1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /* Copyright (c) 2021, Microsoft Corporation. */
3
4 #include <uapi/linux/bpf.h>
5
6 #include <linux/debugfs.h>
7 #include <linux/inetdevice.h>
8 #include <linux/etherdevice.h>
9 #include <linux/ethtool.h>
10 #include <linux/filter.h>
11 #include <linux/mm.h>
12 #include <linux/pci.h>
13
14 #include <net/checksum.h>
15 #include <net/ip6_checksum.h>
16 #include <net/netdev_lock.h>
17 #include <net/page_pool/helpers.h>
18 #include <net/xdp.h>
19
20 #include <net/mana/mana.h>
21 #include <net/mana/mana_auxiliary.h>
22
23 static DEFINE_IDA(mana_adev_ida);
24
mana_adev_idx_alloc(void)25 static int mana_adev_idx_alloc(void)
26 {
27 return ida_alloc(&mana_adev_ida, GFP_KERNEL);
28 }
29
mana_adev_idx_free(int idx)30 static void mana_adev_idx_free(int idx)
31 {
32 ida_free(&mana_adev_ida, idx);
33 }
34
mana_dbg_q_read(struct file * filp,char __user * buf,size_t count,loff_t * pos)35 static ssize_t mana_dbg_q_read(struct file *filp, char __user *buf, size_t count,
36 loff_t *pos)
37 {
38 struct gdma_queue *gdma_q = filp->private_data;
39
40 return simple_read_from_buffer(buf, count, pos, gdma_q->queue_mem_ptr,
41 gdma_q->queue_size);
42 }
43
44 static const struct file_operations mana_dbg_q_fops = {
45 .owner = THIS_MODULE,
46 .open = simple_open,
47 .read = mana_dbg_q_read,
48 };
49
50 /* Microsoft Azure Network Adapter (MANA) functions */
51
mana_open(struct net_device * ndev)52 static int mana_open(struct net_device *ndev)
53 {
54 struct mana_port_context *apc = netdev_priv(ndev);
55 int err;
56 err = mana_alloc_queues(ndev);
57
58 if (err) {
59 netdev_err(ndev, "%s failed to allocate queues: %d\n", __func__, err);
60 return err;
61 }
62
63 apc->port_is_up = true;
64
65 /* Ensure port state updated before txq state */
66 smp_wmb();
67
68 netif_carrier_on(ndev);
69 netif_tx_wake_all_queues(ndev);
70 netdev_dbg(ndev, "%s successful\n", __func__);
71 return 0;
72 }
73
mana_close(struct net_device * ndev)74 static int mana_close(struct net_device *ndev)
75 {
76 struct mana_port_context *apc = netdev_priv(ndev);
77
78 if (!apc->port_is_up)
79 return 0;
80
81 return mana_detach(ndev, true);
82 }
83
mana_can_tx(struct gdma_queue * wq)84 static bool mana_can_tx(struct gdma_queue *wq)
85 {
86 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE;
87 }
88
mana_checksum_info(struct sk_buff * skb)89 static unsigned int mana_checksum_info(struct sk_buff *skb)
90 {
91 if (skb->protocol == htons(ETH_P_IP)) {
92 struct iphdr *ip = ip_hdr(skb);
93
94 if (ip->protocol == IPPROTO_TCP)
95 return IPPROTO_TCP;
96
97 if (ip->protocol == IPPROTO_UDP)
98 return IPPROTO_UDP;
99 } else if (skb->protocol == htons(ETH_P_IPV6)) {
100 struct ipv6hdr *ip6 = ipv6_hdr(skb);
101
102 if (ip6->nexthdr == IPPROTO_TCP)
103 return IPPROTO_TCP;
104
105 if (ip6->nexthdr == IPPROTO_UDP)
106 return IPPROTO_UDP;
107 }
108
109 /* No csum offloading */
110 return 0;
111 }
112
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)113 static void mana_add_sge(struct mana_tx_package *tp, struct mana_skb_head *ash,
114 int sg_i, dma_addr_t da, int sge_len, u32 gpa_mkey)
115 {
116 ash->dma_handle[sg_i] = da;
117 ash->size[sg_i] = sge_len;
118
119 tp->wqe_req.sgl[sg_i].address = da;
120 tp->wqe_req.sgl[sg_i].mem_key = gpa_mkey;
121 tp->wqe_req.sgl[sg_i].size = sge_len;
122 }
123
mana_map_skb(struct sk_buff * skb,struct mana_port_context * apc,struct mana_tx_package * tp,int gso_hs)124 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc,
125 struct mana_tx_package *tp, int gso_hs)
126 {
127 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
128 int hsg = 1; /* num of SGEs of linear part */
129 struct gdma_dev *gd = apc->ac->gdma_dev;
130 int skb_hlen = skb_headlen(skb);
131 int sge0_len, sge1_len = 0;
132 struct gdma_context *gc;
133 struct device *dev;
134 skb_frag_t *frag;
135 dma_addr_t da;
136 int sg_i;
137 int i;
138
139 gc = gd->gdma_context;
140 dev = gc->dev;
141
142 if (gso_hs && gso_hs < skb_hlen) {
143 sge0_len = gso_hs;
144 sge1_len = skb_hlen - gso_hs;
145 } else {
146 sge0_len = skb_hlen;
147 }
148
149 da = dma_map_single(dev, skb->data, sge0_len, DMA_TO_DEVICE);
150 if (dma_mapping_error(dev, da))
151 return -ENOMEM;
152
153 mana_add_sge(tp, ash, 0, da, sge0_len, gd->gpa_mkey);
154
155 if (sge1_len) {
156 sg_i = 1;
157 da = dma_map_single(dev, skb->data + sge0_len, sge1_len,
158 DMA_TO_DEVICE);
159 if (dma_mapping_error(dev, da))
160 goto frag_err;
161
162 mana_add_sge(tp, ash, sg_i, da, sge1_len, gd->gpa_mkey);
163 hsg = 2;
164 }
165
166 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
167 sg_i = hsg + i;
168
169 frag = &skb_shinfo(skb)->frags[i];
170 da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag),
171 DMA_TO_DEVICE);
172 if (dma_mapping_error(dev, da))
173 goto frag_err;
174
175 mana_add_sge(tp, ash, sg_i, da, skb_frag_size(frag),
176 gd->gpa_mkey);
177 }
178
179 return 0;
180
181 frag_err:
182 if (net_ratelimit())
183 netdev_err(apc->ndev, "Failed to map skb of size %u to DMA\n",
184 skb->len);
185 for (i = sg_i - 1; i >= hsg; i--)
186 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
187 DMA_TO_DEVICE);
188
189 for (i = hsg - 1; i >= 0; i--)
190 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i],
191 DMA_TO_DEVICE);
192
193 return -ENOMEM;
194 }
195
196 /* Handle the case when GSO SKB linear length is too large.
197 * MANA NIC requires GSO packets to put only the packet header to SGE0.
198 * So, we need 2 SGEs for the skb linear part which contains more than the
199 * header.
200 * Return a positive value for the number of SGEs, or a negative value
201 * for an error.
202 */
mana_fix_skb_head(struct net_device * ndev,struct sk_buff * skb,int gso_hs)203 static int mana_fix_skb_head(struct net_device *ndev, struct sk_buff *skb,
204 int gso_hs)
205 {
206 int num_sge = 1 + skb_shinfo(skb)->nr_frags;
207 int skb_hlen = skb_headlen(skb);
208
209 if (gso_hs < skb_hlen) {
210 num_sge++;
211 } else if (gso_hs > skb_hlen) {
212 if (net_ratelimit())
213 netdev_err(ndev,
214 "TX nonlinear head: hs:%d, skb_hlen:%d\n",
215 gso_hs, skb_hlen);
216
217 return -EINVAL;
218 }
219
220 return num_sge;
221 }
222
223 /* Get the GSO packet's header size */
mana_get_gso_hs(struct sk_buff * skb)224 static int mana_get_gso_hs(struct sk_buff *skb)
225 {
226 int gso_hs;
227
228 if (skb->encapsulation) {
229 gso_hs = skb_inner_tcp_all_headers(skb);
230 } else {
231 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
232 gso_hs = skb_transport_offset(skb) +
233 sizeof(struct udphdr);
234 } else {
235 gso_hs = skb_tcp_all_headers(skb);
236 }
237 }
238
239 return gso_hs;
240 }
241
mana_start_xmit(struct sk_buff * skb,struct net_device * ndev)242 netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev)
243 {
244 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT;
245 struct mana_port_context *apc = netdev_priv(ndev);
246 int gso_hs = 0; /* zero for non-GSO pkts */
247 u16 txq_idx = skb_get_queue_mapping(skb);
248 struct gdma_dev *gd = apc->ac->gdma_dev;
249 bool ipv4 = false, ipv6 = false;
250 struct mana_tx_package pkg = {};
251 struct netdev_queue *net_txq;
252 struct mana_stats_tx *tx_stats;
253 struct gdma_queue *gdma_sq;
254 unsigned int csum_type;
255 struct mana_txq *txq;
256 struct mana_cq *cq;
257 int err, len;
258
259 if (unlikely(!apc->port_is_up))
260 goto tx_drop;
261
262 if (skb_cow_head(skb, MANA_HEADROOM))
263 goto tx_drop_count;
264
265 if (unlikely(ipv6_hopopt_jumbo_remove(skb)))
266 goto tx_drop_count;
267
268 txq = &apc->tx_qp[txq_idx].txq;
269 gdma_sq = txq->gdma_sq;
270 cq = &apc->tx_qp[txq_idx].tx_cq;
271 tx_stats = &txq->stats;
272
273 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id;
274 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame;
275
276 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) {
277 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset;
278 pkt_fmt = MANA_LONG_PKT_FMT;
279 } else {
280 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset;
281 }
282
283 if (skb_vlan_tag_present(skb)) {
284 pkt_fmt = MANA_LONG_PKT_FMT;
285 pkg.tx_oob.l_oob.inject_vlan_pri_tag = 1;
286 pkg.tx_oob.l_oob.pcp = skb_vlan_tag_get_prio(skb);
287 pkg.tx_oob.l_oob.dei = skb_vlan_tag_get_cfi(skb);
288 pkg.tx_oob.l_oob.vlan_id = skb_vlan_tag_get_id(skb);
289 }
290
291 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt;
292
293 if (pkt_fmt == MANA_SHORT_PKT_FMT) {
294 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob);
295 u64_stats_update_begin(&tx_stats->syncp);
296 tx_stats->short_pkt_fmt++;
297 u64_stats_update_end(&tx_stats->syncp);
298 } else {
299 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob);
300 u64_stats_update_begin(&tx_stats->syncp);
301 tx_stats->long_pkt_fmt++;
302 u64_stats_update_end(&tx_stats->syncp);
303 }
304
305 pkg.wqe_req.inline_oob_data = &pkg.tx_oob;
306 pkg.wqe_req.flags = 0;
307 pkg.wqe_req.client_data_unit = 0;
308
309 pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags;
310
311 if (skb->protocol == htons(ETH_P_IP))
312 ipv4 = true;
313 else if (skb->protocol == htons(ETH_P_IPV6))
314 ipv6 = true;
315
316 if (skb_is_gso(skb)) {
317 int num_sge;
318
319 gso_hs = mana_get_gso_hs(skb);
320
321 num_sge = mana_fix_skb_head(ndev, skb, gso_hs);
322 if (num_sge > 0)
323 pkg.wqe_req.num_sge = num_sge;
324 else
325 goto tx_drop_count;
326
327 u64_stats_update_begin(&tx_stats->syncp);
328 if (skb->encapsulation) {
329 tx_stats->tso_inner_packets++;
330 tx_stats->tso_inner_bytes += skb->len - gso_hs;
331 } else {
332 tx_stats->tso_packets++;
333 tx_stats->tso_bytes += skb->len - gso_hs;
334 }
335 u64_stats_update_end(&tx_stats->syncp);
336
337 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
338 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
339
340 pkg.tx_oob.s_oob.comp_iphdr_csum = 1;
341 pkg.tx_oob.s_oob.comp_tcp_csum = 1;
342 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
343
344 pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size;
345 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0;
346 if (ipv4) {
347 ip_hdr(skb)->tot_len = 0;
348 ip_hdr(skb)->check = 0;
349 tcp_hdr(skb)->check =
350 ~csum_tcpudp_magic(ip_hdr(skb)->saddr,
351 ip_hdr(skb)->daddr, 0,
352 IPPROTO_TCP, 0);
353 } else {
354 ipv6_hdr(skb)->payload_len = 0;
355 tcp_hdr(skb)->check =
356 ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
357 &ipv6_hdr(skb)->daddr, 0,
358 IPPROTO_TCP, 0);
359 }
360 } else if (skb->ip_summed == CHECKSUM_PARTIAL) {
361 csum_type = mana_checksum_info(skb);
362
363 u64_stats_update_begin(&tx_stats->syncp);
364 tx_stats->csum_partial++;
365 u64_stats_update_end(&tx_stats->syncp);
366
367 if (csum_type == IPPROTO_TCP) {
368 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
369 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
370
371 pkg.tx_oob.s_oob.comp_tcp_csum = 1;
372 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
373
374 } else if (csum_type == IPPROTO_UDP) {
375 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
376 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
377
378 pkg.tx_oob.s_oob.comp_udp_csum = 1;
379 } else {
380 /* Can't do offload of this type of checksum */
381 if (skb_checksum_help(skb))
382 goto tx_drop_count;
383 }
384 }
385
386 WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES);
387
388 if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) {
389 pkg.wqe_req.sgl = pkg.sgl_array;
390 } else {
391 pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge,
392 sizeof(struct gdma_sge),
393 GFP_ATOMIC);
394 if (!pkg.sgl_ptr)
395 goto tx_drop_count;
396
397 pkg.wqe_req.sgl = pkg.sgl_ptr;
398 }
399
400 if (mana_map_skb(skb, apc, &pkg, gso_hs)) {
401 u64_stats_update_begin(&tx_stats->syncp);
402 tx_stats->mana_map_err++;
403 u64_stats_update_end(&tx_stats->syncp);
404 goto free_sgl_ptr;
405 }
406
407 skb_queue_tail(&txq->pending_skbs, skb);
408
409 len = skb->len;
410 net_txq = netdev_get_tx_queue(ndev, txq_idx);
411
412 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req,
413 (struct gdma_posted_wqe_info *)skb->cb);
414 if (!mana_can_tx(gdma_sq)) {
415 netif_tx_stop_queue(net_txq);
416 apc->eth_stats.stop_queue++;
417 }
418
419 if (err) {
420 (void)skb_dequeue_tail(&txq->pending_skbs);
421 netdev_warn(ndev, "Failed to post TX OOB: %d\n", err);
422 err = NETDEV_TX_BUSY;
423 goto tx_busy;
424 }
425
426 err = NETDEV_TX_OK;
427 atomic_inc(&txq->pending_sends);
428
429 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq);
430
431 /* skb may be freed after mana_gd_post_work_request. Do not use it. */
432 skb = NULL;
433
434 tx_stats = &txq->stats;
435 u64_stats_update_begin(&tx_stats->syncp);
436 tx_stats->packets++;
437 tx_stats->bytes += len;
438 u64_stats_update_end(&tx_stats->syncp);
439
440 tx_busy:
441 if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) {
442 netif_tx_wake_queue(net_txq);
443 apc->eth_stats.wake_queue++;
444 }
445
446 kfree(pkg.sgl_ptr);
447 return err;
448
449 free_sgl_ptr:
450 kfree(pkg.sgl_ptr);
451 tx_drop_count:
452 ndev->stats.tx_dropped++;
453 tx_drop:
454 dev_kfree_skb_any(skb);
455 return NETDEV_TX_OK;
456 }
457
mana_get_stats64(struct net_device * ndev,struct rtnl_link_stats64 * st)458 static void mana_get_stats64(struct net_device *ndev,
459 struct rtnl_link_stats64 *st)
460 {
461 struct mana_port_context *apc = netdev_priv(ndev);
462 unsigned int num_queues = apc->num_queues;
463 struct mana_stats_rx *rx_stats;
464 struct mana_stats_tx *tx_stats;
465 unsigned int start;
466 u64 packets, bytes;
467 int q;
468
469 if (!apc->port_is_up)
470 return;
471
472 netdev_stats_to_stats64(st, &ndev->stats);
473
474 for (q = 0; q < num_queues; q++) {
475 rx_stats = &apc->rxqs[q]->stats;
476
477 do {
478 start = u64_stats_fetch_begin(&rx_stats->syncp);
479 packets = rx_stats->packets;
480 bytes = rx_stats->bytes;
481 } while (u64_stats_fetch_retry(&rx_stats->syncp, start));
482
483 st->rx_packets += packets;
484 st->rx_bytes += bytes;
485 }
486
487 for (q = 0; q < num_queues; q++) {
488 tx_stats = &apc->tx_qp[q].txq.stats;
489
490 do {
491 start = u64_stats_fetch_begin(&tx_stats->syncp);
492 packets = tx_stats->packets;
493 bytes = tx_stats->bytes;
494 } while (u64_stats_fetch_retry(&tx_stats->syncp, start));
495
496 st->tx_packets += packets;
497 st->tx_bytes += bytes;
498 }
499 }
500
mana_get_tx_queue(struct net_device * ndev,struct sk_buff * skb,int old_q)501 static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb,
502 int old_q)
503 {
504 struct mana_port_context *apc = netdev_priv(ndev);
505 u32 hash = skb_get_hash(skb);
506 struct sock *sk = skb->sk;
507 int txq;
508
509 txq = apc->indir_table[hash & (apc->indir_table_sz - 1)];
510
511 if (txq != old_q && sk && sk_fullsock(sk) &&
512 rcu_access_pointer(sk->sk_dst_cache))
513 sk_tx_queue_set(sk, txq);
514
515 return txq;
516 }
517
mana_select_queue(struct net_device * ndev,struct sk_buff * skb,struct net_device * sb_dev)518 static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb,
519 struct net_device *sb_dev)
520 {
521 int txq;
522
523 if (ndev->real_num_tx_queues == 1)
524 return 0;
525
526 txq = sk_tx_queue_get(skb->sk);
527
528 if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) {
529 if (skb_rx_queue_recorded(skb))
530 txq = skb_get_rx_queue(skb);
531 else
532 txq = mana_get_tx_queue(ndev, skb, txq);
533 }
534
535 return txq;
536 }
537
538 /* Release pre-allocated RX buffers */
mana_pre_dealloc_rxbufs(struct mana_port_context * mpc)539 void mana_pre_dealloc_rxbufs(struct mana_port_context *mpc)
540 {
541 struct device *dev;
542 int i;
543
544 dev = mpc->ac->gdma_dev->gdma_context->dev;
545
546 if (!mpc->rxbufs_pre)
547 goto out1;
548
549 if (!mpc->das_pre)
550 goto out2;
551
552 while (mpc->rxbpre_total) {
553 i = --mpc->rxbpre_total;
554 dma_unmap_single(dev, mpc->das_pre[i], mpc->rxbpre_datasize,
555 DMA_FROM_DEVICE);
556 put_page(virt_to_head_page(mpc->rxbufs_pre[i]));
557 }
558
559 kfree(mpc->das_pre);
560 mpc->das_pre = NULL;
561
562 out2:
563 kfree(mpc->rxbufs_pre);
564 mpc->rxbufs_pre = NULL;
565
566 out1:
567 mpc->rxbpre_datasize = 0;
568 mpc->rxbpre_alloc_size = 0;
569 mpc->rxbpre_headroom = 0;
570 }
571
572 /* Get a buffer from the pre-allocated RX buffers */
mana_get_rxbuf_pre(struct mana_rxq * rxq,dma_addr_t * da)573 static void *mana_get_rxbuf_pre(struct mana_rxq *rxq, dma_addr_t *da)
574 {
575 struct net_device *ndev = rxq->ndev;
576 struct mana_port_context *mpc;
577 void *va;
578
579 mpc = netdev_priv(ndev);
580
581 if (!mpc->rxbufs_pre || !mpc->das_pre || !mpc->rxbpre_total) {
582 netdev_err(ndev, "No RX pre-allocated bufs\n");
583 return NULL;
584 }
585
586 /* Check sizes to catch unexpected coding error */
587 if (mpc->rxbpre_datasize != rxq->datasize) {
588 netdev_err(ndev, "rxbpre_datasize mismatch: %u: %u\n",
589 mpc->rxbpre_datasize, rxq->datasize);
590 return NULL;
591 }
592
593 if (mpc->rxbpre_alloc_size != rxq->alloc_size) {
594 netdev_err(ndev, "rxbpre_alloc_size mismatch: %u: %u\n",
595 mpc->rxbpre_alloc_size, rxq->alloc_size);
596 return NULL;
597 }
598
599 if (mpc->rxbpre_headroom != rxq->headroom) {
600 netdev_err(ndev, "rxbpre_headroom mismatch: %u: %u\n",
601 mpc->rxbpre_headroom, rxq->headroom);
602 return NULL;
603 }
604
605 mpc->rxbpre_total--;
606
607 *da = mpc->das_pre[mpc->rxbpre_total];
608 va = mpc->rxbufs_pre[mpc->rxbpre_total];
609 mpc->rxbufs_pre[mpc->rxbpre_total] = NULL;
610
611 /* Deallocate the array after all buffers are gone */
612 if (!mpc->rxbpre_total)
613 mana_pre_dealloc_rxbufs(mpc);
614
615 return va;
616 }
617
618 /* 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)619 static void mana_get_rxbuf_cfg(int mtu, u32 *datasize, u32 *alloc_size,
620 u32 *headroom)
621 {
622 if (mtu > MANA_XDP_MTU_MAX)
623 *headroom = 0; /* no support for XDP */
624 else
625 *headroom = XDP_PACKET_HEADROOM;
626
627 *alloc_size = SKB_DATA_ALIGN(mtu + MANA_RXBUF_PAD + *headroom);
628
629 /* Using page pool in this case, so alloc_size is PAGE_SIZE */
630 if (*alloc_size < PAGE_SIZE)
631 *alloc_size = PAGE_SIZE;
632
633 *datasize = mtu + ETH_HLEN;
634 }
635
mana_pre_alloc_rxbufs(struct mana_port_context * mpc,int new_mtu,int num_queues)636 int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu, int num_queues)
637 {
638 struct device *dev;
639 struct page *page;
640 dma_addr_t da;
641 int num_rxb;
642 void *va;
643 int i;
644
645 mana_get_rxbuf_cfg(new_mtu, &mpc->rxbpre_datasize,
646 &mpc->rxbpre_alloc_size, &mpc->rxbpre_headroom);
647
648 dev = mpc->ac->gdma_dev->gdma_context->dev;
649
650 num_rxb = num_queues * mpc->rx_queue_size;
651
652 WARN(mpc->rxbufs_pre, "mana rxbufs_pre exists\n");
653 mpc->rxbufs_pre = kmalloc_array(num_rxb, sizeof(void *), GFP_KERNEL);
654 if (!mpc->rxbufs_pre)
655 goto error;
656
657 mpc->das_pre = kmalloc_array(num_rxb, sizeof(dma_addr_t), GFP_KERNEL);
658 if (!mpc->das_pre)
659 goto error;
660
661 mpc->rxbpre_total = 0;
662
663 for (i = 0; i < num_rxb; i++) {
664 page = dev_alloc_pages(get_order(mpc->rxbpre_alloc_size));
665 if (!page)
666 goto error;
667
668 va = page_to_virt(page);
669
670 da = dma_map_single(dev, va + mpc->rxbpre_headroom,
671 mpc->rxbpre_datasize, DMA_FROM_DEVICE);
672 if (dma_mapping_error(dev, da)) {
673 put_page(page);
674 goto error;
675 }
676
677 mpc->rxbufs_pre[i] = va;
678 mpc->das_pre[i] = da;
679 mpc->rxbpre_total = i + 1;
680 }
681
682 return 0;
683
684 error:
685 netdev_err(mpc->ndev, "Failed to pre-allocate RX buffers for %d queues\n", num_queues);
686 mana_pre_dealloc_rxbufs(mpc);
687 return -ENOMEM;
688 }
689
mana_change_mtu(struct net_device * ndev,int new_mtu)690 static int mana_change_mtu(struct net_device *ndev, int new_mtu)
691 {
692 struct mana_port_context *mpc = netdev_priv(ndev);
693 unsigned int old_mtu = ndev->mtu;
694 int err;
695
696 /* Pre-allocate buffers to prevent failure in mana_attach later */
697 err = mana_pre_alloc_rxbufs(mpc, new_mtu, mpc->num_queues);
698 if (err) {
699 netdev_err(ndev, "Insufficient memory for new MTU\n");
700 return err;
701 }
702
703 err = mana_detach(ndev, false);
704 if (err) {
705 netdev_err(ndev, "mana_detach failed: %d\n", err);
706 goto out;
707 }
708
709 WRITE_ONCE(ndev->mtu, new_mtu);
710
711 err = mana_attach(ndev);
712 if (err) {
713 netdev_err(ndev, "mana_attach failed: %d\n", err);
714 WRITE_ONCE(ndev->mtu, old_mtu);
715 }
716
717 out:
718 mana_pre_dealloc_rxbufs(mpc);
719 return err;
720 }
721
722 static const struct net_device_ops mana_devops = {
723 .ndo_open = mana_open,
724 .ndo_stop = mana_close,
725 .ndo_select_queue = mana_select_queue,
726 .ndo_start_xmit = mana_start_xmit,
727 .ndo_validate_addr = eth_validate_addr,
728 .ndo_get_stats64 = mana_get_stats64,
729 .ndo_bpf = mana_bpf,
730 .ndo_xdp_xmit = mana_xdp_xmit,
731 .ndo_change_mtu = mana_change_mtu,
732 };
733
mana_cleanup_port_context(struct mana_port_context * apc)734 static void mana_cleanup_port_context(struct mana_port_context *apc)
735 {
736 /*
737 * make sure subsequent cleanup attempts don't end up removing already
738 * cleaned dentry pointer
739 */
740 debugfs_remove(apc->mana_port_debugfs);
741 apc->mana_port_debugfs = NULL;
742 kfree(apc->rxqs);
743 apc->rxqs = NULL;
744 }
745
mana_cleanup_indir_table(struct mana_port_context * apc)746 static void mana_cleanup_indir_table(struct mana_port_context *apc)
747 {
748 apc->indir_table_sz = 0;
749 kfree(apc->indir_table);
750 kfree(apc->rxobj_table);
751 }
752
mana_init_port_context(struct mana_port_context * apc)753 static int mana_init_port_context(struct mana_port_context *apc)
754 {
755 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *),
756 GFP_KERNEL);
757
758 return !apc->rxqs ? -ENOMEM : 0;
759 }
760
mana_send_request(struct mana_context * ac,void * in_buf,u32 in_len,void * out_buf,u32 out_len)761 static int mana_send_request(struct mana_context *ac, void *in_buf,
762 u32 in_len, void *out_buf, u32 out_len)
763 {
764 struct gdma_context *gc = ac->gdma_dev->gdma_context;
765 struct gdma_resp_hdr *resp = out_buf;
766 struct gdma_req_hdr *req = in_buf;
767 struct device *dev = gc->dev;
768 static atomic_t activity_id;
769 int err;
770
771 req->dev_id = gc->mana.dev_id;
772 req->activity_id = atomic_inc_return(&activity_id);
773
774 err = mana_gd_send_request(gc, in_len, in_buf, out_len,
775 out_buf);
776 if (err || resp->status) {
777 dev_err(dev, "Failed to send mana message: %d, 0x%x\n",
778 err, resp->status);
779 return err ? err : -EPROTO;
780 }
781
782 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 ||
783 req->activity_id != resp->activity_id) {
784 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n",
785 req->dev_id.as_uint32, resp->dev_id.as_uint32,
786 req->activity_id, resp->activity_id);
787 return -EPROTO;
788 }
789
790 return 0;
791 }
792
mana_verify_resp_hdr(const struct gdma_resp_hdr * resp_hdr,const enum mana_command_code expected_code,const u32 min_size)793 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr,
794 const enum mana_command_code expected_code,
795 const u32 min_size)
796 {
797 if (resp_hdr->response.msg_type != expected_code)
798 return -EPROTO;
799
800 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1)
801 return -EPROTO;
802
803 if (resp_hdr->response.msg_size < min_size)
804 return -EPROTO;
805
806 return 0;
807 }
808
mana_pf_register_hw_vport(struct mana_port_context * apc)809 static int mana_pf_register_hw_vport(struct mana_port_context *apc)
810 {
811 struct mana_register_hw_vport_resp resp = {};
812 struct mana_register_hw_vport_req req = {};
813 int err;
814
815 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT,
816 sizeof(req), sizeof(resp));
817 req.attached_gfid = 1;
818 req.is_pf_default_vport = 1;
819 req.allow_all_ether_types = 1;
820
821 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
822 sizeof(resp));
823 if (err) {
824 netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err);
825 return err;
826 }
827
828 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT,
829 sizeof(resp));
830 if (err || resp.hdr.status) {
831 netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n",
832 err, resp.hdr.status);
833 return err ? err : -EPROTO;
834 }
835
836 apc->port_handle = resp.hw_vport_handle;
837 return 0;
838 }
839
mana_pf_deregister_hw_vport(struct mana_port_context * apc)840 static void mana_pf_deregister_hw_vport(struct mana_port_context *apc)
841 {
842 struct mana_deregister_hw_vport_resp resp = {};
843 struct mana_deregister_hw_vport_req req = {};
844 int err;
845
846 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT,
847 sizeof(req), sizeof(resp));
848 req.hw_vport_handle = apc->port_handle;
849
850 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
851 sizeof(resp));
852 if (err) {
853 netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n",
854 err);
855 return;
856 }
857
858 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT,
859 sizeof(resp));
860 if (err || resp.hdr.status)
861 netdev_err(apc->ndev,
862 "Failed to deregister hw vPort: %d, 0x%x\n",
863 err, resp.hdr.status);
864 }
865
mana_pf_register_filter(struct mana_port_context * apc)866 static int mana_pf_register_filter(struct mana_port_context *apc)
867 {
868 struct mana_register_filter_resp resp = {};
869 struct mana_register_filter_req req = {};
870 int err;
871
872 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER,
873 sizeof(req), sizeof(resp));
874 req.vport = apc->port_handle;
875 memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN);
876
877 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
878 sizeof(resp));
879 if (err) {
880 netdev_err(apc->ndev, "Failed to register filter: %d\n", err);
881 return err;
882 }
883
884 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER,
885 sizeof(resp));
886 if (err || resp.hdr.status) {
887 netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n",
888 err, resp.hdr.status);
889 return err ? err : -EPROTO;
890 }
891
892 apc->pf_filter_handle = resp.filter_handle;
893 return 0;
894 }
895
mana_pf_deregister_filter(struct mana_port_context * apc)896 static void mana_pf_deregister_filter(struct mana_port_context *apc)
897 {
898 struct mana_deregister_filter_resp resp = {};
899 struct mana_deregister_filter_req req = {};
900 int err;
901
902 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER,
903 sizeof(req), sizeof(resp));
904 req.filter_handle = apc->pf_filter_handle;
905
906 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
907 sizeof(resp));
908 if (err) {
909 netdev_err(apc->ndev, "Failed to unregister filter: %d\n",
910 err);
911 return;
912 }
913
914 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER,
915 sizeof(resp));
916 if (err || resp.hdr.status)
917 netdev_err(apc->ndev,
918 "Failed to deregister filter: %d, 0x%x\n",
919 err, resp.hdr.status);
920 }
921
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)922 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver,
923 u32 proto_minor_ver, u32 proto_micro_ver,
924 u16 *max_num_vports, u8 *bm_hostmode)
925 {
926 struct gdma_context *gc = ac->gdma_dev->gdma_context;
927 struct mana_query_device_cfg_resp resp = {};
928 struct mana_query_device_cfg_req req = {};
929 struct device *dev = gc->dev;
930 int err = 0;
931
932 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG,
933 sizeof(req), sizeof(resp));
934
935 req.hdr.resp.msg_version = GDMA_MESSAGE_V3;
936
937 req.proto_major_ver = proto_major_ver;
938 req.proto_minor_ver = proto_minor_ver;
939 req.proto_micro_ver = proto_micro_ver;
940
941 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp));
942 if (err) {
943 dev_err(dev, "Failed to query config: %d", err);
944 return err;
945 }
946
947 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG,
948 sizeof(resp));
949 if (err || resp.hdr.status) {
950 dev_err(dev, "Invalid query result: %d, 0x%x\n", err,
951 resp.hdr.status);
952 if (!err)
953 err = -EPROTO;
954 return err;
955 }
956
957 *max_num_vports = resp.max_num_vports;
958
959 if (resp.hdr.response.msg_version >= GDMA_MESSAGE_V2)
960 gc->adapter_mtu = resp.adapter_mtu;
961 else
962 gc->adapter_mtu = ETH_FRAME_LEN;
963
964 if (resp.hdr.response.msg_version >= GDMA_MESSAGE_V3)
965 *bm_hostmode = resp.bm_hostmode;
966 else
967 *bm_hostmode = 0;
968
969 debugfs_create_u16("adapter-MTU", 0400, gc->mana_pci_debugfs, &gc->adapter_mtu);
970
971 return 0;
972 }
973
mana_query_vport_cfg(struct mana_port_context * apc,u32 vport_index,u32 * max_sq,u32 * max_rq,u32 * num_indir_entry)974 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index,
975 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry)
976 {
977 struct mana_query_vport_cfg_resp resp = {};
978 struct mana_query_vport_cfg_req req = {};
979 int err;
980
981 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG,
982 sizeof(req), sizeof(resp));
983
984 req.vport_index = vport_index;
985
986 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
987 sizeof(resp));
988 if (err)
989 return err;
990
991 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG,
992 sizeof(resp));
993 if (err)
994 return err;
995
996 if (resp.hdr.status)
997 return -EPROTO;
998
999 *max_sq = resp.max_num_sq;
1000 *max_rq = resp.max_num_rq;
1001 if (resp.num_indirection_ent > 0 &&
1002 resp.num_indirection_ent <= MANA_INDIRECT_TABLE_MAX_SIZE &&
1003 is_power_of_2(resp.num_indirection_ent)) {
1004 *num_indir_entry = resp.num_indirection_ent;
1005 } else {
1006 netdev_warn(apc->ndev,
1007 "Setting indirection table size to default %d for vPort %d\n",
1008 MANA_INDIRECT_TABLE_DEF_SIZE, apc->port_idx);
1009 *num_indir_entry = MANA_INDIRECT_TABLE_DEF_SIZE;
1010 }
1011
1012 apc->port_handle = resp.vport;
1013 ether_addr_copy(apc->mac_addr, resp.mac_addr);
1014
1015 return 0;
1016 }
1017
mana_uncfg_vport(struct mana_port_context * apc)1018 void mana_uncfg_vport(struct mana_port_context *apc)
1019 {
1020 mutex_lock(&apc->vport_mutex);
1021 apc->vport_use_count--;
1022 WARN_ON(apc->vport_use_count < 0);
1023 mutex_unlock(&apc->vport_mutex);
1024 }
1025 EXPORT_SYMBOL_NS(mana_uncfg_vport, "NET_MANA");
1026
mana_cfg_vport(struct mana_port_context * apc,u32 protection_dom_id,u32 doorbell_pg_id)1027 int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id,
1028 u32 doorbell_pg_id)
1029 {
1030 struct mana_config_vport_resp resp = {};
1031 struct mana_config_vport_req req = {};
1032 int err;
1033
1034 /* This function is used to program the Ethernet port in the hardware
1035 * table. It can be called from the Ethernet driver or the RDMA driver.
1036 *
1037 * For Ethernet usage, the hardware supports only one active user on a
1038 * physical port. The driver checks on the port usage before programming
1039 * the hardware when creating the RAW QP (RDMA driver) or exposing the
1040 * device to kernel NET layer (Ethernet driver).
1041 *
1042 * Because the RDMA driver doesn't know in advance which QP type the
1043 * user will create, it exposes the device with all its ports. The user
1044 * may not be able to create RAW QP on a port if this port is already
1045 * in used by the Ethernet driver from the kernel.
1046 *
1047 * This physical port limitation only applies to the RAW QP. For RC QP,
1048 * the hardware doesn't have this limitation. The user can create RC
1049 * QPs on a physical port up to the hardware limits independent of the
1050 * Ethernet usage on the same port.
1051 */
1052 mutex_lock(&apc->vport_mutex);
1053 if (apc->vport_use_count > 0) {
1054 mutex_unlock(&apc->vport_mutex);
1055 return -EBUSY;
1056 }
1057 apc->vport_use_count++;
1058 mutex_unlock(&apc->vport_mutex);
1059
1060 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX,
1061 sizeof(req), sizeof(resp));
1062 req.vport = apc->port_handle;
1063 req.pdid = protection_dom_id;
1064 req.doorbell_pageid = doorbell_pg_id;
1065
1066 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1067 sizeof(resp));
1068 if (err) {
1069 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err);
1070 goto out;
1071 }
1072
1073 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX,
1074 sizeof(resp));
1075 if (err || resp.hdr.status) {
1076 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n",
1077 err, resp.hdr.status);
1078 if (!err)
1079 err = -EPROTO;
1080
1081 goto out;
1082 }
1083
1084 apc->tx_shortform_allowed = resp.short_form_allowed;
1085 apc->tx_vp_offset = resp.tx_vport_offset;
1086
1087 netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n",
1088 apc->port_handle, protection_dom_id, doorbell_pg_id);
1089 out:
1090 if (err)
1091 mana_uncfg_vport(apc);
1092
1093 return err;
1094 }
1095 EXPORT_SYMBOL_NS(mana_cfg_vport, "NET_MANA");
1096
mana_cfg_vport_steering(struct mana_port_context * apc,enum TRI_STATE rx,bool update_default_rxobj,bool update_key,bool update_tab)1097 static int mana_cfg_vport_steering(struct mana_port_context *apc,
1098 enum TRI_STATE rx,
1099 bool update_default_rxobj, bool update_key,
1100 bool update_tab)
1101 {
1102 struct mana_cfg_rx_steer_req_v2 *req;
1103 struct mana_cfg_rx_steer_resp resp = {};
1104 struct net_device *ndev = apc->ndev;
1105 u32 req_buf_size;
1106 int err;
1107
1108 req_buf_size = struct_size(req, indir_tab, apc->indir_table_sz);
1109 req = kzalloc(req_buf_size, GFP_KERNEL);
1110 if (!req)
1111 return -ENOMEM;
1112
1113 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
1114 sizeof(resp));
1115
1116 req->hdr.req.msg_version = GDMA_MESSAGE_V2;
1117
1118 req->vport = apc->port_handle;
1119 req->num_indir_entries = apc->indir_table_sz;
1120 req->indir_tab_offset = offsetof(struct mana_cfg_rx_steer_req_v2,
1121 indir_tab);
1122 req->rx_enable = rx;
1123 req->rss_enable = apc->rss_state;
1124 req->update_default_rxobj = update_default_rxobj;
1125 req->update_hashkey = update_key;
1126 req->update_indir_tab = update_tab;
1127 req->default_rxobj = apc->default_rxobj;
1128 req->cqe_coalescing_enable = 0;
1129
1130 if (update_key)
1131 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE);
1132
1133 if (update_tab)
1134 memcpy(req->indir_tab, apc->rxobj_table,
1135 flex_array_size(req, indir_tab, req->num_indir_entries));
1136
1137 err = mana_send_request(apc->ac, req, req_buf_size, &resp,
1138 sizeof(resp));
1139 if (err) {
1140 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err);
1141 goto out;
1142 }
1143
1144 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX,
1145 sizeof(resp));
1146 if (err) {
1147 netdev_err(ndev, "vPort RX configuration failed: %d\n", err);
1148 goto out;
1149 }
1150
1151 if (resp.hdr.status) {
1152 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n",
1153 resp.hdr.status);
1154 err = -EPROTO;
1155 }
1156
1157 netdev_info(ndev, "Configured steering vPort %llu entries %u\n",
1158 apc->port_handle, apc->indir_table_sz);
1159 out:
1160 kfree(req);
1161 return err;
1162 }
1163
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)1164 int mana_create_wq_obj(struct mana_port_context *apc,
1165 mana_handle_t vport,
1166 u32 wq_type, struct mana_obj_spec *wq_spec,
1167 struct mana_obj_spec *cq_spec,
1168 mana_handle_t *wq_obj)
1169 {
1170 struct mana_create_wqobj_resp resp = {};
1171 struct mana_create_wqobj_req req = {};
1172 struct net_device *ndev = apc->ndev;
1173 int err;
1174
1175 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ,
1176 sizeof(req), sizeof(resp));
1177 req.vport = vport;
1178 req.wq_type = wq_type;
1179 req.wq_gdma_region = wq_spec->gdma_region;
1180 req.cq_gdma_region = cq_spec->gdma_region;
1181 req.wq_size = wq_spec->queue_size;
1182 req.cq_size = cq_spec->queue_size;
1183 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id;
1184 req.cq_parent_qid = cq_spec->attached_eq;
1185
1186 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1187 sizeof(resp));
1188 if (err) {
1189 netdev_err(ndev, "Failed to create WQ object: %d\n", err);
1190 goto out;
1191 }
1192
1193 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ,
1194 sizeof(resp));
1195 if (err || resp.hdr.status) {
1196 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err,
1197 resp.hdr.status);
1198 if (!err)
1199 err = -EPROTO;
1200 goto out;
1201 }
1202
1203 if (resp.wq_obj == INVALID_MANA_HANDLE) {
1204 netdev_err(ndev, "Got an invalid WQ object handle\n");
1205 err = -EPROTO;
1206 goto out;
1207 }
1208
1209 *wq_obj = resp.wq_obj;
1210 wq_spec->queue_index = resp.wq_id;
1211 cq_spec->queue_index = resp.cq_id;
1212
1213 return 0;
1214 out:
1215 return err;
1216 }
1217 EXPORT_SYMBOL_NS(mana_create_wq_obj, "NET_MANA");
1218
mana_destroy_wq_obj(struct mana_port_context * apc,u32 wq_type,mana_handle_t wq_obj)1219 void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type,
1220 mana_handle_t wq_obj)
1221 {
1222 struct mana_destroy_wqobj_resp resp = {};
1223 struct mana_destroy_wqobj_req req = {};
1224 struct net_device *ndev = apc->ndev;
1225 int err;
1226
1227 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ,
1228 sizeof(req), sizeof(resp));
1229 req.wq_type = wq_type;
1230 req.wq_obj_handle = wq_obj;
1231
1232 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1233 sizeof(resp));
1234 if (err) {
1235 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err);
1236 return;
1237 }
1238
1239 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ,
1240 sizeof(resp));
1241 if (err || resp.hdr.status)
1242 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err,
1243 resp.hdr.status);
1244 }
1245 EXPORT_SYMBOL_NS(mana_destroy_wq_obj, "NET_MANA");
1246
mana_destroy_eq(struct mana_context * ac)1247 static void mana_destroy_eq(struct mana_context *ac)
1248 {
1249 struct gdma_context *gc = ac->gdma_dev->gdma_context;
1250 struct gdma_queue *eq;
1251 int i;
1252
1253 if (!ac->eqs)
1254 return;
1255
1256 debugfs_remove_recursive(ac->mana_eqs_debugfs);
1257 ac->mana_eqs_debugfs = NULL;
1258
1259 for (i = 0; i < gc->max_num_queues; i++) {
1260 eq = ac->eqs[i].eq;
1261 if (!eq)
1262 continue;
1263
1264 mana_gd_destroy_queue(gc, eq);
1265 }
1266
1267 kfree(ac->eqs);
1268 ac->eqs = NULL;
1269 }
1270
mana_create_eq_debugfs(struct mana_context * ac,int i)1271 static void mana_create_eq_debugfs(struct mana_context *ac, int i)
1272 {
1273 struct mana_eq eq = ac->eqs[i];
1274 char eqnum[32];
1275
1276 sprintf(eqnum, "eq%d", i);
1277 eq.mana_eq_debugfs = debugfs_create_dir(eqnum, ac->mana_eqs_debugfs);
1278 debugfs_create_u32("head", 0400, eq.mana_eq_debugfs, &eq.eq->head);
1279 debugfs_create_u32("tail", 0400, eq.mana_eq_debugfs, &eq.eq->tail);
1280 debugfs_create_file("eq_dump", 0400, eq.mana_eq_debugfs, eq.eq, &mana_dbg_q_fops);
1281 }
1282
mana_create_eq(struct mana_context * ac)1283 static int mana_create_eq(struct mana_context *ac)
1284 {
1285 struct gdma_dev *gd = ac->gdma_dev;
1286 struct gdma_context *gc = gd->gdma_context;
1287 struct gdma_queue_spec spec = {};
1288 int err;
1289 int i;
1290
1291 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq),
1292 GFP_KERNEL);
1293 if (!ac->eqs)
1294 return -ENOMEM;
1295
1296 spec.type = GDMA_EQ;
1297 spec.monitor_avl_buf = false;
1298 spec.queue_size = EQ_SIZE;
1299 spec.eq.callback = NULL;
1300 spec.eq.context = ac->eqs;
1301 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE;
1302
1303 ac->mana_eqs_debugfs = debugfs_create_dir("EQs", gc->mana_pci_debugfs);
1304
1305 for (i = 0; i < gc->max_num_queues; i++) {
1306 spec.eq.msix_index = (i + 1) % gc->num_msix_usable;
1307 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq);
1308 if (err) {
1309 dev_err(gc->dev, "Failed to create EQ %d : %d\n", i, err);
1310 goto out;
1311 }
1312 mana_create_eq_debugfs(ac, i);
1313 }
1314
1315 return 0;
1316 out:
1317 mana_destroy_eq(ac);
1318 return err;
1319 }
1320
mana_fence_rq(struct mana_port_context * apc,struct mana_rxq * rxq)1321 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq)
1322 {
1323 struct mana_fence_rq_resp resp = {};
1324 struct mana_fence_rq_req req = {};
1325 int err;
1326
1327 init_completion(&rxq->fence_event);
1328
1329 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ,
1330 sizeof(req), sizeof(resp));
1331 req.wq_obj_handle = rxq->rxobj;
1332
1333 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1334 sizeof(resp));
1335 if (err) {
1336 netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n",
1337 rxq->rxq_idx, err);
1338 return err;
1339 }
1340
1341 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp));
1342 if (err || resp.hdr.status) {
1343 netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n",
1344 rxq->rxq_idx, err, resp.hdr.status);
1345 if (!err)
1346 err = -EPROTO;
1347
1348 return err;
1349 }
1350
1351 if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) {
1352 netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n",
1353 rxq->rxq_idx);
1354 return -ETIMEDOUT;
1355 }
1356
1357 return 0;
1358 }
1359
mana_fence_rqs(struct mana_port_context * apc)1360 static void mana_fence_rqs(struct mana_port_context *apc)
1361 {
1362 unsigned int rxq_idx;
1363 struct mana_rxq *rxq;
1364 int err;
1365
1366 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
1367 rxq = apc->rxqs[rxq_idx];
1368 err = mana_fence_rq(apc, rxq);
1369
1370 /* In case of any error, use sleep instead. */
1371 if (err)
1372 msleep(100);
1373 }
1374 }
1375
mana_move_wq_tail(struct gdma_queue * wq,u32 num_units)1376 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units)
1377 {
1378 u32 used_space_old;
1379 u32 used_space_new;
1380
1381 used_space_old = wq->head - wq->tail;
1382 used_space_new = wq->head - (wq->tail + num_units);
1383
1384 if (WARN_ON_ONCE(used_space_new > used_space_old))
1385 return -ERANGE;
1386
1387 wq->tail += num_units;
1388 return 0;
1389 }
1390
mana_unmap_skb(struct sk_buff * skb,struct mana_port_context * apc)1391 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc)
1392 {
1393 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
1394 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1395 struct device *dev = gc->dev;
1396 int hsg, i;
1397
1398 /* Number of SGEs of linear part */
1399 hsg = (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0]) ? 2 : 1;
1400
1401 for (i = 0; i < hsg; i++)
1402 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i],
1403 DMA_TO_DEVICE);
1404
1405 for (i = hsg; i < skb_shinfo(skb)->nr_frags + hsg; i++)
1406 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
1407 DMA_TO_DEVICE);
1408 }
1409
mana_poll_tx_cq(struct mana_cq * cq)1410 static void mana_poll_tx_cq(struct mana_cq *cq)
1411 {
1412 struct gdma_comp *completions = cq->gdma_comp_buf;
1413 struct gdma_posted_wqe_info *wqe_info;
1414 unsigned int pkt_transmitted = 0;
1415 unsigned int wqe_unit_cnt = 0;
1416 struct mana_txq *txq = cq->txq;
1417 struct mana_port_context *apc;
1418 struct netdev_queue *net_txq;
1419 struct gdma_queue *gdma_wq;
1420 unsigned int avail_space;
1421 struct net_device *ndev;
1422 struct sk_buff *skb;
1423 bool txq_stopped;
1424 int comp_read;
1425 int i;
1426
1427 ndev = txq->ndev;
1428 apc = netdev_priv(ndev);
1429
1430 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions,
1431 CQE_POLLING_BUFFER);
1432
1433 if (comp_read < 1)
1434 return;
1435
1436 for (i = 0; i < comp_read; i++) {
1437 struct mana_tx_comp_oob *cqe_oob;
1438
1439 if (WARN_ON_ONCE(!completions[i].is_sq))
1440 return;
1441
1442 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data;
1443 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type !=
1444 MANA_CQE_COMPLETION))
1445 return;
1446
1447 switch (cqe_oob->cqe_hdr.cqe_type) {
1448 case CQE_TX_OKAY:
1449 break;
1450
1451 case CQE_TX_SA_DROP:
1452 case CQE_TX_MTU_DROP:
1453 case CQE_TX_INVALID_OOB:
1454 case CQE_TX_INVALID_ETH_TYPE:
1455 case CQE_TX_HDR_PROCESSING_ERROR:
1456 case CQE_TX_VF_DISABLED:
1457 case CQE_TX_VPORT_IDX_OUT_OF_RANGE:
1458 case CQE_TX_VPORT_DISABLED:
1459 case CQE_TX_VLAN_TAGGING_VIOLATION:
1460 if (net_ratelimit())
1461 netdev_err(ndev, "TX: CQE error %d\n",
1462 cqe_oob->cqe_hdr.cqe_type);
1463
1464 apc->eth_stats.tx_cqe_err++;
1465 break;
1466
1467 default:
1468 /* If the CQE type is unknown, log an error,
1469 * and still free the SKB, update tail, etc.
1470 */
1471 if (net_ratelimit())
1472 netdev_err(ndev, "TX: unknown CQE type %d\n",
1473 cqe_oob->cqe_hdr.cqe_type);
1474
1475 apc->eth_stats.tx_cqe_unknown_type++;
1476 break;
1477 }
1478
1479 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num))
1480 return;
1481
1482 skb = skb_dequeue(&txq->pending_skbs);
1483 if (WARN_ON_ONCE(!skb))
1484 return;
1485
1486 wqe_info = (struct gdma_posted_wqe_info *)skb->cb;
1487 wqe_unit_cnt += wqe_info->wqe_size_in_bu;
1488
1489 mana_unmap_skb(skb, apc);
1490
1491 napi_consume_skb(skb, cq->budget);
1492
1493 pkt_transmitted++;
1494 }
1495
1496 if (WARN_ON_ONCE(wqe_unit_cnt == 0))
1497 return;
1498
1499 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt);
1500
1501 gdma_wq = txq->gdma_sq;
1502 avail_space = mana_gd_wq_avail_space(gdma_wq);
1503
1504 /* Ensure tail updated before checking q stop */
1505 smp_mb();
1506
1507 net_txq = txq->net_txq;
1508 txq_stopped = netif_tx_queue_stopped(net_txq);
1509
1510 /* Ensure checking txq_stopped before apc->port_is_up. */
1511 smp_rmb();
1512
1513 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) {
1514 netif_tx_wake_queue(net_txq);
1515 apc->eth_stats.wake_queue++;
1516 }
1517
1518 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0)
1519 WARN_ON_ONCE(1);
1520
1521 cq->work_done = pkt_transmitted;
1522 }
1523
mana_post_pkt_rxq(struct mana_rxq * rxq)1524 static void mana_post_pkt_rxq(struct mana_rxq *rxq)
1525 {
1526 struct mana_recv_buf_oob *recv_buf_oob;
1527 u32 curr_index;
1528 int err;
1529
1530 curr_index = rxq->buf_index++;
1531 if (rxq->buf_index == rxq->num_rx_buf)
1532 rxq->buf_index = 0;
1533
1534 recv_buf_oob = &rxq->rx_oobs[curr_index];
1535
1536 err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req,
1537 &recv_buf_oob->wqe_inf);
1538 if (WARN_ON_ONCE(err))
1539 return;
1540
1541 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1);
1542 }
1543
mana_build_skb(struct mana_rxq * rxq,void * buf_va,uint pkt_len,struct xdp_buff * xdp)1544 static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va,
1545 uint pkt_len, struct xdp_buff *xdp)
1546 {
1547 struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size);
1548
1549 if (!skb)
1550 return NULL;
1551
1552 if (xdp->data_hard_start) {
1553 u32 metasize = xdp->data - xdp->data_meta;
1554
1555 skb_reserve(skb, xdp->data - xdp->data_hard_start);
1556 skb_put(skb, xdp->data_end - xdp->data);
1557 if (metasize)
1558 skb_metadata_set(skb, metasize);
1559 return skb;
1560 }
1561
1562 skb_reserve(skb, rxq->headroom);
1563 skb_put(skb, pkt_len);
1564
1565 return skb;
1566 }
1567
mana_rx_skb(void * buf_va,bool from_pool,struct mana_rxcomp_oob * cqe,struct mana_rxq * rxq)1568 static void mana_rx_skb(void *buf_va, bool from_pool,
1569 struct mana_rxcomp_oob *cqe, struct mana_rxq *rxq)
1570 {
1571 struct mana_stats_rx *rx_stats = &rxq->stats;
1572 struct net_device *ndev = rxq->ndev;
1573 uint pkt_len = cqe->ppi[0].pkt_len;
1574 u16 rxq_idx = rxq->rxq_idx;
1575 struct napi_struct *napi;
1576 struct xdp_buff xdp = {};
1577 struct sk_buff *skb;
1578 u32 hash_value;
1579 u32 act;
1580
1581 rxq->rx_cq.work_done++;
1582 napi = &rxq->rx_cq.napi;
1583
1584 if (!buf_va) {
1585 ++ndev->stats.rx_dropped;
1586 return;
1587 }
1588
1589 act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len);
1590
1591 if (act == XDP_REDIRECT && !rxq->xdp_rc)
1592 return;
1593
1594 if (act != XDP_PASS && act != XDP_TX)
1595 goto drop_xdp;
1596
1597 skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp);
1598
1599 if (!skb)
1600 goto drop;
1601
1602 if (from_pool)
1603 skb_mark_for_recycle(skb);
1604
1605 skb->dev = napi->dev;
1606
1607 skb->protocol = eth_type_trans(skb, ndev);
1608 skb_checksum_none_assert(skb);
1609 skb_record_rx_queue(skb, rxq_idx);
1610
1611 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) {
1612 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed)
1613 skb->ip_summed = CHECKSUM_UNNECESSARY;
1614 }
1615
1616 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) {
1617 hash_value = cqe->ppi[0].pkt_hash;
1618
1619 if (cqe->rx_hashtype & MANA_HASH_L4)
1620 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4);
1621 else
1622 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3);
1623 }
1624
1625 if (cqe->rx_vlantag_present) {
1626 u16 vlan_tci = cqe->rx_vlan_id;
1627
1628 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci);
1629 }
1630
1631 u64_stats_update_begin(&rx_stats->syncp);
1632 rx_stats->packets++;
1633 rx_stats->bytes += pkt_len;
1634
1635 if (act == XDP_TX)
1636 rx_stats->xdp_tx++;
1637 u64_stats_update_end(&rx_stats->syncp);
1638
1639 if (act == XDP_TX) {
1640 skb_set_queue_mapping(skb, rxq_idx);
1641 mana_xdp_tx(skb, ndev);
1642 return;
1643 }
1644
1645 napi_gro_receive(napi, skb);
1646
1647 return;
1648
1649 drop_xdp:
1650 u64_stats_update_begin(&rx_stats->syncp);
1651 rx_stats->xdp_drop++;
1652 u64_stats_update_end(&rx_stats->syncp);
1653
1654 drop:
1655 if (from_pool) {
1656 page_pool_recycle_direct(rxq->page_pool,
1657 virt_to_head_page(buf_va));
1658 } else {
1659 WARN_ON_ONCE(rxq->xdp_save_va);
1660 /* Save for reuse */
1661 rxq->xdp_save_va = buf_va;
1662 }
1663
1664 ++ndev->stats.rx_dropped;
1665
1666 return;
1667 }
1668
mana_get_rxfrag(struct mana_rxq * rxq,struct device * dev,dma_addr_t * da,bool * from_pool)1669 static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,
1670 dma_addr_t *da, bool *from_pool)
1671 {
1672 struct page *page;
1673 void *va;
1674
1675 *from_pool = false;
1676
1677 /* Reuse XDP dropped page if available */
1678 if (rxq->xdp_save_va) {
1679 va = rxq->xdp_save_va;
1680 rxq->xdp_save_va = NULL;
1681 } else {
1682 page = page_pool_dev_alloc_pages(rxq->page_pool);
1683 if (!page)
1684 return NULL;
1685
1686 *from_pool = true;
1687 va = page_to_virt(page);
1688 }
1689
1690 *da = dma_map_single(dev, va + rxq->headroom, rxq->datasize,
1691 DMA_FROM_DEVICE);
1692 if (dma_mapping_error(dev, *da)) {
1693 if (*from_pool)
1694 page_pool_put_full_page(rxq->page_pool, page, false);
1695 else
1696 put_page(virt_to_head_page(va));
1697
1698 return NULL;
1699 }
1700
1701 return va;
1702 }
1703
1704 /* 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)1705 static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq,
1706 struct mana_recv_buf_oob *rxoob, void **old_buf,
1707 bool *old_fp)
1708 {
1709 bool from_pool;
1710 dma_addr_t da;
1711 void *va;
1712
1713 va = mana_get_rxfrag(rxq, dev, &da, &from_pool);
1714 if (!va)
1715 return;
1716
1717 dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize,
1718 DMA_FROM_DEVICE);
1719 *old_buf = rxoob->buf_va;
1720 *old_fp = rxoob->from_pool;
1721
1722 rxoob->buf_va = va;
1723 rxoob->sgl[0].address = da;
1724 rxoob->from_pool = from_pool;
1725 }
1726
mana_process_rx_cqe(struct mana_rxq * rxq,struct mana_cq * cq,struct gdma_comp * cqe)1727 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,
1728 struct gdma_comp *cqe)
1729 {
1730 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data;
1731 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1732 struct net_device *ndev = rxq->ndev;
1733 struct mana_recv_buf_oob *rxbuf_oob;
1734 struct mana_port_context *apc;
1735 struct device *dev = gc->dev;
1736 void *old_buf = NULL;
1737 u32 curr, pktlen;
1738 bool old_fp;
1739
1740 apc = netdev_priv(ndev);
1741
1742 switch (oob->cqe_hdr.cqe_type) {
1743 case CQE_RX_OKAY:
1744 break;
1745
1746 case CQE_RX_TRUNCATED:
1747 ++ndev->stats.rx_dropped;
1748 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index];
1749 netdev_warn_once(ndev, "Dropped a truncated packet\n");
1750 goto drop;
1751
1752 case CQE_RX_COALESCED_4:
1753 netdev_err(ndev, "RX coalescing is unsupported\n");
1754 apc->eth_stats.rx_coalesced_err++;
1755 return;
1756
1757 case CQE_RX_OBJECT_FENCE:
1758 complete(&rxq->fence_event);
1759 return;
1760
1761 default:
1762 netdev_err(ndev, "Unknown RX CQE type = %d\n",
1763 oob->cqe_hdr.cqe_type);
1764 apc->eth_stats.rx_cqe_unknown_type++;
1765 return;
1766 }
1767
1768 pktlen = oob->ppi[0].pkt_len;
1769
1770 if (pktlen == 0) {
1771 /* data packets should never have packetlength of zero */
1772 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n",
1773 rxq->gdma_id, cq->gdma_id, rxq->rxobj);
1774 return;
1775 }
1776
1777 curr = rxq->buf_index;
1778 rxbuf_oob = &rxq->rx_oobs[curr];
1779 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1);
1780
1781 mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf, &old_fp);
1782
1783 /* Unsuccessful refill will have old_buf == NULL.
1784 * In this case, mana_rx_skb() will drop the packet.
1785 */
1786 mana_rx_skb(old_buf, old_fp, oob, rxq);
1787
1788 drop:
1789 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu);
1790
1791 mana_post_pkt_rxq(rxq);
1792 }
1793
mana_poll_rx_cq(struct mana_cq * cq)1794 static void mana_poll_rx_cq(struct mana_cq *cq)
1795 {
1796 struct gdma_comp *comp = cq->gdma_comp_buf;
1797 struct mana_rxq *rxq = cq->rxq;
1798 int comp_read, i;
1799
1800 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER);
1801 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER);
1802
1803 rxq->xdp_flush = false;
1804
1805 for (i = 0; i < comp_read; i++) {
1806 if (WARN_ON_ONCE(comp[i].is_sq))
1807 return;
1808
1809 /* verify recv cqe references the right rxq */
1810 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id))
1811 return;
1812
1813 mana_process_rx_cqe(rxq, cq, &comp[i]);
1814 }
1815
1816 if (comp_read > 0) {
1817 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1818
1819 mana_gd_wq_ring_doorbell(gc, rxq->gdma_rq);
1820 }
1821
1822 if (rxq->xdp_flush)
1823 xdp_do_flush();
1824 }
1825
mana_cq_handler(void * context,struct gdma_queue * gdma_queue)1826 static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue)
1827 {
1828 struct mana_cq *cq = context;
1829 int w;
1830
1831 WARN_ON_ONCE(cq->gdma_cq != gdma_queue);
1832
1833 if (cq->type == MANA_CQ_TYPE_RX)
1834 mana_poll_rx_cq(cq);
1835 else
1836 mana_poll_tx_cq(cq);
1837
1838 w = cq->work_done;
1839 cq->work_done_since_doorbell += w;
1840
1841 if (w < cq->budget) {
1842 mana_gd_ring_cq(gdma_queue, SET_ARM_BIT);
1843 cq->work_done_since_doorbell = 0;
1844 napi_complete_done(&cq->napi, w);
1845 } else if (cq->work_done_since_doorbell >
1846 cq->gdma_cq->queue_size / COMP_ENTRY_SIZE * 4) {
1847 /* MANA hardware requires at least one doorbell ring every 8
1848 * wraparounds of CQ even if there is no need to arm the CQ.
1849 * This driver rings the doorbell as soon as we have exceeded
1850 * 4 wraparounds.
1851 */
1852 mana_gd_ring_cq(gdma_queue, 0);
1853 cq->work_done_since_doorbell = 0;
1854 }
1855
1856 return w;
1857 }
1858
mana_poll(struct napi_struct * napi,int budget)1859 static int mana_poll(struct napi_struct *napi, int budget)
1860 {
1861 struct mana_cq *cq = container_of(napi, struct mana_cq, napi);
1862 int w;
1863
1864 cq->work_done = 0;
1865 cq->budget = budget;
1866
1867 w = mana_cq_handler(cq, cq->gdma_cq);
1868
1869 return min(w, budget);
1870 }
1871
mana_schedule_napi(void * context,struct gdma_queue * gdma_queue)1872 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue)
1873 {
1874 struct mana_cq *cq = context;
1875
1876 napi_schedule_irqoff(&cq->napi);
1877 }
1878
mana_deinit_cq(struct mana_port_context * apc,struct mana_cq * cq)1879 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq)
1880 {
1881 struct gdma_dev *gd = apc->ac->gdma_dev;
1882
1883 if (!cq->gdma_cq)
1884 return;
1885
1886 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq);
1887 }
1888
mana_deinit_txq(struct mana_port_context * apc,struct mana_txq * txq)1889 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq)
1890 {
1891 struct gdma_dev *gd = apc->ac->gdma_dev;
1892
1893 if (!txq->gdma_sq)
1894 return;
1895
1896 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq);
1897 }
1898
mana_destroy_txq(struct mana_port_context * apc)1899 static void mana_destroy_txq(struct mana_port_context *apc)
1900 {
1901 struct napi_struct *napi;
1902 int i;
1903
1904 if (!apc->tx_qp)
1905 return;
1906
1907 for (i = 0; i < apc->num_queues; i++) {
1908 debugfs_remove_recursive(apc->tx_qp[i].mana_tx_debugfs);
1909 apc->tx_qp[i].mana_tx_debugfs = NULL;
1910
1911 napi = &apc->tx_qp[i].tx_cq.napi;
1912 if (apc->tx_qp[i].txq.napi_initialized) {
1913 napi_synchronize(napi);
1914 napi_disable(napi);
1915 netif_napi_del(napi);
1916 apc->tx_qp[i].txq.napi_initialized = false;
1917 }
1918 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object);
1919
1920 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq);
1921
1922 mana_deinit_txq(apc, &apc->tx_qp[i].txq);
1923 }
1924
1925 kfree(apc->tx_qp);
1926 apc->tx_qp = NULL;
1927 }
1928
mana_create_txq_debugfs(struct mana_port_context * apc,int idx)1929 static void mana_create_txq_debugfs(struct mana_port_context *apc, int idx)
1930 {
1931 struct mana_tx_qp *tx_qp = &apc->tx_qp[idx];
1932 char qnum[32];
1933
1934 sprintf(qnum, "TX-%d", idx);
1935 tx_qp->mana_tx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs);
1936 debugfs_create_u32("sq_head", 0400, tx_qp->mana_tx_debugfs,
1937 &tx_qp->txq.gdma_sq->head);
1938 debugfs_create_u32("sq_tail", 0400, tx_qp->mana_tx_debugfs,
1939 &tx_qp->txq.gdma_sq->tail);
1940 debugfs_create_u32("sq_pend_skb_qlen", 0400, tx_qp->mana_tx_debugfs,
1941 &tx_qp->txq.pending_skbs.qlen);
1942 debugfs_create_u32("cq_head", 0400, tx_qp->mana_tx_debugfs,
1943 &tx_qp->tx_cq.gdma_cq->head);
1944 debugfs_create_u32("cq_tail", 0400, tx_qp->mana_tx_debugfs,
1945 &tx_qp->tx_cq.gdma_cq->tail);
1946 debugfs_create_u32("cq_budget", 0400, tx_qp->mana_tx_debugfs,
1947 &tx_qp->tx_cq.budget);
1948 debugfs_create_file("txq_dump", 0400, tx_qp->mana_tx_debugfs,
1949 tx_qp->txq.gdma_sq, &mana_dbg_q_fops);
1950 debugfs_create_file("cq_dump", 0400, tx_qp->mana_tx_debugfs,
1951 tx_qp->tx_cq.gdma_cq, &mana_dbg_q_fops);
1952 }
1953
mana_create_txq(struct mana_port_context * apc,struct net_device * net)1954 static int mana_create_txq(struct mana_port_context *apc,
1955 struct net_device *net)
1956 {
1957 struct mana_context *ac = apc->ac;
1958 struct gdma_dev *gd = ac->gdma_dev;
1959 struct mana_obj_spec wq_spec;
1960 struct mana_obj_spec cq_spec;
1961 struct gdma_queue_spec spec;
1962 struct gdma_context *gc;
1963 struct mana_txq *txq;
1964 struct mana_cq *cq;
1965 u32 txq_size;
1966 u32 cq_size;
1967 int err;
1968 int i;
1969
1970 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp),
1971 GFP_KERNEL);
1972 if (!apc->tx_qp)
1973 return -ENOMEM;
1974
1975 /* The minimum size of the WQE is 32 bytes, hence
1976 * apc->tx_queue_size represents the maximum number of WQEs
1977 * the SQ can store. This value is then used to size other queues
1978 * to prevent overflow.
1979 * Also note that the txq_size is always going to be MANA_PAGE_ALIGNED,
1980 * as min val of apc->tx_queue_size is 128 and that would make
1981 * txq_size 128*32 = 4096 and the other higher values of apc->tx_queue_size
1982 * are always power of two
1983 */
1984 txq_size = apc->tx_queue_size * 32;
1985
1986 cq_size = apc->tx_queue_size * COMP_ENTRY_SIZE;
1987
1988 gc = gd->gdma_context;
1989
1990 for (i = 0; i < apc->num_queues; i++) {
1991 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE;
1992
1993 /* Create SQ */
1994 txq = &apc->tx_qp[i].txq;
1995
1996 u64_stats_init(&txq->stats.syncp);
1997 txq->ndev = net;
1998 txq->net_txq = netdev_get_tx_queue(net, i);
1999 txq->vp_offset = apc->tx_vp_offset;
2000 txq->napi_initialized = false;
2001 skb_queue_head_init(&txq->pending_skbs);
2002
2003 memset(&spec, 0, sizeof(spec));
2004 spec.type = GDMA_SQ;
2005 spec.monitor_avl_buf = true;
2006 spec.queue_size = txq_size;
2007 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq);
2008 if (err)
2009 goto out;
2010
2011 /* Create SQ's CQ */
2012 cq = &apc->tx_qp[i].tx_cq;
2013 cq->type = MANA_CQ_TYPE_TX;
2014
2015 cq->txq = txq;
2016
2017 memset(&spec, 0, sizeof(spec));
2018 spec.type = GDMA_CQ;
2019 spec.monitor_avl_buf = false;
2020 spec.queue_size = cq_size;
2021 spec.cq.callback = mana_schedule_napi;
2022 spec.cq.parent_eq = ac->eqs[i].eq;
2023 spec.cq.context = cq;
2024 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2025 if (err)
2026 goto out;
2027
2028 memset(&wq_spec, 0, sizeof(wq_spec));
2029 memset(&cq_spec, 0, sizeof(cq_spec));
2030
2031 wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle;
2032 wq_spec.queue_size = txq->gdma_sq->queue_size;
2033
2034 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2035 cq_spec.queue_size = cq->gdma_cq->queue_size;
2036 cq_spec.modr_ctx_id = 0;
2037 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2038
2039 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ,
2040 &wq_spec, &cq_spec,
2041 &apc->tx_qp[i].tx_object);
2042
2043 if (err)
2044 goto out;
2045
2046 txq->gdma_sq->id = wq_spec.queue_index;
2047 cq->gdma_cq->id = cq_spec.queue_index;
2048
2049 txq->gdma_sq->mem_info.dma_region_handle =
2050 GDMA_INVALID_DMA_REGION;
2051 cq->gdma_cq->mem_info.dma_region_handle =
2052 GDMA_INVALID_DMA_REGION;
2053
2054 txq->gdma_txq_id = txq->gdma_sq->id;
2055
2056 cq->gdma_id = cq->gdma_cq->id;
2057
2058 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
2059 err = -EINVAL;
2060 goto out;
2061 }
2062
2063 gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2064
2065 mana_create_txq_debugfs(apc, i);
2066
2067 netif_napi_add_tx(net, &cq->napi, mana_poll);
2068 napi_enable(&cq->napi);
2069 txq->napi_initialized = true;
2070
2071 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2072 }
2073
2074 return 0;
2075 out:
2076 netdev_err(net, "Failed to create %d TX queues, %d\n",
2077 apc->num_queues, err);
2078 mana_destroy_txq(apc);
2079 return err;
2080 }
2081
mana_destroy_rxq(struct mana_port_context * apc,struct mana_rxq * rxq,bool napi_initialized)2082 static void mana_destroy_rxq(struct mana_port_context *apc,
2083 struct mana_rxq *rxq, bool napi_initialized)
2084
2085 {
2086 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
2087 struct mana_recv_buf_oob *rx_oob;
2088 struct device *dev = gc->dev;
2089 struct napi_struct *napi;
2090 struct page *page;
2091 int i;
2092
2093 if (!rxq)
2094 return;
2095
2096 debugfs_remove_recursive(rxq->mana_rx_debugfs);
2097 rxq->mana_rx_debugfs = NULL;
2098
2099 napi = &rxq->rx_cq.napi;
2100
2101 if (napi_initialized) {
2102 napi_synchronize(napi);
2103
2104 napi_disable(napi);
2105
2106 netif_napi_del(napi);
2107 }
2108 xdp_rxq_info_unreg(&rxq->xdp_rxq);
2109
2110 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj);
2111
2112 mana_deinit_cq(apc, &rxq->rx_cq);
2113
2114 if (rxq->xdp_save_va)
2115 put_page(virt_to_head_page(rxq->xdp_save_va));
2116
2117 for (i = 0; i < rxq->num_rx_buf; i++) {
2118 rx_oob = &rxq->rx_oobs[i];
2119
2120 if (!rx_oob->buf_va)
2121 continue;
2122
2123 dma_unmap_single(dev, rx_oob->sgl[0].address,
2124 rx_oob->sgl[0].size, DMA_FROM_DEVICE);
2125
2126 page = virt_to_head_page(rx_oob->buf_va);
2127
2128 if (rx_oob->from_pool)
2129 page_pool_put_full_page(rxq->page_pool, page, false);
2130 else
2131 put_page(page);
2132
2133 rx_oob->buf_va = NULL;
2134 }
2135
2136 page_pool_destroy(rxq->page_pool);
2137
2138 if (rxq->gdma_rq)
2139 mana_gd_destroy_queue(gc, rxq->gdma_rq);
2140
2141 kfree(rxq);
2142 }
2143
mana_fill_rx_oob(struct mana_recv_buf_oob * rx_oob,u32 mem_key,struct mana_rxq * rxq,struct device * dev)2144 static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key,
2145 struct mana_rxq *rxq, struct device *dev)
2146 {
2147 struct mana_port_context *mpc = netdev_priv(rxq->ndev);
2148 bool from_pool = false;
2149 dma_addr_t da;
2150 void *va;
2151
2152 if (mpc->rxbufs_pre)
2153 va = mana_get_rxbuf_pre(rxq, &da);
2154 else
2155 va = mana_get_rxfrag(rxq, dev, &da, &from_pool);
2156
2157 if (!va)
2158 return -ENOMEM;
2159
2160 rx_oob->buf_va = va;
2161 rx_oob->from_pool = from_pool;
2162
2163 rx_oob->sgl[0].address = da;
2164 rx_oob->sgl[0].size = rxq->datasize;
2165 rx_oob->sgl[0].mem_key = mem_key;
2166
2167 return 0;
2168 }
2169
2170 #define MANA_WQE_HEADER_SIZE 16
2171 #define MANA_WQE_SGE_SIZE 16
2172
mana_alloc_rx_wqe(struct mana_port_context * apc,struct mana_rxq * rxq,u32 * rxq_size,u32 * cq_size)2173 static int mana_alloc_rx_wqe(struct mana_port_context *apc,
2174 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size)
2175 {
2176 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
2177 struct mana_recv_buf_oob *rx_oob;
2178 struct device *dev = gc->dev;
2179 u32 buf_idx;
2180 int ret;
2181
2182 WARN_ON(rxq->datasize == 0);
2183
2184 *rxq_size = 0;
2185 *cq_size = 0;
2186
2187 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2188 rx_oob = &rxq->rx_oobs[buf_idx];
2189 memset(rx_oob, 0, sizeof(*rx_oob));
2190
2191 rx_oob->num_sge = 1;
2192
2193 ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq,
2194 dev);
2195 if (ret)
2196 return ret;
2197
2198 rx_oob->wqe_req.sgl = rx_oob->sgl;
2199 rx_oob->wqe_req.num_sge = rx_oob->num_sge;
2200 rx_oob->wqe_req.inline_oob_size = 0;
2201 rx_oob->wqe_req.inline_oob_data = NULL;
2202 rx_oob->wqe_req.flags = 0;
2203 rx_oob->wqe_req.client_data_unit = 0;
2204
2205 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE +
2206 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32);
2207 *cq_size += COMP_ENTRY_SIZE;
2208 }
2209
2210 return 0;
2211 }
2212
mana_push_wqe(struct mana_rxq * rxq)2213 static int mana_push_wqe(struct mana_rxq *rxq)
2214 {
2215 struct mana_recv_buf_oob *rx_oob;
2216 u32 buf_idx;
2217 int err;
2218
2219 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
2220 rx_oob = &rxq->rx_oobs[buf_idx];
2221
2222 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req,
2223 &rx_oob->wqe_inf);
2224 if (err)
2225 return -ENOSPC;
2226 }
2227
2228 return 0;
2229 }
2230
mana_create_page_pool(struct mana_rxq * rxq,struct gdma_context * gc)2231 static int mana_create_page_pool(struct mana_rxq *rxq, struct gdma_context *gc)
2232 {
2233 struct mana_port_context *mpc = netdev_priv(rxq->ndev);
2234 struct page_pool_params pprm = {};
2235 int ret;
2236
2237 pprm.pool_size = mpc->rx_queue_size;
2238 pprm.nid = gc->numa_node;
2239 pprm.napi = &rxq->rx_cq.napi;
2240 pprm.netdev = rxq->ndev;
2241 pprm.order = get_order(rxq->alloc_size);
2242
2243 rxq->page_pool = page_pool_create(&pprm);
2244
2245 if (IS_ERR(rxq->page_pool)) {
2246 ret = PTR_ERR(rxq->page_pool);
2247 rxq->page_pool = NULL;
2248 return ret;
2249 }
2250
2251 return 0;
2252 }
2253
mana_create_rxq(struct mana_port_context * apc,u32 rxq_idx,struct mana_eq * eq,struct net_device * ndev)2254 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,
2255 u32 rxq_idx, struct mana_eq *eq,
2256 struct net_device *ndev)
2257 {
2258 struct gdma_dev *gd = apc->ac->gdma_dev;
2259 struct mana_obj_spec wq_spec;
2260 struct mana_obj_spec cq_spec;
2261 struct gdma_queue_spec spec;
2262 struct mana_cq *cq = NULL;
2263 struct gdma_context *gc;
2264 u32 cq_size, rq_size;
2265 struct mana_rxq *rxq;
2266 int err;
2267
2268 gc = gd->gdma_context;
2269
2270 rxq = kzalloc(struct_size(rxq, rx_oobs, apc->rx_queue_size),
2271 GFP_KERNEL);
2272 if (!rxq)
2273 return NULL;
2274
2275 rxq->ndev = ndev;
2276 rxq->num_rx_buf = apc->rx_queue_size;
2277 rxq->rxq_idx = rxq_idx;
2278 rxq->rxobj = INVALID_MANA_HANDLE;
2279
2280 mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size,
2281 &rxq->headroom);
2282
2283 /* Create page pool for RX queue */
2284 err = mana_create_page_pool(rxq, gc);
2285 if (err) {
2286 netdev_err(ndev, "Create page pool err:%d\n", err);
2287 goto out;
2288 }
2289
2290 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size);
2291 if (err)
2292 goto out;
2293
2294 rq_size = MANA_PAGE_ALIGN(rq_size);
2295 cq_size = MANA_PAGE_ALIGN(cq_size);
2296
2297 /* Create RQ */
2298 memset(&spec, 0, sizeof(spec));
2299 spec.type = GDMA_RQ;
2300 spec.monitor_avl_buf = true;
2301 spec.queue_size = rq_size;
2302 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq);
2303 if (err)
2304 goto out;
2305
2306 /* Create RQ's CQ */
2307 cq = &rxq->rx_cq;
2308 cq->type = MANA_CQ_TYPE_RX;
2309 cq->rxq = rxq;
2310
2311 memset(&spec, 0, sizeof(spec));
2312 spec.type = GDMA_CQ;
2313 spec.monitor_avl_buf = false;
2314 spec.queue_size = cq_size;
2315 spec.cq.callback = mana_schedule_napi;
2316 spec.cq.parent_eq = eq->eq;
2317 spec.cq.context = cq;
2318 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2319 if (err)
2320 goto out;
2321
2322 memset(&wq_spec, 0, sizeof(wq_spec));
2323 memset(&cq_spec, 0, sizeof(cq_spec));
2324 wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle;
2325 wq_spec.queue_size = rxq->gdma_rq->queue_size;
2326
2327 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2328 cq_spec.queue_size = cq->gdma_cq->queue_size;
2329 cq_spec.modr_ctx_id = 0;
2330 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2331
2332 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ,
2333 &wq_spec, &cq_spec, &rxq->rxobj);
2334 if (err)
2335 goto out;
2336
2337 rxq->gdma_rq->id = wq_spec.queue_index;
2338 cq->gdma_cq->id = cq_spec.queue_index;
2339
2340 rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2341 cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2342
2343 rxq->gdma_id = rxq->gdma_rq->id;
2344 cq->gdma_id = cq->gdma_cq->id;
2345
2346 err = mana_push_wqe(rxq);
2347 if (err)
2348 goto out;
2349
2350 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
2351 err = -EINVAL;
2352 goto out;
2353 }
2354
2355 gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2356
2357 netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1);
2358
2359 WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx,
2360 cq->napi.napi_id));
2361 WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL,
2362 rxq->page_pool));
2363
2364 napi_enable(&cq->napi);
2365
2366 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2367 out:
2368 if (!err)
2369 return rxq;
2370
2371 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err);
2372
2373 mana_destroy_rxq(apc, rxq, false);
2374
2375 if (cq)
2376 mana_deinit_cq(apc, cq);
2377
2378 return NULL;
2379 }
2380
mana_create_rxq_debugfs(struct mana_port_context * apc,int idx)2381 static void mana_create_rxq_debugfs(struct mana_port_context *apc, int idx)
2382 {
2383 struct mana_rxq *rxq;
2384 char qnum[32];
2385
2386 rxq = apc->rxqs[idx];
2387
2388 sprintf(qnum, "RX-%d", idx);
2389 rxq->mana_rx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs);
2390 debugfs_create_u32("rq_head", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->head);
2391 debugfs_create_u32("rq_tail", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->tail);
2392 debugfs_create_u32("rq_nbuf", 0400, rxq->mana_rx_debugfs, &rxq->num_rx_buf);
2393 debugfs_create_u32("cq_head", 0400, rxq->mana_rx_debugfs,
2394 &rxq->rx_cq.gdma_cq->head);
2395 debugfs_create_u32("cq_tail", 0400, rxq->mana_rx_debugfs,
2396 &rxq->rx_cq.gdma_cq->tail);
2397 debugfs_create_u32("cq_budget", 0400, rxq->mana_rx_debugfs, &rxq->rx_cq.budget);
2398 debugfs_create_file("rxq_dump", 0400, rxq->mana_rx_debugfs, rxq->gdma_rq, &mana_dbg_q_fops);
2399 debugfs_create_file("cq_dump", 0400, rxq->mana_rx_debugfs, rxq->rx_cq.gdma_cq,
2400 &mana_dbg_q_fops);
2401 }
2402
mana_add_rx_queues(struct mana_port_context * apc,struct net_device * ndev)2403 static int mana_add_rx_queues(struct mana_port_context *apc,
2404 struct net_device *ndev)
2405 {
2406 struct mana_context *ac = apc->ac;
2407 struct mana_rxq *rxq;
2408 int err = 0;
2409 int i;
2410
2411 for (i = 0; i < apc->num_queues; i++) {
2412 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev);
2413 if (!rxq) {
2414 err = -ENOMEM;
2415 netdev_err(ndev, "Failed to create rxq %d : %d\n", i, err);
2416 goto out;
2417 }
2418
2419 u64_stats_init(&rxq->stats.syncp);
2420
2421 apc->rxqs[i] = rxq;
2422
2423 mana_create_rxq_debugfs(apc, i);
2424 }
2425
2426 apc->default_rxobj = apc->rxqs[0]->rxobj;
2427 out:
2428 return err;
2429 }
2430
mana_destroy_vport(struct mana_port_context * apc)2431 static void mana_destroy_vport(struct mana_port_context *apc)
2432 {
2433 struct gdma_dev *gd = apc->ac->gdma_dev;
2434 struct mana_rxq *rxq;
2435 u32 rxq_idx;
2436
2437 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
2438 rxq = apc->rxqs[rxq_idx];
2439 if (!rxq)
2440 continue;
2441
2442 mana_destroy_rxq(apc, rxq, true);
2443 apc->rxqs[rxq_idx] = NULL;
2444 }
2445
2446 mana_destroy_txq(apc);
2447 mana_uncfg_vport(apc);
2448
2449 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode)
2450 mana_pf_deregister_hw_vport(apc);
2451 }
2452
mana_create_vport(struct mana_port_context * apc,struct net_device * net)2453 static int mana_create_vport(struct mana_port_context *apc,
2454 struct net_device *net)
2455 {
2456 struct gdma_dev *gd = apc->ac->gdma_dev;
2457 int err;
2458
2459 apc->default_rxobj = INVALID_MANA_HANDLE;
2460
2461 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) {
2462 err = mana_pf_register_hw_vport(apc);
2463 if (err)
2464 return err;
2465 }
2466
2467 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell);
2468 if (err)
2469 return err;
2470
2471 return mana_create_txq(apc, net);
2472 }
2473
mana_rss_table_alloc(struct mana_port_context * apc)2474 static int mana_rss_table_alloc(struct mana_port_context *apc)
2475 {
2476 if (!apc->indir_table_sz) {
2477 netdev_err(apc->ndev,
2478 "Indirection table size not set for vPort %d\n",
2479 apc->port_idx);
2480 return -EINVAL;
2481 }
2482
2483 apc->indir_table = kcalloc(apc->indir_table_sz, sizeof(u32), GFP_KERNEL);
2484 if (!apc->indir_table)
2485 return -ENOMEM;
2486
2487 apc->rxobj_table = kcalloc(apc->indir_table_sz, sizeof(mana_handle_t), GFP_KERNEL);
2488 if (!apc->rxobj_table) {
2489 kfree(apc->indir_table);
2490 return -ENOMEM;
2491 }
2492
2493 return 0;
2494 }
2495
mana_rss_table_init(struct mana_port_context * apc)2496 static void mana_rss_table_init(struct mana_port_context *apc)
2497 {
2498 int i;
2499
2500 for (i = 0; i < apc->indir_table_sz; i++)
2501 apc->indir_table[i] =
2502 ethtool_rxfh_indir_default(i, apc->num_queues);
2503 }
2504
mana_config_rss(struct mana_port_context * apc,enum TRI_STATE rx,bool update_hash,bool update_tab)2505 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx,
2506 bool update_hash, bool update_tab)
2507 {
2508 u32 queue_idx;
2509 int err;
2510 int i;
2511
2512 if (update_tab) {
2513 for (i = 0; i < apc->indir_table_sz; i++) {
2514 queue_idx = apc->indir_table[i];
2515 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj;
2516 }
2517 }
2518
2519 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab);
2520 if (err)
2521 return err;
2522
2523 mana_fence_rqs(apc);
2524
2525 return 0;
2526 }
2527
mana_query_gf_stats(struct mana_port_context * apc)2528 void mana_query_gf_stats(struct mana_port_context *apc)
2529 {
2530 struct mana_query_gf_stat_resp resp = {};
2531 struct mana_query_gf_stat_req req = {};
2532 struct net_device *ndev = apc->ndev;
2533 int err;
2534
2535 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_GF_STAT,
2536 sizeof(req), sizeof(resp));
2537 req.hdr.resp.msg_version = GDMA_MESSAGE_V2;
2538 req.req_stats = STATISTICS_FLAGS_RX_DISCARDS_NO_WQE |
2539 STATISTICS_FLAGS_RX_ERRORS_VPORT_DISABLED |
2540 STATISTICS_FLAGS_HC_RX_BYTES |
2541 STATISTICS_FLAGS_HC_RX_UCAST_PACKETS |
2542 STATISTICS_FLAGS_HC_RX_UCAST_BYTES |
2543 STATISTICS_FLAGS_HC_RX_MCAST_PACKETS |
2544 STATISTICS_FLAGS_HC_RX_MCAST_BYTES |
2545 STATISTICS_FLAGS_HC_RX_BCAST_PACKETS |
2546 STATISTICS_FLAGS_HC_RX_BCAST_BYTES |
2547 STATISTICS_FLAGS_TX_ERRORS_GF_DISABLED |
2548 STATISTICS_FLAGS_TX_ERRORS_VPORT_DISABLED |
2549 STATISTICS_FLAGS_TX_ERRORS_INVAL_VPORT_OFFSET_PACKETS |
2550 STATISTICS_FLAGS_TX_ERRORS_VLAN_ENFORCEMENT |
2551 STATISTICS_FLAGS_TX_ERRORS_ETH_TYPE_ENFORCEMENT |
2552 STATISTICS_FLAGS_TX_ERRORS_SA_ENFORCEMENT |
2553 STATISTICS_FLAGS_TX_ERRORS_SQPDID_ENFORCEMENT |
2554 STATISTICS_FLAGS_TX_ERRORS_CQPDID_ENFORCEMENT |
2555 STATISTICS_FLAGS_TX_ERRORS_MTU_VIOLATION |
2556 STATISTICS_FLAGS_TX_ERRORS_INVALID_OOB |
2557 STATISTICS_FLAGS_HC_TX_BYTES |
2558 STATISTICS_FLAGS_HC_TX_UCAST_PACKETS |
2559 STATISTICS_FLAGS_HC_TX_UCAST_BYTES |
2560 STATISTICS_FLAGS_HC_TX_MCAST_PACKETS |
2561 STATISTICS_FLAGS_HC_TX_MCAST_BYTES |
2562 STATISTICS_FLAGS_HC_TX_BCAST_PACKETS |
2563 STATISTICS_FLAGS_HC_TX_BCAST_BYTES |
2564 STATISTICS_FLAGS_TX_ERRORS_GDMA_ERROR;
2565
2566 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
2567 sizeof(resp));
2568 if (err) {
2569 netdev_err(ndev, "Failed to query GF stats: %d\n", err);
2570 return;
2571 }
2572 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_GF_STAT,
2573 sizeof(resp));
2574 if (err || resp.hdr.status) {
2575 netdev_err(ndev, "Failed to query GF stats: %d, 0x%x\n", err,
2576 resp.hdr.status);
2577 return;
2578 }
2579
2580 apc->eth_stats.hc_rx_discards_no_wqe = resp.rx_discards_nowqe;
2581 apc->eth_stats.hc_rx_err_vport_disabled = resp.rx_err_vport_disabled;
2582 apc->eth_stats.hc_rx_bytes = resp.hc_rx_bytes;
2583 apc->eth_stats.hc_rx_ucast_pkts = resp.hc_rx_ucast_pkts;
2584 apc->eth_stats.hc_rx_ucast_bytes = resp.hc_rx_ucast_bytes;
2585 apc->eth_stats.hc_rx_bcast_pkts = resp.hc_rx_bcast_pkts;
2586 apc->eth_stats.hc_rx_bcast_bytes = resp.hc_rx_bcast_bytes;
2587 apc->eth_stats.hc_rx_mcast_pkts = resp.hc_rx_mcast_pkts;
2588 apc->eth_stats.hc_rx_mcast_bytes = resp.hc_rx_mcast_bytes;
2589 apc->eth_stats.hc_tx_err_gf_disabled = resp.tx_err_gf_disabled;
2590 apc->eth_stats.hc_tx_err_vport_disabled = resp.tx_err_vport_disabled;
2591 apc->eth_stats.hc_tx_err_inval_vportoffset_pkt =
2592 resp.tx_err_inval_vport_offset_pkt;
2593 apc->eth_stats.hc_tx_err_vlan_enforcement =
2594 resp.tx_err_vlan_enforcement;
2595 apc->eth_stats.hc_tx_err_eth_type_enforcement =
2596 resp.tx_err_ethtype_enforcement;
2597 apc->eth_stats.hc_tx_err_sa_enforcement = resp.tx_err_SA_enforcement;
2598 apc->eth_stats.hc_tx_err_sqpdid_enforcement =
2599 resp.tx_err_SQPDID_enforcement;
2600 apc->eth_stats.hc_tx_err_cqpdid_enforcement =
2601 resp.tx_err_CQPDID_enforcement;
2602 apc->eth_stats.hc_tx_err_mtu_violation = resp.tx_err_mtu_violation;
2603 apc->eth_stats.hc_tx_err_inval_oob = resp.tx_err_inval_oob;
2604 apc->eth_stats.hc_tx_bytes = resp.hc_tx_bytes;
2605 apc->eth_stats.hc_tx_ucast_pkts = resp.hc_tx_ucast_pkts;
2606 apc->eth_stats.hc_tx_ucast_bytes = resp.hc_tx_ucast_bytes;
2607 apc->eth_stats.hc_tx_bcast_pkts = resp.hc_tx_bcast_pkts;
2608 apc->eth_stats.hc_tx_bcast_bytes = resp.hc_tx_bcast_bytes;
2609 apc->eth_stats.hc_tx_mcast_pkts = resp.hc_tx_mcast_pkts;
2610 apc->eth_stats.hc_tx_mcast_bytes = resp.hc_tx_mcast_bytes;
2611 apc->eth_stats.hc_tx_err_gdma = resp.tx_err_gdma;
2612 }
2613
mana_init_port(struct net_device * ndev)2614 static int mana_init_port(struct net_device *ndev)
2615 {
2616 struct mana_port_context *apc = netdev_priv(ndev);
2617 struct gdma_dev *gd = apc->ac->gdma_dev;
2618 u32 max_txq, max_rxq, max_queues;
2619 int port_idx = apc->port_idx;
2620 struct gdma_context *gc;
2621 char vport[32];
2622 int err;
2623
2624 err = mana_init_port_context(apc);
2625 if (err)
2626 return err;
2627
2628 gc = gd->gdma_context;
2629
2630 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq,
2631 &apc->indir_table_sz);
2632 if (err) {
2633 netdev_err(ndev, "Failed to query info for vPort %d\n",
2634 port_idx);
2635 goto reset_apc;
2636 }
2637
2638 max_queues = min_t(u32, max_txq, max_rxq);
2639 if (apc->max_queues > max_queues)
2640 apc->max_queues = max_queues;
2641
2642 if (apc->num_queues > apc->max_queues)
2643 apc->num_queues = apc->max_queues;
2644
2645 eth_hw_addr_set(ndev, apc->mac_addr);
2646 sprintf(vport, "vport%d", port_idx);
2647 apc->mana_port_debugfs = debugfs_create_dir(vport, gc->mana_pci_debugfs);
2648 return 0;
2649
2650 reset_apc:
2651 mana_cleanup_port_context(apc);
2652 return err;
2653 }
2654
mana_alloc_queues(struct net_device * ndev)2655 int mana_alloc_queues(struct net_device *ndev)
2656 {
2657 struct mana_port_context *apc = netdev_priv(ndev);
2658 struct gdma_dev *gd = apc->ac->gdma_dev;
2659 int err;
2660
2661 err = mana_create_vport(apc, ndev);
2662 if (err) {
2663 netdev_err(ndev, "Failed to create vPort %u : %d\n", apc->port_idx, err);
2664 return err;
2665 }
2666
2667 err = netif_set_real_num_tx_queues(ndev, apc->num_queues);
2668 if (err) {
2669 netdev_err(ndev,
2670 "netif_set_real_num_tx_queues () failed for ndev with num_queues %u : %d\n",
2671 apc->num_queues, err);
2672 goto destroy_vport;
2673 }
2674
2675 err = mana_add_rx_queues(apc, ndev);
2676 if (err)
2677 goto destroy_vport;
2678
2679 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE;
2680
2681 err = netif_set_real_num_rx_queues(ndev, apc->num_queues);
2682 if (err) {
2683 netdev_err(ndev,
2684 "netif_set_real_num_rx_queues () failed for ndev with num_queues %u : %d\n",
2685 apc->num_queues, err);
2686 goto destroy_vport;
2687 }
2688
2689 mana_rss_table_init(apc);
2690
2691 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true);
2692 if (err) {
2693 netdev_err(ndev, "Failed to configure RSS table: %d\n", err);
2694 goto destroy_vport;
2695 }
2696
2697 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) {
2698 err = mana_pf_register_filter(apc);
2699 if (err)
2700 goto destroy_vport;
2701 }
2702
2703 mana_chn_setxdp(apc, mana_xdp_get(apc));
2704
2705 return 0;
2706
2707 destroy_vport:
2708 mana_destroy_vport(apc);
2709 return err;
2710 }
2711
mana_attach(struct net_device * ndev)2712 int mana_attach(struct net_device *ndev)
2713 {
2714 struct mana_port_context *apc = netdev_priv(ndev);
2715 int err;
2716
2717 ASSERT_RTNL();
2718
2719 err = mana_init_port(ndev);
2720 if (err)
2721 return err;
2722
2723 if (apc->port_st_save) {
2724 err = mana_alloc_queues(ndev);
2725 if (err) {
2726 mana_cleanup_port_context(apc);
2727 return err;
2728 }
2729 }
2730
2731 apc->port_is_up = apc->port_st_save;
2732
2733 /* Ensure port state updated before txq state */
2734 smp_wmb();
2735
2736 if (apc->port_is_up)
2737 netif_carrier_on(ndev);
2738
2739 netif_device_attach(ndev);
2740
2741 return 0;
2742 }
2743
mana_dealloc_queues(struct net_device * ndev)2744 static int mana_dealloc_queues(struct net_device *ndev)
2745 {
2746 struct mana_port_context *apc = netdev_priv(ndev);
2747 unsigned long timeout = jiffies + 120 * HZ;
2748 struct gdma_dev *gd = apc->ac->gdma_dev;
2749 struct mana_txq *txq;
2750 struct sk_buff *skb;
2751 int i, err;
2752 u32 tsleep;
2753
2754 if (apc->port_is_up)
2755 return -EINVAL;
2756
2757 mana_chn_setxdp(apc, NULL);
2758
2759 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode)
2760 mana_pf_deregister_filter(apc);
2761
2762 /* No packet can be transmitted now since apc->port_is_up is false.
2763 * There is still a tiny chance that mana_poll_tx_cq() can re-enable
2764 * a txq because it may not timely see apc->port_is_up being cleared
2765 * to false, but it doesn't matter since mana_start_xmit() drops any
2766 * new packets due to apc->port_is_up being false.
2767 *
2768 * Drain all the in-flight TX packets.
2769 * A timeout of 120 seconds for all the queues is used.
2770 * This will break the while loop when h/w is not responding.
2771 * This value of 120 has been decided here considering max
2772 * number of queues.
2773 */
2774
2775 for (i = 0; i < apc->num_queues; i++) {
2776 txq = &apc->tx_qp[i].txq;
2777 tsleep = 1000;
2778 while (atomic_read(&txq->pending_sends) > 0 &&
2779 time_before(jiffies, timeout)) {
2780 usleep_range(tsleep, tsleep + 1000);
2781 tsleep <<= 1;
2782 }
2783 if (atomic_read(&txq->pending_sends)) {
2784 err = pcie_flr(to_pci_dev(gd->gdma_context->dev));
2785 if (err) {
2786 netdev_err(ndev, "flr failed %d with %d pkts pending in txq %u\n",
2787 err, atomic_read(&txq->pending_sends),
2788 txq->gdma_txq_id);
2789 }
2790 break;
2791 }
2792 }
2793
2794 for (i = 0; i < apc->num_queues; i++) {
2795 txq = &apc->tx_qp[i].txq;
2796 while ((skb = skb_dequeue(&txq->pending_skbs))) {
2797 mana_unmap_skb(skb, apc);
2798 dev_kfree_skb_any(skb);
2799 }
2800 atomic_set(&txq->pending_sends, 0);
2801 }
2802 /* We're 100% sure the queues can no longer be woken up, because
2803 * we're sure now mana_poll_tx_cq() can't be running.
2804 */
2805
2806 apc->rss_state = TRI_STATE_FALSE;
2807 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false);
2808 if (err) {
2809 netdev_err(ndev, "Failed to disable vPort: %d\n", err);
2810 return err;
2811 }
2812
2813 mana_destroy_vport(apc);
2814
2815 return 0;
2816 }
2817
mana_detach(struct net_device * ndev,bool from_close)2818 int mana_detach(struct net_device *ndev, bool from_close)
2819 {
2820 struct mana_port_context *apc = netdev_priv(ndev);
2821 int err;
2822
2823 ASSERT_RTNL();
2824
2825 apc->port_st_save = apc->port_is_up;
2826 apc->port_is_up = false;
2827
2828 /* Ensure port state updated before txq state */
2829 smp_wmb();
2830
2831 netif_tx_disable(ndev);
2832 netif_carrier_off(ndev);
2833
2834 if (apc->port_st_save) {
2835 err = mana_dealloc_queues(ndev);
2836 if (err) {
2837 netdev_err(ndev, "%s failed to deallocate queues: %d\n", __func__, err);
2838 return err;
2839 }
2840 }
2841
2842 if (!from_close) {
2843 netif_device_detach(ndev);
2844 mana_cleanup_port_context(apc);
2845 }
2846
2847 return 0;
2848 }
2849
mana_probe_port(struct mana_context * ac,int port_idx,struct net_device ** ndev_storage)2850 static int mana_probe_port(struct mana_context *ac, int port_idx,
2851 struct net_device **ndev_storage)
2852 {
2853 struct gdma_context *gc = ac->gdma_dev->gdma_context;
2854 struct mana_port_context *apc;
2855 struct net_device *ndev;
2856 int err;
2857
2858 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context),
2859 gc->max_num_queues);
2860 if (!ndev)
2861 return -ENOMEM;
2862
2863 *ndev_storage = ndev;
2864
2865 apc = netdev_priv(ndev);
2866 apc->ac = ac;
2867 apc->ndev = ndev;
2868 apc->max_queues = gc->max_num_queues;
2869 apc->num_queues = gc->max_num_queues;
2870 apc->tx_queue_size = DEF_TX_BUFFERS_PER_QUEUE;
2871 apc->rx_queue_size = DEF_RX_BUFFERS_PER_QUEUE;
2872 apc->port_handle = INVALID_MANA_HANDLE;
2873 apc->pf_filter_handle = INVALID_MANA_HANDLE;
2874 apc->port_idx = port_idx;
2875
2876 mutex_init(&apc->vport_mutex);
2877 apc->vport_use_count = 0;
2878
2879 ndev->netdev_ops = &mana_devops;
2880 ndev->ethtool_ops = &mana_ethtool_ops;
2881 ndev->mtu = ETH_DATA_LEN;
2882 ndev->max_mtu = gc->adapter_mtu - ETH_HLEN;
2883 ndev->min_mtu = ETH_MIN_MTU;
2884 ndev->needed_headroom = MANA_HEADROOM;
2885 ndev->dev_port = port_idx;
2886 SET_NETDEV_DEV(ndev, gc->dev);
2887
2888 netif_set_tso_max_size(ndev, GSO_MAX_SIZE);
2889
2890 netif_carrier_off(ndev);
2891
2892 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE);
2893
2894 err = mana_init_port(ndev);
2895 if (err)
2896 goto free_net;
2897
2898 err = mana_rss_table_alloc(apc);
2899 if (err)
2900 goto reset_apc;
2901
2902 netdev_lockdep_set_classes(ndev);
2903
2904 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
2905 ndev->hw_features |= NETIF_F_RXCSUM;
2906 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
2907 ndev->hw_features |= NETIF_F_RXHASH;
2908 ndev->features = ndev->hw_features | NETIF_F_HW_VLAN_CTAG_TX |
2909 NETIF_F_HW_VLAN_CTAG_RX;
2910 ndev->vlan_features = ndev->features;
2911 xdp_set_features_flag(ndev, NETDEV_XDP_ACT_BASIC |
2912 NETDEV_XDP_ACT_REDIRECT |
2913 NETDEV_XDP_ACT_NDO_XMIT);
2914
2915 err = register_netdev(ndev);
2916 if (err) {
2917 netdev_err(ndev, "Unable to register netdev.\n");
2918 goto free_indir;
2919 }
2920
2921 return 0;
2922
2923 free_indir:
2924 mana_cleanup_indir_table(apc);
2925 reset_apc:
2926 mana_cleanup_port_context(apc);
2927 free_net:
2928 *ndev_storage = NULL;
2929 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err);
2930 free_netdev(ndev);
2931 return err;
2932 }
2933
adev_release(struct device * dev)2934 static void adev_release(struct device *dev)
2935 {
2936 struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev);
2937
2938 kfree(madev);
2939 }
2940
remove_adev(struct gdma_dev * gd)2941 static void remove_adev(struct gdma_dev *gd)
2942 {
2943 struct auxiliary_device *adev = gd->adev;
2944 int id = adev->id;
2945
2946 auxiliary_device_delete(adev);
2947 auxiliary_device_uninit(adev);
2948
2949 mana_adev_idx_free(id);
2950 gd->adev = NULL;
2951 }
2952
add_adev(struct gdma_dev * gd,const char * name)2953 static int add_adev(struct gdma_dev *gd, const char *name)
2954 {
2955 struct auxiliary_device *adev;
2956 struct mana_adev *madev;
2957 int ret;
2958
2959 madev = kzalloc(sizeof(*madev), GFP_KERNEL);
2960 if (!madev)
2961 return -ENOMEM;
2962
2963 adev = &madev->adev;
2964 ret = mana_adev_idx_alloc();
2965 if (ret < 0)
2966 goto idx_fail;
2967 adev->id = ret;
2968
2969 adev->name = name;
2970 adev->dev.parent = gd->gdma_context->dev;
2971 adev->dev.release = adev_release;
2972 madev->mdev = gd;
2973
2974 ret = auxiliary_device_init(adev);
2975 if (ret)
2976 goto init_fail;
2977
2978 /* madev is owned by the auxiliary device */
2979 madev = NULL;
2980 ret = auxiliary_device_add(adev);
2981 if (ret)
2982 goto add_fail;
2983
2984 gd->adev = adev;
2985 dev_dbg(gd->gdma_context->dev,
2986 "Auxiliary device added successfully\n");
2987 return 0;
2988
2989 add_fail:
2990 auxiliary_device_uninit(adev);
2991
2992 init_fail:
2993 mana_adev_idx_free(adev->id);
2994
2995 idx_fail:
2996 kfree(madev);
2997
2998 return ret;
2999 }
3000
mana_rdma_service_handle(struct work_struct * work)3001 static void mana_rdma_service_handle(struct work_struct *work)
3002 {
3003 struct mana_service_work *serv_work =
3004 container_of(work, struct mana_service_work, work);
3005 struct gdma_dev *gd = serv_work->gdma_dev;
3006 struct device *dev = gd->gdma_context->dev;
3007 int ret;
3008
3009 if (READ_ONCE(gd->rdma_teardown))
3010 goto out;
3011
3012 switch (serv_work->event) {
3013 case GDMA_SERVICE_TYPE_RDMA_SUSPEND:
3014 if (!gd->adev || gd->is_suspended)
3015 break;
3016
3017 remove_adev(gd);
3018 gd->is_suspended = true;
3019 break;
3020
3021 case GDMA_SERVICE_TYPE_RDMA_RESUME:
3022 if (!gd->is_suspended)
3023 break;
3024
3025 ret = add_adev(gd, "rdma");
3026 if (ret)
3027 dev_err(dev, "Failed to add adev on resume: %d\n", ret);
3028 else
3029 gd->is_suspended = false;
3030 break;
3031
3032 default:
3033 dev_warn(dev, "unknown adev service event %u\n",
3034 serv_work->event);
3035 break;
3036 }
3037
3038 out:
3039 kfree(serv_work);
3040 }
3041
mana_rdma_service_event(struct gdma_context * gc,enum gdma_service_type event)3042 int mana_rdma_service_event(struct gdma_context *gc, enum gdma_service_type event)
3043 {
3044 struct gdma_dev *gd = &gc->mana_ib;
3045 struct mana_service_work *serv_work;
3046
3047 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) {
3048 /* RDMA device is not detected on pci */
3049 return 0;
3050 }
3051
3052 serv_work = kzalloc(sizeof(*serv_work), GFP_ATOMIC);
3053 if (!serv_work)
3054 return -ENOMEM;
3055
3056 serv_work->event = event;
3057 serv_work->gdma_dev = gd;
3058
3059 INIT_WORK(&serv_work->work, mana_rdma_service_handle);
3060 queue_work(gc->service_wq, &serv_work->work);
3061
3062 return 0;
3063 }
3064
mana_probe(struct gdma_dev * gd,bool resuming)3065 int mana_probe(struct gdma_dev *gd, bool resuming)
3066 {
3067 struct gdma_context *gc = gd->gdma_context;
3068 struct mana_context *ac = gd->driver_data;
3069 struct device *dev = gc->dev;
3070 u8 bm_hostmode = 0;
3071 u16 num_ports = 0;
3072 int err;
3073 int i;
3074
3075 dev_info(dev,
3076 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n",
3077 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION);
3078
3079 err = mana_gd_register_device(gd);
3080 if (err)
3081 return err;
3082
3083 if (!resuming) {
3084 ac = kzalloc(sizeof(*ac), GFP_KERNEL);
3085 if (!ac)
3086 return -ENOMEM;
3087
3088 ac->gdma_dev = gd;
3089 gd->driver_data = ac;
3090 }
3091
3092 err = mana_create_eq(ac);
3093 if (err) {
3094 dev_err(dev, "Failed to create EQs: %d\n", err);
3095 goto out;
3096 }
3097
3098 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION,
3099 MANA_MICRO_VERSION, &num_ports, &bm_hostmode);
3100 if (err)
3101 goto out;
3102
3103 ac->bm_hostmode = bm_hostmode;
3104
3105 if (!resuming) {
3106 ac->num_ports = num_ports;
3107 } else {
3108 if (ac->num_ports != num_ports) {
3109 dev_err(dev, "The number of vPorts changed: %d->%d\n",
3110 ac->num_ports, num_ports);
3111 err = -EPROTO;
3112 goto out;
3113 }
3114 }
3115
3116 if (ac->num_ports == 0)
3117 dev_err(dev, "Failed to detect any vPort\n");
3118
3119 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV)
3120 ac->num_ports = MAX_PORTS_IN_MANA_DEV;
3121
3122 if (!resuming) {
3123 for (i = 0; i < ac->num_ports; i++) {
3124 err = mana_probe_port(ac, i, &ac->ports[i]);
3125 /* we log the port for which the probe failed and stop
3126 * probes for subsequent ports.
3127 * Note that we keep running ports, for which the probes
3128 * were successful, unless add_adev fails too
3129 */
3130 if (err) {
3131 dev_err(dev, "Probe Failed for port %d\n", i);
3132 break;
3133 }
3134 }
3135 } else {
3136 for (i = 0; i < ac->num_ports; i++) {
3137 rtnl_lock();
3138 err = mana_attach(ac->ports[i]);
3139 rtnl_unlock();
3140 /* we log the port for which the attach failed and stop
3141 * attach for subsequent ports
3142 * Note that we keep running ports, for which the attach
3143 * were successful, unless add_adev fails too
3144 */
3145 if (err) {
3146 dev_err(dev, "Attach Failed for port %d\n", i);
3147 break;
3148 }
3149 }
3150 }
3151
3152 err = add_adev(gd, "eth");
3153 out:
3154 if (err) {
3155 mana_remove(gd, false);
3156 } else {
3157 dev_dbg(dev, "gd=%p, id=%u, num_ports=%d, type=%u, instance=%u\n",
3158 gd, gd->dev_id.as_uint32, ac->num_ports,
3159 gd->dev_id.type, gd->dev_id.instance);
3160 dev_dbg(dev, "%s succeeded\n", __func__);
3161 }
3162
3163 return err;
3164 }
3165
mana_remove(struct gdma_dev * gd,bool suspending)3166 void mana_remove(struct gdma_dev *gd, bool suspending)
3167 {
3168 struct gdma_context *gc = gd->gdma_context;
3169 struct mana_context *ac = gd->driver_data;
3170 struct mana_port_context *apc;
3171 struct device *dev = gc->dev;
3172 struct net_device *ndev;
3173 int err;
3174 int i;
3175
3176 /* adev currently doesn't support suspending, always remove it */
3177 if (gd->adev)
3178 remove_adev(gd);
3179
3180 for (i = 0; i < ac->num_ports; i++) {
3181 ndev = ac->ports[i];
3182 apc = netdev_priv(ndev);
3183 if (!ndev) {
3184 if (i == 0)
3185 dev_err(dev, "No net device to remove\n");
3186 goto out;
3187 }
3188
3189 /* All cleanup actions should stay after rtnl_lock(), otherwise
3190 * other functions may access partially cleaned up data.
3191 */
3192 rtnl_lock();
3193
3194 err = mana_detach(ndev, false);
3195 if (err)
3196 netdev_err(ndev, "Failed to detach vPort %d: %d\n",
3197 i, err);
3198
3199 if (suspending) {
3200 /* No need to unregister the ndev. */
3201 rtnl_unlock();
3202 continue;
3203 }
3204
3205 unregister_netdevice(ndev);
3206 mana_cleanup_indir_table(apc);
3207
3208 rtnl_unlock();
3209
3210 free_netdev(ndev);
3211 }
3212
3213 mana_destroy_eq(ac);
3214 out:
3215 mana_gd_deregister_device(gd);
3216
3217 if (suspending)
3218 return;
3219
3220 gd->driver_data = NULL;
3221 gd->gdma_context = NULL;
3222 kfree(ac);
3223 dev_dbg(dev, "%s succeeded\n", __func__);
3224 }
3225
mana_rdma_probe(struct gdma_dev * gd)3226 int mana_rdma_probe(struct gdma_dev *gd)
3227 {
3228 int err = 0;
3229
3230 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) {
3231 /* RDMA device is not detected on pci */
3232 return err;
3233 }
3234
3235 err = mana_gd_register_device(gd);
3236 if (err)
3237 return err;
3238
3239 err = add_adev(gd, "rdma");
3240 if (err)
3241 mana_gd_deregister_device(gd);
3242
3243 return err;
3244 }
3245
mana_rdma_remove(struct gdma_dev * gd)3246 void mana_rdma_remove(struct gdma_dev *gd)
3247 {
3248 struct gdma_context *gc = gd->gdma_context;
3249
3250 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) {
3251 /* RDMA device is not detected on pci */
3252 return;
3253 }
3254
3255 WRITE_ONCE(gd->rdma_teardown, true);
3256 flush_workqueue(gc->service_wq);
3257
3258 if (gd->adev)
3259 remove_adev(gd);
3260
3261 mana_gd_deregister_device(gd);
3262 }
3263
mana_get_primary_netdev(struct mana_context * ac,u32 port_index,netdevice_tracker * tracker)3264 struct net_device *mana_get_primary_netdev(struct mana_context *ac,
3265 u32 port_index,
3266 netdevice_tracker *tracker)
3267 {
3268 struct net_device *ndev;
3269
3270 if (port_index >= ac->num_ports)
3271 return NULL;
3272
3273 rcu_read_lock();
3274
3275 /* If mana is used in netvsc, the upper netdevice should be returned. */
3276 ndev = netdev_master_upper_dev_get_rcu(ac->ports[port_index]);
3277
3278 /* If there is no upper device, use the parent Ethernet device */
3279 if (!ndev)
3280 ndev = ac->ports[port_index];
3281
3282 netdev_hold(ndev, tracker, GFP_ATOMIC);
3283 rcu_read_unlock();
3284
3285 return ndev;
3286 }
3287 EXPORT_SYMBOL_NS(mana_get_primary_netdev, "NET_MANA");
3288