Loading...
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/debugfs.h>
7#include <linux/inetdevice.h>
8#include <linux/etherdevice.h>
9#include <linux/ethtool.h>
10#include <linux/filter.h>
11#include <linux/mm.h>
12#include <linux/pci.h>
13
14#include <net/checksum.h>
15#include <net/ip6_checksum.h>
16#include <net/page_pool/helpers.h>
17#include <net/xdp.h>
18
19#include <net/mana/mana.h>
20#include <net/mana/mana_auxiliary.h>
21
22static DEFINE_IDA(mana_adev_ida);
23
24static int mana_adev_idx_alloc(void)
25{
26 return ida_alloc(&mana_adev_ida, GFP_KERNEL);
27}
28
29static void mana_adev_idx_free(int idx)
30{
31 ida_free(&mana_adev_ida, idx);
32}
33
34static ssize_t mana_dbg_q_read(struct file *filp, char __user *buf, size_t count,
35 loff_t *pos)
36{
37 struct gdma_queue *gdma_q = filp->private_data;
38
39 return simple_read_from_buffer(buf, count, pos, gdma_q->queue_mem_ptr,
40 gdma_q->queue_size);
41}
42
43static const struct file_operations mana_dbg_q_fops = {
44 .owner = THIS_MODULE,
45 .open = simple_open,
46 .read = mana_dbg_q_read,
47};
48
49/* Microsoft Azure Network Adapter (MANA) functions */
50
51static int mana_open(struct net_device *ndev)
52{
53 struct mana_port_context *apc = netdev_priv(ndev);
54 int err;
55
56 err = mana_alloc_queues(ndev);
57 if (err)
58 return err;
59
60 apc->port_is_up = true;
61
62 /* Ensure port state updated before txq state */
63 smp_wmb();
64
65 netif_carrier_on(ndev);
66 netif_tx_wake_all_queues(ndev);
67
68 return 0;
69}
70
71static int mana_close(struct net_device *ndev)
72{
73 struct mana_port_context *apc = netdev_priv(ndev);
74
75 if (!apc->port_is_up)
76 return 0;
77
78 return mana_detach(ndev, true);
79}
80
81static bool mana_can_tx(struct gdma_queue *wq)
82{
83 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE;
84}
85
86static unsigned int mana_checksum_info(struct sk_buff *skb)
87{
88 if (skb->protocol == htons(ETH_P_IP)) {
89 struct iphdr *ip = ip_hdr(skb);
90
91 if (ip->protocol == IPPROTO_TCP)
92 return IPPROTO_TCP;
93
94 if (ip->protocol == IPPROTO_UDP)
95 return IPPROTO_UDP;
96 } else if (skb->protocol == htons(ETH_P_IPV6)) {
97 struct ipv6hdr *ip6 = ipv6_hdr(skb);
98
99 if (ip6->nexthdr == IPPROTO_TCP)
100 return IPPROTO_TCP;
101
102 if (ip6->nexthdr == IPPROTO_UDP)
103 return IPPROTO_UDP;
104 }
105
106 /* No csum offloading */
107 return 0;
108}
109
110static void mana_add_sge(struct mana_tx_package *tp, struct mana_skb_head *ash,
111 int sg_i, dma_addr_t da, int sge_len, u32 gpa_mkey)
112{
113 ash->dma_handle[sg_i] = da;
114 ash->size[sg_i] = sge_len;
115
116 tp->wqe_req.sgl[sg_i].address = da;
117 tp->wqe_req.sgl[sg_i].mem_key = gpa_mkey;
118 tp->wqe_req.sgl[sg_i].size = sge_len;
119}
120
121static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc,
122 struct mana_tx_package *tp, int gso_hs)
123{
124 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
125 int hsg = 1; /* num of SGEs of linear part */
126 struct gdma_dev *gd = apc->ac->gdma_dev;
127 int skb_hlen = skb_headlen(skb);
128 int sge0_len, sge1_len = 0;
129 struct gdma_context *gc;
130 struct device *dev;
131 skb_frag_t *frag;
132 dma_addr_t da;
133 int sg_i;
134 int i;
135
136 gc = gd->gdma_context;
137 dev = gc->dev;
138
139 if (gso_hs && gso_hs < skb_hlen) {
140 sge0_len = gso_hs;
141 sge1_len = skb_hlen - gso_hs;
142 } else {
143 sge0_len = skb_hlen;
144 }
145
146 da = dma_map_single(dev, skb->data, sge0_len, DMA_TO_DEVICE);
147 if (dma_mapping_error(dev, da))
148 return -ENOMEM;
149
150 mana_add_sge(tp, ash, 0, da, sge0_len, gd->gpa_mkey);
151
152 if (sge1_len) {
153 sg_i = 1;
154 da = dma_map_single(dev, skb->data + sge0_len, sge1_len,
155 DMA_TO_DEVICE);
156 if (dma_mapping_error(dev, da))
157 goto frag_err;
158
159 mana_add_sge(tp, ash, sg_i, da, sge1_len, gd->gpa_mkey);
160 hsg = 2;
161 }
162
163 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
164 sg_i = hsg + i;
165
166 frag = &skb_shinfo(skb)->frags[i];
167 da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag),
168 DMA_TO_DEVICE);
169 if (dma_mapping_error(dev, da))
170 goto frag_err;
171
172 mana_add_sge(tp, ash, sg_i, da, skb_frag_size(frag),
173 gd->gpa_mkey);
174 }
175
176 return 0;
177
178frag_err:
179 for (i = sg_i - 1; i >= hsg; i--)
180 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
181 DMA_TO_DEVICE);
182
183 for (i = hsg - 1; i >= 0; i--)
184 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i],
185 DMA_TO_DEVICE);
186
187 return -ENOMEM;
188}
189
190/* Handle the case when GSO SKB linear length is too large.
191 * MANA NIC requires GSO packets to put only the packet header to SGE0.
192 * So, we need 2 SGEs for the skb linear part which contains more than the
193 * header.
194 * Return a positive value for the number of SGEs, or a negative value
195 * for an error.
196 */
197static int mana_fix_skb_head(struct net_device *ndev, struct sk_buff *skb,
198 int gso_hs)
199{
200 int num_sge = 1 + skb_shinfo(skb)->nr_frags;
201 int skb_hlen = skb_headlen(skb);
202
203 if (gso_hs < skb_hlen) {
204 num_sge++;
205 } else if (gso_hs > skb_hlen) {
206 if (net_ratelimit())
207 netdev_err(ndev,
208 "TX nonlinear head: hs:%d, skb_hlen:%d\n",
209 gso_hs, skb_hlen);
210
211 return -EINVAL;
212 }
213
214 return num_sge;
215}
216
217/* Get the GSO packet's header size */
218static int mana_get_gso_hs(struct sk_buff *skb)
219{
220 int gso_hs;
221
222 if (skb->encapsulation) {
223 gso_hs = skb_inner_tcp_all_headers(skb);
224 } else {
225 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
226 gso_hs = skb_transport_offset(skb) +
227 sizeof(struct udphdr);
228 } else {
229 gso_hs = skb_tcp_all_headers(skb);
230 }
231 }
232
233 return gso_hs;
234}
235
236netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev)
237{
238 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT;
239 struct mana_port_context *apc = netdev_priv(ndev);
240 int gso_hs = 0; /* zero for non-GSO pkts */
241 u16 txq_idx = skb_get_queue_mapping(skb);
242 struct gdma_dev *gd = apc->ac->gdma_dev;
243 bool ipv4 = false, ipv6 = false;
244 struct mana_tx_package pkg = {};
245 struct netdev_queue *net_txq;
246 struct mana_stats_tx *tx_stats;
247 struct gdma_queue *gdma_sq;
248 unsigned int csum_type;
249 struct mana_txq *txq;
250 struct mana_cq *cq;
251 int err, len;
252
253 if (unlikely(!apc->port_is_up))
254 goto tx_drop;
255
256 if (skb_cow_head(skb, MANA_HEADROOM))
257 goto tx_drop_count;
258
259 txq = &apc->tx_qp[txq_idx].txq;
260 gdma_sq = txq->gdma_sq;
261 cq = &apc->tx_qp[txq_idx].tx_cq;
262 tx_stats = &txq->stats;
263
264 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id;
265 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame;
266
267 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) {
268 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset;
269 pkt_fmt = MANA_LONG_PKT_FMT;
270 } else {
271 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset;
272 }
273
274 if (skb_vlan_tag_present(skb)) {
275 pkt_fmt = MANA_LONG_PKT_FMT;
276 pkg.tx_oob.l_oob.inject_vlan_pri_tag = 1;
277 pkg.tx_oob.l_oob.pcp = skb_vlan_tag_get_prio(skb);
278 pkg.tx_oob.l_oob.dei = skb_vlan_tag_get_cfi(skb);
279 pkg.tx_oob.l_oob.vlan_id = skb_vlan_tag_get_id(skb);
280 }
281
282 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt;
283
284 if (pkt_fmt == MANA_SHORT_PKT_FMT) {
285 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob);
286 u64_stats_update_begin(&tx_stats->syncp);
287 tx_stats->short_pkt_fmt++;
288 u64_stats_update_end(&tx_stats->syncp);
289 } else {
290 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob);
291 u64_stats_update_begin(&tx_stats->syncp);
292 tx_stats->long_pkt_fmt++;
293 u64_stats_update_end(&tx_stats->syncp);
294 }
295
296 pkg.wqe_req.inline_oob_data = &pkg.tx_oob;
297 pkg.wqe_req.flags = 0;
298 pkg.wqe_req.client_data_unit = 0;
299
300 pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags;
301
302 if (skb->protocol == htons(ETH_P_IP))
303 ipv4 = true;
304 else if (skb->protocol == htons(ETH_P_IPV6))
305 ipv6 = true;
306
307 if (skb_is_gso(skb)) {
308 int num_sge;
309
310 gso_hs = mana_get_gso_hs(skb);
311
312 num_sge = mana_fix_skb_head(ndev, skb, gso_hs);
313 if (num_sge > 0)
314 pkg.wqe_req.num_sge = num_sge;
315 else
316 goto tx_drop_count;
317
318 u64_stats_update_begin(&tx_stats->syncp);
319 if (skb->encapsulation) {
320 tx_stats->tso_inner_packets++;
321 tx_stats->tso_inner_bytes += skb->len - gso_hs;
322 } else {
323 tx_stats->tso_packets++;
324 tx_stats->tso_bytes += skb->len - gso_hs;
325 }
326 u64_stats_update_end(&tx_stats->syncp);
327
328 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
329 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
330
331 pkg.tx_oob.s_oob.comp_iphdr_csum = 1;
332 pkg.tx_oob.s_oob.comp_tcp_csum = 1;
333 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
334
335 pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size;
336 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0;
337 if (ipv4) {
338 ip_hdr(skb)->tot_len = 0;
339 ip_hdr(skb)->check = 0;
340 tcp_hdr(skb)->check =
341 ~csum_tcpudp_magic(ip_hdr(skb)->saddr,
342 ip_hdr(skb)->daddr, 0,
343 IPPROTO_TCP, 0);
344 } else {
345 ipv6_hdr(skb)->payload_len = 0;
346 tcp_hdr(skb)->check =
347 ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
348 &ipv6_hdr(skb)->daddr, 0,
349 IPPROTO_TCP, 0);
350 }
351 } else if (skb->ip_summed == CHECKSUM_PARTIAL) {
352 csum_type = mana_checksum_info(skb);
353
354 u64_stats_update_begin(&tx_stats->syncp);
355 tx_stats->csum_partial++;
356 u64_stats_update_end(&tx_stats->syncp);
357
358 if (csum_type == IPPROTO_TCP) {
359 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
360 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
361
362 pkg.tx_oob.s_oob.comp_tcp_csum = 1;
363 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
364
365 } else if (csum_type == IPPROTO_UDP) {
366 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
367 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
368
369 pkg.tx_oob.s_oob.comp_udp_csum = 1;
370 } else {
371 /* Can't do offload of this type of checksum */
372 if (skb_checksum_help(skb))
373 goto tx_drop_count;
374 }
375 }
376
377 WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES);
378
379 if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) {
380 pkg.wqe_req.sgl = pkg.sgl_array;
381 } else {
382 pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge,
383 sizeof(struct gdma_sge),
384 GFP_ATOMIC);
385 if (!pkg.sgl_ptr)
386 goto tx_drop_count;
387
388 pkg.wqe_req.sgl = pkg.sgl_ptr;
389 }
390
391 if (mana_map_skb(skb, apc, &pkg, gso_hs)) {
392 u64_stats_update_begin(&tx_stats->syncp);
393 tx_stats->mana_map_err++;
394 u64_stats_update_end(&tx_stats->syncp);
395 goto free_sgl_ptr;
396 }
397
398 skb_queue_tail(&txq->pending_skbs, skb);
399
400 len = skb->len;
401 net_txq = netdev_get_tx_queue(ndev, txq_idx);
402
403 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req,
404 (struct gdma_posted_wqe_info *)skb->cb);
405 if (!mana_can_tx(gdma_sq)) {
406 netif_tx_stop_queue(net_txq);
407 apc->eth_stats.stop_queue++;
408 }
409
410 if (err) {
411 (void)skb_dequeue_tail(&txq->pending_skbs);
412 netdev_warn(ndev, "Failed to post TX OOB: %d\n", err);
413 err = NETDEV_TX_BUSY;
414 goto tx_busy;
415 }
416
417 err = NETDEV_TX_OK;
418 atomic_inc(&txq->pending_sends);
419
420 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq);
421
422 /* skb may be freed after mana_gd_post_work_request. Do not use it. */
423 skb = NULL;
424
425 tx_stats = &txq->stats;
426 u64_stats_update_begin(&tx_stats->syncp);
427 tx_stats->packets++;
428 tx_stats->bytes += len;
429 u64_stats_update_end(&tx_stats->syncp);
430
431tx_busy:
432 if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) {
433 netif_tx_wake_queue(net_txq);
434 apc->eth_stats.wake_queue++;
435 }
436
437 kfree(pkg.sgl_ptr);
438 return err;
439
440free_sgl_ptr:
441 kfree(pkg.sgl_ptr);
442tx_drop_count:
443 ndev->stats.tx_dropped++;
444tx_drop:
445 dev_kfree_skb_any(skb);
446 return NETDEV_TX_OK;
447}
448
449static void mana_get_stats64(struct net_device *ndev,
450 struct rtnl_link_stats64 *st)
451{
452 struct mana_port_context *apc = netdev_priv(ndev);
453 unsigned int num_queues = apc->num_queues;
454 struct mana_stats_rx *rx_stats;
455 struct mana_stats_tx *tx_stats;
456 unsigned int start;
457 u64 packets, bytes;
458 int q;
459
460 if (!apc->port_is_up)
461 return;
462
463 netdev_stats_to_stats64(st, &ndev->stats);
464
465 for (q = 0; q < num_queues; q++) {
466 rx_stats = &apc->rxqs[q]->stats;
467
468 do {
469 start = u64_stats_fetch_begin(&rx_stats->syncp);
470 packets = rx_stats->packets;
471 bytes = rx_stats->bytes;
472 } while (u64_stats_fetch_retry(&rx_stats->syncp, start));
473
474 st->rx_packets += packets;
475 st->rx_bytes += bytes;
476 }
477
478 for (q = 0; q < num_queues; q++) {
479 tx_stats = &apc->tx_qp[q].txq.stats;
480
481 do {
482 start = u64_stats_fetch_begin(&tx_stats->syncp);
483 packets = tx_stats->packets;
484 bytes = tx_stats->bytes;
485 } while (u64_stats_fetch_retry(&tx_stats->syncp, start));
486
487 st->tx_packets += packets;
488 st->tx_bytes += bytes;
489 }
490}
491
492static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb,
493 int old_q)
494{
495 struct mana_port_context *apc = netdev_priv(ndev);
496 u32 hash = skb_get_hash(skb);
497 struct sock *sk = skb->sk;
498 int txq;
499
500 txq = apc->indir_table[hash & (apc->indir_table_sz - 1)];
501
502 if (txq != old_q && sk && sk_fullsock(sk) &&
503 rcu_access_pointer(sk->sk_dst_cache))
504 sk_tx_queue_set(sk, txq);
505
506 return txq;
507}
508
509static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb,
510 struct net_device *sb_dev)
511{
512 int txq;
513
514 if (ndev->real_num_tx_queues == 1)
515 return 0;
516
517 txq = sk_tx_queue_get(skb->sk);
518
519 if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) {
520 if (skb_rx_queue_recorded(skb))
521 txq = skb_get_rx_queue(skb);
522 else
523 txq = mana_get_tx_queue(ndev, skb, txq);
524 }
525
526 return txq;
527}
528
529/* Release pre-allocated RX buffers */
530void mana_pre_dealloc_rxbufs(struct mana_port_context *mpc)
531{
532 struct device *dev;
533 int i;
534
535 dev = mpc->ac->gdma_dev->gdma_context->dev;
536
537 if (!mpc->rxbufs_pre)
538 goto out1;
539
540 if (!mpc->das_pre)
541 goto out2;
542
543 while (mpc->rxbpre_total) {
544 i = --mpc->rxbpre_total;
545 dma_unmap_single(dev, mpc->das_pre[i], mpc->rxbpre_datasize,
546 DMA_FROM_DEVICE);
547 put_page(virt_to_head_page(mpc->rxbufs_pre[i]));
548 }
549
550 kfree(mpc->das_pre);
551 mpc->das_pre = NULL;
552
553out2:
554 kfree(mpc->rxbufs_pre);
555 mpc->rxbufs_pre = NULL;
556
557out1:
558 mpc->rxbpre_datasize = 0;
559 mpc->rxbpre_alloc_size = 0;
560 mpc->rxbpre_headroom = 0;
561}
562
563/* Get a buffer from the pre-allocated RX buffers */
564static void *mana_get_rxbuf_pre(struct mana_rxq *rxq, dma_addr_t *da)
565{
566 struct net_device *ndev = rxq->ndev;
567 struct mana_port_context *mpc;
568 void *va;
569
570 mpc = netdev_priv(ndev);
571
572 if (!mpc->rxbufs_pre || !mpc->das_pre || !mpc->rxbpre_total) {
573 netdev_err(ndev, "No RX pre-allocated bufs\n");
574 return NULL;
575 }
576
577 /* Check sizes to catch unexpected coding error */
578 if (mpc->rxbpre_datasize != rxq->datasize) {
579 netdev_err(ndev, "rxbpre_datasize mismatch: %u: %u\n",
580 mpc->rxbpre_datasize, rxq->datasize);
581 return NULL;
582 }
583
584 if (mpc->rxbpre_alloc_size != rxq->alloc_size) {
585 netdev_err(ndev, "rxbpre_alloc_size mismatch: %u: %u\n",
586 mpc->rxbpre_alloc_size, rxq->alloc_size);
587 return NULL;
588 }
589
590 if (mpc->rxbpre_headroom != rxq->headroom) {
591 netdev_err(ndev, "rxbpre_headroom mismatch: %u: %u\n",
592 mpc->rxbpre_headroom, rxq->headroom);
593 return NULL;
594 }
595
596 mpc->rxbpre_total--;
597
598 *da = mpc->das_pre[mpc->rxbpre_total];
599 va = mpc->rxbufs_pre[mpc->rxbpre_total];
600 mpc->rxbufs_pre[mpc->rxbpre_total] = NULL;
601
602 /* Deallocate the array after all buffers are gone */
603 if (!mpc->rxbpre_total)
604 mana_pre_dealloc_rxbufs(mpc);
605
606 return va;
607}
608
609/* Get RX buffer's data size, alloc size, XDP headroom based on MTU */
610static void mana_get_rxbuf_cfg(int mtu, u32 *datasize, u32 *alloc_size,
611 u32 *headroom)
612{
613 if (mtu > MANA_XDP_MTU_MAX)
614 *headroom = 0; /* no support for XDP */
615 else
616 *headroom = XDP_PACKET_HEADROOM;
617
618 *alloc_size = SKB_DATA_ALIGN(mtu + MANA_RXBUF_PAD + *headroom);
619
620 /* Using page pool in this case, so alloc_size is PAGE_SIZE */
621 if (*alloc_size < PAGE_SIZE)
622 *alloc_size = PAGE_SIZE;
623
624 *datasize = mtu + ETH_HLEN;
625}
626
627int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu, int num_queues)
628{
629 struct device *dev;
630 struct page *page;
631 dma_addr_t da;
632 int num_rxb;
633 void *va;
634 int i;
635
636 mana_get_rxbuf_cfg(new_mtu, &mpc->rxbpre_datasize,
637 &mpc->rxbpre_alloc_size, &mpc->rxbpre_headroom);
638
639 dev = mpc->ac->gdma_dev->gdma_context->dev;
640
641 num_rxb = num_queues * mpc->rx_queue_size;
642
643 WARN(mpc->rxbufs_pre, "mana rxbufs_pre exists\n");
644 mpc->rxbufs_pre = kmalloc_array(num_rxb, sizeof(void *), GFP_KERNEL);
645 if (!mpc->rxbufs_pre)
646 goto error;
647
648 mpc->das_pre = kmalloc_array(num_rxb, sizeof(dma_addr_t), GFP_KERNEL);
649 if (!mpc->das_pre)
650 goto error;
651
652 mpc->rxbpre_total = 0;
653
654 for (i = 0; i < num_rxb; i++) {
655 if (mpc->rxbpre_alloc_size > PAGE_SIZE) {
656 va = netdev_alloc_frag(mpc->rxbpre_alloc_size);
657 if (!va)
658 goto error;
659
660 page = virt_to_head_page(va);
661 /* Check if the frag falls back to single page */
662 if (compound_order(page) <
663 get_order(mpc->rxbpre_alloc_size)) {
664 put_page(page);
665 goto error;
666 }
667 } else {
668 page = dev_alloc_page();
669 if (!page)
670 goto error;
671
672 va = page_to_virt(page);
673 }
674
675 da = dma_map_single(dev, va + mpc->rxbpre_headroom,
676 mpc->rxbpre_datasize, DMA_FROM_DEVICE);
677 if (dma_mapping_error(dev, da)) {
678 put_page(virt_to_head_page(va));
679 goto error;
680 }
681
682 mpc->rxbufs_pre[i] = va;
683 mpc->das_pre[i] = da;
684 mpc->rxbpre_total = i + 1;
685 }
686
687 return 0;
688
689error:
690 mana_pre_dealloc_rxbufs(mpc);
691 return -ENOMEM;
692}
693
694static int mana_change_mtu(struct net_device *ndev, int new_mtu)
695{
696 struct mana_port_context *mpc = netdev_priv(ndev);
697 unsigned int old_mtu = ndev->mtu;
698 int err;
699
700 /* Pre-allocate buffers to prevent failure in mana_attach later */
701 err = mana_pre_alloc_rxbufs(mpc, new_mtu, mpc->num_queues);
702 if (err) {
703 netdev_err(ndev, "Insufficient memory for new MTU\n");
704 return err;
705 }
706
707 err = mana_detach(ndev, false);
708 if (err) {
709 netdev_err(ndev, "mana_detach failed: %d\n", err);
710 goto out;
711 }
712
713 WRITE_ONCE(ndev->mtu, new_mtu);
714
715 err = mana_attach(ndev);
716 if (err) {
717 netdev_err(ndev, "mana_attach failed: %d\n", err);
718 WRITE_ONCE(ndev->mtu, old_mtu);
719 }
720
721out:
722 mana_pre_dealloc_rxbufs(mpc);
723 return err;
724}
725
726static const struct net_device_ops mana_devops = {
727 .ndo_open = mana_open,
728 .ndo_stop = mana_close,
729 .ndo_select_queue = mana_select_queue,
730 .ndo_start_xmit = mana_start_xmit,
731 .ndo_validate_addr = eth_validate_addr,
732 .ndo_get_stats64 = mana_get_stats64,
733 .ndo_bpf = mana_bpf,
734 .ndo_xdp_xmit = mana_xdp_xmit,
735 .ndo_change_mtu = mana_change_mtu,
736};
737
738static void mana_cleanup_port_context(struct mana_port_context *apc)
739{
740 /*
741 * at this point all dir/files under the vport directory
742 * are already cleaned up.
743 * We are sure the apc->mana_port_debugfs remove will not
744 * cause any freed memory access issues
745 */
746 debugfs_remove(apc->mana_port_debugfs);
747 kfree(apc->rxqs);
748 apc->rxqs = NULL;
749}
750
751static void mana_cleanup_indir_table(struct mana_port_context *apc)
752{
753 apc->indir_table_sz = 0;
754 kfree(apc->indir_table);
755 kfree(apc->rxobj_table);
756}
757
758static int mana_init_port_context(struct mana_port_context *apc)
759{
760 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *),
761 GFP_KERNEL);
762
763 return !apc->rxqs ? -ENOMEM : 0;
764}
765
766static int mana_send_request(struct mana_context *ac, void *in_buf,
767 u32 in_len, void *out_buf, u32 out_len)
768{
769 struct gdma_context *gc = ac->gdma_dev->gdma_context;
770 struct gdma_resp_hdr *resp = out_buf;
771 struct gdma_req_hdr *req = in_buf;
772 struct device *dev = gc->dev;
773 static atomic_t activity_id;
774 int err;
775
776 req->dev_id = gc->mana.dev_id;
777 req->activity_id = atomic_inc_return(&activity_id);
778
779 err = mana_gd_send_request(gc, in_len, in_buf, out_len,
780 out_buf);
781 if (err || resp->status) {
782 dev_err(dev, "Failed to send mana message: %d, 0x%x\n",
783 err, resp->status);
784 return err ? err : -EPROTO;
785 }
786
787 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 ||
788 req->activity_id != resp->activity_id) {
789 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n",
790 req->dev_id.as_uint32, resp->dev_id.as_uint32,
791 req->activity_id, resp->activity_id);
792 return -EPROTO;
793 }
794
795 return 0;
796}
797
798static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr,
799 const enum mana_command_code expected_code,
800 const u32 min_size)
801{
802 if (resp_hdr->response.msg_type != expected_code)
803 return -EPROTO;
804
805 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1)
806 return -EPROTO;
807
808 if (resp_hdr->response.msg_size < min_size)
809 return -EPROTO;
810
811 return 0;
812}
813
814static int mana_pf_register_hw_vport(struct mana_port_context *apc)
815{
816 struct mana_register_hw_vport_resp resp = {};
817 struct mana_register_hw_vport_req req = {};
818 int err;
819
820 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT,
821 sizeof(req), sizeof(resp));
822 req.attached_gfid = 1;
823 req.is_pf_default_vport = 1;
824 req.allow_all_ether_types = 1;
825
826 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
827 sizeof(resp));
828 if (err) {
829 netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err);
830 return err;
831 }
832
833 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT,
834 sizeof(resp));
835 if (err || resp.hdr.status) {
836 netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n",
837 err, resp.hdr.status);
838 return err ? err : -EPROTO;
839 }
840
841 apc->port_handle = resp.hw_vport_handle;
842 return 0;
843}
844
845static void mana_pf_deregister_hw_vport(struct mana_port_context *apc)
846{
847 struct mana_deregister_hw_vport_resp resp = {};
848 struct mana_deregister_hw_vport_req req = {};
849 int err;
850
851 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT,
852 sizeof(req), sizeof(resp));
853 req.hw_vport_handle = apc->port_handle;
854
855 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
856 sizeof(resp));
857 if (err) {
858 netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n",
859 err);
860 return;
861 }
862
863 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT,
864 sizeof(resp));
865 if (err || resp.hdr.status)
866 netdev_err(apc->ndev,
867 "Failed to deregister hw vPort: %d, 0x%x\n",
868 err, resp.hdr.status);
869}
870
871static int mana_pf_register_filter(struct mana_port_context *apc)
872{
873 struct mana_register_filter_resp resp = {};
874 struct mana_register_filter_req req = {};
875 int err;
876
877 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER,
878 sizeof(req), sizeof(resp));
879 req.vport = apc->port_handle;
880 memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN);
881
882 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
883 sizeof(resp));
884 if (err) {
885 netdev_err(apc->ndev, "Failed to register filter: %d\n", err);
886 return err;
887 }
888
889 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER,
890 sizeof(resp));
891 if (err || resp.hdr.status) {
892 netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n",
893 err, resp.hdr.status);
894 return err ? err : -EPROTO;
895 }
896
897 apc->pf_filter_handle = resp.filter_handle;
898 return 0;
899}
900
901static void mana_pf_deregister_filter(struct mana_port_context *apc)
902{
903 struct mana_deregister_filter_resp resp = {};
904 struct mana_deregister_filter_req req = {};
905 int err;
906
907 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER,
908 sizeof(req), sizeof(resp));
909 req.filter_handle = apc->pf_filter_handle;
910
911 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
912 sizeof(resp));
913 if (err) {
914 netdev_err(apc->ndev, "Failed to unregister filter: %d\n",
915 err);
916 return;
917 }
918
919 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER,
920 sizeof(resp));
921 if (err || resp.hdr.status)
922 netdev_err(apc->ndev,
923 "Failed to deregister filter: %d, 0x%x\n",
924 err, resp.hdr.status);
925}
926
927static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver,
928 u32 proto_minor_ver, u32 proto_micro_ver,
929 u16 *max_num_vports)
930{
931 struct gdma_context *gc = ac->gdma_dev->gdma_context;
932 struct mana_query_device_cfg_resp resp = {};
933 struct mana_query_device_cfg_req req = {};
934 struct device *dev = gc->dev;
935 int err = 0;
936
937 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG,
938 sizeof(req), sizeof(resp));
939
940 req.hdr.resp.msg_version = GDMA_MESSAGE_V2;
941
942 req.proto_major_ver = proto_major_ver;
943 req.proto_minor_ver = proto_minor_ver;
944 req.proto_micro_ver = proto_micro_ver;
945
946 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp));
947 if (err) {
948 dev_err(dev, "Failed to query config: %d", err);
949 return err;
950 }
951
952 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG,
953 sizeof(resp));
954 if (err || resp.hdr.status) {
955 dev_err(dev, "Invalid query result: %d, 0x%x\n", err,
956 resp.hdr.status);
957 if (!err)
958 err = -EPROTO;
959 return err;
960 }
961
962 *max_num_vports = resp.max_num_vports;
963
964 if (resp.hdr.response.msg_version == GDMA_MESSAGE_V2)
965 gc->adapter_mtu = resp.adapter_mtu;
966 else
967 gc->adapter_mtu = ETH_FRAME_LEN;
968
969 debugfs_create_u16("adapter-MTU", 0400, gc->mana_pci_debugfs, &gc->adapter_mtu);
970
971 return 0;
972}
973
974static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index,
975 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry)
976{
977 struct mana_query_vport_cfg_resp resp = {};
978 struct mana_query_vport_cfg_req req = {};
979 int err;
980
981 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG,
982 sizeof(req), sizeof(resp));
983
984 req.vport_index = vport_index;
985
986 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
987 sizeof(resp));
988 if (err)
989 return err;
990
991 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG,
992 sizeof(resp));
993 if (err)
994 return err;
995
996 if (resp.hdr.status)
997 return -EPROTO;
998
999 *max_sq = resp.max_num_sq;
1000 *max_rq = resp.max_num_rq;
1001 if (resp.num_indirection_ent > 0 &&
1002 resp.num_indirection_ent <= MANA_INDIRECT_TABLE_MAX_SIZE &&
1003 is_power_of_2(resp.num_indirection_ent)) {
1004 *num_indir_entry = resp.num_indirection_ent;
1005 } else {
1006 netdev_warn(apc->ndev,
1007 "Setting indirection table size to default %d for vPort %d\n",
1008 MANA_INDIRECT_TABLE_DEF_SIZE, apc->port_idx);
1009 *num_indir_entry = MANA_INDIRECT_TABLE_DEF_SIZE;
1010 }
1011
1012 apc->port_handle = resp.vport;
1013 ether_addr_copy(apc->mac_addr, resp.mac_addr);
1014
1015 return 0;
1016}
1017
1018void mana_uncfg_vport(struct mana_port_context *apc)
1019{
1020 mutex_lock(&apc->vport_mutex);
1021 apc->vport_use_count--;
1022 WARN_ON(apc->vport_use_count < 0);
1023 mutex_unlock(&apc->vport_mutex);
1024}
1025EXPORT_SYMBOL_NS(mana_uncfg_vport, "NET_MANA");
1026
1027int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id,
1028 u32 doorbell_pg_id)
1029{
1030 struct mana_config_vport_resp resp = {};
1031 struct mana_config_vport_req req = {};
1032 int err;
1033
1034 /* This function is used to program the Ethernet port in the hardware
1035 * table. It can be called from the Ethernet driver or the RDMA driver.
1036 *
1037 * For Ethernet usage, the hardware supports only one active user on a
1038 * physical port. The driver checks on the port usage before programming
1039 * the hardware when creating the RAW QP (RDMA driver) or exposing the
1040 * device to kernel NET layer (Ethernet driver).
1041 *
1042 * Because the RDMA driver doesn't know in advance which QP type the
1043 * user will create, it exposes the device with all its ports. The user
1044 * may not be able to create RAW QP on a port if this port is already
1045 * in used by the Ethernet driver from the kernel.
1046 *
1047 * This physical port limitation only applies to the RAW QP. For RC QP,
1048 * the hardware doesn't have this limitation. The user can create RC
1049 * QPs on a physical port up to the hardware limits independent of the
1050 * Ethernet usage on the same port.
1051 */
1052 mutex_lock(&apc->vport_mutex);
1053 if (apc->vport_use_count > 0) {
1054 mutex_unlock(&apc->vport_mutex);
1055 return -EBUSY;
1056 }
1057 apc->vport_use_count++;
1058 mutex_unlock(&apc->vport_mutex);
1059
1060 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX,
1061 sizeof(req), sizeof(resp));
1062 req.vport = apc->port_handle;
1063 req.pdid = protection_dom_id;
1064 req.doorbell_pageid = doorbell_pg_id;
1065
1066 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1067 sizeof(resp));
1068 if (err) {
1069 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err);
1070 goto out;
1071 }
1072
1073 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX,
1074 sizeof(resp));
1075 if (err || resp.hdr.status) {
1076 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n",
1077 err, resp.hdr.status);
1078 if (!err)
1079 err = -EPROTO;
1080
1081 goto out;
1082 }
1083
1084 apc->tx_shortform_allowed = resp.short_form_allowed;
1085 apc->tx_vp_offset = resp.tx_vport_offset;
1086
1087 netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n",
1088 apc->port_handle, protection_dom_id, doorbell_pg_id);
1089out:
1090 if (err)
1091 mana_uncfg_vport(apc);
1092
1093 return err;
1094}
1095EXPORT_SYMBOL_NS(mana_cfg_vport, "NET_MANA");
1096
1097static int mana_cfg_vport_steering(struct mana_port_context *apc,
1098 enum TRI_STATE rx,
1099 bool update_default_rxobj, bool update_key,
1100 bool update_tab)
1101{
1102 struct mana_cfg_rx_steer_req_v2 *req;
1103 struct mana_cfg_rx_steer_resp resp = {};
1104 struct net_device *ndev = apc->ndev;
1105 u32 req_buf_size;
1106 int err;
1107
1108 req_buf_size = struct_size(req, indir_tab, apc->indir_table_sz);
1109 req = kzalloc(req_buf_size, GFP_KERNEL);
1110 if (!req)
1111 return -ENOMEM;
1112
1113 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
1114 sizeof(resp));
1115
1116 req->hdr.req.msg_version = GDMA_MESSAGE_V2;
1117
1118 req->vport = apc->port_handle;
1119 req->num_indir_entries = apc->indir_table_sz;
1120 req->indir_tab_offset = offsetof(struct mana_cfg_rx_steer_req_v2,
1121 indir_tab);
1122 req->rx_enable = rx;
1123 req->rss_enable = apc->rss_state;
1124 req->update_default_rxobj = update_default_rxobj;
1125 req->update_hashkey = update_key;
1126 req->update_indir_tab = update_tab;
1127 req->default_rxobj = apc->default_rxobj;
1128 req->cqe_coalescing_enable = 0;
1129
1130 if (update_key)
1131 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE);
1132
1133 if (update_tab)
1134 memcpy(req->indir_tab, apc->rxobj_table,
1135 flex_array_size(req, indir_tab, req->num_indir_entries));
1136
1137 err = mana_send_request(apc->ac, req, req_buf_size, &resp,
1138 sizeof(resp));
1139 if (err) {
1140 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err);
1141 goto out;
1142 }
1143
1144 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX,
1145 sizeof(resp));
1146 if (err) {
1147 netdev_err(ndev, "vPort RX configuration failed: %d\n", err);
1148 goto out;
1149 }
1150
1151 if (resp.hdr.status) {
1152 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n",
1153 resp.hdr.status);
1154 err = -EPROTO;
1155 }
1156
1157 netdev_info(ndev, "Configured steering vPort %llu entries %u\n",
1158 apc->port_handle, apc->indir_table_sz);
1159out:
1160 kfree(req);
1161 return err;
1162}
1163
1164int mana_create_wq_obj(struct mana_port_context *apc,
1165 mana_handle_t vport,
1166 u32 wq_type, struct mana_obj_spec *wq_spec,
1167 struct mana_obj_spec *cq_spec,
1168 mana_handle_t *wq_obj)
1169{
1170 struct mana_create_wqobj_resp resp = {};
1171 struct mana_create_wqobj_req req = {};
1172 struct net_device *ndev = apc->ndev;
1173 int err;
1174
1175 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ,
1176 sizeof(req), sizeof(resp));
1177 req.vport = vport;
1178 req.wq_type = wq_type;
1179 req.wq_gdma_region = wq_spec->gdma_region;
1180 req.cq_gdma_region = cq_spec->gdma_region;
1181 req.wq_size = wq_spec->queue_size;
1182 req.cq_size = cq_spec->queue_size;
1183 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id;
1184 req.cq_parent_qid = cq_spec->attached_eq;
1185
1186 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1187 sizeof(resp));
1188 if (err) {
1189 netdev_err(ndev, "Failed to create WQ object: %d\n", err);
1190 goto out;
1191 }
1192
1193 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ,
1194 sizeof(resp));
1195 if (err || resp.hdr.status) {
1196 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err,
1197 resp.hdr.status);
1198 if (!err)
1199 err = -EPROTO;
1200 goto out;
1201 }
1202
1203 if (resp.wq_obj == INVALID_MANA_HANDLE) {
1204 netdev_err(ndev, "Got an invalid WQ object handle\n");
1205 err = -EPROTO;
1206 goto out;
1207 }
1208
1209 *wq_obj = resp.wq_obj;
1210 wq_spec->queue_index = resp.wq_id;
1211 cq_spec->queue_index = resp.cq_id;
1212
1213 return 0;
1214out:
1215 return err;
1216}
1217EXPORT_SYMBOL_NS(mana_create_wq_obj, "NET_MANA");
1218
1219void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type,
1220 mana_handle_t wq_obj)
1221{
1222 struct mana_destroy_wqobj_resp resp = {};
1223 struct mana_destroy_wqobj_req req = {};
1224 struct net_device *ndev = apc->ndev;
1225 int err;
1226
1227 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ,
1228 sizeof(req), sizeof(resp));
1229 req.wq_type = wq_type;
1230 req.wq_obj_handle = wq_obj;
1231
1232 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1233 sizeof(resp));
1234 if (err) {
1235 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err);
1236 return;
1237 }
1238
1239 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ,
1240 sizeof(resp));
1241 if (err || resp.hdr.status)
1242 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err,
1243 resp.hdr.status);
1244}
1245EXPORT_SYMBOL_NS(mana_destroy_wq_obj, "NET_MANA");
1246
1247static void mana_destroy_eq(struct mana_context *ac)
1248{
1249 struct gdma_context *gc = ac->gdma_dev->gdma_context;
1250 struct gdma_queue *eq;
1251 int i;
1252
1253 if (!ac->eqs)
1254 return;
1255
1256 debugfs_remove_recursive(ac->mana_eqs_debugfs);
1257
1258 for (i = 0; i < gc->max_num_queues; i++) {
1259 eq = ac->eqs[i].eq;
1260 if (!eq)
1261 continue;
1262
1263 mana_gd_destroy_queue(gc, eq);
1264 }
1265
1266 kfree(ac->eqs);
1267 ac->eqs = NULL;
1268}
1269
1270static void mana_create_eq_debugfs(struct mana_context *ac, int i)
1271{
1272 struct mana_eq eq = ac->eqs[i];
1273 char eqnum[32];
1274
1275 sprintf(eqnum, "eq%d", i);
1276 eq.mana_eq_debugfs = debugfs_create_dir(eqnum, ac->mana_eqs_debugfs);
1277 debugfs_create_u32("head", 0400, eq.mana_eq_debugfs, &eq.eq->head);
1278 debugfs_create_u32("tail", 0400, eq.mana_eq_debugfs, &eq.eq->tail);
1279 debugfs_create_file("eq_dump", 0400, eq.mana_eq_debugfs, eq.eq, &mana_dbg_q_fops);
1280}
1281
1282static int mana_create_eq(struct mana_context *ac)
1283{
1284 struct gdma_dev *gd = ac->gdma_dev;
1285 struct gdma_context *gc = gd->gdma_context;
1286 struct gdma_queue_spec spec = {};
1287 int err;
1288 int i;
1289
1290 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq),
1291 GFP_KERNEL);
1292 if (!ac->eqs)
1293 return -ENOMEM;
1294
1295 spec.type = GDMA_EQ;
1296 spec.monitor_avl_buf = false;
1297 spec.queue_size = EQ_SIZE;
1298 spec.eq.callback = NULL;
1299 spec.eq.context = ac->eqs;
1300 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE;
1301
1302 ac->mana_eqs_debugfs = debugfs_create_dir("EQs", gc->mana_pci_debugfs);
1303
1304 for (i = 0; i < gc->max_num_queues; i++) {
1305 spec.eq.msix_index = (i + 1) % gc->num_msix_usable;
1306 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq);
1307 if (err)
1308 goto out;
1309 mana_create_eq_debugfs(ac, i);
1310 }
1311
1312 return 0;
1313out:
1314 mana_destroy_eq(ac);
1315 return err;
1316}
1317
1318static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq)
1319{
1320 struct mana_fence_rq_resp resp = {};
1321 struct mana_fence_rq_req req = {};
1322 int err;
1323
1324 init_completion(&rxq->fence_event);
1325
1326 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ,
1327 sizeof(req), sizeof(resp));
1328 req.wq_obj_handle = rxq->rxobj;
1329
1330 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1331 sizeof(resp));
1332 if (err) {
1333 netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n",
1334 rxq->rxq_idx, err);
1335 return err;
1336 }
1337
1338 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp));
1339 if (err || resp.hdr.status) {
1340 netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n",
1341 rxq->rxq_idx, err, resp.hdr.status);
1342 if (!err)
1343 err = -EPROTO;
1344
1345 return err;
1346 }
1347
1348 if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) {
1349 netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n",
1350 rxq->rxq_idx);
1351 return -ETIMEDOUT;
1352 }
1353
1354 return 0;
1355}
1356
1357static void mana_fence_rqs(struct mana_port_context *apc)
1358{
1359 unsigned int rxq_idx;
1360 struct mana_rxq *rxq;
1361 int err;
1362
1363 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
1364 rxq = apc->rxqs[rxq_idx];
1365 err = mana_fence_rq(apc, rxq);
1366
1367 /* In case of any error, use sleep instead. */
1368 if (err)
1369 msleep(100);
1370 }
1371}
1372
1373static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units)
1374{
1375 u32 used_space_old;
1376 u32 used_space_new;
1377
1378 used_space_old = wq->head - wq->tail;
1379 used_space_new = wq->head - (wq->tail + num_units);
1380
1381 if (WARN_ON_ONCE(used_space_new > used_space_old))
1382 return -ERANGE;
1383
1384 wq->tail += num_units;
1385 return 0;
1386}
1387
1388static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc)
1389{
1390 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
1391 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1392 struct device *dev = gc->dev;
1393 int hsg, i;
1394
1395 /* Number of SGEs of linear part */
1396 hsg = (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0]) ? 2 : 1;
1397
1398 for (i = 0; i < hsg; i++)
1399 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i],
1400 DMA_TO_DEVICE);
1401
1402 for (i = hsg; i < skb_shinfo(skb)->nr_frags + hsg; i++)
1403 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
1404 DMA_TO_DEVICE);
1405}
1406
1407static void mana_poll_tx_cq(struct mana_cq *cq)
1408{
1409 struct gdma_comp *completions = cq->gdma_comp_buf;
1410 struct gdma_posted_wqe_info *wqe_info;
1411 unsigned int pkt_transmitted = 0;
1412 unsigned int wqe_unit_cnt = 0;
1413 struct mana_txq *txq = cq->txq;
1414 struct mana_port_context *apc;
1415 struct netdev_queue *net_txq;
1416 struct gdma_queue *gdma_wq;
1417 unsigned int avail_space;
1418 struct net_device *ndev;
1419 struct sk_buff *skb;
1420 bool txq_stopped;
1421 int comp_read;
1422 int i;
1423
1424 ndev = txq->ndev;
1425 apc = netdev_priv(ndev);
1426
1427 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions,
1428 CQE_POLLING_BUFFER);
1429
1430 if (comp_read < 1)
1431 return;
1432
1433 for (i = 0; i < comp_read; i++) {
1434 struct mana_tx_comp_oob *cqe_oob;
1435
1436 if (WARN_ON_ONCE(!completions[i].is_sq))
1437 return;
1438
1439 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data;
1440 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type !=
1441 MANA_CQE_COMPLETION))
1442 return;
1443
1444 switch (cqe_oob->cqe_hdr.cqe_type) {
1445 case CQE_TX_OKAY:
1446 break;
1447
1448 case CQE_TX_SA_DROP:
1449 case CQE_TX_MTU_DROP:
1450 case CQE_TX_INVALID_OOB:
1451 case CQE_TX_INVALID_ETH_TYPE:
1452 case CQE_TX_HDR_PROCESSING_ERROR:
1453 case CQE_TX_VF_DISABLED:
1454 case CQE_TX_VPORT_IDX_OUT_OF_RANGE:
1455 case CQE_TX_VPORT_DISABLED:
1456 case CQE_TX_VLAN_TAGGING_VIOLATION:
1457 if (net_ratelimit())
1458 netdev_err(ndev, "TX: CQE error %d\n",
1459 cqe_oob->cqe_hdr.cqe_type);
1460
1461 apc->eth_stats.tx_cqe_err++;
1462 break;
1463
1464 default:
1465 /* If the CQE type is unknown, log an error,
1466 * and still free the SKB, update tail, etc.
1467 */
1468 if (net_ratelimit())
1469 netdev_err(ndev, "TX: unknown CQE type %d\n",
1470 cqe_oob->cqe_hdr.cqe_type);
1471
1472 apc->eth_stats.tx_cqe_unknown_type++;
1473 break;
1474 }
1475
1476 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num))
1477 return;
1478
1479 skb = skb_dequeue(&txq->pending_skbs);
1480 if (WARN_ON_ONCE(!skb))
1481 return;
1482
1483 wqe_info = (struct gdma_posted_wqe_info *)skb->cb;
1484 wqe_unit_cnt += wqe_info->wqe_size_in_bu;
1485
1486 mana_unmap_skb(skb, apc);
1487
1488 napi_consume_skb(skb, cq->budget);
1489
1490 pkt_transmitted++;
1491 }
1492
1493 if (WARN_ON_ONCE(wqe_unit_cnt == 0))
1494 return;
1495
1496 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt);
1497
1498 gdma_wq = txq->gdma_sq;
1499 avail_space = mana_gd_wq_avail_space(gdma_wq);
1500
1501 /* Ensure tail updated before checking q stop */
1502 smp_mb();
1503
1504 net_txq = txq->net_txq;
1505 txq_stopped = netif_tx_queue_stopped(net_txq);
1506
1507 /* Ensure checking txq_stopped before apc->port_is_up. */
1508 smp_rmb();
1509
1510 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) {
1511 netif_tx_wake_queue(net_txq);
1512 apc->eth_stats.wake_queue++;
1513 }
1514
1515 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0)
1516 WARN_ON_ONCE(1);
1517
1518 cq->work_done = pkt_transmitted;
1519}
1520
1521static void mana_post_pkt_rxq(struct mana_rxq *rxq)
1522{
1523 struct mana_recv_buf_oob *recv_buf_oob;
1524 u32 curr_index;
1525 int err;
1526
1527 curr_index = rxq->buf_index++;
1528 if (rxq->buf_index == rxq->num_rx_buf)
1529 rxq->buf_index = 0;
1530
1531 recv_buf_oob = &rxq->rx_oobs[curr_index];
1532
1533 err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req,
1534 &recv_buf_oob->wqe_inf);
1535 if (WARN_ON_ONCE(err))
1536 return;
1537
1538 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1);
1539}
1540
1541static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va,
1542 uint pkt_len, struct xdp_buff *xdp)
1543{
1544 struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size);
1545
1546 if (!skb)
1547 return NULL;
1548
1549 if (xdp->data_hard_start) {
1550 skb_reserve(skb, xdp->data - xdp->data_hard_start);
1551 skb_put(skb, xdp->data_end - xdp->data);
1552 return skb;
1553 }
1554
1555 skb_reserve(skb, rxq->headroom);
1556 skb_put(skb, pkt_len);
1557
1558 return skb;
1559}
1560
1561static void mana_rx_skb(void *buf_va, bool from_pool,
1562 struct mana_rxcomp_oob *cqe, struct mana_rxq *rxq)
1563{
1564 struct mana_stats_rx *rx_stats = &rxq->stats;
1565 struct net_device *ndev = rxq->ndev;
1566 uint pkt_len = cqe->ppi[0].pkt_len;
1567 u16 rxq_idx = rxq->rxq_idx;
1568 struct napi_struct *napi;
1569 struct xdp_buff xdp = {};
1570 struct sk_buff *skb;
1571 u32 hash_value;
1572 u32 act;
1573
1574 rxq->rx_cq.work_done++;
1575 napi = &rxq->rx_cq.napi;
1576
1577 if (!buf_va) {
1578 ++ndev->stats.rx_dropped;
1579 return;
1580 }
1581
1582 act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len);
1583
1584 if (act == XDP_REDIRECT && !rxq->xdp_rc)
1585 return;
1586
1587 if (act != XDP_PASS && act != XDP_TX)
1588 goto drop_xdp;
1589
1590 skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp);
1591
1592 if (!skb)
1593 goto drop;
1594
1595 if (from_pool)
1596 skb_mark_for_recycle(skb);
1597
1598 skb->dev = napi->dev;
1599
1600 skb->protocol = eth_type_trans(skb, ndev);
1601 skb_checksum_none_assert(skb);
1602 skb_record_rx_queue(skb, rxq_idx);
1603
1604 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) {
1605 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed)
1606 skb->ip_summed = CHECKSUM_UNNECESSARY;
1607 }
1608
1609 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) {
1610 hash_value = cqe->ppi[0].pkt_hash;
1611
1612 if (cqe->rx_hashtype & MANA_HASH_L4)
1613 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4);
1614 else
1615 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3);
1616 }
1617
1618 if (cqe->rx_vlantag_present) {
1619 u16 vlan_tci = cqe->rx_vlan_id;
1620
1621 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci);
1622 }
1623
1624 u64_stats_update_begin(&rx_stats->syncp);
1625 rx_stats->packets++;
1626 rx_stats->bytes += pkt_len;
1627
1628 if (act == XDP_TX)
1629 rx_stats->xdp_tx++;
1630 u64_stats_update_end(&rx_stats->syncp);
1631
1632 if (act == XDP_TX) {
1633 skb_set_queue_mapping(skb, rxq_idx);
1634 mana_xdp_tx(skb, ndev);
1635 return;
1636 }
1637
1638 napi_gro_receive(napi, skb);
1639
1640 return;
1641
1642drop_xdp:
1643 u64_stats_update_begin(&rx_stats->syncp);
1644 rx_stats->xdp_drop++;
1645 u64_stats_update_end(&rx_stats->syncp);
1646
1647drop:
1648 if (from_pool) {
1649 page_pool_recycle_direct(rxq->page_pool,
1650 virt_to_head_page(buf_va));
1651 } else {
1652 WARN_ON_ONCE(rxq->xdp_save_va);
1653 /* Save for reuse */
1654 rxq->xdp_save_va = buf_va;
1655 }
1656
1657 ++ndev->stats.rx_dropped;
1658
1659 return;
1660}
1661
1662static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,
1663 dma_addr_t *da, bool *from_pool, bool is_napi)
1664{
1665 struct page *page;
1666 void *va;
1667
1668 *from_pool = false;
1669
1670 /* Reuse XDP dropped page if available */
1671 if (rxq->xdp_save_va) {
1672 va = rxq->xdp_save_va;
1673 rxq->xdp_save_va = NULL;
1674 } else if (rxq->alloc_size > PAGE_SIZE) {
1675 if (is_napi)
1676 va = napi_alloc_frag(rxq->alloc_size);
1677 else
1678 va = netdev_alloc_frag(rxq->alloc_size);
1679
1680 if (!va)
1681 return NULL;
1682
1683 page = virt_to_head_page(va);
1684 /* Check if the frag falls back to single page */
1685 if (compound_order(page) < get_order(rxq->alloc_size)) {
1686 put_page(page);
1687 return NULL;
1688 }
1689 } else {
1690 page = page_pool_dev_alloc_pages(rxq->page_pool);
1691 if (!page)
1692 return NULL;
1693
1694 *from_pool = true;
1695 va = page_to_virt(page);
1696 }
1697
1698 *da = dma_map_single(dev, va + rxq->headroom, rxq->datasize,
1699 DMA_FROM_DEVICE);
1700 if (dma_mapping_error(dev, *da)) {
1701 if (*from_pool)
1702 page_pool_put_full_page(rxq->page_pool, page, false);
1703 else
1704 put_page(virt_to_head_page(va));
1705
1706 return NULL;
1707 }
1708
1709 return va;
1710}
1711
1712/* Allocate frag for rx buffer, and save the old buf */
1713static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq,
1714 struct mana_recv_buf_oob *rxoob, void **old_buf,
1715 bool *old_fp)
1716{
1717 bool from_pool;
1718 dma_addr_t da;
1719 void *va;
1720
1721 va = mana_get_rxfrag(rxq, dev, &da, &from_pool, true);
1722 if (!va)
1723 return;
1724
1725 dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize,
1726 DMA_FROM_DEVICE);
1727 *old_buf = rxoob->buf_va;
1728 *old_fp = rxoob->from_pool;
1729
1730 rxoob->buf_va = va;
1731 rxoob->sgl[0].address = da;
1732 rxoob->from_pool = from_pool;
1733}
1734
1735static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,
1736 struct gdma_comp *cqe)
1737{
1738 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data;
1739 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1740 struct net_device *ndev = rxq->ndev;
1741 struct mana_recv_buf_oob *rxbuf_oob;
1742 struct mana_port_context *apc;
1743 struct device *dev = gc->dev;
1744 void *old_buf = NULL;
1745 u32 curr, pktlen;
1746 bool old_fp;
1747
1748 apc = netdev_priv(ndev);
1749
1750 switch (oob->cqe_hdr.cqe_type) {
1751 case CQE_RX_OKAY:
1752 break;
1753
1754 case CQE_RX_TRUNCATED:
1755 ++ndev->stats.rx_dropped;
1756 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index];
1757 netdev_warn_once(ndev, "Dropped a truncated packet\n");
1758 goto drop;
1759
1760 case CQE_RX_COALESCED_4:
1761 netdev_err(ndev, "RX coalescing is unsupported\n");
1762 apc->eth_stats.rx_coalesced_err++;
1763 return;
1764
1765 case CQE_RX_OBJECT_FENCE:
1766 complete(&rxq->fence_event);
1767 return;
1768
1769 default:
1770 netdev_err(ndev, "Unknown RX CQE type = %d\n",
1771 oob->cqe_hdr.cqe_type);
1772 apc->eth_stats.rx_cqe_unknown_type++;
1773 return;
1774 }
1775
1776 pktlen = oob->ppi[0].pkt_len;
1777
1778 if (pktlen == 0) {
1779 /* data packets should never have packetlength of zero */
1780 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n",
1781 rxq->gdma_id, cq->gdma_id, rxq->rxobj);
1782 return;
1783 }
1784
1785 curr = rxq->buf_index;
1786 rxbuf_oob = &rxq->rx_oobs[curr];
1787 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1);
1788
1789 mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf, &old_fp);
1790
1791 /* Unsuccessful refill will have old_buf == NULL.
1792 * In this case, mana_rx_skb() will drop the packet.
1793 */
1794 mana_rx_skb(old_buf, old_fp, oob, rxq);
1795
1796drop:
1797 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu);
1798
1799 mana_post_pkt_rxq(rxq);
1800}
1801
1802static void mana_poll_rx_cq(struct mana_cq *cq)
1803{
1804 struct gdma_comp *comp = cq->gdma_comp_buf;
1805 struct mana_rxq *rxq = cq->rxq;
1806 int comp_read, i;
1807
1808 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER);
1809 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER);
1810
1811 rxq->xdp_flush = false;
1812
1813 for (i = 0; i < comp_read; i++) {
1814 if (WARN_ON_ONCE(comp[i].is_sq))
1815 return;
1816
1817 /* verify recv cqe references the right rxq */
1818 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id))
1819 return;
1820
1821 mana_process_rx_cqe(rxq, cq, &comp[i]);
1822 }
1823
1824 if (comp_read > 0) {
1825 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1826
1827 mana_gd_wq_ring_doorbell(gc, rxq->gdma_rq);
1828 }
1829
1830 if (rxq->xdp_flush)
1831 xdp_do_flush();
1832}
1833
1834static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue)
1835{
1836 struct mana_cq *cq = context;
1837 int w;
1838
1839 WARN_ON_ONCE(cq->gdma_cq != gdma_queue);
1840
1841 if (cq->type == MANA_CQ_TYPE_RX)
1842 mana_poll_rx_cq(cq);
1843 else
1844 mana_poll_tx_cq(cq);
1845
1846 w = cq->work_done;
1847 cq->work_done_since_doorbell += w;
1848
1849 if (w < cq->budget) {
1850 mana_gd_ring_cq(gdma_queue, SET_ARM_BIT);
1851 cq->work_done_since_doorbell = 0;
1852 napi_complete_done(&cq->napi, w);
1853 } else if (cq->work_done_since_doorbell >
1854 cq->gdma_cq->queue_size / COMP_ENTRY_SIZE * 4) {
1855 /* MANA hardware requires at least one doorbell ring every 8
1856 * wraparounds of CQ even if there is no need to arm the CQ.
1857 * This driver rings the doorbell as soon as we have exceeded
1858 * 4 wraparounds.
1859 */
1860 mana_gd_ring_cq(gdma_queue, 0);
1861 cq->work_done_since_doorbell = 0;
1862 }
1863
1864 return w;
1865}
1866
1867static int mana_poll(struct napi_struct *napi, int budget)
1868{
1869 struct mana_cq *cq = container_of(napi, struct mana_cq, napi);
1870 int w;
1871
1872 cq->work_done = 0;
1873 cq->budget = budget;
1874
1875 w = mana_cq_handler(cq, cq->gdma_cq);
1876
1877 return min(w, budget);
1878}
1879
1880static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue)
1881{
1882 struct mana_cq *cq = context;
1883
1884 napi_schedule_irqoff(&cq->napi);
1885}
1886
1887static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq)
1888{
1889 struct gdma_dev *gd = apc->ac->gdma_dev;
1890
1891 if (!cq->gdma_cq)
1892 return;
1893
1894 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq);
1895}
1896
1897static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq)
1898{
1899 struct gdma_dev *gd = apc->ac->gdma_dev;
1900
1901 if (!txq->gdma_sq)
1902 return;
1903
1904 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq);
1905}
1906
1907static void mana_destroy_txq(struct mana_port_context *apc)
1908{
1909 struct napi_struct *napi;
1910 int i;
1911
1912 if (!apc->tx_qp)
1913 return;
1914
1915 for (i = 0; i < apc->num_queues; i++) {
1916 debugfs_remove_recursive(apc->tx_qp[i].mana_tx_debugfs);
1917
1918 napi = &apc->tx_qp[i].tx_cq.napi;
1919 if (apc->tx_qp[i].txq.napi_initialized) {
1920 napi_synchronize(napi);
1921 napi_disable(napi);
1922 netif_napi_del(napi);
1923 apc->tx_qp[i].txq.napi_initialized = false;
1924 }
1925 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object);
1926
1927 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq);
1928
1929 mana_deinit_txq(apc, &apc->tx_qp[i].txq);
1930 }
1931
1932 kfree(apc->tx_qp);
1933 apc->tx_qp = NULL;
1934}
1935
1936static void mana_create_txq_debugfs(struct mana_port_context *apc, int idx)
1937{
1938 struct mana_tx_qp *tx_qp = &apc->tx_qp[idx];
1939 char qnum[32];
1940
1941 sprintf(qnum, "TX-%d", idx);
1942 tx_qp->mana_tx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs);
1943 debugfs_create_u32("sq_head", 0400, tx_qp->mana_tx_debugfs,
1944 &tx_qp->txq.gdma_sq->head);
1945 debugfs_create_u32("sq_tail", 0400, tx_qp->mana_tx_debugfs,
1946 &tx_qp->txq.gdma_sq->tail);
1947 debugfs_create_u32("sq_pend_skb_qlen", 0400, tx_qp->mana_tx_debugfs,
1948 &tx_qp->txq.pending_skbs.qlen);
1949 debugfs_create_u32("cq_head", 0400, tx_qp->mana_tx_debugfs,
1950 &tx_qp->tx_cq.gdma_cq->head);
1951 debugfs_create_u32("cq_tail", 0400, tx_qp->mana_tx_debugfs,
1952 &tx_qp->tx_cq.gdma_cq->tail);
1953 debugfs_create_u32("cq_budget", 0400, tx_qp->mana_tx_debugfs,
1954 &tx_qp->tx_cq.budget);
1955 debugfs_create_file("txq_dump", 0400, tx_qp->mana_tx_debugfs,
1956 tx_qp->txq.gdma_sq, &mana_dbg_q_fops);
1957 debugfs_create_file("cq_dump", 0400, tx_qp->mana_tx_debugfs,
1958 tx_qp->tx_cq.gdma_cq, &mana_dbg_q_fops);
1959}
1960
1961static int mana_create_txq(struct mana_port_context *apc,
1962 struct net_device *net)
1963{
1964 struct mana_context *ac = apc->ac;
1965 struct gdma_dev *gd = ac->gdma_dev;
1966 struct mana_obj_spec wq_spec;
1967 struct mana_obj_spec cq_spec;
1968 struct gdma_queue_spec spec;
1969 struct gdma_context *gc;
1970 struct mana_txq *txq;
1971 struct mana_cq *cq;
1972 u32 txq_size;
1973 u32 cq_size;
1974 int err;
1975 int i;
1976
1977 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp),
1978 GFP_KERNEL);
1979 if (!apc->tx_qp)
1980 return -ENOMEM;
1981
1982 /* The minimum size of the WQE is 32 bytes, hence
1983 * apc->tx_queue_size represents the maximum number of WQEs
1984 * the SQ can store. This value is then used to size other queues
1985 * to prevent overflow.
1986 * Also note that the txq_size is always going to be MANA_PAGE_ALIGNED,
1987 * as min val of apc->tx_queue_size is 128 and that would make
1988 * txq_size 128*32 = 4096 and the other higher values of apc->tx_queue_size
1989 * are always power of two
1990 */
1991 txq_size = apc->tx_queue_size * 32;
1992
1993 cq_size = apc->tx_queue_size * COMP_ENTRY_SIZE;
1994
1995 gc = gd->gdma_context;
1996
1997 for (i = 0; i < apc->num_queues; i++) {
1998 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE;
1999
2000 /* Create SQ */
2001 txq = &apc->tx_qp[i].txq;
2002
2003 u64_stats_init(&txq->stats.syncp);
2004 txq->ndev = net;
2005 txq->net_txq = netdev_get_tx_queue(net, i);
2006 txq->vp_offset = apc->tx_vp_offset;
2007 txq->napi_initialized = false;
2008 skb_queue_head_init(&txq->pending_skbs);
2009
2010 memset(&spec, 0, sizeof(spec));
2011 spec.type = GDMA_SQ;
2012 spec.monitor_avl_buf = true;
2013 spec.queue_size = txq_size;
2014 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq);
2015 if (err)
2016 goto out;
2017
2018 /* Create SQ's CQ */
2019 cq = &apc->tx_qp[i].tx_cq;
2020 cq->type = MANA_CQ_TYPE_TX;
2021
2022 cq->txq = txq;
2023
2024 memset(&spec, 0, sizeof(spec));
2025 spec.type = GDMA_CQ;
2026 spec.monitor_avl_buf = false;
2027 spec.queue_size = cq_size;
2028 spec.cq.callback = mana_schedule_napi;
2029 spec.cq.parent_eq = ac->eqs[i].eq;
2030 spec.cq.context = cq;
2031 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2032 if (err)
2033 goto out;
2034
2035 memset(&wq_spec, 0, sizeof(wq_spec));
2036 memset(&cq_spec, 0, sizeof(cq_spec));
2037
2038 wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle;
2039 wq_spec.queue_size = txq->gdma_sq->queue_size;
2040
2041 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2042 cq_spec.queue_size = cq->gdma_cq->queue_size;
2043 cq_spec.modr_ctx_id = 0;
2044 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2045
2046 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ,
2047 &wq_spec, &cq_spec,
2048 &apc->tx_qp[i].tx_object);
2049
2050 if (err)
2051 goto out;
2052
2053 txq->gdma_sq->id = wq_spec.queue_index;
2054 cq->gdma_cq->id = cq_spec.queue_index;
2055
2056 txq->gdma_sq->mem_info.dma_region_handle =
2057 GDMA_INVALID_DMA_REGION;
2058 cq->gdma_cq->mem_info.dma_region_handle =
2059 GDMA_INVALID_DMA_REGION;
2060
2061 txq->gdma_txq_id = txq->gdma_sq->id;
2062
2063 cq->gdma_id = cq->gdma_cq->id;
2064
2065 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
2066 err = -EINVAL;
2067 goto out;
2068 }
2069
2070 gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2071
2072 mana_create_txq_debugfs(apc, i);
2073
2074 netif_napi_add_tx(net, &cq->napi, mana_poll);
2075 napi_enable(&cq->napi);
2076 txq->napi_initialized = true;
2077
2078 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2079 }
2080
2081 return 0;
2082out:
2083 mana_destroy_txq(apc);
2084 return err;
2085}
2086
2087static void mana_destroy_rxq(struct mana_port_context *apc,
2088 struct mana_rxq *rxq, bool napi_initialized)
2089
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 struct napi_struct *napi;
2095 struct page *page;
2096 int i;
2097
2098 if (!rxq)
2099 return;
2100
2101 debugfs_remove_recursive(rxq->mana_rx_debugfs);
2102
2103 napi = &rxq->rx_cq.napi;
2104
2105 if (napi_initialized) {
2106 napi_synchronize(napi);
2107
2108 napi_disable(napi);
2109
2110 netif_napi_del(napi);
2111 }
2112 xdp_rxq_info_unreg(&rxq->xdp_rxq);
2113
2114 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj);
2115
2116 mana_deinit_cq(apc, &rxq->rx_cq);
2117
2118 if (rxq->xdp_save_va)
2119 put_page(virt_to_head_page(rxq->xdp_save_va));
2120
2121 for (i = 0; i < rxq->num_rx_buf; i++) {
2122 rx_oob = &rxq->rx_oobs[i];
2123
2124 if (!rx_oob->buf_va)
2125 continue;
2126
2127 dma_unmap_single(dev, rx_oob->sgl[0].address,
2128 rx_oob->sgl[0].size, DMA_FROM_DEVICE);
2129
2130 page = virt_to_head_page(rx_oob->buf_va);
2131
2132 if (rx_oob->from_pool)
2133 page_pool_put_full_page(rxq->page_pool, page, false);
2134 else
2135 put_page(page);
2136
2137 rx_oob->buf_va = NULL;
2138 }
2139
2140 page_pool_destroy(rxq->page_pool);
2141
2142 if (rxq->gdma_rq)
2143 mana_gd_destroy_queue(gc, rxq->gdma_rq);
2144
2145 kfree(rxq);
2146}
2147
2148static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key,
2149 struct mana_rxq *rxq, struct device *dev)
2150{
2151 struct mana_port_context *mpc = netdev_priv(rxq->ndev);
2152 bool from_pool = false;
2153 dma_addr_t da;
2154 void *va;
2155
2156 if (mpc->rxbufs_pre)
2157 va = mana_get_rxbuf_pre(rxq, &da);
2158 else
2159 va = mana_get_rxfrag(rxq, dev, &da, &from_pool, false);
2160
2161 if (!va)
2162 return -ENOMEM;
2163
2164 rx_oob->buf_va = va;
2165 rx_oob->from_pool = from_pool;
2166
2167 rx_oob->sgl[0].address = da;
2168 rx_oob->sgl[0].size = rxq->datasize;
2169 rx_oob->sgl[0].mem_key = mem_key;
2170
2171 return 0;
2172}
2173
2174#define MANA_WQE_HEADER_SIZE 16
2175#define MANA_WQE_SGE_SIZE 16
2176
2177static int mana_alloc_rx_wqe(struct mana_port_context *apc,
2178 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size)
2179{
2180 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
2181 struct mana_recv_buf_oob *rx_oob;
2182 struct device *dev = gc->dev;
2183 u32 buf_idx;
2184 int ret;
2185
2186 WARN_ON(rxq->datasize == 0);
2187
2188 *rxq_size = 0;
2189 *cq_size = 0;
2190
2191 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2192 rx_oob = &rxq->rx_oobs[buf_idx];
2193 memset(rx_oob, 0, sizeof(*rx_oob));
2194
2195 rx_oob->num_sge = 1;
2196
2197 ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq,
2198 dev);
2199 if (ret)
2200 return ret;
2201
2202 rx_oob->wqe_req.sgl = rx_oob->sgl;
2203 rx_oob->wqe_req.num_sge = rx_oob->num_sge;
2204 rx_oob->wqe_req.inline_oob_size = 0;
2205 rx_oob->wqe_req.inline_oob_data = NULL;
2206 rx_oob->wqe_req.flags = 0;
2207 rx_oob->wqe_req.client_data_unit = 0;
2208
2209 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE +
2210 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32);
2211 *cq_size += COMP_ENTRY_SIZE;
2212 }
2213
2214 return 0;
2215}
2216
2217static int mana_push_wqe(struct mana_rxq *rxq)
2218{
2219 struct mana_recv_buf_oob *rx_oob;
2220 u32 buf_idx;
2221 int err;
2222
2223 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2224 rx_oob = &rxq->rx_oobs[buf_idx];
2225
2226 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req,
2227 &rx_oob->wqe_inf);
2228 if (err)
2229 return -ENOSPC;
2230 }
2231
2232 return 0;
2233}
2234
2235static int mana_create_page_pool(struct mana_rxq *rxq, struct gdma_context *gc)
2236{
2237 struct mana_port_context *mpc = netdev_priv(rxq->ndev);
2238 struct page_pool_params pprm = {};
2239 int ret;
2240
2241 pprm.pool_size = mpc->rx_queue_size;
2242 pprm.nid = gc->numa_node;
2243 pprm.napi = &rxq->rx_cq.napi;
2244 pprm.netdev = rxq->ndev;
2245
2246 rxq->page_pool = page_pool_create(&pprm);
2247
2248 if (IS_ERR(rxq->page_pool)) {
2249 ret = PTR_ERR(rxq->page_pool);
2250 rxq->page_pool = NULL;
2251 return ret;
2252 }
2253
2254 return 0;
2255}
2256
2257static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,
2258 u32 rxq_idx, struct mana_eq *eq,
2259 struct net_device *ndev)
2260{
2261 struct gdma_dev *gd = apc->ac->gdma_dev;
2262 struct mana_obj_spec wq_spec;
2263 struct mana_obj_spec cq_spec;
2264 struct gdma_queue_spec spec;
2265 struct mana_cq *cq = NULL;
2266 struct gdma_context *gc;
2267 u32 cq_size, rq_size;
2268 struct mana_rxq *rxq;
2269 int err;
2270
2271 gc = gd->gdma_context;
2272
2273 rxq = kzalloc(struct_size(rxq, rx_oobs, apc->rx_queue_size),
2274 GFP_KERNEL);
2275 if (!rxq)
2276 return NULL;
2277
2278 rxq->ndev = ndev;
2279 rxq->num_rx_buf = apc->rx_queue_size;
2280 rxq->rxq_idx = rxq_idx;
2281 rxq->rxobj = INVALID_MANA_HANDLE;
2282
2283 mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size,
2284 &rxq->headroom);
2285
2286 /* Create page pool for RX queue */
2287 err = mana_create_page_pool(rxq, gc);
2288 if (err) {
2289 netdev_err(ndev, "Create page pool err:%d\n", err);
2290 goto out;
2291 }
2292
2293 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size);
2294 if (err)
2295 goto out;
2296
2297 rq_size = MANA_PAGE_ALIGN(rq_size);
2298 cq_size = MANA_PAGE_ALIGN(cq_size);
2299
2300 /* Create RQ */
2301 memset(&spec, 0, sizeof(spec));
2302 spec.type = GDMA_RQ;
2303 spec.monitor_avl_buf = true;
2304 spec.queue_size = rq_size;
2305 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq);
2306 if (err)
2307 goto out;
2308
2309 /* Create RQ's CQ */
2310 cq = &rxq->rx_cq;
2311 cq->type = MANA_CQ_TYPE_RX;
2312 cq->rxq = rxq;
2313
2314 memset(&spec, 0, sizeof(spec));
2315 spec.type = GDMA_CQ;
2316 spec.monitor_avl_buf = false;
2317 spec.queue_size = cq_size;
2318 spec.cq.callback = mana_schedule_napi;
2319 spec.cq.parent_eq = eq->eq;
2320 spec.cq.context = cq;
2321 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2322 if (err)
2323 goto out;
2324
2325 memset(&wq_spec, 0, sizeof(wq_spec));
2326 memset(&cq_spec, 0, sizeof(cq_spec));
2327 wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle;
2328 wq_spec.queue_size = rxq->gdma_rq->queue_size;
2329
2330 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2331 cq_spec.queue_size = cq->gdma_cq->queue_size;
2332 cq_spec.modr_ctx_id = 0;
2333 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2334
2335 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ,
2336 &wq_spec, &cq_spec, &rxq->rxobj);
2337 if (err)
2338 goto out;
2339
2340 rxq->gdma_rq->id = wq_spec.queue_index;
2341 cq->gdma_cq->id = cq_spec.queue_index;
2342
2343 rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2344 cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2345
2346 rxq->gdma_id = rxq->gdma_rq->id;
2347 cq->gdma_id = cq->gdma_cq->id;
2348
2349 err = mana_push_wqe(rxq);
2350 if (err)
2351 goto out;
2352
2353 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
2354 err = -EINVAL;
2355 goto out;
2356 }
2357
2358 gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2359
2360 netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1);
2361
2362 WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx,
2363 cq->napi.napi_id));
2364 WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL,
2365 rxq->page_pool));
2366
2367 napi_enable(&cq->napi);
2368
2369 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2370out:
2371 if (!err)
2372 return rxq;
2373
2374 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err);
2375
2376 mana_destroy_rxq(apc, rxq, false);
2377
2378 if (cq)
2379 mana_deinit_cq(apc, cq);
2380
2381 return NULL;
2382}
2383
2384static void mana_create_rxq_debugfs(struct mana_port_context *apc, int idx)
2385{
2386 struct mana_rxq *rxq;
2387 char qnum[32];
2388
2389 rxq = apc->rxqs[idx];
2390
2391 sprintf(qnum, "RX-%d", idx);
2392 rxq->mana_rx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs);
2393 debugfs_create_u32("rq_head", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->head);
2394 debugfs_create_u32("rq_tail", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->tail);
2395 debugfs_create_u32("rq_nbuf", 0400, rxq->mana_rx_debugfs, &rxq->num_rx_buf);
2396 debugfs_create_u32("cq_head", 0400, rxq->mana_rx_debugfs,
2397 &rxq->rx_cq.gdma_cq->head);
2398 debugfs_create_u32("cq_tail", 0400, rxq->mana_rx_debugfs,
2399 &rxq->rx_cq.gdma_cq->tail);
2400 debugfs_create_u32("cq_budget", 0400, rxq->mana_rx_debugfs, &rxq->rx_cq.budget);
2401 debugfs_create_file("rxq_dump", 0400, rxq->mana_rx_debugfs, rxq->gdma_rq, &mana_dbg_q_fops);
2402 debugfs_create_file("cq_dump", 0400, rxq->mana_rx_debugfs, rxq->rx_cq.gdma_cq,
2403 &mana_dbg_q_fops);
2404}
2405
2406static int mana_add_rx_queues(struct mana_port_context *apc,
2407 struct net_device *ndev)
2408{
2409 struct mana_context *ac = apc->ac;
2410 struct mana_rxq *rxq;
2411 int err = 0;
2412 int i;
2413
2414 for (i = 0; i < apc->num_queues; i++) {
2415 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev);
2416 if (!rxq) {
2417 err = -ENOMEM;
2418 goto out;
2419 }
2420
2421 u64_stats_init(&rxq->stats.syncp);
2422
2423 apc->rxqs[i] = rxq;
2424
2425 mana_create_rxq_debugfs(apc, i);
2426 }
2427
2428 apc->default_rxobj = apc->rxqs[0]->rxobj;
2429out:
2430 return err;
2431}
2432
2433static void mana_destroy_vport(struct mana_port_context *apc)
2434{
2435 struct gdma_dev *gd = apc->ac->gdma_dev;
2436 struct mana_rxq *rxq;
2437 u32 rxq_idx;
2438
2439 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
2440 rxq = apc->rxqs[rxq_idx];
2441 if (!rxq)
2442 continue;
2443
2444 mana_destroy_rxq(apc, rxq, true);
2445 apc->rxqs[rxq_idx] = NULL;
2446 }
2447
2448 mana_destroy_txq(apc);
2449 mana_uncfg_vport(apc);
2450
2451 if (gd->gdma_context->is_pf)
2452 mana_pf_deregister_hw_vport(apc);
2453}
2454
2455static int mana_create_vport(struct mana_port_context *apc,
2456 struct net_device *net)
2457{
2458 struct gdma_dev *gd = apc->ac->gdma_dev;
2459 int err;
2460
2461 apc->default_rxobj = INVALID_MANA_HANDLE;
2462
2463 if (gd->gdma_context->is_pf) {
2464 err = mana_pf_register_hw_vport(apc);
2465 if (err)
2466 return err;
2467 }
2468
2469 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell);
2470 if (err)
2471 return err;
2472
2473 return mana_create_txq(apc, net);
2474}
2475
2476static int mana_rss_table_alloc(struct mana_port_context *apc)
2477{
2478 if (!apc->indir_table_sz) {
2479 netdev_err(apc->ndev,
2480 "Indirection table size not set for vPort %d\n",
2481 apc->port_idx);
2482 return -EINVAL;
2483 }
2484
2485 apc->indir_table = kcalloc(apc->indir_table_sz, sizeof(u32), GFP_KERNEL);
2486 if (!apc->indir_table)
2487 return -ENOMEM;
2488
2489 apc->rxobj_table = kcalloc(apc->indir_table_sz, sizeof(mana_handle_t), GFP_KERNEL);
2490 if (!apc->rxobj_table) {
2491 kfree(apc->indir_table);
2492 return -ENOMEM;
2493 }
2494
2495 return 0;
2496}
2497
2498static void mana_rss_table_init(struct mana_port_context *apc)
2499{
2500 int i;
2501
2502 for (i = 0; i < apc->indir_table_sz; i++)
2503 apc->indir_table[i] =
2504 ethtool_rxfh_indir_default(i, apc->num_queues);
2505}
2506
2507int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx,
2508 bool update_hash, bool update_tab)
2509{
2510 u32 queue_idx;
2511 int err;
2512 int i;
2513
2514 if (update_tab) {
2515 for (i = 0; i < apc->indir_table_sz; i++) {
2516 queue_idx = apc->indir_table[i];
2517 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj;
2518 }
2519 }
2520
2521 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab);
2522 if (err)
2523 return err;
2524
2525 mana_fence_rqs(apc);
2526
2527 return 0;
2528}
2529
2530void mana_query_gf_stats(struct mana_port_context *apc)
2531{
2532 struct mana_query_gf_stat_resp resp = {};
2533 struct mana_query_gf_stat_req req = {};
2534 struct net_device *ndev = apc->ndev;
2535 int err;
2536
2537 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_GF_STAT,
2538 sizeof(req), sizeof(resp));
2539 req.hdr.resp.msg_version = GDMA_MESSAGE_V2;
2540 req.req_stats = STATISTICS_FLAGS_RX_DISCARDS_NO_WQE |
2541 STATISTICS_FLAGS_RX_ERRORS_VPORT_DISABLED |
2542 STATISTICS_FLAGS_HC_RX_BYTES |
2543 STATISTICS_FLAGS_HC_RX_UCAST_PACKETS |
2544 STATISTICS_FLAGS_HC_RX_UCAST_BYTES |
2545 STATISTICS_FLAGS_HC_RX_MCAST_PACKETS |
2546 STATISTICS_FLAGS_HC_RX_MCAST_BYTES |
2547 STATISTICS_FLAGS_HC_RX_BCAST_PACKETS |
2548 STATISTICS_FLAGS_HC_RX_BCAST_BYTES |
2549 STATISTICS_FLAGS_TX_ERRORS_GF_DISABLED |
2550 STATISTICS_FLAGS_TX_ERRORS_VPORT_DISABLED |
2551 STATISTICS_FLAGS_TX_ERRORS_INVAL_VPORT_OFFSET_PACKETS |
2552 STATISTICS_FLAGS_TX_ERRORS_VLAN_ENFORCEMENT |
2553 STATISTICS_FLAGS_TX_ERRORS_ETH_TYPE_ENFORCEMENT |
2554 STATISTICS_FLAGS_TX_ERRORS_SA_ENFORCEMENT |
2555 STATISTICS_FLAGS_TX_ERRORS_SQPDID_ENFORCEMENT |
2556 STATISTICS_FLAGS_TX_ERRORS_CQPDID_ENFORCEMENT |
2557 STATISTICS_FLAGS_TX_ERRORS_MTU_VIOLATION |
2558 STATISTICS_FLAGS_TX_ERRORS_INVALID_OOB |
2559 STATISTICS_FLAGS_HC_TX_BYTES |
2560 STATISTICS_FLAGS_HC_TX_UCAST_PACKETS |
2561 STATISTICS_FLAGS_HC_TX_UCAST_BYTES |
2562 STATISTICS_FLAGS_HC_TX_MCAST_PACKETS |
2563 STATISTICS_FLAGS_HC_TX_MCAST_BYTES |
2564 STATISTICS_FLAGS_HC_TX_BCAST_PACKETS |
2565 STATISTICS_FLAGS_HC_TX_BCAST_BYTES |
2566 STATISTICS_FLAGS_TX_ERRORS_GDMA_ERROR;
2567
2568 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
2569 sizeof(resp));
2570 if (err) {
2571 netdev_err(ndev, "Failed to query GF stats: %d\n", err);
2572 return;
2573 }
2574 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_GF_STAT,
2575 sizeof(resp));
2576 if (err || resp.hdr.status) {
2577 netdev_err(ndev, "Failed to query GF stats: %d, 0x%x\n", err,
2578 resp.hdr.status);
2579 return;
2580 }
2581
2582 apc->eth_stats.hc_rx_discards_no_wqe = resp.rx_discards_nowqe;
2583 apc->eth_stats.hc_rx_err_vport_disabled = resp.rx_err_vport_disabled;
2584 apc->eth_stats.hc_rx_bytes = resp.hc_rx_bytes;
2585 apc->eth_stats.hc_rx_ucast_pkts = resp.hc_rx_ucast_pkts;
2586 apc->eth_stats.hc_rx_ucast_bytes = resp.hc_rx_ucast_bytes;
2587 apc->eth_stats.hc_rx_bcast_pkts = resp.hc_rx_bcast_pkts;
2588 apc->eth_stats.hc_rx_bcast_bytes = resp.hc_rx_bcast_bytes;
2589 apc->eth_stats.hc_rx_mcast_pkts = resp.hc_rx_mcast_pkts;
2590 apc->eth_stats.hc_rx_mcast_bytes = resp.hc_rx_mcast_bytes;
2591 apc->eth_stats.hc_tx_err_gf_disabled = resp.tx_err_gf_disabled;
2592 apc->eth_stats.hc_tx_err_vport_disabled = resp.tx_err_vport_disabled;
2593 apc->eth_stats.hc_tx_err_inval_vportoffset_pkt =
2594 resp.tx_err_inval_vport_offset_pkt;
2595 apc->eth_stats.hc_tx_err_vlan_enforcement =
2596 resp.tx_err_vlan_enforcement;
2597 apc->eth_stats.hc_tx_err_eth_type_enforcement =
2598 resp.tx_err_ethtype_enforcement;
2599 apc->eth_stats.hc_tx_err_sa_enforcement = resp.tx_err_SA_enforcement;
2600 apc->eth_stats.hc_tx_err_sqpdid_enforcement =
2601 resp.tx_err_SQPDID_enforcement;
2602 apc->eth_stats.hc_tx_err_cqpdid_enforcement =
2603 resp.tx_err_CQPDID_enforcement;
2604 apc->eth_stats.hc_tx_err_mtu_violation = resp.tx_err_mtu_violation;
2605 apc->eth_stats.hc_tx_err_inval_oob = resp.tx_err_inval_oob;
2606 apc->eth_stats.hc_tx_bytes = resp.hc_tx_bytes;
2607 apc->eth_stats.hc_tx_ucast_pkts = resp.hc_tx_ucast_pkts;
2608 apc->eth_stats.hc_tx_ucast_bytes = resp.hc_tx_ucast_bytes;
2609 apc->eth_stats.hc_tx_bcast_pkts = resp.hc_tx_bcast_pkts;
2610 apc->eth_stats.hc_tx_bcast_bytes = resp.hc_tx_bcast_bytes;
2611 apc->eth_stats.hc_tx_mcast_pkts = resp.hc_tx_mcast_pkts;
2612 apc->eth_stats.hc_tx_mcast_bytes = resp.hc_tx_mcast_bytes;
2613 apc->eth_stats.hc_tx_err_gdma = resp.tx_err_gdma;
2614}
2615
2616static int mana_init_port(struct net_device *ndev)
2617{
2618 struct mana_port_context *apc = netdev_priv(ndev);
2619 struct gdma_dev *gd = apc->ac->gdma_dev;
2620 u32 max_txq, max_rxq, max_queues;
2621 int port_idx = apc->port_idx;
2622 struct gdma_context *gc;
2623 char vport[32];
2624 int err;
2625
2626 err = mana_init_port_context(apc);
2627 if (err)
2628 return err;
2629
2630 gc = gd->gdma_context;
2631
2632 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq,
2633 &apc->indir_table_sz);
2634 if (err) {
2635 netdev_err(ndev, "Failed to query info for vPort %d\n",
2636 port_idx);
2637 goto reset_apc;
2638 }
2639
2640 max_queues = min_t(u32, max_txq, max_rxq);
2641 if (apc->max_queues > max_queues)
2642 apc->max_queues = max_queues;
2643
2644 if (apc->num_queues > apc->max_queues)
2645 apc->num_queues = apc->max_queues;
2646
2647 eth_hw_addr_set(ndev, apc->mac_addr);
2648 sprintf(vport, "vport%d", port_idx);
2649 apc->mana_port_debugfs = debugfs_create_dir(vport, gc->mana_pci_debugfs);
2650 return 0;
2651
2652reset_apc:
2653 mana_cleanup_port_context(apc);
2654 return err;
2655}
2656
2657int mana_alloc_queues(struct net_device *ndev)
2658{
2659 struct mana_port_context *apc = netdev_priv(ndev);
2660 struct gdma_dev *gd = apc->ac->gdma_dev;
2661 int err;
2662
2663 err = mana_create_vport(apc, ndev);
2664 if (err)
2665 return err;
2666
2667 err = netif_set_real_num_tx_queues(ndev, apc->num_queues);
2668 if (err)
2669 goto destroy_vport;
2670
2671 err = mana_add_rx_queues(apc, ndev);
2672 if (err)
2673 goto destroy_vport;
2674
2675 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE;
2676
2677 err = netif_set_real_num_rx_queues(ndev, apc->num_queues);
2678 if (err)
2679 goto destroy_vport;
2680
2681 mana_rss_table_init(apc);
2682
2683 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true);
2684 if (err)
2685 goto destroy_vport;
2686
2687 if (gd->gdma_context->is_pf) {
2688 err = mana_pf_register_filter(apc);
2689 if (err)
2690 goto destroy_vport;
2691 }
2692
2693 mana_chn_setxdp(apc, mana_xdp_get(apc));
2694
2695 return 0;
2696
2697destroy_vport:
2698 mana_destroy_vport(apc);
2699 return err;
2700}
2701
2702int mana_attach(struct net_device *ndev)
2703{
2704 struct mana_port_context *apc = netdev_priv(ndev);
2705 int err;
2706
2707 ASSERT_RTNL();
2708
2709 err = mana_init_port(ndev);
2710 if (err)
2711 return err;
2712
2713 if (apc->port_st_save) {
2714 err = mana_alloc_queues(ndev);
2715 if (err) {
2716 mana_cleanup_port_context(apc);
2717 return err;
2718 }
2719 }
2720
2721 apc->port_is_up = apc->port_st_save;
2722
2723 /* Ensure port state updated before txq state */
2724 smp_wmb();
2725
2726 if (apc->port_is_up)
2727 netif_carrier_on(ndev);
2728
2729 netif_device_attach(ndev);
2730
2731 return 0;
2732}
2733
2734static int mana_dealloc_queues(struct net_device *ndev)
2735{
2736 struct mana_port_context *apc = netdev_priv(ndev);
2737 unsigned long timeout = jiffies + 120 * HZ;
2738 struct gdma_dev *gd = apc->ac->gdma_dev;
2739 struct mana_txq *txq;
2740 struct sk_buff *skb;
2741 int i, err;
2742 u32 tsleep;
2743
2744 if (apc->port_is_up)
2745 return -EINVAL;
2746
2747 mana_chn_setxdp(apc, NULL);
2748
2749 if (gd->gdma_context->is_pf)
2750 mana_pf_deregister_filter(apc);
2751
2752 /* No packet can be transmitted now since apc->port_is_up is false.
2753 * There is still a tiny chance that mana_poll_tx_cq() can re-enable
2754 * a txq because it may not timely see apc->port_is_up being cleared
2755 * to false, but it doesn't matter since mana_start_xmit() drops any
2756 * new packets due to apc->port_is_up being false.
2757 *
2758 * Drain all the in-flight TX packets.
2759 * A timeout of 120 seconds for all the queues is used.
2760 * This will break the while loop when h/w is not responding.
2761 * This value of 120 has been decided here considering max
2762 * number of queues.
2763 */
2764
2765 for (i = 0; i < apc->num_queues; i++) {
2766 txq = &apc->tx_qp[i].txq;
2767 tsleep = 1000;
2768 while (atomic_read(&txq->pending_sends) > 0 &&
2769 time_before(jiffies, timeout)) {
2770 usleep_range(tsleep, tsleep + 1000);
2771 tsleep <<= 1;
2772 }
2773 if (atomic_read(&txq->pending_sends)) {
2774 err = pcie_flr(to_pci_dev(gd->gdma_context->dev));
2775 if (err) {
2776 netdev_err(ndev, "flr failed %d with %d pkts pending in txq %u\n",
2777 err, atomic_read(&txq->pending_sends),
2778 txq->gdma_txq_id);
2779 }
2780 break;
2781 }
2782 }
2783
2784 for (i = 0; i < apc->num_queues; i++) {
2785 txq = &apc->tx_qp[i].txq;
2786 while ((skb = skb_dequeue(&txq->pending_skbs))) {
2787 mana_unmap_skb(skb, apc);
2788 dev_kfree_skb_any(skb);
2789 }
2790 atomic_set(&txq->pending_sends, 0);
2791 }
2792 /* We're 100% sure the queues can no longer be woken up, because
2793 * we're sure now mana_poll_tx_cq() can't be running.
2794 */
2795
2796 apc->rss_state = TRI_STATE_FALSE;
2797 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false);
2798 if (err) {
2799 netdev_err(ndev, "Failed to disable vPort: %d\n", err);
2800 return err;
2801 }
2802
2803 mana_destroy_vport(apc);
2804
2805 return 0;
2806}
2807
2808int mana_detach(struct net_device *ndev, bool from_close)
2809{
2810 struct mana_port_context *apc = netdev_priv(ndev);
2811 int err;
2812
2813 ASSERT_RTNL();
2814
2815 apc->port_st_save = apc->port_is_up;
2816 apc->port_is_up = false;
2817
2818 /* Ensure port state updated before txq state */
2819 smp_wmb();
2820
2821 netif_tx_disable(ndev);
2822 netif_carrier_off(ndev);
2823
2824 if (apc->port_st_save) {
2825 err = mana_dealloc_queues(ndev);
2826 if (err)
2827 return err;
2828 }
2829
2830 if (!from_close) {
2831 netif_device_detach(ndev);
2832 mana_cleanup_port_context(apc);
2833 }
2834
2835 return 0;
2836}
2837
2838static int mana_probe_port(struct mana_context *ac, int port_idx,
2839 struct net_device **ndev_storage)
2840{
2841 struct gdma_context *gc = ac->gdma_dev->gdma_context;
2842 struct mana_port_context *apc;
2843 struct net_device *ndev;
2844 int err;
2845
2846 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context),
2847 gc->max_num_queues);
2848 if (!ndev)
2849 return -ENOMEM;
2850
2851 *ndev_storage = ndev;
2852
2853 apc = netdev_priv(ndev);
2854 apc->ac = ac;
2855 apc->ndev = ndev;
2856 apc->max_queues = gc->max_num_queues;
2857 apc->num_queues = gc->max_num_queues;
2858 apc->tx_queue_size = DEF_TX_BUFFERS_PER_QUEUE;
2859 apc->rx_queue_size = DEF_RX_BUFFERS_PER_QUEUE;
2860 apc->port_handle = INVALID_MANA_HANDLE;
2861 apc->pf_filter_handle = INVALID_MANA_HANDLE;
2862 apc->port_idx = port_idx;
2863
2864 mutex_init(&apc->vport_mutex);
2865 apc->vport_use_count = 0;
2866
2867 ndev->netdev_ops = &mana_devops;
2868 ndev->ethtool_ops = &mana_ethtool_ops;
2869 ndev->mtu = ETH_DATA_LEN;
2870 ndev->max_mtu = gc->adapter_mtu - ETH_HLEN;
2871 ndev->min_mtu = ETH_MIN_MTU;
2872 ndev->needed_headroom = MANA_HEADROOM;
2873 ndev->dev_port = port_idx;
2874 SET_NETDEV_DEV(ndev, gc->dev);
2875
2876 netif_carrier_off(ndev);
2877
2878 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE);
2879
2880 err = mana_init_port(ndev);
2881 if (err)
2882 goto free_net;
2883
2884 err = mana_rss_table_alloc(apc);
2885 if (err)
2886 goto reset_apc;
2887
2888 netdev_lockdep_set_classes(ndev);
2889
2890 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
2891 ndev->hw_features |= NETIF_F_RXCSUM;
2892 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
2893 ndev->hw_features |= NETIF_F_RXHASH;
2894 ndev->features = ndev->hw_features | NETIF_F_HW_VLAN_CTAG_TX |
2895 NETIF_F_HW_VLAN_CTAG_RX;
2896 ndev->vlan_features = ndev->features;
2897 xdp_set_features_flag(ndev, NETDEV_XDP_ACT_BASIC |
2898 NETDEV_XDP_ACT_REDIRECT |
2899 NETDEV_XDP_ACT_NDO_XMIT);
2900
2901 err = register_netdev(ndev);
2902 if (err) {
2903 netdev_err(ndev, "Unable to register netdev.\n");
2904 goto free_indir;
2905 }
2906
2907 return 0;
2908
2909free_indir:
2910 mana_cleanup_indir_table(apc);
2911reset_apc:
2912 mana_cleanup_port_context(apc);
2913free_net:
2914 *ndev_storage = NULL;
2915 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err);
2916 free_netdev(ndev);
2917 return err;
2918}
2919
2920static void adev_release(struct device *dev)
2921{
2922 struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev);
2923
2924 kfree(madev);
2925}
2926
2927static void remove_adev(struct gdma_dev *gd)
2928{
2929 struct auxiliary_device *adev = gd->adev;
2930 int id = adev->id;
2931
2932 auxiliary_device_delete(adev);
2933 auxiliary_device_uninit(adev);
2934
2935 mana_adev_idx_free(id);
2936 gd->adev = NULL;
2937}
2938
2939static int add_adev(struct gdma_dev *gd)
2940{
2941 struct auxiliary_device *adev;
2942 struct mana_adev *madev;
2943 int ret;
2944
2945 madev = kzalloc(sizeof(*madev), GFP_KERNEL);
2946 if (!madev)
2947 return -ENOMEM;
2948
2949 adev = &madev->adev;
2950 ret = mana_adev_idx_alloc();
2951 if (ret < 0)
2952 goto idx_fail;
2953 adev->id = ret;
2954
2955 adev->name = "rdma";
2956 adev->dev.parent = gd->gdma_context->dev;
2957 adev->dev.release = adev_release;
2958 madev->mdev = gd;
2959
2960 ret = auxiliary_device_init(adev);
2961 if (ret)
2962 goto init_fail;
2963
2964 /* madev is owned by the auxiliary device */
2965 madev = NULL;
2966 ret = auxiliary_device_add(adev);
2967 if (ret)
2968 goto add_fail;
2969
2970 gd->adev = adev;
2971 return 0;
2972
2973add_fail:
2974 auxiliary_device_uninit(adev);
2975
2976init_fail:
2977 mana_adev_idx_free(adev->id);
2978
2979idx_fail:
2980 kfree(madev);
2981
2982 return ret;
2983}
2984
2985int mana_probe(struct gdma_dev *gd, bool resuming)
2986{
2987 struct gdma_context *gc = gd->gdma_context;
2988 struct mana_context *ac = gd->driver_data;
2989 struct device *dev = gc->dev;
2990 u16 num_ports = 0;
2991 int err;
2992 int i;
2993
2994 dev_info(dev,
2995 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n",
2996 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION);
2997
2998 err = mana_gd_register_device(gd);
2999 if (err)
3000 return err;
3001
3002 if (!resuming) {
3003 ac = kzalloc(sizeof(*ac), GFP_KERNEL);
3004 if (!ac)
3005 return -ENOMEM;
3006
3007 ac->gdma_dev = gd;
3008 gd->driver_data = ac;
3009 }
3010
3011 err = mana_create_eq(ac);
3012 if (err)
3013 goto out;
3014
3015 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION,
3016 MANA_MICRO_VERSION, &num_ports);
3017 if (err)
3018 goto out;
3019
3020 if (!resuming) {
3021 ac->num_ports = num_ports;
3022 } else {
3023 if (ac->num_ports != num_ports) {
3024 dev_err(dev, "The number of vPorts changed: %d->%d\n",
3025 ac->num_ports, num_ports);
3026 err = -EPROTO;
3027 goto out;
3028 }
3029 }
3030
3031 if (ac->num_ports == 0)
3032 dev_err(dev, "Failed to detect any vPort\n");
3033
3034 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV)
3035 ac->num_ports = MAX_PORTS_IN_MANA_DEV;
3036
3037 if (!resuming) {
3038 for (i = 0; i < ac->num_ports; i++) {
3039 err = mana_probe_port(ac, i, &ac->ports[i]);
3040 /* we log the port for which the probe failed and stop
3041 * probes for subsequent ports.
3042 * Note that we keep running ports, for which the probes
3043 * were successful, unless add_adev fails too
3044 */
3045 if (err) {
3046 dev_err(dev, "Probe Failed for port %d\n", i);
3047 break;
3048 }
3049 }
3050 } else {
3051 for (i = 0; i < ac->num_ports; i++) {
3052 rtnl_lock();
3053 err = mana_attach(ac->ports[i]);
3054 rtnl_unlock();
3055 /* we log the port for which the attach failed and stop
3056 * attach for subsequent ports
3057 * Note that we keep running ports, for which the attach
3058 * were successful, unless add_adev fails too
3059 */
3060 if (err) {
3061 dev_err(dev, "Attach Failed for port %d\n", i);
3062 break;
3063 }
3064 }
3065 }
3066
3067 err = add_adev(gd);
3068out:
3069 if (err)
3070 mana_remove(gd, false);
3071
3072 return err;
3073}
3074
3075void mana_remove(struct gdma_dev *gd, bool suspending)
3076{
3077 struct gdma_context *gc = gd->gdma_context;
3078 struct mana_context *ac = gd->driver_data;
3079 struct mana_port_context *apc;
3080 struct device *dev = gc->dev;
3081 struct net_device *ndev;
3082 int err;
3083 int i;
3084
3085 /* adev currently doesn't support suspending, always remove it */
3086 if (gd->adev)
3087 remove_adev(gd);
3088
3089 for (i = 0; i < ac->num_ports; i++) {
3090 ndev = ac->ports[i];
3091 apc = netdev_priv(ndev);
3092 if (!ndev) {
3093 if (i == 0)
3094 dev_err(dev, "No net device to remove\n");
3095 goto out;
3096 }
3097
3098 /* All cleanup actions should stay after rtnl_lock(), otherwise
3099 * other functions may access partially cleaned up data.
3100 */
3101 rtnl_lock();
3102
3103 err = mana_detach(ndev, false);
3104 if (err)
3105 netdev_err(ndev, "Failed to detach vPort %d: %d\n",
3106 i, err);
3107
3108 if (suspending) {
3109 /* No need to unregister the ndev. */
3110 rtnl_unlock();
3111 continue;
3112 }
3113
3114 unregister_netdevice(ndev);
3115 mana_cleanup_indir_table(apc);
3116
3117 rtnl_unlock();
3118
3119 free_netdev(ndev);
3120 }
3121
3122 mana_destroy_eq(ac);
3123out:
3124 mana_gd_deregister_device(gd);
3125
3126 if (suspending)
3127 return;
3128
3129 gd->driver_data = NULL;
3130 gd->gdma_context = NULL;
3131 kfree(ac);
3132}
3133
3134struct net_device *mana_get_primary_netdev_rcu(struct mana_context *ac, u32 port_index)
3135{
3136 struct net_device *ndev;
3137
3138 RCU_LOCKDEP_WARN(!rcu_read_lock_held(),
3139 "Taking primary netdev without holding the RCU read lock");
3140 if (port_index >= ac->num_ports)
3141 return NULL;
3142
3143 /* When mana is used in netvsc, the upper netdevice should be returned. */
3144 if (ac->ports[port_index]->flags & IFF_SLAVE)
3145 ndev = netdev_master_upper_dev_get_rcu(ac->ports[port_index]);
3146 else
3147 ndev = ac->ports[port_index];
3148
3149 return ndev;
3150}
3151EXPORT_SYMBOL_NS(mana_get_primary_netdev_rcu, "NET_MANA");
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
21static DEFINE_IDA(mana_adev_ida);
22
23static int mana_adev_idx_alloc(void)
24{
25 return ida_alloc(&mana_adev_ida, GFP_KERNEL);
26}
27
28static 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
35static 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
55static 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
65static bool mana_can_tx(struct gdma_queue *wq)
66{
67 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE;
68}
69
70static 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
94static 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
105static 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
162frag_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 */
181static 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 */
202static 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
220netdev_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
415tx_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
424free_sgl_ptr:
425 kfree(pkg.sgl_ptr);
426tx_drop_count:
427 ndev->stats.tx_dropped++;
428tx_drop:
429 dev_kfree_skb_any(skb);
430 return NETDEV_TX_OK;
431}
432
433static 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
476static 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
493static 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 */
514static 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
537out2:
538 kfree(mpc->rxbufs_pre);
539 mpc->rxbufs_pre = NULL;
540
541out1:
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 */
548static 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 */
594static 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 = mtu + MANA_RXBUF_PAD + *headroom;
603
604 *datasize = mtu + ETH_HLEN;
605}
606
607static int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu)
608{
609 struct device *dev;
610 struct page *page;
611 dma_addr_t da;
612 int num_rxb;
613 void *va;
614 int i;
615
616 mana_get_rxbuf_cfg(new_mtu, &mpc->rxbpre_datasize,
617 &mpc->rxbpre_alloc_size, &mpc->rxbpre_headroom);
618
619 dev = mpc->ac->gdma_dev->gdma_context->dev;
620
621 num_rxb = mpc->num_queues * RX_BUFFERS_PER_QUEUE;
622
623 WARN(mpc->rxbufs_pre, "mana rxbufs_pre exists\n");
624 mpc->rxbufs_pre = kmalloc_array(num_rxb, sizeof(void *), GFP_KERNEL);
625 if (!mpc->rxbufs_pre)
626 goto error;
627
628 mpc->das_pre = kmalloc_array(num_rxb, sizeof(dma_addr_t), GFP_KERNEL);
629 if (!mpc->das_pre)
630 goto error;
631
632 mpc->rxbpre_total = 0;
633
634 for (i = 0; i < num_rxb; i++) {
635 if (mpc->rxbpre_alloc_size > PAGE_SIZE) {
636 va = netdev_alloc_frag(mpc->rxbpre_alloc_size);
637 if (!va)
638 goto error;
639
640 page = virt_to_head_page(va);
641 /* Check if the frag falls back to single page */
642 if (compound_order(page) <
643 get_order(mpc->rxbpre_alloc_size)) {
644 put_page(page);
645 goto error;
646 }
647 } else {
648 page = dev_alloc_page();
649 if (!page)
650 goto error;
651
652 va = page_to_virt(page);
653 }
654
655 da = dma_map_single(dev, va + mpc->rxbpre_headroom,
656 mpc->rxbpre_datasize, DMA_FROM_DEVICE);
657 if (dma_mapping_error(dev, da)) {
658 put_page(virt_to_head_page(va));
659 goto error;
660 }
661
662 mpc->rxbufs_pre[i] = va;
663 mpc->das_pre[i] = da;
664 mpc->rxbpre_total = i + 1;
665 }
666
667 return 0;
668
669error:
670 mana_pre_dealloc_rxbufs(mpc);
671 return -ENOMEM;
672}
673
674static int mana_change_mtu(struct net_device *ndev, int new_mtu)
675{
676 struct mana_port_context *mpc = netdev_priv(ndev);
677 unsigned int old_mtu = ndev->mtu;
678 int err;
679
680 /* Pre-allocate buffers to prevent failure in mana_attach later */
681 err = mana_pre_alloc_rxbufs(mpc, new_mtu);
682 if (err) {
683 netdev_err(ndev, "Insufficient memory for new MTU\n");
684 return err;
685 }
686
687 err = mana_detach(ndev, false);
688 if (err) {
689 netdev_err(ndev, "mana_detach failed: %d\n", err);
690 goto out;
691 }
692
693 ndev->mtu = new_mtu;
694
695 err = mana_attach(ndev);
696 if (err) {
697 netdev_err(ndev, "mana_attach failed: %d\n", err);
698 ndev->mtu = old_mtu;
699 }
700
701out:
702 mana_pre_dealloc_rxbufs(mpc);
703 return err;
704}
705
706static const struct net_device_ops mana_devops = {
707 .ndo_open = mana_open,
708 .ndo_stop = mana_close,
709 .ndo_select_queue = mana_select_queue,
710 .ndo_start_xmit = mana_start_xmit,
711 .ndo_validate_addr = eth_validate_addr,
712 .ndo_get_stats64 = mana_get_stats64,
713 .ndo_bpf = mana_bpf,
714 .ndo_xdp_xmit = mana_xdp_xmit,
715 .ndo_change_mtu = mana_change_mtu,
716};
717
718static void mana_cleanup_port_context(struct mana_port_context *apc)
719{
720 kfree(apc->rxqs);
721 apc->rxqs = NULL;
722}
723
724static int mana_init_port_context(struct mana_port_context *apc)
725{
726 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *),
727 GFP_KERNEL);
728
729 return !apc->rxqs ? -ENOMEM : 0;
730}
731
732static int mana_send_request(struct mana_context *ac, void *in_buf,
733 u32 in_len, void *out_buf, u32 out_len)
734{
735 struct gdma_context *gc = ac->gdma_dev->gdma_context;
736 struct gdma_resp_hdr *resp = out_buf;
737 struct gdma_req_hdr *req = in_buf;
738 struct device *dev = gc->dev;
739 static atomic_t activity_id;
740 int err;
741
742 req->dev_id = gc->mana.dev_id;
743 req->activity_id = atomic_inc_return(&activity_id);
744
745 err = mana_gd_send_request(gc, in_len, in_buf, out_len,
746 out_buf);
747 if (err || resp->status) {
748 dev_err(dev, "Failed to send mana message: %d, 0x%x\n",
749 err, resp->status);
750 return err ? err : -EPROTO;
751 }
752
753 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 ||
754 req->activity_id != resp->activity_id) {
755 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n",
756 req->dev_id.as_uint32, resp->dev_id.as_uint32,
757 req->activity_id, resp->activity_id);
758 return -EPROTO;
759 }
760
761 return 0;
762}
763
764static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr,
765 const enum mana_command_code expected_code,
766 const u32 min_size)
767{
768 if (resp_hdr->response.msg_type != expected_code)
769 return -EPROTO;
770
771 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1)
772 return -EPROTO;
773
774 if (resp_hdr->response.msg_size < min_size)
775 return -EPROTO;
776
777 return 0;
778}
779
780static int mana_pf_register_hw_vport(struct mana_port_context *apc)
781{
782 struct mana_register_hw_vport_resp resp = {};
783 struct mana_register_hw_vport_req req = {};
784 int err;
785
786 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT,
787 sizeof(req), sizeof(resp));
788 req.attached_gfid = 1;
789 req.is_pf_default_vport = 1;
790 req.allow_all_ether_types = 1;
791
792 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
793 sizeof(resp));
794 if (err) {
795 netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err);
796 return err;
797 }
798
799 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT,
800 sizeof(resp));
801 if (err || resp.hdr.status) {
802 netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n",
803 err, resp.hdr.status);
804 return err ? err : -EPROTO;
805 }
806
807 apc->port_handle = resp.hw_vport_handle;
808 return 0;
809}
810
811static void mana_pf_deregister_hw_vport(struct mana_port_context *apc)
812{
813 struct mana_deregister_hw_vport_resp resp = {};
814 struct mana_deregister_hw_vport_req req = {};
815 int err;
816
817 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT,
818 sizeof(req), sizeof(resp));
819 req.hw_vport_handle = apc->port_handle;
820
821 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
822 sizeof(resp));
823 if (err) {
824 netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n",
825 err);
826 return;
827 }
828
829 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT,
830 sizeof(resp));
831 if (err || resp.hdr.status)
832 netdev_err(apc->ndev,
833 "Failed to deregister hw vPort: %d, 0x%x\n",
834 err, resp.hdr.status);
835}
836
837static int mana_pf_register_filter(struct mana_port_context *apc)
838{
839 struct mana_register_filter_resp resp = {};
840 struct mana_register_filter_req req = {};
841 int err;
842
843 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER,
844 sizeof(req), sizeof(resp));
845 req.vport = apc->port_handle;
846 memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN);
847
848 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
849 sizeof(resp));
850 if (err) {
851 netdev_err(apc->ndev, "Failed to register filter: %d\n", err);
852 return err;
853 }
854
855 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER,
856 sizeof(resp));
857 if (err || resp.hdr.status) {
858 netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n",
859 err, resp.hdr.status);
860 return err ? err : -EPROTO;
861 }
862
863 apc->pf_filter_handle = resp.filter_handle;
864 return 0;
865}
866
867static void mana_pf_deregister_filter(struct mana_port_context *apc)
868{
869 struct mana_deregister_filter_resp resp = {};
870 struct mana_deregister_filter_req req = {};
871 int err;
872
873 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER,
874 sizeof(req), sizeof(resp));
875 req.filter_handle = apc->pf_filter_handle;
876
877 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
878 sizeof(resp));
879 if (err) {
880 netdev_err(apc->ndev, "Failed to unregister filter: %d\n",
881 err);
882 return;
883 }
884
885 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER,
886 sizeof(resp));
887 if (err || resp.hdr.status)
888 netdev_err(apc->ndev,
889 "Failed to deregister filter: %d, 0x%x\n",
890 err, resp.hdr.status);
891}
892
893static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver,
894 u32 proto_minor_ver, u32 proto_micro_ver,
895 u16 *max_num_vports)
896{
897 struct gdma_context *gc = ac->gdma_dev->gdma_context;
898 struct mana_query_device_cfg_resp resp = {};
899 struct mana_query_device_cfg_req req = {};
900 struct device *dev = gc->dev;
901 int err = 0;
902
903 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG,
904 sizeof(req), sizeof(resp));
905
906 req.hdr.resp.msg_version = GDMA_MESSAGE_V2;
907
908 req.proto_major_ver = proto_major_ver;
909 req.proto_minor_ver = proto_minor_ver;
910 req.proto_micro_ver = proto_micro_ver;
911
912 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp));
913 if (err) {
914 dev_err(dev, "Failed to query config: %d", err);
915 return err;
916 }
917
918 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG,
919 sizeof(resp));
920 if (err || resp.hdr.status) {
921 dev_err(dev, "Invalid query result: %d, 0x%x\n", err,
922 resp.hdr.status);
923 if (!err)
924 err = -EPROTO;
925 return err;
926 }
927
928 *max_num_vports = resp.max_num_vports;
929
930 if (resp.hdr.response.msg_version == GDMA_MESSAGE_V2)
931 gc->adapter_mtu = resp.adapter_mtu;
932 else
933 gc->adapter_mtu = ETH_FRAME_LEN;
934
935 return 0;
936}
937
938static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index,
939 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry)
940{
941 struct mana_query_vport_cfg_resp resp = {};
942 struct mana_query_vport_cfg_req req = {};
943 int err;
944
945 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG,
946 sizeof(req), sizeof(resp));
947
948 req.vport_index = vport_index;
949
950 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
951 sizeof(resp));
952 if (err)
953 return err;
954
955 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG,
956 sizeof(resp));
957 if (err)
958 return err;
959
960 if (resp.hdr.status)
961 return -EPROTO;
962
963 *max_sq = resp.max_num_sq;
964 *max_rq = resp.max_num_rq;
965 *num_indir_entry = resp.num_indirection_ent;
966
967 apc->port_handle = resp.vport;
968 ether_addr_copy(apc->mac_addr, resp.mac_addr);
969
970 return 0;
971}
972
973void mana_uncfg_vport(struct mana_port_context *apc)
974{
975 mutex_lock(&apc->vport_mutex);
976 apc->vport_use_count--;
977 WARN_ON(apc->vport_use_count < 0);
978 mutex_unlock(&apc->vport_mutex);
979}
980EXPORT_SYMBOL_NS(mana_uncfg_vport, NET_MANA);
981
982int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id,
983 u32 doorbell_pg_id)
984{
985 struct mana_config_vport_resp resp = {};
986 struct mana_config_vport_req req = {};
987 int err;
988
989 /* This function is used to program the Ethernet port in the hardware
990 * table. It can be called from the Ethernet driver or the RDMA driver.
991 *
992 * For Ethernet usage, the hardware supports only one active user on a
993 * physical port. The driver checks on the port usage before programming
994 * the hardware when creating the RAW QP (RDMA driver) or exposing the
995 * device to kernel NET layer (Ethernet driver).
996 *
997 * Because the RDMA driver doesn't know in advance which QP type the
998 * user will create, it exposes the device with all its ports. The user
999 * may not be able to create RAW QP on a port if this port is already
1000 * in used by the Ethernet driver from the kernel.
1001 *
1002 * This physical port limitation only applies to the RAW QP. For RC QP,
1003 * the hardware doesn't have this limitation. The user can create RC
1004 * QPs on a physical port up to the hardware limits independent of the
1005 * Ethernet usage on the same port.
1006 */
1007 mutex_lock(&apc->vport_mutex);
1008 if (apc->vport_use_count > 0) {
1009 mutex_unlock(&apc->vport_mutex);
1010 return -EBUSY;
1011 }
1012 apc->vport_use_count++;
1013 mutex_unlock(&apc->vport_mutex);
1014
1015 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX,
1016 sizeof(req), sizeof(resp));
1017 req.vport = apc->port_handle;
1018 req.pdid = protection_dom_id;
1019 req.doorbell_pageid = doorbell_pg_id;
1020
1021 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1022 sizeof(resp));
1023 if (err) {
1024 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err);
1025 goto out;
1026 }
1027
1028 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX,
1029 sizeof(resp));
1030 if (err || resp.hdr.status) {
1031 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n",
1032 err, resp.hdr.status);
1033 if (!err)
1034 err = -EPROTO;
1035
1036 goto out;
1037 }
1038
1039 apc->tx_shortform_allowed = resp.short_form_allowed;
1040 apc->tx_vp_offset = resp.tx_vport_offset;
1041
1042 netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n",
1043 apc->port_handle, protection_dom_id, doorbell_pg_id);
1044out:
1045 if (err)
1046 mana_uncfg_vport(apc);
1047
1048 return err;
1049}
1050EXPORT_SYMBOL_NS(mana_cfg_vport, NET_MANA);
1051
1052static int mana_cfg_vport_steering(struct mana_port_context *apc,
1053 enum TRI_STATE rx,
1054 bool update_default_rxobj, bool update_key,
1055 bool update_tab)
1056{
1057 u16 num_entries = MANA_INDIRECT_TABLE_SIZE;
1058 struct mana_cfg_rx_steer_req_v2 *req;
1059 struct mana_cfg_rx_steer_resp resp = {};
1060 struct net_device *ndev = apc->ndev;
1061 mana_handle_t *req_indir_tab;
1062 u32 req_buf_size;
1063 int err;
1064
1065 req_buf_size = sizeof(*req) + sizeof(mana_handle_t) * num_entries;
1066 req = kzalloc(req_buf_size, GFP_KERNEL);
1067 if (!req)
1068 return -ENOMEM;
1069
1070 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
1071 sizeof(resp));
1072
1073 req->hdr.req.msg_version = GDMA_MESSAGE_V2;
1074
1075 req->vport = apc->port_handle;
1076 req->num_indir_entries = num_entries;
1077 req->indir_tab_offset = sizeof(*req);
1078 req->rx_enable = rx;
1079 req->rss_enable = apc->rss_state;
1080 req->update_default_rxobj = update_default_rxobj;
1081 req->update_hashkey = update_key;
1082 req->update_indir_tab = update_tab;
1083 req->default_rxobj = apc->default_rxobj;
1084 req->cqe_coalescing_enable = 0;
1085
1086 if (update_key)
1087 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE);
1088
1089 if (update_tab) {
1090 req_indir_tab = (mana_handle_t *)(req + 1);
1091 memcpy(req_indir_tab, apc->rxobj_table,
1092 req->num_indir_entries * sizeof(mana_handle_t));
1093 }
1094
1095 err = mana_send_request(apc->ac, req, req_buf_size, &resp,
1096 sizeof(resp));
1097 if (err) {
1098 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err);
1099 goto out;
1100 }
1101
1102 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX,
1103 sizeof(resp));
1104 if (err) {
1105 netdev_err(ndev, "vPort RX configuration failed: %d\n", err);
1106 goto out;
1107 }
1108
1109 if (resp.hdr.status) {
1110 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n",
1111 resp.hdr.status);
1112 err = -EPROTO;
1113 }
1114
1115 netdev_info(ndev, "Configured steering vPort %llu entries %u\n",
1116 apc->port_handle, num_entries);
1117out:
1118 kfree(req);
1119 return err;
1120}
1121
1122int mana_create_wq_obj(struct mana_port_context *apc,
1123 mana_handle_t vport,
1124 u32 wq_type, struct mana_obj_spec *wq_spec,
1125 struct mana_obj_spec *cq_spec,
1126 mana_handle_t *wq_obj)
1127{
1128 struct mana_create_wqobj_resp resp = {};
1129 struct mana_create_wqobj_req req = {};
1130 struct net_device *ndev = apc->ndev;
1131 int err;
1132
1133 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ,
1134 sizeof(req), sizeof(resp));
1135 req.vport = vport;
1136 req.wq_type = wq_type;
1137 req.wq_gdma_region = wq_spec->gdma_region;
1138 req.cq_gdma_region = cq_spec->gdma_region;
1139 req.wq_size = wq_spec->queue_size;
1140 req.cq_size = cq_spec->queue_size;
1141 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id;
1142 req.cq_parent_qid = cq_spec->attached_eq;
1143
1144 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1145 sizeof(resp));
1146 if (err) {
1147 netdev_err(ndev, "Failed to create WQ object: %d\n", err);
1148 goto out;
1149 }
1150
1151 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ,
1152 sizeof(resp));
1153 if (err || resp.hdr.status) {
1154 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err,
1155 resp.hdr.status);
1156 if (!err)
1157 err = -EPROTO;
1158 goto out;
1159 }
1160
1161 if (resp.wq_obj == INVALID_MANA_HANDLE) {
1162 netdev_err(ndev, "Got an invalid WQ object handle\n");
1163 err = -EPROTO;
1164 goto out;
1165 }
1166
1167 *wq_obj = resp.wq_obj;
1168 wq_spec->queue_index = resp.wq_id;
1169 cq_spec->queue_index = resp.cq_id;
1170
1171 return 0;
1172out:
1173 return err;
1174}
1175EXPORT_SYMBOL_NS(mana_create_wq_obj, NET_MANA);
1176
1177void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type,
1178 mana_handle_t wq_obj)
1179{
1180 struct mana_destroy_wqobj_resp resp = {};
1181 struct mana_destroy_wqobj_req req = {};
1182 struct net_device *ndev = apc->ndev;
1183 int err;
1184
1185 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ,
1186 sizeof(req), sizeof(resp));
1187 req.wq_type = wq_type;
1188 req.wq_obj_handle = wq_obj;
1189
1190 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1191 sizeof(resp));
1192 if (err) {
1193 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err);
1194 return;
1195 }
1196
1197 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ,
1198 sizeof(resp));
1199 if (err || resp.hdr.status)
1200 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err,
1201 resp.hdr.status);
1202}
1203EXPORT_SYMBOL_NS(mana_destroy_wq_obj, NET_MANA);
1204
1205static void mana_destroy_eq(struct mana_context *ac)
1206{
1207 struct gdma_context *gc = ac->gdma_dev->gdma_context;
1208 struct gdma_queue *eq;
1209 int i;
1210
1211 if (!ac->eqs)
1212 return;
1213
1214 for (i = 0; i < gc->max_num_queues; i++) {
1215 eq = ac->eqs[i].eq;
1216 if (!eq)
1217 continue;
1218
1219 mana_gd_destroy_queue(gc, eq);
1220 }
1221
1222 kfree(ac->eqs);
1223 ac->eqs = NULL;
1224}
1225
1226static int mana_create_eq(struct mana_context *ac)
1227{
1228 struct gdma_dev *gd = ac->gdma_dev;
1229 struct gdma_context *gc = gd->gdma_context;
1230 struct gdma_queue_spec spec = {};
1231 int err;
1232 int i;
1233
1234 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq),
1235 GFP_KERNEL);
1236 if (!ac->eqs)
1237 return -ENOMEM;
1238
1239 spec.type = GDMA_EQ;
1240 spec.monitor_avl_buf = false;
1241 spec.queue_size = EQ_SIZE;
1242 spec.eq.callback = NULL;
1243 spec.eq.context = ac->eqs;
1244 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE;
1245
1246 for (i = 0; i < gc->max_num_queues; i++) {
1247 spec.eq.msix_index = (i + 1) % gc->num_msix_usable;
1248 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq);
1249 if (err)
1250 goto out;
1251 }
1252
1253 return 0;
1254out:
1255 mana_destroy_eq(ac);
1256 return err;
1257}
1258
1259static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq)
1260{
1261 struct mana_fence_rq_resp resp = {};
1262 struct mana_fence_rq_req req = {};
1263 int err;
1264
1265 init_completion(&rxq->fence_event);
1266
1267 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ,
1268 sizeof(req), sizeof(resp));
1269 req.wq_obj_handle = rxq->rxobj;
1270
1271 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1272 sizeof(resp));
1273 if (err) {
1274 netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n",
1275 rxq->rxq_idx, err);
1276 return err;
1277 }
1278
1279 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp));
1280 if (err || resp.hdr.status) {
1281 netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n",
1282 rxq->rxq_idx, err, resp.hdr.status);
1283 if (!err)
1284 err = -EPROTO;
1285
1286 return err;
1287 }
1288
1289 if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) {
1290 netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n",
1291 rxq->rxq_idx);
1292 return -ETIMEDOUT;
1293 }
1294
1295 return 0;
1296}
1297
1298static void mana_fence_rqs(struct mana_port_context *apc)
1299{
1300 unsigned int rxq_idx;
1301 struct mana_rxq *rxq;
1302 int err;
1303
1304 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
1305 rxq = apc->rxqs[rxq_idx];
1306 err = mana_fence_rq(apc, rxq);
1307
1308 /* In case of any error, use sleep instead. */
1309 if (err)
1310 msleep(100);
1311 }
1312}
1313
1314static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units)
1315{
1316 u32 used_space_old;
1317 u32 used_space_new;
1318
1319 used_space_old = wq->head - wq->tail;
1320 used_space_new = wq->head - (wq->tail + num_units);
1321
1322 if (WARN_ON_ONCE(used_space_new > used_space_old))
1323 return -ERANGE;
1324
1325 wq->tail += num_units;
1326 return 0;
1327}
1328
1329static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc)
1330{
1331 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
1332 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1333 struct device *dev = gc->dev;
1334 int hsg, i;
1335
1336 /* Number of SGEs of linear part */
1337 hsg = (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0]) ? 2 : 1;
1338
1339 for (i = 0; i < hsg; i++)
1340 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i],
1341 DMA_TO_DEVICE);
1342
1343 for (i = hsg; i < skb_shinfo(skb)->nr_frags + hsg; i++)
1344 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
1345 DMA_TO_DEVICE);
1346}
1347
1348static void mana_poll_tx_cq(struct mana_cq *cq)
1349{
1350 struct gdma_comp *completions = cq->gdma_comp_buf;
1351 struct gdma_posted_wqe_info *wqe_info;
1352 unsigned int pkt_transmitted = 0;
1353 unsigned int wqe_unit_cnt = 0;
1354 struct mana_txq *txq = cq->txq;
1355 struct mana_port_context *apc;
1356 struct netdev_queue *net_txq;
1357 struct gdma_queue *gdma_wq;
1358 unsigned int avail_space;
1359 struct net_device *ndev;
1360 struct sk_buff *skb;
1361 bool txq_stopped;
1362 int comp_read;
1363 int i;
1364
1365 ndev = txq->ndev;
1366 apc = netdev_priv(ndev);
1367
1368 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions,
1369 CQE_POLLING_BUFFER);
1370
1371 if (comp_read < 1)
1372 return;
1373
1374 for (i = 0; i < comp_read; i++) {
1375 struct mana_tx_comp_oob *cqe_oob;
1376
1377 if (WARN_ON_ONCE(!completions[i].is_sq))
1378 return;
1379
1380 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data;
1381 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type !=
1382 MANA_CQE_COMPLETION))
1383 return;
1384
1385 switch (cqe_oob->cqe_hdr.cqe_type) {
1386 case CQE_TX_OKAY:
1387 break;
1388
1389 case CQE_TX_SA_DROP:
1390 case CQE_TX_MTU_DROP:
1391 case CQE_TX_INVALID_OOB:
1392 case CQE_TX_INVALID_ETH_TYPE:
1393 case CQE_TX_HDR_PROCESSING_ERROR:
1394 case CQE_TX_VF_DISABLED:
1395 case CQE_TX_VPORT_IDX_OUT_OF_RANGE:
1396 case CQE_TX_VPORT_DISABLED:
1397 case CQE_TX_VLAN_TAGGING_VIOLATION:
1398 if (net_ratelimit())
1399 netdev_err(ndev, "TX: CQE error %d\n",
1400 cqe_oob->cqe_hdr.cqe_type);
1401
1402 apc->eth_stats.tx_cqe_err++;
1403 break;
1404
1405 default:
1406 /* If the CQE type is unknown, log an error,
1407 * and still free the SKB, update tail, etc.
1408 */
1409 if (net_ratelimit())
1410 netdev_err(ndev, "TX: unknown CQE type %d\n",
1411 cqe_oob->cqe_hdr.cqe_type);
1412
1413 apc->eth_stats.tx_cqe_unknown_type++;
1414 break;
1415 }
1416
1417 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num))
1418 return;
1419
1420 skb = skb_dequeue(&txq->pending_skbs);
1421 if (WARN_ON_ONCE(!skb))
1422 return;
1423
1424 wqe_info = (struct gdma_posted_wqe_info *)skb->cb;
1425 wqe_unit_cnt += wqe_info->wqe_size_in_bu;
1426
1427 mana_unmap_skb(skb, apc);
1428
1429 napi_consume_skb(skb, cq->budget);
1430
1431 pkt_transmitted++;
1432 }
1433
1434 if (WARN_ON_ONCE(wqe_unit_cnt == 0))
1435 return;
1436
1437 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt);
1438
1439 gdma_wq = txq->gdma_sq;
1440 avail_space = mana_gd_wq_avail_space(gdma_wq);
1441
1442 /* Ensure tail updated before checking q stop */
1443 smp_mb();
1444
1445 net_txq = txq->net_txq;
1446 txq_stopped = netif_tx_queue_stopped(net_txq);
1447
1448 /* Ensure checking txq_stopped before apc->port_is_up. */
1449 smp_rmb();
1450
1451 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) {
1452 netif_tx_wake_queue(net_txq);
1453 apc->eth_stats.wake_queue++;
1454 }
1455
1456 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0)
1457 WARN_ON_ONCE(1);
1458
1459 cq->work_done = pkt_transmitted;
1460}
1461
1462static void mana_post_pkt_rxq(struct mana_rxq *rxq)
1463{
1464 struct mana_recv_buf_oob *recv_buf_oob;
1465 u32 curr_index;
1466 int err;
1467
1468 curr_index = rxq->buf_index++;
1469 if (rxq->buf_index == rxq->num_rx_buf)
1470 rxq->buf_index = 0;
1471
1472 recv_buf_oob = &rxq->rx_oobs[curr_index];
1473
1474 err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req,
1475 &recv_buf_oob->wqe_inf);
1476 if (WARN_ON_ONCE(err))
1477 return;
1478
1479 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1);
1480}
1481
1482static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va,
1483 uint pkt_len, struct xdp_buff *xdp)
1484{
1485 struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size);
1486
1487 if (!skb)
1488 return NULL;
1489
1490 if (xdp->data_hard_start) {
1491 skb_reserve(skb, xdp->data - xdp->data_hard_start);
1492 skb_put(skb, xdp->data_end - xdp->data);
1493 return skb;
1494 }
1495
1496 skb_reserve(skb, rxq->headroom);
1497 skb_put(skb, pkt_len);
1498
1499 return skb;
1500}
1501
1502static void mana_rx_skb(void *buf_va, bool from_pool,
1503 struct mana_rxcomp_oob *cqe, struct mana_rxq *rxq)
1504{
1505 struct mana_stats_rx *rx_stats = &rxq->stats;
1506 struct net_device *ndev = rxq->ndev;
1507 uint pkt_len = cqe->ppi[0].pkt_len;
1508 u16 rxq_idx = rxq->rxq_idx;
1509 struct napi_struct *napi;
1510 struct xdp_buff xdp = {};
1511 struct sk_buff *skb;
1512 u32 hash_value;
1513 u32 act;
1514
1515 rxq->rx_cq.work_done++;
1516 napi = &rxq->rx_cq.napi;
1517
1518 if (!buf_va) {
1519 ++ndev->stats.rx_dropped;
1520 return;
1521 }
1522
1523 act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len);
1524
1525 if (act == XDP_REDIRECT && !rxq->xdp_rc)
1526 return;
1527
1528 if (act != XDP_PASS && act != XDP_TX)
1529 goto drop_xdp;
1530
1531 skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp);
1532
1533 if (!skb)
1534 goto drop;
1535
1536 if (from_pool)
1537 skb_mark_for_recycle(skb);
1538
1539 skb->dev = napi->dev;
1540
1541 skb->protocol = eth_type_trans(skb, ndev);
1542 skb_checksum_none_assert(skb);
1543 skb_record_rx_queue(skb, rxq_idx);
1544
1545 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) {
1546 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed)
1547 skb->ip_summed = CHECKSUM_UNNECESSARY;
1548 }
1549
1550 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) {
1551 hash_value = cqe->ppi[0].pkt_hash;
1552
1553 if (cqe->rx_hashtype & MANA_HASH_L4)
1554 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4);
1555 else
1556 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3);
1557 }
1558
1559 if (cqe->rx_vlantag_present) {
1560 u16 vlan_tci = cqe->rx_vlan_id;
1561
1562 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci);
1563 }
1564
1565 u64_stats_update_begin(&rx_stats->syncp);
1566 rx_stats->packets++;
1567 rx_stats->bytes += pkt_len;
1568
1569 if (act == XDP_TX)
1570 rx_stats->xdp_tx++;
1571 u64_stats_update_end(&rx_stats->syncp);
1572
1573 if (act == XDP_TX) {
1574 skb_set_queue_mapping(skb, rxq_idx);
1575 mana_xdp_tx(skb, ndev);
1576 return;
1577 }
1578
1579 napi_gro_receive(napi, skb);
1580
1581 return;
1582
1583drop_xdp:
1584 u64_stats_update_begin(&rx_stats->syncp);
1585 rx_stats->xdp_drop++;
1586 u64_stats_update_end(&rx_stats->syncp);
1587
1588drop:
1589 if (from_pool) {
1590 page_pool_recycle_direct(rxq->page_pool,
1591 virt_to_head_page(buf_va));
1592 } else {
1593 WARN_ON_ONCE(rxq->xdp_save_va);
1594 /* Save for reuse */
1595 rxq->xdp_save_va = buf_va;
1596 }
1597
1598 ++ndev->stats.rx_dropped;
1599
1600 return;
1601}
1602
1603static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,
1604 dma_addr_t *da, bool *from_pool, bool is_napi)
1605{
1606 struct page *page;
1607 void *va;
1608
1609 *from_pool = false;
1610
1611 /* Reuse XDP dropped page if available */
1612 if (rxq->xdp_save_va) {
1613 va = rxq->xdp_save_va;
1614 rxq->xdp_save_va = NULL;
1615 } else if (rxq->alloc_size > PAGE_SIZE) {
1616 if (is_napi)
1617 va = napi_alloc_frag(rxq->alloc_size);
1618 else
1619 va = netdev_alloc_frag(rxq->alloc_size);
1620
1621 if (!va)
1622 return NULL;
1623
1624 page = virt_to_head_page(va);
1625 /* Check if the frag falls back to single page */
1626 if (compound_order(page) < get_order(rxq->alloc_size)) {
1627 put_page(page);
1628 return NULL;
1629 }
1630 } else {
1631 page = page_pool_dev_alloc_pages(rxq->page_pool);
1632 if (!page)
1633 return NULL;
1634
1635 *from_pool = true;
1636 va = page_to_virt(page);
1637 }
1638
1639 *da = dma_map_single(dev, va + rxq->headroom, rxq->datasize,
1640 DMA_FROM_DEVICE);
1641 if (dma_mapping_error(dev, *da)) {
1642 if (*from_pool)
1643 page_pool_put_full_page(rxq->page_pool, page, false);
1644 else
1645 put_page(virt_to_head_page(va));
1646
1647 return NULL;
1648 }
1649
1650 return va;
1651}
1652
1653/* Allocate frag for rx buffer, and save the old buf */
1654static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq,
1655 struct mana_recv_buf_oob *rxoob, void **old_buf,
1656 bool *old_fp)
1657{
1658 bool from_pool;
1659 dma_addr_t da;
1660 void *va;
1661
1662 va = mana_get_rxfrag(rxq, dev, &da, &from_pool, true);
1663 if (!va)
1664 return;
1665
1666 dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize,
1667 DMA_FROM_DEVICE);
1668 *old_buf = rxoob->buf_va;
1669 *old_fp = rxoob->from_pool;
1670
1671 rxoob->buf_va = va;
1672 rxoob->sgl[0].address = da;
1673 rxoob->from_pool = from_pool;
1674}
1675
1676static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,
1677 struct gdma_comp *cqe)
1678{
1679 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data;
1680 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1681 struct net_device *ndev = rxq->ndev;
1682 struct mana_recv_buf_oob *rxbuf_oob;
1683 struct mana_port_context *apc;
1684 struct device *dev = gc->dev;
1685 void *old_buf = NULL;
1686 u32 curr, pktlen;
1687 bool old_fp;
1688
1689 apc = netdev_priv(ndev);
1690
1691 switch (oob->cqe_hdr.cqe_type) {
1692 case CQE_RX_OKAY:
1693 break;
1694
1695 case CQE_RX_TRUNCATED:
1696 ++ndev->stats.rx_dropped;
1697 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index];
1698 netdev_warn_once(ndev, "Dropped a truncated packet\n");
1699 goto drop;
1700
1701 case CQE_RX_COALESCED_4:
1702 netdev_err(ndev, "RX coalescing is unsupported\n");
1703 apc->eth_stats.rx_coalesced_err++;
1704 return;
1705
1706 case CQE_RX_OBJECT_FENCE:
1707 complete(&rxq->fence_event);
1708 return;
1709
1710 default:
1711 netdev_err(ndev, "Unknown RX CQE type = %d\n",
1712 oob->cqe_hdr.cqe_type);
1713 apc->eth_stats.rx_cqe_unknown_type++;
1714 return;
1715 }
1716
1717 pktlen = oob->ppi[0].pkt_len;
1718
1719 if (pktlen == 0) {
1720 /* data packets should never have packetlength of zero */
1721 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n",
1722 rxq->gdma_id, cq->gdma_id, rxq->rxobj);
1723 return;
1724 }
1725
1726 curr = rxq->buf_index;
1727 rxbuf_oob = &rxq->rx_oobs[curr];
1728 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1);
1729
1730 mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf, &old_fp);
1731
1732 /* Unsuccessful refill will have old_buf == NULL.
1733 * In this case, mana_rx_skb() will drop the packet.
1734 */
1735 mana_rx_skb(old_buf, old_fp, oob, rxq);
1736
1737drop:
1738 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu);
1739
1740 mana_post_pkt_rxq(rxq);
1741}
1742
1743static void mana_poll_rx_cq(struct mana_cq *cq)
1744{
1745 struct gdma_comp *comp = cq->gdma_comp_buf;
1746 struct mana_rxq *rxq = cq->rxq;
1747 int comp_read, i;
1748
1749 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER);
1750 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER);
1751
1752 rxq->xdp_flush = false;
1753
1754 for (i = 0; i < comp_read; i++) {
1755 if (WARN_ON_ONCE(comp[i].is_sq))
1756 return;
1757
1758 /* verify recv cqe references the right rxq */
1759 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id))
1760 return;
1761
1762 mana_process_rx_cqe(rxq, cq, &comp[i]);
1763 }
1764
1765 if (comp_read > 0) {
1766 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1767
1768 mana_gd_wq_ring_doorbell(gc, rxq->gdma_rq);
1769 }
1770
1771 if (rxq->xdp_flush)
1772 xdp_do_flush();
1773}
1774
1775static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue)
1776{
1777 struct mana_cq *cq = context;
1778 u8 arm_bit;
1779 int w;
1780
1781 WARN_ON_ONCE(cq->gdma_cq != gdma_queue);
1782
1783 if (cq->type == MANA_CQ_TYPE_RX)
1784 mana_poll_rx_cq(cq);
1785 else
1786 mana_poll_tx_cq(cq);
1787
1788 w = cq->work_done;
1789
1790 if (w < cq->budget &&
1791 napi_complete_done(&cq->napi, w)) {
1792 arm_bit = SET_ARM_BIT;
1793 } else {
1794 arm_bit = 0;
1795 }
1796
1797 mana_gd_ring_cq(gdma_queue, arm_bit);
1798
1799 return w;
1800}
1801
1802static int mana_poll(struct napi_struct *napi, int budget)
1803{
1804 struct mana_cq *cq = container_of(napi, struct mana_cq, napi);
1805 int w;
1806
1807 cq->work_done = 0;
1808 cq->budget = budget;
1809
1810 w = mana_cq_handler(cq, cq->gdma_cq);
1811
1812 return min(w, budget);
1813}
1814
1815static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue)
1816{
1817 struct mana_cq *cq = context;
1818
1819 napi_schedule_irqoff(&cq->napi);
1820}
1821
1822static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq)
1823{
1824 struct gdma_dev *gd = apc->ac->gdma_dev;
1825
1826 if (!cq->gdma_cq)
1827 return;
1828
1829 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq);
1830}
1831
1832static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq)
1833{
1834 struct gdma_dev *gd = apc->ac->gdma_dev;
1835
1836 if (!txq->gdma_sq)
1837 return;
1838
1839 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq);
1840}
1841
1842static void mana_destroy_txq(struct mana_port_context *apc)
1843{
1844 struct napi_struct *napi;
1845 int i;
1846
1847 if (!apc->tx_qp)
1848 return;
1849
1850 for (i = 0; i < apc->num_queues; i++) {
1851 napi = &apc->tx_qp[i].tx_cq.napi;
1852 napi_synchronize(napi);
1853 napi_disable(napi);
1854 netif_napi_del(napi);
1855
1856 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object);
1857
1858 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq);
1859
1860 mana_deinit_txq(apc, &apc->tx_qp[i].txq);
1861 }
1862
1863 kfree(apc->tx_qp);
1864 apc->tx_qp = NULL;
1865}
1866
1867static int mana_create_txq(struct mana_port_context *apc,
1868 struct net_device *net)
1869{
1870 struct mana_context *ac = apc->ac;
1871 struct gdma_dev *gd = ac->gdma_dev;
1872 struct mana_obj_spec wq_spec;
1873 struct mana_obj_spec cq_spec;
1874 struct gdma_queue_spec spec;
1875 struct gdma_context *gc;
1876 struct mana_txq *txq;
1877 struct mana_cq *cq;
1878 u32 txq_size;
1879 u32 cq_size;
1880 int err;
1881 int i;
1882
1883 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp),
1884 GFP_KERNEL);
1885 if (!apc->tx_qp)
1886 return -ENOMEM;
1887
1888 /* The minimum size of the WQE is 32 bytes, hence
1889 * MAX_SEND_BUFFERS_PER_QUEUE represents the maximum number of WQEs
1890 * the SQ can store. This value is then used to size other queues
1891 * to prevent overflow.
1892 */
1893 txq_size = MAX_SEND_BUFFERS_PER_QUEUE * 32;
1894 BUILD_BUG_ON(!PAGE_ALIGNED(txq_size));
1895
1896 cq_size = MAX_SEND_BUFFERS_PER_QUEUE * COMP_ENTRY_SIZE;
1897 cq_size = PAGE_ALIGN(cq_size);
1898
1899 gc = gd->gdma_context;
1900
1901 for (i = 0; i < apc->num_queues; i++) {
1902 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE;
1903
1904 /* Create SQ */
1905 txq = &apc->tx_qp[i].txq;
1906
1907 u64_stats_init(&txq->stats.syncp);
1908 txq->ndev = net;
1909 txq->net_txq = netdev_get_tx_queue(net, i);
1910 txq->vp_offset = apc->tx_vp_offset;
1911 skb_queue_head_init(&txq->pending_skbs);
1912
1913 memset(&spec, 0, sizeof(spec));
1914 spec.type = GDMA_SQ;
1915 spec.monitor_avl_buf = true;
1916 spec.queue_size = txq_size;
1917 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq);
1918 if (err)
1919 goto out;
1920
1921 /* Create SQ's CQ */
1922 cq = &apc->tx_qp[i].tx_cq;
1923 cq->type = MANA_CQ_TYPE_TX;
1924
1925 cq->txq = txq;
1926
1927 memset(&spec, 0, sizeof(spec));
1928 spec.type = GDMA_CQ;
1929 spec.monitor_avl_buf = false;
1930 spec.queue_size = cq_size;
1931 spec.cq.callback = mana_schedule_napi;
1932 spec.cq.parent_eq = ac->eqs[i].eq;
1933 spec.cq.context = cq;
1934 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
1935 if (err)
1936 goto out;
1937
1938 memset(&wq_spec, 0, sizeof(wq_spec));
1939 memset(&cq_spec, 0, sizeof(cq_spec));
1940
1941 wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle;
1942 wq_spec.queue_size = txq->gdma_sq->queue_size;
1943
1944 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
1945 cq_spec.queue_size = cq->gdma_cq->queue_size;
1946 cq_spec.modr_ctx_id = 0;
1947 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
1948
1949 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ,
1950 &wq_spec, &cq_spec,
1951 &apc->tx_qp[i].tx_object);
1952
1953 if (err)
1954 goto out;
1955
1956 txq->gdma_sq->id = wq_spec.queue_index;
1957 cq->gdma_cq->id = cq_spec.queue_index;
1958
1959 txq->gdma_sq->mem_info.dma_region_handle =
1960 GDMA_INVALID_DMA_REGION;
1961 cq->gdma_cq->mem_info.dma_region_handle =
1962 GDMA_INVALID_DMA_REGION;
1963
1964 txq->gdma_txq_id = txq->gdma_sq->id;
1965
1966 cq->gdma_id = cq->gdma_cq->id;
1967
1968 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
1969 err = -EINVAL;
1970 goto out;
1971 }
1972
1973 gc->cq_table[cq->gdma_id] = cq->gdma_cq;
1974
1975 netif_napi_add_tx(net, &cq->napi, mana_poll);
1976 napi_enable(&cq->napi);
1977
1978 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
1979 }
1980
1981 return 0;
1982out:
1983 mana_destroy_txq(apc);
1984 return err;
1985}
1986
1987static void mana_destroy_rxq(struct mana_port_context *apc,
1988 struct mana_rxq *rxq, bool validate_state)
1989
1990{
1991 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1992 struct mana_recv_buf_oob *rx_oob;
1993 struct device *dev = gc->dev;
1994 struct napi_struct *napi;
1995 struct page *page;
1996 int i;
1997
1998 if (!rxq)
1999 return;
2000
2001 napi = &rxq->rx_cq.napi;
2002
2003 if (validate_state)
2004 napi_synchronize(napi);
2005
2006 napi_disable(napi);
2007
2008 xdp_rxq_info_unreg(&rxq->xdp_rxq);
2009
2010 netif_napi_del(napi);
2011
2012 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj);
2013
2014 mana_deinit_cq(apc, &rxq->rx_cq);
2015
2016 if (rxq->xdp_save_va)
2017 put_page(virt_to_head_page(rxq->xdp_save_va));
2018
2019 for (i = 0; i < rxq->num_rx_buf; i++) {
2020 rx_oob = &rxq->rx_oobs[i];
2021
2022 if (!rx_oob->buf_va)
2023 continue;
2024
2025 dma_unmap_single(dev, rx_oob->sgl[0].address,
2026 rx_oob->sgl[0].size, DMA_FROM_DEVICE);
2027
2028 page = virt_to_head_page(rx_oob->buf_va);
2029
2030 if (rx_oob->from_pool)
2031 page_pool_put_full_page(rxq->page_pool, page, false);
2032 else
2033 put_page(page);
2034
2035 rx_oob->buf_va = NULL;
2036 }
2037
2038 page_pool_destroy(rxq->page_pool);
2039
2040 if (rxq->gdma_rq)
2041 mana_gd_destroy_queue(gc, rxq->gdma_rq);
2042
2043 kfree(rxq);
2044}
2045
2046static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key,
2047 struct mana_rxq *rxq, struct device *dev)
2048{
2049 struct mana_port_context *mpc = netdev_priv(rxq->ndev);
2050 bool from_pool = false;
2051 dma_addr_t da;
2052 void *va;
2053
2054 if (mpc->rxbufs_pre)
2055 va = mana_get_rxbuf_pre(rxq, &da);
2056 else
2057 va = mana_get_rxfrag(rxq, dev, &da, &from_pool, false);
2058
2059 if (!va)
2060 return -ENOMEM;
2061
2062 rx_oob->buf_va = va;
2063 rx_oob->from_pool = from_pool;
2064
2065 rx_oob->sgl[0].address = da;
2066 rx_oob->sgl[0].size = rxq->datasize;
2067 rx_oob->sgl[0].mem_key = mem_key;
2068
2069 return 0;
2070}
2071
2072#define MANA_WQE_HEADER_SIZE 16
2073#define MANA_WQE_SGE_SIZE 16
2074
2075static int mana_alloc_rx_wqe(struct mana_port_context *apc,
2076 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size)
2077{
2078 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
2079 struct mana_recv_buf_oob *rx_oob;
2080 struct device *dev = gc->dev;
2081 u32 buf_idx;
2082 int ret;
2083
2084 WARN_ON(rxq->datasize == 0);
2085
2086 *rxq_size = 0;
2087 *cq_size = 0;
2088
2089 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2090 rx_oob = &rxq->rx_oobs[buf_idx];
2091 memset(rx_oob, 0, sizeof(*rx_oob));
2092
2093 rx_oob->num_sge = 1;
2094
2095 ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq,
2096 dev);
2097 if (ret)
2098 return ret;
2099
2100 rx_oob->wqe_req.sgl = rx_oob->sgl;
2101 rx_oob->wqe_req.num_sge = rx_oob->num_sge;
2102 rx_oob->wqe_req.inline_oob_size = 0;
2103 rx_oob->wqe_req.inline_oob_data = NULL;
2104 rx_oob->wqe_req.flags = 0;
2105 rx_oob->wqe_req.client_data_unit = 0;
2106
2107 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE +
2108 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32);
2109 *cq_size += COMP_ENTRY_SIZE;
2110 }
2111
2112 return 0;
2113}
2114
2115static int mana_push_wqe(struct mana_rxq *rxq)
2116{
2117 struct mana_recv_buf_oob *rx_oob;
2118 u32 buf_idx;
2119 int err;
2120
2121 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2122 rx_oob = &rxq->rx_oobs[buf_idx];
2123
2124 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req,
2125 &rx_oob->wqe_inf);
2126 if (err)
2127 return -ENOSPC;
2128 }
2129
2130 return 0;
2131}
2132
2133static int mana_create_page_pool(struct mana_rxq *rxq, struct gdma_context *gc)
2134{
2135 struct page_pool_params pprm = {};
2136 int ret;
2137
2138 pprm.pool_size = RX_BUFFERS_PER_QUEUE;
2139 pprm.nid = gc->numa_node;
2140 pprm.napi = &rxq->rx_cq.napi;
2141 pprm.netdev = rxq->ndev;
2142
2143 rxq->page_pool = page_pool_create(&pprm);
2144
2145 if (IS_ERR(rxq->page_pool)) {
2146 ret = PTR_ERR(rxq->page_pool);
2147 rxq->page_pool = NULL;
2148 return ret;
2149 }
2150
2151 return 0;
2152}
2153
2154static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,
2155 u32 rxq_idx, struct mana_eq *eq,
2156 struct net_device *ndev)
2157{
2158 struct gdma_dev *gd = apc->ac->gdma_dev;
2159 struct mana_obj_spec wq_spec;
2160 struct mana_obj_spec cq_spec;
2161 struct gdma_queue_spec spec;
2162 struct mana_cq *cq = NULL;
2163 struct gdma_context *gc;
2164 u32 cq_size, rq_size;
2165 struct mana_rxq *rxq;
2166 int err;
2167
2168 gc = gd->gdma_context;
2169
2170 rxq = kzalloc(struct_size(rxq, rx_oobs, RX_BUFFERS_PER_QUEUE),
2171 GFP_KERNEL);
2172 if (!rxq)
2173 return NULL;
2174
2175 rxq->ndev = ndev;
2176 rxq->num_rx_buf = RX_BUFFERS_PER_QUEUE;
2177 rxq->rxq_idx = rxq_idx;
2178 rxq->rxobj = INVALID_MANA_HANDLE;
2179
2180 mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size,
2181 &rxq->headroom);
2182
2183 /* Create page pool for RX queue */
2184 err = mana_create_page_pool(rxq, gc);
2185 if (err) {
2186 netdev_err(ndev, "Create page pool err:%d\n", err);
2187 goto out;
2188 }
2189
2190 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size);
2191 if (err)
2192 goto out;
2193
2194 rq_size = PAGE_ALIGN(rq_size);
2195 cq_size = PAGE_ALIGN(cq_size);
2196
2197 /* Create RQ */
2198 memset(&spec, 0, sizeof(spec));
2199 spec.type = GDMA_RQ;
2200 spec.monitor_avl_buf = true;
2201 spec.queue_size = rq_size;
2202 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq);
2203 if (err)
2204 goto out;
2205
2206 /* Create RQ's CQ */
2207 cq = &rxq->rx_cq;
2208 cq->type = MANA_CQ_TYPE_RX;
2209 cq->rxq = rxq;
2210
2211 memset(&spec, 0, sizeof(spec));
2212 spec.type = GDMA_CQ;
2213 spec.monitor_avl_buf = false;
2214 spec.queue_size = cq_size;
2215 spec.cq.callback = mana_schedule_napi;
2216 spec.cq.parent_eq = eq->eq;
2217 spec.cq.context = cq;
2218 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2219 if (err)
2220 goto out;
2221
2222 memset(&wq_spec, 0, sizeof(wq_spec));
2223 memset(&cq_spec, 0, sizeof(cq_spec));
2224 wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle;
2225 wq_spec.queue_size = rxq->gdma_rq->queue_size;
2226
2227 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2228 cq_spec.queue_size = cq->gdma_cq->queue_size;
2229 cq_spec.modr_ctx_id = 0;
2230 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2231
2232 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ,
2233 &wq_spec, &cq_spec, &rxq->rxobj);
2234 if (err)
2235 goto out;
2236
2237 rxq->gdma_rq->id = wq_spec.queue_index;
2238 cq->gdma_cq->id = cq_spec.queue_index;
2239
2240 rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2241 cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2242
2243 rxq->gdma_id = rxq->gdma_rq->id;
2244 cq->gdma_id = cq->gdma_cq->id;
2245
2246 err = mana_push_wqe(rxq);
2247 if (err)
2248 goto out;
2249
2250 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
2251 err = -EINVAL;
2252 goto out;
2253 }
2254
2255 gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2256
2257 netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1);
2258
2259 WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx,
2260 cq->napi.napi_id));
2261 WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL,
2262 rxq->page_pool));
2263
2264 napi_enable(&cq->napi);
2265
2266 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2267out:
2268 if (!err)
2269 return rxq;
2270
2271 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err);
2272
2273 mana_destroy_rxq(apc, rxq, false);
2274
2275 if (cq)
2276 mana_deinit_cq(apc, cq);
2277
2278 return NULL;
2279}
2280
2281static int mana_add_rx_queues(struct mana_port_context *apc,
2282 struct net_device *ndev)
2283{
2284 struct mana_context *ac = apc->ac;
2285 struct mana_rxq *rxq;
2286 int err = 0;
2287 int i;
2288
2289 for (i = 0; i < apc->num_queues; i++) {
2290 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev);
2291 if (!rxq) {
2292 err = -ENOMEM;
2293 goto out;
2294 }
2295
2296 u64_stats_init(&rxq->stats.syncp);
2297
2298 apc->rxqs[i] = rxq;
2299 }
2300
2301 apc->default_rxobj = apc->rxqs[0]->rxobj;
2302out:
2303 return err;
2304}
2305
2306static void mana_destroy_vport(struct mana_port_context *apc)
2307{
2308 struct gdma_dev *gd = apc->ac->gdma_dev;
2309 struct mana_rxq *rxq;
2310 u32 rxq_idx;
2311
2312 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
2313 rxq = apc->rxqs[rxq_idx];
2314 if (!rxq)
2315 continue;
2316
2317 mana_destroy_rxq(apc, rxq, true);
2318 apc->rxqs[rxq_idx] = NULL;
2319 }
2320
2321 mana_destroy_txq(apc);
2322 mana_uncfg_vport(apc);
2323
2324 if (gd->gdma_context->is_pf)
2325 mana_pf_deregister_hw_vport(apc);
2326}
2327
2328static int mana_create_vport(struct mana_port_context *apc,
2329 struct net_device *net)
2330{
2331 struct gdma_dev *gd = apc->ac->gdma_dev;
2332 int err;
2333
2334 apc->default_rxobj = INVALID_MANA_HANDLE;
2335
2336 if (gd->gdma_context->is_pf) {
2337 err = mana_pf_register_hw_vport(apc);
2338 if (err)
2339 return err;
2340 }
2341
2342 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell);
2343 if (err)
2344 return err;
2345
2346 return mana_create_txq(apc, net);
2347}
2348
2349static void mana_rss_table_init(struct mana_port_context *apc)
2350{
2351 int i;
2352
2353 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++)
2354 apc->indir_table[i] =
2355 ethtool_rxfh_indir_default(i, apc->num_queues);
2356}
2357
2358int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx,
2359 bool update_hash, bool update_tab)
2360{
2361 u32 queue_idx;
2362 int err;
2363 int i;
2364
2365 if (update_tab) {
2366 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) {
2367 queue_idx = apc->indir_table[i];
2368 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj;
2369 }
2370 }
2371
2372 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab);
2373 if (err)
2374 return err;
2375
2376 mana_fence_rqs(apc);
2377
2378 return 0;
2379}
2380
2381void mana_query_gf_stats(struct mana_port_context *apc)
2382{
2383 struct mana_query_gf_stat_resp resp = {};
2384 struct mana_query_gf_stat_req req = {};
2385 struct net_device *ndev = apc->ndev;
2386 int err;
2387
2388 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_GF_STAT,
2389 sizeof(req), sizeof(resp));
2390 req.req_stats = STATISTICS_FLAGS_RX_DISCARDS_NO_WQE |
2391 STATISTICS_FLAGS_RX_ERRORS_VPORT_DISABLED |
2392 STATISTICS_FLAGS_HC_RX_BYTES |
2393 STATISTICS_FLAGS_HC_RX_UCAST_PACKETS |
2394 STATISTICS_FLAGS_HC_RX_UCAST_BYTES |
2395 STATISTICS_FLAGS_HC_RX_MCAST_PACKETS |
2396 STATISTICS_FLAGS_HC_RX_MCAST_BYTES |
2397 STATISTICS_FLAGS_HC_RX_BCAST_PACKETS |
2398 STATISTICS_FLAGS_HC_RX_BCAST_BYTES |
2399 STATISTICS_FLAGS_TX_ERRORS_GF_DISABLED |
2400 STATISTICS_FLAGS_TX_ERRORS_VPORT_DISABLED |
2401 STATISTICS_FLAGS_TX_ERRORS_INVAL_VPORT_OFFSET_PACKETS |
2402 STATISTICS_FLAGS_TX_ERRORS_VLAN_ENFORCEMENT |
2403 STATISTICS_FLAGS_TX_ERRORS_ETH_TYPE_ENFORCEMENT |
2404 STATISTICS_FLAGS_TX_ERRORS_SA_ENFORCEMENT |
2405 STATISTICS_FLAGS_TX_ERRORS_SQPDID_ENFORCEMENT |
2406 STATISTICS_FLAGS_TX_ERRORS_CQPDID_ENFORCEMENT |
2407 STATISTICS_FLAGS_TX_ERRORS_MTU_VIOLATION |
2408 STATISTICS_FLAGS_TX_ERRORS_INVALID_OOB |
2409 STATISTICS_FLAGS_HC_TX_BYTES |
2410 STATISTICS_FLAGS_HC_TX_UCAST_PACKETS |
2411 STATISTICS_FLAGS_HC_TX_UCAST_BYTES |
2412 STATISTICS_FLAGS_HC_TX_MCAST_PACKETS |
2413 STATISTICS_FLAGS_HC_TX_MCAST_BYTES |
2414 STATISTICS_FLAGS_HC_TX_BCAST_PACKETS |
2415 STATISTICS_FLAGS_HC_TX_BCAST_BYTES |
2416 STATISTICS_FLAGS_TX_ERRORS_GDMA_ERROR;
2417
2418 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
2419 sizeof(resp));
2420 if (err) {
2421 netdev_err(ndev, "Failed to query GF stats: %d\n", err);
2422 return;
2423 }
2424 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_GF_STAT,
2425 sizeof(resp));
2426 if (err || resp.hdr.status) {
2427 netdev_err(ndev, "Failed to query GF stats: %d, 0x%x\n", err,
2428 resp.hdr.status);
2429 return;
2430 }
2431
2432 apc->eth_stats.hc_rx_discards_no_wqe = resp.rx_discards_nowqe;
2433 apc->eth_stats.hc_rx_err_vport_disabled = resp.rx_err_vport_disabled;
2434 apc->eth_stats.hc_rx_bytes = resp.hc_rx_bytes;
2435 apc->eth_stats.hc_rx_ucast_pkts = resp.hc_rx_ucast_pkts;
2436 apc->eth_stats.hc_rx_ucast_bytes = resp.hc_rx_ucast_bytes;
2437 apc->eth_stats.hc_rx_bcast_pkts = resp.hc_rx_bcast_pkts;
2438 apc->eth_stats.hc_rx_bcast_bytes = resp.hc_rx_bcast_bytes;
2439 apc->eth_stats.hc_rx_mcast_pkts = resp.hc_rx_mcast_pkts;
2440 apc->eth_stats.hc_rx_mcast_bytes = resp.hc_rx_mcast_bytes;
2441 apc->eth_stats.hc_tx_err_gf_disabled = resp.tx_err_gf_disabled;
2442 apc->eth_stats.hc_tx_err_vport_disabled = resp.tx_err_vport_disabled;
2443 apc->eth_stats.hc_tx_err_inval_vportoffset_pkt =
2444 resp.tx_err_inval_vport_offset_pkt;
2445 apc->eth_stats.hc_tx_err_vlan_enforcement =
2446 resp.tx_err_vlan_enforcement;
2447 apc->eth_stats.hc_tx_err_eth_type_enforcement =
2448 resp.tx_err_ethtype_enforcement;
2449 apc->eth_stats.hc_tx_err_sa_enforcement = resp.tx_err_SA_enforcement;
2450 apc->eth_stats.hc_tx_err_sqpdid_enforcement =
2451 resp.tx_err_SQPDID_enforcement;
2452 apc->eth_stats.hc_tx_err_cqpdid_enforcement =
2453 resp.tx_err_CQPDID_enforcement;
2454 apc->eth_stats.hc_tx_err_mtu_violation = resp.tx_err_mtu_violation;
2455 apc->eth_stats.hc_tx_err_inval_oob = resp.tx_err_inval_oob;
2456 apc->eth_stats.hc_tx_bytes = resp.hc_tx_bytes;
2457 apc->eth_stats.hc_tx_ucast_pkts = resp.hc_tx_ucast_pkts;
2458 apc->eth_stats.hc_tx_ucast_bytes = resp.hc_tx_ucast_bytes;
2459 apc->eth_stats.hc_tx_bcast_pkts = resp.hc_tx_bcast_pkts;
2460 apc->eth_stats.hc_tx_bcast_bytes = resp.hc_tx_bcast_bytes;
2461 apc->eth_stats.hc_tx_mcast_pkts = resp.hc_tx_mcast_pkts;
2462 apc->eth_stats.hc_tx_mcast_bytes = resp.hc_tx_mcast_bytes;
2463 apc->eth_stats.hc_tx_err_gdma = resp.tx_err_gdma;
2464}
2465
2466static int mana_init_port(struct net_device *ndev)
2467{
2468 struct mana_port_context *apc = netdev_priv(ndev);
2469 u32 max_txq, max_rxq, max_queues;
2470 int port_idx = apc->port_idx;
2471 u32 num_indirect_entries;
2472 int err;
2473
2474 err = mana_init_port_context(apc);
2475 if (err)
2476 return err;
2477
2478 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq,
2479 &num_indirect_entries);
2480 if (err) {
2481 netdev_err(ndev, "Failed to query info for vPort %d\n",
2482 port_idx);
2483 goto reset_apc;
2484 }
2485
2486 max_queues = min_t(u32, max_txq, max_rxq);
2487 if (apc->max_queues > max_queues)
2488 apc->max_queues = max_queues;
2489
2490 if (apc->num_queues > apc->max_queues)
2491 apc->num_queues = apc->max_queues;
2492
2493 eth_hw_addr_set(ndev, apc->mac_addr);
2494
2495 return 0;
2496
2497reset_apc:
2498 kfree(apc->rxqs);
2499 apc->rxqs = NULL;
2500 return err;
2501}
2502
2503int mana_alloc_queues(struct net_device *ndev)
2504{
2505 struct mana_port_context *apc = netdev_priv(ndev);
2506 struct gdma_dev *gd = apc->ac->gdma_dev;
2507 int err;
2508
2509 err = mana_create_vport(apc, ndev);
2510 if (err)
2511 return err;
2512
2513 err = netif_set_real_num_tx_queues(ndev, apc->num_queues);
2514 if (err)
2515 goto destroy_vport;
2516
2517 err = mana_add_rx_queues(apc, ndev);
2518 if (err)
2519 goto destroy_vport;
2520
2521 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE;
2522
2523 err = netif_set_real_num_rx_queues(ndev, apc->num_queues);
2524 if (err)
2525 goto destroy_vport;
2526
2527 mana_rss_table_init(apc);
2528
2529 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true);
2530 if (err)
2531 goto destroy_vport;
2532
2533 if (gd->gdma_context->is_pf) {
2534 err = mana_pf_register_filter(apc);
2535 if (err)
2536 goto destroy_vport;
2537 }
2538
2539 mana_chn_setxdp(apc, mana_xdp_get(apc));
2540
2541 return 0;
2542
2543destroy_vport:
2544 mana_destroy_vport(apc);
2545 return err;
2546}
2547
2548int mana_attach(struct net_device *ndev)
2549{
2550 struct mana_port_context *apc = netdev_priv(ndev);
2551 int err;
2552
2553 ASSERT_RTNL();
2554
2555 err = mana_init_port(ndev);
2556 if (err)
2557 return err;
2558
2559 if (apc->port_st_save) {
2560 err = mana_alloc_queues(ndev);
2561 if (err) {
2562 mana_cleanup_port_context(apc);
2563 return err;
2564 }
2565 }
2566
2567 apc->port_is_up = apc->port_st_save;
2568
2569 /* Ensure port state updated before txq state */
2570 smp_wmb();
2571
2572 if (apc->port_is_up)
2573 netif_carrier_on(ndev);
2574
2575 netif_device_attach(ndev);
2576
2577 return 0;
2578}
2579
2580static int mana_dealloc_queues(struct net_device *ndev)
2581{
2582 struct mana_port_context *apc = netdev_priv(ndev);
2583 unsigned long timeout = jiffies + 120 * HZ;
2584 struct gdma_dev *gd = apc->ac->gdma_dev;
2585 struct mana_txq *txq;
2586 struct sk_buff *skb;
2587 int i, err;
2588 u32 tsleep;
2589
2590 if (apc->port_is_up)
2591 return -EINVAL;
2592
2593 mana_chn_setxdp(apc, NULL);
2594
2595 if (gd->gdma_context->is_pf)
2596 mana_pf_deregister_filter(apc);
2597
2598 /* No packet can be transmitted now since apc->port_is_up is false.
2599 * There is still a tiny chance that mana_poll_tx_cq() can re-enable
2600 * a txq because it may not timely see apc->port_is_up being cleared
2601 * to false, but it doesn't matter since mana_start_xmit() drops any
2602 * new packets due to apc->port_is_up being false.
2603 *
2604 * Drain all the in-flight TX packets.
2605 * A timeout of 120 seconds for all the queues is used.
2606 * This will break the while loop when h/w is not responding.
2607 * This value of 120 has been decided here considering max
2608 * number of queues.
2609 */
2610
2611 for (i = 0; i < apc->num_queues; i++) {
2612 txq = &apc->tx_qp[i].txq;
2613 tsleep = 1000;
2614 while (atomic_read(&txq->pending_sends) > 0 &&
2615 time_before(jiffies, timeout)) {
2616 usleep_range(tsleep, tsleep + 1000);
2617 tsleep <<= 1;
2618 }
2619 if (atomic_read(&txq->pending_sends)) {
2620 err = pcie_flr(to_pci_dev(gd->gdma_context->dev));
2621 if (err) {
2622 netdev_err(ndev, "flr failed %d with %d pkts pending in txq %u\n",
2623 err, atomic_read(&txq->pending_sends),
2624 txq->gdma_txq_id);
2625 }
2626 break;
2627 }
2628 }
2629
2630 for (i = 0; i < apc->num_queues; i++) {
2631 txq = &apc->tx_qp[i].txq;
2632 while ((skb = skb_dequeue(&txq->pending_skbs))) {
2633 mana_unmap_skb(skb, apc);
2634 dev_kfree_skb_any(skb);
2635 }
2636 atomic_set(&txq->pending_sends, 0);
2637 }
2638 /* We're 100% sure the queues can no longer be woken up, because
2639 * we're sure now mana_poll_tx_cq() can't be running.
2640 */
2641
2642 apc->rss_state = TRI_STATE_FALSE;
2643 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false);
2644 if (err) {
2645 netdev_err(ndev, "Failed to disable vPort: %d\n", err);
2646 return err;
2647 }
2648
2649 mana_destroy_vport(apc);
2650
2651 return 0;
2652}
2653
2654int mana_detach(struct net_device *ndev, bool from_close)
2655{
2656 struct mana_port_context *apc = netdev_priv(ndev);
2657 int err;
2658
2659 ASSERT_RTNL();
2660
2661 apc->port_st_save = apc->port_is_up;
2662 apc->port_is_up = false;
2663
2664 /* Ensure port state updated before txq state */
2665 smp_wmb();
2666
2667 netif_tx_disable(ndev);
2668 netif_carrier_off(ndev);
2669
2670 if (apc->port_st_save) {
2671 err = mana_dealloc_queues(ndev);
2672 if (err)
2673 return err;
2674 }
2675
2676 if (!from_close) {
2677 netif_device_detach(ndev);
2678 mana_cleanup_port_context(apc);
2679 }
2680
2681 return 0;
2682}
2683
2684static int mana_probe_port(struct mana_context *ac, int port_idx,
2685 struct net_device **ndev_storage)
2686{
2687 struct gdma_context *gc = ac->gdma_dev->gdma_context;
2688 struct mana_port_context *apc;
2689 struct net_device *ndev;
2690 int err;
2691
2692 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context),
2693 gc->max_num_queues);
2694 if (!ndev)
2695 return -ENOMEM;
2696
2697 *ndev_storage = ndev;
2698
2699 apc = netdev_priv(ndev);
2700 apc->ac = ac;
2701 apc->ndev = ndev;
2702 apc->max_queues = gc->max_num_queues;
2703 apc->num_queues = gc->max_num_queues;
2704 apc->port_handle = INVALID_MANA_HANDLE;
2705 apc->pf_filter_handle = INVALID_MANA_HANDLE;
2706 apc->port_idx = port_idx;
2707
2708 mutex_init(&apc->vport_mutex);
2709 apc->vport_use_count = 0;
2710
2711 ndev->netdev_ops = &mana_devops;
2712 ndev->ethtool_ops = &mana_ethtool_ops;
2713 ndev->mtu = ETH_DATA_LEN;
2714 ndev->max_mtu = gc->adapter_mtu - ETH_HLEN;
2715 ndev->min_mtu = ETH_MIN_MTU;
2716 ndev->needed_headroom = MANA_HEADROOM;
2717 ndev->dev_port = port_idx;
2718 SET_NETDEV_DEV(ndev, gc->dev);
2719
2720 netif_carrier_off(ndev);
2721
2722 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE);
2723
2724 err = mana_init_port(ndev);
2725 if (err)
2726 goto free_net;
2727
2728 netdev_lockdep_set_classes(ndev);
2729
2730 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
2731 ndev->hw_features |= NETIF_F_RXCSUM;
2732 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
2733 ndev->hw_features |= NETIF_F_RXHASH;
2734 ndev->features = ndev->hw_features | NETIF_F_HW_VLAN_CTAG_TX |
2735 NETIF_F_HW_VLAN_CTAG_RX;
2736 ndev->vlan_features = ndev->features;
2737 xdp_set_features_flag(ndev, NETDEV_XDP_ACT_BASIC |
2738 NETDEV_XDP_ACT_REDIRECT |
2739 NETDEV_XDP_ACT_NDO_XMIT);
2740
2741 err = register_netdev(ndev);
2742 if (err) {
2743 netdev_err(ndev, "Unable to register netdev.\n");
2744 goto reset_apc;
2745 }
2746
2747 return 0;
2748
2749reset_apc:
2750 kfree(apc->rxqs);
2751 apc->rxqs = NULL;
2752free_net:
2753 *ndev_storage = NULL;
2754 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err);
2755 free_netdev(ndev);
2756 return err;
2757}
2758
2759static void adev_release(struct device *dev)
2760{
2761 struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev);
2762
2763 kfree(madev);
2764}
2765
2766static void remove_adev(struct gdma_dev *gd)
2767{
2768 struct auxiliary_device *adev = gd->adev;
2769 int id = adev->id;
2770
2771 auxiliary_device_delete(adev);
2772 auxiliary_device_uninit(adev);
2773
2774 mana_adev_idx_free(id);
2775 gd->adev = NULL;
2776}
2777
2778static int add_adev(struct gdma_dev *gd)
2779{
2780 struct auxiliary_device *adev;
2781 struct mana_adev *madev;
2782 int ret;
2783
2784 madev = kzalloc(sizeof(*madev), GFP_KERNEL);
2785 if (!madev)
2786 return -ENOMEM;
2787
2788 adev = &madev->adev;
2789 ret = mana_adev_idx_alloc();
2790 if (ret < 0)
2791 goto idx_fail;
2792 adev->id = ret;
2793
2794 adev->name = "rdma";
2795 adev->dev.parent = gd->gdma_context->dev;
2796 adev->dev.release = adev_release;
2797 madev->mdev = gd;
2798
2799 ret = auxiliary_device_init(adev);
2800 if (ret)
2801 goto init_fail;
2802
2803 ret = auxiliary_device_add(adev);
2804 if (ret)
2805 goto add_fail;
2806
2807 gd->adev = adev;
2808 return 0;
2809
2810add_fail:
2811 auxiliary_device_uninit(adev);
2812
2813init_fail:
2814 mana_adev_idx_free(adev->id);
2815
2816idx_fail:
2817 kfree(madev);
2818
2819 return ret;
2820}
2821
2822int mana_probe(struct gdma_dev *gd, bool resuming)
2823{
2824 struct gdma_context *gc = gd->gdma_context;
2825 struct mana_context *ac = gd->driver_data;
2826 struct device *dev = gc->dev;
2827 u16 num_ports = 0;
2828 int err;
2829 int i;
2830
2831 dev_info(dev,
2832 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n",
2833 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION);
2834
2835 err = mana_gd_register_device(gd);
2836 if (err)
2837 return err;
2838
2839 if (!resuming) {
2840 ac = kzalloc(sizeof(*ac), GFP_KERNEL);
2841 if (!ac)
2842 return -ENOMEM;
2843
2844 ac->gdma_dev = gd;
2845 gd->driver_data = ac;
2846 }
2847
2848 err = mana_create_eq(ac);
2849 if (err)
2850 goto out;
2851
2852 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION,
2853 MANA_MICRO_VERSION, &num_ports);
2854 if (err)
2855 goto out;
2856
2857 if (!resuming) {
2858 ac->num_ports = num_ports;
2859 } else {
2860 if (ac->num_ports != num_ports) {
2861 dev_err(dev, "The number of vPorts changed: %d->%d\n",
2862 ac->num_ports, num_ports);
2863 err = -EPROTO;
2864 goto out;
2865 }
2866 }
2867
2868 if (ac->num_ports == 0)
2869 dev_err(dev, "Failed to detect any vPort\n");
2870
2871 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV)
2872 ac->num_ports = MAX_PORTS_IN_MANA_DEV;
2873
2874 if (!resuming) {
2875 for (i = 0; i < ac->num_ports; i++) {
2876 err = mana_probe_port(ac, i, &ac->ports[i]);
2877 if (err)
2878 break;
2879 }
2880 } else {
2881 for (i = 0; i < ac->num_ports; i++) {
2882 rtnl_lock();
2883 err = mana_attach(ac->ports[i]);
2884 rtnl_unlock();
2885 if (err)
2886 break;
2887 }
2888 }
2889
2890 err = add_adev(gd);
2891out:
2892 if (err)
2893 mana_remove(gd, false);
2894
2895 return err;
2896}
2897
2898void mana_remove(struct gdma_dev *gd, bool suspending)
2899{
2900 struct gdma_context *gc = gd->gdma_context;
2901 struct mana_context *ac = gd->driver_data;
2902 struct device *dev = gc->dev;
2903 struct net_device *ndev;
2904 int err;
2905 int i;
2906
2907 /* adev currently doesn't support suspending, always remove it */
2908 if (gd->adev)
2909 remove_adev(gd);
2910
2911 for (i = 0; i < ac->num_ports; i++) {
2912 ndev = ac->ports[i];
2913 if (!ndev) {
2914 if (i == 0)
2915 dev_err(dev, "No net device to remove\n");
2916 goto out;
2917 }
2918
2919 /* All cleanup actions should stay after rtnl_lock(), otherwise
2920 * other functions may access partially cleaned up data.
2921 */
2922 rtnl_lock();
2923
2924 err = mana_detach(ndev, false);
2925 if (err)
2926 netdev_err(ndev, "Failed to detach vPort %d: %d\n",
2927 i, err);
2928
2929 if (suspending) {
2930 /* No need to unregister the ndev. */
2931 rtnl_unlock();
2932 continue;
2933 }
2934
2935 unregister_netdevice(ndev);
2936
2937 rtnl_unlock();
2938
2939 free_netdev(ndev);
2940 }
2941
2942 mana_destroy_eq(ac);
2943out:
2944 mana_gd_deregister_device(gd);
2945
2946 if (suspending)
2947 return;
2948
2949 gd->driver_data = NULL;
2950 gd->gdma_context = NULL;
2951 kfree(ac);
2952}