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 #include <linux/export.h>
14
15 #include <net/checksum.h>
16 #include <net/ip6_checksum.h>
17 #include <net/netdev_lock.h>
18 #include <net/page_pool/helpers.h>
19 #include <net/xdp.h>
20
21 #include <net/mana/mana.h>
22 #include <net/mana/mana_auxiliary.h>
23
24 static DEFINE_IDA(mana_adev_ida);
25
mana_adev_idx_alloc(void)26 static int mana_adev_idx_alloc(void)
27 {
28 return ida_alloc(&mana_adev_ida, GFP_KERNEL);
29 }
30
mana_adev_idx_free(int idx)31 static void mana_adev_idx_free(int idx)
32 {
33 ida_free(&mana_adev_ida, idx);
34 }
35
mana_dbg_q_read(struct file * filp,char __user * buf,size_t count,loff_t * pos)36 static ssize_t mana_dbg_q_read(struct file *filp, char __user *buf, size_t count,
37 loff_t *pos)
38 {
39 struct gdma_queue *gdma_q = filp->private_data;
40
41 return simple_read_from_buffer(buf, count, pos, gdma_q->queue_mem_ptr,
42 gdma_q->queue_size);
43 }
44
45 static const struct file_operations mana_dbg_q_fops = {
46 .owner = THIS_MODULE,
47 .open = simple_open,
48 .read = mana_dbg_q_read,
49 };
50
51 /* Microsoft Azure Network Adapter (MANA) functions */
52
mana_open(struct net_device * ndev)53 static int mana_open(struct net_device *ndev)
54 {
55 struct mana_port_context *apc = netdev_priv(ndev);
56 int err;
57 err = mana_alloc_queues(ndev);
58
59 if (err) {
60 netdev_err(ndev, "%s failed to allocate queues: %d\n", __func__, err);
61 return err;
62 }
63
64 apc->port_is_up = true;
65
66 /* Ensure port state updated before txq state */
67 smp_wmb();
68
69 netif_carrier_on(ndev);
70 netif_tx_wake_all_queues(ndev);
71 netdev_dbg(ndev, "%s successful\n", __func__);
72 return 0;
73 }
74
mana_close(struct net_device * ndev)75 static int mana_close(struct net_device *ndev)
76 {
77 struct mana_port_context *apc = netdev_priv(ndev);
78
79 if (!apc->port_is_up)
80 return 0;
81
82 return mana_detach(ndev, true);
83 }
84
mana_can_tx(struct gdma_queue * wq)85 static bool mana_can_tx(struct gdma_queue *wq)
86 {
87 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE;
88 }
89
mana_checksum_info(struct sk_buff * skb)90 static unsigned int mana_checksum_info(struct sk_buff *skb)
91 {
92 if (skb->protocol == htons(ETH_P_IP)) {
93 struct iphdr *ip = ip_hdr(skb);
94
95 if (ip->protocol == IPPROTO_TCP)
96 return IPPROTO_TCP;
97
98 if (ip->protocol == IPPROTO_UDP)
99 return IPPROTO_UDP;
100 } else if (skb->protocol == htons(ETH_P_IPV6)) {
101 struct ipv6hdr *ip6 = ipv6_hdr(skb);
102
103 if (ip6->nexthdr == IPPROTO_TCP)
104 return IPPROTO_TCP;
105
106 if (ip6->nexthdr == IPPROTO_UDP)
107 return IPPROTO_UDP;
108 }
109
110 /* No csum offloading */
111 return 0;
112 }
113
mana_add_sge(struct mana_tx_package * tp,struct mana_skb_head * ash,int sg_i,dma_addr_t da,int sge_len,u32 gpa_mkey)114 static void mana_add_sge(struct mana_tx_package *tp, struct mana_skb_head *ash,
115 int sg_i, dma_addr_t da, int sge_len, u32 gpa_mkey)
116 {
117 ash->dma_handle[sg_i] = da;
118 ash->size[sg_i] = sge_len;
119
120 tp->wqe_req.sgl[sg_i].address = da;
121 tp->wqe_req.sgl[sg_i].mem_key = gpa_mkey;
122 tp->wqe_req.sgl[sg_i].size = sge_len;
123 }
124
mana_map_skb(struct sk_buff * skb,struct mana_port_context * apc,struct mana_tx_package * tp,int gso_hs)125 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc,
126 struct mana_tx_package *tp, int gso_hs)
127 {
128 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
129 int hsg = 1; /* num of SGEs of linear part */
130 struct gdma_dev *gd = apc->ac->gdma_dev;
131 int skb_hlen = skb_headlen(skb);
132 int sge0_len, sge1_len = 0;
133 struct gdma_context *gc;
134 struct device *dev;
135 skb_frag_t *frag;
136 dma_addr_t da;
137 int sg_i;
138 int i;
139
140 gc = gd->gdma_context;
141 dev = gc->dev;
142
143 if (gso_hs && gso_hs < skb_hlen) {
144 sge0_len = gso_hs;
145 sge1_len = skb_hlen - gso_hs;
146 } else {
147 sge0_len = skb_hlen;
148 }
149
150 da = dma_map_single(dev, skb->data, sge0_len, DMA_TO_DEVICE);
151 if (dma_mapping_error(dev, da))
152 return -ENOMEM;
153
154 mana_add_sge(tp, ash, 0, da, sge0_len, gd->gpa_mkey);
155
156 if (sge1_len) {
157 sg_i = 1;
158 da = dma_map_single(dev, skb->data + sge0_len, sge1_len,
159 DMA_TO_DEVICE);
160 if (dma_mapping_error(dev, da))
161 goto frag_err;
162
163 mana_add_sge(tp, ash, sg_i, da, sge1_len, gd->gpa_mkey);
164 hsg = 2;
165 }
166
167 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
168 sg_i = hsg + i;
169
170 frag = &skb_shinfo(skb)->frags[i];
171 da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag),
172 DMA_TO_DEVICE);
173 if (dma_mapping_error(dev, da))
174 goto frag_err;
175
176 mana_add_sge(tp, ash, sg_i, da, skb_frag_size(frag),
177 gd->gpa_mkey);
178 }
179
180 return 0;
181
182 frag_err:
183 if (net_ratelimit())
184 netdev_err(apc->ndev, "Failed to map skb of size %u to DMA\n",
185 skb->len);
186 for (i = sg_i - 1; i >= hsg; i--)
187 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
188 DMA_TO_DEVICE);
189
190 for (i = hsg - 1; i >= 0; i--)
191 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i],
192 DMA_TO_DEVICE);
193
194 return -ENOMEM;
195 }
196
197 /* Handle the case when GSO SKB linear length is too large.
198 * MANA NIC requires GSO packets to put only the packet header to SGE0.
199 * So, we need 2 SGEs for the skb linear part which contains more than the
200 * header.
201 * Return a positive value for the number of SGEs, or a negative value
202 * for an error.
203 */
mana_fix_skb_head(struct net_device * ndev,struct sk_buff * skb,int gso_hs)204 static int mana_fix_skb_head(struct net_device *ndev, struct sk_buff *skb,
205 int gso_hs)
206 {
207 int num_sge = 1 + skb_shinfo(skb)->nr_frags;
208 int skb_hlen = skb_headlen(skb);
209
210 if (gso_hs < skb_hlen) {
211 num_sge++;
212 } else if (gso_hs > skb_hlen) {
213 if (net_ratelimit())
214 netdev_err(ndev,
215 "TX nonlinear head: hs:%d, skb_hlen:%d\n",
216 gso_hs, skb_hlen);
217
218 return -EINVAL;
219 }
220
221 return num_sge;
222 }
223
224 /* Get the GSO packet's header size */
mana_get_gso_hs(struct sk_buff * skb)225 static int mana_get_gso_hs(struct sk_buff *skb)
226 {
227 int gso_hs;
228
229 if (skb->encapsulation) {
230 gso_hs = skb_inner_tcp_all_headers(skb);
231 } else {
232 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
233 gso_hs = skb_transport_offset(skb) +
234 sizeof(struct udphdr);
235 } else {
236 gso_hs = skb_tcp_all_headers(skb);
237 }
238 }
239
240 return gso_hs;
241 }
242
mana_start_xmit(struct sk_buff * skb,struct net_device * ndev)243 netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev)
244 {
245 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT;
246 struct mana_port_context *apc = netdev_priv(ndev);
247 int gso_hs = 0; /* zero for non-GSO pkts */
248 u16 txq_idx = skb_get_queue_mapping(skb);
249 struct gdma_dev *gd = apc->ac->gdma_dev;
250 bool ipv4 = false, ipv6 = false;
251 struct mana_tx_package pkg = {};
252 struct netdev_queue *net_txq;
253 struct mana_stats_tx *tx_stats;
254 struct gdma_queue *gdma_sq;
255 unsigned int csum_type;
256 struct mana_txq *txq;
257 struct mana_cq *cq;
258 int err, len;
259
260 if (unlikely(!apc->port_is_up))
261 goto tx_drop;
262
263 if (skb_cow_head(skb, MANA_HEADROOM))
264 goto tx_drop_count;
265
266 if (unlikely(ipv6_hopopt_jumbo_remove(skb)))
267 goto tx_drop_count;
268
269 txq = &apc->tx_qp[txq_idx].txq;
270 gdma_sq = txq->gdma_sq;
271 cq = &apc->tx_qp[txq_idx].tx_cq;
272 tx_stats = &txq->stats;
273
274 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id;
275 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame;
276
277 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) {
278 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset;
279 pkt_fmt = MANA_LONG_PKT_FMT;
280 } else {
281 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset;
282 }
283
284 if (skb_vlan_tag_present(skb)) {
285 pkt_fmt = MANA_LONG_PKT_FMT;
286 pkg.tx_oob.l_oob.inject_vlan_pri_tag = 1;
287 pkg.tx_oob.l_oob.pcp = skb_vlan_tag_get_prio(skb);
288 pkg.tx_oob.l_oob.dei = skb_vlan_tag_get_cfi(skb);
289 pkg.tx_oob.l_oob.vlan_id = skb_vlan_tag_get_id(skb);
290 }
291
292 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt;
293
294 if (pkt_fmt == MANA_SHORT_PKT_FMT) {
295 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob);
296 u64_stats_update_begin(&tx_stats->syncp);
297 tx_stats->short_pkt_fmt++;
298 u64_stats_update_end(&tx_stats->syncp);
299 } else {
300 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob);
301 u64_stats_update_begin(&tx_stats->syncp);
302 tx_stats->long_pkt_fmt++;
303 u64_stats_update_end(&tx_stats->syncp);
304 }
305
306 pkg.wqe_req.inline_oob_data = &pkg.tx_oob;
307 pkg.wqe_req.flags = 0;
308 pkg.wqe_req.client_data_unit = 0;
309
310 pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags;
311
312 if (skb->protocol == htons(ETH_P_IP))
313 ipv4 = true;
314 else if (skb->protocol == htons(ETH_P_IPV6))
315 ipv6 = true;
316
317 if (skb_is_gso(skb)) {
318 int num_sge;
319
320 gso_hs = mana_get_gso_hs(skb);
321
322 num_sge = mana_fix_skb_head(ndev, skb, gso_hs);
323 if (num_sge > 0)
324 pkg.wqe_req.num_sge = num_sge;
325 else
326 goto tx_drop_count;
327
328 u64_stats_update_begin(&tx_stats->syncp);
329 if (skb->encapsulation) {
330 tx_stats->tso_inner_packets++;
331 tx_stats->tso_inner_bytes += skb->len - gso_hs;
332 } else {
333 tx_stats->tso_packets++;
334 tx_stats->tso_bytes += skb->len - gso_hs;
335 }
336 u64_stats_update_end(&tx_stats->syncp);
337
338 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
339 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
340
341 pkg.tx_oob.s_oob.comp_iphdr_csum = 1;
342 pkg.tx_oob.s_oob.comp_tcp_csum = 1;
343 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
344
345 pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size;
346 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0;
347 if (ipv4) {
348 ip_hdr(skb)->tot_len = 0;
349 ip_hdr(skb)->check = 0;
350 tcp_hdr(skb)->check =
351 ~csum_tcpudp_magic(ip_hdr(skb)->saddr,
352 ip_hdr(skb)->daddr, 0,
353 IPPROTO_TCP, 0);
354 } else {
355 ipv6_hdr(skb)->payload_len = 0;
356 tcp_hdr(skb)->check =
357 ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
358 &ipv6_hdr(skb)->daddr, 0,
359 IPPROTO_TCP, 0);
360 }
361 } else if (skb->ip_summed == CHECKSUM_PARTIAL) {
362 csum_type = mana_checksum_info(skb);
363
364 u64_stats_update_begin(&tx_stats->syncp);
365 tx_stats->csum_partial++;
366 u64_stats_update_end(&tx_stats->syncp);
367
368 if (csum_type == IPPROTO_TCP) {
369 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
370 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
371
372 pkg.tx_oob.s_oob.comp_tcp_csum = 1;
373 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
374
375 } else if (csum_type == IPPROTO_UDP) {
376 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
377 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
378
379 pkg.tx_oob.s_oob.comp_udp_csum = 1;
380 } else {
381 /* Can't do offload of this type of checksum */
382 if (skb_checksum_help(skb))
383 goto tx_drop_count;
384 }
385 }
386
387 WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES);
388
389 if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) {
390 pkg.wqe_req.sgl = pkg.sgl_array;
391 } else {
392 pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge,
393 sizeof(struct gdma_sge),
394 GFP_ATOMIC);
395 if (!pkg.sgl_ptr)
396 goto tx_drop_count;
397
398 pkg.wqe_req.sgl = pkg.sgl_ptr;
399 }
400
401 if (mana_map_skb(skb, apc, &pkg, gso_hs)) {
402 u64_stats_update_begin(&tx_stats->syncp);
403 tx_stats->mana_map_err++;
404 u64_stats_update_end(&tx_stats->syncp);
405 goto free_sgl_ptr;
406 }
407
408 skb_queue_tail(&txq->pending_skbs, skb);
409
410 len = skb->len;
411 net_txq = netdev_get_tx_queue(ndev, txq_idx);
412
413 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req,
414 (struct gdma_posted_wqe_info *)skb->cb);
415 if (!mana_can_tx(gdma_sq)) {
416 netif_tx_stop_queue(net_txq);
417 apc->eth_stats.stop_queue++;
418 }
419
420 if (err) {
421 (void)skb_dequeue_tail(&txq->pending_skbs);
422 netdev_warn(ndev, "Failed to post TX OOB: %d\n", err);
423 err = NETDEV_TX_BUSY;
424 goto tx_busy;
425 }
426
427 err = NETDEV_TX_OK;
428 atomic_inc(&txq->pending_sends);
429
430 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq);
431
432 /* skb may be freed after mana_gd_post_work_request. Do not use it. */
433 skb = NULL;
434
435 tx_stats = &txq->stats;
436 u64_stats_update_begin(&tx_stats->syncp);
437 tx_stats->packets++;
438 tx_stats->bytes += len;
439 u64_stats_update_end(&tx_stats->syncp);
440
441 tx_busy:
442 if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) {
443 netif_tx_wake_queue(net_txq);
444 apc->eth_stats.wake_queue++;
445 }
446
447 kfree(pkg.sgl_ptr);
448 return err;
449
450 free_sgl_ptr:
451 kfree(pkg.sgl_ptr);
452 tx_drop_count:
453 ndev->stats.tx_dropped++;
454 tx_drop:
455 dev_kfree_skb_any(skb);
456 return NETDEV_TX_OK;
457 }
458
mana_get_stats64(struct net_device * ndev,struct rtnl_link_stats64 * st)459 static void mana_get_stats64(struct net_device *ndev,
460 struct rtnl_link_stats64 *st)
461 {
462 struct mana_port_context *apc = netdev_priv(ndev);
463 unsigned int num_queues = apc->num_queues;
464 struct mana_stats_rx *rx_stats;
465 struct mana_stats_tx *tx_stats;
466 unsigned int start;
467 u64 packets, bytes;
468 int q;
469
470 if (!apc->port_is_up)
471 return;
472
473 netdev_stats_to_stats64(st, &ndev->stats);
474
475 for (q = 0; q < num_queues; q++) {
476 rx_stats = &apc->rxqs[q]->stats;
477
478 do {
479 start = u64_stats_fetch_begin(&rx_stats->syncp);
480 packets = rx_stats->packets;
481 bytes = rx_stats->bytes;
482 } while (u64_stats_fetch_retry(&rx_stats->syncp, start));
483
484 st->rx_packets += packets;
485 st->rx_bytes += bytes;
486 }
487
488 for (q = 0; q < num_queues; q++) {
489 tx_stats = &apc->tx_qp[q].txq.stats;
490
491 do {
492 start = u64_stats_fetch_begin(&tx_stats->syncp);
493 packets = tx_stats->packets;
494 bytes = tx_stats->bytes;
495 } while (u64_stats_fetch_retry(&tx_stats->syncp, start));
496
497 st->tx_packets += packets;
498 st->tx_bytes += bytes;
499 }
500 }
501
mana_get_tx_queue(struct net_device * ndev,struct sk_buff * skb,int old_q)502 static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb,
503 int old_q)
504 {
505 struct mana_port_context *apc = netdev_priv(ndev);
506 u32 hash = skb_get_hash(skb);
507 struct sock *sk = skb->sk;
508 int txq;
509
510 txq = apc->indir_table[hash & (apc->indir_table_sz - 1)];
511
512 if (txq != old_q && sk && sk_fullsock(sk) &&
513 rcu_access_pointer(sk->sk_dst_cache))
514 sk_tx_queue_set(sk, txq);
515
516 return txq;
517 }
518
mana_select_queue(struct net_device * ndev,struct sk_buff * skb,struct net_device * sb_dev)519 static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb,
520 struct net_device *sb_dev)
521 {
522 int txq;
523
524 if (ndev->real_num_tx_queues == 1)
525 return 0;
526
527 txq = sk_tx_queue_get(skb->sk);
528
529 if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) {
530 if (skb_rx_queue_recorded(skb))
531 txq = skb_get_rx_queue(skb);
532 else
533 txq = mana_get_tx_queue(ndev, skb, txq);
534 }
535
536 return txq;
537 }
538
539 /* Release pre-allocated RX buffers */
mana_pre_dealloc_rxbufs(struct mana_port_context * mpc)540 void mana_pre_dealloc_rxbufs(struct mana_port_context *mpc)
541 {
542 struct device *dev;
543 int i;
544
545 dev = mpc->ac->gdma_dev->gdma_context->dev;
546
547 if (!mpc->rxbufs_pre)
548 goto out1;
549
550 if (!mpc->das_pre)
551 goto out2;
552
553 while (mpc->rxbpre_total) {
554 i = --mpc->rxbpre_total;
555 dma_unmap_single(dev, mpc->das_pre[i], mpc->rxbpre_datasize,
556 DMA_FROM_DEVICE);
557 put_page(virt_to_head_page(mpc->rxbufs_pre[i]));
558 }
559
560 kfree(mpc->das_pre);
561 mpc->das_pre = NULL;
562
563 out2:
564 kfree(mpc->rxbufs_pre);
565 mpc->rxbufs_pre = NULL;
566
567 out1:
568 mpc->rxbpre_datasize = 0;
569 mpc->rxbpre_alloc_size = 0;
570 mpc->rxbpre_headroom = 0;
571 }
572
573 /* Get a buffer from the pre-allocated RX buffers */
mana_get_rxbuf_pre(struct mana_rxq * rxq,dma_addr_t * da)574 static void *mana_get_rxbuf_pre(struct mana_rxq *rxq, dma_addr_t *da)
575 {
576 struct net_device *ndev = rxq->ndev;
577 struct mana_port_context *mpc;
578 void *va;
579
580 mpc = netdev_priv(ndev);
581
582 if (!mpc->rxbufs_pre || !mpc->das_pre || !mpc->rxbpre_total) {
583 netdev_err(ndev, "No RX pre-allocated bufs\n");
584 return NULL;
585 }
586
587 /* Check sizes to catch unexpected coding error */
588 if (mpc->rxbpre_datasize != rxq->datasize) {
589 netdev_err(ndev, "rxbpre_datasize mismatch: %u: %u\n",
590 mpc->rxbpre_datasize, rxq->datasize);
591 return NULL;
592 }
593
594 if (mpc->rxbpre_alloc_size != rxq->alloc_size) {
595 netdev_err(ndev, "rxbpre_alloc_size mismatch: %u: %u\n",
596 mpc->rxbpre_alloc_size, rxq->alloc_size);
597 return NULL;
598 }
599
600 if (mpc->rxbpre_headroom != rxq->headroom) {
601 netdev_err(ndev, "rxbpre_headroom mismatch: %u: %u\n",
602 mpc->rxbpre_headroom, rxq->headroom);
603 return NULL;
604 }
605
606 mpc->rxbpre_total--;
607
608 *da = mpc->das_pre[mpc->rxbpre_total];
609 va = mpc->rxbufs_pre[mpc->rxbpre_total];
610 mpc->rxbufs_pre[mpc->rxbpre_total] = NULL;
611
612 /* Deallocate the array after all buffers are gone */
613 if (!mpc->rxbpre_total)
614 mana_pre_dealloc_rxbufs(mpc);
615
616 return va;
617 }
618
619 /* Get RX buffer's data size, alloc size, XDP headroom based on MTU */
mana_get_rxbuf_cfg(int mtu,u32 * datasize,u32 * alloc_size,u32 * headroom)620 static void mana_get_rxbuf_cfg(int mtu, u32 *datasize, u32 *alloc_size,
621 u32 *headroom)
622 {
623 if (mtu > MANA_XDP_MTU_MAX)
624 *headroom = 0; /* no support for XDP */
625 else
626 *headroom = XDP_PACKET_HEADROOM;
627
628 *alloc_size = SKB_DATA_ALIGN(mtu + MANA_RXBUF_PAD + *headroom);
629
630 /* Using page pool in this case, so alloc_size is PAGE_SIZE */
631 if (*alloc_size < PAGE_SIZE)
632 *alloc_size = PAGE_SIZE;
633
634 *datasize = mtu + ETH_HLEN;
635 }
636
mana_pre_alloc_rxbufs(struct mana_port_context * mpc,int new_mtu,int num_queues)637 int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu, int num_queues)
638 {
639 struct device *dev;
640 struct page *page;
641 dma_addr_t da;
642 int num_rxb;
643 void *va;
644 int i;
645
646 mana_get_rxbuf_cfg(new_mtu, &mpc->rxbpre_datasize,
647 &mpc->rxbpre_alloc_size, &mpc->rxbpre_headroom);
648
649 dev = mpc->ac->gdma_dev->gdma_context->dev;
650
651 num_rxb = num_queues * mpc->rx_queue_size;
652
653 WARN(mpc->rxbufs_pre, "mana rxbufs_pre exists\n");
654 mpc->rxbufs_pre = kmalloc_array(num_rxb, sizeof(void *), GFP_KERNEL);
655 if (!mpc->rxbufs_pre)
656 goto error;
657
658 mpc->das_pre = kmalloc_array(num_rxb, sizeof(dma_addr_t), GFP_KERNEL);
659 if (!mpc->das_pre)
660 goto error;
661
662 mpc->rxbpre_total = 0;
663
664 for (i = 0; i < num_rxb; i++) {
665 page = dev_alloc_pages(get_order(mpc->rxbpre_alloc_size));
666 if (!page)
667 goto error;
668
669 va = page_to_virt(page);
670
671 da = dma_map_single(dev, va + mpc->rxbpre_headroom,
672 mpc->rxbpre_datasize, DMA_FROM_DEVICE);
673 if (dma_mapping_error(dev, da)) {
674 put_page(page);
675 goto error;
676 }
677
678 mpc->rxbufs_pre[i] = va;
679 mpc->das_pre[i] = da;
680 mpc->rxbpre_total = i + 1;
681 }
682
683 return 0;
684
685 error:
686 netdev_err(mpc->ndev, "Failed to pre-allocate RX buffers for %d queues\n", num_queues);
687 mana_pre_dealloc_rxbufs(mpc);
688 return -ENOMEM;
689 }
690
mana_change_mtu(struct net_device * ndev,int new_mtu)691 static int mana_change_mtu(struct net_device *ndev, int new_mtu)
692 {
693 struct mana_port_context *mpc = netdev_priv(ndev);
694 unsigned int old_mtu = ndev->mtu;
695 int err;
696
697 /* Pre-allocate buffers to prevent failure in mana_attach later */
698 err = mana_pre_alloc_rxbufs(mpc, new_mtu, mpc->num_queues);
699 if (err) {
700 netdev_err(ndev, "Insufficient memory for new MTU\n");
701 return err;
702 }
703
704 err = mana_detach(ndev, false);
705 if (err) {
706 netdev_err(ndev, "mana_detach failed: %d\n", err);
707 goto out;
708 }
709
710 WRITE_ONCE(ndev->mtu, new_mtu);
711
712 err = mana_attach(ndev);
713 if (err) {
714 netdev_err(ndev, "mana_attach failed: %d\n", err);
715 WRITE_ONCE(ndev->mtu, old_mtu);
716 }
717
718 out:
719 mana_pre_dealloc_rxbufs(mpc);
720 return err;
721 }
722
723 static const struct net_device_ops mana_devops = {
724 .ndo_open = mana_open,
725 .ndo_stop = mana_close,
726 .ndo_select_queue = mana_select_queue,
727 .ndo_start_xmit = mana_start_xmit,
728 .ndo_validate_addr = eth_validate_addr,
729 .ndo_get_stats64 = mana_get_stats64,
730 .ndo_bpf = mana_bpf,
731 .ndo_xdp_xmit = mana_xdp_xmit,
732 .ndo_change_mtu = mana_change_mtu,
733 };
734
mana_cleanup_port_context(struct mana_port_context * apc)735 static void mana_cleanup_port_context(struct mana_port_context *apc)
736 {
737 /*
738 * make sure subsequent cleanup attempts don't end up removing already
739 * cleaned dentry pointer
740 */
741 debugfs_remove(apc->mana_port_debugfs);
742 apc->mana_port_debugfs = NULL;
743 kfree(apc->rxqs);
744 apc->rxqs = NULL;
745 }
746
mana_cleanup_indir_table(struct mana_port_context * apc)747 static void mana_cleanup_indir_table(struct mana_port_context *apc)
748 {
749 apc->indir_table_sz = 0;
750 kfree(apc->indir_table);
751 kfree(apc->rxobj_table);
752 }
753
mana_init_port_context(struct mana_port_context * apc)754 static int mana_init_port_context(struct mana_port_context *apc)
755 {
756 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *),
757 GFP_KERNEL);
758
759 return !apc->rxqs ? -ENOMEM : 0;
760 }
761
mana_send_request(struct mana_context * ac,void * in_buf,u32 in_len,void * out_buf,u32 out_len)762 static int mana_send_request(struct mana_context *ac, void *in_buf,
763 u32 in_len, void *out_buf, u32 out_len)
764 {
765 struct gdma_context *gc = ac->gdma_dev->gdma_context;
766 struct gdma_resp_hdr *resp = out_buf;
767 struct gdma_req_hdr *req = in_buf;
768 struct device *dev = gc->dev;
769 static atomic_t activity_id;
770 int err;
771
772 req->dev_id = gc->mana.dev_id;
773 req->activity_id = atomic_inc_return(&activity_id);
774
775 err = mana_gd_send_request(gc, in_len, in_buf, out_len,
776 out_buf);
777 if (err || resp->status) {
778 dev_err(dev, "Failed to send mana message: %d, 0x%x\n",
779 err, resp->status);
780 return err ? err : -EPROTO;
781 }
782
783 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 ||
784 req->activity_id != resp->activity_id) {
785 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n",
786 req->dev_id.as_uint32, resp->dev_id.as_uint32,
787 req->activity_id, resp->activity_id);
788 return -EPROTO;
789 }
790
791 return 0;
792 }
793
mana_verify_resp_hdr(const struct gdma_resp_hdr * resp_hdr,const enum mana_command_code expected_code,const u32 min_size)794 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr,
795 const enum mana_command_code expected_code,
796 const u32 min_size)
797 {
798 if (resp_hdr->response.msg_type != expected_code)
799 return -EPROTO;
800
801 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1)
802 return -EPROTO;
803
804 if (resp_hdr->response.msg_size < min_size)
805 return -EPROTO;
806
807 return 0;
808 }
809
mana_pf_register_hw_vport(struct mana_port_context * apc)810 static int mana_pf_register_hw_vport(struct mana_port_context *apc)
811 {
812 struct mana_register_hw_vport_resp resp = {};
813 struct mana_register_hw_vport_req req = {};
814 int err;
815
816 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT,
817 sizeof(req), sizeof(resp));
818 req.attached_gfid = 1;
819 req.is_pf_default_vport = 1;
820 req.allow_all_ether_types = 1;
821
822 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
823 sizeof(resp));
824 if (err) {
825 netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err);
826 return err;
827 }
828
829 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT,
830 sizeof(resp));
831 if (err || resp.hdr.status) {
832 netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n",
833 err, resp.hdr.status);
834 return err ? err : -EPROTO;
835 }
836
837 apc->port_handle = resp.hw_vport_handle;
838 return 0;
839 }
840
mana_pf_deregister_hw_vport(struct mana_port_context * apc)841 static void mana_pf_deregister_hw_vport(struct mana_port_context *apc)
842 {
843 struct mana_deregister_hw_vport_resp resp = {};
844 struct mana_deregister_hw_vport_req req = {};
845 int err;
846
847 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT,
848 sizeof(req), sizeof(resp));
849 req.hw_vport_handle = apc->port_handle;
850
851 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
852 sizeof(resp));
853 if (err) {
854 netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n",
855 err);
856 return;
857 }
858
859 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT,
860 sizeof(resp));
861 if (err || resp.hdr.status)
862 netdev_err(apc->ndev,
863 "Failed to deregister hw vPort: %d, 0x%x\n",
864 err, resp.hdr.status);
865 }
866
mana_pf_register_filter(struct mana_port_context * apc)867 static int mana_pf_register_filter(struct mana_port_context *apc)
868 {
869 struct mana_register_filter_resp resp = {};
870 struct mana_register_filter_req req = {};
871 int err;
872
873 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER,
874 sizeof(req), sizeof(resp));
875 req.vport = apc->port_handle;
876 memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN);
877
878 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
879 sizeof(resp));
880 if (err) {
881 netdev_err(apc->ndev, "Failed to register filter: %d\n", err);
882 return err;
883 }
884
885 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER,
886 sizeof(resp));
887 if (err || resp.hdr.status) {
888 netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n",
889 err, resp.hdr.status);
890 return err ? err : -EPROTO;
891 }
892
893 apc->pf_filter_handle = resp.filter_handle;
894 return 0;
895 }
896
mana_pf_deregister_filter(struct mana_port_context * apc)897 static void mana_pf_deregister_filter(struct mana_port_context *apc)
898 {
899 struct mana_deregister_filter_resp resp = {};
900 struct mana_deregister_filter_req req = {};
901 int err;
902
903 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER,
904 sizeof(req), sizeof(resp));
905 req.filter_handle = apc->pf_filter_handle;
906
907 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
908 sizeof(resp));
909 if (err) {
910 netdev_err(apc->ndev, "Failed to unregister filter: %d\n",
911 err);
912 return;
913 }
914
915 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER,
916 sizeof(resp));
917 if (err || resp.hdr.status)
918 netdev_err(apc->ndev,
919 "Failed to deregister filter: %d, 0x%x\n",
920 err, resp.hdr.status);
921 }
922
mana_query_device_cfg(struct mana_context * ac,u32 proto_major_ver,u32 proto_minor_ver,u32 proto_micro_ver,u16 * max_num_vports,u8 * bm_hostmode)923 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver,
924 u32 proto_minor_ver, u32 proto_micro_ver,
925 u16 *max_num_vports, u8 *bm_hostmode)
926 {
927 struct gdma_context *gc = ac->gdma_dev->gdma_context;
928 struct mana_query_device_cfg_resp resp = {};
929 struct mana_query_device_cfg_req req = {};
930 struct device *dev = gc->dev;
931 int err = 0;
932
933 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG,
934 sizeof(req), sizeof(resp));
935
936 req.hdr.resp.msg_version = GDMA_MESSAGE_V3;
937
938 req.proto_major_ver = proto_major_ver;
939 req.proto_minor_ver = proto_minor_ver;
940 req.proto_micro_ver = proto_micro_ver;
941
942 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp));
943 if (err) {
944 dev_err(dev, "Failed to query config: %d", err);
945 return err;
946 }
947
948 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG,
949 sizeof(resp));
950 if (err || resp.hdr.status) {
951 dev_err(dev, "Invalid query result: %d, 0x%x\n", err,
952 resp.hdr.status);
953 if (!err)
954 err = -EPROTO;
955 return err;
956 }
957
958 *max_num_vports = resp.max_num_vports;
959
960 if (resp.hdr.response.msg_version >= GDMA_MESSAGE_V2)
961 gc->adapter_mtu = resp.adapter_mtu;
962 else
963 gc->adapter_mtu = ETH_FRAME_LEN;
964
965 if (resp.hdr.response.msg_version >= GDMA_MESSAGE_V3)
966 *bm_hostmode = resp.bm_hostmode;
967 else
968 *bm_hostmode = 0;
969
970 debugfs_create_u16("adapter-MTU", 0400, gc->mana_pci_debugfs, &gc->adapter_mtu);
971
972 return 0;
973 }
974
mana_query_vport_cfg(struct mana_port_context * apc,u32 vport_index,u32 * max_sq,u32 * max_rq,u32 * num_indir_entry)975 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index,
976 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry)
977 {
978 struct mana_query_vport_cfg_resp resp = {};
979 struct mana_query_vport_cfg_req req = {};
980 int err;
981
982 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG,
983 sizeof(req), sizeof(resp));
984
985 req.vport_index = vport_index;
986
987 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
988 sizeof(resp));
989 if (err)
990 return err;
991
992 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG,
993 sizeof(resp));
994 if (err)
995 return err;
996
997 if (resp.hdr.status)
998 return -EPROTO;
999
1000 *max_sq = resp.max_num_sq;
1001 *max_rq = resp.max_num_rq;
1002 if (resp.num_indirection_ent > 0 &&
1003 resp.num_indirection_ent <= MANA_INDIRECT_TABLE_MAX_SIZE &&
1004 is_power_of_2(resp.num_indirection_ent)) {
1005 *num_indir_entry = resp.num_indirection_ent;
1006 } else {
1007 netdev_warn(apc->ndev,
1008 "Setting indirection table size to default %d for vPort %d\n",
1009 MANA_INDIRECT_TABLE_DEF_SIZE, apc->port_idx);
1010 *num_indir_entry = MANA_INDIRECT_TABLE_DEF_SIZE;
1011 }
1012
1013 apc->port_handle = resp.vport;
1014 ether_addr_copy(apc->mac_addr, resp.mac_addr);
1015
1016 return 0;
1017 }
1018
mana_uncfg_vport(struct mana_port_context * apc)1019 void mana_uncfg_vport(struct mana_port_context *apc)
1020 {
1021 mutex_lock(&apc->vport_mutex);
1022 apc->vport_use_count--;
1023 WARN_ON(apc->vport_use_count < 0);
1024 mutex_unlock(&apc->vport_mutex);
1025 }
1026 EXPORT_SYMBOL_NS(mana_uncfg_vport, "NET_MANA");
1027
mana_cfg_vport(struct mana_port_context * apc,u32 protection_dom_id,u32 doorbell_pg_id)1028 int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id,
1029 u32 doorbell_pg_id)
1030 {
1031 struct mana_config_vport_resp resp = {};
1032 struct mana_config_vport_req req = {};
1033 int err;
1034
1035 /* This function is used to program the Ethernet port in the hardware
1036 * table. It can be called from the Ethernet driver or the RDMA driver.
1037 *
1038 * For Ethernet usage, the hardware supports only one active user on a
1039 * physical port. The driver checks on the port usage before programming
1040 * the hardware when creating the RAW QP (RDMA driver) or exposing the
1041 * device to kernel NET layer (Ethernet driver).
1042 *
1043 * Because the RDMA driver doesn't know in advance which QP type the
1044 * user will create, it exposes the device with all its ports. The user
1045 * may not be able to create RAW QP on a port if this port is already
1046 * in used by the Ethernet driver from the kernel.
1047 *
1048 * This physical port limitation only applies to the RAW QP. For RC QP,
1049 * the hardware doesn't have this limitation. The user can create RC
1050 * QPs on a physical port up to the hardware limits independent of the
1051 * Ethernet usage on the same port.
1052 */
1053 mutex_lock(&apc->vport_mutex);
1054 if (apc->vport_use_count > 0) {
1055 mutex_unlock(&apc->vport_mutex);
1056 return -EBUSY;
1057 }
1058 apc->vport_use_count++;
1059 mutex_unlock(&apc->vport_mutex);
1060
1061 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX,
1062 sizeof(req), sizeof(resp));
1063 req.vport = apc->port_handle;
1064 req.pdid = protection_dom_id;
1065 req.doorbell_pageid = doorbell_pg_id;
1066
1067 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1068 sizeof(resp));
1069 if (err) {
1070 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err);
1071 goto out;
1072 }
1073
1074 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX,
1075 sizeof(resp));
1076 if (err || resp.hdr.status) {
1077 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n",
1078 err, resp.hdr.status);
1079 if (!err)
1080 err = -EPROTO;
1081
1082 goto out;
1083 }
1084
1085 apc->tx_shortform_allowed = resp.short_form_allowed;
1086 apc->tx_vp_offset = resp.tx_vport_offset;
1087
1088 netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n",
1089 apc->port_handle, protection_dom_id, doorbell_pg_id);
1090 out:
1091 if (err)
1092 mana_uncfg_vport(apc);
1093
1094 return err;
1095 }
1096 EXPORT_SYMBOL_NS(mana_cfg_vport, "NET_MANA");
1097
mana_cfg_vport_steering(struct mana_port_context * apc,enum TRI_STATE rx,bool update_default_rxobj,bool update_key,bool update_tab)1098 static int mana_cfg_vport_steering(struct mana_port_context *apc,
1099 enum TRI_STATE rx,
1100 bool update_default_rxobj, bool update_key,
1101 bool update_tab)
1102 {
1103 struct mana_cfg_rx_steer_req_v2 *req;
1104 struct mana_cfg_rx_steer_resp resp = {};
1105 struct net_device *ndev = apc->ndev;
1106 u32 req_buf_size;
1107 int err;
1108
1109 req_buf_size = struct_size(req, indir_tab, apc->indir_table_sz);
1110 req = kzalloc(req_buf_size, GFP_KERNEL);
1111 if (!req)
1112 return -ENOMEM;
1113
1114 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
1115 sizeof(resp));
1116
1117 req->hdr.req.msg_version = GDMA_MESSAGE_V2;
1118
1119 req->vport = apc->port_handle;
1120 req->num_indir_entries = apc->indir_table_sz;
1121 req->indir_tab_offset = offsetof(struct mana_cfg_rx_steer_req_v2,
1122 indir_tab);
1123 req->rx_enable = rx;
1124 req->rss_enable = apc->rss_state;
1125 req->update_default_rxobj = update_default_rxobj;
1126 req->update_hashkey = update_key;
1127 req->update_indir_tab = update_tab;
1128 req->default_rxobj = apc->default_rxobj;
1129 req->cqe_coalescing_enable = 0;
1130
1131 if (update_key)
1132 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE);
1133
1134 if (update_tab)
1135 memcpy(req->indir_tab, apc->rxobj_table,
1136 flex_array_size(req, indir_tab, req->num_indir_entries));
1137
1138 err = mana_send_request(apc->ac, req, req_buf_size, &resp,
1139 sizeof(resp));
1140 if (err) {
1141 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err);
1142 goto out;
1143 }
1144
1145 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX,
1146 sizeof(resp));
1147 if (err) {
1148 netdev_err(ndev, "vPort RX configuration failed: %d\n", err);
1149 goto out;
1150 }
1151
1152 if (resp.hdr.status) {
1153 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n",
1154 resp.hdr.status);
1155 err = -EPROTO;
1156 }
1157
1158 netdev_info(ndev, "Configured steering vPort %llu entries %u\n",
1159 apc->port_handle, apc->indir_table_sz);
1160 out:
1161 kfree(req);
1162 return err;
1163 }
1164
mana_create_wq_obj(struct mana_port_context * apc,mana_handle_t vport,u32 wq_type,struct mana_obj_spec * wq_spec,struct mana_obj_spec * cq_spec,mana_handle_t * wq_obj)1165 int mana_create_wq_obj(struct mana_port_context *apc,
1166 mana_handle_t vport,
1167 u32 wq_type, struct mana_obj_spec *wq_spec,
1168 struct mana_obj_spec *cq_spec,
1169 mana_handle_t *wq_obj)
1170 {
1171 struct mana_create_wqobj_resp resp = {};
1172 struct mana_create_wqobj_req req = {};
1173 struct net_device *ndev = apc->ndev;
1174 int err;
1175
1176 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ,
1177 sizeof(req), sizeof(resp));
1178 req.vport = vport;
1179 req.wq_type = wq_type;
1180 req.wq_gdma_region = wq_spec->gdma_region;
1181 req.cq_gdma_region = cq_spec->gdma_region;
1182 req.wq_size = wq_spec->queue_size;
1183 req.cq_size = cq_spec->queue_size;
1184 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id;
1185 req.cq_parent_qid = cq_spec->attached_eq;
1186
1187 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1188 sizeof(resp));
1189 if (err) {
1190 netdev_err(ndev, "Failed to create WQ object: %d\n", err);
1191 goto out;
1192 }
1193
1194 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ,
1195 sizeof(resp));
1196 if (err || resp.hdr.status) {
1197 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err,
1198 resp.hdr.status);
1199 if (!err)
1200 err = -EPROTO;
1201 goto out;
1202 }
1203
1204 if (resp.wq_obj == INVALID_MANA_HANDLE) {
1205 netdev_err(ndev, "Got an invalid WQ object handle\n");
1206 err = -EPROTO;
1207 goto out;
1208 }
1209
1210 *wq_obj = resp.wq_obj;
1211 wq_spec->queue_index = resp.wq_id;
1212 cq_spec->queue_index = resp.cq_id;
1213
1214 return 0;
1215 out:
1216 return err;
1217 }
1218 EXPORT_SYMBOL_NS(mana_create_wq_obj, "NET_MANA");
1219
mana_destroy_wq_obj(struct mana_port_context * apc,u32 wq_type,mana_handle_t wq_obj)1220 void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type,
1221 mana_handle_t wq_obj)
1222 {
1223 struct mana_destroy_wqobj_resp resp = {};
1224 struct mana_destroy_wqobj_req req = {};
1225 struct net_device *ndev = apc->ndev;
1226 int err;
1227
1228 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ,
1229 sizeof(req), sizeof(resp));
1230 req.wq_type = wq_type;
1231 req.wq_obj_handle = wq_obj;
1232
1233 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1234 sizeof(resp));
1235 if (err) {
1236 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err);
1237 return;
1238 }
1239
1240 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ,
1241 sizeof(resp));
1242 if (err || resp.hdr.status)
1243 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err,
1244 resp.hdr.status);
1245 }
1246 EXPORT_SYMBOL_NS(mana_destroy_wq_obj, "NET_MANA");
1247
mana_destroy_eq(struct mana_context * ac)1248 static void mana_destroy_eq(struct mana_context *ac)
1249 {
1250 struct gdma_context *gc = ac->gdma_dev->gdma_context;
1251 struct gdma_queue *eq;
1252 int i;
1253
1254 if (!ac->eqs)
1255 return;
1256
1257 debugfs_remove_recursive(ac->mana_eqs_debugfs);
1258 ac->mana_eqs_debugfs = NULL;
1259
1260 for (i = 0; i < gc->max_num_queues; i++) {
1261 eq = ac->eqs[i].eq;
1262 if (!eq)
1263 continue;
1264
1265 mana_gd_destroy_queue(gc, eq);
1266 }
1267
1268 kfree(ac->eqs);
1269 ac->eqs = NULL;
1270 }
1271
mana_create_eq_debugfs(struct mana_context * ac,int i)1272 static void mana_create_eq_debugfs(struct mana_context *ac, int i)
1273 {
1274 struct mana_eq eq = ac->eqs[i];
1275 char eqnum[32];
1276
1277 sprintf(eqnum, "eq%d", i);
1278 eq.mana_eq_debugfs = debugfs_create_dir(eqnum, ac->mana_eqs_debugfs);
1279 debugfs_create_u32("head", 0400, eq.mana_eq_debugfs, &eq.eq->head);
1280 debugfs_create_u32("tail", 0400, eq.mana_eq_debugfs, &eq.eq->tail);
1281 debugfs_create_file("eq_dump", 0400, eq.mana_eq_debugfs, eq.eq, &mana_dbg_q_fops);
1282 }
1283
mana_create_eq(struct mana_context * ac)1284 static int mana_create_eq(struct mana_context *ac)
1285 {
1286 struct gdma_dev *gd = ac->gdma_dev;
1287 struct gdma_context *gc = gd->gdma_context;
1288 struct gdma_queue_spec spec = {};
1289 int err;
1290 int i;
1291
1292 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq),
1293 GFP_KERNEL);
1294 if (!ac->eqs)
1295 return -ENOMEM;
1296
1297 spec.type = GDMA_EQ;
1298 spec.monitor_avl_buf = false;
1299 spec.queue_size = EQ_SIZE;
1300 spec.eq.callback = NULL;
1301 spec.eq.context = ac->eqs;
1302 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE;
1303
1304 ac->mana_eqs_debugfs = debugfs_create_dir("EQs", gc->mana_pci_debugfs);
1305
1306 for (i = 0; i < gc->max_num_queues; i++) {
1307 spec.eq.msix_index = (i + 1) % gc->num_msix_usable;
1308 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq);
1309 if (err) {
1310 dev_err(gc->dev, "Failed to create EQ %d : %d\n", i, err);
1311 goto out;
1312 }
1313 mana_create_eq_debugfs(ac, i);
1314 }
1315
1316 return 0;
1317 out:
1318 mana_destroy_eq(ac);
1319 return err;
1320 }
1321
mana_fence_rq(struct mana_port_context * apc,struct mana_rxq * rxq)1322 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq)
1323 {
1324 struct mana_fence_rq_resp resp = {};
1325 struct mana_fence_rq_req req = {};
1326 int err;
1327
1328 init_completion(&rxq->fence_event);
1329
1330 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ,
1331 sizeof(req), sizeof(resp));
1332 req.wq_obj_handle = rxq->rxobj;
1333
1334 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1335 sizeof(resp));
1336 if (err) {
1337 netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n",
1338 rxq->rxq_idx, err);
1339 return err;
1340 }
1341
1342 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp));
1343 if (err || resp.hdr.status) {
1344 netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n",
1345 rxq->rxq_idx, err, resp.hdr.status);
1346 if (!err)
1347 err = -EPROTO;
1348
1349 return err;
1350 }
1351
1352 if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) {
1353 netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n",
1354 rxq->rxq_idx);
1355 return -ETIMEDOUT;
1356 }
1357
1358 return 0;
1359 }
1360
mana_fence_rqs(struct mana_port_context * apc)1361 static void mana_fence_rqs(struct mana_port_context *apc)
1362 {
1363 unsigned int rxq_idx;
1364 struct mana_rxq *rxq;
1365 int err;
1366
1367 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
1368 rxq = apc->rxqs[rxq_idx];
1369 err = mana_fence_rq(apc, rxq);
1370
1371 /* In case of any error, use sleep instead. */
1372 if (err)
1373 msleep(100);
1374 }
1375 }
1376
mana_move_wq_tail(struct gdma_queue * wq,u32 num_units)1377 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units)
1378 {
1379 u32 used_space_old;
1380 u32 used_space_new;
1381
1382 used_space_old = wq->head - wq->tail;
1383 used_space_new = wq->head - (wq->tail + num_units);
1384
1385 if (WARN_ON_ONCE(used_space_new > used_space_old))
1386 return -ERANGE;
1387
1388 wq->tail += num_units;
1389 return 0;
1390 }
1391
mana_unmap_skb(struct sk_buff * skb,struct mana_port_context * apc)1392 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc)
1393 {
1394 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
1395 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1396 struct device *dev = gc->dev;
1397 int hsg, i;
1398
1399 /* Number of SGEs of linear part */
1400 hsg = (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0]) ? 2 : 1;
1401
1402 for (i = 0; i < hsg; i++)
1403 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i],
1404 DMA_TO_DEVICE);
1405
1406 for (i = hsg; i < skb_shinfo(skb)->nr_frags + hsg; i++)
1407 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
1408 DMA_TO_DEVICE);
1409 }
1410
mana_poll_tx_cq(struct mana_cq * cq)1411 static void mana_poll_tx_cq(struct mana_cq *cq)
1412 {
1413 struct gdma_comp *completions = cq->gdma_comp_buf;
1414 struct gdma_posted_wqe_info *wqe_info;
1415 unsigned int pkt_transmitted = 0;
1416 unsigned int wqe_unit_cnt = 0;
1417 struct mana_txq *txq = cq->txq;
1418 struct mana_port_context *apc;
1419 struct netdev_queue *net_txq;
1420 struct gdma_queue *gdma_wq;
1421 unsigned int avail_space;
1422 struct net_device *ndev;
1423 struct sk_buff *skb;
1424 bool txq_stopped;
1425 int comp_read;
1426 int i;
1427
1428 ndev = txq->ndev;
1429 apc = netdev_priv(ndev);
1430
1431 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions,
1432 CQE_POLLING_BUFFER);
1433
1434 if (comp_read < 1)
1435 return;
1436
1437 for (i = 0; i < comp_read; i++) {
1438 struct mana_tx_comp_oob *cqe_oob;
1439
1440 if (WARN_ON_ONCE(!completions[i].is_sq))
1441 return;
1442
1443 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data;
1444 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type !=
1445 MANA_CQE_COMPLETION))
1446 return;
1447
1448 switch (cqe_oob->cqe_hdr.cqe_type) {
1449 case CQE_TX_OKAY:
1450 break;
1451
1452 case CQE_TX_SA_DROP:
1453 case CQE_TX_MTU_DROP:
1454 case CQE_TX_INVALID_OOB:
1455 case CQE_TX_INVALID_ETH_TYPE:
1456 case CQE_TX_HDR_PROCESSING_ERROR:
1457 case CQE_TX_VF_DISABLED:
1458 case CQE_TX_VPORT_IDX_OUT_OF_RANGE:
1459 case CQE_TX_VPORT_DISABLED:
1460 case CQE_TX_VLAN_TAGGING_VIOLATION:
1461 if (net_ratelimit())
1462 netdev_err(ndev, "TX: CQE error %d\n",
1463 cqe_oob->cqe_hdr.cqe_type);
1464
1465 apc->eth_stats.tx_cqe_err++;
1466 break;
1467
1468 default:
1469 /* If the CQE type is unknown, log an error,
1470 * and still free the SKB, update tail, etc.
1471 */
1472 if (net_ratelimit())
1473 netdev_err(ndev, "TX: unknown CQE type %d\n",
1474 cqe_oob->cqe_hdr.cqe_type);
1475
1476 apc->eth_stats.tx_cqe_unknown_type++;
1477 break;
1478 }
1479
1480 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num))
1481 return;
1482
1483 skb = skb_dequeue(&txq->pending_skbs);
1484 if (WARN_ON_ONCE(!skb))
1485 return;
1486
1487 wqe_info = (struct gdma_posted_wqe_info *)skb->cb;
1488 wqe_unit_cnt += wqe_info->wqe_size_in_bu;
1489
1490 mana_unmap_skb(skb, apc);
1491
1492 napi_consume_skb(skb, cq->budget);
1493
1494 pkt_transmitted++;
1495 }
1496
1497 if (WARN_ON_ONCE(wqe_unit_cnt == 0))
1498 return;
1499
1500 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt);
1501
1502 gdma_wq = txq->gdma_sq;
1503 avail_space = mana_gd_wq_avail_space(gdma_wq);
1504
1505 /* Ensure tail updated before checking q stop */
1506 smp_mb();
1507
1508 net_txq = txq->net_txq;
1509 txq_stopped = netif_tx_queue_stopped(net_txq);
1510
1511 /* Ensure checking txq_stopped before apc->port_is_up. */
1512 smp_rmb();
1513
1514 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) {
1515 netif_tx_wake_queue(net_txq);
1516 apc->eth_stats.wake_queue++;
1517 }
1518
1519 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0)
1520 WARN_ON_ONCE(1);
1521
1522 cq->work_done = pkt_transmitted;
1523 }
1524
mana_post_pkt_rxq(struct mana_rxq * rxq)1525 static void mana_post_pkt_rxq(struct mana_rxq *rxq)
1526 {
1527 struct mana_recv_buf_oob *recv_buf_oob;
1528 u32 curr_index;
1529 int err;
1530
1531 curr_index = rxq->buf_index++;
1532 if (rxq->buf_index == rxq->num_rx_buf)
1533 rxq->buf_index = 0;
1534
1535 recv_buf_oob = &rxq->rx_oobs[curr_index];
1536
1537 err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req,
1538 &recv_buf_oob->wqe_inf);
1539 if (WARN_ON_ONCE(err))
1540 return;
1541
1542 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1);
1543 }
1544
mana_build_skb(struct mana_rxq * rxq,void * buf_va,uint pkt_len,struct xdp_buff * xdp)1545 static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va,
1546 uint pkt_len, struct xdp_buff *xdp)
1547 {
1548 struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size);
1549
1550 if (!skb)
1551 return NULL;
1552
1553 if (xdp->data_hard_start) {
1554 u32 metasize = xdp->data - xdp->data_meta;
1555
1556 skb_reserve(skb, xdp->data - xdp->data_hard_start);
1557 skb_put(skb, xdp->data_end - xdp->data);
1558 if (metasize)
1559 skb_metadata_set(skb, metasize);
1560 return skb;
1561 }
1562
1563 skb_reserve(skb, rxq->headroom);
1564 skb_put(skb, pkt_len);
1565
1566 return skb;
1567 }
1568
mana_rx_skb(void * buf_va,bool from_pool,struct mana_rxcomp_oob * cqe,struct mana_rxq * rxq)1569 static void mana_rx_skb(void *buf_va, bool from_pool,
1570 struct mana_rxcomp_oob *cqe, struct mana_rxq *rxq)
1571 {
1572 struct mana_stats_rx *rx_stats = &rxq->stats;
1573 struct net_device *ndev = rxq->ndev;
1574 uint pkt_len = cqe->ppi[0].pkt_len;
1575 u16 rxq_idx = rxq->rxq_idx;
1576 struct napi_struct *napi;
1577 struct xdp_buff xdp = {};
1578 struct sk_buff *skb;
1579 u32 hash_value;
1580 u32 act;
1581
1582 rxq->rx_cq.work_done++;
1583 napi = &rxq->rx_cq.napi;
1584
1585 if (!buf_va) {
1586 ++ndev->stats.rx_dropped;
1587 return;
1588 }
1589
1590 act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len);
1591
1592 if (act == XDP_REDIRECT && !rxq->xdp_rc)
1593 return;
1594
1595 if (act != XDP_PASS && act != XDP_TX)
1596 goto drop_xdp;
1597
1598 skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp);
1599
1600 if (!skb)
1601 goto drop;
1602
1603 if (from_pool)
1604 skb_mark_for_recycle(skb);
1605
1606 skb->dev = napi->dev;
1607
1608 skb->protocol = eth_type_trans(skb, ndev);
1609 skb_checksum_none_assert(skb);
1610 skb_record_rx_queue(skb, rxq_idx);
1611
1612 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) {
1613 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed)
1614 skb->ip_summed = CHECKSUM_UNNECESSARY;
1615 }
1616
1617 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) {
1618 hash_value = cqe->ppi[0].pkt_hash;
1619
1620 if (cqe->rx_hashtype & MANA_HASH_L4)
1621 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4);
1622 else
1623 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3);
1624 }
1625
1626 if (cqe->rx_vlantag_present) {
1627 u16 vlan_tci = cqe->rx_vlan_id;
1628
1629 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci);
1630 }
1631
1632 u64_stats_update_begin(&rx_stats->syncp);
1633 rx_stats->packets++;
1634 rx_stats->bytes += pkt_len;
1635
1636 if (act == XDP_TX)
1637 rx_stats->xdp_tx++;
1638 u64_stats_update_end(&rx_stats->syncp);
1639
1640 if (act == XDP_TX) {
1641 skb_set_queue_mapping(skb, rxq_idx);
1642 mana_xdp_tx(skb, ndev);
1643 return;
1644 }
1645
1646 napi_gro_receive(napi, skb);
1647
1648 return;
1649
1650 drop_xdp:
1651 u64_stats_update_begin(&rx_stats->syncp);
1652 rx_stats->xdp_drop++;
1653 u64_stats_update_end(&rx_stats->syncp);
1654
1655 drop:
1656 if (from_pool) {
1657 page_pool_recycle_direct(rxq->page_pool,
1658 virt_to_head_page(buf_va));
1659 } else {
1660 WARN_ON_ONCE(rxq->xdp_save_va);
1661 /* Save for reuse */
1662 rxq->xdp_save_va = buf_va;
1663 }
1664
1665 ++ndev->stats.rx_dropped;
1666
1667 return;
1668 }
1669
mana_get_rxfrag(struct mana_rxq * rxq,struct device * dev,dma_addr_t * da,bool * from_pool)1670 static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,
1671 dma_addr_t *da, bool *from_pool)
1672 {
1673 struct page *page;
1674 void *va;
1675
1676 *from_pool = false;
1677
1678 /* Reuse XDP dropped page if available */
1679 if (rxq->xdp_save_va) {
1680 va = rxq->xdp_save_va;
1681 rxq->xdp_save_va = NULL;
1682 } else {
1683 page = page_pool_dev_alloc_pages(rxq->page_pool);
1684 if (!page)
1685 return NULL;
1686
1687 *from_pool = true;
1688 va = page_to_virt(page);
1689 }
1690
1691 *da = dma_map_single(dev, va + rxq->headroom, rxq->datasize,
1692 DMA_FROM_DEVICE);
1693 if (dma_mapping_error(dev, *da)) {
1694 if (*from_pool)
1695 page_pool_put_full_page(rxq->page_pool, page, false);
1696 else
1697 put_page(virt_to_head_page(va));
1698
1699 return NULL;
1700 }
1701
1702 return va;
1703 }
1704
1705 /* Allocate frag for rx buffer, and save the old buf */
mana_refill_rx_oob(struct device * dev,struct mana_rxq * rxq,struct mana_recv_buf_oob * rxoob,void ** old_buf,bool * old_fp)1706 static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq,
1707 struct mana_recv_buf_oob *rxoob, void **old_buf,
1708 bool *old_fp)
1709 {
1710 bool from_pool;
1711 dma_addr_t da;
1712 void *va;
1713
1714 va = mana_get_rxfrag(rxq, dev, &da, &from_pool);
1715 if (!va)
1716 return;
1717
1718 dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize,
1719 DMA_FROM_DEVICE);
1720 *old_buf = rxoob->buf_va;
1721 *old_fp = rxoob->from_pool;
1722
1723 rxoob->buf_va = va;
1724 rxoob->sgl[0].address = da;
1725 rxoob->from_pool = from_pool;
1726 }
1727
mana_process_rx_cqe(struct mana_rxq * rxq,struct mana_cq * cq,struct gdma_comp * cqe)1728 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,
1729 struct gdma_comp *cqe)
1730 {
1731 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data;
1732 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1733 struct net_device *ndev = rxq->ndev;
1734 struct mana_recv_buf_oob *rxbuf_oob;
1735 struct mana_port_context *apc;
1736 struct device *dev = gc->dev;
1737 void *old_buf = NULL;
1738 u32 curr, pktlen;
1739 bool old_fp;
1740
1741 apc = netdev_priv(ndev);
1742
1743 switch (oob->cqe_hdr.cqe_type) {
1744 case CQE_RX_OKAY:
1745 break;
1746
1747 case CQE_RX_TRUNCATED:
1748 ++ndev->stats.rx_dropped;
1749 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index];
1750 netdev_warn_once(ndev, "Dropped a truncated packet\n");
1751 goto drop;
1752
1753 case CQE_RX_COALESCED_4:
1754 netdev_err(ndev, "RX coalescing is unsupported\n");
1755 apc->eth_stats.rx_coalesced_err++;
1756 return;
1757
1758 case CQE_RX_OBJECT_FENCE:
1759 complete(&rxq->fence_event);
1760 return;
1761
1762 default:
1763 netdev_err(ndev, "Unknown RX CQE type = %d\n",
1764 oob->cqe_hdr.cqe_type);
1765 apc->eth_stats.rx_cqe_unknown_type++;
1766 return;
1767 }
1768
1769 pktlen = oob->ppi[0].pkt_len;
1770
1771 if (pktlen == 0) {
1772 /* data packets should never have packetlength of zero */
1773 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n",
1774 rxq->gdma_id, cq->gdma_id, rxq->rxobj);
1775 return;
1776 }
1777
1778 curr = rxq->buf_index;
1779 rxbuf_oob = &rxq->rx_oobs[curr];
1780 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1);
1781
1782 mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf, &old_fp);
1783
1784 /* Unsuccessful refill will have old_buf == NULL.
1785 * In this case, mana_rx_skb() will drop the packet.
1786 */
1787 mana_rx_skb(old_buf, old_fp, oob, rxq);
1788
1789 drop:
1790 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu);
1791
1792 mana_post_pkt_rxq(rxq);
1793 }
1794
mana_poll_rx_cq(struct mana_cq * cq)1795 static void mana_poll_rx_cq(struct mana_cq *cq)
1796 {
1797 struct gdma_comp *comp = cq->gdma_comp_buf;
1798 struct mana_rxq *rxq = cq->rxq;
1799 int comp_read, i;
1800
1801 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER);
1802 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER);
1803
1804 rxq->xdp_flush = false;
1805
1806 for (i = 0; i < comp_read; i++) {
1807 if (WARN_ON_ONCE(comp[i].is_sq))
1808 return;
1809
1810 /* verify recv cqe references the right rxq */
1811 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id))
1812 return;
1813
1814 mana_process_rx_cqe(rxq, cq, &comp[i]);
1815 }
1816
1817 if (comp_read > 0) {
1818 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1819
1820 mana_gd_wq_ring_doorbell(gc, rxq->gdma_rq);
1821 }
1822
1823 if (rxq->xdp_flush)
1824 xdp_do_flush();
1825 }
1826
mana_cq_handler(void * context,struct gdma_queue * gdma_queue)1827 static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue)
1828 {
1829 struct mana_cq *cq = context;
1830 int w;
1831
1832 WARN_ON_ONCE(cq->gdma_cq != gdma_queue);
1833
1834 if (cq->type == MANA_CQ_TYPE_RX)
1835 mana_poll_rx_cq(cq);
1836 else
1837 mana_poll_tx_cq(cq);
1838
1839 w = cq->work_done;
1840 cq->work_done_since_doorbell += w;
1841
1842 if (w < cq->budget) {
1843 mana_gd_ring_cq(gdma_queue, SET_ARM_BIT);
1844 cq->work_done_since_doorbell = 0;
1845 napi_complete_done(&cq->napi, w);
1846 } else if (cq->work_done_since_doorbell >
1847 cq->gdma_cq->queue_size / COMP_ENTRY_SIZE * 4) {
1848 /* MANA hardware requires at least one doorbell ring every 8
1849 * wraparounds of CQ even if there is no need to arm the CQ.
1850 * This driver rings the doorbell as soon as we have exceeded
1851 * 4 wraparounds.
1852 */
1853 mana_gd_ring_cq(gdma_queue, 0);
1854 cq->work_done_since_doorbell = 0;
1855 }
1856
1857 return w;
1858 }
1859
mana_poll(struct napi_struct * napi,int budget)1860 static int mana_poll(struct napi_struct *napi, int budget)
1861 {
1862 struct mana_cq *cq = container_of(napi, struct mana_cq, napi);
1863 int w;
1864
1865 cq->work_done = 0;
1866 cq->budget = budget;
1867
1868 w = mana_cq_handler(cq, cq->gdma_cq);
1869
1870 return min(w, budget);
1871 }
1872
mana_schedule_napi(void * context,struct gdma_queue * gdma_queue)1873 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue)
1874 {
1875 struct mana_cq *cq = context;
1876
1877 napi_schedule_irqoff(&cq->napi);
1878 }
1879
mana_deinit_cq(struct mana_port_context * apc,struct mana_cq * cq)1880 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq)
1881 {
1882 struct gdma_dev *gd = apc->ac->gdma_dev;
1883
1884 if (!cq->gdma_cq)
1885 return;
1886
1887 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq);
1888 }
1889
mana_deinit_txq(struct mana_port_context * apc,struct mana_txq * txq)1890 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq)
1891 {
1892 struct gdma_dev *gd = apc->ac->gdma_dev;
1893
1894 if (!txq->gdma_sq)
1895 return;
1896
1897 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq);
1898 }
1899
mana_destroy_txq(struct mana_port_context * apc)1900 static void mana_destroy_txq(struct mana_port_context *apc)
1901 {
1902 struct napi_struct *napi;
1903 int i;
1904
1905 if (!apc->tx_qp)
1906 return;
1907
1908 for (i = 0; i < apc->num_queues; i++) {
1909 debugfs_remove_recursive(apc->tx_qp[i].mana_tx_debugfs);
1910 apc->tx_qp[i].mana_tx_debugfs = NULL;
1911
1912 napi = &apc->tx_qp[i].tx_cq.napi;
1913 if (apc->tx_qp[i].txq.napi_initialized) {
1914 napi_synchronize(napi);
1915 napi_disable(napi);
1916 netif_napi_del(napi);
1917 apc->tx_qp[i].txq.napi_initialized = false;
1918 }
1919 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object);
1920
1921 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq);
1922
1923 mana_deinit_txq(apc, &apc->tx_qp[i].txq);
1924 }
1925
1926 kfree(apc->tx_qp);
1927 apc->tx_qp = NULL;
1928 }
1929
mana_create_txq_debugfs(struct mana_port_context * apc,int idx)1930 static void mana_create_txq_debugfs(struct mana_port_context *apc, int idx)
1931 {
1932 struct mana_tx_qp *tx_qp = &apc->tx_qp[idx];
1933 char qnum[32];
1934
1935 sprintf(qnum, "TX-%d", idx);
1936 tx_qp->mana_tx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs);
1937 debugfs_create_u32("sq_head", 0400, tx_qp->mana_tx_debugfs,
1938 &tx_qp->txq.gdma_sq->head);
1939 debugfs_create_u32("sq_tail", 0400, tx_qp->mana_tx_debugfs,
1940 &tx_qp->txq.gdma_sq->tail);
1941 debugfs_create_u32("sq_pend_skb_qlen", 0400, tx_qp->mana_tx_debugfs,
1942 &tx_qp->txq.pending_skbs.qlen);
1943 debugfs_create_u32("cq_head", 0400, tx_qp->mana_tx_debugfs,
1944 &tx_qp->tx_cq.gdma_cq->head);
1945 debugfs_create_u32("cq_tail", 0400, tx_qp->mana_tx_debugfs,
1946 &tx_qp->tx_cq.gdma_cq->tail);
1947 debugfs_create_u32("cq_budget", 0400, tx_qp->mana_tx_debugfs,
1948 &tx_qp->tx_cq.budget);
1949 debugfs_create_file("txq_dump", 0400, tx_qp->mana_tx_debugfs,
1950 tx_qp->txq.gdma_sq, &mana_dbg_q_fops);
1951 debugfs_create_file("cq_dump", 0400, tx_qp->mana_tx_debugfs,
1952 tx_qp->tx_cq.gdma_cq, &mana_dbg_q_fops);
1953 }
1954
mana_create_txq(struct mana_port_context * apc,struct net_device * net)1955 static int mana_create_txq(struct mana_port_context *apc,
1956 struct net_device *net)
1957 {
1958 struct mana_context *ac = apc->ac;
1959 struct gdma_dev *gd = ac->gdma_dev;
1960 struct mana_obj_spec wq_spec;
1961 struct mana_obj_spec cq_spec;
1962 struct gdma_queue_spec spec;
1963 struct gdma_context *gc;
1964 struct mana_txq *txq;
1965 struct mana_cq *cq;
1966 u32 txq_size;
1967 u32 cq_size;
1968 int err;
1969 int i;
1970
1971 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp),
1972 GFP_KERNEL);
1973 if (!apc->tx_qp)
1974 return -ENOMEM;
1975
1976 /* The minimum size of the WQE is 32 bytes, hence
1977 * apc->tx_queue_size represents the maximum number of WQEs
1978 * the SQ can store. This value is then used to size other queues
1979 * to prevent overflow.
1980 * Also note that the txq_size is always going to be MANA_PAGE_ALIGNED,
1981 * as min val of apc->tx_queue_size is 128 and that would make
1982 * txq_size 128*32 = 4096 and the other higher values of apc->tx_queue_size
1983 * are always power of two
1984 */
1985 txq_size = apc->tx_queue_size * 32;
1986
1987 cq_size = apc->tx_queue_size * COMP_ENTRY_SIZE;
1988
1989 gc = gd->gdma_context;
1990
1991 for (i = 0; i < apc->num_queues; i++) {
1992 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE;
1993
1994 /* Create SQ */
1995 txq = &apc->tx_qp[i].txq;
1996
1997 u64_stats_init(&txq->stats.syncp);
1998 txq->ndev = net;
1999 txq->net_txq = netdev_get_tx_queue(net, i);
2000 txq->vp_offset = apc->tx_vp_offset;
2001 txq->napi_initialized = false;
2002 skb_queue_head_init(&txq->pending_skbs);
2003
2004 memset(&spec, 0, sizeof(spec));
2005 spec.type = GDMA_SQ;
2006 spec.monitor_avl_buf = true;
2007 spec.queue_size = txq_size;
2008 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq);
2009 if (err)
2010 goto out;
2011
2012 /* Create SQ's CQ */
2013 cq = &apc->tx_qp[i].tx_cq;
2014 cq->type = MANA_CQ_TYPE_TX;
2015
2016 cq->txq = txq;
2017
2018 memset(&spec, 0, sizeof(spec));
2019 spec.type = GDMA_CQ;
2020 spec.monitor_avl_buf = false;
2021 spec.queue_size = cq_size;
2022 spec.cq.callback = mana_schedule_napi;
2023 spec.cq.parent_eq = ac->eqs[i].eq;
2024 spec.cq.context = cq;
2025 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2026 if (err)
2027 goto out;
2028
2029 memset(&wq_spec, 0, sizeof(wq_spec));
2030 memset(&cq_spec, 0, sizeof(cq_spec));
2031
2032 wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle;
2033 wq_spec.queue_size = txq->gdma_sq->queue_size;
2034
2035 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2036 cq_spec.queue_size = cq->gdma_cq->queue_size;
2037 cq_spec.modr_ctx_id = 0;
2038 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2039
2040 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ,
2041 &wq_spec, &cq_spec,
2042 &apc->tx_qp[i].tx_object);
2043
2044 if (err)
2045 goto out;
2046
2047 txq->gdma_sq->id = wq_spec.queue_index;
2048 cq->gdma_cq->id = cq_spec.queue_index;
2049
2050 txq->gdma_sq->mem_info.dma_region_handle =
2051 GDMA_INVALID_DMA_REGION;
2052 cq->gdma_cq->mem_info.dma_region_handle =
2053 GDMA_INVALID_DMA_REGION;
2054
2055 txq->gdma_txq_id = txq->gdma_sq->id;
2056
2057 cq->gdma_id = cq->gdma_cq->id;
2058
2059 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
2060 err = -EINVAL;
2061 goto out;
2062 }
2063
2064 gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2065
2066 mana_create_txq_debugfs(apc, i);
2067
2068 netif_napi_add_tx(net, &cq->napi, mana_poll);
2069 napi_enable(&cq->napi);
2070 txq->napi_initialized = true;
2071
2072 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2073 }
2074
2075 return 0;
2076 out:
2077 netdev_err(net, "Failed to create %d TX queues, %d\n",
2078 apc->num_queues, err);
2079 mana_destroy_txq(apc);
2080 return err;
2081 }
2082
mana_destroy_rxq(struct mana_port_context * apc,struct mana_rxq * rxq,bool napi_initialized)2083 static void mana_destroy_rxq(struct mana_port_context *apc,
2084 struct mana_rxq *rxq, bool napi_initialized)
2085
2086 {
2087 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
2088 struct mana_recv_buf_oob *rx_oob;
2089 struct device *dev = gc->dev;
2090 struct napi_struct *napi;
2091 struct page *page;
2092 int i;
2093
2094 if (!rxq)
2095 return;
2096
2097 debugfs_remove_recursive(rxq->mana_rx_debugfs);
2098 rxq->mana_rx_debugfs = NULL;
2099
2100 napi = &rxq->rx_cq.napi;
2101
2102 if (napi_initialized) {
2103 napi_synchronize(napi);
2104
2105 napi_disable(napi);
2106
2107 netif_napi_del(napi);
2108 }
2109 xdp_rxq_info_unreg(&rxq->xdp_rxq);
2110
2111 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj);
2112
2113 mana_deinit_cq(apc, &rxq->rx_cq);
2114
2115 if (rxq->xdp_save_va)
2116 put_page(virt_to_head_page(rxq->xdp_save_va));
2117
2118 for (i = 0; i < rxq->num_rx_buf; i++) {
2119 rx_oob = &rxq->rx_oobs[i];
2120
2121 if (!rx_oob->buf_va)
2122 continue;
2123
2124 dma_unmap_single(dev, rx_oob->sgl[0].address,
2125 rx_oob->sgl[0].size, DMA_FROM_DEVICE);
2126
2127 page = virt_to_head_page(rx_oob->buf_va);
2128
2129 if (rx_oob->from_pool)
2130 page_pool_put_full_page(rxq->page_pool, page, false);
2131 else
2132 put_page(page);
2133
2134 rx_oob->buf_va = NULL;
2135 }
2136
2137 page_pool_destroy(rxq->page_pool);
2138
2139 if (rxq->gdma_rq)
2140 mana_gd_destroy_queue(gc, rxq->gdma_rq);
2141
2142 kfree(rxq);
2143 }
2144
mana_fill_rx_oob(struct mana_recv_buf_oob * rx_oob,u32 mem_key,struct mana_rxq * rxq,struct device * dev)2145 static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key,
2146 struct mana_rxq *rxq, struct device *dev)
2147 {
2148 struct mana_port_context *mpc = netdev_priv(rxq->ndev);
2149 bool from_pool = false;
2150 dma_addr_t da;
2151 void *va;
2152
2153 if (mpc->rxbufs_pre)
2154 va = mana_get_rxbuf_pre(rxq, &da);
2155 else
2156 va = mana_get_rxfrag(rxq, dev, &da, &from_pool);
2157
2158 if (!va)
2159 return -ENOMEM;
2160
2161 rx_oob->buf_va = va;
2162 rx_oob->from_pool = from_pool;
2163
2164 rx_oob->sgl[0].address = da;
2165 rx_oob->sgl[0].size = rxq->datasize;
2166 rx_oob->sgl[0].mem_key = mem_key;
2167
2168 return 0;
2169 }
2170
2171 #define MANA_WQE_HEADER_SIZE 16
2172 #define MANA_WQE_SGE_SIZE 16
2173
mana_alloc_rx_wqe(struct mana_port_context * apc,struct mana_rxq * rxq,u32 * rxq_size,u32 * cq_size)2174 static int mana_alloc_rx_wqe(struct mana_port_context *apc,
2175 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size)
2176 {
2177 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
2178 struct mana_recv_buf_oob *rx_oob;
2179 struct device *dev = gc->dev;
2180 u32 buf_idx;
2181 int ret;
2182
2183 WARN_ON(rxq->datasize == 0);
2184
2185 *rxq_size = 0;
2186 *cq_size = 0;
2187
2188 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2189 rx_oob = &rxq->rx_oobs[buf_idx];
2190 memset(rx_oob, 0, sizeof(*rx_oob));
2191
2192 rx_oob->num_sge = 1;
2193
2194 ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq,
2195 dev);
2196 if (ret)
2197 return ret;
2198
2199 rx_oob->wqe_req.sgl = rx_oob->sgl;
2200 rx_oob->wqe_req.num_sge = rx_oob->num_sge;
2201 rx_oob->wqe_req.inline_oob_size = 0;
2202 rx_oob->wqe_req.inline_oob_data = NULL;
2203 rx_oob->wqe_req.flags = 0;
2204 rx_oob->wqe_req.client_data_unit = 0;
2205
2206 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE +
2207 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32);
2208 *cq_size += COMP_ENTRY_SIZE;
2209 }
2210
2211 return 0;
2212 }
2213
mana_push_wqe(struct mana_rxq * rxq)2214 static int mana_push_wqe(struct mana_rxq *rxq)
2215 {
2216 struct mana_recv_buf_oob *rx_oob;
2217 u32 buf_idx;
2218 int err;
2219
2220 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2221 rx_oob = &rxq->rx_oobs[buf_idx];
2222
2223 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req,
2224 &rx_oob->wqe_inf);
2225 if (err)
2226 return -ENOSPC;
2227 }
2228
2229 return 0;
2230 }
2231
mana_create_page_pool(struct mana_rxq * rxq,struct gdma_context * gc)2232 static int mana_create_page_pool(struct mana_rxq *rxq, struct gdma_context *gc)
2233 {
2234 struct mana_port_context *mpc = netdev_priv(rxq->ndev);
2235 struct page_pool_params pprm = {};
2236 int ret;
2237
2238 pprm.pool_size = mpc->rx_queue_size;
2239 pprm.nid = gc->numa_node;
2240 pprm.napi = &rxq->rx_cq.napi;
2241 pprm.netdev = rxq->ndev;
2242 pprm.order = get_order(rxq->alloc_size);
2243
2244 rxq->page_pool = page_pool_create(&pprm);
2245
2246 if (IS_ERR(rxq->page_pool)) {
2247 ret = PTR_ERR(rxq->page_pool);
2248 rxq->page_pool = NULL;
2249 return ret;
2250 }
2251
2252 return 0;
2253 }
2254
mana_create_rxq(struct mana_port_context * apc,u32 rxq_idx,struct mana_eq * eq,struct net_device * ndev)2255 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,
2256 u32 rxq_idx, struct mana_eq *eq,
2257 struct net_device *ndev)
2258 {
2259 struct gdma_dev *gd = apc->ac->gdma_dev;
2260 struct mana_obj_spec wq_spec;
2261 struct mana_obj_spec cq_spec;
2262 struct gdma_queue_spec spec;
2263 struct mana_cq *cq = NULL;
2264 struct gdma_context *gc;
2265 u32 cq_size, rq_size;
2266 struct mana_rxq *rxq;
2267 int err;
2268
2269 gc = gd->gdma_context;
2270
2271 rxq = kzalloc(struct_size(rxq, rx_oobs, apc->rx_queue_size),
2272 GFP_KERNEL);
2273 if (!rxq)
2274 return NULL;
2275
2276 rxq->ndev = ndev;
2277 rxq->num_rx_buf = apc->rx_queue_size;
2278 rxq->rxq_idx = rxq_idx;
2279 rxq->rxobj = INVALID_MANA_HANDLE;
2280
2281 mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size,
2282 &rxq->headroom);
2283
2284 /* Create page pool for RX queue */
2285 err = mana_create_page_pool(rxq, gc);
2286 if (err) {
2287 netdev_err(ndev, "Create page pool err:%d\n", err);
2288 goto out;
2289 }
2290
2291 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size);
2292 if (err)
2293 goto out;
2294
2295 rq_size = MANA_PAGE_ALIGN(rq_size);
2296 cq_size = MANA_PAGE_ALIGN(cq_size);
2297
2298 /* Create RQ */
2299 memset(&spec, 0, sizeof(spec));
2300 spec.type = GDMA_RQ;
2301 spec.monitor_avl_buf = true;
2302 spec.queue_size = rq_size;
2303 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq);
2304 if (err)
2305 goto out;
2306
2307 /* Create RQ's CQ */
2308 cq = &rxq->rx_cq;
2309 cq->type = MANA_CQ_TYPE_RX;
2310 cq->rxq = rxq;
2311
2312 memset(&spec, 0, sizeof(spec));
2313 spec.type = GDMA_CQ;
2314 spec.monitor_avl_buf = false;
2315 spec.queue_size = cq_size;
2316 spec.cq.callback = mana_schedule_napi;
2317 spec.cq.parent_eq = eq->eq;
2318 spec.cq.context = cq;
2319 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2320 if (err)
2321 goto out;
2322
2323 memset(&wq_spec, 0, sizeof(wq_spec));
2324 memset(&cq_spec, 0, sizeof(cq_spec));
2325 wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle;
2326 wq_spec.queue_size = rxq->gdma_rq->queue_size;
2327
2328 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2329 cq_spec.queue_size = cq->gdma_cq->queue_size;
2330 cq_spec.modr_ctx_id = 0;
2331 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2332
2333 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ,
2334 &wq_spec, &cq_spec, &rxq->rxobj);
2335 if (err)
2336 goto out;
2337
2338 rxq->gdma_rq->id = wq_spec.queue_index;
2339 cq->gdma_cq->id = cq_spec.queue_index;
2340
2341 rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2342 cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2343
2344 rxq->gdma_id = rxq->gdma_rq->id;
2345 cq->gdma_id = cq->gdma_cq->id;
2346
2347 err = mana_push_wqe(rxq);
2348 if (err)
2349 goto out;
2350
2351 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
2352 err = -EINVAL;
2353 goto out;
2354 }
2355
2356 gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2357
2358 netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1);
2359
2360 WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx,
2361 cq->napi.napi_id));
2362 WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL,
2363 rxq->page_pool));
2364
2365 napi_enable(&cq->napi);
2366
2367 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2368 out:
2369 if (!err)
2370 return rxq;
2371
2372 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err);
2373
2374 mana_destroy_rxq(apc, rxq, false);
2375
2376 if (cq)
2377 mana_deinit_cq(apc, cq);
2378
2379 return NULL;
2380 }
2381
mana_create_rxq_debugfs(struct mana_port_context * apc,int idx)2382 static void mana_create_rxq_debugfs(struct mana_port_context *apc, int idx)
2383 {
2384 struct mana_rxq *rxq;
2385 char qnum[32];
2386
2387 rxq = apc->rxqs[idx];
2388
2389 sprintf(qnum, "RX-%d", idx);
2390 rxq->mana_rx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs);
2391 debugfs_create_u32("rq_head", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->head);
2392 debugfs_create_u32("rq_tail", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->tail);
2393 debugfs_create_u32("rq_nbuf", 0400, rxq->mana_rx_debugfs, &rxq->num_rx_buf);
2394 debugfs_create_u32("cq_head", 0400, rxq->mana_rx_debugfs,
2395 &rxq->rx_cq.gdma_cq->head);
2396 debugfs_create_u32("cq_tail", 0400, rxq->mana_rx_debugfs,
2397 &rxq->rx_cq.gdma_cq->tail);
2398 debugfs_create_u32("cq_budget", 0400, rxq->mana_rx_debugfs, &rxq->rx_cq.budget);
2399 debugfs_create_file("rxq_dump", 0400, rxq->mana_rx_debugfs, rxq->gdma_rq, &mana_dbg_q_fops);
2400 debugfs_create_file("cq_dump", 0400, rxq->mana_rx_debugfs, rxq->rx_cq.gdma_cq,
2401 &mana_dbg_q_fops);
2402 }
2403
mana_add_rx_queues(struct mana_port_context * apc,struct net_device * ndev)2404 static int mana_add_rx_queues(struct mana_port_context *apc,
2405 struct net_device *ndev)
2406 {
2407 struct mana_context *ac = apc->ac;
2408 struct mana_rxq *rxq;
2409 int err = 0;
2410 int i;
2411
2412 for (i = 0; i < apc->num_queues; i++) {
2413 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev);
2414 if (!rxq) {
2415 err = -ENOMEM;
2416 netdev_err(ndev, "Failed to create rxq %d : %d\n", i, err);
2417 goto out;
2418 }
2419
2420 u64_stats_init(&rxq->stats.syncp);
2421
2422 apc->rxqs[i] = rxq;
2423
2424 mana_create_rxq_debugfs(apc, i);
2425 }
2426
2427 apc->default_rxobj = apc->rxqs[0]->rxobj;
2428 out:
2429 return err;
2430 }
2431
mana_destroy_vport(struct mana_port_context * apc)2432 static void mana_destroy_vport(struct mana_port_context *apc)
2433 {
2434 struct gdma_dev *gd = apc->ac->gdma_dev;
2435 struct mana_rxq *rxq;
2436 u32 rxq_idx;
2437
2438 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
2439 rxq = apc->rxqs[rxq_idx];
2440 if (!rxq)
2441 continue;
2442
2443 mana_destroy_rxq(apc, rxq, true);
2444 apc->rxqs[rxq_idx] = NULL;
2445 }
2446
2447 mana_destroy_txq(apc);
2448 mana_uncfg_vport(apc);
2449
2450 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode)
2451 mana_pf_deregister_hw_vport(apc);
2452 }
2453
mana_create_vport(struct mana_port_context * apc,struct net_device * net)2454 static int mana_create_vport(struct mana_port_context *apc,
2455 struct net_device *net)
2456 {
2457 struct gdma_dev *gd = apc->ac->gdma_dev;
2458 int err;
2459
2460 apc->default_rxobj = INVALID_MANA_HANDLE;
2461
2462 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) {
2463 err = mana_pf_register_hw_vport(apc);
2464 if (err)
2465 return err;
2466 }
2467
2468 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell);
2469 if (err)
2470 return err;
2471
2472 return mana_create_txq(apc, net);
2473 }
2474
mana_rss_table_alloc(struct mana_port_context * apc)2475 static int mana_rss_table_alloc(struct mana_port_context *apc)
2476 {
2477 if (!apc->indir_table_sz) {
2478 netdev_err(apc->ndev,
2479 "Indirection table size not set for vPort %d\n",
2480 apc->port_idx);
2481 return -EINVAL;
2482 }
2483
2484 apc->indir_table = kcalloc(apc->indir_table_sz, sizeof(u32), GFP_KERNEL);
2485 if (!apc->indir_table)
2486 return -ENOMEM;
2487
2488 apc->rxobj_table = kcalloc(apc->indir_table_sz, sizeof(mana_handle_t), GFP_KERNEL);
2489 if (!apc->rxobj_table) {
2490 kfree(apc->indir_table);
2491 return -ENOMEM;
2492 }
2493
2494 return 0;
2495 }
2496
mana_rss_table_init(struct mana_port_context * apc)2497 static void mana_rss_table_init(struct mana_port_context *apc)
2498 {
2499 int i;
2500
2501 for (i = 0; i < apc->indir_table_sz; i++)
2502 apc->indir_table[i] =
2503 ethtool_rxfh_indir_default(i, apc->num_queues);
2504 }
2505
mana_config_rss(struct mana_port_context * apc,enum TRI_STATE rx,bool update_hash,bool update_tab)2506 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx,
2507 bool update_hash, bool update_tab)
2508 {
2509 u32 queue_idx;
2510 int err;
2511 int i;
2512
2513 if (update_tab) {
2514 for (i = 0; i < apc->indir_table_sz; i++) {
2515 queue_idx = apc->indir_table[i];
2516 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj;
2517 }
2518 }
2519
2520 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab);
2521 if (err)
2522 return err;
2523
2524 mana_fence_rqs(apc);
2525
2526 return 0;
2527 }
2528
mana_query_gf_stats(struct mana_port_context * apc)2529 void mana_query_gf_stats(struct mana_port_context *apc)
2530 {
2531 struct mana_query_gf_stat_resp resp = {};
2532 struct mana_query_gf_stat_req req = {};
2533 struct net_device *ndev = apc->ndev;
2534 int err;
2535
2536 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_GF_STAT,
2537 sizeof(req), sizeof(resp));
2538 req.hdr.resp.msg_version = GDMA_MESSAGE_V2;
2539 req.req_stats = STATISTICS_FLAGS_RX_DISCARDS_NO_WQE |
2540 STATISTICS_FLAGS_RX_ERRORS_VPORT_DISABLED |
2541 STATISTICS_FLAGS_HC_RX_BYTES |
2542 STATISTICS_FLAGS_HC_RX_UCAST_PACKETS |
2543 STATISTICS_FLAGS_HC_RX_UCAST_BYTES |
2544 STATISTICS_FLAGS_HC_RX_MCAST_PACKETS |
2545 STATISTICS_FLAGS_HC_RX_MCAST_BYTES |
2546 STATISTICS_FLAGS_HC_RX_BCAST_PACKETS |
2547 STATISTICS_FLAGS_HC_RX_BCAST_BYTES |
2548 STATISTICS_FLAGS_TX_ERRORS_GF_DISABLED |
2549 STATISTICS_FLAGS_TX_ERRORS_VPORT_DISABLED |
2550 STATISTICS_FLAGS_TX_ERRORS_INVAL_VPORT_OFFSET_PACKETS |
2551 STATISTICS_FLAGS_TX_ERRORS_VLAN_ENFORCEMENT |
2552 STATISTICS_FLAGS_TX_ERRORS_ETH_TYPE_ENFORCEMENT |
2553 STATISTICS_FLAGS_TX_ERRORS_SA_ENFORCEMENT |
2554 STATISTICS_FLAGS_TX_ERRORS_SQPDID_ENFORCEMENT |
2555 STATISTICS_FLAGS_TX_ERRORS_CQPDID_ENFORCEMENT |
2556 STATISTICS_FLAGS_TX_ERRORS_MTU_VIOLATION |
2557 STATISTICS_FLAGS_TX_ERRORS_INVALID_OOB |
2558 STATISTICS_FLAGS_HC_TX_BYTES |
2559 STATISTICS_FLAGS_HC_TX_UCAST_PACKETS |
2560 STATISTICS_FLAGS_HC_TX_UCAST_BYTES |
2561 STATISTICS_FLAGS_HC_TX_MCAST_PACKETS |
2562 STATISTICS_FLAGS_HC_TX_MCAST_BYTES |
2563 STATISTICS_FLAGS_HC_TX_BCAST_PACKETS |
2564 STATISTICS_FLAGS_HC_TX_BCAST_BYTES |
2565 STATISTICS_FLAGS_TX_ERRORS_GDMA_ERROR;
2566
2567 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
2568 sizeof(resp));
2569 if (err) {
2570 netdev_err(ndev, "Failed to query GF stats: %d\n", err);
2571 return;
2572 }
2573 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_GF_STAT,
2574 sizeof(resp));
2575 if (err || resp.hdr.status) {
2576 netdev_err(ndev, "Failed to query GF stats: %d, 0x%x\n", err,
2577 resp.hdr.status);
2578 return;
2579 }
2580
2581 apc->eth_stats.hc_rx_discards_no_wqe = resp.rx_discards_nowqe;
2582 apc->eth_stats.hc_rx_err_vport_disabled = resp.rx_err_vport_disabled;
2583 apc->eth_stats.hc_rx_bytes = resp.hc_rx_bytes;
2584 apc->eth_stats.hc_rx_ucast_pkts = resp.hc_rx_ucast_pkts;
2585 apc->eth_stats.hc_rx_ucast_bytes = resp.hc_rx_ucast_bytes;
2586 apc->eth_stats.hc_rx_bcast_pkts = resp.hc_rx_bcast_pkts;
2587 apc->eth_stats.hc_rx_bcast_bytes = resp.hc_rx_bcast_bytes;
2588 apc->eth_stats.hc_rx_mcast_pkts = resp.hc_rx_mcast_pkts;
2589 apc->eth_stats.hc_rx_mcast_bytes = resp.hc_rx_mcast_bytes;
2590 apc->eth_stats.hc_tx_err_gf_disabled = resp.tx_err_gf_disabled;
2591 apc->eth_stats.hc_tx_err_vport_disabled = resp.tx_err_vport_disabled;
2592 apc->eth_stats.hc_tx_err_inval_vportoffset_pkt =
2593 resp.tx_err_inval_vport_offset_pkt;
2594 apc->eth_stats.hc_tx_err_vlan_enforcement =
2595 resp.tx_err_vlan_enforcement;
2596 apc->eth_stats.hc_tx_err_eth_type_enforcement =
2597 resp.tx_err_ethtype_enforcement;
2598 apc->eth_stats.hc_tx_err_sa_enforcement = resp.tx_err_SA_enforcement;
2599 apc->eth_stats.hc_tx_err_sqpdid_enforcement =
2600 resp.tx_err_SQPDID_enforcement;
2601 apc->eth_stats.hc_tx_err_cqpdid_enforcement =
2602 resp.tx_err_CQPDID_enforcement;
2603 apc->eth_stats.hc_tx_err_mtu_violation = resp.tx_err_mtu_violation;
2604 apc->eth_stats.hc_tx_err_inval_oob = resp.tx_err_inval_oob;
2605 apc->eth_stats.hc_tx_bytes = resp.hc_tx_bytes;
2606 apc->eth_stats.hc_tx_ucast_pkts = resp.hc_tx_ucast_pkts;
2607 apc->eth_stats.hc_tx_ucast_bytes = resp.hc_tx_ucast_bytes;
2608 apc->eth_stats.hc_tx_bcast_pkts = resp.hc_tx_bcast_pkts;
2609 apc->eth_stats.hc_tx_bcast_bytes = resp.hc_tx_bcast_bytes;
2610 apc->eth_stats.hc_tx_mcast_pkts = resp.hc_tx_mcast_pkts;
2611 apc->eth_stats.hc_tx_mcast_bytes = resp.hc_tx_mcast_bytes;
2612 apc->eth_stats.hc_tx_err_gdma = resp.tx_err_gdma;
2613 }
2614
mana_init_port(struct net_device * ndev)2615 static int mana_init_port(struct net_device *ndev)
2616 {
2617 struct mana_port_context *apc = netdev_priv(ndev);
2618 struct gdma_dev *gd = apc->ac->gdma_dev;
2619 u32 max_txq, max_rxq, max_queues;
2620 int port_idx = apc->port_idx;
2621 struct gdma_context *gc;
2622 char vport[32];
2623 int err;
2624
2625 err = mana_init_port_context(apc);
2626 if (err)
2627 return err;
2628
2629 gc = gd->gdma_context;
2630
2631 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq,
2632 &apc->indir_table_sz);
2633 if (err) {
2634 netdev_err(ndev, "Failed to query info for vPort %d\n",
2635 port_idx);
2636 goto reset_apc;
2637 }
2638
2639 max_queues = min_t(u32, max_txq, max_rxq);
2640 if (apc->max_queues > max_queues)
2641 apc->max_queues = max_queues;
2642
2643 if (apc->num_queues > apc->max_queues)
2644 apc->num_queues = apc->max_queues;
2645
2646 eth_hw_addr_set(ndev, apc->mac_addr);
2647 sprintf(vport, "vport%d", port_idx);
2648 apc->mana_port_debugfs = debugfs_create_dir(vport, gc->mana_pci_debugfs);
2649 return 0;
2650
2651 reset_apc:
2652 mana_cleanup_port_context(apc);
2653 return err;
2654 }
2655
mana_alloc_queues(struct net_device * ndev)2656 int mana_alloc_queues(struct net_device *ndev)
2657 {
2658 struct mana_port_context *apc = netdev_priv(ndev);
2659 struct gdma_dev *gd = apc->ac->gdma_dev;
2660 int err;
2661
2662 err = mana_create_vport(apc, ndev);
2663 if (err) {
2664 netdev_err(ndev, "Failed to create vPort %u : %d\n", apc->port_idx, err);
2665 return err;
2666 }
2667
2668 err = netif_set_real_num_tx_queues(ndev, apc->num_queues);
2669 if (err) {
2670 netdev_err(ndev,
2671 "netif_set_real_num_tx_queues () failed for ndev with num_queues %u : %d\n",
2672 apc->num_queues, err);
2673 goto destroy_vport;
2674 }
2675
2676 err = mana_add_rx_queues(apc, ndev);
2677 if (err)
2678 goto destroy_vport;
2679
2680 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE;
2681
2682 err = netif_set_real_num_rx_queues(ndev, apc->num_queues);
2683 if (err) {
2684 netdev_err(ndev,
2685 "netif_set_real_num_rx_queues () failed for ndev with num_queues %u : %d\n",
2686 apc->num_queues, err);
2687 goto destroy_vport;
2688 }
2689
2690 mana_rss_table_init(apc);
2691
2692 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true);
2693 if (err) {
2694 netdev_err(ndev, "Failed to configure RSS table: %d\n", err);
2695 goto destroy_vport;
2696 }
2697
2698 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) {
2699 err = mana_pf_register_filter(apc);
2700 if (err)
2701 goto destroy_vport;
2702 }
2703
2704 mana_chn_setxdp(apc, mana_xdp_get(apc));
2705
2706 return 0;
2707
2708 destroy_vport:
2709 mana_destroy_vport(apc);
2710 return err;
2711 }
2712
mana_attach(struct net_device * ndev)2713 int mana_attach(struct net_device *ndev)
2714 {
2715 struct mana_port_context *apc = netdev_priv(ndev);
2716 int err;
2717
2718 ASSERT_RTNL();
2719
2720 err = mana_init_port(ndev);
2721 if (err)
2722 return err;
2723
2724 if (apc->port_st_save) {
2725 err = mana_alloc_queues(ndev);
2726 if (err) {
2727 mana_cleanup_port_context(apc);
2728 return err;
2729 }
2730 }
2731
2732 apc->port_is_up = apc->port_st_save;
2733
2734 /* Ensure port state updated before txq state */
2735 smp_wmb();
2736
2737 if (apc->port_is_up)
2738 netif_carrier_on(ndev);
2739
2740 netif_device_attach(ndev);
2741
2742 return 0;
2743 }
2744
mana_dealloc_queues(struct net_device * ndev)2745 static int mana_dealloc_queues(struct net_device *ndev)
2746 {
2747 struct mana_port_context *apc = netdev_priv(ndev);
2748 unsigned long timeout = jiffies + 120 * HZ;
2749 struct gdma_dev *gd = apc->ac->gdma_dev;
2750 struct mana_txq *txq;
2751 struct sk_buff *skb;
2752 int i, err;
2753 u32 tsleep;
2754
2755 if (apc->port_is_up)
2756 return -EINVAL;
2757
2758 mana_chn_setxdp(apc, NULL);
2759
2760 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode)
2761 mana_pf_deregister_filter(apc);
2762
2763 /* No packet can be transmitted now since apc->port_is_up is false.
2764 * There is still a tiny chance that mana_poll_tx_cq() can re-enable
2765 * a txq because it may not timely see apc->port_is_up being cleared
2766 * to false, but it doesn't matter since mana_start_xmit() drops any
2767 * new packets due to apc->port_is_up being false.
2768 *
2769 * Drain all the in-flight TX packets.
2770 * A timeout of 120 seconds for all the queues is used.
2771 * This will break the while loop when h/w is not responding.
2772 * This value of 120 has been decided here considering max
2773 * number of queues.
2774 */
2775
2776 for (i = 0; i < apc->num_queues; i++) {
2777 txq = &apc->tx_qp[i].txq;
2778 tsleep = 1000;
2779 while (atomic_read(&txq->pending_sends) > 0 &&
2780 time_before(jiffies, timeout)) {
2781 usleep_range(tsleep, tsleep + 1000);
2782 tsleep <<= 1;
2783 }
2784 if (atomic_read(&txq->pending_sends)) {
2785 err = pcie_flr(to_pci_dev(gd->gdma_context->dev));
2786 if (err) {
2787 netdev_err(ndev, "flr failed %d with %d pkts pending in txq %u\n",
2788 err, atomic_read(&txq->pending_sends),
2789 txq->gdma_txq_id);
2790 }
2791 break;
2792 }
2793 }
2794
2795 for (i = 0; i < apc->num_queues; i++) {
2796 txq = &apc->tx_qp[i].txq;
2797 while ((skb = skb_dequeue(&txq->pending_skbs))) {
2798 mana_unmap_skb(skb, apc);
2799 dev_kfree_skb_any(skb);
2800 }
2801 atomic_set(&txq->pending_sends, 0);
2802 }
2803 /* We're 100% sure the queues can no longer be woken up, because
2804 * we're sure now mana_poll_tx_cq() can't be running.
2805 */
2806
2807 apc->rss_state = TRI_STATE_FALSE;
2808 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false);
2809 if (err) {
2810 netdev_err(ndev, "Failed to disable vPort: %d\n", err);
2811 return err;
2812 }
2813
2814 mana_destroy_vport(apc);
2815
2816 return 0;
2817 }
2818
mana_detach(struct net_device * ndev,bool from_close)2819 int mana_detach(struct net_device *ndev, bool from_close)
2820 {
2821 struct mana_port_context *apc = netdev_priv(ndev);
2822 int err;
2823
2824 ASSERT_RTNL();
2825
2826 apc->port_st_save = apc->port_is_up;
2827 apc->port_is_up = false;
2828
2829 /* Ensure port state updated before txq state */
2830 smp_wmb();
2831
2832 netif_tx_disable(ndev);
2833 netif_carrier_off(ndev);
2834
2835 if (apc->port_st_save) {
2836 err = mana_dealloc_queues(ndev);
2837 if (err) {
2838 netdev_err(ndev, "%s failed to deallocate queues: %d\n", __func__, err);
2839 return err;
2840 }
2841 }
2842
2843 if (!from_close) {
2844 netif_device_detach(ndev);
2845 mana_cleanup_port_context(apc);
2846 }
2847
2848 return 0;
2849 }
2850
mana_probe_port(struct mana_context * ac,int port_idx,struct net_device ** ndev_storage)2851 static int mana_probe_port(struct mana_context *ac, int port_idx,
2852 struct net_device **ndev_storage)
2853 {
2854 struct gdma_context *gc = ac->gdma_dev->gdma_context;
2855 struct mana_port_context *apc;
2856 struct net_device *ndev;
2857 int err;
2858
2859 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context),
2860 gc->max_num_queues);
2861 if (!ndev)
2862 return -ENOMEM;
2863
2864 *ndev_storage = ndev;
2865
2866 apc = netdev_priv(ndev);
2867 apc->ac = ac;
2868 apc->ndev = ndev;
2869 apc->max_queues = gc->max_num_queues;
2870 apc->num_queues = gc->max_num_queues;
2871 apc->tx_queue_size = DEF_TX_BUFFERS_PER_QUEUE;
2872 apc->rx_queue_size = DEF_RX_BUFFERS_PER_QUEUE;
2873 apc->port_handle = INVALID_MANA_HANDLE;
2874 apc->pf_filter_handle = INVALID_MANA_HANDLE;
2875 apc->port_idx = port_idx;
2876
2877 mutex_init(&apc->vport_mutex);
2878 apc->vport_use_count = 0;
2879
2880 ndev->netdev_ops = &mana_devops;
2881 ndev->ethtool_ops = &mana_ethtool_ops;
2882 ndev->mtu = ETH_DATA_LEN;
2883 ndev->max_mtu = gc->adapter_mtu - ETH_HLEN;
2884 ndev->min_mtu = ETH_MIN_MTU;
2885 ndev->needed_headroom = MANA_HEADROOM;
2886 ndev->dev_port = port_idx;
2887 SET_NETDEV_DEV(ndev, gc->dev);
2888
2889 netif_set_tso_max_size(ndev, GSO_MAX_SIZE);
2890
2891 netif_carrier_off(ndev);
2892
2893 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE);
2894
2895 err = mana_init_port(ndev);
2896 if (err)
2897 goto free_net;
2898
2899 err = mana_rss_table_alloc(apc);
2900 if (err)
2901 goto reset_apc;
2902
2903 netdev_lockdep_set_classes(ndev);
2904
2905 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
2906 ndev->hw_features |= NETIF_F_RXCSUM;
2907 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
2908 ndev->hw_features |= NETIF_F_RXHASH;
2909 ndev->features = ndev->hw_features | NETIF_F_HW_VLAN_CTAG_TX |
2910 NETIF_F_HW_VLAN_CTAG_RX;
2911 ndev->vlan_features = ndev->features;
2912 xdp_set_features_flag(ndev, NETDEV_XDP_ACT_BASIC |
2913 NETDEV_XDP_ACT_REDIRECT |
2914 NETDEV_XDP_ACT_NDO_XMIT);
2915
2916 err = register_netdev(ndev);
2917 if (err) {
2918 netdev_err(ndev, "Unable to register netdev.\n");
2919 goto free_indir;
2920 }
2921
2922 return 0;
2923
2924 free_indir:
2925 mana_cleanup_indir_table(apc);
2926 reset_apc:
2927 mana_cleanup_port_context(apc);
2928 free_net:
2929 *ndev_storage = NULL;
2930 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err);
2931 free_netdev(ndev);
2932 return err;
2933 }
2934
adev_release(struct device * dev)2935 static void adev_release(struct device *dev)
2936 {
2937 struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev);
2938
2939 kfree(madev);
2940 }
2941
remove_adev(struct gdma_dev * gd)2942 static void remove_adev(struct gdma_dev *gd)
2943 {
2944 struct auxiliary_device *adev = gd->adev;
2945 int id = adev->id;
2946
2947 auxiliary_device_delete(adev);
2948 auxiliary_device_uninit(adev);
2949
2950 mana_adev_idx_free(id);
2951 gd->adev = NULL;
2952 }
2953
add_adev(struct gdma_dev * gd,const char * name)2954 static int add_adev(struct gdma_dev *gd, const char *name)
2955 {
2956 struct auxiliary_device *adev;
2957 struct mana_adev *madev;
2958 int ret;
2959
2960 madev = kzalloc(sizeof(*madev), GFP_KERNEL);
2961 if (!madev)
2962 return -ENOMEM;
2963
2964 adev = &madev->adev;
2965 ret = mana_adev_idx_alloc();
2966 if (ret < 0)
2967 goto idx_fail;
2968 adev->id = ret;
2969
2970 adev->name = name;
2971 adev->dev.parent = gd->gdma_context->dev;
2972 adev->dev.release = adev_release;
2973 madev->mdev = gd;
2974
2975 ret = auxiliary_device_init(adev);
2976 if (ret)
2977 goto init_fail;
2978
2979 /* madev is owned by the auxiliary device */
2980 madev = NULL;
2981 ret = auxiliary_device_add(adev);
2982 if (ret)
2983 goto add_fail;
2984
2985 gd->adev = adev;
2986 dev_dbg(gd->gdma_context->dev,
2987 "Auxiliary device added successfully\n");
2988 return 0;
2989
2990 add_fail:
2991 auxiliary_device_uninit(adev);
2992
2993 init_fail:
2994 mana_adev_idx_free(adev->id);
2995
2996 idx_fail:
2997 kfree(madev);
2998
2999 return ret;
3000 }
3001
mana_rdma_service_handle(struct work_struct * work)3002 static void mana_rdma_service_handle(struct work_struct *work)
3003 {
3004 struct mana_service_work *serv_work =
3005 container_of(work, struct mana_service_work, work);
3006 struct gdma_dev *gd = serv_work->gdma_dev;
3007 struct device *dev = gd->gdma_context->dev;
3008 int ret;
3009
3010 if (READ_ONCE(gd->rdma_teardown))
3011 goto out;
3012
3013 switch (serv_work->event) {
3014 case GDMA_SERVICE_TYPE_RDMA_SUSPEND:
3015 if (!gd->adev || gd->is_suspended)
3016 break;
3017
3018 remove_adev(gd);
3019 gd->is_suspended = true;
3020 break;
3021
3022 case GDMA_SERVICE_TYPE_RDMA_RESUME:
3023 if (!gd->is_suspended)
3024 break;
3025
3026 ret = add_adev(gd, "rdma");
3027 if (ret)
3028 dev_err(dev, "Failed to add adev on resume: %d\n", ret);
3029 else
3030 gd->is_suspended = false;
3031 break;
3032
3033 default:
3034 dev_warn(dev, "unknown adev service event %u\n",
3035 serv_work->event);
3036 break;
3037 }
3038
3039 out:
3040 kfree(serv_work);
3041 }
3042
mana_rdma_service_event(struct gdma_context * gc,enum gdma_service_type event)3043 int mana_rdma_service_event(struct gdma_context *gc, enum gdma_service_type event)
3044 {
3045 struct gdma_dev *gd = &gc->mana_ib;
3046 struct mana_service_work *serv_work;
3047
3048 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) {
3049 /* RDMA device is not detected on pci */
3050 return 0;
3051 }
3052
3053 serv_work = kzalloc(sizeof(*serv_work), GFP_ATOMIC);
3054 if (!serv_work)
3055 return -ENOMEM;
3056
3057 serv_work->event = event;
3058 serv_work->gdma_dev = gd;
3059
3060 INIT_WORK(&serv_work->work, mana_rdma_service_handle);
3061 queue_work(gc->service_wq, &serv_work->work);
3062
3063 return 0;
3064 }
3065
mana_probe(struct gdma_dev * gd,bool resuming)3066 int mana_probe(struct gdma_dev *gd, bool resuming)
3067 {
3068 struct gdma_context *gc = gd->gdma_context;
3069 struct mana_context *ac = gd->driver_data;
3070 struct device *dev = gc->dev;
3071 u8 bm_hostmode = 0;
3072 u16 num_ports = 0;
3073 int err;
3074 int i;
3075
3076 dev_info(dev,
3077 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n",
3078 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION);
3079
3080 err = mana_gd_register_device(gd);
3081 if (err)
3082 return err;
3083
3084 if (!resuming) {
3085 ac = kzalloc(sizeof(*ac), GFP_KERNEL);
3086 if (!ac)
3087 return -ENOMEM;
3088
3089 ac->gdma_dev = gd;
3090 gd->driver_data = ac;
3091 }
3092
3093 err = mana_create_eq(ac);
3094 if (err) {
3095 dev_err(dev, "Failed to create EQs: %d\n", err);
3096 goto out;
3097 }
3098
3099 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION,
3100 MANA_MICRO_VERSION, &num_ports, &bm_hostmode);
3101 if (err)
3102 goto out;
3103
3104 ac->bm_hostmode = bm_hostmode;
3105
3106 if (!resuming) {
3107 ac->num_ports = num_ports;
3108 } else {
3109 if (ac->num_ports != num_ports) {
3110 dev_err(dev, "The number of vPorts changed: %d->%d\n",
3111 ac->num_ports, num_ports);
3112 err = -EPROTO;
3113 goto out;
3114 }
3115 }
3116
3117 if (ac->num_ports == 0)
3118 dev_err(dev, "Failed to detect any vPort\n");
3119
3120 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV)
3121 ac->num_ports = MAX_PORTS_IN_MANA_DEV;
3122
3123 if (!resuming) {
3124 for (i = 0; i < ac->num_ports; i++) {
3125 err = mana_probe_port(ac, i, &ac->ports[i]);
3126 /* we log the port for which the probe failed and stop
3127 * probes for subsequent ports.
3128 * Note that we keep running ports, for which the probes
3129 * were successful, unless add_adev fails too
3130 */
3131 if (err) {
3132 dev_err(dev, "Probe Failed for port %d\n", i);
3133 break;
3134 }
3135 }
3136 } else {
3137 for (i = 0; i < ac->num_ports; i++) {
3138 rtnl_lock();
3139 err = mana_attach(ac->ports[i]);
3140 rtnl_unlock();
3141 /* we log the port for which the attach failed and stop
3142 * attach for subsequent ports
3143 * Note that we keep running ports, for which the attach
3144 * were successful, unless add_adev fails too
3145 */
3146 if (err) {
3147 dev_err(dev, "Attach Failed for port %d\n", i);
3148 break;
3149 }
3150 }
3151 }
3152
3153 err = add_adev(gd, "eth");
3154 out:
3155 if (err) {
3156 mana_remove(gd, false);
3157 } else {
3158 dev_dbg(dev, "gd=%p, id=%u, num_ports=%d, type=%u, instance=%u\n",
3159 gd, gd->dev_id.as_uint32, ac->num_ports,
3160 gd->dev_id.type, gd->dev_id.instance);
3161 dev_dbg(dev, "%s succeeded\n", __func__);
3162 }
3163
3164 return err;
3165 }
3166
mana_remove(struct gdma_dev * gd,bool suspending)3167 void mana_remove(struct gdma_dev *gd, bool suspending)
3168 {
3169 struct gdma_context *gc = gd->gdma_context;
3170 struct mana_context *ac = gd->driver_data;
3171 struct mana_port_context *apc;
3172 struct device *dev = gc->dev;
3173 struct net_device *ndev;
3174 int err;
3175 int i;
3176
3177 /* adev currently doesn't support suspending, always remove it */
3178 if (gd->adev)
3179 remove_adev(gd);
3180
3181 for (i = 0; i < ac->num_ports; i++) {
3182 ndev = ac->ports[i];
3183 apc = netdev_priv(ndev);
3184 if (!ndev) {
3185 if (i == 0)
3186 dev_err(dev, "No net device to remove\n");
3187 goto out;
3188 }
3189
3190 /* All cleanup actions should stay after rtnl_lock(), otherwise
3191 * other functions may access partially cleaned up data.
3192 */
3193 rtnl_lock();
3194
3195 err = mana_detach(ndev, false);
3196 if (err)
3197 netdev_err(ndev, "Failed to detach vPort %d: %d\n",
3198 i, err);
3199
3200 if (suspending) {
3201 /* No need to unregister the ndev. */
3202 rtnl_unlock();
3203 continue;
3204 }
3205
3206 unregister_netdevice(ndev);
3207 mana_cleanup_indir_table(apc);
3208
3209 rtnl_unlock();
3210
3211 free_netdev(ndev);
3212 }
3213
3214 mana_destroy_eq(ac);
3215 out:
3216 mana_gd_deregister_device(gd);
3217
3218 if (suspending)
3219 return;
3220
3221 gd->driver_data = NULL;
3222 gd->gdma_context = NULL;
3223 kfree(ac);
3224 dev_dbg(dev, "%s succeeded\n", __func__);
3225 }
3226
mana_rdma_probe(struct gdma_dev * gd)3227 int mana_rdma_probe(struct gdma_dev *gd)
3228 {
3229 int err = 0;
3230
3231 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) {
3232 /* RDMA device is not detected on pci */
3233 return err;
3234 }
3235
3236 err = mana_gd_register_device(gd);
3237 if (err)
3238 return err;
3239
3240 err = add_adev(gd, "rdma");
3241 if (err)
3242 mana_gd_deregister_device(gd);
3243
3244 return err;
3245 }
3246
mana_rdma_remove(struct gdma_dev * gd)3247 void mana_rdma_remove(struct gdma_dev *gd)
3248 {
3249 struct gdma_context *gc = gd->gdma_context;
3250
3251 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) {
3252 /* RDMA device is not detected on pci */
3253 return;
3254 }
3255
3256 WRITE_ONCE(gd->rdma_teardown, true);
3257 flush_workqueue(gc->service_wq);
3258
3259 if (gd->adev)
3260 remove_adev(gd);
3261
3262 mana_gd_deregister_device(gd);
3263 }
3264
mana_get_primary_netdev(struct mana_context * ac,u32 port_index,netdevice_tracker * tracker)3265 struct net_device *mana_get_primary_netdev(struct mana_context *ac,
3266 u32 port_index,
3267 netdevice_tracker *tracker)
3268 {
3269 struct net_device *ndev;
3270
3271 if (port_index >= ac->num_ports)
3272 return NULL;
3273
3274 rcu_read_lock();
3275
3276 /* If mana is used in netvsc, the upper netdevice should be returned. */
3277 ndev = netdev_master_upper_dev_get_rcu(ac->ports[port_index]);
3278
3279 /* If there is no upper device, use the parent Ethernet device */
3280 if (!ndev)
3281 ndev = ac->ports[port_index];
3282
3283 netdev_hold(ndev, tracker, GFP_ATOMIC);
3284 rcu_read_unlock();
3285
3286 return ndev;
3287 }
3288 EXPORT_SYMBOL_NS(mana_get_primary_netdev, "NET_MANA");
3289