1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause 2 /* Copyright (c) 2021, Microsoft Corporation. */ 3 4 #include <uapi/linux/bpf.h> 5 6 #include <linux/inetdevice.h> 7 #include <linux/etherdevice.h> 8 #include <linux/ethtool.h> 9 #include <linux/filter.h> 10 #include <linux/mm.h> 11 #include <linux/pci.h> 12 13 #include <net/checksum.h> 14 #include <net/ip6_checksum.h> 15 #include <net/page_pool/helpers.h> 16 #include <net/xdp.h> 17 18 #include <net/mana/mana.h> 19 #include <net/mana/mana_auxiliary.h> 20 21 static DEFINE_IDA(mana_adev_ida); 22 23 static int mana_adev_idx_alloc(void) 24 { 25 return ida_alloc(&mana_adev_ida, GFP_KERNEL); 26 } 27 28 static void mana_adev_idx_free(int idx) 29 { 30 ida_free(&mana_adev_ida, idx); 31 } 32 33 /* Microsoft Azure Network Adapter (MANA) functions */ 34 35 static int mana_open(struct net_device *ndev) 36 { 37 struct mana_port_context *apc = netdev_priv(ndev); 38 int err; 39 40 err = mana_alloc_queues(ndev); 41 if (err) 42 return err; 43 44 apc->port_is_up = true; 45 46 /* Ensure port state updated before txq state */ 47 smp_wmb(); 48 49 netif_carrier_on(ndev); 50 netif_tx_wake_all_queues(ndev); 51 52 return 0; 53 } 54 55 static int mana_close(struct net_device *ndev) 56 { 57 struct mana_port_context *apc = netdev_priv(ndev); 58 59 if (!apc->port_is_up) 60 return 0; 61 62 return mana_detach(ndev, true); 63 } 64 65 static bool mana_can_tx(struct gdma_queue *wq) 66 { 67 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE; 68 } 69 70 static unsigned int mana_checksum_info(struct sk_buff *skb) 71 { 72 if (skb->protocol == htons(ETH_P_IP)) { 73 struct iphdr *ip = ip_hdr(skb); 74 75 if (ip->protocol == IPPROTO_TCP) 76 return IPPROTO_TCP; 77 78 if (ip->protocol == IPPROTO_UDP) 79 return IPPROTO_UDP; 80 } else if (skb->protocol == htons(ETH_P_IPV6)) { 81 struct ipv6hdr *ip6 = ipv6_hdr(skb); 82 83 if (ip6->nexthdr == IPPROTO_TCP) 84 return IPPROTO_TCP; 85 86 if (ip6->nexthdr == IPPROTO_UDP) 87 return IPPROTO_UDP; 88 } 89 90 /* No csum offloading */ 91 return 0; 92 } 93 94 static void mana_add_sge(struct mana_tx_package *tp, struct mana_skb_head *ash, 95 int sg_i, dma_addr_t da, int sge_len, u32 gpa_mkey) 96 { 97 ash->dma_handle[sg_i] = da; 98 ash->size[sg_i] = sge_len; 99 100 tp->wqe_req.sgl[sg_i].address = da; 101 tp->wqe_req.sgl[sg_i].mem_key = gpa_mkey; 102 tp->wqe_req.sgl[sg_i].size = sge_len; 103 } 104 105 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc, 106 struct mana_tx_package *tp, int gso_hs) 107 { 108 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 109 int hsg = 1; /* num of SGEs of linear part */ 110 struct gdma_dev *gd = apc->ac->gdma_dev; 111 int skb_hlen = skb_headlen(skb); 112 int sge0_len, sge1_len = 0; 113 struct gdma_context *gc; 114 struct device *dev; 115 skb_frag_t *frag; 116 dma_addr_t da; 117 int sg_i; 118 int i; 119 120 gc = gd->gdma_context; 121 dev = gc->dev; 122 123 if (gso_hs && gso_hs < skb_hlen) { 124 sge0_len = gso_hs; 125 sge1_len = skb_hlen - gso_hs; 126 } else { 127 sge0_len = skb_hlen; 128 } 129 130 da = dma_map_single(dev, skb->data, sge0_len, DMA_TO_DEVICE); 131 if (dma_mapping_error(dev, da)) 132 return -ENOMEM; 133 134 mana_add_sge(tp, ash, 0, da, sge0_len, gd->gpa_mkey); 135 136 if (sge1_len) { 137 sg_i = 1; 138 da = dma_map_single(dev, skb->data + sge0_len, sge1_len, 139 DMA_TO_DEVICE); 140 if (dma_mapping_error(dev, da)) 141 goto frag_err; 142 143 mana_add_sge(tp, ash, sg_i, da, sge1_len, gd->gpa_mkey); 144 hsg = 2; 145 } 146 147 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 148 sg_i = hsg + i; 149 150 frag = &skb_shinfo(skb)->frags[i]; 151 da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag), 152 DMA_TO_DEVICE); 153 if (dma_mapping_error(dev, da)) 154 goto frag_err; 155 156 mana_add_sge(tp, ash, sg_i, da, skb_frag_size(frag), 157 gd->gpa_mkey); 158 } 159 160 return 0; 161 162 frag_err: 163 for (i = sg_i - 1; i >= hsg; i--) 164 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i], 165 DMA_TO_DEVICE); 166 167 for (i = hsg - 1; i >= 0; i--) 168 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i], 169 DMA_TO_DEVICE); 170 171 return -ENOMEM; 172 } 173 174 /* Handle the case when GSO SKB linear length is too large. 175 * MANA NIC requires GSO packets to put only the packet header to SGE0. 176 * So, we need 2 SGEs for the skb linear part which contains more than the 177 * header. 178 * Return a positive value for the number of SGEs, or a negative value 179 * for an error. 180 */ 181 static int mana_fix_skb_head(struct net_device *ndev, struct sk_buff *skb, 182 int gso_hs) 183 { 184 int num_sge = 1 + skb_shinfo(skb)->nr_frags; 185 int skb_hlen = skb_headlen(skb); 186 187 if (gso_hs < skb_hlen) { 188 num_sge++; 189 } else if (gso_hs > skb_hlen) { 190 if (net_ratelimit()) 191 netdev_err(ndev, 192 "TX nonlinear head: hs:%d, skb_hlen:%d\n", 193 gso_hs, skb_hlen); 194 195 return -EINVAL; 196 } 197 198 return num_sge; 199 } 200 201 /* Get the GSO packet's header size */ 202 static int mana_get_gso_hs(struct sk_buff *skb) 203 { 204 int gso_hs; 205 206 if (skb->encapsulation) { 207 gso_hs = skb_inner_tcp_all_headers(skb); 208 } else { 209 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) { 210 gso_hs = skb_transport_offset(skb) + 211 sizeof(struct udphdr); 212 } else { 213 gso_hs = skb_tcp_all_headers(skb); 214 } 215 } 216 217 return gso_hs; 218 } 219 220 netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev) 221 { 222 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT; 223 struct mana_port_context *apc = netdev_priv(ndev); 224 int gso_hs = 0; /* zero for non-GSO pkts */ 225 u16 txq_idx = skb_get_queue_mapping(skb); 226 struct gdma_dev *gd = apc->ac->gdma_dev; 227 bool ipv4 = false, ipv6 = false; 228 struct mana_tx_package pkg = {}; 229 struct netdev_queue *net_txq; 230 struct mana_stats_tx *tx_stats; 231 struct gdma_queue *gdma_sq; 232 unsigned int csum_type; 233 struct mana_txq *txq; 234 struct mana_cq *cq; 235 int err, len; 236 237 if (unlikely(!apc->port_is_up)) 238 goto tx_drop; 239 240 if (skb_cow_head(skb, MANA_HEADROOM)) 241 goto tx_drop_count; 242 243 txq = &apc->tx_qp[txq_idx].txq; 244 gdma_sq = txq->gdma_sq; 245 cq = &apc->tx_qp[txq_idx].tx_cq; 246 tx_stats = &txq->stats; 247 248 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id; 249 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame; 250 251 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) { 252 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset; 253 pkt_fmt = MANA_LONG_PKT_FMT; 254 } else { 255 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset; 256 } 257 258 if (skb_vlan_tag_present(skb)) { 259 pkt_fmt = MANA_LONG_PKT_FMT; 260 pkg.tx_oob.l_oob.inject_vlan_pri_tag = 1; 261 pkg.tx_oob.l_oob.pcp = skb_vlan_tag_get_prio(skb); 262 pkg.tx_oob.l_oob.dei = skb_vlan_tag_get_cfi(skb); 263 pkg.tx_oob.l_oob.vlan_id = skb_vlan_tag_get_id(skb); 264 } 265 266 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt; 267 268 if (pkt_fmt == MANA_SHORT_PKT_FMT) { 269 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob); 270 u64_stats_update_begin(&tx_stats->syncp); 271 tx_stats->short_pkt_fmt++; 272 u64_stats_update_end(&tx_stats->syncp); 273 } else { 274 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob); 275 u64_stats_update_begin(&tx_stats->syncp); 276 tx_stats->long_pkt_fmt++; 277 u64_stats_update_end(&tx_stats->syncp); 278 } 279 280 pkg.wqe_req.inline_oob_data = &pkg.tx_oob; 281 pkg.wqe_req.flags = 0; 282 pkg.wqe_req.client_data_unit = 0; 283 284 pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags; 285 286 if (skb->protocol == htons(ETH_P_IP)) 287 ipv4 = true; 288 else if (skb->protocol == htons(ETH_P_IPV6)) 289 ipv6 = true; 290 291 if (skb_is_gso(skb)) { 292 int num_sge; 293 294 gso_hs = mana_get_gso_hs(skb); 295 296 num_sge = mana_fix_skb_head(ndev, skb, gso_hs); 297 if (num_sge > 0) 298 pkg.wqe_req.num_sge = num_sge; 299 else 300 goto tx_drop_count; 301 302 u64_stats_update_begin(&tx_stats->syncp); 303 if (skb->encapsulation) { 304 tx_stats->tso_inner_packets++; 305 tx_stats->tso_inner_bytes += skb->len - gso_hs; 306 } else { 307 tx_stats->tso_packets++; 308 tx_stats->tso_bytes += skb->len - gso_hs; 309 } 310 u64_stats_update_end(&tx_stats->syncp); 311 312 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 313 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 314 315 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 316 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 317 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 318 319 pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size; 320 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0; 321 if (ipv4) { 322 ip_hdr(skb)->tot_len = 0; 323 ip_hdr(skb)->check = 0; 324 tcp_hdr(skb)->check = 325 ~csum_tcpudp_magic(ip_hdr(skb)->saddr, 326 ip_hdr(skb)->daddr, 0, 327 IPPROTO_TCP, 0); 328 } else { 329 ipv6_hdr(skb)->payload_len = 0; 330 tcp_hdr(skb)->check = 331 ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr, 332 &ipv6_hdr(skb)->daddr, 0, 333 IPPROTO_TCP, 0); 334 } 335 } else if (skb->ip_summed == CHECKSUM_PARTIAL) { 336 csum_type = mana_checksum_info(skb); 337 338 u64_stats_update_begin(&tx_stats->syncp); 339 tx_stats->csum_partial++; 340 u64_stats_update_end(&tx_stats->syncp); 341 342 if (csum_type == IPPROTO_TCP) { 343 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 344 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 345 346 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 347 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 348 349 } else if (csum_type == IPPROTO_UDP) { 350 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 351 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 352 353 pkg.tx_oob.s_oob.comp_udp_csum = 1; 354 } else { 355 /* Can't do offload of this type of checksum */ 356 if (skb_checksum_help(skb)) 357 goto tx_drop_count; 358 } 359 } 360 361 WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES); 362 363 if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) { 364 pkg.wqe_req.sgl = pkg.sgl_array; 365 } else { 366 pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge, 367 sizeof(struct gdma_sge), 368 GFP_ATOMIC); 369 if (!pkg.sgl_ptr) 370 goto tx_drop_count; 371 372 pkg.wqe_req.sgl = pkg.sgl_ptr; 373 } 374 375 if (mana_map_skb(skb, apc, &pkg, gso_hs)) { 376 u64_stats_update_begin(&tx_stats->syncp); 377 tx_stats->mana_map_err++; 378 u64_stats_update_end(&tx_stats->syncp); 379 goto free_sgl_ptr; 380 } 381 382 skb_queue_tail(&txq->pending_skbs, skb); 383 384 len = skb->len; 385 net_txq = netdev_get_tx_queue(ndev, txq_idx); 386 387 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req, 388 (struct gdma_posted_wqe_info *)skb->cb); 389 if (!mana_can_tx(gdma_sq)) { 390 netif_tx_stop_queue(net_txq); 391 apc->eth_stats.stop_queue++; 392 } 393 394 if (err) { 395 (void)skb_dequeue_tail(&txq->pending_skbs); 396 netdev_warn(ndev, "Failed to post TX OOB: %d\n", err); 397 err = NETDEV_TX_BUSY; 398 goto tx_busy; 399 } 400 401 err = NETDEV_TX_OK; 402 atomic_inc(&txq->pending_sends); 403 404 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq); 405 406 /* skb may be freed after mana_gd_post_work_request. Do not use it. */ 407 skb = NULL; 408 409 tx_stats = &txq->stats; 410 u64_stats_update_begin(&tx_stats->syncp); 411 tx_stats->packets++; 412 tx_stats->bytes += len; 413 u64_stats_update_end(&tx_stats->syncp); 414 415 tx_busy: 416 if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) { 417 netif_tx_wake_queue(net_txq); 418 apc->eth_stats.wake_queue++; 419 } 420 421 kfree(pkg.sgl_ptr); 422 return err; 423 424 free_sgl_ptr: 425 kfree(pkg.sgl_ptr); 426 tx_drop_count: 427 ndev->stats.tx_dropped++; 428 tx_drop: 429 dev_kfree_skb_any(skb); 430 return NETDEV_TX_OK; 431 } 432 433 static void mana_get_stats64(struct net_device *ndev, 434 struct rtnl_link_stats64 *st) 435 { 436 struct mana_port_context *apc = netdev_priv(ndev); 437 unsigned int num_queues = apc->num_queues; 438 struct mana_stats_rx *rx_stats; 439 struct mana_stats_tx *tx_stats; 440 unsigned int start; 441 u64 packets, bytes; 442 int q; 443 444 if (!apc->port_is_up) 445 return; 446 447 netdev_stats_to_stats64(st, &ndev->stats); 448 449 for (q = 0; q < num_queues; q++) { 450 rx_stats = &apc->rxqs[q]->stats; 451 452 do { 453 start = u64_stats_fetch_begin(&rx_stats->syncp); 454 packets = rx_stats->packets; 455 bytes = rx_stats->bytes; 456 } while (u64_stats_fetch_retry(&rx_stats->syncp, start)); 457 458 st->rx_packets += packets; 459 st->rx_bytes += bytes; 460 } 461 462 for (q = 0; q < num_queues; q++) { 463 tx_stats = &apc->tx_qp[q].txq.stats; 464 465 do { 466 start = u64_stats_fetch_begin(&tx_stats->syncp); 467 packets = tx_stats->packets; 468 bytes = tx_stats->bytes; 469 } while (u64_stats_fetch_retry(&tx_stats->syncp, start)); 470 471 st->tx_packets += packets; 472 st->tx_bytes += bytes; 473 } 474 } 475 476 static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb, 477 int old_q) 478 { 479 struct mana_port_context *apc = netdev_priv(ndev); 480 u32 hash = skb_get_hash(skb); 481 struct sock *sk = skb->sk; 482 int txq; 483 484 txq = apc->indir_table[hash & (apc->indir_table_sz - 1)]; 485 486 if (txq != old_q && sk && sk_fullsock(sk) && 487 rcu_access_pointer(sk->sk_dst_cache)) 488 sk_tx_queue_set(sk, txq); 489 490 return txq; 491 } 492 493 static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb, 494 struct net_device *sb_dev) 495 { 496 int txq; 497 498 if (ndev->real_num_tx_queues == 1) 499 return 0; 500 501 txq = sk_tx_queue_get(skb->sk); 502 503 if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) { 504 if (skb_rx_queue_recorded(skb)) 505 txq = skb_get_rx_queue(skb); 506 else 507 txq = mana_get_tx_queue(ndev, skb, txq); 508 } 509 510 return txq; 511 } 512 513 /* Release pre-allocated RX buffers */ 514 static void mana_pre_dealloc_rxbufs(struct mana_port_context *mpc) 515 { 516 struct device *dev; 517 int i; 518 519 dev = mpc->ac->gdma_dev->gdma_context->dev; 520 521 if (!mpc->rxbufs_pre) 522 goto out1; 523 524 if (!mpc->das_pre) 525 goto out2; 526 527 while (mpc->rxbpre_total) { 528 i = --mpc->rxbpre_total; 529 dma_unmap_single(dev, mpc->das_pre[i], mpc->rxbpre_datasize, 530 DMA_FROM_DEVICE); 531 put_page(virt_to_head_page(mpc->rxbufs_pre[i])); 532 } 533 534 kfree(mpc->das_pre); 535 mpc->das_pre = NULL; 536 537 out2: 538 kfree(mpc->rxbufs_pre); 539 mpc->rxbufs_pre = NULL; 540 541 out1: 542 mpc->rxbpre_datasize = 0; 543 mpc->rxbpre_alloc_size = 0; 544 mpc->rxbpre_headroom = 0; 545 } 546 547 /* Get a buffer from the pre-allocated RX buffers */ 548 static void *mana_get_rxbuf_pre(struct mana_rxq *rxq, dma_addr_t *da) 549 { 550 struct net_device *ndev = rxq->ndev; 551 struct mana_port_context *mpc; 552 void *va; 553 554 mpc = netdev_priv(ndev); 555 556 if (!mpc->rxbufs_pre || !mpc->das_pre || !mpc->rxbpre_total) { 557 netdev_err(ndev, "No RX pre-allocated bufs\n"); 558 return NULL; 559 } 560 561 /* Check sizes to catch unexpected coding error */ 562 if (mpc->rxbpre_datasize != rxq->datasize) { 563 netdev_err(ndev, "rxbpre_datasize mismatch: %u: %u\n", 564 mpc->rxbpre_datasize, rxq->datasize); 565 return NULL; 566 } 567 568 if (mpc->rxbpre_alloc_size != rxq->alloc_size) { 569 netdev_err(ndev, "rxbpre_alloc_size mismatch: %u: %u\n", 570 mpc->rxbpre_alloc_size, rxq->alloc_size); 571 return NULL; 572 } 573 574 if (mpc->rxbpre_headroom != rxq->headroom) { 575 netdev_err(ndev, "rxbpre_headroom mismatch: %u: %u\n", 576 mpc->rxbpre_headroom, rxq->headroom); 577 return NULL; 578 } 579 580 mpc->rxbpre_total--; 581 582 *da = mpc->das_pre[mpc->rxbpre_total]; 583 va = mpc->rxbufs_pre[mpc->rxbpre_total]; 584 mpc->rxbufs_pre[mpc->rxbpre_total] = NULL; 585 586 /* Deallocate the array after all buffers are gone */ 587 if (!mpc->rxbpre_total) 588 mana_pre_dealloc_rxbufs(mpc); 589 590 return va; 591 } 592 593 /* Get RX buffer's data size, alloc size, XDP headroom based on MTU */ 594 static void mana_get_rxbuf_cfg(int mtu, u32 *datasize, u32 *alloc_size, 595 u32 *headroom) 596 { 597 if (mtu > MANA_XDP_MTU_MAX) 598 *headroom = 0; /* no support for XDP */ 599 else 600 *headroom = XDP_PACKET_HEADROOM; 601 602 *alloc_size = mtu + MANA_RXBUF_PAD + *headroom; 603 604 *datasize = mtu + ETH_HLEN; 605 } 606 607 static int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu) 608 { 609 struct device *dev; 610 struct page *page; 611 dma_addr_t da; 612 int num_rxb; 613 void *va; 614 int i; 615 616 mana_get_rxbuf_cfg(new_mtu, &mpc->rxbpre_datasize, 617 &mpc->rxbpre_alloc_size, &mpc->rxbpre_headroom); 618 619 dev = mpc->ac->gdma_dev->gdma_context->dev; 620 621 num_rxb = mpc->num_queues * RX_BUFFERS_PER_QUEUE; 622 623 WARN(mpc->rxbufs_pre, "mana rxbufs_pre exists\n"); 624 mpc->rxbufs_pre = kmalloc_array(num_rxb, sizeof(void *), GFP_KERNEL); 625 if (!mpc->rxbufs_pre) 626 goto error; 627 628 mpc->das_pre = kmalloc_array(num_rxb, sizeof(dma_addr_t), GFP_KERNEL); 629 if (!mpc->das_pre) 630 goto error; 631 632 mpc->rxbpre_total = 0; 633 634 for (i = 0; i < num_rxb; i++) { 635 if (mpc->rxbpre_alloc_size > PAGE_SIZE) { 636 va = netdev_alloc_frag(mpc->rxbpre_alloc_size); 637 if (!va) 638 goto error; 639 640 page = virt_to_head_page(va); 641 /* Check if the frag falls back to single page */ 642 if (compound_order(page) < 643 get_order(mpc->rxbpre_alloc_size)) { 644 put_page(page); 645 goto error; 646 } 647 } else { 648 page = dev_alloc_page(); 649 if (!page) 650 goto error; 651 652 va = page_to_virt(page); 653 } 654 655 da = dma_map_single(dev, va + mpc->rxbpre_headroom, 656 mpc->rxbpre_datasize, DMA_FROM_DEVICE); 657 if (dma_mapping_error(dev, da)) { 658 put_page(virt_to_head_page(va)); 659 goto error; 660 } 661 662 mpc->rxbufs_pre[i] = va; 663 mpc->das_pre[i] = da; 664 mpc->rxbpre_total = i + 1; 665 } 666 667 return 0; 668 669 error: 670 mana_pre_dealloc_rxbufs(mpc); 671 return -ENOMEM; 672 } 673 674 static int mana_change_mtu(struct net_device *ndev, int new_mtu) 675 { 676 struct mana_port_context *mpc = netdev_priv(ndev); 677 unsigned int old_mtu = ndev->mtu; 678 int err; 679 680 /* Pre-allocate buffers to prevent failure in mana_attach later */ 681 err = mana_pre_alloc_rxbufs(mpc, new_mtu); 682 if (err) { 683 netdev_err(ndev, "Insufficient memory for new MTU\n"); 684 return err; 685 } 686 687 err = mana_detach(ndev, false); 688 if (err) { 689 netdev_err(ndev, "mana_detach failed: %d\n", err); 690 goto out; 691 } 692 693 WRITE_ONCE(ndev->mtu, new_mtu); 694 695 err = mana_attach(ndev); 696 if (err) { 697 netdev_err(ndev, "mana_attach failed: %d\n", err); 698 WRITE_ONCE(ndev->mtu, old_mtu); 699 } 700 701 out: 702 mana_pre_dealloc_rxbufs(mpc); 703 return err; 704 } 705 706 static const struct net_device_ops mana_devops = { 707 .ndo_open = mana_open, 708 .ndo_stop = mana_close, 709 .ndo_select_queue = mana_select_queue, 710 .ndo_start_xmit = mana_start_xmit, 711 .ndo_validate_addr = eth_validate_addr, 712 .ndo_get_stats64 = mana_get_stats64, 713 .ndo_bpf = mana_bpf, 714 .ndo_xdp_xmit = mana_xdp_xmit, 715 .ndo_change_mtu = mana_change_mtu, 716 }; 717 718 static void mana_cleanup_port_context(struct mana_port_context *apc) 719 { 720 kfree(apc->rxqs); 721 apc->rxqs = NULL; 722 } 723 724 static void mana_cleanup_indir_table(struct mana_port_context *apc) 725 { 726 apc->indir_table_sz = 0; 727 kfree(apc->indir_table); 728 kfree(apc->rxobj_table); 729 } 730 731 static int mana_init_port_context(struct mana_port_context *apc) 732 { 733 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *), 734 GFP_KERNEL); 735 736 return !apc->rxqs ? -ENOMEM : 0; 737 } 738 739 static int mana_send_request(struct mana_context *ac, void *in_buf, 740 u32 in_len, void *out_buf, u32 out_len) 741 { 742 struct gdma_context *gc = ac->gdma_dev->gdma_context; 743 struct gdma_resp_hdr *resp = out_buf; 744 struct gdma_req_hdr *req = in_buf; 745 struct device *dev = gc->dev; 746 static atomic_t activity_id; 747 int err; 748 749 req->dev_id = gc->mana.dev_id; 750 req->activity_id = atomic_inc_return(&activity_id); 751 752 err = mana_gd_send_request(gc, in_len, in_buf, out_len, 753 out_buf); 754 if (err || resp->status) { 755 dev_err(dev, "Failed to send mana message: %d, 0x%x\n", 756 err, resp->status); 757 return err ? err : -EPROTO; 758 } 759 760 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 || 761 req->activity_id != resp->activity_id) { 762 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n", 763 req->dev_id.as_uint32, resp->dev_id.as_uint32, 764 req->activity_id, resp->activity_id); 765 return -EPROTO; 766 } 767 768 return 0; 769 } 770 771 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr, 772 const enum mana_command_code expected_code, 773 const u32 min_size) 774 { 775 if (resp_hdr->response.msg_type != expected_code) 776 return -EPROTO; 777 778 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1) 779 return -EPROTO; 780 781 if (resp_hdr->response.msg_size < min_size) 782 return -EPROTO; 783 784 return 0; 785 } 786 787 static int mana_pf_register_hw_vport(struct mana_port_context *apc) 788 { 789 struct mana_register_hw_vport_resp resp = {}; 790 struct mana_register_hw_vport_req req = {}; 791 int err; 792 793 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT, 794 sizeof(req), sizeof(resp)); 795 req.attached_gfid = 1; 796 req.is_pf_default_vport = 1; 797 req.allow_all_ether_types = 1; 798 799 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 800 sizeof(resp)); 801 if (err) { 802 netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err); 803 return err; 804 } 805 806 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT, 807 sizeof(resp)); 808 if (err || resp.hdr.status) { 809 netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n", 810 err, resp.hdr.status); 811 return err ? err : -EPROTO; 812 } 813 814 apc->port_handle = resp.hw_vport_handle; 815 return 0; 816 } 817 818 static void mana_pf_deregister_hw_vport(struct mana_port_context *apc) 819 { 820 struct mana_deregister_hw_vport_resp resp = {}; 821 struct mana_deregister_hw_vport_req req = {}; 822 int err; 823 824 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT, 825 sizeof(req), sizeof(resp)); 826 req.hw_vport_handle = apc->port_handle; 827 828 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 829 sizeof(resp)); 830 if (err) { 831 netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n", 832 err); 833 return; 834 } 835 836 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT, 837 sizeof(resp)); 838 if (err || resp.hdr.status) 839 netdev_err(apc->ndev, 840 "Failed to deregister hw vPort: %d, 0x%x\n", 841 err, resp.hdr.status); 842 } 843 844 static int mana_pf_register_filter(struct mana_port_context *apc) 845 { 846 struct mana_register_filter_resp resp = {}; 847 struct mana_register_filter_req req = {}; 848 int err; 849 850 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER, 851 sizeof(req), sizeof(resp)); 852 req.vport = apc->port_handle; 853 memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN); 854 855 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 856 sizeof(resp)); 857 if (err) { 858 netdev_err(apc->ndev, "Failed to register filter: %d\n", err); 859 return err; 860 } 861 862 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER, 863 sizeof(resp)); 864 if (err || resp.hdr.status) { 865 netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n", 866 err, resp.hdr.status); 867 return err ? err : -EPROTO; 868 } 869 870 apc->pf_filter_handle = resp.filter_handle; 871 return 0; 872 } 873 874 static void mana_pf_deregister_filter(struct mana_port_context *apc) 875 { 876 struct mana_deregister_filter_resp resp = {}; 877 struct mana_deregister_filter_req req = {}; 878 int err; 879 880 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER, 881 sizeof(req), sizeof(resp)); 882 req.filter_handle = apc->pf_filter_handle; 883 884 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 885 sizeof(resp)); 886 if (err) { 887 netdev_err(apc->ndev, "Failed to unregister filter: %d\n", 888 err); 889 return; 890 } 891 892 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER, 893 sizeof(resp)); 894 if (err || resp.hdr.status) 895 netdev_err(apc->ndev, 896 "Failed to deregister filter: %d, 0x%x\n", 897 err, resp.hdr.status); 898 } 899 900 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver, 901 u32 proto_minor_ver, u32 proto_micro_ver, 902 u16 *max_num_vports) 903 { 904 struct gdma_context *gc = ac->gdma_dev->gdma_context; 905 struct mana_query_device_cfg_resp resp = {}; 906 struct mana_query_device_cfg_req req = {}; 907 struct device *dev = gc->dev; 908 int err = 0; 909 910 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG, 911 sizeof(req), sizeof(resp)); 912 913 req.hdr.resp.msg_version = GDMA_MESSAGE_V2; 914 915 req.proto_major_ver = proto_major_ver; 916 req.proto_minor_ver = proto_minor_ver; 917 req.proto_micro_ver = proto_micro_ver; 918 919 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp)); 920 if (err) { 921 dev_err(dev, "Failed to query config: %d", err); 922 return err; 923 } 924 925 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG, 926 sizeof(resp)); 927 if (err || resp.hdr.status) { 928 dev_err(dev, "Invalid query result: %d, 0x%x\n", err, 929 resp.hdr.status); 930 if (!err) 931 err = -EPROTO; 932 return err; 933 } 934 935 *max_num_vports = resp.max_num_vports; 936 937 if (resp.hdr.response.msg_version == GDMA_MESSAGE_V2) 938 gc->adapter_mtu = resp.adapter_mtu; 939 else 940 gc->adapter_mtu = ETH_FRAME_LEN; 941 942 return 0; 943 } 944 945 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index, 946 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry) 947 { 948 struct mana_query_vport_cfg_resp resp = {}; 949 struct mana_query_vport_cfg_req req = {}; 950 int err; 951 952 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG, 953 sizeof(req), sizeof(resp)); 954 955 req.vport_index = vport_index; 956 957 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 958 sizeof(resp)); 959 if (err) 960 return err; 961 962 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG, 963 sizeof(resp)); 964 if (err) 965 return err; 966 967 if (resp.hdr.status) 968 return -EPROTO; 969 970 *max_sq = resp.max_num_sq; 971 *max_rq = resp.max_num_rq; 972 if (resp.num_indirection_ent > 0 && 973 resp.num_indirection_ent <= MANA_INDIRECT_TABLE_MAX_SIZE && 974 is_power_of_2(resp.num_indirection_ent)) { 975 *num_indir_entry = resp.num_indirection_ent; 976 } else { 977 netdev_warn(apc->ndev, 978 "Setting indirection table size to default %d for vPort %d\n", 979 MANA_INDIRECT_TABLE_DEF_SIZE, apc->port_idx); 980 *num_indir_entry = MANA_INDIRECT_TABLE_DEF_SIZE; 981 } 982 983 apc->port_handle = resp.vport; 984 ether_addr_copy(apc->mac_addr, resp.mac_addr); 985 986 return 0; 987 } 988 989 void mana_uncfg_vport(struct mana_port_context *apc) 990 { 991 mutex_lock(&apc->vport_mutex); 992 apc->vport_use_count--; 993 WARN_ON(apc->vport_use_count < 0); 994 mutex_unlock(&apc->vport_mutex); 995 } 996 EXPORT_SYMBOL_NS(mana_uncfg_vport, NET_MANA); 997 998 int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id, 999 u32 doorbell_pg_id) 1000 { 1001 struct mana_config_vport_resp resp = {}; 1002 struct mana_config_vport_req req = {}; 1003 int err; 1004 1005 /* This function is used to program the Ethernet port in the hardware 1006 * table. It can be called from the Ethernet driver or the RDMA driver. 1007 * 1008 * For Ethernet usage, the hardware supports only one active user on a 1009 * physical port. The driver checks on the port usage before programming 1010 * the hardware when creating the RAW QP (RDMA driver) or exposing the 1011 * device to kernel NET layer (Ethernet driver). 1012 * 1013 * Because the RDMA driver doesn't know in advance which QP type the 1014 * user will create, it exposes the device with all its ports. The user 1015 * may not be able to create RAW QP on a port if this port is already 1016 * in used by the Ethernet driver from the kernel. 1017 * 1018 * This physical port limitation only applies to the RAW QP. For RC QP, 1019 * the hardware doesn't have this limitation. The user can create RC 1020 * QPs on a physical port up to the hardware limits independent of the 1021 * Ethernet usage on the same port. 1022 */ 1023 mutex_lock(&apc->vport_mutex); 1024 if (apc->vport_use_count > 0) { 1025 mutex_unlock(&apc->vport_mutex); 1026 return -EBUSY; 1027 } 1028 apc->vport_use_count++; 1029 mutex_unlock(&apc->vport_mutex); 1030 1031 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX, 1032 sizeof(req), sizeof(resp)); 1033 req.vport = apc->port_handle; 1034 req.pdid = protection_dom_id; 1035 req.doorbell_pageid = doorbell_pg_id; 1036 1037 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1038 sizeof(resp)); 1039 if (err) { 1040 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err); 1041 goto out; 1042 } 1043 1044 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX, 1045 sizeof(resp)); 1046 if (err || resp.hdr.status) { 1047 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n", 1048 err, resp.hdr.status); 1049 if (!err) 1050 err = -EPROTO; 1051 1052 goto out; 1053 } 1054 1055 apc->tx_shortform_allowed = resp.short_form_allowed; 1056 apc->tx_vp_offset = resp.tx_vport_offset; 1057 1058 netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n", 1059 apc->port_handle, protection_dom_id, doorbell_pg_id); 1060 out: 1061 if (err) 1062 mana_uncfg_vport(apc); 1063 1064 return err; 1065 } 1066 EXPORT_SYMBOL_NS(mana_cfg_vport, NET_MANA); 1067 1068 static int mana_cfg_vport_steering(struct mana_port_context *apc, 1069 enum TRI_STATE rx, 1070 bool update_default_rxobj, bool update_key, 1071 bool update_tab) 1072 { 1073 struct mana_cfg_rx_steer_req_v2 *req; 1074 struct mana_cfg_rx_steer_resp resp = {}; 1075 struct net_device *ndev = apc->ndev; 1076 u32 req_buf_size; 1077 int err; 1078 1079 req_buf_size = struct_size(req, indir_tab, apc->indir_table_sz); 1080 req = kzalloc(req_buf_size, GFP_KERNEL); 1081 if (!req) 1082 return -ENOMEM; 1083 1084 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size, 1085 sizeof(resp)); 1086 1087 req->hdr.req.msg_version = GDMA_MESSAGE_V2; 1088 1089 req->vport = apc->port_handle; 1090 req->num_indir_entries = apc->indir_table_sz; 1091 req->indir_tab_offset = offsetof(struct mana_cfg_rx_steer_req_v2, 1092 indir_tab); 1093 req->rx_enable = rx; 1094 req->rss_enable = apc->rss_state; 1095 req->update_default_rxobj = update_default_rxobj; 1096 req->update_hashkey = update_key; 1097 req->update_indir_tab = update_tab; 1098 req->default_rxobj = apc->default_rxobj; 1099 req->cqe_coalescing_enable = 0; 1100 1101 if (update_key) 1102 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE); 1103 1104 if (update_tab) 1105 memcpy(req->indir_tab, apc->rxobj_table, 1106 flex_array_size(req, indir_tab, req->num_indir_entries)); 1107 1108 err = mana_send_request(apc->ac, req, req_buf_size, &resp, 1109 sizeof(resp)); 1110 if (err) { 1111 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err); 1112 goto out; 1113 } 1114 1115 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX, 1116 sizeof(resp)); 1117 if (err) { 1118 netdev_err(ndev, "vPort RX configuration failed: %d\n", err); 1119 goto out; 1120 } 1121 1122 if (resp.hdr.status) { 1123 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n", 1124 resp.hdr.status); 1125 err = -EPROTO; 1126 } 1127 1128 netdev_info(ndev, "Configured steering vPort %llu entries %u\n", 1129 apc->port_handle, apc->indir_table_sz); 1130 out: 1131 kfree(req); 1132 return err; 1133 } 1134 1135 int mana_create_wq_obj(struct mana_port_context *apc, 1136 mana_handle_t vport, 1137 u32 wq_type, struct mana_obj_spec *wq_spec, 1138 struct mana_obj_spec *cq_spec, 1139 mana_handle_t *wq_obj) 1140 { 1141 struct mana_create_wqobj_resp resp = {}; 1142 struct mana_create_wqobj_req req = {}; 1143 struct net_device *ndev = apc->ndev; 1144 int err; 1145 1146 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ, 1147 sizeof(req), sizeof(resp)); 1148 req.vport = vport; 1149 req.wq_type = wq_type; 1150 req.wq_gdma_region = wq_spec->gdma_region; 1151 req.cq_gdma_region = cq_spec->gdma_region; 1152 req.wq_size = wq_spec->queue_size; 1153 req.cq_size = cq_spec->queue_size; 1154 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id; 1155 req.cq_parent_qid = cq_spec->attached_eq; 1156 1157 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1158 sizeof(resp)); 1159 if (err) { 1160 netdev_err(ndev, "Failed to create WQ object: %d\n", err); 1161 goto out; 1162 } 1163 1164 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ, 1165 sizeof(resp)); 1166 if (err || resp.hdr.status) { 1167 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err, 1168 resp.hdr.status); 1169 if (!err) 1170 err = -EPROTO; 1171 goto out; 1172 } 1173 1174 if (resp.wq_obj == INVALID_MANA_HANDLE) { 1175 netdev_err(ndev, "Got an invalid WQ object handle\n"); 1176 err = -EPROTO; 1177 goto out; 1178 } 1179 1180 *wq_obj = resp.wq_obj; 1181 wq_spec->queue_index = resp.wq_id; 1182 cq_spec->queue_index = resp.cq_id; 1183 1184 return 0; 1185 out: 1186 return err; 1187 } 1188 EXPORT_SYMBOL_NS(mana_create_wq_obj, NET_MANA); 1189 1190 void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type, 1191 mana_handle_t wq_obj) 1192 { 1193 struct mana_destroy_wqobj_resp resp = {}; 1194 struct mana_destroy_wqobj_req req = {}; 1195 struct net_device *ndev = apc->ndev; 1196 int err; 1197 1198 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ, 1199 sizeof(req), sizeof(resp)); 1200 req.wq_type = wq_type; 1201 req.wq_obj_handle = wq_obj; 1202 1203 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1204 sizeof(resp)); 1205 if (err) { 1206 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err); 1207 return; 1208 } 1209 1210 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ, 1211 sizeof(resp)); 1212 if (err || resp.hdr.status) 1213 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err, 1214 resp.hdr.status); 1215 } 1216 EXPORT_SYMBOL_NS(mana_destroy_wq_obj, NET_MANA); 1217 1218 static void mana_destroy_eq(struct mana_context *ac) 1219 { 1220 struct gdma_context *gc = ac->gdma_dev->gdma_context; 1221 struct gdma_queue *eq; 1222 int i; 1223 1224 if (!ac->eqs) 1225 return; 1226 1227 for (i = 0; i < gc->max_num_queues; i++) { 1228 eq = ac->eqs[i].eq; 1229 if (!eq) 1230 continue; 1231 1232 mana_gd_destroy_queue(gc, eq); 1233 } 1234 1235 kfree(ac->eqs); 1236 ac->eqs = NULL; 1237 } 1238 1239 static int mana_create_eq(struct mana_context *ac) 1240 { 1241 struct gdma_dev *gd = ac->gdma_dev; 1242 struct gdma_context *gc = gd->gdma_context; 1243 struct gdma_queue_spec spec = {}; 1244 int err; 1245 int i; 1246 1247 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq), 1248 GFP_KERNEL); 1249 if (!ac->eqs) 1250 return -ENOMEM; 1251 1252 spec.type = GDMA_EQ; 1253 spec.monitor_avl_buf = false; 1254 spec.queue_size = EQ_SIZE; 1255 spec.eq.callback = NULL; 1256 spec.eq.context = ac->eqs; 1257 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE; 1258 1259 for (i = 0; i < gc->max_num_queues; i++) { 1260 spec.eq.msix_index = (i + 1) % gc->num_msix_usable; 1261 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq); 1262 if (err) 1263 goto out; 1264 } 1265 1266 return 0; 1267 out: 1268 mana_destroy_eq(ac); 1269 return err; 1270 } 1271 1272 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq) 1273 { 1274 struct mana_fence_rq_resp resp = {}; 1275 struct mana_fence_rq_req req = {}; 1276 int err; 1277 1278 init_completion(&rxq->fence_event); 1279 1280 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ, 1281 sizeof(req), sizeof(resp)); 1282 req.wq_obj_handle = rxq->rxobj; 1283 1284 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1285 sizeof(resp)); 1286 if (err) { 1287 netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n", 1288 rxq->rxq_idx, err); 1289 return err; 1290 } 1291 1292 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp)); 1293 if (err || resp.hdr.status) { 1294 netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n", 1295 rxq->rxq_idx, err, resp.hdr.status); 1296 if (!err) 1297 err = -EPROTO; 1298 1299 return err; 1300 } 1301 1302 if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) { 1303 netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n", 1304 rxq->rxq_idx); 1305 return -ETIMEDOUT; 1306 } 1307 1308 return 0; 1309 } 1310 1311 static void mana_fence_rqs(struct mana_port_context *apc) 1312 { 1313 unsigned int rxq_idx; 1314 struct mana_rxq *rxq; 1315 int err; 1316 1317 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 1318 rxq = apc->rxqs[rxq_idx]; 1319 err = mana_fence_rq(apc, rxq); 1320 1321 /* In case of any error, use sleep instead. */ 1322 if (err) 1323 msleep(100); 1324 } 1325 } 1326 1327 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units) 1328 { 1329 u32 used_space_old; 1330 u32 used_space_new; 1331 1332 used_space_old = wq->head - wq->tail; 1333 used_space_new = wq->head - (wq->tail + num_units); 1334 1335 if (WARN_ON_ONCE(used_space_new > used_space_old)) 1336 return -ERANGE; 1337 1338 wq->tail += num_units; 1339 return 0; 1340 } 1341 1342 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc) 1343 { 1344 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 1345 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 1346 struct device *dev = gc->dev; 1347 int hsg, i; 1348 1349 /* Number of SGEs of linear part */ 1350 hsg = (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0]) ? 2 : 1; 1351 1352 for (i = 0; i < hsg; i++) 1353 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i], 1354 DMA_TO_DEVICE); 1355 1356 for (i = hsg; i < skb_shinfo(skb)->nr_frags + hsg; i++) 1357 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i], 1358 DMA_TO_DEVICE); 1359 } 1360 1361 static void mana_poll_tx_cq(struct mana_cq *cq) 1362 { 1363 struct gdma_comp *completions = cq->gdma_comp_buf; 1364 struct gdma_posted_wqe_info *wqe_info; 1365 unsigned int pkt_transmitted = 0; 1366 unsigned int wqe_unit_cnt = 0; 1367 struct mana_txq *txq = cq->txq; 1368 struct mana_port_context *apc; 1369 struct netdev_queue *net_txq; 1370 struct gdma_queue *gdma_wq; 1371 unsigned int avail_space; 1372 struct net_device *ndev; 1373 struct sk_buff *skb; 1374 bool txq_stopped; 1375 int comp_read; 1376 int i; 1377 1378 ndev = txq->ndev; 1379 apc = netdev_priv(ndev); 1380 1381 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions, 1382 CQE_POLLING_BUFFER); 1383 1384 if (comp_read < 1) 1385 return; 1386 1387 for (i = 0; i < comp_read; i++) { 1388 struct mana_tx_comp_oob *cqe_oob; 1389 1390 if (WARN_ON_ONCE(!completions[i].is_sq)) 1391 return; 1392 1393 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data; 1394 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type != 1395 MANA_CQE_COMPLETION)) 1396 return; 1397 1398 switch (cqe_oob->cqe_hdr.cqe_type) { 1399 case CQE_TX_OKAY: 1400 break; 1401 1402 case CQE_TX_SA_DROP: 1403 case CQE_TX_MTU_DROP: 1404 case CQE_TX_INVALID_OOB: 1405 case CQE_TX_INVALID_ETH_TYPE: 1406 case CQE_TX_HDR_PROCESSING_ERROR: 1407 case CQE_TX_VF_DISABLED: 1408 case CQE_TX_VPORT_IDX_OUT_OF_RANGE: 1409 case CQE_TX_VPORT_DISABLED: 1410 case CQE_TX_VLAN_TAGGING_VIOLATION: 1411 if (net_ratelimit()) 1412 netdev_err(ndev, "TX: CQE error %d\n", 1413 cqe_oob->cqe_hdr.cqe_type); 1414 1415 apc->eth_stats.tx_cqe_err++; 1416 break; 1417 1418 default: 1419 /* If the CQE type is unknown, log an error, 1420 * and still free the SKB, update tail, etc. 1421 */ 1422 if (net_ratelimit()) 1423 netdev_err(ndev, "TX: unknown CQE type %d\n", 1424 cqe_oob->cqe_hdr.cqe_type); 1425 1426 apc->eth_stats.tx_cqe_unknown_type++; 1427 break; 1428 } 1429 1430 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num)) 1431 return; 1432 1433 skb = skb_dequeue(&txq->pending_skbs); 1434 if (WARN_ON_ONCE(!skb)) 1435 return; 1436 1437 wqe_info = (struct gdma_posted_wqe_info *)skb->cb; 1438 wqe_unit_cnt += wqe_info->wqe_size_in_bu; 1439 1440 mana_unmap_skb(skb, apc); 1441 1442 napi_consume_skb(skb, cq->budget); 1443 1444 pkt_transmitted++; 1445 } 1446 1447 if (WARN_ON_ONCE(wqe_unit_cnt == 0)) 1448 return; 1449 1450 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt); 1451 1452 gdma_wq = txq->gdma_sq; 1453 avail_space = mana_gd_wq_avail_space(gdma_wq); 1454 1455 /* Ensure tail updated before checking q stop */ 1456 smp_mb(); 1457 1458 net_txq = txq->net_txq; 1459 txq_stopped = netif_tx_queue_stopped(net_txq); 1460 1461 /* Ensure checking txq_stopped before apc->port_is_up. */ 1462 smp_rmb(); 1463 1464 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) { 1465 netif_tx_wake_queue(net_txq); 1466 apc->eth_stats.wake_queue++; 1467 } 1468 1469 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0) 1470 WARN_ON_ONCE(1); 1471 1472 cq->work_done = pkt_transmitted; 1473 } 1474 1475 static void mana_post_pkt_rxq(struct mana_rxq *rxq) 1476 { 1477 struct mana_recv_buf_oob *recv_buf_oob; 1478 u32 curr_index; 1479 int err; 1480 1481 curr_index = rxq->buf_index++; 1482 if (rxq->buf_index == rxq->num_rx_buf) 1483 rxq->buf_index = 0; 1484 1485 recv_buf_oob = &rxq->rx_oobs[curr_index]; 1486 1487 err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req, 1488 &recv_buf_oob->wqe_inf); 1489 if (WARN_ON_ONCE(err)) 1490 return; 1491 1492 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1); 1493 } 1494 1495 static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va, 1496 uint pkt_len, struct xdp_buff *xdp) 1497 { 1498 struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size); 1499 1500 if (!skb) 1501 return NULL; 1502 1503 if (xdp->data_hard_start) { 1504 skb_reserve(skb, xdp->data - xdp->data_hard_start); 1505 skb_put(skb, xdp->data_end - xdp->data); 1506 return skb; 1507 } 1508 1509 skb_reserve(skb, rxq->headroom); 1510 skb_put(skb, pkt_len); 1511 1512 return skb; 1513 } 1514 1515 static void mana_rx_skb(void *buf_va, bool from_pool, 1516 struct mana_rxcomp_oob *cqe, struct mana_rxq *rxq) 1517 { 1518 struct mana_stats_rx *rx_stats = &rxq->stats; 1519 struct net_device *ndev = rxq->ndev; 1520 uint pkt_len = cqe->ppi[0].pkt_len; 1521 u16 rxq_idx = rxq->rxq_idx; 1522 struct napi_struct *napi; 1523 struct xdp_buff xdp = {}; 1524 struct sk_buff *skb; 1525 u32 hash_value; 1526 u32 act; 1527 1528 rxq->rx_cq.work_done++; 1529 napi = &rxq->rx_cq.napi; 1530 1531 if (!buf_va) { 1532 ++ndev->stats.rx_dropped; 1533 return; 1534 } 1535 1536 act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len); 1537 1538 if (act == XDP_REDIRECT && !rxq->xdp_rc) 1539 return; 1540 1541 if (act != XDP_PASS && act != XDP_TX) 1542 goto drop_xdp; 1543 1544 skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp); 1545 1546 if (!skb) 1547 goto drop; 1548 1549 if (from_pool) 1550 skb_mark_for_recycle(skb); 1551 1552 skb->dev = napi->dev; 1553 1554 skb->protocol = eth_type_trans(skb, ndev); 1555 skb_checksum_none_assert(skb); 1556 skb_record_rx_queue(skb, rxq_idx); 1557 1558 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) { 1559 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed) 1560 skb->ip_summed = CHECKSUM_UNNECESSARY; 1561 } 1562 1563 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) { 1564 hash_value = cqe->ppi[0].pkt_hash; 1565 1566 if (cqe->rx_hashtype & MANA_HASH_L4) 1567 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4); 1568 else 1569 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3); 1570 } 1571 1572 if (cqe->rx_vlantag_present) { 1573 u16 vlan_tci = cqe->rx_vlan_id; 1574 1575 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci); 1576 } 1577 1578 u64_stats_update_begin(&rx_stats->syncp); 1579 rx_stats->packets++; 1580 rx_stats->bytes += pkt_len; 1581 1582 if (act == XDP_TX) 1583 rx_stats->xdp_tx++; 1584 u64_stats_update_end(&rx_stats->syncp); 1585 1586 if (act == XDP_TX) { 1587 skb_set_queue_mapping(skb, rxq_idx); 1588 mana_xdp_tx(skb, ndev); 1589 return; 1590 } 1591 1592 napi_gro_receive(napi, skb); 1593 1594 return; 1595 1596 drop_xdp: 1597 u64_stats_update_begin(&rx_stats->syncp); 1598 rx_stats->xdp_drop++; 1599 u64_stats_update_end(&rx_stats->syncp); 1600 1601 drop: 1602 if (from_pool) { 1603 page_pool_recycle_direct(rxq->page_pool, 1604 virt_to_head_page(buf_va)); 1605 } else { 1606 WARN_ON_ONCE(rxq->xdp_save_va); 1607 /* Save for reuse */ 1608 rxq->xdp_save_va = buf_va; 1609 } 1610 1611 ++ndev->stats.rx_dropped; 1612 1613 return; 1614 } 1615 1616 static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev, 1617 dma_addr_t *da, bool *from_pool, bool is_napi) 1618 { 1619 struct page *page; 1620 void *va; 1621 1622 *from_pool = false; 1623 1624 /* Reuse XDP dropped page if available */ 1625 if (rxq->xdp_save_va) { 1626 va = rxq->xdp_save_va; 1627 rxq->xdp_save_va = NULL; 1628 } else if (rxq->alloc_size > PAGE_SIZE) { 1629 if (is_napi) 1630 va = napi_alloc_frag(rxq->alloc_size); 1631 else 1632 va = netdev_alloc_frag(rxq->alloc_size); 1633 1634 if (!va) 1635 return NULL; 1636 1637 page = virt_to_head_page(va); 1638 /* Check if the frag falls back to single page */ 1639 if (compound_order(page) < get_order(rxq->alloc_size)) { 1640 put_page(page); 1641 return NULL; 1642 } 1643 } else { 1644 page = page_pool_dev_alloc_pages(rxq->page_pool); 1645 if (!page) 1646 return NULL; 1647 1648 *from_pool = true; 1649 va = page_to_virt(page); 1650 } 1651 1652 *da = dma_map_single(dev, va + rxq->headroom, rxq->datasize, 1653 DMA_FROM_DEVICE); 1654 if (dma_mapping_error(dev, *da)) { 1655 if (*from_pool) 1656 page_pool_put_full_page(rxq->page_pool, page, false); 1657 else 1658 put_page(virt_to_head_page(va)); 1659 1660 return NULL; 1661 } 1662 1663 return va; 1664 } 1665 1666 /* Allocate frag for rx buffer, and save the old buf */ 1667 static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq, 1668 struct mana_recv_buf_oob *rxoob, void **old_buf, 1669 bool *old_fp) 1670 { 1671 bool from_pool; 1672 dma_addr_t da; 1673 void *va; 1674 1675 va = mana_get_rxfrag(rxq, dev, &da, &from_pool, true); 1676 if (!va) 1677 return; 1678 1679 dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize, 1680 DMA_FROM_DEVICE); 1681 *old_buf = rxoob->buf_va; 1682 *old_fp = rxoob->from_pool; 1683 1684 rxoob->buf_va = va; 1685 rxoob->sgl[0].address = da; 1686 rxoob->from_pool = from_pool; 1687 } 1688 1689 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq, 1690 struct gdma_comp *cqe) 1691 { 1692 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data; 1693 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context; 1694 struct net_device *ndev = rxq->ndev; 1695 struct mana_recv_buf_oob *rxbuf_oob; 1696 struct mana_port_context *apc; 1697 struct device *dev = gc->dev; 1698 void *old_buf = NULL; 1699 u32 curr, pktlen; 1700 bool old_fp; 1701 1702 apc = netdev_priv(ndev); 1703 1704 switch (oob->cqe_hdr.cqe_type) { 1705 case CQE_RX_OKAY: 1706 break; 1707 1708 case CQE_RX_TRUNCATED: 1709 ++ndev->stats.rx_dropped; 1710 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index]; 1711 netdev_warn_once(ndev, "Dropped a truncated packet\n"); 1712 goto drop; 1713 1714 case CQE_RX_COALESCED_4: 1715 netdev_err(ndev, "RX coalescing is unsupported\n"); 1716 apc->eth_stats.rx_coalesced_err++; 1717 return; 1718 1719 case CQE_RX_OBJECT_FENCE: 1720 complete(&rxq->fence_event); 1721 return; 1722 1723 default: 1724 netdev_err(ndev, "Unknown RX CQE type = %d\n", 1725 oob->cqe_hdr.cqe_type); 1726 apc->eth_stats.rx_cqe_unknown_type++; 1727 return; 1728 } 1729 1730 pktlen = oob->ppi[0].pkt_len; 1731 1732 if (pktlen == 0) { 1733 /* data packets should never have packetlength of zero */ 1734 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n", 1735 rxq->gdma_id, cq->gdma_id, rxq->rxobj); 1736 return; 1737 } 1738 1739 curr = rxq->buf_index; 1740 rxbuf_oob = &rxq->rx_oobs[curr]; 1741 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1); 1742 1743 mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf, &old_fp); 1744 1745 /* Unsuccessful refill will have old_buf == NULL. 1746 * In this case, mana_rx_skb() will drop the packet. 1747 */ 1748 mana_rx_skb(old_buf, old_fp, oob, rxq); 1749 1750 drop: 1751 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu); 1752 1753 mana_post_pkt_rxq(rxq); 1754 } 1755 1756 static void mana_poll_rx_cq(struct mana_cq *cq) 1757 { 1758 struct gdma_comp *comp = cq->gdma_comp_buf; 1759 struct mana_rxq *rxq = cq->rxq; 1760 int comp_read, i; 1761 1762 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER); 1763 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER); 1764 1765 rxq->xdp_flush = false; 1766 1767 for (i = 0; i < comp_read; i++) { 1768 if (WARN_ON_ONCE(comp[i].is_sq)) 1769 return; 1770 1771 /* verify recv cqe references the right rxq */ 1772 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id)) 1773 return; 1774 1775 mana_process_rx_cqe(rxq, cq, &comp[i]); 1776 } 1777 1778 if (comp_read > 0) { 1779 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context; 1780 1781 mana_gd_wq_ring_doorbell(gc, rxq->gdma_rq); 1782 } 1783 1784 if (rxq->xdp_flush) 1785 xdp_do_flush(); 1786 } 1787 1788 static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue) 1789 { 1790 struct mana_cq *cq = context; 1791 u8 arm_bit; 1792 int w; 1793 1794 WARN_ON_ONCE(cq->gdma_cq != gdma_queue); 1795 1796 if (cq->type == MANA_CQ_TYPE_RX) 1797 mana_poll_rx_cq(cq); 1798 else 1799 mana_poll_tx_cq(cq); 1800 1801 w = cq->work_done; 1802 1803 if (w < cq->budget && 1804 napi_complete_done(&cq->napi, w)) { 1805 arm_bit = SET_ARM_BIT; 1806 } else { 1807 arm_bit = 0; 1808 } 1809 1810 mana_gd_ring_cq(gdma_queue, arm_bit); 1811 1812 return w; 1813 } 1814 1815 static int mana_poll(struct napi_struct *napi, int budget) 1816 { 1817 struct mana_cq *cq = container_of(napi, struct mana_cq, napi); 1818 int w; 1819 1820 cq->work_done = 0; 1821 cq->budget = budget; 1822 1823 w = mana_cq_handler(cq, cq->gdma_cq); 1824 1825 return min(w, budget); 1826 } 1827 1828 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue) 1829 { 1830 struct mana_cq *cq = context; 1831 1832 napi_schedule_irqoff(&cq->napi); 1833 } 1834 1835 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq) 1836 { 1837 struct gdma_dev *gd = apc->ac->gdma_dev; 1838 1839 if (!cq->gdma_cq) 1840 return; 1841 1842 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq); 1843 } 1844 1845 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq) 1846 { 1847 struct gdma_dev *gd = apc->ac->gdma_dev; 1848 1849 if (!txq->gdma_sq) 1850 return; 1851 1852 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq); 1853 } 1854 1855 static void mana_destroy_txq(struct mana_port_context *apc) 1856 { 1857 struct napi_struct *napi; 1858 int i; 1859 1860 if (!apc->tx_qp) 1861 return; 1862 1863 for (i = 0; i < apc->num_queues; i++) { 1864 napi = &apc->tx_qp[i].tx_cq.napi; 1865 napi_synchronize(napi); 1866 napi_disable(napi); 1867 netif_napi_del(napi); 1868 1869 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object); 1870 1871 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq); 1872 1873 mana_deinit_txq(apc, &apc->tx_qp[i].txq); 1874 } 1875 1876 kfree(apc->tx_qp); 1877 apc->tx_qp = NULL; 1878 } 1879 1880 static int mana_create_txq(struct mana_port_context *apc, 1881 struct net_device *net) 1882 { 1883 struct mana_context *ac = apc->ac; 1884 struct gdma_dev *gd = ac->gdma_dev; 1885 struct mana_obj_spec wq_spec; 1886 struct mana_obj_spec cq_spec; 1887 struct gdma_queue_spec spec; 1888 struct gdma_context *gc; 1889 struct mana_txq *txq; 1890 struct mana_cq *cq; 1891 u32 txq_size; 1892 u32 cq_size; 1893 int err; 1894 int i; 1895 1896 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp), 1897 GFP_KERNEL); 1898 if (!apc->tx_qp) 1899 return -ENOMEM; 1900 1901 /* The minimum size of the WQE is 32 bytes, hence 1902 * MAX_SEND_BUFFERS_PER_QUEUE represents the maximum number of WQEs 1903 * the SQ can store. This value is then used to size other queues 1904 * to prevent overflow. 1905 */ 1906 txq_size = MAX_SEND_BUFFERS_PER_QUEUE * 32; 1907 BUILD_BUG_ON(!MANA_PAGE_ALIGNED(txq_size)); 1908 1909 cq_size = MAX_SEND_BUFFERS_PER_QUEUE * COMP_ENTRY_SIZE; 1910 cq_size = MANA_PAGE_ALIGN(cq_size); 1911 1912 gc = gd->gdma_context; 1913 1914 for (i = 0; i < apc->num_queues; i++) { 1915 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE; 1916 1917 /* Create SQ */ 1918 txq = &apc->tx_qp[i].txq; 1919 1920 u64_stats_init(&txq->stats.syncp); 1921 txq->ndev = net; 1922 txq->net_txq = netdev_get_tx_queue(net, i); 1923 txq->vp_offset = apc->tx_vp_offset; 1924 skb_queue_head_init(&txq->pending_skbs); 1925 1926 memset(&spec, 0, sizeof(spec)); 1927 spec.type = GDMA_SQ; 1928 spec.monitor_avl_buf = true; 1929 spec.queue_size = txq_size; 1930 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq); 1931 if (err) 1932 goto out; 1933 1934 /* Create SQ's CQ */ 1935 cq = &apc->tx_qp[i].tx_cq; 1936 cq->type = MANA_CQ_TYPE_TX; 1937 1938 cq->txq = txq; 1939 1940 memset(&spec, 0, sizeof(spec)); 1941 spec.type = GDMA_CQ; 1942 spec.monitor_avl_buf = false; 1943 spec.queue_size = cq_size; 1944 spec.cq.callback = mana_schedule_napi; 1945 spec.cq.parent_eq = ac->eqs[i].eq; 1946 spec.cq.context = cq; 1947 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 1948 if (err) 1949 goto out; 1950 1951 memset(&wq_spec, 0, sizeof(wq_spec)); 1952 memset(&cq_spec, 0, sizeof(cq_spec)); 1953 1954 wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle; 1955 wq_spec.queue_size = txq->gdma_sq->queue_size; 1956 1957 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 1958 cq_spec.queue_size = cq->gdma_cq->queue_size; 1959 cq_spec.modr_ctx_id = 0; 1960 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 1961 1962 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ, 1963 &wq_spec, &cq_spec, 1964 &apc->tx_qp[i].tx_object); 1965 1966 if (err) 1967 goto out; 1968 1969 txq->gdma_sq->id = wq_spec.queue_index; 1970 cq->gdma_cq->id = cq_spec.queue_index; 1971 1972 txq->gdma_sq->mem_info.dma_region_handle = 1973 GDMA_INVALID_DMA_REGION; 1974 cq->gdma_cq->mem_info.dma_region_handle = 1975 GDMA_INVALID_DMA_REGION; 1976 1977 txq->gdma_txq_id = txq->gdma_sq->id; 1978 1979 cq->gdma_id = cq->gdma_cq->id; 1980 1981 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 1982 err = -EINVAL; 1983 goto out; 1984 } 1985 1986 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 1987 1988 netif_napi_add_tx(net, &cq->napi, mana_poll); 1989 napi_enable(&cq->napi); 1990 1991 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 1992 } 1993 1994 return 0; 1995 out: 1996 mana_destroy_txq(apc); 1997 return err; 1998 } 1999 2000 static void mana_destroy_rxq(struct mana_port_context *apc, 2001 struct mana_rxq *rxq, bool validate_state) 2002 2003 { 2004 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 2005 struct mana_recv_buf_oob *rx_oob; 2006 struct device *dev = gc->dev; 2007 struct napi_struct *napi; 2008 struct page *page; 2009 int i; 2010 2011 if (!rxq) 2012 return; 2013 2014 napi = &rxq->rx_cq.napi; 2015 2016 if (validate_state) 2017 napi_synchronize(napi); 2018 2019 napi_disable(napi); 2020 2021 xdp_rxq_info_unreg(&rxq->xdp_rxq); 2022 2023 netif_napi_del(napi); 2024 2025 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj); 2026 2027 mana_deinit_cq(apc, &rxq->rx_cq); 2028 2029 if (rxq->xdp_save_va) 2030 put_page(virt_to_head_page(rxq->xdp_save_va)); 2031 2032 for (i = 0; i < rxq->num_rx_buf; i++) { 2033 rx_oob = &rxq->rx_oobs[i]; 2034 2035 if (!rx_oob->buf_va) 2036 continue; 2037 2038 dma_unmap_single(dev, rx_oob->sgl[0].address, 2039 rx_oob->sgl[0].size, DMA_FROM_DEVICE); 2040 2041 page = virt_to_head_page(rx_oob->buf_va); 2042 2043 if (rx_oob->from_pool) 2044 page_pool_put_full_page(rxq->page_pool, page, false); 2045 else 2046 put_page(page); 2047 2048 rx_oob->buf_va = NULL; 2049 } 2050 2051 page_pool_destroy(rxq->page_pool); 2052 2053 if (rxq->gdma_rq) 2054 mana_gd_destroy_queue(gc, rxq->gdma_rq); 2055 2056 kfree(rxq); 2057 } 2058 2059 static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key, 2060 struct mana_rxq *rxq, struct device *dev) 2061 { 2062 struct mana_port_context *mpc = netdev_priv(rxq->ndev); 2063 bool from_pool = false; 2064 dma_addr_t da; 2065 void *va; 2066 2067 if (mpc->rxbufs_pre) 2068 va = mana_get_rxbuf_pre(rxq, &da); 2069 else 2070 va = mana_get_rxfrag(rxq, dev, &da, &from_pool, false); 2071 2072 if (!va) 2073 return -ENOMEM; 2074 2075 rx_oob->buf_va = va; 2076 rx_oob->from_pool = from_pool; 2077 2078 rx_oob->sgl[0].address = da; 2079 rx_oob->sgl[0].size = rxq->datasize; 2080 rx_oob->sgl[0].mem_key = mem_key; 2081 2082 return 0; 2083 } 2084 2085 #define MANA_WQE_HEADER_SIZE 16 2086 #define MANA_WQE_SGE_SIZE 16 2087 2088 static int mana_alloc_rx_wqe(struct mana_port_context *apc, 2089 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size) 2090 { 2091 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 2092 struct mana_recv_buf_oob *rx_oob; 2093 struct device *dev = gc->dev; 2094 u32 buf_idx; 2095 int ret; 2096 2097 WARN_ON(rxq->datasize == 0); 2098 2099 *rxq_size = 0; 2100 *cq_size = 0; 2101 2102 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2103 rx_oob = &rxq->rx_oobs[buf_idx]; 2104 memset(rx_oob, 0, sizeof(*rx_oob)); 2105 2106 rx_oob->num_sge = 1; 2107 2108 ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq, 2109 dev); 2110 if (ret) 2111 return ret; 2112 2113 rx_oob->wqe_req.sgl = rx_oob->sgl; 2114 rx_oob->wqe_req.num_sge = rx_oob->num_sge; 2115 rx_oob->wqe_req.inline_oob_size = 0; 2116 rx_oob->wqe_req.inline_oob_data = NULL; 2117 rx_oob->wqe_req.flags = 0; 2118 rx_oob->wqe_req.client_data_unit = 0; 2119 2120 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE + 2121 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32); 2122 *cq_size += COMP_ENTRY_SIZE; 2123 } 2124 2125 return 0; 2126 } 2127 2128 static int mana_push_wqe(struct mana_rxq *rxq) 2129 { 2130 struct mana_recv_buf_oob *rx_oob; 2131 u32 buf_idx; 2132 int err; 2133 2134 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2135 rx_oob = &rxq->rx_oobs[buf_idx]; 2136 2137 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req, 2138 &rx_oob->wqe_inf); 2139 if (err) 2140 return -ENOSPC; 2141 } 2142 2143 return 0; 2144 } 2145 2146 static int mana_create_page_pool(struct mana_rxq *rxq, struct gdma_context *gc) 2147 { 2148 struct page_pool_params pprm = {}; 2149 int ret; 2150 2151 pprm.pool_size = RX_BUFFERS_PER_QUEUE; 2152 pprm.nid = gc->numa_node; 2153 pprm.napi = &rxq->rx_cq.napi; 2154 pprm.netdev = rxq->ndev; 2155 2156 rxq->page_pool = page_pool_create(&pprm); 2157 2158 if (IS_ERR(rxq->page_pool)) { 2159 ret = PTR_ERR(rxq->page_pool); 2160 rxq->page_pool = NULL; 2161 return ret; 2162 } 2163 2164 return 0; 2165 } 2166 2167 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc, 2168 u32 rxq_idx, struct mana_eq *eq, 2169 struct net_device *ndev) 2170 { 2171 struct gdma_dev *gd = apc->ac->gdma_dev; 2172 struct mana_obj_spec wq_spec; 2173 struct mana_obj_spec cq_spec; 2174 struct gdma_queue_spec spec; 2175 struct mana_cq *cq = NULL; 2176 struct gdma_context *gc; 2177 u32 cq_size, rq_size; 2178 struct mana_rxq *rxq; 2179 int err; 2180 2181 gc = gd->gdma_context; 2182 2183 rxq = kzalloc(struct_size(rxq, rx_oobs, RX_BUFFERS_PER_QUEUE), 2184 GFP_KERNEL); 2185 if (!rxq) 2186 return NULL; 2187 2188 rxq->ndev = ndev; 2189 rxq->num_rx_buf = RX_BUFFERS_PER_QUEUE; 2190 rxq->rxq_idx = rxq_idx; 2191 rxq->rxobj = INVALID_MANA_HANDLE; 2192 2193 mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size, 2194 &rxq->headroom); 2195 2196 /* Create page pool for RX queue */ 2197 err = mana_create_page_pool(rxq, gc); 2198 if (err) { 2199 netdev_err(ndev, "Create page pool err:%d\n", err); 2200 goto out; 2201 } 2202 2203 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size); 2204 if (err) 2205 goto out; 2206 2207 rq_size = MANA_PAGE_ALIGN(rq_size); 2208 cq_size = MANA_PAGE_ALIGN(cq_size); 2209 2210 /* Create RQ */ 2211 memset(&spec, 0, sizeof(spec)); 2212 spec.type = GDMA_RQ; 2213 spec.monitor_avl_buf = true; 2214 spec.queue_size = rq_size; 2215 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq); 2216 if (err) 2217 goto out; 2218 2219 /* Create RQ's CQ */ 2220 cq = &rxq->rx_cq; 2221 cq->type = MANA_CQ_TYPE_RX; 2222 cq->rxq = rxq; 2223 2224 memset(&spec, 0, sizeof(spec)); 2225 spec.type = GDMA_CQ; 2226 spec.monitor_avl_buf = false; 2227 spec.queue_size = cq_size; 2228 spec.cq.callback = mana_schedule_napi; 2229 spec.cq.parent_eq = eq->eq; 2230 spec.cq.context = cq; 2231 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 2232 if (err) 2233 goto out; 2234 2235 memset(&wq_spec, 0, sizeof(wq_spec)); 2236 memset(&cq_spec, 0, sizeof(cq_spec)); 2237 wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle; 2238 wq_spec.queue_size = rxq->gdma_rq->queue_size; 2239 2240 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 2241 cq_spec.queue_size = cq->gdma_cq->queue_size; 2242 cq_spec.modr_ctx_id = 0; 2243 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 2244 2245 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ, 2246 &wq_spec, &cq_spec, &rxq->rxobj); 2247 if (err) 2248 goto out; 2249 2250 rxq->gdma_rq->id = wq_spec.queue_index; 2251 cq->gdma_cq->id = cq_spec.queue_index; 2252 2253 rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2254 cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2255 2256 rxq->gdma_id = rxq->gdma_rq->id; 2257 cq->gdma_id = cq->gdma_cq->id; 2258 2259 err = mana_push_wqe(rxq); 2260 if (err) 2261 goto out; 2262 2263 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 2264 err = -EINVAL; 2265 goto out; 2266 } 2267 2268 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2269 2270 netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1); 2271 2272 WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx, 2273 cq->napi.napi_id)); 2274 WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL, 2275 rxq->page_pool)); 2276 2277 napi_enable(&cq->napi); 2278 2279 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2280 out: 2281 if (!err) 2282 return rxq; 2283 2284 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err); 2285 2286 mana_destroy_rxq(apc, rxq, false); 2287 2288 if (cq) 2289 mana_deinit_cq(apc, cq); 2290 2291 return NULL; 2292 } 2293 2294 static int mana_add_rx_queues(struct mana_port_context *apc, 2295 struct net_device *ndev) 2296 { 2297 struct mana_context *ac = apc->ac; 2298 struct mana_rxq *rxq; 2299 int err = 0; 2300 int i; 2301 2302 for (i = 0; i < apc->num_queues; i++) { 2303 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev); 2304 if (!rxq) { 2305 err = -ENOMEM; 2306 goto out; 2307 } 2308 2309 u64_stats_init(&rxq->stats.syncp); 2310 2311 apc->rxqs[i] = rxq; 2312 } 2313 2314 apc->default_rxobj = apc->rxqs[0]->rxobj; 2315 out: 2316 return err; 2317 } 2318 2319 static void mana_destroy_vport(struct mana_port_context *apc) 2320 { 2321 struct gdma_dev *gd = apc->ac->gdma_dev; 2322 struct mana_rxq *rxq; 2323 u32 rxq_idx; 2324 2325 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 2326 rxq = apc->rxqs[rxq_idx]; 2327 if (!rxq) 2328 continue; 2329 2330 mana_destroy_rxq(apc, rxq, true); 2331 apc->rxqs[rxq_idx] = NULL; 2332 } 2333 2334 mana_destroy_txq(apc); 2335 mana_uncfg_vport(apc); 2336 2337 if (gd->gdma_context->is_pf) 2338 mana_pf_deregister_hw_vport(apc); 2339 } 2340 2341 static int mana_create_vport(struct mana_port_context *apc, 2342 struct net_device *net) 2343 { 2344 struct gdma_dev *gd = apc->ac->gdma_dev; 2345 int err; 2346 2347 apc->default_rxobj = INVALID_MANA_HANDLE; 2348 2349 if (gd->gdma_context->is_pf) { 2350 err = mana_pf_register_hw_vport(apc); 2351 if (err) 2352 return err; 2353 } 2354 2355 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell); 2356 if (err) 2357 return err; 2358 2359 return mana_create_txq(apc, net); 2360 } 2361 2362 static int mana_rss_table_alloc(struct mana_port_context *apc) 2363 { 2364 if (!apc->indir_table_sz) { 2365 netdev_err(apc->ndev, 2366 "Indirection table size not set for vPort %d\n", 2367 apc->port_idx); 2368 return -EINVAL; 2369 } 2370 2371 apc->indir_table = kcalloc(apc->indir_table_sz, sizeof(u32), GFP_KERNEL); 2372 if (!apc->indir_table) 2373 return -ENOMEM; 2374 2375 apc->rxobj_table = kcalloc(apc->indir_table_sz, sizeof(mana_handle_t), GFP_KERNEL); 2376 if (!apc->rxobj_table) { 2377 kfree(apc->indir_table); 2378 return -ENOMEM; 2379 } 2380 2381 return 0; 2382 } 2383 2384 static void mana_rss_table_init(struct mana_port_context *apc) 2385 { 2386 int i; 2387 2388 for (i = 0; i < apc->indir_table_sz; i++) 2389 apc->indir_table[i] = 2390 ethtool_rxfh_indir_default(i, apc->num_queues); 2391 } 2392 2393 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx, 2394 bool update_hash, bool update_tab) 2395 { 2396 u32 queue_idx; 2397 int err; 2398 int i; 2399 2400 if (update_tab) { 2401 for (i = 0; i < apc->indir_table_sz; i++) { 2402 queue_idx = apc->indir_table[i]; 2403 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj; 2404 } 2405 } 2406 2407 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab); 2408 if (err) 2409 return err; 2410 2411 mana_fence_rqs(apc); 2412 2413 return 0; 2414 } 2415 2416 void mana_query_gf_stats(struct mana_port_context *apc) 2417 { 2418 struct mana_query_gf_stat_resp resp = {}; 2419 struct mana_query_gf_stat_req req = {}; 2420 struct net_device *ndev = apc->ndev; 2421 int err; 2422 2423 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_GF_STAT, 2424 sizeof(req), sizeof(resp)); 2425 req.req_stats = STATISTICS_FLAGS_RX_DISCARDS_NO_WQE | 2426 STATISTICS_FLAGS_RX_ERRORS_VPORT_DISABLED | 2427 STATISTICS_FLAGS_HC_RX_BYTES | 2428 STATISTICS_FLAGS_HC_RX_UCAST_PACKETS | 2429 STATISTICS_FLAGS_HC_RX_UCAST_BYTES | 2430 STATISTICS_FLAGS_HC_RX_MCAST_PACKETS | 2431 STATISTICS_FLAGS_HC_RX_MCAST_BYTES | 2432 STATISTICS_FLAGS_HC_RX_BCAST_PACKETS | 2433 STATISTICS_FLAGS_HC_RX_BCAST_BYTES | 2434 STATISTICS_FLAGS_TX_ERRORS_GF_DISABLED | 2435 STATISTICS_FLAGS_TX_ERRORS_VPORT_DISABLED | 2436 STATISTICS_FLAGS_TX_ERRORS_INVAL_VPORT_OFFSET_PACKETS | 2437 STATISTICS_FLAGS_TX_ERRORS_VLAN_ENFORCEMENT | 2438 STATISTICS_FLAGS_TX_ERRORS_ETH_TYPE_ENFORCEMENT | 2439 STATISTICS_FLAGS_TX_ERRORS_SA_ENFORCEMENT | 2440 STATISTICS_FLAGS_TX_ERRORS_SQPDID_ENFORCEMENT | 2441 STATISTICS_FLAGS_TX_ERRORS_CQPDID_ENFORCEMENT | 2442 STATISTICS_FLAGS_TX_ERRORS_MTU_VIOLATION | 2443 STATISTICS_FLAGS_TX_ERRORS_INVALID_OOB | 2444 STATISTICS_FLAGS_HC_TX_BYTES | 2445 STATISTICS_FLAGS_HC_TX_UCAST_PACKETS | 2446 STATISTICS_FLAGS_HC_TX_UCAST_BYTES | 2447 STATISTICS_FLAGS_HC_TX_MCAST_PACKETS | 2448 STATISTICS_FLAGS_HC_TX_MCAST_BYTES | 2449 STATISTICS_FLAGS_HC_TX_BCAST_PACKETS | 2450 STATISTICS_FLAGS_HC_TX_BCAST_BYTES | 2451 STATISTICS_FLAGS_TX_ERRORS_GDMA_ERROR; 2452 2453 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 2454 sizeof(resp)); 2455 if (err) { 2456 netdev_err(ndev, "Failed to query GF stats: %d\n", err); 2457 return; 2458 } 2459 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_GF_STAT, 2460 sizeof(resp)); 2461 if (err || resp.hdr.status) { 2462 netdev_err(ndev, "Failed to query GF stats: %d, 0x%x\n", err, 2463 resp.hdr.status); 2464 return; 2465 } 2466 2467 apc->eth_stats.hc_rx_discards_no_wqe = resp.rx_discards_nowqe; 2468 apc->eth_stats.hc_rx_err_vport_disabled = resp.rx_err_vport_disabled; 2469 apc->eth_stats.hc_rx_bytes = resp.hc_rx_bytes; 2470 apc->eth_stats.hc_rx_ucast_pkts = resp.hc_rx_ucast_pkts; 2471 apc->eth_stats.hc_rx_ucast_bytes = resp.hc_rx_ucast_bytes; 2472 apc->eth_stats.hc_rx_bcast_pkts = resp.hc_rx_bcast_pkts; 2473 apc->eth_stats.hc_rx_bcast_bytes = resp.hc_rx_bcast_bytes; 2474 apc->eth_stats.hc_rx_mcast_pkts = resp.hc_rx_mcast_pkts; 2475 apc->eth_stats.hc_rx_mcast_bytes = resp.hc_rx_mcast_bytes; 2476 apc->eth_stats.hc_tx_err_gf_disabled = resp.tx_err_gf_disabled; 2477 apc->eth_stats.hc_tx_err_vport_disabled = resp.tx_err_vport_disabled; 2478 apc->eth_stats.hc_tx_err_inval_vportoffset_pkt = 2479 resp.tx_err_inval_vport_offset_pkt; 2480 apc->eth_stats.hc_tx_err_vlan_enforcement = 2481 resp.tx_err_vlan_enforcement; 2482 apc->eth_stats.hc_tx_err_eth_type_enforcement = 2483 resp.tx_err_ethtype_enforcement; 2484 apc->eth_stats.hc_tx_err_sa_enforcement = resp.tx_err_SA_enforcement; 2485 apc->eth_stats.hc_tx_err_sqpdid_enforcement = 2486 resp.tx_err_SQPDID_enforcement; 2487 apc->eth_stats.hc_tx_err_cqpdid_enforcement = 2488 resp.tx_err_CQPDID_enforcement; 2489 apc->eth_stats.hc_tx_err_mtu_violation = resp.tx_err_mtu_violation; 2490 apc->eth_stats.hc_tx_err_inval_oob = resp.tx_err_inval_oob; 2491 apc->eth_stats.hc_tx_bytes = resp.hc_tx_bytes; 2492 apc->eth_stats.hc_tx_ucast_pkts = resp.hc_tx_ucast_pkts; 2493 apc->eth_stats.hc_tx_ucast_bytes = resp.hc_tx_ucast_bytes; 2494 apc->eth_stats.hc_tx_bcast_pkts = resp.hc_tx_bcast_pkts; 2495 apc->eth_stats.hc_tx_bcast_bytes = resp.hc_tx_bcast_bytes; 2496 apc->eth_stats.hc_tx_mcast_pkts = resp.hc_tx_mcast_pkts; 2497 apc->eth_stats.hc_tx_mcast_bytes = resp.hc_tx_mcast_bytes; 2498 apc->eth_stats.hc_tx_err_gdma = resp.tx_err_gdma; 2499 } 2500 2501 static int mana_init_port(struct net_device *ndev) 2502 { 2503 struct mana_port_context *apc = netdev_priv(ndev); 2504 u32 max_txq, max_rxq, max_queues; 2505 int port_idx = apc->port_idx; 2506 int err; 2507 2508 err = mana_init_port_context(apc); 2509 if (err) 2510 return err; 2511 2512 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq, 2513 &apc->indir_table_sz); 2514 if (err) { 2515 netdev_err(ndev, "Failed to query info for vPort %d\n", 2516 port_idx); 2517 goto reset_apc; 2518 } 2519 2520 max_queues = min_t(u32, max_txq, max_rxq); 2521 if (apc->max_queues > max_queues) 2522 apc->max_queues = max_queues; 2523 2524 if (apc->num_queues > apc->max_queues) 2525 apc->num_queues = apc->max_queues; 2526 2527 eth_hw_addr_set(ndev, apc->mac_addr); 2528 2529 return 0; 2530 2531 reset_apc: 2532 mana_cleanup_port_context(apc); 2533 return err; 2534 } 2535 2536 int mana_alloc_queues(struct net_device *ndev) 2537 { 2538 struct mana_port_context *apc = netdev_priv(ndev); 2539 struct gdma_dev *gd = apc->ac->gdma_dev; 2540 int err; 2541 2542 err = mana_create_vport(apc, ndev); 2543 if (err) 2544 return err; 2545 2546 err = netif_set_real_num_tx_queues(ndev, apc->num_queues); 2547 if (err) 2548 goto destroy_vport; 2549 2550 err = mana_add_rx_queues(apc, ndev); 2551 if (err) 2552 goto destroy_vport; 2553 2554 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE; 2555 2556 err = netif_set_real_num_rx_queues(ndev, apc->num_queues); 2557 if (err) 2558 goto destroy_vport; 2559 2560 mana_rss_table_init(apc); 2561 2562 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true); 2563 if (err) 2564 goto destroy_vport; 2565 2566 if (gd->gdma_context->is_pf) { 2567 err = mana_pf_register_filter(apc); 2568 if (err) 2569 goto destroy_vport; 2570 } 2571 2572 mana_chn_setxdp(apc, mana_xdp_get(apc)); 2573 2574 return 0; 2575 2576 destroy_vport: 2577 mana_destroy_vport(apc); 2578 return err; 2579 } 2580 2581 int mana_attach(struct net_device *ndev) 2582 { 2583 struct mana_port_context *apc = netdev_priv(ndev); 2584 int err; 2585 2586 ASSERT_RTNL(); 2587 2588 err = mana_init_port(ndev); 2589 if (err) 2590 return err; 2591 2592 if (apc->port_st_save) { 2593 err = mana_alloc_queues(ndev); 2594 if (err) { 2595 mana_cleanup_port_context(apc); 2596 return err; 2597 } 2598 } 2599 2600 apc->port_is_up = apc->port_st_save; 2601 2602 /* Ensure port state updated before txq state */ 2603 smp_wmb(); 2604 2605 if (apc->port_is_up) 2606 netif_carrier_on(ndev); 2607 2608 netif_device_attach(ndev); 2609 2610 return 0; 2611 } 2612 2613 static int mana_dealloc_queues(struct net_device *ndev) 2614 { 2615 struct mana_port_context *apc = netdev_priv(ndev); 2616 unsigned long timeout = jiffies + 120 * HZ; 2617 struct gdma_dev *gd = apc->ac->gdma_dev; 2618 struct mana_txq *txq; 2619 struct sk_buff *skb; 2620 int i, err; 2621 u32 tsleep; 2622 2623 if (apc->port_is_up) 2624 return -EINVAL; 2625 2626 mana_chn_setxdp(apc, NULL); 2627 2628 if (gd->gdma_context->is_pf) 2629 mana_pf_deregister_filter(apc); 2630 2631 /* No packet can be transmitted now since apc->port_is_up is false. 2632 * There is still a tiny chance that mana_poll_tx_cq() can re-enable 2633 * a txq because it may not timely see apc->port_is_up being cleared 2634 * to false, but it doesn't matter since mana_start_xmit() drops any 2635 * new packets due to apc->port_is_up being false. 2636 * 2637 * Drain all the in-flight TX packets. 2638 * A timeout of 120 seconds for all the queues is used. 2639 * This will break the while loop when h/w is not responding. 2640 * This value of 120 has been decided here considering max 2641 * number of queues. 2642 */ 2643 2644 for (i = 0; i < apc->num_queues; i++) { 2645 txq = &apc->tx_qp[i].txq; 2646 tsleep = 1000; 2647 while (atomic_read(&txq->pending_sends) > 0 && 2648 time_before(jiffies, timeout)) { 2649 usleep_range(tsleep, tsleep + 1000); 2650 tsleep <<= 1; 2651 } 2652 if (atomic_read(&txq->pending_sends)) { 2653 err = pcie_flr(to_pci_dev(gd->gdma_context->dev)); 2654 if (err) { 2655 netdev_err(ndev, "flr failed %d with %d pkts pending in txq %u\n", 2656 err, atomic_read(&txq->pending_sends), 2657 txq->gdma_txq_id); 2658 } 2659 break; 2660 } 2661 } 2662 2663 for (i = 0; i < apc->num_queues; i++) { 2664 txq = &apc->tx_qp[i].txq; 2665 while ((skb = skb_dequeue(&txq->pending_skbs))) { 2666 mana_unmap_skb(skb, apc); 2667 dev_kfree_skb_any(skb); 2668 } 2669 atomic_set(&txq->pending_sends, 0); 2670 } 2671 /* We're 100% sure the queues can no longer be woken up, because 2672 * we're sure now mana_poll_tx_cq() can't be running. 2673 */ 2674 2675 apc->rss_state = TRI_STATE_FALSE; 2676 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false); 2677 if (err) { 2678 netdev_err(ndev, "Failed to disable vPort: %d\n", err); 2679 return err; 2680 } 2681 2682 mana_destroy_vport(apc); 2683 2684 return 0; 2685 } 2686 2687 int mana_detach(struct net_device *ndev, bool from_close) 2688 { 2689 struct mana_port_context *apc = netdev_priv(ndev); 2690 int err; 2691 2692 ASSERT_RTNL(); 2693 2694 apc->port_st_save = apc->port_is_up; 2695 apc->port_is_up = false; 2696 2697 /* Ensure port state updated before txq state */ 2698 smp_wmb(); 2699 2700 netif_tx_disable(ndev); 2701 netif_carrier_off(ndev); 2702 2703 if (apc->port_st_save) { 2704 err = mana_dealloc_queues(ndev); 2705 if (err) 2706 return err; 2707 } 2708 2709 if (!from_close) { 2710 netif_device_detach(ndev); 2711 mana_cleanup_port_context(apc); 2712 } 2713 2714 return 0; 2715 } 2716 2717 static int mana_probe_port(struct mana_context *ac, int port_idx, 2718 struct net_device **ndev_storage) 2719 { 2720 struct gdma_context *gc = ac->gdma_dev->gdma_context; 2721 struct mana_port_context *apc; 2722 struct net_device *ndev; 2723 int err; 2724 2725 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context), 2726 gc->max_num_queues); 2727 if (!ndev) 2728 return -ENOMEM; 2729 2730 *ndev_storage = ndev; 2731 2732 apc = netdev_priv(ndev); 2733 apc->ac = ac; 2734 apc->ndev = ndev; 2735 apc->max_queues = gc->max_num_queues; 2736 apc->num_queues = gc->max_num_queues; 2737 apc->port_handle = INVALID_MANA_HANDLE; 2738 apc->pf_filter_handle = INVALID_MANA_HANDLE; 2739 apc->port_idx = port_idx; 2740 2741 mutex_init(&apc->vport_mutex); 2742 apc->vport_use_count = 0; 2743 2744 ndev->netdev_ops = &mana_devops; 2745 ndev->ethtool_ops = &mana_ethtool_ops; 2746 ndev->mtu = ETH_DATA_LEN; 2747 ndev->max_mtu = gc->adapter_mtu - ETH_HLEN; 2748 ndev->min_mtu = ETH_MIN_MTU; 2749 ndev->needed_headroom = MANA_HEADROOM; 2750 ndev->dev_port = port_idx; 2751 SET_NETDEV_DEV(ndev, gc->dev); 2752 2753 netif_carrier_off(ndev); 2754 2755 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE); 2756 2757 err = mana_init_port(ndev); 2758 if (err) 2759 goto free_net; 2760 2761 err = mana_rss_table_alloc(apc); 2762 if (err) 2763 goto reset_apc; 2764 2765 netdev_lockdep_set_classes(ndev); 2766 2767 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 2768 ndev->hw_features |= NETIF_F_RXCSUM; 2769 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6; 2770 ndev->hw_features |= NETIF_F_RXHASH; 2771 ndev->features = ndev->hw_features | NETIF_F_HW_VLAN_CTAG_TX | 2772 NETIF_F_HW_VLAN_CTAG_RX; 2773 ndev->vlan_features = ndev->features; 2774 xdp_set_features_flag(ndev, NETDEV_XDP_ACT_BASIC | 2775 NETDEV_XDP_ACT_REDIRECT | 2776 NETDEV_XDP_ACT_NDO_XMIT); 2777 2778 err = register_netdev(ndev); 2779 if (err) { 2780 netdev_err(ndev, "Unable to register netdev.\n"); 2781 goto free_indir; 2782 } 2783 2784 return 0; 2785 2786 free_indir: 2787 mana_cleanup_indir_table(apc); 2788 reset_apc: 2789 mana_cleanup_port_context(apc); 2790 free_net: 2791 *ndev_storage = NULL; 2792 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err); 2793 free_netdev(ndev); 2794 return err; 2795 } 2796 2797 static void adev_release(struct device *dev) 2798 { 2799 struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev); 2800 2801 kfree(madev); 2802 } 2803 2804 static void remove_adev(struct gdma_dev *gd) 2805 { 2806 struct auxiliary_device *adev = gd->adev; 2807 int id = adev->id; 2808 2809 auxiliary_device_delete(adev); 2810 auxiliary_device_uninit(adev); 2811 2812 mana_adev_idx_free(id); 2813 gd->adev = NULL; 2814 } 2815 2816 static int add_adev(struct gdma_dev *gd) 2817 { 2818 struct auxiliary_device *adev; 2819 struct mana_adev *madev; 2820 int ret; 2821 2822 madev = kzalloc(sizeof(*madev), GFP_KERNEL); 2823 if (!madev) 2824 return -ENOMEM; 2825 2826 adev = &madev->adev; 2827 ret = mana_adev_idx_alloc(); 2828 if (ret < 0) 2829 goto idx_fail; 2830 adev->id = ret; 2831 2832 adev->name = "rdma"; 2833 adev->dev.parent = gd->gdma_context->dev; 2834 adev->dev.release = adev_release; 2835 madev->mdev = gd; 2836 2837 ret = auxiliary_device_init(adev); 2838 if (ret) 2839 goto init_fail; 2840 2841 /* madev is owned by the auxiliary device */ 2842 madev = NULL; 2843 ret = auxiliary_device_add(adev); 2844 if (ret) 2845 goto add_fail; 2846 2847 gd->adev = adev; 2848 return 0; 2849 2850 add_fail: 2851 auxiliary_device_uninit(adev); 2852 2853 init_fail: 2854 mana_adev_idx_free(adev->id); 2855 2856 idx_fail: 2857 kfree(madev); 2858 2859 return ret; 2860 } 2861 2862 int mana_probe(struct gdma_dev *gd, bool resuming) 2863 { 2864 struct gdma_context *gc = gd->gdma_context; 2865 struct mana_context *ac = gd->driver_data; 2866 struct device *dev = gc->dev; 2867 u16 num_ports = 0; 2868 int err; 2869 int i; 2870 2871 dev_info(dev, 2872 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n", 2873 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION); 2874 2875 err = mana_gd_register_device(gd); 2876 if (err) 2877 return err; 2878 2879 if (!resuming) { 2880 ac = kzalloc(sizeof(*ac), GFP_KERNEL); 2881 if (!ac) 2882 return -ENOMEM; 2883 2884 ac->gdma_dev = gd; 2885 gd->driver_data = ac; 2886 } 2887 2888 err = mana_create_eq(ac); 2889 if (err) 2890 goto out; 2891 2892 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION, 2893 MANA_MICRO_VERSION, &num_ports); 2894 if (err) 2895 goto out; 2896 2897 if (!resuming) { 2898 ac->num_ports = num_ports; 2899 } else { 2900 if (ac->num_ports != num_ports) { 2901 dev_err(dev, "The number of vPorts changed: %d->%d\n", 2902 ac->num_ports, num_ports); 2903 err = -EPROTO; 2904 goto out; 2905 } 2906 } 2907 2908 if (ac->num_ports == 0) 2909 dev_err(dev, "Failed to detect any vPort\n"); 2910 2911 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV) 2912 ac->num_ports = MAX_PORTS_IN_MANA_DEV; 2913 2914 if (!resuming) { 2915 for (i = 0; i < ac->num_ports; i++) { 2916 err = mana_probe_port(ac, i, &ac->ports[i]); 2917 /* we log the port for which the probe failed and stop 2918 * probes for subsequent ports. 2919 * Note that we keep running ports, for which the probes 2920 * were successful, unless add_adev fails too 2921 */ 2922 if (err) { 2923 dev_err(dev, "Probe Failed for port %d\n", i); 2924 break; 2925 } 2926 } 2927 } else { 2928 for (i = 0; i < ac->num_ports; i++) { 2929 rtnl_lock(); 2930 err = mana_attach(ac->ports[i]); 2931 rtnl_unlock(); 2932 /* we log the port for which the attach failed and stop 2933 * attach for subsequent ports 2934 * Note that we keep running ports, for which the attach 2935 * were successful, unless add_adev fails too 2936 */ 2937 if (err) { 2938 dev_err(dev, "Attach Failed for port %d\n", i); 2939 break; 2940 } 2941 } 2942 } 2943 2944 err = add_adev(gd); 2945 out: 2946 if (err) 2947 mana_remove(gd, false); 2948 2949 return err; 2950 } 2951 2952 void mana_remove(struct gdma_dev *gd, bool suspending) 2953 { 2954 struct gdma_context *gc = gd->gdma_context; 2955 struct mana_context *ac = gd->driver_data; 2956 struct mana_port_context *apc; 2957 struct device *dev = gc->dev; 2958 struct net_device *ndev; 2959 int err; 2960 int i; 2961 2962 /* adev currently doesn't support suspending, always remove it */ 2963 if (gd->adev) 2964 remove_adev(gd); 2965 2966 for (i = 0; i < ac->num_ports; i++) { 2967 ndev = ac->ports[i]; 2968 apc = netdev_priv(ndev); 2969 if (!ndev) { 2970 if (i == 0) 2971 dev_err(dev, "No net device to remove\n"); 2972 goto out; 2973 } 2974 2975 /* All cleanup actions should stay after rtnl_lock(), otherwise 2976 * other functions may access partially cleaned up data. 2977 */ 2978 rtnl_lock(); 2979 2980 err = mana_detach(ndev, false); 2981 if (err) 2982 netdev_err(ndev, "Failed to detach vPort %d: %d\n", 2983 i, err); 2984 2985 if (suspending) { 2986 /* No need to unregister the ndev. */ 2987 rtnl_unlock(); 2988 continue; 2989 } 2990 2991 unregister_netdevice(ndev); 2992 mana_cleanup_indir_table(apc); 2993 2994 rtnl_unlock(); 2995 2996 free_netdev(ndev); 2997 } 2998 2999 mana_destroy_eq(ac); 3000 out: 3001 mana_gd_deregister_device(gd); 3002 3003 if (suspending) 3004 return; 3005 3006 gd->driver_data = NULL; 3007 gd->gdma_context = NULL; 3008 kfree(ac); 3009 } 3010 3011 struct net_device *mana_get_primary_netdev_rcu(struct mana_context *ac, u32 port_index) 3012 { 3013 struct net_device *ndev; 3014 3015 RCU_LOCKDEP_WARN(!rcu_read_lock_held(), 3016 "Taking primary netdev without holding the RCU read lock"); 3017 if (port_index >= ac->num_ports) 3018 return NULL; 3019 3020 /* When mana is used in netvsc, the upper netdevice should be returned. */ 3021 if (ac->ports[port_index]->flags & IFF_SLAVE) 3022 ndev = netdev_master_upper_dev_get_rcu(ac->ports[port_index]); 3023 else 3024 ndev = ac->ports[port_index]; 3025 3026 return ndev; 3027 } 3028 EXPORT_SYMBOL_NS(mana_get_primary_netdev_rcu, NET_MANA); 3029