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 
mana_adev_idx_alloc(void)23 static int mana_adev_idx_alloc(void)
24 {
25 	return ida_alloc(&mana_adev_ida, GFP_KERNEL);
26 }
27 
mana_adev_idx_free(int idx)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 
mana_open(struct net_device * ndev)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 
mana_close(struct net_device * ndev)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 
mana_can_tx(struct gdma_queue * wq)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 
mana_checksum_info(struct sk_buff * skb)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 
mana_add_sge(struct mana_tx_package * tp,struct mana_skb_head * ash,int sg_i,dma_addr_t da,int sge_len,u32 gpa_mkey)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 
mana_map_skb(struct sk_buff * skb,struct mana_port_context * apc,struct mana_tx_package * tp,int gso_hs)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  */
mana_fix_skb_head(struct net_device * ndev,struct sk_buff * skb,int gso_hs)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 */
mana_get_gso_hs(struct sk_buff * skb)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 
mana_start_xmit(struct sk_buff * skb,struct net_device * ndev)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 
mana_get_stats64(struct net_device * ndev,struct rtnl_link_stats64 * st)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 
mana_get_tx_queue(struct net_device * ndev,struct sk_buff * skb,int old_q)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 & MANA_INDIRECT_TABLE_MASK];
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 
mana_select_queue(struct net_device * ndev,struct sk_buff * skb,struct net_device * sb_dev)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 */
mana_pre_dealloc_rxbufs(struct mana_port_context * mpc)514 static 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 */
mana_get_rxbuf_pre(struct mana_rxq * rxq,dma_addr_t * da)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 */
mana_get_rxbuf_cfg(int mtu,u32 * datasize,u32 * alloc_size,u32 * headroom)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 
mana_pre_alloc_rxbufs(struct mana_port_context * mpc,int new_mtu)611 static int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu)
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 = mpc->num_queues * RX_BUFFERS_PER_QUEUE;
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 
mana_change_mtu(struct net_device * ndev,int new_mtu)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);
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 	ndev->mtu = new_mtu;
698 
699 	err = mana_attach(ndev);
700 	if (err) {
701 		netdev_err(ndev, "mana_attach failed: %d\n", err);
702 		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 
mana_cleanup_port_context(struct mana_port_context * apc)722 static void mana_cleanup_port_context(struct mana_port_context *apc)
723 {
724 	kfree(apc->rxqs);
725 	apc->rxqs = NULL;
726 }
727 
mana_init_port_context(struct mana_port_context * apc)728 static int mana_init_port_context(struct mana_port_context *apc)
729 {
730 	apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *),
731 			    GFP_KERNEL);
732 
733 	return !apc->rxqs ? -ENOMEM : 0;
734 }
735 
mana_send_request(struct mana_context * ac,void * in_buf,u32 in_len,void * out_buf,u32 out_len)736 static int mana_send_request(struct mana_context *ac, void *in_buf,
737 			     u32 in_len, void *out_buf, u32 out_len)
738 {
739 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
740 	struct gdma_resp_hdr *resp = out_buf;
741 	struct gdma_req_hdr *req = in_buf;
742 	struct device *dev = gc->dev;
743 	static atomic_t activity_id;
744 	int err;
745 
746 	req->dev_id = gc->mana.dev_id;
747 	req->activity_id = atomic_inc_return(&activity_id);
748 
749 	err = mana_gd_send_request(gc, in_len, in_buf, out_len,
750 				   out_buf);
751 	if (err || resp->status) {
752 		dev_err(dev, "Failed to send mana message: %d, 0x%x\n",
753 			err, resp->status);
754 		return err ? err : -EPROTO;
755 	}
756 
757 	if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 ||
758 	    req->activity_id != resp->activity_id) {
759 		dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n",
760 			req->dev_id.as_uint32, resp->dev_id.as_uint32,
761 			req->activity_id, resp->activity_id);
762 		return -EPROTO;
763 	}
764 
765 	return 0;
766 }
767 
mana_verify_resp_hdr(const struct gdma_resp_hdr * resp_hdr,const enum mana_command_code expected_code,const u32 min_size)768 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr,
769 				const enum mana_command_code expected_code,
770 				const u32 min_size)
771 {
772 	if (resp_hdr->response.msg_type != expected_code)
773 		return -EPROTO;
774 
775 	if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1)
776 		return -EPROTO;
777 
778 	if (resp_hdr->response.msg_size < min_size)
779 		return -EPROTO;
780 
781 	return 0;
782 }
783 
mana_pf_register_hw_vport(struct mana_port_context * apc)784 static int mana_pf_register_hw_vport(struct mana_port_context *apc)
785 {
786 	struct mana_register_hw_vport_resp resp = {};
787 	struct mana_register_hw_vport_req req = {};
788 	int err;
789 
790 	mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT,
791 			     sizeof(req), sizeof(resp));
792 	req.attached_gfid = 1;
793 	req.is_pf_default_vport = 1;
794 	req.allow_all_ether_types = 1;
795 
796 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
797 				sizeof(resp));
798 	if (err) {
799 		netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err);
800 		return err;
801 	}
802 
803 	err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT,
804 				   sizeof(resp));
805 	if (err || resp.hdr.status) {
806 		netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n",
807 			   err, resp.hdr.status);
808 		return err ? err : -EPROTO;
809 	}
810 
811 	apc->port_handle = resp.hw_vport_handle;
812 	return 0;
813 }
814 
mana_pf_deregister_hw_vport(struct mana_port_context * apc)815 static void mana_pf_deregister_hw_vport(struct mana_port_context *apc)
816 {
817 	struct mana_deregister_hw_vport_resp resp = {};
818 	struct mana_deregister_hw_vport_req req = {};
819 	int err;
820 
821 	mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT,
822 			     sizeof(req), sizeof(resp));
823 	req.hw_vport_handle = apc->port_handle;
824 
825 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
826 				sizeof(resp));
827 	if (err) {
828 		netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n",
829 			   err);
830 		return;
831 	}
832 
833 	err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT,
834 				   sizeof(resp));
835 	if (err || resp.hdr.status)
836 		netdev_err(apc->ndev,
837 			   "Failed to deregister hw vPort: %d, 0x%x\n",
838 			   err, resp.hdr.status);
839 }
840 
mana_pf_register_filter(struct mana_port_context * apc)841 static int mana_pf_register_filter(struct mana_port_context *apc)
842 {
843 	struct mana_register_filter_resp resp = {};
844 	struct mana_register_filter_req req = {};
845 	int err;
846 
847 	mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER,
848 			     sizeof(req), sizeof(resp));
849 	req.vport = apc->port_handle;
850 	memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN);
851 
852 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
853 				sizeof(resp));
854 	if (err) {
855 		netdev_err(apc->ndev, "Failed to register filter: %d\n", err);
856 		return err;
857 	}
858 
859 	err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER,
860 				   sizeof(resp));
861 	if (err || resp.hdr.status) {
862 		netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n",
863 			   err, resp.hdr.status);
864 		return err ? err : -EPROTO;
865 	}
866 
867 	apc->pf_filter_handle = resp.filter_handle;
868 	return 0;
869 }
870 
mana_pf_deregister_filter(struct mana_port_context * apc)871 static void mana_pf_deregister_filter(struct mana_port_context *apc)
872 {
873 	struct mana_deregister_filter_resp resp = {};
874 	struct mana_deregister_filter_req req = {};
875 	int err;
876 
877 	mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER,
878 			     sizeof(req), sizeof(resp));
879 	req.filter_handle = apc->pf_filter_handle;
880 
881 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
882 				sizeof(resp));
883 	if (err) {
884 		netdev_err(apc->ndev, "Failed to unregister filter: %d\n",
885 			   err);
886 		return;
887 	}
888 
889 	err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER,
890 				   sizeof(resp));
891 	if (err || resp.hdr.status)
892 		netdev_err(apc->ndev,
893 			   "Failed to deregister filter: %d, 0x%x\n",
894 			   err, resp.hdr.status);
895 }
896 
mana_query_device_cfg(struct mana_context * ac,u32 proto_major_ver,u32 proto_minor_ver,u32 proto_micro_ver,u16 * max_num_vports)897 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver,
898 				 u32 proto_minor_ver, u32 proto_micro_ver,
899 				 u16 *max_num_vports)
900 {
901 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
902 	struct mana_query_device_cfg_resp resp = {};
903 	struct mana_query_device_cfg_req req = {};
904 	struct device *dev = gc->dev;
905 	int err = 0;
906 
907 	mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG,
908 			     sizeof(req), sizeof(resp));
909 
910 	req.hdr.resp.msg_version = GDMA_MESSAGE_V2;
911 
912 	req.proto_major_ver = proto_major_ver;
913 	req.proto_minor_ver = proto_minor_ver;
914 	req.proto_micro_ver = proto_micro_ver;
915 
916 	err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp));
917 	if (err) {
918 		dev_err(dev, "Failed to query config: %d", err);
919 		return err;
920 	}
921 
922 	err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG,
923 				   sizeof(resp));
924 	if (err || resp.hdr.status) {
925 		dev_err(dev, "Invalid query result: %d, 0x%x\n", err,
926 			resp.hdr.status);
927 		if (!err)
928 			err = -EPROTO;
929 		return err;
930 	}
931 
932 	*max_num_vports = resp.max_num_vports;
933 
934 	if (resp.hdr.response.msg_version == GDMA_MESSAGE_V2)
935 		gc->adapter_mtu = resp.adapter_mtu;
936 	else
937 		gc->adapter_mtu = ETH_FRAME_LEN;
938 
939 	return 0;
940 }
941 
mana_query_vport_cfg(struct mana_port_context * apc,u32 vport_index,u32 * max_sq,u32 * max_rq,u32 * num_indir_entry)942 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index,
943 				u32 *max_sq, u32 *max_rq, u32 *num_indir_entry)
944 {
945 	struct mana_query_vport_cfg_resp resp = {};
946 	struct mana_query_vport_cfg_req req = {};
947 	int err;
948 
949 	mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG,
950 			     sizeof(req), sizeof(resp));
951 
952 	req.vport_index = vport_index;
953 
954 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
955 				sizeof(resp));
956 	if (err)
957 		return err;
958 
959 	err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG,
960 				   sizeof(resp));
961 	if (err)
962 		return err;
963 
964 	if (resp.hdr.status)
965 		return -EPROTO;
966 
967 	*max_sq = resp.max_num_sq;
968 	*max_rq = resp.max_num_rq;
969 	*num_indir_entry = resp.num_indirection_ent;
970 
971 	apc->port_handle = resp.vport;
972 	ether_addr_copy(apc->mac_addr, resp.mac_addr);
973 
974 	return 0;
975 }
976 
mana_uncfg_vport(struct mana_port_context * apc)977 void mana_uncfg_vport(struct mana_port_context *apc)
978 {
979 	mutex_lock(&apc->vport_mutex);
980 	apc->vport_use_count--;
981 	WARN_ON(apc->vport_use_count < 0);
982 	mutex_unlock(&apc->vport_mutex);
983 }
984 EXPORT_SYMBOL_NS(mana_uncfg_vport, NET_MANA);
985 
mana_cfg_vport(struct mana_port_context * apc,u32 protection_dom_id,u32 doorbell_pg_id)986 int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id,
987 		   u32 doorbell_pg_id)
988 {
989 	struct mana_config_vport_resp resp = {};
990 	struct mana_config_vport_req req = {};
991 	int err;
992 
993 	/* This function is used to program the Ethernet port in the hardware
994 	 * table. It can be called from the Ethernet driver or the RDMA driver.
995 	 *
996 	 * For Ethernet usage, the hardware supports only one active user on a
997 	 * physical port. The driver checks on the port usage before programming
998 	 * the hardware when creating the RAW QP (RDMA driver) or exposing the
999 	 * device to kernel NET layer (Ethernet driver).
1000 	 *
1001 	 * Because the RDMA driver doesn't know in advance which QP type the
1002 	 * user will create, it exposes the device with all its ports. The user
1003 	 * may not be able to create RAW QP on a port if this port is already
1004 	 * in used by the Ethernet driver from the kernel.
1005 	 *
1006 	 * This physical port limitation only applies to the RAW QP. For RC QP,
1007 	 * the hardware doesn't have this limitation. The user can create RC
1008 	 * QPs on a physical port up to the hardware limits independent of the
1009 	 * Ethernet usage on the same port.
1010 	 */
1011 	mutex_lock(&apc->vport_mutex);
1012 	if (apc->vport_use_count > 0) {
1013 		mutex_unlock(&apc->vport_mutex);
1014 		return -EBUSY;
1015 	}
1016 	apc->vport_use_count++;
1017 	mutex_unlock(&apc->vport_mutex);
1018 
1019 	mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX,
1020 			     sizeof(req), sizeof(resp));
1021 	req.vport = apc->port_handle;
1022 	req.pdid = protection_dom_id;
1023 	req.doorbell_pageid = doorbell_pg_id;
1024 
1025 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1026 				sizeof(resp));
1027 	if (err) {
1028 		netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err);
1029 		goto out;
1030 	}
1031 
1032 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX,
1033 				   sizeof(resp));
1034 	if (err || resp.hdr.status) {
1035 		netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n",
1036 			   err, resp.hdr.status);
1037 		if (!err)
1038 			err = -EPROTO;
1039 
1040 		goto out;
1041 	}
1042 
1043 	apc->tx_shortform_allowed = resp.short_form_allowed;
1044 	apc->tx_vp_offset = resp.tx_vport_offset;
1045 
1046 	netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n",
1047 		    apc->port_handle, protection_dom_id, doorbell_pg_id);
1048 out:
1049 	if (err)
1050 		mana_uncfg_vport(apc);
1051 
1052 	return err;
1053 }
1054 EXPORT_SYMBOL_NS(mana_cfg_vport, NET_MANA);
1055 
mana_cfg_vport_steering(struct mana_port_context * apc,enum TRI_STATE rx,bool update_default_rxobj,bool update_key,bool update_tab)1056 static int mana_cfg_vport_steering(struct mana_port_context *apc,
1057 				   enum TRI_STATE rx,
1058 				   bool update_default_rxobj, bool update_key,
1059 				   bool update_tab)
1060 {
1061 	u16 num_entries = MANA_INDIRECT_TABLE_SIZE;
1062 	struct mana_cfg_rx_steer_req_v2 *req;
1063 	struct mana_cfg_rx_steer_resp resp = {};
1064 	struct net_device *ndev = apc->ndev;
1065 	mana_handle_t *req_indir_tab;
1066 	u32 req_buf_size;
1067 	int err;
1068 
1069 	req_buf_size = sizeof(*req) + sizeof(mana_handle_t) * num_entries;
1070 	req = kzalloc(req_buf_size, GFP_KERNEL);
1071 	if (!req)
1072 		return -ENOMEM;
1073 
1074 	mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
1075 			     sizeof(resp));
1076 
1077 	req->hdr.req.msg_version = GDMA_MESSAGE_V2;
1078 
1079 	req->vport = apc->port_handle;
1080 	req->num_indir_entries = num_entries;
1081 	req->indir_tab_offset = sizeof(*req);
1082 	req->rx_enable = rx;
1083 	req->rss_enable = apc->rss_state;
1084 	req->update_default_rxobj = update_default_rxobj;
1085 	req->update_hashkey = update_key;
1086 	req->update_indir_tab = update_tab;
1087 	req->default_rxobj = apc->default_rxobj;
1088 	req->cqe_coalescing_enable = 0;
1089 
1090 	if (update_key)
1091 		memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE);
1092 
1093 	if (update_tab) {
1094 		req_indir_tab = (mana_handle_t *)(req + 1);
1095 		memcpy(req_indir_tab, apc->rxobj_table,
1096 		       req->num_indir_entries * sizeof(mana_handle_t));
1097 	}
1098 
1099 	err = mana_send_request(apc->ac, req, req_buf_size, &resp,
1100 				sizeof(resp));
1101 	if (err) {
1102 		netdev_err(ndev, "Failed to configure vPort RX: %d\n", err);
1103 		goto out;
1104 	}
1105 
1106 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX,
1107 				   sizeof(resp));
1108 	if (err) {
1109 		netdev_err(ndev, "vPort RX configuration failed: %d\n", err);
1110 		goto out;
1111 	}
1112 
1113 	if (resp.hdr.status) {
1114 		netdev_err(ndev, "vPort RX configuration failed: 0x%x\n",
1115 			   resp.hdr.status);
1116 		err = -EPROTO;
1117 	}
1118 
1119 	netdev_info(ndev, "Configured steering vPort %llu entries %u\n",
1120 		    apc->port_handle, num_entries);
1121 out:
1122 	kfree(req);
1123 	return err;
1124 }
1125 
mana_create_wq_obj(struct mana_port_context * apc,mana_handle_t vport,u32 wq_type,struct mana_obj_spec * wq_spec,struct mana_obj_spec * cq_spec,mana_handle_t * wq_obj)1126 int mana_create_wq_obj(struct mana_port_context *apc,
1127 		       mana_handle_t vport,
1128 		       u32 wq_type, struct mana_obj_spec *wq_spec,
1129 		       struct mana_obj_spec *cq_spec,
1130 		       mana_handle_t *wq_obj)
1131 {
1132 	struct mana_create_wqobj_resp resp = {};
1133 	struct mana_create_wqobj_req req = {};
1134 	struct net_device *ndev = apc->ndev;
1135 	int err;
1136 
1137 	mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ,
1138 			     sizeof(req), sizeof(resp));
1139 	req.vport = vport;
1140 	req.wq_type = wq_type;
1141 	req.wq_gdma_region = wq_spec->gdma_region;
1142 	req.cq_gdma_region = cq_spec->gdma_region;
1143 	req.wq_size = wq_spec->queue_size;
1144 	req.cq_size = cq_spec->queue_size;
1145 	req.cq_moderation_ctx_id = cq_spec->modr_ctx_id;
1146 	req.cq_parent_qid = cq_spec->attached_eq;
1147 
1148 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1149 				sizeof(resp));
1150 	if (err) {
1151 		netdev_err(ndev, "Failed to create WQ object: %d\n", err);
1152 		goto out;
1153 	}
1154 
1155 	err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ,
1156 				   sizeof(resp));
1157 	if (err || resp.hdr.status) {
1158 		netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err,
1159 			   resp.hdr.status);
1160 		if (!err)
1161 			err = -EPROTO;
1162 		goto out;
1163 	}
1164 
1165 	if (resp.wq_obj == INVALID_MANA_HANDLE) {
1166 		netdev_err(ndev, "Got an invalid WQ object handle\n");
1167 		err = -EPROTO;
1168 		goto out;
1169 	}
1170 
1171 	*wq_obj = resp.wq_obj;
1172 	wq_spec->queue_index = resp.wq_id;
1173 	cq_spec->queue_index = resp.cq_id;
1174 
1175 	return 0;
1176 out:
1177 	return err;
1178 }
1179 EXPORT_SYMBOL_NS(mana_create_wq_obj, NET_MANA);
1180 
mana_destroy_wq_obj(struct mana_port_context * apc,u32 wq_type,mana_handle_t wq_obj)1181 void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type,
1182 			 mana_handle_t wq_obj)
1183 {
1184 	struct mana_destroy_wqobj_resp resp = {};
1185 	struct mana_destroy_wqobj_req req = {};
1186 	struct net_device *ndev = apc->ndev;
1187 	int err;
1188 
1189 	mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ,
1190 			     sizeof(req), sizeof(resp));
1191 	req.wq_type = wq_type;
1192 	req.wq_obj_handle = wq_obj;
1193 
1194 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1195 				sizeof(resp));
1196 	if (err) {
1197 		netdev_err(ndev, "Failed to destroy WQ object: %d\n", err);
1198 		return;
1199 	}
1200 
1201 	err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ,
1202 				   sizeof(resp));
1203 	if (err || resp.hdr.status)
1204 		netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err,
1205 			   resp.hdr.status);
1206 }
1207 EXPORT_SYMBOL_NS(mana_destroy_wq_obj, NET_MANA);
1208 
mana_destroy_eq(struct mana_context * ac)1209 static void mana_destroy_eq(struct mana_context *ac)
1210 {
1211 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
1212 	struct gdma_queue *eq;
1213 	int i;
1214 
1215 	if (!ac->eqs)
1216 		return;
1217 
1218 	for (i = 0; i < gc->max_num_queues; i++) {
1219 		eq = ac->eqs[i].eq;
1220 		if (!eq)
1221 			continue;
1222 
1223 		mana_gd_destroy_queue(gc, eq);
1224 	}
1225 
1226 	kfree(ac->eqs);
1227 	ac->eqs = NULL;
1228 }
1229 
mana_create_eq(struct mana_context * ac)1230 static int mana_create_eq(struct mana_context *ac)
1231 {
1232 	struct gdma_dev *gd = ac->gdma_dev;
1233 	struct gdma_context *gc = gd->gdma_context;
1234 	struct gdma_queue_spec spec = {};
1235 	int err;
1236 	int i;
1237 
1238 	ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq),
1239 			  GFP_KERNEL);
1240 	if (!ac->eqs)
1241 		return -ENOMEM;
1242 
1243 	spec.type = GDMA_EQ;
1244 	spec.monitor_avl_buf = false;
1245 	spec.queue_size = EQ_SIZE;
1246 	spec.eq.callback = NULL;
1247 	spec.eq.context = ac->eqs;
1248 	spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE;
1249 
1250 	for (i = 0; i < gc->max_num_queues; i++) {
1251 		err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq);
1252 		if (err)
1253 			goto out;
1254 	}
1255 
1256 	return 0;
1257 out:
1258 	mana_destroy_eq(ac);
1259 	return err;
1260 }
1261 
mana_fence_rq(struct mana_port_context * apc,struct mana_rxq * rxq)1262 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq)
1263 {
1264 	struct mana_fence_rq_resp resp = {};
1265 	struct mana_fence_rq_req req = {};
1266 	int err;
1267 
1268 	init_completion(&rxq->fence_event);
1269 
1270 	mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ,
1271 			     sizeof(req), sizeof(resp));
1272 	req.wq_obj_handle =  rxq->rxobj;
1273 
1274 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1275 				sizeof(resp));
1276 	if (err) {
1277 		netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n",
1278 			   rxq->rxq_idx, err);
1279 		return err;
1280 	}
1281 
1282 	err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp));
1283 	if (err || resp.hdr.status) {
1284 		netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n",
1285 			   rxq->rxq_idx, err, resp.hdr.status);
1286 		if (!err)
1287 			err = -EPROTO;
1288 
1289 		return err;
1290 	}
1291 
1292 	if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) {
1293 		netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n",
1294 			   rxq->rxq_idx);
1295 		return -ETIMEDOUT;
1296 	}
1297 
1298 	return 0;
1299 }
1300 
mana_fence_rqs(struct mana_port_context * apc)1301 static void mana_fence_rqs(struct mana_port_context *apc)
1302 {
1303 	unsigned int rxq_idx;
1304 	struct mana_rxq *rxq;
1305 	int err;
1306 
1307 	for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
1308 		rxq = apc->rxqs[rxq_idx];
1309 		err = mana_fence_rq(apc, rxq);
1310 
1311 		/* In case of any error, use sleep instead. */
1312 		if (err)
1313 			msleep(100);
1314 	}
1315 }
1316 
mana_move_wq_tail(struct gdma_queue * wq,u32 num_units)1317 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units)
1318 {
1319 	u32 used_space_old;
1320 	u32 used_space_new;
1321 
1322 	used_space_old = wq->head - wq->tail;
1323 	used_space_new = wq->head - (wq->tail + num_units);
1324 
1325 	if (WARN_ON_ONCE(used_space_new > used_space_old))
1326 		return -ERANGE;
1327 
1328 	wq->tail += num_units;
1329 	return 0;
1330 }
1331 
mana_unmap_skb(struct sk_buff * skb,struct mana_port_context * apc)1332 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc)
1333 {
1334 	struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
1335 	struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1336 	struct device *dev = gc->dev;
1337 	int hsg, i;
1338 
1339 	/* Number of SGEs of linear part */
1340 	hsg = (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0]) ? 2 : 1;
1341 
1342 	for (i = 0; i < hsg; i++)
1343 		dma_unmap_single(dev, ash->dma_handle[i], ash->size[i],
1344 				 DMA_TO_DEVICE);
1345 
1346 	for (i = hsg; i < skb_shinfo(skb)->nr_frags + hsg; i++)
1347 		dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
1348 			       DMA_TO_DEVICE);
1349 }
1350 
mana_poll_tx_cq(struct mana_cq * cq)1351 static void mana_poll_tx_cq(struct mana_cq *cq)
1352 {
1353 	struct gdma_comp *completions = cq->gdma_comp_buf;
1354 	struct gdma_posted_wqe_info *wqe_info;
1355 	unsigned int pkt_transmitted = 0;
1356 	unsigned int wqe_unit_cnt = 0;
1357 	struct mana_txq *txq = cq->txq;
1358 	struct mana_port_context *apc;
1359 	struct netdev_queue *net_txq;
1360 	struct gdma_queue *gdma_wq;
1361 	unsigned int avail_space;
1362 	struct net_device *ndev;
1363 	struct sk_buff *skb;
1364 	bool txq_stopped;
1365 	int comp_read;
1366 	int i;
1367 
1368 	ndev = txq->ndev;
1369 	apc = netdev_priv(ndev);
1370 
1371 	comp_read = mana_gd_poll_cq(cq->gdma_cq, completions,
1372 				    CQE_POLLING_BUFFER);
1373 
1374 	if (comp_read < 1)
1375 		return;
1376 
1377 	for (i = 0; i < comp_read; i++) {
1378 		struct mana_tx_comp_oob *cqe_oob;
1379 
1380 		if (WARN_ON_ONCE(!completions[i].is_sq))
1381 			return;
1382 
1383 		cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data;
1384 		if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type !=
1385 				 MANA_CQE_COMPLETION))
1386 			return;
1387 
1388 		switch (cqe_oob->cqe_hdr.cqe_type) {
1389 		case CQE_TX_OKAY:
1390 			break;
1391 
1392 		case CQE_TX_SA_DROP:
1393 		case CQE_TX_MTU_DROP:
1394 		case CQE_TX_INVALID_OOB:
1395 		case CQE_TX_INVALID_ETH_TYPE:
1396 		case CQE_TX_HDR_PROCESSING_ERROR:
1397 		case CQE_TX_VF_DISABLED:
1398 		case CQE_TX_VPORT_IDX_OUT_OF_RANGE:
1399 		case CQE_TX_VPORT_DISABLED:
1400 		case CQE_TX_VLAN_TAGGING_VIOLATION:
1401 			if (net_ratelimit())
1402 				netdev_err(ndev, "TX: CQE error %d\n",
1403 					   cqe_oob->cqe_hdr.cqe_type);
1404 
1405 			apc->eth_stats.tx_cqe_err++;
1406 			break;
1407 
1408 		default:
1409 			/* If the CQE type is unknown, log an error,
1410 			 * and still free the SKB, update tail, etc.
1411 			 */
1412 			if (net_ratelimit())
1413 				netdev_err(ndev, "TX: unknown CQE type %d\n",
1414 					   cqe_oob->cqe_hdr.cqe_type);
1415 
1416 			apc->eth_stats.tx_cqe_unknown_type++;
1417 			break;
1418 		}
1419 
1420 		if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num))
1421 			return;
1422 
1423 		skb = skb_dequeue(&txq->pending_skbs);
1424 		if (WARN_ON_ONCE(!skb))
1425 			return;
1426 
1427 		wqe_info = (struct gdma_posted_wqe_info *)skb->cb;
1428 		wqe_unit_cnt += wqe_info->wqe_size_in_bu;
1429 
1430 		mana_unmap_skb(skb, apc);
1431 
1432 		napi_consume_skb(skb, cq->budget);
1433 
1434 		pkt_transmitted++;
1435 	}
1436 
1437 	if (WARN_ON_ONCE(wqe_unit_cnt == 0))
1438 		return;
1439 
1440 	mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt);
1441 
1442 	gdma_wq = txq->gdma_sq;
1443 	avail_space = mana_gd_wq_avail_space(gdma_wq);
1444 
1445 	/* Ensure tail updated before checking q stop */
1446 	smp_mb();
1447 
1448 	net_txq = txq->net_txq;
1449 	txq_stopped = netif_tx_queue_stopped(net_txq);
1450 
1451 	/* Ensure checking txq_stopped before apc->port_is_up. */
1452 	smp_rmb();
1453 
1454 	if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) {
1455 		netif_tx_wake_queue(net_txq);
1456 		apc->eth_stats.wake_queue++;
1457 	}
1458 
1459 	if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0)
1460 		WARN_ON_ONCE(1);
1461 
1462 	cq->work_done = pkt_transmitted;
1463 }
1464 
mana_post_pkt_rxq(struct mana_rxq * rxq)1465 static void mana_post_pkt_rxq(struct mana_rxq *rxq)
1466 {
1467 	struct mana_recv_buf_oob *recv_buf_oob;
1468 	u32 curr_index;
1469 	int err;
1470 
1471 	curr_index = rxq->buf_index++;
1472 	if (rxq->buf_index == rxq->num_rx_buf)
1473 		rxq->buf_index = 0;
1474 
1475 	recv_buf_oob = &rxq->rx_oobs[curr_index];
1476 
1477 	err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req,
1478 					&recv_buf_oob->wqe_inf);
1479 	if (WARN_ON_ONCE(err))
1480 		return;
1481 
1482 	WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1);
1483 }
1484 
mana_build_skb(struct mana_rxq * rxq,void * buf_va,uint pkt_len,struct xdp_buff * xdp)1485 static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va,
1486 				      uint pkt_len, struct xdp_buff *xdp)
1487 {
1488 	struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size);
1489 
1490 	if (!skb)
1491 		return NULL;
1492 
1493 	if (xdp->data_hard_start) {
1494 		skb_reserve(skb, xdp->data - xdp->data_hard_start);
1495 		skb_put(skb, xdp->data_end - xdp->data);
1496 		return skb;
1497 	}
1498 
1499 	skb_reserve(skb, rxq->headroom);
1500 	skb_put(skb, pkt_len);
1501 
1502 	return skb;
1503 }
1504 
mana_rx_skb(void * buf_va,bool from_pool,struct mana_rxcomp_oob * cqe,struct mana_rxq * rxq)1505 static void mana_rx_skb(void *buf_va, bool from_pool,
1506 			struct mana_rxcomp_oob *cqe, struct mana_rxq *rxq)
1507 {
1508 	struct mana_stats_rx *rx_stats = &rxq->stats;
1509 	struct net_device *ndev = rxq->ndev;
1510 	uint pkt_len = cqe->ppi[0].pkt_len;
1511 	u16 rxq_idx = rxq->rxq_idx;
1512 	struct napi_struct *napi;
1513 	struct xdp_buff xdp = {};
1514 	struct sk_buff *skb;
1515 	u32 hash_value;
1516 	u32 act;
1517 
1518 	rxq->rx_cq.work_done++;
1519 	napi = &rxq->rx_cq.napi;
1520 
1521 	if (!buf_va) {
1522 		++ndev->stats.rx_dropped;
1523 		return;
1524 	}
1525 
1526 	act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len);
1527 
1528 	if (act == XDP_REDIRECT && !rxq->xdp_rc)
1529 		return;
1530 
1531 	if (act != XDP_PASS && act != XDP_TX)
1532 		goto drop_xdp;
1533 
1534 	skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp);
1535 
1536 	if (!skb)
1537 		goto drop;
1538 
1539 	if (from_pool)
1540 		skb_mark_for_recycle(skb);
1541 
1542 	skb->dev = napi->dev;
1543 
1544 	skb->protocol = eth_type_trans(skb, ndev);
1545 	skb_checksum_none_assert(skb);
1546 	skb_record_rx_queue(skb, rxq_idx);
1547 
1548 	if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) {
1549 		if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed)
1550 			skb->ip_summed = CHECKSUM_UNNECESSARY;
1551 	}
1552 
1553 	if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) {
1554 		hash_value = cqe->ppi[0].pkt_hash;
1555 
1556 		if (cqe->rx_hashtype & MANA_HASH_L4)
1557 			skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4);
1558 		else
1559 			skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3);
1560 	}
1561 
1562 	if (cqe->rx_vlantag_present) {
1563 		u16 vlan_tci = cqe->rx_vlan_id;
1564 
1565 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci);
1566 	}
1567 
1568 	u64_stats_update_begin(&rx_stats->syncp);
1569 	rx_stats->packets++;
1570 	rx_stats->bytes += pkt_len;
1571 
1572 	if (act == XDP_TX)
1573 		rx_stats->xdp_tx++;
1574 	u64_stats_update_end(&rx_stats->syncp);
1575 
1576 	if (act == XDP_TX) {
1577 		skb_set_queue_mapping(skb, rxq_idx);
1578 		mana_xdp_tx(skb, ndev);
1579 		return;
1580 	}
1581 
1582 	napi_gro_receive(napi, skb);
1583 
1584 	return;
1585 
1586 drop_xdp:
1587 	u64_stats_update_begin(&rx_stats->syncp);
1588 	rx_stats->xdp_drop++;
1589 	u64_stats_update_end(&rx_stats->syncp);
1590 
1591 drop:
1592 	if (from_pool) {
1593 		page_pool_recycle_direct(rxq->page_pool,
1594 					 virt_to_head_page(buf_va));
1595 	} else {
1596 		WARN_ON_ONCE(rxq->xdp_save_va);
1597 		/* Save for reuse */
1598 		rxq->xdp_save_va = buf_va;
1599 	}
1600 
1601 	++ndev->stats.rx_dropped;
1602 
1603 	return;
1604 }
1605 
mana_get_rxfrag(struct mana_rxq * rxq,struct device * dev,dma_addr_t * da,bool * from_pool,bool is_napi)1606 static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,
1607 			     dma_addr_t *da, bool *from_pool, bool is_napi)
1608 {
1609 	struct page *page;
1610 	void *va;
1611 
1612 	*from_pool = false;
1613 
1614 	/* Reuse XDP dropped page if available */
1615 	if (rxq->xdp_save_va) {
1616 		va = rxq->xdp_save_va;
1617 		rxq->xdp_save_va = NULL;
1618 	} else if (rxq->alloc_size > PAGE_SIZE) {
1619 		if (is_napi)
1620 			va = napi_alloc_frag(rxq->alloc_size);
1621 		else
1622 			va = netdev_alloc_frag(rxq->alloc_size);
1623 
1624 		if (!va)
1625 			return NULL;
1626 
1627 		page = virt_to_head_page(va);
1628 		/* Check if the frag falls back to single page */
1629 		if (compound_order(page) < get_order(rxq->alloc_size)) {
1630 			put_page(page);
1631 			return NULL;
1632 		}
1633 	} else {
1634 		page = page_pool_dev_alloc_pages(rxq->page_pool);
1635 		if (!page)
1636 			return NULL;
1637 
1638 		*from_pool = true;
1639 		va = page_to_virt(page);
1640 	}
1641 
1642 	*da = dma_map_single(dev, va + rxq->headroom, rxq->datasize,
1643 			     DMA_FROM_DEVICE);
1644 	if (dma_mapping_error(dev, *da)) {
1645 		if (*from_pool)
1646 			page_pool_put_full_page(rxq->page_pool, page, false);
1647 		else
1648 			put_page(virt_to_head_page(va));
1649 
1650 		return NULL;
1651 	}
1652 
1653 	return va;
1654 }
1655 
1656 /* Allocate frag for rx buffer, and save the old buf */
mana_refill_rx_oob(struct device * dev,struct mana_rxq * rxq,struct mana_recv_buf_oob * rxoob,void ** old_buf,bool * old_fp)1657 static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq,
1658 			       struct mana_recv_buf_oob *rxoob, void **old_buf,
1659 			       bool *old_fp)
1660 {
1661 	bool from_pool;
1662 	dma_addr_t da;
1663 	void *va;
1664 
1665 	va = mana_get_rxfrag(rxq, dev, &da, &from_pool, true);
1666 	if (!va)
1667 		return;
1668 
1669 	dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize,
1670 			 DMA_FROM_DEVICE);
1671 	*old_buf = rxoob->buf_va;
1672 	*old_fp = rxoob->from_pool;
1673 
1674 	rxoob->buf_va = va;
1675 	rxoob->sgl[0].address = da;
1676 	rxoob->from_pool = from_pool;
1677 }
1678 
mana_process_rx_cqe(struct mana_rxq * rxq,struct mana_cq * cq,struct gdma_comp * cqe)1679 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,
1680 				struct gdma_comp *cqe)
1681 {
1682 	struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data;
1683 	struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1684 	struct net_device *ndev = rxq->ndev;
1685 	struct mana_recv_buf_oob *rxbuf_oob;
1686 	struct mana_port_context *apc;
1687 	struct device *dev = gc->dev;
1688 	void *old_buf = NULL;
1689 	u32 curr, pktlen;
1690 	bool old_fp;
1691 
1692 	apc = netdev_priv(ndev);
1693 
1694 	switch (oob->cqe_hdr.cqe_type) {
1695 	case CQE_RX_OKAY:
1696 		break;
1697 
1698 	case CQE_RX_TRUNCATED:
1699 		++ndev->stats.rx_dropped;
1700 		rxbuf_oob = &rxq->rx_oobs[rxq->buf_index];
1701 		netdev_warn_once(ndev, "Dropped a truncated packet\n");
1702 		goto drop;
1703 
1704 	case CQE_RX_COALESCED_4:
1705 		netdev_err(ndev, "RX coalescing is unsupported\n");
1706 		apc->eth_stats.rx_coalesced_err++;
1707 		return;
1708 
1709 	case CQE_RX_OBJECT_FENCE:
1710 		complete(&rxq->fence_event);
1711 		return;
1712 
1713 	default:
1714 		netdev_err(ndev, "Unknown RX CQE type = %d\n",
1715 			   oob->cqe_hdr.cqe_type);
1716 		apc->eth_stats.rx_cqe_unknown_type++;
1717 		return;
1718 	}
1719 
1720 	pktlen = oob->ppi[0].pkt_len;
1721 
1722 	if (pktlen == 0) {
1723 		/* data packets should never have packetlength of zero */
1724 		netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n",
1725 			   rxq->gdma_id, cq->gdma_id, rxq->rxobj);
1726 		return;
1727 	}
1728 
1729 	curr = rxq->buf_index;
1730 	rxbuf_oob = &rxq->rx_oobs[curr];
1731 	WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1);
1732 
1733 	mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf, &old_fp);
1734 
1735 	/* Unsuccessful refill will have old_buf == NULL.
1736 	 * In this case, mana_rx_skb() will drop the packet.
1737 	 */
1738 	mana_rx_skb(old_buf, old_fp, oob, rxq);
1739 
1740 drop:
1741 	mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu);
1742 
1743 	mana_post_pkt_rxq(rxq);
1744 }
1745 
mana_poll_rx_cq(struct mana_cq * cq)1746 static void mana_poll_rx_cq(struct mana_cq *cq)
1747 {
1748 	struct gdma_comp *comp = cq->gdma_comp_buf;
1749 	struct mana_rxq *rxq = cq->rxq;
1750 	int comp_read, i;
1751 
1752 	comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER);
1753 	WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER);
1754 
1755 	rxq->xdp_flush = false;
1756 
1757 	for (i = 0; i < comp_read; i++) {
1758 		if (WARN_ON_ONCE(comp[i].is_sq))
1759 			return;
1760 
1761 		/* verify recv cqe references the right rxq */
1762 		if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id))
1763 			return;
1764 
1765 		mana_process_rx_cqe(rxq, cq, &comp[i]);
1766 	}
1767 
1768 	if (comp_read > 0) {
1769 		struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1770 
1771 		mana_gd_wq_ring_doorbell(gc, rxq->gdma_rq);
1772 	}
1773 
1774 	if (rxq->xdp_flush)
1775 		xdp_do_flush();
1776 }
1777 
mana_cq_handler(void * context,struct gdma_queue * gdma_queue)1778 static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue)
1779 {
1780 	struct mana_cq *cq = context;
1781 	int w;
1782 
1783 	WARN_ON_ONCE(cq->gdma_cq != gdma_queue);
1784 
1785 	if (cq->type == MANA_CQ_TYPE_RX)
1786 		mana_poll_rx_cq(cq);
1787 	else
1788 		mana_poll_tx_cq(cq);
1789 
1790 	w = cq->work_done;
1791 	cq->work_done_since_doorbell += w;
1792 
1793 	if (w < cq->budget) {
1794 		mana_gd_ring_cq(gdma_queue, SET_ARM_BIT);
1795 		cq->work_done_since_doorbell = 0;
1796 		napi_complete_done(&cq->napi, w);
1797 	} else if (cq->work_done_since_doorbell >
1798 		   cq->gdma_cq->queue_size / COMP_ENTRY_SIZE * 4) {
1799 		/* MANA hardware requires at least one doorbell ring every 8
1800 		 * wraparounds of CQ even if there is no need to arm the CQ.
1801 		 * This driver rings the doorbell as soon as we have exceeded
1802 		 * 4 wraparounds.
1803 		 */
1804 		mana_gd_ring_cq(gdma_queue, 0);
1805 		cq->work_done_since_doorbell = 0;
1806 	}
1807 
1808 	return w;
1809 }
1810 
mana_poll(struct napi_struct * napi,int budget)1811 static int mana_poll(struct napi_struct *napi, int budget)
1812 {
1813 	struct mana_cq *cq = container_of(napi, struct mana_cq, napi);
1814 	int w;
1815 
1816 	cq->work_done = 0;
1817 	cq->budget = budget;
1818 
1819 	w = mana_cq_handler(cq, cq->gdma_cq);
1820 
1821 	return min(w, budget);
1822 }
1823 
mana_schedule_napi(void * context,struct gdma_queue * gdma_queue)1824 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue)
1825 {
1826 	struct mana_cq *cq = context;
1827 
1828 	napi_schedule_irqoff(&cq->napi);
1829 }
1830 
mana_deinit_cq(struct mana_port_context * apc,struct mana_cq * cq)1831 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq)
1832 {
1833 	struct gdma_dev *gd = apc->ac->gdma_dev;
1834 
1835 	if (!cq->gdma_cq)
1836 		return;
1837 
1838 	mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq);
1839 }
1840 
mana_deinit_txq(struct mana_port_context * apc,struct mana_txq * txq)1841 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq)
1842 {
1843 	struct gdma_dev *gd = apc->ac->gdma_dev;
1844 
1845 	if (!txq->gdma_sq)
1846 		return;
1847 
1848 	mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq);
1849 }
1850 
mana_destroy_txq(struct mana_port_context * apc)1851 static void mana_destroy_txq(struct mana_port_context *apc)
1852 {
1853 	struct napi_struct *napi;
1854 	int i;
1855 
1856 	if (!apc->tx_qp)
1857 		return;
1858 
1859 	for (i = 0; i < apc->num_queues; i++) {
1860 		napi = &apc->tx_qp[i].tx_cq.napi;
1861 		if (apc->tx_qp[i].txq.napi_initialized) {
1862 			napi_synchronize(napi);
1863 			napi_disable(napi);
1864 			netif_napi_del(napi);
1865 			apc->tx_qp[i].txq.napi_initialized = false;
1866 		}
1867 		mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object);
1868 
1869 		mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq);
1870 
1871 		mana_deinit_txq(apc, &apc->tx_qp[i].txq);
1872 	}
1873 
1874 	kfree(apc->tx_qp);
1875 	apc->tx_qp = NULL;
1876 }
1877 
mana_create_txq(struct mana_port_context * apc,struct net_device * net)1878 static int mana_create_txq(struct mana_port_context *apc,
1879 			   struct net_device *net)
1880 {
1881 	struct mana_context *ac = apc->ac;
1882 	struct gdma_dev *gd = ac->gdma_dev;
1883 	struct mana_obj_spec wq_spec;
1884 	struct mana_obj_spec cq_spec;
1885 	struct gdma_queue_spec spec;
1886 	struct gdma_context *gc;
1887 	struct mana_txq *txq;
1888 	struct mana_cq *cq;
1889 	u32 txq_size;
1890 	u32 cq_size;
1891 	int err;
1892 	int i;
1893 
1894 	apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp),
1895 			     GFP_KERNEL);
1896 	if (!apc->tx_qp)
1897 		return -ENOMEM;
1898 
1899 	/*  The minimum size of the WQE is 32 bytes, hence
1900 	 *  MAX_SEND_BUFFERS_PER_QUEUE represents the maximum number of WQEs
1901 	 *  the SQ can store. This value is then used to size other queues
1902 	 *  to prevent overflow.
1903 	 */
1904 	txq_size = MAX_SEND_BUFFERS_PER_QUEUE * 32;
1905 	BUILD_BUG_ON(!MANA_PAGE_ALIGNED(txq_size));
1906 
1907 	cq_size = MAX_SEND_BUFFERS_PER_QUEUE * COMP_ENTRY_SIZE;
1908 	cq_size = MANA_PAGE_ALIGN(cq_size);
1909 
1910 	gc = gd->gdma_context;
1911 
1912 	for (i = 0; i < apc->num_queues; i++) {
1913 		apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE;
1914 
1915 		/* Create SQ */
1916 		txq = &apc->tx_qp[i].txq;
1917 
1918 		u64_stats_init(&txq->stats.syncp);
1919 		txq->ndev = net;
1920 		txq->net_txq = netdev_get_tx_queue(net, i);
1921 		txq->vp_offset = apc->tx_vp_offset;
1922 		txq->napi_initialized = false;
1923 		skb_queue_head_init(&txq->pending_skbs);
1924 
1925 		memset(&spec, 0, sizeof(spec));
1926 		spec.type = GDMA_SQ;
1927 		spec.monitor_avl_buf = true;
1928 		spec.queue_size = txq_size;
1929 		err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq);
1930 		if (err)
1931 			goto out;
1932 
1933 		/* Create SQ's CQ */
1934 		cq = &apc->tx_qp[i].tx_cq;
1935 		cq->type = MANA_CQ_TYPE_TX;
1936 
1937 		cq->txq = txq;
1938 
1939 		memset(&spec, 0, sizeof(spec));
1940 		spec.type = GDMA_CQ;
1941 		spec.monitor_avl_buf = false;
1942 		spec.queue_size = cq_size;
1943 		spec.cq.callback = mana_schedule_napi;
1944 		spec.cq.parent_eq = ac->eqs[i].eq;
1945 		spec.cq.context = cq;
1946 		err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
1947 		if (err)
1948 			goto out;
1949 
1950 		memset(&wq_spec, 0, sizeof(wq_spec));
1951 		memset(&cq_spec, 0, sizeof(cq_spec));
1952 
1953 		wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle;
1954 		wq_spec.queue_size = txq->gdma_sq->queue_size;
1955 
1956 		cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
1957 		cq_spec.queue_size = cq->gdma_cq->queue_size;
1958 		cq_spec.modr_ctx_id = 0;
1959 		cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
1960 
1961 		err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ,
1962 					 &wq_spec, &cq_spec,
1963 					 &apc->tx_qp[i].tx_object);
1964 
1965 		if (err)
1966 			goto out;
1967 
1968 		txq->gdma_sq->id = wq_spec.queue_index;
1969 		cq->gdma_cq->id = cq_spec.queue_index;
1970 
1971 		txq->gdma_sq->mem_info.dma_region_handle =
1972 			GDMA_INVALID_DMA_REGION;
1973 		cq->gdma_cq->mem_info.dma_region_handle =
1974 			GDMA_INVALID_DMA_REGION;
1975 
1976 		txq->gdma_txq_id = txq->gdma_sq->id;
1977 
1978 		cq->gdma_id = cq->gdma_cq->id;
1979 
1980 		if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
1981 			err = -EINVAL;
1982 			goto out;
1983 		}
1984 
1985 		gc->cq_table[cq->gdma_id] = cq->gdma_cq;
1986 
1987 		netif_napi_add_tx(net, &cq->napi, mana_poll);
1988 		napi_enable(&cq->napi);
1989 		txq->napi_initialized = true;
1990 
1991 		mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
1992 	}
1993 
1994 	return 0;
1995 out:
1996 	mana_destroy_txq(apc);
1997 	return err;
1998 }
1999 
mana_destroy_rxq(struct mana_port_context * apc,struct mana_rxq * rxq,bool napi_initialized)2000 static void mana_destroy_rxq(struct mana_port_context *apc,
2001 			     struct mana_rxq *rxq, bool napi_initialized)
2002 
2003 {
2004 	struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
2005 	struct mana_recv_buf_oob *rx_oob;
2006 	struct device *dev = gc->dev;
2007 	struct napi_struct *napi;
2008 	struct page *page;
2009 	int i;
2010 
2011 	if (!rxq)
2012 		return;
2013 
2014 	napi = &rxq->rx_cq.napi;
2015 
2016 	if (napi_initialized) {
2017 		napi_synchronize(napi);
2018 
2019 		napi_disable(napi);
2020 
2021 		netif_napi_del(napi);
2022 	}
2023 	xdp_rxq_info_unreg(&rxq->xdp_rxq);
2024 
2025 	mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj);
2026 
2027 	mana_deinit_cq(apc, &rxq->rx_cq);
2028 
2029 	if (rxq->xdp_save_va)
2030 		put_page(virt_to_head_page(rxq->xdp_save_va));
2031 
2032 	for (i = 0; i < rxq->num_rx_buf; i++) {
2033 		rx_oob = &rxq->rx_oobs[i];
2034 
2035 		if (!rx_oob->buf_va)
2036 			continue;
2037 
2038 		dma_unmap_single(dev, rx_oob->sgl[0].address,
2039 				 rx_oob->sgl[0].size, DMA_FROM_DEVICE);
2040 
2041 		page = virt_to_head_page(rx_oob->buf_va);
2042 
2043 		if (rx_oob->from_pool)
2044 			page_pool_put_full_page(rxq->page_pool, page, false);
2045 		else
2046 			put_page(page);
2047 
2048 		rx_oob->buf_va = NULL;
2049 	}
2050 
2051 	page_pool_destroy(rxq->page_pool);
2052 
2053 	if (rxq->gdma_rq)
2054 		mana_gd_destroy_queue(gc, rxq->gdma_rq);
2055 
2056 	kfree(rxq);
2057 }
2058 
mana_fill_rx_oob(struct mana_recv_buf_oob * rx_oob,u32 mem_key,struct mana_rxq * rxq,struct device * dev)2059 static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key,
2060 			    struct mana_rxq *rxq, struct device *dev)
2061 {
2062 	struct mana_port_context *mpc = netdev_priv(rxq->ndev);
2063 	bool from_pool = false;
2064 	dma_addr_t da;
2065 	void *va;
2066 
2067 	if (mpc->rxbufs_pre)
2068 		va = mana_get_rxbuf_pre(rxq, &da);
2069 	else
2070 		va = mana_get_rxfrag(rxq, dev, &da, &from_pool, false);
2071 
2072 	if (!va)
2073 		return -ENOMEM;
2074 
2075 	rx_oob->buf_va = va;
2076 	rx_oob->from_pool = from_pool;
2077 
2078 	rx_oob->sgl[0].address = da;
2079 	rx_oob->sgl[0].size = rxq->datasize;
2080 	rx_oob->sgl[0].mem_key = mem_key;
2081 
2082 	return 0;
2083 }
2084 
2085 #define MANA_WQE_HEADER_SIZE 16
2086 #define MANA_WQE_SGE_SIZE 16
2087 
mana_alloc_rx_wqe(struct mana_port_context * apc,struct mana_rxq * rxq,u32 * rxq_size,u32 * cq_size)2088 static int mana_alloc_rx_wqe(struct mana_port_context *apc,
2089 			     struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size)
2090 {
2091 	struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
2092 	struct mana_recv_buf_oob *rx_oob;
2093 	struct device *dev = gc->dev;
2094 	u32 buf_idx;
2095 	int ret;
2096 
2097 	WARN_ON(rxq->datasize == 0);
2098 
2099 	*rxq_size = 0;
2100 	*cq_size = 0;
2101 
2102 	for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2103 		rx_oob = &rxq->rx_oobs[buf_idx];
2104 		memset(rx_oob, 0, sizeof(*rx_oob));
2105 
2106 		rx_oob->num_sge = 1;
2107 
2108 		ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq,
2109 				       dev);
2110 		if (ret)
2111 			return ret;
2112 
2113 		rx_oob->wqe_req.sgl = rx_oob->sgl;
2114 		rx_oob->wqe_req.num_sge = rx_oob->num_sge;
2115 		rx_oob->wqe_req.inline_oob_size = 0;
2116 		rx_oob->wqe_req.inline_oob_data = NULL;
2117 		rx_oob->wqe_req.flags = 0;
2118 		rx_oob->wqe_req.client_data_unit = 0;
2119 
2120 		*rxq_size += ALIGN(MANA_WQE_HEADER_SIZE +
2121 				   MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32);
2122 		*cq_size += COMP_ENTRY_SIZE;
2123 	}
2124 
2125 	return 0;
2126 }
2127 
mana_push_wqe(struct mana_rxq * rxq)2128 static int mana_push_wqe(struct mana_rxq *rxq)
2129 {
2130 	struct mana_recv_buf_oob *rx_oob;
2131 	u32 buf_idx;
2132 	int err;
2133 
2134 	for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2135 		rx_oob = &rxq->rx_oobs[buf_idx];
2136 
2137 		err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req,
2138 					    &rx_oob->wqe_inf);
2139 		if (err)
2140 			return -ENOSPC;
2141 	}
2142 
2143 	return 0;
2144 }
2145 
mana_create_page_pool(struct mana_rxq * rxq,struct gdma_context * gc)2146 static int mana_create_page_pool(struct mana_rxq *rxq, struct gdma_context *gc)
2147 {
2148 	struct page_pool_params pprm = {};
2149 	int ret;
2150 
2151 	pprm.pool_size = RX_BUFFERS_PER_QUEUE;
2152 	pprm.nid = gc->numa_node;
2153 	pprm.napi = &rxq->rx_cq.napi;
2154 
2155 	rxq->page_pool = page_pool_create(&pprm);
2156 
2157 	if (IS_ERR(rxq->page_pool)) {
2158 		ret = PTR_ERR(rxq->page_pool);
2159 		rxq->page_pool = NULL;
2160 		return ret;
2161 	}
2162 
2163 	return 0;
2164 }
2165 
mana_create_rxq(struct mana_port_context * apc,u32 rxq_idx,struct mana_eq * eq,struct net_device * ndev)2166 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,
2167 					u32 rxq_idx, struct mana_eq *eq,
2168 					struct net_device *ndev)
2169 {
2170 	struct gdma_dev *gd = apc->ac->gdma_dev;
2171 	struct mana_obj_spec wq_spec;
2172 	struct mana_obj_spec cq_spec;
2173 	struct gdma_queue_spec spec;
2174 	struct mana_cq *cq = NULL;
2175 	struct gdma_context *gc;
2176 	u32 cq_size, rq_size;
2177 	struct mana_rxq *rxq;
2178 	int err;
2179 
2180 	gc = gd->gdma_context;
2181 
2182 	rxq = kzalloc(struct_size(rxq, rx_oobs, RX_BUFFERS_PER_QUEUE),
2183 		      GFP_KERNEL);
2184 	if (!rxq)
2185 		return NULL;
2186 
2187 	rxq->ndev = ndev;
2188 	rxq->num_rx_buf = RX_BUFFERS_PER_QUEUE;
2189 	rxq->rxq_idx = rxq_idx;
2190 	rxq->rxobj = INVALID_MANA_HANDLE;
2191 
2192 	mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size,
2193 			   &rxq->headroom);
2194 
2195 	/* Create page pool for RX queue */
2196 	err = mana_create_page_pool(rxq, gc);
2197 	if (err) {
2198 		netdev_err(ndev, "Create page pool err:%d\n", err);
2199 		goto out;
2200 	}
2201 
2202 	err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size);
2203 	if (err)
2204 		goto out;
2205 
2206 	rq_size = MANA_PAGE_ALIGN(rq_size);
2207 	cq_size = MANA_PAGE_ALIGN(cq_size);
2208 
2209 	/* Create RQ */
2210 	memset(&spec, 0, sizeof(spec));
2211 	spec.type = GDMA_RQ;
2212 	spec.monitor_avl_buf = true;
2213 	spec.queue_size = rq_size;
2214 	err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq);
2215 	if (err)
2216 		goto out;
2217 
2218 	/* Create RQ's CQ */
2219 	cq = &rxq->rx_cq;
2220 	cq->type = MANA_CQ_TYPE_RX;
2221 	cq->rxq = rxq;
2222 
2223 	memset(&spec, 0, sizeof(spec));
2224 	spec.type = GDMA_CQ;
2225 	spec.monitor_avl_buf = false;
2226 	spec.queue_size = cq_size;
2227 	spec.cq.callback = mana_schedule_napi;
2228 	spec.cq.parent_eq = eq->eq;
2229 	spec.cq.context = cq;
2230 	err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2231 	if (err)
2232 		goto out;
2233 
2234 	memset(&wq_spec, 0, sizeof(wq_spec));
2235 	memset(&cq_spec, 0, sizeof(cq_spec));
2236 	wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle;
2237 	wq_spec.queue_size = rxq->gdma_rq->queue_size;
2238 
2239 	cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2240 	cq_spec.queue_size = cq->gdma_cq->queue_size;
2241 	cq_spec.modr_ctx_id = 0;
2242 	cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2243 
2244 	err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ,
2245 				 &wq_spec, &cq_spec, &rxq->rxobj);
2246 	if (err)
2247 		goto out;
2248 
2249 	rxq->gdma_rq->id = wq_spec.queue_index;
2250 	cq->gdma_cq->id = cq_spec.queue_index;
2251 
2252 	rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2253 	cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2254 
2255 	rxq->gdma_id = rxq->gdma_rq->id;
2256 	cq->gdma_id = cq->gdma_cq->id;
2257 
2258 	err = mana_push_wqe(rxq);
2259 	if (err)
2260 		goto out;
2261 
2262 	if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
2263 		err = -EINVAL;
2264 		goto out;
2265 	}
2266 
2267 	gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2268 
2269 	netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1);
2270 
2271 	WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx,
2272 				 cq->napi.napi_id));
2273 	WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL,
2274 					   rxq->page_pool));
2275 
2276 	napi_enable(&cq->napi);
2277 
2278 	mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2279 out:
2280 	if (!err)
2281 		return rxq;
2282 
2283 	netdev_err(ndev, "Failed to create RXQ: err = %d\n", err);
2284 
2285 	mana_destroy_rxq(apc, rxq, false);
2286 
2287 	if (cq)
2288 		mana_deinit_cq(apc, cq);
2289 
2290 	return NULL;
2291 }
2292 
mana_add_rx_queues(struct mana_port_context * apc,struct net_device * ndev)2293 static int mana_add_rx_queues(struct mana_port_context *apc,
2294 			      struct net_device *ndev)
2295 {
2296 	struct mana_context *ac = apc->ac;
2297 	struct mana_rxq *rxq;
2298 	int err = 0;
2299 	int i;
2300 
2301 	for (i = 0; i < apc->num_queues; i++) {
2302 		rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev);
2303 		if (!rxq) {
2304 			err = -ENOMEM;
2305 			goto out;
2306 		}
2307 
2308 		u64_stats_init(&rxq->stats.syncp);
2309 
2310 		apc->rxqs[i] = rxq;
2311 	}
2312 
2313 	apc->default_rxobj = apc->rxqs[0]->rxobj;
2314 out:
2315 	return err;
2316 }
2317 
mana_destroy_vport(struct mana_port_context * apc)2318 static void mana_destroy_vport(struct mana_port_context *apc)
2319 {
2320 	struct gdma_dev *gd = apc->ac->gdma_dev;
2321 	struct mana_rxq *rxq;
2322 	u32 rxq_idx;
2323 
2324 	for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
2325 		rxq = apc->rxqs[rxq_idx];
2326 		if (!rxq)
2327 			continue;
2328 
2329 		mana_destroy_rxq(apc, rxq, true);
2330 		apc->rxqs[rxq_idx] = NULL;
2331 	}
2332 
2333 	mana_destroy_txq(apc);
2334 	mana_uncfg_vport(apc);
2335 
2336 	if (gd->gdma_context->is_pf)
2337 		mana_pf_deregister_hw_vport(apc);
2338 }
2339 
mana_create_vport(struct mana_port_context * apc,struct net_device * net)2340 static int mana_create_vport(struct mana_port_context *apc,
2341 			     struct net_device *net)
2342 {
2343 	struct gdma_dev *gd = apc->ac->gdma_dev;
2344 	int err;
2345 
2346 	apc->default_rxobj = INVALID_MANA_HANDLE;
2347 
2348 	if (gd->gdma_context->is_pf) {
2349 		err = mana_pf_register_hw_vport(apc);
2350 		if (err)
2351 			return err;
2352 	}
2353 
2354 	err = mana_cfg_vport(apc, gd->pdid, gd->doorbell);
2355 	if (err)
2356 		return err;
2357 
2358 	return mana_create_txq(apc, net);
2359 }
2360 
mana_rss_table_init(struct mana_port_context * apc)2361 static void mana_rss_table_init(struct mana_port_context *apc)
2362 {
2363 	int i;
2364 
2365 	for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++)
2366 		apc->indir_table[i] =
2367 			ethtool_rxfh_indir_default(i, apc->num_queues);
2368 }
2369 
mana_config_rss(struct mana_port_context * apc,enum TRI_STATE rx,bool update_hash,bool update_tab)2370 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx,
2371 		    bool update_hash, bool update_tab)
2372 {
2373 	u32 queue_idx;
2374 	int err;
2375 	int i;
2376 
2377 	if (update_tab) {
2378 		for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) {
2379 			queue_idx = apc->indir_table[i];
2380 			apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj;
2381 		}
2382 	}
2383 
2384 	err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab);
2385 	if (err)
2386 		return err;
2387 
2388 	mana_fence_rqs(apc);
2389 
2390 	return 0;
2391 }
2392 
mana_query_gf_stats(struct mana_port_context * apc)2393 void mana_query_gf_stats(struct mana_port_context *apc)
2394 {
2395 	struct mana_query_gf_stat_resp resp = {};
2396 	struct mana_query_gf_stat_req req = {};
2397 	struct net_device *ndev = apc->ndev;
2398 	int err;
2399 
2400 	mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_GF_STAT,
2401 			     sizeof(req), sizeof(resp));
2402 	req.req_stats = STATISTICS_FLAGS_HC_TX_BYTES |
2403 			STATISTICS_FLAGS_HC_TX_UCAST_PACKETS |
2404 			STATISTICS_FLAGS_HC_TX_UCAST_BYTES |
2405 			STATISTICS_FLAGS_HC_TX_MCAST_PACKETS |
2406 			STATISTICS_FLAGS_HC_TX_MCAST_BYTES |
2407 			STATISTICS_FLAGS_HC_TX_BCAST_PACKETS |
2408 			STATISTICS_FLAGS_HC_TX_BCAST_BYTES;
2409 
2410 	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
2411 				sizeof(resp));
2412 	if (err) {
2413 		netdev_err(ndev, "Failed to query GF stats: %d\n", err);
2414 		return;
2415 	}
2416 	err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_GF_STAT,
2417 				   sizeof(resp));
2418 	if (err || resp.hdr.status) {
2419 		netdev_err(ndev, "Failed to query GF stats: %d, 0x%x\n", err,
2420 			   resp.hdr.status);
2421 		return;
2422 	}
2423 
2424 	apc->eth_stats.hc_tx_bytes = resp.hc_tx_bytes;
2425 	apc->eth_stats.hc_tx_ucast_pkts = resp.hc_tx_ucast_pkts;
2426 	apc->eth_stats.hc_tx_ucast_bytes = resp.hc_tx_ucast_bytes;
2427 	apc->eth_stats.hc_tx_bcast_pkts = resp.hc_tx_bcast_pkts;
2428 	apc->eth_stats.hc_tx_bcast_bytes = resp.hc_tx_bcast_bytes;
2429 	apc->eth_stats.hc_tx_mcast_pkts = resp.hc_tx_mcast_pkts;
2430 	apc->eth_stats.hc_tx_mcast_bytes = resp.hc_tx_mcast_bytes;
2431 }
2432 
mana_init_port(struct net_device * ndev)2433 static int mana_init_port(struct net_device *ndev)
2434 {
2435 	struct mana_port_context *apc = netdev_priv(ndev);
2436 	u32 max_txq, max_rxq, max_queues;
2437 	int port_idx = apc->port_idx;
2438 	u32 num_indirect_entries;
2439 	int err;
2440 
2441 	err = mana_init_port_context(apc);
2442 	if (err)
2443 		return err;
2444 
2445 	err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq,
2446 				   &num_indirect_entries);
2447 	if (err) {
2448 		netdev_err(ndev, "Failed to query info for vPort %d\n",
2449 			   port_idx);
2450 		goto reset_apc;
2451 	}
2452 
2453 	max_queues = min_t(u32, max_txq, max_rxq);
2454 	if (apc->max_queues > max_queues)
2455 		apc->max_queues = max_queues;
2456 
2457 	if (apc->num_queues > apc->max_queues)
2458 		apc->num_queues = apc->max_queues;
2459 
2460 	eth_hw_addr_set(ndev, apc->mac_addr);
2461 
2462 	return 0;
2463 
2464 reset_apc:
2465 	kfree(apc->rxqs);
2466 	apc->rxqs = NULL;
2467 	return err;
2468 }
2469 
mana_alloc_queues(struct net_device * ndev)2470 int mana_alloc_queues(struct net_device *ndev)
2471 {
2472 	struct mana_port_context *apc = netdev_priv(ndev);
2473 	struct gdma_dev *gd = apc->ac->gdma_dev;
2474 	int err;
2475 
2476 	err = mana_create_vport(apc, ndev);
2477 	if (err)
2478 		return err;
2479 
2480 	err = netif_set_real_num_tx_queues(ndev, apc->num_queues);
2481 	if (err)
2482 		goto destroy_vport;
2483 
2484 	err = mana_add_rx_queues(apc, ndev);
2485 	if (err)
2486 		goto destroy_vport;
2487 
2488 	apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE;
2489 
2490 	err = netif_set_real_num_rx_queues(ndev, apc->num_queues);
2491 	if (err)
2492 		goto destroy_vport;
2493 
2494 	mana_rss_table_init(apc);
2495 
2496 	err = mana_config_rss(apc, TRI_STATE_TRUE, true, true);
2497 	if (err)
2498 		goto destroy_vport;
2499 
2500 	if (gd->gdma_context->is_pf) {
2501 		err = mana_pf_register_filter(apc);
2502 		if (err)
2503 			goto destroy_vport;
2504 	}
2505 
2506 	mana_chn_setxdp(apc, mana_xdp_get(apc));
2507 
2508 	return 0;
2509 
2510 destroy_vport:
2511 	mana_destroy_vport(apc);
2512 	return err;
2513 }
2514 
mana_attach(struct net_device * ndev)2515 int mana_attach(struct net_device *ndev)
2516 {
2517 	struct mana_port_context *apc = netdev_priv(ndev);
2518 	int err;
2519 
2520 	ASSERT_RTNL();
2521 
2522 	err = mana_init_port(ndev);
2523 	if (err)
2524 		return err;
2525 
2526 	if (apc->port_st_save) {
2527 		err = mana_alloc_queues(ndev);
2528 		if (err) {
2529 			mana_cleanup_port_context(apc);
2530 			return err;
2531 		}
2532 	}
2533 
2534 	apc->port_is_up = apc->port_st_save;
2535 
2536 	/* Ensure port state updated before txq state */
2537 	smp_wmb();
2538 
2539 	if (apc->port_is_up)
2540 		netif_carrier_on(ndev);
2541 
2542 	netif_device_attach(ndev);
2543 
2544 	return 0;
2545 }
2546 
mana_dealloc_queues(struct net_device * ndev)2547 static int mana_dealloc_queues(struct net_device *ndev)
2548 {
2549 	struct mana_port_context *apc = netdev_priv(ndev);
2550 	unsigned long timeout = jiffies + 120 * HZ;
2551 	struct gdma_dev *gd = apc->ac->gdma_dev;
2552 	struct mana_txq *txq;
2553 	struct sk_buff *skb;
2554 	int i, err;
2555 	u32 tsleep;
2556 
2557 	if (apc->port_is_up)
2558 		return -EINVAL;
2559 
2560 	mana_chn_setxdp(apc, NULL);
2561 
2562 	if (gd->gdma_context->is_pf)
2563 		mana_pf_deregister_filter(apc);
2564 
2565 	/* No packet can be transmitted now since apc->port_is_up is false.
2566 	 * There is still a tiny chance that mana_poll_tx_cq() can re-enable
2567 	 * a txq because it may not timely see apc->port_is_up being cleared
2568 	 * to false, but it doesn't matter since mana_start_xmit() drops any
2569 	 * new packets due to apc->port_is_up being false.
2570 	 *
2571 	 * Drain all the in-flight TX packets.
2572 	 * A timeout of 120 seconds for all the queues is used.
2573 	 * This will break the while loop when h/w is not responding.
2574 	 * This value of 120 has been decided here considering max
2575 	 * number of queues.
2576 	 */
2577 
2578 	for (i = 0; i < apc->num_queues; i++) {
2579 		txq = &apc->tx_qp[i].txq;
2580 		tsleep = 1000;
2581 		while (atomic_read(&txq->pending_sends) > 0 &&
2582 		       time_before(jiffies, timeout)) {
2583 			usleep_range(tsleep, tsleep + 1000);
2584 			tsleep <<= 1;
2585 		}
2586 		if (atomic_read(&txq->pending_sends)) {
2587 			err = pcie_flr(to_pci_dev(gd->gdma_context->dev));
2588 			if (err) {
2589 				netdev_err(ndev, "flr failed %d with %d pkts pending in txq %u\n",
2590 					   err, atomic_read(&txq->pending_sends),
2591 					   txq->gdma_txq_id);
2592 			}
2593 			break;
2594 		}
2595 	}
2596 
2597 	for (i = 0; i < apc->num_queues; i++) {
2598 		txq = &apc->tx_qp[i].txq;
2599 		while ((skb = skb_dequeue(&txq->pending_skbs))) {
2600 			mana_unmap_skb(skb, apc);
2601 			dev_kfree_skb_any(skb);
2602 		}
2603 		atomic_set(&txq->pending_sends, 0);
2604 	}
2605 	/* We're 100% sure the queues can no longer be woken up, because
2606 	 * we're sure now mana_poll_tx_cq() can't be running.
2607 	 */
2608 
2609 	apc->rss_state = TRI_STATE_FALSE;
2610 	err = mana_config_rss(apc, TRI_STATE_FALSE, false, false);
2611 	if (err) {
2612 		netdev_err(ndev, "Failed to disable vPort: %d\n", err);
2613 		return err;
2614 	}
2615 
2616 	mana_destroy_vport(apc);
2617 
2618 	return 0;
2619 }
2620 
mana_detach(struct net_device * ndev,bool from_close)2621 int mana_detach(struct net_device *ndev, bool from_close)
2622 {
2623 	struct mana_port_context *apc = netdev_priv(ndev);
2624 	int err;
2625 
2626 	ASSERT_RTNL();
2627 
2628 	apc->port_st_save = apc->port_is_up;
2629 	apc->port_is_up = false;
2630 
2631 	/* Ensure port state updated before txq state */
2632 	smp_wmb();
2633 
2634 	netif_tx_disable(ndev);
2635 	netif_carrier_off(ndev);
2636 
2637 	if (apc->port_st_save) {
2638 		err = mana_dealloc_queues(ndev);
2639 		if (err)
2640 			return err;
2641 	}
2642 
2643 	if (!from_close) {
2644 		netif_device_detach(ndev);
2645 		mana_cleanup_port_context(apc);
2646 	}
2647 
2648 	return 0;
2649 }
2650 
mana_probe_port(struct mana_context * ac,int port_idx,struct net_device ** ndev_storage)2651 static int mana_probe_port(struct mana_context *ac, int port_idx,
2652 			   struct net_device **ndev_storage)
2653 {
2654 	struct gdma_context *gc = ac->gdma_dev->gdma_context;
2655 	struct mana_port_context *apc;
2656 	struct net_device *ndev;
2657 	int err;
2658 
2659 	ndev = alloc_etherdev_mq(sizeof(struct mana_port_context),
2660 				 gc->max_num_queues);
2661 	if (!ndev)
2662 		return -ENOMEM;
2663 
2664 	*ndev_storage = ndev;
2665 
2666 	apc = netdev_priv(ndev);
2667 	apc->ac = ac;
2668 	apc->ndev = ndev;
2669 	apc->max_queues = gc->max_num_queues;
2670 	apc->num_queues = gc->max_num_queues;
2671 	apc->port_handle = INVALID_MANA_HANDLE;
2672 	apc->pf_filter_handle = INVALID_MANA_HANDLE;
2673 	apc->port_idx = port_idx;
2674 
2675 	mutex_init(&apc->vport_mutex);
2676 	apc->vport_use_count = 0;
2677 
2678 	ndev->netdev_ops = &mana_devops;
2679 	ndev->ethtool_ops = &mana_ethtool_ops;
2680 	ndev->mtu = ETH_DATA_LEN;
2681 	ndev->max_mtu = gc->adapter_mtu - ETH_HLEN;
2682 	ndev->min_mtu = ETH_MIN_MTU;
2683 	ndev->needed_headroom = MANA_HEADROOM;
2684 	ndev->dev_port = port_idx;
2685 	SET_NETDEV_DEV(ndev, gc->dev);
2686 
2687 	netif_carrier_off(ndev);
2688 
2689 	netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE);
2690 
2691 	err = mana_init_port(ndev);
2692 	if (err)
2693 		goto free_net;
2694 
2695 	netdev_lockdep_set_classes(ndev);
2696 
2697 	ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
2698 	ndev->hw_features |= NETIF_F_RXCSUM;
2699 	ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
2700 	ndev->hw_features |= NETIF_F_RXHASH;
2701 	ndev->features = ndev->hw_features | NETIF_F_HW_VLAN_CTAG_TX |
2702 			 NETIF_F_HW_VLAN_CTAG_RX;
2703 	ndev->vlan_features = ndev->features;
2704 	ndev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT |
2705 			     NETDEV_XDP_ACT_NDO_XMIT;
2706 
2707 	err = register_netdev(ndev);
2708 	if (err) {
2709 		netdev_err(ndev, "Unable to register netdev.\n");
2710 		goto reset_apc;
2711 	}
2712 
2713 	return 0;
2714 
2715 reset_apc:
2716 	kfree(apc->rxqs);
2717 	apc->rxqs = NULL;
2718 free_net:
2719 	*ndev_storage = NULL;
2720 	netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err);
2721 	free_netdev(ndev);
2722 	return err;
2723 }
2724 
adev_release(struct device * dev)2725 static void adev_release(struct device *dev)
2726 {
2727 	struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev);
2728 
2729 	kfree(madev);
2730 }
2731 
remove_adev(struct gdma_dev * gd)2732 static void remove_adev(struct gdma_dev *gd)
2733 {
2734 	struct auxiliary_device *adev = gd->adev;
2735 	int id = adev->id;
2736 
2737 	auxiliary_device_delete(adev);
2738 	auxiliary_device_uninit(adev);
2739 
2740 	mana_adev_idx_free(id);
2741 	gd->adev = NULL;
2742 }
2743 
add_adev(struct gdma_dev * gd)2744 static int add_adev(struct gdma_dev *gd)
2745 {
2746 	struct auxiliary_device *adev;
2747 	struct mana_adev *madev;
2748 	int ret;
2749 
2750 	madev = kzalloc(sizeof(*madev), GFP_KERNEL);
2751 	if (!madev)
2752 		return -ENOMEM;
2753 
2754 	adev = &madev->adev;
2755 	ret = mana_adev_idx_alloc();
2756 	if (ret < 0)
2757 		goto idx_fail;
2758 	adev->id = ret;
2759 
2760 	adev->name = "rdma";
2761 	adev->dev.parent = gd->gdma_context->dev;
2762 	adev->dev.release = adev_release;
2763 	madev->mdev = gd;
2764 
2765 	ret = auxiliary_device_init(adev);
2766 	if (ret)
2767 		goto init_fail;
2768 
2769 	/* madev is owned by the auxiliary device */
2770 	madev = NULL;
2771 	ret = auxiliary_device_add(adev);
2772 	if (ret)
2773 		goto add_fail;
2774 
2775 	gd->adev = adev;
2776 	return 0;
2777 
2778 add_fail:
2779 	auxiliary_device_uninit(adev);
2780 
2781 init_fail:
2782 	mana_adev_idx_free(adev->id);
2783 
2784 idx_fail:
2785 	kfree(madev);
2786 
2787 	return ret;
2788 }
2789 
mana_probe(struct gdma_dev * gd,bool resuming)2790 int mana_probe(struct gdma_dev *gd, bool resuming)
2791 {
2792 	struct gdma_context *gc = gd->gdma_context;
2793 	struct mana_context *ac = gd->driver_data;
2794 	struct device *dev = gc->dev;
2795 	u16 num_ports = 0;
2796 	int err;
2797 	int i;
2798 
2799 	dev_info(dev,
2800 		 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n",
2801 		 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION);
2802 
2803 	err = mana_gd_register_device(gd);
2804 	if (err)
2805 		return err;
2806 
2807 	if (!resuming) {
2808 		ac = kzalloc(sizeof(*ac), GFP_KERNEL);
2809 		if (!ac)
2810 			return -ENOMEM;
2811 
2812 		ac->gdma_dev = gd;
2813 		gd->driver_data = ac;
2814 	}
2815 
2816 	err = mana_create_eq(ac);
2817 	if (err)
2818 		goto out;
2819 
2820 	err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION,
2821 				    MANA_MICRO_VERSION, &num_ports);
2822 	if (err)
2823 		goto out;
2824 
2825 	if (!resuming) {
2826 		ac->num_ports = num_ports;
2827 	} else {
2828 		if (ac->num_ports != num_ports) {
2829 			dev_err(dev, "The number of vPorts changed: %d->%d\n",
2830 				ac->num_ports, num_ports);
2831 			err = -EPROTO;
2832 			goto out;
2833 		}
2834 	}
2835 
2836 	if (ac->num_ports == 0)
2837 		dev_err(dev, "Failed to detect any vPort\n");
2838 
2839 	if (ac->num_ports > MAX_PORTS_IN_MANA_DEV)
2840 		ac->num_ports = MAX_PORTS_IN_MANA_DEV;
2841 
2842 	if (!resuming) {
2843 		for (i = 0; i < ac->num_ports; i++) {
2844 			err = mana_probe_port(ac, i, &ac->ports[i]);
2845 			if (err)
2846 				break;
2847 		}
2848 	} else {
2849 		for (i = 0; i < ac->num_ports; i++) {
2850 			rtnl_lock();
2851 			err = mana_attach(ac->ports[i]);
2852 			rtnl_unlock();
2853 			if (err)
2854 				break;
2855 		}
2856 	}
2857 
2858 	err = add_adev(gd);
2859 out:
2860 	if (err)
2861 		mana_remove(gd, false);
2862 
2863 	return err;
2864 }
2865 
mana_remove(struct gdma_dev * gd,bool suspending)2866 void mana_remove(struct gdma_dev *gd, bool suspending)
2867 {
2868 	struct gdma_context *gc = gd->gdma_context;
2869 	struct mana_context *ac = gd->driver_data;
2870 	struct device *dev = gc->dev;
2871 	struct net_device *ndev;
2872 	int err;
2873 	int i;
2874 
2875 	/* adev currently doesn't support suspending, always remove it */
2876 	if (gd->adev)
2877 		remove_adev(gd);
2878 
2879 	for (i = 0; i < ac->num_ports; i++) {
2880 		ndev = ac->ports[i];
2881 		if (!ndev) {
2882 			if (i == 0)
2883 				dev_err(dev, "No net device to remove\n");
2884 			goto out;
2885 		}
2886 
2887 		/* All cleanup actions should stay after rtnl_lock(), otherwise
2888 		 * other functions may access partially cleaned up data.
2889 		 */
2890 		rtnl_lock();
2891 
2892 		err = mana_detach(ndev, false);
2893 		if (err)
2894 			netdev_err(ndev, "Failed to detach vPort %d: %d\n",
2895 				   i, err);
2896 
2897 		if (suspending) {
2898 			/* No need to unregister the ndev. */
2899 			rtnl_unlock();
2900 			continue;
2901 		}
2902 
2903 		unregister_netdevice(ndev);
2904 
2905 		rtnl_unlock();
2906 
2907 		free_netdev(ndev);
2908 	}
2909 
2910 	mana_destroy_eq(ac);
2911 out:
2912 	mana_gd_deregister_device(gd);
2913 
2914 	if (suspending)
2915 		return;
2916 
2917 	gd->driver_data = NULL;
2918 	gd->gdma_context = NULL;
2919 	kfree(ac);
2920 }
2921