xref: /linux/drivers/net/ethernet/microsoft/mana/mana_en.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 #include <linux/pci.h>
12 
13 #include <net/checksum.h>
14 #include <net/ip6_checksum.h>
15 #include <net/page_pool/helpers.h>
16 #include <net/xdp.h>
17 
18 #include <net/mana/mana.h>
19 #include <net/mana/mana_auxiliary.h>
20 
21 static DEFINE_IDA(mana_adev_ida);
22 
23 static int mana_adev_idx_alloc(void)
24 {
25 	return ida_alloc(&mana_adev_ida, GFP_KERNEL);
26 }
27 
28 static void mana_adev_idx_free(int idx)
29 {
30 	ida_free(&mana_adev_ida, idx);
31 }
32 
33 /* Microsoft Azure Network Adapter (MANA) functions */
34 
35 static int mana_open(struct net_device *ndev)
36 {
37 	struct mana_port_context *apc = netdev_priv(ndev);
38 	int err;
39 
40 	err = mana_alloc_queues(ndev);
41 	if (err)
42 		return err;
43 
44 	apc->port_is_up = true;
45 
46 	/* Ensure port state updated before txq state */
47 	smp_wmb();
48 
49 	netif_carrier_on(ndev);
50 	netif_tx_wake_all_queues(ndev);
51 
52 	return 0;
53 }
54 
55 static int mana_close(struct net_device *ndev)
56 {
57 	struct mana_port_context *apc = netdev_priv(ndev);
58 
59 	if (!apc->port_is_up)
60 		return 0;
61 
62 	return mana_detach(ndev, true);
63 }
64 
65 static bool mana_can_tx(struct gdma_queue *wq)
66 {
67 	return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE;
68 }
69 
70 static unsigned int mana_checksum_info(struct sk_buff *skb)
71 {
72 	if (skb->protocol == htons(ETH_P_IP)) {
73 		struct iphdr *ip = ip_hdr(skb);
74 
75 		if (ip->protocol == IPPROTO_TCP)
76 			return IPPROTO_TCP;
77 
78 		if (ip->protocol == IPPROTO_UDP)
79 			return IPPROTO_UDP;
80 	} else if (skb->protocol == htons(ETH_P_IPV6)) {
81 		struct ipv6hdr *ip6 = ipv6_hdr(skb);
82 
83 		if (ip6->nexthdr == IPPROTO_TCP)
84 			return IPPROTO_TCP;
85 
86 		if (ip6->nexthdr == IPPROTO_UDP)
87 			return IPPROTO_UDP;
88 	}
89 
90 	/* No csum offloading */
91 	return 0;
92 }
93 
94 static void mana_add_sge(struct mana_tx_package *tp, struct mana_skb_head *ash,
95 			 int sg_i, dma_addr_t da, int sge_len, u32 gpa_mkey)
96 {
97 	ash->dma_handle[sg_i] = da;
98 	ash->size[sg_i] = sge_len;
99 
100 	tp->wqe_req.sgl[sg_i].address = da;
101 	tp->wqe_req.sgl[sg_i].mem_key = gpa_mkey;
102 	tp->wqe_req.sgl[sg_i].size = sge_len;
103 }
104 
105 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc,
106 			struct mana_tx_package *tp, int gso_hs)
107 {
108 	struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
109 	int hsg = 1; /* num of SGEs of linear part */
110 	struct gdma_dev *gd = apc->ac->gdma_dev;
111 	int skb_hlen = skb_headlen(skb);
112 	int sge0_len, sge1_len = 0;
113 	struct gdma_context *gc;
114 	struct device *dev;
115 	skb_frag_t *frag;
116 	dma_addr_t da;
117 	int sg_i;
118 	int i;
119 
120 	gc = gd->gdma_context;
121 	dev = gc->dev;
122 
123 	if (gso_hs && gso_hs < skb_hlen) {
124 		sge0_len = gso_hs;
125 		sge1_len = skb_hlen - gso_hs;
126 	} else {
127 		sge0_len = skb_hlen;
128 	}
129 
130 	da = dma_map_single(dev, skb->data, sge0_len, DMA_TO_DEVICE);
131 	if (dma_mapping_error(dev, da))
132 		return -ENOMEM;
133 
134 	mana_add_sge(tp, ash, 0, da, sge0_len, gd->gpa_mkey);
135 
136 	if (sge1_len) {
137 		sg_i = 1;
138 		da = dma_map_single(dev, skb->data + sge0_len, sge1_len,
139 				    DMA_TO_DEVICE);
140 		if (dma_mapping_error(dev, da))
141 			goto frag_err;
142 
143 		mana_add_sge(tp, ash, sg_i, da, sge1_len, gd->gpa_mkey);
144 		hsg = 2;
145 	}
146 
147 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
148 		sg_i = hsg + i;
149 
150 		frag = &skb_shinfo(skb)->frags[i];
151 		da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag),
152 				      DMA_TO_DEVICE);
153 		if (dma_mapping_error(dev, da))
154 			goto frag_err;
155 
156 		mana_add_sge(tp, ash, sg_i, da, skb_frag_size(frag),
157 			     gd->gpa_mkey);
158 	}
159 
160 	return 0;
161 
162 frag_err:
163 	for (i = sg_i - 1; i >= hsg; i--)
164 		dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
165 			       DMA_TO_DEVICE);
166 
167 	for (i = hsg - 1; i >= 0; i--)
168 		dma_unmap_single(dev, ash->dma_handle[i], ash->size[i],
169 				 DMA_TO_DEVICE);
170 
171 	return -ENOMEM;
172 }
173 
174 /* Handle the case when GSO SKB linear length is too large.
175  * MANA NIC requires GSO packets to put only the packet header to SGE0.
176  * So, we need 2 SGEs for the skb linear part which contains more than the
177  * header.
178  * Return a positive value for the number of SGEs, or a negative value
179  * for an error.
180  */
181 static int mana_fix_skb_head(struct net_device *ndev, struct sk_buff *skb,
182 			     int gso_hs)
183 {
184 	int num_sge = 1 + skb_shinfo(skb)->nr_frags;
185 	int skb_hlen = skb_headlen(skb);
186 
187 	if (gso_hs < skb_hlen) {
188 		num_sge++;
189 	} else if (gso_hs > skb_hlen) {
190 		if (net_ratelimit())
191 			netdev_err(ndev,
192 				   "TX nonlinear head: hs:%d, skb_hlen:%d\n",
193 				   gso_hs, skb_hlen);
194 
195 		return -EINVAL;
196 	}
197 
198 	return num_sge;
199 }
200 
201 /* Get the GSO packet's header size */
202 static int mana_get_gso_hs(struct sk_buff *skb)
203 {
204 	int gso_hs;
205 
206 	if (skb->encapsulation) {
207 		gso_hs = skb_inner_tcp_all_headers(skb);
208 	} else {
209 		if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
210 			gso_hs = skb_transport_offset(skb) +
211 				 sizeof(struct udphdr);
212 		} else {
213 			gso_hs = skb_tcp_all_headers(skb);
214 		}
215 	}
216 
217 	return gso_hs;
218 }
219 
220 netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev)
221 {
222 	enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT;
223 	struct mana_port_context *apc = netdev_priv(ndev);
224 	int gso_hs = 0; /* zero for non-GSO pkts */
225 	u16 txq_idx = skb_get_queue_mapping(skb);
226 	struct gdma_dev *gd = apc->ac->gdma_dev;
227 	bool ipv4 = false, ipv6 = false;
228 	struct mana_tx_package pkg = {};
229 	struct netdev_queue *net_txq;
230 	struct mana_stats_tx *tx_stats;
231 	struct gdma_queue *gdma_sq;
232 	unsigned int csum_type;
233 	struct mana_txq *txq;
234 	struct mana_cq *cq;
235 	int err, len;
236 
237 	if (unlikely(!apc->port_is_up))
238 		goto tx_drop;
239 
240 	if (skb_cow_head(skb, MANA_HEADROOM))
241 		goto tx_drop_count;
242 
243 	txq = &apc->tx_qp[txq_idx].txq;
244 	gdma_sq = txq->gdma_sq;
245 	cq = &apc->tx_qp[txq_idx].tx_cq;
246 	tx_stats = &txq->stats;
247 
248 	pkg.tx_oob.s_oob.vcq_num = cq->gdma_id;
249 	pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame;
250 
251 	if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) {
252 		pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset;
253 		pkt_fmt = MANA_LONG_PKT_FMT;
254 	} else {
255 		pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset;
256 	}
257 
258 	if (skb_vlan_tag_present(skb)) {
259 		pkt_fmt = MANA_LONG_PKT_FMT;
260 		pkg.tx_oob.l_oob.inject_vlan_pri_tag = 1;
261 		pkg.tx_oob.l_oob.pcp = skb_vlan_tag_get_prio(skb);
262 		pkg.tx_oob.l_oob.dei = skb_vlan_tag_get_cfi(skb);
263 		pkg.tx_oob.l_oob.vlan_id = skb_vlan_tag_get_id(skb);
264 	}
265 
266 	pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt;
267 
268 	if (pkt_fmt == MANA_SHORT_PKT_FMT) {
269 		pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob);
270 		u64_stats_update_begin(&tx_stats->syncp);
271 		tx_stats->short_pkt_fmt++;
272 		u64_stats_update_end(&tx_stats->syncp);
273 	} else {
274 		pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob);
275 		u64_stats_update_begin(&tx_stats->syncp);
276 		tx_stats->long_pkt_fmt++;
277 		u64_stats_update_end(&tx_stats->syncp);
278 	}
279 
280 	pkg.wqe_req.inline_oob_data = &pkg.tx_oob;
281 	pkg.wqe_req.flags = 0;
282 	pkg.wqe_req.client_data_unit = 0;
283 
284 	pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags;
285 
286 	if (skb->protocol == htons(ETH_P_IP))
287 		ipv4 = true;
288 	else if (skb->protocol == htons(ETH_P_IPV6))
289 		ipv6 = true;
290 
291 	if (skb_is_gso(skb)) {
292 		int num_sge;
293 
294 		gso_hs = mana_get_gso_hs(skb);
295 
296 		num_sge = mana_fix_skb_head(ndev, skb, gso_hs);
297 		if (num_sge > 0)
298 			pkg.wqe_req.num_sge = num_sge;
299 		else
300 			goto tx_drop_count;
301 
302 		u64_stats_update_begin(&tx_stats->syncp);
303 		if (skb->encapsulation) {
304 			tx_stats->tso_inner_packets++;
305 			tx_stats->tso_inner_bytes += skb->len - gso_hs;
306 		} else {
307 			tx_stats->tso_packets++;
308 			tx_stats->tso_bytes += skb->len - gso_hs;
309 		}
310 		u64_stats_update_end(&tx_stats->syncp);
311 
312 		pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
313 		pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
314 
315 		pkg.tx_oob.s_oob.comp_iphdr_csum = 1;
316 		pkg.tx_oob.s_oob.comp_tcp_csum = 1;
317 		pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
318 
319 		pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size;
320 		pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0;
321 		if (ipv4) {
322 			ip_hdr(skb)->tot_len = 0;
323 			ip_hdr(skb)->check = 0;
324 			tcp_hdr(skb)->check =
325 				~csum_tcpudp_magic(ip_hdr(skb)->saddr,
326 						   ip_hdr(skb)->daddr, 0,
327 						   IPPROTO_TCP, 0);
328 		} else {
329 			ipv6_hdr(skb)->payload_len = 0;
330 			tcp_hdr(skb)->check =
331 				~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
332 						 &ipv6_hdr(skb)->daddr, 0,
333 						 IPPROTO_TCP, 0);
334 		}
335 	} else if (skb->ip_summed == CHECKSUM_PARTIAL) {
336 		csum_type = mana_checksum_info(skb);
337 
338 		u64_stats_update_begin(&tx_stats->syncp);
339 		tx_stats->csum_partial++;
340 		u64_stats_update_end(&tx_stats->syncp);
341 
342 		if (csum_type == IPPROTO_TCP) {
343 			pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
344 			pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
345 
346 			pkg.tx_oob.s_oob.comp_tcp_csum = 1;
347 			pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
348 
349 		} else if (csum_type == IPPROTO_UDP) {
350 			pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
351 			pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
352 
353 			pkg.tx_oob.s_oob.comp_udp_csum = 1;
354 		} else {
355 			/* Can't do offload of this type of checksum */
356 			if (skb_checksum_help(skb))
357 				goto tx_drop_count;
358 		}
359 	}
360 
361 	WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES);
362 
363 	if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) {
364 		pkg.wqe_req.sgl = pkg.sgl_array;
365 	} else {
366 		pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge,
367 					    sizeof(struct gdma_sge),
368 					    GFP_ATOMIC);
369 		if (!pkg.sgl_ptr)
370 			goto tx_drop_count;
371 
372 		pkg.wqe_req.sgl = pkg.sgl_ptr;
373 	}
374 
375 	if (mana_map_skb(skb, apc, &pkg, gso_hs)) {
376 		u64_stats_update_begin(&tx_stats->syncp);
377 		tx_stats->mana_map_err++;
378 		u64_stats_update_end(&tx_stats->syncp);
379 		goto free_sgl_ptr;
380 	}
381 
382 	skb_queue_tail(&txq->pending_skbs, skb);
383 
384 	len = skb->len;
385 	net_txq = netdev_get_tx_queue(ndev, txq_idx);
386 
387 	err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req,
388 					(struct gdma_posted_wqe_info *)skb->cb);
389 	if (!mana_can_tx(gdma_sq)) {
390 		netif_tx_stop_queue(net_txq);
391 		apc->eth_stats.stop_queue++;
392 	}
393 
394 	if (err) {
395 		(void)skb_dequeue_tail(&txq->pending_skbs);
396 		netdev_warn(ndev, "Failed to post TX OOB: %d\n", err);
397 		err = NETDEV_TX_BUSY;
398 		goto tx_busy;
399 	}
400 
401 	err = NETDEV_TX_OK;
402 	atomic_inc(&txq->pending_sends);
403 
404 	mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq);
405 
406 	/* skb may be freed after mana_gd_post_work_request. Do not use it. */
407 	skb = NULL;
408 
409 	tx_stats = &txq->stats;
410 	u64_stats_update_begin(&tx_stats->syncp);
411 	tx_stats->packets++;
412 	tx_stats->bytes += len;
413 	u64_stats_update_end(&tx_stats->syncp);
414 
415 tx_busy:
416 	if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) {
417 		netif_tx_wake_queue(net_txq);
418 		apc->eth_stats.wake_queue++;
419 	}
420 
421 	kfree(pkg.sgl_ptr);
422 	return err;
423 
424 free_sgl_ptr:
425 	kfree(pkg.sgl_ptr);
426 tx_drop_count:
427 	ndev->stats.tx_dropped++;
428 tx_drop:
429 	dev_kfree_skb_any(skb);
430 	return NETDEV_TX_OK;
431 }
432 
433 static void mana_get_stats64(struct net_device *ndev,
434 			     struct rtnl_link_stats64 *st)
435 {
436 	struct mana_port_context *apc = netdev_priv(ndev);
437 	unsigned int num_queues = apc->num_queues;
438 	struct mana_stats_rx *rx_stats;
439 	struct mana_stats_tx *tx_stats;
440 	unsigned int start;
441 	u64 packets, bytes;
442 	int q;
443 
444 	if (!apc->port_is_up)
445 		return;
446 
447 	netdev_stats_to_stats64(st, &ndev->stats);
448 
449 	for (q = 0; q < num_queues; q++) {
450 		rx_stats = &apc->rxqs[q]->stats;
451 
452 		do {
453 			start = u64_stats_fetch_begin(&rx_stats->syncp);
454 			packets = rx_stats->packets;
455 			bytes = rx_stats->bytes;
456 		} while (u64_stats_fetch_retry(&rx_stats->syncp, start));
457 
458 		st->rx_packets += packets;
459 		st->rx_bytes += bytes;
460 	}
461 
462 	for (q = 0; q < num_queues; q++) {
463 		tx_stats = &apc->tx_qp[q].txq.stats;
464 
465 		do {
466 			start = u64_stats_fetch_begin(&tx_stats->syncp);
467 			packets = tx_stats->packets;
468 			bytes = tx_stats->bytes;
469 		} while (u64_stats_fetch_retry(&tx_stats->syncp, start));
470 
471 		st->tx_packets += packets;
472 		st->tx_bytes += bytes;
473 	}
474 }
475 
476 static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb,
477 			     int old_q)
478 {
479 	struct mana_port_context *apc = netdev_priv(ndev);
480 	u32 hash = skb_get_hash(skb);
481 	struct sock *sk = skb->sk;
482 	int txq;
483 
484 	txq = apc->indir_table[hash & (apc->indir_table_sz - 1)];
485 
486 	if (txq != old_q && sk && sk_fullsock(sk) &&
487 	    rcu_access_pointer(sk->sk_dst_cache))
488 		sk_tx_queue_set(sk, txq);
489 
490 	return txq;
491 }
492 
493 static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb,
494 			     struct net_device *sb_dev)
495 {
496 	int txq;
497 
498 	if (ndev->real_num_tx_queues == 1)
499 		return 0;
500 
501 	txq = sk_tx_queue_get(skb->sk);
502 
503 	if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) {
504 		if (skb_rx_queue_recorded(skb))
505 			txq = skb_get_rx_queue(skb);
506 		else
507 			txq = mana_get_tx_queue(ndev, skb, txq);
508 	}
509 
510 	return txq;
511 }
512 
513 /* Release pre-allocated RX buffers */
514 void mana_pre_dealloc_rxbufs(struct mana_port_context *mpc)
515 {
516 	struct device *dev;
517 	int i;
518 
519 	dev = mpc->ac->gdma_dev->gdma_context->dev;
520 
521 	if (!mpc->rxbufs_pre)
522 		goto out1;
523 
524 	if (!mpc->das_pre)
525 		goto out2;
526 
527 	while (mpc->rxbpre_total) {
528 		i = --mpc->rxbpre_total;
529 		dma_unmap_single(dev, mpc->das_pre[i], mpc->rxbpre_datasize,
530 				 DMA_FROM_DEVICE);
531 		put_page(virt_to_head_page(mpc->rxbufs_pre[i]));
532 	}
533 
534 	kfree(mpc->das_pre);
535 	mpc->das_pre = NULL;
536 
537 out2:
538 	kfree(mpc->rxbufs_pre);
539 	mpc->rxbufs_pre = NULL;
540 
541 out1:
542 	mpc->rxbpre_datasize = 0;
543 	mpc->rxbpre_alloc_size = 0;
544 	mpc->rxbpre_headroom = 0;
545 }
546 
547 /* Get a buffer from the pre-allocated RX buffers */
548 static void *mana_get_rxbuf_pre(struct mana_rxq *rxq, dma_addr_t *da)
549 {
550 	struct net_device *ndev = rxq->ndev;
551 	struct mana_port_context *mpc;
552 	void *va;
553 
554 	mpc = netdev_priv(ndev);
555 
556 	if (!mpc->rxbufs_pre || !mpc->das_pre || !mpc->rxbpre_total) {
557 		netdev_err(ndev, "No RX pre-allocated bufs\n");
558 		return NULL;
559 	}
560 
561 	/* Check sizes to catch unexpected coding error */
562 	if (mpc->rxbpre_datasize != rxq->datasize) {
563 		netdev_err(ndev, "rxbpre_datasize mismatch: %u: %u\n",
564 			   mpc->rxbpre_datasize, rxq->datasize);
565 		return NULL;
566 	}
567 
568 	if (mpc->rxbpre_alloc_size != rxq->alloc_size) {
569 		netdev_err(ndev, "rxbpre_alloc_size mismatch: %u: %u\n",
570 			   mpc->rxbpre_alloc_size, rxq->alloc_size);
571 		return NULL;
572 	}
573 
574 	if (mpc->rxbpre_headroom != rxq->headroom) {
575 		netdev_err(ndev, "rxbpre_headroom mismatch: %u: %u\n",
576 			   mpc->rxbpre_headroom, rxq->headroom);
577 		return NULL;
578 	}
579 
580 	mpc->rxbpre_total--;
581 
582 	*da = mpc->das_pre[mpc->rxbpre_total];
583 	va = mpc->rxbufs_pre[mpc->rxbpre_total];
584 	mpc->rxbufs_pre[mpc->rxbpre_total] = NULL;
585 
586 	/* Deallocate the array after all buffers are gone */
587 	if (!mpc->rxbpre_total)
588 		mana_pre_dealloc_rxbufs(mpc);
589 
590 	return va;
591 }
592 
593 /* Get RX buffer's data size, alloc size, XDP headroom based on MTU */
594 static void mana_get_rxbuf_cfg(int mtu, u32 *datasize, u32 *alloc_size,
595 			       u32 *headroom)
596 {
597 	if (mtu > MANA_XDP_MTU_MAX)
598 		*headroom = 0; /* no support for XDP */
599 	else
600 		*headroom = XDP_PACKET_HEADROOM;
601 
602 	*alloc_size = SKB_DATA_ALIGN(mtu + MANA_RXBUF_PAD + *headroom);
603 
604 	/* Using page pool in this case, so alloc_size is PAGE_SIZE */
605 	if (*alloc_size < PAGE_SIZE)
606 		*alloc_size = PAGE_SIZE;
607 
608 	*datasize = mtu + ETH_HLEN;
609 }
610 
611 int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu, int num_queues)
612 {
613 	struct device *dev;
614 	struct page *page;
615 	dma_addr_t da;
616 	int num_rxb;
617 	void *va;
618 	int i;
619 
620 	mana_get_rxbuf_cfg(new_mtu, &mpc->rxbpre_datasize,
621 			   &mpc->rxbpre_alloc_size, &mpc->rxbpre_headroom);
622 
623 	dev = mpc->ac->gdma_dev->gdma_context->dev;
624 
625 	num_rxb = num_queues * mpc->rx_queue_size;
626 
627 	WARN(mpc->rxbufs_pre, "mana rxbufs_pre exists\n");
628 	mpc->rxbufs_pre = kmalloc_array(num_rxb, sizeof(void *), GFP_KERNEL);
629 	if (!mpc->rxbufs_pre)
630 		goto error;
631 
632 	mpc->das_pre = kmalloc_array(num_rxb, sizeof(dma_addr_t), GFP_KERNEL);
633 	if (!mpc->das_pre)
634 		goto error;
635 
636 	mpc->rxbpre_total = 0;
637 
638 	for (i = 0; i < num_rxb; i++) {
639 		if (mpc->rxbpre_alloc_size > PAGE_SIZE) {
640 			va = netdev_alloc_frag(mpc->rxbpre_alloc_size);
641 			if (!va)
642 				goto error;
643 
644 			page = virt_to_head_page(va);
645 			/* Check if the frag falls back to single page */
646 			if (compound_order(page) <
647 			    get_order(mpc->rxbpre_alloc_size)) {
648 				put_page(page);
649 				goto error;
650 			}
651 		} else {
652 			page = dev_alloc_page();
653 			if (!page)
654 				goto error;
655 
656 			va = page_to_virt(page);
657 		}
658 
659 		da = dma_map_single(dev, va + mpc->rxbpre_headroom,
660 				    mpc->rxbpre_datasize, DMA_FROM_DEVICE);
661 		if (dma_mapping_error(dev, da)) {
662 			put_page(virt_to_head_page(va));
663 			goto error;
664 		}
665 
666 		mpc->rxbufs_pre[i] = va;
667 		mpc->das_pre[i] = da;
668 		mpc->rxbpre_total = i + 1;
669 	}
670 
671 	return 0;
672 
673 error:
674 	mana_pre_dealloc_rxbufs(mpc);
675 	return -ENOMEM;
676 }
677 
678 static int mana_change_mtu(struct net_device *ndev, int new_mtu)
679 {
680 	struct mana_port_context *mpc = netdev_priv(ndev);
681 	unsigned int old_mtu = ndev->mtu;
682 	int err;
683 
684 	/* Pre-allocate buffers to prevent failure in mana_attach later */
685 	err = mana_pre_alloc_rxbufs(mpc, new_mtu, mpc->num_queues);
686 	if (err) {
687 		netdev_err(ndev, "Insufficient memory for new MTU\n");
688 		return err;
689 	}
690 
691 	err = mana_detach(ndev, false);
692 	if (err) {
693 		netdev_err(ndev, "mana_detach failed: %d\n", err);
694 		goto out;
695 	}
696 
697 	WRITE_ONCE(ndev->mtu, new_mtu);
698 
699 	err = mana_attach(ndev);
700 	if (err) {
701 		netdev_err(ndev, "mana_attach failed: %d\n", err);
702 		WRITE_ONCE(ndev->mtu, old_mtu);
703 	}
704 
705 out:
706 	mana_pre_dealloc_rxbufs(mpc);
707 	return err;
708 }
709 
710 static const struct net_device_ops mana_devops = {
711 	.ndo_open		= mana_open,
712 	.ndo_stop		= mana_close,
713 	.ndo_select_queue	= mana_select_queue,
714 	.ndo_start_xmit		= mana_start_xmit,
715 	.ndo_validate_addr	= eth_validate_addr,
716 	.ndo_get_stats64	= mana_get_stats64,
717 	.ndo_bpf		= mana_bpf,
718 	.ndo_xdp_xmit		= mana_xdp_xmit,
719 	.ndo_change_mtu		= mana_change_mtu,
720 };
721 
722 static void mana_cleanup_port_context(struct mana_port_context *apc)
723 {
724 	kfree(apc->rxqs);
725 	apc->rxqs = NULL;
726 }
727 
728 static void mana_cleanup_indir_table(struct mana_port_context *apc)
729 {
730 	apc->indir_table_sz = 0;
731 	kfree(apc->indir_table);
732 	kfree(apc->rxobj_table);
733 }
734 
735 static int mana_init_port_context(struct mana_port_context *apc)
736 {
737 	apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *),
738 			    GFP_KERNEL);
739 
740 	return !apc->rxqs ? -ENOMEM : 0;
741 }
742 
743 static int mana_send_request(struct mana_context *ac, void *in_buf,
744 			     u32 in_len, void *out_buf, u32 out_len)
745 {
746 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
747 	struct gdma_resp_hdr *resp = out_buf;
748 	struct gdma_req_hdr *req = in_buf;
749 	struct device *dev = gc->dev;
750 	static atomic_t activity_id;
751 	int err;
752 
753 	req->dev_id = gc->mana.dev_id;
754 	req->activity_id = atomic_inc_return(&activity_id);
755 
756 	err = mana_gd_send_request(gc, in_len, in_buf, out_len,
757 				   out_buf);
758 	if (err || resp->status) {
759 		dev_err(dev, "Failed to send mana message: %d, 0x%x\n",
760 			err, resp->status);
761 		return err ? err : -EPROTO;
762 	}
763 
764 	if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 ||
765 	    req->activity_id != resp->activity_id) {
766 		dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n",
767 			req->dev_id.as_uint32, resp->dev_id.as_uint32,
768 			req->activity_id, resp->activity_id);
769 		return -EPROTO;
770 	}
771 
772 	return 0;
773 }
774 
775 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr,
776 				const enum mana_command_code expected_code,
777 				const u32 min_size)
778 {
779 	if (resp_hdr->response.msg_type != expected_code)
780 		return -EPROTO;
781 
782 	if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1)
783 		return -EPROTO;
784 
785 	if (resp_hdr->response.msg_size < min_size)
786 		return -EPROTO;
787 
788 	return 0;
789 }
790 
791 static int mana_pf_register_hw_vport(struct mana_port_context *apc)
792 {
793 	struct mana_register_hw_vport_resp resp = {};
794 	struct mana_register_hw_vport_req req = {};
795 	int err;
796 
797 	mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT,
798 			     sizeof(req), sizeof(resp));
799 	req.attached_gfid = 1;
800 	req.is_pf_default_vport = 1;
801 	req.allow_all_ether_types = 1;
802 
803 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
804 				sizeof(resp));
805 	if (err) {
806 		netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err);
807 		return err;
808 	}
809 
810 	err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT,
811 				   sizeof(resp));
812 	if (err || resp.hdr.status) {
813 		netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n",
814 			   err, resp.hdr.status);
815 		return err ? err : -EPROTO;
816 	}
817 
818 	apc->port_handle = resp.hw_vport_handle;
819 	return 0;
820 }
821 
822 static void mana_pf_deregister_hw_vport(struct mana_port_context *apc)
823 {
824 	struct mana_deregister_hw_vport_resp resp = {};
825 	struct mana_deregister_hw_vport_req req = {};
826 	int err;
827 
828 	mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT,
829 			     sizeof(req), sizeof(resp));
830 	req.hw_vport_handle = apc->port_handle;
831 
832 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
833 				sizeof(resp));
834 	if (err) {
835 		netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n",
836 			   err);
837 		return;
838 	}
839 
840 	err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT,
841 				   sizeof(resp));
842 	if (err || resp.hdr.status)
843 		netdev_err(apc->ndev,
844 			   "Failed to deregister hw vPort: %d, 0x%x\n",
845 			   err, resp.hdr.status);
846 }
847 
848 static int mana_pf_register_filter(struct mana_port_context *apc)
849 {
850 	struct mana_register_filter_resp resp = {};
851 	struct mana_register_filter_req req = {};
852 	int err;
853 
854 	mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER,
855 			     sizeof(req), sizeof(resp));
856 	req.vport = apc->port_handle;
857 	memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN);
858 
859 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
860 				sizeof(resp));
861 	if (err) {
862 		netdev_err(apc->ndev, "Failed to register filter: %d\n", err);
863 		return err;
864 	}
865 
866 	err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER,
867 				   sizeof(resp));
868 	if (err || resp.hdr.status) {
869 		netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n",
870 			   err, resp.hdr.status);
871 		return err ? err : -EPROTO;
872 	}
873 
874 	apc->pf_filter_handle = resp.filter_handle;
875 	return 0;
876 }
877 
878 static void mana_pf_deregister_filter(struct mana_port_context *apc)
879 {
880 	struct mana_deregister_filter_resp resp = {};
881 	struct mana_deregister_filter_req req = {};
882 	int err;
883 
884 	mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER,
885 			     sizeof(req), sizeof(resp));
886 	req.filter_handle = apc->pf_filter_handle;
887 
888 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
889 				sizeof(resp));
890 	if (err) {
891 		netdev_err(apc->ndev, "Failed to unregister filter: %d\n",
892 			   err);
893 		return;
894 	}
895 
896 	err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER,
897 				   sizeof(resp));
898 	if (err || resp.hdr.status)
899 		netdev_err(apc->ndev,
900 			   "Failed to deregister filter: %d, 0x%x\n",
901 			   err, resp.hdr.status);
902 }
903 
904 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver,
905 				 u32 proto_minor_ver, u32 proto_micro_ver,
906 				 u16 *max_num_vports)
907 {
908 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
909 	struct mana_query_device_cfg_resp resp = {};
910 	struct mana_query_device_cfg_req req = {};
911 	struct device *dev = gc->dev;
912 	int err = 0;
913 
914 	mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG,
915 			     sizeof(req), sizeof(resp));
916 
917 	req.hdr.resp.msg_version = GDMA_MESSAGE_V2;
918 
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 		dev_err(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 		dev_err(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 	if (resp.hdr.response.msg_version == GDMA_MESSAGE_V2)
942 		gc->adapter_mtu = resp.adapter_mtu;
943 	else
944 		gc->adapter_mtu = ETH_FRAME_LEN;
945 
946 	return 0;
947 }
948 
949 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index,
950 				u32 *max_sq, u32 *max_rq, u32 *num_indir_entry)
951 {
952 	struct mana_query_vport_cfg_resp resp = {};
953 	struct mana_query_vport_cfg_req req = {};
954 	int err;
955 
956 	mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG,
957 			     sizeof(req), sizeof(resp));
958 
959 	req.vport_index = vport_index;
960 
961 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
962 				sizeof(resp));
963 	if (err)
964 		return err;
965 
966 	err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG,
967 				   sizeof(resp));
968 	if (err)
969 		return err;
970 
971 	if (resp.hdr.status)
972 		return -EPROTO;
973 
974 	*max_sq = resp.max_num_sq;
975 	*max_rq = resp.max_num_rq;
976 	if (resp.num_indirection_ent > 0 &&
977 	    resp.num_indirection_ent <= MANA_INDIRECT_TABLE_MAX_SIZE &&
978 	    is_power_of_2(resp.num_indirection_ent)) {
979 		*num_indir_entry = resp.num_indirection_ent;
980 	} else {
981 		netdev_warn(apc->ndev,
982 			    "Setting indirection table size to default %d for vPort %d\n",
983 			    MANA_INDIRECT_TABLE_DEF_SIZE, apc->port_idx);
984 		*num_indir_entry = MANA_INDIRECT_TABLE_DEF_SIZE;
985 	}
986 
987 	apc->port_handle = resp.vport;
988 	ether_addr_copy(apc->mac_addr, resp.mac_addr);
989 
990 	return 0;
991 }
992 
993 void mana_uncfg_vport(struct mana_port_context *apc)
994 {
995 	mutex_lock(&apc->vport_mutex);
996 	apc->vport_use_count--;
997 	WARN_ON(apc->vport_use_count < 0);
998 	mutex_unlock(&apc->vport_mutex);
999 }
1000 EXPORT_SYMBOL_NS(mana_uncfg_vport, NET_MANA);
1001 
1002 int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id,
1003 		   u32 doorbell_pg_id)
1004 {
1005 	struct mana_config_vport_resp resp = {};
1006 	struct mana_config_vport_req req = {};
1007 	int err;
1008 
1009 	/* This function is used to program the Ethernet port in the hardware
1010 	 * table. It can be called from the Ethernet driver or the RDMA driver.
1011 	 *
1012 	 * For Ethernet usage, the hardware supports only one active user on a
1013 	 * physical port. The driver checks on the port usage before programming
1014 	 * the hardware when creating the RAW QP (RDMA driver) or exposing the
1015 	 * device to kernel NET layer (Ethernet driver).
1016 	 *
1017 	 * Because the RDMA driver doesn't know in advance which QP type the
1018 	 * user will create, it exposes the device with all its ports. The user
1019 	 * may not be able to create RAW QP on a port if this port is already
1020 	 * in used by the Ethernet driver from the kernel.
1021 	 *
1022 	 * This physical port limitation only applies to the RAW QP. For RC QP,
1023 	 * the hardware doesn't have this limitation. The user can create RC
1024 	 * QPs on a physical port up to the hardware limits independent of the
1025 	 * Ethernet usage on the same port.
1026 	 */
1027 	mutex_lock(&apc->vport_mutex);
1028 	if (apc->vport_use_count > 0) {
1029 		mutex_unlock(&apc->vport_mutex);
1030 		return -EBUSY;
1031 	}
1032 	apc->vport_use_count++;
1033 	mutex_unlock(&apc->vport_mutex);
1034 
1035 	mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX,
1036 			     sizeof(req), sizeof(resp));
1037 	req.vport = apc->port_handle;
1038 	req.pdid = protection_dom_id;
1039 	req.doorbell_pageid = doorbell_pg_id;
1040 
1041 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1042 				sizeof(resp));
1043 	if (err) {
1044 		netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err);
1045 		goto out;
1046 	}
1047 
1048 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX,
1049 				   sizeof(resp));
1050 	if (err || resp.hdr.status) {
1051 		netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n",
1052 			   err, resp.hdr.status);
1053 		if (!err)
1054 			err = -EPROTO;
1055 
1056 		goto out;
1057 	}
1058 
1059 	apc->tx_shortform_allowed = resp.short_form_allowed;
1060 	apc->tx_vp_offset = resp.tx_vport_offset;
1061 
1062 	netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n",
1063 		    apc->port_handle, protection_dom_id, doorbell_pg_id);
1064 out:
1065 	if (err)
1066 		mana_uncfg_vport(apc);
1067 
1068 	return err;
1069 }
1070 EXPORT_SYMBOL_NS(mana_cfg_vport, NET_MANA);
1071 
1072 static int mana_cfg_vport_steering(struct mana_port_context *apc,
1073 				   enum TRI_STATE rx,
1074 				   bool update_default_rxobj, bool update_key,
1075 				   bool update_tab)
1076 {
1077 	struct mana_cfg_rx_steer_req_v2 *req;
1078 	struct mana_cfg_rx_steer_resp resp = {};
1079 	struct net_device *ndev = apc->ndev;
1080 	u32 req_buf_size;
1081 	int err;
1082 
1083 	req_buf_size = struct_size(req, indir_tab, apc->indir_table_sz);
1084 	req = kzalloc(req_buf_size, GFP_KERNEL);
1085 	if (!req)
1086 		return -ENOMEM;
1087 
1088 	mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
1089 			     sizeof(resp));
1090 
1091 	req->hdr.req.msg_version = GDMA_MESSAGE_V2;
1092 
1093 	req->vport = apc->port_handle;
1094 	req->num_indir_entries = apc->indir_table_sz;
1095 	req->indir_tab_offset = offsetof(struct mana_cfg_rx_steer_req_v2,
1096 					 indir_tab);
1097 	req->rx_enable = rx;
1098 	req->rss_enable = apc->rss_state;
1099 	req->update_default_rxobj = update_default_rxobj;
1100 	req->update_hashkey = update_key;
1101 	req->update_indir_tab = update_tab;
1102 	req->default_rxobj = apc->default_rxobj;
1103 	req->cqe_coalescing_enable = 0;
1104 
1105 	if (update_key)
1106 		memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE);
1107 
1108 	if (update_tab)
1109 		memcpy(req->indir_tab, apc->rxobj_table,
1110 		       flex_array_size(req, indir_tab, req->num_indir_entries));
1111 
1112 	err = mana_send_request(apc->ac, req, req_buf_size, &resp,
1113 				sizeof(resp));
1114 	if (err) {
1115 		netdev_err(ndev, "Failed to configure vPort RX: %d\n", err);
1116 		goto out;
1117 	}
1118 
1119 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX,
1120 				   sizeof(resp));
1121 	if (err) {
1122 		netdev_err(ndev, "vPort RX configuration failed: %d\n", err);
1123 		goto out;
1124 	}
1125 
1126 	if (resp.hdr.status) {
1127 		netdev_err(ndev, "vPort RX configuration failed: 0x%x\n",
1128 			   resp.hdr.status);
1129 		err = -EPROTO;
1130 	}
1131 
1132 	netdev_info(ndev, "Configured steering vPort %llu entries %u\n",
1133 		    apc->port_handle, apc->indir_table_sz);
1134 out:
1135 	kfree(req);
1136 	return err;
1137 }
1138 
1139 int mana_create_wq_obj(struct mana_port_context *apc,
1140 		       mana_handle_t vport,
1141 		       u32 wq_type, struct mana_obj_spec *wq_spec,
1142 		       struct mana_obj_spec *cq_spec,
1143 		       mana_handle_t *wq_obj)
1144 {
1145 	struct mana_create_wqobj_resp resp = {};
1146 	struct mana_create_wqobj_req req = {};
1147 	struct net_device *ndev = apc->ndev;
1148 	int err;
1149 
1150 	mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ,
1151 			     sizeof(req), sizeof(resp));
1152 	req.vport = vport;
1153 	req.wq_type = wq_type;
1154 	req.wq_gdma_region = wq_spec->gdma_region;
1155 	req.cq_gdma_region = cq_spec->gdma_region;
1156 	req.wq_size = wq_spec->queue_size;
1157 	req.cq_size = cq_spec->queue_size;
1158 	req.cq_moderation_ctx_id = cq_spec->modr_ctx_id;
1159 	req.cq_parent_qid = cq_spec->attached_eq;
1160 
1161 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1162 				sizeof(resp));
1163 	if (err) {
1164 		netdev_err(ndev, "Failed to create WQ object: %d\n", err);
1165 		goto out;
1166 	}
1167 
1168 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ,
1169 				   sizeof(resp));
1170 	if (err || resp.hdr.status) {
1171 		netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err,
1172 			   resp.hdr.status);
1173 		if (!err)
1174 			err = -EPROTO;
1175 		goto out;
1176 	}
1177 
1178 	if (resp.wq_obj == INVALID_MANA_HANDLE) {
1179 		netdev_err(ndev, "Got an invalid WQ object handle\n");
1180 		err = -EPROTO;
1181 		goto out;
1182 	}
1183 
1184 	*wq_obj = resp.wq_obj;
1185 	wq_spec->queue_index = resp.wq_id;
1186 	cq_spec->queue_index = resp.cq_id;
1187 
1188 	return 0;
1189 out:
1190 	return err;
1191 }
1192 EXPORT_SYMBOL_NS(mana_create_wq_obj, NET_MANA);
1193 
1194 void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type,
1195 			 mana_handle_t wq_obj)
1196 {
1197 	struct mana_destroy_wqobj_resp resp = {};
1198 	struct mana_destroy_wqobj_req req = {};
1199 	struct net_device *ndev = apc->ndev;
1200 	int err;
1201 
1202 	mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ,
1203 			     sizeof(req), sizeof(resp));
1204 	req.wq_type = wq_type;
1205 	req.wq_obj_handle = wq_obj;
1206 
1207 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1208 				sizeof(resp));
1209 	if (err) {
1210 		netdev_err(ndev, "Failed to destroy WQ object: %d\n", err);
1211 		return;
1212 	}
1213 
1214 	err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ,
1215 				   sizeof(resp));
1216 	if (err || resp.hdr.status)
1217 		netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err,
1218 			   resp.hdr.status);
1219 }
1220 EXPORT_SYMBOL_NS(mana_destroy_wq_obj, NET_MANA);
1221 
1222 static void mana_destroy_eq(struct mana_context *ac)
1223 {
1224 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
1225 	struct gdma_queue *eq;
1226 	int i;
1227 
1228 	if (!ac->eqs)
1229 		return;
1230 
1231 	for (i = 0; i < gc->max_num_queues; i++) {
1232 		eq = ac->eqs[i].eq;
1233 		if (!eq)
1234 			continue;
1235 
1236 		mana_gd_destroy_queue(gc, eq);
1237 	}
1238 
1239 	kfree(ac->eqs);
1240 	ac->eqs = NULL;
1241 }
1242 
1243 static int mana_create_eq(struct mana_context *ac)
1244 {
1245 	struct gdma_dev *gd = ac->gdma_dev;
1246 	struct gdma_context *gc = gd->gdma_context;
1247 	struct gdma_queue_spec spec = {};
1248 	int err;
1249 	int i;
1250 
1251 	ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq),
1252 			  GFP_KERNEL);
1253 	if (!ac->eqs)
1254 		return -ENOMEM;
1255 
1256 	spec.type = GDMA_EQ;
1257 	spec.monitor_avl_buf = false;
1258 	spec.queue_size = EQ_SIZE;
1259 	spec.eq.callback = NULL;
1260 	spec.eq.context = ac->eqs;
1261 	spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE;
1262 
1263 	for (i = 0; i < gc->max_num_queues; i++) {
1264 		spec.eq.msix_index = (i + 1) % gc->num_msix_usable;
1265 		err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq);
1266 		if (err)
1267 			goto out;
1268 	}
1269 
1270 	return 0;
1271 out:
1272 	mana_destroy_eq(ac);
1273 	return err;
1274 }
1275 
1276 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq)
1277 {
1278 	struct mana_fence_rq_resp resp = {};
1279 	struct mana_fence_rq_req req = {};
1280 	int err;
1281 
1282 	init_completion(&rxq->fence_event);
1283 
1284 	mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ,
1285 			     sizeof(req), sizeof(resp));
1286 	req.wq_obj_handle =  rxq->rxobj;
1287 
1288 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1289 				sizeof(resp));
1290 	if (err) {
1291 		netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n",
1292 			   rxq->rxq_idx, err);
1293 		return err;
1294 	}
1295 
1296 	err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp));
1297 	if (err || resp.hdr.status) {
1298 		netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n",
1299 			   rxq->rxq_idx, err, resp.hdr.status);
1300 		if (!err)
1301 			err = -EPROTO;
1302 
1303 		return err;
1304 	}
1305 
1306 	if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) {
1307 		netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n",
1308 			   rxq->rxq_idx);
1309 		return -ETIMEDOUT;
1310 	}
1311 
1312 	return 0;
1313 }
1314 
1315 static void 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 			msleep(100);
1328 	}
1329 }
1330 
1331 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units)
1332 {
1333 	u32 used_space_old;
1334 	u32 used_space_new;
1335 
1336 	used_space_old = wq->head - wq->tail;
1337 	used_space_new = wq->head - (wq->tail + num_units);
1338 
1339 	if (WARN_ON_ONCE(used_space_new > used_space_old))
1340 		return -ERANGE;
1341 
1342 	wq->tail += num_units;
1343 	return 0;
1344 }
1345 
1346 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc)
1347 {
1348 	struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
1349 	struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1350 	struct device *dev = gc->dev;
1351 	int hsg, i;
1352 
1353 	/* Number of SGEs of linear part */
1354 	hsg = (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0]) ? 2 : 1;
1355 
1356 	for (i = 0; i < hsg; i++)
1357 		dma_unmap_single(dev, ash->dma_handle[i], ash->size[i],
1358 				 DMA_TO_DEVICE);
1359 
1360 	for (i = hsg; i < skb_shinfo(skb)->nr_frags + hsg; i++)
1361 		dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
1362 			       DMA_TO_DEVICE);
1363 }
1364 
1365 static void mana_poll_tx_cq(struct mana_cq *cq)
1366 {
1367 	struct gdma_comp *completions = cq->gdma_comp_buf;
1368 	struct gdma_posted_wqe_info *wqe_info;
1369 	unsigned int pkt_transmitted = 0;
1370 	unsigned int wqe_unit_cnt = 0;
1371 	struct mana_txq *txq = cq->txq;
1372 	struct mana_port_context *apc;
1373 	struct netdev_queue *net_txq;
1374 	struct gdma_queue *gdma_wq;
1375 	unsigned int avail_space;
1376 	struct net_device *ndev;
1377 	struct sk_buff *skb;
1378 	bool txq_stopped;
1379 	int comp_read;
1380 	int i;
1381 
1382 	ndev = txq->ndev;
1383 	apc = netdev_priv(ndev);
1384 
1385 	comp_read = mana_gd_poll_cq(cq->gdma_cq, completions,
1386 				    CQE_POLLING_BUFFER);
1387 
1388 	if (comp_read < 1)
1389 		return;
1390 
1391 	for (i = 0; i < comp_read; i++) {
1392 		struct mana_tx_comp_oob *cqe_oob;
1393 
1394 		if (WARN_ON_ONCE(!completions[i].is_sq))
1395 			return;
1396 
1397 		cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data;
1398 		if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type !=
1399 				 MANA_CQE_COMPLETION))
1400 			return;
1401 
1402 		switch (cqe_oob->cqe_hdr.cqe_type) {
1403 		case CQE_TX_OKAY:
1404 			break;
1405 
1406 		case CQE_TX_SA_DROP:
1407 		case CQE_TX_MTU_DROP:
1408 		case CQE_TX_INVALID_OOB:
1409 		case CQE_TX_INVALID_ETH_TYPE:
1410 		case CQE_TX_HDR_PROCESSING_ERROR:
1411 		case CQE_TX_VF_DISABLED:
1412 		case CQE_TX_VPORT_IDX_OUT_OF_RANGE:
1413 		case CQE_TX_VPORT_DISABLED:
1414 		case CQE_TX_VLAN_TAGGING_VIOLATION:
1415 			if (net_ratelimit())
1416 				netdev_err(ndev, "TX: CQE error %d\n",
1417 					   cqe_oob->cqe_hdr.cqe_type);
1418 
1419 			apc->eth_stats.tx_cqe_err++;
1420 			break;
1421 
1422 		default:
1423 			/* If the CQE type is unknown, log an error,
1424 			 * and still free the SKB, update tail, etc.
1425 			 */
1426 			if (net_ratelimit())
1427 				netdev_err(ndev, "TX: unknown CQE type %d\n",
1428 					   cqe_oob->cqe_hdr.cqe_type);
1429 
1430 			apc->eth_stats.tx_cqe_unknown_type++;
1431 			break;
1432 		}
1433 
1434 		if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num))
1435 			return;
1436 
1437 		skb = skb_dequeue(&txq->pending_skbs);
1438 		if (WARN_ON_ONCE(!skb))
1439 			return;
1440 
1441 		wqe_info = (struct gdma_posted_wqe_info *)skb->cb;
1442 		wqe_unit_cnt += wqe_info->wqe_size_in_bu;
1443 
1444 		mana_unmap_skb(skb, apc);
1445 
1446 		napi_consume_skb(skb, cq->budget);
1447 
1448 		pkt_transmitted++;
1449 	}
1450 
1451 	if (WARN_ON_ONCE(wqe_unit_cnt == 0))
1452 		return;
1453 
1454 	mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt);
1455 
1456 	gdma_wq = txq->gdma_sq;
1457 	avail_space = mana_gd_wq_avail_space(gdma_wq);
1458 
1459 	/* Ensure tail updated before checking q stop */
1460 	smp_mb();
1461 
1462 	net_txq = txq->net_txq;
1463 	txq_stopped = netif_tx_queue_stopped(net_txq);
1464 
1465 	/* Ensure checking txq_stopped before apc->port_is_up. */
1466 	smp_rmb();
1467 
1468 	if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) {
1469 		netif_tx_wake_queue(net_txq);
1470 		apc->eth_stats.wake_queue++;
1471 	}
1472 
1473 	if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0)
1474 		WARN_ON_ONCE(1);
1475 
1476 	cq->work_done = pkt_transmitted;
1477 }
1478 
1479 static void mana_post_pkt_rxq(struct mana_rxq *rxq)
1480 {
1481 	struct mana_recv_buf_oob *recv_buf_oob;
1482 	u32 curr_index;
1483 	int err;
1484 
1485 	curr_index = rxq->buf_index++;
1486 	if (rxq->buf_index == rxq->num_rx_buf)
1487 		rxq->buf_index = 0;
1488 
1489 	recv_buf_oob = &rxq->rx_oobs[curr_index];
1490 
1491 	err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req,
1492 					&recv_buf_oob->wqe_inf);
1493 	if (WARN_ON_ONCE(err))
1494 		return;
1495 
1496 	WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1);
1497 }
1498 
1499 static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va,
1500 				      uint pkt_len, struct xdp_buff *xdp)
1501 {
1502 	struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size);
1503 
1504 	if (!skb)
1505 		return NULL;
1506 
1507 	if (xdp->data_hard_start) {
1508 		skb_reserve(skb, xdp->data - xdp->data_hard_start);
1509 		skb_put(skb, xdp->data_end - xdp->data);
1510 		return skb;
1511 	}
1512 
1513 	skb_reserve(skb, rxq->headroom);
1514 	skb_put(skb, pkt_len);
1515 
1516 	return skb;
1517 }
1518 
1519 static void mana_rx_skb(void *buf_va, bool from_pool,
1520 			struct mana_rxcomp_oob *cqe, struct mana_rxq *rxq)
1521 {
1522 	struct mana_stats_rx *rx_stats = &rxq->stats;
1523 	struct net_device *ndev = rxq->ndev;
1524 	uint pkt_len = cqe->ppi[0].pkt_len;
1525 	u16 rxq_idx = rxq->rxq_idx;
1526 	struct napi_struct *napi;
1527 	struct xdp_buff xdp = {};
1528 	struct sk_buff *skb;
1529 	u32 hash_value;
1530 	u32 act;
1531 
1532 	rxq->rx_cq.work_done++;
1533 	napi = &rxq->rx_cq.napi;
1534 
1535 	if (!buf_va) {
1536 		++ndev->stats.rx_dropped;
1537 		return;
1538 	}
1539 
1540 	act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len);
1541 
1542 	if (act == XDP_REDIRECT && !rxq->xdp_rc)
1543 		return;
1544 
1545 	if (act != XDP_PASS && act != XDP_TX)
1546 		goto drop_xdp;
1547 
1548 	skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp);
1549 
1550 	if (!skb)
1551 		goto drop;
1552 
1553 	if (from_pool)
1554 		skb_mark_for_recycle(skb);
1555 
1556 	skb->dev = napi->dev;
1557 
1558 	skb->protocol = eth_type_trans(skb, ndev);
1559 	skb_checksum_none_assert(skb);
1560 	skb_record_rx_queue(skb, rxq_idx);
1561 
1562 	if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) {
1563 		if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed)
1564 			skb->ip_summed = CHECKSUM_UNNECESSARY;
1565 	}
1566 
1567 	if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) {
1568 		hash_value = cqe->ppi[0].pkt_hash;
1569 
1570 		if (cqe->rx_hashtype & MANA_HASH_L4)
1571 			skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4);
1572 		else
1573 			skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3);
1574 	}
1575 
1576 	if (cqe->rx_vlantag_present) {
1577 		u16 vlan_tci = cqe->rx_vlan_id;
1578 
1579 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci);
1580 	}
1581 
1582 	u64_stats_update_begin(&rx_stats->syncp);
1583 	rx_stats->packets++;
1584 	rx_stats->bytes += pkt_len;
1585 
1586 	if (act == XDP_TX)
1587 		rx_stats->xdp_tx++;
1588 	u64_stats_update_end(&rx_stats->syncp);
1589 
1590 	if (act == XDP_TX) {
1591 		skb_set_queue_mapping(skb, rxq_idx);
1592 		mana_xdp_tx(skb, ndev);
1593 		return;
1594 	}
1595 
1596 	napi_gro_receive(napi, skb);
1597 
1598 	return;
1599 
1600 drop_xdp:
1601 	u64_stats_update_begin(&rx_stats->syncp);
1602 	rx_stats->xdp_drop++;
1603 	u64_stats_update_end(&rx_stats->syncp);
1604 
1605 drop:
1606 	if (from_pool) {
1607 		page_pool_recycle_direct(rxq->page_pool,
1608 					 virt_to_head_page(buf_va));
1609 	} else {
1610 		WARN_ON_ONCE(rxq->xdp_save_va);
1611 		/* Save for reuse */
1612 		rxq->xdp_save_va = buf_va;
1613 	}
1614 
1615 	++ndev->stats.rx_dropped;
1616 
1617 	return;
1618 }
1619 
1620 static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,
1621 			     dma_addr_t *da, bool *from_pool, bool is_napi)
1622 {
1623 	struct page *page;
1624 	void *va;
1625 
1626 	*from_pool = false;
1627 
1628 	/* Reuse XDP dropped page if available */
1629 	if (rxq->xdp_save_va) {
1630 		va = rxq->xdp_save_va;
1631 		rxq->xdp_save_va = NULL;
1632 	} else if (rxq->alloc_size > PAGE_SIZE) {
1633 		if (is_napi)
1634 			va = napi_alloc_frag(rxq->alloc_size);
1635 		else
1636 			va = netdev_alloc_frag(rxq->alloc_size);
1637 
1638 		if (!va)
1639 			return NULL;
1640 
1641 		page = virt_to_head_page(va);
1642 		/* Check if the frag falls back to single page */
1643 		if (compound_order(page) < get_order(rxq->alloc_size)) {
1644 			put_page(page);
1645 			return NULL;
1646 		}
1647 	} else {
1648 		page = page_pool_dev_alloc_pages(rxq->page_pool);
1649 		if (!page)
1650 			return NULL;
1651 
1652 		*from_pool = true;
1653 		va = page_to_virt(page);
1654 	}
1655 
1656 	*da = dma_map_single(dev, va + rxq->headroom, rxq->datasize,
1657 			     DMA_FROM_DEVICE);
1658 	if (dma_mapping_error(dev, *da)) {
1659 		if (*from_pool)
1660 			page_pool_put_full_page(rxq->page_pool, page, false);
1661 		else
1662 			put_page(virt_to_head_page(va));
1663 
1664 		return NULL;
1665 	}
1666 
1667 	return va;
1668 }
1669 
1670 /* Allocate frag for rx buffer, and save the old buf */
1671 static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq,
1672 			       struct mana_recv_buf_oob *rxoob, void **old_buf,
1673 			       bool *old_fp)
1674 {
1675 	bool from_pool;
1676 	dma_addr_t da;
1677 	void *va;
1678 
1679 	va = mana_get_rxfrag(rxq, dev, &da, &from_pool, true);
1680 	if (!va)
1681 		return;
1682 
1683 	dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize,
1684 			 DMA_FROM_DEVICE);
1685 	*old_buf = rxoob->buf_va;
1686 	*old_fp = rxoob->from_pool;
1687 
1688 	rxoob->buf_va = va;
1689 	rxoob->sgl[0].address = da;
1690 	rxoob->from_pool = from_pool;
1691 }
1692 
1693 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,
1694 				struct gdma_comp *cqe)
1695 {
1696 	struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data;
1697 	struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1698 	struct net_device *ndev = rxq->ndev;
1699 	struct mana_recv_buf_oob *rxbuf_oob;
1700 	struct mana_port_context *apc;
1701 	struct device *dev = gc->dev;
1702 	void *old_buf = NULL;
1703 	u32 curr, pktlen;
1704 	bool old_fp;
1705 
1706 	apc = netdev_priv(ndev);
1707 
1708 	switch (oob->cqe_hdr.cqe_type) {
1709 	case CQE_RX_OKAY:
1710 		break;
1711 
1712 	case CQE_RX_TRUNCATED:
1713 		++ndev->stats.rx_dropped;
1714 		rxbuf_oob = &rxq->rx_oobs[rxq->buf_index];
1715 		netdev_warn_once(ndev, "Dropped a truncated packet\n");
1716 		goto drop;
1717 
1718 	case CQE_RX_COALESCED_4:
1719 		netdev_err(ndev, "RX coalescing is unsupported\n");
1720 		apc->eth_stats.rx_coalesced_err++;
1721 		return;
1722 
1723 	case CQE_RX_OBJECT_FENCE:
1724 		complete(&rxq->fence_event);
1725 		return;
1726 
1727 	default:
1728 		netdev_err(ndev, "Unknown RX CQE type = %d\n",
1729 			   oob->cqe_hdr.cqe_type);
1730 		apc->eth_stats.rx_cqe_unknown_type++;
1731 		return;
1732 	}
1733 
1734 	pktlen = oob->ppi[0].pkt_len;
1735 
1736 	if (pktlen == 0) {
1737 		/* data packets should never have packetlength of zero */
1738 		netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n",
1739 			   rxq->gdma_id, cq->gdma_id, rxq->rxobj);
1740 		return;
1741 	}
1742 
1743 	curr = rxq->buf_index;
1744 	rxbuf_oob = &rxq->rx_oobs[curr];
1745 	WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1);
1746 
1747 	mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf, &old_fp);
1748 
1749 	/* Unsuccessful refill will have old_buf == NULL.
1750 	 * In this case, mana_rx_skb() will drop the packet.
1751 	 */
1752 	mana_rx_skb(old_buf, old_fp, oob, rxq);
1753 
1754 drop:
1755 	mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu);
1756 
1757 	mana_post_pkt_rxq(rxq);
1758 }
1759 
1760 static void mana_poll_rx_cq(struct mana_cq *cq)
1761 {
1762 	struct gdma_comp *comp = cq->gdma_comp_buf;
1763 	struct mana_rxq *rxq = cq->rxq;
1764 	int comp_read, i;
1765 
1766 	comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER);
1767 	WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER);
1768 
1769 	rxq->xdp_flush = false;
1770 
1771 	for (i = 0; i < comp_read; i++) {
1772 		if (WARN_ON_ONCE(comp[i].is_sq))
1773 			return;
1774 
1775 		/* verify recv cqe references the right rxq */
1776 		if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id))
1777 			return;
1778 
1779 		mana_process_rx_cqe(rxq, cq, &comp[i]);
1780 	}
1781 
1782 	if (comp_read > 0) {
1783 		struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1784 
1785 		mana_gd_wq_ring_doorbell(gc, rxq->gdma_rq);
1786 	}
1787 
1788 	if (rxq->xdp_flush)
1789 		xdp_do_flush();
1790 }
1791 
1792 static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue)
1793 {
1794 	struct mana_cq *cq = context;
1795 	int w;
1796 
1797 	WARN_ON_ONCE(cq->gdma_cq != gdma_queue);
1798 
1799 	if (cq->type == MANA_CQ_TYPE_RX)
1800 		mana_poll_rx_cq(cq);
1801 	else
1802 		mana_poll_tx_cq(cq);
1803 
1804 	w = cq->work_done;
1805 	cq->work_done_since_doorbell += w;
1806 
1807 	if (w < cq->budget) {
1808 		mana_gd_ring_cq(gdma_queue, SET_ARM_BIT);
1809 		cq->work_done_since_doorbell = 0;
1810 		napi_complete_done(&cq->napi, w);
1811 	} else if (cq->work_done_since_doorbell >
1812 		   cq->gdma_cq->queue_size / COMP_ENTRY_SIZE * 4) {
1813 		/* MANA hardware requires at least one doorbell ring every 8
1814 		 * wraparounds of CQ even if there is no need to arm the CQ.
1815 		 * This driver rings the doorbell as soon as we have exceeded
1816 		 * 4 wraparounds.
1817 		 */
1818 		mana_gd_ring_cq(gdma_queue, 0);
1819 		cq->work_done_since_doorbell = 0;
1820 	}
1821 
1822 	return w;
1823 }
1824 
1825 static int mana_poll(struct napi_struct *napi, int budget)
1826 {
1827 	struct mana_cq *cq = container_of(napi, struct mana_cq, napi);
1828 	int w;
1829 
1830 	cq->work_done = 0;
1831 	cq->budget = budget;
1832 
1833 	w = mana_cq_handler(cq, cq->gdma_cq);
1834 
1835 	return min(w, budget);
1836 }
1837 
1838 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue)
1839 {
1840 	struct mana_cq *cq = context;
1841 
1842 	napi_schedule_irqoff(&cq->napi);
1843 }
1844 
1845 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq)
1846 {
1847 	struct gdma_dev *gd = apc->ac->gdma_dev;
1848 
1849 	if (!cq->gdma_cq)
1850 		return;
1851 
1852 	mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq);
1853 }
1854 
1855 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq)
1856 {
1857 	struct gdma_dev *gd = apc->ac->gdma_dev;
1858 
1859 	if (!txq->gdma_sq)
1860 		return;
1861 
1862 	mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq);
1863 }
1864 
1865 static void mana_destroy_txq(struct mana_port_context *apc)
1866 {
1867 	struct napi_struct *napi;
1868 	int i;
1869 
1870 	if (!apc->tx_qp)
1871 		return;
1872 
1873 	for (i = 0; i < apc->num_queues; i++) {
1874 		napi = &apc->tx_qp[i].tx_cq.napi;
1875 		if (apc->tx_qp[i].txq.napi_initialized) {
1876 			napi_synchronize(napi);
1877 			napi_disable(napi);
1878 			netif_napi_del(napi);
1879 			apc->tx_qp[i].txq.napi_initialized = false;
1880 		}
1881 		mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object);
1882 
1883 		mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq);
1884 
1885 		mana_deinit_txq(apc, &apc->tx_qp[i].txq);
1886 	}
1887 
1888 	kfree(apc->tx_qp);
1889 	apc->tx_qp = NULL;
1890 }
1891 
1892 static int mana_create_txq(struct mana_port_context *apc,
1893 			   struct net_device *net)
1894 {
1895 	struct mana_context *ac = apc->ac;
1896 	struct gdma_dev *gd = ac->gdma_dev;
1897 	struct mana_obj_spec wq_spec;
1898 	struct mana_obj_spec cq_spec;
1899 	struct gdma_queue_spec spec;
1900 	struct gdma_context *gc;
1901 	struct mana_txq *txq;
1902 	struct mana_cq *cq;
1903 	u32 txq_size;
1904 	u32 cq_size;
1905 	int err;
1906 	int i;
1907 
1908 	apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp),
1909 			     GFP_KERNEL);
1910 	if (!apc->tx_qp)
1911 		return -ENOMEM;
1912 
1913 	/*  The minimum size of the WQE is 32 bytes, hence
1914 	 *  apc->tx_queue_size represents the maximum number of WQEs
1915 	 *  the SQ can store. This value is then used to size other queues
1916 	 *  to prevent overflow.
1917 	 *  Also note that the txq_size is always going to be MANA_PAGE_ALIGNED,
1918 	 *  as min val of apc->tx_queue_size is 128 and that would make
1919 	 *  txq_size 128*32 = 4096 and the other higher values of apc->tx_queue_size
1920 	 *  are always power of two
1921 	 */
1922 	txq_size = apc->tx_queue_size * 32;
1923 
1924 	cq_size = apc->tx_queue_size * COMP_ENTRY_SIZE;
1925 
1926 	gc = gd->gdma_context;
1927 
1928 	for (i = 0; i < apc->num_queues; i++) {
1929 		apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE;
1930 
1931 		/* Create SQ */
1932 		txq = &apc->tx_qp[i].txq;
1933 
1934 		u64_stats_init(&txq->stats.syncp);
1935 		txq->ndev = net;
1936 		txq->net_txq = netdev_get_tx_queue(net, i);
1937 		txq->vp_offset = apc->tx_vp_offset;
1938 		txq->napi_initialized = false;
1939 		skb_queue_head_init(&txq->pending_skbs);
1940 
1941 		memset(&spec, 0, sizeof(spec));
1942 		spec.type = GDMA_SQ;
1943 		spec.monitor_avl_buf = true;
1944 		spec.queue_size = txq_size;
1945 		err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq);
1946 		if (err)
1947 			goto out;
1948 
1949 		/* Create SQ's CQ */
1950 		cq = &apc->tx_qp[i].tx_cq;
1951 		cq->type = MANA_CQ_TYPE_TX;
1952 
1953 		cq->txq = txq;
1954 
1955 		memset(&spec, 0, sizeof(spec));
1956 		spec.type = GDMA_CQ;
1957 		spec.monitor_avl_buf = false;
1958 		spec.queue_size = cq_size;
1959 		spec.cq.callback = mana_schedule_napi;
1960 		spec.cq.parent_eq = ac->eqs[i].eq;
1961 		spec.cq.context = cq;
1962 		err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
1963 		if (err)
1964 			goto out;
1965 
1966 		memset(&wq_spec, 0, sizeof(wq_spec));
1967 		memset(&cq_spec, 0, sizeof(cq_spec));
1968 
1969 		wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle;
1970 		wq_spec.queue_size = txq->gdma_sq->queue_size;
1971 
1972 		cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
1973 		cq_spec.queue_size = cq->gdma_cq->queue_size;
1974 		cq_spec.modr_ctx_id = 0;
1975 		cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
1976 
1977 		err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ,
1978 					 &wq_spec, &cq_spec,
1979 					 &apc->tx_qp[i].tx_object);
1980 
1981 		if (err)
1982 			goto out;
1983 
1984 		txq->gdma_sq->id = wq_spec.queue_index;
1985 		cq->gdma_cq->id = cq_spec.queue_index;
1986 
1987 		txq->gdma_sq->mem_info.dma_region_handle =
1988 			GDMA_INVALID_DMA_REGION;
1989 		cq->gdma_cq->mem_info.dma_region_handle =
1990 			GDMA_INVALID_DMA_REGION;
1991 
1992 		txq->gdma_txq_id = txq->gdma_sq->id;
1993 
1994 		cq->gdma_id = cq->gdma_cq->id;
1995 
1996 		if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
1997 			err = -EINVAL;
1998 			goto out;
1999 		}
2000 
2001 		gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2002 
2003 		netif_napi_add_tx(net, &cq->napi, mana_poll);
2004 		napi_enable(&cq->napi);
2005 		txq->napi_initialized = true;
2006 
2007 		mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2008 	}
2009 
2010 	return 0;
2011 out:
2012 	mana_destroy_txq(apc);
2013 	return err;
2014 }
2015 
2016 static void mana_destroy_rxq(struct mana_port_context *apc,
2017 			     struct mana_rxq *rxq, bool napi_initialized)
2018 
2019 {
2020 	struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
2021 	struct mana_recv_buf_oob *rx_oob;
2022 	struct device *dev = gc->dev;
2023 	struct napi_struct *napi;
2024 	struct page *page;
2025 	int i;
2026 
2027 	if (!rxq)
2028 		return;
2029 
2030 	napi = &rxq->rx_cq.napi;
2031 
2032 	if (napi_initialized) {
2033 		napi_synchronize(napi);
2034 
2035 		napi_disable(napi);
2036 
2037 		netif_napi_del(napi);
2038 	}
2039 	xdp_rxq_info_unreg(&rxq->xdp_rxq);
2040 
2041 	mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj);
2042 
2043 	mana_deinit_cq(apc, &rxq->rx_cq);
2044 
2045 	if (rxq->xdp_save_va)
2046 		put_page(virt_to_head_page(rxq->xdp_save_va));
2047 
2048 	for (i = 0; i < rxq->num_rx_buf; i++) {
2049 		rx_oob = &rxq->rx_oobs[i];
2050 
2051 		if (!rx_oob->buf_va)
2052 			continue;
2053 
2054 		dma_unmap_single(dev, rx_oob->sgl[0].address,
2055 				 rx_oob->sgl[0].size, DMA_FROM_DEVICE);
2056 
2057 		page = virt_to_head_page(rx_oob->buf_va);
2058 
2059 		if (rx_oob->from_pool)
2060 			page_pool_put_full_page(rxq->page_pool, page, false);
2061 		else
2062 			put_page(page);
2063 
2064 		rx_oob->buf_va = NULL;
2065 	}
2066 
2067 	page_pool_destroy(rxq->page_pool);
2068 
2069 	if (rxq->gdma_rq)
2070 		mana_gd_destroy_queue(gc, rxq->gdma_rq);
2071 
2072 	kfree(rxq);
2073 }
2074 
2075 static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key,
2076 			    struct mana_rxq *rxq, struct device *dev)
2077 {
2078 	struct mana_port_context *mpc = netdev_priv(rxq->ndev);
2079 	bool from_pool = false;
2080 	dma_addr_t da;
2081 	void *va;
2082 
2083 	if (mpc->rxbufs_pre)
2084 		va = mana_get_rxbuf_pre(rxq, &da);
2085 	else
2086 		va = mana_get_rxfrag(rxq, dev, &da, &from_pool, false);
2087 
2088 	if (!va)
2089 		return -ENOMEM;
2090 
2091 	rx_oob->buf_va = va;
2092 	rx_oob->from_pool = from_pool;
2093 
2094 	rx_oob->sgl[0].address = da;
2095 	rx_oob->sgl[0].size = rxq->datasize;
2096 	rx_oob->sgl[0].mem_key = mem_key;
2097 
2098 	return 0;
2099 }
2100 
2101 #define MANA_WQE_HEADER_SIZE 16
2102 #define MANA_WQE_SGE_SIZE 16
2103 
2104 static int mana_alloc_rx_wqe(struct mana_port_context *apc,
2105 			     struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size)
2106 {
2107 	struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
2108 	struct mana_recv_buf_oob *rx_oob;
2109 	struct device *dev = gc->dev;
2110 	u32 buf_idx;
2111 	int ret;
2112 
2113 	WARN_ON(rxq->datasize == 0);
2114 
2115 	*rxq_size = 0;
2116 	*cq_size = 0;
2117 
2118 	for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2119 		rx_oob = &rxq->rx_oobs[buf_idx];
2120 		memset(rx_oob, 0, sizeof(*rx_oob));
2121 
2122 		rx_oob->num_sge = 1;
2123 
2124 		ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq,
2125 				       dev);
2126 		if (ret)
2127 			return ret;
2128 
2129 		rx_oob->wqe_req.sgl = rx_oob->sgl;
2130 		rx_oob->wqe_req.num_sge = rx_oob->num_sge;
2131 		rx_oob->wqe_req.inline_oob_size = 0;
2132 		rx_oob->wqe_req.inline_oob_data = NULL;
2133 		rx_oob->wqe_req.flags = 0;
2134 		rx_oob->wqe_req.client_data_unit = 0;
2135 
2136 		*rxq_size += ALIGN(MANA_WQE_HEADER_SIZE +
2137 				   MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32);
2138 		*cq_size += COMP_ENTRY_SIZE;
2139 	}
2140 
2141 	return 0;
2142 }
2143 
2144 static int mana_push_wqe(struct mana_rxq *rxq)
2145 {
2146 	struct mana_recv_buf_oob *rx_oob;
2147 	u32 buf_idx;
2148 	int err;
2149 
2150 	for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2151 		rx_oob = &rxq->rx_oobs[buf_idx];
2152 
2153 		err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req,
2154 					    &rx_oob->wqe_inf);
2155 		if (err)
2156 			return -ENOSPC;
2157 	}
2158 
2159 	return 0;
2160 }
2161 
2162 static int mana_create_page_pool(struct mana_rxq *rxq, struct gdma_context *gc)
2163 {
2164 	struct mana_port_context *mpc = netdev_priv(rxq->ndev);
2165 	struct page_pool_params pprm = {};
2166 	int ret;
2167 
2168 	pprm.pool_size = mpc->rx_queue_size;
2169 	pprm.nid = gc->numa_node;
2170 	pprm.napi = &rxq->rx_cq.napi;
2171 	pprm.netdev = rxq->ndev;
2172 
2173 	rxq->page_pool = page_pool_create(&pprm);
2174 
2175 	if (IS_ERR(rxq->page_pool)) {
2176 		ret = PTR_ERR(rxq->page_pool);
2177 		rxq->page_pool = NULL;
2178 		return ret;
2179 	}
2180 
2181 	return 0;
2182 }
2183 
2184 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,
2185 					u32 rxq_idx, struct mana_eq *eq,
2186 					struct net_device *ndev)
2187 {
2188 	struct gdma_dev *gd = apc->ac->gdma_dev;
2189 	struct mana_obj_spec wq_spec;
2190 	struct mana_obj_spec cq_spec;
2191 	struct gdma_queue_spec spec;
2192 	struct mana_cq *cq = NULL;
2193 	struct gdma_context *gc;
2194 	u32 cq_size, rq_size;
2195 	struct mana_rxq *rxq;
2196 	int err;
2197 
2198 	gc = gd->gdma_context;
2199 
2200 	rxq = kzalloc(struct_size(rxq, rx_oobs, apc->rx_queue_size),
2201 		      GFP_KERNEL);
2202 	if (!rxq)
2203 		return NULL;
2204 
2205 	rxq->ndev = ndev;
2206 	rxq->num_rx_buf = apc->rx_queue_size;
2207 	rxq->rxq_idx = rxq_idx;
2208 	rxq->rxobj = INVALID_MANA_HANDLE;
2209 
2210 	mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size,
2211 			   &rxq->headroom);
2212 
2213 	/* Create page pool for RX queue */
2214 	err = mana_create_page_pool(rxq, gc);
2215 	if (err) {
2216 		netdev_err(ndev, "Create page pool err:%d\n", err);
2217 		goto out;
2218 	}
2219 
2220 	err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size);
2221 	if (err)
2222 		goto out;
2223 
2224 	rq_size = MANA_PAGE_ALIGN(rq_size);
2225 	cq_size = MANA_PAGE_ALIGN(cq_size);
2226 
2227 	/* Create RQ */
2228 	memset(&spec, 0, sizeof(spec));
2229 	spec.type = GDMA_RQ;
2230 	spec.monitor_avl_buf = true;
2231 	spec.queue_size = rq_size;
2232 	err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq);
2233 	if (err)
2234 		goto out;
2235 
2236 	/* Create RQ's CQ */
2237 	cq = &rxq->rx_cq;
2238 	cq->type = MANA_CQ_TYPE_RX;
2239 	cq->rxq = rxq;
2240 
2241 	memset(&spec, 0, sizeof(spec));
2242 	spec.type = GDMA_CQ;
2243 	spec.monitor_avl_buf = false;
2244 	spec.queue_size = cq_size;
2245 	spec.cq.callback = mana_schedule_napi;
2246 	spec.cq.parent_eq = eq->eq;
2247 	spec.cq.context = cq;
2248 	err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2249 	if (err)
2250 		goto out;
2251 
2252 	memset(&wq_spec, 0, sizeof(wq_spec));
2253 	memset(&cq_spec, 0, sizeof(cq_spec));
2254 	wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle;
2255 	wq_spec.queue_size = rxq->gdma_rq->queue_size;
2256 
2257 	cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2258 	cq_spec.queue_size = cq->gdma_cq->queue_size;
2259 	cq_spec.modr_ctx_id = 0;
2260 	cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2261 
2262 	err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ,
2263 				 &wq_spec, &cq_spec, &rxq->rxobj);
2264 	if (err)
2265 		goto out;
2266 
2267 	rxq->gdma_rq->id = wq_spec.queue_index;
2268 	cq->gdma_cq->id = cq_spec.queue_index;
2269 
2270 	rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2271 	cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2272 
2273 	rxq->gdma_id = rxq->gdma_rq->id;
2274 	cq->gdma_id = cq->gdma_cq->id;
2275 
2276 	err = mana_push_wqe(rxq);
2277 	if (err)
2278 		goto out;
2279 
2280 	if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
2281 		err = -EINVAL;
2282 		goto out;
2283 	}
2284 
2285 	gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2286 
2287 	netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1);
2288 
2289 	WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx,
2290 				 cq->napi.napi_id));
2291 	WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL,
2292 					   rxq->page_pool));
2293 
2294 	napi_enable(&cq->napi);
2295 
2296 	mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2297 out:
2298 	if (!err)
2299 		return rxq;
2300 
2301 	netdev_err(ndev, "Failed to create RXQ: err = %d\n", err);
2302 
2303 	mana_destroy_rxq(apc, rxq, false);
2304 
2305 	if (cq)
2306 		mana_deinit_cq(apc, cq);
2307 
2308 	return NULL;
2309 }
2310 
2311 static int mana_add_rx_queues(struct mana_port_context *apc,
2312 			      struct net_device *ndev)
2313 {
2314 	struct mana_context *ac = apc->ac;
2315 	struct mana_rxq *rxq;
2316 	int err = 0;
2317 	int i;
2318 
2319 	for (i = 0; i < apc->num_queues; i++) {
2320 		rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev);
2321 		if (!rxq) {
2322 			err = -ENOMEM;
2323 			goto out;
2324 		}
2325 
2326 		u64_stats_init(&rxq->stats.syncp);
2327 
2328 		apc->rxqs[i] = rxq;
2329 	}
2330 
2331 	apc->default_rxobj = apc->rxqs[0]->rxobj;
2332 out:
2333 	return err;
2334 }
2335 
2336 static void mana_destroy_vport(struct mana_port_context *apc)
2337 {
2338 	struct gdma_dev *gd = apc->ac->gdma_dev;
2339 	struct mana_rxq *rxq;
2340 	u32 rxq_idx;
2341 
2342 	for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
2343 		rxq = apc->rxqs[rxq_idx];
2344 		if (!rxq)
2345 			continue;
2346 
2347 		mana_destroy_rxq(apc, rxq, true);
2348 		apc->rxqs[rxq_idx] = NULL;
2349 	}
2350 
2351 	mana_destroy_txq(apc);
2352 	mana_uncfg_vport(apc);
2353 
2354 	if (gd->gdma_context->is_pf)
2355 		mana_pf_deregister_hw_vport(apc);
2356 }
2357 
2358 static int mana_create_vport(struct mana_port_context *apc,
2359 			     struct net_device *net)
2360 {
2361 	struct gdma_dev *gd = apc->ac->gdma_dev;
2362 	int err;
2363 
2364 	apc->default_rxobj = INVALID_MANA_HANDLE;
2365 
2366 	if (gd->gdma_context->is_pf) {
2367 		err = mana_pf_register_hw_vport(apc);
2368 		if (err)
2369 			return err;
2370 	}
2371 
2372 	err = mana_cfg_vport(apc, gd->pdid, gd->doorbell);
2373 	if (err)
2374 		return err;
2375 
2376 	return mana_create_txq(apc, net);
2377 }
2378 
2379 static int mana_rss_table_alloc(struct mana_port_context *apc)
2380 {
2381 	if (!apc->indir_table_sz) {
2382 		netdev_err(apc->ndev,
2383 			   "Indirection table size not set for vPort %d\n",
2384 			   apc->port_idx);
2385 		return -EINVAL;
2386 	}
2387 
2388 	apc->indir_table = kcalloc(apc->indir_table_sz, sizeof(u32), GFP_KERNEL);
2389 	if (!apc->indir_table)
2390 		return -ENOMEM;
2391 
2392 	apc->rxobj_table = kcalloc(apc->indir_table_sz, sizeof(mana_handle_t), GFP_KERNEL);
2393 	if (!apc->rxobj_table) {
2394 		kfree(apc->indir_table);
2395 		return -ENOMEM;
2396 	}
2397 
2398 	return 0;
2399 }
2400 
2401 static void mana_rss_table_init(struct mana_port_context *apc)
2402 {
2403 	int i;
2404 
2405 	for (i = 0; i < apc->indir_table_sz; i++)
2406 		apc->indir_table[i] =
2407 			ethtool_rxfh_indir_default(i, apc->num_queues);
2408 }
2409 
2410 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx,
2411 		    bool update_hash, bool update_tab)
2412 {
2413 	u32 queue_idx;
2414 	int err;
2415 	int i;
2416 
2417 	if (update_tab) {
2418 		for (i = 0; i < apc->indir_table_sz; i++) {
2419 			queue_idx = apc->indir_table[i];
2420 			apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj;
2421 		}
2422 	}
2423 
2424 	err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab);
2425 	if (err)
2426 		return err;
2427 
2428 	mana_fence_rqs(apc);
2429 
2430 	return 0;
2431 }
2432 
2433 void mana_query_gf_stats(struct mana_port_context *apc)
2434 {
2435 	struct mana_query_gf_stat_resp resp = {};
2436 	struct mana_query_gf_stat_req req = {};
2437 	struct net_device *ndev = apc->ndev;
2438 	int err;
2439 
2440 	mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_GF_STAT,
2441 			     sizeof(req), sizeof(resp));
2442 	req.req_stats = STATISTICS_FLAGS_RX_DISCARDS_NO_WQE |
2443 			STATISTICS_FLAGS_RX_ERRORS_VPORT_DISABLED |
2444 			STATISTICS_FLAGS_HC_RX_BYTES |
2445 			STATISTICS_FLAGS_HC_RX_UCAST_PACKETS |
2446 			STATISTICS_FLAGS_HC_RX_UCAST_BYTES |
2447 			STATISTICS_FLAGS_HC_RX_MCAST_PACKETS |
2448 			STATISTICS_FLAGS_HC_RX_MCAST_BYTES |
2449 			STATISTICS_FLAGS_HC_RX_BCAST_PACKETS |
2450 			STATISTICS_FLAGS_HC_RX_BCAST_BYTES |
2451 			STATISTICS_FLAGS_TX_ERRORS_GF_DISABLED |
2452 			STATISTICS_FLAGS_TX_ERRORS_VPORT_DISABLED |
2453 			STATISTICS_FLAGS_TX_ERRORS_INVAL_VPORT_OFFSET_PACKETS |
2454 			STATISTICS_FLAGS_TX_ERRORS_VLAN_ENFORCEMENT |
2455 			STATISTICS_FLAGS_TX_ERRORS_ETH_TYPE_ENFORCEMENT |
2456 			STATISTICS_FLAGS_TX_ERRORS_SA_ENFORCEMENT |
2457 			STATISTICS_FLAGS_TX_ERRORS_SQPDID_ENFORCEMENT |
2458 			STATISTICS_FLAGS_TX_ERRORS_CQPDID_ENFORCEMENT |
2459 			STATISTICS_FLAGS_TX_ERRORS_MTU_VIOLATION |
2460 			STATISTICS_FLAGS_TX_ERRORS_INVALID_OOB |
2461 			STATISTICS_FLAGS_HC_TX_BYTES |
2462 			STATISTICS_FLAGS_HC_TX_UCAST_PACKETS |
2463 			STATISTICS_FLAGS_HC_TX_UCAST_BYTES |
2464 			STATISTICS_FLAGS_HC_TX_MCAST_PACKETS |
2465 			STATISTICS_FLAGS_HC_TX_MCAST_BYTES |
2466 			STATISTICS_FLAGS_HC_TX_BCAST_PACKETS |
2467 			STATISTICS_FLAGS_HC_TX_BCAST_BYTES |
2468 			STATISTICS_FLAGS_TX_ERRORS_GDMA_ERROR;
2469 
2470 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
2471 				sizeof(resp));
2472 	if (err) {
2473 		netdev_err(ndev, "Failed to query GF stats: %d\n", err);
2474 		return;
2475 	}
2476 	err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_GF_STAT,
2477 				   sizeof(resp));
2478 	if (err || resp.hdr.status) {
2479 		netdev_err(ndev, "Failed to query GF stats: %d, 0x%x\n", err,
2480 			   resp.hdr.status);
2481 		return;
2482 	}
2483 
2484 	apc->eth_stats.hc_rx_discards_no_wqe = resp.rx_discards_nowqe;
2485 	apc->eth_stats.hc_rx_err_vport_disabled = resp.rx_err_vport_disabled;
2486 	apc->eth_stats.hc_rx_bytes = resp.hc_rx_bytes;
2487 	apc->eth_stats.hc_rx_ucast_pkts = resp.hc_rx_ucast_pkts;
2488 	apc->eth_stats.hc_rx_ucast_bytes = resp.hc_rx_ucast_bytes;
2489 	apc->eth_stats.hc_rx_bcast_pkts = resp.hc_rx_bcast_pkts;
2490 	apc->eth_stats.hc_rx_bcast_bytes = resp.hc_rx_bcast_bytes;
2491 	apc->eth_stats.hc_rx_mcast_pkts = resp.hc_rx_mcast_pkts;
2492 	apc->eth_stats.hc_rx_mcast_bytes = resp.hc_rx_mcast_bytes;
2493 	apc->eth_stats.hc_tx_err_gf_disabled = resp.tx_err_gf_disabled;
2494 	apc->eth_stats.hc_tx_err_vport_disabled = resp.tx_err_vport_disabled;
2495 	apc->eth_stats.hc_tx_err_inval_vportoffset_pkt =
2496 					     resp.tx_err_inval_vport_offset_pkt;
2497 	apc->eth_stats.hc_tx_err_vlan_enforcement =
2498 					     resp.tx_err_vlan_enforcement;
2499 	apc->eth_stats.hc_tx_err_eth_type_enforcement =
2500 					     resp.tx_err_ethtype_enforcement;
2501 	apc->eth_stats.hc_tx_err_sa_enforcement = resp.tx_err_SA_enforcement;
2502 	apc->eth_stats.hc_tx_err_sqpdid_enforcement =
2503 					     resp.tx_err_SQPDID_enforcement;
2504 	apc->eth_stats.hc_tx_err_cqpdid_enforcement =
2505 					     resp.tx_err_CQPDID_enforcement;
2506 	apc->eth_stats.hc_tx_err_mtu_violation = resp.tx_err_mtu_violation;
2507 	apc->eth_stats.hc_tx_err_inval_oob = resp.tx_err_inval_oob;
2508 	apc->eth_stats.hc_tx_bytes = resp.hc_tx_bytes;
2509 	apc->eth_stats.hc_tx_ucast_pkts = resp.hc_tx_ucast_pkts;
2510 	apc->eth_stats.hc_tx_ucast_bytes = resp.hc_tx_ucast_bytes;
2511 	apc->eth_stats.hc_tx_bcast_pkts = resp.hc_tx_bcast_pkts;
2512 	apc->eth_stats.hc_tx_bcast_bytes = resp.hc_tx_bcast_bytes;
2513 	apc->eth_stats.hc_tx_mcast_pkts = resp.hc_tx_mcast_pkts;
2514 	apc->eth_stats.hc_tx_mcast_bytes = resp.hc_tx_mcast_bytes;
2515 	apc->eth_stats.hc_tx_err_gdma = resp.tx_err_gdma;
2516 }
2517 
2518 static int mana_init_port(struct net_device *ndev)
2519 {
2520 	struct mana_port_context *apc = netdev_priv(ndev);
2521 	u32 max_txq, max_rxq, max_queues;
2522 	int port_idx = apc->port_idx;
2523 	int err;
2524 
2525 	err = mana_init_port_context(apc);
2526 	if (err)
2527 		return err;
2528 
2529 	err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq,
2530 				   &apc->indir_table_sz);
2531 	if (err) {
2532 		netdev_err(ndev, "Failed to query info for vPort %d\n",
2533 			   port_idx);
2534 		goto reset_apc;
2535 	}
2536 
2537 	max_queues = min_t(u32, max_txq, max_rxq);
2538 	if (apc->max_queues > max_queues)
2539 		apc->max_queues = max_queues;
2540 
2541 	if (apc->num_queues > apc->max_queues)
2542 		apc->num_queues = apc->max_queues;
2543 
2544 	eth_hw_addr_set(ndev, apc->mac_addr);
2545 
2546 	return 0;
2547 
2548 reset_apc:
2549 	mana_cleanup_port_context(apc);
2550 	return err;
2551 }
2552 
2553 int mana_alloc_queues(struct net_device *ndev)
2554 {
2555 	struct mana_port_context *apc = netdev_priv(ndev);
2556 	struct gdma_dev *gd = apc->ac->gdma_dev;
2557 	int err;
2558 
2559 	err = mana_create_vport(apc, ndev);
2560 	if (err)
2561 		return err;
2562 
2563 	err = netif_set_real_num_tx_queues(ndev, apc->num_queues);
2564 	if (err)
2565 		goto destroy_vport;
2566 
2567 	err = mana_add_rx_queues(apc, ndev);
2568 	if (err)
2569 		goto destroy_vport;
2570 
2571 	apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE;
2572 
2573 	err = netif_set_real_num_rx_queues(ndev, apc->num_queues);
2574 	if (err)
2575 		goto destroy_vport;
2576 
2577 	mana_rss_table_init(apc);
2578 
2579 	err = mana_config_rss(apc, TRI_STATE_TRUE, true, true);
2580 	if (err)
2581 		goto destroy_vport;
2582 
2583 	if (gd->gdma_context->is_pf) {
2584 		err = mana_pf_register_filter(apc);
2585 		if (err)
2586 			goto destroy_vport;
2587 	}
2588 
2589 	mana_chn_setxdp(apc, mana_xdp_get(apc));
2590 
2591 	return 0;
2592 
2593 destroy_vport:
2594 	mana_destroy_vport(apc);
2595 	return err;
2596 }
2597 
2598 int mana_attach(struct net_device *ndev)
2599 {
2600 	struct mana_port_context *apc = netdev_priv(ndev);
2601 	int err;
2602 
2603 	ASSERT_RTNL();
2604 
2605 	err = mana_init_port(ndev);
2606 	if (err)
2607 		return err;
2608 
2609 	if (apc->port_st_save) {
2610 		err = mana_alloc_queues(ndev);
2611 		if (err) {
2612 			mana_cleanup_port_context(apc);
2613 			return err;
2614 		}
2615 	}
2616 
2617 	apc->port_is_up = apc->port_st_save;
2618 
2619 	/* Ensure port state updated before txq state */
2620 	smp_wmb();
2621 
2622 	if (apc->port_is_up)
2623 		netif_carrier_on(ndev);
2624 
2625 	netif_device_attach(ndev);
2626 
2627 	return 0;
2628 }
2629 
2630 static int mana_dealloc_queues(struct net_device *ndev)
2631 {
2632 	struct mana_port_context *apc = netdev_priv(ndev);
2633 	unsigned long timeout = jiffies + 120 * HZ;
2634 	struct gdma_dev *gd = apc->ac->gdma_dev;
2635 	struct mana_txq *txq;
2636 	struct sk_buff *skb;
2637 	int i, err;
2638 	u32 tsleep;
2639 
2640 	if (apc->port_is_up)
2641 		return -EINVAL;
2642 
2643 	mana_chn_setxdp(apc, NULL);
2644 
2645 	if (gd->gdma_context->is_pf)
2646 		mana_pf_deregister_filter(apc);
2647 
2648 	/* No packet can be transmitted now since apc->port_is_up is false.
2649 	 * There is still a tiny chance that mana_poll_tx_cq() can re-enable
2650 	 * a txq because it may not timely see apc->port_is_up being cleared
2651 	 * to false, but it doesn't matter since mana_start_xmit() drops any
2652 	 * new packets due to apc->port_is_up being false.
2653 	 *
2654 	 * Drain all the in-flight TX packets.
2655 	 * A timeout of 120 seconds for all the queues is used.
2656 	 * This will break the while loop when h/w is not responding.
2657 	 * This value of 120 has been decided here considering max
2658 	 * number of queues.
2659 	 */
2660 
2661 	for (i = 0; i < apc->num_queues; i++) {
2662 		txq = &apc->tx_qp[i].txq;
2663 		tsleep = 1000;
2664 		while (atomic_read(&txq->pending_sends) > 0 &&
2665 		       time_before(jiffies, timeout)) {
2666 			usleep_range(tsleep, tsleep + 1000);
2667 			tsleep <<= 1;
2668 		}
2669 		if (atomic_read(&txq->pending_sends)) {
2670 			err = pcie_flr(to_pci_dev(gd->gdma_context->dev));
2671 			if (err) {
2672 				netdev_err(ndev, "flr failed %d with %d pkts pending in txq %u\n",
2673 					   err, atomic_read(&txq->pending_sends),
2674 					   txq->gdma_txq_id);
2675 			}
2676 			break;
2677 		}
2678 	}
2679 
2680 	for (i = 0; i < apc->num_queues; i++) {
2681 		txq = &apc->tx_qp[i].txq;
2682 		while ((skb = skb_dequeue(&txq->pending_skbs))) {
2683 			mana_unmap_skb(skb, apc);
2684 			dev_kfree_skb_any(skb);
2685 		}
2686 		atomic_set(&txq->pending_sends, 0);
2687 	}
2688 	/* We're 100% sure the queues can no longer be woken up, because
2689 	 * we're sure now mana_poll_tx_cq() can't be running.
2690 	 */
2691 
2692 	apc->rss_state = TRI_STATE_FALSE;
2693 	err = mana_config_rss(apc, TRI_STATE_FALSE, false, false);
2694 	if (err) {
2695 		netdev_err(ndev, "Failed to disable vPort: %d\n", err);
2696 		return err;
2697 	}
2698 
2699 	mana_destroy_vport(apc);
2700 
2701 	return 0;
2702 }
2703 
2704 int mana_detach(struct net_device *ndev, bool from_close)
2705 {
2706 	struct mana_port_context *apc = netdev_priv(ndev);
2707 	int err;
2708 
2709 	ASSERT_RTNL();
2710 
2711 	apc->port_st_save = apc->port_is_up;
2712 	apc->port_is_up = false;
2713 
2714 	/* Ensure port state updated before txq state */
2715 	smp_wmb();
2716 
2717 	netif_tx_disable(ndev);
2718 	netif_carrier_off(ndev);
2719 
2720 	if (apc->port_st_save) {
2721 		err = mana_dealloc_queues(ndev);
2722 		if (err)
2723 			return err;
2724 	}
2725 
2726 	if (!from_close) {
2727 		netif_device_detach(ndev);
2728 		mana_cleanup_port_context(apc);
2729 	}
2730 
2731 	return 0;
2732 }
2733 
2734 static int mana_probe_port(struct mana_context *ac, int port_idx,
2735 			   struct net_device **ndev_storage)
2736 {
2737 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
2738 	struct mana_port_context *apc;
2739 	struct net_device *ndev;
2740 	int err;
2741 
2742 	ndev = alloc_etherdev_mq(sizeof(struct mana_port_context),
2743 				 gc->max_num_queues);
2744 	if (!ndev)
2745 		return -ENOMEM;
2746 
2747 	*ndev_storage = ndev;
2748 
2749 	apc = netdev_priv(ndev);
2750 	apc->ac = ac;
2751 	apc->ndev = ndev;
2752 	apc->max_queues = gc->max_num_queues;
2753 	apc->num_queues = gc->max_num_queues;
2754 	apc->tx_queue_size = DEF_TX_BUFFERS_PER_QUEUE;
2755 	apc->rx_queue_size = DEF_RX_BUFFERS_PER_QUEUE;
2756 	apc->port_handle = INVALID_MANA_HANDLE;
2757 	apc->pf_filter_handle = INVALID_MANA_HANDLE;
2758 	apc->port_idx = port_idx;
2759 
2760 	mutex_init(&apc->vport_mutex);
2761 	apc->vport_use_count = 0;
2762 
2763 	ndev->netdev_ops = &mana_devops;
2764 	ndev->ethtool_ops = &mana_ethtool_ops;
2765 	ndev->mtu = ETH_DATA_LEN;
2766 	ndev->max_mtu = gc->adapter_mtu - ETH_HLEN;
2767 	ndev->min_mtu = ETH_MIN_MTU;
2768 	ndev->needed_headroom = MANA_HEADROOM;
2769 	ndev->dev_port = port_idx;
2770 	SET_NETDEV_DEV(ndev, gc->dev);
2771 
2772 	netif_carrier_off(ndev);
2773 
2774 	netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE);
2775 
2776 	err = mana_init_port(ndev);
2777 	if (err)
2778 		goto free_net;
2779 
2780 	err = mana_rss_table_alloc(apc);
2781 	if (err)
2782 		goto reset_apc;
2783 
2784 	netdev_lockdep_set_classes(ndev);
2785 
2786 	ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
2787 	ndev->hw_features |= NETIF_F_RXCSUM;
2788 	ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
2789 	ndev->hw_features |= NETIF_F_RXHASH;
2790 	ndev->features = ndev->hw_features | NETIF_F_HW_VLAN_CTAG_TX |
2791 			 NETIF_F_HW_VLAN_CTAG_RX;
2792 	ndev->vlan_features = ndev->features;
2793 	xdp_set_features_flag(ndev, NETDEV_XDP_ACT_BASIC |
2794 			      NETDEV_XDP_ACT_REDIRECT |
2795 			      NETDEV_XDP_ACT_NDO_XMIT);
2796 
2797 	err = register_netdev(ndev);
2798 	if (err) {
2799 		netdev_err(ndev, "Unable to register netdev.\n");
2800 		goto free_indir;
2801 	}
2802 
2803 	return 0;
2804 
2805 free_indir:
2806 	mana_cleanup_indir_table(apc);
2807 reset_apc:
2808 	mana_cleanup_port_context(apc);
2809 free_net:
2810 	*ndev_storage = NULL;
2811 	netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err);
2812 	free_netdev(ndev);
2813 	return err;
2814 }
2815 
2816 static void adev_release(struct device *dev)
2817 {
2818 	struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev);
2819 
2820 	kfree(madev);
2821 }
2822 
2823 static void remove_adev(struct gdma_dev *gd)
2824 {
2825 	struct auxiliary_device *adev = gd->adev;
2826 	int id = adev->id;
2827 
2828 	auxiliary_device_delete(adev);
2829 	auxiliary_device_uninit(adev);
2830 
2831 	mana_adev_idx_free(id);
2832 	gd->adev = NULL;
2833 }
2834 
2835 static int add_adev(struct gdma_dev *gd)
2836 {
2837 	struct auxiliary_device *adev;
2838 	struct mana_adev *madev;
2839 	int ret;
2840 
2841 	madev = kzalloc(sizeof(*madev), GFP_KERNEL);
2842 	if (!madev)
2843 		return -ENOMEM;
2844 
2845 	adev = &madev->adev;
2846 	ret = mana_adev_idx_alloc();
2847 	if (ret < 0)
2848 		goto idx_fail;
2849 	adev->id = ret;
2850 
2851 	adev->name = "rdma";
2852 	adev->dev.parent = gd->gdma_context->dev;
2853 	adev->dev.release = adev_release;
2854 	madev->mdev = gd;
2855 
2856 	ret = auxiliary_device_init(adev);
2857 	if (ret)
2858 		goto init_fail;
2859 
2860 	/* madev is owned by the auxiliary device */
2861 	madev = NULL;
2862 	ret = auxiliary_device_add(adev);
2863 	if (ret)
2864 		goto add_fail;
2865 
2866 	gd->adev = adev;
2867 	return 0;
2868 
2869 add_fail:
2870 	auxiliary_device_uninit(adev);
2871 
2872 init_fail:
2873 	mana_adev_idx_free(adev->id);
2874 
2875 idx_fail:
2876 	kfree(madev);
2877 
2878 	return ret;
2879 }
2880 
2881 int mana_probe(struct gdma_dev *gd, bool resuming)
2882 {
2883 	struct gdma_context *gc = gd->gdma_context;
2884 	struct mana_context *ac = gd->driver_data;
2885 	struct device *dev = gc->dev;
2886 	u16 num_ports = 0;
2887 	int err;
2888 	int i;
2889 
2890 	dev_info(dev,
2891 		 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n",
2892 		 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION);
2893 
2894 	err = mana_gd_register_device(gd);
2895 	if (err)
2896 		return err;
2897 
2898 	if (!resuming) {
2899 		ac = kzalloc(sizeof(*ac), GFP_KERNEL);
2900 		if (!ac)
2901 			return -ENOMEM;
2902 
2903 		ac->gdma_dev = gd;
2904 		gd->driver_data = ac;
2905 	}
2906 
2907 	err = mana_create_eq(ac);
2908 	if (err)
2909 		goto out;
2910 
2911 	err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION,
2912 				    MANA_MICRO_VERSION, &num_ports);
2913 	if (err)
2914 		goto out;
2915 
2916 	if (!resuming) {
2917 		ac->num_ports = num_ports;
2918 	} else {
2919 		if (ac->num_ports != num_ports) {
2920 			dev_err(dev, "The number of vPorts changed: %d->%d\n",
2921 				ac->num_ports, num_ports);
2922 			err = -EPROTO;
2923 			goto out;
2924 		}
2925 	}
2926 
2927 	if (ac->num_ports == 0)
2928 		dev_err(dev, "Failed to detect any vPort\n");
2929 
2930 	if (ac->num_ports > MAX_PORTS_IN_MANA_DEV)
2931 		ac->num_ports = MAX_PORTS_IN_MANA_DEV;
2932 
2933 	if (!resuming) {
2934 		for (i = 0; i < ac->num_ports; i++) {
2935 			err = mana_probe_port(ac, i, &ac->ports[i]);
2936 			/* we log the port for which the probe failed and stop
2937 			 * probes for subsequent ports.
2938 			 * Note that we keep running ports, for which the probes
2939 			 * were successful, unless add_adev fails too
2940 			 */
2941 			if (err) {
2942 				dev_err(dev, "Probe Failed for port %d\n", i);
2943 				break;
2944 			}
2945 		}
2946 	} else {
2947 		for (i = 0; i < ac->num_ports; i++) {
2948 			rtnl_lock();
2949 			err = mana_attach(ac->ports[i]);
2950 			rtnl_unlock();
2951 			/* we log the port for which the attach failed and stop
2952 			 * attach for subsequent ports
2953 			 * Note that we keep running ports, for which the attach
2954 			 * were successful, unless add_adev fails too
2955 			 */
2956 			if (err) {
2957 				dev_err(dev, "Attach Failed for port %d\n", i);
2958 				break;
2959 			}
2960 		}
2961 	}
2962 
2963 	err = add_adev(gd);
2964 out:
2965 	if (err)
2966 		mana_remove(gd, false);
2967 
2968 	return err;
2969 }
2970 
2971 void mana_remove(struct gdma_dev *gd, bool suspending)
2972 {
2973 	struct gdma_context *gc = gd->gdma_context;
2974 	struct mana_context *ac = gd->driver_data;
2975 	struct mana_port_context *apc;
2976 	struct device *dev = gc->dev;
2977 	struct net_device *ndev;
2978 	int err;
2979 	int i;
2980 
2981 	/* adev currently doesn't support suspending, always remove it */
2982 	if (gd->adev)
2983 		remove_adev(gd);
2984 
2985 	for (i = 0; i < ac->num_ports; i++) {
2986 		ndev = ac->ports[i];
2987 		apc = netdev_priv(ndev);
2988 		if (!ndev) {
2989 			if (i == 0)
2990 				dev_err(dev, "No net device to remove\n");
2991 			goto out;
2992 		}
2993 
2994 		/* All cleanup actions should stay after rtnl_lock(), otherwise
2995 		 * other functions may access partially cleaned up data.
2996 		 */
2997 		rtnl_lock();
2998 
2999 		err = mana_detach(ndev, false);
3000 		if (err)
3001 			netdev_err(ndev, "Failed to detach vPort %d: %d\n",
3002 				   i, err);
3003 
3004 		if (suspending) {
3005 			/* No need to unregister the ndev. */
3006 			rtnl_unlock();
3007 			continue;
3008 		}
3009 
3010 		unregister_netdevice(ndev);
3011 		mana_cleanup_indir_table(apc);
3012 
3013 		rtnl_unlock();
3014 
3015 		free_netdev(ndev);
3016 	}
3017 
3018 	mana_destroy_eq(ac);
3019 out:
3020 	mana_gd_deregister_device(gd);
3021 
3022 	if (suspending)
3023 		return;
3024 
3025 	gd->driver_data = NULL;
3026 	gd->gdma_context = NULL;
3027 	kfree(ac);
3028 }
3029 
3030 struct net_device *mana_get_primary_netdev_rcu(struct mana_context *ac, u32 port_index)
3031 {
3032 	struct net_device *ndev;
3033 
3034 	RCU_LOCKDEP_WARN(!rcu_read_lock_held(),
3035 			 "Taking primary netdev without holding the RCU read lock");
3036 	if (port_index >= ac->num_ports)
3037 		return NULL;
3038 
3039 	/* When mana is used in netvsc, the upper netdevice should be returned. */
3040 	if (ac->ports[port_index]->flags & IFF_SLAVE)
3041 		ndev = netdev_master_upper_dev_get_rcu(ac->ports[port_index]);
3042 	else
3043 		ndev = ac->ports[port_index];
3044 
3045 	return ndev;
3046 }
3047 EXPORT_SYMBOL_NS(mana_get_primary_netdev_rcu, NET_MANA);
3048