1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /* Copyright (c) 2021, Microsoft Corporation. */
4 #include <uapi/linux/bpf.h>
6 #include <linux/inetdevice.h>
7 #include <linux/etherdevice.h>
8 #include <linux/ethtool.h>
9 #include <linux/filter.h>
12 #include <net/checksum.h>
13 #include <net/ip6_checksum.h>
15 #include <net/mana/mana.h>
16 #include <net/mana/mana_auxiliary.h>
18 static DEFINE_IDA(mana_adev_ida);
20 static int mana_adev_idx_alloc(void)
22 return ida_alloc(&mana_adev_ida, GFP_KERNEL);
25 static void mana_adev_idx_free(int idx)
27 ida_free(&mana_adev_ida, idx);
30 /* Microsoft Azure Network Adapter (MANA) functions */
32 static int mana_open(struct net_device *ndev)
34 struct mana_port_context *apc = netdev_priv(ndev);
37 err = mana_alloc_queues(ndev);
41 apc->port_is_up = true;
43 /* Ensure port state updated before txq state */
46 netif_carrier_on(ndev);
47 netif_tx_wake_all_queues(ndev);
52 static int mana_close(struct net_device *ndev)
54 struct mana_port_context *apc = netdev_priv(ndev);
59 return mana_detach(ndev, true);
62 static bool mana_can_tx(struct gdma_queue *wq)
64 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE;
67 static unsigned int mana_checksum_info(struct sk_buff *skb)
69 if (skb->protocol == htons(ETH_P_IP)) {
70 struct iphdr *ip = ip_hdr(skb);
72 if (ip->protocol == IPPROTO_TCP)
75 if (ip->protocol == IPPROTO_UDP)
77 } else if (skb->protocol == htons(ETH_P_IPV6)) {
78 struct ipv6hdr *ip6 = ipv6_hdr(skb);
80 if (ip6->nexthdr == IPPROTO_TCP)
83 if (ip6->nexthdr == IPPROTO_UDP)
87 /* No csum offloading */
91 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc,
92 struct mana_tx_package *tp)
94 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
95 struct gdma_dev *gd = apc->ac->gdma_dev;
96 struct gdma_context *gc;
102 gc = gd->gdma_context;
104 da = dma_map_single(dev, skb->data, skb_headlen(skb), DMA_TO_DEVICE);
106 if (dma_mapping_error(dev, da))
109 ash->dma_handle[0] = da;
110 ash->size[0] = skb_headlen(skb);
112 tp->wqe_req.sgl[0].address = ash->dma_handle[0];
113 tp->wqe_req.sgl[0].mem_key = gd->gpa_mkey;
114 tp->wqe_req.sgl[0].size = ash->size[0];
116 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
117 frag = &skb_shinfo(skb)->frags[i];
118 da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag),
121 if (dma_mapping_error(dev, da))
124 ash->dma_handle[i + 1] = da;
125 ash->size[i + 1] = skb_frag_size(frag);
127 tp->wqe_req.sgl[i + 1].address = ash->dma_handle[i + 1];
128 tp->wqe_req.sgl[i + 1].mem_key = gd->gpa_mkey;
129 tp->wqe_req.sgl[i + 1].size = ash->size[i + 1];
135 for (i = i - 1; i >= 0; i--)
136 dma_unmap_page(dev, ash->dma_handle[i + 1], ash->size[i + 1],
139 dma_unmap_single(dev, ash->dma_handle[0], ash->size[0], DMA_TO_DEVICE);
144 netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev)
146 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT;
147 struct mana_port_context *apc = netdev_priv(ndev);
148 u16 txq_idx = skb_get_queue_mapping(skb);
149 struct gdma_dev *gd = apc->ac->gdma_dev;
150 bool ipv4 = false, ipv6 = false;
151 struct mana_tx_package pkg = {};
152 struct netdev_queue *net_txq;
153 struct mana_stats_tx *tx_stats;
154 struct gdma_queue *gdma_sq;
155 unsigned int csum_type;
156 struct mana_txq *txq;
161 if (unlikely(!apc->port_is_up))
164 if (skb_cow_head(skb, MANA_HEADROOM))
167 txq = &apc->tx_qp[txq_idx].txq;
168 gdma_sq = txq->gdma_sq;
169 cq = &apc->tx_qp[txq_idx].tx_cq;
170 tx_stats = &txq->stats;
172 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id;
173 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame;
175 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) {
176 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset;
177 pkt_fmt = MANA_LONG_PKT_FMT;
179 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset;
182 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt;
184 if (pkt_fmt == MANA_SHORT_PKT_FMT) {
185 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob);
186 u64_stats_update_begin(&tx_stats->syncp);
187 tx_stats->short_pkt_fmt++;
188 u64_stats_update_end(&tx_stats->syncp);
190 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob);
191 u64_stats_update_begin(&tx_stats->syncp);
192 tx_stats->long_pkt_fmt++;
193 u64_stats_update_end(&tx_stats->syncp);
196 pkg.wqe_req.inline_oob_data = &pkg.tx_oob;
197 pkg.wqe_req.flags = 0;
198 pkg.wqe_req.client_data_unit = 0;
200 pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags;
201 WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES);
203 if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) {
204 pkg.wqe_req.sgl = pkg.sgl_array;
206 pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge,
207 sizeof(struct gdma_sge),
212 pkg.wqe_req.sgl = pkg.sgl_ptr;
215 if (skb->protocol == htons(ETH_P_IP))
217 else if (skb->protocol == htons(ETH_P_IPV6))
220 if (skb_is_gso(skb)) {
221 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
222 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
224 pkg.tx_oob.s_oob.comp_iphdr_csum = 1;
225 pkg.tx_oob.s_oob.comp_tcp_csum = 1;
226 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
228 pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size;
229 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0;
231 ip_hdr(skb)->tot_len = 0;
232 ip_hdr(skb)->check = 0;
233 tcp_hdr(skb)->check =
234 ~csum_tcpudp_magic(ip_hdr(skb)->saddr,
235 ip_hdr(skb)->daddr, 0,
238 ipv6_hdr(skb)->payload_len = 0;
239 tcp_hdr(skb)->check =
240 ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
241 &ipv6_hdr(skb)->daddr, 0,
245 if (skb->encapsulation) {
246 ihs = skb_inner_tcp_all_headers(skb);
247 u64_stats_update_begin(&tx_stats->syncp);
248 tx_stats->tso_inner_packets++;
249 tx_stats->tso_inner_bytes += skb->len - ihs;
250 u64_stats_update_end(&tx_stats->syncp);
252 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
253 ihs = skb_transport_offset(skb) + sizeof(struct udphdr);
255 ihs = skb_tcp_all_headers(skb);
256 if (ipv6_has_hopopt_jumbo(skb))
257 ihs -= sizeof(struct hop_jumbo_hdr);
260 u64_stats_update_begin(&tx_stats->syncp);
261 tx_stats->tso_packets++;
262 tx_stats->tso_bytes += skb->len - ihs;
263 u64_stats_update_end(&tx_stats->syncp);
266 } else if (skb->ip_summed == CHECKSUM_PARTIAL) {
267 csum_type = mana_checksum_info(skb);
269 u64_stats_update_begin(&tx_stats->syncp);
270 tx_stats->csum_partial++;
271 u64_stats_update_end(&tx_stats->syncp);
273 if (csum_type == IPPROTO_TCP) {
274 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
275 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
277 pkg.tx_oob.s_oob.comp_tcp_csum = 1;
278 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb);
280 } else if (csum_type == IPPROTO_UDP) {
281 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4;
282 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6;
284 pkg.tx_oob.s_oob.comp_udp_csum = 1;
286 /* Can't do offload of this type of checksum */
287 if (skb_checksum_help(skb))
292 if (mana_map_skb(skb, apc, &pkg)) {
293 u64_stats_update_begin(&tx_stats->syncp);
294 tx_stats->mana_map_err++;
295 u64_stats_update_end(&tx_stats->syncp);
299 skb_queue_tail(&txq->pending_skbs, skb);
302 net_txq = netdev_get_tx_queue(ndev, txq_idx);
304 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req,
305 (struct gdma_posted_wqe_info *)skb->cb);
306 if (!mana_can_tx(gdma_sq)) {
307 netif_tx_stop_queue(net_txq);
308 apc->eth_stats.stop_queue++;
312 (void)skb_dequeue_tail(&txq->pending_skbs);
313 netdev_warn(ndev, "Failed to post TX OOB: %d\n", err);
314 err = NETDEV_TX_BUSY;
319 atomic_inc(&txq->pending_sends);
321 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq);
323 /* skb may be freed after mana_gd_post_work_request. Do not use it. */
326 tx_stats = &txq->stats;
327 u64_stats_update_begin(&tx_stats->syncp);
329 tx_stats->bytes += len;
330 u64_stats_update_end(&tx_stats->syncp);
333 if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) {
334 netif_tx_wake_queue(net_txq);
335 apc->eth_stats.wake_queue++;
344 ndev->stats.tx_dropped++;
346 dev_kfree_skb_any(skb);
350 static void mana_get_stats64(struct net_device *ndev,
351 struct rtnl_link_stats64 *st)
353 struct mana_port_context *apc = netdev_priv(ndev);
354 unsigned int num_queues = apc->num_queues;
355 struct mana_stats_rx *rx_stats;
356 struct mana_stats_tx *tx_stats;
361 if (!apc->port_is_up)
364 netdev_stats_to_stats64(st, &ndev->stats);
366 for (q = 0; q < num_queues; q++) {
367 rx_stats = &apc->rxqs[q]->stats;
370 start = u64_stats_fetch_begin(&rx_stats->syncp);
371 packets = rx_stats->packets;
372 bytes = rx_stats->bytes;
373 } while (u64_stats_fetch_retry(&rx_stats->syncp, start));
375 st->rx_packets += packets;
376 st->rx_bytes += bytes;
379 for (q = 0; q < num_queues; q++) {
380 tx_stats = &apc->tx_qp[q].txq.stats;
383 start = u64_stats_fetch_begin(&tx_stats->syncp);
384 packets = tx_stats->packets;
385 bytes = tx_stats->bytes;
386 } while (u64_stats_fetch_retry(&tx_stats->syncp, start));
388 st->tx_packets += packets;
389 st->tx_bytes += bytes;
393 static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb,
396 struct mana_port_context *apc = netdev_priv(ndev);
397 u32 hash = skb_get_hash(skb);
398 struct sock *sk = skb->sk;
401 txq = apc->indir_table[hash & MANA_INDIRECT_TABLE_MASK];
403 if (txq != old_q && sk && sk_fullsock(sk) &&
404 rcu_access_pointer(sk->sk_dst_cache))
405 sk_tx_queue_set(sk, txq);
410 static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb,
411 struct net_device *sb_dev)
415 if (ndev->real_num_tx_queues == 1)
418 txq = sk_tx_queue_get(skb->sk);
420 if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) {
421 if (skb_rx_queue_recorded(skb))
422 txq = skb_get_rx_queue(skb);
424 txq = mana_get_tx_queue(ndev, skb, txq);
430 /* Release pre-allocated RX buffers */
431 static void mana_pre_dealloc_rxbufs(struct mana_port_context *mpc)
436 dev = mpc->ac->gdma_dev->gdma_context->dev;
438 if (!mpc->rxbufs_pre)
444 while (mpc->rxbpre_total) {
445 i = --mpc->rxbpre_total;
446 dma_unmap_single(dev, mpc->das_pre[i], mpc->rxbpre_datasize,
448 put_page(virt_to_head_page(mpc->rxbufs_pre[i]));
455 kfree(mpc->rxbufs_pre);
456 mpc->rxbufs_pre = NULL;
459 mpc->rxbpre_datasize = 0;
460 mpc->rxbpre_alloc_size = 0;
461 mpc->rxbpre_headroom = 0;
464 /* Get a buffer from the pre-allocated RX buffers */
465 static void *mana_get_rxbuf_pre(struct mana_rxq *rxq, dma_addr_t *da)
467 struct net_device *ndev = rxq->ndev;
468 struct mana_port_context *mpc;
471 mpc = netdev_priv(ndev);
473 if (!mpc->rxbufs_pre || !mpc->das_pre || !mpc->rxbpre_total) {
474 netdev_err(ndev, "No RX pre-allocated bufs\n");
478 /* Check sizes to catch unexpected coding error */
479 if (mpc->rxbpre_datasize != rxq->datasize) {
480 netdev_err(ndev, "rxbpre_datasize mismatch: %u: %u\n",
481 mpc->rxbpre_datasize, rxq->datasize);
485 if (mpc->rxbpre_alloc_size != rxq->alloc_size) {
486 netdev_err(ndev, "rxbpre_alloc_size mismatch: %u: %u\n",
487 mpc->rxbpre_alloc_size, rxq->alloc_size);
491 if (mpc->rxbpre_headroom != rxq->headroom) {
492 netdev_err(ndev, "rxbpre_headroom mismatch: %u: %u\n",
493 mpc->rxbpre_headroom, rxq->headroom);
499 *da = mpc->das_pre[mpc->rxbpre_total];
500 va = mpc->rxbufs_pre[mpc->rxbpre_total];
501 mpc->rxbufs_pre[mpc->rxbpre_total] = NULL;
503 /* Deallocate the array after all buffers are gone */
504 if (!mpc->rxbpre_total)
505 mana_pre_dealloc_rxbufs(mpc);
510 /* Get RX buffer's data size, alloc size, XDP headroom based on MTU */
511 static void mana_get_rxbuf_cfg(int mtu, u32 *datasize, u32 *alloc_size,
514 if (mtu > MANA_XDP_MTU_MAX)
515 *headroom = 0; /* no support for XDP */
517 *headroom = XDP_PACKET_HEADROOM;
519 *alloc_size = mtu + MANA_RXBUF_PAD + *headroom;
521 *datasize = ALIGN(mtu + ETH_HLEN, MANA_RX_DATA_ALIGN);
524 static int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu)
533 mana_get_rxbuf_cfg(new_mtu, &mpc->rxbpre_datasize,
534 &mpc->rxbpre_alloc_size, &mpc->rxbpre_headroom);
536 dev = mpc->ac->gdma_dev->gdma_context->dev;
538 num_rxb = mpc->num_queues * RX_BUFFERS_PER_QUEUE;
540 WARN(mpc->rxbufs_pre, "mana rxbufs_pre exists\n");
541 mpc->rxbufs_pre = kmalloc_array(num_rxb, sizeof(void *), GFP_KERNEL);
542 if (!mpc->rxbufs_pre)
545 mpc->das_pre = kmalloc_array(num_rxb, sizeof(dma_addr_t), GFP_KERNEL);
549 mpc->rxbpre_total = 0;
551 for (i = 0; i < num_rxb; i++) {
552 if (mpc->rxbpre_alloc_size > PAGE_SIZE) {
553 va = netdev_alloc_frag(mpc->rxbpre_alloc_size);
557 page = virt_to_head_page(va);
558 /* Check if the frag falls back to single page */
559 if (compound_order(page) <
560 get_order(mpc->rxbpre_alloc_size)) {
565 page = dev_alloc_page();
569 va = page_to_virt(page);
572 da = dma_map_single(dev, va + mpc->rxbpre_headroom,
573 mpc->rxbpre_datasize, DMA_FROM_DEVICE);
574 if (dma_mapping_error(dev, da)) {
575 put_page(virt_to_head_page(va));
579 mpc->rxbufs_pre[i] = va;
580 mpc->das_pre[i] = da;
581 mpc->rxbpre_total = i + 1;
587 mana_pre_dealloc_rxbufs(mpc);
591 static int mana_change_mtu(struct net_device *ndev, int new_mtu)
593 struct mana_port_context *mpc = netdev_priv(ndev);
594 unsigned int old_mtu = ndev->mtu;
597 /* Pre-allocate buffers to prevent failure in mana_attach later */
598 err = mana_pre_alloc_rxbufs(mpc, new_mtu);
600 netdev_err(ndev, "Insufficient memory for new MTU\n");
604 err = mana_detach(ndev, false);
606 netdev_err(ndev, "mana_detach failed: %d\n", err);
612 err = mana_attach(ndev);
614 netdev_err(ndev, "mana_attach failed: %d\n", err);
619 mana_pre_dealloc_rxbufs(mpc);
623 static const struct net_device_ops mana_devops = {
624 .ndo_open = mana_open,
625 .ndo_stop = mana_close,
626 .ndo_select_queue = mana_select_queue,
627 .ndo_start_xmit = mana_start_xmit,
628 .ndo_validate_addr = eth_validate_addr,
629 .ndo_get_stats64 = mana_get_stats64,
631 .ndo_xdp_xmit = mana_xdp_xmit,
632 .ndo_change_mtu = mana_change_mtu,
635 static void mana_cleanup_port_context(struct mana_port_context *apc)
641 static int mana_init_port_context(struct mana_port_context *apc)
643 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *),
646 return !apc->rxqs ? -ENOMEM : 0;
649 static int mana_send_request(struct mana_context *ac, void *in_buf,
650 u32 in_len, void *out_buf, u32 out_len)
652 struct gdma_context *gc = ac->gdma_dev->gdma_context;
653 struct gdma_resp_hdr *resp = out_buf;
654 struct gdma_req_hdr *req = in_buf;
655 struct device *dev = gc->dev;
656 static atomic_t activity_id;
659 req->dev_id = gc->mana.dev_id;
660 req->activity_id = atomic_inc_return(&activity_id);
662 err = mana_gd_send_request(gc, in_len, in_buf, out_len,
664 if (err || resp->status) {
665 dev_err(dev, "Failed to send mana message: %d, 0x%x\n",
667 return err ? err : -EPROTO;
670 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 ||
671 req->activity_id != resp->activity_id) {
672 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n",
673 req->dev_id.as_uint32, resp->dev_id.as_uint32,
674 req->activity_id, resp->activity_id);
681 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr,
682 const enum mana_command_code expected_code,
685 if (resp_hdr->response.msg_type != expected_code)
688 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1)
691 if (resp_hdr->response.msg_size < min_size)
697 static int mana_pf_register_hw_vport(struct mana_port_context *apc)
699 struct mana_register_hw_vport_resp resp = {};
700 struct mana_register_hw_vport_req req = {};
703 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT,
704 sizeof(req), sizeof(resp));
705 req.attached_gfid = 1;
706 req.is_pf_default_vport = 1;
707 req.allow_all_ether_types = 1;
709 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
712 netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err);
716 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT,
718 if (err || resp.hdr.status) {
719 netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n",
720 err, resp.hdr.status);
721 return err ? err : -EPROTO;
724 apc->port_handle = resp.hw_vport_handle;
728 static void mana_pf_deregister_hw_vport(struct mana_port_context *apc)
730 struct mana_deregister_hw_vport_resp resp = {};
731 struct mana_deregister_hw_vport_req req = {};
734 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT,
735 sizeof(req), sizeof(resp));
736 req.hw_vport_handle = apc->port_handle;
738 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
741 netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n",
746 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT,
748 if (err || resp.hdr.status)
749 netdev_err(apc->ndev,
750 "Failed to deregister hw vPort: %d, 0x%x\n",
751 err, resp.hdr.status);
754 static int mana_pf_register_filter(struct mana_port_context *apc)
756 struct mana_register_filter_resp resp = {};
757 struct mana_register_filter_req req = {};
760 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER,
761 sizeof(req), sizeof(resp));
762 req.vport = apc->port_handle;
763 memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN);
765 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
768 netdev_err(apc->ndev, "Failed to register filter: %d\n", err);
772 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER,
774 if (err || resp.hdr.status) {
775 netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n",
776 err, resp.hdr.status);
777 return err ? err : -EPROTO;
780 apc->pf_filter_handle = resp.filter_handle;
784 static void mana_pf_deregister_filter(struct mana_port_context *apc)
786 struct mana_deregister_filter_resp resp = {};
787 struct mana_deregister_filter_req req = {};
790 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER,
791 sizeof(req), sizeof(resp));
792 req.filter_handle = apc->pf_filter_handle;
794 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
797 netdev_err(apc->ndev, "Failed to unregister filter: %d\n",
802 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER,
804 if (err || resp.hdr.status)
805 netdev_err(apc->ndev,
806 "Failed to deregister filter: %d, 0x%x\n",
807 err, resp.hdr.status);
810 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver,
811 u32 proto_minor_ver, u32 proto_micro_ver,
814 struct gdma_context *gc = ac->gdma_dev->gdma_context;
815 struct mana_query_device_cfg_resp resp = {};
816 struct mana_query_device_cfg_req req = {};
817 struct device *dev = gc->dev;
820 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG,
821 sizeof(req), sizeof(resp));
823 req.hdr.resp.msg_version = GDMA_MESSAGE_V2;
825 req.proto_major_ver = proto_major_ver;
826 req.proto_minor_ver = proto_minor_ver;
827 req.proto_micro_ver = proto_micro_ver;
829 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp));
831 dev_err(dev, "Failed to query config: %d", err);
835 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG,
837 if (err || resp.hdr.status) {
838 dev_err(dev, "Invalid query result: %d, 0x%x\n", err,
845 *max_num_vports = resp.max_num_vports;
847 if (resp.hdr.response.msg_version == GDMA_MESSAGE_V2)
848 gc->adapter_mtu = resp.adapter_mtu;
850 gc->adapter_mtu = ETH_FRAME_LEN;
855 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index,
856 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry)
858 struct mana_query_vport_cfg_resp resp = {};
859 struct mana_query_vport_cfg_req req = {};
862 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG,
863 sizeof(req), sizeof(resp));
865 req.vport_index = vport_index;
867 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
872 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG,
880 *max_sq = resp.max_num_sq;
881 *max_rq = resp.max_num_rq;
882 *num_indir_entry = resp.num_indirection_ent;
884 apc->port_handle = resp.vport;
885 ether_addr_copy(apc->mac_addr, resp.mac_addr);
890 void mana_uncfg_vport(struct mana_port_context *apc)
892 mutex_lock(&apc->vport_mutex);
893 apc->vport_use_count--;
894 WARN_ON(apc->vport_use_count < 0);
895 mutex_unlock(&apc->vport_mutex);
897 EXPORT_SYMBOL_NS(mana_uncfg_vport, NET_MANA);
899 int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id,
902 struct mana_config_vport_resp resp = {};
903 struct mana_config_vport_req req = {};
906 /* This function is used to program the Ethernet port in the hardware
907 * table. It can be called from the Ethernet driver or the RDMA driver.
909 * For Ethernet usage, the hardware supports only one active user on a
910 * physical port. The driver checks on the port usage before programming
911 * the hardware when creating the RAW QP (RDMA driver) or exposing the
912 * device to kernel NET layer (Ethernet driver).
914 * Because the RDMA driver doesn't know in advance which QP type the
915 * user will create, it exposes the device with all its ports. The user
916 * may not be able to create RAW QP on a port if this port is already
917 * in used by the Ethernet driver from the kernel.
919 * This physical port limitation only applies to the RAW QP. For RC QP,
920 * the hardware doesn't have this limitation. The user can create RC
921 * QPs on a physical port up to the hardware limits independent of the
922 * Ethernet usage on the same port.
924 mutex_lock(&apc->vport_mutex);
925 if (apc->vport_use_count > 0) {
926 mutex_unlock(&apc->vport_mutex);
929 apc->vport_use_count++;
930 mutex_unlock(&apc->vport_mutex);
932 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX,
933 sizeof(req), sizeof(resp));
934 req.vport = apc->port_handle;
935 req.pdid = protection_dom_id;
936 req.doorbell_pageid = doorbell_pg_id;
938 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
941 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err);
945 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX,
947 if (err || resp.hdr.status) {
948 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n",
949 err, resp.hdr.status);
956 apc->tx_shortform_allowed = resp.short_form_allowed;
957 apc->tx_vp_offset = resp.tx_vport_offset;
959 netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n",
960 apc->port_handle, protection_dom_id, doorbell_pg_id);
963 mana_uncfg_vport(apc);
967 EXPORT_SYMBOL_NS(mana_cfg_vport, NET_MANA);
969 static int mana_cfg_vport_steering(struct mana_port_context *apc,
971 bool update_default_rxobj, bool update_key,
974 u16 num_entries = MANA_INDIRECT_TABLE_SIZE;
975 struct mana_cfg_rx_steer_req *req = NULL;
976 struct mana_cfg_rx_steer_resp resp = {};
977 struct net_device *ndev = apc->ndev;
978 mana_handle_t *req_indir_tab;
982 req_buf_size = sizeof(*req) + sizeof(mana_handle_t) * num_entries;
983 req = kzalloc(req_buf_size, GFP_KERNEL);
987 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
990 req->vport = apc->port_handle;
991 req->num_indir_entries = num_entries;
992 req->indir_tab_offset = sizeof(*req);
994 req->rss_enable = apc->rss_state;
995 req->update_default_rxobj = update_default_rxobj;
996 req->update_hashkey = update_key;
997 req->update_indir_tab = update_tab;
998 req->default_rxobj = apc->default_rxobj;
1001 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE);
1004 req_indir_tab = (mana_handle_t *)(req + 1);
1005 memcpy(req_indir_tab, apc->rxobj_table,
1006 req->num_indir_entries * sizeof(mana_handle_t));
1009 err = mana_send_request(apc->ac, req, req_buf_size, &resp,
1012 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err);
1016 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX,
1019 netdev_err(ndev, "vPort RX configuration failed: %d\n", err);
1023 if (resp.hdr.status) {
1024 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n",
1029 netdev_info(ndev, "Configured steering vPort %llu entries %u\n",
1030 apc->port_handle, num_entries);
1036 int mana_create_wq_obj(struct mana_port_context *apc,
1037 mana_handle_t vport,
1038 u32 wq_type, struct mana_obj_spec *wq_spec,
1039 struct mana_obj_spec *cq_spec,
1040 mana_handle_t *wq_obj)
1042 struct mana_create_wqobj_resp resp = {};
1043 struct mana_create_wqobj_req req = {};
1044 struct net_device *ndev = apc->ndev;
1047 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ,
1048 sizeof(req), sizeof(resp));
1050 req.wq_type = wq_type;
1051 req.wq_gdma_region = wq_spec->gdma_region;
1052 req.cq_gdma_region = cq_spec->gdma_region;
1053 req.wq_size = wq_spec->queue_size;
1054 req.cq_size = cq_spec->queue_size;
1055 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id;
1056 req.cq_parent_qid = cq_spec->attached_eq;
1058 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1061 netdev_err(ndev, "Failed to create WQ object: %d\n", err);
1065 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ,
1067 if (err || resp.hdr.status) {
1068 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err,
1075 if (resp.wq_obj == INVALID_MANA_HANDLE) {
1076 netdev_err(ndev, "Got an invalid WQ object handle\n");
1081 *wq_obj = resp.wq_obj;
1082 wq_spec->queue_index = resp.wq_id;
1083 cq_spec->queue_index = resp.cq_id;
1089 EXPORT_SYMBOL_NS(mana_create_wq_obj, NET_MANA);
1091 void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type,
1092 mana_handle_t wq_obj)
1094 struct mana_destroy_wqobj_resp resp = {};
1095 struct mana_destroy_wqobj_req req = {};
1096 struct net_device *ndev = apc->ndev;
1099 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ,
1100 sizeof(req), sizeof(resp));
1101 req.wq_type = wq_type;
1102 req.wq_obj_handle = wq_obj;
1104 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1107 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err);
1111 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ,
1113 if (err || resp.hdr.status)
1114 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err,
1117 EXPORT_SYMBOL_NS(mana_destroy_wq_obj, NET_MANA);
1119 static void mana_destroy_eq(struct mana_context *ac)
1121 struct gdma_context *gc = ac->gdma_dev->gdma_context;
1122 struct gdma_queue *eq;
1128 for (i = 0; i < gc->max_num_queues; i++) {
1133 mana_gd_destroy_queue(gc, eq);
1140 static int mana_create_eq(struct mana_context *ac)
1142 struct gdma_dev *gd = ac->gdma_dev;
1143 struct gdma_context *gc = gd->gdma_context;
1144 struct gdma_queue_spec spec = {};
1148 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq),
1153 spec.type = GDMA_EQ;
1154 spec.monitor_avl_buf = false;
1155 spec.queue_size = EQ_SIZE;
1156 spec.eq.callback = NULL;
1157 spec.eq.context = ac->eqs;
1158 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE;
1160 for (i = 0; i < gc->max_num_queues; i++) {
1161 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq);
1168 mana_destroy_eq(ac);
1172 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq)
1174 struct mana_fence_rq_resp resp = {};
1175 struct mana_fence_rq_req req = {};
1178 init_completion(&rxq->fence_event);
1180 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ,
1181 sizeof(req), sizeof(resp));
1182 req.wq_obj_handle = rxq->rxobj;
1184 err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
1187 netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n",
1192 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp));
1193 if (err || resp.hdr.status) {
1194 netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n",
1195 rxq->rxq_idx, err, resp.hdr.status);
1202 if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) {
1203 netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n",
1211 static void mana_fence_rqs(struct mana_port_context *apc)
1213 unsigned int rxq_idx;
1214 struct mana_rxq *rxq;
1217 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
1218 rxq = apc->rxqs[rxq_idx];
1219 err = mana_fence_rq(apc, rxq);
1221 /* In case of any error, use sleep instead. */
1227 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units)
1232 used_space_old = wq->head - wq->tail;
1233 used_space_new = wq->head - (wq->tail + num_units);
1235 if (WARN_ON_ONCE(used_space_new > used_space_old))
1238 wq->tail += num_units;
1242 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc)
1244 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
1245 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1246 struct device *dev = gc->dev;
1249 dma_unmap_single(dev, ash->dma_handle[0], ash->size[0], DMA_TO_DEVICE);
1251 for (i = 1; i < skb_shinfo(skb)->nr_frags + 1; i++)
1252 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i],
1256 static void mana_poll_tx_cq(struct mana_cq *cq)
1258 struct gdma_comp *completions = cq->gdma_comp_buf;
1259 struct gdma_posted_wqe_info *wqe_info;
1260 unsigned int pkt_transmitted = 0;
1261 unsigned int wqe_unit_cnt = 0;
1262 struct mana_txq *txq = cq->txq;
1263 struct mana_port_context *apc;
1264 struct netdev_queue *net_txq;
1265 struct gdma_queue *gdma_wq;
1266 unsigned int avail_space;
1267 struct net_device *ndev;
1268 struct sk_buff *skb;
1274 apc = netdev_priv(ndev);
1276 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions,
1277 CQE_POLLING_BUFFER);
1282 apc->eth_stats.tx_cqes = comp_read;
1284 for (i = 0; i < comp_read; i++) {
1285 struct mana_tx_comp_oob *cqe_oob;
1287 if (WARN_ON_ONCE(!completions[i].is_sq))
1290 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data;
1291 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type !=
1292 MANA_CQE_COMPLETION))
1295 switch (cqe_oob->cqe_hdr.cqe_type) {
1299 case CQE_TX_SA_DROP:
1300 case CQE_TX_MTU_DROP:
1301 case CQE_TX_INVALID_OOB:
1302 case CQE_TX_INVALID_ETH_TYPE:
1303 case CQE_TX_HDR_PROCESSING_ERROR:
1304 case CQE_TX_VF_DISABLED:
1305 case CQE_TX_VPORT_IDX_OUT_OF_RANGE:
1306 case CQE_TX_VPORT_DISABLED:
1307 case CQE_TX_VLAN_TAGGING_VIOLATION:
1308 WARN_ONCE(1, "TX: CQE error %d: ignored.\n",
1309 cqe_oob->cqe_hdr.cqe_type);
1310 apc->eth_stats.tx_cqe_err++;
1314 /* If the CQE type is unexpected, log an error, assert,
1315 * and go through the error path.
1317 WARN_ONCE(1, "TX: Unexpected CQE type %d: HW BUG?\n",
1318 cqe_oob->cqe_hdr.cqe_type);
1319 apc->eth_stats.tx_cqe_unknown_type++;
1323 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num))
1326 skb = skb_dequeue(&txq->pending_skbs);
1327 if (WARN_ON_ONCE(!skb))
1330 wqe_info = (struct gdma_posted_wqe_info *)skb->cb;
1331 wqe_unit_cnt += wqe_info->wqe_size_in_bu;
1333 mana_unmap_skb(skb, apc);
1335 napi_consume_skb(skb, cq->budget);
1340 if (WARN_ON_ONCE(wqe_unit_cnt == 0))
1343 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt);
1345 gdma_wq = txq->gdma_sq;
1346 avail_space = mana_gd_wq_avail_space(gdma_wq);
1348 /* Ensure tail updated before checking q stop */
1351 net_txq = txq->net_txq;
1352 txq_stopped = netif_tx_queue_stopped(net_txq);
1354 /* Ensure checking txq_stopped before apc->port_is_up. */
1357 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) {
1358 netif_tx_wake_queue(net_txq);
1359 apc->eth_stats.wake_queue++;
1362 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0)
1365 cq->work_done = pkt_transmitted;
1367 apc->eth_stats.tx_cqes -= pkt_transmitted;
1370 static void mana_post_pkt_rxq(struct mana_rxq *rxq)
1372 struct mana_recv_buf_oob *recv_buf_oob;
1376 curr_index = rxq->buf_index++;
1377 if (rxq->buf_index == rxq->num_rx_buf)
1380 recv_buf_oob = &rxq->rx_oobs[curr_index];
1382 err = mana_gd_post_and_ring(rxq->gdma_rq, &recv_buf_oob->wqe_req,
1383 &recv_buf_oob->wqe_inf);
1384 if (WARN_ON_ONCE(err))
1387 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1);
1390 static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va,
1391 uint pkt_len, struct xdp_buff *xdp)
1393 struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size);
1398 if (xdp->data_hard_start) {
1399 skb_reserve(skb, xdp->data - xdp->data_hard_start);
1400 skb_put(skb, xdp->data_end - xdp->data);
1404 skb_reserve(skb, rxq->headroom);
1405 skb_put(skb, pkt_len);
1410 static void mana_rx_skb(void *buf_va, struct mana_rxcomp_oob *cqe,
1411 struct mana_rxq *rxq)
1413 struct mana_stats_rx *rx_stats = &rxq->stats;
1414 struct net_device *ndev = rxq->ndev;
1415 uint pkt_len = cqe->ppi[0].pkt_len;
1416 u16 rxq_idx = rxq->rxq_idx;
1417 struct napi_struct *napi;
1418 struct xdp_buff xdp = {};
1419 struct sk_buff *skb;
1423 rxq->rx_cq.work_done++;
1424 napi = &rxq->rx_cq.napi;
1427 ++ndev->stats.rx_dropped;
1431 act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len);
1433 if (act == XDP_REDIRECT && !rxq->xdp_rc)
1436 if (act != XDP_PASS && act != XDP_TX)
1439 skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp);
1444 skb->dev = napi->dev;
1446 skb->protocol = eth_type_trans(skb, ndev);
1447 skb_checksum_none_assert(skb);
1448 skb_record_rx_queue(skb, rxq_idx);
1450 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) {
1451 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed)
1452 skb->ip_summed = CHECKSUM_UNNECESSARY;
1455 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) {
1456 hash_value = cqe->ppi[0].pkt_hash;
1458 if (cqe->rx_hashtype & MANA_HASH_L4)
1459 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4);
1461 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3);
1464 u64_stats_update_begin(&rx_stats->syncp);
1465 rx_stats->packets++;
1466 rx_stats->bytes += pkt_len;
1470 u64_stats_update_end(&rx_stats->syncp);
1472 if (act == XDP_TX) {
1473 skb_set_queue_mapping(skb, rxq_idx);
1474 mana_xdp_tx(skb, ndev);
1478 napi_gro_receive(napi, skb);
1483 u64_stats_update_begin(&rx_stats->syncp);
1484 rx_stats->xdp_drop++;
1485 u64_stats_update_end(&rx_stats->syncp);
1488 WARN_ON_ONCE(rxq->xdp_save_va);
1489 /* Save for reuse */
1490 rxq->xdp_save_va = buf_va;
1492 ++ndev->stats.rx_dropped;
1497 static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,
1498 dma_addr_t *da, bool is_napi)
1503 /* Reuse XDP dropped page if available */
1504 if (rxq->xdp_save_va) {
1505 va = rxq->xdp_save_va;
1506 rxq->xdp_save_va = NULL;
1507 } else if (rxq->alloc_size > PAGE_SIZE) {
1509 va = napi_alloc_frag(rxq->alloc_size);
1511 va = netdev_alloc_frag(rxq->alloc_size);
1516 page = virt_to_head_page(va);
1517 /* Check if the frag falls back to single page */
1518 if (compound_order(page) < get_order(rxq->alloc_size)) {
1523 page = dev_alloc_page();
1527 va = page_to_virt(page);
1530 *da = dma_map_single(dev, va + rxq->headroom, rxq->datasize,
1532 if (dma_mapping_error(dev, *da)) {
1533 put_page(virt_to_head_page(va));
1540 /* Allocate frag for rx buffer, and save the old buf */
1541 static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq,
1542 struct mana_recv_buf_oob *rxoob, void **old_buf)
1547 va = mana_get_rxfrag(rxq, dev, &da, true);
1551 dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize,
1553 *old_buf = rxoob->buf_va;
1556 rxoob->sgl[0].address = da;
1559 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,
1560 struct gdma_comp *cqe)
1562 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data;
1563 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context;
1564 struct net_device *ndev = rxq->ndev;
1565 struct mana_recv_buf_oob *rxbuf_oob;
1566 struct mana_port_context *apc;
1567 struct device *dev = gc->dev;
1568 void *old_buf = NULL;
1571 apc = netdev_priv(ndev);
1573 switch (oob->cqe_hdr.cqe_type) {
1577 case CQE_RX_TRUNCATED:
1578 ++ndev->stats.rx_dropped;
1579 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index];
1580 netdev_warn_once(ndev, "Dropped a truncated packet\n");
1583 case CQE_RX_COALESCED_4:
1584 netdev_err(ndev, "RX coalescing is unsupported\n");
1585 apc->eth_stats.rx_coalesced_err++;
1588 case CQE_RX_OBJECT_FENCE:
1589 complete(&rxq->fence_event);
1593 netdev_err(ndev, "Unknown RX CQE type = %d\n",
1594 oob->cqe_hdr.cqe_type);
1595 apc->eth_stats.rx_cqe_unknown_type++;
1599 pktlen = oob->ppi[0].pkt_len;
1602 /* data packets should never have packetlength of zero */
1603 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n",
1604 rxq->gdma_id, cq->gdma_id, rxq->rxobj);
1608 curr = rxq->buf_index;
1609 rxbuf_oob = &rxq->rx_oobs[curr];
1610 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1);
1612 mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf);
1614 /* Unsuccessful refill will have old_buf == NULL.
1615 * In this case, mana_rx_skb() will drop the packet.
1617 mana_rx_skb(old_buf, oob, rxq);
1620 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu);
1622 mana_post_pkt_rxq(rxq);
1625 static void mana_poll_rx_cq(struct mana_cq *cq)
1627 struct gdma_comp *comp = cq->gdma_comp_buf;
1628 struct mana_rxq *rxq = cq->rxq;
1629 struct mana_port_context *apc;
1632 apc = netdev_priv(rxq->ndev);
1634 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER);
1635 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER);
1637 apc->eth_stats.rx_cqes = comp_read;
1638 rxq->xdp_flush = false;
1640 for (i = 0; i < comp_read; i++) {
1641 if (WARN_ON_ONCE(comp[i].is_sq))
1644 /* verify recv cqe references the right rxq */
1645 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id))
1648 mana_process_rx_cqe(rxq, cq, &comp[i]);
1650 apc->eth_stats.rx_cqes--;
1657 static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue)
1659 struct mana_cq *cq = context;
1663 WARN_ON_ONCE(cq->gdma_cq != gdma_queue);
1665 if (cq->type == MANA_CQ_TYPE_RX)
1666 mana_poll_rx_cq(cq);
1668 mana_poll_tx_cq(cq);
1672 if (w < cq->budget &&
1673 napi_complete_done(&cq->napi, w)) {
1674 arm_bit = SET_ARM_BIT;
1679 mana_gd_ring_cq(gdma_queue, arm_bit);
1684 static int mana_poll(struct napi_struct *napi, int budget)
1686 struct mana_cq *cq = container_of(napi, struct mana_cq, napi);
1690 cq->budget = budget;
1692 w = mana_cq_handler(cq, cq->gdma_cq);
1694 return min(w, budget);
1697 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue)
1699 struct mana_cq *cq = context;
1701 napi_schedule_irqoff(&cq->napi);
1704 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq)
1706 struct gdma_dev *gd = apc->ac->gdma_dev;
1711 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq);
1714 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq)
1716 struct gdma_dev *gd = apc->ac->gdma_dev;
1721 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq);
1724 static void mana_destroy_txq(struct mana_port_context *apc)
1726 struct napi_struct *napi;
1732 for (i = 0; i < apc->num_queues; i++) {
1733 napi = &apc->tx_qp[i].tx_cq.napi;
1734 napi_synchronize(napi);
1736 netif_napi_del(napi);
1738 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object);
1740 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq);
1742 mana_deinit_txq(apc, &apc->tx_qp[i].txq);
1749 static int mana_create_txq(struct mana_port_context *apc,
1750 struct net_device *net)
1752 struct mana_context *ac = apc->ac;
1753 struct gdma_dev *gd = ac->gdma_dev;
1754 struct mana_obj_spec wq_spec;
1755 struct mana_obj_spec cq_spec;
1756 struct gdma_queue_spec spec;
1757 struct gdma_context *gc;
1758 struct mana_txq *txq;
1765 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp),
1770 /* The minimum size of the WQE is 32 bytes, hence
1771 * MAX_SEND_BUFFERS_PER_QUEUE represents the maximum number of WQEs
1772 * the SQ can store. This value is then used to size other queues
1773 * to prevent overflow.
1775 txq_size = MAX_SEND_BUFFERS_PER_QUEUE * 32;
1776 BUILD_BUG_ON(!PAGE_ALIGNED(txq_size));
1778 cq_size = MAX_SEND_BUFFERS_PER_QUEUE * COMP_ENTRY_SIZE;
1779 cq_size = PAGE_ALIGN(cq_size);
1781 gc = gd->gdma_context;
1783 for (i = 0; i < apc->num_queues; i++) {
1784 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE;
1787 txq = &apc->tx_qp[i].txq;
1789 u64_stats_init(&txq->stats.syncp);
1791 txq->net_txq = netdev_get_tx_queue(net, i);
1792 txq->vp_offset = apc->tx_vp_offset;
1793 skb_queue_head_init(&txq->pending_skbs);
1795 memset(&spec, 0, sizeof(spec));
1796 spec.type = GDMA_SQ;
1797 spec.monitor_avl_buf = true;
1798 spec.queue_size = txq_size;
1799 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq);
1803 /* Create SQ's CQ */
1804 cq = &apc->tx_qp[i].tx_cq;
1805 cq->type = MANA_CQ_TYPE_TX;
1809 memset(&spec, 0, sizeof(spec));
1810 spec.type = GDMA_CQ;
1811 spec.monitor_avl_buf = false;
1812 spec.queue_size = cq_size;
1813 spec.cq.callback = mana_schedule_napi;
1814 spec.cq.parent_eq = ac->eqs[i].eq;
1815 spec.cq.context = cq;
1816 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
1820 memset(&wq_spec, 0, sizeof(wq_spec));
1821 memset(&cq_spec, 0, sizeof(cq_spec));
1823 wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle;
1824 wq_spec.queue_size = txq->gdma_sq->queue_size;
1826 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
1827 cq_spec.queue_size = cq->gdma_cq->queue_size;
1828 cq_spec.modr_ctx_id = 0;
1829 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
1831 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ,
1833 &apc->tx_qp[i].tx_object);
1838 txq->gdma_sq->id = wq_spec.queue_index;
1839 cq->gdma_cq->id = cq_spec.queue_index;
1841 txq->gdma_sq->mem_info.dma_region_handle =
1842 GDMA_INVALID_DMA_REGION;
1843 cq->gdma_cq->mem_info.dma_region_handle =
1844 GDMA_INVALID_DMA_REGION;
1846 txq->gdma_txq_id = txq->gdma_sq->id;
1848 cq->gdma_id = cq->gdma_cq->id;
1850 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
1855 gc->cq_table[cq->gdma_id] = cq->gdma_cq;
1857 netif_napi_add_tx(net, &cq->napi, mana_poll);
1858 napi_enable(&cq->napi);
1860 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
1865 mana_destroy_txq(apc);
1869 static void mana_destroy_rxq(struct mana_port_context *apc,
1870 struct mana_rxq *rxq, bool validate_state)
1873 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1874 struct mana_recv_buf_oob *rx_oob;
1875 struct device *dev = gc->dev;
1876 struct napi_struct *napi;
1882 napi = &rxq->rx_cq.napi;
1885 napi_synchronize(napi);
1889 xdp_rxq_info_unreg(&rxq->xdp_rxq);
1891 netif_napi_del(napi);
1893 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj);
1895 mana_deinit_cq(apc, &rxq->rx_cq);
1897 if (rxq->xdp_save_va)
1898 put_page(virt_to_head_page(rxq->xdp_save_va));
1900 for (i = 0; i < rxq->num_rx_buf; i++) {
1901 rx_oob = &rxq->rx_oobs[i];
1903 if (!rx_oob->buf_va)
1906 dma_unmap_single(dev, rx_oob->sgl[0].address,
1907 rx_oob->sgl[0].size, DMA_FROM_DEVICE);
1909 put_page(virt_to_head_page(rx_oob->buf_va));
1910 rx_oob->buf_va = NULL;
1914 mana_gd_destroy_queue(gc, rxq->gdma_rq);
1919 static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key,
1920 struct mana_rxq *rxq, struct device *dev)
1922 struct mana_port_context *mpc = netdev_priv(rxq->ndev);
1926 if (mpc->rxbufs_pre)
1927 va = mana_get_rxbuf_pre(rxq, &da);
1929 va = mana_get_rxfrag(rxq, dev, &da, false);
1934 rx_oob->buf_va = va;
1936 rx_oob->sgl[0].address = da;
1937 rx_oob->sgl[0].size = rxq->datasize;
1938 rx_oob->sgl[0].mem_key = mem_key;
1943 #define MANA_WQE_HEADER_SIZE 16
1944 #define MANA_WQE_SGE_SIZE 16
1946 static int mana_alloc_rx_wqe(struct mana_port_context *apc,
1947 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size)
1949 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
1950 struct mana_recv_buf_oob *rx_oob;
1951 struct device *dev = gc->dev;
1955 WARN_ON(rxq->datasize == 0);
1960 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
1961 rx_oob = &rxq->rx_oobs[buf_idx];
1962 memset(rx_oob, 0, sizeof(*rx_oob));
1964 rx_oob->num_sge = 1;
1966 ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq,
1971 rx_oob->wqe_req.sgl = rx_oob->sgl;
1972 rx_oob->wqe_req.num_sge = rx_oob->num_sge;
1973 rx_oob->wqe_req.inline_oob_size = 0;
1974 rx_oob->wqe_req.inline_oob_data = NULL;
1975 rx_oob->wqe_req.flags = 0;
1976 rx_oob->wqe_req.client_data_unit = 0;
1978 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE +
1979 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32);
1980 *cq_size += COMP_ENTRY_SIZE;
1986 static int mana_push_wqe(struct mana_rxq *rxq)
1988 struct mana_recv_buf_oob *rx_oob;
1992 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) {
1993 rx_oob = &rxq->rx_oobs[buf_idx];
1995 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req,
2004 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,
2005 u32 rxq_idx, struct mana_eq *eq,
2006 struct net_device *ndev)
2008 struct gdma_dev *gd = apc->ac->gdma_dev;
2009 struct mana_obj_spec wq_spec;
2010 struct mana_obj_spec cq_spec;
2011 struct gdma_queue_spec spec;
2012 struct mana_cq *cq = NULL;
2013 struct gdma_context *gc;
2014 u32 cq_size, rq_size;
2015 struct mana_rxq *rxq;
2018 gc = gd->gdma_context;
2020 rxq = kzalloc(struct_size(rxq, rx_oobs, RX_BUFFERS_PER_QUEUE),
2026 rxq->num_rx_buf = RX_BUFFERS_PER_QUEUE;
2027 rxq->rxq_idx = rxq_idx;
2028 rxq->rxobj = INVALID_MANA_HANDLE;
2030 mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size,
2033 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size);
2037 rq_size = PAGE_ALIGN(rq_size);
2038 cq_size = PAGE_ALIGN(cq_size);
2041 memset(&spec, 0, sizeof(spec));
2042 spec.type = GDMA_RQ;
2043 spec.monitor_avl_buf = true;
2044 spec.queue_size = rq_size;
2045 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq);
2049 /* Create RQ's CQ */
2051 cq->type = MANA_CQ_TYPE_RX;
2054 memset(&spec, 0, sizeof(spec));
2055 spec.type = GDMA_CQ;
2056 spec.monitor_avl_buf = false;
2057 spec.queue_size = cq_size;
2058 spec.cq.callback = mana_schedule_napi;
2059 spec.cq.parent_eq = eq->eq;
2060 spec.cq.context = cq;
2061 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq);
2065 memset(&wq_spec, 0, sizeof(wq_spec));
2066 memset(&cq_spec, 0, sizeof(cq_spec));
2067 wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle;
2068 wq_spec.queue_size = rxq->gdma_rq->queue_size;
2070 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle;
2071 cq_spec.queue_size = cq->gdma_cq->queue_size;
2072 cq_spec.modr_ctx_id = 0;
2073 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id;
2075 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ,
2076 &wq_spec, &cq_spec, &rxq->rxobj);
2080 rxq->gdma_rq->id = wq_spec.queue_index;
2081 cq->gdma_cq->id = cq_spec.queue_index;
2083 rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2084 cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
2086 rxq->gdma_id = rxq->gdma_rq->id;
2087 cq->gdma_id = cq->gdma_cq->id;
2089 err = mana_push_wqe(rxq);
2093 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) {
2098 gc->cq_table[cq->gdma_id] = cq->gdma_cq;
2100 netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1);
2102 WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx,
2104 WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq,
2105 MEM_TYPE_PAGE_SHARED, NULL));
2107 napi_enable(&cq->napi);
2109 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT);
2114 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err);
2116 mana_destroy_rxq(apc, rxq, false);
2119 mana_deinit_cq(apc, cq);
2124 static int mana_add_rx_queues(struct mana_port_context *apc,
2125 struct net_device *ndev)
2127 struct mana_context *ac = apc->ac;
2128 struct mana_rxq *rxq;
2132 for (i = 0; i < apc->num_queues; i++) {
2133 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev);
2139 u64_stats_init(&rxq->stats.syncp);
2144 apc->default_rxobj = apc->rxqs[0]->rxobj;
2149 static void mana_destroy_vport(struct mana_port_context *apc)
2151 struct gdma_dev *gd = apc->ac->gdma_dev;
2152 struct mana_rxq *rxq;
2155 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
2156 rxq = apc->rxqs[rxq_idx];
2160 mana_destroy_rxq(apc, rxq, true);
2161 apc->rxqs[rxq_idx] = NULL;
2164 mana_destroy_txq(apc);
2165 mana_uncfg_vport(apc);
2167 if (gd->gdma_context->is_pf)
2168 mana_pf_deregister_hw_vport(apc);
2171 static int mana_create_vport(struct mana_port_context *apc,
2172 struct net_device *net)
2174 struct gdma_dev *gd = apc->ac->gdma_dev;
2177 apc->default_rxobj = INVALID_MANA_HANDLE;
2179 if (gd->gdma_context->is_pf) {
2180 err = mana_pf_register_hw_vport(apc);
2185 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell);
2189 return mana_create_txq(apc, net);
2192 static void mana_rss_table_init(struct mana_port_context *apc)
2196 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++)
2197 apc->indir_table[i] =
2198 ethtool_rxfh_indir_default(i, apc->num_queues);
2201 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx,
2202 bool update_hash, bool update_tab)
2209 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) {
2210 queue_idx = apc->indir_table[i];
2211 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj;
2215 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab);
2219 mana_fence_rqs(apc);
2224 static int mana_init_port(struct net_device *ndev)
2226 struct mana_port_context *apc = netdev_priv(ndev);
2227 u32 max_txq, max_rxq, max_queues;
2228 int port_idx = apc->port_idx;
2229 u32 num_indirect_entries;
2232 err = mana_init_port_context(apc);
2236 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq,
2237 &num_indirect_entries);
2239 netdev_err(ndev, "Failed to query info for vPort %d\n",
2244 max_queues = min_t(u32, max_txq, max_rxq);
2245 if (apc->max_queues > max_queues)
2246 apc->max_queues = max_queues;
2248 if (apc->num_queues > apc->max_queues)
2249 apc->num_queues = apc->max_queues;
2251 eth_hw_addr_set(ndev, apc->mac_addr);
2261 int mana_alloc_queues(struct net_device *ndev)
2263 struct mana_port_context *apc = netdev_priv(ndev);
2264 struct gdma_dev *gd = apc->ac->gdma_dev;
2267 err = mana_create_vport(apc, ndev);
2271 err = netif_set_real_num_tx_queues(ndev, apc->num_queues);
2275 err = mana_add_rx_queues(apc, ndev);
2279 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE;
2281 err = netif_set_real_num_rx_queues(ndev, apc->num_queues);
2285 mana_rss_table_init(apc);
2287 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true);
2291 if (gd->gdma_context->is_pf) {
2292 err = mana_pf_register_filter(apc);
2297 mana_chn_setxdp(apc, mana_xdp_get(apc));
2302 mana_destroy_vport(apc);
2306 int mana_attach(struct net_device *ndev)
2308 struct mana_port_context *apc = netdev_priv(ndev);
2313 err = mana_init_port(ndev);
2317 if (apc->port_st_save) {
2318 err = mana_alloc_queues(ndev);
2320 mana_cleanup_port_context(apc);
2325 apc->port_is_up = apc->port_st_save;
2327 /* Ensure port state updated before txq state */
2330 if (apc->port_is_up)
2331 netif_carrier_on(ndev);
2333 netif_device_attach(ndev);
2338 static int mana_dealloc_queues(struct net_device *ndev)
2340 struct mana_port_context *apc = netdev_priv(ndev);
2341 struct gdma_dev *gd = apc->ac->gdma_dev;
2342 struct mana_txq *txq;
2345 if (apc->port_is_up)
2348 mana_chn_setxdp(apc, NULL);
2350 if (gd->gdma_context->is_pf)
2351 mana_pf_deregister_filter(apc);
2353 /* No packet can be transmitted now since apc->port_is_up is false.
2354 * There is still a tiny chance that mana_poll_tx_cq() can re-enable
2355 * a txq because it may not timely see apc->port_is_up being cleared
2356 * to false, but it doesn't matter since mana_start_xmit() drops any
2357 * new packets due to apc->port_is_up being false.
2359 * Drain all the in-flight TX packets
2361 for (i = 0; i < apc->num_queues; i++) {
2362 txq = &apc->tx_qp[i].txq;
2364 while (atomic_read(&txq->pending_sends) > 0)
2365 usleep_range(1000, 2000);
2368 /* We're 100% sure the queues can no longer be woken up, because
2369 * we're sure now mana_poll_tx_cq() can't be running.
2372 apc->rss_state = TRI_STATE_FALSE;
2373 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false);
2375 netdev_err(ndev, "Failed to disable vPort: %d\n", err);
2379 mana_destroy_vport(apc);
2384 int mana_detach(struct net_device *ndev, bool from_close)
2386 struct mana_port_context *apc = netdev_priv(ndev);
2391 apc->port_st_save = apc->port_is_up;
2392 apc->port_is_up = false;
2394 /* Ensure port state updated before txq state */
2397 netif_tx_disable(ndev);
2398 netif_carrier_off(ndev);
2400 if (apc->port_st_save) {
2401 err = mana_dealloc_queues(ndev);
2407 netif_device_detach(ndev);
2408 mana_cleanup_port_context(apc);
2414 static int mana_probe_port(struct mana_context *ac, int port_idx,
2415 struct net_device **ndev_storage)
2417 struct gdma_context *gc = ac->gdma_dev->gdma_context;
2418 struct mana_port_context *apc;
2419 struct net_device *ndev;
2422 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context),
2423 gc->max_num_queues);
2427 *ndev_storage = ndev;
2429 apc = netdev_priv(ndev);
2432 apc->max_queues = gc->max_num_queues;
2433 apc->num_queues = gc->max_num_queues;
2434 apc->port_handle = INVALID_MANA_HANDLE;
2435 apc->pf_filter_handle = INVALID_MANA_HANDLE;
2436 apc->port_idx = port_idx;
2438 mutex_init(&apc->vport_mutex);
2439 apc->vport_use_count = 0;
2441 ndev->netdev_ops = &mana_devops;
2442 ndev->ethtool_ops = &mana_ethtool_ops;
2443 ndev->mtu = ETH_DATA_LEN;
2444 ndev->max_mtu = gc->adapter_mtu - ETH_HLEN;
2445 ndev->min_mtu = ETH_MIN_MTU;
2446 ndev->needed_headroom = MANA_HEADROOM;
2447 ndev->dev_port = port_idx;
2448 SET_NETDEV_DEV(ndev, gc->dev);
2450 netif_carrier_off(ndev);
2452 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE);
2454 err = mana_init_port(ndev);
2458 netdev_lockdep_set_classes(ndev);
2460 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
2461 ndev->hw_features |= NETIF_F_RXCSUM;
2462 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
2463 ndev->hw_features |= NETIF_F_RXHASH;
2464 ndev->features = ndev->hw_features;
2465 ndev->vlan_features = 0;
2466 ndev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT |
2467 NETDEV_XDP_ACT_NDO_XMIT;
2469 err = register_netdev(ndev);
2471 netdev_err(ndev, "Unable to register netdev.\n");
2481 *ndev_storage = NULL;
2482 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err);
2487 static void adev_release(struct device *dev)
2489 struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev);
2494 static void remove_adev(struct gdma_dev *gd)
2496 struct auxiliary_device *adev = gd->adev;
2499 auxiliary_device_delete(adev);
2500 auxiliary_device_uninit(adev);
2502 mana_adev_idx_free(id);
2506 static int add_adev(struct gdma_dev *gd)
2508 struct auxiliary_device *adev;
2509 struct mana_adev *madev;
2512 madev = kzalloc(sizeof(*madev), GFP_KERNEL);
2516 adev = &madev->adev;
2517 ret = mana_adev_idx_alloc();
2522 adev->name = "rdma";
2523 adev->dev.parent = gd->gdma_context->dev;
2524 adev->dev.release = adev_release;
2527 ret = auxiliary_device_init(adev);
2531 ret = auxiliary_device_add(adev);
2539 auxiliary_device_uninit(adev);
2542 mana_adev_idx_free(adev->id);
2550 int mana_probe(struct gdma_dev *gd, bool resuming)
2552 struct gdma_context *gc = gd->gdma_context;
2553 struct mana_context *ac = gd->driver_data;
2554 struct device *dev = gc->dev;
2560 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n",
2561 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION);
2563 err = mana_gd_register_device(gd);
2568 ac = kzalloc(sizeof(*ac), GFP_KERNEL);
2573 gd->driver_data = ac;
2576 err = mana_create_eq(ac);
2580 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION,
2581 MANA_MICRO_VERSION, &num_ports);
2586 ac->num_ports = num_ports;
2588 if (ac->num_ports != num_ports) {
2589 dev_err(dev, "The number of vPorts changed: %d->%d\n",
2590 ac->num_ports, num_ports);
2596 if (ac->num_ports == 0)
2597 dev_err(dev, "Failed to detect any vPort\n");
2599 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV)
2600 ac->num_ports = MAX_PORTS_IN_MANA_DEV;
2603 for (i = 0; i < ac->num_ports; i++) {
2604 err = mana_probe_port(ac, i, &ac->ports[i]);
2609 for (i = 0; i < ac->num_ports; i++) {
2611 err = mana_attach(ac->ports[i]);
2621 mana_remove(gd, false);
2626 void mana_remove(struct gdma_dev *gd, bool suspending)
2628 struct gdma_context *gc = gd->gdma_context;
2629 struct mana_context *ac = gd->driver_data;
2630 struct device *dev = gc->dev;
2631 struct net_device *ndev;
2635 /* adev currently doesn't support suspending, always remove it */
2639 for (i = 0; i < ac->num_ports; i++) {
2640 ndev = ac->ports[i];
2643 dev_err(dev, "No net device to remove\n");
2647 /* All cleanup actions should stay after rtnl_lock(), otherwise
2648 * other functions may access partially cleaned up data.
2652 err = mana_detach(ndev, false);
2654 netdev_err(ndev, "Failed to detach vPort %d: %d\n",
2658 /* No need to unregister the ndev. */
2663 unregister_netdevice(ndev);
2670 mana_destroy_eq(ac);
2672 mana_gd_deregister_device(gd);
2677 gd->driver_data = NULL;
2678 gd->gdma_context = NULL;