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