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/mm.h> 10 11 #include <net/checksum.h> 12 #include <net/ip6_checksum.h> 13 14 #include "mana.h" 15 16 /* Microsoft Azure Network Adapter (MANA) functions */ 17 18 static int mana_open(struct net_device *ndev) 19 { 20 struct mana_port_context *apc = netdev_priv(ndev); 21 int err; 22 23 err = mana_alloc_queues(ndev); 24 if (err) 25 return err; 26 27 apc->port_is_up = true; 28 29 /* Ensure port state updated before txq state */ 30 smp_wmb(); 31 32 netif_carrier_on(ndev); 33 netif_tx_wake_all_queues(ndev); 34 35 return 0; 36 } 37 38 static int mana_close(struct net_device *ndev) 39 { 40 struct mana_port_context *apc = netdev_priv(ndev); 41 42 if (!apc->port_is_up) 43 return 0; 44 45 return mana_detach(ndev, true); 46 } 47 48 static bool mana_can_tx(struct gdma_queue *wq) 49 { 50 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE; 51 } 52 53 static unsigned int mana_checksum_info(struct sk_buff *skb) 54 { 55 if (skb->protocol == htons(ETH_P_IP)) { 56 struct iphdr *ip = ip_hdr(skb); 57 58 if (ip->protocol == IPPROTO_TCP) 59 return IPPROTO_TCP; 60 61 if (ip->protocol == IPPROTO_UDP) 62 return IPPROTO_UDP; 63 } else if (skb->protocol == htons(ETH_P_IPV6)) { 64 struct ipv6hdr *ip6 = ipv6_hdr(skb); 65 66 if (ip6->nexthdr == IPPROTO_TCP) 67 return IPPROTO_TCP; 68 69 if (ip6->nexthdr == IPPROTO_UDP) 70 return IPPROTO_UDP; 71 } 72 73 /* No csum offloading */ 74 return 0; 75 } 76 77 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc, 78 struct mana_tx_package *tp) 79 { 80 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 81 struct gdma_dev *gd = apc->ac->gdma_dev; 82 struct gdma_context *gc; 83 struct device *dev; 84 skb_frag_t *frag; 85 dma_addr_t da; 86 int i; 87 88 gc = gd->gdma_context; 89 dev = gc->dev; 90 da = dma_map_single(dev, skb->data, skb_headlen(skb), DMA_TO_DEVICE); 91 92 if (dma_mapping_error(dev, da)) 93 return -ENOMEM; 94 95 ash->dma_handle[0] = da; 96 ash->size[0] = skb_headlen(skb); 97 98 tp->wqe_req.sgl[0].address = ash->dma_handle[0]; 99 tp->wqe_req.sgl[0].mem_key = gd->gpa_mkey; 100 tp->wqe_req.sgl[0].size = ash->size[0]; 101 102 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 103 frag = &skb_shinfo(skb)->frags[i]; 104 da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag), 105 DMA_TO_DEVICE); 106 107 if (dma_mapping_error(dev, da)) 108 goto frag_err; 109 110 ash->dma_handle[i + 1] = da; 111 ash->size[i + 1] = skb_frag_size(frag); 112 113 tp->wqe_req.sgl[i + 1].address = ash->dma_handle[i + 1]; 114 tp->wqe_req.sgl[i + 1].mem_key = gd->gpa_mkey; 115 tp->wqe_req.sgl[i + 1].size = ash->size[i + 1]; 116 } 117 118 return 0; 119 120 frag_err: 121 for (i = i - 1; i >= 0; i--) 122 dma_unmap_page(dev, ash->dma_handle[i + 1], ash->size[i + 1], 123 DMA_TO_DEVICE); 124 125 dma_unmap_single(dev, ash->dma_handle[0], ash->size[0], DMA_TO_DEVICE); 126 127 return -ENOMEM; 128 } 129 130 int mana_start_xmit(struct sk_buff *skb, struct net_device *ndev) 131 { 132 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT; 133 struct mana_port_context *apc = netdev_priv(ndev); 134 u16 txq_idx = skb_get_queue_mapping(skb); 135 struct gdma_dev *gd = apc->ac->gdma_dev; 136 bool ipv4 = false, ipv6 = false; 137 struct mana_tx_package pkg = {}; 138 struct netdev_queue *net_txq; 139 struct mana_stats_tx *tx_stats; 140 struct gdma_queue *gdma_sq; 141 unsigned int csum_type; 142 struct mana_txq *txq; 143 struct mana_cq *cq; 144 int err, len; 145 146 if (unlikely(!apc->port_is_up)) 147 goto tx_drop; 148 149 if (skb_cow_head(skb, MANA_HEADROOM)) 150 goto tx_drop_count; 151 152 txq = &apc->tx_qp[txq_idx].txq; 153 gdma_sq = txq->gdma_sq; 154 cq = &apc->tx_qp[txq_idx].tx_cq; 155 156 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id; 157 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame; 158 159 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) { 160 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset; 161 pkt_fmt = MANA_LONG_PKT_FMT; 162 } else { 163 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset; 164 } 165 166 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt; 167 168 if (pkt_fmt == MANA_SHORT_PKT_FMT) 169 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob); 170 else 171 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob); 172 173 pkg.wqe_req.inline_oob_data = &pkg.tx_oob; 174 pkg.wqe_req.flags = 0; 175 pkg.wqe_req.client_data_unit = 0; 176 177 pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags; 178 WARN_ON_ONCE(pkg.wqe_req.num_sge > 30); 179 180 if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) { 181 pkg.wqe_req.sgl = pkg.sgl_array; 182 } else { 183 pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge, 184 sizeof(struct gdma_sge), 185 GFP_ATOMIC); 186 if (!pkg.sgl_ptr) 187 goto tx_drop_count; 188 189 pkg.wqe_req.sgl = pkg.sgl_ptr; 190 } 191 192 if (skb->protocol == htons(ETH_P_IP)) 193 ipv4 = true; 194 else if (skb->protocol == htons(ETH_P_IPV6)) 195 ipv6 = true; 196 197 if (skb_is_gso(skb)) { 198 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 199 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 200 201 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 202 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 203 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 204 205 pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size; 206 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0; 207 if (ipv4) { 208 ip_hdr(skb)->tot_len = 0; 209 ip_hdr(skb)->check = 0; 210 tcp_hdr(skb)->check = 211 ~csum_tcpudp_magic(ip_hdr(skb)->saddr, 212 ip_hdr(skb)->daddr, 0, 213 IPPROTO_TCP, 0); 214 } else { 215 ipv6_hdr(skb)->payload_len = 0; 216 tcp_hdr(skb)->check = 217 ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr, 218 &ipv6_hdr(skb)->daddr, 0, 219 IPPROTO_TCP, 0); 220 } 221 } else if (skb->ip_summed == CHECKSUM_PARTIAL) { 222 csum_type = mana_checksum_info(skb); 223 224 if (csum_type == IPPROTO_TCP) { 225 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 226 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 227 228 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 229 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 230 231 } else if (csum_type == IPPROTO_UDP) { 232 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 233 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 234 235 pkg.tx_oob.s_oob.comp_udp_csum = 1; 236 } else { 237 /* Can't do offload of this type of checksum */ 238 if (skb_checksum_help(skb)) 239 goto free_sgl_ptr; 240 } 241 } 242 243 if (mana_map_skb(skb, apc, &pkg)) 244 goto free_sgl_ptr; 245 246 skb_queue_tail(&txq->pending_skbs, skb); 247 248 len = skb->len; 249 net_txq = netdev_get_tx_queue(ndev, txq_idx); 250 251 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req, 252 (struct gdma_posted_wqe_info *)skb->cb); 253 if (!mana_can_tx(gdma_sq)) { 254 netif_tx_stop_queue(net_txq); 255 apc->eth_stats.stop_queue++; 256 } 257 258 if (err) { 259 (void)skb_dequeue_tail(&txq->pending_skbs); 260 netdev_warn(ndev, "Failed to post TX OOB: %d\n", err); 261 err = NETDEV_TX_BUSY; 262 goto tx_busy; 263 } 264 265 err = NETDEV_TX_OK; 266 atomic_inc(&txq->pending_sends); 267 268 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq); 269 270 /* skb may be freed after mana_gd_post_work_request. Do not use it. */ 271 skb = NULL; 272 273 tx_stats = &txq->stats; 274 u64_stats_update_begin(&tx_stats->syncp); 275 tx_stats->packets++; 276 tx_stats->bytes += len; 277 u64_stats_update_end(&tx_stats->syncp); 278 279 tx_busy: 280 if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) { 281 netif_tx_wake_queue(net_txq); 282 apc->eth_stats.wake_queue++; 283 } 284 285 kfree(pkg.sgl_ptr); 286 return err; 287 288 free_sgl_ptr: 289 kfree(pkg.sgl_ptr); 290 tx_drop_count: 291 ndev->stats.tx_dropped++; 292 tx_drop: 293 dev_kfree_skb_any(skb); 294 return NETDEV_TX_OK; 295 } 296 297 static void mana_get_stats64(struct net_device *ndev, 298 struct rtnl_link_stats64 *st) 299 { 300 struct mana_port_context *apc = netdev_priv(ndev); 301 unsigned int num_queues = apc->num_queues; 302 struct mana_stats_rx *rx_stats; 303 struct mana_stats_tx *tx_stats; 304 unsigned int start; 305 u64 packets, bytes; 306 int q; 307 308 if (!apc->port_is_up) 309 return; 310 311 netdev_stats_to_stats64(st, &ndev->stats); 312 313 for (q = 0; q < num_queues; q++) { 314 rx_stats = &apc->rxqs[q]->stats; 315 316 do { 317 start = u64_stats_fetch_begin_irq(&rx_stats->syncp); 318 packets = rx_stats->packets; 319 bytes = rx_stats->bytes; 320 } while (u64_stats_fetch_retry_irq(&rx_stats->syncp, start)); 321 322 st->rx_packets += packets; 323 st->rx_bytes += bytes; 324 } 325 326 for (q = 0; q < num_queues; q++) { 327 tx_stats = &apc->tx_qp[q].txq.stats; 328 329 do { 330 start = u64_stats_fetch_begin_irq(&tx_stats->syncp); 331 packets = tx_stats->packets; 332 bytes = tx_stats->bytes; 333 } while (u64_stats_fetch_retry_irq(&tx_stats->syncp, start)); 334 335 st->tx_packets += packets; 336 st->tx_bytes += bytes; 337 } 338 } 339 340 static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb, 341 int old_q) 342 { 343 struct mana_port_context *apc = netdev_priv(ndev); 344 u32 hash = skb_get_hash(skb); 345 struct sock *sk = skb->sk; 346 int txq; 347 348 txq = apc->indir_table[hash & MANA_INDIRECT_TABLE_MASK]; 349 350 if (txq != old_q && sk && sk_fullsock(sk) && 351 rcu_access_pointer(sk->sk_dst_cache)) 352 sk_tx_queue_set(sk, txq); 353 354 return txq; 355 } 356 357 static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb, 358 struct net_device *sb_dev) 359 { 360 int txq; 361 362 if (ndev->real_num_tx_queues == 1) 363 return 0; 364 365 txq = sk_tx_queue_get(skb->sk); 366 367 if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) { 368 if (skb_rx_queue_recorded(skb)) 369 txq = skb_get_rx_queue(skb); 370 else 371 txq = mana_get_tx_queue(ndev, skb, txq); 372 } 373 374 return txq; 375 } 376 377 static const struct net_device_ops mana_devops = { 378 .ndo_open = mana_open, 379 .ndo_stop = mana_close, 380 .ndo_select_queue = mana_select_queue, 381 .ndo_start_xmit = mana_start_xmit, 382 .ndo_validate_addr = eth_validate_addr, 383 .ndo_get_stats64 = mana_get_stats64, 384 .ndo_bpf = mana_bpf, 385 }; 386 387 static void mana_cleanup_port_context(struct mana_port_context *apc) 388 { 389 kfree(apc->rxqs); 390 apc->rxqs = NULL; 391 } 392 393 static int mana_init_port_context(struct mana_port_context *apc) 394 { 395 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *), 396 GFP_KERNEL); 397 398 return !apc->rxqs ? -ENOMEM : 0; 399 } 400 401 static int mana_send_request(struct mana_context *ac, void *in_buf, 402 u32 in_len, void *out_buf, u32 out_len) 403 { 404 struct gdma_context *gc = ac->gdma_dev->gdma_context; 405 struct gdma_resp_hdr *resp = out_buf; 406 struct gdma_req_hdr *req = in_buf; 407 struct device *dev = gc->dev; 408 static atomic_t activity_id; 409 int err; 410 411 req->dev_id = gc->mana.dev_id; 412 req->activity_id = atomic_inc_return(&activity_id); 413 414 err = mana_gd_send_request(gc, in_len, in_buf, out_len, 415 out_buf); 416 if (err || resp->status) { 417 dev_err(dev, "Failed to send mana message: %d, 0x%x\n", 418 err, resp->status); 419 return err ? err : -EPROTO; 420 } 421 422 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 || 423 req->activity_id != resp->activity_id) { 424 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n", 425 req->dev_id.as_uint32, resp->dev_id.as_uint32, 426 req->activity_id, resp->activity_id); 427 return -EPROTO; 428 } 429 430 return 0; 431 } 432 433 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr, 434 const enum mana_command_code expected_code, 435 const u32 min_size) 436 { 437 if (resp_hdr->response.msg_type != expected_code) 438 return -EPROTO; 439 440 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1) 441 return -EPROTO; 442 443 if (resp_hdr->response.msg_size < min_size) 444 return -EPROTO; 445 446 return 0; 447 } 448 449 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver, 450 u32 proto_minor_ver, u32 proto_micro_ver, 451 u16 *max_num_vports) 452 { 453 struct gdma_context *gc = ac->gdma_dev->gdma_context; 454 struct mana_query_device_cfg_resp resp = {}; 455 struct mana_query_device_cfg_req req = {}; 456 struct device *dev = gc->dev; 457 int err = 0; 458 459 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG, 460 sizeof(req), sizeof(resp)); 461 req.proto_major_ver = proto_major_ver; 462 req.proto_minor_ver = proto_minor_ver; 463 req.proto_micro_ver = proto_micro_ver; 464 465 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp)); 466 if (err) { 467 dev_err(dev, "Failed to query config: %d", err); 468 return err; 469 } 470 471 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG, 472 sizeof(resp)); 473 if (err || resp.hdr.status) { 474 dev_err(dev, "Invalid query result: %d, 0x%x\n", err, 475 resp.hdr.status); 476 if (!err) 477 err = -EPROTO; 478 return err; 479 } 480 481 *max_num_vports = resp.max_num_vports; 482 483 return 0; 484 } 485 486 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index, 487 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry) 488 { 489 struct mana_query_vport_cfg_resp resp = {}; 490 struct mana_query_vport_cfg_req req = {}; 491 int err; 492 493 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG, 494 sizeof(req), sizeof(resp)); 495 496 req.vport_index = vport_index; 497 498 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 499 sizeof(resp)); 500 if (err) 501 return err; 502 503 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG, 504 sizeof(resp)); 505 if (err) 506 return err; 507 508 if (resp.hdr.status) 509 return -EPROTO; 510 511 *max_sq = resp.max_num_sq; 512 *max_rq = resp.max_num_rq; 513 *num_indir_entry = resp.num_indirection_ent; 514 515 apc->port_handle = resp.vport; 516 ether_addr_copy(apc->mac_addr, resp.mac_addr); 517 518 return 0; 519 } 520 521 static int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id, 522 u32 doorbell_pg_id) 523 { 524 struct mana_config_vport_resp resp = {}; 525 struct mana_config_vport_req req = {}; 526 int err; 527 528 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX, 529 sizeof(req), sizeof(resp)); 530 req.vport = apc->port_handle; 531 req.pdid = protection_dom_id; 532 req.doorbell_pageid = doorbell_pg_id; 533 534 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 535 sizeof(resp)); 536 if (err) { 537 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err); 538 goto out; 539 } 540 541 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX, 542 sizeof(resp)); 543 if (err || resp.hdr.status) { 544 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n", 545 err, resp.hdr.status); 546 if (!err) 547 err = -EPROTO; 548 549 goto out; 550 } 551 552 apc->tx_shortform_allowed = resp.short_form_allowed; 553 apc->tx_vp_offset = resp.tx_vport_offset; 554 out: 555 return err; 556 } 557 558 static int mana_cfg_vport_steering(struct mana_port_context *apc, 559 enum TRI_STATE rx, 560 bool update_default_rxobj, bool update_key, 561 bool update_tab) 562 { 563 u16 num_entries = MANA_INDIRECT_TABLE_SIZE; 564 struct mana_cfg_rx_steer_req *req = NULL; 565 struct mana_cfg_rx_steer_resp resp = {}; 566 struct net_device *ndev = apc->ndev; 567 mana_handle_t *req_indir_tab; 568 u32 req_buf_size; 569 int err; 570 571 req_buf_size = sizeof(*req) + sizeof(mana_handle_t) * num_entries; 572 req = kzalloc(req_buf_size, GFP_KERNEL); 573 if (!req) 574 return -ENOMEM; 575 576 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size, 577 sizeof(resp)); 578 579 req->vport = apc->port_handle; 580 req->num_indir_entries = num_entries; 581 req->indir_tab_offset = sizeof(*req); 582 req->rx_enable = rx; 583 req->rss_enable = apc->rss_state; 584 req->update_default_rxobj = update_default_rxobj; 585 req->update_hashkey = update_key; 586 req->update_indir_tab = update_tab; 587 req->default_rxobj = apc->default_rxobj; 588 589 if (update_key) 590 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE); 591 592 if (update_tab) { 593 req_indir_tab = (mana_handle_t *)(req + 1); 594 memcpy(req_indir_tab, apc->rxobj_table, 595 req->num_indir_entries * sizeof(mana_handle_t)); 596 } 597 598 err = mana_send_request(apc->ac, req, req_buf_size, &resp, 599 sizeof(resp)); 600 if (err) { 601 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err); 602 goto out; 603 } 604 605 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX, 606 sizeof(resp)); 607 if (err) { 608 netdev_err(ndev, "vPort RX configuration failed: %d\n", err); 609 goto out; 610 } 611 612 if (resp.hdr.status) { 613 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n", 614 resp.hdr.status); 615 err = -EPROTO; 616 } 617 out: 618 kfree(req); 619 return err; 620 } 621 622 static int mana_create_wq_obj(struct mana_port_context *apc, 623 mana_handle_t vport, 624 u32 wq_type, struct mana_obj_spec *wq_spec, 625 struct mana_obj_spec *cq_spec, 626 mana_handle_t *wq_obj) 627 { 628 struct mana_create_wqobj_resp resp = {}; 629 struct mana_create_wqobj_req req = {}; 630 struct net_device *ndev = apc->ndev; 631 int err; 632 633 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ, 634 sizeof(req), sizeof(resp)); 635 req.vport = vport; 636 req.wq_type = wq_type; 637 req.wq_gdma_region = wq_spec->gdma_region; 638 req.cq_gdma_region = cq_spec->gdma_region; 639 req.wq_size = wq_spec->queue_size; 640 req.cq_size = cq_spec->queue_size; 641 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id; 642 req.cq_parent_qid = cq_spec->attached_eq; 643 644 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 645 sizeof(resp)); 646 if (err) { 647 netdev_err(ndev, "Failed to create WQ object: %d\n", err); 648 goto out; 649 } 650 651 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ, 652 sizeof(resp)); 653 if (err || resp.hdr.status) { 654 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err, 655 resp.hdr.status); 656 if (!err) 657 err = -EPROTO; 658 goto out; 659 } 660 661 if (resp.wq_obj == INVALID_MANA_HANDLE) { 662 netdev_err(ndev, "Got an invalid WQ object handle\n"); 663 err = -EPROTO; 664 goto out; 665 } 666 667 *wq_obj = resp.wq_obj; 668 wq_spec->queue_index = resp.wq_id; 669 cq_spec->queue_index = resp.cq_id; 670 671 return 0; 672 out: 673 return err; 674 } 675 676 static void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type, 677 mana_handle_t wq_obj) 678 { 679 struct mana_destroy_wqobj_resp resp = {}; 680 struct mana_destroy_wqobj_req req = {}; 681 struct net_device *ndev = apc->ndev; 682 int err; 683 684 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ, 685 sizeof(req), sizeof(resp)); 686 req.wq_type = wq_type; 687 req.wq_obj_handle = wq_obj; 688 689 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 690 sizeof(resp)); 691 if (err) { 692 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err); 693 return; 694 } 695 696 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ, 697 sizeof(resp)); 698 if (err || resp.hdr.status) 699 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err, 700 resp.hdr.status); 701 } 702 703 static void mana_destroy_eq(struct mana_context *ac) 704 { 705 struct gdma_context *gc = ac->gdma_dev->gdma_context; 706 struct gdma_queue *eq; 707 int i; 708 709 if (!ac->eqs) 710 return; 711 712 for (i = 0; i < gc->max_num_queues; i++) { 713 eq = ac->eqs[i].eq; 714 if (!eq) 715 continue; 716 717 mana_gd_destroy_queue(gc, eq); 718 } 719 720 kfree(ac->eqs); 721 ac->eqs = NULL; 722 } 723 724 static int mana_create_eq(struct mana_context *ac) 725 { 726 struct gdma_dev *gd = ac->gdma_dev; 727 struct gdma_context *gc = gd->gdma_context; 728 struct gdma_queue_spec spec = {}; 729 int err; 730 int i; 731 732 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq), 733 GFP_KERNEL); 734 if (!ac->eqs) 735 return -ENOMEM; 736 737 spec.type = GDMA_EQ; 738 spec.monitor_avl_buf = false; 739 spec.queue_size = EQ_SIZE; 740 spec.eq.callback = NULL; 741 spec.eq.context = ac->eqs; 742 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE; 743 744 for (i = 0; i < gc->max_num_queues; i++) { 745 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq); 746 if (err) 747 goto out; 748 } 749 750 return 0; 751 out: 752 mana_destroy_eq(ac); 753 return err; 754 } 755 756 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq) 757 { 758 struct mana_fence_rq_resp resp = {}; 759 struct mana_fence_rq_req req = {}; 760 int err; 761 762 init_completion(&rxq->fence_event); 763 764 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ, 765 sizeof(req), sizeof(resp)); 766 req.wq_obj_handle = rxq->rxobj; 767 768 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 769 sizeof(resp)); 770 if (err) { 771 netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n", 772 rxq->rxq_idx, err); 773 return err; 774 } 775 776 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp)); 777 if (err || resp.hdr.status) { 778 netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n", 779 rxq->rxq_idx, err, resp.hdr.status); 780 if (!err) 781 err = -EPROTO; 782 783 return err; 784 } 785 786 if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) { 787 netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n", 788 rxq->rxq_idx); 789 return -ETIMEDOUT; 790 } 791 792 return 0; 793 } 794 795 static void mana_fence_rqs(struct mana_port_context *apc) 796 { 797 unsigned int rxq_idx; 798 struct mana_rxq *rxq; 799 int err; 800 801 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 802 rxq = apc->rxqs[rxq_idx]; 803 err = mana_fence_rq(apc, rxq); 804 805 /* In case of any error, use sleep instead. */ 806 if (err) 807 msleep(100); 808 } 809 } 810 811 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units) 812 { 813 u32 used_space_old; 814 u32 used_space_new; 815 816 used_space_old = wq->head - wq->tail; 817 used_space_new = wq->head - (wq->tail + num_units); 818 819 if (WARN_ON_ONCE(used_space_new > used_space_old)) 820 return -ERANGE; 821 822 wq->tail += num_units; 823 return 0; 824 } 825 826 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc) 827 { 828 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 829 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 830 struct device *dev = gc->dev; 831 int i; 832 833 dma_unmap_single(dev, ash->dma_handle[0], ash->size[0], DMA_TO_DEVICE); 834 835 for (i = 1; i < skb_shinfo(skb)->nr_frags + 1; i++) 836 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i], 837 DMA_TO_DEVICE); 838 } 839 840 static void mana_poll_tx_cq(struct mana_cq *cq) 841 { 842 struct gdma_comp *completions = cq->gdma_comp_buf; 843 struct gdma_posted_wqe_info *wqe_info; 844 unsigned int pkt_transmitted = 0; 845 unsigned int wqe_unit_cnt = 0; 846 struct mana_txq *txq = cq->txq; 847 struct mana_port_context *apc; 848 struct netdev_queue *net_txq; 849 struct gdma_queue *gdma_wq; 850 unsigned int avail_space; 851 struct net_device *ndev; 852 struct sk_buff *skb; 853 bool txq_stopped; 854 int comp_read; 855 int i; 856 857 ndev = txq->ndev; 858 apc = netdev_priv(ndev); 859 860 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions, 861 CQE_POLLING_BUFFER); 862 863 if (comp_read < 1) 864 return; 865 866 for (i = 0; i < comp_read; i++) { 867 struct mana_tx_comp_oob *cqe_oob; 868 869 if (WARN_ON_ONCE(!completions[i].is_sq)) 870 return; 871 872 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data; 873 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type != 874 MANA_CQE_COMPLETION)) 875 return; 876 877 switch (cqe_oob->cqe_hdr.cqe_type) { 878 case CQE_TX_OKAY: 879 break; 880 881 case CQE_TX_SA_DROP: 882 case CQE_TX_MTU_DROP: 883 case CQE_TX_INVALID_OOB: 884 case CQE_TX_INVALID_ETH_TYPE: 885 case CQE_TX_HDR_PROCESSING_ERROR: 886 case CQE_TX_VF_DISABLED: 887 case CQE_TX_VPORT_IDX_OUT_OF_RANGE: 888 case CQE_TX_VPORT_DISABLED: 889 case CQE_TX_VLAN_TAGGING_VIOLATION: 890 WARN_ONCE(1, "TX: CQE error %d: ignored.\n", 891 cqe_oob->cqe_hdr.cqe_type); 892 break; 893 894 default: 895 /* If the CQE type is unexpected, log an error, assert, 896 * and go through the error path. 897 */ 898 WARN_ONCE(1, "TX: Unexpected CQE type %d: HW BUG?\n", 899 cqe_oob->cqe_hdr.cqe_type); 900 return; 901 } 902 903 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num)) 904 return; 905 906 skb = skb_dequeue(&txq->pending_skbs); 907 if (WARN_ON_ONCE(!skb)) 908 return; 909 910 wqe_info = (struct gdma_posted_wqe_info *)skb->cb; 911 wqe_unit_cnt += wqe_info->wqe_size_in_bu; 912 913 mana_unmap_skb(skb, apc); 914 915 napi_consume_skb(skb, cq->budget); 916 917 pkt_transmitted++; 918 } 919 920 if (WARN_ON_ONCE(wqe_unit_cnt == 0)) 921 return; 922 923 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt); 924 925 gdma_wq = txq->gdma_sq; 926 avail_space = mana_gd_wq_avail_space(gdma_wq); 927 928 /* Ensure tail updated before checking q stop */ 929 smp_mb(); 930 931 net_txq = txq->net_txq; 932 txq_stopped = netif_tx_queue_stopped(net_txq); 933 934 /* Ensure checking txq_stopped before apc->port_is_up. */ 935 smp_rmb(); 936 937 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) { 938 netif_tx_wake_queue(net_txq); 939 apc->eth_stats.wake_queue++; 940 } 941 942 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0) 943 WARN_ON_ONCE(1); 944 945 cq->work_done = pkt_transmitted; 946 } 947 948 static void mana_post_pkt_rxq(struct mana_rxq *rxq) 949 { 950 struct mana_recv_buf_oob *recv_buf_oob; 951 u32 curr_index; 952 int err; 953 954 curr_index = rxq->buf_index++; 955 if (rxq->buf_index == rxq->num_rx_buf) 956 rxq->buf_index = 0; 957 958 recv_buf_oob = &rxq->rx_oobs[curr_index]; 959 960 err = mana_gd_post_and_ring(rxq->gdma_rq, &recv_buf_oob->wqe_req, 961 &recv_buf_oob->wqe_inf); 962 if (WARN_ON_ONCE(err)) 963 return; 964 965 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1); 966 } 967 968 static struct sk_buff *mana_build_skb(void *buf_va, uint pkt_len, 969 struct xdp_buff *xdp) 970 { 971 struct sk_buff *skb = build_skb(buf_va, PAGE_SIZE); 972 973 if (!skb) 974 return NULL; 975 976 if (xdp->data_hard_start) { 977 skb_reserve(skb, xdp->data - xdp->data_hard_start); 978 skb_put(skb, xdp->data_end - xdp->data); 979 } else { 980 skb_reserve(skb, XDP_PACKET_HEADROOM); 981 skb_put(skb, pkt_len); 982 } 983 984 return skb; 985 } 986 987 static void mana_rx_skb(void *buf_va, struct mana_rxcomp_oob *cqe, 988 struct mana_rxq *rxq) 989 { 990 struct mana_stats_rx *rx_stats = &rxq->stats; 991 struct net_device *ndev = rxq->ndev; 992 uint pkt_len = cqe->ppi[0].pkt_len; 993 u16 rxq_idx = rxq->rxq_idx; 994 struct napi_struct *napi; 995 struct xdp_buff xdp = {}; 996 struct sk_buff *skb; 997 u32 hash_value; 998 u32 act; 999 1000 rxq->rx_cq.work_done++; 1001 napi = &rxq->rx_cq.napi; 1002 1003 if (!buf_va) { 1004 ++ndev->stats.rx_dropped; 1005 return; 1006 } 1007 1008 act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len); 1009 1010 if (act != XDP_PASS && act != XDP_TX) 1011 goto drop_xdp; 1012 1013 skb = mana_build_skb(buf_va, pkt_len, &xdp); 1014 1015 if (!skb) 1016 goto drop; 1017 1018 skb->dev = napi->dev; 1019 1020 skb->protocol = eth_type_trans(skb, ndev); 1021 skb_checksum_none_assert(skb); 1022 skb_record_rx_queue(skb, rxq_idx); 1023 1024 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) { 1025 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed) 1026 skb->ip_summed = CHECKSUM_UNNECESSARY; 1027 } 1028 1029 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) { 1030 hash_value = cqe->ppi[0].pkt_hash; 1031 1032 if (cqe->rx_hashtype & MANA_HASH_L4) 1033 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4); 1034 else 1035 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3); 1036 } 1037 1038 u64_stats_update_begin(&rx_stats->syncp); 1039 rx_stats->packets++; 1040 rx_stats->bytes += pkt_len; 1041 1042 if (act == XDP_TX) 1043 rx_stats->xdp_tx++; 1044 u64_stats_update_end(&rx_stats->syncp); 1045 1046 if (act == XDP_TX) { 1047 skb_set_queue_mapping(skb, rxq_idx); 1048 mana_xdp_tx(skb, ndev); 1049 return; 1050 } 1051 1052 napi_gro_receive(napi, skb); 1053 1054 return; 1055 1056 drop_xdp: 1057 u64_stats_update_begin(&rx_stats->syncp); 1058 rx_stats->xdp_drop++; 1059 u64_stats_update_end(&rx_stats->syncp); 1060 1061 drop: 1062 WARN_ON_ONCE(rxq->xdp_save_page); 1063 rxq->xdp_save_page = virt_to_page(buf_va); 1064 1065 ++ndev->stats.rx_dropped; 1066 1067 return; 1068 } 1069 1070 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq, 1071 struct gdma_comp *cqe) 1072 { 1073 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data; 1074 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context; 1075 struct net_device *ndev = rxq->ndev; 1076 struct mana_recv_buf_oob *rxbuf_oob; 1077 struct device *dev = gc->dev; 1078 void *new_buf, *old_buf; 1079 struct page *new_page; 1080 u32 curr, pktlen; 1081 dma_addr_t da; 1082 1083 switch (oob->cqe_hdr.cqe_type) { 1084 case CQE_RX_OKAY: 1085 break; 1086 1087 case CQE_RX_TRUNCATED: 1088 ++ndev->stats.rx_dropped; 1089 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index]; 1090 netdev_warn_once(ndev, "Dropped a truncated packet\n"); 1091 goto drop; 1092 1093 case CQE_RX_COALESCED_4: 1094 netdev_err(ndev, "RX coalescing is unsupported\n"); 1095 return; 1096 1097 case CQE_RX_OBJECT_FENCE: 1098 complete(&rxq->fence_event); 1099 return; 1100 1101 default: 1102 netdev_err(ndev, "Unknown RX CQE type = %d\n", 1103 oob->cqe_hdr.cqe_type); 1104 return; 1105 } 1106 1107 pktlen = oob->ppi[0].pkt_len; 1108 1109 if (pktlen == 0) { 1110 /* data packets should never have packetlength of zero */ 1111 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n", 1112 rxq->gdma_id, cq->gdma_id, rxq->rxobj); 1113 return; 1114 } 1115 1116 curr = rxq->buf_index; 1117 rxbuf_oob = &rxq->rx_oobs[curr]; 1118 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1); 1119 1120 /* Reuse XDP dropped page if available */ 1121 if (rxq->xdp_save_page) { 1122 new_page = rxq->xdp_save_page; 1123 rxq->xdp_save_page = NULL; 1124 } else { 1125 new_page = alloc_page(GFP_ATOMIC); 1126 } 1127 1128 if (new_page) { 1129 da = dma_map_page(dev, new_page, XDP_PACKET_HEADROOM, rxq->datasize, 1130 DMA_FROM_DEVICE); 1131 1132 if (dma_mapping_error(dev, da)) { 1133 __free_page(new_page); 1134 new_page = NULL; 1135 } 1136 } 1137 1138 new_buf = new_page ? page_to_virt(new_page) : NULL; 1139 1140 if (new_buf) { 1141 dma_unmap_page(dev, rxbuf_oob->buf_dma_addr, rxq->datasize, 1142 DMA_FROM_DEVICE); 1143 1144 old_buf = rxbuf_oob->buf_va; 1145 1146 /* refresh the rxbuf_oob with the new page */ 1147 rxbuf_oob->buf_va = new_buf; 1148 rxbuf_oob->buf_dma_addr = da; 1149 rxbuf_oob->sgl[0].address = rxbuf_oob->buf_dma_addr; 1150 } else { 1151 old_buf = NULL; /* drop the packet if no memory */ 1152 } 1153 1154 mana_rx_skb(old_buf, oob, rxq); 1155 1156 drop: 1157 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu); 1158 1159 mana_post_pkt_rxq(rxq); 1160 } 1161 1162 static void mana_poll_rx_cq(struct mana_cq *cq) 1163 { 1164 struct gdma_comp *comp = cq->gdma_comp_buf; 1165 int comp_read, i; 1166 1167 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER); 1168 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER); 1169 1170 for (i = 0; i < comp_read; i++) { 1171 if (WARN_ON_ONCE(comp[i].is_sq)) 1172 return; 1173 1174 /* verify recv cqe references the right rxq */ 1175 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id)) 1176 return; 1177 1178 mana_process_rx_cqe(cq->rxq, cq, &comp[i]); 1179 } 1180 } 1181 1182 static void mana_cq_handler(void *context, struct gdma_queue *gdma_queue) 1183 { 1184 struct mana_cq *cq = context; 1185 u8 arm_bit; 1186 1187 WARN_ON_ONCE(cq->gdma_cq != gdma_queue); 1188 1189 if (cq->type == MANA_CQ_TYPE_RX) 1190 mana_poll_rx_cq(cq); 1191 else 1192 mana_poll_tx_cq(cq); 1193 1194 if (cq->work_done < cq->budget && 1195 napi_complete_done(&cq->napi, cq->work_done)) { 1196 arm_bit = SET_ARM_BIT; 1197 } else { 1198 arm_bit = 0; 1199 } 1200 1201 mana_gd_ring_cq(gdma_queue, arm_bit); 1202 } 1203 1204 static int mana_poll(struct napi_struct *napi, int budget) 1205 { 1206 struct mana_cq *cq = container_of(napi, struct mana_cq, napi); 1207 1208 cq->work_done = 0; 1209 cq->budget = budget; 1210 1211 mana_cq_handler(cq, cq->gdma_cq); 1212 1213 return min(cq->work_done, budget); 1214 } 1215 1216 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue) 1217 { 1218 struct mana_cq *cq = context; 1219 1220 napi_schedule_irqoff(&cq->napi); 1221 } 1222 1223 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq) 1224 { 1225 struct gdma_dev *gd = apc->ac->gdma_dev; 1226 1227 if (!cq->gdma_cq) 1228 return; 1229 1230 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq); 1231 } 1232 1233 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq) 1234 { 1235 struct gdma_dev *gd = apc->ac->gdma_dev; 1236 1237 if (!txq->gdma_sq) 1238 return; 1239 1240 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq); 1241 } 1242 1243 static void mana_destroy_txq(struct mana_port_context *apc) 1244 { 1245 struct napi_struct *napi; 1246 int i; 1247 1248 if (!apc->tx_qp) 1249 return; 1250 1251 for (i = 0; i < apc->num_queues; i++) { 1252 napi = &apc->tx_qp[i].tx_cq.napi; 1253 napi_synchronize(napi); 1254 napi_disable(napi); 1255 netif_napi_del(napi); 1256 1257 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object); 1258 1259 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq); 1260 1261 mana_deinit_txq(apc, &apc->tx_qp[i].txq); 1262 } 1263 1264 kfree(apc->tx_qp); 1265 apc->tx_qp = NULL; 1266 } 1267 1268 static int mana_create_txq(struct mana_port_context *apc, 1269 struct net_device *net) 1270 { 1271 struct mana_context *ac = apc->ac; 1272 struct gdma_dev *gd = ac->gdma_dev; 1273 struct mana_obj_spec wq_spec; 1274 struct mana_obj_spec cq_spec; 1275 struct gdma_queue_spec spec; 1276 struct gdma_context *gc; 1277 struct mana_txq *txq; 1278 struct mana_cq *cq; 1279 u32 txq_size; 1280 u32 cq_size; 1281 int err; 1282 int i; 1283 1284 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp), 1285 GFP_KERNEL); 1286 if (!apc->tx_qp) 1287 return -ENOMEM; 1288 1289 /* The minimum size of the WQE is 32 bytes, hence 1290 * MAX_SEND_BUFFERS_PER_QUEUE represents the maximum number of WQEs 1291 * the SQ can store. This value is then used to size other queues 1292 * to prevent overflow. 1293 */ 1294 txq_size = MAX_SEND_BUFFERS_PER_QUEUE * 32; 1295 BUILD_BUG_ON(!PAGE_ALIGNED(txq_size)); 1296 1297 cq_size = MAX_SEND_BUFFERS_PER_QUEUE * COMP_ENTRY_SIZE; 1298 cq_size = PAGE_ALIGN(cq_size); 1299 1300 gc = gd->gdma_context; 1301 1302 for (i = 0; i < apc->num_queues; i++) { 1303 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE; 1304 1305 /* Create SQ */ 1306 txq = &apc->tx_qp[i].txq; 1307 1308 u64_stats_init(&txq->stats.syncp); 1309 txq->ndev = net; 1310 txq->net_txq = netdev_get_tx_queue(net, i); 1311 txq->vp_offset = apc->tx_vp_offset; 1312 skb_queue_head_init(&txq->pending_skbs); 1313 1314 memset(&spec, 0, sizeof(spec)); 1315 spec.type = GDMA_SQ; 1316 spec.monitor_avl_buf = true; 1317 spec.queue_size = txq_size; 1318 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq); 1319 if (err) 1320 goto out; 1321 1322 /* Create SQ's CQ */ 1323 cq = &apc->tx_qp[i].tx_cq; 1324 cq->type = MANA_CQ_TYPE_TX; 1325 1326 cq->txq = txq; 1327 1328 memset(&spec, 0, sizeof(spec)); 1329 spec.type = GDMA_CQ; 1330 spec.monitor_avl_buf = false; 1331 spec.queue_size = cq_size; 1332 spec.cq.callback = mana_schedule_napi; 1333 spec.cq.parent_eq = ac->eqs[i].eq; 1334 spec.cq.context = cq; 1335 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 1336 if (err) 1337 goto out; 1338 1339 memset(&wq_spec, 0, sizeof(wq_spec)); 1340 memset(&cq_spec, 0, sizeof(cq_spec)); 1341 1342 wq_spec.gdma_region = txq->gdma_sq->mem_info.gdma_region; 1343 wq_spec.queue_size = txq->gdma_sq->queue_size; 1344 1345 cq_spec.gdma_region = cq->gdma_cq->mem_info.gdma_region; 1346 cq_spec.queue_size = cq->gdma_cq->queue_size; 1347 cq_spec.modr_ctx_id = 0; 1348 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 1349 1350 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ, 1351 &wq_spec, &cq_spec, 1352 &apc->tx_qp[i].tx_object); 1353 1354 if (err) 1355 goto out; 1356 1357 txq->gdma_sq->id = wq_spec.queue_index; 1358 cq->gdma_cq->id = cq_spec.queue_index; 1359 1360 txq->gdma_sq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 1361 cq->gdma_cq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 1362 1363 txq->gdma_txq_id = txq->gdma_sq->id; 1364 1365 cq->gdma_id = cq->gdma_cq->id; 1366 1367 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 1368 err = -EINVAL; 1369 goto out; 1370 } 1371 1372 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 1373 1374 netif_napi_add_tx(net, &cq->napi, mana_poll); 1375 napi_enable(&cq->napi); 1376 1377 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 1378 } 1379 1380 return 0; 1381 out: 1382 mana_destroy_txq(apc); 1383 return err; 1384 } 1385 1386 static void mana_destroy_rxq(struct mana_port_context *apc, 1387 struct mana_rxq *rxq, bool validate_state) 1388 1389 { 1390 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 1391 struct mana_recv_buf_oob *rx_oob; 1392 struct device *dev = gc->dev; 1393 struct napi_struct *napi; 1394 int i; 1395 1396 if (!rxq) 1397 return; 1398 1399 napi = &rxq->rx_cq.napi; 1400 1401 if (validate_state) 1402 napi_synchronize(napi); 1403 1404 napi_disable(napi); 1405 1406 xdp_rxq_info_unreg(&rxq->xdp_rxq); 1407 1408 netif_napi_del(napi); 1409 1410 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj); 1411 1412 mana_deinit_cq(apc, &rxq->rx_cq); 1413 1414 if (rxq->xdp_save_page) 1415 __free_page(rxq->xdp_save_page); 1416 1417 for (i = 0; i < rxq->num_rx_buf; i++) { 1418 rx_oob = &rxq->rx_oobs[i]; 1419 1420 if (!rx_oob->buf_va) 1421 continue; 1422 1423 dma_unmap_page(dev, rx_oob->buf_dma_addr, rxq->datasize, 1424 DMA_FROM_DEVICE); 1425 1426 free_page((unsigned long)rx_oob->buf_va); 1427 rx_oob->buf_va = NULL; 1428 } 1429 1430 if (rxq->gdma_rq) 1431 mana_gd_destroy_queue(gc, rxq->gdma_rq); 1432 1433 kfree(rxq); 1434 } 1435 1436 #define MANA_WQE_HEADER_SIZE 16 1437 #define MANA_WQE_SGE_SIZE 16 1438 1439 static int mana_alloc_rx_wqe(struct mana_port_context *apc, 1440 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size) 1441 { 1442 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 1443 struct mana_recv_buf_oob *rx_oob; 1444 struct device *dev = gc->dev; 1445 struct page *page; 1446 dma_addr_t da; 1447 u32 buf_idx; 1448 1449 WARN_ON(rxq->datasize == 0 || rxq->datasize > PAGE_SIZE); 1450 1451 *rxq_size = 0; 1452 *cq_size = 0; 1453 1454 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 1455 rx_oob = &rxq->rx_oobs[buf_idx]; 1456 memset(rx_oob, 0, sizeof(*rx_oob)); 1457 1458 page = alloc_page(GFP_KERNEL); 1459 if (!page) 1460 return -ENOMEM; 1461 1462 da = dma_map_page(dev, page, XDP_PACKET_HEADROOM, rxq->datasize, 1463 DMA_FROM_DEVICE); 1464 1465 if (dma_mapping_error(dev, da)) { 1466 __free_page(page); 1467 return -ENOMEM; 1468 } 1469 1470 rx_oob->buf_va = page_to_virt(page); 1471 rx_oob->buf_dma_addr = da; 1472 1473 rx_oob->num_sge = 1; 1474 rx_oob->sgl[0].address = rx_oob->buf_dma_addr; 1475 rx_oob->sgl[0].size = rxq->datasize; 1476 rx_oob->sgl[0].mem_key = apc->ac->gdma_dev->gpa_mkey; 1477 1478 rx_oob->wqe_req.sgl = rx_oob->sgl; 1479 rx_oob->wqe_req.num_sge = rx_oob->num_sge; 1480 rx_oob->wqe_req.inline_oob_size = 0; 1481 rx_oob->wqe_req.inline_oob_data = NULL; 1482 rx_oob->wqe_req.flags = 0; 1483 rx_oob->wqe_req.client_data_unit = 0; 1484 1485 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE + 1486 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32); 1487 *cq_size += COMP_ENTRY_SIZE; 1488 } 1489 1490 return 0; 1491 } 1492 1493 static int mana_push_wqe(struct mana_rxq *rxq) 1494 { 1495 struct mana_recv_buf_oob *rx_oob; 1496 u32 buf_idx; 1497 int err; 1498 1499 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 1500 rx_oob = &rxq->rx_oobs[buf_idx]; 1501 1502 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req, 1503 &rx_oob->wqe_inf); 1504 if (err) 1505 return -ENOSPC; 1506 } 1507 1508 return 0; 1509 } 1510 1511 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc, 1512 u32 rxq_idx, struct mana_eq *eq, 1513 struct net_device *ndev) 1514 { 1515 struct gdma_dev *gd = apc->ac->gdma_dev; 1516 struct mana_obj_spec wq_spec; 1517 struct mana_obj_spec cq_spec; 1518 struct gdma_queue_spec spec; 1519 struct mana_cq *cq = NULL; 1520 struct gdma_context *gc; 1521 u32 cq_size, rq_size; 1522 struct mana_rxq *rxq; 1523 int err; 1524 1525 gc = gd->gdma_context; 1526 1527 rxq = kzalloc(struct_size(rxq, rx_oobs, RX_BUFFERS_PER_QUEUE), 1528 GFP_KERNEL); 1529 if (!rxq) 1530 return NULL; 1531 1532 rxq->ndev = ndev; 1533 rxq->num_rx_buf = RX_BUFFERS_PER_QUEUE; 1534 rxq->rxq_idx = rxq_idx; 1535 rxq->datasize = ALIGN(MAX_FRAME_SIZE, 64); 1536 rxq->rxobj = INVALID_MANA_HANDLE; 1537 1538 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size); 1539 if (err) 1540 goto out; 1541 1542 rq_size = PAGE_ALIGN(rq_size); 1543 cq_size = PAGE_ALIGN(cq_size); 1544 1545 /* Create RQ */ 1546 memset(&spec, 0, sizeof(spec)); 1547 spec.type = GDMA_RQ; 1548 spec.monitor_avl_buf = true; 1549 spec.queue_size = rq_size; 1550 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq); 1551 if (err) 1552 goto out; 1553 1554 /* Create RQ's CQ */ 1555 cq = &rxq->rx_cq; 1556 cq->type = MANA_CQ_TYPE_RX; 1557 cq->rxq = rxq; 1558 1559 memset(&spec, 0, sizeof(spec)); 1560 spec.type = GDMA_CQ; 1561 spec.monitor_avl_buf = false; 1562 spec.queue_size = cq_size; 1563 spec.cq.callback = mana_schedule_napi; 1564 spec.cq.parent_eq = eq->eq; 1565 spec.cq.context = cq; 1566 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 1567 if (err) 1568 goto out; 1569 1570 memset(&wq_spec, 0, sizeof(wq_spec)); 1571 memset(&cq_spec, 0, sizeof(cq_spec)); 1572 wq_spec.gdma_region = rxq->gdma_rq->mem_info.gdma_region; 1573 wq_spec.queue_size = rxq->gdma_rq->queue_size; 1574 1575 cq_spec.gdma_region = cq->gdma_cq->mem_info.gdma_region; 1576 cq_spec.queue_size = cq->gdma_cq->queue_size; 1577 cq_spec.modr_ctx_id = 0; 1578 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 1579 1580 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ, 1581 &wq_spec, &cq_spec, &rxq->rxobj); 1582 if (err) 1583 goto out; 1584 1585 rxq->gdma_rq->id = wq_spec.queue_index; 1586 cq->gdma_cq->id = cq_spec.queue_index; 1587 1588 rxq->gdma_rq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 1589 cq->gdma_cq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 1590 1591 rxq->gdma_id = rxq->gdma_rq->id; 1592 cq->gdma_id = cq->gdma_cq->id; 1593 1594 err = mana_push_wqe(rxq); 1595 if (err) 1596 goto out; 1597 1598 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 1599 err = -EINVAL; 1600 goto out; 1601 } 1602 1603 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 1604 1605 netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1); 1606 1607 WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx, 1608 cq->napi.napi_id)); 1609 WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, 1610 MEM_TYPE_PAGE_SHARED, NULL)); 1611 1612 napi_enable(&cq->napi); 1613 1614 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 1615 out: 1616 if (!err) 1617 return rxq; 1618 1619 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err); 1620 1621 mana_destroy_rxq(apc, rxq, false); 1622 1623 if (cq) 1624 mana_deinit_cq(apc, cq); 1625 1626 return NULL; 1627 } 1628 1629 static int mana_add_rx_queues(struct mana_port_context *apc, 1630 struct net_device *ndev) 1631 { 1632 struct mana_context *ac = apc->ac; 1633 struct mana_rxq *rxq; 1634 int err = 0; 1635 int i; 1636 1637 for (i = 0; i < apc->num_queues; i++) { 1638 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev); 1639 if (!rxq) { 1640 err = -ENOMEM; 1641 goto out; 1642 } 1643 1644 u64_stats_init(&rxq->stats.syncp); 1645 1646 apc->rxqs[i] = rxq; 1647 } 1648 1649 apc->default_rxobj = apc->rxqs[0]->rxobj; 1650 out: 1651 return err; 1652 } 1653 1654 static void mana_destroy_vport(struct mana_port_context *apc) 1655 { 1656 struct mana_rxq *rxq; 1657 u32 rxq_idx; 1658 1659 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 1660 rxq = apc->rxqs[rxq_idx]; 1661 if (!rxq) 1662 continue; 1663 1664 mana_destroy_rxq(apc, rxq, true); 1665 apc->rxqs[rxq_idx] = NULL; 1666 } 1667 1668 mana_destroy_txq(apc); 1669 } 1670 1671 static int mana_create_vport(struct mana_port_context *apc, 1672 struct net_device *net) 1673 { 1674 struct gdma_dev *gd = apc->ac->gdma_dev; 1675 int err; 1676 1677 apc->default_rxobj = INVALID_MANA_HANDLE; 1678 1679 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell); 1680 if (err) 1681 return err; 1682 1683 return mana_create_txq(apc, net); 1684 } 1685 1686 static void mana_rss_table_init(struct mana_port_context *apc) 1687 { 1688 int i; 1689 1690 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) 1691 apc->indir_table[i] = 1692 ethtool_rxfh_indir_default(i, apc->num_queues); 1693 } 1694 1695 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx, 1696 bool update_hash, bool update_tab) 1697 { 1698 u32 queue_idx; 1699 int err; 1700 int i; 1701 1702 if (update_tab) { 1703 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) { 1704 queue_idx = apc->indir_table[i]; 1705 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj; 1706 } 1707 } 1708 1709 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab); 1710 if (err) 1711 return err; 1712 1713 mana_fence_rqs(apc); 1714 1715 return 0; 1716 } 1717 1718 static int mana_init_port(struct net_device *ndev) 1719 { 1720 struct mana_port_context *apc = netdev_priv(ndev); 1721 u32 max_txq, max_rxq, max_queues; 1722 int port_idx = apc->port_idx; 1723 u32 num_indirect_entries; 1724 int err; 1725 1726 err = mana_init_port_context(apc); 1727 if (err) 1728 return err; 1729 1730 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq, 1731 &num_indirect_entries); 1732 if (err) { 1733 netdev_err(ndev, "Failed to query info for vPort %d\n", 1734 port_idx); 1735 goto reset_apc; 1736 } 1737 1738 max_queues = min_t(u32, max_txq, max_rxq); 1739 if (apc->max_queues > max_queues) 1740 apc->max_queues = max_queues; 1741 1742 if (apc->num_queues > apc->max_queues) 1743 apc->num_queues = apc->max_queues; 1744 1745 eth_hw_addr_set(ndev, apc->mac_addr); 1746 1747 return 0; 1748 1749 reset_apc: 1750 kfree(apc->rxqs); 1751 apc->rxqs = NULL; 1752 return err; 1753 } 1754 1755 int mana_alloc_queues(struct net_device *ndev) 1756 { 1757 struct mana_port_context *apc = netdev_priv(ndev); 1758 int err; 1759 1760 err = mana_create_vport(apc, ndev); 1761 if (err) 1762 return err; 1763 1764 err = netif_set_real_num_tx_queues(ndev, apc->num_queues); 1765 if (err) 1766 goto destroy_vport; 1767 1768 err = mana_add_rx_queues(apc, ndev); 1769 if (err) 1770 goto destroy_vport; 1771 1772 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE; 1773 1774 err = netif_set_real_num_rx_queues(ndev, apc->num_queues); 1775 if (err) 1776 goto destroy_vport; 1777 1778 mana_rss_table_init(apc); 1779 1780 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true); 1781 if (err) 1782 goto destroy_vport; 1783 1784 mana_chn_setxdp(apc, mana_xdp_get(apc)); 1785 1786 return 0; 1787 1788 destroy_vport: 1789 mana_destroy_vport(apc); 1790 return err; 1791 } 1792 1793 int mana_attach(struct net_device *ndev) 1794 { 1795 struct mana_port_context *apc = netdev_priv(ndev); 1796 int err; 1797 1798 ASSERT_RTNL(); 1799 1800 err = mana_init_port(ndev); 1801 if (err) 1802 return err; 1803 1804 if (apc->port_st_save) { 1805 err = mana_alloc_queues(ndev); 1806 if (err) { 1807 mana_cleanup_port_context(apc); 1808 return err; 1809 } 1810 } 1811 1812 apc->port_is_up = apc->port_st_save; 1813 1814 /* Ensure port state updated before txq state */ 1815 smp_wmb(); 1816 1817 if (apc->port_is_up) 1818 netif_carrier_on(ndev); 1819 1820 netif_device_attach(ndev); 1821 1822 return 0; 1823 } 1824 1825 static int mana_dealloc_queues(struct net_device *ndev) 1826 { 1827 struct mana_port_context *apc = netdev_priv(ndev); 1828 struct mana_txq *txq; 1829 int i, err; 1830 1831 if (apc->port_is_up) 1832 return -EINVAL; 1833 1834 mana_chn_setxdp(apc, NULL); 1835 1836 /* No packet can be transmitted now since apc->port_is_up is false. 1837 * There is still a tiny chance that mana_poll_tx_cq() can re-enable 1838 * a txq because it may not timely see apc->port_is_up being cleared 1839 * to false, but it doesn't matter since mana_start_xmit() drops any 1840 * new packets due to apc->port_is_up being false. 1841 * 1842 * Drain all the in-flight TX packets 1843 */ 1844 for (i = 0; i < apc->num_queues; i++) { 1845 txq = &apc->tx_qp[i].txq; 1846 1847 while (atomic_read(&txq->pending_sends) > 0) 1848 usleep_range(1000, 2000); 1849 } 1850 1851 /* We're 100% sure the queues can no longer be woken up, because 1852 * we're sure now mana_poll_tx_cq() can't be running. 1853 */ 1854 1855 apc->rss_state = TRI_STATE_FALSE; 1856 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false); 1857 if (err) { 1858 netdev_err(ndev, "Failed to disable vPort: %d\n", err); 1859 return err; 1860 } 1861 1862 mana_destroy_vport(apc); 1863 1864 return 0; 1865 } 1866 1867 int mana_detach(struct net_device *ndev, bool from_close) 1868 { 1869 struct mana_port_context *apc = netdev_priv(ndev); 1870 int err; 1871 1872 ASSERT_RTNL(); 1873 1874 apc->port_st_save = apc->port_is_up; 1875 apc->port_is_up = false; 1876 1877 /* Ensure port state updated before txq state */ 1878 smp_wmb(); 1879 1880 netif_tx_disable(ndev); 1881 netif_carrier_off(ndev); 1882 1883 if (apc->port_st_save) { 1884 err = mana_dealloc_queues(ndev); 1885 if (err) 1886 return err; 1887 } 1888 1889 if (!from_close) { 1890 netif_device_detach(ndev); 1891 mana_cleanup_port_context(apc); 1892 } 1893 1894 return 0; 1895 } 1896 1897 static int mana_probe_port(struct mana_context *ac, int port_idx, 1898 struct net_device **ndev_storage) 1899 { 1900 struct gdma_context *gc = ac->gdma_dev->gdma_context; 1901 struct mana_port_context *apc; 1902 struct net_device *ndev; 1903 int err; 1904 1905 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context), 1906 gc->max_num_queues); 1907 if (!ndev) 1908 return -ENOMEM; 1909 1910 *ndev_storage = ndev; 1911 1912 apc = netdev_priv(ndev); 1913 apc->ac = ac; 1914 apc->ndev = ndev; 1915 apc->max_queues = gc->max_num_queues; 1916 apc->num_queues = gc->max_num_queues; 1917 apc->port_handle = INVALID_MANA_HANDLE; 1918 apc->port_idx = port_idx; 1919 1920 ndev->netdev_ops = &mana_devops; 1921 ndev->ethtool_ops = &mana_ethtool_ops; 1922 ndev->mtu = ETH_DATA_LEN; 1923 ndev->max_mtu = ndev->mtu; 1924 ndev->min_mtu = ndev->mtu; 1925 ndev->needed_headroom = MANA_HEADROOM; 1926 SET_NETDEV_DEV(ndev, gc->dev); 1927 1928 netif_carrier_off(ndev); 1929 1930 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE); 1931 1932 err = mana_init_port(ndev); 1933 if (err) 1934 goto free_net; 1935 1936 netdev_lockdep_set_classes(ndev); 1937 1938 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 1939 ndev->hw_features |= NETIF_F_RXCSUM; 1940 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6; 1941 ndev->hw_features |= NETIF_F_RXHASH; 1942 ndev->features = ndev->hw_features; 1943 ndev->vlan_features = 0; 1944 1945 err = register_netdev(ndev); 1946 if (err) { 1947 netdev_err(ndev, "Unable to register netdev.\n"); 1948 goto reset_apc; 1949 } 1950 1951 return 0; 1952 1953 reset_apc: 1954 kfree(apc->rxqs); 1955 apc->rxqs = NULL; 1956 free_net: 1957 *ndev_storage = NULL; 1958 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err); 1959 free_netdev(ndev); 1960 return err; 1961 } 1962 1963 int mana_probe(struct gdma_dev *gd, bool resuming) 1964 { 1965 struct gdma_context *gc = gd->gdma_context; 1966 struct mana_context *ac = gd->driver_data; 1967 struct device *dev = gc->dev; 1968 u16 num_ports = 0; 1969 int err; 1970 int i; 1971 1972 dev_info(dev, 1973 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n", 1974 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION); 1975 1976 err = mana_gd_register_device(gd); 1977 if (err) 1978 return err; 1979 1980 if (!resuming) { 1981 ac = kzalloc(sizeof(*ac), GFP_KERNEL); 1982 if (!ac) 1983 return -ENOMEM; 1984 1985 ac->gdma_dev = gd; 1986 gd->driver_data = ac; 1987 } 1988 1989 err = mana_create_eq(ac); 1990 if (err) 1991 goto out; 1992 1993 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION, 1994 MANA_MICRO_VERSION, &num_ports); 1995 if (err) 1996 goto out; 1997 1998 if (!resuming) { 1999 ac->num_ports = num_ports; 2000 } else { 2001 if (ac->num_ports != num_ports) { 2002 dev_err(dev, "The number of vPorts changed: %d->%d\n", 2003 ac->num_ports, num_ports); 2004 err = -EPROTO; 2005 goto out; 2006 } 2007 } 2008 2009 if (ac->num_ports == 0) 2010 dev_err(dev, "Failed to detect any vPort\n"); 2011 2012 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV) 2013 ac->num_ports = MAX_PORTS_IN_MANA_DEV; 2014 2015 if (!resuming) { 2016 for (i = 0; i < ac->num_ports; i++) { 2017 err = mana_probe_port(ac, i, &ac->ports[i]); 2018 if (err) 2019 break; 2020 } 2021 } else { 2022 for (i = 0; i < ac->num_ports; i++) { 2023 rtnl_lock(); 2024 err = mana_attach(ac->ports[i]); 2025 rtnl_unlock(); 2026 if (err) 2027 break; 2028 } 2029 } 2030 out: 2031 if (err) 2032 mana_remove(gd, false); 2033 2034 return err; 2035 } 2036 2037 void mana_remove(struct gdma_dev *gd, bool suspending) 2038 { 2039 struct gdma_context *gc = gd->gdma_context; 2040 struct mana_context *ac = gd->driver_data; 2041 struct device *dev = gc->dev; 2042 struct net_device *ndev; 2043 int err; 2044 int i; 2045 2046 for (i = 0; i < ac->num_ports; i++) { 2047 ndev = ac->ports[i]; 2048 if (!ndev) { 2049 if (i == 0) 2050 dev_err(dev, "No net device to remove\n"); 2051 goto out; 2052 } 2053 2054 /* All cleanup actions should stay after rtnl_lock(), otherwise 2055 * other functions may access partially cleaned up data. 2056 */ 2057 rtnl_lock(); 2058 2059 err = mana_detach(ndev, false); 2060 if (err) 2061 netdev_err(ndev, "Failed to detach vPort %d: %d\n", 2062 i, err); 2063 2064 if (suspending) { 2065 /* No need to unregister the ndev. */ 2066 rtnl_unlock(); 2067 continue; 2068 } 2069 2070 unregister_netdevice(ndev); 2071 2072 rtnl_unlock(); 2073 2074 free_netdev(ndev); 2075 } 2076 2077 mana_destroy_eq(ac); 2078 2079 out: 2080 mana_gd_deregister_device(gd); 2081 2082 if (suspending) 2083 return; 2084 2085 gd->driver_data = NULL; 2086 gd->gdma_context = NULL; 2087 kfree(ac); 2088 } 2089