1 /* 2 * Virtual network driver for conversing with remote driver backends. 3 * 4 * Copyright (c) 2002-2005, K A Fraser 5 * Copyright (c) 2005, XenSource Ltd 6 * 7 * This program is free software; you can redistribute it and/or 8 * modify it under the terms of the GNU General Public License version 2 9 * as published by the Free Software Foundation; or, when distributed 10 * separately from the Linux kernel or incorporated into other 11 * software packages, subject to the following license: 12 * 13 * Permission is hereby granted, free of charge, to any person obtaining a copy 14 * of this source file (the "Software"), to deal in the Software without 15 * restriction, including without limitation the rights to use, copy, modify, 16 * merge, publish, distribute, sublicense, and/or sell copies of the Software, 17 * and to permit persons to whom the Software is furnished to do so, subject to 18 * the following conditions: 19 * 20 * The above copyright notice and this permission notice shall be included in 21 * all copies or substantial portions of the Software. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 24 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 25 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 26 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 27 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 28 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 29 * IN THE SOFTWARE. 30 */ 31 32 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 33 34 #include <linux/module.h> 35 #include <linux/kernel.h> 36 #include <linux/netdevice.h> 37 #include <linux/etherdevice.h> 38 #include <linux/skbuff.h> 39 #include <linux/ethtool.h> 40 #include <linux/if_ether.h> 41 #include <net/tcp.h> 42 #include <linux/udp.h> 43 #include <linux/moduleparam.h> 44 #include <linux/mm.h> 45 #include <linux/slab.h> 46 #include <net/ip.h> 47 #include <linux/bpf.h> 48 #include <net/page_pool.h> 49 #include <linux/bpf_trace.h> 50 51 #include <xen/xen.h> 52 #include <xen/xenbus.h> 53 #include <xen/events.h> 54 #include <xen/page.h> 55 #include <xen/platform_pci.h> 56 #include <xen/grant_table.h> 57 58 #include <xen/interface/io/netif.h> 59 #include <xen/interface/memory.h> 60 #include <xen/interface/grant_table.h> 61 62 /* Module parameters */ 63 #define MAX_QUEUES_DEFAULT 8 64 static unsigned int xennet_max_queues; 65 module_param_named(max_queues, xennet_max_queues, uint, 0644); 66 MODULE_PARM_DESC(max_queues, 67 "Maximum number of queues per virtual interface"); 68 69 #define XENNET_TIMEOUT (5 * HZ) 70 71 static const struct ethtool_ops xennet_ethtool_ops; 72 73 struct netfront_cb { 74 int pull_to; 75 }; 76 77 #define NETFRONT_SKB_CB(skb) ((struct netfront_cb *)((skb)->cb)) 78 79 #define RX_COPY_THRESHOLD 256 80 81 #define GRANT_INVALID_REF 0 82 83 #define NET_TX_RING_SIZE __CONST_RING_SIZE(xen_netif_tx, XEN_PAGE_SIZE) 84 #define NET_RX_RING_SIZE __CONST_RING_SIZE(xen_netif_rx, XEN_PAGE_SIZE) 85 86 /* Minimum number of Rx slots (includes slot for GSO metadata). */ 87 #define NET_RX_SLOTS_MIN (XEN_NETIF_NR_SLOTS_MIN + 1) 88 89 /* Queue name is interface name with "-qNNN" appended */ 90 #define QUEUE_NAME_SIZE (IFNAMSIZ + 6) 91 92 /* IRQ name is queue name with "-tx" or "-rx" appended */ 93 #define IRQ_NAME_SIZE (QUEUE_NAME_SIZE + 3) 94 95 static DECLARE_WAIT_QUEUE_HEAD(module_wq); 96 97 struct netfront_stats { 98 u64 packets; 99 u64 bytes; 100 struct u64_stats_sync syncp; 101 }; 102 103 struct netfront_info; 104 105 struct netfront_queue { 106 unsigned int id; /* Queue ID, 0-based */ 107 char name[QUEUE_NAME_SIZE]; /* DEVNAME-qN */ 108 struct netfront_info *info; 109 110 struct bpf_prog __rcu *xdp_prog; 111 112 struct napi_struct napi; 113 114 /* Split event channels support, tx_* == rx_* when using 115 * single event channel. 116 */ 117 unsigned int tx_evtchn, rx_evtchn; 118 unsigned int tx_irq, rx_irq; 119 /* Only used when split event channels support is enabled */ 120 char tx_irq_name[IRQ_NAME_SIZE]; /* DEVNAME-qN-tx */ 121 char rx_irq_name[IRQ_NAME_SIZE]; /* DEVNAME-qN-rx */ 122 123 spinlock_t tx_lock; 124 struct xen_netif_tx_front_ring tx; 125 int tx_ring_ref; 126 127 /* 128 * {tx,rx}_skbs store outstanding skbuffs. Free tx_skb entries 129 * are linked from tx_skb_freelist through tx_link. 130 */ 131 struct sk_buff *tx_skbs[NET_TX_RING_SIZE]; 132 unsigned short tx_link[NET_TX_RING_SIZE]; 133 #define TX_LINK_NONE 0xffff 134 #define TX_PENDING 0xfffe 135 grant_ref_t gref_tx_head; 136 grant_ref_t grant_tx_ref[NET_TX_RING_SIZE]; 137 struct page *grant_tx_page[NET_TX_RING_SIZE]; 138 unsigned tx_skb_freelist; 139 unsigned int tx_pend_queue; 140 141 spinlock_t rx_lock ____cacheline_aligned_in_smp; 142 struct xen_netif_rx_front_ring rx; 143 int rx_ring_ref; 144 145 struct timer_list rx_refill_timer; 146 147 struct sk_buff *rx_skbs[NET_RX_RING_SIZE]; 148 grant_ref_t gref_rx_head; 149 grant_ref_t grant_rx_ref[NET_RX_RING_SIZE]; 150 151 struct page_pool *page_pool; 152 struct xdp_rxq_info xdp_rxq; 153 }; 154 155 struct netfront_info { 156 struct list_head list; 157 struct net_device *netdev; 158 159 struct xenbus_device *xbdev; 160 161 /* Multi-queue support */ 162 struct netfront_queue *queues; 163 164 /* Statistics */ 165 struct netfront_stats __percpu *rx_stats; 166 struct netfront_stats __percpu *tx_stats; 167 168 /* XDP state */ 169 bool netback_has_xdp_headroom; 170 bool netfront_xdp_enabled; 171 172 /* Is device behaving sane? */ 173 bool broken; 174 175 atomic_t rx_gso_checksum_fixup; 176 }; 177 178 struct netfront_rx_info { 179 struct xen_netif_rx_response rx; 180 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1]; 181 }; 182 183 /* 184 * Access macros for acquiring freeing slots in tx_skbs[]. 185 */ 186 187 static void add_id_to_list(unsigned *head, unsigned short *list, 188 unsigned short id) 189 { 190 list[id] = *head; 191 *head = id; 192 } 193 194 static unsigned short get_id_from_list(unsigned *head, unsigned short *list) 195 { 196 unsigned int id = *head; 197 198 if (id != TX_LINK_NONE) { 199 *head = list[id]; 200 list[id] = TX_LINK_NONE; 201 } 202 return id; 203 } 204 205 static int xennet_rxidx(RING_IDX idx) 206 { 207 return idx & (NET_RX_RING_SIZE - 1); 208 } 209 210 static struct sk_buff *xennet_get_rx_skb(struct netfront_queue *queue, 211 RING_IDX ri) 212 { 213 int i = xennet_rxidx(ri); 214 struct sk_buff *skb = queue->rx_skbs[i]; 215 queue->rx_skbs[i] = NULL; 216 return skb; 217 } 218 219 static grant_ref_t xennet_get_rx_ref(struct netfront_queue *queue, 220 RING_IDX ri) 221 { 222 int i = xennet_rxidx(ri); 223 grant_ref_t ref = queue->grant_rx_ref[i]; 224 queue->grant_rx_ref[i] = GRANT_INVALID_REF; 225 return ref; 226 } 227 228 #ifdef CONFIG_SYSFS 229 static const struct attribute_group xennet_dev_group; 230 #endif 231 232 static bool xennet_can_sg(struct net_device *dev) 233 { 234 return dev->features & NETIF_F_SG; 235 } 236 237 238 static void rx_refill_timeout(struct timer_list *t) 239 { 240 struct netfront_queue *queue = from_timer(queue, t, rx_refill_timer); 241 napi_schedule(&queue->napi); 242 } 243 244 static int netfront_tx_slot_available(struct netfront_queue *queue) 245 { 246 return (queue->tx.req_prod_pvt - queue->tx.rsp_cons) < 247 (NET_TX_RING_SIZE - XEN_NETIF_NR_SLOTS_MIN - 1); 248 } 249 250 static void xennet_maybe_wake_tx(struct netfront_queue *queue) 251 { 252 struct net_device *dev = queue->info->netdev; 253 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, queue->id); 254 255 if (unlikely(netif_tx_queue_stopped(dev_queue)) && 256 netfront_tx_slot_available(queue) && 257 likely(netif_running(dev))) 258 netif_tx_wake_queue(netdev_get_tx_queue(dev, queue->id)); 259 } 260 261 262 static struct sk_buff *xennet_alloc_one_rx_buffer(struct netfront_queue *queue) 263 { 264 struct sk_buff *skb; 265 struct page *page; 266 267 skb = __netdev_alloc_skb(queue->info->netdev, 268 RX_COPY_THRESHOLD + NET_IP_ALIGN, 269 GFP_ATOMIC | __GFP_NOWARN); 270 if (unlikely(!skb)) 271 return NULL; 272 273 page = page_pool_dev_alloc_pages(queue->page_pool); 274 if (unlikely(!page)) { 275 kfree_skb(skb); 276 return NULL; 277 } 278 skb_add_rx_frag(skb, 0, page, 0, 0, PAGE_SIZE); 279 280 /* Align ip header to a 16 bytes boundary */ 281 skb_reserve(skb, NET_IP_ALIGN); 282 skb->dev = queue->info->netdev; 283 284 return skb; 285 } 286 287 288 static void xennet_alloc_rx_buffers(struct netfront_queue *queue) 289 { 290 RING_IDX req_prod = queue->rx.req_prod_pvt; 291 int notify; 292 int err = 0; 293 294 if (unlikely(!netif_carrier_ok(queue->info->netdev))) 295 return; 296 297 for (req_prod = queue->rx.req_prod_pvt; 298 req_prod - queue->rx.rsp_cons < NET_RX_RING_SIZE; 299 req_prod++) { 300 struct sk_buff *skb; 301 unsigned short id; 302 grant_ref_t ref; 303 struct page *page; 304 struct xen_netif_rx_request *req; 305 306 skb = xennet_alloc_one_rx_buffer(queue); 307 if (!skb) { 308 err = -ENOMEM; 309 break; 310 } 311 312 id = xennet_rxidx(req_prod); 313 314 BUG_ON(queue->rx_skbs[id]); 315 queue->rx_skbs[id] = skb; 316 317 ref = gnttab_claim_grant_reference(&queue->gref_rx_head); 318 WARN_ON_ONCE(IS_ERR_VALUE((unsigned long)(int)ref)); 319 queue->grant_rx_ref[id] = ref; 320 321 page = skb_frag_page(&skb_shinfo(skb)->frags[0]); 322 323 req = RING_GET_REQUEST(&queue->rx, req_prod); 324 gnttab_page_grant_foreign_access_ref_one(ref, 325 queue->info->xbdev->otherend_id, 326 page, 327 0); 328 req->id = id; 329 req->gref = ref; 330 } 331 332 queue->rx.req_prod_pvt = req_prod; 333 334 /* Try again later if there are not enough requests or skb allocation 335 * failed. 336 * Enough requests is quantified as the sum of newly created slots and 337 * the unconsumed slots at the backend. 338 */ 339 if (req_prod - queue->rx.rsp_cons < NET_RX_SLOTS_MIN || 340 unlikely(err)) { 341 mod_timer(&queue->rx_refill_timer, jiffies + (HZ/10)); 342 return; 343 } 344 345 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&queue->rx, notify); 346 if (notify) 347 notify_remote_via_irq(queue->rx_irq); 348 } 349 350 static int xennet_open(struct net_device *dev) 351 { 352 struct netfront_info *np = netdev_priv(dev); 353 unsigned int num_queues = dev->real_num_tx_queues; 354 unsigned int i = 0; 355 struct netfront_queue *queue = NULL; 356 357 if (!np->queues || np->broken) 358 return -ENODEV; 359 360 for (i = 0; i < num_queues; ++i) { 361 queue = &np->queues[i]; 362 napi_enable(&queue->napi); 363 364 spin_lock_bh(&queue->rx_lock); 365 if (netif_carrier_ok(dev)) { 366 xennet_alloc_rx_buffers(queue); 367 queue->rx.sring->rsp_event = queue->rx.rsp_cons + 1; 368 if (RING_HAS_UNCONSUMED_RESPONSES(&queue->rx)) 369 napi_schedule(&queue->napi); 370 } 371 spin_unlock_bh(&queue->rx_lock); 372 } 373 374 netif_tx_start_all_queues(dev); 375 376 return 0; 377 } 378 379 static void xennet_tx_buf_gc(struct netfront_queue *queue) 380 { 381 RING_IDX cons, prod; 382 unsigned short id; 383 struct sk_buff *skb; 384 bool more_to_do; 385 const struct device *dev = &queue->info->netdev->dev; 386 387 BUG_ON(!netif_carrier_ok(queue->info->netdev)); 388 389 do { 390 prod = queue->tx.sring->rsp_prod; 391 if (RING_RESPONSE_PROD_OVERFLOW(&queue->tx, prod)) { 392 dev_alert(dev, "Illegal number of responses %u\n", 393 prod - queue->tx.rsp_cons); 394 goto err; 395 } 396 rmb(); /* Ensure we see responses up to 'rp'. */ 397 398 for (cons = queue->tx.rsp_cons; cons != prod; cons++) { 399 struct xen_netif_tx_response txrsp; 400 401 RING_COPY_RESPONSE(&queue->tx, cons, &txrsp); 402 if (txrsp.status == XEN_NETIF_RSP_NULL) 403 continue; 404 405 id = txrsp.id; 406 if (id >= RING_SIZE(&queue->tx)) { 407 dev_alert(dev, 408 "Response has incorrect id (%u)\n", 409 id); 410 goto err; 411 } 412 if (queue->tx_link[id] != TX_PENDING) { 413 dev_alert(dev, 414 "Response for inactive request\n"); 415 goto err; 416 } 417 418 queue->tx_link[id] = TX_LINK_NONE; 419 skb = queue->tx_skbs[id]; 420 queue->tx_skbs[id] = NULL; 421 if (unlikely(gnttab_query_foreign_access( 422 queue->grant_tx_ref[id]) != 0)) { 423 dev_alert(dev, 424 "Grant still in use by backend domain\n"); 425 goto err; 426 } 427 gnttab_end_foreign_access_ref( 428 queue->grant_tx_ref[id], GNTMAP_readonly); 429 gnttab_release_grant_reference( 430 &queue->gref_tx_head, queue->grant_tx_ref[id]); 431 queue->grant_tx_ref[id] = GRANT_INVALID_REF; 432 queue->grant_tx_page[id] = NULL; 433 add_id_to_list(&queue->tx_skb_freelist, queue->tx_link, id); 434 dev_kfree_skb_irq(skb); 435 } 436 437 queue->tx.rsp_cons = prod; 438 439 RING_FINAL_CHECK_FOR_RESPONSES(&queue->tx, more_to_do); 440 } while (more_to_do); 441 442 xennet_maybe_wake_tx(queue); 443 444 return; 445 446 err: 447 queue->info->broken = true; 448 dev_alert(dev, "Disabled for further use\n"); 449 } 450 451 struct xennet_gnttab_make_txreq { 452 struct netfront_queue *queue; 453 struct sk_buff *skb; 454 struct page *page; 455 struct xen_netif_tx_request *tx; /* Last request on ring page */ 456 struct xen_netif_tx_request tx_local; /* Last request local copy*/ 457 unsigned int size; 458 }; 459 460 static void xennet_tx_setup_grant(unsigned long gfn, unsigned int offset, 461 unsigned int len, void *data) 462 { 463 struct xennet_gnttab_make_txreq *info = data; 464 unsigned int id; 465 struct xen_netif_tx_request *tx; 466 grant_ref_t ref; 467 /* convenient aliases */ 468 struct page *page = info->page; 469 struct netfront_queue *queue = info->queue; 470 struct sk_buff *skb = info->skb; 471 472 id = get_id_from_list(&queue->tx_skb_freelist, queue->tx_link); 473 tx = RING_GET_REQUEST(&queue->tx, queue->tx.req_prod_pvt++); 474 ref = gnttab_claim_grant_reference(&queue->gref_tx_head); 475 WARN_ON_ONCE(IS_ERR_VALUE((unsigned long)(int)ref)); 476 477 gnttab_grant_foreign_access_ref(ref, queue->info->xbdev->otherend_id, 478 gfn, GNTMAP_readonly); 479 480 queue->tx_skbs[id] = skb; 481 queue->grant_tx_page[id] = page; 482 queue->grant_tx_ref[id] = ref; 483 484 info->tx_local.id = id; 485 info->tx_local.gref = ref; 486 info->tx_local.offset = offset; 487 info->tx_local.size = len; 488 info->tx_local.flags = 0; 489 490 *tx = info->tx_local; 491 492 /* 493 * Put the request in the pending queue, it will be set to be pending 494 * when the producer index is about to be raised. 495 */ 496 add_id_to_list(&queue->tx_pend_queue, queue->tx_link, id); 497 498 info->tx = tx; 499 info->size += info->tx_local.size; 500 } 501 502 static struct xen_netif_tx_request *xennet_make_first_txreq( 503 struct xennet_gnttab_make_txreq *info, 504 unsigned int offset, unsigned int len) 505 { 506 info->size = 0; 507 508 gnttab_for_one_grant(info->page, offset, len, xennet_tx_setup_grant, info); 509 510 return info->tx; 511 } 512 513 static void xennet_make_one_txreq(unsigned long gfn, unsigned int offset, 514 unsigned int len, void *data) 515 { 516 struct xennet_gnttab_make_txreq *info = data; 517 518 info->tx->flags |= XEN_NETTXF_more_data; 519 skb_get(info->skb); 520 xennet_tx_setup_grant(gfn, offset, len, data); 521 } 522 523 static void xennet_make_txreqs( 524 struct xennet_gnttab_make_txreq *info, 525 struct page *page, 526 unsigned int offset, unsigned int len) 527 { 528 /* Skip unused frames from start of page */ 529 page += offset >> PAGE_SHIFT; 530 offset &= ~PAGE_MASK; 531 532 while (len) { 533 info->page = page; 534 info->size = 0; 535 536 gnttab_foreach_grant_in_range(page, offset, len, 537 xennet_make_one_txreq, 538 info); 539 540 page++; 541 offset = 0; 542 len -= info->size; 543 } 544 } 545 546 /* 547 * Count how many ring slots are required to send this skb. Each frag 548 * might be a compound page. 549 */ 550 static int xennet_count_skb_slots(struct sk_buff *skb) 551 { 552 int i, frags = skb_shinfo(skb)->nr_frags; 553 int slots; 554 555 slots = gnttab_count_grant(offset_in_page(skb->data), 556 skb_headlen(skb)); 557 558 for (i = 0; i < frags; i++) { 559 skb_frag_t *frag = skb_shinfo(skb)->frags + i; 560 unsigned long size = skb_frag_size(frag); 561 unsigned long offset = skb_frag_off(frag); 562 563 /* Skip unused frames from start of page */ 564 offset &= ~PAGE_MASK; 565 566 slots += gnttab_count_grant(offset, size); 567 } 568 569 return slots; 570 } 571 572 static u16 xennet_select_queue(struct net_device *dev, struct sk_buff *skb, 573 struct net_device *sb_dev) 574 { 575 unsigned int num_queues = dev->real_num_tx_queues; 576 u32 hash; 577 u16 queue_idx; 578 579 /* First, check if there is only one queue */ 580 if (num_queues == 1) { 581 queue_idx = 0; 582 } else { 583 hash = skb_get_hash(skb); 584 queue_idx = hash % num_queues; 585 } 586 587 return queue_idx; 588 } 589 590 static void xennet_mark_tx_pending(struct netfront_queue *queue) 591 { 592 unsigned int i; 593 594 while ((i = get_id_from_list(&queue->tx_pend_queue, queue->tx_link)) != 595 TX_LINK_NONE) 596 queue->tx_link[i] = TX_PENDING; 597 } 598 599 static int xennet_xdp_xmit_one(struct net_device *dev, 600 struct netfront_queue *queue, 601 struct xdp_frame *xdpf) 602 { 603 struct netfront_info *np = netdev_priv(dev); 604 struct netfront_stats *tx_stats = this_cpu_ptr(np->tx_stats); 605 struct xennet_gnttab_make_txreq info = { 606 .queue = queue, 607 .skb = NULL, 608 .page = virt_to_page(xdpf->data), 609 }; 610 int notify; 611 612 xennet_make_first_txreq(&info, 613 offset_in_page(xdpf->data), 614 xdpf->len); 615 616 xennet_mark_tx_pending(queue); 617 618 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&queue->tx, notify); 619 if (notify) 620 notify_remote_via_irq(queue->tx_irq); 621 622 u64_stats_update_begin(&tx_stats->syncp); 623 tx_stats->bytes += xdpf->len; 624 tx_stats->packets++; 625 u64_stats_update_end(&tx_stats->syncp); 626 627 xennet_tx_buf_gc(queue); 628 629 return 0; 630 } 631 632 static int xennet_xdp_xmit(struct net_device *dev, int n, 633 struct xdp_frame **frames, u32 flags) 634 { 635 unsigned int num_queues = dev->real_num_tx_queues; 636 struct netfront_info *np = netdev_priv(dev); 637 struct netfront_queue *queue = NULL; 638 unsigned long irq_flags; 639 int nxmit = 0; 640 int i; 641 642 if (unlikely(np->broken)) 643 return -ENODEV; 644 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) 645 return -EINVAL; 646 647 queue = &np->queues[smp_processor_id() % num_queues]; 648 649 spin_lock_irqsave(&queue->tx_lock, irq_flags); 650 for (i = 0; i < n; i++) { 651 struct xdp_frame *xdpf = frames[i]; 652 653 if (!xdpf) 654 continue; 655 if (xennet_xdp_xmit_one(dev, queue, xdpf)) 656 break; 657 nxmit++; 658 } 659 spin_unlock_irqrestore(&queue->tx_lock, irq_flags); 660 661 return nxmit; 662 } 663 664 665 #define MAX_XEN_SKB_FRAGS (65536 / XEN_PAGE_SIZE + 1) 666 667 static netdev_tx_t xennet_start_xmit(struct sk_buff *skb, struct net_device *dev) 668 { 669 struct netfront_info *np = netdev_priv(dev); 670 struct netfront_stats *tx_stats = this_cpu_ptr(np->tx_stats); 671 struct xen_netif_tx_request *first_tx; 672 unsigned int i; 673 int notify; 674 int slots; 675 struct page *page; 676 unsigned int offset; 677 unsigned int len; 678 unsigned long flags; 679 struct netfront_queue *queue = NULL; 680 struct xennet_gnttab_make_txreq info = { }; 681 unsigned int num_queues = dev->real_num_tx_queues; 682 u16 queue_index; 683 struct sk_buff *nskb; 684 685 /* Drop the packet if no queues are set up */ 686 if (num_queues < 1) 687 goto drop; 688 if (unlikely(np->broken)) 689 goto drop; 690 /* Determine which queue to transmit this SKB on */ 691 queue_index = skb_get_queue_mapping(skb); 692 queue = &np->queues[queue_index]; 693 694 /* If skb->len is too big for wire format, drop skb and alert 695 * user about misconfiguration. 696 */ 697 if (unlikely(skb->len > XEN_NETIF_MAX_TX_SIZE)) { 698 net_alert_ratelimited( 699 "xennet: skb->len = %u, too big for wire format\n", 700 skb->len); 701 goto drop; 702 } 703 704 slots = xennet_count_skb_slots(skb); 705 if (unlikely(slots > MAX_XEN_SKB_FRAGS + 1)) { 706 net_dbg_ratelimited("xennet: skb rides the rocket: %d slots, %d bytes\n", 707 slots, skb->len); 708 if (skb_linearize(skb)) 709 goto drop; 710 } 711 712 page = virt_to_page(skb->data); 713 offset = offset_in_page(skb->data); 714 715 /* The first req should be at least ETH_HLEN size or the packet will be 716 * dropped by netback. 717 */ 718 if (unlikely(PAGE_SIZE - offset < ETH_HLEN)) { 719 nskb = skb_copy(skb, GFP_ATOMIC); 720 if (!nskb) 721 goto drop; 722 dev_consume_skb_any(skb); 723 skb = nskb; 724 page = virt_to_page(skb->data); 725 offset = offset_in_page(skb->data); 726 } 727 728 len = skb_headlen(skb); 729 730 spin_lock_irqsave(&queue->tx_lock, flags); 731 732 if (unlikely(!netif_carrier_ok(dev) || 733 (slots > 1 && !xennet_can_sg(dev)) || 734 netif_needs_gso(skb, netif_skb_features(skb)))) { 735 spin_unlock_irqrestore(&queue->tx_lock, flags); 736 goto drop; 737 } 738 739 /* First request for the linear area. */ 740 info.queue = queue; 741 info.skb = skb; 742 info.page = page; 743 first_tx = xennet_make_first_txreq(&info, offset, len); 744 offset += info.tx_local.size; 745 if (offset == PAGE_SIZE) { 746 page++; 747 offset = 0; 748 } 749 len -= info.tx_local.size; 750 751 if (skb->ip_summed == CHECKSUM_PARTIAL) 752 /* local packet? */ 753 first_tx->flags |= XEN_NETTXF_csum_blank | 754 XEN_NETTXF_data_validated; 755 else if (skb->ip_summed == CHECKSUM_UNNECESSARY) 756 /* remote but checksummed. */ 757 first_tx->flags |= XEN_NETTXF_data_validated; 758 759 /* Optional extra info after the first request. */ 760 if (skb_shinfo(skb)->gso_size) { 761 struct xen_netif_extra_info *gso; 762 763 gso = (struct xen_netif_extra_info *) 764 RING_GET_REQUEST(&queue->tx, queue->tx.req_prod_pvt++); 765 766 first_tx->flags |= XEN_NETTXF_extra_info; 767 768 gso->u.gso.size = skb_shinfo(skb)->gso_size; 769 gso->u.gso.type = (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) ? 770 XEN_NETIF_GSO_TYPE_TCPV6 : 771 XEN_NETIF_GSO_TYPE_TCPV4; 772 gso->u.gso.pad = 0; 773 gso->u.gso.features = 0; 774 775 gso->type = XEN_NETIF_EXTRA_TYPE_GSO; 776 gso->flags = 0; 777 } 778 779 /* Requests for the rest of the linear area. */ 780 xennet_make_txreqs(&info, page, offset, len); 781 782 /* Requests for all the frags. */ 783 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 784 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 785 xennet_make_txreqs(&info, skb_frag_page(frag), 786 skb_frag_off(frag), 787 skb_frag_size(frag)); 788 } 789 790 /* First request has the packet length. */ 791 first_tx->size = skb->len; 792 793 /* timestamp packet in software */ 794 skb_tx_timestamp(skb); 795 796 xennet_mark_tx_pending(queue); 797 798 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&queue->tx, notify); 799 if (notify) 800 notify_remote_via_irq(queue->tx_irq); 801 802 u64_stats_update_begin(&tx_stats->syncp); 803 tx_stats->bytes += skb->len; 804 tx_stats->packets++; 805 u64_stats_update_end(&tx_stats->syncp); 806 807 /* Note: It is not safe to access skb after xennet_tx_buf_gc()! */ 808 xennet_tx_buf_gc(queue); 809 810 if (!netfront_tx_slot_available(queue)) 811 netif_tx_stop_queue(netdev_get_tx_queue(dev, queue->id)); 812 813 spin_unlock_irqrestore(&queue->tx_lock, flags); 814 815 return NETDEV_TX_OK; 816 817 drop: 818 dev->stats.tx_dropped++; 819 dev_kfree_skb_any(skb); 820 return NETDEV_TX_OK; 821 } 822 823 static int xennet_close(struct net_device *dev) 824 { 825 struct netfront_info *np = netdev_priv(dev); 826 unsigned int num_queues = dev->real_num_tx_queues; 827 unsigned int i; 828 struct netfront_queue *queue; 829 netif_tx_stop_all_queues(np->netdev); 830 for (i = 0; i < num_queues; ++i) { 831 queue = &np->queues[i]; 832 napi_disable(&queue->napi); 833 } 834 return 0; 835 } 836 837 static void xennet_move_rx_slot(struct netfront_queue *queue, struct sk_buff *skb, 838 grant_ref_t ref) 839 { 840 int new = xennet_rxidx(queue->rx.req_prod_pvt); 841 842 BUG_ON(queue->rx_skbs[new]); 843 queue->rx_skbs[new] = skb; 844 queue->grant_rx_ref[new] = ref; 845 RING_GET_REQUEST(&queue->rx, queue->rx.req_prod_pvt)->id = new; 846 RING_GET_REQUEST(&queue->rx, queue->rx.req_prod_pvt)->gref = ref; 847 queue->rx.req_prod_pvt++; 848 } 849 850 static int xennet_get_extras(struct netfront_queue *queue, 851 struct xen_netif_extra_info *extras, 852 RING_IDX rp) 853 854 { 855 struct xen_netif_extra_info extra; 856 struct device *dev = &queue->info->netdev->dev; 857 RING_IDX cons = queue->rx.rsp_cons; 858 int err = 0; 859 860 do { 861 struct sk_buff *skb; 862 grant_ref_t ref; 863 864 if (unlikely(cons + 1 == rp)) { 865 if (net_ratelimit()) 866 dev_warn(dev, "Missing extra info\n"); 867 err = -EBADR; 868 break; 869 } 870 871 RING_COPY_RESPONSE(&queue->rx, ++cons, &extra); 872 873 if (unlikely(!extra.type || 874 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) { 875 if (net_ratelimit()) 876 dev_warn(dev, "Invalid extra type: %d\n", 877 extra.type); 878 err = -EINVAL; 879 } else { 880 extras[extra.type - 1] = extra; 881 } 882 883 skb = xennet_get_rx_skb(queue, cons); 884 ref = xennet_get_rx_ref(queue, cons); 885 xennet_move_rx_slot(queue, skb, ref); 886 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE); 887 888 queue->rx.rsp_cons = cons; 889 return err; 890 } 891 892 static u32 xennet_run_xdp(struct netfront_queue *queue, struct page *pdata, 893 struct xen_netif_rx_response *rx, struct bpf_prog *prog, 894 struct xdp_buff *xdp, bool *need_xdp_flush) 895 { 896 struct xdp_frame *xdpf; 897 u32 len = rx->status; 898 u32 act; 899 int err; 900 901 xdp_init_buff(xdp, XEN_PAGE_SIZE - XDP_PACKET_HEADROOM, 902 &queue->xdp_rxq); 903 xdp_prepare_buff(xdp, page_address(pdata), XDP_PACKET_HEADROOM, 904 len, false); 905 906 act = bpf_prog_run_xdp(prog, xdp); 907 switch (act) { 908 case XDP_TX: 909 get_page(pdata); 910 xdpf = xdp_convert_buff_to_frame(xdp); 911 err = xennet_xdp_xmit(queue->info->netdev, 1, &xdpf, 0); 912 if (unlikely(!err)) 913 xdp_return_frame_rx_napi(xdpf); 914 else if (unlikely(err < 0)) 915 trace_xdp_exception(queue->info->netdev, prog, act); 916 break; 917 case XDP_REDIRECT: 918 get_page(pdata); 919 err = xdp_do_redirect(queue->info->netdev, xdp, prog); 920 *need_xdp_flush = true; 921 if (unlikely(err)) 922 trace_xdp_exception(queue->info->netdev, prog, act); 923 break; 924 case XDP_PASS: 925 case XDP_DROP: 926 break; 927 928 case XDP_ABORTED: 929 trace_xdp_exception(queue->info->netdev, prog, act); 930 break; 931 932 default: 933 bpf_warn_invalid_xdp_action(act); 934 } 935 936 return act; 937 } 938 939 static int xennet_get_responses(struct netfront_queue *queue, 940 struct netfront_rx_info *rinfo, RING_IDX rp, 941 struct sk_buff_head *list, 942 bool *need_xdp_flush) 943 { 944 struct xen_netif_rx_response *rx = &rinfo->rx, rx_local; 945 int max = XEN_NETIF_NR_SLOTS_MIN + (rx->status <= RX_COPY_THRESHOLD); 946 RING_IDX cons = queue->rx.rsp_cons; 947 struct sk_buff *skb = xennet_get_rx_skb(queue, cons); 948 struct xen_netif_extra_info *extras = rinfo->extras; 949 grant_ref_t ref = xennet_get_rx_ref(queue, cons); 950 struct device *dev = &queue->info->netdev->dev; 951 struct bpf_prog *xdp_prog; 952 struct xdp_buff xdp; 953 unsigned long ret; 954 int slots = 1; 955 int err = 0; 956 u32 verdict; 957 958 if (rx->flags & XEN_NETRXF_extra_info) { 959 err = xennet_get_extras(queue, extras, rp); 960 if (!err) { 961 if (extras[XEN_NETIF_EXTRA_TYPE_XDP - 1].type) { 962 struct xen_netif_extra_info *xdp; 963 964 xdp = &extras[XEN_NETIF_EXTRA_TYPE_XDP - 1]; 965 rx->offset = xdp->u.xdp.headroom; 966 } 967 } 968 cons = queue->rx.rsp_cons; 969 } 970 971 for (;;) { 972 if (unlikely(rx->status < 0 || 973 rx->offset + rx->status > XEN_PAGE_SIZE)) { 974 if (net_ratelimit()) 975 dev_warn(dev, "rx->offset: %u, size: %d\n", 976 rx->offset, rx->status); 977 xennet_move_rx_slot(queue, skb, ref); 978 err = -EINVAL; 979 goto next; 980 } 981 982 /* 983 * This definitely indicates a bug, either in this driver or in 984 * the backend driver. In future this should flag the bad 985 * situation to the system controller to reboot the backend. 986 */ 987 if (ref == GRANT_INVALID_REF) { 988 if (net_ratelimit()) 989 dev_warn(dev, "Bad rx response id %d.\n", 990 rx->id); 991 err = -EINVAL; 992 goto next; 993 } 994 995 ret = gnttab_end_foreign_access_ref(ref, 0); 996 BUG_ON(!ret); 997 998 gnttab_release_grant_reference(&queue->gref_rx_head, ref); 999 1000 rcu_read_lock(); 1001 xdp_prog = rcu_dereference(queue->xdp_prog); 1002 if (xdp_prog) { 1003 if (!(rx->flags & XEN_NETRXF_more_data)) { 1004 /* currently only a single page contains data */ 1005 verdict = xennet_run_xdp(queue, 1006 skb_frag_page(&skb_shinfo(skb)->frags[0]), 1007 rx, xdp_prog, &xdp, need_xdp_flush); 1008 if (verdict != XDP_PASS) 1009 err = -EINVAL; 1010 } else { 1011 /* drop the frame */ 1012 err = -EINVAL; 1013 } 1014 } 1015 rcu_read_unlock(); 1016 next: 1017 __skb_queue_tail(list, skb); 1018 if (!(rx->flags & XEN_NETRXF_more_data)) 1019 break; 1020 1021 if (cons + slots == rp) { 1022 if (net_ratelimit()) 1023 dev_warn(dev, "Need more slots\n"); 1024 err = -ENOENT; 1025 break; 1026 } 1027 1028 RING_COPY_RESPONSE(&queue->rx, cons + slots, &rx_local); 1029 rx = &rx_local; 1030 skb = xennet_get_rx_skb(queue, cons + slots); 1031 ref = xennet_get_rx_ref(queue, cons + slots); 1032 slots++; 1033 } 1034 1035 if (unlikely(slots > max)) { 1036 if (net_ratelimit()) 1037 dev_warn(dev, "Too many slots\n"); 1038 err = -E2BIG; 1039 } 1040 1041 if (unlikely(err)) 1042 queue->rx.rsp_cons = cons + slots; 1043 1044 return err; 1045 } 1046 1047 static int xennet_set_skb_gso(struct sk_buff *skb, 1048 struct xen_netif_extra_info *gso) 1049 { 1050 if (!gso->u.gso.size) { 1051 if (net_ratelimit()) 1052 pr_warn("GSO size must not be zero\n"); 1053 return -EINVAL; 1054 } 1055 1056 if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4 && 1057 gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV6) { 1058 if (net_ratelimit()) 1059 pr_warn("Bad GSO type %d\n", gso->u.gso.type); 1060 return -EINVAL; 1061 } 1062 1063 skb_shinfo(skb)->gso_size = gso->u.gso.size; 1064 skb_shinfo(skb)->gso_type = 1065 (gso->u.gso.type == XEN_NETIF_GSO_TYPE_TCPV4) ? 1066 SKB_GSO_TCPV4 : 1067 SKB_GSO_TCPV6; 1068 1069 /* Header must be checked, and gso_segs computed. */ 1070 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY; 1071 skb_shinfo(skb)->gso_segs = 0; 1072 1073 return 0; 1074 } 1075 1076 static int xennet_fill_frags(struct netfront_queue *queue, 1077 struct sk_buff *skb, 1078 struct sk_buff_head *list) 1079 { 1080 RING_IDX cons = queue->rx.rsp_cons; 1081 struct sk_buff *nskb; 1082 1083 while ((nskb = __skb_dequeue(list))) { 1084 struct xen_netif_rx_response rx; 1085 skb_frag_t *nfrag = &skb_shinfo(nskb)->frags[0]; 1086 1087 RING_COPY_RESPONSE(&queue->rx, ++cons, &rx); 1088 1089 if (skb_shinfo(skb)->nr_frags == MAX_SKB_FRAGS) { 1090 unsigned int pull_to = NETFRONT_SKB_CB(skb)->pull_to; 1091 1092 BUG_ON(pull_to < skb_headlen(skb)); 1093 __pskb_pull_tail(skb, pull_to - skb_headlen(skb)); 1094 } 1095 if (unlikely(skb_shinfo(skb)->nr_frags >= MAX_SKB_FRAGS)) { 1096 queue->rx.rsp_cons = ++cons + skb_queue_len(list); 1097 kfree_skb(nskb); 1098 return -ENOENT; 1099 } 1100 1101 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, 1102 skb_frag_page(nfrag), 1103 rx.offset, rx.status, PAGE_SIZE); 1104 1105 skb_shinfo(nskb)->nr_frags = 0; 1106 kfree_skb(nskb); 1107 } 1108 1109 queue->rx.rsp_cons = cons; 1110 1111 return 0; 1112 } 1113 1114 static int checksum_setup(struct net_device *dev, struct sk_buff *skb) 1115 { 1116 bool recalculate_partial_csum = false; 1117 1118 /* 1119 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy 1120 * peers can fail to set NETRXF_csum_blank when sending a GSO 1121 * frame. In this case force the SKB to CHECKSUM_PARTIAL and 1122 * recalculate the partial checksum. 1123 */ 1124 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) { 1125 struct netfront_info *np = netdev_priv(dev); 1126 atomic_inc(&np->rx_gso_checksum_fixup); 1127 skb->ip_summed = CHECKSUM_PARTIAL; 1128 recalculate_partial_csum = true; 1129 } 1130 1131 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */ 1132 if (skb->ip_summed != CHECKSUM_PARTIAL) 1133 return 0; 1134 1135 return skb_checksum_setup(skb, recalculate_partial_csum); 1136 } 1137 1138 static int handle_incoming_queue(struct netfront_queue *queue, 1139 struct sk_buff_head *rxq) 1140 { 1141 struct netfront_stats *rx_stats = this_cpu_ptr(queue->info->rx_stats); 1142 int packets_dropped = 0; 1143 struct sk_buff *skb; 1144 1145 while ((skb = __skb_dequeue(rxq)) != NULL) { 1146 int pull_to = NETFRONT_SKB_CB(skb)->pull_to; 1147 1148 if (pull_to > skb_headlen(skb)) 1149 __pskb_pull_tail(skb, pull_to - skb_headlen(skb)); 1150 1151 /* Ethernet work: Delayed to here as it peeks the header. */ 1152 skb->protocol = eth_type_trans(skb, queue->info->netdev); 1153 skb_reset_network_header(skb); 1154 1155 if (checksum_setup(queue->info->netdev, skb)) { 1156 kfree_skb(skb); 1157 packets_dropped++; 1158 queue->info->netdev->stats.rx_errors++; 1159 continue; 1160 } 1161 1162 u64_stats_update_begin(&rx_stats->syncp); 1163 rx_stats->packets++; 1164 rx_stats->bytes += skb->len; 1165 u64_stats_update_end(&rx_stats->syncp); 1166 1167 /* Pass it up. */ 1168 napi_gro_receive(&queue->napi, skb); 1169 } 1170 1171 return packets_dropped; 1172 } 1173 1174 static int xennet_poll(struct napi_struct *napi, int budget) 1175 { 1176 struct netfront_queue *queue = container_of(napi, struct netfront_queue, napi); 1177 struct net_device *dev = queue->info->netdev; 1178 struct sk_buff *skb; 1179 struct netfront_rx_info rinfo; 1180 struct xen_netif_rx_response *rx = &rinfo.rx; 1181 struct xen_netif_extra_info *extras = rinfo.extras; 1182 RING_IDX i, rp; 1183 int work_done; 1184 struct sk_buff_head rxq; 1185 struct sk_buff_head errq; 1186 struct sk_buff_head tmpq; 1187 int err; 1188 bool need_xdp_flush = false; 1189 1190 spin_lock(&queue->rx_lock); 1191 1192 skb_queue_head_init(&rxq); 1193 skb_queue_head_init(&errq); 1194 skb_queue_head_init(&tmpq); 1195 1196 rp = queue->rx.sring->rsp_prod; 1197 if (RING_RESPONSE_PROD_OVERFLOW(&queue->rx, rp)) { 1198 dev_alert(&dev->dev, "Illegal number of responses %u\n", 1199 rp - queue->rx.rsp_cons); 1200 queue->info->broken = true; 1201 spin_unlock(&queue->rx_lock); 1202 return 0; 1203 } 1204 rmb(); /* Ensure we see queued responses up to 'rp'. */ 1205 1206 i = queue->rx.rsp_cons; 1207 work_done = 0; 1208 while ((i != rp) && (work_done < budget)) { 1209 RING_COPY_RESPONSE(&queue->rx, i, rx); 1210 memset(extras, 0, sizeof(rinfo.extras)); 1211 1212 err = xennet_get_responses(queue, &rinfo, rp, &tmpq, 1213 &need_xdp_flush); 1214 1215 if (unlikely(err)) { 1216 err: 1217 while ((skb = __skb_dequeue(&tmpq))) 1218 __skb_queue_tail(&errq, skb); 1219 dev->stats.rx_errors++; 1220 i = queue->rx.rsp_cons; 1221 continue; 1222 } 1223 1224 skb = __skb_dequeue(&tmpq); 1225 1226 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) { 1227 struct xen_netif_extra_info *gso; 1228 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1]; 1229 1230 if (unlikely(xennet_set_skb_gso(skb, gso))) { 1231 __skb_queue_head(&tmpq, skb); 1232 queue->rx.rsp_cons += skb_queue_len(&tmpq); 1233 goto err; 1234 } 1235 } 1236 1237 NETFRONT_SKB_CB(skb)->pull_to = rx->status; 1238 if (NETFRONT_SKB_CB(skb)->pull_to > RX_COPY_THRESHOLD) 1239 NETFRONT_SKB_CB(skb)->pull_to = RX_COPY_THRESHOLD; 1240 1241 skb_frag_off_set(&skb_shinfo(skb)->frags[0], rx->offset); 1242 skb_frag_size_set(&skb_shinfo(skb)->frags[0], rx->status); 1243 skb->data_len = rx->status; 1244 skb->len += rx->status; 1245 1246 if (unlikely(xennet_fill_frags(queue, skb, &tmpq))) 1247 goto err; 1248 1249 if (rx->flags & XEN_NETRXF_csum_blank) 1250 skb->ip_summed = CHECKSUM_PARTIAL; 1251 else if (rx->flags & XEN_NETRXF_data_validated) 1252 skb->ip_summed = CHECKSUM_UNNECESSARY; 1253 1254 __skb_queue_tail(&rxq, skb); 1255 1256 i = ++queue->rx.rsp_cons; 1257 work_done++; 1258 } 1259 if (need_xdp_flush) 1260 xdp_do_flush(); 1261 1262 __skb_queue_purge(&errq); 1263 1264 work_done -= handle_incoming_queue(queue, &rxq); 1265 1266 xennet_alloc_rx_buffers(queue); 1267 1268 if (work_done < budget) { 1269 int more_to_do = 0; 1270 1271 napi_complete_done(napi, work_done); 1272 1273 RING_FINAL_CHECK_FOR_RESPONSES(&queue->rx, more_to_do); 1274 if (more_to_do) 1275 napi_schedule(napi); 1276 } 1277 1278 spin_unlock(&queue->rx_lock); 1279 1280 return work_done; 1281 } 1282 1283 static int xennet_change_mtu(struct net_device *dev, int mtu) 1284 { 1285 int max = xennet_can_sg(dev) ? XEN_NETIF_MAX_TX_SIZE : ETH_DATA_LEN; 1286 1287 if (mtu > max) 1288 return -EINVAL; 1289 dev->mtu = mtu; 1290 return 0; 1291 } 1292 1293 static void xennet_get_stats64(struct net_device *dev, 1294 struct rtnl_link_stats64 *tot) 1295 { 1296 struct netfront_info *np = netdev_priv(dev); 1297 int cpu; 1298 1299 for_each_possible_cpu(cpu) { 1300 struct netfront_stats *rx_stats = per_cpu_ptr(np->rx_stats, cpu); 1301 struct netfront_stats *tx_stats = per_cpu_ptr(np->tx_stats, cpu); 1302 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 1303 unsigned int start; 1304 1305 do { 1306 start = u64_stats_fetch_begin_irq(&tx_stats->syncp); 1307 tx_packets = tx_stats->packets; 1308 tx_bytes = tx_stats->bytes; 1309 } while (u64_stats_fetch_retry_irq(&tx_stats->syncp, start)); 1310 1311 do { 1312 start = u64_stats_fetch_begin_irq(&rx_stats->syncp); 1313 rx_packets = rx_stats->packets; 1314 rx_bytes = rx_stats->bytes; 1315 } while (u64_stats_fetch_retry_irq(&rx_stats->syncp, start)); 1316 1317 tot->rx_packets += rx_packets; 1318 tot->tx_packets += tx_packets; 1319 tot->rx_bytes += rx_bytes; 1320 tot->tx_bytes += tx_bytes; 1321 } 1322 1323 tot->rx_errors = dev->stats.rx_errors; 1324 tot->tx_dropped = dev->stats.tx_dropped; 1325 } 1326 1327 static void xennet_release_tx_bufs(struct netfront_queue *queue) 1328 { 1329 struct sk_buff *skb; 1330 int i; 1331 1332 for (i = 0; i < NET_TX_RING_SIZE; i++) { 1333 /* Skip over entries which are actually freelist references */ 1334 if (!queue->tx_skbs[i]) 1335 continue; 1336 1337 skb = queue->tx_skbs[i]; 1338 queue->tx_skbs[i] = NULL; 1339 get_page(queue->grant_tx_page[i]); 1340 gnttab_end_foreign_access(queue->grant_tx_ref[i], 1341 GNTMAP_readonly, 1342 (unsigned long)page_address(queue->grant_tx_page[i])); 1343 queue->grant_tx_page[i] = NULL; 1344 queue->grant_tx_ref[i] = GRANT_INVALID_REF; 1345 add_id_to_list(&queue->tx_skb_freelist, queue->tx_link, i); 1346 dev_kfree_skb_irq(skb); 1347 } 1348 } 1349 1350 static void xennet_release_rx_bufs(struct netfront_queue *queue) 1351 { 1352 int id, ref; 1353 1354 spin_lock_bh(&queue->rx_lock); 1355 1356 for (id = 0; id < NET_RX_RING_SIZE; id++) { 1357 struct sk_buff *skb; 1358 struct page *page; 1359 1360 skb = queue->rx_skbs[id]; 1361 if (!skb) 1362 continue; 1363 1364 ref = queue->grant_rx_ref[id]; 1365 if (ref == GRANT_INVALID_REF) 1366 continue; 1367 1368 page = skb_frag_page(&skb_shinfo(skb)->frags[0]); 1369 1370 /* gnttab_end_foreign_access() needs a page ref until 1371 * foreign access is ended (which may be deferred). 1372 */ 1373 get_page(page); 1374 gnttab_end_foreign_access(ref, 0, 1375 (unsigned long)page_address(page)); 1376 queue->grant_rx_ref[id] = GRANT_INVALID_REF; 1377 1378 kfree_skb(skb); 1379 } 1380 1381 spin_unlock_bh(&queue->rx_lock); 1382 } 1383 1384 static netdev_features_t xennet_fix_features(struct net_device *dev, 1385 netdev_features_t features) 1386 { 1387 struct netfront_info *np = netdev_priv(dev); 1388 1389 if (features & NETIF_F_SG && 1390 !xenbus_read_unsigned(np->xbdev->otherend, "feature-sg", 0)) 1391 features &= ~NETIF_F_SG; 1392 1393 if (features & NETIF_F_IPV6_CSUM && 1394 !xenbus_read_unsigned(np->xbdev->otherend, 1395 "feature-ipv6-csum-offload", 0)) 1396 features &= ~NETIF_F_IPV6_CSUM; 1397 1398 if (features & NETIF_F_TSO && 1399 !xenbus_read_unsigned(np->xbdev->otherend, "feature-gso-tcpv4", 0)) 1400 features &= ~NETIF_F_TSO; 1401 1402 if (features & NETIF_F_TSO6 && 1403 !xenbus_read_unsigned(np->xbdev->otherend, "feature-gso-tcpv6", 0)) 1404 features &= ~NETIF_F_TSO6; 1405 1406 return features; 1407 } 1408 1409 static int xennet_set_features(struct net_device *dev, 1410 netdev_features_t features) 1411 { 1412 if (!(features & NETIF_F_SG) && dev->mtu > ETH_DATA_LEN) { 1413 netdev_info(dev, "Reducing MTU because no SG offload"); 1414 dev->mtu = ETH_DATA_LEN; 1415 } 1416 1417 return 0; 1418 } 1419 1420 static irqreturn_t xennet_tx_interrupt(int irq, void *dev_id) 1421 { 1422 struct netfront_queue *queue = dev_id; 1423 unsigned long flags; 1424 1425 if (queue->info->broken) 1426 return IRQ_HANDLED; 1427 1428 spin_lock_irqsave(&queue->tx_lock, flags); 1429 xennet_tx_buf_gc(queue); 1430 spin_unlock_irqrestore(&queue->tx_lock, flags); 1431 1432 return IRQ_HANDLED; 1433 } 1434 1435 static irqreturn_t xennet_rx_interrupt(int irq, void *dev_id) 1436 { 1437 struct netfront_queue *queue = dev_id; 1438 struct net_device *dev = queue->info->netdev; 1439 1440 if (queue->info->broken) 1441 return IRQ_HANDLED; 1442 1443 if (likely(netif_carrier_ok(dev) && 1444 RING_HAS_UNCONSUMED_RESPONSES(&queue->rx))) 1445 napi_schedule(&queue->napi); 1446 1447 return IRQ_HANDLED; 1448 } 1449 1450 static irqreturn_t xennet_interrupt(int irq, void *dev_id) 1451 { 1452 xennet_tx_interrupt(irq, dev_id); 1453 xennet_rx_interrupt(irq, dev_id); 1454 return IRQ_HANDLED; 1455 } 1456 1457 #ifdef CONFIG_NET_POLL_CONTROLLER 1458 static void xennet_poll_controller(struct net_device *dev) 1459 { 1460 /* Poll each queue */ 1461 struct netfront_info *info = netdev_priv(dev); 1462 unsigned int num_queues = dev->real_num_tx_queues; 1463 unsigned int i; 1464 1465 if (info->broken) 1466 return; 1467 1468 for (i = 0; i < num_queues; ++i) 1469 xennet_interrupt(0, &info->queues[i]); 1470 } 1471 #endif 1472 1473 #define NETBACK_XDP_HEADROOM_DISABLE 0 1474 #define NETBACK_XDP_HEADROOM_ENABLE 1 1475 1476 static int talk_to_netback_xdp(struct netfront_info *np, int xdp) 1477 { 1478 int err; 1479 unsigned short headroom; 1480 1481 headroom = xdp ? XDP_PACKET_HEADROOM : 0; 1482 err = xenbus_printf(XBT_NIL, np->xbdev->nodename, 1483 "xdp-headroom", "%hu", 1484 headroom); 1485 if (err) 1486 pr_warn("Error writing xdp-headroom\n"); 1487 1488 return err; 1489 } 1490 1491 static int xennet_xdp_set(struct net_device *dev, struct bpf_prog *prog, 1492 struct netlink_ext_ack *extack) 1493 { 1494 unsigned long max_mtu = XEN_PAGE_SIZE - XDP_PACKET_HEADROOM; 1495 struct netfront_info *np = netdev_priv(dev); 1496 struct bpf_prog *old_prog; 1497 unsigned int i, err; 1498 1499 if (dev->mtu > max_mtu) { 1500 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_mtu); 1501 return -EINVAL; 1502 } 1503 1504 if (!np->netback_has_xdp_headroom) 1505 return 0; 1506 1507 xenbus_switch_state(np->xbdev, XenbusStateReconfiguring); 1508 1509 err = talk_to_netback_xdp(np, prog ? NETBACK_XDP_HEADROOM_ENABLE : 1510 NETBACK_XDP_HEADROOM_DISABLE); 1511 if (err) 1512 return err; 1513 1514 /* avoid the race with XDP headroom adjustment */ 1515 wait_event(module_wq, 1516 xenbus_read_driver_state(np->xbdev->otherend) == 1517 XenbusStateReconfigured); 1518 np->netfront_xdp_enabled = true; 1519 1520 old_prog = rtnl_dereference(np->queues[0].xdp_prog); 1521 1522 if (prog) 1523 bpf_prog_add(prog, dev->real_num_tx_queues); 1524 1525 for (i = 0; i < dev->real_num_tx_queues; ++i) 1526 rcu_assign_pointer(np->queues[i].xdp_prog, prog); 1527 1528 if (old_prog) 1529 for (i = 0; i < dev->real_num_tx_queues; ++i) 1530 bpf_prog_put(old_prog); 1531 1532 xenbus_switch_state(np->xbdev, XenbusStateConnected); 1533 1534 return 0; 1535 } 1536 1537 static int xennet_xdp(struct net_device *dev, struct netdev_bpf *xdp) 1538 { 1539 struct netfront_info *np = netdev_priv(dev); 1540 1541 if (np->broken) 1542 return -ENODEV; 1543 1544 switch (xdp->command) { 1545 case XDP_SETUP_PROG: 1546 return xennet_xdp_set(dev, xdp->prog, xdp->extack); 1547 default: 1548 return -EINVAL; 1549 } 1550 } 1551 1552 static const struct net_device_ops xennet_netdev_ops = { 1553 .ndo_open = xennet_open, 1554 .ndo_stop = xennet_close, 1555 .ndo_start_xmit = xennet_start_xmit, 1556 .ndo_change_mtu = xennet_change_mtu, 1557 .ndo_get_stats64 = xennet_get_stats64, 1558 .ndo_set_mac_address = eth_mac_addr, 1559 .ndo_validate_addr = eth_validate_addr, 1560 .ndo_fix_features = xennet_fix_features, 1561 .ndo_set_features = xennet_set_features, 1562 .ndo_select_queue = xennet_select_queue, 1563 .ndo_bpf = xennet_xdp, 1564 .ndo_xdp_xmit = xennet_xdp_xmit, 1565 #ifdef CONFIG_NET_POLL_CONTROLLER 1566 .ndo_poll_controller = xennet_poll_controller, 1567 #endif 1568 }; 1569 1570 static void xennet_free_netdev(struct net_device *netdev) 1571 { 1572 struct netfront_info *np = netdev_priv(netdev); 1573 1574 free_percpu(np->rx_stats); 1575 free_percpu(np->tx_stats); 1576 free_netdev(netdev); 1577 } 1578 1579 static struct net_device *xennet_create_dev(struct xenbus_device *dev) 1580 { 1581 int err; 1582 struct net_device *netdev; 1583 struct netfront_info *np; 1584 1585 netdev = alloc_etherdev_mq(sizeof(struct netfront_info), xennet_max_queues); 1586 if (!netdev) 1587 return ERR_PTR(-ENOMEM); 1588 1589 np = netdev_priv(netdev); 1590 np->xbdev = dev; 1591 1592 np->queues = NULL; 1593 1594 err = -ENOMEM; 1595 np->rx_stats = netdev_alloc_pcpu_stats(struct netfront_stats); 1596 if (np->rx_stats == NULL) 1597 goto exit; 1598 np->tx_stats = netdev_alloc_pcpu_stats(struct netfront_stats); 1599 if (np->tx_stats == NULL) 1600 goto exit; 1601 1602 netdev->netdev_ops = &xennet_netdev_ops; 1603 1604 netdev->features = NETIF_F_IP_CSUM | NETIF_F_RXCSUM | 1605 NETIF_F_GSO_ROBUST; 1606 netdev->hw_features = NETIF_F_SG | 1607 NETIF_F_IPV6_CSUM | 1608 NETIF_F_TSO | NETIF_F_TSO6; 1609 1610 /* 1611 * Assume that all hw features are available for now. This set 1612 * will be adjusted by the call to netdev_update_features() in 1613 * xennet_connect() which is the earliest point where we can 1614 * negotiate with the backend regarding supported features. 1615 */ 1616 netdev->features |= netdev->hw_features; 1617 1618 netdev->ethtool_ops = &xennet_ethtool_ops; 1619 netdev->min_mtu = ETH_MIN_MTU; 1620 netdev->max_mtu = XEN_NETIF_MAX_TX_SIZE; 1621 SET_NETDEV_DEV(netdev, &dev->dev); 1622 1623 np->netdev = netdev; 1624 np->netfront_xdp_enabled = false; 1625 1626 netif_carrier_off(netdev); 1627 1628 do { 1629 xenbus_switch_state(dev, XenbusStateInitialising); 1630 err = wait_event_timeout(module_wq, 1631 xenbus_read_driver_state(dev->otherend) != 1632 XenbusStateClosed && 1633 xenbus_read_driver_state(dev->otherend) != 1634 XenbusStateUnknown, XENNET_TIMEOUT); 1635 } while (!err); 1636 1637 return netdev; 1638 1639 exit: 1640 xennet_free_netdev(netdev); 1641 return ERR_PTR(err); 1642 } 1643 1644 /* 1645 * Entry point to this code when a new device is created. Allocate the basic 1646 * structures and the ring buffers for communication with the backend, and 1647 * inform the backend of the appropriate details for those. 1648 */ 1649 static int netfront_probe(struct xenbus_device *dev, 1650 const struct xenbus_device_id *id) 1651 { 1652 int err; 1653 struct net_device *netdev; 1654 struct netfront_info *info; 1655 1656 netdev = xennet_create_dev(dev); 1657 if (IS_ERR(netdev)) { 1658 err = PTR_ERR(netdev); 1659 xenbus_dev_fatal(dev, err, "creating netdev"); 1660 return err; 1661 } 1662 1663 info = netdev_priv(netdev); 1664 dev_set_drvdata(&dev->dev, info); 1665 #ifdef CONFIG_SYSFS 1666 info->netdev->sysfs_groups[0] = &xennet_dev_group; 1667 #endif 1668 1669 return 0; 1670 } 1671 1672 static void xennet_end_access(int ref, void *page) 1673 { 1674 /* This frees the page as a side-effect */ 1675 if (ref != GRANT_INVALID_REF) 1676 gnttab_end_foreign_access(ref, 0, (unsigned long)page); 1677 } 1678 1679 static void xennet_disconnect_backend(struct netfront_info *info) 1680 { 1681 unsigned int i = 0; 1682 unsigned int num_queues = info->netdev->real_num_tx_queues; 1683 1684 netif_carrier_off(info->netdev); 1685 1686 for (i = 0; i < num_queues && info->queues; ++i) { 1687 struct netfront_queue *queue = &info->queues[i]; 1688 1689 del_timer_sync(&queue->rx_refill_timer); 1690 1691 if (queue->tx_irq && (queue->tx_irq == queue->rx_irq)) 1692 unbind_from_irqhandler(queue->tx_irq, queue); 1693 if (queue->tx_irq && (queue->tx_irq != queue->rx_irq)) { 1694 unbind_from_irqhandler(queue->tx_irq, queue); 1695 unbind_from_irqhandler(queue->rx_irq, queue); 1696 } 1697 queue->tx_evtchn = queue->rx_evtchn = 0; 1698 queue->tx_irq = queue->rx_irq = 0; 1699 1700 if (netif_running(info->netdev)) 1701 napi_synchronize(&queue->napi); 1702 1703 xennet_release_tx_bufs(queue); 1704 xennet_release_rx_bufs(queue); 1705 gnttab_free_grant_references(queue->gref_tx_head); 1706 gnttab_free_grant_references(queue->gref_rx_head); 1707 1708 /* End access and free the pages */ 1709 xennet_end_access(queue->tx_ring_ref, queue->tx.sring); 1710 xennet_end_access(queue->rx_ring_ref, queue->rx.sring); 1711 1712 queue->tx_ring_ref = GRANT_INVALID_REF; 1713 queue->rx_ring_ref = GRANT_INVALID_REF; 1714 queue->tx.sring = NULL; 1715 queue->rx.sring = NULL; 1716 1717 page_pool_destroy(queue->page_pool); 1718 } 1719 } 1720 1721 /* 1722 * We are reconnecting to the backend, due to a suspend/resume, or a backend 1723 * driver restart. We tear down our netif structure and recreate it, but 1724 * leave the device-layer structures intact so that this is transparent to the 1725 * rest of the kernel. 1726 */ 1727 static int netfront_resume(struct xenbus_device *dev) 1728 { 1729 struct netfront_info *info = dev_get_drvdata(&dev->dev); 1730 1731 dev_dbg(&dev->dev, "%s\n", dev->nodename); 1732 1733 netif_tx_lock_bh(info->netdev); 1734 netif_device_detach(info->netdev); 1735 netif_tx_unlock_bh(info->netdev); 1736 1737 xennet_disconnect_backend(info); 1738 return 0; 1739 } 1740 1741 static int xen_net_read_mac(struct xenbus_device *dev, u8 mac[]) 1742 { 1743 char *s, *e, *macstr; 1744 int i; 1745 1746 macstr = s = xenbus_read(XBT_NIL, dev->nodename, "mac", NULL); 1747 if (IS_ERR(macstr)) 1748 return PTR_ERR(macstr); 1749 1750 for (i = 0; i < ETH_ALEN; i++) { 1751 mac[i] = simple_strtoul(s, &e, 16); 1752 if ((s == e) || (*e != ((i == ETH_ALEN-1) ? '\0' : ':'))) { 1753 kfree(macstr); 1754 return -ENOENT; 1755 } 1756 s = e+1; 1757 } 1758 1759 kfree(macstr); 1760 return 0; 1761 } 1762 1763 static int setup_netfront_single(struct netfront_queue *queue) 1764 { 1765 int err; 1766 1767 err = xenbus_alloc_evtchn(queue->info->xbdev, &queue->tx_evtchn); 1768 if (err < 0) 1769 goto fail; 1770 1771 err = bind_evtchn_to_irqhandler(queue->tx_evtchn, 1772 xennet_interrupt, 1773 0, queue->info->netdev->name, queue); 1774 if (err < 0) 1775 goto bind_fail; 1776 queue->rx_evtchn = queue->tx_evtchn; 1777 queue->rx_irq = queue->tx_irq = err; 1778 1779 return 0; 1780 1781 bind_fail: 1782 xenbus_free_evtchn(queue->info->xbdev, queue->tx_evtchn); 1783 queue->tx_evtchn = 0; 1784 fail: 1785 return err; 1786 } 1787 1788 static int setup_netfront_split(struct netfront_queue *queue) 1789 { 1790 int err; 1791 1792 err = xenbus_alloc_evtchn(queue->info->xbdev, &queue->tx_evtchn); 1793 if (err < 0) 1794 goto fail; 1795 err = xenbus_alloc_evtchn(queue->info->xbdev, &queue->rx_evtchn); 1796 if (err < 0) 1797 goto alloc_rx_evtchn_fail; 1798 1799 snprintf(queue->tx_irq_name, sizeof(queue->tx_irq_name), 1800 "%s-tx", queue->name); 1801 err = bind_evtchn_to_irqhandler(queue->tx_evtchn, 1802 xennet_tx_interrupt, 1803 0, queue->tx_irq_name, queue); 1804 if (err < 0) 1805 goto bind_tx_fail; 1806 queue->tx_irq = err; 1807 1808 snprintf(queue->rx_irq_name, sizeof(queue->rx_irq_name), 1809 "%s-rx", queue->name); 1810 err = bind_evtchn_to_irqhandler(queue->rx_evtchn, 1811 xennet_rx_interrupt, 1812 0, queue->rx_irq_name, queue); 1813 if (err < 0) 1814 goto bind_rx_fail; 1815 queue->rx_irq = err; 1816 1817 return 0; 1818 1819 bind_rx_fail: 1820 unbind_from_irqhandler(queue->tx_irq, queue); 1821 queue->tx_irq = 0; 1822 bind_tx_fail: 1823 xenbus_free_evtchn(queue->info->xbdev, queue->rx_evtchn); 1824 queue->rx_evtchn = 0; 1825 alloc_rx_evtchn_fail: 1826 xenbus_free_evtchn(queue->info->xbdev, queue->tx_evtchn); 1827 queue->tx_evtchn = 0; 1828 fail: 1829 return err; 1830 } 1831 1832 static int setup_netfront(struct xenbus_device *dev, 1833 struct netfront_queue *queue, unsigned int feature_split_evtchn) 1834 { 1835 struct xen_netif_tx_sring *txs; 1836 struct xen_netif_rx_sring *rxs; 1837 grant_ref_t gref; 1838 int err; 1839 1840 queue->tx_ring_ref = GRANT_INVALID_REF; 1841 queue->rx_ring_ref = GRANT_INVALID_REF; 1842 queue->rx.sring = NULL; 1843 queue->tx.sring = NULL; 1844 1845 txs = (struct xen_netif_tx_sring *)get_zeroed_page(GFP_NOIO | __GFP_HIGH); 1846 if (!txs) { 1847 err = -ENOMEM; 1848 xenbus_dev_fatal(dev, err, "allocating tx ring page"); 1849 goto fail; 1850 } 1851 SHARED_RING_INIT(txs); 1852 FRONT_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE); 1853 1854 err = xenbus_grant_ring(dev, txs, 1, &gref); 1855 if (err < 0) 1856 goto grant_tx_ring_fail; 1857 queue->tx_ring_ref = gref; 1858 1859 rxs = (struct xen_netif_rx_sring *)get_zeroed_page(GFP_NOIO | __GFP_HIGH); 1860 if (!rxs) { 1861 err = -ENOMEM; 1862 xenbus_dev_fatal(dev, err, "allocating rx ring page"); 1863 goto alloc_rx_ring_fail; 1864 } 1865 SHARED_RING_INIT(rxs); 1866 FRONT_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE); 1867 1868 err = xenbus_grant_ring(dev, rxs, 1, &gref); 1869 if (err < 0) 1870 goto grant_rx_ring_fail; 1871 queue->rx_ring_ref = gref; 1872 1873 if (feature_split_evtchn) 1874 err = setup_netfront_split(queue); 1875 /* setup single event channel if 1876 * a) feature-split-event-channels == 0 1877 * b) feature-split-event-channels == 1 but failed to setup 1878 */ 1879 if (!feature_split_evtchn || err) 1880 err = setup_netfront_single(queue); 1881 1882 if (err) 1883 goto alloc_evtchn_fail; 1884 1885 return 0; 1886 1887 /* If we fail to setup netfront, it is safe to just revoke access to 1888 * granted pages because backend is not accessing it at this point. 1889 */ 1890 alloc_evtchn_fail: 1891 gnttab_end_foreign_access_ref(queue->rx_ring_ref, 0); 1892 grant_rx_ring_fail: 1893 free_page((unsigned long)rxs); 1894 alloc_rx_ring_fail: 1895 gnttab_end_foreign_access_ref(queue->tx_ring_ref, 0); 1896 grant_tx_ring_fail: 1897 free_page((unsigned long)txs); 1898 fail: 1899 return err; 1900 } 1901 1902 /* Queue-specific initialisation 1903 * This used to be done in xennet_create_dev() but must now 1904 * be run per-queue. 1905 */ 1906 static int xennet_init_queue(struct netfront_queue *queue) 1907 { 1908 unsigned short i; 1909 int err = 0; 1910 char *devid; 1911 1912 spin_lock_init(&queue->tx_lock); 1913 spin_lock_init(&queue->rx_lock); 1914 1915 timer_setup(&queue->rx_refill_timer, rx_refill_timeout, 0); 1916 1917 devid = strrchr(queue->info->xbdev->nodename, '/') + 1; 1918 snprintf(queue->name, sizeof(queue->name), "vif%s-q%u", 1919 devid, queue->id); 1920 1921 /* Initialise tx_skb_freelist as a free chain containing every entry. */ 1922 queue->tx_skb_freelist = 0; 1923 queue->tx_pend_queue = TX_LINK_NONE; 1924 for (i = 0; i < NET_TX_RING_SIZE; i++) { 1925 queue->tx_link[i] = i + 1; 1926 queue->grant_tx_ref[i] = GRANT_INVALID_REF; 1927 queue->grant_tx_page[i] = NULL; 1928 } 1929 queue->tx_link[NET_TX_RING_SIZE - 1] = TX_LINK_NONE; 1930 1931 /* Clear out rx_skbs */ 1932 for (i = 0; i < NET_RX_RING_SIZE; i++) { 1933 queue->rx_skbs[i] = NULL; 1934 queue->grant_rx_ref[i] = GRANT_INVALID_REF; 1935 } 1936 1937 /* A grant for every tx ring slot */ 1938 if (gnttab_alloc_grant_references(NET_TX_RING_SIZE, 1939 &queue->gref_tx_head) < 0) { 1940 pr_alert("can't alloc tx grant refs\n"); 1941 err = -ENOMEM; 1942 goto exit; 1943 } 1944 1945 /* A grant for every rx ring slot */ 1946 if (gnttab_alloc_grant_references(NET_RX_RING_SIZE, 1947 &queue->gref_rx_head) < 0) { 1948 pr_alert("can't alloc rx grant refs\n"); 1949 err = -ENOMEM; 1950 goto exit_free_tx; 1951 } 1952 1953 return 0; 1954 1955 exit_free_tx: 1956 gnttab_free_grant_references(queue->gref_tx_head); 1957 exit: 1958 return err; 1959 } 1960 1961 static int write_queue_xenstore_keys(struct netfront_queue *queue, 1962 struct xenbus_transaction *xbt, int write_hierarchical) 1963 { 1964 /* Write the queue-specific keys into XenStore in the traditional 1965 * way for a single queue, or in a queue subkeys for multiple 1966 * queues. 1967 */ 1968 struct xenbus_device *dev = queue->info->xbdev; 1969 int err; 1970 const char *message; 1971 char *path; 1972 size_t pathsize; 1973 1974 /* Choose the correct place to write the keys */ 1975 if (write_hierarchical) { 1976 pathsize = strlen(dev->nodename) + 10; 1977 path = kzalloc(pathsize, GFP_KERNEL); 1978 if (!path) { 1979 err = -ENOMEM; 1980 message = "out of memory while writing ring references"; 1981 goto error; 1982 } 1983 snprintf(path, pathsize, "%s/queue-%u", 1984 dev->nodename, queue->id); 1985 } else { 1986 path = (char *)dev->nodename; 1987 } 1988 1989 /* Write ring references */ 1990 err = xenbus_printf(*xbt, path, "tx-ring-ref", "%u", 1991 queue->tx_ring_ref); 1992 if (err) { 1993 message = "writing tx-ring-ref"; 1994 goto error; 1995 } 1996 1997 err = xenbus_printf(*xbt, path, "rx-ring-ref", "%u", 1998 queue->rx_ring_ref); 1999 if (err) { 2000 message = "writing rx-ring-ref"; 2001 goto error; 2002 } 2003 2004 /* Write event channels; taking into account both shared 2005 * and split event channel scenarios. 2006 */ 2007 if (queue->tx_evtchn == queue->rx_evtchn) { 2008 /* Shared event channel */ 2009 err = xenbus_printf(*xbt, path, 2010 "event-channel", "%u", queue->tx_evtchn); 2011 if (err) { 2012 message = "writing event-channel"; 2013 goto error; 2014 } 2015 } else { 2016 /* Split event channels */ 2017 err = xenbus_printf(*xbt, path, 2018 "event-channel-tx", "%u", queue->tx_evtchn); 2019 if (err) { 2020 message = "writing event-channel-tx"; 2021 goto error; 2022 } 2023 2024 err = xenbus_printf(*xbt, path, 2025 "event-channel-rx", "%u", queue->rx_evtchn); 2026 if (err) { 2027 message = "writing event-channel-rx"; 2028 goto error; 2029 } 2030 } 2031 2032 if (write_hierarchical) 2033 kfree(path); 2034 return 0; 2035 2036 error: 2037 if (write_hierarchical) 2038 kfree(path); 2039 xenbus_dev_fatal(dev, err, "%s", message); 2040 return err; 2041 } 2042 2043 static void xennet_destroy_queues(struct netfront_info *info) 2044 { 2045 unsigned int i; 2046 2047 for (i = 0; i < info->netdev->real_num_tx_queues; i++) { 2048 struct netfront_queue *queue = &info->queues[i]; 2049 2050 if (netif_running(info->netdev)) 2051 napi_disable(&queue->napi); 2052 netif_napi_del(&queue->napi); 2053 } 2054 2055 kfree(info->queues); 2056 info->queues = NULL; 2057 } 2058 2059 2060 2061 static int xennet_create_page_pool(struct netfront_queue *queue) 2062 { 2063 int err; 2064 struct page_pool_params pp_params = { 2065 .order = 0, 2066 .flags = 0, 2067 .pool_size = NET_RX_RING_SIZE, 2068 .nid = NUMA_NO_NODE, 2069 .dev = &queue->info->netdev->dev, 2070 .offset = XDP_PACKET_HEADROOM, 2071 .max_len = XEN_PAGE_SIZE - XDP_PACKET_HEADROOM, 2072 }; 2073 2074 queue->page_pool = page_pool_create(&pp_params); 2075 if (IS_ERR(queue->page_pool)) { 2076 err = PTR_ERR(queue->page_pool); 2077 queue->page_pool = NULL; 2078 return err; 2079 } 2080 2081 err = xdp_rxq_info_reg(&queue->xdp_rxq, queue->info->netdev, 2082 queue->id, 0); 2083 if (err) { 2084 netdev_err(queue->info->netdev, "xdp_rxq_info_reg failed\n"); 2085 goto err_free_pp; 2086 } 2087 2088 err = xdp_rxq_info_reg_mem_model(&queue->xdp_rxq, 2089 MEM_TYPE_PAGE_POOL, queue->page_pool); 2090 if (err) { 2091 netdev_err(queue->info->netdev, "xdp_rxq_info_reg_mem_model failed\n"); 2092 goto err_unregister_rxq; 2093 } 2094 return 0; 2095 2096 err_unregister_rxq: 2097 xdp_rxq_info_unreg(&queue->xdp_rxq); 2098 err_free_pp: 2099 page_pool_destroy(queue->page_pool); 2100 queue->page_pool = NULL; 2101 return err; 2102 } 2103 2104 static int xennet_create_queues(struct netfront_info *info, 2105 unsigned int *num_queues) 2106 { 2107 unsigned int i; 2108 int ret; 2109 2110 info->queues = kcalloc(*num_queues, sizeof(struct netfront_queue), 2111 GFP_KERNEL); 2112 if (!info->queues) 2113 return -ENOMEM; 2114 2115 for (i = 0; i < *num_queues; i++) { 2116 struct netfront_queue *queue = &info->queues[i]; 2117 2118 queue->id = i; 2119 queue->info = info; 2120 2121 ret = xennet_init_queue(queue); 2122 if (ret < 0) { 2123 dev_warn(&info->xbdev->dev, 2124 "only created %d queues\n", i); 2125 *num_queues = i; 2126 break; 2127 } 2128 2129 /* use page pool recycling instead of buddy allocator */ 2130 ret = xennet_create_page_pool(queue); 2131 if (ret < 0) { 2132 dev_err(&info->xbdev->dev, "can't allocate page pool\n"); 2133 *num_queues = i; 2134 return ret; 2135 } 2136 2137 netif_napi_add(queue->info->netdev, &queue->napi, 2138 xennet_poll, 64); 2139 if (netif_running(info->netdev)) 2140 napi_enable(&queue->napi); 2141 } 2142 2143 netif_set_real_num_tx_queues(info->netdev, *num_queues); 2144 2145 if (*num_queues == 0) { 2146 dev_err(&info->xbdev->dev, "no queues\n"); 2147 return -EINVAL; 2148 } 2149 return 0; 2150 } 2151 2152 /* Common code used when first setting up, and when resuming. */ 2153 static int talk_to_netback(struct xenbus_device *dev, 2154 struct netfront_info *info) 2155 { 2156 const char *message; 2157 struct xenbus_transaction xbt; 2158 int err; 2159 unsigned int feature_split_evtchn; 2160 unsigned int i = 0; 2161 unsigned int max_queues = 0; 2162 struct netfront_queue *queue = NULL; 2163 unsigned int num_queues = 1; 2164 u8 addr[ETH_ALEN]; 2165 2166 info->netdev->irq = 0; 2167 2168 /* Check if backend supports multiple queues */ 2169 max_queues = xenbus_read_unsigned(info->xbdev->otherend, 2170 "multi-queue-max-queues", 1); 2171 num_queues = min(max_queues, xennet_max_queues); 2172 2173 /* Check feature-split-event-channels */ 2174 feature_split_evtchn = xenbus_read_unsigned(info->xbdev->otherend, 2175 "feature-split-event-channels", 0); 2176 2177 /* Read mac addr. */ 2178 err = xen_net_read_mac(dev, addr); 2179 if (err) { 2180 xenbus_dev_fatal(dev, err, "parsing %s/mac", dev->nodename); 2181 goto out_unlocked; 2182 } 2183 eth_hw_addr_set(info->netdev, addr); 2184 2185 info->netback_has_xdp_headroom = xenbus_read_unsigned(info->xbdev->otherend, 2186 "feature-xdp-headroom", 0); 2187 if (info->netback_has_xdp_headroom) { 2188 /* set the current xen-netfront xdp state */ 2189 err = talk_to_netback_xdp(info, info->netfront_xdp_enabled ? 2190 NETBACK_XDP_HEADROOM_ENABLE : 2191 NETBACK_XDP_HEADROOM_DISABLE); 2192 if (err) 2193 goto out_unlocked; 2194 } 2195 2196 rtnl_lock(); 2197 if (info->queues) 2198 xennet_destroy_queues(info); 2199 2200 /* For the case of a reconnect reset the "broken" indicator. */ 2201 info->broken = false; 2202 2203 err = xennet_create_queues(info, &num_queues); 2204 if (err < 0) { 2205 xenbus_dev_fatal(dev, err, "creating queues"); 2206 kfree(info->queues); 2207 info->queues = NULL; 2208 goto out; 2209 } 2210 rtnl_unlock(); 2211 2212 /* Create shared ring, alloc event channel -- for each queue */ 2213 for (i = 0; i < num_queues; ++i) { 2214 queue = &info->queues[i]; 2215 err = setup_netfront(dev, queue, feature_split_evtchn); 2216 if (err) 2217 goto destroy_ring; 2218 } 2219 2220 again: 2221 err = xenbus_transaction_start(&xbt); 2222 if (err) { 2223 xenbus_dev_fatal(dev, err, "starting transaction"); 2224 goto destroy_ring; 2225 } 2226 2227 if (xenbus_exists(XBT_NIL, 2228 info->xbdev->otherend, "multi-queue-max-queues")) { 2229 /* Write the number of queues */ 2230 err = xenbus_printf(xbt, dev->nodename, 2231 "multi-queue-num-queues", "%u", num_queues); 2232 if (err) { 2233 message = "writing multi-queue-num-queues"; 2234 goto abort_transaction_no_dev_fatal; 2235 } 2236 } 2237 2238 if (num_queues == 1) { 2239 err = write_queue_xenstore_keys(&info->queues[0], &xbt, 0); /* flat */ 2240 if (err) 2241 goto abort_transaction_no_dev_fatal; 2242 } else { 2243 /* Write the keys for each queue */ 2244 for (i = 0; i < num_queues; ++i) { 2245 queue = &info->queues[i]; 2246 err = write_queue_xenstore_keys(queue, &xbt, 1); /* hierarchical */ 2247 if (err) 2248 goto abort_transaction_no_dev_fatal; 2249 } 2250 } 2251 2252 /* The remaining keys are not queue-specific */ 2253 err = xenbus_printf(xbt, dev->nodename, "request-rx-copy", "%u", 2254 1); 2255 if (err) { 2256 message = "writing request-rx-copy"; 2257 goto abort_transaction; 2258 } 2259 2260 err = xenbus_printf(xbt, dev->nodename, "feature-rx-notify", "%d", 1); 2261 if (err) { 2262 message = "writing feature-rx-notify"; 2263 goto abort_transaction; 2264 } 2265 2266 err = xenbus_printf(xbt, dev->nodename, "feature-sg", "%d", 1); 2267 if (err) { 2268 message = "writing feature-sg"; 2269 goto abort_transaction; 2270 } 2271 2272 err = xenbus_printf(xbt, dev->nodename, "feature-gso-tcpv4", "%d", 1); 2273 if (err) { 2274 message = "writing feature-gso-tcpv4"; 2275 goto abort_transaction; 2276 } 2277 2278 err = xenbus_write(xbt, dev->nodename, "feature-gso-tcpv6", "1"); 2279 if (err) { 2280 message = "writing feature-gso-tcpv6"; 2281 goto abort_transaction; 2282 } 2283 2284 err = xenbus_write(xbt, dev->nodename, "feature-ipv6-csum-offload", 2285 "1"); 2286 if (err) { 2287 message = "writing feature-ipv6-csum-offload"; 2288 goto abort_transaction; 2289 } 2290 2291 err = xenbus_transaction_end(xbt, 0); 2292 if (err) { 2293 if (err == -EAGAIN) 2294 goto again; 2295 xenbus_dev_fatal(dev, err, "completing transaction"); 2296 goto destroy_ring; 2297 } 2298 2299 return 0; 2300 2301 abort_transaction: 2302 xenbus_dev_fatal(dev, err, "%s", message); 2303 abort_transaction_no_dev_fatal: 2304 xenbus_transaction_end(xbt, 1); 2305 destroy_ring: 2306 xennet_disconnect_backend(info); 2307 rtnl_lock(); 2308 xennet_destroy_queues(info); 2309 out: 2310 rtnl_unlock(); 2311 out_unlocked: 2312 device_unregister(&dev->dev); 2313 return err; 2314 } 2315 2316 static int xennet_connect(struct net_device *dev) 2317 { 2318 struct netfront_info *np = netdev_priv(dev); 2319 unsigned int num_queues = 0; 2320 int err; 2321 unsigned int j = 0; 2322 struct netfront_queue *queue = NULL; 2323 2324 if (!xenbus_read_unsigned(np->xbdev->otherend, "feature-rx-copy", 0)) { 2325 dev_info(&dev->dev, 2326 "backend does not support copying receive path\n"); 2327 return -ENODEV; 2328 } 2329 2330 err = talk_to_netback(np->xbdev, np); 2331 if (err) 2332 return err; 2333 if (np->netback_has_xdp_headroom) 2334 pr_info("backend supports XDP headroom\n"); 2335 2336 /* talk_to_netback() sets the correct number of queues */ 2337 num_queues = dev->real_num_tx_queues; 2338 2339 if (dev->reg_state == NETREG_UNINITIALIZED) { 2340 err = register_netdev(dev); 2341 if (err) { 2342 pr_warn("%s: register_netdev err=%d\n", __func__, err); 2343 device_unregister(&np->xbdev->dev); 2344 return err; 2345 } 2346 } 2347 2348 rtnl_lock(); 2349 netdev_update_features(dev); 2350 rtnl_unlock(); 2351 2352 /* 2353 * All public and private state should now be sane. Get 2354 * ready to start sending and receiving packets and give the driver 2355 * domain a kick because we've probably just requeued some 2356 * packets. 2357 */ 2358 netif_tx_lock_bh(np->netdev); 2359 netif_device_attach(np->netdev); 2360 netif_tx_unlock_bh(np->netdev); 2361 2362 netif_carrier_on(np->netdev); 2363 for (j = 0; j < num_queues; ++j) { 2364 queue = &np->queues[j]; 2365 2366 notify_remote_via_irq(queue->tx_irq); 2367 if (queue->tx_irq != queue->rx_irq) 2368 notify_remote_via_irq(queue->rx_irq); 2369 2370 spin_lock_irq(&queue->tx_lock); 2371 xennet_tx_buf_gc(queue); 2372 spin_unlock_irq(&queue->tx_lock); 2373 2374 spin_lock_bh(&queue->rx_lock); 2375 xennet_alloc_rx_buffers(queue); 2376 spin_unlock_bh(&queue->rx_lock); 2377 } 2378 2379 return 0; 2380 } 2381 2382 /* 2383 * Callback received when the backend's state changes. 2384 */ 2385 static void netback_changed(struct xenbus_device *dev, 2386 enum xenbus_state backend_state) 2387 { 2388 struct netfront_info *np = dev_get_drvdata(&dev->dev); 2389 struct net_device *netdev = np->netdev; 2390 2391 dev_dbg(&dev->dev, "%s\n", xenbus_strstate(backend_state)); 2392 2393 wake_up_all(&module_wq); 2394 2395 switch (backend_state) { 2396 case XenbusStateInitialising: 2397 case XenbusStateInitialised: 2398 case XenbusStateReconfiguring: 2399 case XenbusStateReconfigured: 2400 case XenbusStateUnknown: 2401 break; 2402 2403 case XenbusStateInitWait: 2404 if (dev->state != XenbusStateInitialising) 2405 break; 2406 if (xennet_connect(netdev) != 0) 2407 break; 2408 xenbus_switch_state(dev, XenbusStateConnected); 2409 break; 2410 2411 case XenbusStateConnected: 2412 netdev_notify_peers(netdev); 2413 break; 2414 2415 case XenbusStateClosed: 2416 if (dev->state == XenbusStateClosed) 2417 break; 2418 fallthrough; /* Missed the backend's CLOSING state */ 2419 case XenbusStateClosing: 2420 xenbus_frontend_closed(dev); 2421 break; 2422 } 2423 } 2424 2425 static const struct xennet_stat { 2426 char name[ETH_GSTRING_LEN]; 2427 u16 offset; 2428 } xennet_stats[] = { 2429 { 2430 "rx_gso_checksum_fixup", 2431 offsetof(struct netfront_info, rx_gso_checksum_fixup) 2432 }, 2433 }; 2434 2435 static int xennet_get_sset_count(struct net_device *dev, int string_set) 2436 { 2437 switch (string_set) { 2438 case ETH_SS_STATS: 2439 return ARRAY_SIZE(xennet_stats); 2440 default: 2441 return -EINVAL; 2442 } 2443 } 2444 2445 static void xennet_get_ethtool_stats(struct net_device *dev, 2446 struct ethtool_stats *stats, u64 * data) 2447 { 2448 void *np = netdev_priv(dev); 2449 int i; 2450 2451 for (i = 0; i < ARRAY_SIZE(xennet_stats); i++) 2452 data[i] = atomic_read((atomic_t *)(np + xennet_stats[i].offset)); 2453 } 2454 2455 static void xennet_get_strings(struct net_device *dev, u32 stringset, u8 * data) 2456 { 2457 int i; 2458 2459 switch (stringset) { 2460 case ETH_SS_STATS: 2461 for (i = 0; i < ARRAY_SIZE(xennet_stats); i++) 2462 memcpy(data + i * ETH_GSTRING_LEN, 2463 xennet_stats[i].name, ETH_GSTRING_LEN); 2464 break; 2465 } 2466 } 2467 2468 static const struct ethtool_ops xennet_ethtool_ops = 2469 { 2470 .get_link = ethtool_op_get_link, 2471 2472 .get_sset_count = xennet_get_sset_count, 2473 .get_ethtool_stats = xennet_get_ethtool_stats, 2474 .get_strings = xennet_get_strings, 2475 .get_ts_info = ethtool_op_get_ts_info, 2476 }; 2477 2478 #ifdef CONFIG_SYSFS 2479 static ssize_t show_rxbuf(struct device *dev, 2480 struct device_attribute *attr, char *buf) 2481 { 2482 return sprintf(buf, "%lu\n", NET_RX_RING_SIZE); 2483 } 2484 2485 static ssize_t store_rxbuf(struct device *dev, 2486 struct device_attribute *attr, 2487 const char *buf, size_t len) 2488 { 2489 char *endp; 2490 2491 if (!capable(CAP_NET_ADMIN)) 2492 return -EPERM; 2493 2494 simple_strtoul(buf, &endp, 0); 2495 if (endp == buf) 2496 return -EBADMSG; 2497 2498 /* rxbuf_min and rxbuf_max are no longer configurable. */ 2499 2500 return len; 2501 } 2502 2503 static DEVICE_ATTR(rxbuf_min, 0644, show_rxbuf, store_rxbuf); 2504 static DEVICE_ATTR(rxbuf_max, 0644, show_rxbuf, store_rxbuf); 2505 static DEVICE_ATTR(rxbuf_cur, 0444, show_rxbuf, NULL); 2506 2507 static struct attribute *xennet_dev_attrs[] = { 2508 &dev_attr_rxbuf_min.attr, 2509 &dev_attr_rxbuf_max.attr, 2510 &dev_attr_rxbuf_cur.attr, 2511 NULL 2512 }; 2513 2514 static const struct attribute_group xennet_dev_group = { 2515 .attrs = xennet_dev_attrs 2516 }; 2517 #endif /* CONFIG_SYSFS */ 2518 2519 static void xennet_bus_close(struct xenbus_device *dev) 2520 { 2521 int ret; 2522 2523 if (xenbus_read_driver_state(dev->otherend) == XenbusStateClosed) 2524 return; 2525 do { 2526 xenbus_switch_state(dev, XenbusStateClosing); 2527 ret = wait_event_timeout(module_wq, 2528 xenbus_read_driver_state(dev->otherend) == 2529 XenbusStateClosing || 2530 xenbus_read_driver_state(dev->otherend) == 2531 XenbusStateClosed || 2532 xenbus_read_driver_state(dev->otherend) == 2533 XenbusStateUnknown, 2534 XENNET_TIMEOUT); 2535 } while (!ret); 2536 2537 if (xenbus_read_driver_state(dev->otherend) == XenbusStateClosed) 2538 return; 2539 2540 do { 2541 xenbus_switch_state(dev, XenbusStateClosed); 2542 ret = wait_event_timeout(module_wq, 2543 xenbus_read_driver_state(dev->otherend) == 2544 XenbusStateClosed || 2545 xenbus_read_driver_state(dev->otherend) == 2546 XenbusStateUnknown, 2547 XENNET_TIMEOUT); 2548 } while (!ret); 2549 } 2550 2551 static int xennet_remove(struct xenbus_device *dev) 2552 { 2553 struct netfront_info *info = dev_get_drvdata(&dev->dev); 2554 2555 xennet_bus_close(dev); 2556 xennet_disconnect_backend(info); 2557 2558 if (info->netdev->reg_state == NETREG_REGISTERED) 2559 unregister_netdev(info->netdev); 2560 2561 if (info->queues) { 2562 rtnl_lock(); 2563 xennet_destroy_queues(info); 2564 rtnl_unlock(); 2565 } 2566 xennet_free_netdev(info->netdev); 2567 2568 return 0; 2569 } 2570 2571 static const struct xenbus_device_id netfront_ids[] = { 2572 { "vif" }, 2573 { "" } 2574 }; 2575 2576 static struct xenbus_driver netfront_driver = { 2577 .ids = netfront_ids, 2578 .probe = netfront_probe, 2579 .remove = xennet_remove, 2580 .resume = netfront_resume, 2581 .otherend_changed = netback_changed, 2582 }; 2583 2584 static int __init netif_init(void) 2585 { 2586 if (!xen_domain()) 2587 return -ENODEV; 2588 2589 if (!xen_has_pv_nic_devices()) 2590 return -ENODEV; 2591 2592 pr_info("Initialising Xen virtual ethernet driver\n"); 2593 2594 /* Allow as many queues as there are CPUs inut max. 8 if user has not 2595 * specified a value. 2596 */ 2597 if (xennet_max_queues == 0) 2598 xennet_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT, 2599 num_online_cpus()); 2600 2601 return xenbus_register_frontend(&netfront_driver); 2602 } 2603 module_init(netif_init); 2604 2605 2606 static void __exit netif_exit(void) 2607 { 2608 xenbus_unregister_driver(&netfront_driver); 2609 } 2610 module_exit(netif_exit); 2611 2612 MODULE_DESCRIPTION("Xen virtual network device frontend"); 2613 MODULE_LICENSE("GPL"); 2614 MODULE_ALIAS("xen:vif"); 2615 MODULE_ALIAS("xennet"); 2616