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 12 #include <net/checksum.h> 13 #include <net/ip6_checksum.h> 14 15 #include <net/mana/mana.h> 16 #include <net/mana/mana_auxiliary.h> 17 18 static DEFINE_IDA(mana_adev_ida); 19 20 static int mana_adev_idx_alloc(void) 21 { 22 return ida_alloc(&mana_adev_ida, GFP_KERNEL); 23 } 24 25 static void mana_adev_idx_free(int idx) 26 { 27 ida_free(&mana_adev_ida, idx); 28 } 29 30 /* Microsoft Azure Network Adapter (MANA) functions */ 31 32 static int mana_open(struct net_device *ndev) 33 { 34 struct mana_port_context *apc = netdev_priv(ndev); 35 int err; 36 37 err = mana_alloc_queues(ndev); 38 if (err) 39 return err; 40 41 apc->port_is_up = true; 42 43 /* Ensure port state updated before txq state */ 44 smp_wmb(); 45 46 netif_carrier_on(ndev); 47 netif_tx_wake_all_queues(ndev); 48 49 return 0; 50 } 51 52 static int mana_close(struct net_device *ndev) 53 { 54 struct mana_port_context *apc = netdev_priv(ndev); 55 56 if (!apc->port_is_up) 57 return 0; 58 59 return mana_detach(ndev, true); 60 } 61 62 static bool mana_can_tx(struct gdma_queue *wq) 63 { 64 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE; 65 } 66 67 static unsigned int mana_checksum_info(struct sk_buff *skb) 68 { 69 if (skb->protocol == htons(ETH_P_IP)) { 70 struct iphdr *ip = ip_hdr(skb); 71 72 if (ip->protocol == IPPROTO_TCP) 73 return IPPROTO_TCP; 74 75 if (ip->protocol == IPPROTO_UDP) 76 return IPPROTO_UDP; 77 } else if (skb->protocol == htons(ETH_P_IPV6)) { 78 struct ipv6hdr *ip6 = ipv6_hdr(skb); 79 80 if (ip6->nexthdr == IPPROTO_TCP) 81 return IPPROTO_TCP; 82 83 if (ip6->nexthdr == IPPROTO_UDP) 84 return IPPROTO_UDP; 85 } 86 87 /* No csum offloading */ 88 return 0; 89 } 90 91 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc, 92 struct mana_tx_package *tp) 93 { 94 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 95 struct gdma_dev *gd = apc->ac->gdma_dev; 96 struct gdma_context *gc; 97 struct device *dev; 98 skb_frag_t *frag; 99 dma_addr_t da; 100 int i; 101 102 gc = gd->gdma_context; 103 dev = gc->dev; 104 da = dma_map_single(dev, skb->data, skb_headlen(skb), DMA_TO_DEVICE); 105 106 if (dma_mapping_error(dev, da)) 107 return -ENOMEM; 108 109 ash->dma_handle[0] = da; 110 ash->size[0] = skb_headlen(skb); 111 112 tp->wqe_req.sgl[0].address = ash->dma_handle[0]; 113 tp->wqe_req.sgl[0].mem_key = gd->gpa_mkey; 114 tp->wqe_req.sgl[0].size = ash->size[0]; 115 116 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 117 frag = &skb_shinfo(skb)->frags[i]; 118 da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag), 119 DMA_TO_DEVICE); 120 121 if (dma_mapping_error(dev, da)) 122 goto frag_err; 123 124 ash->dma_handle[i + 1] = da; 125 ash->size[i + 1] = skb_frag_size(frag); 126 127 tp->wqe_req.sgl[i + 1].address = ash->dma_handle[i + 1]; 128 tp->wqe_req.sgl[i + 1].mem_key = gd->gpa_mkey; 129 tp->wqe_req.sgl[i + 1].size = ash->size[i + 1]; 130 } 131 132 return 0; 133 134 frag_err: 135 for (i = i - 1; i >= 0; i--) 136 dma_unmap_page(dev, ash->dma_handle[i + 1], ash->size[i + 1], 137 DMA_TO_DEVICE); 138 139 dma_unmap_single(dev, ash->dma_handle[0], ash->size[0], DMA_TO_DEVICE); 140 141 return -ENOMEM; 142 } 143 144 netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev) 145 { 146 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT; 147 struct mana_port_context *apc = netdev_priv(ndev); 148 u16 txq_idx = skb_get_queue_mapping(skb); 149 struct gdma_dev *gd = apc->ac->gdma_dev; 150 bool ipv4 = false, ipv6 = false; 151 struct mana_tx_package pkg = {}; 152 struct netdev_queue *net_txq; 153 struct mana_stats_tx *tx_stats; 154 struct gdma_queue *gdma_sq; 155 unsigned int csum_type; 156 struct mana_txq *txq; 157 struct mana_cq *cq; 158 int err, len; 159 u16 ihs; 160 161 if (unlikely(!apc->port_is_up)) 162 goto tx_drop; 163 164 if (skb_cow_head(skb, MANA_HEADROOM)) 165 goto tx_drop_count; 166 167 txq = &apc->tx_qp[txq_idx].txq; 168 gdma_sq = txq->gdma_sq; 169 cq = &apc->tx_qp[txq_idx].tx_cq; 170 tx_stats = &txq->stats; 171 172 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id; 173 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame; 174 175 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) { 176 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset; 177 pkt_fmt = MANA_LONG_PKT_FMT; 178 } else { 179 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset; 180 } 181 182 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt; 183 184 if (pkt_fmt == MANA_SHORT_PKT_FMT) { 185 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob); 186 u64_stats_update_begin(&tx_stats->syncp); 187 tx_stats->short_pkt_fmt++; 188 u64_stats_update_end(&tx_stats->syncp); 189 } else { 190 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob); 191 u64_stats_update_begin(&tx_stats->syncp); 192 tx_stats->long_pkt_fmt++; 193 u64_stats_update_end(&tx_stats->syncp); 194 } 195 196 pkg.wqe_req.inline_oob_data = &pkg.tx_oob; 197 pkg.wqe_req.flags = 0; 198 pkg.wqe_req.client_data_unit = 0; 199 200 pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags; 201 WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES); 202 203 if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) { 204 pkg.wqe_req.sgl = pkg.sgl_array; 205 } else { 206 pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge, 207 sizeof(struct gdma_sge), 208 GFP_ATOMIC); 209 if (!pkg.sgl_ptr) 210 goto tx_drop_count; 211 212 pkg.wqe_req.sgl = pkg.sgl_ptr; 213 } 214 215 if (skb->protocol == htons(ETH_P_IP)) 216 ipv4 = true; 217 else if (skb->protocol == htons(ETH_P_IPV6)) 218 ipv6 = true; 219 220 if (skb_is_gso(skb)) { 221 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 222 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 223 224 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 225 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 226 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 227 228 pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size; 229 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0; 230 if (ipv4) { 231 ip_hdr(skb)->tot_len = 0; 232 ip_hdr(skb)->check = 0; 233 tcp_hdr(skb)->check = 234 ~csum_tcpudp_magic(ip_hdr(skb)->saddr, 235 ip_hdr(skb)->daddr, 0, 236 IPPROTO_TCP, 0); 237 } else { 238 ipv6_hdr(skb)->payload_len = 0; 239 tcp_hdr(skb)->check = 240 ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr, 241 &ipv6_hdr(skb)->daddr, 0, 242 IPPROTO_TCP, 0); 243 } 244 245 if (skb->encapsulation) { 246 ihs = skb_inner_tcp_all_headers(skb); 247 u64_stats_update_begin(&tx_stats->syncp); 248 tx_stats->tso_inner_packets++; 249 tx_stats->tso_inner_bytes += skb->len - ihs; 250 u64_stats_update_end(&tx_stats->syncp); 251 } else { 252 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) { 253 ihs = skb_transport_offset(skb) + sizeof(struct udphdr); 254 } else { 255 ihs = skb_tcp_all_headers(skb); 256 if (ipv6_has_hopopt_jumbo(skb)) 257 ihs -= sizeof(struct hop_jumbo_hdr); 258 } 259 260 u64_stats_update_begin(&tx_stats->syncp); 261 tx_stats->tso_packets++; 262 tx_stats->tso_bytes += skb->len - ihs; 263 u64_stats_update_end(&tx_stats->syncp); 264 } 265 266 } else if (skb->ip_summed == CHECKSUM_PARTIAL) { 267 csum_type = mana_checksum_info(skb); 268 269 u64_stats_update_begin(&tx_stats->syncp); 270 tx_stats->csum_partial++; 271 u64_stats_update_end(&tx_stats->syncp); 272 273 if (csum_type == IPPROTO_TCP) { 274 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 275 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 276 277 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 278 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 279 280 } else if (csum_type == IPPROTO_UDP) { 281 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 282 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 283 284 pkg.tx_oob.s_oob.comp_udp_csum = 1; 285 } else { 286 /* Can't do offload of this type of checksum */ 287 if (skb_checksum_help(skb)) 288 goto free_sgl_ptr; 289 } 290 } 291 292 if (mana_map_skb(skb, apc, &pkg)) { 293 u64_stats_update_begin(&tx_stats->syncp); 294 tx_stats->mana_map_err++; 295 u64_stats_update_end(&tx_stats->syncp); 296 goto free_sgl_ptr; 297 } 298 299 skb_queue_tail(&txq->pending_skbs, skb); 300 301 len = skb->len; 302 net_txq = netdev_get_tx_queue(ndev, txq_idx); 303 304 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req, 305 (struct gdma_posted_wqe_info *)skb->cb); 306 if (!mana_can_tx(gdma_sq)) { 307 netif_tx_stop_queue(net_txq); 308 apc->eth_stats.stop_queue++; 309 } 310 311 if (err) { 312 (void)skb_dequeue_tail(&txq->pending_skbs); 313 netdev_warn(ndev, "Failed to post TX OOB: %d\n", err); 314 err = NETDEV_TX_BUSY; 315 goto tx_busy; 316 } 317 318 err = NETDEV_TX_OK; 319 atomic_inc(&txq->pending_sends); 320 321 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq); 322 323 /* skb may be freed after mana_gd_post_work_request. Do not use it. */ 324 skb = NULL; 325 326 tx_stats = &txq->stats; 327 u64_stats_update_begin(&tx_stats->syncp); 328 tx_stats->packets++; 329 tx_stats->bytes += len; 330 u64_stats_update_end(&tx_stats->syncp); 331 332 tx_busy: 333 if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) { 334 netif_tx_wake_queue(net_txq); 335 apc->eth_stats.wake_queue++; 336 } 337 338 kfree(pkg.sgl_ptr); 339 return err; 340 341 free_sgl_ptr: 342 kfree(pkg.sgl_ptr); 343 tx_drop_count: 344 ndev->stats.tx_dropped++; 345 tx_drop: 346 dev_kfree_skb_any(skb); 347 return NETDEV_TX_OK; 348 } 349 350 static void mana_get_stats64(struct net_device *ndev, 351 struct rtnl_link_stats64 *st) 352 { 353 struct mana_port_context *apc = netdev_priv(ndev); 354 unsigned int num_queues = apc->num_queues; 355 struct mana_stats_rx *rx_stats; 356 struct mana_stats_tx *tx_stats; 357 unsigned int start; 358 u64 packets, bytes; 359 int q; 360 361 if (!apc->port_is_up) 362 return; 363 364 netdev_stats_to_stats64(st, &ndev->stats); 365 366 for (q = 0; q < num_queues; q++) { 367 rx_stats = &apc->rxqs[q]->stats; 368 369 do { 370 start = u64_stats_fetch_begin(&rx_stats->syncp); 371 packets = rx_stats->packets; 372 bytes = rx_stats->bytes; 373 } while (u64_stats_fetch_retry(&rx_stats->syncp, start)); 374 375 st->rx_packets += packets; 376 st->rx_bytes += bytes; 377 } 378 379 for (q = 0; q < num_queues; q++) { 380 tx_stats = &apc->tx_qp[q].txq.stats; 381 382 do { 383 start = u64_stats_fetch_begin(&tx_stats->syncp); 384 packets = tx_stats->packets; 385 bytes = tx_stats->bytes; 386 } while (u64_stats_fetch_retry(&tx_stats->syncp, start)); 387 388 st->tx_packets += packets; 389 st->tx_bytes += bytes; 390 } 391 } 392 393 static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb, 394 int old_q) 395 { 396 struct mana_port_context *apc = netdev_priv(ndev); 397 u32 hash = skb_get_hash(skb); 398 struct sock *sk = skb->sk; 399 int txq; 400 401 txq = apc->indir_table[hash & MANA_INDIRECT_TABLE_MASK]; 402 403 if (txq != old_q && sk && sk_fullsock(sk) && 404 rcu_access_pointer(sk->sk_dst_cache)) 405 sk_tx_queue_set(sk, txq); 406 407 return txq; 408 } 409 410 static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb, 411 struct net_device *sb_dev) 412 { 413 int txq; 414 415 if (ndev->real_num_tx_queues == 1) 416 return 0; 417 418 txq = sk_tx_queue_get(skb->sk); 419 420 if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) { 421 if (skb_rx_queue_recorded(skb)) 422 txq = skb_get_rx_queue(skb); 423 else 424 txq = mana_get_tx_queue(ndev, skb, txq); 425 } 426 427 return txq; 428 } 429 430 /* Release pre-allocated RX buffers */ 431 static void mana_pre_dealloc_rxbufs(struct mana_port_context *mpc) 432 { 433 struct device *dev; 434 int i; 435 436 dev = mpc->ac->gdma_dev->gdma_context->dev; 437 438 if (!mpc->rxbufs_pre) 439 goto out1; 440 441 if (!mpc->das_pre) 442 goto out2; 443 444 while (mpc->rxbpre_total) { 445 i = --mpc->rxbpre_total; 446 dma_unmap_single(dev, mpc->das_pre[i], mpc->rxbpre_datasize, 447 DMA_FROM_DEVICE); 448 put_page(virt_to_head_page(mpc->rxbufs_pre[i])); 449 } 450 451 kfree(mpc->das_pre); 452 mpc->das_pre = NULL; 453 454 out2: 455 kfree(mpc->rxbufs_pre); 456 mpc->rxbufs_pre = NULL; 457 458 out1: 459 mpc->rxbpre_datasize = 0; 460 mpc->rxbpre_alloc_size = 0; 461 mpc->rxbpre_headroom = 0; 462 } 463 464 /* Get a buffer from the pre-allocated RX buffers */ 465 static void *mana_get_rxbuf_pre(struct mana_rxq *rxq, dma_addr_t *da) 466 { 467 struct net_device *ndev = rxq->ndev; 468 struct mana_port_context *mpc; 469 void *va; 470 471 mpc = netdev_priv(ndev); 472 473 if (!mpc->rxbufs_pre || !mpc->das_pre || !mpc->rxbpre_total) { 474 netdev_err(ndev, "No RX pre-allocated bufs\n"); 475 return NULL; 476 } 477 478 /* Check sizes to catch unexpected coding error */ 479 if (mpc->rxbpre_datasize != rxq->datasize) { 480 netdev_err(ndev, "rxbpre_datasize mismatch: %u: %u\n", 481 mpc->rxbpre_datasize, rxq->datasize); 482 return NULL; 483 } 484 485 if (mpc->rxbpre_alloc_size != rxq->alloc_size) { 486 netdev_err(ndev, "rxbpre_alloc_size mismatch: %u: %u\n", 487 mpc->rxbpre_alloc_size, rxq->alloc_size); 488 return NULL; 489 } 490 491 if (mpc->rxbpre_headroom != rxq->headroom) { 492 netdev_err(ndev, "rxbpre_headroom mismatch: %u: %u\n", 493 mpc->rxbpre_headroom, rxq->headroom); 494 return NULL; 495 } 496 497 mpc->rxbpre_total--; 498 499 *da = mpc->das_pre[mpc->rxbpre_total]; 500 va = mpc->rxbufs_pre[mpc->rxbpre_total]; 501 mpc->rxbufs_pre[mpc->rxbpre_total] = NULL; 502 503 /* Deallocate the array after all buffers are gone */ 504 if (!mpc->rxbpre_total) 505 mana_pre_dealloc_rxbufs(mpc); 506 507 return va; 508 } 509 510 /* Get RX buffer's data size, alloc size, XDP headroom based on MTU */ 511 static void mana_get_rxbuf_cfg(int mtu, u32 *datasize, u32 *alloc_size, 512 u32 *headroom) 513 { 514 if (mtu > MANA_XDP_MTU_MAX) 515 *headroom = 0; /* no support for XDP */ 516 else 517 *headroom = XDP_PACKET_HEADROOM; 518 519 *alloc_size = mtu + MANA_RXBUF_PAD + *headroom; 520 521 *datasize = ALIGN(mtu + ETH_HLEN, MANA_RX_DATA_ALIGN); 522 } 523 524 static int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu) 525 { 526 struct device *dev; 527 struct page *page; 528 dma_addr_t da; 529 int num_rxb; 530 void *va; 531 int i; 532 533 mana_get_rxbuf_cfg(new_mtu, &mpc->rxbpre_datasize, 534 &mpc->rxbpre_alloc_size, &mpc->rxbpre_headroom); 535 536 dev = mpc->ac->gdma_dev->gdma_context->dev; 537 538 num_rxb = mpc->num_queues * RX_BUFFERS_PER_QUEUE; 539 540 WARN(mpc->rxbufs_pre, "mana rxbufs_pre exists\n"); 541 mpc->rxbufs_pre = kmalloc_array(num_rxb, sizeof(void *), GFP_KERNEL); 542 if (!mpc->rxbufs_pre) 543 goto error; 544 545 mpc->das_pre = kmalloc_array(num_rxb, sizeof(dma_addr_t), GFP_KERNEL); 546 if (!mpc->das_pre) 547 goto error; 548 549 mpc->rxbpre_total = 0; 550 551 for (i = 0; i < num_rxb; i++) { 552 if (mpc->rxbpre_alloc_size > PAGE_SIZE) { 553 va = netdev_alloc_frag(mpc->rxbpre_alloc_size); 554 if (!va) 555 goto error; 556 557 page = virt_to_head_page(va); 558 /* Check if the frag falls back to single page */ 559 if (compound_order(page) < 560 get_order(mpc->rxbpre_alloc_size)) { 561 put_page(page); 562 goto error; 563 } 564 } else { 565 page = dev_alloc_page(); 566 if (!page) 567 goto error; 568 569 va = page_to_virt(page); 570 } 571 572 da = dma_map_single(dev, va + mpc->rxbpre_headroom, 573 mpc->rxbpre_datasize, DMA_FROM_DEVICE); 574 if (dma_mapping_error(dev, da)) { 575 put_page(virt_to_head_page(va)); 576 goto error; 577 } 578 579 mpc->rxbufs_pre[i] = va; 580 mpc->das_pre[i] = da; 581 mpc->rxbpre_total = i + 1; 582 } 583 584 return 0; 585 586 error: 587 mana_pre_dealloc_rxbufs(mpc); 588 return -ENOMEM; 589 } 590 591 static int mana_change_mtu(struct net_device *ndev, int new_mtu) 592 { 593 struct mana_port_context *mpc = netdev_priv(ndev); 594 unsigned int old_mtu = ndev->mtu; 595 int err; 596 597 /* Pre-allocate buffers to prevent failure in mana_attach later */ 598 err = mana_pre_alloc_rxbufs(mpc, new_mtu); 599 if (err) { 600 netdev_err(ndev, "Insufficient memory for new MTU\n"); 601 return err; 602 } 603 604 err = mana_detach(ndev, false); 605 if (err) { 606 netdev_err(ndev, "mana_detach failed: %d\n", err); 607 goto out; 608 } 609 610 ndev->mtu = new_mtu; 611 612 err = mana_attach(ndev); 613 if (err) { 614 netdev_err(ndev, "mana_attach failed: %d\n", err); 615 ndev->mtu = old_mtu; 616 } 617 618 out: 619 mana_pre_dealloc_rxbufs(mpc); 620 return err; 621 } 622 623 static const struct net_device_ops mana_devops = { 624 .ndo_open = mana_open, 625 .ndo_stop = mana_close, 626 .ndo_select_queue = mana_select_queue, 627 .ndo_start_xmit = mana_start_xmit, 628 .ndo_validate_addr = eth_validate_addr, 629 .ndo_get_stats64 = mana_get_stats64, 630 .ndo_bpf = mana_bpf, 631 .ndo_xdp_xmit = mana_xdp_xmit, 632 .ndo_change_mtu = mana_change_mtu, 633 }; 634 635 static void mana_cleanup_port_context(struct mana_port_context *apc) 636 { 637 kfree(apc->rxqs); 638 apc->rxqs = NULL; 639 } 640 641 static int mana_init_port_context(struct mana_port_context *apc) 642 { 643 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *), 644 GFP_KERNEL); 645 646 return !apc->rxqs ? -ENOMEM : 0; 647 } 648 649 static int mana_send_request(struct mana_context *ac, void *in_buf, 650 u32 in_len, void *out_buf, u32 out_len) 651 { 652 struct gdma_context *gc = ac->gdma_dev->gdma_context; 653 struct gdma_resp_hdr *resp = out_buf; 654 struct gdma_req_hdr *req = in_buf; 655 struct device *dev = gc->dev; 656 static atomic_t activity_id; 657 int err; 658 659 req->dev_id = gc->mana.dev_id; 660 req->activity_id = atomic_inc_return(&activity_id); 661 662 err = mana_gd_send_request(gc, in_len, in_buf, out_len, 663 out_buf); 664 if (err || resp->status) { 665 dev_err(dev, "Failed to send mana message: %d, 0x%x\n", 666 err, resp->status); 667 return err ? err : -EPROTO; 668 } 669 670 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 || 671 req->activity_id != resp->activity_id) { 672 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n", 673 req->dev_id.as_uint32, resp->dev_id.as_uint32, 674 req->activity_id, resp->activity_id); 675 return -EPROTO; 676 } 677 678 return 0; 679 } 680 681 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr, 682 const enum mana_command_code expected_code, 683 const u32 min_size) 684 { 685 if (resp_hdr->response.msg_type != expected_code) 686 return -EPROTO; 687 688 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1) 689 return -EPROTO; 690 691 if (resp_hdr->response.msg_size < min_size) 692 return -EPROTO; 693 694 return 0; 695 } 696 697 static int mana_pf_register_hw_vport(struct mana_port_context *apc) 698 { 699 struct mana_register_hw_vport_resp resp = {}; 700 struct mana_register_hw_vport_req req = {}; 701 int err; 702 703 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT, 704 sizeof(req), sizeof(resp)); 705 req.attached_gfid = 1; 706 req.is_pf_default_vport = 1; 707 req.allow_all_ether_types = 1; 708 709 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 710 sizeof(resp)); 711 if (err) { 712 netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err); 713 return err; 714 } 715 716 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT, 717 sizeof(resp)); 718 if (err || resp.hdr.status) { 719 netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n", 720 err, resp.hdr.status); 721 return err ? err : -EPROTO; 722 } 723 724 apc->port_handle = resp.hw_vport_handle; 725 return 0; 726 } 727 728 static void mana_pf_deregister_hw_vport(struct mana_port_context *apc) 729 { 730 struct mana_deregister_hw_vport_resp resp = {}; 731 struct mana_deregister_hw_vport_req req = {}; 732 int err; 733 734 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT, 735 sizeof(req), sizeof(resp)); 736 req.hw_vport_handle = apc->port_handle; 737 738 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 739 sizeof(resp)); 740 if (err) { 741 netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n", 742 err); 743 return; 744 } 745 746 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT, 747 sizeof(resp)); 748 if (err || resp.hdr.status) 749 netdev_err(apc->ndev, 750 "Failed to deregister hw vPort: %d, 0x%x\n", 751 err, resp.hdr.status); 752 } 753 754 static int mana_pf_register_filter(struct mana_port_context *apc) 755 { 756 struct mana_register_filter_resp resp = {}; 757 struct mana_register_filter_req req = {}; 758 int err; 759 760 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER, 761 sizeof(req), sizeof(resp)); 762 req.vport = apc->port_handle; 763 memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN); 764 765 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 766 sizeof(resp)); 767 if (err) { 768 netdev_err(apc->ndev, "Failed to register filter: %d\n", err); 769 return err; 770 } 771 772 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER, 773 sizeof(resp)); 774 if (err || resp.hdr.status) { 775 netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n", 776 err, resp.hdr.status); 777 return err ? err : -EPROTO; 778 } 779 780 apc->pf_filter_handle = resp.filter_handle; 781 return 0; 782 } 783 784 static void mana_pf_deregister_filter(struct mana_port_context *apc) 785 { 786 struct mana_deregister_filter_resp resp = {}; 787 struct mana_deregister_filter_req req = {}; 788 int err; 789 790 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER, 791 sizeof(req), sizeof(resp)); 792 req.filter_handle = apc->pf_filter_handle; 793 794 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 795 sizeof(resp)); 796 if (err) { 797 netdev_err(apc->ndev, "Failed to unregister filter: %d\n", 798 err); 799 return; 800 } 801 802 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER, 803 sizeof(resp)); 804 if (err || resp.hdr.status) 805 netdev_err(apc->ndev, 806 "Failed to deregister filter: %d, 0x%x\n", 807 err, resp.hdr.status); 808 } 809 810 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver, 811 u32 proto_minor_ver, u32 proto_micro_ver, 812 u16 *max_num_vports) 813 { 814 struct gdma_context *gc = ac->gdma_dev->gdma_context; 815 struct mana_query_device_cfg_resp resp = {}; 816 struct mana_query_device_cfg_req req = {}; 817 struct device *dev = gc->dev; 818 int err = 0; 819 820 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG, 821 sizeof(req), sizeof(resp)); 822 823 req.hdr.resp.msg_version = GDMA_MESSAGE_V2; 824 825 req.proto_major_ver = proto_major_ver; 826 req.proto_minor_ver = proto_minor_ver; 827 req.proto_micro_ver = proto_micro_ver; 828 829 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp)); 830 if (err) { 831 dev_err(dev, "Failed to query config: %d", err); 832 return err; 833 } 834 835 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG, 836 sizeof(resp)); 837 if (err || resp.hdr.status) { 838 dev_err(dev, "Invalid query result: %d, 0x%x\n", err, 839 resp.hdr.status); 840 if (!err) 841 err = -EPROTO; 842 return err; 843 } 844 845 *max_num_vports = resp.max_num_vports; 846 847 if (resp.hdr.response.msg_version == GDMA_MESSAGE_V2) 848 gc->adapter_mtu = resp.adapter_mtu; 849 else 850 gc->adapter_mtu = ETH_FRAME_LEN; 851 852 return 0; 853 } 854 855 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index, 856 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry) 857 { 858 struct mana_query_vport_cfg_resp resp = {}; 859 struct mana_query_vport_cfg_req req = {}; 860 int err; 861 862 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG, 863 sizeof(req), sizeof(resp)); 864 865 req.vport_index = vport_index; 866 867 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 868 sizeof(resp)); 869 if (err) 870 return err; 871 872 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG, 873 sizeof(resp)); 874 if (err) 875 return err; 876 877 if (resp.hdr.status) 878 return -EPROTO; 879 880 *max_sq = resp.max_num_sq; 881 *max_rq = resp.max_num_rq; 882 *num_indir_entry = resp.num_indirection_ent; 883 884 apc->port_handle = resp.vport; 885 ether_addr_copy(apc->mac_addr, resp.mac_addr); 886 887 return 0; 888 } 889 890 void mana_uncfg_vport(struct mana_port_context *apc) 891 { 892 mutex_lock(&apc->vport_mutex); 893 apc->vport_use_count--; 894 WARN_ON(apc->vport_use_count < 0); 895 mutex_unlock(&apc->vport_mutex); 896 } 897 EXPORT_SYMBOL_NS(mana_uncfg_vport, NET_MANA); 898 899 int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id, 900 u32 doorbell_pg_id) 901 { 902 struct mana_config_vport_resp resp = {}; 903 struct mana_config_vport_req req = {}; 904 int err; 905 906 /* This function is used to program the Ethernet port in the hardware 907 * table. It can be called from the Ethernet driver or the RDMA driver. 908 * 909 * For Ethernet usage, the hardware supports only one active user on a 910 * physical port. The driver checks on the port usage before programming 911 * the hardware when creating the RAW QP (RDMA driver) or exposing the 912 * device to kernel NET layer (Ethernet driver). 913 * 914 * Because the RDMA driver doesn't know in advance which QP type the 915 * user will create, it exposes the device with all its ports. The user 916 * may not be able to create RAW QP on a port if this port is already 917 * in used by the Ethernet driver from the kernel. 918 * 919 * This physical port limitation only applies to the RAW QP. For RC QP, 920 * the hardware doesn't have this limitation. The user can create RC 921 * QPs on a physical port up to the hardware limits independent of the 922 * Ethernet usage on the same port. 923 */ 924 mutex_lock(&apc->vport_mutex); 925 if (apc->vport_use_count > 0) { 926 mutex_unlock(&apc->vport_mutex); 927 return -EBUSY; 928 } 929 apc->vport_use_count++; 930 mutex_unlock(&apc->vport_mutex); 931 932 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX, 933 sizeof(req), sizeof(resp)); 934 req.vport = apc->port_handle; 935 req.pdid = protection_dom_id; 936 req.doorbell_pageid = doorbell_pg_id; 937 938 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 939 sizeof(resp)); 940 if (err) { 941 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err); 942 goto out; 943 } 944 945 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX, 946 sizeof(resp)); 947 if (err || resp.hdr.status) { 948 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n", 949 err, resp.hdr.status); 950 if (!err) 951 err = -EPROTO; 952 953 goto out; 954 } 955 956 apc->tx_shortform_allowed = resp.short_form_allowed; 957 apc->tx_vp_offset = resp.tx_vport_offset; 958 959 netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n", 960 apc->port_handle, protection_dom_id, doorbell_pg_id); 961 out: 962 if (err) 963 mana_uncfg_vport(apc); 964 965 return err; 966 } 967 EXPORT_SYMBOL_NS(mana_cfg_vport, NET_MANA); 968 969 static int mana_cfg_vport_steering(struct mana_port_context *apc, 970 enum TRI_STATE rx, 971 bool update_default_rxobj, bool update_key, 972 bool update_tab) 973 { 974 u16 num_entries = MANA_INDIRECT_TABLE_SIZE; 975 struct mana_cfg_rx_steer_req *req = NULL; 976 struct mana_cfg_rx_steer_resp resp = {}; 977 struct net_device *ndev = apc->ndev; 978 mana_handle_t *req_indir_tab; 979 u32 req_buf_size; 980 int err; 981 982 req_buf_size = sizeof(*req) + sizeof(mana_handle_t) * num_entries; 983 req = kzalloc(req_buf_size, GFP_KERNEL); 984 if (!req) 985 return -ENOMEM; 986 987 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size, 988 sizeof(resp)); 989 990 req->vport = apc->port_handle; 991 req->num_indir_entries = num_entries; 992 req->indir_tab_offset = sizeof(*req); 993 req->rx_enable = rx; 994 req->rss_enable = apc->rss_state; 995 req->update_default_rxobj = update_default_rxobj; 996 req->update_hashkey = update_key; 997 req->update_indir_tab = update_tab; 998 req->default_rxobj = apc->default_rxobj; 999 1000 if (update_key) 1001 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE); 1002 1003 if (update_tab) { 1004 req_indir_tab = (mana_handle_t *)(req + 1); 1005 memcpy(req_indir_tab, apc->rxobj_table, 1006 req->num_indir_entries * sizeof(mana_handle_t)); 1007 } 1008 1009 err = mana_send_request(apc->ac, req, req_buf_size, &resp, 1010 sizeof(resp)); 1011 if (err) { 1012 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err); 1013 goto out; 1014 } 1015 1016 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX, 1017 sizeof(resp)); 1018 if (err) { 1019 netdev_err(ndev, "vPort RX configuration failed: %d\n", err); 1020 goto out; 1021 } 1022 1023 if (resp.hdr.status) { 1024 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n", 1025 resp.hdr.status); 1026 err = -EPROTO; 1027 } 1028 1029 netdev_info(ndev, "Configured steering vPort %llu entries %u\n", 1030 apc->port_handle, num_entries); 1031 out: 1032 kfree(req); 1033 return err; 1034 } 1035 1036 int mana_create_wq_obj(struct mana_port_context *apc, 1037 mana_handle_t vport, 1038 u32 wq_type, struct mana_obj_spec *wq_spec, 1039 struct mana_obj_spec *cq_spec, 1040 mana_handle_t *wq_obj) 1041 { 1042 struct mana_create_wqobj_resp resp = {}; 1043 struct mana_create_wqobj_req req = {}; 1044 struct net_device *ndev = apc->ndev; 1045 int err; 1046 1047 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ, 1048 sizeof(req), sizeof(resp)); 1049 req.vport = vport; 1050 req.wq_type = wq_type; 1051 req.wq_gdma_region = wq_spec->gdma_region; 1052 req.cq_gdma_region = cq_spec->gdma_region; 1053 req.wq_size = wq_spec->queue_size; 1054 req.cq_size = cq_spec->queue_size; 1055 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id; 1056 req.cq_parent_qid = cq_spec->attached_eq; 1057 1058 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1059 sizeof(resp)); 1060 if (err) { 1061 netdev_err(ndev, "Failed to create WQ object: %d\n", err); 1062 goto out; 1063 } 1064 1065 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ, 1066 sizeof(resp)); 1067 if (err || resp.hdr.status) { 1068 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err, 1069 resp.hdr.status); 1070 if (!err) 1071 err = -EPROTO; 1072 goto out; 1073 } 1074 1075 if (resp.wq_obj == INVALID_MANA_HANDLE) { 1076 netdev_err(ndev, "Got an invalid WQ object handle\n"); 1077 err = -EPROTO; 1078 goto out; 1079 } 1080 1081 *wq_obj = resp.wq_obj; 1082 wq_spec->queue_index = resp.wq_id; 1083 cq_spec->queue_index = resp.cq_id; 1084 1085 return 0; 1086 out: 1087 return err; 1088 } 1089 EXPORT_SYMBOL_NS(mana_create_wq_obj, NET_MANA); 1090 1091 void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type, 1092 mana_handle_t wq_obj) 1093 { 1094 struct mana_destroy_wqobj_resp resp = {}; 1095 struct mana_destroy_wqobj_req req = {}; 1096 struct net_device *ndev = apc->ndev; 1097 int err; 1098 1099 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ, 1100 sizeof(req), sizeof(resp)); 1101 req.wq_type = wq_type; 1102 req.wq_obj_handle = wq_obj; 1103 1104 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1105 sizeof(resp)); 1106 if (err) { 1107 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err); 1108 return; 1109 } 1110 1111 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ, 1112 sizeof(resp)); 1113 if (err || resp.hdr.status) 1114 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err, 1115 resp.hdr.status); 1116 } 1117 EXPORT_SYMBOL_NS(mana_destroy_wq_obj, NET_MANA); 1118 1119 static void mana_destroy_eq(struct mana_context *ac) 1120 { 1121 struct gdma_context *gc = ac->gdma_dev->gdma_context; 1122 struct gdma_queue *eq; 1123 int i; 1124 1125 if (!ac->eqs) 1126 return; 1127 1128 for (i = 0; i < gc->max_num_queues; i++) { 1129 eq = ac->eqs[i].eq; 1130 if (!eq) 1131 continue; 1132 1133 mana_gd_destroy_queue(gc, eq); 1134 } 1135 1136 kfree(ac->eqs); 1137 ac->eqs = NULL; 1138 } 1139 1140 static int mana_create_eq(struct mana_context *ac) 1141 { 1142 struct gdma_dev *gd = ac->gdma_dev; 1143 struct gdma_context *gc = gd->gdma_context; 1144 struct gdma_queue_spec spec = {}; 1145 int err; 1146 int i; 1147 1148 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq), 1149 GFP_KERNEL); 1150 if (!ac->eqs) 1151 return -ENOMEM; 1152 1153 spec.type = GDMA_EQ; 1154 spec.monitor_avl_buf = false; 1155 spec.queue_size = EQ_SIZE; 1156 spec.eq.callback = NULL; 1157 spec.eq.context = ac->eqs; 1158 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE; 1159 1160 for (i = 0; i < gc->max_num_queues; i++) { 1161 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq); 1162 if (err) 1163 goto out; 1164 } 1165 1166 return 0; 1167 out: 1168 mana_destroy_eq(ac); 1169 return err; 1170 } 1171 1172 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq) 1173 { 1174 struct mana_fence_rq_resp resp = {}; 1175 struct mana_fence_rq_req req = {}; 1176 int err; 1177 1178 init_completion(&rxq->fence_event); 1179 1180 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ, 1181 sizeof(req), sizeof(resp)); 1182 req.wq_obj_handle = rxq->rxobj; 1183 1184 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1185 sizeof(resp)); 1186 if (err) { 1187 netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n", 1188 rxq->rxq_idx, err); 1189 return err; 1190 } 1191 1192 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp)); 1193 if (err || resp.hdr.status) { 1194 netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n", 1195 rxq->rxq_idx, err, resp.hdr.status); 1196 if (!err) 1197 err = -EPROTO; 1198 1199 return err; 1200 } 1201 1202 if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) { 1203 netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n", 1204 rxq->rxq_idx); 1205 return -ETIMEDOUT; 1206 } 1207 1208 return 0; 1209 } 1210 1211 static void mana_fence_rqs(struct mana_port_context *apc) 1212 { 1213 unsigned int rxq_idx; 1214 struct mana_rxq *rxq; 1215 int err; 1216 1217 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 1218 rxq = apc->rxqs[rxq_idx]; 1219 err = mana_fence_rq(apc, rxq); 1220 1221 /* In case of any error, use sleep instead. */ 1222 if (err) 1223 msleep(100); 1224 } 1225 } 1226 1227 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units) 1228 { 1229 u32 used_space_old; 1230 u32 used_space_new; 1231 1232 used_space_old = wq->head - wq->tail; 1233 used_space_new = wq->head - (wq->tail + num_units); 1234 1235 if (WARN_ON_ONCE(used_space_new > used_space_old)) 1236 return -ERANGE; 1237 1238 wq->tail += num_units; 1239 return 0; 1240 } 1241 1242 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc) 1243 { 1244 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 1245 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 1246 struct device *dev = gc->dev; 1247 int i; 1248 1249 dma_unmap_single(dev, ash->dma_handle[0], ash->size[0], DMA_TO_DEVICE); 1250 1251 for (i = 1; i < skb_shinfo(skb)->nr_frags + 1; i++) 1252 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i], 1253 DMA_TO_DEVICE); 1254 } 1255 1256 static void mana_poll_tx_cq(struct mana_cq *cq) 1257 { 1258 struct gdma_comp *completions = cq->gdma_comp_buf; 1259 struct gdma_posted_wqe_info *wqe_info; 1260 unsigned int pkt_transmitted = 0; 1261 unsigned int wqe_unit_cnt = 0; 1262 struct mana_txq *txq = cq->txq; 1263 struct mana_port_context *apc; 1264 struct netdev_queue *net_txq; 1265 struct gdma_queue *gdma_wq; 1266 unsigned int avail_space; 1267 struct net_device *ndev; 1268 struct sk_buff *skb; 1269 bool txq_stopped; 1270 int comp_read; 1271 int i; 1272 1273 ndev = txq->ndev; 1274 apc = netdev_priv(ndev); 1275 1276 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions, 1277 CQE_POLLING_BUFFER); 1278 1279 if (comp_read < 1) 1280 return; 1281 1282 apc->eth_stats.tx_cqes = comp_read; 1283 1284 for (i = 0; i < comp_read; i++) { 1285 struct mana_tx_comp_oob *cqe_oob; 1286 1287 if (WARN_ON_ONCE(!completions[i].is_sq)) 1288 return; 1289 1290 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data; 1291 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type != 1292 MANA_CQE_COMPLETION)) 1293 return; 1294 1295 switch (cqe_oob->cqe_hdr.cqe_type) { 1296 case CQE_TX_OKAY: 1297 break; 1298 1299 case CQE_TX_SA_DROP: 1300 case CQE_TX_MTU_DROP: 1301 case CQE_TX_INVALID_OOB: 1302 case CQE_TX_INVALID_ETH_TYPE: 1303 case CQE_TX_HDR_PROCESSING_ERROR: 1304 case CQE_TX_VF_DISABLED: 1305 case CQE_TX_VPORT_IDX_OUT_OF_RANGE: 1306 case CQE_TX_VPORT_DISABLED: 1307 case CQE_TX_VLAN_TAGGING_VIOLATION: 1308 WARN_ONCE(1, "TX: CQE error %d: ignored.\n", 1309 cqe_oob->cqe_hdr.cqe_type); 1310 apc->eth_stats.tx_cqe_err++; 1311 break; 1312 1313 default: 1314 /* If the CQE type is unexpected, log an error, assert, 1315 * and go through the error path. 1316 */ 1317 WARN_ONCE(1, "TX: Unexpected CQE type %d: HW BUG?\n", 1318 cqe_oob->cqe_hdr.cqe_type); 1319 apc->eth_stats.tx_cqe_unknown_type++; 1320 return; 1321 } 1322 1323 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num)) 1324 return; 1325 1326 skb = skb_dequeue(&txq->pending_skbs); 1327 if (WARN_ON_ONCE(!skb)) 1328 return; 1329 1330 wqe_info = (struct gdma_posted_wqe_info *)skb->cb; 1331 wqe_unit_cnt += wqe_info->wqe_size_in_bu; 1332 1333 mana_unmap_skb(skb, apc); 1334 1335 napi_consume_skb(skb, cq->budget); 1336 1337 pkt_transmitted++; 1338 } 1339 1340 if (WARN_ON_ONCE(wqe_unit_cnt == 0)) 1341 return; 1342 1343 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt); 1344 1345 gdma_wq = txq->gdma_sq; 1346 avail_space = mana_gd_wq_avail_space(gdma_wq); 1347 1348 /* Ensure tail updated before checking q stop */ 1349 smp_mb(); 1350 1351 net_txq = txq->net_txq; 1352 txq_stopped = netif_tx_queue_stopped(net_txq); 1353 1354 /* Ensure checking txq_stopped before apc->port_is_up. */ 1355 smp_rmb(); 1356 1357 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) { 1358 netif_tx_wake_queue(net_txq); 1359 apc->eth_stats.wake_queue++; 1360 } 1361 1362 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0) 1363 WARN_ON_ONCE(1); 1364 1365 cq->work_done = pkt_transmitted; 1366 1367 apc->eth_stats.tx_cqes -= pkt_transmitted; 1368 } 1369 1370 static void mana_post_pkt_rxq(struct mana_rxq *rxq) 1371 { 1372 struct mana_recv_buf_oob *recv_buf_oob; 1373 u32 curr_index; 1374 int err; 1375 1376 curr_index = rxq->buf_index++; 1377 if (rxq->buf_index == rxq->num_rx_buf) 1378 rxq->buf_index = 0; 1379 1380 recv_buf_oob = &rxq->rx_oobs[curr_index]; 1381 1382 err = mana_gd_post_and_ring(rxq->gdma_rq, &recv_buf_oob->wqe_req, 1383 &recv_buf_oob->wqe_inf); 1384 if (WARN_ON_ONCE(err)) 1385 return; 1386 1387 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1); 1388 } 1389 1390 static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va, 1391 uint pkt_len, struct xdp_buff *xdp) 1392 { 1393 struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size); 1394 1395 if (!skb) 1396 return NULL; 1397 1398 if (xdp->data_hard_start) { 1399 skb_reserve(skb, xdp->data - xdp->data_hard_start); 1400 skb_put(skb, xdp->data_end - xdp->data); 1401 return skb; 1402 } 1403 1404 skb_reserve(skb, rxq->headroom); 1405 skb_put(skb, pkt_len); 1406 1407 return skb; 1408 } 1409 1410 static void mana_rx_skb(void *buf_va, struct mana_rxcomp_oob *cqe, 1411 struct mana_rxq *rxq) 1412 { 1413 struct mana_stats_rx *rx_stats = &rxq->stats; 1414 struct net_device *ndev = rxq->ndev; 1415 uint pkt_len = cqe->ppi[0].pkt_len; 1416 u16 rxq_idx = rxq->rxq_idx; 1417 struct napi_struct *napi; 1418 struct xdp_buff xdp = {}; 1419 struct sk_buff *skb; 1420 u32 hash_value; 1421 u32 act; 1422 1423 rxq->rx_cq.work_done++; 1424 napi = &rxq->rx_cq.napi; 1425 1426 if (!buf_va) { 1427 ++ndev->stats.rx_dropped; 1428 return; 1429 } 1430 1431 act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len); 1432 1433 if (act == XDP_REDIRECT && !rxq->xdp_rc) 1434 return; 1435 1436 if (act != XDP_PASS && act != XDP_TX) 1437 goto drop_xdp; 1438 1439 skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp); 1440 1441 if (!skb) 1442 goto drop; 1443 1444 skb->dev = napi->dev; 1445 1446 skb->protocol = eth_type_trans(skb, ndev); 1447 skb_checksum_none_assert(skb); 1448 skb_record_rx_queue(skb, rxq_idx); 1449 1450 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) { 1451 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed) 1452 skb->ip_summed = CHECKSUM_UNNECESSARY; 1453 } 1454 1455 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) { 1456 hash_value = cqe->ppi[0].pkt_hash; 1457 1458 if (cqe->rx_hashtype & MANA_HASH_L4) 1459 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4); 1460 else 1461 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3); 1462 } 1463 1464 u64_stats_update_begin(&rx_stats->syncp); 1465 rx_stats->packets++; 1466 rx_stats->bytes += pkt_len; 1467 1468 if (act == XDP_TX) 1469 rx_stats->xdp_tx++; 1470 u64_stats_update_end(&rx_stats->syncp); 1471 1472 if (act == XDP_TX) { 1473 skb_set_queue_mapping(skb, rxq_idx); 1474 mana_xdp_tx(skb, ndev); 1475 return; 1476 } 1477 1478 napi_gro_receive(napi, skb); 1479 1480 return; 1481 1482 drop_xdp: 1483 u64_stats_update_begin(&rx_stats->syncp); 1484 rx_stats->xdp_drop++; 1485 u64_stats_update_end(&rx_stats->syncp); 1486 1487 drop: 1488 WARN_ON_ONCE(rxq->xdp_save_va); 1489 /* Save for reuse */ 1490 rxq->xdp_save_va = buf_va; 1491 1492 ++ndev->stats.rx_dropped; 1493 1494 return; 1495 } 1496 1497 static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev, 1498 dma_addr_t *da, bool is_napi) 1499 { 1500 struct page *page; 1501 void *va; 1502 1503 /* Reuse XDP dropped page if available */ 1504 if (rxq->xdp_save_va) { 1505 va = rxq->xdp_save_va; 1506 rxq->xdp_save_va = NULL; 1507 } else if (rxq->alloc_size > PAGE_SIZE) { 1508 if (is_napi) 1509 va = napi_alloc_frag(rxq->alloc_size); 1510 else 1511 va = netdev_alloc_frag(rxq->alloc_size); 1512 1513 if (!va) 1514 return NULL; 1515 1516 page = virt_to_head_page(va); 1517 /* Check if the frag falls back to single page */ 1518 if (compound_order(page) < get_order(rxq->alloc_size)) { 1519 put_page(page); 1520 return NULL; 1521 } 1522 } else { 1523 page = dev_alloc_page(); 1524 if (!page) 1525 return NULL; 1526 1527 va = page_to_virt(page); 1528 } 1529 1530 *da = dma_map_single(dev, va + rxq->headroom, rxq->datasize, 1531 DMA_FROM_DEVICE); 1532 if (dma_mapping_error(dev, *da)) { 1533 put_page(virt_to_head_page(va)); 1534 return NULL; 1535 } 1536 1537 return va; 1538 } 1539 1540 /* Allocate frag for rx buffer, and save the old buf */ 1541 static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq, 1542 struct mana_recv_buf_oob *rxoob, void **old_buf) 1543 { 1544 dma_addr_t da; 1545 void *va; 1546 1547 va = mana_get_rxfrag(rxq, dev, &da, true); 1548 if (!va) 1549 return; 1550 1551 dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize, 1552 DMA_FROM_DEVICE); 1553 *old_buf = rxoob->buf_va; 1554 1555 rxoob->buf_va = va; 1556 rxoob->sgl[0].address = da; 1557 } 1558 1559 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq, 1560 struct gdma_comp *cqe) 1561 { 1562 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data; 1563 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context; 1564 struct net_device *ndev = rxq->ndev; 1565 struct mana_recv_buf_oob *rxbuf_oob; 1566 struct mana_port_context *apc; 1567 struct device *dev = gc->dev; 1568 void *old_buf = NULL; 1569 u32 curr, pktlen; 1570 1571 apc = netdev_priv(ndev); 1572 1573 switch (oob->cqe_hdr.cqe_type) { 1574 case CQE_RX_OKAY: 1575 break; 1576 1577 case CQE_RX_TRUNCATED: 1578 ++ndev->stats.rx_dropped; 1579 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index]; 1580 netdev_warn_once(ndev, "Dropped a truncated packet\n"); 1581 goto drop; 1582 1583 case CQE_RX_COALESCED_4: 1584 netdev_err(ndev, "RX coalescing is unsupported\n"); 1585 apc->eth_stats.rx_coalesced_err++; 1586 return; 1587 1588 case CQE_RX_OBJECT_FENCE: 1589 complete(&rxq->fence_event); 1590 return; 1591 1592 default: 1593 netdev_err(ndev, "Unknown RX CQE type = %d\n", 1594 oob->cqe_hdr.cqe_type); 1595 apc->eth_stats.rx_cqe_unknown_type++; 1596 return; 1597 } 1598 1599 pktlen = oob->ppi[0].pkt_len; 1600 1601 if (pktlen == 0) { 1602 /* data packets should never have packetlength of zero */ 1603 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n", 1604 rxq->gdma_id, cq->gdma_id, rxq->rxobj); 1605 return; 1606 } 1607 1608 curr = rxq->buf_index; 1609 rxbuf_oob = &rxq->rx_oobs[curr]; 1610 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1); 1611 1612 mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf); 1613 1614 /* Unsuccessful refill will have old_buf == NULL. 1615 * In this case, mana_rx_skb() will drop the packet. 1616 */ 1617 mana_rx_skb(old_buf, oob, rxq); 1618 1619 drop: 1620 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu); 1621 1622 mana_post_pkt_rxq(rxq); 1623 } 1624 1625 static void mana_poll_rx_cq(struct mana_cq *cq) 1626 { 1627 struct gdma_comp *comp = cq->gdma_comp_buf; 1628 struct mana_rxq *rxq = cq->rxq; 1629 struct mana_port_context *apc; 1630 int comp_read, i; 1631 1632 apc = netdev_priv(rxq->ndev); 1633 1634 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER); 1635 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER); 1636 1637 apc->eth_stats.rx_cqes = comp_read; 1638 rxq->xdp_flush = false; 1639 1640 for (i = 0; i < comp_read; i++) { 1641 if (WARN_ON_ONCE(comp[i].is_sq)) 1642 return; 1643 1644 /* verify recv cqe references the right rxq */ 1645 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id)) 1646 return; 1647 1648 mana_process_rx_cqe(rxq, cq, &comp[i]); 1649 1650 apc->eth_stats.rx_cqes--; 1651 } 1652 1653 if (rxq->xdp_flush) 1654 xdp_do_flush(); 1655 } 1656 1657 static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue) 1658 { 1659 struct mana_cq *cq = context; 1660 u8 arm_bit; 1661 int w; 1662 1663 WARN_ON_ONCE(cq->gdma_cq != gdma_queue); 1664 1665 if (cq->type == MANA_CQ_TYPE_RX) 1666 mana_poll_rx_cq(cq); 1667 else 1668 mana_poll_tx_cq(cq); 1669 1670 w = cq->work_done; 1671 1672 if (w < cq->budget && 1673 napi_complete_done(&cq->napi, w)) { 1674 arm_bit = SET_ARM_BIT; 1675 } else { 1676 arm_bit = 0; 1677 } 1678 1679 mana_gd_ring_cq(gdma_queue, arm_bit); 1680 1681 return w; 1682 } 1683 1684 static int mana_poll(struct napi_struct *napi, int budget) 1685 { 1686 struct mana_cq *cq = container_of(napi, struct mana_cq, napi); 1687 int w; 1688 1689 cq->work_done = 0; 1690 cq->budget = budget; 1691 1692 w = mana_cq_handler(cq, cq->gdma_cq); 1693 1694 return min(w, budget); 1695 } 1696 1697 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue) 1698 { 1699 struct mana_cq *cq = context; 1700 1701 napi_schedule_irqoff(&cq->napi); 1702 } 1703 1704 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq) 1705 { 1706 struct gdma_dev *gd = apc->ac->gdma_dev; 1707 1708 if (!cq->gdma_cq) 1709 return; 1710 1711 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq); 1712 } 1713 1714 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq) 1715 { 1716 struct gdma_dev *gd = apc->ac->gdma_dev; 1717 1718 if (!txq->gdma_sq) 1719 return; 1720 1721 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq); 1722 } 1723 1724 static void mana_destroy_txq(struct mana_port_context *apc) 1725 { 1726 struct napi_struct *napi; 1727 int i; 1728 1729 if (!apc->tx_qp) 1730 return; 1731 1732 for (i = 0; i < apc->num_queues; i++) { 1733 napi = &apc->tx_qp[i].tx_cq.napi; 1734 napi_synchronize(napi); 1735 napi_disable(napi); 1736 netif_napi_del(napi); 1737 1738 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object); 1739 1740 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq); 1741 1742 mana_deinit_txq(apc, &apc->tx_qp[i].txq); 1743 } 1744 1745 kfree(apc->tx_qp); 1746 apc->tx_qp = NULL; 1747 } 1748 1749 static int mana_create_txq(struct mana_port_context *apc, 1750 struct net_device *net) 1751 { 1752 struct mana_context *ac = apc->ac; 1753 struct gdma_dev *gd = ac->gdma_dev; 1754 struct mana_obj_spec wq_spec; 1755 struct mana_obj_spec cq_spec; 1756 struct gdma_queue_spec spec; 1757 struct gdma_context *gc; 1758 struct mana_txq *txq; 1759 struct mana_cq *cq; 1760 u32 txq_size; 1761 u32 cq_size; 1762 int err; 1763 int i; 1764 1765 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp), 1766 GFP_KERNEL); 1767 if (!apc->tx_qp) 1768 return -ENOMEM; 1769 1770 /* The minimum size of the WQE is 32 bytes, hence 1771 * MAX_SEND_BUFFERS_PER_QUEUE represents the maximum number of WQEs 1772 * the SQ can store. This value is then used to size other queues 1773 * to prevent overflow. 1774 */ 1775 txq_size = MAX_SEND_BUFFERS_PER_QUEUE * 32; 1776 BUILD_BUG_ON(!PAGE_ALIGNED(txq_size)); 1777 1778 cq_size = MAX_SEND_BUFFERS_PER_QUEUE * COMP_ENTRY_SIZE; 1779 cq_size = PAGE_ALIGN(cq_size); 1780 1781 gc = gd->gdma_context; 1782 1783 for (i = 0; i < apc->num_queues; i++) { 1784 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE; 1785 1786 /* Create SQ */ 1787 txq = &apc->tx_qp[i].txq; 1788 1789 u64_stats_init(&txq->stats.syncp); 1790 txq->ndev = net; 1791 txq->net_txq = netdev_get_tx_queue(net, i); 1792 txq->vp_offset = apc->tx_vp_offset; 1793 skb_queue_head_init(&txq->pending_skbs); 1794 1795 memset(&spec, 0, sizeof(spec)); 1796 spec.type = GDMA_SQ; 1797 spec.monitor_avl_buf = true; 1798 spec.queue_size = txq_size; 1799 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq); 1800 if (err) 1801 goto out; 1802 1803 /* Create SQ's CQ */ 1804 cq = &apc->tx_qp[i].tx_cq; 1805 cq->type = MANA_CQ_TYPE_TX; 1806 1807 cq->txq = txq; 1808 1809 memset(&spec, 0, sizeof(spec)); 1810 spec.type = GDMA_CQ; 1811 spec.monitor_avl_buf = false; 1812 spec.queue_size = cq_size; 1813 spec.cq.callback = mana_schedule_napi; 1814 spec.cq.parent_eq = ac->eqs[i].eq; 1815 spec.cq.context = cq; 1816 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 1817 if (err) 1818 goto out; 1819 1820 memset(&wq_spec, 0, sizeof(wq_spec)); 1821 memset(&cq_spec, 0, sizeof(cq_spec)); 1822 1823 wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle; 1824 wq_spec.queue_size = txq->gdma_sq->queue_size; 1825 1826 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 1827 cq_spec.queue_size = cq->gdma_cq->queue_size; 1828 cq_spec.modr_ctx_id = 0; 1829 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 1830 1831 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ, 1832 &wq_spec, &cq_spec, 1833 &apc->tx_qp[i].tx_object); 1834 1835 if (err) 1836 goto out; 1837 1838 txq->gdma_sq->id = wq_spec.queue_index; 1839 cq->gdma_cq->id = cq_spec.queue_index; 1840 1841 txq->gdma_sq->mem_info.dma_region_handle = 1842 GDMA_INVALID_DMA_REGION; 1843 cq->gdma_cq->mem_info.dma_region_handle = 1844 GDMA_INVALID_DMA_REGION; 1845 1846 txq->gdma_txq_id = txq->gdma_sq->id; 1847 1848 cq->gdma_id = cq->gdma_cq->id; 1849 1850 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 1851 err = -EINVAL; 1852 goto out; 1853 } 1854 1855 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 1856 1857 netif_napi_add_tx(net, &cq->napi, mana_poll); 1858 napi_enable(&cq->napi); 1859 1860 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 1861 } 1862 1863 return 0; 1864 out: 1865 mana_destroy_txq(apc); 1866 return err; 1867 } 1868 1869 static void mana_destroy_rxq(struct mana_port_context *apc, 1870 struct mana_rxq *rxq, bool validate_state) 1871 1872 { 1873 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 1874 struct mana_recv_buf_oob *rx_oob; 1875 struct device *dev = gc->dev; 1876 struct napi_struct *napi; 1877 int i; 1878 1879 if (!rxq) 1880 return; 1881 1882 napi = &rxq->rx_cq.napi; 1883 1884 if (validate_state) 1885 napi_synchronize(napi); 1886 1887 napi_disable(napi); 1888 1889 xdp_rxq_info_unreg(&rxq->xdp_rxq); 1890 1891 netif_napi_del(napi); 1892 1893 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj); 1894 1895 mana_deinit_cq(apc, &rxq->rx_cq); 1896 1897 if (rxq->xdp_save_va) 1898 put_page(virt_to_head_page(rxq->xdp_save_va)); 1899 1900 for (i = 0; i < rxq->num_rx_buf; i++) { 1901 rx_oob = &rxq->rx_oobs[i]; 1902 1903 if (!rx_oob->buf_va) 1904 continue; 1905 1906 dma_unmap_single(dev, rx_oob->sgl[0].address, 1907 rx_oob->sgl[0].size, DMA_FROM_DEVICE); 1908 1909 put_page(virt_to_head_page(rx_oob->buf_va)); 1910 rx_oob->buf_va = NULL; 1911 } 1912 1913 if (rxq->gdma_rq) 1914 mana_gd_destroy_queue(gc, rxq->gdma_rq); 1915 1916 kfree(rxq); 1917 } 1918 1919 static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key, 1920 struct mana_rxq *rxq, struct device *dev) 1921 { 1922 struct mana_port_context *mpc = netdev_priv(rxq->ndev); 1923 dma_addr_t da; 1924 void *va; 1925 1926 if (mpc->rxbufs_pre) 1927 va = mana_get_rxbuf_pre(rxq, &da); 1928 else 1929 va = mana_get_rxfrag(rxq, dev, &da, false); 1930 1931 if (!va) 1932 return -ENOMEM; 1933 1934 rx_oob->buf_va = va; 1935 1936 rx_oob->sgl[0].address = da; 1937 rx_oob->sgl[0].size = rxq->datasize; 1938 rx_oob->sgl[0].mem_key = mem_key; 1939 1940 return 0; 1941 } 1942 1943 #define MANA_WQE_HEADER_SIZE 16 1944 #define MANA_WQE_SGE_SIZE 16 1945 1946 static int mana_alloc_rx_wqe(struct mana_port_context *apc, 1947 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size) 1948 { 1949 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 1950 struct mana_recv_buf_oob *rx_oob; 1951 struct device *dev = gc->dev; 1952 u32 buf_idx; 1953 int ret; 1954 1955 WARN_ON(rxq->datasize == 0); 1956 1957 *rxq_size = 0; 1958 *cq_size = 0; 1959 1960 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 1961 rx_oob = &rxq->rx_oobs[buf_idx]; 1962 memset(rx_oob, 0, sizeof(*rx_oob)); 1963 1964 rx_oob->num_sge = 1; 1965 1966 ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq, 1967 dev); 1968 if (ret) 1969 return ret; 1970 1971 rx_oob->wqe_req.sgl = rx_oob->sgl; 1972 rx_oob->wqe_req.num_sge = rx_oob->num_sge; 1973 rx_oob->wqe_req.inline_oob_size = 0; 1974 rx_oob->wqe_req.inline_oob_data = NULL; 1975 rx_oob->wqe_req.flags = 0; 1976 rx_oob->wqe_req.client_data_unit = 0; 1977 1978 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE + 1979 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32); 1980 *cq_size += COMP_ENTRY_SIZE; 1981 } 1982 1983 return 0; 1984 } 1985 1986 static int mana_push_wqe(struct mana_rxq *rxq) 1987 { 1988 struct mana_recv_buf_oob *rx_oob; 1989 u32 buf_idx; 1990 int err; 1991 1992 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 1993 rx_oob = &rxq->rx_oobs[buf_idx]; 1994 1995 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req, 1996 &rx_oob->wqe_inf); 1997 if (err) 1998 return -ENOSPC; 1999 } 2000 2001 return 0; 2002 } 2003 2004 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc, 2005 u32 rxq_idx, struct mana_eq *eq, 2006 struct net_device *ndev) 2007 { 2008 struct gdma_dev *gd = apc->ac->gdma_dev; 2009 struct mana_obj_spec wq_spec; 2010 struct mana_obj_spec cq_spec; 2011 struct gdma_queue_spec spec; 2012 struct mana_cq *cq = NULL; 2013 struct gdma_context *gc; 2014 u32 cq_size, rq_size; 2015 struct mana_rxq *rxq; 2016 int err; 2017 2018 gc = gd->gdma_context; 2019 2020 rxq = kzalloc(struct_size(rxq, rx_oobs, RX_BUFFERS_PER_QUEUE), 2021 GFP_KERNEL); 2022 if (!rxq) 2023 return NULL; 2024 2025 rxq->ndev = ndev; 2026 rxq->num_rx_buf = RX_BUFFERS_PER_QUEUE; 2027 rxq->rxq_idx = rxq_idx; 2028 rxq->rxobj = INVALID_MANA_HANDLE; 2029 2030 mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size, 2031 &rxq->headroom); 2032 2033 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size); 2034 if (err) 2035 goto out; 2036 2037 rq_size = PAGE_ALIGN(rq_size); 2038 cq_size = PAGE_ALIGN(cq_size); 2039 2040 /* Create RQ */ 2041 memset(&spec, 0, sizeof(spec)); 2042 spec.type = GDMA_RQ; 2043 spec.monitor_avl_buf = true; 2044 spec.queue_size = rq_size; 2045 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq); 2046 if (err) 2047 goto out; 2048 2049 /* Create RQ's CQ */ 2050 cq = &rxq->rx_cq; 2051 cq->type = MANA_CQ_TYPE_RX; 2052 cq->rxq = rxq; 2053 2054 memset(&spec, 0, sizeof(spec)); 2055 spec.type = GDMA_CQ; 2056 spec.monitor_avl_buf = false; 2057 spec.queue_size = cq_size; 2058 spec.cq.callback = mana_schedule_napi; 2059 spec.cq.parent_eq = eq->eq; 2060 spec.cq.context = cq; 2061 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 2062 if (err) 2063 goto out; 2064 2065 memset(&wq_spec, 0, sizeof(wq_spec)); 2066 memset(&cq_spec, 0, sizeof(cq_spec)); 2067 wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle; 2068 wq_spec.queue_size = rxq->gdma_rq->queue_size; 2069 2070 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 2071 cq_spec.queue_size = cq->gdma_cq->queue_size; 2072 cq_spec.modr_ctx_id = 0; 2073 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 2074 2075 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ, 2076 &wq_spec, &cq_spec, &rxq->rxobj); 2077 if (err) 2078 goto out; 2079 2080 rxq->gdma_rq->id = wq_spec.queue_index; 2081 cq->gdma_cq->id = cq_spec.queue_index; 2082 2083 rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2084 cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2085 2086 rxq->gdma_id = rxq->gdma_rq->id; 2087 cq->gdma_id = cq->gdma_cq->id; 2088 2089 err = mana_push_wqe(rxq); 2090 if (err) 2091 goto out; 2092 2093 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 2094 err = -EINVAL; 2095 goto out; 2096 } 2097 2098 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2099 2100 netif_napi_add_weight(ndev, &cq->napi, mana_poll, 1); 2101 2102 WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx, 2103 cq->napi.napi_id)); 2104 WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, 2105 MEM_TYPE_PAGE_SHARED, NULL)); 2106 2107 napi_enable(&cq->napi); 2108 2109 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2110 out: 2111 if (!err) 2112 return rxq; 2113 2114 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err); 2115 2116 mana_destroy_rxq(apc, rxq, false); 2117 2118 if (cq) 2119 mana_deinit_cq(apc, cq); 2120 2121 return NULL; 2122 } 2123 2124 static int mana_add_rx_queues(struct mana_port_context *apc, 2125 struct net_device *ndev) 2126 { 2127 struct mana_context *ac = apc->ac; 2128 struct mana_rxq *rxq; 2129 int err = 0; 2130 int i; 2131 2132 for (i = 0; i < apc->num_queues; i++) { 2133 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev); 2134 if (!rxq) { 2135 err = -ENOMEM; 2136 goto out; 2137 } 2138 2139 u64_stats_init(&rxq->stats.syncp); 2140 2141 apc->rxqs[i] = rxq; 2142 } 2143 2144 apc->default_rxobj = apc->rxqs[0]->rxobj; 2145 out: 2146 return err; 2147 } 2148 2149 static void mana_destroy_vport(struct mana_port_context *apc) 2150 { 2151 struct gdma_dev *gd = apc->ac->gdma_dev; 2152 struct mana_rxq *rxq; 2153 u32 rxq_idx; 2154 2155 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 2156 rxq = apc->rxqs[rxq_idx]; 2157 if (!rxq) 2158 continue; 2159 2160 mana_destroy_rxq(apc, rxq, true); 2161 apc->rxqs[rxq_idx] = NULL; 2162 } 2163 2164 mana_destroy_txq(apc); 2165 mana_uncfg_vport(apc); 2166 2167 if (gd->gdma_context->is_pf) 2168 mana_pf_deregister_hw_vport(apc); 2169 } 2170 2171 static int mana_create_vport(struct mana_port_context *apc, 2172 struct net_device *net) 2173 { 2174 struct gdma_dev *gd = apc->ac->gdma_dev; 2175 int err; 2176 2177 apc->default_rxobj = INVALID_MANA_HANDLE; 2178 2179 if (gd->gdma_context->is_pf) { 2180 err = mana_pf_register_hw_vport(apc); 2181 if (err) 2182 return err; 2183 } 2184 2185 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell); 2186 if (err) 2187 return err; 2188 2189 return mana_create_txq(apc, net); 2190 } 2191 2192 static void mana_rss_table_init(struct mana_port_context *apc) 2193 { 2194 int i; 2195 2196 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) 2197 apc->indir_table[i] = 2198 ethtool_rxfh_indir_default(i, apc->num_queues); 2199 } 2200 2201 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx, 2202 bool update_hash, bool update_tab) 2203 { 2204 u32 queue_idx; 2205 int err; 2206 int i; 2207 2208 if (update_tab) { 2209 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) { 2210 queue_idx = apc->indir_table[i]; 2211 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj; 2212 } 2213 } 2214 2215 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab); 2216 if (err) 2217 return err; 2218 2219 mana_fence_rqs(apc); 2220 2221 return 0; 2222 } 2223 2224 static int mana_init_port(struct net_device *ndev) 2225 { 2226 struct mana_port_context *apc = netdev_priv(ndev); 2227 u32 max_txq, max_rxq, max_queues; 2228 int port_idx = apc->port_idx; 2229 u32 num_indirect_entries; 2230 int err; 2231 2232 err = mana_init_port_context(apc); 2233 if (err) 2234 return err; 2235 2236 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq, 2237 &num_indirect_entries); 2238 if (err) { 2239 netdev_err(ndev, "Failed to query info for vPort %d\n", 2240 port_idx); 2241 goto reset_apc; 2242 } 2243 2244 max_queues = min_t(u32, max_txq, max_rxq); 2245 if (apc->max_queues > max_queues) 2246 apc->max_queues = max_queues; 2247 2248 if (apc->num_queues > apc->max_queues) 2249 apc->num_queues = apc->max_queues; 2250 2251 eth_hw_addr_set(ndev, apc->mac_addr); 2252 2253 return 0; 2254 2255 reset_apc: 2256 kfree(apc->rxqs); 2257 apc->rxqs = NULL; 2258 return err; 2259 } 2260 2261 int mana_alloc_queues(struct net_device *ndev) 2262 { 2263 struct mana_port_context *apc = netdev_priv(ndev); 2264 struct gdma_dev *gd = apc->ac->gdma_dev; 2265 int err; 2266 2267 err = mana_create_vport(apc, ndev); 2268 if (err) 2269 return err; 2270 2271 err = netif_set_real_num_tx_queues(ndev, apc->num_queues); 2272 if (err) 2273 goto destroy_vport; 2274 2275 err = mana_add_rx_queues(apc, ndev); 2276 if (err) 2277 goto destroy_vport; 2278 2279 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE; 2280 2281 err = netif_set_real_num_rx_queues(ndev, apc->num_queues); 2282 if (err) 2283 goto destroy_vport; 2284 2285 mana_rss_table_init(apc); 2286 2287 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true); 2288 if (err) 2289 goto destroy_vport; 2290 2291 if (gd->gdma_context->is_pf) { 2292 err = mana_pf_register_filter(apc); 2293 if (err) 2294 goto destroy_vport; 2295 } 2296 2297 mana_chn_setxdp(apc, mana_xdp_get(apc)); 2298 2299 return 0; 2300 2301 destroy_vport: 2302 mana_destroy_vport(apc); 2303 return err; 2304 } 2305 2306 int mana_attach(struct net_device *ndev) 2307 { 2308 struct mana_port_context *apc = netdev_priv(ndev); 2309 int err; 2310 2311 ASSERT_RTNL(); 2312 2313 err = mana_init_port(ndev); 2314 if (err) 2315 return err; 2316 2317 if (apc->port_st_save) { 2318 err = mana_alloc_queues(ndev); 2319 if (err) { 2320 mana_cleanup_port_context(apc); 2321 return err; 2322 } 2323 } 2324 2325 apc->port_is_up = apc->port_st_save; 2326 2327 /* Ensure port state updated before txq state */ 2328 smp_wmb(); 2329 2330 if (apc->port_is_up) 2331 netif_carrier_on(ndev); 2332 2333 netif_device_attach(ndev); 2334 2335 return 0; 2336 } 2337 2338 static int mana_dealloc_queues(struct net_device *ndev) 2339 { 2340 struct mana_port_context *apc = netdev_priv(ndev); 2341 struct gdma_dev *gd = apc->ac->gdma_dev; 2342 struct mana_txq *txq; 2343 int i, err; 2344 2345 if (apc->port_is_up) 2346 return -EINVAL; 2347 2348 mana_chn_setxdp(apc, NULL); 2349 2350 if (gd->gdma_context->is_pf) 2351 mana_pf_deregister_filter(apc); 2352 2353 /* No packet can be transmitted now since apc->port_is_up is false. 2354 * There is still a tiny chance that mana_poll_tx_cq() can re-enable 2355 * a txq because it may not timely see apc->port_is_up being cleared 2356 * to false, but it doesn't matter since mana_start_xmit() drops any 2357 * new packets due to apc->port_is_up being false. 2358 * 2359 * Drain all the in-flight TX packets 2360 */ 2361 for (i = 0; i < apc->num_queues; i++) { 2362 txq = &apc->tx_qp[i].txq; 2363 2364 while (atomic_read(&txq->pending_sends) > 0) 2365 usleep_range(1000, 2000); 2366 } 2367 2368 /* We're 100% sure the queues can no longer be woken up, because 2369 * we're sure now mana_poll_tx_cq() can't be running. 2370 */ 2371 2372 apc->rss_state = TRI_STATE_FALSE; 2373 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false); 2374 if (err) { 2375 netdev_err(ndev, "Failed to disable vPort: %d\n", err); 2376 return err; 2377 } 2378 2379 mana_destroy_vport(apc); 2380 2381 return 0; 2382 } 2383 2384 int mana_detach(struct net_device *ndev, bool from_close) 2385 { 2386 struct mana_port_context *apc = netdev_priv(ndev); 2387 int err; 2388 2389 ASSERT_RTNL(); 2390 2391 apc->port_st_save = apc->port_is_up; 2392 apc->port_is_up = false; 2393 2394 /* Ensure port state updated before txq state */ 2395 smp_wmb(); 2396 2397 netif_tx_disable(ndev); 2398 netif_carrier_off(ndev); 2399 2400 if (apc->port_st_save) { 2401 err = mana_dealloc_queues(ndev); 2402 if (err) 2403 return err; 2404 } 2405 2406 if (!from_close) { 2407 netif_device_detach(ndev); 2408 mana_cleanup_port_context(apc); 2409 } 2410 2411 return 0; 2412 } 2413 2414 static int mana_probe_port(struct mana_context *ac, int port_idx, 2415 struct net_device **ndev_storage) 2416 { 2417 struct gdma_context *gc = ac->gdma_dev->gdma_context; 2418 struct mana_port_context *apc; 2419 struct net_device *ndev; 2420 int err; 2421 2422 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context), 2423 gc->max_num_queues); 2424 if (!ndev) 2425 return -ENOMEM; 2426 2427 *ndev_storage = ndev; 2428 2429 apc = netdev_priv(ndev); 2430 apc->ac = ac; 2431 apc->ndev = ndev; 2432 apc->max_queues = gc->max_num_queues; 2433 apc->num_queues = gc->max_num_queues; 2434 apc->port_handle = INVALID_MANA_HANDLE; 2435 apc->pf_filter_handle = INVALID_MANA_HANDLE; 2436 apc->port_idx = port_idx; 2437 2438 mutex_init(&apc->vport_mutex); 2439 apc->vport_use_count = 0; 2440 2441 ndev->netdev_ops = &mana_devops; 2442 ndev->ethtool_ops = &mana_ethtool_ops; 2443 ndev->mtu = ETH_DATA_LEN; 2444 ndev->max_mtu = gc->adapter_mtu - ETH_HLEN; 2445 ndev->min_mtu = ETH_MIN_MTU; 2446 ndev->needed_headroom = MANA_HEADROOM; 2447 ndev->dev_port = port_idx; 2448 SET_NETDEV_DEV(ndev, gc->dev); 2449 2450 netif_carrier_off(ndev); 2451 2452 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE); 2453 2454 err = mana_init_port(ndev); 2455 if (err) 2456 goto free_net; 2457 2458 netdev_lockdep_set_classes(ndev); 2459 2460 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 2461 ndev->hw_features |= NETIF_F_RXCSUM; 2462 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6; 2463 ndev->hw_features |= NETIF_F_RXHASH; 2464 ndev->features = ndev->hw_features; 2465 ndev->vlan_features = 0; 2466 ndev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT | 2467 NETDEV_XDP_ACT_NDO_XMIT; 2468 2469 err = register_netdev(ndev); 2470 if (err) { 2471 netdev_err(ndev, "Unable to register netdev.\n"); 2472 goto reset_apc; 2473 } 2474 2475 return 0; 2476 2477 reset_apc: 2478 kfree(apc->rxqs); 2479 apc->rxqs = NULL; 2480 free_net: 2481 *ndev_storage = NULL; 2482 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err); 2483 free_netdev(ndev); 2484 return err; 2485 } 2486 2487 static void adev_release(struct device *dev) 2488 { 2489 struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev); 2490 2491 kfree(madev); 2492 } 2493 2494 static void remove_adev(struct gdma_dev *gd) 2495 { 2496 struct auxiliary_device *adev = gd->adev; 2497 int id = adev->id; 2498 2499 auxiliary_device_delete(adev); 2500 auxiliary_device_uninit(adev); 2501 2502 mana_adev_idx_free(id); 2503 gd->adev = NULL; 2504 } 2505 2506 static int add_adev(struct gdma_dev *gd) 2507 { 2508 struct auxiliary_device *adev; 2509 struct mana_adev *madev; 2510 int ret; 2511 2512 madev = kzalloc(sizeof(*madev), GFP_KERNEL); 2513 if (!madev) 2514 return -ENOMEM; 2515 2516 adev = &madev->adev; 2517 ret = mana_adev_idx_alloc(); 2518 if (ret < 0) 2519 goto idx_fail; 2520 adev->id = ret; 2521 2522 adev->name = "rdma"; 2523 adev->dev.parent = gd->gdma_context->dev; 2524 adev->dev.release = adev_release; 2525 madev->mdev = gd; 2526 2527 ret = auxiliary_device_init(adev); 2528 if (ret) 2529 goto init_fail; 2530 2531 ret = auxiliary_device_add(adev); 2532 if (ret) 2533 goto add_fail; 2534 2535 gd->adev = adev; 2536 return 0; 2537 2538 add_fail: 2539 auxiliary_device_uninit(adev); 2540 2541 init_fail: 2542 mana_adev_idx_free(adev->id); 2543 2544 idx_fail: 2545 kfree(madev); 2546 2547 return ret; 2548 } 2549 2550 int mana_probe(struct gdma_dev *gd, bool resuming) 2551 { 2552 struct gdma_context *gc = gd->gdma_context; 2553 struct mana_context *ac = gd->driver_data; 2554 struct device *dev = gc->dev; 2555 u16 num_ports = 0; 2556 int err; 2557 int i; 2558 2559 dev_info(dev, 2560 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n", 2561 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION); 2562 2563 err = mana_gd_register_device(gd); 2564 if (err) 2565 return err; 2566 2567 if (!resuming) { 2568 ac = kzalloc(sizeof(*ac), GFP_KERNEL); 2569 if (!ac) 2570 return -ENOMEM; 2571 2572 ac->gdma_dev = gd; 2573 gd->driver_data = ac; 2574 } 2575 2576 err = mana_create_eq(ac); 2577 if (err) 2578 goto out; 2579 2580 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION, 2581 MANA_MICRO_VERSION, &num_ports); 2582 if (err) 2583 goto out; 2584 2585 if (!resuming) { 2586 ac->num_ports = num_ports; 2587 } else { 2588 if (ac->num_ports != num_ports) { 2589 dev_err(dev, "The number of vPorts changed: %d->%d\n", 2590 ac->num_ports, num_ports); 2591 err = -EPROTO; 2592 goto out; 2593 } 2594 } 2595 2596 if (ac->num_ports == 0) 2597 dev_err(dev, "Failed to detect any vPort\n"); 2598 2599 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV) 2600 ac->num_ports = MAX_PORTS_IN_MANA_DEV; 2601 2602 if (!resuming) { 2603 for (i = 0; i < ac->num_ports; i++) { 2604 err = mana_probe_port(ac, i, &ac->ports[i]); 2605 if (err) 2606 break; 2607 } 2608 } else { 2609 for (i = 0; i < ac->num_ports; i++) { 2610 rtnl_lock(); 2611 err = mana_attach(ac->ports[i]); 2612 rtnl_unlock(); 2613 if (err) 2614 break; 2615 } 2616 } 2617 2618 err = add_adev(gd); 2619 out: 2620 if (err) 2621 mana_remove(gd, false); 2622 2623 return err; 2624 } 2625 2626 void mana_remove(struct gdma_dev *gd, bool suspending) 2627 { 2628 struct gdma_context *gc = gd->gdma_context; 2629 struct mana_context *ac = gd->driver_data; 2630 struct device *dev = gc->dev; 2631 struct net_device *ndev; 2632 int err; 2633 int i; 2634 2635 /* adev currently doesn't support suspending, always remove it */ 2636 if (gd->adev) 2637 remove_adev(gd); 2638 2639 for (i = 0; i < ac->num_ports; i++) { 2640 ndev = ac->ports[i]; 2641 if (!ndev) { 2642 if (i == 0) 2643 dev_err(dev, "No net device to remove\n"); 2644 goto out; 2645 } 2646 2647 /* All cleanup actions should stay after rtnl_lock(), otherwise 2648 * other functions may access partially cleaned up data. 2649 */ 2650 rtnl_lock(); 2651 2652 err = mana_detach(ndev, false); 2653 if (err) 2654 netdev_err(ndev, "Failed to detach vPort %d: %d\n", 2655 i, err); 2656 2657 if (suspending) { 2658 /* No need to unregister the ndev. */ 2659 rtnl_unlock(); 2660 continue; 2661 } 2662 2663 unregister_netdevice(ndev); 2664 2665 rtnl_unlock(); 2666 2667 free_netdev(ndev); 2668 } 2669 2670 mana_destroy_eq(ac); 2671 out: 2672 mana_gd_deregister_device(gd); 2673 2674 if (suspending) 2675 return; 2676 2677 gd->driver_data = NULL; 2678 gd->gdma_context = NULL; 2679 kfree(ac); 2680 } 2681