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 #include <linux/export.h> 14 15 #include <net/checksum.h> 16 #include <net/ip6_checksum.h> 17 #include <net/netdev_lock.h> 18 #include <net/page_pool/helpers.h> 19 #include <net/xdp.h> 20 21 #include <net/mana/mana.h> 22 #include <net/mana/mana_auxiliary.h> 23 24 static DEFINE_IDA(mana_adev_ida); 25 26 static int mana_adev_idx_alloc(void) 27 { 28 return ida_alloc(&mana_adev_ida, GFP_KERNEL); 29 } 30 31 static void mana_adev_idx_free(int idx) 32 { 33 ida_free(&mana_adev_ida, idx); 34 } 35 36 static ssize_t mana_dbg_q_read(struct file *filp, char __user *buf, size_t count, 37 loff_t *pos) 38 { 39 struct gdma_queue *gdma_q = filp->private_data; 40 41 return simple_read_from_buffer(buf, count, pos, gdma_q->queue_mem_ptr, 42 gdma_q->queue_size); 43 } 44 45 static const struct file_operations mana_dbg_q_fops = { 46 .owner = THIS_MODULE, 47 .open = simple_open, 48 .read = mana_dbg_q_read, 49 }; 50 51 static bool mana_en_need_log(struct mana_port_context *apc, int err) 52 { 53 if (apc && apc->ac && apc->ac->gdma_dev && 54 apc->ac->gdma_dev->gdma_context) 55 return mana_need_log(apc->ac->gdma_dev->gdma_context, err); 56 else 57 return true; 58 } 59 60 /* Microsoft Azure Network Adapter (MANA) functions */ 61 62 static int mana_open(struct net_device *ndev) 63 { 64 struct mana_port_context *apc = netdev_priv(ndev); 65 int err; 66 err = mana_alloc_queues(ndev); 67 68 if (err) { 69 netdev_err(ndev, "%s failed to allocate queues: %d\n", __func__, err); 70 return err; 71 } 72 73 apc->port_is_up = true; 74 75 /* Ensure port state updated before txq state */ 76 smp_wmb(); 77 78 netif_carrier_on(ndev); 79 netif_tx_wake_all_queues(ndev); 80 netdev_dbg(ndev, "%s successful\n", __func__); 81 return 0; 82 } 83 84 static int mana_close(struct net_device *ndev) 85 { 86 struct mana_port_context *apc = netdev_priv(ndev); 87 88 if (!apc->port_is_up) 89 return 0; 90 91 return mana_detach(ndev, true); 92 } 93 94 static bool mana_can_tx(struct gdma_queue *wq) 95 { 96 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE; 97 } 98 99 static unsigned int mana_checksum_info(struct sk_buff *skb) 100 { 101 if (skb->protocol == htons(ETH_P_IP)) { 102 struct iphdr *ip = ip_hdr(skb); 103 104 if (ip->protocol == IPPROTO_TCP) 105 return IPPROTO_TCP; 106 107 if (ip->protocol == IPPROTO_UDP) 108 return IPPROTO_UDP; 109 } else if (skb->protocol == htons(ETH_P_IPV6)) { 110 struct ipv6hdr *ip6 = ipv6_hdr(skb); 111 112 if (ip6->nexthdr == IPPROTO_TCP) 113 return IPPROTO_TCP; 114 115 if (ip6->nexthdr == IPPROTO_UDP) 116 return IPPROTO_UDP; 117 } 118 119 /* No csum offloading */ 120 return 0; 121 } 122 123 static void mana_add_sge(struct mana_tx_package *tp, struct mana_skb_head *ash, 124 int sg_i, dma_addr_t da, int sge_len, u32 gpa_mkey) 125 { 126 ash->dma_handle[sg_i] = da; 127 ash->size[sg_i] = sge_len; 128 129 tp->wqe_req.sgl[sg_i].address = da; 130 tp->wqe_req.sgl[sg_i].mem_key = gpa_mkey; 131 tp->wqe_req.sgl[sg_i].size = sge_len; 132 } 133 134 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc, 135 struct mana_tx_package *tp, int gso_hs) 136 { 137 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 138 int hsg = 1; /* num of SGEs of linear part */ 139 struct gdma_dev *gd = apc->ac->gdma_dev; 140 int skb_hlen = skb_headlen(skb); 141 int sge0_len, sge1_len = 0; 142 struct gdma_context *gc; 143 struct device *dev; 144 skb_frag_t *frag; 145 dma_addr_t da; 146 int sg_i; 147 int i; 148 149 gc = gd->gdma_context; 150 dev = gc->dev; 151 152 if (gso_hs && gso_hs < skb_hlen) { 153 sge0_len = gso_hs; 154 sge1_len = skb_hlen - gso_hs; 155 } else { 156 sge0_len = skb_hlen; 157 } 158 159 da = dma_map_single(dev, skb->data, sge0_len, DMA_TO_DEVICE); 160 if (dma_mapping_error(dev, da)) 161 return -ENOMEM; 162 163 mana_add_sge(tp, ash, 0, da, sge0_len, gd->gpa_mkey); 164 165 if (sge1_len) { 166 sg_i = 1; 167 da = dma_map_single(dev, skb->data + sge0_len, sge1_len, 168 DMA_TO_DEVICE); 169 if (dma_mapping_error(dev, da)) 170 goto frag_err; 171 172 mana_add_sge(tp, ash, sg_i, da, sge1_len, gd->gpa_mkey); 173 hsg = 2; 174 } 175 176 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 177 sg_i = hsg + i; 178 179 frag = &skb_shinfo(skb)->frags[i]; 180 da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag), 181 DMA_TO_DEVICE); 182 if (dma_mapping_error(dev, da)) 183 goto frag_err; 184 185 mana_add_sge(tp, ash, sg_i, da, skb_frag_size(frag), 186 gd->gpa_mkey); 187 } 188 189 return 0; 190 191 frag_err: 192 if (net_ratelimit()) 193 netdev_err(apc->ndev, "Failed to map skb of size %u to DMA\n", 194 skb->len); 195 for (i = sg_i - 1; i >= hsg; i--) 196 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i], 197 DMA_TO_DEVICE); 198 199 for (i = hsg - 1; i >= 0; i--) 200 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i], 201 DMA_TO_DEVICE); 202 203 return -ENOMEM; 204 } 205 206 /* Handle the case when GSO SKB linear length is too large. 207 * MANA NIC requires GSO packets to put only the packet header to SGE0. 208 * So, we need 2 SGEs for the skb linear part which contains more than the 209 * header. 210 * Return a positive value for the number of SGEs, or a negative value 211 * for an error. 212 */ 213 static int mana_fix_skb_head(struct net_device *ndev, struct sk_buff *skb, 214 int gso_hs) 215 { 216 int num_sge = 1 + skb_shinfo(skb)->nr_frags; 217 int skb_hlen = skb_headlen(skb); 218 219 if (gso_hs < skb_hlen) { 220 num_sge++; 221 } else if (gso_hs > skb_hlen) { 222 if (net_ratelimit()) 223 netdev_err(ndev, 224 "TX nonlinear head: hs:%d, skb_hlen:%d\n", 225 gso_hs, skb_hlen); 226 227 return -EINVAL; 228 } 229 230 return num_sge; 231 } 232 233 /* Get the GSO packet's header size */ 234 static int mana_get_gso_hs(struct sk_buff *skb) 235 { 236 int gso_hs; 237 238 if (skb->encapsulation) { 239 gso_hs = skb_inner_tcp_all_headers(skb); 240 } else { 241 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) { 242 gso_hs = skb_transport_offset(skb) + 243 sizeof(struct udphdr); 244 } else { 245 gso_hs = skb_tcp_all_headers(skb); 246 } 247 } 248 249 return gso_hs; 250 } 251 252 netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct net_device *ndev) 253 { 254 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT; 255 struct mana_port_context *apc = netdev_priv(ndev); 256 int gso_hs = 0; /* zero for non-GSO pkts */ 257 u16 txq_idx = skb_get_queue_mapping(skb); 258 struct gdma_dev *gd = apc->ac->gdma_dev; 259 bool ipv4 = false, ipv6 = false; 260 struct mana_tx_package pkg = {}; 261 struct netdev_queue *net_txq; 262 struct mana_stats_tx *tx_stats; 263 struct gdma_queue *gdma_sq; 264 int err, len, num_gso_seg; 265 unsigned int csum_type; 266 struct mana_txq *txq; 267 struct mana_cq *cq; 268 269 if (unlikely(!apc->port_is_up)) 270 goto tx_drop; 271 272 if (skb_cow_head(skb, MANA_HEADROOM)) 273 goto tx_drop_count; 274 275 if (unlikely(ipv6_hopopt_jumbo_remove(skb))) 276 goto tx_drop_count; 277 278 txq = &apc->tx_qp[txq_idx].txq; 279 gdma_sq = txq->gdma_sq; 280 cq = &apc->tx_qp[txq_idx].tx_cq; 281 tx_stats = &txq->stats; 282 283 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id; 284 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame; 285 286 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) { 287 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset; 288 pkt_fmt = MANA_LONG_PKT_FMT; 289 } else { 290 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset; 291 } 292 293 if (skb_vlan_tag_present(skb)) { 294 pkt_fmt = MANA_LONG_PKT_FMT; 295 pkg.tx_oob.l_oob.inject_vlan_pri_tag = 1; 296 pkg.tx_oob.l_oob.pcp = skb_vlan_tag_get_prio(skb); 297 pkg.tx_oob.l_oob.dei = skb_vlan_tag_get_cfi(skb); 298 pkg.tx_oob.l_oob.vlan_id = skb_vlan_tag_get_id(skb); 299 } 300 301 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt; 302 303 if (pkt_fmt == MANA_SHORT_PKT_FMT) { 304 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob); 305 u64_stats_update_begin(&tx_stats->syncp); 306 tx_stats->short_pkt_fmt++; 307 u64_stats_update_end(&tx_stats->syncp); 308 } else { 309 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob); 310 u64_stats_update_begin(&tx_stats->syncp); 311 tx_stats->long_pkt_fmt++; 312 u64_stats_update_end(&tx_stats->syncp); 313 } 314 315 pkg.wqe_req.inline_oob_data = &pkg.tx_oob; 316 pkg.wqe_req.flags = 0; 317 pkg.wqe_req.client_data_unit = 0; 318 319 pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags; 320 321 if (skb->protocol == htons(ETH_P_IP)) 322 ipv4 = true; 323 else if (skb->protocol == htons(ETH_P_IPV6)) 324 ipv6 = true; 325 326 if (skb_is_gso(skb)) { 327 int num_sge; 328 329 gso_hs = mana_get_gso_hs(skb); 330 331 num_sge = mana_fix_skb_head(ndev, skb, gso_hs); 332 if (num_sge > 0) 333 pkg.wqe_req.num_sge = num_sge; 334 else 335 goto tx_drop_count; 336 337 u64_stats_update_begin(&tx_stats->syncp); 338 if (skb->encapsulation) { 339 tx_stats->tso_inner_packets++; 340 tx_stats->tso_inner_bytes += skb->len - gso_hs; 341 } else { 342 tx_stats->tso_packets++; 343 tx_stats->tso_bytes += skb->len - gso_hs; 344 } 345 u64_stats_update_end(&tx_stats->syncp); 346 347 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 348 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 349 350 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 351 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 352 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 353 354 pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size; 355 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0; 356 if (ipv4) { 357 ip_hdr(skb)->tot_len = 0; 358 ip_hdr(skb)->check = 0; 359 tcp_hdr(skb)->check = 360 ~csum_tcpudp_magic(ip_hdr(skb)->saddr, 361 ip_hdr(skb)->daddr, 0, 362 IPPROTO_TCP, 0); 363 } else { 364 ipv6_hdr(skb)->payload_len = 0; 365 tcp_hdr(skb)->check = 366 ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr, 367 &ipv6_hdr(skb)->daddr, 0, 368 IPPROTO_TCP, 0); 369 } 370 } else if (skb->ip_summed == CHECKSUM_PARTIAL) { 371 csum_type = mana_checksum_info(skb); 372 373 u64_stats_update_begin(&tx_stats->syncp); 374 tx_stats->csum_partial++; 375 u64_stats_update_end(&tx_stats->syncp); 376 377 if (csum_type == IPPROTO_TCP) { 378 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 379 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 380 381 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 382 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 383 384 } else if (csum_type == IPPROTO_UDP) { 385 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 386 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 387 388 pkg.tx_oob.s_oob.comp_udp_csum = 1; 389 } else { 390 /* Can't do offload of this type of checksum */ 391 if (skb_checksum_help(skb)) 392 goto tx_drop_count; 393 } 394 } 395 396 WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES); 397 398 if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) { 399 pkg.wqe_req.sgl = pkg.sgl_array; 400 } else { 401 pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge, 402 sizeof(struct gdma_sge), 403 GFP_ATOMIC); 404 if (!pkg.sgl_ptr) 405 goto tx_drop_count; 406 407 pkg.wqe_req.sgl = pkg.sgl_ptr; 408 } 409 410 if (mana_map_skb(skb, apc, &pkg, gso_hs)) { 411 u64_stats_update_begin(&tx_stats->syncp); 412 tx_stats->mana_map_err++; 413 u64_stats_update_end(&tx_stats->syncp); 414 goto free_sgl_ptr; 415 } 416 417 skb_queue_tail(&txq->pending_skbs, skb); 418 419 len = skb->len; 420 num_gso_seg = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1; 421 net_txq = netdev_get_tx_queue(ndev, txq_idx); 422 423 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req, 424 (struct gdma_posted_wqe_info *)skb->cb); 425 if (!mana_can_tx(gdma_sq)) { 426 netif_tx_stop_queue(net_txq); 427 apc->eth_stats.stop_queue++; 428 } 429 430 if (err) { 431 (void)skb_dequeue_tail(&txq->pending_skbs); 432 netdev_warn(ndev, "Failed to post TX OOB: %d\n", err); 433 err = NETDEV_TX_BUSY; 434 goto tx_busy; 435 } 436 437 err = NETDEV_TX_OK; 438 atomic_inc(&txq->pending_sends); 439 440 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq); 441 442 /* skb may be freed after mana_gd_post_work_request. Do not use it. */ 443 skb = NULL; 444 445 /* Populated the packet and bytes counters based on post GSO packet 446 * calculations 447 */ 448 tx_stats = &txq->stats; 449 u64_stats_update_begin(&tx_stats->syncp); 450 tx_stats->packets += num_gso_seg; 451 tx_stats->bytes += len + ((num_gso_seg - 1) * gso_hs); 452 u64_stats_update_end(&tx_stats->syncp); 453 454 tx_busy: 455 if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) { 456 netif_tx_wake_queue(net_txq); 457 apc->eth_stats.wake_queue++; 458 } 459 460 kfree(pkg.sgl_ptr); 461 return err; 462 463 free_sgl_ptr: 464 kfree(pkg.sgl_ptr); 465 tx_drop_count: 466 ndev->stats.tx_dropped++; 467 tx_drop: 468 dev_kfree_skb_any(skb); 469 return NETDEV_TX_OK; 470 } 471 472 static void mana_get_stats64(struct net_device *ndev, 473 struct rtnl_link_stats64 *st) 474 { 475 struct mana_port_context *apc = netdev_priv(ndev); 476 unsigned int num_queues = apc->num_queues; 477 struct mana_stats_rx *rx_stats; 478 struct mana_stats_tx *tx_stats; 479 unsigned int start; 480 u64 packets, bytes; 481 int q; 482 483 if (!apc->port_is_up) 484 return; 485 486 netdev_stats_to_stats64(st, &ndev->stats); 487 488 for (q = 0; q < num_queues; q++) { 489 rx_stats = &apc->rxqs[q]->stats; 490 491 do { 492 start = u64_stats_fetch_begin(&rx_stats->syncp); 493 packets = rx_stats->packets; 494 bytes = rx_stats->bytes; 495 } while (u64_stats_fetch_retry(&rx_stats->syncp, start)); 496 497 st->rx_packets += packets; 498 st->rx_bytes += bytes; 499 } 500 501 for (q = 0; q < num_queues; q++) { 502 tx_stats = &apc->tx_qp[q].txq.stats; 503 504 do { 505 start = u64_stats_fetch_begin(&tx_stats->syncp); 506 packets = tx_stats->packets; 507 bytes = tx_stats->bytes; 508 } while (u64_stats_fetch_retry(&tx_stats->syncp, start)); 509 510 st->tx_packets += packets; 511 st->tx_bytes += bytes; 512 } 513 } 514 515 static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb, 516 int old_q) 517 { 518 struct mana_port_context *apc = netdev_priv(ndev); 519 u32 hash = skb_get_hash(skb); 520 struct sock *sk = skb->sk; 521 int txq; 522 523 txq = apc->indir_table[hash & (apc->indir_table_sz - 1)]; 524 525 if (txq != old_q && sk && sk_fullsock(sk) && 526 rcu_access_pointer(sk->sk_dst_cache)) 527 sk_tx_queue_set(sk, txq); 528 529 return txq; 530 } 531 532 static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb, 533 struct net_device *sb_dev) 534 { 535 int txq; 536 537 if (ndev->real_num_tx_queues == 1) 538 return 0; 539 540 txq = sk_tx_queue_get(skb->sk); 541 542 if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) { 543 if (skb_rx_queue_recorded(skb)) 544 txq = skb_get_rx_queue(skb); 545 else 546 txq = mana_get_tx_queue(ndev, skb, txq); 547 } 548 549 return txq; 550 } 551 552 /* Release pre-allocated RX buffers */ 553 void mana_pre_dealloc_rxbufs(struct mana_port_context *mpc) 554 { 555 struct device *dev; 556 int i; 557 558 dev = mpc->ac->gdma_dev->gdma_context->dev; 559 560 if (!mpc->rxbufs_pre) 561 goto out1; 562 563 if (!mpc->das_pre) 564 goto out2; 565 566 while (mpc->rxbpre_total) { 567 i = --mpc->rxbpre_total; 568 dma_unmap_single(dev, mpc->das_pre[i], mpc->rxbpre_datasize, 569 DMA_FROM_DEVICE); 570 put_page(virt_to_head_page(mpc->rxbufs_pre[i])); 571 } 572 573 kfree(mpc->das_pre); 574 mpc->das_pre = NULL; 575 576 out2: 577 kfree(mpc->rxbufs_pre); 578 mpc->rxbufs_pre = NULL; 579 580 out1: 581 mpc->rxbpre_datasize = 0; 582 mpc->rxbpre_alloc_size = 0; 583 mpc->rxbpre_headroom = 0; 584 } 585 586 /* Get a buffer from the pre-allocated RX buffers */ 587 static void *mana_get_rxbuf_pre(struct mana_rxq *rxq, dma_addr_t *da) 588 { 589 struct net_device *ndev = rxq->ndev; 590 struct mana_port_context *mpc; 591 void *va; 592 593 mpc = netdev_priv(ndev); 594 595 if (!mpc->rxbufs_pre || !mpc->das_pre || !mpc->rxbpre_total) { 596 netdev_err(ndev, "No RX pre-allocated bufs\n"); 597 return NULL; 598 } 599 600 /* Check sizes to catch unexpected coding error */ 601 if (mpc->rxbpre_datasize != rxq->datasize) { 602 netdev_err(ndev, "rxbpre_datasize mismatch: %u: %u\n", 603 mpc->rxbpre_datasize, rxq->datasize); 604 return NULL; 605 } 606 607 if (mpc->rxbpre_alloc_size != rxq->alloc_size) { 608 netdev_err(ndev, "rxbpre_alloc_size mismatch: %u: %u\n", 609 mpc->rxbpre_alloc_size, rxq->alloc_size); 610 return NULL; 611 } 612 613 if (mpc->rxbpre_headroom != rxq->headroom) { 614 netdev_err(ndev, "rxbpre_headroom mismatch: %u: %u\n", 615 mpc->rxbpre_headroom, rxq->headroom); 616 return NULL; 617 } 618 619 mpc->rxbpre_total--; 620 621 *da = mpc->das_pre[mpc->rxbpre_total]; 622 va = mpc->rxbufs_pre[mpc->rxbpre_total]; 623 mpc->rxbufs_pre[mpc->rxbpre_total] = NULL; 624 625 /* Deallocate the array after all buffers are gone */ 626 if (!mpc->rxbpre_total) 627 mana_pre_dealloc_rxbufs(mpc); 628 629 return va; 630 } 631 632 /* Get RX buffer's data size, alloc size, XDP headroom based on MTU */ 633 static void mana_get_rxbuf_cfg(int mtu, u32 *datasize, u32 *alloc_size, 634 u32 *headroom) 635 { 636 if (mtu > MANA_XDP_MTU_MAX) 637 *headroom = 0; /* no support for XDP */ 638 else 639 *headroom = XDP_PACKET_HEADROOM; 640 641 *alloc_size = SKB_DATA_ALIGN(mtu + MANA_RXBUF_PAD + *headroom); 642 643 /* Using page pool in this case, so alloc_size is PAGE_SIZE */ 644 if (*alloc_size < PAGE_SIZE) 645 *alloc_size = PAGE_SIZE; 646 647 *datasize = mtu + ETH_HLEN; 648 } 649 650 int mana_pre_alloc_rxbufs(struct mana_port_context *mpc, int new_mtu, int num_queues) 651 { 652 struct device *dev; 653 struct page *page; 654 dma_addr_t da; 655 int num_rxb; 656 void *va; 657 int i; 658 659 mana_get_rxbuf_cfg(new_mtu, &mpc->rxbpre_datasize, 660 &mpc->rxbpre_alloc_size, &mpc->rxbpre_headroom); 661 662 dev = mpc->ac->gdma_dev->gdma_context->dev; 663 664 num_rxb = num_queues * mpc->rx_queue_size; 665 666 WARN(mpc->rxbufs_pre, "mana rxbufs_pre exists\n"); 667 mpc->rxbufs_pre = kmalloc_array(num_rxb, sizeof(void *), GFP_KERNEL); 668 if (!mpc->rxbufs_pre) 669 goto error; 670 671 mpc->das_pre = kmalloc_array(num_rxb, sizeof(dma_addr_t), GFP_KERNEL); 672 if (!mpc->das_pre) 673 goto error; 674 675 mpc->rxbpre_total = 0; 676 677 for (i = 0; i < num_rxb; i++) { 678 page = dev_alloc_pages(get_order(mpc->rxbpre_alloc_size)); 679 if (!page) 680 goto error; 681 682 va = page_to_virt(page); 683 684 da = dma_map_single(dev, va + mpc->rxbpre_headroom, 685 mpc->rxbpre_datasize, DMA_FROM_DEVICE); 686 if (dma_mapping_error(dev, da)) { 687 put_page(page); 688 goto error; 689 } 690 691 mpc->rxbufs_pre[i] = va; 692 mpc->das_pre[i] = da; 693 mpc->rxbpre_total = i + 1; 694 } 695 696 return 0; 697 698 error: 699 netdev_err(mpc->ndev, "Failed to pre-allocate RX buffers for %d queues\n", num_queues); 700 mana_pre_dealloc_rxbufs(mpc); 701 return -ENOMEM; 702 } 703 704 static int mana_change_mtu(struct net_device *ndev, int new_mtu) 705 { 706 struct mana_port_context *mpc = netdev_priv(ndev); 707 unsigned int old_mtu = ndev->mtu; 708 int err; 709 710 /* Pre-allocate buffers to prevent failure in mana_attach later */ 711 err = mana_pre_alloc_rxbufs(mpc, new_mtu, mpc->num_queues); 712 if (err) { 713 netdev_err(ndev, "Insufficient memory for new MTU\n"); 714 return err; 715 } 716 717 err = mana_detach(ndev, false); 718 if (err) { 719 netdev_err(ndev, "mana_detach failed: %d\n", err); 720 goto out; 721 } 722 723 WRITE_ONCE(ndev->mtu, new_mtu); 724 725 err = mana_attach(ndev); 726 if (err) { 727 netdev_err(ndev, "mana_attach failed: %d\n", err); 728 WRITE_ONCE(ndev->mtu, old_mtu); 729 } 730 731 out: 732 mana_pre_dealloc_rxbufs(mpc); 733 return err; 734 } 735 736 static int mana_shaper_set(struct net_shaper_binding *binding, 737 const struct net_shaper *shaper, 738 struct netlink_ext_ack *extack) 739 { 740 struct mana_port_context *apc = netdev_priv(binding->netdev); 741 u32 old_speed, rate; 742 int err; 743 744 if (shaper->handle.scope != NET_SHAPER_SCOPE_NETDEV) { 745 NL_SET_ERR_MSG_MOD(extack, "net shaper scope should be netdev"); 746 return -EINVAL; 747 } 748 749 if (apc->handle.id && shaper->handle.id != apc->handle.id) { 750 NL_SET_ERR_MSG_MOD(extack, "Cannot create multiple shapers"); 751 return -EOPNOTSUPP; 752 } 753 754 if (!shaper->bw_max || (shaper->bw_max % 100000000)) { 755 NL_SET_ERR_MSG_MOD(extack, "Please use multiples of 100Mbps for bandwidth"); 756 return -EINVAL; 757 } 758 759 rate = div_u64(shaper->bw_max, 1000); /* Convert bps to Kbps */ 760 rate = div_u64(rate, 1000); /* Convert Kbps to Mbps */ 761 762 /* Get current speed */ 763 err = mana_query_link_cfg(apc); 764 old_speed = (err) ? SPEED_UNKNOWN : apc->speed; 765 766 if (!err) { 767 err = mana_set_bw_clamp(apc, rate, TRI_STATE_TRUE); 768 apc->speed = (err) ? old_speed : rate; 769 apc->handle = (err) ? apc->handle : shaper->handle; 770 } 771 772 return err; 773 } 774 775 static int mana_shaper_del(struct net_shaper_binding *binding, 776 const struct net_shaper_handle *handle, 777 struct netlink_ext_ack *extack) 778 { 779 struct mana_port_context *apc = netdev_priv(binding->netdev); 780 int err; 781 782 err = mana_set_bw_clamp(apc, 0, TRI_STATE_FALSE); 783 784 if (!err) { 785 /* Reset mana port context parameters */ 786 apc->handle.id = 0; 787 apc->handle.scope = NET_SHAPER_SCOPE_UNSPEC; 788 apc->speed = 0; 789 } 790 791 return err; 792 } 793 794 static void mana_shaper_cap(struct net_shaper_binding *binding, 795 enum net_shaper_scope scope, 796 unsigned long *flags) 797 { 798 *flags = BIT(NET_SHAPER_A_CAPS_SUPPORT_BW_MAX) | 799 BIT(NET_SHAPER_A_CAPS_SUPPORT_METRIC_BPS); 800 } 801 802 static const struct net_shaper_ops mana_shaper_ops = { 803 .set = mana_shaper_set, 804 .delete = mana_shaper_del, 805 .capabilities = mana_shaper_cap, 806 }; 807 808 static const struct net_device_ops mana_devops = { 809 .ndo_open = mana_open, 810 .ndo_stop = mana_close, 811 .ndo_select_queue = mana_select_queue, 812 .ndo_start_xmit = mana_start_xmit, 813 .ndo_validate_addr = eth_validate_addr, 814 .ndo_get_stats64 = mana_get_stats64, 815 .ndo_bpf = mana_bpf, 816 .ndo_xdp_xmit = mana_xdp_xmit, 817 .ndo_change_mtu = mana_change_mtu, 818 .net_shaper_ops = &mana_shaper_ops, 819 }; 820 821 static void mana_cleanup_port_context(struct mana_port_context *apc) 822 { 823 /* 824 * make sure subsequent cleanup attempts don't end up removing already 825 * cleaned dentry pointer 826 */ 827 debugfs_remove(apc->mana_port_debugfs); 828 apc->mana_port_debugfs = NULL; 829 kfree(apc->rxqs); 830 apc->rxqs = NULL; 831 } 832 833 static void mana_cleanup_indir_table(struct mana_port_context *apc) 834 { 835 apc->indir_table_sz = 0; 836 kfree(apc->indir_table); 837 kfree(apc->rxobj_table); 838 } 839 840 static int mana_init_port_context(struct mana_port_context *apc) 841 { 842 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *), 843 GFP_KERNEL); 844 845 return !apc->rxqs ? -ENOMEM : 0; 846 } 847 848 static int mana_send_request(struct mana_context *ac, void *in_buf, 849 u32 in_len, void *out_buf, u32 out_len) 850 { 851 struct gdma_context *gc = ac->gdma_dev->gdma_context; 852 struct gdma_resp_hdr *resp = out_buf; 853 struct gdma_req_hdr *req = in_buf; 854 struct device *dev = gc->dev; 855 static atomic_t activity_id; 856 int err; 857 858 req->dev_id = gc->mana.dev_id; 859 req->activity_id = atomic_inc_return(&activity_id); 860 861 err = mana_gd_send_request(gc, in_len, in_buf, out_len, 862 out_buf); 863 if (err || resp->status) { 864 if (err == -EOPNOTSUPP) 865 return err; 866 867 if (req->req.msg_type != MANA_QUERY_PHY_STAT && 868 mana_need_log(gc, err)) 869 dev_err(dev, "Failed to send mana message: %d, 0x%x\n", 870 err, resp->status); 871 return err ? err : -EPROTO; 872 } 873 874 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 || 875 req->activity_id != resp->activity_id) { 876 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n", 877 req->dev_id.as_uint32, resp->dev_id.as_uint32, 878 req->activity_id, resp->activity_id); 879 return -EPROTO; 880 } 881 882 return 0; 883 } 884 885 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr, 886 const enum mana_command_code expected_code, 887 const u32 min_size) 888 { 889 if (resp_hdr->response.msg_type != expected_code) 890 return -EPROTO; 891 892 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1) 893 return -EPROTO; 894 895 if (resp_hdr->response.msg_size < min_size) 896 return -EPROTO; 897 898 return 0; 899 } 900 901 static int mana_pf_register_hw_vport(struct mana_port_context *apc) 902 { 903 struct mana_register_hw_vport_resp resp = {}; 904 struct mana_register_hw_vport_req req = {}; 905 int err; 906 907 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_HW_PORT, 908 sizeof(req), sizeof(resp)); 909 req.attached_gfid = 1; 910 req.is_pf_default_vport = 1; 911 req.allow_all_ether_types = 1; 912 913 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 914 sizeof(resp)); 915 if (err) { 916 netdev_err(apc->ndev, "Failed to register hw vPort: %d\n", err); 917 return err; 918 } 919 920 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_HW_PORT, 921 sizeof(resp)); 922 if (err || resp.hdr.status) { 923 netdev_err(apc->ndev, "Failed to register hw vPort: %d, 0x%x\n", 924 err, resp.hdr.status); 925 return err ? err : -EPROTO; 926 } 927 928 apc->port_handle = resp.hw_vport_handle; 929 return 0; 930 } 931 932 static void mana_pf_deregister_hw_vport(struct mana_port_context *apc) 933 { 934 struct mana_deregister_hw_vport_resp resp = {}; 935 struct mana_deregister_hw_vport_req req = {}; 936 int err; 937 938 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_HW_PORT, 939 sizeof(req), sizeof(resp)); 940 req.hw_vport_handle = apc->port_handle; 941 942 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 943 sizeof(resp)); 944 if (err) { 945 if (mana_en_need_log(apc, err)) 946 netdev_err(apc->ndev, "Failed to unregister hw vPort: %d\n", 947 err); 948 949 return; 950 } 951 952 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_HW_PORT, 953 sizeof(resp)); 954 if (err || resp.hdr.status) 955 netdev_err(apc->ndev, 956 "Failed to deregister hw vPort: %d, 0x%x\n", 957 err, resp.hdr.status); 958 } 959 960 static int mana_pf_register_filter(struct mana_port_context *apc) 961 { 962 struct mana_register_filter_resp resp = {}; 963 struct mana_register_filter_req req = {}; 964 int err; 965 966 mana_gd_init_req_hdr(&req.hdr, MANA_REGISTER_FILTER, 967 sizeof(req), sizeof(resp)); 968 req.vport = apc->port_handle; 969 memcpy(req.mac_addr, apc->mac_addr, ETH_ALEN); 970 971 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 972 sizeof(resp)); 973 if (err) { 974 netdev_err(apc->ndev, "Failed to register filter: %d\n", err); 975 return err; 976 } 977 978 err = mana_verify_resp_hdr(&resp.hdr, MANA_REGISTER_FILTER, 979 sizeof(resp)); 980 if (err || resp.hdr.status) { 981 netdev_err(apc->ndev, "Failed to register filter: %d, 0x%x\n", 982 err, resp.hdr.status); 983 return err ? err : -EPROTO; 984 } 985 986 apc->pf_filter_handle = resp.filter_handle; 987 return 0; 988 } 989 990 static void mana_pf_deregister_filter(struct mana_port_context *apc) 991 { 992 struct mana_deregister_filter_resp resp = {}; 993 struct mana_deregister_filter_req req = {}; 994 int err; 995 996 mana_gd_init_req_hdr(&req.hdr, MANA_DEREGISTER_FILTER, 997 sizeof(req), sizeof(resp)); 998 req.filter_handle = apc->pf_filter_handle; 999 1000 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1001 sizeof(resp)); 1002 if (err) { 1003 if (mana_en_need_log(apc, err)) 1004 netdev_err(apc->ndev, "Failed to unregister filter: %d\n", 1005 err); 1006 1007 return; 1008 } 1009 1010 err = mana_verify_resp_hdr(&resp.hdr, MANA_DEREGISTER_FILTER, 1011 sizeof(resp)); 1012 if (err || resp.hdr.status) 1013 netdev_err(apc->ndev, 1014 "Failed to deregister filter: %d, 0x%x\n", 1015 err, resp.hdr.status); 1016 } 1017 1018 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver, 1019 u32 proto_minor_ver, u32 proto_micro_ver, 1020 u16 *max_num_vports, u8 *bm_hostmode) 1021 { 1022 struct gdma_context *gc = ac->gdma_dev->gdma_context; 1023 struct mana_query_device_cfg_resp resp = {}; 1024 struct mana_query_device_cfg_req req = {}; 1025 struct device *dev = gc->dev; 1026 int err = 0; 1027 1028 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG, 1029 sizeof(req), sizeof(resp)); 1030 1031 req.hdr.resp.msg_version = GDMA_MESSAGE_V3; 1032 1033 req.proto_major_ver = proto_major_ver; 1034 req.proto_minor_ver = proto_minor_ver; 1035 req.proto_micro_ver = proto_micro_ver; 1036 1037 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp)); 1038 if (err) { 1039 dev_err(dev, "Failed to query config: %d", err); 1040 return err; 1041 } 1042 1043 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG, 1044 sizeof(resp)); 1045 if (err || resp.hdr.status) { 1046 dev_err(dev, "Invalid query result: %d, 0x%x\n", err, 1047 resp.hdr.status); 1048 if (!err) 1049 err = -EPROTO; 1050 return err; 1051 } 1052 1053 *max_num_vports = resp.max_num_vports; 1054 1055 if (resp.hdr.response.msg_version >= GDMA_MESSAGE_V2) 1056 gc->adapter_mtu = resp.adapter_mtu; 1057 else 1058 gc->adapter_mtu = ETH_FRAME_LEN; 1059 1060 if (resp.hdr.response.msg_version >= GDMA_MESSAGE_V3) 1061 *bm_hostmode = resp.bm_hostmode; 1062 else 1063 *bm_hostmode = 0; 1064 1065 debugfs_create_u16("adapter-MTU", 0400, gc->mana_pci_debugfs, &gc->adapter_mtu); 1066 1067 return 0; 1068 } 1069 1070 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index, 1071 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry) 1072 { 1073 struct mana_query_vport_cfg_resp resp = {}; 1074 struct mana_query_vport_cfg_req req = {}; 1075 int err; 1076 1077 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG, 1078 sizeof(req), sizeof(resp)); 1079 1080 req.vport_index = vport_index; 1081 1082 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1083 sizeof(resp)); 1084 if (err) 1085 return err; 1086 1087 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG, 1088 sizeof(resp)); 1089 if (err) 1090 return err; 1091 1092 if (resp.hdr.status) 1093 return -EPROTO; 1094 1095 *max_sq = resp.max_num_sq; 1096 *max_rq = resp.max_num_rq; 1097 if (resp.num_indirection_ent > 0 && 1098 resp.num_indirection_ent <= MANA_INDIRECT_TABLE_MAX_SIZE && 1099 is_power_of_2(resp.num_indirection_ent)) { 1100 *num_indir_entry = resp.num_indirection_ent; 1101 } else { 1102 netdev_warn(apc->ndev, 1103 "Setting indirection table size to default %d for vPort %d\n", 1104 MANA_INDIRECT_TABLE_DEF_SIZE, apc->port_idx); 1105 *num_indir_entry = MANA_INDIRECT_TABLE_DEF_SIZE; 1106 } 1107 1108 apc->port_handle = resp.vport; 1109 ether_addr_copy(apc->mac_addr, resp.mac_addr); 1110 1111 return 0; 1112 } 1113 1114 void mana_uncfg_vport(struct mana_port_context *apc) 1115 { 1116 mutex_lock(&apc->vport_mutex); 1117 apc->vport_use_count--; 1118 WARN_ON(apc->vport_use_count < 0); 1119 mutex_unlock(&apc->vport_mutex); 1120 } 1121 EXPORT_SYMBOL_NS(mana_uncfg_vport, "NET_MANA"); 1122 1123 int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id, 1124 u32 doorbell_pg_id) 1125 { 1126 struct mana_config_vport_resp resp = {}; 1127 struct mana_config_vport_req req = {}; 1128 int err; 1129 1130 /* This function is used to program the Ethernet port in the hardware 1131 * table. It can be called from the Ethernet driver or the RDMA driver. 1132 * 1133 * For Ethernet usage, the hardware supports only one active user on a 1134 * physical port. The driver checks on the port usage before programming 1135 * the hardware when creating the RAW QP (RDMA driver) or exposing the 1136 * device to kernel NET layer (Ethernet driver). 1137 * 1138 * Because the RDMA driver doesn't know in advance which QP type the 1139 * user will create, it exposes the device with all its ports. The user 1140 * may not be able to create RAW QP on a port if this port is already 1141 * in used by the Ethernet driver from the kernel. 1142 * 1143 * This physical port limitation only applies to the RAW QP. For RC QP, 1144 * the hardware doesn't have this limitation. The user can create RC 1145 * QPs on a physical port up to the hardware limits independent of the 1146 * Ethernet usage on the same port. 1147 */ 1148 mutex_lock(&apc->vport_mutex); 1149 if (apc->vport_use_count > 0) { 1150 mutex_unlock(&apc->vport_mutex); 1151 return -EBUSY; 1152 } 1153 apc->vport_use_count++; 1154 mutex_unlock(&apc->vport_mutex); 1155 1156 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX, 1157 sizeof(req), sizeof(resp)); 1158 req.vport = apc->port_handle; 1159 req.pdid = protection_dom_id; 1160 req.doorbell_pageid = doorbell_pg_id; 1161 1162 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1163 sizeof(resp)); 1164 if (err) { 1165 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err); 1166 goto out; 1167 } 1168 1169 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX, 1170 sizeof(resp)); 1171 if (err || resp.hdr.status) { 1172 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n", 1173 err, resp.hdr.status); 1174 if (!err) 1175 err = -EPROTO; 1176 1177 goto out; 1178 } 1179 1180 apc->tx_shortform_allowed = resp.short_form_allowed; 1181 apc->tx_vp_offset = resp.tx_vport_offset; 1182 1183 netdev_info(apc->ndev, "Configured vPort %llu PD %u DB %u\n", 1184 apc->port_handle, protection_dom_id, doorbell_pg_id); 1185 out: 1186 if (err) 1187 mana_uncfg_vport(apc); 1188 1189 return err; 1190 } 1191 EXPORT_SYMBOL_NS(mana_cfg_vport, "NET_MANA"); 1192 1193 static int mana_cfg_vport_steering(struct mana_port_context *apc, 1194 enum TRI_STATE rx, 1195 bool update_default_rxobj, bool update_key, 1196 bool update_tab) 1197 { 1198 struct mana_cfg_rx_steer_req_v2 *req; 1199 struct mana_cfg_rx_steer_resp resp = {}; 1200 struct net_device *ndev = apc->ndev; 1201 u32 req_buf_size; 1202 int err; 1203 1204 req_buf_size = struct_size(req, indir_tab, apc->indir_table_sz); 1205 req = kzalloc(req_buf_size, GFP_KERNEL); 1206 if (!req) 1207 return -ENOMEM; 1208 1209 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size, 1210 sizeof(resp)); 1211 1212 req->hdr.req.msg_version = GDMA_MESSAGE_V2; 1213 1214 req->vport = apc->port_handle; 1215 req->num_indir_entries = apc->indir_table_sz; 1216 req->indir_tab_offset = offsetof(struct mana_cfg_rx_steer_req_v2, 1217 indir_tab); 1218 req->rx_enable = rx; 1219 req->rss_enable = apc->rss_state; 1220 req->update_default_rxobj = update_default_rxobj; 1221 req->update_hashkey = update_key; 1222 req->update_indir_tab = update_tab; 1223 req->default_rxobj = apc->default_rxobj; 1224 req->cqe_coalescing_enable = 0; 1225 1226 if (update_key) 1227 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE); 1228 1229 if (update_tab) 1230 memcpy(req->indir_tab, apc->rxobj_table, 1231 flex_array_size(req, indir_tab, req->num_indir_entries)); 1232 1233 err = mana_send_request(apc->ac, req, req_buf_size, &resp, 1234 sizeof(resp)); 1235 if (err) { 1236 if (mana_en_need_log(apc, err)) 1237 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err); 1238 1239 goto out; 1240 } 1241 1242 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX, 1243 sizeof(resp)); 1244 if (err) { 1245 netdev_err(ndev, "vPort RX configuration failed: %d\n", err); 1246 goto out; 1247 } 1248 1249 if (resp.hdr.status) { 1250 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n", 1251 resp.hdr.status); 1252 err = -EPROTO; 1253 } 1254 1255 netdev_info(ndev, "Configured steering vPort %llu entries %u\n", 1256 apc->port_handle, apc->indir_table_sz); 1257 out: 1258 kfree(req); 1259 return err; 1260 } 1261 1262 int mana_query_link_cfg(struct mana_port_context *apc) 1263 { 1264 struct net_device *ndev = apc->ndev; 1265 struct mana_query_link_config_resp resp = {}; 1266 struct mana_query_link_config_req req = {}; 1267 int err; 1268 1269 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_LINK_CONFIG, 1270 sizeof(req), sizeof(resp)); 1271 1272 req.vport = apc->port_handle; 1273 req.hdr.resp.msg_version = GDMA_MESSAGE_V2; 1274 1275 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1276 sizeof(resp)); 1277 1278 if (err) { 1279 if (err == -EOPNOTSUPP) { 1280 netdev_info_once(ndev, "MANA_QUERY_LINK_CONFIG not supported\n"); 1281 return err; 1282 } 1283 netdev_err(ndev, "Failed to query link config: %d\n", err); 1284 return err; 1285 } 1286 1287 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_LINK_CONFIG, 1288 sizeof(resp)); 1289 1290 if (err || resp.hdr.status) { 1291 netdev_err(ndev, "Failed to query link config: %d, 0x%x\n", err, 1292 resp.hdr.status); 1293 if (!err) 1294 err = -EOPNOTSUPP; 1295 return err; 1296 } 1297 1298 if (resp.qos_unconfigured) { 1299 err = -EINVAL; 1300 return err; 1301 } 1302 apc->speed = resp.link_speed_mbps; 1303 apc->max_speed = resp.qos_speed_mbps; 1304 return 0; 1305 } 1306 1307 int mana_set_bw_clamp(struct mana_port_context *apc, u32 speed, 1308 int enable_clamping) 1309 { 1310 struct mana_set_bw_clamp_resp resp = {}; 1311 struct mana_set_bw_clamp_req req = {}; 1312 struct net_device *ndev = apc->ndev; 1313 int err; 1314 1315 mana_gd_init_req_hdr(&req.hdr, MANA_SET_BW_CLAMP, 1316 sizeof(req), sizeof(resp)); 1317 req.vport = apc->port_handle; 1318 req.link_speed_mbps = speed; 1319 req.enable_clamping = enable_clamping; 1320 1321 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1322 sizeof(resp)); 1323 1324 if (err) { 1325 if (err == -EOPNOTSUPP) { 1326 netdev_info_once(ndev, "MANA_SET_BW_CLAMP not supported\n"); 1327 return err; 1328 } 1329 netdev_err(ndev, "Failed to set bandwidth clamp for speed %u, err = %d", 1330 speed, err); 1331 return err; 1332 } 1333 1334 err = mana_verify_resp_hdr(&resp.hdr, MANA_SET_BW_CLAMP, 1335 sizeof(resp)); 1336 1337 if (err || resp.hdr.status) { 1338 netdev_err(ndev, "Failed to set bandwidth clamp: %d, 0x%x\n", err, 1339 resp.hdr.status); 1340 if (!err) 1341 err = -EOPNOTSUPP; 1342 return err; 1343 } 1344 1345 if (resp.qos_unconfigured) 1346 netdev_info(ndev, "QoS is unconfigured\n"); 1347 1348 return 0; 1349 } 1350 1351 int mana_create_wq_obj(struct mana_port_context *apc, 1352 mana_handle_t vport, 1353 u32 wq_type, struct mana_obj_spec *wq_spec, 1354 struct mana_obj_spec *cq_spec, 1355 mana_handle_t *wq_obj) 1356 { 1357 struct mana_create_wqobj_resp resp = {}; 1358 struct mana_create_wqobj_req req = {}; 1359 struct net_device *ndev = apc->ndev; 1360 int err; 1361 1362 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ, 1363 sizeof(req), sizeof(resp)); 1364 req.vport = vport; 1365 req.wq_type = wq_type; 1366 req.wq_gdma_region = wq_spec->gdma_region; 1367 req.cq_gdma_region = cq_spec->gdma_region; 1368 req.wq_size = wq_spec->queue_size; 1369 req.cq_size = cq_spec->queue_size; 1370 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id; 1371 req.cq_parent_qid = cq_spec->attached_eq; 1372 1373 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1374 sizeof(resp)); 1375 if (err) { 1376 netdev_err(ndev, "Failed to create WQ object: %d\n", err); 1377 goto out; 1378 } 1379 1380 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ, 1381 sizeof(resp)); 1382 if (err || resp.hdr.status) { 1383 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err, 1384 resp.hdr.status); 1385 if (!err) 1386 err = -EPROTO; 1387 goto out; 1388 } 1389 1390 if (resp.wq_obj == INVALID_MANA_HANDLE) { 1391 netdev_err(ndev, "Got an invalid WQ object handle\n"); 1392 err = -EPROTO; 1393 goto out; 1394 } 1395 1396 *wq_obj = resp.wq_obj; 1397 wq_spec->queue_index = resp.wq_id; 1398 cq_spec->queue_index = resp.cq_id; 1399 1400 return 0; 1401 out: 1402 return err; 1403 } 1404 EXPORT_SYMBOL_NS(mana_create_wq_obj, "NET_MANA"); 1405 1406 void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type, 1407 mana_handle_t wq_obj) 1408 { 1409 struct mana_destroy_wqobj_resp resp = {}; 1410 struct mana_destroy_wqobj_req req = {}; 1411 struct net_device *ndev = apc->ndev; 1412 int err; 1413 1414 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ, 1415 sizeof(req), sizeof(resp)); 1416 req.wq_type = wq_type; 1417 req.wq_obj_handle = wq_obj; 1418 1419 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1420 sizeof(resp)); 1421 if (err) { 1422 if (mana_en_need_log(apc, err)) 1423 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err); 1424 1425 return; 1426 } 1427 1428 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ, 1429 sizeof(resp)); 1430 if (err || resp.hdr.status) 1431 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err, 1432 resp.hdr.status); 1433 } 1434 EXPORT_SYMBOL_NS(mana_destroy_wq_obj, "NET_MANA"); 1435 1436 static void mana_destroy_eq(struct mana_context *ac) 1437 { 1438 struct gdma_context *gc = ac->gdma_dev->gdma_context; 1439 struct gdma_queue *eq; 1440 int i; 1441 1442 if (!ac->eqs) 1443 return; 1444 1445 debugfs_remove_recursive(ac->mana_eqs_debugfs); 1446 ac->mana_eqs_debugfs = NULL; 1447 1448 for (i = 0; i < gc->max_num_queues; i++) { 1449 eq = ac->eqs[i].eq; 1450 if (!eq) 1451 continue; 1452 1453 mana_gd_destroy_queue(gc, eq); 1454 } 1455 1456 kfree(ac->eqs); 1457 ac->eqs = NULL; 1458 } 1459 1460 static void mana_create_eq_debugfs(struct mana_context *ac, int i) 1461 { 1462 struct mana_eq eq = ac->eqs[i]; 1463 char eqnum[32]; 1464 1465 sprintf(eqnum, "eq%d", i); 1466 eq.mana_eq_debugfs = debugfs_create_dir(eqnum, ac->mana_eqs_debugfs); 1467 debugfs_create_u32("head", 0400, eq.mana_eq_debugfs, &eq.eq->head); 1468 debugfs_create_u32("tail", 0400, eq.mana_eq_debugfs, &eq.eq->tail); 1469 debugfs_create_file("eq_dump", 0400, eq.mana_eq_debugfs, eq.eq, &mana_dbg_q_fops); 1470 } 1471 1472 static int mana_create_eq(struct mana_context *ac) 1473 { 1474 struct gdma_dev *gd = ac->gdma_dev; 1475 struct gdma_context *gc = gd->gdma_context; 1476 struct gdma_queue_spec spec = {}; 1477 int err; 1478 int i; 1479 1480 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq), 1481 GFP_KERNEL); 1482 if (!ac->eqs) 1483 return -ENOMEM; 1484 1485 spec.type = GDMA_EQ; 1486 spec.monitor_avl_buf = false; 1487 spec.queue_size = EQ_SIZE; 1488 spec.eq.callback = NULL; 1489 spec.eq.context = ac->eqs; 1490 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE; 1491 1492 ac->mana_eqs_debugfs = debugfs_create_dir("EQs", gc->mana_pci_debugfs); 1493 1494 for (i = 0; i < gc->max_num_queues; i++) { 1495 spec.eq.msix_index = (i + 1) % gc->num_msix_usable; 1496 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq); 1497 if (err) { 1498 dev_err(gc->dev, "Failed to create EQ %d : %d\n", i, err); 1499 goto out; 1500 } 1501 mana_create_eq_debugfs(ac, i); 1502 } 1503 1504 return 0; 1505 out: 1506 mana_destroy_eq(ac); 1507 return err; 1508 } 1509 1510 static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq) 1511 { 1512 struct mana_fence_rq_resp resp = {}; 1513 struct mana_fence_rq_req req = {}; 1514 int err; 1515 1516 init_completion(&rxq->fence_event); 1517 1518 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ, 1519 sizeof(req), sizeof(resp)); 1520 req.wq_obj_handle = rxq->rxobj; 1521 1522 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1523 sizeof(resp)); 1524 if (err) { 1525 netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n", 1526 rxq->rxq_idx, err); 1527 return err; 1528 } 1529 1530 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp)); 1531 if (err || resp.hdr.status) { 1532 netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n", 1533 rxq->rxq_idx, err, resp.hdr.status); 1534 if (!err) 1535 err = -EPROTO; 1536 1537 return err; 1538 } 1539 1540 if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) { 1541 netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n", 1542 rxq->rxq_idx); 1543 return -ETIMEDOUT; 1544 } 1545 1546 return 0; 1547 } 1548 1549 static void mana_fence_rqs(struct mana_port_context *apc) 1550 { 1551 unsigned int rxq_idx; 1552 struct mana_rxq *rxq; 1553 int err; 1554 1555 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 1556 rxq = apc->rxqs[rxq_idx]; 1557 err = mana_fence_rq(apc, rxq); 1558 1559 /* In case of any error, use sleep instead. */ 1560 if (err) 1561 msleep(100); 1562 } 1563 } 1564 1565 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units) 1566 { 1567 u32 used_space_old; 1568 u32 used_space_new; 1569 1570 used_space_old = wq->head - wq->tail; 1571 used_space_new = wq->head - (wq->tail + num_units); 1572 1573 if (WARN_ON_ONCE(used_space_new > used_space_old)) 1574 return -ERANGE; 1575 1576 wq->tail += num_units; 1577 return 0; 1578 } 1579 1580 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc) 1581 { 1582 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 1583 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 1584 struct device *dev = gc->dev; 1585 int hsg, i; 1586 1587 /* Number of SGEs of linear part */ 1588 hsg = (skb_is_gso(skb) && skb_headlen(skb) > ash->size[0]) ? 2 : 1; 1589 1590 for (i = 0; i < hsg; i++) 1591 dma_unmap_single(dev, ash->dma_handle[i], ash->size[i], 1592 DMA_TO_DEVICE); 1593 1594 for (i = hsg; i < skb_shinfo(skb)->nr_frags + hsg; i++) 1595 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i], 1596 DMA_TO_DEVICE); 1597 } 1598 1599 static void mana_poll_tx_cq(struct mana_cq *cq) 1600 { 1601 struct gdma_comp *completions = cq->gdma_comp_buf; 1602 struct gdma_posted_wqe_info *wqe_info; 1603 unsigned int pkt_transmitted = 0; 1604 unsigned int wqe_unit_cnt = 0; 1605 struct mana_txq *txq = cq->txq; 1606 struct mana_port_context *apc; 1607 struct netdev_queue *net_txq; 1608 struct gdma_queue *gdma_wq; 1609 unsigned int avail_space; 1610 struct net_device *ndev; 1611 struct sk_buff *skb; 1612 bool txq_stopped; 1613 int comp_read; 1614 int i; 1615 1616 ndev = txq->ndev; 1617 apc = netdev_priv(ndev); 1618 1619 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions, 1620 CQE_POLLING_BUFFER); 1621 1622 if (comp_read < 1) 1623 return; 1624 1625 for (i = 0; i < comp_read; i++) { 1626 struct mana_tx_comp_oob *cqe_oob; 1627 1628 if (WARN_ON_ONCE(!completions[i].is_sq)) 1629 return; 1630 1631 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data; 1632 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type != 1633 MANA_CQE_COMPLETION)) 1634 return; 1635 1636 switch (cqe_oob->cqe_hdr.cqe_type) { 1637 case CQE_TX_OKAY: 1638 break; 1639 1640 case CQE_TX_SA_DROP: 1641 case CQE_TX_MTU_DROP: 1642 case CQE_TX_INVALID_OOB: 1643 case CQE_TX_INVALID_ETH_TYPE: 1644 case CQE_TX_HDR_PROCESSING_ERROR: 1645 case CQE_TX_VF_DISABLED: 1646 case CQE_TX_VPORT_IDX_OUT_OF_RANGE: 1647 case CQE_TX_VPORT_DISABLED: 1648 case CQE_TX_VLAN_TAGGING_VIOLATION: 1649 if (net_ratelimit()) 1650 netdev_err(ndev, "TX: CQE error %d\n", 1651 cqe_oob->cqe_hdr.cqe_type); 1652 1653 apc->eth_stats.tx_cqe_err++; 1654 break; 1655 1656 default: 1657 /* If the CQE type is unknown, log an error, 1658 * and still free the SKB, update tail, etc. 1659 */ 1660 if (net_ratelimit()) 1661 netdev_err(ndev, "TX: unknown CQE type %d\n", 1662 cqe_oob->cqe_hdr.cqe_type); 1663 1664 apc->eth_stats.tx_cqe_unknown_type++; 1665 break; 1666 } 1667 1668 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num)) 1669 return; 1670 1671 skb = skb_dequeue(&txq->pending_skbs); 1672 if (WARN_ON_ONCE(!skb)) 1673 return; 1674 1675 wqe_info = (struct gdma_posted_wqe_info *)skb->cb; 1676 wqe_unit_cnt += wqe_info->wqe_size_in_bu; 1677 1678 mana_unmap_skb(skb, apc); 1679 1680 napi_consume_skb(skb, cq->budget); 1681 1682 pkt_transmitted++; 1683 } 1684 1685 if (WARN_ON_ONCE(wqe_unit_cnt == 0)) 1686 return; 1687 1688 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt); 1689 1690 gdma_wq = txq->gdma_sq; 1691 avail_space = mana_gd_wq_avail_space(gdma_wq); 1692 1693 /* Ensure tail updated before checking q stop */ 1694 smp_mb(); 1695 1696 net_txq = txq->net_txq; 1697 txq_stopped = netif_tx_queue_stopped(net_txq); 1698 1699 /* Ensure checking txq_stopped before apc->port_is_up. */ 1700 smp_rmb(); 1701 1702 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) { 1703 netif_tx_wake_queue(net_txq); 1704 apc->eth_stats.wake_queue++; 1705 } 1706 1707 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0) 1708 WARN_ON_ONCE(1); 1709 1710 cq->work_done = pkt_transmitted; 1711 } 1712 1713 static void mana_post_pkt_rxq(struct mana_rxq *rxq) 1714 { 1715 struct mana_recv_buf_oob *recv_buf_oob; 1716 u32 curr_index; 1717 int err; 1718 1719 curr_index = rxq->buf_index++; 1720 if (rxq->buf_index == rxq->num_rx_buf) 1721 rxq->buf_index = 0; 1722 1723 recv_buf_oob = &rxq->rx_oobs[curr_index]; 1724 1725 err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req, 1726 &recv_buf_oob->wqe_inf); 1727 if (WARN_ON_ONCE(err)) 1728 return; 1729 1730 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1); 1731 } 1732 1733 static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va, 1734 uint pkt_len, struct xdp_buff *xdp) 1735 { 1736 struct sk_buff *skb = napi_build_skb(buf_va, rxq->alloc_size); 1737 1738 if (!skb) 1739 return NULL; 1740 1741 if (xdp->data_hard_start) { 1742 u32 metasize = xdp->data - xdp->data_meta; 1743 1744 skb_reserve(skb, xdp->data - xdp->data_hard_start); 1745 skb_put(skb, xdp->data_end - xdp->data); 1746 if (metasize) 1747 skb_metadata_set(skb, metasize); 1748 return skb; 1749 } 1750 1751 skb_reserve(skb, rxq->headroom); 1752 skb_put(skb, pkt_len); 1753 1754 return skb; 1755 } 1756 1757 static void mana_rx_skb(void *buf_va, bool from_pool, 1758 struct mana_rxcomp_oob *cqe, struct mana_rxq *rxq) 1759 { 1760 struct mana_stats_rx *rx_stats = &rxq->stats; 1761 struct net_device *ndev = rxq->ndev; 1762 uint pkt_len = cqe->ppi[0].pkt_len; 1763 u16 rxq_idx = rxq->rxq_idx; 1764 struct napi_struct *napi; 1765 struct xdp_buff xdp = {}; 1766 struct sk_buff *skb; 1767 u32 hash_value; 1768 u32 act; 1769 1770 rxq->rx_cq.work_done++; 1771 napi = &rxq->rx_cq.napi; 1772 1773 if (!buf_va) { 1774 ++ndev->stats.rx_dropped; 1775 return; 1776 } 1777 1778 act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len); 1779 1780 if (act == XDP_REDIRECT && !rxq->xdp_rc) 1781 return; 1782 1783 if (act != XDP_PASS && act != XDP_TX) 1784 goto drop_xdp; 1785 1786 skb = mana_build_skb(rxq, buf_va, pkt_len, &xdp); 1787 1788 if (!skb) 1789 goto drop; 1790 1791 if (from_pool) 1792 skb_mark_for_recycle(skb); 1793 1794 skb->dev = napi->dev; 1795 1796 skb->protocol = eth_type_trans(skb, ndev); 1797 skb_checksum_none_assert(skb); 1798 skb_record_rx_queue(skb, rxq_idx); 1799 1800 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) { 1801 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed) 1802 skb->ip_summed = CHECKSUM_UNNECESSARY; 1803 } 1804 1805 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) { 1806 hash_value = cqe->ppi[0].pkt_hash; 1807 1808 if (cqe->rx_hashtype & MANA_HASH_L4) 1809 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4); 1810 else 1811 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3); 1812 } 1813 1814 if (cqe->rx_vlantag_present) { 1815 u16 vlan_tci = cqe->rx_vlan_id; 1816 1817 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci); 1818 } 1819 1820 u64_stats_update_begin(&rx_stats->syncp); 1821 rx_stats->packets++; 1822 rx_stats->bytes += pkt_len; 1823 1824 if (act == XDP_TX) 1825 rx_stats->xdp_tx++; 1826 u64_stats_update_end(&rx_stats->syncp); 1827 1828 if (act == XDP_TX) { 1829 skb_set_queue_mapping(skb, rxq_idx); 1830 mana_xdp_tx(skb, ndev); 1831 return; 1832 } 1833 1834 napi_gro_receive(napi, skb); 1835 1836 return; 1837 1838 drop_xdp: 1839 u64_stats_update_begin(&rx_stats->syncp); 1840 rx_stats->xdp_drop++; 1841 u64_stats_update_end(&rx_stats->syncp); 1842 1843 drop: 1844 if (from_pool) { 1845 page_pool_recycle_direct(rxq->page_pool, 1846 virt_to_head_page(buf_va)); 1847 } else { 1848 WARN_ON_ONCE(rxq->xdp_save_va); 1849 /* Save for reuse */ 1850 rxq->xdp_save_va = buf_va; 1851 } 1852 1853 ++ndev->stats.rx_dropped; 1854 1855 return; 1856 } 1857 1858 static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev, 1859 dma_addr_t *da, bool *from_pool) 1860 { 1861 struct page *page; 1862 void *va; 1863 1864 *from_pool = false; 1865 1866 /* Reuse XDP dropped page if available */ 1867 if (rxq->xdp_save_va) { 1868 va = rxq->xdp_save_va; 1869 rxq->xdp_save_va = NULL; 1870 } else { 1871 page = page_pool_dev_alloc_pages(rxq->page_pool); 1872 if (!page) 1873 return NULL; 1874 1875 *from_pool = true; 1876 va = page_to_virt(page); 1877 } 1878 1879 *da = dma_map_single(dev, va + rxq->headroom, rxq->datasize, 1880 DMA_FROM_DEVICE); 1881 if (dma_mapping_error(dev, *da)) { 1882 if (*from_pool) 1883 page_pool_put_full_page(rxq->page_pool, page, false); 1884 else 1885 put_page(virt_to_head_page(va)); 1886 1887 return NULL; 1888 } 1889 1890 return va; 1891 } 1892 1893 /* Allocate frag for rx buffer, and save the old buf */ 1894 static void mana_refill_rx_oob(struct device *dev, struct mana_rxq *rxq, 1895 struct mana_recv_buf_oob *rxoob, void **old_buf, 1896 bool *old_fp) 1897 { 1898 bool from_pool; 1899 dma_addr_t da; 1900 void *va; 1901 1902 va = mana_get_rxfrag(rxq, dev, &da, &from_pool); 1903 if (!va) 1904 return; 1905 1906 dma_unmap_single(dev, rxoob->sgl[0].address, rxq->datasize, 1907 DMA_FROM_DEVICE); 1908 *old_buf = rxoob->buf_va; 1909 *old_fp = rxoob->from_pool; 1910 1911 rxoob->buf_va = va; 1912 rxoob->sgl[0].address = da; 1913 rxoob->from_pool = from_pool; 1914 } 1915 1916 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq, 1917 struct gdma_comp *cqe) 1918 { 1919 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data; 1920 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context; 1921 struct net_device *ndev = rxq->ndev; 1922 struct mana_recv_buf_oob *rxbuf_oob; 1923 struct mana_port_context *apc; 1924 struct device *dev = gc->dev; 1925 void *old_buf = NULL; 1926 u32 curr, pktlen; 1927 bool old_fp; 1928 1929 apc = netdev_priv(ndev); 1930 1931 switch (oob->cqe_hdr.cqe_type) { 1932 case CQE_RX_OKAY: 1933 break; 1934 1935 case CQE_RX_TRUNCATED: 1936 ++ndev->stats.rx_dropped; 1937 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index]; 1938 netdev_warn_once(ndev, "Dropped a truncated packet\n"); 1939 goto drop; 1940 1941 case CQE_RX_COALESCED_4: 1942 netdev_err(ndev, "RX coalescing is unsupported\n"); 1943 apc->eth_stats.rx_coalesced_err++; 1944 return; 1945 1946 case CQE_RX_OBJECT_FENCE: 1947 complete(&rxq->fence_event); 1948 return; 1949 1950 default: 1951 netdev_err(ndev, "Unknown RX CQE type = %d\n", 1952 oob->cqe_hdr.cqe_type); 1953 apc->eth_stats.rx_cqe_unknown_type++; 1954 return; 1955 } 1956 1957 pktlen = oob->ppi[0].pkt_len; 1958 1959 if (pktlen == 0) { 1960 /* data packets should never have packetlength of zero */ 1961 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n", 1962 rxq->gdma_id, cq->gdma_id, rxq->rxobj); 1963 return; 1964 } 1965 1966 curr = rxq->buf_index; 1967 rxbuf_oob = &rxq->rx_oobs[curr]; 1968 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1); 1969 1970 mana_refill_rx_oob(dev, rxq, rxbuf_oob, &old_buf, &old_fp); 1971 1972 /* Unsuccessful refill will have old_buf == NULL. 1973 * In this case, mana_rx_skb() will drop the packet. 1974 */ 1975 mana_rx_skb(old_buf, old_fp, oob, rxq); 1976 1977 drop: 1978 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu); 1979 1980 mana_post_pkt_rxq(rxq); 1981 } 1982 1983 static void mana_poll_rx_cq(struct mana_cq *cq) 1984 { 1985 struct gdma_comp *comp = cq->gdma_comp_buf; 1986 struct mana_rxq *rxq = cq->rxq; 1987 int comp_read, i; 1988 1989 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER); 1990 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER); 1991 1992 rxq->xdp_flush = false; 1993 1994 for (i = 0; i < comp_read; i++) { 1995 if (WARN_ON_ONCE(comp[i].is_sq)) 1996 return; 1997 1998 /* verify recv cqe references the right rxq */ 1999 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id)) 2000 return; 2001 2002 mana_process_rx_cqe(rxq, cq, &comp[i]); 2003 } 2004 2005 if (comp_read > 0) { 2006 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context; 2007 2008 mana_gd_wq_ring_doorbell(gc, rxq->gdma_rq); 2009 } 2010 2011 if (rxq->xdp_flush) 2012 xdp_do_flush(); 2013 } 2014 2015 static int mana_cq_handler(void *context, struct gdma_queue *gdma_queue) 2016 { 2017 struct mana_cq *cq = context; 2018 int w; 2019 2020 WARN_ON_ONCE(cq->gdma_cq != gdma_queue); 2021 2022 if (cq->type == MANA_CQ_TYPE_RX) 2023 mana_poll_rx_cq(cq); 2024 else 2025 mana_poll_tx_cq(cq); 2026 2027 w = cq->work_done; 2028 cq->work_done_since_doorbell += w; 2029 2030 if (w < cq->budget) { 2031 mana_gd_ring_cq(gdma_queue, SET_ARM_BIT); 2032 cq->work_done_since_doorbell = 0; 2033 napi_complete_done(&cq->napi, w); 2034 } else if (cq->work_done_since_doorbell > 2035 cq->gdma_cq->queue_size / COMP_ENTRY_SIZE * 4) { 2036 /* MANA hardware requires at least one doorbell ring every 8 2037 * wraparounds of CQ even if there is no need to arm the CQ. 2038 * This driver rings the doorbell as soon as we have exceeded 2039 * 4 wraparounds. 2040 */ 2041 mana_gd_ring_cq(gdma_queue, 0); 2042 cq->work_done_since_doorbell = 0; 2043 } 2044 2045 return w; 2046 } 2047 2048 static int mana_poll(struct napi_struct *napi, int budget) 2049 { 2050 struct mana_cq *cq = container_of(napi, struct mana_cq, napi); 2051 int w; 2052 2053 cq->work_done = 0; 2054 cq->budget = budget; 2055 2056 w = mana_cq_handler(cq, cq->gdma_cq); 2057 2058 return min(w, budget); 2059 } 2060 2061 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue) 2062 { 2063 struct mana_cq *cq = context; 2064 2065 napi_schedule_irqoff(&cq->napi); 2066 } 2067 2068 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq) 2069 { 2070 struct gdma_dev *gd = apc->ac->gdma_dev; 2071 2072 if (!cq->gdma_cq) 2073 return; 2074 2075 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq); 2076 } 2077 2078 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq) 2079 { 2080 struct gdma_dev *gd = apc->ac->gdma_dev; 2081 2082 if (!txq->gdma_sq) 2083 return; 2084 2085 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq); 2086 } 2087 2088 static void mana_destroy_txq(struct mana_port_context *apc) 2089 { 2090 struct napi_struct *napi; 2091 int i; 2092 2093 if (!apc->tx_qp) 2094 return; 2095 2096 for (i = 0; i < apc->num_queues; i++) { 2097 debugfs_remove_recursive(apc->tx_qp[i].mana_tx_debugfs); 2098 apc->tx_qp[i].mana_tx_debugfs = NULL; 2099 2100 napi = &apc->tx_qp[i].tx_cq.napi; 2101 if (apc->tx_qp[i].txq.napi_initialized) { 2102 napi_synchronize(napi); 2103 netdev_lock_ops_to_full(napi->dev); 2104 napi_disable_locked(napi); 2105 netif_napi_del_locked(napi); 2106 netdev_unlock_full_to_ops(napi->dev); 2107 apc->tx_qp[i].txq.napi_initialized = false; 2108 } 2109 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object); 2110 2111 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq); 2112 2113 mana_deinit_txq(apc, &apc->tx_qp[i].txq); 2114 } 2115 2116 kfree(apc->tx_qp); 2117 apc->tx_qp = NULL; 2118 } 2119 2120 static void mana_create_txq_debugfs(struct mana_port_context *apc, int idx) 2121 { 2122 struct mana_tx_qp *tx_qp = &apc->tx_qp[idx]; 2123 char qnum[32]; 2124 2125 sprintf(qnum, "TX-%d", idx); 2126 tx_qp->mana_tx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs); 2127 debugfs_create_u32("sq_head", 0400, tx_qp->mana_tx_debugfs, 2128 &tx_qp->txq.gdma_sq->head); 2129 debugfs_create_u32("sq_tail", 0400, tx_qp->mana_tx_debugfs, 2130 &tx_qp->txq.gdma_sq->tail); 2131 debugfs_create_u32("sq_pend_skb_qlen", 0400, tx_qp->mana_tx_debugfs, 2132 &tx_qp->txq.pending_skbs.qlen); 2133 debugfs_create_u32("cq_head", 0400, tx_qp->mana_tx_debugfs, 2134 &tx_qp->tx_cq.gdma_cq->head); 2135 debugfs_create_u32("cq_tail", 0400, tx_qp->mana_tx_debugfs, 2136 &tx_qp->tx_cq.gdma_cq->tail); 2137 debugfs_create_u32("cq_budget", 0400, tx_qp->mana_tx_debugfs, 2138 &tx_qp->tx_cq.budget); 2139 debugfs_create_file("txq_dump", 0400, tx_qp->mana_tx_debugfs, 2140 tx_qp->txq.gdma_sq, &mana_dbg_q_fops); 2141 debugfs_create_file("cq_dump", 0400, tx_qp->mana_tx_debugfs, 2142 tx_qp->tx_cq.gdma_cq, &mana_dbg_q_fops); 2143 } 2144 2145 static int mana_create_txq(struct mana_port_context *apc, 2146 struct net_device *net) 2147 { 2148 struct mana_context *ac = apc->ac; 2149 struct gdma_dev *gd = ac->gdma_dev; 2150 struct mana_obj_spec wq_spec; 2151 struct mana_obj_spec cq_spec; 2152 struct gdma_queue_spec spec; 2153 struct gdma_context *gc; 2154 struct mana_txq *txq; 2155 struct mana_cq *cq; 2156 u32 txq_size; 2157 u32 cq_size; 2158 int err; 2159 int i; 2160 2161 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp), 2162 GFP_KERNEL); 2163 if (!apc->tx_qp) 2164 return -ENOMEM; 2165 2166 /* The minimum size of the WQE is 32 bytes, hence 2167 * apc->tx_queue_size represents the maximum number of WQEs 2168 * the SQ can store. This value is then used to size other queues 2169 * to prevent overflow. 2170 * Also note that the txq_size is always going to be MANA_PAGE_ALIGNED, 2171 * as min val of apc->tx_queue_size is 128 and that would make 2172 * txq_size 128*32 = 4096 and the other higher values of apc->tx_queue_size 2173 * are always power of two 2174 */ 2175 txq_size = apc->tx_queue_size * 32; 2176 2177 cq_size = apc->tx_queue_size * COMP_ENTRY_SIZE; 2178 2179 gc = gd->gdma_context; 2180 2181 for (i = 0; i < apc->num_queues; i++) { 2182 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE; 2183 2184 /* Create SQ */ 2185 txq = &apc->tx_qp[i].txq; 2186 2187 u64_stats_init(&txq->stats.syncp); 2188 txq->ndev = net; 2189 txq->net_txq = netdev_get_tx_queue(net, i); 2190 txq->vp_offset = apc->tx_vp_offset; 2191 txq->napi_initialized = false; 2192 skb_queue_head_init(&txq->pending_skbs); 2193 2194 memset(&spec, 0, sizeof(spec)); 2195 spec.type = GDMA_SQ; 2196 spec.monitor_avl_buf = true; 2197 spec.queue_size = txq_size; 2198 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq); 2199 if (err) 2200 goto out; 2201 2202 /* Create SQ's CQ */ 2203 cq = &apc->tx_qp[i].tx_cq; 2204 cq->type = MANA_CQ_TYPE_TX; 2205 2206 cq->txq = txq; 2207 2208 memset(&spec, 0, sizeof(spec)); 2209 spec.type = GDMA_CQ; 2210 spec.monitor_avl_buf = false; 2211 spec.queue_size = cq_size; 2212 spec.cq.callback = mana_schedule_napi; 2213 spec.cq.parent_eq = ac->eqs[i].eq; 2214 spec.cq.context = cq; 2215 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 2216 if (err) 2217 goto out; 2218 2219 memset(&wq_spec, 0, sizeof(wq_spec)); 2220 memset(&cq_spec, 0, sizeof(cq_spec)); 2221 2222 wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle; 2223 wq_spec.queue_size = txq->gdma_sq->queue_size; 2224 2225 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 2226 cq_spec.queue_size = cq->gdma_cq->queue_size; 2227 cq_spec.modr_ctx_id = 0; 2228 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 2229 2230 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ, 2231 &wq_spec, &cq_spec, 2232 &apc->tx_qp[i].tx_object); 2233 2234 if (err) 2235 goto out; 2236 2237 txq->gdma_sq->id = wq_spec.queue_index; 2238 cq->gdma_cq->id = cq_spec.queue_index; 2239 2240 txq->gdma_sq->mem_info.dma_region_handle = 2241 GDMA_INVALID_DMA_REGION; 2242 cq->gdma_cq->mem_info.dma_region_handle = 2243 GDMA_INVALID_DMA_REGION; 2244 2245 txq->gdma_txq_id = txq->gdma_sq->id; 2246 2247 cq->gdma_id = cq->gdma_cq->id; 2248 2249 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 2250 err = -EINVAL; 2251 goto out; 2252 } 2253 2254 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2255 2256 mana_create_txq_debugfs(apc, i); 2257 2258 set_bit(NAPI_STATE_NO_BUSY_POLL, &cq->napi.state); 2259 netdev_lock_ops_to_full(net); 2260 netif_napi_add_locked(net, &cq->napi, mana_poll); 2261 napi_enable_locked(&cq->napi); 2262 netdev_unlock_full_to_ops(net); 2263 txq->napi_initialized = true; 2264 2265 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2266 } 2267 2268 return 0; 2269 out: 2270 netdev_err(net, "Failed to create %d TX queues, %d\n", 2271 apc->num_queues, err); 2272 mana_destroy_txq(apc); 2273 return err; 2274 } 2275 2276 static void mana_destroy_rxq(struct mana_port_context *apc, 2277 struct mana_rxq *rxq, bool napi_initialized) 2278 2279 { 2280 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 2281 struct mana_recv_buf_oob *rx_oob; 2282 struct device *dev = gc->dev; 2283 struct napi_struct *napi; 2284 struct page *page; 2285 int i; 2286 2287 if (!rxq) 2288 return; 2289 2290 debugfs_remove_recursive(rxq->mana_rx_debugfs); 2291 rxq->mana_rx_debugfs = NULL; 2292 2293 napi = &rxq->rx_cq.napi; 2294 2295 if (napi_initialized) { 2296 napi_synchronize(napi); 2297 2298 netdev_lock_ops_to_full(napi->dev); 2299 napi_disable_locked(napi); 2300 netif_napi_del_locked(napi); 2301 netdev_unlock_full_to_ops(napi->dev); 2302 } 2303 xdp_rxq_info_unreg(&rxq->xdp_rxq); 2304 2305 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj); 2306 2307 mana_deinit_cq(apc, &rxq->rx_cq); 2308 2309 if (rxq->xdp_save_va) 2310 put_page(virt_to_head_page(rxq->xdp_save_va)); 2311 2312 for (i = 0; i < rxq->num_rx_buf; i++) { 2313 rx_oob = &rxq->rx_oobs[i]; 2314 2315 if (!rx_oob->buf_va) 2316 continue; 2317 2318 dma_unmap_single(dev, rx_oob->sgl[0].address, 2319 rx_oob->sgl[0].size, DMA_FROM_DEVICE); 2320 2321 page = virt_to_head_page(rx_oob->buf_va); 2322 2323 if (rx_oob->from_pool) 2324 page_pool_put_full_page(rxq->page_pool, page, false); 2325 else 2326 put_page(page); 2327 2328 rx_oob->buf_va = NULL; 2329 } 2330 2331 page_pool_destroy(rxq->page_pool); 2332 2333 if (rxq->gdma_rq) 2334 mana_gd_destroy_queue(gc, rxq->gdma_rq); 2335 2336 kfree(rxq); 2337 } 2338 2339 static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key, 2340 struct mana_rxq *rxq, struct device *dev) 2341 { 2342 struct mana_port_context *mpc = netdev_priv(rxq->ndev); 2343 bool from_pool = false; 2344 dma_addr_t da; 2345 void *va; 2346 2347 if (mpc->rxbufs_pre) 2348 va = mana_get_rxbuf_pre(rxq, &da); 2349 else 2350 va = mana_get_rxfrag(rxq, dev, &da, &from_pool); 2351 2352 if (!va) 2353 return -ENOMEM; 2354 2355 rx_oob->buf_va = va; 2356 rx_oob->from_pool = from_pool; 2357 2358 rx_oob->sgl[0].address = da; 2359 rx_oob->sgl[0].size = rxq->datasize; 2360 rx_oob->sgl[0].mem_key = mem_key; 2361 2362 return 0; 2363 } 2364 2365 #define MANA_WQE_HEADER_SIZE 16 2366 #define MANA_WQE_SGE_SIZE 16 2367 2368 static int mana_alloc_rx_wqe(struct mana_port_context *apc, 2369 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size) 2370 { 2371 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 2372 struct mana_recv_buf_oob *rx_oob; 2373 struct device *dev = gc->dev; 2374 u32 buf_idx; 2375 int ret; 2376 2377 WARN_ON(rxq->datasize == 0); 2378 2379 *rxq_size = 0; 2380 *cq_size = 0; 2381 2382 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2383 rx_oob = &rxq->rx_oobs[buf_idx]; 2384 memset(rx_oob, 0, sizeof(*rx_oob)); 2385 2386 rx_oob->num_sge = 1; 2387 2388 ret = mana_fill_rx_oob(rx_oob, apc->ac->gdma_dev->gpa_mkey, rxq, 2389 dev); 2390 if (ret) 2391 return ret; 2392 2393 rx_oob->wqe_req.sgl = rx_oob->sgl; 2394 rx_oob->wqe_req.num_sge = rx_oob->num_sge; 2395 rx_oob->wqe_req.inline_oob_size = 0; 2396 rx_oob->wqe_req.inline_oob_data = NULL; 2397 rx_oob->wqe_req.flags = 0; 2398 rx_oob->wqe_req.client_data_unit = 0; 2399 2400 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE + 2401 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32); 2402 *cq_size += COMP_ENTRY_SIZE; 2403 } 2404 2405 return 0; 2406 } 2407 2408 static int mana_push_wqe(struct mana_rxq *rxq) 2409 { 2410 struct mana_recv_buf_oob *rx_oob; 2411 u32 buf_idx; 2412 int err; 2413 2414 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2415 rx_oob = &rxq->rx_oobs[buf_idx]; 2416 2417 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req, 2418 &rx_oob->wqe_inf); 2419 if (err) 2420 return -ENOSPC; 2421 } 2422 2423 return 0; 2424 } 2425 2426 static int mana_create_page_pool(struct mana_rxq *rxq, struct gdma_context *gc) 2427 { 2428 struct mana_port_context *mpc = netdev_priv(rxq->ndev); 2429 struct page_pool_params pprm = {}; 2430 int ret; 2431 2432 pprm.pool_size = mpc->rx_queue_size; 2433 pprm.nid = gc->numa_node; 2434 pprm.napi = &rxq->rx_cq.napi; 2435 pprm.netdev = rxq->ndev; 2436 pprm.order = get_order(rxq->alloc_size); 2437 2438 rxq->page_pool = page_pool_create(&pprm); 2439 2440 if (IS_ERR(rxq->page_pool)) { 2441 ret = PTR_ERR(rxq->page_pool); 2442 rxq->page_pool = NULL; 2443 return ret; 2444 } 2445 2446 return 0; 2447 } 2448 2449 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc, 2450 u32 rxq_idx, struct mana_eq *eq, 2451 struct net_device *ndev) 2452 { 2453 struct gdma_dev *gd = apc->ac->gdma_dev; 2454 struct mana_obj_spec wq_spec; 2455 struct mana_obj_spec cq_spec; 2456 struct gdma_queue_spec spec; 2457 struct mana_cq *cq = NULL; 2458 struct gdma_context *gc; 2459 u32 cq_size, rq_size; 2460 struct mana_rxq *rxq; 2461 int err; 2462 2463 gc = gd->gdma_context; 2464 2465 rxq = kzalloc(struct_size(rxq, rx_oobs, apc->rx_queue_size), 2466 GFP_KERNEL); 2467 if (!rxq) 2468 return NULL; 2469 2470 rxq->ndev = ndev; 2471 rxq->num_rx_buf = apc->rx_queue_size; 2472 rxq->rxq_idx = rxq_idx; 2473 rxq->rxobj = INVALID_MANA_HANDLE; 2474 2475 mana_get_rxbuf_cfg(ndev->mtu, &rxq->datasize, &rxq->alloc_size, 2476 &rxq->headroom); 2477 2478 /* Create page pool for RX queue */ 2479 err = mana_create_page_pool(rxq, gc); 2480 if (err) { 2481 netdev_err(ndev, "Create page pool err:%d\n", err); 2482 goto out; 2483 } 2484 2485 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size); 2486 if (err) 2487 goto out; 2488 2489 rq_size = MANA_PAGE_ALIGN(rq_size); 2490 cq_size = MANA_PAGE_ALIGN(cq_size); 2491 2492 /* Create RQ */ 2493 memset(&spec, 0, sizeof(spec)); 2494 spec.type = GDMA_RQ; 2495 spec.monitor_avl_buf = true; 2496 spec.queue_size = rq_size; 2497 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq); 2498 if (err) 2499 goto out; 2500 2501 /* Create RQ's CQ */ 2502 cq = &rxq->rx_cq; 2503 cq->type = MANA_CQ_TYPE_RX; 2504 cq->rxq = rxq; 2505 2506 memset(&spec, 0, sizeof(spec)); 2507 spec.type = GDMA_CQ; 2508 spec.monitor_avl_buf = false; 2509 spec.queue_size = cq_size; 2510 spec.cq.callback = mana_schedule_napi; 2511 spec.cq.parent_eq = eq->eq; 2512 spec.cq.context = cq; 2513 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 2514 if (err) 2515 goto out; 2516 2517 memset(&wq_spec, 0, sizeof(wq_spec)); 2518 memset(&cq_spec, 0, sizeof(cq_spec)); 2519 wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle; 2520 wq_spec.queue_size = rxq->gdma_rq->queue_size; 2521 2522 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 2523 cq_spec.queue_size = cq->gdma_cq->queue_size; 2524 cq_spec.modr_ctx_id = 0; 2525 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 2526 2527 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ, 2528 &wq_spec, &cq_spec, &rxq->rxobj); 2529 if (err) 2530 goto out; 2531 2532 rxq->gdma_rq->id = wq_spec.queue_index; 2533 cq->gdma_cq->id = cq_spec.queue_index; 2534 2535 rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2536 cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2537 2538 rxq->gdma_id = rxq->gdma_rq->id; 2539 cq->gdma_id = cq->gdma_cq->id; 2540 2541 err = mana_push_wqe(rxq); 2542 if (err) 2543 goto out; 2544 2545 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 2546 err = -EINVAL; 2547 goto out; 2548 } 2549 2550 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2551 2552 netdev_lock_ops_to_full(ndev); 2553 netif_napi_add_weight_locked(ndev, &cq->napi, mana_poll, 1); 2554 netdev_unlock_full_to_ops(ndev); 2555 2556 WARN_ON(xdp_rxq_info_reg(&rxq->xdp_rxq, ndev, rxq_idx, 2557 cq->napi.napi_id)); 2558 WARN_ON(xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL, 2559 rxq->page_pool)); 2560 2561 netdev_lock_ops_to_full(ndev); 2562 napi_enable_locked(&cq->napi); 2563 netdev_unlock_full_to_ops(ndev); 2564 2565 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2566 out: 2567 if (!err) 2568 return rxq; 2569 2570 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err); 2571 2572 mana_destroy_rxq(apc, rxq, false); 2573 2574 if (cq) 2575 mana_deinit_cq(apc, cq); 2576 2577 return NULL; 2578 } 2579 2580 static void mana_create_rxq_debugfs(struct mana_port_context *apc, int idx) 2581 { 2582 struct mana_rxq *rxq; 2583 char qnum[32]; 2584 2585 rxq = apc->rxqs[idx]; 2586 2587 sprintf(qnum, "RX-%d", idx); 2588 rxq->mana_rx_debugfs = debugfs_create_dir(qnum, apc->mana_port_debugfs); 2589 debugfs_create_u32("rq_head", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->head); 2590 debugfs_create_u32("rq_tail", 0400, rxq->mana_rx_debugfs, &rxq->gdma_rq->tail); 2591 debugfs_create_u32("rq_nbuf", 0400, rxq->mana_rx_debugfs, &rxq->num_rx_buf); 2592 debugfs_create_u32("cq_head", 0400, rxq->mana_rx_debugfs, 2593 &rxq->rx_cq.gdma_cq->head); 2594 debugfs_create_u32("cq_tail", 0400, rxq->mana_rx_debugfs, 2595 &rxq->rx_cq.gdma_cq->tail); 2596 debugfs_create_u32("cq_budget", 0400, rxq->mana_rx_debugfs, &rxq->rx_cq.budget); 2597 debugfs_create_file("rxq_dump", 0400, rxq->mana_rx_debugfs, rxq->gdma_rq, &mana_dbg_q_fops); 2598 debugfs_create_file("cq_dump", 0400, rxq->mana_rx_debugfs, rxq->rx_cq.gdma_cq, 2599 &mana_dbg_q_fops); 2600 } 2601 2602 static int mana_add_rx_queues(struct mana_port_context *apc, 2603 struct net_device *ndev) 2604 { 2605 struct mana_context *ac = apc->ac; 2606 struct mana_rxq *rxq; 2607 int err = 0; 2608 int i; 2609 2610 for (i = 0; i < apc->num_queues; i++) { 2611 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev); 2612 if (!rxq) { 2613 err = -ENOMEM; 2614 netdev_err(ndev, "Failed to create rxq %d : %d\n", i, err); 2615 goto out; 2616 } 2617 2618 u64_stats_init(&rxq->stats.syncp); 2619 2620 apc->rxqs[i] = rxq; 2621 2622 mana_create_rxq_debugfs(apc, i); 2623 } 2624 2625 apc->default_rxobj = apc->rxqs[0]->rxobj; 2626 out: 2627 return err; 2628 } 2629 2630 static void mana_destroy_vport(struct mana_port_context *apc) 2631 { 2632 struct gdma_dev *gd = apc->ac->gdma_dev; 2633 struct mana_rxq *rxq; 2634 u32 rxq_idx; 2635 2636 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 2637 rxq = apc->rxqs[rxq_idx]; 2638 if (!rxq) 2639 continue; 2640 2641 mana_destroy_rxq(apc, rxq, true); 2642 apc->rxqs[rxq_idx] = NULL; 2643 } 2644 2645 mana_destroy_txq(apc); 2646 mana_uncfg_vport(apc); 2647 2648 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) 2649 mana_pf_deregister_hw_vport(apc); 2650 } 2651 2652 static int mana_create_vport(struct mana_port_context *apc, 2653 struct net_device *net) 2654 { 2655 struct gdma_dev *gd = apc->ac->gdma_dev; 2656 int err; 2657 2658 apc->default_rxobj = INVALID_MANA_HANDLE; 2659 2660 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) { 2661 err = mana_pf_register_hw_vport(apc); 2662 if (err) 2663 return err; 2664 } 2665 2666 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell); 2667 if (err) 2668 return err; 2669 2670 return mana_create_txq(apc, net); 2671 } 2672 2673 static int mana_rss_table_alloc(struct mana_port_context *apc) 2674 { 2675 if (!apc->indir_table_sz) { 2676 netdev_err(apc->ndev, 2677 "Indirection table size not set for vPort %d\n", 2678 apc->port_idx); 2679 return -EINVAL; 2680 } 2681 2682 apc->indir_table = kcalloc(apc->indir_table_sz, sizeof(u32), GFP_KERNEL); 2683 if (!apc->indir_table) 2684 return -ENOMEM; 2685 2686 apc->rxobj_table = kcalloc(apc->indir_table_sz, sizeof(mana_handle_t), GFP_KERNEL); 2687 if (!apc->rxobj_table) { 2688 kfree(apc->indir_table); 2689 return -ENOMEM; 2690 } 2691 2692 return 0; 2693 } 2694 2695 static void mana_rss_table_init(struct mana_port_context *apc) 2696 { 2697 int i; 2698 2699 for (i = 0; i < apc->indir_table_sz; i++) 2700 apc->indir_table[i] = 2701 ethtool_rxfh_indir_default(i, apc->num_queues); 2702 } 2703 2704 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx, 2705 bool update_hash, bool update_tab) 2706 { 2707 u32 queue_idx; 2708 int err; 2709 int i; 2710 2711 if (update_tab) { 2712 for (i = 0; i < apc->indir_table_sz; i++) { 2713 queue_idx = apc->indir_table[i]; 2714 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj; 2715 } 2716 } 2717 2718 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab); 2719 if (err) 2720 return err; 2721 2722 mana_fence_rqs(apc); 2723 2724 return 0; 2725 } 2726 2727 void mana_query_gf_stats(struct mana_port_context *apc) 2728 { 2729 struct mana_query_gf_stat_resp resp = {}; 2730 struct mana_query_gf_stat_req req = {}; 2731 struct net_device *ndev = apc->ndev; 2732 int err; 2733 2734 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_GF_STAT, 2735 sizeof(req), sizeof(resp)); 2736 req.hdr.resp.msg_version = GDMA_MESSAGE_V2; 2737 req.req_stats = STATISTICS_FLAGS_RX_DISCARDS_NO_WQE | 2738 STATISTICS_FLAGS_RX_ERRORS_VPORT_DISABLED | 2739 STATISTICS_FLAGS_HC_RX_BYTES | 2740 STATISTICS_FLAGS_HC_RX_UCAST_PACKETS | 2741 STATISTICS_FLAGS_HC_RX_UCAST_BYTES | 2742 STATISTICS_FLAGS_HC_RX_MCAST_PACKETS | 2743 STATISTICS_FLAGS_HC_RX_MCAST_BYTES | 2744 STATISTICS_FLAGS_HC_RX_BCAST_PACKETS | 2745 STATISTICS_FLAGS_HC_RX_BCAST_BYTES | 2746 STATISTICS_FLAGS_TX_ERRORS_GF_DISABLED | 2747 STATISTICS_FLAGS_TX_ERRORS_VPORT_DISABLED | 2748 STATISTICS_FLAGS_TX_ERRORS_INVAL_VPORT_OFFSET_PACKETS | 2749 STATISTICS_FLAGS_TX_ERRORS_VLAN_ENFORCEMENT | 2750 STATISTICS_FLAGS_TX_ERRORS_ETH_TYPE_ENFORCEMENT | 2751 STATISTICS_FLAGS_TX_ERRORS_SA_ENFORCEMENT | 2752 STATISTICS_FLAGS_TX_ERRORS_SQPDID_ENFORCEMENT | 2753 STATISTICS_FLAGS_TX_ERRORS_CQPDID_ENFORCEMENT | 2754 STATISTICS_FLAGS_TX_ERRORS_MTU_VIOLATION | 2755 STATISTICS_FLAGS_TX_ERRORS_INVALID_OOB | 2756 STATISTICS_FLAGS_HC_TX_BYTES | 2757 STATISTICS_FLAGS_HC_TX_UCAST_PACKETS | 2758 STATISTICS_FLAGS_HC_TX_UCAST_BYTES | 2759 STATISTICS_FLAGS_HC_TX_MCAST_PACKETS | 2760 STATISTICS_FLAGS_HC_TX_MCAST_BYTES | 2761 STATISTICS_FLAGS_HC_TX_BCAST_PACKETS | 2762 STATISTICS_FLAGS_HC_TX_BCAST_BYTES | 2763 STATISTICS_FLAGS_TX_ERRORS_GDMA_ERROR; 2764 2765 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 2766 sizeof(resp)); 2767 if (err) { 2768 netdev_err(ndev, "Failed to query GF stats: %d\n", err); 2769 return; 2770 } 2771 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_GF_STAT, 2772 sizeof(resp)); 2773 if (err || resp.hdr.status) { 2774 netdev_err(ndev, "Failed to query GF stats: %d, 0x%x\n", err, 2775 resp.hdr.status); 2776 return; 2777 } 2778 2779 apc->eth_stats.hc_rx_discards_no_wqe = resp.rx_discards_nowqe; 2780 apc->eth_stats.hc_rx_err_vport_disabled = resp.rx_err_vport_disabled; 2781 apc->eth_stats.hc_rx_bytes = resp.hc_rx_bytes; 2782 apc->eth_stats.hc_rx_ucast_pkts = resp.hc_rx_ucast_pkts; 2783 apc->eth_stats.hc_rx_ucast_bytes = resp.hc_rx_ucast_bytes; 2784 apc->eth_stats.hc_rx_bcast_pkts = resp.hc_rx_bcast_pkts; 2785 apc->eth_stats.hc_rx_bcast_bytes = resp.hc_rx_bcast_bytes; 2786 apc->eth_stats.hc_rx_mcast_pkts = resp.hc_rx_mcast_pkts; 2787 apc->eth_stats.hc_rx_mcast_bytes = resp.hc_rx_mcast_bytes; 2788 apc->eth_stats.hc_tx_err_gf_disabled = resp.tx_err_gf_disabled; 2789 apc->eth_stats.hc_tx_err_vport_disabled = resp.tx_err_vport_disabled; 2790 apc->eth_stats.hc_tx_err_inval_vportoffset_pkt = 2791 resp.tx_err_inval_vport_offset_pkt; 2792 apc->eth_stats.hc_tx_err_vlan_enforcement = 2793 resp.tx_err_vlan_enforcement; 2794 apc->eth_stats.hc_tx_err_eth_type_enforcement = 2795 resp.tx_err_ethtype_enforcement; 2796 apc->eth_stats.hc_tx_err_sa_enforcement = resp.tx_err_SA_enforcement; 2797 apc->eth_stats.hc_tx_err_sqpdid_enforcement = 2798 resp.tx_err_SQPDID_enforcement; 2799 apc->eth_stats.hc_tx_err_cqpdid_enforcement = 2800 resp.tx_err_CQPDID_enforcement; 2801 apc->eth_stats.hc_tx_err_mtu_violation = resp.tx_err_mtu_violation; 2802 apc->eth_stats.hc_tx_err_inval_oob = resp.tx_err_inval_oob; 2803 apc->eth_stats.hc_tx_bytes = resp.hc_tx_bytes; 2804 apc->eth_stats.hc_tx_ucast_pkts = resp.hc_tx_ucast_pkts; 2805 apc->eth_stats.hc_tx_ucast_bytes = resp.hc_tx_ucast_bytes; 2806 apc->eth_stats.hc_tx_bcast_pkts = resp.hc_tx_bcast_pkts; 2807 apc->eth_stats.hc_tx_bcast_bytes = resp.hc_tx_bcast_bytes; 2808 apc->eth_stats.hc_tx_mcast_pkts = resp.hc_tx_mcast_pkts; 2809 apc->eth_stats.hc_tx_mcast_bytes = resp.hc_tx_mcast_bytes; 2810 apc->eth_stats.hc_tx_err_gdma = resp.tx_err_gdma; 2811 } 2812 2813 void mana_query_phy_stats(struct mana_port_context *apc) 2814 { 2815 struct mana_query_phy_stat_resp resp = {}; 2816 struct mana_query_phy_stat_req req = {}; 2817 struct net_device *ndev = apc->ndev; 2818 int err; 2819 2820 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_PHY_STAT, 2821 sizeof(req), sizeof(resp)); 2822 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 2823 sizeof(resp)); 2824 if (err) 2825 return; 2826 2827 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_PHY_STAT, 2828 sizeof(resp)); 2829 if (err || resp.hdr.status) { 2830 netdev_err(ndev, 2831 "Failed to query PHY stats: %d, resp:0x%x\n", 2832 err, resp.hdr.status); 2833 return; 2834 } 2835 2836 /* Aggregate drop counters */ 2837 apc->phy_stats.rx_pkt_drop_phy = resp.rx_pkt_drop_phy; 2838 apc->phy_stats.tx_pkt_drop_phy = resp.tx_pkt_drop_phy; 2839 2840 /* Per TC traffic Counters */ 2841 apc->phy_stats.rx_pkt_tc0_phy = resp.rx_pkt_tc0_phy; 2842 apc->phy_stats.tx_pkt_tc0_phy = resp.tx_pkt_tc0_phy; 2843 apc->phy_stats.rx_pkt_tc1_phy = resp.rx_pkt_tc1_phy; 2844 apc->phy_stats.tx_pkt_tc1_phy = resp.tx_pkt_tc1_phy; 2845 apc->phy_stats.rx_pkt_tc2_phy = resp.rx_pkt_tc2_phy; 2846 apc->phy_stats.tx_pkt_tc2_phy = resp.tx_pkt_tc2_phy; 2847 apc->phy_stats.rx_pkt_tc3_phy = resp.rx_pkt_tc3_phy; 2848 apc->phy_stats.tx_pkt_tc3_phy = resp.tx_pkt_tc3_phy; 2849 apc->phy_stats.rx_pkt_tc4_phy = resp.rx_pkt_tc4_phy; 2850 apc->phy_stats.tx_pkt_tc4_phy = resp.tx_pkt_tc4_phy; 2851 apc->phy_stats.rx_pkt_tc5_phy = resp.rx_pkt_tc5_phy; 2852 apc->phy_stats.tx_pkt_tc5_phy = resp.tx_pkt_tc5_phy; 2853 apc->phy_stats.rx_pkt_tc6_phy = resp.rx_pkt_tc6_phy; 2854 apc->phy_stats.tx_pkt_tc6_phy = resp.tx_pkt_tc6_phy; 2855 apc->phy_stats.rx_pkt_tc7_phy = resp.rx_pkt_tc7_phy; 2856 apc->phy_stats.tx_pkt_tc7_phy = resp.tx_pkt_tc7_phy; 2857 2858 /* Per TC byte Counters */ 2859 apc->phy_stats.rx_byte_tc0_phy = resp.rx_byte_tc0_phy; 2860 apc->phy_stats.tx_byte_tc0_phy = resp.tx_byte_tc0_phy; 2861 apc->phy_stats.rx_byte_tc1_phy = resp.rx_byte_tc1_phy; 2862 apc->phy_stats.tx_byte_tc1_phy = resp.tx_byte_tc1_phy; 2863 apc->phy_stats.rx_byte_tc2_phy = resp.rx_byte_tc2_phy; 2864 apc->phy_stats.tx_byte_tc2_phy = resp.tx_byte_tc2_phy; 2865 apc->phy_stats.rx_byte_tc3_phy = resp.rx_byte_tc3_phy; 2866 apc->phy_stats.tx_byte_tc3_phy = resp.tx_byte_tc3_phy; 2867 apc->phy_stats.rx_byte_tc4_phy = resp.rx_byte_tc4_phy; 2868 apc->phy_stats.tx_byte_tc4_phy = resp.tx_byte_tc4_phy; 2869 apc->phy_stats.rx_byte_tc5_phy = resp.rx_byte_tc5_phy; 2870 apc->phy_stats.tx_byte_tc5_phy = resp.tx_byte_tc5_phy; 2871 apc->phy_stats.rx_byte_tc6_phy = resp.rx_byte_tc6_phy; 2872 apc->phy_stats.tx_byte_tc6_phy = resp.tx_byte_tc6_phy; 2873 apc->phy_stats.rx_byte_tc7_phy = resp.rx_byte_tc7_phy; 2874 apc->phy_stats.tx_byte_tc7_phy = resp.tx_byte_tc7_phy; 2875 2876 /* Per TC pause Counters */ 2877 apc->phy_stats.rx_pause_tc0_phy = resp.rx_pause_tc0_phy; 2878 apc->phy_stats.tx_pause_tc0_phy = resp.tx_pause_tc0_phy; 2879 apc->phy_stats.rx_pause_tc1_phy = resp.rx_pause_tc1_phy; 2880 apc->phy_stats.tx_pause_tc1_phy = resp.tx_pause_tc1_phy; 2881 apc->phy_stats.rx_pause_tc2_phy = resp.rx_pause_tc2_phy; 2882 apc->phy_stats.tx_pause_tc2_phy = resp.tx_pause_tc2_phy; 2883 apc->phy_stats.rx_pause_tc3_phy = resp.rx_pause_tc3_phy; 2884 apc->phy_stats.tx_pause_tc3_phy = resp.tx_pause_tc3_phy; 2885 apc->phy_stats.rx_pause_tc4_phy = resp.rx_pause_tc4_phy; 2886 apc->phy_stats.tx_pause_tc4_phy = resp.tx_pause_tc4_phy; 2887 apc->phy_stats.rx_pause_tc5_phy = resp.rx_pause_tc5_phy; 2888 apc->phy_stats.tx_pause_tc5_phy = resp.tx_pause_tc5_phy; 2889 apc->phy_stats.rx_pause_tc6_phy = resp.rx_pause_tc6_phy; 2890 apc->phy_stats.tx_pause_tc6_phy = resp.tx_pause_tc6_phy; 2891 apc->phy_stats.rx_pause_tc7_phy = resp.rx_pause_tc7_phy; 2892 apc->phy_stats.tx_pause_tc7_phy = resp.tx_pause_tc7_phy; 2893 } 2894 2895 static int mana_init_port(struct net_device *ndev) 2896 { 2897 struct mana_port_context *apc = netdev_priv(ndev); 2898 struct gdma_dev *gd = apc->ac->gdma_dev; 2899 u32 max_txq, max_rxq, max_queues; 2900 int port_idx = apc->port_idx; 2901 struct gdma_context *gc; 2902 char vport[32]; 2903 int err; 2904 2905 err = mana_init_port_context(apc); 2906 if (err) 2907 return err; 2908 2909 gc = gd->gdma_context; 2910 2911 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq, 2912 &apc->indir_table_sz); 2913 if (err) { 2914 netdev_err(ndev, "Failed to query info for vPort %d\n", 2915 port_idx); 2916 goto reset_apc; 2917 } 2918 2919 max_queues = min_t(u32, max_txq, max_rxq); 2920 if (apc->max_queues > max_queues) 2921 apc->max_queues = max_queues; 2922 2923 if (apc->num_queues > apc->max_queues) 2924 apc->num_queues = apc->max_queues; 2925 2926 eth_hw_addr_set(ndev, apc->mac_addr); 2927 sprintf(vport, "vport%d", port_idx); 2928 apc->mana_port_debugfs = debugfs_create_dir(vport, gc->mana_pci_debugfs); 2929 return 0; 2930 2931 reset_apc: 2932 mana_cleanup_port_context(apc); 2933 return err; 2934 } 2935 2936 int mana_alloc_queues(struct net_device *ndev) 2937 { 2938 struct mana_port_context *apc = netdev_priv(ndev); 2939 struct gdma_dev *gd = apc->ac->gdma_dev; 2940 int err; 2941 2942 err = mana_create_vport(apc, ndev); 2943 if (err) { 2944 netdev_err(ndev, "Failed to create vPort %u : %d\n", apc->port_idx, err); 2945 return err; 2946 } 2947 2948 err = netif_set_real_num_tx_queues(ndev, apc->num_queues); 2949 if (err) { 2950 netdev_err(ndev, 2951 "netif_set_real_num_tx_queues () failed for ndev with num_queues %u : %d\n", 2952 apc->num_queues, err); 2953 goto destroy_vport; 2954 } 2955 2956 err = mana_add_rx_queues(apc, ndev); 2957 if (err) 2958 goto destroy_vport; 2959 2960 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE; 2961 2962 err = netif_set_real_num_rx_queues(ndev, apc->num_queues); 2963 if (err) { 2964 netdev_err(ndev, 2965 "netif_set_real_num_rx_queues () failed for ndev with num_queues %u : %d\n", 2966 apc->num_queues, err); 2967 goto destroy_vport; 2968 } 2969 2970 mana_rss_table_init(apc); 2971 2972 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true); 2973 if (err) { 2974 netdev_err(ndev, "Failed to configure RSS table: %d\n", err); 2975 goto destroy_vport; 2976 } 2977 2978 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) { 2979 err = mana_pf_register_filter(apc); 2980 if (err) 2981 goto destroy_vport; 2982 } 2983 2984 mana_chn_setxdp(apc, mana_xdp_get(apc)); 2985 2986 return 0; 2987 2988 destroy_vport: 2989 mana_destroy_vport(apc); 2990 return err; 2991 } 2992 2993 int mana_attach(struct net_device *ndev) 2994 { 2995 struct mana_port_context *apc = netdev_priv(ndev); 2996 int err; 2997 2998 ASSERT_RTNL(); 2999 3000 err = mana_init_port(ndev); 3001 if (err) 3002 return err; 3003 3004 if (apc->port_st_save) { 3005 err = mana_alloc_queues(ndev); 3006 if (err) { 3007 mana_cleanup_port_context(apc); 3008 return err; 3009 } 3010 } 3011 3012 apc->port_is_up = apc->port_st_save; 3013 3014 /* Ensure port state updated before txq state */ 3015 smp_wmb(); 3016 3017 if (apc->port_is_up) 3018 netif_carrier_on(ndev); 3019 3020 netif_device_attach(ndev); 3021 3022 return 0; 3023 } 3024 3025 static int mana_dealloc_queues(struct net_device *ndev) 3026 { 3027 struct mana_port_context *apc = netdev_priv(ndev); 3028 unsigned long timeout = jiffies + 120 * HZ; 3029 struct gdma_dev *gd = apc->ac->gdma_dev; 3030 struct mana_txq *txq; 3031 struct sk_buff *skb; 3032 int i, err; 3033 u32 tsleep; 3034 3035 if (apc->port_is_up) 3036 return -EINVAL; 3037 3038 mana_chn_setxdp(apc, NULL); 3039 3040 if (gd->gdma_context->is_pf && !apc->ac->bm_hostmode) 3041 mana_pf_deregister_filter(apc); 3042 3043 /* No packet can be transmitted now since apc->port_is_up is false. 3044 * There is still a tiny chance that mana_poll_tx_cq() can re-enable 3045 * a txq because it may not timely see apc->port_is_up being cleared 3046 * to false, but it doesn't matter since mana_start_xmit() drops any 3047 * new packets due to apc->port_is_up being false. 3048 * 3049 * Drain all the in-flight TX packets. 3050 * A timeout of 120 seconds for all the queues is used. 3051 * This will break the while loop when h/w is not responding. 3052 * This value of 120 has been decided here considering max 3053 * number of queues. 3054 */ 3055 3056 for (i = 0; i < apc->num_queues; i++) { 3057 txq = &apc->tx_qp[i].txq; 3058 tsleep = 1000; 3059 while (atomic_read(&txq->pending_sends) > 0 && 3060 time_before(jiffies, timeout)) { 3061 usleep_range(tsleep, tsleep + 1000); 3062 tsleep <<= 1; 3063 } 3064 if (atomic_read(&txq->pending_sends)) { 3065 err = pcie_flr(to_pci_dev(gd->gdma_context->dev)); 3066 if (err) { 3067 netdev_err(ndev, "flr failed %d with %d pkts pending in txq %u\n", 3068 err, atomic_read(&txq->pending_sends), 3069 txq->gdma_txq_id); 3070 } 3071 break; 3072 } 3073 } 3074 3075 for (i = 0; i < apc->num_queues; i++) { 3076 txq = &apc->tx_qp[i].txq; 3077 while ((skb = skb_dequeue(&txq->pending_skbs))) { 3078 mana_unmap_skb(skb, apc); 3079 dev_kfree_skb_any(skb); 3080 } 3081 atomic_set(&txq->pending_sends, 0); 3082 } 3083 /* We're 100% sure the queues can no longer be woken up, because 3084 * we're sure now mana_poll_tx_cq() can't be running. 3085 */ 3086 3087 apc->rss_state = TRI_STATE_FALSE; 3088 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false); 3089 if (err && mana_en_need_log(apc, err)) 3090 netdev_err(ndev, "Failed to disable vPort: %d\n", err); 3091 3092 /* Even in err case, still need to cleanup the vPort */ 3093 mana_destroy_vport(apc); 3094 3095 return 0; 3096 } 3097 3098 int mana_detach(struct net_device *ndev, bool from_close) 3099 { 3100 struct mana_port_context *apc = netdev_priv(ndev); 3101 int err; 3102 3103 ASSERT_RTNL(); 3104 3105 apc->port_st_save = apc->port_is_up; 3106 apc->port_is_up = false; 3107 3108 /* Ensure port state updated before txq state */ 3109 smp_wmb(); 3110 3111 netif_tx_disable(ndev); 3112 netif_carrier_off(ndev); 3113 3114 if (apc->port_st_save) { 3115 err = mana_dealloc_queues(ndev); 3116 if (err) { 3117 netdev_err(ndev, "%s failed to deallocate queues: %d\n", __func__, err); 3118 return err; 3119 } 3120 } 3121 3122 if (!from_close) { 3123 netif_device_detach(ndev); 3124 mana_cleanup_port_context(apc); 3125 } 3126 3127 return 0; 3128 } 3129 3130 static int mana_probe_port(struct mana_context *ac, int port_idx, 3131 struct net_device **ndev_storage) 3132 { 3133 struct gdma_context *gc = ac->gdma_dev->gdma_context; 3134 struct mana_port_context *apc; 3135 struct net_device *ndev; 3136 int err; 3137 3138 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context), 3139 gc->max_num_queues); 3140 if (!ndev) 3141 return -ENOMEM; 3142 3143 *ndev_storage = ndev; 3144 3145 apc = netdev_priv(ndev); 3146 apc->ac = ac; 3147 apc->ndev = ndev; 3148 apc->max_queues = gc->max_num_queues; 3149 apc->num_queues = gc->max_num_queues; 3150 apc->tx_queue_size = DEF_TX_BUFFERS_PER_QUEUE; 3151 apc->rx_queue_size = DEF_RX_BUFFERS_PER_QUEUE; 3152 apc->port_handle = INVALID_MANA_HANDLE; 3153 apc->pf_filter_handle = INVALID_MANA_HANDLE; 3154 apc->port_idx = port_idx; 3155 3156 mutex_init(&apc->vport_mutex); 3157 apc->vport_use_count = 0; 3158 3159 ndev->netdev_ops = &mana_devops; 3160 ndev->ethtool_ops = &mana_ethtool_ops; 3161 ndev->mtu = ETH_DATA_LEN; 3162 ndev->max_mtu = gc->adapter_mtu - ETH_HLEN; 3163 ndev->min_mtu = ETH_MIN_MTU; 3164 ndev->needed_headroom = MANA_HEADROOM; 3165 ndev->dev_port = port_idx; 3166 SET_NETDEV_DEV(ndev, gc->dev); 3167 3168 netif_set_tso_max_size(ndev, GSO_MAX_SIZE); 3169 3170 netif_carrier_off(ndev); 3171 3172 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE); 3173 3174 err = mana_init_port(ndev); 3175 if (err) 3176 goto free_net; 3177 3178 err = mana_rss_table_alloc(apc); 3179 if (err) 3180 goto reset_apc; 3181 3182 netdev_lockdep_set_classes(ndev); 3183 3184 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 3185 ndev->hw_features |= NETIF_F_RXCSUM; 3186 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6; 3187 ndev->hw_features |= NETIF_F_RXHASH; 3188 ndev->features = ndev->hw_features | NETIF_F_HW_VLAN_CTAG_TX | 3189 NETIF_F_HW_VLAN_CTAG_RX; 3190 ndev->vlan_features = ndev->features; 3191 xdp_set_features_flag(ndev, NETDEV_XDP_ACT_BASIC | 3192 NETDEV_XDP_ACT_REDIRECT | 3193 NETDEV_XDP_ACT_NDO_XMIT); 3194 3195 err = register_netdev(ndev); 3196 if (err) { 3197 netdev_err(ndev, "Unable to register netdev.\n"); 3198 goto free_indir; 3199 } 3200 3201 debugfs_create_u32("current_speed", 0400, apc->mana_port_debugfs, &apc->speed); 3202 3203 return 0; 3204 3205 free_indir: 3206 mana_cleanup_indir_table(apc); 3207 reset_apc: 3208 mana_cleanup_port_context(apc); 3209 free_net: 3210 *ndev_storage = NULL; 3211 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err); 3212 free_netdev(ndev); 3213 return err; 3214 } 3215 3216 static void adev_release(struct device *dev) 3217 { 3218 struct mana_adev *madev = container_of(dev, struct mana_adev, adev.dev); 3219 3220 kfree(madev); 3221 } 3222 3223 static void remove_adev(struct gdma_dev *gd) 3224 { 3225 struct auxiliary_device *adev = gd->adev; 3226 int id = adev->id; 3227 3228 auxiliary_device_delete(adev); 3229 auxiliary_device_uninit(adev); 3230 3231 mana_adev_idx_free(id); 3232 gd->adev = NULL; 3233 } 3234 3235 static int add_adev(struct gdma_dev *gd, const char *name) 3236 { 3237 struct auxiliary_device *adev; 3238 struct mana_adev *madev; 3239 int ret; 3240 3241 madev = kzalloc(sizeof(*madev), GFP_KERNEL); 3242 if (!madev) 3243 return -ENOMEM; 3244 3245 adev = &madev->adev; 3246 ret = mana_adev_idx_alloc(); 3247 if (ret < 0) 3248 goto idx_fail; 3249 adev->id = ret; 3250 3251 adev->name = name; 3252 adev->dev.parent = gd->gdma_context->dev; 3253 adev->dev.release = adev_release; 3254 madev->mdev = gd; 3255 3256 ret = auxiliary_device_init(adev); 3257 if (ret) 3258 goto init_fail; 3259 3260 /* madev is owned by the auxiliary device */ 3261 madev = NULL; 3262 ret = auxiliary_device_add(adev); 3263 if (ret) 3264 goto add_fail; 3265 3266 gd->adev = adev; 3267 dev_dbg(gd->gdma_context->dev, 3268 "Auxiliary device added successfully\n"); 3269 return 0; 3270 3271 add_fail: 3272 auxiliary_device_uninit(adev); 3273 3274 init_fail: 3275 mana_adev_idx_free(adev->id); 3276 3277 idx_fail: 3278 kfree(madev); 3279 3280 return ret; 3281 } 3282 3283 static void mana_rdma_service_handle(struct work_struct *work) 3284 { 3285 struct mana_service_work *serv_work = 3286 container_of(work, struct mana_service_work, work); 3287 struct gdma_dev *gd = serv_work->gdma_dev; 3288 struct device *dev = gd->gdma_context->dev; 3289 int ret; 3290 3291 if (READ_ONCE(gd->rdma_teardown)) 3292 goto out; 3293 3294 switch (serv_work->event) { 3295 case GDMA_SERVICE_TYPE_RDMA_SUSPEND: 3296 if (!gd->adev || gd->is_suspended) 3297 break; 3298 3299 remove_adev(gd); 3300 gd->is_suspended = true; 3301 break; 3302 3303 case GDMA_SERVICE_TYPE_RDMA_RESUME: 3304 if (!gd->is_suspended) 3305 break; 3306 3307 ret = add_adev(gd, "rdma"); 3308 if (ret) 3309 dev_err(dev, "Failed to add adev on resume: %d\n", ret); 3310 else 3311 gd->is_suspended = false; 3312 break; 3313 3314 default: 3315 dev_warn(dev, "unknown adev service event %u\n", 3316 serv_work->event); 3317 break; 3318 } 3319 3320 out: 3321 kfree(serv_work); 3322 } 3323 3324 int mana_rdma_service_event(struct gdma_context *gc, enum gdma_service_type event) 3325 { 3326 struct gdma_dev *gd = &gc->mana_ib; 3327 struct mana_service_work *serv_work; 3328 3329 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) { 3330 /* RDMA device is not detected on pci */ 3331 return 0; 3332 } 3333 3334 serv_work = kzalloc(sizeof(*serv_work), GFP_ATOMIC); 3335 if (!serv_work) 3336 return -ENOMEM; 3337 3338 serv_work->event = event; 3339 serv_work->gdma_dev = gd; 3340 3341 INIT_WORK(&serv_work->work, mana_rdma_service_handle); 3342 queue_work(gc->service_wq, &serv_work->work); 3343 3344 return 0; 3345 } 3346 3347 int mana_probe(struct gdma_dev *gd, bool resuming) 3348 { 3349 struct gdma_context *gc = gd->gdma_context; 3350 struct mana_context *ac = gd->driver_data; 3351 struct device *dev = gc->dev; 3352 u8 bm_hostmode = 0; 3353 u16 num_ports = 0; 3354 int err; 3355 int i; 3356 3357 dev_info(dev, 3358 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n", 3359 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION); 3360 3361 err = mana_gd_register_device(gd); 3362 if (err) 3363 return err; 3364 3365 if (!resuming) { 3366 ac = kzalloc(sizeof(*ac), GFP_KERNEL); 3367 if (!ac) 3368 return -ENOMEM; 3369 3370 ac->gdma_dev = gd; 3371 gd->driver_data = ac; 3372 } 3373 3374 err = mana_create_eq(ac); 3375 if (err) { 3376 dev_err(dev, "Failed to create EQs: %d\n", err); 3377 goto out; 3378 } 3379 3380 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION, 3381 MANA_MICRO_VERSION, &num_ports, &bm_hostmode); 3382 if (err) 3383 goto out; 3384 3385 ac->bm_hostmode = bm_hostmode; 3386 3387 if (!resuming) { 3388 ac->num_ports = num_ports; 3389 } else { 3390 if (ac->num_ports != num_ports) { 3391 dev_err(dev, "The number of vPorts changed: %d->%d\n", 3392 ac->num_ports, num_ports); 3393 err = -EPROTO; 3394 goto out; 3395 } 3396 } 3397 3398 if (ac->num_ports == 0) 3399 dev_err(dev, "Failed to detect any vPort\n"); 3400 3401 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV) 3402 ac->num_ports = MAX_PORTS_IN_MANA_DEV; 3403 3404 if (!resuming) { 3405 for (i = 0; i < ac->num_ports; i++) { 3406 err = mana_probe_port(ac, i, &ac->ports[i]); 3407 /* we log the port for which the probe failed and stop 3408 * probes for subsequent ports. 3409 * Note that we keep running ports, for which the probes 3410 * were successful, unless add_adev fails too 3411 */ 3412 if (err) { 3413 dev_err(dev, "Probe Failed for port %d\n", i); 3414 break; 3415 } 3416 } 3417 } else { 3418 for (i = 0; i < ac->num_ports; i++) { 3419 rtnl_lock(); 3420 err = mana_attach(ac->ports[i]); 3421 rtnl_unlock(); 3422 /* we log the port for which the attach failed and stop 3423 * attach for subsequent ports 3424 * Note that we keep running ports, for which the attach 3425 * were successful, unless add_adev fails too 3426 */ 3427 if (err) { 3428 dev_err(dev, "Attach Failed for port %d\n", i); 3429 break; 3430 } 3431 } 3432 } 3433 3434 err = add_adev(gd, "eth"); 3435 out: 3436 if (err) { 3437 mana_remove(gd, false); 3438 } else { 3439 dev_dbg(dev, "gd=%p, id=%u, num_ports=%d, type=%u, instance=%u\n", 3440 gd, gd->dev_id.as_uint32, ac->num_ports, 3441 gd->dev_id.type, gd->dev_id.instance); 3442 dev_dbg(dev, "%s succeeded\n", __func__); 3443 } 3444 3445 return err; 3446 } 3447 3448 void mana_remove(struct gdma_dev *gd, bool suspending) 3449 { 3450 struct gdma_context *gc = gd->gdma_context; 3451 struct mana_context *ac = gd->driver_data; 3452 struct mana_port_context *apc; 3453 struct device *dev = gc->dev; 3454 struct net_device *ndev; 3455 int err; 3456 int i; 3457 3458 /* adev currently doesn't support suspending, always remove it */ 3459 if (gd->adev) 3460 remove_adev(gd); 3461 3462 for (i = 0; i < ac->num_ports; i++) { 3463 ndev = ac->ports[i]; 3464 apc = netdev_priv(ndev); 3465 if (!ndev) { 3466 if (i == 0) 3467 dev_err(dev, "No net device to remove\n"); 3468 goto out; 3469 } 3470 3471 /* All cleanup actions should stay after rtnl_lock(), otherwise 3472 * other functions may access partially cleaned up data. 3473 */ 3474 rtnl_lock(); 3475 3476 err = mana_detach(ndev, false); 3477 if (err) 3478 netdev_err(ndev, "Failed to detach vPort %d: %d\n", 3479 i, err); 3480 3481 if (suspending) { 3482 /* No need to unregister the ndev. */ 3483 rtnl_unlock(); 3484 continue; 3485 } 3486 3487 unregister_netdevice(ndev); 3488 mana_cleanup_indir_table(apc); 3489 3490 rtnl_unlock(); 3491 3492 free_netdev(ndev); 3493 } 3494 3495 mana_destroy_eq(ac); 3496 out: 3497 mana_gd_deregister_device(gd); 3498 3499 if (suspending) 3500 return; 3501 3502 gd->driver_data = NULL; 3503 gd->gdma_context = NULL; 3504 kfree(ac); 3505 dev_dbg(dev, "%s succeeded\n", __func__); 3506 } 3507 3508 int mana_rdma_probe(struct gdma_dev *gd) 3509 { 3510 int err = 0; 3511 3512 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) { 3513 /* RDMA device is not detected on pci */ 3514 return err; 3515 } 3516 3517 err = mana_gd_register_device(gd); 3518 if (err) 3519 return err; 3520 3521 err = add_adev(gd, "rdma"); 3522 if (err) 3523 mana_gd_deregister_device(gd); 3524 3525 return err; 3526 } 3527 3528 void mana_rdma_remove(struct gdma_dev *gd) 3529 { 3530 struct gdma_context *gc = gd->gdma_context; 3531 3532 if (gd->dev_id.type != GDMA_DEVICE_MANA_IB) { 3533 /* RDMA device is not detected on pci */ 3534 return; 3535 } 3536 3537 WRITE_ONCE(gd->rdma_teardown, true); 3538 flush_workqueue(gc->service_wq); 3539 3540 if (gd->adev) 3541 remove_adev(gd); 3542 3543 mana_gd_deregister_device(gd); 3544 } 3545 3546 struct net_device *mana_get_primary_netdev(struct mana_context *ac, 3547 u32 port_index, 3548 netdevice_tracker *tracker) 3549 { 3550 struct net_device *ndev; 3551 3552 if (port_index >= ac->num_ports) 3553 return NULL; 3554 3555 rcu_read_lock(); 3556 3557 /* If mana is used in netvsc, the upper netdevice should be returned. */ 3558 ndev = netdev_master_upper_dev_get_rcu(ac->ports[port_index]); 3559 3560 /* If there is no upper device, use the parent Ethernet device */ 3561 if (!ndev) 3562 ndev = ac->ports[port_index]; 3563 3564 netdev_hold(ndev, tracker, GFP_ATOMIC); 3565 rcu_read_unlock(); 3566 3567 return ndev; 3568 } 3569 EXPORT_SYMBOL_NS(mana_get_primary_netdev, "NET_MANA"); 3570