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