	    apc->port_handle, num_entries);
1130 
1131 out:
1132 	free(req, M_DEVBUF);
1133 	return err;
1134 }
1135 
1136 int
1137 mana_create_wq_obj(struct mana_port_context *apc,
1138     mana_handle_t vport,
1139     uint32_t wq_type, struct mana_obj_spec *wq_spec,
1140     struct mana_obj_spec *cq_spec,
1141     mana_handle_t *wq_obj)
1142 {
1143 	struct mana_create_wqobj_resp resp = {};
1144 	struct mana_create_wqobj_req req = {};
1145 	struct ifnet *ndev = apc->ndev;
1146 	int err;
1147 
1148 	mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ,
1149 	    sizeof(req), sizeof(resp));
1150 	req.vport = vport;
1151 	req.wq_type = wq_type;
1152 	req.wq_gdma_region = wq_spec->gdma_region;
1153 	req.cq_gdma_region = cq_spec->gdma_region;
1154 	req.wq_size = wq_spec->queue_size;
1155 	req.cq_size = cq_spec->queue_size;
1156 	req.cq_moderation_ctx_id = cq_spec->modr_ctx_id;
1157 	req.cq_parent_qid = cq_spec->attached_eq;
1158 
1159 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1160 	    sizeof(resp));
1161 	if (err) {
1162 		if_printf(ndev, "Failed to create WQ object: %d\n", err);
1163 		goto out;
1164 	}
1165 
1166 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ,
1167 	    sizeof(resp));
1168 	if (err || resp.hdr.status) {
1169 		if_printf(ndev, "Failed to create WQ object: %d, 0x%x\n", err,
1170 		    resp.hdr.status);
1171 		if (!err)
1172 			err = EPROTO;
1173 		goto out;
1174 	}
1175 
1176 	if (resp.wq_obj == INVALID_MANA_HANDLE) {
1177 		if_printf(ndev, "Got an invalid WQ object handle\n");
1178 		err = EPROTO;
1179 		goto out;
1180 	}
1181 
1182 	*wq_obj = resp.wq_obj;
1183 	wq_spec->queue_index = resp.wq_id;
1184 	cq_spec->queue_index = resp.cq_id;
1185 
1186 	return 0;
1187 out:
1188 	return err;
1189 }
1190 
1191 void
1192 mana_destroy_wq_obj(struct mana_port_context *apc, uint32_t wq_type,
1193     mana_handle_t wq_obj)
1194 {
1195 	struct mana_destroy_wqobj_resp resp = {};
1196 	struct mana_destroy_wqobj_req req = {};
1197 	struct ifnet *ndev = apc->ndev;
1198 	int err;
1199 
1200 	mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ,
1201 	    sizeof(req), sizeof(resp));
1202 	req.wq_type = wq_type;
1203 	req.wq_obj_handle = wq_obj;
1204 
1205 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1206 	    sizeof(resp));
1207 	if (err) {
1208 		if_printf(ndev, "Failed to destroy WQ object: %d\n", err);
1209 		return;
1210 	}
1211 
1212 	err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ,
1213 	    sizeof(resp));
1214 	if (err || resp.hdr.status)
1215 		if_printf(ndev, "Failed to destroy WQ object: %d, 0x%x\n",
1216 		    err, resp.hdr.status);
1217 }
1218 
1219 static void
1220 mana_destroy_eq(struct mana_context *ac)
1221 {
1222 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
1223 	struct gdma_queue *eq;
1224 	int i;
1225 
1226 	if (!ac->eqs)
1227 		return;
1228 
1229 	for (i = 0; i < gc->max_num_queues; i++) {
1230 		eq = ac->eqs[i].eq;
1231 		if (!eq)
1232 			continue;
1233 
1234 		mana_gd_destroy_queue(gc, eq);
1235 	}
1236 
1237 	free(ac->eqs, M_DEVBUF);
1238 	ac->eqs = NULL;
1239 }
1240 
1241 static int
1242 mana_create_eq(struct mana_context *ac)
1243 {
1244 	struct gdma_dev *gd = ac->gdma_dev;
1245 	struct gdma_context *gc = gd->gdma_context;
1246 	struct gdma_queue_spec spec = {};
1247 	int err;
1248 	int i;
1249 
1250 	ac->eqs = mallocarray(gc->max_num_queues, sizeof(struct mana_eq),
1251 	    M_DEVBUF, M_WAITOK | M_ZERO);
1252 	if (!ac->eqs)
1253 		return ENOMEM;
1254 
1255 	spec.type = GDMA_EQ;
1256 	spec.monitor_avl_buf = false;
1257 	spec.queue_size = EQ_SIZE;
1258 	spec.eq.callback = NULL;
1259 	spec.eq.context = ac->eqs;
1260 	spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE;
1261 
1262 	for (i = 0; i < gc->max_num_queues; i++) {
1263 		err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq);
1264 		if (err)
1265 			goto out;
1266 	}
1267 
1268 	return 0;
1269 out:
1270 	mana_destroy_eq(ac);
1271 	return err;
1272 }
1273 
1274 static int
1275 mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq)
1276 {
1277 	struct mana_fence_rq_resp resp = {};
1278 	struct mana_fence_rq_req req = {};
1279 	int err;
1280 
1281 	init_completion(&rxq->fence_event);
1282 
1283 	mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ,
1284 	    sizeof(req), sizeof(resp));
1285 	req.wq_obj_handle = rxq->rxobj;
1286 
1287 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1288 	    sizeof(resp));
1289 	if (err) {
1290 		if_printf(apc->ndev, "Failed to fence RQ %u: %d\n",
1291 		    rxq->rxq_idx, err);
1292 		return err;
1293 	}
1294 
1295 	err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp));
1296 	if (err || resp.hdr.status) {
1297 		if_printf(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n",
1298 		    rxq->rxq_idx, err, resp.hdr.status);
1299 		if (!err)
1300 			err = EPROTO;
1301 
1302 		return err;
1303 	}
1304 
1305 	if (wait_for_completion_timeout(&rxq->fence_event, 10 * hz)) {
1306 		if_printf(apc->ndev, "Failed to fence RQ %u: timed out\n",
1307 		    rxq->rxq_idx);
1308 		return ETIMEDOUT;
1309         }
1310 
1311 	return 0;
1312 }
1313 
1314 static void
1315 mana_fence_rqs(struct mana_port_context *apc)
1316 {
1317 	unsigned int rxq_idx;
1318 	struct mana_rxq *rxq;
1319 	int err;
1320 
1321 	for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
1322 		rxq = apc->rxqs[rxq_idx];
1323 		err = mana_fence_rq(apc, rxq);
1324 
1325 		/* In case of any error, use sleep instead. */
1326 		if (err)
1327 			gdma_msleep(100);
1328 	}
1329 }
1330 
1331 static int
1332 mana_move_wq_tail(struct gdma_queue *wq, uint32_t num_units)
1333 {
1334 	uint32_t used_space_old;
1335 	uint32_t used_space_new;
1336 
1337 	used_space_old = wq->head - wq->tail;
1338 	used_space_new = wq->head - (wq->tail + num_units);
1339 
1340 	if (used_space_new > used_space_old) {
1341 		mana_err(NULL,
1342 		    "WARNING: new used space %u greater than old one %u\n",
1343 		    used_space_new, used_space_old);
1344 		return ERANGE;
1345 	}
1346 
1347 	wq->tail += num_units;
1348 	return 0;
1349 }
1350 
1351 static void
1352 mana_poll_tx_cq(struct mana_cq *cq)
1353 {
1354 	struct gdma_comp *completions = cq->gdma_comp_buf;
1355 	struct gdma_posted_wqe_info *wqe_info;
1356 	struct mana_send_buf_info *tx_info;
1357 	unsigned int pkt_transmitted = 0;
1358 	unsigned int wqe_unit_cnt = 0;
1359 	struct mana_txq *txq = cq->txq;
1360 	struct mana_port_context *apc;
1361 	uint16_t next_to_complete;
1362 	struct ifnet *ndev;
1363 	int comp_read;
1364 	int txq_idx = txq->idx;;
1365 	int i;
1366 	int sa_drop = 0;
1367 
1368 	struct gdma_queue *gdma_wq;
1369 	unsigned int avail_space;
1370 	bool txq_full = false;
1371 
1372 	ndev = txq->ndev;
1373 	apc = if_getsoftc(ndev);
1374 
1375 	comp_read = mana_gd_poll_cq(cq->gdma_cq, completions,
1376 	    CQE_POLLING_BUFFER);
1377 
1378 	if (comp_read < 1)
1379 		return;
1380 
1381 	next_to_complete = txq->next_to_complete;
1382 
1383 	for (i = 0; i < comp_read; i++) {
1384 		struct mana_tx_comp_oob *cqe_oob;
1385 
1386 		if (!completions[i].is_sq) {
1387 			mana_err(NULL, "WARNING: Not for SQ\n");
1388 			return;
1389 		}
1390 
1391 		cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data;
1392 		if (cqe_oob->cqe_hdr.client_type !=
1393 				 MANA_CQE_COMPLETION) {
1394 			mana_err(NULL,
1395 			    "WARNING: Invalid CQE client type %u\n",
1396 			    cqe_oob->cqe_hdr.client_type);
1397 			return;
1398 		}
1399 
1400 		switch (cqe_oob->cqe_hdr.cqe_type) {
1401 		case CQE_TX_OKAY:
1402 			break;
1403 
1404 		case CQE_TX_SA_DROP:
1405 		case CQE_TX_MTU_DROP:
1406 		case CQE_TX_INVALID_OOB:
1407 		case CQE_TX_INVALID_ETH_TYPE:
1408 		case CQE_TX_HDR_PROCESSING_ERROR:
1409 		case CQE_TX_VF_DISABLED:
1410 		case CQE_TX_VPORT_IDX_OUT_OF_RANGE:
1411 		case CQE_TX_VPORT_DISABLED:
1412 		case CQE_TX_VLAN_TAGGING_VIOLATION:
1413 			sa_drop ++;
1414 			mana_err(NULL,
1415 			    "TX: txq %d CQE error %d, ntc = %d, "
1416 			    "pending sends = %d: err ignored.\n",
1417 			    txq_idx, cqe_oob->cqe_hdr.cqe_type,
1418 			    next_to_complete, txq->pending_sends);
1419 			break;
1420 
1421 		default:
1422 			/* If the CQE type is unexpected, log an error,
1423 			 * and go through the error path.
1424 			 */
1425 			mana_err(NULL,
1426 			    "ERROR: TX: Unexpected CQE type %d: HW BUG?\n",
1427 			    cqe_oob->cqe_hdr.cqe_type);
1428 			return;
1429 		}
1430 		if (txq->gdma_txq_id != completions[i].wq_num) {
1431 			mana_dbg(NULL,
1432 			    "txq gdma id not match completion wq num: "
1433 			    "%d != %d\n",
1434 			    txq->gdma_txq_id, completions[i].wq_num);
1435 			break;
1436 		}
1437 
1438 		tx_info = &txq->tx_buf_info[next_to_complete];
1439 		if (!tx_info->mbuf) {
1440 			mana_err(NULL,
1441 			    "WARNING: txq %d Empty mbuf on tx_info: %u, "
1442 			    "ntu = %u, pending_sends = %d, "
1443 			    "transmitted = %d, sa_drop = %d, i = %d, comp_read = %d\n",
1444 			    txq_idx, next_to_complete, txq->next_to_use,
1445 			    txq->pending_sends, pkt_transmitted, sa_drop,
1446 			    i, comp_read);
1447 			break;
1448 		}
1449 
1450 		wqe_info = &tx_info->wqe_inf;
1451 		wqe_unit_cnt += wqe_info->wqe_size_in_bu;
1452 
1453 		mana_tx_unmap_mbuf(apc, tx_info);
1454 		mb();
1455 
1456 		next_to_complete =
1457 		    (next_to_complete + 1) % MAX_SEND_BUFFERS_PER_QUEUE;
1458 
1459 		pkt_transmitted++;
1460 	}
1461 
1462 	txq->next_to_complete = next_to_complete;
1463 
1464 	if (wqe_unit_cnt == 0) {
1465 		mana_err(NULL,
1466 		    "WARNING: TX ring not proceeding!\n");
1467 		return;
1468 	}
1469 
1470 	mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt);
1471 
1472 	/* Ensure tail updated before checking q stop */
1473 	wmb();
1474 
1475 	gdma_wq = txq->gdma_sq;
1476 	avail_space = mana_gd_wq_avail_space(gdma_wq);
1477 
1478 
1479 	if ((if_getdrvflags(ndev) & MANA_TXQ_FULL) == MANA_TXQ_FULL) {
1480 		txq_full = true;
1481 	}
1482 
1483 	/* Ensure checking txq_full before apc->port_is_up. */
1484 	rmb();
1485 
1486 	if (txq_full && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) {
1487 		/* Grab the txq lock and re-test */
1488 		mtx_lock(&txq->txq_mtx);
1489 		avail_space = mana_gd_wq_avail_space(gdma_wq);
1490 
1491 		if ((if_getdrvflags(ndev) & MANA_TXQ_FULL) == MANA_TXQ_FULL &&
1492 		    apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) {
1493 			/* Clear the Q full flag */
1494 			if_setdrvflagbits(apc->ndev, IFF_DRV_RUNNING,
1495 			    IFF_DRV_OACTIVE);
1496 			counter_u64_add(txq->stats.wakeup, 1);
1497 			if (txq->alt_txq_idx != txq->idx) {
1498 				uint64_t stops = counter_u64_fetch(txq->stats.stop);
1499 				uint64_t wakeups = counter_u64_fetch(txq->stats.wakeup);
1500 				/* Reset alt_txq_idx back if it is not overloaded */
1501 				if (stops < wakeups) {
1502 					txq->alt_txq_idx = txq->idx;
1503 					counter_u64_add(txq->stats.alt_reset, 1);
1504 				}
1505 			}
1506 			rmb();
1507 			/* Schedule a tx enqueue task */
1508 			taskqueue_enqueue(txq->enqueue_tq, &txq->enqueue_task);
1509 		}
1510 		mtx_unlock(&txq->txq_mtx);
1511 	}
1512 
1513 	if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0)
1514 		mana_err(NULL,
1515 		    "WARNING: TX %d pending_sends error: %d\n",
1516 		    txq->idx, txq->pending_sends);
1517 
1518 	cq->work_done = pkt_transmitted;
1519 }
1520 
1521 static void
1522 mana_post_pkt_rxq(struct mana_rxq *rxq)
1523 {
1524 	struct mana_recv_buf_oob *recv_buf_oob;
1525 	uint32_t curr_index;
1526 	int err;
1527 
1528 	curr_index = rxq->buf_index++;
1529 	if (rxq->buf_index == rxq->num_rx_buf)
1530 		rxq->buf_index = 0;
1531 
1532 	recv_buf_oob = &rxq->rx_oobs[curr_index];
1533 
1534 	err = mana_gd_post_and_ring(rxq->gdma_rq, &recv_buf_oob->wqe_req,
1535 	    &recv_buf_oob->wqe_inf);
1536 	if (err) {
1537 		mana_err(NULL, "WARNING: rxq %u post pkt err %d\n",
1538 		    rxq->rxq_idx, err);
1539 		return;
1540 	}
1541 
1542 	if (recv_buf_oob->wqe_inf.wqe_size_in_bu != 1) {
1543 		mana_err(NULL, "WARNING: rxq %u wqe_size_in_bu %u\n",
1544 		    rxq->rxq_idx, recv_buf_oob->wqe_inf.wqe_size_in_bu);
1545 	}
1546 }
1547 
1548 static void
1549 mana_rx_mbuf(struct mbuf *mbuf, struct mana_rxcomp_oob *cqe,
1550     struct mana_rxq *rxq)
1551 {
1552 	struct mana_stats *rx_stats = &rxq->stats;
1553 	struct ifnet *ndev = rxq->ndev;
1554 	uint32_t pkt_len = cqe->ppi[0].pkt_len;
1555 	uint16_t rxq_idx = rxq->rxq_idx;
1556 	struct mana_port_context *apc;
1557 	bool do_lro = false;
1558 	bool do_if_input;
1559 
1560 	apc = if_getsoftc(ndev);
1561 	rxq->rx_cq.work_done++;
1562 
1563 	if (!mbuf) {
1564 		return;
1565 	}
1566 
1567 	mbuf->m_flags |= M_PKTHDR;
1568 	mbuf->m_pkthdr.len = pkt_len;
1569 	mbuf->m_len = pkt_len;
1570 	mbuf->m_pkthdr.rcvif = ndev;
1571 
1572 	if ((ndev->if_capenable & IFCAP_RXCSUM ||
1573 	    ndev->if_capenable & IFCAP_RXCSUM_IPV6) &&
1574 	    (cqe->rx_iphdr_csum_succeed)) {
1575 		mbuf->m_pkthdr.csum_flags = CSUM_IP_CHECKED;
1576 		mbuf->m_pkthdr.csum_flags |= CSUM_IP_VALID;
1577 		if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed) {
1578 			mbuf->m_pkthdr.csum_flags |=
1579 			    (CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1580 			mbuf->m_pkthdr.csum_data = 0xffff;
1581 
1582 			if (cqe->rx_tcp_csum_succeed)
1583 				do_lro = true;
1584 		}
1585 	}
1586 
1587 	if (cqe->rx_hashtype != 0) {
1588 		mbuf->m_pkthdr.flowid = cqe->ppi[0].pkt_hash;
1589 
1590 		uint16_t hashtype = cqe->rx_hashtype;
1591 		if (hashtype & NDIS_HASH_IPV4_MASK) {
1592 			hashtype &= NDIS_HASH_IPV4_MASK;
1593 			switch (hashtype) {
1594 			case NDIS_HASH_TCP_IPV4:
1595 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_TCP_IPV4);
1596 				break;
1597 			case NDIS_HASH_UDP_IPV4:
1598 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_UDP_IPV4);
1599 				break;
1600 			default:
1601 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_IPV4);
1602 			}
1603 		} else if (hashtype & NDIS_HASH_IPV6_MASK) {
1604 			hashtype &= NDIS_HASH_IPV6_MASK;
1605 			switch (hashtype) {
1606 			case NDIS_HASH_TCP_IPV6:
1607 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_TCP_IPV6);
1608 				break;
1609 			case NDIS_HASH_TCP_IPV6_EX:
1610 				M_HASHTYPE_SET(mbuf,
1611 				    M_HASHTYPE_RSS_TCP_IPV6_EX);
1612 				break;
1613 			case NDIS_HASH_UDP_IPV6:
1614 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_UDP_IPV6);
1615 				break;
1616 			case NDIS_HASH_UDP_IPV6_EX:
1617 				M_HASHTYPE_SET(mbuf,
1618 				    M_HASHTYPE_RSS_UDP_IPV6_EX);
1619 				break;
1620 			default:
1621 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_IPV6);
1622 			}
1623 		} else {
1624 			M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE_HASH);
1625 		}
1626 	} else {
1627 		mbuf->m_pkthdr.flowid = rxq_idx;
1628 		M_HASHTYPE_SET(mbuf, M_HASHTYPE_NONE);
1629 	}
1630 
1631 	do_if_input = true;
1632 	if ((ndev->if_capenable & IFCAP_LRO) && do_lro) {
1633 		if (rxq->lro.lro_cnt != 0 &&
1634 		    tcp_lro_rx(&rxq->lro, mbuf, 0) == 0)
1635 			do_if_input = false;
1636 	}
1637 	if (do_if_input) {
1638 		ndev->if_input(ndev, mbuf);
1639 	}
1640 
1641 	counter_enter();
1642 	counter_u64_add_protected(rx_stats->packets, 1);
1643 	counter_u64_add_protected(apc->port_stats.rx_packets, 1);
1644 	counter_u64_add_protected(rx_stats->bytes, pkt_len);
1645 	counter_u64_add_protected(apc->port_stats.rx_bytes, pkt_len);
1646 	counter_exit();
1647 }
1648 
1649 static void
1650 mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,
1651     struct gdma_comp *cqe)
1652 {
1653 	struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data;
1654 	struct mana_recv_buf_oob *rxbuf_oob;
1655 	struct ifnet *ndev = rxq->ndev;
1656 	struct mana_port_context *apc;
1657 	struct mbuf *old_mbuf;
1658 	uint32_t curr, pktlen;
1659 	int err;
1660 
1661 	switch (oob->cqe_hdr.cqe_type) {
1662 	case CQE_RX_OKAY:
1663 		break;
1664 
1665 	case CQE_RX_TRUNCATED:
1666 		apc = if_getsoftc(ndev);
1667 		counter_u64_add(apc->port_stats.rx_drops, 1);
1668 		rxbuf_oob = &rxq->rx_oobs[rxq->buf_index];
1669 		if_printf(ndev, "Dropped a truncated packet\n");
1670 		goto drop;
1671 
1672 	case CQE_RX_COALESCED_4:
1673 		if_printf(ndev, "RX coalescing is unsupported\n");
1674 		return;
1675 
1676 	case CQE_RX_OBJECT_FENCE:
1677 		complete(&rxq->fence_event);
1678 		return;
1679 
1680 	default:
1681 		if_printf(ndev, "Unknown RX CQE type = %d\n",
1682 		    oob->cqe_hdr.cqe_type);
1683 		return;
1684 	}
1685 
1686 	if (oob->cqe_hdr.cqe_type != CQE_RX_OKAY)
1687 		return;
1688 
1689 	pktlen = oob->ppi[0].pkt_len;
1690 
1691 	if (pktlen == 0) {
1692 		/* data packets should never have packetlength of zero */
1693 		if_printf(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%jx\n",
1694 		    rxq->gdma_id, cq->gdma_id, rxq->rxobj);
1695 		return;
1696 	}
1697 
1698 	curr = rxq->buf_index;
1699 	rxbuf_oob = &rxq->rx_oobs[curr];
1700 	if (rxbuf_oob->wqe_inf.wqe_size_in_bu != 1) {
1701 		mana_err(NULL, "WARNING: Rx Incorrect complete "
1702 		    "WQE size %u\n",
1703 		    rxbuf_oob->wqe_inf.wqe_size_in_bu);
1704 	}
1705 
1706 	apc = if_getsoftc(ndev);
1707 
1708 	old_mbuf = rxbuf_oob->mbuf;
1709 
1710 	/* Unload DMA map for the old mbuf */
1711 	mana_unload_rx_mbuf(apc, rxq, rxbuf_oob, false);
1712 
1713 	/* Load a new mbuf to replace the old one */
1714 	err = mana_load_rx_mbuf(apc, rxq, rxbuf_oob, true);
1715 	if (err) {
1716 		mana_dbg(NULL,
1717 		    "failed to load rx mbuf, err = %d, packet dropped.\n",
1718 		    err);
1719 		counter_u64_add(rxq->stats.mbuf_alloc_fail, 1);
1720 		/*
1721 		 * Failed to load new mbuf, rxbuf_oob->mbuf is still
1722 		 * pointing to the old one. Drop the packet.
1723 		 */
1724 		 old_mbuf = NULL;
1725 		 /* Reload the existing mbuf */
1726 		 mana_load_rx_mbuf(apc, rxq, rxbuf_oob, false);
1727 	}
1728 
1729 	mana_rx_mbuf(old_mbuf, oob, rxq);
1730 
1731 drop:
1732 	mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu);
1733 
1734 	mana_post_pkt_rxq(rxq);
1735 }
1736 
1737 static void
1738 mana_poll_rx_cq(struct mana_cq *cq)
1739 {
1740 	struct gdma_comp *comp = cq->gdma_comp_buf;
1741 	int comp_read, i;
1742 
1743 	comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER);
1744 	KASSERT(comp_read <= CQE_POLLING_BUFFER,
1745 	    ("comp_read %d great than buf size %d",
1746 	    comp_read, CQE_POLLING_BUFFER));
1747 
1748 	for (i = 0; i < comp_read; i++) {
1749 		if (comp[i].is_sq == true) {
1750 			mana_err(NULL,
1751 			    "WARNING: CQE not for receive queue\n");
1752 			return;
1753 		}
1754 
1755 		/* verify recv cqe references the right rxq */
1756 		if (comp[i].wq_num != cq->rxq->gdma_id) {
1757 			mana_err(NULL,
1758 			    "WARNING: Received CQE %d  not for "
1759 			    "this receive queue %d\n",
1760 			    comp[i].wq_num,  cq->rxq->gdma_id);
1761 			return;
1762 		}
1763 
1764 		mana_process_rx_cqe(cq->rxq, cq, &comp[i]);
1765 	}
1766 
1767 	tcp_lro_flush_all(&cq->rxq->lro);
1768 }
1769 
1770 static void
1771 mana_cq_handler(void *context, struct gdma_queue *gdma_queue)
1772 {
1773 	struct mana_cq *cq = context;
1774 	uint8_t arm_bit;
1775 
1776 	KASSERT(cq->gdma_cq == gdma_queue,
1777 	    ("cq do not match %p, %p", cq->gdma_cq, gdma_queue));
1778 
1779 	if (cq->type == MANA_CQ_TYPE_RX) {
1780 		mana_poll_rx_cq(cq);
1781 	} else {
1782 		mana_poll_tx_cq(cq);
1783 	}
1784 
1785 	if (cq->work_done < cq->budget && cq->do_not_ring_db == false)
1786 		arm_bit = SET_ARM_BIT;
1787 	else
1788 		arm_bit = 0;
1789 
1790 	mana_gd_ring_cq(gdma_queue, arm_bit);
1791 }
1792 
1793 #define MANA_POLL_BUDGET	8
1794 #define MANA_RX_BUDGET		256
1795 #define MANA_TX_BUDGET		MAX_SEND_BUFFERS_PER_QUEUE
1796 
1797 static void
1798 mana_poll(void *arg, int pending)
1799 {
1800 	struct mana_cq *cq = arg;
1801 	int i;
1802 
1803 	cq->work_done = 0;
1804 	if (cq->type == MANA_CQ_TYPE_RX) {
1805 		cq->budget = MANA_RX_BUDGET;
1806 	} else {
1807 		cq->budget = MANA_TX_BUDGET;
1808 	}
1809 
1810 	for (i = 0; i < MANA_POLL_BUDGET; i++) {
1811 		/*
1812 		 * If this is the last loop, set the budget big enough
1813 		 * so it will arm the CQ any way.
1814 		 */
1815 		if (i == (MANA_POLL_BUDGET - 1))
1816 			cq->budget = CQE_POLLING_BUFFER + 1;
1817 
1818 		mana_cq_handler(cq, cq->gdma_cq);
1819 
1820 		if (cq->work_done < cq->budget)
1821 			break;
1822 
1823 		cq->work_done = 0;
1824 	}
1825 }
1826 
1827 static void
1828 mana_schedule_task(void *arg, struct gdma_queue *gdma_queue)
1829 {
1830 	struct mana_cq *cq = arg;
1831 
1832 	taskqueue_enqueue(cq->cleanup_tq, &cq->cleanup_task);
1833 }
1834 
1835 static void
1836 mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq)
1837 {
1838 	struct gdma_dev *gd = apc->ac->gdma_dev;
1839 
1840 	if (!cq->gdma_cq)
1841 		return;
1842 
1843 	/* Drain cleanup taskqueue */
1844 	if (cq->cleanup_tq) {
1845 		while (taskqueue_cancel(cq->cleanup_tq,
1846 		    &cq->cleanup_task, NULL)) {
1847 			taskqueue_drain(cq->cleanup_tq,
1848 			    &cq->cleanup_task);
1849 		}
1850 
1851 		taskqueue_free(cq->cleanup_tq);
1852 	}
1853 
1854 	mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq);
1855 }
1856 
1857 static void
1858 mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq)
1859 {
1860 	struct gdma_dev *gd = apc->ac->gdma_dev;
1861 	struct mana_send_buf_info *txbuf_info;
1862 	uint32_t pending_sends;
1863 	int i;
1864 
1865 	if (!txq->gdma_sq)
1866 		return;
1867 
1868 	if ((pending_sends = atomic_read(&txq->pending_sends)) > 0) {
1869 		mana_err(NULL,
1870 		    "WARNING: txq pending sends not zero: %u\n",
1871 		    pending_sends);
1872 	}
1873 
1874 	if (txq->next_to_use != txq->next_to_complete) {
1875 		mana_err(NULL,
1876 		    "WARNING: txq buf not completed, "
1877 		    "next use %u, next complete %u\n",
1878 		    txq->next_to_use, txq->next_to_complete);
1879 	}
1880 
1881 	/* Flush buf ring. Grab txq mtx lock */
1882 	if (txq->txq_br) {
1883 		mtx_lock(&txq->txq_mtx);
1884 		drbr_flush(apc->ndev, txq->txq_br);
1885 		mtx_unlock(&txq->txq_mtx);
1886 		buf_ring_free(txq->txq_br, M_DEVBUF);
1887 	}
1888 
1889 	/* Drain taskqueue */
1890 	if (txq->enqueue_tq) {
1891 		while (taskqueue_cancel(txq->enqueue_tq,
1892 		    &txq->enqueue_task, NULL)) {
1893 			taskqueue_drain(txq->enqueue_tq,
1894 			    &txq->enqueue_task);
1895 		}
1896 
1897 		taskqueue_free(txq->enqueue_tq);
1898 	}
1899 
1900 	if (txq->tx_buf_info) {
1901 		/* Free all mbufs which are still in-flight */
1902 		for (i = 0; i < MAX_SEND_BUFFERS_PER_QUEUE; i++) {
1903 			txbuf_info = &txq->tx_buf_info[i];
1904 			if (txbuf_info->mbuf) {
1905 				mana_tx_unmap_mbuf(apc, txbuf_info);
1906 			}
1907 		}
1908 
1909 		free(txq->tx_buf_info, M_DEVBUF);
1910 	}
1911 
1912 	mana_free_counters((counter_u64_t *)&txq->stats,
1913 	    sizeof(txq->stats));
1914 
1915 	mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq);
1916 
1917 	mtx_destroy(&txq->txq_mtx);
1918 }
1919 
1920 static void
1921 mana_destroy_txq(struct mana_port_context *apc)
1922 {
1923 	int i;
1924 
1925 	if (!apc->tx_qp)
1926 		return;
1927 
1928 	for (i = 0; i < apc->num_queues; i++) {
1929 		mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object);
1930 
1931 		mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq);
1932 
1933 		mana_deinit_txq(apc, &apc->tx_qp[i].txq);
1934 	}
1935 
1936 	free(apc->tx_qp, M_DEVBUF);
1937 	apc->tx_qp = NULL;
1938 }
1939 
1940 static int
1941 mana_create_txq(struct mana_port_context *apc, struct ifnet *net)
1942 {
1943 	struct mana_context *ac = apc->ac;
1944 	struct gdma_dev *gd = ac->gdma_dev;
1945 	struct mana_obj_spec wq_spec;
1946 	struct mana_obj_spec cq_spec;
1947 	struct gdma_queue_spec spec;
1948 	struct gdma_context *gc;
1949 	struct mana_txq *txq;
1950 	struct mana_cq *cq;
1951 	uint32_t txq_size;
1952 	uint32_t cq_size;
1953 	int err;
1954 	int i;
1955 
1956 	apc->tx_qp = mallocarray(apc->num_queues, sizeof(struct mana_tx_qp),
1957 	    M_DEVBUF, M_WAITOK | M_ZERO);
1958 	if (!apc->tx_qp)
1959 		return ENOMEM;
1960 
1961 	/*  The minimum size of the WQE is 32 bytes, hence
1962 	 *  MAX_SEND_BUFFERS_PER_QUEUE represents the maximum number of WQEs
1963 	 *  the SQ can store. This value is then used to size other queues
1964 	 *  to prevent overflow.
1965 	 */
1966 	txq_size = MAX_SEND_BUFFERS_PER_QUEUE * 32;
1967 	KASSERT(IS_ALIGNED(txq_size, PAGE_SIZE),
1968 	    ("txq size not page aligned"));
1969 
1970 	cq_size = MAX_SEND_BUFFERS_PER_QUEUE * COMP_ENTRY_SIZE;
1971 	cq_size = ALIGN(cq_size, PAGE_SIZE);
1972 
1973 	gc = gd->gdma_context;
1974 
1975 	for (i = 0; i < apc->num_queues; i++) {
1976 		apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE;
1977 
1978 		/* Create SQ */
1979 		txq = &apc->tx_qp[i].txq;
1980 
1981 		txq->ndev = net;
1982 		txq->vp_offset = apc->tx_vp_offset;
1983 		txq->idx = i;
1984 		txq->alt_txq_idx = i;
1985 
1986 		memset(&spec, 0, sizeof(spec));
1987 		spec.type = GDMA_SQ;
1988 		spec.monitor_avl_buf = true;
1989 		spec.queue_size = txq_size;
1990 		err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq);
1991 		if (err)
1992 			goto out;
1993 
1994 		/* Create SQ's CQ */
1995 		cq = &apc->tx_qp[i].tx_cq;
1996 		cq->type = MANA_CQ_TYPE_TX;
1997 
1998 		cq->txq = txq;
1999 
2000 		memset(&spec, 0, sizeof(spec));
2001 		spec.type = GDMA_CQ;
2002 		spec.monitor_avl_buf = false;
2003 		spec.queue_size = cq_size;
2004 		spec.cq.callback = mana_schedule_task;
2005 		spec.cq.parent_eq = ac->eqs[i].eq;
2006 		spec.cq.context = cq;
2007 		err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2008 		if (err)
2009 			goto out;
2010 
2011 		memset(&wq_spec, 0, sizeof(wq_spec));
2012 		memset(&cq_spec, 0, sizeof(cq_spec));
2013 
2014 		wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle;
2015 		wq_spec.queue_size = txq->gdma_sq->queue_size;
2016 
2017 		cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2018 		cq_spec.queue_size = cq->gdma_cq->queue_size;
2019 		cq_spec.modr_ctx_id = 0;
2020 		cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2021 
2022 		err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ,
2023 		    &wq_spec, &cq_spec, &apc->tx_qp[i].tx_object);
2024 
2025 		if (err)
2026 			goto out;
2027 
2028 		txq->gdma_sq->id = wq_spec.queue_index;
2029 		cq->gdma_cq->id = cq_spec.queue_index;
2030 
2031 		txq->gdma_sq->mem_info.dma_region_handle =
2032 		    GDMA_INVALID_DMA_REGION;
2033 		cq->gdma_cq->mem_info.dma_region_handle =
2034 		    GDMA_INVALID_DMA_REGION;
2035 
2036 		txq->gdma_txq_id = txq->gdma_sq->id;
2037 
2038 		cq->gdma_id = cq->gdma_cq->id;
2039 
2040 		mana_dbg(NULL,
2041 		    "txq %d, txq gdma id %d, txq cq gdma id %d\n",
2042 		    i, txq->gdma_txq_id, cq->gdma_id);;
2043 
2044 		if (cq->gdma_id >= gc->max_num_cqs) {
2045 			if_printf(net, "CQ id %u too large.\n", cq->gdma_id);
2046 			err = EINVAL;
2047 			goto out;
2048 		}
2049 
2050 		gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2051 
2052 		/* Initialize tx specific data */
2053 		txq->tx_buf_info = malloc(MAX_SEND_BUFFERS_PER_QUEUE *
2054 		    sizeof(struct mana_send_buf_info),
2055 		    M_DEVBUF, M_WAITOK | M_ZERO);
2056 		if (unlikely(txq->tx_buf_info == NULL)) {
2057 			if_printf(net,
2058 			    "Failed to allocate tx buf info for SQ %u\n",
2059 			    txq->gdma_sq->id);
2060 			err = ENOMEM;
2061 			goto out;
2062 		}
2063 
2064 
2065 		snprintf(txq->txq_mtx_name, nitems(txq->txq_mtx_name),
2066 		    "mana:tx(%d)", i);
2067 		mtx_init(&txq->txq_mtx, txq->txq_mtx_name, NULL, MTX_DEF);
2068 
2069 		txq->txq_br = buf_ring_alloc(4 * MAX_SEND_BUFFERS_PER_QUEUE,
2070 		    M_DEVBUF, M_WAITOK, &txq->txq_mtx);
2071 		if (unlikely(txq->txq_br == NULL)) {
2072 			if_printf(net,
2073 			    "Failed to allocate buf ring for SQ %u\n",
2074 			    txq->gdma_sq->id);
2075 			err = ENOMEM;
2076 			goto out;
2077 		}
2078 
2079 		/* Allocate taskqueue for deferred send */
2080 		TASK_INIT(&txq->enqueue_task, 0, mana_xmit_taskfunc, txq);
2081 		txq->enqueue_tq = taskqueue_create_fast("mana_tx_enque",
2082 		    M_NOWAIT, taskqueue_thread_enqueue, &txq->enqueue_tq);
2083 		if (unlikely(txq->enqueue_tq == NULL)) {
2084 			if_printf(net,
2085 			    "Unable to create tx %d enqueue task queue\n", i);
2086 			err = ENOMEM;
2087 			goto out;
2088 		}
2089 		taskqueue_start_threads(&txq->enqueue_tq, 1, PI_NET,
2090 		    "mana txq p%u-tx%d", apc->port_idx, i);
2091 
2092 		mana_alloc_counters((counter_u64_t *)&txq->stats,
2093 		    sizeof(txq->stats));
2094 
2095 		/* Allocate and start the cleanup task on CQ */
2096 		cq->do_not_ring_db = false;
2097 
2098 		NET_TASK_INIT(&cq->cleanup_task, 0, mana_poll, cq);
2099 		cq->cleanup_tq =
2100 		    taskqueue_create_fast("mana tx cq cleanup",
2101 		    M_WAITOK, taskqueue_thread_enqueue,
2102 		    &cq->cleanup_tq);
2103 
2104 		if (apc->last_tx_cq_bind_cpu < 0)
2105 			apc->last_tx_cq_bind_cpu = CPU_FIRST();
2106 		cq->cpu = apc->last_tx_cq_bind_cpu;
2107 		apc->last_tx_cq_bind_cpu = CPU_NEXT(apc->last_tx_cq_bind_cpu);
2108 
2109 		if (apc->bind_cleanup_thread_cpu) {
2110 			cpuset_t cpu_mask;
2111 			CPU_SETOF(cq->cpu, &cpu_mask);
2112 			taskqueue_start_threads_cpuset(&cq->cleanup_tq,
2113 			    1, PI_NET, &cpu_mask,
2114 			    "mana cq p%u-tx%u-cpu%d",
2115 			    apc->port_idx, txq->idx, cq->cpu);
2116 		} else {
2117 			taskqueue_start_threads(&cq->cleanup_tq, 1,
2118 			    PI_NET, "mana cq p%u-tx%u",
2119 			    apc->port_idx, txq->idx);
2120 		}
2121 
2122 		mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2123 	}
2124 
2125 	return 0;
2126 out:
2127 	mana_destroy_txq(apc);
2128 	return err;
2129 }
2130 
2131 static void
2132 mana_destroy_rxq(struct mana_port_context *apc, struct mana_rxq *rxq,
2133     bool validate_state)
2134 {
2135 	struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
2136 	struct mana_recv_buf_oob *rx_oob;
2137 	int i;
2138 
2139 	if (!rxq)
2140 		return;
2141 
2142 	if (validate_state) {
2143 		/*
2144 		 * XXX Cancel and drain cleanup task queue here.
2145 		 */
2146 		;
2147 	}
2148 
2149 	mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj);
2150 
2151 	mana_deinit_cq(apc, &rxq->rx_cq);
2152 
2153 	mana_free_counters((counter_u64_t *)&rxq->stats,
2154 	    sizeof(rxq->stats));
2155 
2156 	/* Free LRO resources */
2157 	tcp_lro_free(&rxq->lro);
2158 
2159 	for (i = 0; i < rxq->num_rx_buf; i++) {
2160 		rx_oob = &rxq->rx_oobs[i];
2161 
2162 		if (rx_oob->mbuf)
2163 			mana_unload_rx_mbuf(apc, rxq, rx_oob, true);
2164 
2165 		bus_dmamap_destroy(apc->rx_buf_tag, rx_oob->dma_map);
2166 	}
2167 
2168 	if (rxq->gdma_rq)
2169 		mana_gd_destroy_queue(gc, rxq->gdma_rq);
2170 
2171 	free(rxq, M_DEVBUF);
2172 }
2173 
2174 #define MANA_WQE_HEADER_SIZE 16
2175 #define MANA_WQE_SGE_SIZE 16
2176 
2177 static int
2178 mana_alloc_rx_wqe(struct mana_port_context *apc,
2179     struct mana_rxq *rxq, uint32_t *rxq_size, uint32_t *cq_size)
2180 {
2181 	struct mana_recv_buf_oob *rx_oob;
2182 	uint32_t buf_idx;
2183 	int err;
2184 
2185 	if (rxq->datasize == 0 || rxq->datasize > PAGE_SIZE) {
2186 		mana_err(NULL,
2187 		    "WARNING: Invalid rxq datasize %u\n", rxq->datasize);
2188 	}
2189 
2190 	*rxq_size = 0;
2191 	*cq_size = 0;
2192 
2193 	for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2194 		rx_oob = &rxq->rx_oobs[buf_idx];
2195 		memset(rx_oob, 0, sizeof(*rx_oob));
2196 
2197 		err = bus_dmamap_create(apc->rx_buf_tag, 0,
2198 		    &rx_oob->dma_map);
2199 		if (err) {
2200 			mana_err(NULL,
2201 			    "Failed to  create rx DMA map for buf %d\n",
2202 			    buf_idx);
2203 			return err;
2204 		}
2205 
2206 		err = mana_load_rx_mbuf(apc, rxq, rx_oob, true);
2207 		if (err) {
2208 			mana_err(NULL,
2209 			    "Failed to  create rx DMA map for buf %d\n",
2210 			    buf_idx);
2211 			bus_dmamap_destroy(apc->rx_buf_tag, rx_oob->dma_map);
2212 			return err;
2213 		}
2214 
2215 		rx_oob->wqe_req.sgl = rx_oob->sgl;
2216 		rx_oob->wqe_req.num_sge = rx_oob->num_sge;
2217 		rx_oob->wqe_req.inline_oob_size = 0;
2218 		rx_oob->wqe_req.inline_oob_data = NULL;
2219 		rx_oob->wqe_req.flags = 0;
2220 		rx_oob->wqe_req.client_data_unit = 0;
2221 
2222 		*rxq_size += ALIGN(MANA_WQE_HEADER_SIZE +
2223 				   MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32);
2224 		*cq_size += COMP_ENTRY_SIZE;
2225 	}
2226 
2227 	return 0;
2228 }
2229 
2230 static int
2231 mana_push_wqe(struct mana_rxq *rxq)
2232 {
2233 	struct mana_recv_buf_oob *rx_oob;
2234 	uint32_t buf_idx;
2235 	int err;
2236 
2237 	for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2238 		rx_oob = &rxq->rx_oobs[buf_idx];
2239 
2240 		err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req,
2241 		    &rx_oob->wqe_inf);
2242 		if (err)
2243 			return ENOSPC;
2244 	}
2245 
2246 	return 0;
2247 }
2248 
2249 static struct mana_rxq *
2250 mana_create_rxq(struct mana_port_context *apc, uint32_t rxq_idx,
2251     struct mana_eq *eq, struct ifnet *ndev)
2252 {
2253 	struct gdma_dev *gd = apc->ac->gdma_dev;
2254 	struct mana_obj_spec wq_spec;
2255 	struct mana_obj_spec cq_spec;
2256 	struct gdma_queue_spec spec;
2257 	struct mana_cq *cq = NULL;
2258 	uint32_t cq_size, rq_size;
2259 	struct gdma_context *gc;
2260 	struct mana_rxq *rxq;
2261 	int err;
2262 
2263 	gc = gd->gdma_context;
2264 
2265 	rxq = malloc(sizeof(*rxq) +
2266 	    RX_BUFFERS_PER_QUEUE * sizeof(struct mana_recv_buf_oob),
2267 	    M_DEVBUF, M_WAITOK | M_ZERO);
2268 	if (!rxq)
2269 		return NULL;
2270 
2271 	rxq->ndev = ndev;
2272 	rxq->num_rx_buf = RX_BUFFERS_PER_QUEUE;
2273 	rxq->rxq_idx = rxq_idx;
2274 	/*
2275 	 * Minimum size is MCLBYTES(2048) bytes for a mbuf cluster.
2276 	 * Now we just allow maximum size of 4096.
2277 	 */
2278 	rxq->datasize = ALIGN(apc->frame_size, MCLBYTES);
2279 	if (rxq->datasize > MAX_FRAME_SIZE)
2280 		rxq->datasize = MAX_FRAME_SIZE;
2281 
2282 	mana_dbg(NULL, "Setting rxq %d datasize %d\n",
2283 	    rxq_idx, rxq->datasize);
2284 
2285 	rxq->rxobj = INVALID_MANA_HANDLE;
2286 
2287 	err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size);
2288 	if (err)
2289 		goto out;
2290 
2291 	/* Create LRO for the RQ */
2292 	if (ndev->if_capenable & IFCAP_LRO) {
2293 		err = tcp_lro_init(&rxq->lro);
2294 		if (err) {
2295 			if_printf(ndev, "Failed to create LRO for rxq %d\n",
2296 			    rxq_idx);
2297 		} else {
2298 			rxq->lro.ifp = ndev;
2299 		}
2300 	}
2301 
2302 	mana_alloc_counters((counter_u64_t *)&rxq->stats,
2303 	    sizeof(rxq->stats));
2304 
2305 	rq_size = ALIGN(rq_size, PAGE_SIZE);
2306 	cq_size = ALIGN(cq_size, PAGE_SIZE);
2307 
2308 	/* Create RQ */
2309 	memset(&spec, 0, sizeof(spec));
2310 	spec.type = GDMA_RQ;
2311 	spec.monitor_avl_buf = true;
2312 	spec.queue_size = rq_size;
2313 	err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq);
2314 	if (err)
2315 		goto out;
2316 
2317 	/* Create RQ's CQ */
2318 	cq = &rxq->rx_cq;
2319 	cq->type = MANA_CQ_TYPE_RX;
2320 	cq->rxq = rxq;
2321 
2322 	memset(&spec, 0, sizeof(spec));
2323 	spec.type = GDMA_CQ;
2324 	spec.monitor_avl_buf = false;
2325 	spec.queue_size = cq_size;
2326 	spec.cq.callback = mana_schedule_task;
2327 	spec.cq.parent_eq = eq->eq;
2328 	spec.cq.context = cq;
2329 	err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2330 	if (err)
2331 		goto out;
2332 
2333 	memset(&wq_spec, 0, sizeof(wq_spec));
2334 	memset(&cq_spec, 0, sizeof(cq_spec));
2335 	wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle;
2336 	wq_spec.queue_size = rxq->gdma_rq->queue_size;
2337 
2338 	cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2339 	cq_spec.queue_size = cq->gdma_cq->queue_size;
2340 	cq_spec.modr_ctx_id = 0;
2341 	cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2342 
2343 	err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ,
2344 	    &wq_spec, &cq_spec, &rxq->rxobj);
2345 	if (err)
2346 		goto out;
2347 
2348 	rxq->gdma_rq->id = wq_spec.queue_index;
2349 	cq->gdma_cq->id = cq_spec.queue_index;
2350 
2351 	rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2352 	cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2353 
2354 	rxq->gdma_id = rxq->gdma_rq->id;
2355 	cq->gdma_id = cq->gdma_cq->id;
2356 
2357 	err = mana_push_wqe(rxq);
2358 	if (err)
2359 		goto out;
2360 
2361 	if (cq->gdma_id >= gc->max_num_cqs) {
2362 		err = EINVAL;
2363 		goto out;
2364 	}
2365 
2366 	gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2367 
2368 	/* Allocate and start the cleanup task on CQ */
2369 	cq->do_not_ring_db = false;
2370 
2371 	NET_TASK_INIT(&cq->cleanup_task, 0, mana_poll, cq);
2372 	cq->cleanup_tq =
2373 	    taskqueue_create_fast("mana rx cq cleanup",
2374 	    M_WAITOK, taskqueue_thread_enqueue,
2375 	    &cq->cleanup_tq);
2376 
2377 	if (apc->last_rx_cq_bind_cpu < 0)
2378 		apc->last_rx_cq_bind_cpu = CPU_FIRST();
2379 	cq->cpu = apc->last_rx_cq_bind_cpu;
2380 	apc->last_rx_cq_bind_cpu = CPU_NEXT(apc->last_rx_cq_bind_cpu);
2381 
2382 	if (apc->bind_cleanup_thread_cpu) {
2383 		cpuset_t cpu_mask;
2384 		CPU_SETOF(cq->cpu, &cpu_mask);
2385 		taskqueue_start_threads_cpuset(&cq->cleanup_tq,
2386 		    1, PI_NET, &cpu_mask,
2387 		    "mana cq p%u-rx%u-cpu%d",
2388 		    apc->port_idx, rxq->rxq_idx, cq->cpu);
2389 	} else {
2390 		taskqueue_start_threads(&cq->cleanup_tq, 1,
2391 		    PI_NET, "mana cq p%u-rx%u",
2392 		    apc->port_idx, rxq->rxq_idx);
2393 	}
2394 
2395 	mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2396 out:
2397 	if (!err)
2398 		return rxq;
2399 
2400 	if_printf(ndev, "Failed to create RXQ: err = %d\n", err);
2401 
2402 	mana_destroy_rxq(apc, rxq, false);
2403 
2404 	if (cq)
2405 		mana_deinit_cq(apc, cq);
2406 
2407 	return NULL;
2408 }
2409 
2410 static int
2411 mana_add_rx_queues(struct mana_port_context *apc, struct ifnet *ndev)
2412 {
2413 	struct mana_context *ac = apc->ac;
2414 	struct mana_rxq *rxq;
2415 	int err = 0;
2416 	int i;
2417 
2418 	for (i = 0; i < apc->num_queues; i++) {
2419 		rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev);
2420 		if (!rxq) {
2421 			err = ENOMEM;
2422 			goto out;
2423 		}
2424 
2425 		apc->rxqs[i] = rxq;
2426 	}
2427 
2428 	apc->default_rxobj = apc->rxqs[0]->rxobj;
2429 out:
2430 	return err;
2431 }
2432 
2433 static void
2434 mana_destroy_vport(struct mana_port_context *apc)
2435 {
2436 	struct mana_rxq *rxq;
2437 	uint32_t rxq_idx;
2438 
2439 	for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
2440 		rxq = apc->rxqs[rxq_idx];
2441 		if (!rxq)
2442 			continue;
2443 
2444 		mana_destroy_rxq(apc, rxq, true);
2445 		apc->rxqs[rxq_idx] = NULL;
2446 	}
2447 
2448 	mana_destroy_txq(apc);
2449 
2450 	mana_uncfg_vport(apc);
2451 }
2452 
2453 static int
2454 mana_create_vport(struct mana_port_context *apc, struct ifnet *net)
2455 {
2456 	struct gdma_dev *gd = apc->ac->gdma_dev;
2457 	int err;
2458 
2459 	apc->default_rxobj = INVALID_MANA_HANDLE;
2460 
2461 	err = mana_cfg_vport(apc, gd->pdid, gd->doorbell);
2462 	if (err)
2463 		return err;
2464 
2465 	return mana_create_txq(apc, net);
2466 }
2467 
2468 
2469 static void mana_rss_table_init(struct mana_port_context *apc)
2470 {
2471 	int i;
2472 
2473 	for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++)
2474 		apc->indir_table[i] = i % apc->num_queues;
2475 }
2476 
2477 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx,
2478 		    bool update_hash, bool update_tab)
2479 {
2480 	uint32_t queue_idx;
2481 	int err;
2482 	int i;
2483 
2484 	if (update_tab) {
2485 		for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) {
2486 			queue_idx = apc->indir_table[i];
2487 			apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj;
2488 		}
2489 	}
2490 
2491 	err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab);
2492 	if (err)
2493 		return err;
2494 
2495 	mana_fence_rqs(apc);
2496 
2497 	return 0;
2498 }
2499 
2500 static int
2501 mana_init_port(struct ifnet *ndev)
2502 {
2503 	struct mana_port_context *apc = if_getsoftc(ndev);
2504 	uint32_t max_txq, max_rxq, max_queues;
2505 	int port_idx = apc->port_idx;
2506 	uint32_t num_indirect_entries;
2507 	int err;
2508 
2509 	err = mana_init_port_context(apc);
2510 	if (err)
2511 		return err;
2512 
2513 	err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq,
2514 	    &num_indirect_entries);
2515 	if (err) {
2516 		if_printf(ndev, "Failed to query info for vPort %d\n",
2517 		    port_idx);
2518 		goto reset_apc;
2519 	}
2520 
2521 	max_queues = min_t(uint32_t, max_txq, max_rxq);
2522 	if (apc->max_queues > max_queues)
2523 		apc->max_queues = max_queues;
2524 
2525 	if (apc->num_queues > apc->max_queues)
2526 		apc->num_queues = apc->max_queues;
2527 
2528 	return 0;
2529 
2530 reset_apc:
2531 	bus_dma_tag_destroy(apc->rx_buf_tag);
2532 	apc->rx_buf_tag = NULL;
2533 	free(apc->rxqs, M_DEVBUF);
2534 	apc->rxqs = NULL;
2535 	return err;
2536 }
2537 
2538 int
2539 mana_alloc_queues(struct ifnet *ndev)
2540 {
2541 	struct mana_port_context *apc = if_getsoftc(ndev);
2542 	int err;
2543 
2544 	err = mana_create_vport(apc, ndev);
2545 	if (err)
2546 		return err;
2547 
2548 	err = mana_add_rx_queues(apc, ndev);
2549 	if (err)
2550 		goto destroy_vport;
2551 
2552 	apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE;
2553 
2554 	mana_rss_table_init(apc);
2555 
2556 	err = mana_config_rss(apc, TRI_STATE_TRUE, true, true);
2557 	if (err)
2558 		goto destroy_vport;
2559 
2560 	return 0;
2561 
2562 destroy_vport:
2563 	mana_destroy_vport(apc);
2564 	return err;
2565 }
2566 
2567 static int
2568 mana_up(struct mana_port_context *apc)
2569 {
2570 	int err;
2571 
2572 	mana_dbg(NULL, "mana_up called\n");
2573 
2574 	err = mana_alloc_queues(apc->ndev);
2575 	if (err) {
2576 		mana_err(NULL, "Faile alloc mana queues: %d\n", err);
2577 		return err;
2578 	}
2579 
2580 	/* Add queue specific sysctl */
2581 	mana_sysctl_add_queues(apc);
2582 
2583 	apc->port_is_up = true;
2584 
2585 	/* Ensure port state updated before txq state */
2586 	wmb();
2587 
2588 	if_link_state_change(apc->ndev, LINK_STATE_UP);
2589 	if_setdrvflagbits(apc->ndev, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2590 
2591 	return 0;
2592 }
2593 
2594 
2595 static void
2596 mana_init(void *arg)
2597 {
2598 	struct mana_port_context *apc = (struct mana_port_context *)arg;
2599 
2600 	MANA_APC_LOCK_LOCK(apc);
2601 	if (!apc->port_is_up) {
2602 		mana_up(apc);
2603 	}
2604 	MANA_APC_LOCK_UNLOCK(apc);
2605 }
2606 
2607 static int
2608 mana_dealloc_queues(struct ifnet *ndev)
2609 {
2610 	struct mana_port_context *apc = if_getsoftc(ndev);
2611 	struct mana_txq *txq;
2612 	int i, err;
2613 
2614 	if (apc->port_is_up)
2615 		return EINVAL;
2616 
2617 	/* No packet can be transmitted now since apc->port_is_up is false.
2618 	 * There is still a tiny chance that mana_poll_tx_cq() can re-enable
2619 	 * a txq because it may not timely see apc->port_is_up being cleared
2620 	 * to false, but it doesn't matter since mana_start_xmit() drops any
2621 	 * new packets due to apc->port_is_up being false.
2622 	 *
2623 	 * Drain all the in-flight TX packets
2624 	 */
2625 	for (i = 0; i < apc->num_queues; i++) {
2626 		txq = &apc->tx_qp[i].txq;
2627 
2628 		struct mana_cq *tx_cq = &apc->tx_qp[i].tx_cq;
2629 		struct mana_cq *rx_cq = &(apc->rxqs[i]->rx_cq);
2630 
2631 		tx_cq->do_not_ring_db = true;
2632 		rx_cq->do_not_ring_db = true;
2633 
2634 		/* Schedule a cleanup task */
2635 		taskqueue_enqueue(tx_cq->cleanup_tq, &tx_cq->cleanup_task);
2636 
2637 		while (atomic_read(&txq->pending_sends) > 0)
2638 			usleep_range(1000, 2000);
2639 	}
2640 
2641 	/* We're 100% sure the queues can no longer be woken up, because
2642 	 * we're sure now mana_poll_tx_cq() can't be running.
2643 	 */
2644 
2645 	apc->rss_state = TRI_STATE_FALSE;
2646 	err = mana_config_rss(apc, TRI_STATE_FALSE, false, false);
2647 	if (err) {
2648 		if_printf(ndev, "Failed to disable vPort: %d\n", err);
2649 		return err;
2650 	}
2651 
2652 	mana_destroy_vport(apc);
2653 
2654 	return 0;
2655 }
2656 
2657 static int
2658 mana_down(struct mana_port_context *apc)
2659 {
2660 	int err = 0;
2661 
2662 	apc->port_st_save = apc->port_is_up;
2663 	apc->port_is_up = false;
2664 
2665 	/* Ensure port state updated before txq state */
2666 	wmb();
2667 
2668 	if (apc->port_st_save) {
2669 		if_setdrvflagbits(apc->ndev, IFF_DRV_OACTIVE,
2670 		    IFF_DRV_RUNNING);
2671 		if_link_state_change(apc->ndev, LINK_STATE_DOWN);
2672 
2673 		mana_sysctl_free_queues(apc);
2674 
2675 		err = mana_dealloc_queues(apc->ndev);
2676 		if (err) {
2677 			if_printf(apc->ndev,
2678 			    "Failed to bring down mana interface: %d\n", err);
2679 		}
2680 	}
2681 
2682 	return err;
2683 }
2684 
2685 int
2686 mana_detach(struct ifnet *ndev)
2687 {
2688 	struct mana_port_context *apc = if_getsoftc(ndev);
2689 	int err;
2690 
2691 	ether_ifdetach(ndev);
2692 
2693 	if (!apc)
2694 		return 0;
2695 
2696 	MANA_APC_LOCK_LOCK(apc);
2697 	err = mana_down(apc);
2698 	MANA_APC_LOCK_UNLOCK(apc);
2699 
2700 	mana_cleanup_port_context(apc);
2701 
2702 	MANA_APC_LOCK_DESTROY(apc);
2703 
2704 	free(apc, M_DEVBUF);
2705 
2706 	return err;
2707 }
2708 
2709 static int
2710 mana_probe_port(struct mana_context *ac, int port_idx,
2711     struct ifnet **ndev_storage)
2712 {
2713 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
2714 	struct mana_port_context *apc;
2715 	struct ifnet *ndev;
2716 	int err;
2717 
2718 	ndev = if_alloc_dev(IFT_ETHER, gc->dev);
2719 	if (!ndev) {
2720 		mana_err(NULL, "Failed to allocate ifnet struct\n");
2721 		return ENOMEM;
2722 	}
2723 
2724 	*ndev_storage = ndev;
2725 
2726 	apc = malloc(sizeof(*apc), M_DEVBUF, M_WAITOK | M_ZERO);
2727 	if (!apc) {
2728 		mana_err(NULL, "Failed to allocate port context\n");
2729 		err = ENOMEM;
2730 		goto free_net;
2731 	}
2732 
2733 	apc->ac = ac;
2734 	apc->ndev = ndev;
2735 	apc->max_queues = gc->max_num_queues;
2736 	apc->num_queues = min_t(unsigned int,
2737 	    gc->max_num_queues, MANA_MAX_NUM_QUEUES);
2738 	apc->port_handle = INVALID_MANA_HANDLE;
2739 	apc->port_idx = port_idx;
2740 	apc->frame_size = DEFAULT_FRAME_SIZE;
2741 	apc->last_tx_cq_bind_cpu = -1;
2742 	apc->last_rx_cq_bind_cpu = -1;
2743 	apc->vport_use_count = 0;
2744 
2745 	MANA_APC_LOCK_INIT(apc);
2746 
2747 	if_initname(ndev, device_get_name(gc->dev), port_idx);
2748 	if_setdev(ndev,gc->dev);
2749 	if_setsoftc(ndev, apc);
2750 
2751 	if_setflags(ndev, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
2752 	if_setinitfn(ndev, mana_init);
2753 	if_settransmitfn(ndev, mana_start_xmit);
2754 	if_setqflushfn(ndev, mana_qflush);
2755 	if_setioctlfn(ndev, mana_ioctl);
2756 	if_setgetcounterfn(ndev, mana_get_counter);
2757 
2758 	if_setmtu(ndev, ETHERMTU);
2759 	if_setbaudrate(ndev, IF_Gbps(100));
2760 
2761 	mana_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE);
2762 
2763 	err = mana_init_port(ndev);
2764 	if (err)
2765 		goto reset_apc;
2766 
2767 	ndev->if_capabilities |= IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6;
2768 	ndev->if_capabilities |= IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6;
2769 	ndev->if_capabilities |= IFCAP_TSO4 | IFCAP_TSO6;
2770 
2771 	ndev->if_capabilities |= IFCAP_LRO | IFCAP_LINKSTATE;
2772 
2773 	/* Enable all available capabilities by default. */
2774 	ndev->if_capenable = ndev->if_capabilities;
2775 
2776 	/* TSO parameters */
2777 	ndev->if_hw_tsomax = MAX_MBUF_FRAGS * MANA_TSO_MAXSEG_SZ -
2778 	    (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
2779 	ndev->if_hw_tsomaxsegcount = MAX_MBUF_FRAGS;
2780 	ndev->if_hw_tsomaxsegsize = PAGE_SIZE;
2781 
2782 	ifmedia_init(&apc->media, IFM_IMASK,
2783 	    mana_ifmedia_change, mana_ifmedia_status);
2784 	ifmedia_add(&apc->media, IFM_ETHER | IFM_AUTO, 0, NULL);
2785 	ifmedia_set(&apc->media, IFM_ETHER | IFM_AUTO);
2786 
2787 	ether_ifattach(ndev, apc->mac_addr);
2788 
2789 	/* Initialize statistics */
2790 	mana_alloc_counters((counter_u64_t *)&apc->port_stats,
2791 	    sizeof(struct mana_port_stats));
2792 	mana_sysctl_add_port(apc);
2793 
2794 	/* Tell the stack that the interface is not active */
2795 	if_setdrvflagbits(ndev, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2796 
2797 	return 0;
2798 
2799 reset_apc:
2800 	free(apc, M_DEVBUF);
2801 free_net:
2802 	*ndev_storage = NULL;
2803 	if_printf(ndev, "Failed to probe vPort %d: %d\n", port_idx, err);
2804 	if_free(ndev);
2805 	return err;
2806 }
2807 
2808 int mana_probe(struct gdma_dev *gd)
2809 {
2810 	struct gdma_context *gc = gd->gdma_context;
2811 	device_t dev = gc->dev;
2812 	struct mana_context *ac;
2813 	int err;
2814 	int i;
2815 
2816 	device_printf(dev, "%s protocol version: %d.%d.%d\n", DEVICE_NAME,
2817 		 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION);
2818 
2819 	err = mana_gd_register_device(gd);
2820 	if (err)
2821 		return err;
2822 
2823 	ac = malloc(sizeof(*ac), M_DEVBUF, M_WAITOK | M_ZERO);
2824 	if (!ac)
2825 		return ENOMEM;
2826 
2827 	ac->gdma_dev = gd;
2828 	ac->num_ports = 1;
2829 	gd->driver_data = ac;
2830 
2831 	err = mana_create_eq(ac);
2832 	if (err)
2833 		goto out;
2834 
2835 	err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION,
2836 	    MANA_MICRO_VERSION, &ac->num_ports);
2837 	if (err)
2838 		goto out;
2839 
2840 	if (ac->num_ports > MAX_PORTS_IN_MANA_DEV)
2841 		ac->num_ports = MAX_PORTS_IN_MANA_DEV;
2842 
2843 	for (i = 0; i < ac->num_ports; i++) {
2844 		err = mana_probe_port(ac, i, &ac->ports[i]);
2845 		if (err) {
2846 			device_printf(dev,
2847 			    "Failed to probe mana port %d\n", i);
2848 			break;
2849 		}
2850 	}
2851 
2852 out:
2853 	if (err)
2854 		mana_remove(gd);
2855 
2856 	return err;
2857 }
2858 
2859 void
2860 mana_remove(struct gdma_dev *gd)
2861 {
2862 	struct gdma_context *gc = gd->gdma_context;
2863 	struct mana_context *ac = gd->driver_data;
2864 	device_t dev = gc->dev;
2865 	struct ifnet *ndev;
2866 	int i;
2867 
2868 	for (i = 0; i < ac->num_ports; i++) {
2869 		ndev = ac->ports[i];
2870 		if (!ndev) {
2871 			if (i == 0)
2872 				device_printf(dev, "No net device to remove\n");
2873 			goto out;
2874 		}
2875 
2876 		mana_detach(ndev);
2877 
2878 		if_free(ndev);
2879 	}
2880 
2881 	mana_destroy_eq(ac);
2882 
2883 out:
2884 	mana_gd_deregister_device(gd);
2885 	gd->driver_data = NULL;
2886 	gd->gdma_context = NULL;
2887 	free(ac, M_DEVBUF);
2888 }
2889