1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /* Copyright (c) 2021, Microsoft Corporation. */
3 
4 #include <uapi/linux/bpf.h>
5 
6 #include <linux/inetdevice.h>
7 #include <linux/etherdevice.h>
8 #include <linux/ethtool.h>
9 #include <linux/filter.h>
10 #include <linux/mm.h>
11 
12 #include <net/checksum.h>
13 #include <net/ip6_checksum.h>
14 #include <net/page_pool/helpers.h>
15 #include <net/xdp.h>
16 
17 #include <net/mana/mana.h>
18 #include <net/mana/mana_auxiliary.h>
19 
20 static DEFINE_IDA(mana_adev_ida);
21 
22 static int mana_adev_idx_alloc(void)
23 {
24 	return ida_alloc(&mana_adev_ida, GFP_KERNEL);
25 }
26 
27 static void mana_adev_idx_free(int idx)
28 {
29 	ida_free(&mana_adev_ida, idx);
30 }
31 
32 /* Microsoft Azure Network Adapter (MANA) functions */
33 
34 static int mana_open(struct net_device *ndev)
35 {
36 	struct mana_port_context *apc = netdev_priv(ndev);
37 	int err;
38 
39 	err = mana_alloc_queues(ndev);
40 	if (err)
41 		return err;
42 
43 	apc->port_is_up = true;
44 
45 	/* Ensure port state updated before txq state */
46 	smp_wmb();
47 
48 	netif_carrier_on(ndev);
49 	netif_tx_wake_all_queues(ndev);
50 
51 	return 0;
52 }
53 
54 static int mana_close(struct net_device *ndev)
55 {
56 	struct mana_port_context *apc = netdev_priv(ndev);
57 
58 	if (!apc->port_is_up)
59 		return 0;
60 
61 	return mana_detach(ndev, true);
62 }
63 
64 static bool mana_can_tx(struct gdma_queue *wq)
65 {
66 	return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE;
67 }
68 
69 static unsigned int mana_checksum_info(struct sk_buff *skb)
70 {
71 	if (skb->protocol == htons(ETH_P_IP)) {
72 		struct iphdr *ip = ip_hdr(skb);
73 
74 		if (ip->protocol == IPPROTO_TCP)
75 			return IPPROTO_TCP;
76 
77 		if (ip->protocol == IPPROTO_UDP)
78 			return IPPROTO_UDP;
79 	} else if (skb->protocol == htons(ETH_P_IPV6)) {
80 		struct ipv6hdr *ip6 = ipv6_hdr(skb);
81 
82 		if (ip6->nexthdr == IPPROTO_TCP)
83 			return IPPROTO_TCP;
84 
85 		if (ip6->nexthdr == IPPROTO_UDP)
86 			return IPPROTO_UDP;
87 	}
88 
89 	/* No csum offloading */
90 	return 0;
91 }
92 
93 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc,
94 			struct mana_tx_package *tp)
95 {
96 	struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
97 	struct gdma_dev *gd = apc->ac->gdma_dev;
98 	struct gdma_context *gc;
99 	struct device *dev;
100 	skb_frag_t *frag;
101 	dma_addr_t da;
102 	int i;
103 
104 	gc = gd->gdma_context;
105 	dev = gc->dev;
106 	da = dma_map_single(dev, skb->data, skb_headlen(skb), DMA_TO_DEVICE);
107 
108 	if (dma_mapping_error(dev, da))
109 		return -ENOMEM;
110 
111 	ash->dma_handle[0] = da;
112 	ash->size[0] = skb_headlen(skb);
113 
114 	tp->wqe_req.sgl[0].address = ash->dma_handle[0];
115 	tp->wqe_req.sgl[0].mem_key = gd->gpa_mkey;
116 	tp->wqe_req.sgl[0].size = ash->size[0];
117 
118 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
119 		frag = &skb_shinfo(skb)->frags[i];
120 		da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag),
121 				      DMA_TO_DEVICE);
122 
123 		if (dma_mapping_error(dev, da))
124 			goto frag_err;
125 
126 		ash->dma_handle[i + 1] = da;
127 		ash->size[i + 1] = skb_frag_size(frag);
128 
129 		tp->wqe_req.sgl[i + 1].address = ash->dma_handle[i + 1];
130 		tp->wqe_req.sgl[i + 1].mem_key = gd->gpa_mkey;
131 		tp->wqe_req.sgl[i + 1].size = ash->size[i + 1];
132 	}
133 
134 	return 0;
135 
136 frag_err:
137 	for (i = i - 1; i >= 0; i--)
138 		dma_unmap_page(dev, ash->dma_handle[i + 1], ash->size[i + 1],
139 			       DMA_TO_DEVICE);
140 
141 	dma_unmap_single(dev, ash->dma_handle[0], ash->size[0], DMA_TO_DEVICE);
142 
143 	return -ENOMEM;
144 }
145 
146 netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev)
147 {
148 	enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT;
149 	struct mana_port_context *apc = netdev_priv(ndev);
150 	u16 txq_idx = skb_get_queue_mapping(skb);
151 	struct gdma_dev *gd = apc->ac->gdma_dev;
152 	bool ipv4 = false, ipv6 = false;
153 	struct mana_tx_package pkg = {};
154 	struct netdev_queue *net_txq;
155 	struct mana_stats_tx *tx_stats;
156 	struct gdma_queue *gdma_sq;
157 	unsigned int csum_type;
158 	struct mana_txq *txq;
159 	struct mana_cq *cq;
160 	int err, len;
161 	u16 ihs;
162 
163 	if (unlikely(!apc->port_is_up))
164 		goto tx_drop;
165 
166 	if (skb_cow_head(skb, MANA_HEADROOM))
167 		goto tx_drop_count;
168 
169 	txq = &apc->tx_qp[txq_idx].txq;
170 	gdma_sq = txq->gdma_sq;
171 	cq = &apc->tx_qp[txq_idx].tx_cq;
172 	tx_stats = &txq->stats;
173 
174 	pkg.tx_oob.s_oob.vcq_num = cq->gdma_id;
175 	pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame;
176 
177 	if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) {
178 		pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset;
179 		pkt_fmt = MANA_LONG_PKT_FMT;
180 	} else {
181 		pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset;
182 	}
183 
184 	if (skb_vlan_tag_present(skb)) {
185 		pkt_fmt = MANA_LONG_PKT_FMT;
186 		pkg.tx_oob.l_oob.inject_vlan_pri_tag = 1;
187 		pkg.tx_oob.l_oob.pcp = skb_vlan_tag_get_prio(skb);
188 		pkg.tx_oob.l_oob.dei = skb_vlan_tag_get_cfi(skb);
189 		pkg.tx_oob.l_oob.vlan_id = skb_vlan_tag_get_id(skb);
190 	}
191 
192 	pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt;
193 
194 	if (pkt_fmt == MANA_SHORT_PKT_FMT) {
195 		pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob);
196 		u64_stats_update_begin(&tx_stats->syncp);
197 		tx_stats->short_pkt_fmt++;
198 		u64_stats_update_end(&tx_stats->syncp);
199 	} else {
200 		pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob);
201 		u64_stats_update_begin(&tx_stats->syncp);
202 		tx_stats->long_pkt_fmt++;
203 		u64_stats_update_end(&tx_stats->syncp);
204 	}
205 
206 	pkg.wqe_req.inline_oob_data = &pkg.tx_oob;
207 	pkg.wqe_req.flags = 0;
208 	pkg.wqe_req.client_data_unit = 0;
209 
210 	pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags;
211 	WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES);
212 
213 	if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) {
214 		pkg.wqe_req.sgl = pkg.sgl_array;
215 	} else {
216 		pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge,
217 					    sizeof(struct gdma_sge),
218 					    GFP_ATOMIC);
219 		if (!pkg.sgl_ptr)
220 			goto tx_drop_count;
221 
222 		pkg.wqe_req.sgl = pkg.sgl_ptr;
223 	}
224 
225 	if (skb->protocol == htons(ETH_P_IP))
226 		ipv4 = true;
227 	else if (skb->protocol == htons(ETH_P_IPV6))
228 		ipv6 = true;
229 
230 	if (skb_is_gso(skb)) {
231 		pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
232 		pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
233 
234 		pkg.tx_oob.s_oob.comp_iphdr_csum = 1;
235 		pkg.tx_oob.s_oob.comp_tcp_csum = 1;
236 		pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
237 
238 		pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size;
239 		pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0;
240 		if (ipv4) {
241 			ip_hdr(skb)->tot_len = 0;
242 			ip_hdr(skb)->check = 0;
243 			tcp_hdr(skb)->check =
244 				~csum_tcpudp_magic(ip_hdr(skb)->saddr,
245 						   ip_hdr(skb)->daddr, 0,
246 						   IPPROTO_TCP, 0);
247 		} else {
248 			ipv6_hdr(skb)->payload_len = 0;
249 			tcp_hdr(skb)->check =
250 				~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
251 						 &ipv6_hdr(skb)->daddr, 0,
252 						 IPPROTO_TCP, 0);
253 		}
254 
255 		if (skb->encapsulation) {
256 			ihs = skb_inner_tcp_all_headers(skb);
257 			u64_stats_update_begin(&tx_stats->syncp);
258 			tx_stats->tso_inner_packets++;
259 			tx_stats->tso_inner_bytes += skb->len - ihs;
260 			u64_stats_update_end(&tx_stats->syncp);
261 		} else {
262 			if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
263 				ihs = skb_transport_offset(skb) + sizeof(struct udphdr);
264 			} else {
265 				ihs = skb_tcp_all_headers(skb);
266 				if (ipv6_has_hopopt_jumbo(skb))
267 					ihs -= sizeof(struct hop_jumbo_hdr);
268 			}
269 
270 			u64_stats_update_begin(&tx_stats->syncp);
271 			tx_stats->tso_packets++;
272 			tx_stats->tso_bytes += skb->len - ihs;
273 			u64_stats_update_end(&tx_stats->syncp);
274 		}
275 
276 	} else if (skb->ip_summed == CHECKSUM_PARTIAL) {
277 		csum_type = mana_checksum_info(skb);
278 
279 		u64_stats_update_begin(&tx_stats->syncp);
280 		tx_stats->csum_partial++;
281 		u64_stats_update_end(&tx_stats->syncp);
282 
283 		if (csum_type == IPPROTO_TCP) {
284 			pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
285 			pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
286 
287 			pkg.tx_oob.s_oob.comp_tcp_csum = 1;
288 			pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
289 
290 		} else if (csum_type == IPPROTO_UDP) {
291 			pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
292 			pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
293 
294 			pkg.tx_oob.s_oob.comp_udp_csum = 1;
295 		} else {
296 			/* Can't do offload of this type of checksum */
297 			if (skb_checksum_help(skb))
298 				goto free_sgl_ptr;
299 		}
300 	}
301 
302 	if (mana_map_skb(skb, apc, &pkg)) {
303 		u64_stats_update_begin(&tx_stats->syncp);
304 		tx_stats->mana_map_err++;
305 		u64_stats_update_end(&tx_stats->syncp);
306 		goto free_sgl_ptr;
307 	}
308 
309 	skb_queue_tail(&txq->pending_skbs, skb);
310 
311 	len = skb->len;
312 	net_txq = netdev_get_tx_queue(ndev, txq_idx);
313 
314 	err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req,
315 					(struct gdma_posted_wqe_info *)skb->cb);
316 	if (!mana_can_tx(gdma_sq)) {
317 		netif_tx_stop_queue(net_txq);
318 		apc->eth_stats.stop_queue++;
319 	}
320 
321 	if (err) {
322 		(void)skb_dequeue_tail(&txq->pending_skbs);
323 		netdev_warn(ndev, "Failed to post TX OOB: %d\n", err);
324 		err = NETDEV_TX_BUSY;
325 		goto tx_busy;
326 	}
327 
328 	err = NETDEV_TX_OK;
329 	atomic_inc(&txq->pending_sends);
330 
331 	mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq);
332 
333 	/* skb may be freed after mana_gd_post_work_request. Do not use it. */
334 	skb = NULL;
335 
336 	tx_stats = &txq->stats;
337 	u64_stats_update_begin(&tx_stats->syncp);
338 	tx_stats->packets++;
339 	tx_stats->bytes += len;
340 	u64_stats_update_end(&tx_stats->syncp);
341 
342 tx_busy:
343 	if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) {
344 		netif_tx_wake_queue(net_txq);
345 		apc->eth_stats.wake_queue++;
346 	}
347 
348 	kfree(pkg.sgl_ptr);
349 	return err;
350 
351 free_sgl_ptr:
352 	kfree(pkg.sgl_ptr);
353 tx_drop_count:
354 	ndev->stats.tx_dropped++;
355 tx_drop:
356 	dev_kfree_skb_any(skb);
357 	return NETDEV_TX_OK;
358 }
359 
360 static void mana_get_stats64(struct net_device *ndev,
361 			     struct rtnl_link_stats64 *st)
362 {
363 	struct mana_port_context *apc = netdev_priv(ndev);
364 	unsigned int num_queues = apc->num_queues;
365 	struct mana_stats_rx *rx_stats;
366 	struct mana_stats_tx *tx_stats;
367 	unsigned int start;
368 	u64 packets, bytes;
369 	int q;
370 
371 	if (!apc->port_is_up)
372 		return;
373 
374 	netdev_stats_to_stats64(st, &ndev->stats);
375 
376 	for (q = 0; q < num_queues; q++) {
377 		rx_stats = &apc->rxqs[q]->stats;
378 
379 		do {
380 			start = u64_stats_fetch_begin(&rx_stats->syncp);
381 			packets = rx_stats->packets;
382 			bytes = rx_stats->bytes;
383 		} while (u64_stats_fetch_retry(&rx_stats->syncp, start));
384 
385 		st->rx_packets += packets;
386 		st->rx_bytes += bytes;
387 	}
388 
389 	for (q = 0; q < num_queues; q++) {
390 		tx_stats = &apc->tx_qp[q].txq.stats;
391 
392 		do {
393 			start = u64_stats_fetch_begin(&tx_stats->syncp);
394 			packets = tx_stats->packets;
395 			bytes = tx_stats->bytes;
396 		} while (u64_stats_fetch_retry(&tx_stats->syncp, start));
397 
398 		st->tx_packets += packets;
399 		st->tx_bytes += bytes;
400 	}
401 }
402 
403 static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb,
404 			     int old_q)
405 {
406 	struct mana_port_context *apc = netdev_priv(ndev);
407 	u32 hash = skb_get_hash(skb);
408 	struct sock *sk = skb->sk;
409 	int txq;
410 
411 	txq = apc->indir_table[hash & MANA_INDIRECT_TABLE_MASK];
412 
413 	if (txq != old_q && sk && sk_fullsock(sk) &&
414 	    rcu_access_pointer(sk->sk_dst_cache))
415 		sk_tx_queue_set(sk, txq);
416 
417 	return txq;
418 }
419 
420 static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb,
421 			     struct net_device *sb_dev)
422 {
423 	int txq;
424 
425 	if (ndev->real_num_tx_queues == 1)
426 		return 0;
427 
428 	txq = sk_tx_queue_get(skb->sk);
429 
430 	if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) {
431 		if (skb_rx_queue_recorded(skb))
432 			txq = skb_get_rx_queue(skb);
433 		else
434 			txq = mana_get_tx_queue(ndev, skb, txq);
435 	}
436 
437 	return txq;
438 }
439 
440 /* Release pre-allocated RX buffers */
441 static void mana_pre_dealloc_rxbufs(struct mana_port_context *mpc)
442 {
443 	struct device *dev;
444 	int i;
445 
446 	dev = mpc->ac->gdma_dev->gdma_context->dev;
447 
448 	if (!mpc->rxbufs_pre)
449 		goto out1;
450 
451 	if (!mpc->das_pre)
452 		goto out2;
453 
454 	while (mpc->rxbpre_total) {
455 		i = --mpc->rxbpre_total;
456 		dma_unmap_single(dev, mpc->das_pre[i], mpc->rxbpre_datasize,
457 				 DMA_FROM_DEVICE);
458 		put_page(virt_to_head_page(mpc->rxbufs_pre[i]));
459 	}
460 
461 	kfree(mpc->das_pre);
462 	mpc->das_pre = NULL;
463 
464 out2:
465 	kfree(mpc->rxbufs_pre);
466 	mpc->rxbufs_pre = NULL;
467 
468 out1:
469 	mpc->rxbpre_datasize = 0;
470 	mpc->rxbpre_alloc_size = 0;
471 	mpc->rxbpre_headroom = 0;
472 }
473 
474 /* Get a buffer from the pre-allocated RX buffers */
475 static void *mana_get_rxbuf_pre(struct mana_rxq *rxq, dma_addr_t *da)
476 {
477 	struct net_device *ndev = rxq->ndev;
478 	struct mana_port_context *mpc;
479 	void *va;
480 
481 	mpc = netdev_priv(ndev);
482 
483 	if (!mpc->rxbufs_pre || !mpc->das_pre || !mpc->rxbpre_total) {
484 		netdev_err(ndev, "No RX pre-allocated bufs\n");
485 		return NULL;
486 	}
487 
488 	/* Check sizes to catch unexpected coding error */
489 	if (mpc->rxbpre_datasize != rxq->datasize) {
490 		netdev_err(ndev, "rxbpre_datasize mismatch: %u: %u\n",
491 			   mpc->rxbpre_datasize, rxq->datasize);
492 		return NULL;
493 	}
494 
495 	if (mpc->rxbpre_alloc_size != rxq->alloc_size) {
496 		netdev_err(ndev, "rxbpre_alloc_size mismatch: %u: %u\n",
497 			   mpc->rxbpre_alloc_size, rxq->alloc_size);
498 		return NULL;
499 	}
500 
501 	if (mpc->rxbpre_headroom != rxq->headroom) {
502 		netdev_err(ndev, "rxbpre_headroom mismatch: %u: %u\n",
503 			   mpc->rxbpre_headroom, rxq->headroom);
504 		return NULL;
505 	}
506 
507 	mpc->rxbpre_total--;
508 
509 	*da = mpc->das_pre[mpc->rxbpre_total];
510 	va = mpc->rxbufs_pre[mpc->rxbpre_total];
511 	mpc->rxbufs_pre[mpc->rxbpre_total] = NULL;
512 
513 	/* Deallocate the array after all buffers are gone */
514 	if (!mpc->rxbpre_total)
515 		mana_pre_dealloc_rxbufs(mpc);
516 
517 	return va;
518 }
519 
520 /* Get RX buffer's data size, alloc size, XDP headroom based on MTU */
521 static void mana_get_rxbuf_cfg(int mtu, u32 *datasize, u32 *alloc_size,
522 			       u32 *headroom)
523 {
524 	if (mtu > MANA_XDP_MTU_MAX)
525 		*headroom = 0; /* no support for XDP */
526 	else
527 		*headroom = XDP_PACKET_HEADROOM;
528 
529 	*alloc_size = mtu + MANA_RXBUF_PAD + *headroom;
530 
531 	*datasize = ALIGN(mtu + ETH_HLEN, MANA_RX_DATA_ALIGN);
532 }
533 
534 static int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu)
535 {
536 	struct device *dev;
537 	struct page *page;
538 	dma_addr_t da;
539 	int num_rxb;
540 	void *va;
541 	int i;
542 
543 	mana_get_rxbuf_cfg(new_mtu, &mpc->rxbpre_datasize,
544 			   &mpc->rxbpre_alloc_size, &mpc->rxbpre_headroom);
545 
546 	dev = mpc->ac->gdma_dev->gdma_context->dev;
547 
548 	num_rxb = mpc->num_queues * RX_BUFFERS_PER_QUEUE;
549 
550 	WARN(mpc->rxbufs_pre, "mana rxbufs_pre exists\n");
551 	mpc->rxbufs_pre = kmalloc_array(num_rxb, sizeof(void *), GFP_KERNEL);
552 	if (!mpc->rxbufs_pre)
553 		goto error;
554 
555 	mpc->das_pre = kmalloc_array(num_rxb, sizeof(dma_addr_t), GFP_KERNEL);
556 	if (!mpc->das_pre)
557 		goto error;
558 
559 	mpc->rxbpre_total = 0;
560 
561 	for (i = 0; i < num_rxb; i++) {
562 		if (mpc->rxbpre_alloc_size > PAGE_SIZE) {
563 			va = netdev_alloc_frag(mpc->rxbpre_alloc_size);
564 			if (!va)
565 				goto error;
566 
567 			page = virt_to_head_page(va);
568 			/* Check if the frag falls back to single page */
569 			if (compound_order(page) <
570 			    get_order(mpc->rxbpre_alloc_size)) {
571 				put_page(page);
572 				goto error;
573 			}
574 		} else {
575 			page = dev_alloc_page();
576 			if (!page)
577 				goto error;
578 
579 			va = page_to_virt(page);
580 		}
581 
582 		da = dma_map_single(dev, va + mpc->rxbpre_headroom,
583 				    mpc->rxbpre_datasize, DMA_FROM_DEVICE);
584 		if (dma_mapping_error(dev, da)) {
585 			put_page(virt_to_head_page(va));
586 			goto error;
587 		}
588 
589 		mpc->rxbufs_pre[i] = va;
590 		mpc->das_pre[i] = da;
591 		mpc->rxbpre_total = i + 1;
592 	}
593 
594 	return 0;
595 
596 error:
597 	mana_pre_dealloc_rxbufs(mpc);
598 	return -ENOMEM;
599 }
600 
601 static int mana_change_mtu(struct net_device *ndev, int new_mtu)
602 {
603 	struct mana_port_context *mpc = netdev_priv(ndev);
604 	unsigned int old_mtu = ndev->mtu;
605 	int err;
606 
607 	/* Pre-allocate buffers to prevent failure in mana_attach later */
608 	err = mana_pre_alloc_rxbufs(mpc, new_mtu);
609 	if (err) {
610 		netdev_err(ndev, "Insufficient memory for new MTU\n");
611 		return err;
612 	}
613 
614 	err = mana_detach(ndev, false);
615 	if (err) {
616 		netdev_err(ndev, "mana_detach failed: %d\n", err);
617 		goto out;
618 	}
619 
620 	ndev->mtu = new_mtu;
621 
622 	err = mana_attach(ndev);
623 	if (err) {
624 		netdev_err(ndev, "mana_attach failed: %d\n", err);
625 		ndev->mtu = old_mtu;
626 	}
627 
628 out:
629 	mana_pre_dealloc_rxbufs(mpc);
630 	return err;
631 }
632 
633 static const struct net_device_ops mana_devops = {
634 	.ndo_open		= mana_open,
635 	.ndo_stop		= mana_close,
636 	.ndo_select_queue	= mana_select_queue,
637 	.ndo_start_xmit		= mana_start_xmit,
638 	.ndo_validate_addr	= eth_validate_addr,
639 	.ndo_get_stats64	= mana_get_stats64,
640 	.ndo_bpf		= mana_bpf,
641 	.ndo_xdp_xmit		= mana_xdp_xmit,
642 	.ndo_change_mtu		= mana_change_mtu,
643 };
644 
645 static void mana_cleanup_port_context(struct mana_port_context *apc)
646 {
647 	kfree(apc->rxqs);
648 	apc->rxqs = NULL;
649 }
650 
651 static int mana_init_port_context(struct mana_port_context *apc)
652 {
653 	apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *),
654 			    GFP_KERNEL);
655 
656 	return !apc->rxqs ? -ENOMEM : 0;
657 }
658 
659 static int mana_send_request(struct mana_context *ac, void *in_buf,
660 			     u32 in_len, void *out_buf, u32 out_len)
661 {
662 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
663 	struct gdma_resp_hdr *resp = out_buf;
664 	struct gdma_req_hdr *req = in_buf;
665 	struct device *dev = gc->dev;
666 	static atomic_t activity_id;
667 	int err;
668 
669 	req->dev_id = gc->mana.dev_id;
670 	req->activity_id = atomic_inc_return(&activity_id);
671 
672 	err = mana_gd_send_request(gc, in_len, in_buf, out_len,
673 				   out_buf);
674 	if (err || resp->status) {
675 		dev_err(dev, "Failed to send mana message: %d, 0x%x\n",
676 			err, resp->status);
677 		return err ? err : -EPROTO;
678 	}
679 
680 	if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 ||
681 	    req->activity_id != resp->activity_id) {
682 		dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n",
683 			req->dev_id.as_uint32, resp->dev_id.as_uint32,
684 			req->activity_id, resp->activity_id);
685 		return -EPROTO;
686 	}
687 
688 	return 0;
689 }
690 
691 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr,
692 				const enum mana_command_code expected_code,
693 				const u32 min_size)
694 {
695 	if (resp_hdr->response.msg_type != expected_code)
696 		return -EPROTO;
697 
698 	if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1)
699 		return -EPROTO;
700 
701 	if (resp_hdr->response.msg_size < min_size)
702 		return -EPROTO;
703 
704 	return 0;
705 }
706 
707 static int mana_pf_register_hw_vport(struct mana_port_context *apc)
708 {
709 	struct mana_register_hw_vport_resp resp = {};
710 	struct mana_register_hw_vport_req req = {};
711 	int err;
712 
713 	mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT,
714 			     sizeof(req), sizeof(resp));
715 	req.attached_gfid = 1;
716 	req.is_pf_default_vport = 1;
717 	req.allow_all_ether_types = 1;
718 
719 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
720 				sizeof(resp));
721 	if (err) {
722 		netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err);
723 		return err;
724 	}
725 
726 	err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT,
727 				   sizeof(resp));
728 	if (err || resp.hdr.status) {
729 		netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n",
730 			   err, resp.hdr.status);
731 		return err ? err : -EPROTO;
732 	}
733 
734 	apc->port_handle = resp.hw_vport_handle;
735 	return 0;
736 }
737 
738 static void mana_pf_deregister_hw_vport(struct mana_port_context *apc)
739 {
740 	struct mana_deregister_hw_vport_resp resp = {};
741 	struct mana_deregister_hw_vport_req req = {};
742 	int err;
743 
744 	mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT,
745 			     sizeof(req), sizeof(resp));
746 	req.hw_vport_handle = apc->port_handle;
747 
748 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
749 				sizeof(resp));
750 	if (err) {
751 		netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n",
752 			   err);
753 		return;
754 	}
755 
756 	err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT,
757 				   sizeof(resp));
758 	if (err || resp.hdr.status)
759 		netdev_err(apc->ndev,
760 			   "Failed to deregister hw vPort: %d, 0x%x\n",
761 			   err, resp.hdr.status);
762 }
763 
764 static int mana_pf_register_filter(struct mana_port_context *apc)
765 {
766 	struct mana_register_filter_resp resp = {};
767 	struct mana_register_filter_req req = {};
768 	int err;
769 
770 	mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER,
771 			     sizeof(req), sizeof(resp));
772 	req.vport = apc->port_handle;
773 	memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN);
774 
775 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
776 				sizeof(resp));
777 	if (err) {
778 		netdev_err(apc->ndev, "Failed to register filter: %d\n", err);
779 		return err;
780 	}
781 
782 	err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER,
783 				   sizeof(resp));
784 	if (err || resp.hdr.status) {
785 		netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n",
786 			   err, resp.hdr.status);
787 		return err ? err : -EPROTO;
788 	}
789 
790 	apc->pf_filter_handle = resp.filter_handle;
791 	return 0;
792 }
793 
794 static void mana_pf_deregister_filter(struct mana_port_context *apc)
795 {
796 	struct mana_deregister_filter_resp resp = {};
797 	struct mana_deregister_filter_req req = {};
798 	int err;
799 
800 	mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER,
801 			     sizeof(req), sizeof(resp));
802 	req.filter_handle = apc->pf_filter_handle;
803 
804 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
805 				sizeof(resp));
806 	if (err) {
807 		netdev_err(apc->ndev, "Failed to unregister filter: %d\n",
808 			   err);
809 		return;
810 	}
811 
812 	err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER,
813 				   sizeof(resp));
814 	if (err || resp.hdr.status)
815 		netdev_err(apc->ndev,
816 			   "Failed to deregister filter: %d, 0x%x\n",
817 			   err, resp.hdr.status);
818 }
819 
820 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver,
821 				 u32 proto_minor_ver, u32 proto_micro_ver,
822 				 u16 *max_num_vports)
823 {
824 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
825 	struct mana_query_device_cfg_resp resp = {};
826 	struct mana_query_device_cfg_req req = {};
827 	struct device *dev = gc->dev;
828 	int err = 0;
829 
830 	mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG,
831 			     sizeof(req), sizeof(resp));
832 
833 	req.hdr.resp.msg_version = GDMA_MESSAGE_V2;
834 
835 	req.proto_major_ver = proto_major_ver;
836 	req.proto_minor_ver = proto_minor_ver;
837 	req.proto_micro_ver = proto_micro_ver;
838 
839 	err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp));
840 	if (err) {
841 		dev_err(dev, "Failed to query config: %d", err);
842 		return err;
843 	}
844 
845 	err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG,
846 				   sizeof(resp));
847 	if (err || resp.hdr.status) {
848 		dev_err(dev, "Invalid query result: %d, 0x%x\n", err,
849 			resp.hdr.status);
850 		if (!err)
851 			err = -EPROTO;
852 		return err;
853 	}
854 
855 	*max_num_vports = resp.max_num_vports;
856 
857 	if (resp.hdr.response.msg_version == GDMA_MESSAGE_V2)
858 		gc->adapter_mtu = resp.adapter_mtu;
859 	else
860 		gc->adapter_mtu = ETH_FRAME_LEN;
861 
862 	return 0;
863 }
864 
865 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index,
866 				u32 *max_sq, u32 *max_rq, u32 *num_indir_entry)
867 {
868 	struct mana_query_vport_cfg_resp resp = {};
869 	struct mana_query_vport_cfg_req req = {};
870 	int err;
871 
872 	mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG,
873 			     sizeof(req), sizeof(resp));
874 
875 	req.vport_index = vport_index;
876 
877 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
878 				sizeof(resp));
879 	if (err)
880 		return err;
881 
882 	err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG,
883 				   sizeof(resp));
884 	if (err)
885 		return err;
886 
887 	if (resp.hdr.status)
888 		return -EPROTO;
889 
890 	*max_sq = resp.max_num_sq;
891 	*max_rq = resp.max_num_rq;
892 	*num_indir_entry = resp.num_indirection_ent;
893 
894 	apc->port_handle = resp.vport;
895 	ether_addr_copy(apc->mac_addr, resp.mac_addr);
896 
897 	return 0;
898 }
899 
900 void mana_uncfg_vport(struct mana_port_context *apc)
901 {
902 	mutex_lock(&apc->vport_mutex);
903 	apc->vport_use_count--;
904 	WARN_ON(apc->vport_use_count < 0);
905 	mutex_unlock(&apc->vport_mutex);
906 }
907 EXPORT_SYMBOL_NS(mana_uncfg_vport, NET_MANA);
908 
909 int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id,
910 		   u32 doorbell_pg_id)
911 {
912 	struct mana_config_vport_resp resp = {};
913 	struct mana_config_vport_req req = {};
914 	int err;
915 
916 	/* This function is used to program the Ethernet port in the hardware
917 	 * table. It can be called from the Ethernet driver or the RDMA driver.
918 	 *
919 	 * For Ethernet usage, the hardware supports only one active user on a
920 	 * physical port. The driver checks on the port usage before programming
921 	 * the hardware when creating the RAW QP (RDMA driver) or exposing the
922 	 * device to kernel NET layer (Ethernet driver).
923 	 *
924 	 * Because the RDMA driver doesn't know in advance which QP type the
925 	 * user will create, it exposes the device with all its ports. The user
926 	 * may not be able to create RAW QP on a port if this port is already
927 	 * in used by the Ethernet driver from the kernel.
928 	 *
929 	 * This physical port limitation only applies to the RAW QP. For RC QP,
930 	 * the hardware doesn't have this limitation. The user can create RC
931 	 * QPs on a physical port up to the hardware limits independent of the
932 	 * Ethernet usage on the same port.
933 	 */
934 	mutex_lock(&apc->vport_mutex);
935 	if (apc->vport_use_count > 0) {
936 		mutex_unlock(&apc->vport_mutex);
937 		return -EBUSY;
938 	}
939 	apc->vport_use_count++;
940 	mutex_unlock(&apc->vport_mutex);
941 
942 	mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX,
943 			     sizeof(req), sizeof(resp));
944 	req.vport = apc->port_handle;
945 	req.pdid = protection_dom_id;
946 	req.doorbell_pageid = doorbell_pg_id;
947 
948 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
949 				sizeof(resp));
950 	if (err) {
951 		netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err);
952 		goto out;
953 	}
954 
955 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX,
956 				   sizeof(resp));
957 	if (err || resp.hdr.status) {
958 		netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n",
959 			   err, resp.hdr.status);
960 		if (!err)
961 			err = -EPROTO;
962 
963 		goto out;
964 	}
965 
966 	apc->tx_shortform_allowed = resp.short_form_allowed;
967 	apc->tx_vp_offset = resp.tx_vport_offset;
968 
969 	netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n",
970 		    apc->port_handle, protection_dom_id, doorbell_pg_id);
971 out:
972 	if (err)
973 		mana_uncfg_vport(apc);
974 
975 	return err;
976 }
977 EXPORT_SYMBOL_NS(mana_cfg_vport, NET_MANA);
978 
979 static int mana_cfg_vport_steering(struct mana_port_context *apc,
980 				   enum TRI_STATE rx,
981 				   bool update_default_rxobj, bool update_key,
982 				   bool update_tab)
983 {
984 	u16 num_entries = MANA_INDIRECT_TABLE_SIZE;
985 	struct mana_cfg_rx_steer_req_v2 *req;
986 	struct mana_cfg_rx_steer_resp resp = {};
987 	struct net_device *ndev = apc->ndev;
988 	mana_handle_t *req_indir_tab;
989 	u32 req_buf_size;
990 	int err;
991 
992 	req_buf_size = sizeof(*req) + sizeof(mana_handle_t) * num_entries;
993 	req = kzalloc(req_buf_size, GFP_KERNEL);
994 	if (!req)
995 		return -ENOMEM;
996 
997 	mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
998 			     sizeof(resp));
999 
1000 	req->hdr.req.msg_version = GDMA_MESSAGE_V2;
1001 
1002 	req->vport = apc->port_handle;
1003 	req->num_indir_entries = num_entries;
1004 	req->indir_tab_offset = sizeof(*req);
1005 	req->rx_enable = rx;
1006 	req->rss_enable = apc->rss_state;
1007 	req->update_default_rxobj = update_default_rxobj;
1008 	req->update_hashkey = update_key;
1009 	req->update_indir_tab = update_tab;
1010 	req->default_rxobj = apc->default_rxobj;
1011 	req->cqe_coalescing_enable = 0;
1012 
1013 	if (update_key)
1014 		memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE);
1015 
1016 	if (update_tab) {
1017 		req_indir_tab = (mana_handle_t *)(req + 1);
1018 		memcpy(req_indir_tab, apc->rxobj_table,
1019 		       req->num_indir_entries * sizeof(mana_handle_t));
1020 	}
1021 
1022 	err = mana_send_request(apc->ac, req, req_buf_size, &resp,
1023 				sizeof(resp));
1024 	if (err) {
1025 		netdev_err(ndev, "Failed to configure vPort RX: %d\n", err);
1026 		goto out;
1027 	}
1028 
1029 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX,
1030 				   sizeof(resp));
1031 	if (err) {
1032 		netdev_err(ndev, "vPort RX configuration failed: %d\n", err);
1033 		goto out;
1034 	}
1035 
1036 	if (resp.hdr.status) {
1037 		netdev_err(ndev, "vPort RX configuration failed: 0x%x\n",
1038 			   resp.hdr.status);
1039 		err = -EPROTO;
1040 	}
1041 
1042 	netdev_info(ndev, "Configured steering vPort %llu entries %u\n",
1043 		    apc->port_handle, num_entries);
1044 out:
1045 	kfree(req);
1046 	return err;
1047 }
1048 
1049 int mana_create_wq_obj(struct mana_port_context *apc,
1050 		       mana_handle_t vport,
1051 		       u32 wq_type, struct mana_obj_spec *wq_spec,
1052 		       struct mana_obj_spec *cq_spec,
1053 		       mana_handle_t *wq_obj)
1054 {
1055 	struct mana_create_wqobj_resp resp = {};
1056 	struct mana_create_wqobj_req req = {};
1057 	struct net_device *ndev = apc->ndev;
1058 	int err;
1059 
1060 	mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ,
1061 			     sizeof(req), sizeof(resp));
1062 	req.vport = vport;
1063 	req.wq_type = wq_type;
1064 	req.wq_gdma_region = wq_spec->gdma_region;
1065 	req.cq_gdma_region = cq_spec->gdma_region;
1066 	req.wq_size = wq_spec->queue_size;
1067 	req.cq_size = cq_spec->queue_size;
1068 	req.cq_moderation_ctx_id = cq_spec->modr_ctx_id;
1069 	req.cq_parent_qid = cq_spec->attached_eq;
1070 
1071 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1072 				sizeof(resp));
1073 	if (err) {
1074 		netdev_err(ndev, "Failed to create WQ object: %d\n", err);
1075 		goto out;
1076 	}
1077 
1078 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ,
1079 				   sizeof(resp));
1080 	if (err || resp.hdr.status) {
1081 		netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err,
1082 			   resp.hdr.status);
1083 		if (!err)
1084 			err = -EPROTO;
1085 		goto out;
1086 	}
1087 
1088 	if (resp.wq_obj == INVALID_MANA_HANDLE) {
1089 		netdev_err(ndev, "Got an invalid WQ object handle\n");
1090 		err = -EPROTO;
1091 		goto out;
1092 	}
1093 
1094 	*wq_obj = resp.wq_obj;
1095 	wq_spec->queue_index = resp.wq_id;
1096 	cq_spec->queue_index = resp.cq_id;
1097 
1098 	return 0;
1099 out:
1100 	return err;
1101 }
1102 EXPORT_SYMBOL_NS(mana_create_wq_obj, NET_MANA);
1103 
1104 void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type,
1105 			 mana_handle_t wq_obj)
1106 {
1107 	struct mana_destroy_wqobj_resp resp = {};
1108 	struct mana_destroy_wqobj_req req = {};
1109 	struct net_device *ndev = apc->ndev;
1110 	int err;
1111 
1112 	mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ,
1113 			     sizeof(req), sizeof(resp));
1114 	req.wq_type = wq_type;
1115 	req.wq_obj_handle = wq_obj;
1116 
1117 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1118 				sizeof(resp));
1119 	if (err) {
1120 		netdev_err(ndev, "Failed to destroy WQ object: %d\n", err);
1121 		return;
1122 	}
1123 
1124 	err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ,
1125 				   sizeof(resp));
1126 	if (err || resp.hdr.status)
1127 		netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err,
1128 			   resp.hdr.status);
1129 }
1130 EXPORT_SYMBOL_NS(mana_destroy_wq_obj, NET_MANA);
1131 
1132 static void mana_destroy_eq(struct mana_context *ac)
1133 {
1134 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
1135 	struct gdma_queue *eq;
1136 	int i;
1137 
1138 	if (!ac->eqs)
1139 		return;
1140 
1141 	for (i = 0; i < gc->max_num_queues; i++) {
1142 		eq = ac->eqs[i].eq;
1143 		if (!eq)
1144 			continue;
1145 
1146 		mana_gd_destroy_queue(gc, eq);
1147 	}
1148 
1149 	kfree(ac->eqs);
1150 	ac->eqs = NULL;
1151 }
1152 
1153 static int mana_create_eq(struct mana_context *ac)
1154 {
1155 	struct gdma_dev *gd = ac->gdma_dev;
1156 	struct gdma_context *gc = gd->gdma_context;
1157 	struct gdma_queue_spec spec = {};
1158 	int err;
1159 	int i;
1160 
1161 	ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq),
1162 			  GFP_KERNEL);
1163 	if (!ac->eqs)
1164 		return -ENOMEM;
1165 
1166 	spec.type = GDMA_EQ;
1167 	spec.monitor_avl_buf = false;
1168 	spec.queue_size = EQ_SIZE;
1169 	spec.eq.callback = NULL;
1170 	spec.eq.context = ac->eqs;
1171 	spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE;
1172 
1173 	for (i = 0; i < gc->max_num_queues; i++) {
1174 		err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq);
1175 		if (err)
1176 			goto out;
1177 	}
1178 
1179 	return 0;
1180 out:
1181 	mana_destroy_eq(ac);
1182 	return err;
1183 }
1184 
1185 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq)
1186 {
1187 	struct mana_fence_rq_resp resp = {};
1188 	struct mana_fence_rq_req req = {};
1189 	int err;
1190 
1191 	init_completion(&rxq->fence_event);
1192 
1193 	mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ,
1194 			     sizeof(req), sizeof(resp));
1195 	req.wq_obj_handle =  rxq->rxobj;
1196 
1197 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1198 				sizeof(resp));
1199 	if (err) {
1200 		netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n",
1201 			   rxq->rxq_idx, err);
1202 		return err;
1203 	}
1204 
1205 	err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp));
1206 	if (err || resp.hdr.status) {
1207 		netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n",
1208 			   rxq->rxq_idx, err, resp.hdr.status);
1209 		if (!err)
1210 			err = -EPROTO;
1211 
1212 		return err;
1213 	}
1214 
1215 	if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) {
1216 		netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n",
1217 			   rxq->rxq_idx);
1218 		return -ETIMEDOUT;
1219 	}
1220 
1221 	return 0;
1222 }
1223 
1224 static void mana_fence_rqs(struct mana_port_context *apc)
1225 {
1226 	unsigned int rxq_idx;
1227 	struct mana_rxq *rxq;
1228 	int err;
1229 
1230 	for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
1231 		rxq = apc->rxqs[rxq_idx];
1232 		err = mana_fence_rq(apc, rxq);
1233 
1234 		/* In case of any error, use sleep instead. */
1235 		if (err)
1236 			msleep(100);
1237 	}
1238 }
1239 
1240 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units)
1241 {
1242 	u32 used_space_old;
1243 	u32 used_space_new;
1244 
1245 	used_space_old = wq->head - wq->tail;
1246 	used_space_new = wq->head - (wq->tail + num_units);
1247 
1248 	if (WARN_ON_ONCE(used_space_new > used_space_old))
1249 		return -ERANGE;
1250 
1251 	wq->tail += num_units;
1252 	return 0;
1253 }
1254 
1255 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc)
1256 {
1257 	struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
1258 	struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1259 	struct device *dev = gc->dev;
1260 	int i;
1261 
1262 	dma_unmap_single(dev, ash->dma_handle[0], ash->size[0], DMA_TO_DEVICE);
1263 
1264 	for (i = 1; i < skb_shinfo(skb)->nr_frags + 1; i++)
1265 		dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
1266 			       DMA_TO_DEVICE);
1267 }
1268 
1269 static void mana_poll_tx_cq(struct mana_cq *cq)
1270 {
1271 	struct gdma_comp *completions = cq->gdma_comp_buf;
1272 	struct gdma_posted_wqe_info *wqe_info;
1273 	unsigned int pkt_transmitted = 0;
1274 	unsigned int wqe_unit_cnt = 0;
1275 	struct mana_txq *txq = cq->txq;
1276 	struct mana_port_context *apc;
1277 	struct netdev_queue *net_txq;
1278 	struct gdma_queue *gdma_wq;
1279 	unsigned int avail_space;
1280 	struct net_device *ndev;
1281 	struct sk_buff *skb;
1282 	bool txq_stopped;
1283 	int comp_read;
1284 	int i;
1285 
1286 	ndev = txq->ndev;
1287 	apc = netdev_priv(ndev);
1288 
1289 	comp_read = mana_gd_poll_cq(cq->gdma_cq, completions,
1290 				    CQE_POLLING_BUFFER);
1291 
1292 	if (comp_read < 1)
1293 		return;
1294 
1295 	for (i = 0; i < comp_read; i++) {
1296 		struct mana_tx_comp_oob *cqe_oob;
1297 
1298 		if (WARN_ON_ONCE(!completions[i].is_sq))
1299 			return;
1300 
1301 		cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data;
1302 		if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type !=
1303 				 MANA_CQE_COMPLETION))
1304 			return;
1305 
1306 		switch (cqe_oob->cqe_hdr.cqe_type) {
1307 		case CQE_TX_OKAY:
1308 			break;
1309 
1310 		case CQE_TX_SA_DROP:
1311 		case CQE_TX_MTU_DROP:
1312 		case CQE_TX_INVALID_OOB:
1313 		case CQE_TX_INVALID_ETH_TYPE:
1314 		case CQE_TX_HDR_PROCESSING_ERROR:
1315 		case CQE_TX_VF_DISABLED:
1316 		case CQE_TX_VPORT_IDX_OUT_OF_RANGE:
1317 		case CQE_TX_VPORT_DISABLED:
1318 		case CQE_TX_VLAN_TAGGING_VIOLATION:
1319 			WARN_ONCE(1, "TX: CQE error %d: ignored.\n",
1320 				  cqe_oob->cqe_hdr.cqe_type);
1321 			apc->eth_stats.tx_cqe_err++;
1322 			break;
1323 
1324 		default:
1325 			/* If the CQE type is unexpected, log an error, assert,
1326 			 * and go through the error path.
1327 			 */
1328 			WARN_ONCE(1, "TX: Unexpected CQE type %d: HW BUG?\n",
1329 				  cqe_oob->cqe_hdr.cqe_type);
1330 			apc->eth_stats.tx_cqe_unknown_type++;
1331 			return;
1332 		}
1333 
1334 		if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num))
1335 			return;
1336 
1337 		skb = skb_dequeue(&txq->pending_skbs);
1338 		if (WARN_ON_ONCE(!skb))
1339 			return;
1340 
1341 		wqe_info = (struct gdma_posted_wqe_info *)skb->cb;
1342 		wqe_unit_cnt += wqe_info->wqe_size_in_bu;
1343 
1344 		mana_unmap_skb(skb, apc);
1345 
1346 		napi_consume_skb(skb, cq->budget);
1347 
1348 		pkt_transmitted++;
1349 	}
1350 
1351 	if (WARN_ON_ONCE(wqe_unit_cnt == 0))
1352 		return;
1353 
1354 	mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt);
1355 
1356 	gdma_wq = txq->gdma_sq;
1357 	avail_space = mana_gd_wq_avail_space(gdma_wq);
1358 
1359 	/* Ensure tail updated before checking q stop */
1360 	smp_mb();
1361 
1362 	net_txq = txq->net_txq;
1363 	txq_stopped = netif_tx_queue_stopped(net_txq);
1364 
1365 	/* Ensure checking txq_stopped before apc->port_is_up. */
1366 	smp_rmb();
1367 
1368 	if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) {
1369 		netif_tx_wake_queue(net_txq);
1370 		apc->eth_stats.wake_queue++;
1371 	}
1372 
1373 	if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0)
1374 		WARN_ON_ONCE(1);
1375 
1376 	cq->work_done = pkt_transmitted;
1377 }
1378 
1379 static void mana_post_pkt_rxq(struct mana_rxq *rxq)
1380 {
1381 	struct mana_recv_buf_oob *recv_buf_oob;
1382 	u32 curr_index;
1383 	int err;
1384 
1385 	curr_index = rxq->buf_index++;
1386 	if (rxq->buf_index == rxq->num_rx_buf)
1387 		rxq->buf_index = 0;
1388 
1389 	recv_buf_oob = &rxq->rx_oobs[curr_index];
1390 
1391 	err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req,
1392 					&recv_buf_oob->wqe_inf);
1393 	if (WARN_ON_ONCE(err))
1394 		return;
1395 
1396 	WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1);
1397 }
1398 
1399 static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va,
1400 				      uint pkt_len, struct xdp_buff *xdp)
1401 {
1402 	struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size);
1403 
1404 	if (!skb)
1405 		return NULL;
1406 
1407 	if (xdp->data_hard_start) {
1408 		skb_reserve(skb, xdp->data - xdp->data_hard_start);
1409 		skb_put(skb, xdp->data_end - xdp->data);
1410 		return skb;
1411 	}
1412 
1413 	skb_reserve(skb, rxq->headroom);
1414 	skb_put(skb, pkt_len);
1415 
1416 	return skb;
1417 }
1418 
1419 static void mana_rx_skb(void *buf_va, bool from_pool,
1420 			struct mana_rxcomp_oob *cqe, struct mana_rxq *rxq)
1421 {
1422 	struct mana_stats_rx *rx_stats = &rxq->stats;
1423 	struct net_device *ndev = rxq->ndev;
1424 	uint pkt_len = cqe->ppi[0].pkt_len;
1425 	u16 rxq_idx = rxq->rxq_idx;
1426 	struct napi_struct *napi;
1427 	struct xdp_buff xdp = {};
1428 	struct sk_buff *skb;
1429 	u32 hash_value;
1430 	u32 act;
1431 
1432 	rxq->rx_cq.work_done++;
1433 	napi = &rxq->rx_cq.napi;
1434 
1435 	if (!buf_va) {
1436 		++ndev->stats.rx_dropped;
1437 		return;
1438 	}
1439 
1440 	act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len);
1441 
1442 	if (act == XDP_REDIRECT && !rxq->xdp_rc)
1443 		return;
1444 
1445 	if (act != XDP_PASS && act != XDP_TX)
1446 		goto drop_xdp;
1447 
1448 	skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp);
1449 
1450 	if (!skb)
1451 		goto drop;
1452 
1453 	if (from_pool)
1454 		skb_mark_for_recycle(skb);
1455 
1456 	skb->dev = napi->dev;
1457 
1458 	skb->protocol = eth_type_trans(skb, ndev);
1459 	skb_checksum_none_assert(skb);
1460 	skb_record_rx_queue(skb, rxq_idx);
1461 
1462 	if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) {
1463 		if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed)
1464 			skb->ip_summed = CHECKSUM_UNNECESSARY;
1465 	}
1466 
1467 	if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) {
1468 		hash_value = cqe->ppi[0].pkt_hash;
1469 
1470 		if (cqe->rx_hashtype & MANA_HASH_L4)
1471 			skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4);
1472 		else
1473 			skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3);
1474 	}
1475 
1476 	if (cqe->rx_vlantag_present) {
1477 		u16 vlan_tci = cqe->rx_vlan_id;
1478 
1479 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci);
1480 	}
1481 
1482 	u64_stats_update_begin(&rx_stats->syncp);
1483 	rx_stats->packets++;
1484 	rx_stats->bytes += pkt_len;
1485 
1486 	if (act == XDP_TX)
1487 		rx_stats->xdp_tx++;
1488 	u64_stats_update_end(&rx_stats->syncp);
1489 
1490 	if (act == XDP_TX) {
1491 		skb_set_queue_mapping(skb, rxq_idx);
1492 		mana_xdp_tx(skb, ndev);
1493 		return;
1494 	}
1495 
1496 	napi_gro_receive(napi, skb);
1497 
1498 	return;
1499 
1500 drop_xdp:
1501 	u64_stats_update_begin(&rx_stats->syncp);
1502 	rx_stats->xdp_drop++;
1503 	u64_stats_update_end(&rx_stats->syncp);
1504 
1505 drop:
1506 	if (from_pool) {
1507 		page_pool_recycle_direct(rxq->page_pool,
1508 					 virt_to_head_page(buf_va));
1509 	} else {
1510 		WARN_ON_ONCE(rxq->xdp_save_va);
1511 		/* Save for reuse */
1512 		rxq->xdp_save_va = buf_va;
1513 	}
1514 
1515 	++ndev->stats.rx_dropped;
1516 
1517 	return;
1518 }
1519 
1520 static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,
1521 			     dma_addr_t *da, bool *from_pool, bool is_napi)
1522 {
1523 	struct page *page;
1524 	void *va;
1525 
1526 	*from_pool = false;
1527 
1528 	/* Reuse XDP dropped page if available */
1529 	if (rxq->xdp_save_va) {
1530 		va = rxq->xdp_save_va;
1531 		rxq->xdp_save_va = NULL;
1532 	} else if (rxq->alloc_size > PAGE_SIZE) {
1533 		if (is_napi)
1534 			va = napi_alloc_frag(rxq->alloc_size);
1535 		else
1536 			va = netdev_alloc_frag(rxq->alloc_size);
1537 
1538 		if (!va)
1539 			return NULL;
1540 
1541 		page = virt_to_head_page(va);
1542 		/* Check if the frag falls back to single page */
1543 		if (compound_order(page) < get_order(rxq->alloc_size)) {
1544 			put_page(page);
1545 			return NULL;
1546 		}
1547 	} else {
1548 		page = page_pool_dev_alloc_pages(rxq->page_pool);
1549 		if (!page)
1550 			return NULL;
1551 
1552 		*from_pool = true;
1553 		va = page_to_virt(page);
1554 	}
1555 
1556 	*da = dma_map_single(dev, va + rxq->headroom, rxq->datasize,
1557 			     DMA_FROM_DEVICE);
1558 	if (dma_mapping_error(dev, *da)) {
1559 		if (*from_pool)
1560 			page_pool_put_full_page(rxq->page_pool, page, false);
1561 		else
1562 			put_page(virt_to_head_page(va));
1563 
1564 		return NULL;
1565 	}
1566 
1567 	return va;
1568 }
1569 
1570 /* Allocate frag for rx buffer, and save the old buf */
1571 static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq,
1572 			       struct mana_recv_buf_oob *rxoob, void **old_buf,
1573 			       bool *old_fp)
1574 {
1575 	bool from_pool;
1576 	dma_addr_t da;
1577 	void *va;
1578 
1579 	va = mana_get_rxfrag(rxq, dev, &da, &from_pool, true);
1580 	if (!va)
1581 		return;
1582 
1583 	dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize,
1584 			 DMA_FROM_DEVICE);
1585 	*old_buf = rxoob->buf_va;
1586 	*old_fp = rxoob->from_pool;
1587 
1588 	rxoob->buf_va = va;
1589 	rxoob->sgl[0].address = da;
1590 	rxoob->from_pool = from_pool;
1591 }
1592 
1593 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,
1594 				struct gdma_comp *cqe)
1595 {
1596 	struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data;
1597 	struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1598 	struct net_device *ndev = rxq->ndev;
1599 	struct mana_recv_buf_oob *rxbuf_oob;
1600 	struct mana_port_context *apc;
1601 	struct device *dev = gc->dev;
1602 	void *old_buf = NULL;
1603 	u32 curr, pktlen;
1604 	bool old_fp;
1605 
1606 	apc = netdev_priv(ndev);
1607 
1608 	switch (oob->cqe_hdr.cqe_type) {
1609 	case CQE_RX_OKAY:
1610 		break;
1611 
1612 	case CQE_RX_TRUNCATED:
1613 		++ndev->stats.rx_dropped;
1614 		rxbuf_oob = &rxq->rx_oobs[rxq->buf_index];
1615 		netdev_warn_once(ndev, "Dropped a truncated packet\n");
1616 		goto drop;
1617 
1618 	case CQE_RX_COALESCED_4:
1619 		netdev_err(ndev, "RX coalescing is unsupported\n");
1620 		apc->eth_stats.rx_coalesced_err++;
1621 		return;
1622 
1623 	case CQE_RX_OBJECT_FENCE:
1624 		complete(&rxq->fence_event);
1625 		return;
1626 
1627 	default:
1628 		netdev_err(ndev, "Unknown RX CQE type = %d\n",
1629 			   oob->cqe_hdr.cqe_type);
1630 		apc->eth_stats.rx_cqe_unknown_type++;
1631 		return;
1632 	}
1633 
1634 	pktlen = oob->ppi[0].pkt_len;
1635 
1636 	if (pktlen == 0) {
1637 		/* data packets should never have packetlength of zero */
1638 		netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n",
1639 			   rxq->gdma_id, cq->gdma_id, rxq->rxobj);
1640 		return;
1641 	}
1642 
1643 	curr = rxq->buf_index;
1644 	rxbuf_oob = &rxq->rx_oobs[curr];
1645 	WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1);
1646 
1647 	mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf, &old_fp);
1648 
1649 	/* Unsuccessful refill will have old_buf == NULL.
1650 	 * In this case, mana_rx_skb() will drop the packet.
1651 	 */
1652 	mana_rx_skb(old_buf, old_fp, oob, rxq);
1653 
1654 drop:
1655 	mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu);
1656 
1657 	mana_post_pkt_rxq(rxq);
1658 }
1659 
1660 static void mana_poll_rx_cq(struct mana_cq *cq)
1661 {
1662 	struct gdma_comp *comp = cq->gdma_comp_buf;
1663 	struct mana_rxq *rxq = cq->rxq;
1664 	int comp_read, i;
1665 
1666 	comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER);
1667 	WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER);
1668 
1669 	rxq->xdp_flush = false;
1670 
1671 	for (i = 0; i < comp_read; i++) {
1672 		if (WARN_ON_ONCE(comp[i].is_sq))
1673 			return;
1674 
1675 		/* verify recv cqe references the right rxq */
1676 		if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id))
1677 			return;
1678 
1679 		mana_process_rx_cqe(rxq, cq, &comp[i]);
1680 	}
1681 
1682 	if (comp_read > 0) {
1683 		struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1684 
1685 		mana_gd_wq_ring_doorbell(gc, rxq->gdma_rq);
1686 	}
1687 
1688 	if (rxq->xdp_flush)
1689 		xdp_do_flush();
1690 }
1691 
1692 static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue)
1693 {
1694 	struct mana_cq *cq = context;
1695 	u8 arm_bit;
1696 	int w;
1697 
1698 	WARN_ON_ONCE(cq->gdma_cq != gdma_queue);
1699 
1700 	if (cq->type == MANA_CQ_TYPE_RX)
1701 		mana_poll_rx_cq(cq);
1702 	else
1703 		mana_poll_tx_cq(cq);
1704 
1705 	w = cq->work_done;
1706 
1707 	if (w < cq->budget &&
1708 	    napi_complete_done(&cq->napi, w)) {
1709 		arm_bit = SET_ARM_BIT;
1710 	} else {
1711 		arm_bit = 0;
1712 	}
1713 
1714 	mana_gd_ring_cq(gdma_queue, arm_bit);
1715 
1716 	return w;
1717 }
1718 
1719 static int mana_poll(struct napi_struct *napi, int budget)
1720 {
1721 	struct mana_cq *cq = container_of(napi, struct mana_cq, napi);
1722 	int w;
1723 
1724 	cq->work_done = 0;
1725 	cq->budget = budget;
1726 
1727 	w = mana_cq_handler(cq, cq->gdma_cq);
1728 
1729 	return min(w, budget);
1730 }
1731 
1732 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue)
1733 {
1734 	struct mana_cq *cq = context;
1735 
1736 	napi_schedule_irqoff(&cq->napi);
1737 }
1738 
1739 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq)
1740 {
1741 	struct gdma_dev *gd = apc->ac->gdma_dev;
1742 
1743 	if (!cq->gdma_cq)
1744 		return;
1745 
1746 	mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq);
1747 }
1748 
1749 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq)
1750 {
1751 	struct gdma_dev *gd = apc->ac->gdma_dev;
1752 
1753 	if (!txq->gdma_sq)
1754 		return;
1755 
1756 	mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq);
1757 }
1758 
1759 static void mana_destroy_txq(struct mana_port_context *apc)
1760 {
1761 	struct napi_struct *napi;
1762 	int i;
1763 
1764 	if (!apc->tx_qp)
1765 		return;
1766 
1767 	for (i = 0; i < apc->num_queues; i++) {
1768 		napi = &apc->tx_qp[i].tx_cq.napi;
1769 		napi_synchronize(napi);
1770 		napi_disable(napi);
1771 		netif_napi_del(napi);
1772 
1773 		mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object);
1774 
1775 		mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq);
1776 
1777 		mana_deinit_txq(apc, &apc->tx_qp[i].txq);
1778 	}
1779 
1780 	kfree(apc->tx_qp);
1781 	apc->tx_qp = NULL;
1782 }
1783 
1784 static int mana_create_txq(struct mana_port_context *apc,
1785 			   struct net_device *net)
1786 {
1787 	struct mana_context *ac = apc->ac;
1788 	struct gdma_dev *gd = ac->gdma_dev;
1789 	struct mana_obj_spec wq_spec;
1790 	struct mana_obj_spec cq_spec;
1791 	struct gdma_queue_spec spec;
1792 	struct gdma_context *gc;
1793 	struct mana_txq *txq;
1794 	struct mana_cq *cq;
1795 	u32 txq_size;
1796 	u32 cq_size;
1797 	int err;
1798 	int i;
1799 
1800 	apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp),
1801 			     GFP_KERNEL);
1802 	if (!apc->tx_qp)
1803 		return -ENOMEM;
1804 
1805 	/*  The minimum size of the WQE is 32 bytes, hence
1806 	 *  MAX_SEND_BUFFERS_PER_QUEUE represents the maximum number of WQEs
1807 	 *  the SQ can store. This value is then used to size other queues
1808 	 *  to prevent overflow.
1809 	 */
1810 	txq_size = MAX_SEND_BUFFERS_PER_QUEUE * 32;
1811 	BUILD_BUG_ON(!PAGE_ALIGNED(txq_size));
1812 
1813 	cq_size = MAX_SEND_BUFFERS_PER_QUEUE * COMP_ENTRY_SIZE;
1814 	cq_size = PAGE_ALIGN(cq_size);
1815 
1816 	gc = gd->gdma_context;
1817 
1818 	for (i = 0; i < apc->num_queues; i++) {
1819 		apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE;
1820 
1821 		/* Create SQ */
1822 		txq = &apc->tx_qp[i].txq;
1823 
1824 		u64_stats_init(&txq->stats.syncp);
1825 		txq->ndev = net;
1826 		txq->net_txq = netdev_get_tx_queue(net, i);
1827 		txq->vp_offset = apc->tx_vp_offset;
1828 		skb_queue_head_init(&txq->pending_skbs);
1829 
1830 		memset(&spec, 0, sizeof(spec));
1831 		spec.type = GDMA_SQ;
1832 		spec.monitor_avl_buf = true;
1833 		spec.queue_size = txq_size;
1834 		err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq);
1835 		if (err)
1836 			goto out;
1837 
1838 		/* Create SQ's CQ */
1839 		cq = &apc->tx_qp[i].tx_cq;
1840 		cq->type = MANA_CQ_TYPE_TX;
1841 
1842 		cq->txq = txq;
1843 
1844 		memset(&spec, 0, sizeof(spec));
1845 		spec.type = GDMA_CQ;
1846 		spec.monitor_avl_buf = false;
1847 		spec.queue_size = cq_size;
1848 		spec.cq.callback = mana_schedule_napi;
1849 		spec.cq.parent_eq = ac->eqs[i].eq;
1850 		spec.cq.context = cq;
1851 		err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
1852 		if (err)
1853 			goto out;
1854 
1855 		memset(&wq_spec, 0, sizeof(wq_spec));
1856 		memset(&cq_spec, 0, sizeof(cq_spec));
1857 
1858 		wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle;
1859 		wq_spec.queue_size = txq->gdma_sq->queue_size;
1860 
1861 		cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
1862 		cq_spec.queue_size = cq->gdma_cq->queue_size;
1863 		cq_spec.modr_ctx_id = 0;
1864 		cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
1865 
1866 		err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ,
1867 					 &wq_spec, &cq_spec,
1868 					 &apc->tx_qp[i].tx_object);
1869 
1870 		if (err)
1871 			goto out;
1872 
1873 		txq->gdma_sq->id = wq_spec.queue_index;
1874 		cq->gdma_cq->id = cq_spec.queue_index;
1875 
1876 		txq->gdma_sq->mem_info.dma_region_handle =
1877 			GDMA_INVALID_DMA_REGION;
1878 		cq->gdma_cq->mem_info.dma_region_handle =
1879 			GDMA_INVALID_DMA_REGION;
1880 
1881 		txq->gdma_txq_id = txq->gdma_sq->id;
1882 
1883 		cq->gdma_id = cq->gdma_cq->id;
1884 
1885 		if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
1886 			err = -EINVAL;
1887 			goto out;
1888 		}
1889 
1890 		gc->cq_table[cq->gdma_id] = cq->gdma_cq;
1891 
1892 		netif_napi_add_tx(net, &cq->napi, mana_poll);
1893 		napi_enable(&cq->napi);
1894 
1895 		mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
1896 	}
1897 
1898 	return 0;
1899 out:
1900 	mana_destroy_txq(apc);
1901 	return err;
1902 }
1903 
1904 static void mana_destroy_rxq(struct mana_port_context *apc,
1905 			     struct mana_rxq *rxq, bool validate_state)
1906 
1907 {
1908 	struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1909 	struct mana_recv_buf_oob *rx_oob;
1910 	struct device *dev = gc->dev;
1911 	struct napi_struct *napi;
1912 	struct page *page;
1913 	int i;
1914 
1915 	if (!rxq)
1916 		return;
1917 
1918 	napi = &rxq->rx_cq.napi;
1919 
1920 	if (validate_state)
1921 		napi_synchronize(napi);
1922 
1923 	napi_disable(napi);
1924 
1925 	xdp_rxq_info_unreg(&rxq->xdp_rxq);
1926 
1927 	netif_napi_del(napi);
1928 
1929 	mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj);
1930 
1931 	mana_deinit_cq(apc, &rxq->rx_cq);
1932 
1933 	if (rxq->xdp_save_va)
1934 		put_page(virt_to_head_page(rxq->xdp_save_va));
1935 
1936 	for (i = 0; i < rxq->num_rx_buf; i++) {
1937 		rx_oob = &rxq->rx_oobs[i];
1938 
1939 		if (!rx_oob->buf_va)
1940 			continue;
1941 
1942 		dma_unmap_single(dev, rx_oob->sgl[0].address,
1943 				 rx_oob->sgl[0].size, DMA_FROM_DEVICE);
1944 
1945 		page = virt_to_head_page(rx_oob->buf_va);
1946 
1947 		if (rx_oob->from_pool)
1948 			page_pool_put_full_page(rxq->page_pool, page, false);
1949 		else
1950 			put_page(page);
1951 
1952 		rx_oob->buf_va = NULL;
1953 	}
1954 
1955 	page_pool_destroy(rxq->page_pool);
1956 
1957 	if (rxq->gdma_rq)
1958 		mana_gd_destroy_queue(gc, rxq->gdma_rq);
1959 
1960 	kfree(rxq);
1961 }
1962 
1963 static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key,
1964 			    struct mana_rxq *rxq, struct device *dev)
1965 {
1966 	struct mana_port_context *mpc = netdev_priv(rxq->ndev);
1967 	bool from_pool = false;
1968 	dma_addr_t da;
1969 	void *va;
1970 
1971 	if (mpc->rxbufs_pre)
1972 		va = mana_get_rxbuf_pre(rxq, &da);
1973 	else
1974 		va = mana_get_rxfrag(rxq, dev, &da, &from_pool, false);
1975 
1976 	if (!va)
1977 		return -ENOMEM;
1978 
1979 	rx_oob->buf_va = va;
1980 	rx_oob->from_pool = from_pool;
1981 
1982 	rx_oob->sgl[0].address = da;
1983 	rx_oob->sgl[0].size = rxq->datasize;
1984 	rx_oob->sgl[0].mem_key = mem_key;
1985 
1986 	return 0;
1987 }
1988 
1989 #define MANA_WQE_HEADER_SIZE 16
1990 #define MANA_WQE_SGE_SIZE 16
1991 
1992 static int mana_alloc_rx_wqe(struct mana_port_context *apc,
1993 			     struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size)
1994 {
1995 	struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1996 	struct mana_recv_buf_oob *rx_oob;
1997 	struct device *dev = gc->dev;
1998 	u32 buf_idx;
1999 	int ret;
2000 
2001 	WARN_ON(rxq->datasize == 0);
2002 
2003 	*rxq_size = 0;
2004 	*cq_size = 0;
2005 
2006 	for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2007 		rx_oob = &rxq->rx_oobs[buf_idx];
2008 		memset(rx_oob, 0, sizeof(*rx_oob));
2009 
2010 		rx_oob->num_sge = 1;
2011 
2012 		ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq,
2013 				       dev);
2014 		if (ret)
2015 			return ret;
2016 
2017 		rx_oob->wqe_req.sgl = rx_oob->sgl;
2018 		rx_oob->wqe_req.num_sge = rx_oob->num_sge;
2019 		rx_oob->wqe_req.inline_oob_size = 0;
2020 		rx_oob->wqe_req.inline_oob_data = NULL;
2021 		rx_oob->wqe_req.flags = 0;
2022 		rx_oob->wqe_req.client_data_unit = 0;
2023 
2024 		*rxq_size += ALIGN(MANA_WQE_HEADER_SIZE +
2025 				   MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32);
2026 		*cq_size += COMP_ENTRY_SIZE;
2027 	}
2028 
2029 	return 0;
2030 }
2031 
2032 static int mana_push_wqe(struct mana_rxq *rxq)
2033 {
2034 	struct mana_recv_buf_oob *rx_oob;
2035 	u32 buf_idx;
2036 	int err;
2037 
2038 	for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2039 		rx_oob = &rxq->rx_oobs[buf_idx];
2040 
2041 		err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req,
2042 					    &rx_oob->wqe_inf);
2043 		if (err)
2044 			return -ENOSPC;
2045 	}
2046 
2047 	return 0;
2048 }
2049 
2050 static int mana_create_page_pool(struct mana_rxq *rxq, struct gdma_context *gc)
2051 {
2052 	struct page_pool_params pprm = {};
2053 	int ret;
2054 
2055 	pprm.pool_size = RX_BUFFERS_PER_QUEUE;
2056 	pprm.nid = gc->numa_node;
2057 	pprm.napi = &rxq->rx_cq.napi;
2058 
2059 	rxq->page_pool = page_pool_create(&pprm);
2060 
2061 	if (IS_ERR(rxq->page_pool)) {
2062 		ret = PTR_ERR(rxq->page_pool);
2063 		rxq->page_pool = NULL;
2064 		return ret;
2065 	}
2066 
2067 	return 0;
2068 }
2069 
2070 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,
2071 					u32 rxq_idx, struct mana_eq *eq,
2072 					struct net_device *ndev)
2073 {
2074 	struct gdma_dev *gd = apc->ac->gdma_dev;
2075 	struct mana_obj_spec wq_spec;
2076 	struct mana_obj_spec cq_spec;
2077 	struct gdma_queue_spec spec;
2078 	struct mana_cq *cq = NULL;
2079 	struct gdma_context *gc;
2080 	u32 cq_size, rq_size;
2081 	struct mana_rxq *rxq;
2082 	int err;
2083 
2084 	gc = gd->gdma_context;
2085 
2086 	rxq = kzalloc(struct_size(rxq, rx_oobs, RX_BUFFERS_PER_QUEUE),
2087 		      GFP_KERNEL);
2088 	if (!rxq)
2089 		return NULL;
2090 
2091 	rxq->ndev = ndev;
2092 	rxq->num_rx_buf = RX_BUFFERS_PER_QUEUE;
2093 	rxq->rxq_idx = rxq_idx;
2094 	rxq->rxobj = INVALID_MANA_HANDLE;
2095 
2096 	mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size,
2097 			   &rxq->headroom);
2098 
2099 	/* Create page pool for RX queue */
2100 	err = mana_create_page_pool(rxq, gc);
2101 	if (err) {
2102 		netdev_err(ndev, "Create page pool err:%d\n", err);
2103 		goto out;
2104 	}
2105 
2106 	err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size);
2107 	if (err)
2108 		goto out;
2109 
2110 	rq_size = PAGE_ALIGN(rq_size);
2111 	cq_size = PAGE_ALIGN(cq_size);
2112 
2113 	/* Create RQ */
2114 	memset(&spec, 0, sizeof(spec));
2115 	spec.type = GDMA_RQ;
2116 	spec.monitor_avl_buf = true;
2117 	spec.queue_size = rq_size;
2118 	err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq);
2119 	if (err)
2120 		goto out;
2121 
2122 	/* Create RQ's CQ */
2123 	cq = &rxq->rx_cq;
2124 	cq->type = MANA_CQ_TYPE_RX;
2125 	cq->rxq = rxq;
2126 
2127 	memset(&spec, 0, sizeof(spec));
2128 	spec.type = GDMA_CQ;
2129 	spec.monitor_avl_buf = false;
2130 	spec.queue_size = cq_size;
2131 	spec.cq.callback = mana_schedule_napi;
2132 	spec.cq.parent_eq = eq->eq;
2133 	spec.cq.context = cq;
2134 	err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2135 	if (err)
2136 		goto out;
2137 
2138 	memset(&wq_spec, 0, sizeof(wq_spec));
2139 	memset(&cq_spec, 0, sizeof(cq_spec));
2140 	wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle;
2141 	wq_spec.queue_size = rxq->gdma_rq->queue_size;
2142 
2143 	cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2144 	cq_spec.queue_size = cq->gdma_cq->queue_size;
2145 	cq_spec.modr_ctx_id = 0;
2146 	cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2147 
2148 	err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ,
2149 				 &wq_spec, &cq_spec, &rxq->rxobj);
2150 	if (err)
2151 		goto out;
2152 
2153 	rxq->gdma_rq->id = wq_spec.queue_index;
2154 	cq->gdma_cq->id = cq_spec.queue_index;
2155 
2156 	rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2157 	cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2158 
2159 	rxq->gdma_id = rxq->gdma_rq->id;
2160 	cq->gdma_id = cq->gdma_cq->id;
2161 
2162 	err = mana_push_wqe(rxq);
2163 	if (err)
2164 		goto out;
2165 
2166 	if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
2167 		err = -EINVAL;
2168 		goto out;
2169 	}
2170 
2171 	gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2172 
2173 	netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1);
2174 
2175 	WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx,
2176 				 cq->napi.napi_id));
2177 	WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL,
2178 					   rxq->page_pool));
2179 
2180 	napi_enable(&cq->napi);
2181 
2182 	mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2183 out:
2184 	if (!err)
2185 		return rxq;
2186 
2187 	netdev_err(ndev, "Failed to create RXQ: err = %d\n", err);
2188 
2189 	mana_destroy_rxq(apc, rxq, false);
2190 
2191 	if (cq)
2192 		mana_deinit_cq(apc, cq);
2193 
2194 	return NULL;
2195 }
2196 
2197 static int mana_add_rx_queues(struct mana_port_context *apc,
2198 			      struct net_device *ndev)
2199 {
2200 	struct mana_context *ac = apc->ac;
2201 	struct mana_rxq *rxq;
2202 	int err = 0;
2203 	int i;
2204 
2205 	for (i = 0; i < apc->num_queues; i++) {
2206 		rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev);
2207 		if (!rxq) {
2208 			err = -ENOMEM;
2209 			goto out;
2210 		}
2211 
2212 		u64_stats_init(&rxq->stats.syncp);
2213 
2214 		apc->rxqs[i] = rxq;
2215 	}
2216 
2217 	apc->default_rxobj = apc->rxqs[0]->rxobj;
2218 out:
2219 	return err;
2220 }
2221 
2222 static void mana_destroy_vport(struct mana_port_context *apc)
2223 {
2224 	struct gdma_dev *gd = apc->ac->gdma_dev;
2225 	struct mana_rxq *rxq;
2226 	u32 rxq_idx;
2227 
2228 	for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
2229 		rxq = apc->rxqs[rxq_idx];
2230 		if (!rxq)
2231 			continue;
2232 
2233 		mana_destroy_rxq(apc, rxq, true);
2234 		apc->rxqs[rxq_idx] = NULL;
2235 	}
2236 
2237 	mana_destroy_txq(apc);
2238 	mana_uncfg_vport(apc);
2239 
2240 	if (gd->gdma_context->is_pf)
2241 		mana_pf_deregister_hw_vport(apc);
2242 }
2243 
2244 static int mana_create_vport(struct mana_port_context *apc,
2245 			     struct net_device *net)
2246 {
2247 	struct gdma_dev *gd = apc->ac->gdma_dev;
2248 	int err;
2249 
2250 	apc->default_rxobj = INVALID_MANA_HANDLE;
2251 
2252 	if (gd->gdma_context->is_pf) {
2253 		err = mana_pf_register_hw_vport(apc);
2254 		if (err)
2255 			return err;
2256 	}
2257 
2258 	err = mana_cfg_vport(apc, gd->pdid, gd->doorbell);
2259 	if (err)
2260 		return err;
2261 
2262 	return mana_create_txq(apc, net);
2263 }
2264 
2265 static void mana_rss_table_init(struct mana_port_context *apc)
2266 {
2267 	int i;
2268 
2269 	for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++)
2270 		apc->indir_table[i] =
2271 			ethtool_rxfh_indir_default(i, apc->num_queues);
2272 }
2273 
2274 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx,
2275 		    bool update_hash, bool update_tab)
2276 {
2277 	u32 queue_idx;
2278 	int err;
2279 	int i;
2280 
2281 	if (update_tab) {
2282 		for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) {
2283 			queue_idx = apc->indir_table[i];
2284 			apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj;
2285 		}
2286 	}
2287 
2288 	err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab);
2289 	if (err)
2290 		return err;
2291 
2292 	mana_fence_rqs(apc);
2293 
2294 	return 0;
2295 }
2296 
2297 static int mana_init_port(struct net_device *ndev)
2298 {
2299 	struct mana_port_context *apc = netdev_priv(ndev);
2300 	u32 max_txq, max_rxq, max_queues;
2301 	int port_idx = apc->port_idx;
2302 	u32 num_indirect_entries;
2303 	int err;
2304 
2305 	err = mana_init_port_context(apc);
2306 	if (err)
2307 		return err;
2308 
2309 	err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq,
2310 				   &num_indirect_entries);
2311 	if (err) {
2312 		netdev_err(ndev, "Failed to query info for vPort %d\n",
2313 			   port_idx);
2314 		goto reset_apc;
2315 	}
2316 
2317 	max_queues = min_t(u32, max_txq, max_rxq);
2318 	if (apc->max_queues > max_queues)
2319 		apc->max_queues = max_queues;
2320 
2321 	if (apc->num_queues > apc->max_queues)
2322 		apc->num_queues = apc->max_queues;
2323 
2324 	eth_hw_addr_set(ndev, apc->mac_addr);
2325 
2326 	return 0;
2327 
2328 reset_apc:
2329 	kfree(apc->rxqs);
2330 	apc->rxqs = NULL;
2331 	return err;
2332 }
2333 
2334 int mana_alloc_queues(struct net_device *ndev)
2335 {
2336 	struct mana_port_context *apc = netdev_priv(ndev);
2337 	struct gdma_dev *gd = apc->ac->gdma_dev;
2338 	int err;
2339 
2340 	err = mana_create_vport(apc, ndev);
2341 	if (err)
2342 		return err;
2343 
2344 	err = netif_set_real_num_tx_queues(ndev, apc->num_queues);
2345 	if (err)
2346 		goto destroy_vport;
2347 
2348 	err = mana_add_rx_queues(apc, ndev);
2349 	if (err)
2350 		goto destroy_vport;
2351 
2352 	apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE;
2353 
2354 	err = netif_set_real_num_rx_queues(ndev, apc->num_queues);
2355 	if (err)
2356 		goto destroy_vport;
2357 
2358 	mana_rss_table_init(apc);
2359 
2360 	err = mana_config_rss(apc, TRI_STATE_TRUE, true, true);
2361 	if (err)
2362 		goto destroy_vport;
2363 
2364 	if (gd->gdma_context->is_pf) {
2365 		err = mana_pf_register_filter(apc);
2366 		if (err)
2367 			goto destroy_vport;
2368 	}
2369 
2370 	mana_chn_setxdp(apc, mana_xdp_get(apc));
2371 
2372 	return 0;
2373 
2374 destroy_vport:
2375 	mana_destroy_vport(apc);
2376 	return err;
2377 }
2378 
2379 int mana_attach(struct net_device *ndev)
2380 {
2381 	struct mana_port_context *apc = netdev_priv(ndev);
2382 	int err;
2383 
2384 	ASSERT_RTNL();
2385 
2386 	err = mana_init_port(ndev);
2387 	if (err)
2388 		return err;
2389 
2390 	if (apc->port_st_save) {
2391 		err = mana_alloc_queues(ndev);
2392 		if (err) {
2393 			mana_cleanup_port_context(apc);
2394 			return err;
2395 		}
2396 	}
2397 
2398 	apc->port_is_up = apc->port_st_save;
2399 
2400 	/* Ensure port state updated before txq state */
2401 	smp_wmb();
2402 
2403 	if (apc->port_is_up)
2404 		netif_carrier_on(ndev);
2405 
2406 	netif_device_attach(ndev);
2407 
2408 	return 0;
2409 }
2410 
2411 static int mana_dealloc_queues(struct net_device *ndev)
2412 {
2413 	struct mana_port_context *apc = netdev_priv(ndev);
2414 	struct gdma_dev *gd = apc->ac->gdma_dev;
2415 	struct mana_txq *txq;
2416 	int i, err;
2417 
2418 	if (apc->port_is_up)
2419 		return -EINVAL;
2420 
2421 	mana_chn_setxdp(apc, NULL);
2422 
2423 	if (gd->gdma_context->is_pf)
2424 		mana_pf_deregister_filter(apc);
2425 
2426 	/* No packet can be transmitted now since apc->port_is_up is false.
2427 	 * There is still a tiny chance that mana_poll_tx_cq() can re-enable
2428 	 * a txq because it may not timely see apc->port_is_up being cleared
2429 	 * to false, but it doesn't matter since mana_start_xmit() drops any
2430 	 * new packets due to apc->port_is_up being false.
2431 	 *
2432 	 * Drain all the in-flight TX packets
2433 	 */
2434 	for (i = 0; i < apc->num_queues; i++) {
2435 		txq = &apc->tx_qp[i].txq;
2436 
2437 		while (atomic_read(&txq->pending_sends) > 0)
2438 			usleep_range(1000, 2000);
2439 	}
2440 
2441 	/* We're 100% sure the queues can no longer be woken up, because
2442 	 * we're sure now mana_poll_tx_cq() can't be running.
2443 	 */
2444 
2445 	apc->rss_state = TRI_STATE_FALSE;
2446 	err = mana_config_rss(apc, TRI_STATE_FALSE, false, false);
2447 	if (err) {
2448 		netdev_err(ndev, "Failed to disable vPort: %d\n", err);
2449 		return err;
2450 	}
2451 
2452 	mana_destroy_vport(apc);
2453 
2454 	return 0;
2455 }
2456 
2457 int mana_detach(struct net_device *ndev, bool from_close)
2458 {
2459 	struct mana_port_context *apc = netdev_priv(ndev);
2460 	int err;
2461 
2462 	ASSERT_RTNL();
2463 
2464 	apc->port_st_save = apc->port_is_up;
2465 	apc->port_is_up = false;
2466 
2467 	/* Ensure port state updated before txq state */
2468 	smp_wmb();
2469 
2470 	netif_tx_disable(ndev);
2471 	netif_carrier_off(ndev);
2472 
2473 	if (apc->port_st_save) {
2474 		err = mana_dealloc_queues(ndev);
2475 		if (err)
2476 			return err;
2477 	}
2478 
2479 	if (!from_close) {
2480 		netif_device_detach(ndev);
2481 		mana_cleanup_port_context(apc);
2482 	}
2483 
2484 	return 0;
2485 }
2486 
2487 static int mana_probe_port(struct mana_context *ac, int port_idx,
2488 			   struct net_device **ndev_storage)
2489 {
2490 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
2491 	struct mana_port_context *apc;
2492 	struct net_device *ndev;
2493 	int err;
2494 
2495 	ndev = alloc_etherdev_mq(sizeof(struct mana_port_context),
2496 				 gc->max_num_queues);
2497 	if (!ndev)
2498 		return -ENOMEM;
2499 
2500 	*ndev_storage = ndev;
2501 
2502 	apc = netdev_priv(ndev);
2503 	apc->ac = ac;
2504 	apc->ndev = ndev;
2505 	apc->max_queues = gc->max_num_queues;
2506 	apc->num_queues = gc->max_num_queues;
2507 	apc->port_handle = INVALID_MANA_HANDLE;
2508 	apc->pf_filter_handle = INVALID_MANA_HANDLE;
2509 	apc->port_idx = port_idx;
2510 
2511 	mutex_init(&apc->vport_mutex);
2512 	apc->vport_use_count = 0;
2513 
2514 	ndev->netdev_ops = &mana_devops;
2515 	ndev->ethtool_ops = &mana_ethtool_ops;
2516 	ndev->mtu = ETH_DATA_LEN;
2517 	ndev->max_mtu = gc->adapter_mtu - ETH_HLEN;
2518 	ndev->min_mtu = ETH_MIN_MTU;
2519 	ndev->needed_headroom = MANA_HEADROOM;
2520 	ndev->dev_port = port_idx;
2521 	SET_NETDEV_DEV(ndev, gc->dev);
2522 
2523 	netif_carrier_off(ndev);
2524 
2525 	netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE);
2526 
2527 	err = mana_init_port(ndev);
2528 	if (err)
2529 		goto free_net;
2530 
2531 	netdev_lockdep_set_classes(ndev);
2532 
2533 	ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
2534 	ndev->hw_features |= NETIF_F_RXCSUM;
2535 	ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
2536 	ndev->hw_features |= NETIF_F_RXHASH;
2537 	ndev->features = ndev->hw_features | NETIF_F_HW_VLAN_CTAG_TX |
2538 			 NETIF_F_HW_VLAN_CTAG_RX;
2539 	ndev->vlan_features = ndev->features;
2540 	ndev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT |
2541 			     NETDEV_XDP_ACT_NDO_XMIT;
2542 
2543 	err = register_netdev(ndev);
2544 	if (err) {
2545 		netdev_err(ndev, "Unable to register netdev.\n");
2546 		goto reset_apc;
2547 	}
2548 
2549 	return 0;
2550 
2551 reset_apc:
2552 	kfree(apc->rxqs);
2553 	apc->rxqs = NULL;
2554 free_net:
2555 	*ndev_storage = NULL;
2556 	netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err);
2557 	free_netdev(ndev);
2558 	return err;
2559 }
2560 
2561 static void adev_release(struct device *dev)
2562 {
2563 	struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev);
2564 
2565 	kfree(madev);
2566 }
2567 
2568 static void remove_adev(struct gdma_dev *gd)
2569 {
2570 	struct auxiliary_device *adev = gd->adev;
2571 	int id = adev->id;
2572 
2573 	auxiliary_device_delete(adev);
2574 	auxiliary_device_uninit(adev);
2575 
2576 	mana_adev_idx_free(id);
2577 	gd->adev = NULL;
2578 }
2579 
2580 static int add_adev(struct gdma_dev *gd)
2581 {
2582 	struct auxiliary_device *adev;
2583 	struct mana_adev *madev;
2584 	int ret;
2585 
2586 	madev = kzalloc(sizeof(*madev), GFP_KERNEL);
2587 	if (!madev)
2588 		return -ENOMEM;
2589 
2590 	adev = &madev->adev;
2591 	ret = mana_adev_idx_alloc();
2592 	if (ret < 0)
2593 		goto idx_fail;
2594 	adev->id = ret;
2595 
2596 	adev->name = "rdma";
2597 	adev->dev.parent = gd->gdma_context->dev;
2598 	adev->dev.release = adev_release;
2599 	madev->mdev = gd;
2600 
2601 	ret = auxiliary_device_init(adev);
2602 	if (ret)
2603 		goto init_fail;
2604 
2605 	ret = auxiliary_device_add(adev);
2606 	if (ret)
2607 		goto add_fail;
2608 
2609 	gd->adev = adev;
2610 	return 0;
2611 
2612 add_fail:
2613 	auxiliary_device_uninit(adev);
2614 
2615 init_fail:
2616 	mana_adev_idx_free(adev->id);
2617 
2618 idx_fail:
2619 	kfree(madev);
2620 
2621 	return ret;
2622 }
2623 
2624 int mana_probe(struct gdma_dev *gd, bool resuming)
2625 {
2626 	struct gdma_context *gc = gd->gdma_context;
2627 	struct mana_context *ac = gd->driver_data;
2628 	struct device *dev = gc->dev;
2629 	u16 num_ports = 0;
2630 	int err;
2631 	int i;
2632 
2633 	dev_info(dev,
2634 		 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n",
2635 		 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION);
2636 
2637 	err = mana_gd_register_device(gd);
2638 	if (err)
2639 		return err;
2640 
2641 	if (!resuming) {
2642 		ac = kzalloc(sizeof(*ac), GFP_KERNEL);
2643 		if (!ac)
2644 			return -ENOMEM;
2645 
2646 		ac->gdma_dev = gd;
2647 		gd->driver_data = ac;
2648 	}
2649 
2650 	err = mana_create_eq(ac);
2651 	if (err)
2652 		goto out;
2653 
2654 	err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION,
2655 				    MANA_MICRO_VERSION, &num_ports);
2656 	if (err)
2657 		goto out;
2658 
2659 	if (!resuming) {
2660 		ac->num_ports = num_ports;
2661 	} else {
2662 		if (ac->num_ports != num_ports) {
2663 			dev_err(dev, "The number of vPorts changed: %d->%d\n",
2664 				ac->num_ports, num_ports);
2665 			err = -EPROTO;
2666 			goto out;
2667 		}
2668 	}
2669 
2670 	if (ac->num_ports == 0)
2671 		dev_err(dev, "Failed to detect any vPort\n");
2672 
2673 	if (ac->num_ports > MAX_PORTS_IN_MANA_DEV)
2674 		ac->num_ports = MAX_PORTS_IN_MANA_DEV;
2675 
2676 	if (!resuming) {
2677 		for (i = 0; i < ac->num_ports; i++) {
2678 			err = mana_probe_port(ac, i, &ac->ports[i]);
2679 			if (err)
2680 				break;
2681 		}
2682 	} else {
2683 		for (i = 0; i < ac->num_ports; i++) {
2684 			rtnl_lock();
2685 			err = mana_attach(ac->ports[i]);
2686 			rtnl_unlock();
2687 			if (err)
2688 				break;
2689 		}
2690 	}
2691 
2692 	err = add_adev(gd);
2693 out:
2694 	if (err)
2695 		mana_remove(gd, false);
2696 
2697 	return err;
2698 }
2699 
2700 void mana_remove(struct gdma_dev *gd, bool suspending)
2701 {
2702 	struct gdma_context *gc = gd->gdma_context;
2703 	struct mana_context *ac = gd->driver_data;
2704 	struct device *dev = gc->dev;
2705 	struct net_device *ndev;
2706 	int err;
2707 	int i;
2708 
2709 	/* adev currently doesn't support suspending, always remove it */
2710 	if (gd->adev)
2711 		remove_adev(gd);
2712 
2713 	for (i = 0; i < ac->num_ports; i++) {
2714 		ndev = ac->ports[i];
2715 		if (!ndev) {
2716 			if (i == 0)
2717 				dev_err(dev, "No net device to remove\n");
2718 			goto out;
2719 		}
2720 
2721 		/* All cleanup actions should stay after rtnl_lock(), otherwise
2722 		 * other functions may access partially cleaned up data.
2723 		 */
2724 		rtnl_lock();
2725 
2726 		err = mana_detach(ndev, false);
2727 		if (err)
2728 			netdev_err(ndev, "Failed to detach vPort %d: %d\n",
2729 				   i, err);
2730 
2731 		if (suspending) {
2732 			/* No need to unregister the ndev. */
2733 			rtnl_unlock();
2734 			continue;
2735 		}
2736 
2737 		unregister_netdevice(ndev);
2738 
2739 		rtnl_unlock();
2740 
2741 		free_netdev(ndev);
2742 	}
2743 
2744 	mana_destroy_eq(ac);
2745 out:
2746 	mana_gd_deregister_device(gd);
2747 
2748 	if (suspending)
2749 		return;
2750 
2751 	gd->driver_data = NULL;
2752 	gd->gdma_context = NULL;
2753 	kfree(ac);
2754 }
2755