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