1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause 2 /* Copyright (c) 2021, Microsoft Corporation. */ 3 4 #include <uapi/linux/bpf.h> 5 6 #include <linux/debugfs.h> 7 #include <linux/inetdevice.h> 8 #include <linux/etherdevice.h> 9 #include <linux/ethtool.h> 10 #include <linux/filter.h> 11 #include <linux/mm.h> 12 #include <linux/pci.h> 13 14 #include <net/checksum.h> 15 #include <net/ip6_checksum.h> 16 #include <net/netdev_lock.h> 17 #include <net/page_pool/helpers.h> 18 #include <net/xdp.h> 19 20 #include <net/mana/mana.h> 21 #include <net/mana/mana_auxiliary.h> 22 23 static DEFINE_IDA(mana_adev_ida); 24 25 static int mana_adev_idx_alloc(void) 26 { 27 return ida_alloc(&mana_adev_ida, GFP_KERNEL); 28 } 29 30 static void mana_adev_idx_free(int idx) 31 { 32 ida_free(&mana_adev_ida, idx); 33 } 34 35 static ssize_t mana_dbg_q_read(struct file *filp, char __user *buf, size_t count, 36 loff_t *pos) 37 { 38 struct gdma_queue *gdma_q = filp->private_data; 39 40 return simple_read_from_buffer(buf, count, pos, gdma_q->queue_mem_ptr, 41 gdma_q->queue_size); 42 } 43 44 static const struct file_operations mana_dbg_q_fops = { 45 .owner = THIS_MODULE, 46 .open = simple_open, 47 .read = mana_dbg_q_read, 48 }; 49 50 static bool mana_en_need_log(struct mana_port_context *apc, int err) 51 { 52 if (apc && apc->ac && apc->ac->gdma_dev && 53 apc->ac->gdma_dev->gdma_context) 54 return mana_need_log(apc->ac->gdma_dev->gdma_context, err); 55 else 56 return true; 57 } 58 59 /* Microsoft Azure Network Adapter (MANA) functions */ 60 61 static int mana_open(struct net_device *ndev) 62 { 63 struct mana_port_context *apc = netdev_priv(ndev); 64 int err; 65 err = mana_alloc_queues(ndev); 66 67 if (err) { 68 netdev_err(ndev, "%s failed to allocate queues: %d\n", __func__, err); 69 return err; 70 } 71 72 apc->port_is_up = true; 73 74 /* Ensure port state updated before txq state */ 75 smp_wmb(); 76 77 netif_carrier_on(ndev); 78 netif_tx_wake_all_queues(ndev); 79 netdev_dbg(ndev, "%s successful\n", __func__); 80 return 0; 81 } 82 83 static int mana_close(struct net_device *ndev) 84 { 85 struct mana_port_context *apc = netdev_priv(ndev); 86 87 if (!apc->port_is_up) 88 return 0; 89 90 return mana_detach(ndev, true); 91 } 92 93 static bool mana_can_tx(struct gdma_queue *wq) 94 { 95 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE; 96 } 97 98 static unsigned int mana_checksum_info(struct sk_buff *skb) 99 { 100 if (skb->protocol == htons(ETH_P_IP)) { 101 struct iphdr *ip = ip_hdr(skb); 102 103 if (ip->protocol == IPPROTO_TCP) 104 return IPPROTO_TCP; 105 106 if (ip->protocol == IPPROTO_UDP) 107 return IPPROTO_UDP; 108 } else if (skb->protocol == htons(ETH_P_IPV6)) { 109 struct ipv6hdr *ip6 = ipv6_hdr(skb); 110 111 if (ip6->nexthdr == IPPROTO_TCP) 112 return IPPROTO_TCP; 113 114 if (ip6->nexthdr == IPPROTO_UDP) 115 return IPPROTO_UDP; 116 } 117 118 /* No csum offloading */ 119 return 0; 120 } 121 122 static void mana_add_sge(struct mana_tx_package *tp, struct mana_skb_head *ash, 123 int sg_i, dma_addr_t da, int sge_len, u32 gpa_mkey) 124 { 125 ash->dma_handle[sg_i] = da; 126 ash->size[sg_i] = sge_len; 127 128 tp->wqe_req.sgl[sg_i].address = da; 129 tp->wqe_req.sgl[sg_i].mem_key = gpa_mkey; 130 tp->wqe_req.sgl[sg_i].size = sge_len; 131 } 132 133 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc, 134 struct mana_tx_package *tp, int gso_hs) 135 { 136 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 137 int hsg = 1; /* num of SGEs of linear part */ 138 struct gdma_dev *gd = apc->ac->gdma_dev; 139 int skb_hlen = skb_headlen(skb); 140 int sge0_len, sge1_len = 0; 141 struct gdma_context *gc; 142 struct device *dev; 143 skb_frag_t *frag; 144 dma_addr_t da; 145 int sg_i; 146 int i; 147 148 gc = gd->gdma_context; 149 dev = gc->dev; 150 151 if (gso_hs && gso_hs < skb_hlen) { 152 sge0_len = gso_hs; 153 sge1_len = skb_hlen - gso_hs; 154 } else { 155 sge0_len = skb_hlen; 156 } 157 158 da = dma_map_single(dev, skb->data, sge0_len, DMA_TO_DEVICE); 159 if (dma_mapping_error(dev, da)) 160 return -ENOMEM; 161 162 mana_add_sge(tp, ash, 0, da, sge0_len, gd->gpa_mkey); 163 164 if (sge1_len) { 165 sg_i = 1; 166 da = dma_map_single(dev, skb->data + sge0_len, sge1_len, 167 DMA_TO_DEVICE); 168 if (dma_mapping_error(dev, da)) 169 goto frag_err; 170 171 mana_add_sge(tp, ash, sg_i, da, sge1_len, gd->gpa_mkey); 172 hsg = 2; 173 } 174 175 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 176 sg_i = hsg + i; 177 178 frag = &skb_shinfo(skb)->frags[i]; 179 da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag), 180 DMA_TO_DEVICE); 181 if (dma_mapping_error(dev, da)) 182 goto frag_err; 183 184 mana_add_sge(tp, ash, sg_i, da, skb_frag_size(frag), 185 gd->gpa_mkey); 186 } 187 188 return 0; 189 190 frag_err: 191 if (net_ratelimit()) 192 netdev_err(apc->ndev, "Failed to map skb of size %u to DMA\n", 193 skb->len); 194 for (i = sg_i - 1; i >= hsg; i--) 195 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i], 196 DMA_TO_DEVICE); 197 198 for (i = hsg - 1; i >= 0; i--) 199 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i], 200 DMA_TO_DEVICE); 201 202 return -ENOMEM; 203 } 204 205 /* Handle the case when GSO SKB linear length is too large. 206 * MANA NIC requires GSO packets to put only the packet header to SGE0. 207 * So, we need 2 SGEs for the skb linear part which contains more than the 208 * header. 209 * Return a positive value for the number of SGEs, or a negative value 210 * for an error. 211 */ 212 static int mana_fix_skb_head(struct net_device *ndev, struct sk_buff *skb, 213 int gso_hs) 214 { 215 int num_sge = 1 + skb_shinfo(skb)->nr_frags; 216 int skb_hlen = skb_headlen(skb); 217 218 if (gso_hs < skb_hlen) { 219 num_sge++; 220 } else if (gso_hs > skb_hlen) { 221 if (net_ratelimit()) 222 netdev_err(ndev, 223 "TX nonlinear head: hs:%d, skb_hlen:%d\n", 224 gso_hs, skb_hlen); 225 226 return -EINVAL; 227 } 228 229 return num_sge; 230 } 231 232 /* Get the GSO packet's header size */ 233 static int mana_get_gso_hs(struct sk_buff *skb) 234 { 235 int gso_hs; 236 237 if (skb->encapsulation) { 238 gso_hs = skb_inner_tcp_all_headers(skb); 239 } else { 240 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) { 241 gso_hs = skb_transport_offset(skb) + 242 sizeof(struct udphdr); 243 } else { 244 gso_hs = skb_tcp_all_headers(skb); 245 } 246 } 247 248 return gso_hs; 249 } 250 251 netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev) 252 { 253 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT; 254 struct mana_port_context *apc = netdev_priv(ndev); 255 int gso_hs = 0; /* zero for non-GSO pkts */ 256 u16 txq_idx = skb_get_queue_mapping(skb); 257 struct gdma_dev *gd = apc->ac->gdma_dev; 258 bool ipv4 = false, ipv6 = false; 259 struct mana_tx_package pkg = {}; 260 struct netdev_queue *net_txq; 261 struct mana_stats_tx *tx_stats; 262 struct gdma_queue *gdma_sq; 263 int err, len, num_gso_seg; 264 unsigned int csum_type; 265 struct mana_txq *txq; 266 struct mana_cq *cq; 267 268 if (unlikely(!apc->port_is_up)) 269 goto tx_drop; 270 271 if (skb_cow_head(skb, MANA_HEADROOM)) 272 goto tx_drop_count; 273 274 if (unlikely(ipv6_hopopt_jumbo_remove(skb))) 275 goto tx_drop_count; 276 277 txq = &apc->tx_qp[txq_idx].txq; 278 gdma_sq = txq->gdma_sq; 279 cq = &apc->tx_qp[txq_idx].tx_cq; 280 tx_stats = &txq->stats; 281 282 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id; 283 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame; 284 285 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) { 286 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset; 287 pkt_fmt = MANA_LONG_PKT_FMT; 288 } else { 289 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset; 290 } 291 292 if (skb_vlan_tag_present(skb)) { 293 pkt_fmt = MANA_LONG_PKT_FMT; 294 pkg.tx_oob.l_oob.inject_vlan_pri_tag = 1; 295 pkg.tx_oob.l_oob.pcp = skb_vlan_tag_get_prio(skb); 296 pkg.tx_oob.l_oob.dei = skb_vlan_tag_get_cfi(skb); 297 pkg.tx_oob.l_oob.vlan_id = skb_vlan_tag_get_id(skb); 298 } 299 300 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt; 301 302 if (pkt_fmt == MANA_SHORT_PKT_FMT) { 303 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob); 304 u64_stats_update_begin(&tx_stats->syncp); 305 tx_stats->short_pkt_fmt++; 306 u64_stats_update_end(&tx_stats->syncp); 307 } else { 308 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob); 309 u64_stats_update_begin(&tx_stats->syncp); 310 tx_stats->long_pkt_fmt++; 311 u64_stats_update_end(&tx_stats->syncp); 312 } 313 314 pkg.wqe_req.inline_oob_data = &pkg.tx_oob; 315 pkg.wqe_req.flags = 0; 316 pkg.wqe_req.client_data_unit = 0; 317 318 pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags; 319 320 if (skb->protocol == htons(ETH_P_IP)) 321 ipv4 = true; 322 else if (skb->protocol == htons(ETH_P_IPV6)) 323 ipv6 = true; 324 325 if (skb_is_gso(skb)) { 326 int num_sge; 327 328 gso_hs = mana_get_gso_hs(skb); 329 330 num_sge = mana_fix_skb_head(ndev, skb, gso_hs); 331 if (num_sge > 0) 332 pkg.wqe_req.num_sge = num_sge; 333 else 334 goto tx_drop_count; 335 336 u64_stats_update_begin(&tx_stats->syncp); 337 if (skb->encapsulation) { 338 tx_stats->tso_inner_packets++; 339 tx_stats->tso_inner_bytes += skb->len - gso_hs; 340 } else { 341 tx_stats->tso_packets++; 342 tx_stats->tso_bytes += skb->len - gso_hs; 343 } 344 u64_stats_update_end(&tx_stats->syncp); 345 346 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 347 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 348 349 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 350 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 351 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 352 353 pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size; 354 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0; 355 if (ipv4) { 356 ip_hdr(skb)->tot_len = 0; 357 ip_hdr(skb)->check = 0; 358 tcp_hdr(skb)->check = 359 ~csum_tcpudp_magic(ip_hdr(skb)->saddr, 360 ip_hdr(skb)->daddr, 0, 361 IPPROTO_TCP, 0); 362 } else { 363 ipv6_hdr(skb)->payload_len = 0; 364 tcp_hdr(skb)->check = 365 ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr, 366 &ipv6_hdr(skb)->daddr, 0, 367 IPPROTO_TCP, 0); 368 } 369 } else if (skb->ip_summed == CHECKSUM_PARTIAL) { 370 csum_type = mana_checksum_info(skb); 371 372 u64_stats_update_begin(&tx_stats->syncp); 373 tx_stats->csum_partial++; 374 u64_stats_update_end(&tx_stats->syncp); 375 376 if (csum_type == IPPROTO_TCP) { 377 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 378 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 379 380 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 381 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 382 383 } else if (csum_type == IPPROTO_UDP) { 384 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 385 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 386 387 pkg.tx_oob.s_oob.comp_udp_csum = 1; 388 } else { 389 /* Can't do offload of this type of checksum */ 390 if (skb_checksum_help(skb)) 391 goto tx_drop_count; 392 } 393 } 394 395 WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES); 396 397 if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) { 398 pkg.wqe_req.sgl = pkg.sgl_array; 399 } else { 400 pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge, 401 sizeof(struct gdma_sge), 402 GFP_ATOMIC); 403 if (!pkg.sgl_ptr) 404 goto tx_drop_count; 405 406 pkg.wqe_req.sgl = pkg.sgl_ptr; 407 } 408 409 if (mana_map_skb(skb, apc, &pkg, gso_hs)) { 410 u64_stats_update_begin(&tx_stats->syncp); 411 tx_stats->mana_map_err++; 412 u64_stats_update_end(&tx_stats->syncp); 413 goto free_sgl_ptr; 414 } 415 416 skb_queue_tail(&txq->pending_skbs, skb); 417 418 len = skb->len; 419 num_gso_seg = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1; 420 net_txq = netdev_get_tx_queue(ndev, txq_idx); 421 422 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req, 423 (struct gdma_posted_wqe_info *)skb->cb); 424 if (!mana_can_tx(gdma_sq)) { 425 netif_tx_stop_queue(net_txq); 426 apc->eth_stats.stop_queue++; 427 } 428 429 if (err) { 430 (void)skb_dequeue_tail(&txq->pending_skbs); 431 netdev_warn(ndev, "Failed to post TX OOB: %d\n", err); 432 err = NETDEV_TX_BUSY; 433 goto tx_busy; 434 } 435 436 err = NETDEV_TX_OK; 437 atomic_inc(&txq->pending_sends); 438 439 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq); 440 441 /* skb may be freed after mana_gd_post_work_request. Do not use it. */ 442 skb = NULL; 443 444 /* Populated the packet and bytes counters based on post GSO packet 445 * calculations 446 */ 447 tx_stats = &txq->stats; 448 u64_stats_update_begin(&tx_stats->syncp); 449 tx_stats->packets += num_gso_seg; 450 tx_stats->bytes += len + ((num_gso_seg - 1) * gso_hs); 451 u64_stats_update_end(&tx_stats->syncp); 452 453 tx_busy: 454 if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) { 455 netif_tx_wake_queue(net_txq); 456 apc->eth_stats.wake_queue++; 457 } 458 459 kfree(pkg.sgl_ptr); 460 return err; 461 462 free_sgl_ptr: 463 kfree(pkg.sgl_ptr); 464 tx_drop_count: 465 ndev->stats.tx_dropped++; 466 tx_drop: 467 dev_kfree_skb_any(skb); 468 return NETDEV_TX_OK; 469 } 470 471 static void mana_get_stats64(struct net_device *ndev, 472 struct rtnl_link_stats64 *st) 473 { 474 struct mana_port_context *apc = netdev_priv(ndev); 475 unsigned int num_queues = apc->num_queues; 476 struct mana_stats_rx *rx_stats; 477 struct mana_stats_tx *tx_stats; 478 unsigned int start; 479 u64 packets, bytes; 480 int q; 481 482 if (!apc->port_is_up) 483 return; 484 485 netdev_stats_to_stats64(st, &ndev->stats); 486 487 for (q = 0; q < num_queues; q++) { 488 rx_stats = &apc->rxqs[q]->stats; 489 490 do { 491 start = u64_stats_fetch_begin(&rx_stats->syncp); 492 packets = rx_stats->packets; 493 bytes = rx_stats->bytes; 494 } while (u64_stats_fetch_retry(&rx_stats->syncp, start)); 495 496 st->rx_packets += packets; 497 st->rx_bytes += bytes; 498 } 499 500 for (q = 0; q < num_queues; q++) { 501 tx_stats = &apc->tx_qp[q].txq.stats; 502 503 do { 504 start = u64_stats_fetch_begin(&tx_stats->syncp); 505 packets = tx_stats->packets; 506 bytes = tx_stats->bytes; 507 } while (u64_stats_fetch_retry(&tx_stats->syncp, start)); 508 509 st->tx_packets += packets; 510 st->tx_bytes += bytes; 511 } 512 } 513 514 static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb, 515 int old_q) 516 { 517 struct mana_port_context *apc = netdev_priv(ndev); 518 u32 hash = skb_get_hash(skb); 519 struct sock *sk = skb->sk; 520 int txq; 521 522 txq = apc->indir_table[hash & (apc->indir_table_sz - 1)]; 523 524 if (txq != old_q && sk && sk_fullsock(sk) && 525 rcu_access_pointer(sk->sk_dst_cache)) 526 sk_tx_queue_set(sk, txq); 527 528 return txq; 529 } 530 531 static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb, 532 struct net_device *sb_dev) 533 { 534 int txq; 535 536 if (ndev->real_num_tx_queues == 1) 537 return 0; 538 539 txq = sk_tx_queue_get(skb->sk); 540 541 if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) { 542 if (skb_rx_queue_recorded(skb)) 543 txq = skb_get_rx_queue(skb); 544 else 545 txq = mana_get_tx_queue(ndev, skb, txq); 546 } 547 548 return txq; 549 } 550 551 /* Release pre-allocated RX buffers */ 552 void mana_pre_dealloc_rxbufs(struct mana_port_context *mpc) 553 { 554 struct device *dev; 555 int i; 556 557 dev = mpc->ac->gdma_dev->gdma_context->dev; 558 559 if (!mpc->rxbufs_pre) 560 goto out1; 561 562 if (!mpc->das_pre) 563 goto out2; 564 565 while (mpc->rxbpre_total) { 566 i = --mpc->rxbpre_total; 567 dma_unmap_single(dev, mpc->das_pre[i], mpc->rxbpre_datasize, 568 DMA_FROM_DEVICE); 569 put_page(virt_to_head_page(mpc->rxbufs_pre[i])); 570 } 571 572 kfree(mpc->das_pre); 573 mpc->das_pre = NULL; 574 575 out2: 576 kfree(mpc->rxbufs_pre); 577 mpc->rxbufs_pre = NULL; 578 579 out1: 580 mpc->rxbpre_datasize = 0; 581 mpc->rxbpre_alloc_size = 0; 582 mpc->rxbpre_headroom = 0; 583 } 584 585 /* Get a buffer from the pre-allocated RX buffers */ 586 static void *mana_get_rxbuf_pre(struct mana_rxq *rxq, dma_addr_t *da) 587 { 588 struct net_device *ndev = rxq->ndev; 589 struct mana_port_context *mpc; 590 void *va; 591 592 mpc = netdev_priv(ndev); 593 594 if (!mpc->rxbufs_pre || !mpc->das_pre || !mpc->rxbpre_total) { 595 netdev_err(ndev, "No RX pre-allocated bufs\n"); 596 return NULL; 597 } 598 599 /* Check sizes to catch unexpected coding error */ 600 if (mpc->rxbpre_datasize != rxq->datasize) { 601 netdev_err(ndev, "rxbpre_datasize mismatch: %u: %u\n", 602 mpc->rxbpre_datasize, rxq->datasize); 603 return NULL; 604 } 605 606 if (mpc->rxbpre_alloc_size != rxq->alloc_size) { 607 netdev_err(ndev, "rxbpre_alloc_size mismatch: %u: %u\n", 608 mpc->rxbpre_alloc_size, rxq->alloc_size); 609 return NULL; 610 } 611 612 if (mpc->rxbpre_headroom != rxq->headroom) { 613 netdev_err(ndev, "rxbpre_headroom mismatch: %u: %u\n", 614 mpc->rxbpre_headroom, rxq->headroom); 615 return NULL; 616 } 617 618 mpc->rxbpre_total--; 619 620 *da = mpc->das_pre[mpc->rxbpre_total]; 621 va = mpc->rxbufs_pre[mpc->rxbpre_total]; 622 mpc->rxbufs_pre[mpc->rxbpre_total] = NULL; 623 624 /* Deallocate the array after all buffers are gone */ 625 if (!mpc->rxbpre_total) 626 mana_pre_dealloc_rxbufs(mpc); 627 628 return va; 629 } 630 631 /* Get RX buffer's data size, alloc size, XDP headroom based on MTU */ 632 static void mana_get_rxbuf_cfg(int mtu, u32 *datasize, u32 *alloc_size, 633 u32 *headroom) 634 { 635 if (mtu > MANA_XDP_MTU_MAX) 636 *headroom = 0; /* no support for XDP */ 637 else 638 *headroom = XDP_PACKET_HEADROOM; 639 640 *alloc_size = SKB_DATA_ALIGN(mtu + MANA_RXBUF_PAD + *headroom); 641 642 /* Using page pool in this case, so alloc_size is PAGE_SIZE */ 643 if (*alloc_size < PAGE_SIZE) 644 *alloc_size = PAGE_SIZE; 645 646 *datasize = mtu + ETH_HLEN; 647 } 648 649 int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu, int num_queues) 650 { 651 struct device *dev; 652 struct page *page; 653 dma_addr_t da; 654 int num_rxb; 655 void *va; 656 int i; 657 658 mana_get_rxbuf_cfg(new_mtu, &mpc->rxbpre_datasize, 659 &mpc->rxbpre_alloc_size, &mpc->rxbpre_headroom); 660 661 dev = mpc->ac->gdma_dev->gdma_context->dev; 662 663 num_rxb = num_queues * mpc->rx_queue_size; 664 665 WARN(mpc->rxbufs_pre, "mana rxbufs_pre exists\n"); 666 mpc->rxbufs_pre = kmalloc_array(num_rxb, sizeof(void *), GFP_KERNEL); 667 if (!mpc->rxbufs_pre) 668 goto error; 669 670 mpc->das_pre = kmalloc_array(num_rxb, sizeof(dma_addr_t), GFP_KERNEL); 671 if (!mpc->das_pre) 672 goto error; 673 674 mpc->rxbpre_total = 0; 675 676 for (i = 0; i < num_rxb; i++) { 677 page = dev_alloc_pages(get_order(mpc->rxbpre_alloc_size)); 678 if (!page) 679 goto error; 680 681 va = page_to_virt(page); 682 683 da = dma_map_single(dev, va + mpc->rxbpre_headroom, 684 mpc->rxbpre_datasize, DMA_FROM_DEVICE); 685 if (dma_mapping_error(dev, da)) { 686 put_page(page); 687 goto error; 688 } 689 690 mpc->rxbufs_pre[i] = va; 691 mpc->das_pre[i] = da; 692 mpc->rxbpre_total = i + 1; 693 } 694 695 return 0; 696 697 error: 698 netdev_err(mpc->ndev, "Failed to pre-allocate RX buffers for %d queues\n", num_queues); 699 mana_pre_dealloc_rxbufs(mpc); 700 return -ENOMEM; 701 } 702 703 static int mana_change_mtu(struct net_device *ndev, int new_mtu) 704 { 705 struct mana_port_context *mpc = netdev_priv(ndev); 706 unsigned int old_mtu = ndev->mtu; 707 int err; 708 709 /* Pre-allocate buffers to prevent failure in mana_attach later */ 710 err = mana_pre_alloc_rxbufs(mpc, new_mtu, mpc->num_queues); 711 if (err) { 712 netdev_err(ndev, "Insufficient memory for new MTU\n"); 713 return err; 714 } 715 716 err = mana_detach(ndev, false); 717 if (err) { 718 netdev_err(ndev, "mana_detach failed: %d\n", err); 719 goto out; 720 } 721 722 WRITE_ONCE(ndev->mtu, new_mtu); 723 724 err = mana_attach(ndev); 725 if (err) { 726 netdev_err(ndev, "mana_attach failed: %d\n", err); 727 WRITE_ONCE(ndev->mtu, old_mtu); 728 } 729 730 out: 731 mana_pre_dealloc_rxbufs(mpc); 732 return err; 733 } 734 735 static int mana_shaper_set(struct net_shaper_binding *binding, 736 const struct net_shaper *shaper, 737 struct netlink_ext_ack *extack) 738 { 739 struct mana_port_context *apc = netdev_priv(binding->netdev); 740 u32 old_speed, rate; 741 int err; 742 743 if (shaper->handle.scope != NET_SHAPER_SCOPE_NETDEV) { 744 NL_SET_ERR_MSG_MOD(extack, "net shaper scope should be netdev"); 745 return -EINVAL; 746 } 747 748 if (apc->handle.id && shaper->handle.id != apc->handle.id) { 749 NL_SET_ERR_MSG_MOD(extack, "Cannot create multiple shapers"); 750 return -EOPNOTSUPP; 751 } 752 753 if (!shaper->bw_max || (shaper->bw_max % 100000000)) { 754 NL_SET_ERR_MSG_MOD(extack, "Please use multiples of 100Mbps for bandwidth"); 755 return -EINVAL; 756 } 757 758 rate = div_u64(shaper->bw_max, 1000); /* Convert bps to Kbps */ 759 rate = div_u64(rate, 1000); /* Convert Kbps to Mbps */ 760 761 /* Get current speed */ 762 err = mana_query_link_cfg(apc); 763 old_speed = (err) ? SPEED_UNKNOWN : apc->speed; 764 765 if (!err) { 766 err = mana_set_bw_clamp(apc, rate, TRI_STATE_TRUE); 767 apc->speed = (err) ? old_speed : rate; 768 apc->handle = (err) ? apc->handle : shaper->handle; 769 } 770 771 return err; 772 } 773 774 static int mana_shaper_del(struct net_shaper_binding *binding, 775 const struct net_shaper_handle *handle, 776 struct netlink_ext_ack *extack) 777 { 778 struct mana_port_context *apc = netdev_priv(binding->netdev); 779 int err; 780 781 err = mana_set_bw_clamp(apc, 0, TRI_STATE_FALSE); 782 783 if (!err) { 784 /* Reset mana port context parameters */ 785 apc->handle.id = 0; 786 apc->handle.scope = NET_SHAPER_SCOPE_UNSPEC; 787 apc->speed = 0; 788 } 789 790 return err; 791 } 792 793 static void mana_shaper_cap(struct net_shaper_binding *binding, 794 enum net_shaper_scope scope, 795 unsigned long *flags) 796 { 797 *flags = BIT(NET_SHAPER_A_CAPS_SUPPORT_BW_MAX) | 798 BIT(NET_SHAPER_A_CAPS_SUPPORT_METRIC_BPS); 799 } 800 801 static const struct net_shaper_ops mana_shaper_ops = { 802 .set = mana_shaper_set, 803 .delete = mana_shaper_del, 804 .capabilities = mana_shaper_cap, 805 }; 806 807 static const struct net_device_ops mana_devops = { 808 .ndo_open = mana_open, 809 .ndo_stop = mana_close, 810 .ndo_select_queue = mana_select_queue, 811 .ndo_start_xmit = mana_start_xmit, 812 .ndo_validate_addr = eth_validate_addr, 813 .ndo_get_stats64 = mana_get_stats64, 814 .ndo_bpf = mana_bpf, 815 .ndo_xdp_xmit = mana_xdp_xmit, 816 .ndo_change_mtu = mana_change_mtu, 817 .net_shaper_ops = &mana_shaper_ops, 818 }; 819 820 static void mana_cleanup_port_context(struct mana_port_context *apc) 821 { 822 /* 823 * make sure subsequent cleanup attempts don't end up removing already 824 * cleaned dentry pointer 825 */ 826 debugfs_remove(apc->mana_port_debugfs); 827 apc->mana_port_debugfs = NULL; 828 kfree(apc->rxqs); 829 apc->rxqs = NULL; 830 } 831 832 static void mana_cleanup_indir_table(struct mana_port_context *apc) 833 { 834 apc->indir_table_sz = 0; 835 kfree(apc->indir_table); 836 kfree(apc->rxobj_table); 837 } 838 839 static int mana_init_port_context(struct mana_port_context *apc) 840 { 841 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *), 842 GFP_KERNEL); 843 844 return !apc->rxqs ? -ENOMEM : 0; 845 } 846 847 static int mana_send_request(struct mana_context *ac, void *in_buf, 848 u32 in_len, void *out_buf, u32 out_len) 849 { 850 struct gdma_context *gc = ac->gdma_dev->gdma_context; 851 struct gdma_resp_hdr *resp = out_buf; 852 struct gdma_req_hdr *req = in_buf; 853 struct device *dev = gc->dev; 854 static atomic_t activity_id; 855 int err; 856 857 req->dev_id = gc->mana.dev_id; 858 req->activity_id = atomic_inc_return(&activity_id); 859 860 err = mana_gd_send_request(gc, in_len, in_buf, out_len, 861 out_buf); 862 if (err || resp->status) { 863 if (err == -EOPNOTSUPP) 864 return err; 865 866 if (req->req.msg_type != MANA_QUERY_PHY_STAT && 867 mana_need_log(gc, err)) 868 dev_err(dev, "Failed to send mana message: %d, 0x%x\n", 869 err, resp->status); 870 return err ? err : -EPROTO; 871 } 872 873 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 || 874 req->activity_id != resp->activity_id) { 875 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n", 876 req->dev_id.as_uint32, resp->dev_id.as_uint32, 877 req->activity_id, resp->activity_id); 878 return -EPROTO; 879 } 880 881 return 0; 882 } 883 884 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr, 885 const enum mana_command_code expected_code, 886 const u32 min_size) 887 { 888 if (resp_hdr->response.msg_type != expected_code) 889 return -EPROTO; 890 891 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1) 892 return -EPROTO; 893 894 if (resp_hdr->response.msg_size < min_size) 895 return -EPROTO; 896 897 return 0; 898 } 899 900 static int mana_pf_register_hw_vport(struct mana_port_context *apc) 901 { 902 struct mana_register_hw_vport_resp resp = {}; 903 struct mana_register_hw_vport_req req = {}; 904 int err; 905 906 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT, 907 sizeof(req), sizeof(resp)); 908 req.attached_gfid = 1; 909 req.is_pf_default_vport = 1; 910 req.allow_all_ether_types = 1; 911 912 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 913 sizeof(resp)); 914 if (err) { 915 netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err); 916 return err; 917 } 918 919 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT, 920 sizeof(resp)); 921 if (err || resp.hdr.status) { 922 netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n", 923 err, resp.hdr.status); 924 return err ? err : -EPROTO; 925 } 926 927 apc->port_handle = resp.hw_vport_handle; 928 return 0; 929 } 930 931 static void mana_pf_deregister_hw_vport(struct mana_port_context *apc) 932 { 933 struct mana_deregister_hw_vport_resp resp = {}; 934 struct mana_deregister_hw_vport_req req = {}; 935 int err; 936 937 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT, 938 sizeof(req), sizeof(resp)); 939 req.hw_vport_handle = apc->port_handle; 940 941 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 942 sizeof(resp)); 943 if (err) { 944 if (mana_en_need_log(apc, err)) 945 netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n", 946 err); 947 948 return; 949 } 950 951 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT, 952 sizeof(resp)); 953 if (err || resp.hdr.status) 954 netdev_err(apc->ndev, 955 "Failed to deregister hw vPort: %d, 0x%x\n", 956 err, resp.hdr.status); 957 } 958 959 static int mana_pf_register_filter(struct mana_port_context *apc) 960 { 961 struct mana_register_filter_resp resp = {}; 962 struct mana_register_filter_req req = {}; 963 int err; 964 965 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER, 966 sizeof(req), sizeof(resp)); 967 req.vport = apc->port_handle; 968 memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN); 969 970 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 971 sizeof(resp)); 972 if (err) { 973 netdev_err(apc->ndev, "Failed to register filter: %d\n", err); 974 return err; 975 } 976 977 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER, 978 sizeof(resp)); 979 if (err || resp.hdr.status) { 980 netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n", 981 err, resp.hdr.status); 982 return err ? err : -EPROTO; 983 } 984 985 apc->pf_filter_handle = resp.filter_handle; 986 return 0; 987 } 988 989 static void mana_pf_deregister_filter(struct mana_port_context *apc) 990 { 991 struct mana_deregister_filter_resp resp = {}; 992 struct mana_deregister_filter_req req = {}; 993 int err; 994 995 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER, 996 sizeof(req), sizeof(resp)); 997 req.filter_handle = apc->pf_filter_handle; 998 999 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1000 sizeof(resp)); 1001 if (err) { 1002 if (mana_en_need_log(apc, err)) 1003 netdev_err(apc->ndev, "Failed to unregister filter: %d\n", 1004 err); 1005 1006 return; 1007 } 1008 1009 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER, 1010 sizeof(resp)); 1011 if (err || resp.hdr.status) 1012 netdev_err(apc->ndev, 1013 "Failed to deregister filter: %d, 0x%x\n", 1014 err, resp.hdr.status); 1015 } 1016 1017 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver, 1018 u32 proto_minor_ver, u32 proto_micro_ver, 1019 u16 *max_num_vports, u8 *bm_hostmode) 1020 { 1021 struct gdma_context *gc = ac->gdma_dev->gdma_context; 1022 struct mana_query_device_cfg_resp resp = {}; 1023 struct mana_query_device_cfg_req req = {}; 1024 struct device *dev = gc->dev; 1025 int err = 0; 1026 1027 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG, 1028 sizeof(req), sizeof(resp)); 1029 1030 req.hdr.resp.msg_version = GDMA_MESSAGE_V3; 1031 1032 req.proto_major_ver = proto_major_ver; 1033 req.proto_minor_ver = proto_minor_ver; 1034 req.proto_micro_ver = proto_micro_ver; 1035 1036 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp)); 1037 if (err) { 1038 dev_err(dev, "Failed to query config: %d", err); 1039 return err; 1040 } 1041 1042 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG, 1043 sizeof(resp)); 1044 if (err || resp.hdr.status) { 1045 dev_err(dev, "Invalid query result: %d, 0x%x\n", err, 1046 resp.hdr.status); 1047 if (!err) 1048 err = -EPROTO; 1049 return err; 1050 } 1051 1052 *max_num_vports = resp.max_num_vports; 1053 1054 if (resp.hdr.response.msg_version >= GDMA_MESSAGE_V2) 1055 gc->adapter_mtu = resp.adapter_mtu; 1056 else 1057 gc->adapter_mtu = ETH_FRAME_LEN; 1058 1059 if (resp.hdr.response.msg_version >= GDMA_MESSAGE_V3) 1060 *bm_hostmode = resp.bm_hostmode; 1061 else 1062 *bm_hostmode = 0; 1063 1064 debugfs_create_u16("adapter-MTU", 0400, gc->mana_pci_debugfs, &gc->adapter_mtu); 1065 1066 return 0; 1067 } 1068 1069 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index, 1070 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry) 1071 { 1072 struct mana_query_vport_cfg_resp resp = {}; 1073 struct mana_query_vport_cfg_req req = {}; 1074 int err; 1075 1076 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG, 1077 sizeof(req), sizeof(resp)); 1078 1079 req.vport_index = vport_index; 1080 1081 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1082 sizeof(resp)); 1083 if (err) 1084 return err; 1085 1086 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG, 1087 sizeof(resp)); 1088 if (err) 1089 return err; 1090 1091 if (resp.hdr.status) 1092 return -EPROTO; 1093 1094 *max_sq = resp.max_num_sq; 1095 *max_rq = resp.max_num_rq; 1096 if (resp.num_indirection_ent > 0 && 1097 resp.num_indirection_ent <= MANA_INDIRECT_TABLE_MAX_SIZE && 1098 is_power_of_2(resp.num_indirection_ent)) { 1099 *num_indir_entry = resp.num_indirection_ent; 1100 } else { 1101 netdev_warn(apc->ndev, 1102 "Setting indirection table size to default %d for vPort %d\n", 1103 MANA_INDIRECT_TABLE_DEF_SIZE, apc->port_idx); 1104 *num_indir_entry = MANA_INDIRECT_TABLE_DEF_SIZE; 1105 } 1106 1107 apc->port_handle = resp.vport; 1108 ether_addr_copy(apc->mac_addr, resp.mac_addr); 1109 1110 return 0; 1111 } 1112 1113 void mana_uncfg_vport(struct mana_port_context *apc) 1114 { 1115 mutex_lock(&apc->vport_mutex); 1116 apc->vport_use_count--; 1117 WARN_ON(apc->vport_use_count < 0); 1118 mutex_unlock(&apc->vport_mutex); 1119 } 1120 EXPORT_SYMBOL_NS(mana_uncfg_vport, "NET_MANA"); 1121 1122 int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id, 1123 u32 doorbell_pg_id) 1124 { 1125 struct mana_config_vport_resp resp = {}; 1126 struct mana_config_vport_req req = {}; 1127 int err; 1128 1129 /* This function is used to program the Ethernet port in the hardware 1130 * table. It can be called from the Ethernet driver or the RDMA driver. 1131 * 1132 * For Ethernet usage, the hardware supports only one active user on a 1133 * physical port. The driver checks on the port usage before programming 1134 * the hardware when creating the RAW QP (RDMA driver) or exposing the 1135 * device to kernel NET layer (Ethernet driver). 1136 * 1137 * Because the RDMA driver doesn't know in advance which QP type the 1138 * user will create, it exposes the device with all its ports. The user 1139 * may not be able to create RAW QP on a port if this port is already 1140 * in used by the Ethernet driver from the kernel. 1141 * 1142 * This physical port limitation only applies to the RAW QP. For RC QP, 1143 * the hardware doesn't have this limitation. The user can create RC 1144 * QPs on a physical port up to the hardware limits independent of the 1145 * Ethernet usage on the same port. 1146 */ 1147 mutex_lock(&apc->vport_mutex); 1148 if (apc->vport_use_count > 0) { 1149 mutex_unlock(&apc->vport_mutex); 1150 return -EBUSY; 1151 } 1152 apc->vport_use_count++; 1153 mutex_unlock(&apc->vport_mutex); 1154 1155 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX, 1156 sizeof(req), sizeof(resp)); 1157 req.vport = apc->port_handle; 1158 req.pdid = protection_dom_id; 1159 req.doorbell_pageid = doorbell_pg_id; 1160 1161 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1162 sizeof(resp)); 1163 if (err) { 1164 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err); 1165 goto out; 1166 } 1167 1168 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX, 1169 sizeof(resp)); 1170 if (err || resp.hdr.status) { 1171 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n", 1172 err, resp.hdr.status); 1173 if (!err) 1174 err = -EPROTO; 1175 1176 goto out; 1177 } 1178 1179 apc->tx_shortform_allowed = resp.short_form_allowed; 1180 apc->tx_vp_offset = resp.tx_vport_offset; 1181 1182 netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n", 1183 apc->port_handle, protection_dom_id, doorbell_pg_id); 1184 out: 1185 if (err) 1186 mana_uncfg_vport(apc); 1187 1188 return err; 1189 } 1190 EXPORT_SYMBOL_NS(mana_cfg_vport, "NET_MANA"); 1191 1192 static int mana_cfg_vport_steering(struct mana_port_context *apc, 1193 enum TRI_STATE rx, 1194 bool update_default_rxobj, bool update_key, 1195 bool update_tab) 1196 { 1197 struct mana_cfg_rx_steer_req_v2 *req; 1198 struct mana_cfg_rx_steer_resp resp = {}; 1199 struct net_device *ndev = apc->ndev; 1200 u32 req_buf_size; 1201 int err; 1202 1203 req_buf_size = struct_size(req, indir_tab, apc->indir_table_sz); 1204 req = kzalloc(req_buf_size, GFP_KERNEL); 1205 if (!req) 1206 return -ENOMEM; 1207 1208 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size, 1209 sizeof(resp)); 1210 1211 req->hdr.req.msg_version = GDMA_MESSAGE_V2; 1212 1213 req->vport = apc->port_handle; 1214 req->num_indir_entries = apc->indir_table_sz; 1215 req->indir_tab_offset = offsetof(struct mana_cfg_rx_steer_req_v2, 1216 indir_tab); 1217 req->rx_enable = rx; 1218 req->rss_enable = apc->rss_state; 1219 req->update_default_rxobj = update_default_rxobj; 1220 req->update_hashkey = update_key; 1221 req->update_indir_tab = update_tab; 1222 req->default_rxobj = apc->default_rxobj; 1223 req->cqe_coalescing_enable = 0; 1224 1225 if (update_key) 1226 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE); 1227 1228 if (update_tab) 1229 memcpy(req->indir_tab, apc->rxobj_table, 1230 flex_array_size(req, indir_tab, req->num_indir_entries)); 1231 1232 err = mana_send_request(apc->ac, req, req_buf_size, &resp, 1233 sizeof(resp)); 1234 if (err) { 1235 if (mana_en_need_log(apc, err)) 1236 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err); 1237 1238 goto out; 1239 } 1240 1241 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX, 1242 sizeof(resp)); 1243 if (err) { 1244 netdev_err(ndev, "vPort RX configuration failed: %d\n", err); 1245 goto out; 1246 } 1247 1248 if (resp.hdr.status) { 1249 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n", 1250 resp.hdr.status); 1251 err = -EPROTO; 1252 } 1253 1254 netdev_info(ndev, "Configured steering vPort %llu entries %u\n", 1255 apc->port_handle, apc->indir_table_sz); 1256 out: 1257 kfree(req); 1258 return err; 1259 } 1260 1261 int mana_query_link_cfg(struct mana_port_context *apc) 1262 { 1263 struct net_device *ndev = apc->ndev; 1264 struct mana_query_link_config_resp resp = {}; 1265 struct mana_query_link_config_req req = {}; 1266 int err; 1267 1268 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_LINK_CONFIG, 1269 sizeof(req), sizeof(resp)); 1270 1271 req.vport = apc->port_handle; 1272 req.hdr.resp.msg_version = GDMA_MESSAGE_V2; 1273 1274 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1275 sizeof(resp)); 1276 1277 if (err) { 1278 if (err == -EOPNOTSUPP) { 1279 netdev_info_once(ndev, "MANA_QUERY_LINK_CONFIG not supported\n"); 1280 return err; 1281 } 1282 netdev_err(ndev, "Failed to query link config: %d\n", err); 1283 return err; 1284 } 1285 1286 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_LINK_CONFIG, 1287 sizeof(resp)); 1288 1289 if (err || resp.hdr.status) { 1290 netdev_err(ndev, "Failed to query link config: %d, 0x%x\n", err, 1291 resp.hdr.status); 1292 if (!err) 1293 err = -EOPNOTSUPP; 1294 return err; 1295 } 1296 1297 if (resp.qos_unconfigured) { 1298 err = -EINVAL; 1299 return err; 1300 } 1301 apc->speed = resp.link_speed_mbps; 1302 apc->max_speed = resp.qos_speed_mbps; 1303 return 0; 1304 } 1305 1306 int mana_set_bw_clamp(struct mana_port_context *apc, u32 speed, 1307 int enable_clamping) 1308 { 1309 struct mana_set_bw_clamp_resp resp = {}; 1310 struct mana_set_bw_clamp_req req = {}; 1311 struct net_device *ndev = apc->ndev; 1312 int err; 1313 1314 mana_gd_init_req_hdr(&req.hdr, MANA_SET_BW_CLAMP, 1315 sizeof(req), sizeof(resp)); 1316 req.vport = apc->port_handle; 1317 req.link_speed_mbps = speed; 1318 req.enable_clamping = enable_clamping; 1319 1320 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1321 sizeof(resp)); 1322 1323 if (err) { 1324 if (err == -EOPNOTSUPP) { 1325 netdev_info_once(ndev, "MANA_SET_BW_CLAMP not supported\n"); 1326 return err; 1327 } 1328 netdev_err(ndev, "Failed to set bandwidth clamp for speed %u, err = %d", 1329 speed, err); 1330 return err; 1331 } 1332 1333 err = mana_verify_resp_hdr(&resp.hdr, MANA_SET_BW_CLAMP, 1334 sizeof(resp)); 1335 1336 if (err || resp.hdr.status) { 1337 netdev_err(ndev, "Failed to set bandwidth clamp: %d, 0x%x\n", err, 1338 resp.hdr.status); 1339 if (!err) 1340 err = -EOPNOTSUPP; 1341 return err; 1342 } 1343 1344 if (resp.qos_unconfigured) 1345 netdev_info(ndev, "QoS is unconfigured\n"); 1346 1347 return 0; 1348 } 1349 1350 int mana_create_wq_obj(struct mana_port_context *apc, 1351 mana_handle_t vport, 1352 u32 wq_type, struct mana_obj_spec *wq_spec, 1353 struct mana_obj_spec *cq_spec, 1354 mana_handle_t *wq_obj) 1355 { 1356 struct mana_create_wqobj_resp resp = {}; 1357 struct mana_create_wqobj_req req = {}; 1358 struct net_device *ndev = apc->ndev; 1359 int err; 1360 1361 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ, 1362 sizeof(req), sizeof(resp)); 1363 req.vport = vport; 1364 req.wq_type = wq_type; 1365 req.wq_gdma_region = wq_spec->gdma_region; 1366 req.cq_gdma_region = cq_spec->gdma_region; 1367 req.wq_size = wq_spec->queue_size; 1368 req.cq_size = cq_spec->queue_size; 1369 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id; 1370 req.cq_parent_qid = cq_spec->attached_eq; 1371 1372 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1373 sizeof(resp)); 1374 if (err) { 1375 netdev_err(ndev, "Failed to create WQ object: %d\n", err); 1376 goto out; 1377 } 1378 1379 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ, 1380 sizeof(resp)); 1381 if (err || resp.hdr.status) { 1382 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err, 1383 resp.hdr.status); 1384 if (!err) 1385 err = -EPROTO; 1386 goto out; 1387 } 1388 1389 if (resp.wq_obj == INVALID_MANA_HANDLE) { 1390 netdev_err(ndev, "Got an invalid WQ object handle\n"); 1391 err = -EPROTO; 1392 goto out; 1393 } 1394 1395 *wq_obj = resp.wq_obj; 1396 wq_spec->queue_index = resp.wq_id; 1397 cq_spec->queue_index = resp.cq_id; 1398 1399 return 0; 1400 out: 1401 return err; 1402 } 1403 EXPORT_SYMBOL_NS(mana_create_wq_obj, "NET_MANA"); 1404 1405 void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type, 1406 mana_handle_t wq_obj) 1407 { 1408 struct mana_destroy_wqobj_resp resp = {}; 1409 struct mana_destroy_wqobj_req req = {}; 1410 struct net_device *ndev = apc->ndev; 1411 int err; 1412 1413 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ, 1414 sizeof(req), sizeof(resp)); 1415 req.wq_type = wq_type; 1416 req.wq_obj_handle = wq_obj; 1417 1418 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1419 sizeof(resp)); 1420 if (err) { 1421 if (mana_en_need_log(apc, err)) 1422 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err); 1423 1424 return; 1425 } 1426 1427 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ, 1428 sizeof(resp)); 1429 if (err || resp.hdr.status) 1430 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err, 1431 resp.hdr.status); 1432 } 1433 EXPORT_SYMBOL_NS(mana_destroy_wq_obj, "NET_MANA"); 1434 1435 static void mana_destroy_eq(struct mana_context *ac) 1436 { 1437 struct gdma_context *gc = ac->gdma_dev->gdma_context; 1438 struct gdma_queue *eq; 1439 int i; 1440 1441 if (!ac->eqs) 1442 return; 1443 1444 debugfs_remove_recursive(ac->mana_eqs_debugfs); 1445 ac->mana_eqs_debugfs = NULL; 1446 1447 for (i = 0; i < gc->max_num_queues; i++) { 1448 eq = ac->eqs[i].eq; 1449 if (!eq) 1450 continue; 1451 1452 mana_gd_destroy_queue(gc, eq); 1453 } 1454 1455 kfree(ac->eqs); 1456 ac->eqs = NULL; 1457 } 1458 1459 static void mana_create_eq_debugfs(struct mana_context *ac, int i) 1460 { 1461 struct mana_eq eq = ac->eqs[i]; 1462 char eqnum[32]; 1463 1464 sprintf(eqnum, "eq%d", i); 1465 eq.mana_eq_debugfs = debugfs_create_dir(eqnum, ac->mana_eqs_debugfs); 1466 debugfs_create_u32("head", 0400, eq.mana_eq_debugfs, &eq.eq->head); 1467 debugfs_create_u32("tail", 0400, eq.mana_eq_debugfs, &eq.eq->tail); 1468 debugfs_create_file("eq_dump", 0400, eq.mana_eq_debugfs, eq.eq, &mana_dbg_q_fops); 1469 } 1470 1471 static int mana_create_eq(struct mana_context *ac) 1472 { 1473 struct gdma_dev *gd = ac->gdma_dev; 1474 struct gdma_context *gc = gd->gdma_context; 1475 struct gdma_queue_spec spec = {}; 1476 int err; 1477 int i; 1478 1479 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq), 1480 GFP_KERNEL); 1481 if (!ac->eqs) 1482 return -ENOMEM; 1483 1484 spec.type = GDMA_EQ; 1485 spec.monitor_avl_buf = false; 1486 spec.queue_size = EQ_SIZE; 1487 spec.eq.callback = NULL; 1488 spec.eq.context = ac->eqs; 1489 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE; 1490 1491 ac->mana_eqs_debugfs = debugfs_create_dir("EQs", gc->mana_pci_debugfs); 1492 1493 for (i = 0; i < gc->max_num_queues; i++) { 1494 spec.eq.msix_index = (i + 1) % gc->num_msix_usable; 1495 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq); 1496 if (err) { 1497 dev_err(gc->dev, "Failed to create EQ %d : %d\n", i, err); 1498 goto out; 1499 } 1500 mana_create_eq_debugfs(ac, i); 1501 } 1502 1503 return 0; 1504 out: 1505 mana_destroy_eq(ac); 1506 return err; 1507 } 1508 1509 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq) 1510 { 1511 struct mana_fence_rq_resp resp = {}; 1512 struct mana_fence_rq_req req = {}; 1513 int err; 1514 1515 init_completion(&rxq->fence_event); 1516 1517 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ, 1518 sizeof(req), sizeof(resp)); 1519 req.wq_obj_handle = rxq->rxobj; 1520 1521 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1522 sizeof(resp)); 1523 if (err) { 1524 netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n", 1525 rxq->rxq_idx, err); 1526 return err; 1527 } 1528 1529 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp)); 1530 if (err || resp.hdr.status) { 1531 netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n", 1532 rxq->rxq_idx, err, resp.hdr.status); 1533 if (!err) 1534 err = -EPROTO; 1535 1536 return err; 1537 } 1538 1539 if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) { 1540 netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n", 1541 rxq->rxq_idx); 1542 return -ETIMEDOUT; 1543 } 1544 1545 return 0; 1546 } 1547 1548 static void mana_fence_rqs(struct mana_port_context *apc) 1549 { 1550 unsigned int rxq_idx; 1551 struct mana_rxq *rxq; 1552 int err; 1553 1554 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 1555 rxq = apc->rxqs[rxq_idx]; 1556 err = mana_fence_rq(apc, rxq); 1557 1558 /* In case of any error, use sleep instead. */ 1559 if (err) 1560 msleep(100); 1561 } 1562 } 1563 1564 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units) 1565 { 1566 u32 used_space_old; 1567 u32 used_space_new; 1568 1569 used_space_old = wq->head - wq->tail; 1570 used_space_new = wq->head - (wq->tail + num_units); 1571 1572 if (WARN_ON_ONCE(used_space_new > used_space_old)) 1573 return -ERANGE; 1574 1575 wq->tail += num_units; 1576 return 0; 1577 } 1578 1579 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc) 1580 { 1581 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 1582 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 1583 struct device *dev = gc->dev; 1584 int hsg, i; 1585 1586 /* Number of SGEs of linear part */ 1587 hsg = (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0]) ? 2 : 1; 1588 1589 for (i = 0; i < hsg; i++) 1590 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i], 1591 DMA_TO_DEVICE); 1592 1593 for (i = hsg; i < skb_shinfo(skb)->nr_frags + hsg; i++) 1594 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i], 1595 DMA_TO_DEVICE); 1596 } 1597 1598 static void mana_poll_tx_cq(struct mana_cq *cq) 1599 { 1600 struct gdma_comp *completions = cq->gdma_comp_buf; 1601 struct gdma_posted_wqe_info *wqe_info; 1602 unsigned int pkt_transmitted = 0; 1603 unsigned int wqe_unit_cnt = 0; 1604 struct mana_txq *txq = cq->txq; 1605 struct mana_port_context *apc; 1606 struct netdev_queue *net_txq; 1607 struct gdma_queue *gdma_wq; 1608 unsigned int avail_space; 1609 struct net_device *ndev; 1610 struct sk_buff *skb; 1611 bool txq_stopped; 1612 int comp_read; 1613 int i; 1614 1615 ndev = txq->ndev; 1616 apc = netdev_priv(ndev); 1617 1618 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions, 1619 CQE_POLLING_BUFFER); 1620 1621 if (comp_read < 1) 1622 return; 1623 1624 for (i = 0; i < comp_read; i++) { 1625 struct mana_tx_comp_oob *cqe_oob; 1626 1627 if (WARN_ON_ONCE(!completions[i].is_sq)) 1628 return; 1629 1630 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data; 1631 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type != 1632 MANA_CQE_COMPLETION)) 1633 return; 1634 1635 switch (cqe_oob->cqe_hdr.cqe_type) { 1636 case CQE_TX_OKAY: 1637 break; 1638 1639 case CQE_TX_SA_DROP: 1640 case CQE_TX_MTU_DROP: 1641 case CQE_TX_INVALID_OOB: 1642 case CQE_TX_INVALID_ETH_TYPE: 1643 case CQE_TX_HDR_PROCESSING_ERROR: 1644 case CQE_TX_VF_DISABLED: 1645 case CQE_TX_VPORT_IDX_OUT_OF_RANGE: 1646 case CQE_TX_VPORT_DISABLED: 1647 case CQE_TX_VLAN_TAGGING_VIOLATION: 1648 if (net_ratelimit()) 1649 netdev_err(ndev, "TX: CQE error %d\n", 1650 cqe_oob->cqe_hdr.cqe_type); 1651 1652 apc->eth_stats.tx_cqe_err++; 1653 break; 1654 1655 default: 1656 /* If the CQE type is unknown, log an error, 1657 * and still free the SKB, update tail, etc. 1658 */ 1659 if (net_ratelimit()) 1660 netdev_err(ndev, "TX: unknown CQE type %d\n", 1661 cqe_oob->cqe_hdr.cqe_type); 1662 1663 apc->eth_stats.tx_cqe_unknown_type++; 1664 break; 1665 } 1666 1667 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num)) 1668 return; 1669 1670 skb = skb_dequeue(&txq->pending_skbs); 1671 if (WARN_ON_ONCE(!skb)) 1672 return; 1673 1674 wqe_info = (struct gdma_posted_wqe_info *)skb->cb; 1675 wqe_unit_cnt += wqe_info->wqe_size_in_bu; 1676 1677 mana_unmap_skb(skb, apc); 1678 1679 napi_consume_skb(skb, cq->budget); 1680 1681 pkt_transmitted++; 1682 } 1683 1684 if (WARN_ON_ONCE(wqe_unit_cnt == 0)) 1685 return; 1686 1687 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt); 1688 1689 gdma_wq = txq->gdma_sq; 1690 avail_space = mana_gd_wq_avail_space(gdma_wq); 1691 1692 /* Ensure tail updated before checking q stop */ 1693 smp_mb(); 1694 1695 net_txq = txq->net_txq; 1696 txq_stopped = netif_tx_queue_stopped(net_txq); 1697 1698 /* Ensure checking txq_stopped before apc->port_is_up. */ 1699 smp_rmb(); 1700 1701 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) { 1702 netif_tx_wake_queue(net_txq); 1703 apc->eth_stats.wake_queue++; 1704 } 1705 1706 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0) 1707 WARN_ON_ONCE(1); 1708 1709 cq->work_done = pkt_transmitted; 1710 } 1711 1712 static void mana_post_pkt_rxq(struct mana_rxq *rxq) 1713 { 1714 struct mana_recv_buf_oob *recv_buf_oob; 1715 u32 curr_index; 1716 int err; 1717 1718 curr_index = rxq->buf_index++; 1719 if (rxq->buf_index == rxq->num_rx_buf) 1720 rxq->buf_index = 0; 1721 1722 recv_buf_oob = &rxq->rx_oobs[curr_index]; 1723 1724 err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req, 1725 &recv_buf_oob->wqe_inf); 1726 if (WARN_ON_ONCE(err)) 1727 return; 1728 1729 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1); 1730 } 1731 1732 static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va, 1733 uint pkt_len, struct xdp_buff *xdp) 1734 { 1735 struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size); 1736 1737 if (!skb) 1738 return NULL; 1739 1740 if (xdp->data_hard_start) { 1741 u32 metasize = xdp->data - xdp->data_meta; 1742 1743 skb_reserve(skb, xdp->data - xdp->data_hard_start); 1744 skb_put(skb, xdp->data_end - xdp->data); 1745 if (metasize) 1746 skb_metadata_set(skb, metasize); 1747 return skb; 1748 } 1749 1750 skb_reserve(skb, rxq->headroom); 1751 skb_put(skb, pkt_len); 1752 1753 return skb; 1754 } 1755 1756 static void mana_rx_skb(void *buf_va, bool from_pool, 1757 struct mana_rxcomp_oob *cqe, struct mana_rxq *rxq) 1758 { 1759 struct mana_stats_rx *rx_stats = &rxq->stats; 1760 struct net_device *ndev = rxq->ndev; 1761 uint pkt_len = cqe->ppi[0].pkt_len; 1762 u16 rxq_idx = rxq->rxq_idx; 1763 struct napi_struct *napi; 1764 struct xdp_buff xdp = {}; 1765 struct sk_buff *skb; 1766 u32 hash_value; 1767 u32 act; 1768 1769 rxq->rx_cq.work_done++; 1770 napi = &rxq->rx_cq.napi; 1771 1772 if (!buf_va) { 1773 ++ndev->stats.rx_dropped; 1774 return; 1775 } 1776 1777 act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len); 1778 1779 if (act == XDP_REDIRECT && !rxq->xdp_rc) 1780 return; 1781 1782 if (act != XDP_PASS && act != XDP_TX) 1783 goto drop_xdp; 1784 1785 skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp); 1786 1787 if (!skb) 1788 goto drop; 1789 1790 if (from_pool) 1791 skb_mark_for_recycle(skb); 1792 1793 skb->dev = napi->dev; 1794 1795 skb->protocol = eth_type_trans(skb, ndev); 1796 skb_checksum_none_assert(skb); 1797 skb_record_rx_queue(skb, rxq_idx); 1798 1799 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) { 1800 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed) 1801 skb->ip_summed = CHECKSUM_UNNECESSARY; 1802 } 1803 1804 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) { 1805 hash_value = cqe->ppi[0].pkt_hash; 1806 1807 if (cqe->rx_hashtype & MANA_HASH_L4) 1808 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4); 1809 else 1810 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3); 1811 } 1812 1813 if (cqe->rx_vlantag_present) { 1814 u16 vlan_tci = cqe->rx_vlan_id; 1815 1816 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci); 1817 } 1818 1819 u64_stats_update_begin(&rx_stats->syncp); 1820 rx_stats->packets++; 1821 rx_stats->bytes += pkt_len; 1822 1823 if (act == XDP_TX) 1824 rx_stats->xdp_tx++; 1825 u64_stats_update_end(&rx_stats->syncp); 1826 1827 if (act == XDP_TX) { 1828 skb_set_queue_mapping(skb, rxq_idx); 1829 mana_xdp_tx(skb, ndev); 1830 return; 1831 } 1832 1833 napi_gro_receive(napi, skb); 1834 1835 return; 1836 1837 drop_xdp: 1838 u64_stats_update_begin(&rx_stats->syncp); 1839 rx_stats->xdp_drop++; 1840 u64_stats_update_end(&rx_stats->syncp); 1841 1842 drop: 1843 if (from_pool) { 1844 page_pool_recycle_direct(rxq->page_pool, 1845 virt_to_head_page(buf_va)); 1846 } else { 1847 WARN_ON_ONCE(rxq->xdp_save_va); 1848 /* Save for reuse */ 1849 rxq->xdp_save_va = buf_va; 1850 } 1851 1852 ++ndev->stats.rx_dropped; 1853 1854 return; 1855 } 1856 1857 static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev, 1858 dma_addr_t *da, bool *from_pool) 1859 { 1860 struct page *page; 1861 void *va; 1862 1863 *from_pool = false; 1864 1865 /* Reuse XDP dropped page if available */ 1866 if (rxq->xdp_save_va) { 1867 va = rxq->xdp_save_va; 1868 rxq->xdp_save_va = NULL; 1869 } else { 1870 page = page_pool_dev_alloc_pages(rxq->page_pool); 1871 if (!page) 1872 return NULL; 1873 1874 *from_pool = true; 1875 va = page_to_virt(page); 1876 } 1877 1878 *da = dma_map_single(dev, va + rxq->headroom, rxq->datasize, 1879 DMA_FROM_DEVICE); 1880 if (dma_mapping_error(dev, *da)) { 1881 if (*from_pool) 1882 page_pool_put_full_page(rxq->page_pool, page, false); 1883 else 1884 put_page(virt_to_head_page(va)); 1885 1886 return NULL; 1887 } 1888 1889 return va; 1890 } 1891 1892 /* Allocate frag for rx buffer, and save the old buf */ 1893 static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq, 1894 struct mana_recv_buf_oob *rxoob, void **old_buf, 1895 bool *old_fp) 1896 { 1897 bool from_pool; 1898 dma_addr_t da; 1899 void *va; 1900 1901 va = mana_get_rxfrag(rxq, dev, &da, &from_pool); 1902 if (!va) 1903 return; 1904 1905 dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize, 1906 DMA_FROM_DEVICE); 1907 *old_buf = rxoob->buf_va; 1908 *old_fp = rxoob->from_pool; 1909 1910 rxoob->buf_va = va; 1911 rxoob->sgl[0].address = da; 1912 rxoob->from_pool = from_pool; 1913 } 1914 1915 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq, 1916 struct gdma_comp *cqe) 1917 { 1918 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data; 1919 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context; 1920 struct net_device *ndev = rxq->ndev; 1921 struct mana_recv_buf_oob *rxbuf_oob; 1922 struct mana_port_context *apc; 1923 struct device *dev = gc->dev; 1924 void *old_buf = NULL; 1925 u32 curr, pktlen; 1926 bool old_fp; 1927 1928 apc = netdev_priv(ndev); 1929 1930 switch (oob->cqe_hdr.cqe_type) { 1931 case CQE_RX_OKAY: 1932 break; 1933 1934 case CQE_RX_TRUNCATED: 1935 ++ndev->stats.rx_dropped; 1936 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index]; 1937 netdev_warn_once(ndev, "Dropped a truncated packet\n"); 1938 goto drop; 1939 1940 case CQE_RX_COALESCED_4: 1941 netdev_err(ndev, "RX coalescing is unsupported\n"); 1942 apc->eth_stats.rx_coalesced_err++; 1943 return; 1944 1945 case CQE_RX_OBJECT_FENCE: 1946 complete(&rxq->fence_event); 1947 return; 1948 1949 default: 1950 netdev_err(ndev, "Unknown RX CQE type = %d\n", 1951 oob->cqe_hdr.cqe_type); 1952 apc->eth_stats.rx_cqe_unknown_type++; 1953 return; 1954 } 1955 1956 pktlen = oob->ppi[0].pkt_len; 1957 1958 if (pktlen == 0) { 1959 /* data packets should never have packetlength of zero */ 1960 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n", 1961 rxq->gdma_id, cq->gdma_id, rxq->rxobj); 1962 return; 1963 } 1964 1965 curr = rxq->buf_index; 1966 rxbuf_oob = &rxq->rx_oobs[curr]; 1967 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1); 1968 1969 mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf, &old_fp); 1970 1971 /* Unsuccessful refill will have old_buf == NULL. 1972 * In this case, mana_rx_skb() will drop the packet. 1973 */ 1974 mana_rx_skb(old_buf, old_fp, oob, rxq); 1975 1976 drop: 1977 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu); 1978 1979 mana_post_pkt_rxq(rxq); 1980 } 1981 1982 static void mana_poll_rx_cq(struct mana_cq *cq) 1983 { 1984 struct gdma_comp *comp = cq->gdma_comp_buf; 1985 struct mana_rxq *rxq = cq->rxq; 1986 int comp_read, i; 1987 1988 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER); 1989 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER); 1990 1991 rxq->xdp_flush = false; 1992 1993 for (i = 0; i < comp_read; i++) { 1994 if (WARN_ON_ONCE(comp[i].is_sq)) 1995 return; 1996 1997 /* verify recv cqe references the right rxq */ 1998 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id)) 1999 return; 2000 2001 mana_process_rx_cqe(rxq, cq, &comp[i]); 2002 } 2003 2004 if (comp_read > 0) { 2005 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context; 2006 2007 mana_gd_wq_ring_doorbell(gc, rxq->gdma_rq); 2008 } 2009 2010 if (rxq->xdp_flush) 2011 xdp_do_flush(); 2012 } 2013 2014 static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue) 2015 { 2016 struct mana_cq *cq = context; 2017 int w; 2018 2019 WARN_ON_ONCE(cq->gdma_cq != gdma_queue); 2020 2021 if (cq->type == MANA_CQ_TYPE_RX) 2022 mana_poll_rx_cq(cq); 2023 else 2024 mana_poll_tx_cq(cq); 2025 2026 w = cq->work_done; 2027 cq->work_done_since_doorbell += w; 2028 2029 if (w < cq->budget) { 2030 mana_gd_ring_cq(gdma_queue, SET_ARM_BIT); 2031 cq->work_done_since_doorbell = 0; 2032 napi_complete_done(&cq->napi, w); 2033 } else if (cq->work_done_since_doorbell > 2034 cq->gdma_cq->queue_size / COMP_ENTRY_SIZE * 4) { 2035 /* MANA hardware requires at least one doorbell ring every 8 2036 * wraparounds of CQ even if there is no need to arm the CQ. 2037 * This driver rings the doorbell as soon as we have exceeded 2038 * 4 wraparounds. 2039 */ 2040 mana_gd_ring_cq(gdma_queue, 0); 2041 cq->work_done_since_doorbell = 0; 2042 } 2043 2044 return w; 2045 } 2046 2047 static int mana_poll(struct napi_struct *napi, int budget) 2048 { 2049 struct mana_cq *cq = container_of(napi, struct mana_cq, napi); 2050 int w; 2051 2052 cq->work_done = 0; 2053 cq->budget = budget; 2054 2055 w = mana_cq_handler(cq, cq->gdma_cq); 2056 2057 return min(w, budget); 2058 } 2059 2060 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue) 2061 { 2062 struct mana_cq *cq = context; 2063 2064 napi_schedule_irqoff(&cq->napi); 2065 } 2066 2067 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq) 2068 { 2069 struct gdma_dev *gd = apc->ac->gdma_dev; 2070 2071 if (!cq->gdma_cq) 2072 return; 2073 2074 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq); 2075 } 2076 2077 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq) 2078 { 2079 struct gdma_dev *gd = apc->ac->gdma_dev; 2080 2081 if (!txq->gdma_sq) 2082 return; 2083 2084 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq); 2085 } 2086 2087 static void mana_destroy_txq(struct mana_port_context *apc) 2088 { 2089 struct napi_struct *napi; 2090 int i; 2091 2092 if (!apc->tx_qp) 2093 return; 2094 2095 for (i = 0; i < apc->num_queues; i++) { 2096 debugfs_remove_recursive(apc->tx_qp[i].mana_tx_debugfs); 2097 apc->tx_qp[i].mana_tx_debugfs = NULL; 2098 2099 napi = &apc->tx_qp[i].tx_cq.napi; 2100 if (apc->tx_qp[i].txq.napi_initialized) { 2101 napi_synchronize(napi); 2102 netdev_lock_ops_to_full(napi->dev); 2103 napi_disable_locked(napi); 2104 netif_napi_del_locked(napi); 2105 netdev_unlock_full_to_ops(napi->dev); 2106 apc->tx_qp[i].txq.napi_initialized = false; 2107 } 2108 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object); 2109 2110 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq); 2111 2112 mana_deinit_txq(apc, &apc->tx_qp[i].txq); 2113 } 2114 2115 kfree(apc->tx_qp); 2116 apc->tx_qp = NULL; 2117 } 2118 2119 static void mana_create_txq_debugfs(struct mana_port_context *apc, int idx) 2120 { 2121 struct mana_tx_qp *tx_qp = &apc->tx_qp[idx]; 2122 char qnum[32]; 2123 2124 sprintf(qnum, "TX-%d", idx); 2125 tx_qp->mana_tx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs); 2126 debugfs_create_u32("sq_head", 0400, tx_qp->mana_tx_debugfs, 2127 &tx_qp->txq.gdma_sq->head); 2128 debugfs_create_u32("sq_tail", 0400, tx_qp->mana_tx_debugfs, 2129 &tx_qp->txq.gdma_sq->tail); 2130 debugfs_create_u32("sq_pend_skb_qlen", 0400, tx_qp->mana_tx_debugfs, 2131 &tx_qp->txq.pending_skbs.qlen); 2132 debugfs_create_u32("cq_head", 0400, tx_qp->mana_tx_debugfs, 2133 &tx_qp->tx_cq.gdma_cq->head); 2134 debugfs_create_u32("cq_tail", 0400, tx_qp->mana_tx_debugfs, 2135 &tx_qp->tx_cq.gdma_cq->tail); 2136 debugfs_create_u32("cq_budget", 0400, tx_qp->mana_tx_debugfs, 2137 &tx_qp->tx_cq.budget); 2138 debugfs_create_file("txq_dump", 0400, tx_qp->mana_tx_debugfs, 2139 tx_qp->txq.gdma_sq, &mana_dbg_q_fops); 2140 debugfs_create_file("cq_dump", 0400, tx_qp->mana_tx_debugfs, 2141 tx_qp->tx_cq.gdma_cq, &mana_dbg_q_fops); 2142 } 2143 2144 static int mana_create_txq(struct mana_port_context *apc, 2145 struct net_device *net) 2146 { 2147 struct mana_context *ac = apc->ac; 2148 struct gdma_dev *gd = ac->gdma_dev; 2149 struct mana_obj_spec wq_spec; 2150 struct mana_obj_spec cq_spec; 2151 struct gdma_queue_spec spec; 2152 struct gdma_context *gc; 2153 struct mana_txq *txq; 2154 struct mana_cq *cq; 2155 u32 txq_size; 2156 u32 cq_size; 2157 int err; 2158 int i; 2159 2160 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp), 2161 GFP_KERNEL); 2162 if (!apc->tx_qp) 2163 return -ENOMEM; 2164 2165 /* The minimum size of the WQE is 32 bytes, hence 2166 * apc->tx_queue_size represents the maximum number of WQEs 2167 * the SQ can store. This value is then used to size other queues 2168 * to prevent overflow. 2169 * Also note that the txq_size is always going to be MANA_PAGE_ALIGNED, 2170 * as min val of apc->tx_queue_size is 128 and that would make 2171 * txq_size 128*32 = 4096 and the other higher values of apc->tx_queue_size 2172 * are always power of two 2173 */ 2174 txq_size = apc->tx_queue_size * 32; 2175 2176 cq_size = apc->tx_queue_size * COMP_ENTRY_SIZE; 2177 2178 gc = gd->gdma_context; 2179 2180 for (i = 0; i < apc->num_queues; i++) { 2181 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE; 2182 2183 /* Create SQ */ 2184 txq = &apc->tx_qp[i].txq; 2185 2186 u64_stats_init(&txq->stats.syncp); 2187 txq->ndev = net; 2188 txq->net_txq = netdev_get_tx_queue(net, i); 2189 txq->vp_offset = apc->tx_vp_offset; 2190 txq->napi_initialized = false; 2191 skb_queue_head_init(&txq->pending_skbs); 2192 2193 memset(&spec, 0, sizeof(spec)); 2194 spec.type = GDMA_SQ; 2195 spec.monitor_avl_buf = true; 2196 spec.queue_size = txq_size; 2197 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq); 2198 if (err) 2199 goto out; 2200 2201 /* Create SQ's CQ */ 2202 cq = &apc->tx_qp[i].tx_cq; 2203 cq->type = MANA_CQ_TYPE_TX; 2204 2205 cq->txq = txq; 2206 2207 memset(&spec, 0, sizeof(spec)); 2208 spec.type = GDMA_CQ; 2209 spec.monitor_avl_buf = false; 2210 spec.queue_size = cq_size; 2211 spec.cq.callback = mana_schedule_napi; 2212 spec.cq.parent_eq = ac->eqs[i].eq; 2213 spec.cq.context = cq; 2214 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 2215 if (err) 2216 goto out; 2217 2218 memset(&wq_spec, 0, sizeof(wq_spec)); 2219 memset(&cq_spec, 0, sizeof(cq_spec)); 2220 2221 wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle; 2222 wq_spec.queue_size = txq->gdma_sq->queue_size; 2223 2224 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 2225 cq_spec.queue_size = cq->gdma_cq->queue_size; 2226 cq_spec.modr_ctx_id = 0; 2227 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 2228 2229 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ, 2230 &wq_spec, &cq_spec, 2231 &apc->tx_qp[i].tx_object); 2232 2233 if (err) 2234 goto out; 2235 2236 txq->gdma_sq->id = wq_spec.queue_index; 2237 cq->gdma_cq->id = cq_spec.queue_index; 2238 2239 txq->gdma_sq->mem_info.dma_region_handle = 2240 GDMA_INVALID_DMA_REGION; 2241 cq->gdma_cq->mem_info.dma_region_handle = 2242 GDMA_INVALID_DMA_REGION; 2243 2244 txq->gdma_txq_id = txq->gdma_sq->id; 2245 2246 cq->gdma_id = cq->gdma_cq->id; 2247 2248 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 2249 err = -EINVAL; 2250 goto out; 2251 } 2252 2253 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2254 2255 mana_create_txq_debugfs(apc, i); 2256 2257 set_bit(NAPI_STATE_NO_BUSY_POLL, &cq->napi.state); 2258 netdev_lock_ops_to_full(net); 2259 netif_napi_add_locked(net, &cq->napi, mana_poll); 2260 napi_enable_locked(&cq->napi); 2261 netdev_unlock_full_to_ops(net); 2262 txq->napi_initialized = true; 2263 2264 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2265 } 2266 2267 return 0; 2268 out: 2269 netdev_err(net, "Failed to create %d TX queues, %d\n", 2270 apc->num_queues, err); 2271 mana_destroy_txq(apc); 2272 return err; 2273 } 2274 2275 static void mana_destroy_rxq(struct mana_port_context *apc, 2276 struct mana_rxq *rxq, bool napi_initialized) 2277 2278 { 2279 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 2280 struct mana_recv_buf_oob *rx_oob; 2281 struct device *dev = gc->dev; 2282 struct napi_struct *napi; 2283 struct page *page; 2284 int i; 2285 2286 if (!rxq) 2287 return; 2288 2289 debugfs_remove_recursive(rxq->mana_rx_debugfs); 2290 rxq->mana_rx_debugfs = NULL; 2291 2292 napi = &rxq->rx_cq.napi; 2293 2294 if (napi_initialized) { 2295 napi_synchronize(napi); 2296 2297 netdev_lock_ops_to_full(napi->dev); 2298 napi_disable_locked(napi); 2299 netif_napi_del_locked(napi); 2300 netdev_unlock_full_to_ops(napi->dev); 2301 } 2302 xdp_rxq_info_unreg(&rxq->xdp_rxq); 2303 2304 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj); 2305 2306 mana_deinit_cq(apc, &rxq->rx_cq); 2307 2308 if (rxq->xdp_save_va) 2309 put_page(virt_to_head_page(rxq->xdp_save_va)); 2310 2311 for (i = 0; i < rxq->num_rx_buf; i++) { 2312 rx_oob = &rxq->rx_oobs[i]; 2313 2314 if (!rx_oob->buf_va) 2315 continue; 2316 2317 dma_unmap_single(dev, rx_oob->sgl[0].address, 2318 rx_oob->sgl[0].size, DMA_FROM_DEVICE); 2319 2320 page = virt_to_head_page(rx_oob->buf_va); 2321 2322 if (rx_oob->from_pool) 2323 page_pool_put_full_page(rxq->page_pool, page, false); 2324 else 2325 put_page(page); 2326 2327 rx_oob->buf_va = NULL; 2328 } 2329 2330 page_pool_destroy(rxq->page_pool); 2331 2332 if (rxq->gdma_rq) 2333 mana_gd_destroy_queue(gc, rxq->gdma_rq); 2334 2335 kfree(rxq); 2336 } 2337 2338 static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key, 2339 struct mana_rxq *rxq, struct device *dev) 2340 { 2341 struct mana_port_context *mpc = netdev_priv(rxq->ndev); 2342 bool from_pool = false; 2343 dma_addr_t da; 2344 void *va; 2345 2346 if (mpc->rxbufs_pre) 2347 va = mana_get_rxbuf_pre(rxq, &da); 2348 else 2349 va = mana_get_rxfrag(rxq, dev, &da, &from_pool); 2350 2351 if (!va) 2352 return -ENOMEM; 2353 2354 rx_oob->buf_va = va; 2355 rx_oob->from_pool = from_pool; 2356 2357 rx_oob->sgl[0].address = da; 2358 rx_oob->sgl[0].size = rxq->datasize; 2359 rx_oob->sgl[0].mem_key = mem_key; 2360 2361 return 0; 2362 } 2363 2364 #define MANA_WQE_HEADER_SIZE 16 2365 #define MANA_WQE_SGE_SIZE 16 2366 2367 static int mana_alloc_rx_wqe(struct mana_port_context *apc, 2368 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size) 2369 { 2370 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 2371 struct mana_recv_buf_oob *rx_oob; 2372 struct device *dev = gc->dev; 2373 u32 buf_idx; 2374 int ret; 2375 2376 WARN_ON(rxq->datasize == 0); 2377 2378 *rxq_size = 0; 2379 *cq_size = 0; 2380 2381 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2382 rx_oob = &rxq->rx_oobs[buf_idx]; 2383 memset(rx_oob, 0, sizeof(*rx_oob)); 2384 2385 rx_oob->num_sge = 1; 2386 2387 ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq, 2388 dev); 2389 if (ret) 2390 return ret; 2391 2392 rx_oob->wqe_req.sgl = rx_oob->sgl; 2393 rx_oob->wqe_req.num_sge = rx_oob->num_sge; 2394 rx_oob->wqe_req.inline_oob_size = 0; 2395 rx_oob->wqe_req.inline_oob_data = NULL; 2396 rx_oob->wqe_req.flags = 0; 2397 rx_oob->wqe_req.client_data_unit = 0; 2398 2399 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE + 2400 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32); 2401 *cq_size += COMP_ENTRY_SIZE; 2402 } 2403 2404 return 0; 2405 } 2406 2407 static int mana_push_wqe(struct mana_rxq *rxq) 2408 { 2409 struct mana_recv_buf_oob *rx_oob; 2410 u32 buf_idx; 2411 int err; 2412 2413 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2414 rx_oob = &rxq->rx_oobs[buf_idx]; 2415 2416 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req, 2417 &rx_oob->wqe_inf); 2418 if (err) 2419 return -ENOSPC; 2420 } 2421 2422 return 0; 2423 } 2424 2425 static int mana_create_page_pool(struct mana_rxq *rxq, struct gdma_context *gc) 2426 { 2427 struct mana_port_context *mpc = netdev_priv(rxq->ndev); 2428 struct page_pool_params pprm = {}; 2429 int ret; 2430 2431 pprm.pool_size = mpc->rx_queue_size; 2432 pprm.nid = gc->numa_node; 2433 pprm.napi = &rxq->rx_cq.napi; 2434 pprm.netdev = rxq->ndev; 2435 pprm.order = get_order(rxq->alloc_size); 2436 2437 rxq->page_pool = page_pool_create(&pprm); 2438 2439 if (IS_ERR(rxq->page_pool)) { 2440 ret = PTR_ERR(rxq->page_pool); 2441 rxq->page_pool = NULL; 2442 return ret; 2443 } 2444 2445 return 0; 2446 } 2447 2448 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc, 2449 u32 rxq_idx, struct mana_eq *eq, 2450 struct net_device *ndev) 2451 { 2452 struct gdma_dev *gd = apc->ac->gdma_dev; 2453 struct mana_obj_spec wq_spec; 2454 struct mana_obj_spec cq_spec; 2455 struct gdma_queue_spec spec; 2456 struct mana_cq *cq = NULL; 2457 struct gdma_context *gc; 2458 u32 cq_size, rq_size; 2459 struct mana_rxq *rxq; 2460 int err; 2461 2462 gc = gd->gdma_context; 2463 2464 rxq = kzalloc(struct_size(rxq, rx_oobs, apc->rx_queue_size), 2465 GFP_KERNEL); 2466 if (!rxq) 2467 return NULL; 2468 2469 rxq->ndev = ndev; 2470 rxq->num_rx_buf = apc->rx_queue_size; 2471 rxq->rxq_idx = rxq_idx; 2472 rxq->rxobj = INVALID_MANA_HANDLE; 2473 2474 mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size, 2475 &rxq->headroom); 2476 2477 /* Create page pool for RX queue */ 2478 err = mana_create_page_pool(rxq, gc); 2479 if (err) { 2480 netdev_err(ndev, "Create page pool err:%d\n", err); 2481 goto out; 2482 } 2483 2484 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size); 2485 if (err) 2486 goto out; 2487 2488 rq_size = MANA_PAGE_ALIGN(rq_size); 2489 cq_size = MANA_PAGE_ALIGN(cq_size); 2490 2491 /* Create RQ */ 2492 memset(&spec, 0, sizeof(spec)); 2493 spec.type = GDMA_RQ; 2494 spec.monitor_avl_buf = true; 2495 spec.queue_size = rq_size; 2496 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq); 2497 if (err) 2498 goto out; 2499 2500 /* Create RQ's CQ */ 2501 cq = &rxq->rx_cq; 2502 cq->type = MANA_CQ_TYPE_RX; 2503 cq->rxq = rxq; 2504 2505 memset(&spec, 0, sizeof(spec)); 2506 spec.type = GDMA_CQ; 2507 spec.monitor_avl_buf = false; 2508 spec.queue_size = cq_size; 2509 spec.cq.callback = mana_schedule_napi; 2510 spec.cq.parent_eq = eq->eq; 2511 spec.cq.context = cq; 2512 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 2513 if (err) 2514 goto out; 2515 2516 memset(&wq_spec, 0, sizeof(wq_spec)); 2517 memset(&cq_spec, 0, sizeof(cq_spec)); 2518 wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle; 2519 wq_spec.queue_size = rxq->gdma_rq->queue_size; 2520 2521 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 2522 cq_spec.queue_size = cq->gdma_cq->queue_size; 2523 cq_spec.modr_ctx_id = 0; 2524 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 2525 2526 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ, 2527 &wq_spec, &cq_spec, &rxq->rxobj); 2528 if (err) 2529 goto out; 2530 2531 rxq->gdma_rq->id = wq_spec.queue_index; 2532 cq->gdma_cq->id = cq_spec.queue_index; 2533 2534 rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2535 cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2536 2537 rxq->gdma_id = rxq->gdma_rq->id; 2538 cq->gdma_id = cq->gdma_cq->id; 2539 2540 err = mana_push_wqe(rxq); 2541 if (err) 2542 goto out; 2543 2544 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 2545 err = -EINVAL; 2546 goto out; 2547 } 2548 2549 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2550 2551 netdev_lock_ops_to_full(ndev); 2552 netif_napi_add_weight_locked(ndev, &cq->napi, mana_poll, 1); 2553 netdev_unlock_full_to_ops(ndev); 2554 2555 WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx, 2556 cq->napi.napi_id)); 2557 WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL, 2558 rxq->page_pool)); 2559 2560 netdev_lock_ops_to_full(ndev); 2561 napi_enable_locked(&cq->napi); 2562 netdev_unlock_full_to_ops(ndev); 2563 2564 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2565 out: 2566 if (!err) 2567 return rxq; 2568 2569 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err); 2570 2571 mana_destroy_rxq(apc, rxq, false); 2572 2573 if (cq) 2574 mana_deinit_cq(apc, cq); 2575 2576 return NULL; 2577 } 2578 2579 static void mana_create_rxq_debugfs(struct mana_port_context *apc, int idx) 2580 { 2581 struct mana_rxq *rxq; 2582 char qnum[32]; 2583 2584 rxq = apc->rxqs[idx]; 2585 2586 sprintf(qnum, "RX-%d", idx); 2587 rxq->mana_rx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs); 2588 debugfs_create_u32("rq_head", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->head); 2589 debugfs_create_u32("rq_tail", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->tail); 2590 debugfs_create_u32("rq_nbuf", 0400, rxq->mana_rx_debugfs, &rxq->num_rx_buf); 2591 debugfs_create_u32("cq_head", 0400, rxq->mana_rx_debugfs, 2592 &rxq->rx_cq.gdma_cq->head); 2593 debugfs_create_u32("cq_tail", 0400, rxq->mana_rx_debugfs, 2594 &rxq->rx_cq.gdma_cq->tail); 2595 debugfs_create_u32("cq_budget", 0400, rxq->mana_rx_debugfs, &rxq->rx_cq.budget); 2596 debugfs_create_file("rxq_dump", 0400, rxq->mana_rx_debugfs, rxq->gdma_rq, &mana_dbg_q_fops); 2597 debugfs_create_file("cq_dump", 0400, rxq->mana_rx_debugfs, rxq->rx_cq.gdma_cq, 2598 &mana_dbg_q_fops); 2599 } 2600 2601 static int mana_add_rx_queues(struct mana_port_context *apc, 2602 struct net_device *ndev) 2603 { 2604 struct mana_context *ac = apc->ac; 2605 struct mana_rxq *rxq; 2606 int err = 0; 2607 int i; 2608 2609 for (i = 0; i < apc->num_queues; i++) { 2610 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev); 2611 if (!rxq) { 2612 err = -ENOMEM; 2613 netdev_err(ndev, "Failed to create rxq %d : %d\n", i, err); 2614 goto out; 2615 } 2616 2617 u64_stats_init(&rxq->stats.syncp); 2618 2619 apc->rxqs[i] = rxq; 2620 2621 mana_create_rxq_debugfs(apc, i); 2622 } 2623 2624 apc->default_rxobj = apc->rxqs[0]->rxobj; 2625 out: 2626 return err; 2627 } 2628 2629 static void mana_destroy_vport(struct mana_port_context *apc) 2630 { 2631 struct gdma_dev *gd = apc->ac->gdma_dev; 2632 struct mana_rxq *rxq; 2633 u32 rxq_idx; 2634 2635 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 2636 rxq = apc->rxqs[rxq_idx]; 2637 if (!rxq) 2638 continue; 2639 2640 mana_destroy_rxq(apc, rxq, true); 2641 apc->rxqs[rxq_idx] = NULL; 2642 } 2643 2644 mana_destroy_txq(apc); 2645 mana_uncfg_vport(apc); 2646 2647 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) 2648 mana_pf_deregister_hw_vport(apc); 2649 } 2650 2651 static int mana_create_vport(struct mana_port_context *apc, 2652 struct net_device *net) 2653 { 2654 struct gdma_dev *gd = apc->ac->gdma_dev; 2655 int err; 2656 2657 apc->default_rxobj = INVALID_MANA_HANDLE; 2658 2659 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) { 2660 err = mana_pf_register_hw_vport(apc); 2661 if (err) 2662 return err; 2663 } 2664 2665 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell); 2666 if (err) 2667 return err; 2668 2669 return mana_create_txq(apc, net); 2670 } 2671 2672 static int mana_rss_table_alloc(struct mana_port_context *apc) 2673 { 2674 if (!apc->indir_table_sz) { 2675 netdev_err(apc->ndev, 2676 "Indirection table size not set for vPort %d\n", 2677 apc->port_idx); 2678 return -EINVAL; 2679 } 2680 2681 apc->indir_table = kcalloc(apc->indir_table_sz, sizeof(u32), GFP_KERNEL); 2682 if (!apc->indir_table) 2683 return -ENOMEM; 2684 2685 apc->rxobj_table = kcalloc(apc->indir_table_sz, sizeof(mana_handle_t), GFP_KERNEL); 2686 if (!apc->rxobj_table) { 2687 kfree(apc->indir_table); 2688 return -ENOMEM; 2689 } 2690 2691 return 0; 2692 } 2693 2694 static void mana_rss_table_init(struct mana_port_context *apc) 2695 { 2696 int i; 2697 2698 for (i = 0; i < apc->indir_table_sz; i++) 2699 apc->indir_table[i] = 2700 ethtool_rxfh_indir_default(i, apc->num_queues); 2701 } 2702 2703 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx, 2704 bool update_hash, bool update_tab) 2705 { 2706 u32 queue_idx; 2707 int err; 2708 int i; 2709 2710 if (update_tab) { 2711 for (i = 0; i < apc->indir_table_sz; i++) { 2712 queue_idx = apc->indir_table[i]; 2713 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj; 2714 } 2715 } 2716 2717 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab); 2718 if (err) 2719 return err; 2720 2721 mana_fence_rqs(apc); 2722 2723 return 0; 2724 } 2725 2726 void mana_query_gf_stats(struct mana_port_context *apc) 2727 { 2728 struct mana_query_gf_stat_resp resp = {}; 2729 struct mana_query_gf_stat_req req = {}; 2730 struct net_device *ndev = apc->ndev; 2731 int err; 2732 2733 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_GF_STAT, 2734 sizeof(req), sizeof(resp)); 2735 req.hdr.resp.msg_version = GDMA_MESSAGE_V2; 2736 req.req_stats = STATISTICS_FLAGS_RX_DISCARDS_NO_WQE | 2737 STATISTICS_FLAGS_RX_ERRORS_VPORT_DISABLED | 2738 STATISTICS_FLAGS_HC_RX_BYTES | 2739 STATISTICS_FLAGS_HC_RX_UCAST_PACKETS | 2740 STATISTICS_FLAGS_HC_RX_UCAST_BYTES | 2741 STATISTICS_FLAGS_HC_RX_MCAST_PACKETS | 2742 STATISTICS_FLAGS_HC_RX_MCAST_BYTES | 2743 STATISTICS_FLAGS_HC_RX_BCAST_PACKETS | 2744 STATISTICS_FLAGS_HC_RX_BCAST_BYTES | 2745 STATISTICS_FLAGS_TX_ERRORS_GF_DISABLED | 2746 STATISTICS_FLAGS_TX_ERRORS_VPORT_DISABLED | 2747 STATISTICS_FLAGS_TX_ERRORS_INVAL_VPORT_OFFSET_PACKETS | 2748 STATISTICS_FLAGS_TX_ERRORS_VLAN_ENFORCEMENT | 2749 STATISTICS_FLAGS_TX_ERRORS_ETH_TYPE_ENFORCEMENT | 2750 STATISTICS_FLAGS_TX_ERRORS_SA_ENFORCEMENT | 2751 STATISTICS_FLAGS_TX_ERRORS_SQPDID_ENFORCEMENT | 2752 STATISTICS_FLAGS_TX_ERRORS_CQPDID_ENFORCEMENT | 2753 STATISTICS_FLAGS_TX_ERRORS_MTU_VIOLATION | 2754 STATISTICS_FLAGS_TX_ERRORS_INVALID_OOB | 2755 STATISTICS_FLAGS_HC_TX_BYTES | 2756 STATISTICS_FLAGS_HC_TX_UCAST_PACKETS | 2757 STATISTICS_FLAGS_HC_TX_UCAST_BYTES | 2758 STATISTICS_FLAGS_HC_TX_MCAST_PACKETS | 2759 STATISTICS_FLAGS_HC_TX_MCAST_BYTES | 2760 STATISTICS_FLAGS_HC_TX_BCAST_PACKETS | 2761 STATISTICS_FLAGS_HC_TX_BCAST_BYTES | 2762 STATISTICS_FLAGS_TX_ERRORS_GDMA_ERROR; 2763 2764 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 2765 sizeof(resp)); 2766 if (err) { 2767 netdev_err(ndev, "Failed to query GF stats: %d\n", err); 2768 return; 2769 } 2770 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_GF_STAT, 2771 sizeof(resp)); 2772 if (err || resp.hdr.status) { 2773 netdev_err(ndev, "Failed to query GF stats: %d, 0x%x\n", err, 2774 resp.hdr.status); 2775 return; 2776 } 2777 2778 apc->eth_stats.hc_rx_discards_no_wqe = resp.rx_discards_nowqe; 2779 apc->eth_stats.hc_rx_err_vport_disabled = resp.rx_err_vport_disabled; 2780 apc->eth_stats.hc_rx_bytes = resp.hc_rx_bytes; 2781 apc->eth_stats.hc_rx_ucast_pkts = resp.hc_rx_ucast_pkts; 2782 apc->eth_stats.hc_rx_ucast_bytes = resp.hc_rx_ucast_bytes; 2783 apc->eth_stats.hc_rx_bcast_pkts = resp.hc_rx_bcast_pkts; 2784 apc->eth_stats.hc_rx_bcast_bytes = resp.hc_rx_bcast_bytes; 2785 apc->eth_stats.hc_rx_mcast_pkts = resp.hc_rx_mcast_pkts; 2786 apc->eth_stats.hc_rx_mcast_bytes = resp.hc_rx_mcast_bytes; 2787 apc->eth_stats.hc_tx_err_gf_disabled = resp.tx_err_gf_disabled; 2788 apc->eth_stats.hc_tx_err_vport_disabled = resp.tx_err_vport_disabled; 2789 apc->eth_stats.hc_tx_err_inval_vportoffset_pkt = 2790 resp.tx_err_inval_vport_offset_pkt; 2791 apc->eth_stats.hc_tx_err_vlan_enforcement = 2792 resp.tx_err_vlan_enforcement; 2793 apc->eth_stats.hc_tx_err_eth_type_enforcement = 2794 resp.tx_err_ethtype_enforcement; 2795 apc->eth_stats.hc_tx_err_sa_enforcement = resp.tx_err_SA_enforcement; 2796 apc->eth_stats.hc_tx_err_sqpdid_enforcement = 2797 resp.tx_err_SQPDID_enforcement; 2798 apc->eth_stats.hc_tx_err_cqpdid_enforcement = 2799 resp.tx_err_CQPDID_enforcement; 2800 apc->eth_stats.hc_tx_err_mtu_violation = resp.tx_err_mtu_violation; 2801 apc->eth_stats.hc_tx_err_inval_oob = resp.tx_err_inval_oob; 2802 apc->eth_stats.hc_tx_bytes = resp.hc_tx_bytes; 2803 apc->eth_stats.hc_tx_ucast_pkts = resp.hc_tx_ucast_pkts; 2804 apc->eth_stats.hc_tx_ucast_bytes = resp.hc_tx_ucast_bytes; 2805 apc->eth_stats.hc_tx_bcast_pkts = resp.hc_tx_bcast_pkts; 2806 apc->eth_stats.hc_tx_bcast_bytes = resp.hc_tx_bcast_bytes; 2807 apc->eth_stats.hc_tx_mcast_pkts = resp.hc_tx_mcast_pkts; 2808 apc->eth_stats.hc_tx_mcast_bytes = resp.hc_tx_mcast_bytes; 2809 apc->eth_stats.hc_tx_err_gdma = resp.tx_err_gdma; 2810 } 2811 2812 void mana_query_phy_stats(struct mana_port_context *apc) 2813 { 2814 struct mana_query_phy_stat_resp resp = {}; 2815 struct mana_query_phy_stat_req req = {}; 2816 struct net_device *ndev = apc->ndev; 2817 int err; 2818 2819 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_PHY_STAT, 2820 sizeof(req), sizeof(resp)); 2821 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 2822 sizeof(resp)); 2823 if (err) 2824 return; 2825 2826 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_PHY_STAT, 2827 sizeof(resp)); 2828 if (err || resp.hdr.status) { 2829 netdev_err(ndev, 2830 "Failed to query PHY stats: %d, resp:0x%x\n", 2831 err, resp.hdr.status); 2832 return; 2833 } 2834 2835 /* Aggregate drop counters */ 2836 apc->phy_stats.rx_pkt_drop_phy = resp.rx_pkt_drop_phy; 2837 apc->phy_stats.tx_pkt_drop_phy = resp.tx_pkt_drop_phy; 2838 2839 /* Per TC traffic Counters */ 2840 apc->phy_stats.rx_pkt_tc0_phy = resp.rx_pkt_tc0_phy; 2841 apc->phy_stats.tx_pkt_tc0_phy = resp.tx_pkt_tc0_phy; 2842 apc->phy_stats.rx_pkt_tc1_phy = resp.rx_pkt_tc1_phy; 2843 apc->phy_stats.tx_pkt_tc1_phy = resp.tx_pkt_tc1_phy; 2844 apc->phy_stats.rx_pkt_tc2_phy = resp.rx_pkt_tc2_phy; 2845 apc->phy_stats.tx_pkt_tc2_phy = resp.tx_pkt_tc2_phy; 2846 apc->phy_stats.rx_pkt_tc3_phy = resp.rx_pkt_tc3_phy; 2847 apc->phy_stats.tx_pkt_tc3_phy = resp.tx_pkt_tc3_phy; 2848 apc->phy_stats.rx_pkt_tc4_phy = resp.rx_pkt_tc4_phy; 2849 apc->phy_stats.tx_pkt_tc4_phy = resp.tx_pkt_tc4_phy; 2850 apc->phy_stats.rx_pkt_tc5_phy = resp.rx_pkt_tc5_phy; 2851 apc->phy_stats.tx_pkt_tc5_phy = resp.tx_pkt_tc5_phy; 2852 apc->phy_stats.rx_pkt_tc6_phy = resp.rx_pkt_tc6_phy; 2853 apc->phy_stats.tx_pkt_tc6_phy = resp.tx_pkt_tc6_phy; 2854 apc->phy_stats.rx_pkt_tc7_phy = resp.rx_pkt_tc7_phy; 2855 apc->phy_stats.tx_pkt_tc7_phy = resp.tx_pkt_tc7_phy; 2856 2857 /* Per TC byte Counters */ 2858 apc->phy_stats.rx_byte_tc0_phy = resp.rx_byte_tc0_phy; 2859 apc->phy_stats.tx_byte_tc0_phy = resp.tx_byte_tc0_phy; 2860 apc->phy_stats.rx_byte_tc1_phy = resp.rx_byte_tc1_phy; 2861 apc->phy_stats.tx_byte_tc1_phy = resp.tx_byte_tc1_phy; 2862 apc->phy_stats.rx_byte_tc2_phy = resp.rx_byte_tc2_phy; 2863 apc->phy_stats.tx_byte_tc2_phy = resp.tx_byte_tc2_phy; 2864 apc->phy_stats.rx_byte_tc3_phy = resp.rx_byte_tc3_phy; 2865 apc->phy_stats.tx_byte_tc3_phy = resp.tx_byte_tc3_phy; 2866 apc->phy_stats.rx_byte_tc4_phy = resp.rx_byte_tc4_phy; 2867 apc->phy_stats.tx_byte_tc4_phy = resp.tx_byte_tc4_phy; 2868 apc->phy_stats.rx_byte_tc5_phy = resp.rx_byte_tc5_phy; 2869 apc->phy_stats.tx_byte_tc5_phy = resp.tx_byte_tc5_phy; 2870 apc->phy_stats.rx_byte_tc6_phy = resp.rx_byte_tc6_phy; 2871 apc->phy_stats.tx_byte_tc6_phy = resp.tx_byte_tc6_phy; 2872 apc->phy_stats.rx_byte_tc7_phy = resp.rx_byte_tc7_phy; 2873 apc->phy_stats.tx_byte_tc7_phy = resp.tx_byte_tc7_phy; 2874 2875 /* Per TC pause Counters */ 2876 apc->phy_stats.rx_pause_tc0_phy = resp.rx_pause_tc0_phy; 2877 apc->phy_stats.tx_pause_tc0_phy = resp.tx_pause_tc0_phy; 2878 apc->phy_stats.rx_pause_tc1_phy = resp.rx_pause_tc1_phy; 2879 apc->phy_stats.tx_pause_tc1_phy = resp.tx_pause_tc1_phy; 2880 apc->phy_stats.rx_pause_tc2_phy = resp.rx_pause_tc2_phy; 2881 apc->phy_stats.tx_pause_tc2_phy = resp.tx_pause_tc2_phy; 2882 apc->phy_stats.rx_pause_tc3_phy = resp.rx_pause_tc3_phy; 2883 apc->phy_stats.tx_pause_tc3_phy = resp.tx_pause_tc3_phy; 2884 apc->phy_stats.rx_pause_tc4_phy = resp.rx_pause_tc4_phy; 2885 apc->phy_stats.tx_pause_tc4_phy = resp.tx_pause_tc4_phy; 2886 apc->phy_stats.rx_pause_tc5_phy = resp.rx_pause_tc5_phy; 2887 apc->phy_stats.tx_pause_tc5_phy = resp.tx_pause_tc5_phy; 2888 apc->phy_stats.rx_pause_tc6_phy = resp.rx_pause_tc6_phy; 2889 apc->phy_stats.tx_pause_tc6_phy = resp.tx_pause_tc6_phy; 2890 apc->phy_stats.rx_pause_tc7_phy = resp.rx_pause_tc7_phy; 2891 apc->phy_stats.tx_pause_tc7_phy = resp.tx_pause_tc7_phy; 2892 } 2893 2894 static int mana_init_port(struct net_device *ndev) 2895 { 2896 struct mana_port_context *apc = netdev_priv(ndev); 2897 struct gdma_dev *gd = apc->ac->gdma_dev; 2898 u32 max_txq, max_rxq, max_queues; 2899 int port_idx = apc->port_idx; 2900 struct gdma_context *gc; 2901 char vport[32]; 2902 int err; 2903 2904 err = mana_init_port_context(apc); 2905 if (err) 2906 return err; 2907 2908 gc = gd->gdma_context; 2909 2910 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq, 2911 &apc->indir_table_sz); 2912 if (err) { 2913 netdev_err(ndev, "Failed to query info for vPort %d\n", 2914 port_idx); 2915 goto reset_apc; 2916 } 2917 2918 max_queues = min_t(u32, max_txq, max_rxq); 2919 if (apc->max_queues > max_queues) 2920 apc->max_queues = max_queues; 2921 2922 if (apc->num_queues > apc->max_queues) 2923 apc->num_queues = apc->max_queues; 2924 2925 eth_hw_addr_set(ndev, apc->mac_addr); 2926 sprintf(vport, "vport%d", port_idx); 2927 apc->mana_port_debugfs = debugfs_create_dir(vport, gc->mana_pci_debugfs); 2928 return 0; 2929 2930 reset_apc: 2931 mana_cleanup_port_context(apc); 2932 return err; 2933 } 2934 2935 int mana_alloc_queues(struct net_device *ndev) 2936 { 2937 struct mana_port_context *apc = netdev_priv(ndev); 2938 struct gdma_dev *gd = apc->ac->gdma_dev; 2939 int err; 2940 2941 err = mana_create_vport(apc, ndev); 2942 if (err) { 2943 netdev_err(ndev, "Failed to create vPort %u : %d\n", apc->port_idx, err); 2944 return err; 2945 } 2946 2947 err = netif_set_real_num_tx_queues(ndev, apc->num_queues); 2948 if (err) { 2949 netdev_err(ndev, 2950 "netif_set_real_num_tx_queues () failed for ndev with num_queues %u : %d\n", 2951 apc->num_queues, err); 2952 goto destroy_vport; 2953 } 2954 2955 err = mana_add_rx_queues(apc, ndev); 2956 if (err) 2957 goto destroy_vport; 2958 2959 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE; 2960 2961 err = netif_set_real_num_rx_queues(ndev, apc->num_queues); 2962 if (err) { 2963 netdev_err(ndev, 2964 "netif_set_real_num_rx_queues () failed for ndev with num_queues %u : %d\n", 2965 apc->num_queues, err); 2966 goto destroy_vport; 2967 } 2968 2969 mana_rss_table_init(apc); 2970 2971 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true); 2972 if (err) { 2973 netdev_err(ndev, "Failed to configure RSS table: %d\n", err); 2974 goto destroy_vport; 2975 } 2976 2977 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) { 2978 err = mana_pf_register_filter(apc); 2979 if (err) 2980 goto destroy_vport; 2981 } 2982 2983 mana_chn_setxdp(apc, mana_xdp_get(apc)); 2984 2985 return 0; 2986 2987 destroy_vport: 2988 mana_destroy_vport(apc); 2989 return err; 2990 } 2991 2992 int mana_attach(struct net_device *ndev) 2993 { 2994 struct mana_port_context *apc = netdev_priv(ndev); 2995 int err; 2996 2997 ASSERT_RTNL(); 2998 2999 err = mana_init_port(ndev); 3000 if (err) 3001 return err; 3002 3003 if (apc->port_st_save) { 3004 err = mana_alloc_queues(ndev); 3005 if (err) { 3006 mana_cleanup_port_context(apc); 3007 return err; 3008 } 3009 } 3010 3011 apc->port_is_up = apc->port_st_save; 3012 3013 /* Ensure port state updated before txq state */ 3014 smp_wmb(); 3015 3016 if (apc->port_is_up) 3017 netif_carrier_on(ndev); 3018 3019 netif_device_attach(ndev); 3020 3021 return 0; 3022 } 3023 3024 static int mana_dealloc_queues(struct net_device *ndev) 3025 { 3026 struct mana_port_context *apc = netdev_priv(ndev); 3027 unsigned long timeout = jiffies + 120 * HZ; 3028 struct gdma_dev *gd = apc->ac->gdma_dev; 3029 struct mana_txq *txq; 3030 struct sk_buff *skb; 3031 int i, err; 3032 u32 tsleep; 3033 3034 if (apc->port_is_up) 3035 return -EINVAL; 3036 3037 mana_chn_setxdp(apc, NULL); 3038 3039 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) 3040 mana_pf_deregister_filter(apc); 3041 3042 /* No packet can be transmitted now since apc->port_is_up is false. 3043 * There is still a tiny chance that mana_poll_tx_cq() can re-enable 3044 * a txq because it may not timely see apc->port_is_up being cleared 3045 * to false, but it doesn't matter since mana_start_xmit() drops any 3046 * new packets due to apc->port_is_up being false. 3047 * 3048 * Drain all the in-flight TX packets. 3049 * A timeout of 120 seconds for all the queues is used. 3050 * This will break the while loop when h/w is not responding. 3051 * This value of 120 has been decided here considering max 3052 * number of queues. 3053 */ 3054 3055 for (i = 0; i < apc->num_queues; i++) { 3056 txq = &apc->tx_qp[i].txq; 3057 tsleep = 1000; 3058 while (atomic_read(&txq->pending_sends) > 0 && 3059 time_before(jiffies, timeout)) { 3060 usleep_range(tsleep, tsleep + 1000); 3061 tsleep <<= 1; 3062 } 3063 if (atomic_read(&txq->pending_sends)) { 3064 err = pcie_flr(to_pci_dev(gd->gdma_context->dev)); 3065 if (err) { 3066 netdev_err(ndev, "flr failed %d with %d pkts pending in txq %u\n", 3067 err, atomic_read(&txq->pending_sends), 3068 txq->gdma_txq_id); 3069 } 3070 break; 3071 } 3072 } 3073 3074 for (i = 0; i < apc->num_queues; i++) { 3075 txq = &apc->tx_qp[i].txq; 3076 while ((skb = skb_dequeue(&txq->pending_skbs))) { 3077 mana_unmap_skb(skb, apc); 3078 dev_kfree_skb_any(skb); 3079 } 3080 atomic_set(&txq->pending_sends, 0); 3081 } 3082 /* We're 100% sure the queues can no longer be woken up, because 3083 * we're sure now mana_poll_tx_cq() can't be running. 3084 */ 3085 3086 apc->rss_state = TRI_STATE_FALSE; 3087 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false); 3088 if (err && mana_en_need_log(apc, err)) 3089 netdev_err(ndev, "Failed to disable vPort: %d\n", err); 3090 3091 /* Even in err case, still need to cleanup the vPort */ 3092 mana_destroy_vport(apc); 3093 3094 return 0; 3095 } 3096 3097 int mana_detach(struct net_device *ndev, bool from_close) 3098 { 3099 struct mana_port_context *apc = netdev_priv(ndev); 3100 int err; 3101 3102 ASSERT_RTNL(); 3103 3104 apc->port_st_save = apc->port_is_up; 3105 apc->port_is_up = false; 3106 3107 /* Ensure port state updated before txq state */ 3108 smp_wmb(); 3109 3110 netif_tx_disable(ndev); 3111 netif_carrier_off(ndev); 3112 3113 if (apc->port_st_save) { 3114 err = mana_dealloc_queues(ndev); 3115 if (err) { 3116 netdev_err(ndev, "%s failed to deallocate queues: %d\n", __func__, err); 3117 return err; 3118 } 3119 } 3120 3121 if (!from_close) { 3122 netif_device_detach(ndev); 3123 mana_cleanup_port_context(apc); 3124 } 3125 3126 return 0; 3127 } 3128 3129 static int mana_probe_port(struct mana_context *ac, int port_idx, 3130 struct net_device **ndev_storage) 3131 { 3132 struct gdma_context *gc = ac->gdma_dev->gdma_context; 3133 struct mana_port_context *apc; 3134 struct net_device *ndev; 3135 int err; 3136 3137 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context), 3138 gc->max_num_queues); 3139 if (!ndev) 3140 return -ENOMEM; 3141 3142 *ndev_storage = ndev; 3143 3144 apc = netdev_priv(ndev); 3145 apc->ac = ac; 3146 apc->ndev = ndev; 3147 apc->max_queues = gc->max_num_queues; 3148 apc->num_queues = gc->max_num_queues; 3149 apc->tx_queue_size = DEF_TX_BUFFERS_PER_QUEUE; 3150 apc->rx_queue_size = DEF_RX_BUFFERS_PER_QUEUE; 3151 apc->port_handle = INVALID_MANA_HANDLE; 3152 apc->pf_filter_handle = INVALID_MANA_HANDLE; 3153 apc->port_idx = port_idx; 3154 3155 mutex_init(&apc->vport_mutex); 3156 apc->vport_use_count = 0; 3157 3158 ndev->netdev_ops = &mana_devops; 3159 ndev->ethtool_ops = &mana_ethtool_ops; 3160 ndev->mtu = ETH_DATA_LEN; 3161 ndev->max_mtu = gc->adapter_mtu - ETH_HLEN; 3162 ndev->min_mtu = ETH_MIN_MTU; 3163 ndev->needed_headroom = MANA_HEADROOM; 3164 ndev->dev_port = port_idx; 3165 SET_NETDEV_DEV(ndev, gc->dev); 3166 3167 netif_set_tso_max_size(ndev, GSO_MAX_SIZE); 3168 3169 netif_carrier_off(ndev); 3170 3171 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE); 3172 3173 err = mana_init_port(ndev); 3174 if (err) 3175 goto free_net; 3176 3177 err = mana_rss_table_alloc(apc); 3178 if (err) 3179 goto reset_apc; 3180 3181 netdev_lockdep_set_classes(ndev); 3182 3183 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 3184 ndev->hw_features |= NETIF_F_RXCSUM; 3185 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6; 3186 ndev->hw_features |= NETIF_F_RXHASH; 3187 ndev->features = ndev->hw_features | NETIF_F_HW_VLAN_CTAG_TX | 3188 NETIF_F_HW_VLAN_CTAG_RX; 3189 ndev->vlan_features = ndev->features; 3190 xdp_set_features_flag(ndev, NETDEV_XDP_ACT_BASIC | 3191 NETDEV_XDP_ACT_REDIRECT | 3192 NETDEV_XDP_ACT_NDO_XMIT); 3193 3194 err = register_netdev(ndev); 3195 if (err) { 3196 netdev_err(ndev, "Unable to register netdev.\n"); 3197 goto free_indir; 3198 } 3199 3200 debugfs_create_u32("current_speed", 0400, apc->mana_port_debugfs, &apc->speed); 3201 3202 return 0; 3203 3204 free_indir: 3205 mana_cleanup_indir_table(apc); 3206 reset_apc: 3207 mana_cleanup_port_context(apc); 3208 free_net: 3209 *ndev_storage = NULL; 3210 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err); 3211 free_netdev(ndev); 3212 return err; 3213 } 3214 3215 static void adev_release(struct device *dev) 3216 { 3217 struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev); 3218 3219 kfree(madev); 3220 } 3221 3222 static void remove_adev(struct gdma_dev *gd) 3223 { 3224 struct auxiliary_device *adev = gd->adev; 3225 int id = adev->id; 3226 3227 auxiliary_device_delete(adev); 3228 auxiliary_device_uninit(adev); 3229 3230 mana_adev_idx_free(id); 3231 gd->adev = NULL; 3232 } 3233 3234 static int add_adev(struct gdma_dev *gd, const char *name) 3235 { 3236 struct auxiliary_device *adev; 3237 struct mana_adev *madev; 3238 int ret; 3239 3240 madev = kzalloc(sizeof(*madev), GFP_KERNEL); 3241 if (!madev) 3242 return -ENOMEM; 3243 3244 adev = &madev->adev; 3245 ret = mana_adev_idx_alloc(); 3246 if (ret < 0) 3247 goto idx_fail; 3248 adev->id = ret; 3249 3250 adev->name = name; 3251 adev->dev.parent = gd->gdma_context->dev; 3252 adev->dev.release = adev_release; 3253 madev->mdev = gd; 3254 3255 ret = auxiliary_device_init(adev); 3256 if (ret) 3257 goto init_fail; 3258 3259 /* madev is owned by the auxiliary device */ 3260 madev = NULL; 3261 ret = auxiliary_device_add(adev); 3262 if (ret) 3263 goto add_fail; 3264 3265 gd->adev = adev; 3266 dev_dbg(gd->gdma_context->dev, 3267 "Auxiliary device added successfully\n"); 3268 return 0; 3269 3270 add_fail: 3271 auxiliary_device_uninit(adev); 3272 3273 init_fail: 3274 mana_adev_idx_free(adev->id); 3275 3276 idx_fail: 3277 kfree(madev); 3278 3279 return ret; 3280 } 3281 3282 static void mana_rdma_service_handle(struct work_struct *work) 3283 { 3284 struct mana_service_work *serv_work = 3285 container_of(work, struct mana_service_work, work); 3286 struct gdma_dev *gd = serv_work->gdma_dev; 3287 struct device *dev = gd->gdma_context->dev; 3288 int ret; 3289 3290 if (READ_ONCE(gd->rdma_teardown)) 3291 goto out; 3292 3293 switch (serv_work->event) { 3294 case GDMA_SERVICE_TYPE_RDMA_SUSPEND: 3295 if (!gd->adev || gd->is_suspended) 3296 break; 3297 3298 remove_adev(gd); 3299 gd->is_suspended = true; 3300 break; 3301 3302 case GDMA_SERVICE_TYPE_RDMA_RESUME: 3303 if (!gd->is_suspended) 3304 break; 3305 3306 ret = add_adev(gd, "rdma"); 3307 if (ret) 3308 dev_err(dev, "Failed to add adev on resume: %d\n", ret); 3309 else 3310 gd->is_suspended = false; 3311 break; 3312 3313 default: 3314 dev_warn(dev, "unknown adev service event %u\n", 3315 serv_work->event); 3316 break; 3317 } 3318 3319 out: 3320 kfree(serv_work); 3321 } 3322 3323 int mana_rdma_service_event(struct gdma_context *gc, enum gdma_service_type event) 3324 { 3325 struct gdma_dev *gd = &gc->mana_ib; 3326 struct mana_service_work *serv_work; 3327 3328 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) { 3329 /* RDMA device is not detected on pci */ 3330 return 0; 3331 } 3332 3333 serv_work = kzalloc(sizeof(*serv_work), GFP_ATOMIC); 3334 if (!serv_work) 3335 return -ENOMEM; 3336 3337 serv_work->event = event; 3338 serv_work->gdma_dev = gd; 3339 3340 INIT_WORK(&serv_work->work, mana_rdma_service_handle); 3341 queue_work(gc->service_wq, &serv_work->work); 3342 3343 return 0; 3344 } 3345 3346 int mana_probe(struct gdma_dev *gd, bool resuming) 3347 { 3348 struct gdma_context *gc = gd->gdma_context; 3349 struct mana_context *ac = gd->driver_data; 3350 struct device *dev = gc->dev; 3351 u8 bm_hostmode = 0; 3352 u16 num_ports = 0; 3353 int err; 3354 int i; 3355 3356 dev_info(dev, 3357 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n", 3358 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION); 3359 3360 err = mana_gd_register_device(gd); 3361 if (err) 3362 return err; 3363 3364 if (!resuming) { 3365 ac = kzalloc(sizeof(*ac), GFP_KERNEL); 3366 if (!ac) 3367 return -ENOMEM; 3368 3369 ac->gdma_dev = gd; 3370 gd->driver_data = ac; 3371 } 3372 3373 err = mana_create_eq(ac); 3374 if (err) { 3375 dev_err(dev, "Failed to create EQs: %d\n", err); 3376 goto out; 3377 } 3378 3379 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION, 3380 MANA_MICRO_VERSION, &num_ports, &bm_hostmode); 3381 if (err) 3382 goto out; 3383 3384 ac->bm_hostmode = bm_hostmode; 3385 3386 if (!resuming) { 3387 ac->num_ports = num_ports; 3388 } else { 3389 if (ac->num_ports != num_ports) { 3390 dev_err(dev, "The number of vPorts changed: %d->%d\n", 3391 ac->num_ports, num_ports); 3392 err = -EPROTO; 3393 goto out; 3394 } 3395 } 3396 3397 if (ac->num_ports == 0) 3398 dev_err(dev, "Failed to detect any vPort\n"); 3399 3400 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV) 3401 ac->num_ports = MAX_PORTS_IN_MANA_DEV; 3402 3403 if (!resuming) { 3404 for (i = 0; i < ac->num_ports; i++) { 3405 err = mana_probe_port(ac, i, &ac->ports[i]); 3406 /* we log the port for which the probe failed and stop 3407 * probes for subsequent ports. 3408 * Note that we keep running ports, for which the probes 3409 * were successful, unless add_adev fails too 3410 */ 3411 if (err) { 3412 dev_err(dev, "Probe Failed for port %d\n", i); 3413 break; 3414 } 3415 } 3416 } else { 3417 for (i = 0; i < ac->num_ports; i++) { 3418 rtnl_lock(); 3419 err = mana_attach(ac->ports[i]); 3420 rtnl_unlock(); 3421 /* we log the port for which the attach failed and stop 3422 * attach for subsequent ports 3423 * Note that we keep running ports, for which the attach 3424 * were successful, unless add_adev fails too 3425 */ 3426 if (err) { 3427 dev_err(dev, "Attach Failed for port %d\n", i); 3428 break; 3429 } 3430 } 3431 } 3432 3433 err = add_adev(gd, "eth"); 3434 out: 3435 if (err) { 3436 mana_remove(gd, false); 3437 } else { 3438 dev_dbg(dev, "gd=%p, id=%u, num_ports=%d, type=%u, instance=%u\n", 3439 gd, gd->dev_id.as_uint32, ac->num_ports, 3440 gd->dev_id.type, gd->dev_id.instance); 3441 dev_dbg(dev, "%s succeeded\n", __func__); 3442 } 3443 3444 return err; 3445 } 3446 3447 void mana_remove(struct gdma_dev *gd, bool suspending) 3448 { 3449 struct gdma_context *gc = gd->gdma_context; 3450 struct mana_context *ac = gd->driver_data; 3451 struct mana_port_context *apc; 3452 struct device *dev = gc->dev; 3453 struct net_device *ndev; 3454 int err; 3455 int i; 3456 3457 /* adev currently doesn't support suspending, always remove it */ 3458 if (gd->adev) 3459 remove_adev(gd); 3460 3461 for (i = 0; i < ac->num_ports; i++) { 3462 ndev = ac->ports[i]; 3463 apc = netdev_priv(ndev); 3464 if (!ndev) { 3465 if (i == 0) 3466 dev_err(dev, "No net device to remove\n"); 3467 goto out; 3468 } 3469 3470 /* All cleanup actions should stay after rtnl_lock(), otherwise 3471 * other functions may access partially cleaned up data. 3472 */ 3473 rtnl_lock(); 3474 3475 err = mana_detach(ndev, false); 3476 if (err) 3477 netdev_err(ndev, "Failed to detach vPort %d: %d\n", 3478 i, err); 3479 3480 if (suspending) { 3481 /* No need to unregister the ndev. */ 3482 rtnl_unlock(); 3483 continue; 3484 } 3485 3486 unregister_netdevice(ndev); 3487 mana_cleanup_indir_table(apc); 3488 3489 rtnl_unlock(); 3490 3491 free_netdev(ndev); 3492 } 3493 3494 mana_destroy_eq(ac); 3495 out: 3496 mana_gd_deregister_device(gd); 3497 3498 if (suspending) 3499 return; 3500 3501 gd->driver_data = NULL; 3502 gd->gdma_context = NULL; 3503 kfree(ac); 3504 dev_dbg(dev, "%s succeeded\n", __func__); 3505 } 3506 3507 int mana_rdma_probe(struct gdma_dev *gd) 3508 { 3509 int err = 0; 3510 3511 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) { 3512 /* RDMA device is not detected on pci */ 3513 return err; 3514 } 3515 3516 err = mana_gd_register_device(gd); 3517 if (err) 3518 return err; 3519 3520 err = add_adev(gd, "rdma"); 3521 if (err) 3522 mana_gd_deregister_device(gd); 3523 3524 return err; 3525 } 3526 3527 void mana_rdma_remove(struct gdma_dev *gd) 3528 { 3529 struct gdma_context *gc = gd->gdma_context; 3530 3531 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) { 3532 /* RDMA device is not detected on pci */ 3533 return; 3534 } 3535 3536 WRITE_ONCE(gd->rdma_teardown, true); 3537 flush_workqueue(gc->service_wq); 3538 3539 if (gd->adev) 3540 remove_adev(gd); 3541 3542 mana_gd_deregister_device(gd); 3543 } 3544 3545 struct net_device *mana_get_primary_netdev(struct mana_context *ac, 3546 u32 port_index, 3547 netdevice_tracker *tracker) 3548 { 3549 struct net_device *ndev; 3550 3551 if (port_index >= ac->num_ports) 3552 return NULL; 3553 3554 rcu_read_lock(); 3555 3556 /* If mana is used in netvsc, the upper netdevice should be returned. */ 3557 ndev = netdev_master_upper_dev_get_rcu(ac->ports[port_index]); 3558 3559 /* If there is no upper device, use the parent Ethernet device */ 3560 if (!ndev) 3561 ndev = ac->ports[port_index]; 3562 3563 netdev_hold(ndev, tracker, GFP_ATOMIC); 3564 rcu_read_unlock(); 3565 3566 return ndev; 3567 } 3568 EXPORT_SYMBOL_NS(mana_get_primary_netdev, "NET_MANA"); 3569