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 page = dev_alloc_pages(get_order(mpc->rxbpre_alloc_size)); 665 if (!page) 666 goto error; 667 668 va = page_to_virt(page); 669 670 da = dma_map_single(dev, va + mpc->rxbpre_headroom, 671 mpc->rxbpre_datasize, DMA_FROM_DEVICE); 672 if (dma_mapping_error(dev, da)) { 673 put_page(page); 674 goto error; 675 } 676 677 mpc->rxbufs_pre[i] = va; 678 mpc->das_pre[i] = da; 679 mpc->rxbpre_total = i + 1; 680 } 681 682 return 0; 683 684 error: 685 netdev_err(mpc->ndev, "Failed to pre-allocate RX buffers for %d queues\n", num_queues); 686 mana_pre_dealloc_rxbufs(mpc); 687 return -ENOMEM; 688 } 689 690 static int mana_change_mtu(struct net_device *ndev, int new_mtu) 691 { 692 struct mana_port_context *mpc = netdev_priv(ndev); 693 unsigned int old_mtu = ndev->mtu; 694 int err; 695 696 /* Pre-allocate buffers to prevent failure in mana_attach later */ 697 err = mana_pre_alloc_rxbufs(mpc, new_mtu, mpc->num_queues); 698 if (err) { 699 netdev_err(ndev, "Insufficient memory for new MTU\n"); 700 return err; 701 } 702 703 err = mana_detach(ndev, false); 704 if (err) { 705 netdev_err(ndev, "mana_detach failed: %d\n", err); 706 goto out; 707 } 708 709 WRITE_ONCE(ndev->mtu, new_mtu); 710 711 err = mana_attach(ndev); 712 if (err) { 713 netdev_err(ndev, "mana_attach failed: %d\n", err); 714 WRITE_ONCE(ndev->mtu, old_mtu); 715 } 716 717 out: 718 mana_pre_dealloc_rxbufs(mpc); 719 return err; 720 } 721 722 static const struct net_device_ops mana_devops = { 723 .ndo_open = mana_open, 724 .ndo_stop = mana_close, 725 .ndo_select_queue = mana_select_queue, 726 .ndo_start_xmit = mana_start_xmit, 727 .ndo_validate_addr = eth_validate_addr, 728 .ndo_get_stats64 = mana_get_stats64, 729 .ndo_bpf = mana_bpf, 730 .ndo_xdp_xmit = mana_xdp_xmit, 731 .ndo_change_mtu = mana_change_mtu, 732 }; 733 734 static void mana_cleanup_port_context(struct mana_port_context *apc) 735 { 736 /* 737 * make sure subsequent cleanup attempts don't end up removing already 738 * cleaned dentry pointer 739 */ 740 debugfs_remove(apc->mana_port_debugfs); 741 apc->mana_port_debugfs = NULL; 742 kfree(apc->rxqs); 743 apc->rxqs = NULL; 744 } 745 746 static void mana_cleanup_indir_table(struct mana_port_context *apc) 747 { 748 apc->indir_table_sz = 0; 749 kfree(apc->indir_table); 750 kfree(apc->rxobj_table); 751 } 752 753 static int mana_init_port_context(struct mana_port_context *apc) 754 { 755 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *), 756 GFP_KERNEL); 757 758 return !apc->rxqs ? -ENOMEM : 0; 759 } 760 761 static int mana_send_request(struct mana_context *ac, void *in_buf, 762 u32 in_len, void *out_buf, u32 out_len) 763 { 764 struct gdma_context *gc = ac->gdma_dev->gdma_context; 765 struct gdma_resp_hdr *resp = out_buf; 766 struct gdma_req_hdr *req = in_buf; 767 struct device *dev = gc->dev; 768 static atomic_t activity_id; 769 int err; 770 771 req->dev_id = gc->mana.dev_id; 772 req->activity_id = atomic_inc_return(&activity_id); 773 774 err = mana_gd_send_request(gc, in_len, in_buf, out_len, 775 out_buf); 776 if (err || resp->status) { 777 dev_err(dev, "Failed to send mana message: %d, 0x%x\n", 778 err, resp->status); 779 return err ? err : -EPROTO; 780 } 781 782 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 || 783 req->activity_id != resp->activity_id) { 784 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n", 785 req->dev_id.as_uint32, resp->dev_id.as_uint32, 786 req->activity_id, resp->activity_id); 787 return -EPROTO; 788 } 789 790 return 0; 791 } 792 793 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr, 794 const enum mana_command_code expected_code, 795 const u32 min_size) 796 { 797 if (resp_hdr->response.msg_type != expected_code) 798 return -EPROTO; 799 800 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1) 801 return -EPROTO; 802 803 if (resp_hdr->response.msg_size < min_size) 804 return -EPROTO; 805 806 return 0; 807 } 808 809 static int mana_pf_register_hw_vport(struct mana_port_context *apc) 810 { 811 struct mana_register_hw_vport_resp resp = {}; 812 struct mana_register_hw_vport_req req = {}; 813 int err; 814 815 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT, 816 sizeof(req), sizeof(resp)); 817 req.attached_gfid = 1; 818 req.is_pf_default_vport = 1; 819 req.allow_all_ether_types = 1; 820 821 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 822 sizeof(resp)); 823 if (err) { 824 netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err); 825 return err; 826 } 827 828 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT, 829 sizeof(resp)); 830 if (err || resp.hdr.status) { 831 netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n", 832 err, resp.hdr.status); 833 return err ? err : -EPROTO; 834 } 835 836 apc->port_handle = resp.hw_vport_handle; 837 return 0; 838 } 839 840 static void mana_pf_deregister_hw_vport(struct mana_port_context *apc) 841 { 842 struct mana_deregister_hw_vport_resp resp = {}; 843 struct mana_deregister_hw_vport_req req = {}; 844 int err; 845 846 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT, 847 sizeof(req), sizeof(resp)); 848 req.hw_vport_handle = apc->port_handle; 849 850 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 851 sizeof(resp)); 852 if (err) { 853 netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n", 854 err); 855 return; 856 } 857 858 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT, 859 sizeof(resp)); 860 if (err || resp.hdr.status) 861 netdev_err(apc->ndev, 862 "Failed to deregister hw vPort: %d, 0x%x\n", 863 err, resp.hdr.status); 864 } 865 866 static int mana_pf_register_filter(struct mana_port_context *apc) 867 { 868 struct mana_register_filter_resp resp = {}; 869 struct mana_register_filter_req req = {}; 870 int err; 871 872 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER, 873 sizeof(req), sizeof(resp)); 874 req.vport = apc->port_handle; 875 memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN); 876 877 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 878 sizeof(resp)); 879 if (err) { 880 netdev_err(apc->ndev, "Failed to register filter: %d\n", err); 881 return err; 882 } 883 884 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER, 885 sizeof(resp)); 886 if (err || resp.hdr.status) { 887 netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n", 888 err, resp.hdr.status); 889 return err ? err : -EPROTO; 890 } 891 892 apc->pf_filter_handle = resp.filter_handle; 893 return 0; 894 } 895 896 static void mana_pf_deregister_filter(struct mana_port_context *apc) 897 { 898 struct mana_deregister_filter_resp resp = {}; 899 struct mana_deregister_filter_req req = {}; 900 int err; 901 902 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER, 903 sizeof(req), sizeof(resp)); 904 req.filter_handle = apc->pf_filter_handle; 905 906 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 907 sizeof(resp)); 908 if (err) { 909 netdev_err(apc->ndev, "Failed to unregister filter: %d\n", 910 err); 911 return; 912 } 913 914 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER, 915 sizeof(resp)); 916 if (err || resp.hdr.status) 917 netdev_err(apc->ndev, 918 "Failed to deregister filter: %d, 0x%x\n", 919 err, resp.hdr.status); 920 } 921 922 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver, 923 u32 proto_minor_ver, u32 proto_micro_ver, 924 u16 *max_num_vports) 925 { 926 struct gdma_context *gc = ac->gdma_dev->gdma_context; 927 struct mana_query_device_cfg_resp resp = {}; 928 struct mana_query_device_cfg_req req = {}; 929 struct device *dev = gc->dev; 930 int err = 0; 931 932 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG, 933 sizeof(req), sizeof(resp)); 934 935 req.hdr.resp.msg_version = GDMA_MESSAGE_V2; 936 937 req.proto_major_ver = proto_major_ver; 938 req.proto_minor_ver = proto_minor_ver; 939 req.proto_micro_ver = proto_micro_ver; 940 941 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp)); 942 if (err) { 943 dev_err(dev, "Failed to query config: %d", err); 944 return err; 945 } 946 947 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG, 948 sizeof(resp)); 949 if (err || resp.hdr.status) { 950 dev_err(dev, "Invalid query result: %d, 0x%x\n", err, 951 resp.hdr.status); 952 if (!err) 953 err = -EPROTO; 954 return err; 955 } 956 957 *max_num_vports = resp.max_num_vports; 958 959 if (resp.hdr.response.msg_version == GDMA_MESSAGE_V2) 960 gc->adapter_mtu = resp.adapter_mtu; 961 else 962 gc->adapter_mtu = ETH_FRAME_LEN; 963 964 debugfs_create_u16("adapter-MTU", 0400, gc->mana_pci_debugfs, &gc->adapter_mtu); 965 966 return 0; 967 } 968 969 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index, 970 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry) 971 { 972 struct mana_query_vport_cfg_resp resp = {}; 973 struct mana_query_vport_cfg_req req = {}; 974 int err; 975 976 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG, 977 sizeof(req), sizeof(resp)); 978 979 req.vport_index = vport_index; 980 981 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 982 sizeof(resp)); 983 if (err) 984 return err; 985 986 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG, 987 sizeof(resp)); 988 if (err) 989 return err; 990 991 if (resp.hdr.status) 992 return -EPROTO; 993 994 *max_sq = resp.max_num_sq; 995 *max_rq = resp.max_num_rq; 996 if (resp.num_indirection_ent > 0 && 997 resp.num_indirection_ent <= MANA_INDIRECT_TABLE_MAX_SIZE && 998 is_power_of_2(resp.num_indirection_ent)) { 999 *num_indir_entry = resp.num_indirection_ent; 1000 } else { 1001 netdev_warn(apc->ndev, 1002 "Setting indirection table size to default %d for vPort %d\n", 1003 MANA_INDIRECT_TABLE_DEF_SIZE, apc->port_idx); 1004 *num_indir_entry = MANA_INDIRECT_TABLE_DEF_SIZE; 1005 } 1006 1007 apc->port_handle = resp.vport; 1008 ether_addr_copy(apc->mac_addr, resp.mac_addr); 1009 1010 return 0; 1011 } 1012 1013 void mana_uncfg_vport(struct mana_port_context *apc) 1014 { 1015 mutex_lock(&apc->vport_mutex); 1016 apc->vport_use_count--; 1017 WARN_ON(apc->vport_use_count < 0); 1018 mutex_unlock(&apc->vport_mutex); 1019 } 1020 EXPORT_SYMBOL_NS(mana_uncfg_vport, "NET_MANA"); 1021 1022 int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id, 1023 u32 doorbell_pg_id) 1024 { 1025 struct mana_config_vport_resp resp = {}; 1026 struct mana_config_vport_req req = {}; 1027 int err; 1028 1029 /* This function is used to program the Ethernet port in the hardware 1030 * table. It can be called from the Ethernet driver or the RDMA driver. 1031 * 1032 * For Ethernet usage, the hardware supports only one active user on a 1033 * physical port. The driver checks on the port usage before programming 1034 * the hardware when creating the RAW QP (RDMA driver) or exposing the 1035 * device to kernel NET layer (Ethernet driver). 1036 * 1037 * Because the RDMA driver doesn't know in advance which QP type the 1038 * user will create, it exposes the device with all its ports. The user 1039 * may not be able to create RAW QP on a port if this port is already 1040 * in used by the Ethernet driver from the kernel. 1041 * 1042 * This physical port limitation only applies to the RAW QP. For RC QP, 1043 * the hardware doesn't have this limitation. The user can create RC 1044 * QPs on a physical port up to the hardware limits independent of the 1045 * Ethernet usage on the same port. 1046 */ 1047 mutex_lock(&apc->vport_mutex); 1048 if (apc->vport_use_count > 0) { 1049 mutex_unlock(&apc->vport_mutex); 1050 return -EBUSY; 1051 } 1052 apc->vport_use_count++; 1053 mutex_unlock(&apc->vport_mutex); 1054 1055 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX, 1056 sizeof(req), sizeof(resp)); 1057 req.vport = apc->port_handle; 1058 req.pdid = protection_dom_id; 1059 req.doorbell_pageid = doorbell_pg_id; 1060 1061 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1062 sizeof(resp)); 1063 if (err) { 1064 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err); 1065 goto out; 1066 } 1067 1068 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX, 1069 sizeof(resp)); 1070 if (err || resp.hdr.status) { 1071 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n", 1072 err, resp.hdr.status); 1073 if (!err) 1074 err = -EPROTO; 1075 1076 goto out; 1077 } 1078 1079 apc->tx_shortform_allowed = resp.short_form_allowed; 1080 apc->tx_vp_offset = resp.tx_vport_offset; 1081 1082 netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n", 1083 apc->port_handle, protection_dom_id, doorbell_pg_id); 1084 out: 1085 if (err) 1086 mana_uncfg_vport(apc); 1087 1088 return err; 1089 } 1090 EXPORT_SYMBOL_NS(mana_cfg_vport, "NET_MANA"); 1091 1092 static int mana_cfg_vport_steering(struct mana_port_context *apc, 1093 enum TRI_STATE rx, 1094 bool update_default_rxobj, bool update_key, 1095 bool update_tab) 1096 { 1097 struct mana_cfg_rx_steer_req_v2 *req; 1098 struct mana_cfg_rx_steer_resp resp = {}; 1099 struct net_device *ndev = apc->ndev; 1100 u32 req_buf_size; 1101 int err; 1102 1103 req_buf_size = struct_size(req, indir_tab, apc->indir_table_sz); 1104 req = kzalloc(req_buf_size, GFP_KERNEL); 1105 if (!req) 1106 return -ENOMEM; 1107 1108 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size, 1109 sizeof(resp)); 1110 1111 req->hdr.req.msg_version = GDMA_MESSAGE_V2; 1112 1113 req->vport = apc->port_handle; 1114 req->num_indir_entries = apc->indir_table_sz; 1115 req->indir_tab_offset = offsetof(struct mana_cfg_rx_steer_req_v2, 1116 indir_tab); 1117 req->rx_enable = rx; 1118 req->rss_enable = apc->rss_state; 1119 req->update_default_rxobj = update_default_rxobj; 1120 req->update_hashkey = update_key; 1121 req->update_indir_tab = update_tab; 1122 req->default_rxobj = apc->default_rxobj; 1123 req->cqe_coalescing_enable = 0; 1124 1125 if (update_key) 1126 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE); 1127 1128 if (update_tab) 1129 memcpy(req->indir_tab, apc->rxobj_table, 1130 flex_array_size(req, indir_tab, req->num_indir_entries)); 1131 1132 err = mana_send_request(apc->ac, req, req_buf_size, &resp, 1133 sizeof(resp)); 1134 if (err) { 1135 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err); 1136 goto out; 1137 } 1138 1139 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX, 1140 sizeof(resp)); 1141 if (err) { 1142 netdev_err(ndev, "vPort RX configuration failed: %d\n", err); 1143 goto out; 1144 } 1145 1146 if (resp.hdr.status) { 1147 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n", 1148 resp.hdr.status); 1149 err = -EPROTO; 1150 } 1151 1152 netdev_info(ndev, "Configured steering vPort %llu entries %u\n", 1153 apc->port_handle, apc->indir_table_sz); 1154 out: 1155 kfree(req); 1156 return err; 1157 } 1158 1159 int mana_create_wq_obj(struct mana_port_context *apc, 1160 mana_handle_t vport, 1161 u32 wq_type, struct mana_obj_spec *wq_spec, 1162 struct mana_obj_spec *cq_spec, 1163 mana_handle_t *wq_obj) 1164 { 1165 struct mana_create_wqobj_resp resp = {}; 1166 struct mana_create_wqobj_req req = {}; 1167 struct net_device *ndev = apc->ndev; 1168 int err; 1169 1170 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ, 1171 sizeof(req), sizeof(resp)); 1172 req.vport = vport; 1173 req.wq_type = wq_type; 1174 req.wq_gdma_region = wq_spec->gdma_region; 1175 req.cq_gdma_region = cq_spec->gdma_region; 1176 req.wq_size = wq_spec->queue_size; 1177 req.cq_size = cq_spec->queue_size; 1178 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id; 1179 req.cq_parent_qid = cq_spec->attached_eq; 1180 1181 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1182 sizeof(resp)); 1183 if (err) { 1184 netdev_err(ndev, "Failed to create WQ object: %d\n", err); 1185 goto out; 1186 } 1187 1188 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ, 1189 sizeof(resp)); 1190 if (err || resp.hdr.status) { 1191 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err, 1192 resp.hdr.status); 1193 if (!err) 1194 err = -EPROTO; 1195 goto out; 1196 } 1197 1198 if (resp.wq_obj == INVALID_MANA_HANDLE) { 1199 netdev_err(ndev, "Got an invalid WQ object handle\n"); 1200 err = -EPROTO; 1201 goto out; 1202 } 1203 1204 *wq_obj = resp.wq_obj; 1205 wq_spec->queue_index = resp.wq_id; 1206 cq_spec->queue_index = resp.cq_id; 1207 1208 return 0; 1209 out: 1210 return err; 1211 } 1212 EXPORT_SYMBOL_NS(mana_create_wq_obj, "NET_MANA"); 1213 1214 void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type, 1215 mana_handle_t wq_obj) 1216 { 1217 struct mana_destroy_wqobj_resp resp = {}; 1218 struct mana_destroy_wqobj_req req = {}; 1219 struct net_device *ndev = apc->ndev; 1220 int err; 1221 1222 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ, 1223 sizeof(req), sizeof(resp)); 1224 req.wq_type = wq_type; 1225 req.wq_obj_handle = wq_obj; 1226 1227 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1228 sizeof(resp)); 1229 if (err) { 1230 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err); 1231 return; 1232 } 1233 1234 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ, 1235 sizeof(resp)); 1236 if (err || resp.hdr.status) 1237 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err, 1238 resp.hdr.status); 1239 } 1240 EXPORT_SYMBOL_NS(mana_destroy_wq_obj, "NET_MANA"); 1241 1242 static void mana_destroy_eq(struct mana_context *ac) 1243 { 1244 struct gdma_context *gc = ac->gdma_dev->gdma_context; 1245 struct gdma_queue *eq; 1246 int i; 1247 1248 if (!ac->eqs) 1249 return; 1250 1251 debugfs_remove_recursive(ac->mana_eqs_debugfs); 1252 ac->mana_eqs_debugfs = NULL; 1253 1254 for (i = 0; i < gc->max_num_queues; i++) { 1255 eq = ac->eqs[i].eq; 1256 if (!eq) 1257 continue; 1258 1259 mana_gd_destroy_queue(gc, eq); 1260 } 1261 1262 kfree(ac->eqs); 1263 ac->eqs = NULL; 1264 } 1265 1266 static void mana_create_eq_debugfs(struct mana_context *ac, int i) 1267 { 1268 struct mana_eq eq = ac->eqs[i]; 1269 char eqnum[32]; 1270 1271 sprintf(eqnum, "eq%d", i); 1272 eq.mana_eq_debugfs = debugfs_create_dir(eqnum, ac->mana_eqs_debugfs); 1273 debugfs_create_u32("head", 0400, eq.mana_eq_debugfs, &eq.eq->head); 1274 debugfs_create_u32("tail", 0400, eq.mana_eq_debugfs, &eq.eq->tail); 1275 debugfs_create_file("eq_dump", 0400, eq.mana_eq_debugfs, eq.eq, &mana_dbg_q_fops); 1276 } 1277 1278 static int mana_create_eq(struct mana_context *ac) 1279 { 1280 struct gdma_dev *gd = ac->gdma_dev; 1281 struct gdma_context *gc = gd->gdma_context; 1282 struct gdma_queue_spec spec = {}; 1283 int err; 1284 int i; 1285 1286 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq), 1287 GFP_KERNEL); 1288 if (!ac->eqs) 1289 return -ENOMEM; 1290 1291 spec.type = GDMA_EQ; 1292 spec.monitor_avl_buf = false; 1293 spec.queue_size = EQ_SIZE; 1294 spec.eq.callback = NULL; 1295 spec.eq.context = ac->eqs; 1296 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE; 1297 1298 ac->mana_eqs_debugfs = debugfs_create_dir("EQs", gc->mana_pci_debugfs); 1299 1300 for (i = 0; i < gc->max_num_queues; i++) { 1301 spec.eq.msix_index = (i + 1) % gc->num_msix_usable; 1302 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq); 1303 if (err) { 1304 dev_err(gc->dev, "Failed to create EQ %d : %d\n", i, err); 1305 goto out; 1306 } 1307 mana_create_eq_debugfs(ac, i); 1308 } 1309 1310 return 0; 1311 out: 1312 mana_destroy_eq(ac); 1313 return err; 1314 } 1315 1316 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq) 1317 { 1318 struct mana_fence_rq_resp resp = {}; 1319 struct mana_fence_rq_req req = {}; 1320 int err; 1321 1322 init_completion(&rxq->fence_event); 1323 1324 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ, 1325 sizeof(req), sizeof(resp)); 1326 req.wq_obj_handle = rxq->rxobj; 1327 1328 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1329 sizeof(resp)); 1330 if (err) { 1331 netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n", 1332 rxq->rxq_idx, err); 1333 return err; 1334 } 1335 1336 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp)); 1337 if (err || resp.hdr.status) { 1338 netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n", 1339 rxq->rxq_idx, err, resp.hdr.status); 1340 if (!err) 1341 err = -EPROTO; 1342 1343 return err; 1344 } 1345 1346 if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) { 1347 netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n", 1348 rxq->rxq_idx); 1349 return -ETIMEDOUT; 1350 } 1351 1352 return 0; 1353 } 1354 1355 static void mana_fence_rqs(struct mana_port_context *apc) 1356 { 1357 unsigned int rxq_idx; 1358 struct mana_rxq *rxq; 1359 int err; 1360 1361 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 1362 rxq = apc->rxqs[rxq_idx]; 1363 err = mana_fence_rq(apc, rxq); 1364 1365 /* In case of any error, use sleep instead. */ 1366 if (err) 1367 msleep(100); 1368 } 1369 } 1370 1371 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units) 1372 { 1373 u32 used_space_old; 1374 u32 used_space_new; 1375 1376 used_space_old = wq->head - wq->tail; 1377 used_space_new = wq->head - (wq->tail + num_units); 1378 1379 if (WARN_ON_ONCE(used_space_new > used_space_old)) 1380 return -ERANGE; 1381 1382 wq->tail += num_units; 1383 return 0; 1384 } 1385 1386 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc) 1387 { 1388 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 1389 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 1390 struct device *dev = gc->dev; 1391 int hsg, i; 1392 1393 /* Number of SGEs of linear part */ 1394 hsg = (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0]) ? 2 : 1; 1395 1396 for (i = 0; i < hsg; i++) 1397 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i], 1398 DMA_TO_DEVICE); 1399 1400 for (i = hsg; i < skb_shinfo(skb)->nr_frags + hsg; i++) 1401 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i], 1402 DMA_TO_DEVICE); 1403 } 1404 1405 static void mana_poll_tx_cq(struct mana_cq *cq) 1406 { 1407 struct gdma_comp *completions = cq->gdma_comp_buf; 1408 struct gdma_posted_wqe_info *wqe_info; 1409 unsigned int pkt_transmitted = 0; 1410 unsigned int wqe_unit_cnt = 0; 1411 struct mana_txq *txq = cq->txq; 1412 struct mana_port_context *apc; 1413 struct netdev_queue *net_txq; 1414 struct gdma_queue *gdma_wq; 1415 unsigned int avail_space; 1416 struct net_device *ndev; 1417 struct sk_buff *skb; 1418 bool txq_stopped; 1419 int comp_read; 1420 int i; 1421 1422 ndev = txq->ndev; 1423 apc = netdev_priv(ndev); 1424 1425 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions, 1426 CQE_POLLING_BUFFER); 1427 1428 if (comp_read < 1) 1429 return; 1430 1431 for (i = 0; i < comp_read; i++) { 1432 struct mana_tx_comp_oob *cqe_oob; 1433 1434 if (WARN_ON_ONCE(!completions[i].is_sq)) 1435 return; 1436 1437 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data; 1438 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type != 1439 MANA_CQE_COMPLETION)) 1440 return; 1441 1442 switch (cqe_oob->cqe_hdr.cqe_type) { 1443 case CQE_TX_OKAY: 1444 break; 1445 1446 case CQE_TX_SA_DROP: 1447 case CQE_TX_MTU_DROP: 1448 case CQE_TX_INVALID_OOB: 1449 case CQE_TX_INVALID_ETH_TYPE: 1450 case CQE_TX_HDR_PROCESSING_ERROR: 1451 case CQE_TX_VF_DISABLED: 1452 case CQE_TX_VPORT_IDX_OUT_OF_RANGE: 1453 case CQE_TX_VPORT_DISABLED: 1454 case CQE_TX_VLAN_TAGGING_VIOLATION: 1455 if (net_ratelimit()) 1456 netdev_err(ndev, "TX: CQE error %d\n", 1457 cqe_oob->cqe_hdr.cqe_type); 1458 1459 apc->eth_stats.tx_cqe_err++; 1460 break; 1461 1462 default: 1463 /* If the CQE type is unknown, log an error, 1464 * and still free the SKB, update tail, etc. 1465 */ 1466 if (net_ratelimit()) 1467 netdev_err(ndev, "TX: unknown CQE type %d\n", 1468 cqe_oob->cqe_hdr.cqe_type); 1469 1470 apc->eth_stats.tx_cqe_unknown_type++; 1471 break; 1472 } 1473 1474 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num)) 1475 return; 1476 1477 skb = skb_dequeue(&txq->pending_skbs); 1478 if (WARN_ON_ONCE(!skb)) 1479 return; 1480 1481 wqe_info = (struct gdma_posted_wqe_info *)skb->cb; 1482 wqe_unit_cnt += wqe_info->wqe_size_in_bu; 1483 1484 mana_unmap_skb(skb, apc); 1485 1486 napi_consume_skb(skb, cq->budget); 1487 1488 pkt_transmitted++; 1489 } 1490 1491 if (WARN_ON_ONCE(wqe_unit_cnt == 0)) 1492 return; 1493 1494 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt); 1495 1496 gdma_wq = txq->gdma_sq; 1497 avail_space = mana_gd_wq_avail_space(gdma_wq); 1498 1499 /* Ensure tail updated before checking q stop */ 1500 smp_mb(); 1501 1502 net_txq = txq->net_txq; 1503 txq_stopped = netif_tx_queue_stopped(net_txq); 1504 1505 /* Ensure checking txq_stopped before apc->port_is_up. */ 1506 smp_rmb(); 1507 1508 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) { 1509 netif_tx_wake_queue(net_txq); 1510 apc->eth_stats.wake_queue++; 1511 } 1512 1513 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0) 1514 WARN_ON_ONCE(1); 1515 1516 cq->work_done = pkt_transmitted; 1517 } 1518 1519 static void mana_post_pkt_rxq(struct mana_rxq *rxq) 1520 { 1521 struct mana_recv_buf_oob *recv_buf_oob; 1522 u32 curr_index; 1523 int err; 1524 1525 curr_index = rxq->buf_index++; 1526 if (rxq->buf_index == rxq->num_rx_buf) 1527 rxq->buf_index = 0; 1528 1529 recv_buf_oob = &rxq->rx_oobs[curr_index]; 1530 1531 err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req, 1532 &recv_buf_oob->wqe_inf); 1533 if (WARN_ON_ONCE(err)) 1534 return; 1535 1536 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1); 1537 } 1538 1539 static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va, 1540 uint pkt_len, struct xdp_buff *xdp) 1541 { 1542 struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size); 1543 1544 if (!skb) 1545 return NULL; 1546 1547 if (xdp->data_hard_start) { 1548 u32 metasize = xdp->data - xdp->data_meta; 1549 1550 skb_reserve(skb, xdp->data - xdp->data_hard_start); 1551 skb_put(skb, xdp->data_end - xdp->data); 1552 if (metasize) 1553 skb_metadata_set(skb, metasize); 1554 return skb; 1555 } 1556 1557 skb_reserve(skb, rxq->headroom); 1558 skb_put(skb, pkt_len); 1559 1560 return skb; 1561 } 1562 1563 static void mana_rx_skb(void *buf_va, bool from_pool, 1564 struct mana_rxcomp_oob *cqe, struct mana_rxq *rxq) 1565 { 1566 struct mana_stats_rx *rx_stats = &rxq->stats; 1567 struct net_device *ndev = rxq->ndev; 1568 uint pkt_len = cqe->ppi[0].pkt_len; 1569 u16 rxq_idx = rxq->rxq_idx; 1570 struct napi_struct *napi; 1571 struct xdp_buff xdp = {}; 1572 struct sk_buff *skb; 1573 u32 hash_value; 1574 u32 act; 1575 1576 rxq->rx_cq.work_done++; 1577 napi = &rxq->rx_cq.napi; 1578 1579 if (!buf_va) { 1580 ++ndev->stats.rx_dropped; 1581 return; 1582 } 1583 1584 act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len); 1585 1586 if (act == XDP_REDIRECT && !rxq->xdp_rc) 1587 return; 1588 1589 if (act != XDP_PASS && act != XDP_TX) 1590 goto drop_xdp; 1591 1592 skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp); 1593 1594 if (!skb) 1595 goto drop; 1596 1597 if (from_pool) 1598 skb_mark_for_recycle(skb); 1599 1600 skb->dev = napi->dev; 1601 1602 skb->protocol = eth_type_trans(skb, ndev); 1603 skb_checksum_none_assert(skb); 1604 skb_record_rx_queue(skb, rxq_idx); 1605 1606 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) { 1607 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed) 1608 skb->ip_summed = CHECKSUM_UNNECESSARY; 1609 } 1610 1611 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) { 1612 hash_value = cqe->ppi[0].pkt_hash; 1613 1614 if (cqe->rx_hashtype & MANA_HASH_L4) 1615 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4); 1616 else 1617 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3); 1618 } 1619 1620 if (cqe->rx_vlantag_present) { 1621 u16 vlan_tci = cqe->rx_vlan_id; 1622 1623 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci); 1624 } 1625 1626 u64_stats_update_begin(&rx_stats->syncp); 1627 rx_stats->packets++; 1628 rx_stats->bytes += pkt_len; 1629 1630 if (act == XDP_TX) 1631 rx_stats->xdp_tx++; 1632 u64_stats_update_end(&rx_stats->syncp); 1633 1634 if (act == XDP_TX) { 1635 skb_set_queue_mapping(skb, rxq_idx); 1636 mana_xdp_tx(skb, ndev); 1637 return; 1638 } 1639 1640 napi_gro_receive(napi, skb); 1641 1642 return; 1643 1644 drop_xdp: 1645 u64_stats_update_begin(&rx_stats->syncp); 1646 rx_stats->xdp_drop++; 1647 u64_stats_update_end(&rx_stats->syncp); 1648 1649 drop: 1650 if (from_pool) { 1651 page_pool_recycle_direct(rxq->page_pool, 1652 virt_to_head_page(buf_va)); 1653 } else { 1654 WARN_ON_ONCE(rxq->xdp_save_va); 1655 /* Save for reuse */ 1656 rxq->xdp_save_va = buf_va; 1657 } 1658 1659 ++ndev->stats.rx_dropped; 1660 1661 return; 1662 } 1663 1664 static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev, 1665 dma_addr_t *da, bool *from_pool) 1666 { 1667 struct page *page; 1668 void *va; 1669 1670 *from_pool = false; 1671 1672 /* Reuse XDP dropped page if available */ 1673 if (rxq->xdp_save_va) { 1674 va = rxq->xdp_save_va; 1675 rxq->xdp_save_va = NULL; 1676 } else { 1677 page = page_pool_dev_alloc_pages(rxq->page_pool); 1678 if (!page) 1679 return NULL; 1680 1681 *from_pool = true; 1682 va = page_to_virt(page); 1683 } 1684 1685 *da = dma_map_single(dev, va + rxq->headroom, rxq->datasize, 1686 DMA_FROM_DEVICE); 1687 if (dma_mapping_error(dev, *da)) { 1688 if (*from_pool) 1689 page_pool_put_full_page(rxq->page_pool, page, false); 1690 else 1691 put_page(virt_to_head_page(va)); 1692 1693 return NULL; 1694 } 1695 1696 return va; 1697 } 1698 1699 /* Allocate frag for rx buffer, and save the old buf */ 1700 static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq, 1701 struct mana_recv_buf_oob *rxoob, void **old_buf, 1702 bool *old_fp) 1703 { 1704 bool from_pool; 1705 dma_addr_t da; 1706 void *va; 1707 1708 va = mana_get_rxfrag(rxq, dev, &da, &from_pool); 1709 if (!va) 1710 return; 1711 1712 dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize, 1713 DMA_FROM_DEVICE); 1714 *old_buf = rxoob->buf_va; 1715 *old_fp = rxoob->from_pool; 1716 1717 rxoob->buf_va = va; 1718 rxoob->sgl[0].address = da; 1719 rxoob->from_pool = from_pool; 1720 } 1721 1722 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq, 1723 struct gdma_comp *cqe) 1724 { 1725 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data; 1726 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context; 1727 struct net_device *ndev = rxq->ndev; 1728 struct mana_recv_buf_oob *rxbuf_oob; 1729 struct mana_port_context *apc; 1730 struct device *dev = gc->dev; 1731 void *old_buf = NULL; 1732 u32 curr, pktlen; 1733 bool old_fp; 1734 1735 apc = netdev_priv(ndev); 1736 1737 switch (oob->cqe_hdr.cqe_type) { 1738 case CQE_RX_OKAY: 1739 break; 1740 1741 case CQE_RX_TRUNCATED: 1742 ++ndev->stats.rx_dropped; 1743 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index]; 1744 netdev_warn_once(ndev, "Dropped a truncated packet\n"); 1745 goto drop; 1746 1747 case CQE_RX_COALESCED_4: 1748 netdev_err(ndev, "RX coalescing is unsupported\n"); 1749 apc->eth_stats.rx_coalesced_err++; 1750 return; 1751 1752 case CQE_RX_OBJECT_FENCE: 1753 complete(&rxq->fence_event); 1754 return; 1755 1756 default: 1757 netdev_err(ndev, "Unknown RX CQE type = %d\n", 1758 oob->cqe_hdr.cqe_type); 1759 apc->eth_stats.rx_cqe_unknown_type++; 1760 return; 1761 } 1762 1763 pktlen = oob->ppi[0].pkt_len; 1764 1765 if (pktlen == 0) { 1766 /* data packets should never have packetlength of zero */ 1767 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n", 1768 rxq->gdma_id, cq->gdma_id, rxq->rxobj); 1769 return; 1770 } 1771 1772 curr = rxq->buf_index; 1773 rxbuf_oob = &rxq->rx_oobs[curr]; 1774 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1); 1775 1776 mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf, &old_fp); 1777 1778 /* Unsuccessful refill will have old_buf == NULL. 1779 * In this case, mana_rx_skb() will drop the packet. 1780 */ 1781 mana_rx_skb(old_buf, old_fp, oob, rxq); 1782 1783 drop: 1784 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu); 1785 1786 mana_post_pkt_rxq(rxq); 1787 } 1788 1789 static void mana_poll_rx_cq(struct mana_cq *cq) 1790 { 1791 struct gdma_comp *comp = cq->gdma_comp_buf; 1792 struct mana_rxq *rxq = cq->rxq; 1793 int comp_read, i; 1794 1795 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER); 1796 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER); 1797 1798 rxq->xdp_flush = false; 1799 1800 for (i = 0; i < comp_read; i++) { 1801 if (WARN_ON_ONCE(comp[i].is_sq)) 1802 return; 1803 1804 /* verify recv cqe references the right rxq */ 1805 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id)) 1806 return; 1807 1808 mana_process_rx_cqe(rxq, cq, &comp[i]); 1809 } 1810 1811 if (comp_read > 0) { 1812 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context; 1813 1814 mana_gd_wq_ring_doorbell(gc, rxq->gdma_rq); 1815 } 1816 1817 if (rxq->xdp_flush) 1818 xdp_do_flush(); 1819 } 1820 1821 static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue) 1822 { 1823 struct mana_cq *cq = context; 1824 int w; 1825 1826 WARN_ON_ONCE(cq->gdma_cq != gdma_queue); 1827 1828 if (cq->type == MANA_CQ_TYPE_RX) 1829 mana_poll_rx_cq(cq); 1830 else 1831 mana_poll_tx_cq(cq); 1832 1833 w = cq->work_done; 1834 cq->work_done_since_doorbell += w; 1835 1836 if (w < cq->budget) { 1837 mana_gd_ring_cq(gdma_queue, SET_ARM_BIT); 1838 cq->work_done_since_doorbell = 0; 1839 napi_complete_done(&cq->napi, w); 1840 } else if (cq->work_done_since_doorbell > 1841 cq->gdma_cq->queue_size / COMP_ENTRY_SIZE * 4) { 1842 /* MANA hardware requires at least one doorbell ring every 8 1843 * wraparounds of CQ even if there is no need to arm the CQ. 1844 * This driver rings the doorbell as soon as we have exceeded 1845 * 4 wraparounds. 1846 */ 1847 mana_gd_ring_cq(gdma_queue, 0); 1848 cq->work_done_since_doorbell = 0; 1849 } 1850 1851 return w; 1852 } 1853 1854 static int mana_poll(struct napi_struct *napi, int budget) 1855 { 1856 struct mana_cq *cq = container_of(napi, struct mana_cq, napi); 1857 int w; 1858 1859 cq->work_done = 0; 1860 cq->budget = budget; 1861 1862 w = mana_cq_handler(cq, cq->gdma_cq); 1863 1864 return min(w, budget); 1865 } 1866 1867 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue) 1868 { 1869 struct mana_cq *cq = context; 1870 1871 napi_schedule_irqoff(&cq->napi); 1872 } 1873 1874 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq) 1875 { 1876 struct gdma_dev *gd = apc->ac->gdma_dev; 1877 1878 if (!cq->gdma_cq) 1879 return; 1880 1881 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq); 1882 } 1883 1884 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq) 1885 { 1886 struct gdma_dev *gd = apc->ac->gdma_dev; 1887 1888 if (!txq->gdma_sq) 1889 return; 1890 1891 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq); 1892 } 1893 1894 static void mana_destroy_txq(struct mana_port_context *apc) 1895 { 1896 struct napi_struct *napi; 1897 int i; 1898 1899 if (!apc->tx_qp) 1900 return; 1901 1902 for (i = 0; i < apc->num_queues; i++) { 1903 debugfs_remove_recursive(apc->tx_qp[i].mana_tx_debugfs); 1904 apc->tx_qp[i].mana_tx_debugfs = NULL; 1905 1906 napi = &apc->tx_qp[i].tx_cq.napi; 1907 if (apc->tx_qp[i].txq.napi_initialized) { 1908 napi_synchronize(napi); 1909 napi_disable(napi); 1910 netif_napi_del(napi); 1911 apc->tx_qp[i].txq.napi_initialized = false; 1912 } 1913 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object); 1914 1915 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq); 1916 1917 mana_deinit_txq(apc, &apc->tx_qp[i].txq); 1918 } 1919 1920 kfree(apc->tx_qp); 1921 apc->tx_qp = NULL; 1922 } 1923 1924 static void mana_create_txq_debugfs(struct mana_port_context *apc, int idx) 1925 { 1926 struct mana_tx_qp *tx_qp = &apc->tx_qp[idx]; 1927 char qnum[32]; 1928 1929 sprintf(qnum, "TX-%d", idx); 1930 tx_qp->mana_tx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs); 1931 debugfs_create_u32("sq_head", 0400, tx_qp->mana_tx_debugfs, 1932 &tx_qp->txq.gdma_sq->head); 1933 debugfs_create_u32("sq_tail", 0400, tx_qp->mana_tx_debugfs, 1934 &tx_qp->txq.gdma_sq->tail); 1935 debugfs_create_u32("sq_pend_skb_qlen", 0400, tx_qp->mana_tx_debugfs, 1936 &tx_qp->txq.pending_skbs.qlen); 1937 debugfs_create_u32("cq_head", 0400, tx_qp->mana_tx_debugfs, 1938 &tx_qp->tx_cq.gdma_cq->head); 1939 debugfs_create_u32("cq_tail", 0400, tx_qp->mana_tx_debugfs, 1940 &tx_qp->tx_cq.gdma_cq->tail); 1941 debugfs_create_u32("cq_budget", 0400, tx_qp->mana_tx_debugfs, 1942 &tx_qp->tx_cq.budget); 1943 debugfs_create_file("txq_dump", 0400, tx_qp->mana_tx_debugfs, 1944 tx_qp->txq.gdma_sq, &mana_dbg_q_fops); 1945 debugfs_create_file("cq_dump", 0400, tx_qp->mana_tx_debugfs, 1946 tx_qp->tx_cq.gdma_cq, &mana_dbg_q_fops); 1947 } 1948 1949 static int mana_create_txq(struct mana_port_context *apc, 1950 struct net_device *net) 1951 { 1952 struct mana_context *ac = apc->ac; 1953 struct gdma_dev *gd = ac->gdma_dev; 1954 struct mana_obj_spec wq_spec; 1955 struct mana_obj_spec cq_spec; 1956 struct gdma_queue_spec spec; 1957 struct gdma_context *gc; 1958 struct mana_txq *txq; 1959 struct mana_cq *cq; 1960 u32 txq_size; 1961 u32 cq_size; 1962 int err; 1963 int i; 1964 1965 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp), 1966 GFP_KERNEL); 1967 if (!apc->tx_qp) 1968 return -ENOMEM; 1969 1970 /* The minimum size of the WQE is 32 bytes, hence 1971 * apc->tx_queue_size represents the maximum number of WQEs 1972 * the SQ can store. This value is then used to size other queues 1973 * to prevent overflow. 1974 * Also note that the txq_size is always going to be MANA_PAGE_ALIGNED, 1975 * as min val of apc->tx_queue_size is 128 and that would make 1976 * txq_size 128*32 = 4096 and the other higher values of apc->tx_queue_size 1977 * are always power of two 1978 */ 1979 txq_size = apc->tx_queue_size * 32; 1980 1981 cq_size = apc->tx_queue_size * COMP_ENTRY_SIZE; 1982 1983 gc = gd->gdma_context; 1984 1985 for (i = 0; i < apc->num_queues; i++) { 1986 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE; 1987 1988 /* Create SQ */ 1989 txq = &apc->tx_qp[i].txq; 1990 1991 u64_stats_init(&txq->stats.syncp); 1992 txq->ndev = net; 1993 txq->net_txq = netdev_get_tx_queue(net, i); 1994 txq->vp_offset = apc->tx_vp_offset; 1995 txq->napi_initialized = false; 1996 skb_queue_head_init(&txq->pending_skbs); 1997 1998 memset(&spec, 0, sizeof(spec)); 1999 spec.type = GDMA_SQ; 2000 spec.monitor_avl_buf = true; 2001 spec.queue_size = txq_size; 2002 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq); 2003 if (err) 2004 goto out; 2005 2006 /* Create SQ's CQ */ 2007 cq = &apc->tx_qp[i].tx_cq; 2008 cq->type = MANA_CQ_TYPE_TX; 2009 2010 cq->txq = txq; 2011 2012 memset(&spec, 0, sizeof(spec)); 2013 spec.type = GDMA_CQ; 2014 spec.monitor_avl_buf = false; 2015 spec.queue_size = cq_size; 2016 spec.cq.callback = mana_schedule_napi; 2017 spec.cq.parent_eq = ac->eqs[i].eq; 2018 spec.cq.context = cq; 2019 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 2020 if (err) 2021 goto out; 2022 2023 memset(&wq_spec, 0, sizeof(wq_spec)); 2024 memset(&cq_spec, 0, sizeof(cq_spec)); 2025 2026 wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle; 2027 wq_spec.queue_size = txq->gdma_sq->queue_size; 2028 2029 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 2030 cq_spec.queue_size = cq->gdma_cq->queue_size; 2031 cq_spec.modr_ctx_id = 0; 2032 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 2033 2034 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ, 2035 &wq_spec, &cq_spec, 2036 &apc->tx_qp[i].tx_object); 2037 2038 if (err) 2039 goto out; 2040 2041 txq->gdma_sq->id = wq_spec.queue_index; 2042 cq->gdma_cq->id = cq_spec.queue_index; 2043 2044 txq->gdma_sq->mem_info.dma_region_handle = 2045 GDMA_INVALID_DMA_REGION; 2046 cq->gdma_cq->mem_info.dma_region_handle = 2047 GDMA_INVALID_DMA_REGION; 2048 2049 txq->gdma_txq_id = txq->gdma_sq->id; 2050 2051 cq->gdma_id = cq->gdma_cq->id; 2052 2053 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 2054 err = -EINVAL; 2055 goto out; 2056 } 2057 2058 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2059 2060 mana_create_txq_debugfs(apc, i); 2061 2062 netif_napi_add_tx(net, &cq->napi, mana_poll); 2063 napi_enable(&cq->napi); 2064 txq->napi_initialized = true; 2065 2066 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2067 } 2068 2069 return 0; 2070 out: 2071 netdev_err(net, "Failed to create %d TX queues, %d\n", 2072 apc->num_queues, err); 2073 mana_destroy_txq(apc); 2074 return err; 2075 } 2076 2077 static void mana_destroy_rxq(struct mana_port_context *apc, 2078 struct mana_rxq *rxq, bool napi_initialized) 2079 2080 { 2081 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 2082 struct mana_recv_buf_oob *rx_oob; 2083 struct device *dev = gc->dev; 2084 struct napi_struct *napi; 2085 struct page *page; 2086 int i; 2087 2088 if (!rxq) 2089 return; 2090 2091 debugfs_remove_recursive(rxq->mana_rx_debugfs); 2092 rxq->mana_rx_debugfs = NULL; 2093 2094 napi = &rxq->rx_cq.napi; 2095 2096 if (napi_initialized) { 2097 napi_synchronize(napi); 2098 2099 napi_disable(napi); 2100 2101 netif_napi_del(napi); 2102 } 2103 xdp_rxq_info_unreg(&rxq->xdp_rxq); 2104 2105 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj); 2106 2107 mana_deinit_cq(apc, &rxq->rx_cq); 2108 2109 if (rxq->xdp_save_va) 2110 put_page(virt_to_head_page(rxq->xdp_save_va)); 2111 2112 for (i = 0; i < rxq->num_rx_buf; i++) { 2113 rx_oob = &rxq->rx_oobs[i]; 2114 2115 if (!rx_oob->buf_va) 2116 continue; 2117 2118 dma_unmap_single(dev, rx_oob->sgl[0].address, 2119 rx_oob->sgl[0].size, DMA_FROM_DEVICE); 2120 2121 page = virt_to_head_page(rx_oob->buf_va); 2122 2123 if (rx_oob->from_pool) 2124 page_pool_put_full_page(rxq->page_pool, page, false); 2125 else 2126 put_page(page); 2127 2128 rx_oob->buf_va = NULL; 2129 } 2130 2131 page_pool_destroy(rxq->page_pool); 2132 2133 if (rxq->gdma_rq) 2134 mana_gd_destroy_queue(gc, rxq->gdma_rq); 2135 2136 kfree(rxq); 2137 } 2138 2139 static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key, 2140 struct mana_rxq *rxq, struct device *dev) 2141 { 2142 struct mana_port_context *mpc = netdev_priv(rxq->ndev); 2143 bool from_pool = false; 2144 dma_addr_t da; 2145 void *va; 2146 2147 if (mpc->rxbufs_pre) 2148 va = mana_get_rxbuf_pre(rxq, &da); 2149 else 2150 va = mana_get_rxfrag(rxq, dev, &da, &from_pool); 2151 2152 if (!va) 2153 return -ENOMEM; 2154 2155 rx_oob->buf_va = va; 2156 rx_oob->from_pool = from_pool; 2157 2158 rx_oob->sgl[0].address = da; 2159 rx_oob->sgl[0].size = rxq->datasize; 2160 rx_oob->sgl[0].mem_key = mem_key; 2161 2162 return 0; 2163 } 2164 2165 #define MANA_WQE_HEADER_SIZE 16 2166 #define MANA_WQE_SGE_SIZE 16 2167 2168 static int mana_alloc_rx_wqe(struct mana_port_context *apc, 2169 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size) 2170 { 2171 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 2172 struct mana_recv_buf_oob *rx_oob; 2173 struct device *dev = gc->dev; 2174 u32 buf_idx; 2175 int ret; 2176 2177 WARN_ON(rxq->datasize == 0); 2178 2179 *rxq_size = 0; 2180 *cq_size = 0; 2181 2182 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2183 rx_oob = &rxq->rx_oobs[buf_idx]; 2184 memset(rx_oob, 0, sizeof(*rx_oob)); 2185 2186 rx_oob->num_sge = 1; 2187 2188 ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq, 2189 dev); 2190 if (ret) 2191 return ret; 2192 2193 rx_oob->wqe_req.sgl = rx_oob->sgl; 2194 rx_oob->wqe_req.num_sge = rx_oob->num_sge; 2195 rx_oob->wqe_req.inline_oob_size = 0; 2196 rx_oob->wqe_req.inline_oob_data = NULL; 2197 rx_oob->wqe_req.flags = 0; 2198 rx_oob->wqe_req.client_data_unit = 0; 2199 2200 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE + 2201 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32); 2202 *cq_size += COMP_ENTRY_SIZE; 2203 } 2204 2205 return 0; 2206 } 2207 2208 static int mana_push_wqe(struct mana_rxq *rxq) 2209 { 2210 struct mana_recv_buf_oob *rx_oob; 2211 u32 buf_idx; 2212 int err; 2213 2214 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2215 rx_oob = &rxq->rx_oobs[buf_idx]; 2216 2217 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req, 2218 &rx_oob->wqe_inf); 2219 if (err) 2220 return -ENOSPC; 2221 } 2222 2223 return 0; 2224 } 2225 2226 static int mana_create_page_pool(struct mana_rxq *rxq, struct gdma_context *gc) 2227 { 2228 struct mana_port_context *mpc = netdev_priv(rxq->ndev); 2229 struct page_pool_params pprm = {}; 2230 int ret; 2231 2232 pprm.pool_size = mpc->rx_queue_size; 2233 pprm.nid = gc->numa_node; 2234 pprm.napi = &rxq->rx_cq.napi; 2235 pprm.netdev = rxq->ndev; 2236 pprm.order = get_order(rxq->alloc_size); 2237 2238 rxq->page_pool = page_pool_create(&pprm); 2239 2240 if (IS_ERR(rxq->page_pool)) { 2241 ret = PTR_ERR(rxq->page_pool); 2242 rxq->page_pool = NULL; 2243 return ret; 2244 } 2245 2246 return 0; 2247 } 2248 2249 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc, 2250 u32 rxq_idx, struct mana_eq *eq, 2251 struct net_device *ndev) 2252 { 2253 struct gdma_dev *gd = apc->ac->gdma_dev; 2254 struct mana_obj_spec wq_spec; 2255 struct mana_obj_spec cq_spec; 2256 struct gdma_queue_spec spec; 2257 struct mana_cq *cq = NULL; 2258 struct gdma_context *gc; 2259 u32 cq_size, rq_size; 2260 struct mana_rxq *rxq; 2261 int err; 2262 2263 gc = gd->gdma_context; 2264 2265 rxq = kzalloc(struct_size(rxq, rx_oobs, apc->rx_queue_size), 2266 GFP_KERNEL); 2267 if (!rxq) 2268 return NULL; 2269 2270 rxq->ndev = ndev; 2271 rxq->num_rx_buf = apc->rx_queue_size; 2272 rxq->rxq_idx = rxq_idx; 2273 rxq->rxobj = INVALID_MANA_HANDLE; 2274 2275 mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size, 2276 &rxq->headroom); 2277 2278 /* Create page pool for RX queue */ 2279 err = mana_create_page_pool(rxq, gc); 2280 if (err) { 2281 netdev_err(ndev, "Create page pool err:%d\n", err); 2282 goto out; 2283 } 2284 2285 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size); 2286 if (err) 2287 goto out; 2288 2289 rq_size = MANA_PAGE_ALIGN(rq_size); 2290 cq_size = MANA_PAGE_ALIGN(cq_size); 2291 2292 /* Create RQ */ 2293 memset(&spec, 0, sizeof(spec)); 2294 spec.type = GDMA_RQ; 2295 spec.monitor_avl_buf = true; 2296 spec.queue_size = rq_size; 2297 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq); 2298 if (err) 2299 goto out; 2300 2301 /* Create RQ's CQ */ 2302 cq = &rxq->rx_cq; 2303 cq->type = MANA_CQ_TYPE_RX; 2304 cq->rxq = rxq; 2305 2306 memset(&spec, 0, sizeof(spec)); 2307 spec.type = GDMA_CQ; 2308 spec.monitor_avl_buf = false; 2309 spec.queue_size = cq_size; 2310 spec.cq.callback = mana_schedule_napi; 2311 spec.cq.parent_eq = eq->eq; 2312 spec.cq.context = cq; 2313 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 2314 if (err) 2315 goto out; 2316 2317 memset(&wq_spec, 0, sizeof(wq_spec)); 2318 memset(&cq_spec, 0, sizeof(cq_spec)); 2319 wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle; 2320 wq_spec.queue_size = rxq->gdma_rq->queue_size; 2321 2322 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 2323 cq_spec.queue_size = cq->gdma_cq->queue_size; 2324 cq_spec.modr_ctx_id = 0; 2325 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 2326 2327 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ, 2328 &wq_spec, &cq_spec, &rxq->rxobj); 2329 if (err) 2330 goto out; 2331 2332 rxq->gdma_rq->id = wq_spec.queue_index; 2333 cq->gdma_cq->id = cq_spec.queue_index; 2334 2335 rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2336 cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2337 2338 rxq->gdma_id = rxq->gdma_rq->id; 2339 cq->gdma_id = cq->gdma_cq->id; 2340 2341 err = mana_push_wqe(rxq); 2342 if (err) 2343 goto out; 2344 2345 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 2346 err = -EINVAL; 2347 goto out; 2348 } 2349 2350 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2351 2352 netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1); 2353 2354 WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx, 2355 cq->napi.napi_id)); 2356 WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL, 2357 rxq->page_pool)); 2358 2359 napi_enable(&cq->napi); 2360 2361 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2362 out: 2363 if (!err) 2364 return rxq; 2365 2366 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err); 2367 2368 mana_destroy_rxq(apc, rxq, false); 2369 2370 if (cq) 2371 mana_deinit_cq(apc, cq); 2372 2373 return NULL; 2374 } 2375 2376 static void mana_create_rxq_debugfs(struct mana_port_context *apc, int idx) 2377 { 2378 struct mana_rxq *rxq; 2379 char qnum[32]; 2380 2381 rxq = apc->rxqs[idx]; 2382 2383 sprintf(qnum, "RX-%d", idx); 2384 rxq->mana_rx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs); 2385 debugfs_create_u32("rq_head", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->head); 2386 debugfs_create_u32("rq_tail", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->tail); 2387 debugfs_create_u32("rq_nbuf", 0400, rxq->mana_rx_debugfs, &rxq->num_rx_buf); 2388 debugfs_create_u32("cq_head", 0400, rxq->mana_rx_debugfs, 2389 &rxq->rx_cq.gdma_cq->head); 2390 debugfs_create_u32("cq_tail", 0400, rxq->mana_rx_debugfs, 2391 &rxq->rx_cq.gdma_cq->tail); 2392 debugfs_create_u32("cq_budget", 0400, rxq->mana_rx_debugfs, &rxq->rx_cq.budget); 2393 debugfs_create_file("rxq_dump", 0400, rxq->mana_rx_debugfs, rxq->gdma_rq, &mana_dbg_q_fops); 2394 debugfs_create_file("cq_dump", 0400, rxq->mana_rx_debugfs, rxq->rx_cq.gdma_cq, 2395 &mana_dbg_q_fops); 2396 } 2397 2398 static int mana_add_rx_queues(struct mana_port_context *apc, 2399 struct net_device *ndev) 2400 { 2401 struct mana_context *ac = apc->ac; 2402 struct mana_rxq *rxq; 2403 int err = 0; 2404 int i; 2405 2406 for (i = 0; i < apc->num_queues; i++) { 2407 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev); 2408 if (!rxq) { 2409 err = -ENOMEM; 2410 netdev_err(ndev, "Failed to create rxq %d : %d\n", i, err); 2411 goto out; 2412 } 2413 2414 u64_stats_init(&rxq->stats.syncp); 2415 2416 apc->rxqs[i] = rxq; 2417 2418 mana_create_rxq_debugfs(apc, i); 2419 } 2420 2421 apc->default_rxobj = apc->rxqs[0]->rxobj; 2422 out: 2423 return err; 2424 } 2425 2426 static void mana_destroy_vport(struct mana_port_context *apc) 2427 { 2428 struct gdma_dev *gd = apc->ac->gdma_dev; 2429 struct mana_rxq *rxq; 2430 u32 rxq_idx; 2431 2432 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 2433 rxq = apc->rxqs[rxq_idx]; 2434 if (!rxq) 2435 continue; 2436 2437 mana_destroy_rxq(apc, rxq, true); 2438 apc->rxqs[rxq_idx] = NULL; 2439 } 2440 2441 mana_destroy_txq(apc); 2442 mana_uncfg_vport(apc); 2443 2444 if (gd->gdma_context->is_pf) 2445 mana_pf_deregister_hw_vport(apc); 2446 } 2447 2448 static int mana_create_vport(struct mana_port_context *apc, 2449 struct net_device *net) 2450 { 2451 struct gdma_dev *gd = apc->ac->gdma_dev; 2452 int err; 2453 2454 apc->default_rxobj = INVALID_MANA_HANDLE; 2455 2456 if (gd->gdma_context->is_pf) { 2457 err = mana_pf_register_hw_vport(apc); 2458 if (err) 2459 return err; 2460 } 2461 2462 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell); 2463 if (err) 2464 return err; 2465 2466 return mana_create_txq(apc, net); 2467 } 2468 2469 static int mana_rss_table_alloc(struct mana_port_context *apc) 2470 { 2471 if (!apc->indir_table_sz) { 2472 netdev_err(apc->ndev, 2473 "Indirection table size not set for vPort %d\n", 2474 apc->port_idx); 2475 return -EINVAL; 2476 } 2477 2478 apc->indir_table = kcalloc(apc->indir_table_sz, sizeof(u32), GFP_KERNEL); 2479 if (!apc->indir_table) 2480 return -ENOMEM; 2481 2482 apc->rxobj_table = kcalloc(apc->indir_table_sz, sizeof(mana_handle_t), GFP_KERNEL); 2483 if (!apc->rxobj_table) { 2484 kfree(apc->indir_table); 2485 return -ENOMEM; 2486 } 2487 2488 return 0; 2489 } 2490 2491 static void mana_rss_table_init(struct mana_port_context *apc) 2492 { 2493 int i; 2494 2495 for (i = 0; i < apc->indir_table_sz; i++) 2496 apc->indir_table[i] = 2497 ethtool_rxfh_indir_default(i, apc->num_queues); 2498 } 2499 2500 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx, 2501 bool update_hash, bool update_tab) 2502 { 2503 u32 queue_idx; 2504 int err; 2505 int i; 2506 2507 if (update_tab) { 2508 for (i = 0; i < apc->indir_table_sz; i++) { 2509 queue_idx = apc->indir_table[i]; 2510 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj; 2511 } 2512 } 2513 2514 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab); 2515 if (err) 2516 return err; 2517 2518 mana_fence_rqs(apc); 2519 2520 return 0; 2521 } 2522 2523 void mana_query_gf_stats(struct mana_port_context *apc) 2524 { 2525 struct mana_query_gf_stat_resp resp = {}; 2526 struct mana_query_gf_stat_req req = {}; 2527 struct net_device *ndev = apc->ndev; 2528 int err; 2529 2530 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_GF_STAT, 2531 sizeof(req), sizeof(resp)); 2532 req.hdr.resp.msg_version = GDMA_MESSAGE_V2; 2533 req.req_stats = STATISTICS_FLAGS_RX_DISCARDS_NO_WQE | 2534 STATISTICS_FLAGS_RX_ERRORS_VPORT_DISABLED | 2535 STATISTICS_FLAGS_HC_RX_BYTES | 2536 STATISTICS_FLAGS_HC_RX_UCAST_PACKETS | 2537 STATISTICS_FLAGS_HC_RX_UCAST_BYTES | 2538 STATISTICS_FLAGS_HC_RX_MCAST_PACKETS | 2539 STATISTICS_FLAGS_HC_RX_MCAST_BYTES | 2540 STATISTICS_FLAGS_HC_RX_BCAST_PACKETS | 2541 STATISTICS_FLAGS_HC_RX_BCAST_BYTES | 2542 STATISTICS_FLAGS_TX_ERRORS_GF_DISABLED | 2543 STATISTICS_FLAGS_TX_ERRORS_VPORT_DISABLED | 2544 STATISTICS_FLAGS_TX_ERRORS_INVAL_VPORT_OFFSET_PACKETS | 2545 STATISTICS_FLAGS_TX_ERRORS_VLAN_ENFORCEMENT | 2546 STATISTICS_FLAGS_TX_ERRORS_ETH_TYPE_ENFORCEMENT | 2547 STATISTICS_FLAGS_TX_ERRORS_SA_ENFORCEMENT | 2548 STATISTICS_FLAGS_TX_ERRORS_SQPDID_ENFORCEMENT | 2549 STATISTICS_FLAGS_TX_ERRORS_CQPDID_ENFORCEMENT | 2550 STATISTICS_FLAGS_TX_ERRORS_MTU_VIOLATION | 2551 STATISTICS_FLAGS_TX_ERRORS_INVALID_OOB | 2552 STATISTICS_FLAGS_HC_TX_BYTES | 2553 STATISTICS_FLAGS_HC_TX_UCAST_PACKETS | 2554 STATISTICS_FLAGS_HC_TX_UCAST_BYTES | 2555 STATISTICS_FLAGS_HC_TX_MCAST_PACKETS | 2556 STATISTICS_FLAGS_HC_TX_MCAST_BYTES | 2557 STATISTICS_FLAGS_HC_TX_BCAST_PACKETS | 2558 STATISTICS_FLAGS_HC_TX_BCAST_BYTES | 2559 STATISTICS_FLAGS_TX_ERRORS_GDMA_ERROR; 2560 2561 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 2562 sizeof(resp)); 2563 if (err) { 2564 netdev_err(ndev, "Failed to query GF stats: %d\n", err); 2565 return; 2566 } 2567 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_GF_STAT, 2568 sizeof(resp)); 2569 if (err || resp.hdr.status) { 2570 netdev_err(ndev, "Failed to query GF stats: %d, 0x%x\n", err, 2571 resp.hdr.status); 2572 return; 2573 } 2574 2575 apc->eth_stats.hc_rx_discards_no_wqe = resp.rx_discards_nowqe; 2576 apc->eth_stats.hc_rx_err_vport_disabled = resp.rx_err_vport_disabled; 2577 apc->eth_stats.hc_rx_bytes = resp.hc_rx_bytes; 2578 apc->eth_stats.hc_rx_ucast_pkts = resp.hc_rx_ucast_pkts; 2579 apc->eth_stats.hc_rx_ucast_bytes = resp.hc_rx_ucast_bytes; 2580 apc->eth_stats.hc_rx_bcast_pkts = resp.hc_rx_bcast_pkts; 2581 apc->eth_stats.hc_rx_bcast_bytes = resp.hc_rx_bcast_bytes; 2582 apc->eth_stats.hc_rx_mcast_pkts = resp.hc_rx_mcast_pkts; 2583 apc->eth_stats.hc_rx_mcast_bytes = resp.hc_rx_mcast_bytes; 2584 apc->eth_stats.hc_tx_err_gf_disabled = resp.tx_err_gf_disabled; 2585 apc->eth_stats.hc_tx_err_vport_disabled = resp.tx_err_vport_disabled; 2586 apc->eth_stats.hc_tx_err_inval_vportoffset_pkt = 2587 resp.tx_err_inval_vport_offset_pkt; 2588 apc->eth_stats.hc_tx_err_vlan_enforcement = 2589 resp.tx_err_vlan_enforcement; 2590 apc->eth_stats.hc_tx_err_eth_type_enforcement = 2591 resp.tx_err_ethtype_enforcement; 2592 apc->eth_stats.hc_tx_err_sa_enforcement = resp.tx_err_SA_enforcement; 2593 apc->eth_stats.hc_tx_err_sqpdid_enforcement = 2594 resp.tx_err_SQPDID_enforcement; 2595 apc->eth_stats.hc_tx_err_cqpdid_enforcement = 2596 resp.tx_err_CQPDID_enforcement; 2597 apc->eth_stats.hc_tx_err_mtu_violation = resp.tx_err_mtu_violation; 2598 apc->eth_stats.hc_tx_err_inval_oob = resp.tx_err_inval_oob; 2599 apc->eth_stats.hc_tx_bytes = resp.hc_tx_bytes; 2600 apc->eth_stats.hc_tx_ucast_pkts = resp.hc_tx_ucast_pkts; 2601 apc->eth_stats.hc_tx_ucast_bytes = resp.hc_tx_ucast_bytes; 2602 apc->eth_stats.hc_tx_bcast_pkts = resp.hc_tx_bcast_pkts; 2603 apc->eth_stats.hc_tx_bcast_bytes = resp.hc_tx_bcast_bytes; 2604 apc->eth_stats.hc_tx_mcast_pkts = resp.hc_tx_mcast_pkts; 2605 apc->eth_stats.hc_tx_mcast_bytes = resp.hc_tx_mcast_bytes; 2606 apc->eth_stats.hc_tx_err_gdma = resp.tx_err_gdma; 2607 } 2608 2609 static int mana_init_port(struct net_device *ndev) 2610 { 2611 struct mana_port_context *apc = netdev_priv(ndev); 2612 struct gdma_dev *gd = apc->ac->gdma_dev; 2613 u32 max_txq, max_rxq, max_queues; 2614 int port_idx = apc->port_idx; 2615 struct gdma_context *gc; 2616 char vport[32]; 2617 int err; 2618 2619 err = mana_init_port_context(apc); 2620 if (err) 2621 return err; 2622 2623 gc = gd->gdma_context; 2624 2625 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq, 2626 &apc->indir_table_sz); 2627 if (err) { 2628 netdev_err(ndev, "Failed to query info for vPort %d\n", 2629 port_idx); 2630 goto reset_apc; 2631 } 2632 2633 max_queues = min_t(u32, max_txq, max_rxq); 2634 if (apc->max_queues > max_queues) 2635 apc->max_queues = max_queues; 2636 2637 if (apc->num_queues > apc->max_queues) 2638 apc->num_queues = apc->max_queues; 2639 2640 eth_hw_addr_set(ndev, apc->mac_addr); 2641 sprintf(vport, "vport%d", port_idx); 2642 apc->mana_port_debugfs = debugfs_create_dir(vport, gc->mana_pci_debugfs); 2643 return 0; 2644 2645 reset_apc: 2646 mana_cleanup_port_context(apc); 2647 return err; 2648 } 2649 2650 int mana_alloc_queues(struct net_device *ndev) 2651 { 2652 struct mana_port_context *apc = netdev_priv(ndev); 2653 struct gdma_dev *gd = apc->ac->gdma_dev; 2654 int err; 2655 2656 err = mana_create_vport(apc, ndev); 2657 if (err) { 2658 netdev_err(ndev, "Failed to create vPort %u : %d\n", apc->port_idx, err); 2659 return err; 2660 } 2661 2662 err = netif_set_real_num_tx_queues(ndev, apc->num_queues); 2663 if (err) { 2664 netdev_err(ndev, 2665 "netif_set_real_num_tx_queues () failed for ndev with num_queues %u : %d\n", 2666 apc->num_queues, err); 2667 goto destroy_vport; 2668 } 2669 2670 err = mana_add_rx_queues(apc, ndev); 2671 if (err) 2672 goto destroy_vport; 2673 2674 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE; 2675 2676 err = netif_set_real_num_rx_queues(ndev, apc->num_queues); 2677 if (err) { 2678 netdev_err(ndev, 2679 "netif_set_real_num_rx_queues () failed for ndev with num_queues %u : %d\n", 2680 apc->num_queues, err); 2681 goto destroy_vport; 2682 } 2683 2684 mana_rss_table_init(apc); 2685 2686 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true); 2687 if (err) { 2688 netdev_err(ndev, "Failed to configure RSS table: %d\n", err); 2689 goto destroy_vport; 2690 } 2691 2692 if (gd->gdma_context->is_pf) { 2693 err = mana_pf_register_filter(apc); 2694 if (err) 2695 goto destroy_vport; 2696 } 2697 2698 mana_chn_setxdp(apc, mana_xdp_get(apc)); 2699 2700 return 0; 2701 2702 destroy_vport: 2703 mana_destroy_vport(apc); 2704 return err; 2705 } 2706 2707 int mana_attach(struct net_device *ndev) 2708 { 2709 struct mana_port_context *apc = netdev_priv(ndev); 2710 int err; 2711 2712 ASSERT_RTNL(); 2713 2714 err = mana_init_port(ndev); 2715 if (err) 2716 return err; 2717 2718 if (apc->port_st_save) { 2719 err = mana_alloc_queues(ndev); 2720 if (err) { 2721 mana_cleanup_port_context(apc); 2722 return err; 2723 } 2724 } 2725 2726 apc->port_is_up = apc->port_st_save; 2727 2728 /* Ensure port state updated before txq state */ 2729 smp_wmb(); 2730 2731 if (apc->port_is_up) 2732 netif_carrier_on(ndev); 2733 2734 netif_device_attach(ndev); 2735 2736 return 0; 2737 } 2738 2739 static int mana_dealloc_queues(struct net_device *ndev) 2740 { 2741 struct mana_port_context *apc = netdev_priv(ndev); 2742 unsigned long timeout = jiffies + 120 * HZ; 2743 struct gdma_dev *gd = apc->ac->gdma_dev; 2744 struct mana_txq *txq; 2745 struct sk_buff *skb; 2746 int i, err; 2747 u32 tsleep; 2748 2749 if (apc->port_is_up) 2750 return -EINVAL; 2751 2752 mana_chn_setxdp(apc, NULL); 2753 2754 if (gd->gdma_context->is_pf) 2755 mana_pf_deregister_filter(apc); 2756 2757 /* No packet can be transmitted now since apc->port_is_up is false. 2758 * There is still a tiny chance that mana_poll_tx_cq() can re-enable 2759 * a txq because it may not timely see apc->port_is_up being cleared 2760 * to false, but it doesn't matter since mana_start_xmit() drops any 2761 * new packets due to apc->port_is_up being false. 2762 * 2763 * Drain all the in-flight TX packets. 2764 * A timeout of 120 seconds for all the queues is used. 2765 * This will break the while loop when h/w is not responding. 2766 * This value of 120 has been decided here considering max 2767 * number of queues. 2768 */ 2769 2770 for (i = 0; i < apc->num_queues; i++) { 2771 txq = &apc->tx_qp[i].txq; 2772 tsleep = 1000; 2773 while (atomic_read(&txq->pending_sends) > 0 && 2774 time_before(jiffies, timeout)) { 2775 usleep_range(tsleep, tsleep + 1000); 2776 tsleep <<= 1; 2777 } 2778 if (atomic_read(&txq->pending_sends)) { 2779 err = pcie_flr(to_pci_dev(gd->gdma_context->dev)); 2780 if (err) { 2781 netdev_err(ndev, "flr failed %d with %d pkts pending in txq %u\n", 2782 err, atomic_read(&txq->pending_sends), 2783 txq->gdma_txq_id); 2784 } 2785 break; 2786 } 2787 } 2788 2789 for (i = 0; i < apc->num_queues; i++) { 2790 txq = &apc->tx_qp[i].txq; 2791 while ((skb = skb_dequeue(&txq->pending_skbs))) { 2792 mana_unmap_skb(skb, apc); 2793 dev_kfree_skb_any(skb); 2794 } 2795 atomic_set(&txq->pending_sends, 0); 2796 } 2797 /* We're 100% sure the queues can no longer be woken up, because 2798 * we're sure now mana_poll_tx_cq() can't be running. 2799 */ 2800 2801 apc->rss_state = TRI_STATE_FALSE; 2802 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false); 2803 if (err) { 2804 netdev_err(ndev, "Failed to disable vPort: %d\n", err); 2805 return err; 2806 } 2807 2808 mana_destroy_vport(apc); 2809 2810 return 0; 2811 } 2812 2813 int mana_detach(struct net_device *ndev, bool from_close) 2814 { 2815 struct mana_port_context *apc = netdev_priv(ndev); 2816 int err; 2817 2818 ASSERT_RTNL(); 2819 2820 apc->port_st_save = apc->port_is_up; 2821 apc->port_is_up = false; 2822 2823 /* Ensure port state updated before txq state */ 2824 smp_wmb(); 2825 2826 netif_tx_disable(ndev); 2827 netif_carrier_off(ndev); 2828 2829 if (apc->port_st_save) { 2830 err = mana_dealloc_queues(ndev); 2831 if (err) { 2832 netdev_err(ndev, "%s failed to deallocate queues: %d\n", __func__, err); 2833 return err; 2834 } 2835 } 2836 2837 if (!from_close) { 2838 netif_device_detach(ndev); 2839 mana_cleanup_port_context(apc); 2840 } 2841 2842 return 0; 2843 } 2844 2845 static int mana_probe_port(struct mana_context *ac, int port_idx, 2846 struct net_device **ndev_storage) 2847 { 2848 struct gdma_context *gc = ac->gdma_dev->gdma_context; 2849 struct mana_port_context *apc; 2850 struct net_device *ndev; 2851 int err; 2852 2853 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context), 2854 gc->max_num_queues); 2855 if (!ndev) 2856 return -ENOMEM; 2857 2858 *ndev_storage = ndev; 2859 2860 apc = netdev_priv(ndev); 2861 apc->ac = ac; 2862 apc->ndev = ndev; 2863 apc->max_queues = gc->max_num_queues; 2864 apc->num_queues = gc->max_num_queues; 2865 apc->tx_queue_size = DEF_TX_BUFFERS_PER_QUEUE; 2866 apc->rx_queue_size = DEF_RX_BUFFERS_PER_QUEUE; 2867 apc->port_handle = INVALID_MANA_HANDLE; 2868 apc->pf_filter_handle = INVALID_MANA_HANDLE; 2869 apc->port_idx = port_idx; 2870 2871 mutex_init(&apc->vport_mutex); 2872 apc->vport_use_count = 0; 2873 2874 ndev->netdev_ops = &mana_devops; 2875 ndev->ethtool_ops = &mana_ethtool_ops; 2876 ndev->mtu = ETH_DATA_LEN; 2877 ndev->max_mtu = gc->adapter_mtu - ETH_HLEN; 2878 ndev->min_mtu = ETH_MIN_MTU; 2879 ndev->needed_headroom = MANA_HEADROOM; 2880 ndev->dev_port = port_idx; 2881 SET_NETDEV_DEV(ndev, gc->dev); 2882 2883 netif_set_tso_max_size(ndev, GSO_MAX_SIZE); 2884 2885 netif_carrier_off(ndev); 2886 2887 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE); 2888 2889 err = mana_init_port(ndev); 2890 if (err) 2891 goto free_net; 2892 2893 err = mana_rss_table_alloc(apc); 2894 if (err) 2895 goto reset_apc; 2896 2897 netdev_lockdep_set_classes(ndev); 2898 2899 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 2900 ndev->hw_features |= NETIF_F_RXCSUM; 2901 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6; 2902 ndev->hw_features |= NETIF_F_RXHASH; 2903 ndev->features = ndev->hw_features | NETIF_F_HW_VLAN_CTAG_TX | 2904 NETIF_F_HW_VLAN_CTAG_RX; 2905 ndev->vlan_features = ndev->features; 2906 xdp_set_features_flag(ndev, NETDEV_XDP_ACT_BASIC | 2907 NETDEV_XDP_ACT_REDIRECT | 2908 NETDEV_XDP_ACT_NDO_XMIT); 2909 2910 err = register_netdev(ndev); 2911 if (err) { 2912 netdev_err(ndev, "Unable to register netdev.\n"); 2913 goto free_indir; 2914 } 2915 2916 return 0; 2917 2918 free_indir: 2919 mana_cleanup_indir_table(apc); 2920 reset_apc: 2921 mana_cleanup_port_context(apc); 2922 free_net: 2923 *ndev_storage = NULL; 2924 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err); 2925 free_netdev(ndev); 2926 return err; 2927 } 2928 2929 static void adev_release(struct device *dev) 2930 { 2931 struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev); 2932 2933 kfree(madev); 2934 } 2935 2936 static void remove_adev(struct gdma_dev *gd) 2937 { 2938 struct auxiliary_device *adev = gd->adev; 2939 int id = adev->id; 2940 2941 auxiliary_device_delete(adev); 2942 auxiliary_device_uninit(adev); 2943 2944 mana_adev_idx_free(id); 2945 gd->adev = NULL; 2946 } 2947 2948 static int add_adev(struct gdma_dev *gd) 2949 { 2950 struct auxiliary_device *adev; 2951 struct mana_adev *madev; 2952 int ret; 2953 2954 madev = kzalloc(sizeof(*madev), GFP_KERNEL); 2955 if (!madev) 2956 return -ENOMEM; 2957 2958 adev = &madev->adev; 2959 ret = mana_adev_idx_alloc(); 2960 if (ret < 0) 2961 goto idx_fail; 2962 adev->id = ret; 2963 2964 adev->name = "rdma"; 2965 adev->dev.parent = gd->gdma_context->dev; 2966 adev->dev.release = adev_release; 2967 madev->mdev = gd; 2968 2969 ret = auxiliary_device_init(adev); 2970 if (ret) 2971 goto init_fail; 2972 2973 /* madev is owned by the auxiliary device */ 2974 madev = NULL; 2975 ret = auxiliary_device_add(adev); 2976 if (ret) 2977 goto add_fail; 2978 2979 gd->adev = adev; 2980 dev_dbg(gd->gdma_context->dev, 2981 "Auxiliary device added successfully\n"); 2982 return 0; 2983 2984 add_fail: 2985 auxiliary_device_uninit(adev); 2986 2987 init_fail: 2988 mana_adev_idx_free(adev->id); 2989 2990 idx_fail: 2991 kfree(madev); 2992 2993 return ret; 2994 } 2995 2996 int mana_probe(struct gdma_dev *gd, bool resuming) 2997 { 2998 struct gdma_context *gc = gd->gdma_context; 2999 struct mana_context *ac = gd->driver_data; 3000 struct device *dev = gc->dev; 3001 u16 num_ports = 0; 3002 int err; 3003 int i; 3004 3005 dev_info(dev, 3006 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n", 3007 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION); 3008 3009 err = mana_gd_register_device(gd); 3010 if (err) 3011 return err; 3012 3013 if (!resuming) { 3014 ac = kzalloc(sizeof(*ac), GFP_KERNEL); 3015 if (!ac) 3016 return -ENOMEM; 3017 3018 ac->gdma_dev = gd; 3019 gd->driver_data = ac; 3020 } 3021 3022 err = mana_create_eq(ac); 3023 if (err) { 3024 dev_err(dev, "Failed to create EQs: %d\n", err); 3025 goto out; 3026 } 3027 3028 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION, 3029 MANA_MICRO_VERSION, &num_ports); 3030 if (err) 3031 goto out; 3032 3033 if (!resuming) { 3034 ac->num_ports = num_ports; 3035 } else { 3036 if (ac->num_ports != num_ports) { 3037 dev_err(dev, "The number of vPorts changed: %d->%d\n", 3038 ac->num_ports, num_ports); 3039 err = -EPROTO; 3040 goto out; 3041 } 3042 } 3043 3044 if (ac->num_ports == 0) 3045 dev_err(dev, "Failed to detect any vPort\n"); 3046 3047 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV) 3048 ac->num_ports = MAX_PORTS_IN_MANA_DEV; 3049 3050 if (!resuming) { 3051 for (i = 0; i < ac->num_ports; i++) { 3052 err = mana_probe_port(ac, i, &ac->ports[i]); 3053 /* we log the port for which the probe failed and stop 3054 * probes for subsequent ports. 3055 * Note that we keep running ports, for which the probes 3056 * were successful, unless add_adev fails too 3057 */ 3058 if (err) { 3059 dev_err(dev, "Probe Failed for port %d\n", i); 3060 break; 3061 } 3062 } 3063 } else { 3064 for (i = 0; i < ac->num_ports; i++) { 3065 rtnl_lock(); 3066 err = mana_attach(ac->ports[i]); 3067 rtnl_unlock(); 3068 /* we log the port for which the attach failed and stop 3069 * attach for subsequent ports 3070 * Note that we keep running ports, for which the attach 3071 * were successful, unless add_adev fails too 3072 */ 3073 if (err) { 3074 dev_err(dev, "Attach Failed for port %d\n", i); 3075 break; 3076 } 3077 } 3078 } 3079 3080 err = add_adev(gd); 3081 out: 3082 if (err) { 3083 mana_remove(gd, false); 3084 } else { 3085 dev_dbg(dev, "gd=%p, id=%u, num_ports=%d, type=%u, instance=%u\n", 3086 gd, gd->dev_id.as_uint32, ac->num_ports, 3087 gd->dev_id.type, gd->dev_id.instance); 3088 dev_dbg(dev, "%s succeeded\n", __func__); 3089 } 3090 3091 return err; 3092 } 3093 3094 void mana_remove(struct gdma_dev *gd, bool suspending) 3095 { 3096 struct gdma_context *gc = gd->gdma_context; 3097 struct mana_context *ac = gd->driver_data; 3098 struct mana_port_context *apc; 3099 struct device *dev = gc->dev; 3100 struct net_device *ndev; 3101 int err; 3102 int i; 3103 3104 /* adev currently doesn't support suspending, always remove it */ 3105 if (gd->adev) 3106 remove_adev(gd); 3107 3108 for (i = 0; i < ac->num_ports; i++) { 3109 ndev = ac->ports[i]; 3110 apc = netdev_priv(ndev); 3111 if (!ndev) { 3112 if (i == 0) 3113 dev_err(dev, "No net device to remove\n"); 3114 goto out; 3115 } 3116 3117 /* All cleanup actions should stay after rtnl_lock(), otherwise 3118 * other functions may access partially cleaned up data. 3119 */ 3120 rtnl_lock(); 3121 3122 err = mana_detach(ndev, false); 3123 if (err) 3124 netdev_err(ndev, "Failed to detach vPort %d: %d\n", 3125 i, err); 3126 3127 if (suspending) { 3128 /* No need to unregister the ndev. */ 3129 rtnl_unlock(); 3130 continue; 3131 } 3132 3133 unregister_netdevice(ndev); 3134 mana_cleanup_indir_table(apc); 3135 3136 rtnl_unlock(); 3137 3138 free_netdev(ndev); 3139 } 3140 3141 mana_destroy_eq(ac); 3142 out: 3143 mana_gd_deregister_device(gd); 3144 3145 if (suspending) 3146 return; 3147 3148 gd->driver_data = NULL; 3149 gd->gdma_context = NULL; 3150 kfree(ac); 3151 dev_dbg(dev, "%s succeeded\n", __func__); 3152 } 3153 3154 struct net_device *mana_get_primary_netdev(struct mana_context *ac, 3155 u32 port_index, 3156 netdevice_tracker *tracker) 3157 { 3158 struct net_device *ndev; 3159 3160 if (port_index >= ac->num_ports) 3161 return NULL; 3162 3163 rcu_read_lock(); 3164 3165 /* If mana is used in netvsc, the upper netdevice should be returned. */ 3166 ndev = netdev_master_upper_dev_get_rcu(ac->ports[port_index]); 3167 3168 /* If there is no upper device, use the parent Ethernet device */ 3169 if (!ndev) 3170 ndev = ac->ports[port_index]; 3171 3172 netdev_hold(ndev, tracker, GFP_ATOMIC); 3173 rcu_read_unlock(); 3174 3175 return ndev; 3176 } 3177 EXPORT_SYMBOL_NS(mana_get_primary_netdev, "NET_MANA"); 3178