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