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 #include <linux/module.h> 33 #include <linux/kernel.h> 34 #include <linux/netdevice.h> 35 #include <linux/etherdevice.h> 36 #include <linux/skbuff.h> 37 #include <linux/ethtool.h> 38 #include <linux/if_ether.h> 39 #include <linux/tcp.h> 40 #include <linux/udp.h> 41 #include <linux/moduleparam.h> 42 #include <linux/mm.h> 43 #include <linux/slab.h> 44 #include <net/ip.h> 45 46 #include <asm/xen/page.h> 47 #include <xen/xen.h> 48 #include <xen/xenbus.h> 49 #include <xen/events.h> 50 #include <xen/page.h> 51 #include <xen/platform_pci.h> 52 #include <xen/grant_table.h> 53 54 #include <xen/interface/io/netif.h> 55 #include <xen/interface/memory.h> 56 #include <xen/interface/grant_table.h> 57 58 static const struct ethtool_ops xennet_ethtool_ops; 59 60 struct netfront_cb { 61 int pull_to; 62 }; 63 64 #define NETFRONT_SKB_CB(skb) ((struct netfront_cb *)((skb)->cb)) 65 66 #define RX_COPY_THRESHOLD 256 67 68 #define GRANT_INVALID_REF 0 69 70 #define NET_TX_RING_SIZE __CONST_RING_SIZE(xen_netif_tx, PAGE_SIZE) 71 #define NET_RX_RING_SIZE __CONST_RING_SIZE(xen_netif_rx, PAGE_SIZE) 72 #define TX_MAX_TARGET min_t(int, NET_TX_RING_SIZE, 256) 73 74 struct netfront_stats { 75 u64 rx_packets; 76 u64 tx_packets; 77 u64 rx_bytes; 78 u64 tx_bytes; 79 struct u64_stats_sync syncp; 80 }; 81 82 struct netfront_info { 83 struct list_head list; 84 struct net_device *netdev; 85 86 struct napi_struct napi; 87 88 unsigned int evtchn; 89 struct xenbus_device *xbdev; 90 91 spinlock_t tx_lock; 92 struct xen_netif_tx_front_ring tx; 93 int tx_ring_ref; 94 95 /* 96 * {tx,rx}_skbs store outstanding skbuffs. Free tx_skb entries 97 * are linked from tx_skb_freelist through skb_entry.link. 98 * 99 * NB. Freelist index entries are always going to be less than 100 * PAGE_OFFSET, whereas pointers to skbs will always be equal or 101 * greater than PAGE_OFFSET: we use this property to distinguish 102 * them. 103 */ 104 union skb_entry { 105 struct sk_buff *skb; 106 unsigned long link; 107 } tx_skbs[NET_TX_RING_SIZE]; 108 grant_ref_t gref_tx_head; 109 grant_ref_t grant_tx_ref[NET_TX_RING_SIZE]; 110 unsigned tx_skb_freelist; 111 112 spinlock_t rx_lock ____cacheline_aligned_in_smp; 113 struct xen_netif_rx_front_ring rx; 114 int rx_ring_ref; 115 116 /* Receive-ring batched refills. */ 117 #define RX_MIN_TARGET 8 118 #define RX_DFL_MIN_TARGET 64 119 #define RX_MAX_TARGET min_t(int, NET_RX_RING_SIZE, 256) 120 unsigned rx_min_target, rx_max_target, rx_target; 121 struct sk_buff_head rx_batch; 122 123 struct timer_list rx_refill_timer; 124 125 struct sk_buff *rx_skbs[NET_RX_RING_SIZE]; 126 grant_ref_t gref_rx_head; 127 grant_ref_t grant_rx_ref[NET_RX_RING_SIZE]; 128 129 unsigned long rx_pfn_array[NET_RX_RING_SIZE]; 130 struct multicall_entry rx_mcl[NET_RX_RING_SIZE+1]; 131 struct mmu_update rx_mmu[NET_RX_RING_SIZE]; 132 133 /* Statistics */ 134 struct netfront_stats __percpu *stats; 135 136 unsigned long rx_gso_checksum_fixup; 137 }; 138 139 struct netfront_rx_info { 140 struct xen_netif_rx_response rx; 141 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1]; 142 }; 143 144 static void skb_entry_set_link(union skb_entry *list, unsigned short id) 145 { 146 list->link = id; 147 } 148 149 static int skb_entry_is_link(const union skb_entry *list) 150 { 151 BUILD_BUG_ON(sizeof(list->skb) != sizeof(list->link)); 152 return (unsigned long)list->skb < PAGE_OFFSET; 153 } 154 155 /* 156 * Access macros for acquiring freeing slots in tx_skbs[]. 157 */ 158 159 static void add_id_to_freelist(unsigned *head, union skb_entry *list, 160 unsigned short id) 161 { 162 skb_entry_set_link(&list[id], *head); 163 *head = id; 164 } 165 166 static unsigned short get_id_from_freelist(unsigned *head, 167 union skb_entry *list) 168 { 169 unsigned int id = *head; 170 *head = list[id].link; 171 return id; 172 } 173 174 static int xennet_rxidx(RING_IDX idx) 175 { 176 return idx & (NET_RX_RING_SIZE - 1); 177 } 178 179 static struct sk_buff *xennet_get_rx_skb(struct netfront_info *np, 180 RING_IDX ri) 181 { 182 int i = xennet_rxidx(ri); 183 struct sk_buff *skb = np->rx_skbs[i]; 184 np->rx_skbs[i] = NULL; 185 return skb; 186 } 187 188 static grant_ref_t xennet_get_rx_ref(struct netfront_info *np, 189 RING_IDX ri) 190 { 191 int i = xennet_rxidx(ri); 192 grant_ref_t ref = np->grant_rx_ref[i]; 193 np->grant_rx_ref[i] = GRANT_INVALID_REF; 194 return ref; 195 } 196 197 #ifdef CONFIG_SYSFS 198 static int xennet_sysfs_addif(struct net_device *netdev); 199 static void xennet_sysfs_delif(struct net_device *netdev); 200 #else /* !CONFIG_SYSFS */ 201 #define xennet_sysfs_addif(dev) (0) 202 #define xennet_sysfs_delif(dev) do { } while (0) 203 #endif 204 205 static bool xennet_can_sg(struct net_device *dev) 206 { 207 return dev->features & NETIF_F_SG; 208 } 209 210 211 static void rx_refill_timeout(unsigned long data) 212 { 213 struct net_device *dev = (struct net_device *)data; 214 struct netfront_info *np = netdev_priv(dev); 215 napi_schedule(&np->napi); 216 } 217 218 static int netfront_tx_slot_available(struct netfront_info *np) 219 { 220 return (np->tx.req_prod_pvt - np->tx.rsp_cons) < 221 (TX_MAX_TARGET - MAX_SKB_FRAGS - 2); 222 } 223 224 static void xennet_maybe_wake_tx(struct net_device *dev) 225 { 226 struct netfront_info *np = netdev_priv(dev); 227 228 if (unlikely(netif_queue_stopped(dev)) && 229 netfront_tx_slot_available(np) && 230 likely(netif_running(dev))) 231 netif_wake_queue(dev); 232 } 233 234 static void xennet_alloc_rx_buffers(struct net_device *dev) 235 { 236 unsigned short id; 237 struct netfront_info *np = netdev_priv(dev); 238 struct sk_buff *skb; 239 struct page *page; 240 int i, batch_target, notify; 241 RING_IDX req_prod = np->rx.req_prod_pvt; 242 grant_ref_t ref; 243 unsigned long pfn; 244 void *vaddr; 245 struct xen_netif_rx_request *req; 246 247 if (unlikely(!netif_carrier_ok(dev))) 248 return; 249 250 /* 251 * Allocate skbuffs greedily, even though we batch updates to the 252 * receive ring. This creates a less bursty demand on the memory 253 * allocator, so should reduce the chance of failed allocation requests 254 * both for ourself and for other kernel subsystems. 255 */ 256 batch_target = np->rx_target - (req_prod - np->rx.rsp_cons); 257 for (i = skb_queue_len(&np->rx_batch); i < batch_target; i++) { 258 skb = __netdev_alloc_skb(dev, RX_COPY_THRESHOLD + NET_IP_ALIGN, 259 GFP_ATOMIC | __GFP_NOWARN); 260 if (unlikely(!skb)) 261 goto no_skb; 262 263 /* Align ip header to a 16 bytes boundary */ 264 skb_reserve(skb, NET_IP_ALIGN); 265 266 page = alloc_page(GFP_ATOMIC | __GFP_NOWARN); 267 if (!page) { 268 kfree_skb(skb); 269 no_skb: 270 /* Any skbuffs queued for refill? Force them out. */ 271 if (i != 0) 272 goto refill; 273 /* Could not allocate any skbuffs. Try again later. */ 274 mod_timer(&np->rx_refill_timer, 275 jiffies + (HZ/10)); 276 break; 277 } 278 279 __skb_fill_page_desc(skb, 0, page, 0, 0); 280 skb_shinfo(skb)->nr_frags = 1; 281 __skb_queue_tail(&np->rx_batch, skb); 282 } 283 284 /* Is the batch large enough to be worthwhile? */ 285 if (i < (np->rx_target/2)) { 286 if (req_prod > np->rx.sring->req_prod) 287 goto push; 288 return; 289 } 290 291 /* Adjust our fill target if we risked running out of buffers. */ 292 if (((req_prod - np->rx.sring->rsp_prod) < (np->rx_target / 4)) && 293 ((np->rx_target *= 2) > np->rx_max_target)) 294 np->rx_target = np->rx_max_target; 295 296 refill: 297 for (i = 0; ; i++) { 298 skb = __skb_dequeue(&np->rx_batch); 299 if (skb == NULL) 300 break; 301 302 skb->dev = dev; 303 304 id = xennet_rxidx(req_prod + i); 305 306 BUG_ON(np->rx_skbs[id]); 307 np->rx_skbs[id] = skb; 308 309 ref = gnttab_claim_grant_reference(&np->gref_rx_head); 310 BUG_ON((signed short)ref < 0); 311 np->grant_rx_ref[id] = ref; 312 313 pfn = page_to_pfn(skb_frag_page(&skb_shinfo(skb)->frags[0])); 314 vaddr = page_address(skb_frag_page(&skb_shinfo(skb)->frags[0])); 315 316 req = RING_GET_REQUEST(&np->rx, req_prod + i); 317 gnttab_grant_foreign_access_ref(ref, 318 np->xbdev->otherend_id, 319 pfn_to_mfn(pfn), 320 0); 321 322 req->id = id; 323 req->gref = ref; 324 } 325 326 wmb(); /* barrier so backend seens requests */ 327 328 /* Above is a suitable barrier to ensure backend will see requests. */ 329 np->rx.req_prod_pvt = req_prod + i; 330 push: 331 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&np->rx, notify); 332 if (notify) 333 notify_remote_via_irq(np->netdev->irq); 334 } 335 336 static int xennet_open(struct net_device *dev) 337 { 338 struct netfront_info *np = netdev_priv(dev); 339 340 napi_enable(&np->napi); 341 342 spin_lock_bh(&np->rx_lock); 343 if (netif_carrier_ok(dev)) { 344 xennet_alloc_rx_buffers(dev); 345 np->rx.sring->rsp_event = np->rx.rsp_cons + 1; 346 if (RING_HAS_UNCONSUMED_RESPONSES(&np->rx)) 347 napi_schedule(&np->napi); 348 } 349 spin_unlock_bh(&np->rx_lock); 350 351 netif_start_queue(dev); 352 353 return 0; 354 } 355 356 static void xennet_tx_buf_gc(struct net_device *dev) 357 { 358 RING_IDX cons, prod; 359 unsigned short id; 360 struct netfront_info *np = netdev_priv(dev); 361 struct sk_buff *skb; 362 363 BUG_ON(!netif_carrier_ok(dev)); 364 365 do { 366 prod = np->tx.sring->rsp_prod; 367 rmb(); /* Ensure we see responses up to 'rp'. */ 368 369 for (cons = np->tx.rsp_cons; cons != prod; cons++) { 370 struct xen_netif_tx_response *txrsp; 371 372 txrsp = RING_GET_RESPONSE(&np->tx, cons); 373 if (txrsp->status == XEN_NETIF_RSP_NULL) 374 continue; 375 376 id = txrsp->id; 377 skb = np->tx_skbs[id].skb; 378 if (unlikely(gnttab_query_foreign_access( 379 np->grant_tx_ref[id]) != 0)) { 380 printk(KERN_ALERT "xennet_tx_buf_gc: warning " 381 "-- grant still in use by backend " 382 "domain.\n"); 383 BUG(); 384 } 385 gnttab_end_foreign_access_ref( 386 np->grant_tx_ref[id], GNTMAP_readonly); 387 gnttab_release_grant_reference( 388 &np->gref_tx_head, np->grant_tx_ref[id]); 389 np->grant_tx_ref[id] = GRANT_INVALID_REF; 390 add_id_to_freelist(&np->tx_skb_freelist, np->tx_skbs, id); 391 dev_kfree_skb_irq(skb); 392 } 393 394 np->tx.rsp_cons = prod; 395 396 /* 397 * Set a new event, then check for race with update of tx_cons. 398 * Note that it is essential to schedule a callback, no matter 399 * how few buffers are pending. Even if there is space in the 400 * transmit ring, higher layers may be blocked because too much 401 * data is outstanding: in such cases notification from Xen is 402 * likely to be the only kick that we'll get. 403 */ 404 np->tx.sring->rsp_event = 405 prod + ((np->tx.sring->req_prod - prod) >> 1) + 1; 406 mb(); /* update shared area */ 407 } while ((cons == prod) && (prod != np->tx.sring->rsp_prod)); 408 409 xennet_maybe_wake_tx(dev); 410 } 411 412 static void xennet_make_frags(struct sk_buff *skb, struct net_device *dev, 413 struct xen_netif_tx_request *tx) 414 { 415 struct netfront_info *np = netdev_priv(dev); 416 char *data = skb->data; 417 unsigned long mfn; 418 RING_IDX prod = np->tx.req_prod_pvt; 419 int frags = skb_shinfo(skb)->nr_frags; 420 unsigned int offset = offset_in_page(data); 421 unsigned int len = skb_headlen(skb); 422 unsigned int id; 423 grant_ref_t ref; 424 int i; 425 426 /* While the header overlaps a page boundary (including being 427 larger than a page), split it it into page-sized chunks. */ 428 while (len > PAGE_SIZE - offset) { 429 tx->size = PAGE_SIZE - offset; 430 tx->flags |= XEN_NETTXF_more_data; 431 len -= tx->size; 432 data += tx->size; 433 offset = 0; 434 435 id = get_id_from_freelist(&np->tx_skb_freelist, np->tx_skbs); 436 np->tx_skbs[id].skb = skb_get(skb); 437 tx = RING_GET_REQUEST(&np->tx, prod++); 438 tx->id = id; 439 ref = gnttab_claim_grant_reference(&np->gref_tx_head); 440 BUG_ON((signed short)ref < 0); 441 442 mfn = virt_to_mfn(data); 443 gnttab_grant_foreign_access_ref(ref, np->xbdev->otherend_id, 444 mfn, GNTMAP_readonly); 445 446 tx->gref = np->grant_tx_ref[id] = ref; 447 tx->offset = offset; 448 tx->size = len; 449 tx->flags = 0; 450 } 451 452 /* Grant backend access to each skb fragment page. */ 453 for (i = 0; i < frags; i++) { 454 skb_frag_t *frag = skb_shinfo(skb)->frags + i; 455 struct page *page = skb_frag_page(frag); 456 457 len = skb_frag_size(frag); 458 offset = frag->page_offset; 459 460 /* Data must not cross a page boundary. */ 461 BUG_ON(len + offset > PAGE_SIZE<<compound_order(page)); 462 463 /* Skip unused frames from start of page */ 464 page += offset >> PAGE_SHIFT; 465 offset &= ~PAGE_MASK; 466 467 while (len > 0) { 468 unsigned long bytes; 469 470 BUG_ON(offset >= PAGE_SIZE); 471 472 bytes = PAGE_SIZE - offset; 473 if (bytes > len) 474 bytes = len; 475 476 tx->flags |= XEN_NETTXF_more_data; 477 478 id = get_id_from_freelist(&np->tx_skb_freelist, 479 np->tx_skbs); 480 np->tx_skbs[id].skb = skb_get(skb); 481 tx = RING_GET_REQUEST(&np->tx, prod++); 482 tx->id = id; 483 ref = gnttab_claim_grant_reference(&np->gref_tx_head); 484 BUG_ON((signed short)ref < 0); 485 486 mfn = pfn_to_mfn(page_to_pfn(page)); 487 gnttab_grant_foreign_access_ref(ref, 488 np->xbdev->otherend_id, 489 mfn, GNTMAP_readonly); 490 491 tx->gref = np->grant_tx_ref[id] = ref; 492 tx->offset = offset; 493 tx->size = bytes; 494 tx->flags = 0; 495 496 offset += bytes; 497 len -= bytes; 498 499 /* Next frame */ 500 if (offset == PAGE_SIZE && len) { 501 BUG_ON(!PageCompound(page)); 502 page++; 503 offset = 0; 504 } 505 } 506 } 507 508 np->tx.req_prod_pvt = prod; 509 } 510 511 /* 512 * Count how many ring slots are required to send the frags of this 513 * skb. Each frag might be a compound page. 514 */ 515 static int xennet_count_skb_frag_slots(struct sk_buff *skb) 516 { 517 int i, frags = skb_shinfo(skb)->nr_frags; 518 int pages = 0; 519 520 for (i = 0; i < frags; i++) { 521 skb_frag_t *frag = skb_shinfo(skb)->frags + i; 522 unsigned long size = skb_frag_size(frag); 523 unsigned long offset = frag->page_offset; 524 525 /* Skip unused frames from start of page */ 526 offset &= ~PAGE_MASK; 527 528 pages += PFN_UP(offset + size); 529 } 530 531 return pages; 532 } 533 534 static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev) 535 { 536 unsigned short id; 537 struct netfront_info *np = netdev_priv(dev); 538 struct netfront_stats *stats = this_cpu_ptr(np->stats); 539 struct xen_netif_tx_request *tx; 540 struct xen_netif_extra_info *extra; 541 char *data = skb->data; 542 RING_IDX i; 543 grant_ref_t ref; 544 unsigned long mfn; 545 int notify; 546 int slots; 547 unsigned int offset = offset_in_page(data); 548 unsigned int len = skb_headlen(skb); 549 unsigned long flags; 550 551 slots = DIV_ROUND_UP(offset + len, PAGE_SIZE) + 552 xennet_count_skb_frag_slots(skb); 553 if (unlikely(slots > MAX_SKB_FRAGS + 1)) { 554 net_alert_ratelimited( 555 "xennet: skb rides the rocket: %d slots\n", slots); 556 goto drop; 557 } 558 559 spin_lock_irqsave(&np->tx_lock, flags); 560 561 if (unlikely(!netif_carrier_ok(dev) || 562 (slots > 1 && !xennet_can_sg(dev)) || 563 netif_needs_gso(skb, netif_skb_features(skb)))) { 564 spin_unlock_irqrestore(&np->tx_lock, flags); 565 goto drop; 566 } 567 568 i = np->tx.req_prod_pvt; 569 570 id = get_id_from_freelist(&np->tx_skb_freelist, np->tx_skbs); 571 np->tx_skbs[id].skb = skb; 572 573 tx = RING_GET_REQUEST(&np->tx, i); 574 575 tx->id = id; 576 ref = gnttab_claim_grant_reference(&np->gref_tx_head); 577 BUG_ON((signed short)ref < 0); 578 mfn = virt_to_mfn(data); 579 gnttab_grant_foreign_access_ref( 580 ref, np->xbdev->otherend_id, mfn, GNTMAP_readonly); 581 tx->gref = np->grant_tx_ref[id] = ref; 582 tx->offset = offset; 583 tx->size = len; 584 extra = NULL; 585 586 tx->flags = 0; 587 if (skb->ip_summed == CHECKSUM_PARTIAL) 588 /* local packet? */ 589 tx->flags |= XEN_NETTXF_csum_blank | XEN_NETTXF_data_validated; 590 else if (skb->ip_summed == CHECKSUM_UNNECESSARY) 591 /* remote but checksummed. */ 592 tx->flags |= XEN_NETTXF_data_validated; 593 594 if (skb_shinfo(skb)->gso_size) { 595 struct xen_netif_extra_info *gso; 596 597 gso = (struct xen_netif_extra_info *) 598 RING_GET_REQUEST(&np->tx, ++i); 599 600 if (extra) 601 extra->flags |= XEN_NETIF_EXTRA_FLAG_MORE; 602 else 603 tx->flags |= XEN_NETTXF_extra_info; 604 605 gso->u.gso.size = skb_shinfo(skb)->gso_size; 606 gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4; 607 gso->u.gso.pad = 0; 608 gso->u.gso.features = 0; 609 610 gso->type = XEN_NETIF_EXTRA_TYPE_GSO; 611 gso->flags = 0; 612 extra = gso; 613 } 614 615 np->tx.req_prod_pvt = i + 1; 616 617 xennet_make_frags(skb, dev, tx); 618 tx->size = skb->len; 619 620 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&np->tx, notify); 621 if (notify) 622 notify_remote_via_irq(np->netdev->irq); 623 624 u64_stats_update_begin(&stats->syncp); 625 stats->tx_bytes += skb->len; 626 stats->tx_packets++; 627 u64_stats_update_end(&stats->syncp); 628 629 /* Note: It is not safe to access skb after xennet_tx_buf_gc()! */ 630 xennet_tx_buf_gc(dev); 631 632 if (!netfront_tx_slot_available(np)) 633 netif_stop_queue(dev); 634 635 spin_unlock_irqrestore(&np->tx_lock, flags); 636 637 return NETDEV_TX_OK; 638 639 drop: 640 dev->stats.tx_dropped++; 641 dev_kfree_skb(skb); 642 return NETDEV_TX_OK; 643 } 644 645 static int xennet_close(struct net_device *dev) 646 { 647 struct netfront_info *np = netdev_priv(dev); 648 netif_stop_queue(np->netdev); 649 napi_disable(&np->napi); 650 return 0; 651 } 652 653 static void xennet_move_rx_slot(struct netfront_info *np, struct sk_buff *skb, 654 grant_ref_t ref) 655 { 656 int new = xennet_rxidx(np->rx.req_prod_pvt); 657 658 BUG_ON(np->rx_skbs[new]); 659 np->rx_skbs[new] = skb; 660 np->grant_rx_ref[new] = ref; 661 RING_GET_REQUEST(&np->rx, np->rx.req_prod_pvt)->id = new; 662 RING_GET_REQUEST(&np->rx, np->rx.req_prod_pvt)->gref = ref; 663 np->rx.req_prod_pvt++; 664 } 665 666 static int xennet_get_extras(struct netfront_info *np, 667 struct xen_netif_extra_info *extras, 668 RING_IDX rp) 669 670 { 671 struct xen_netif_extra_info *extra; 672 struct device *dev = &np->netdev->dev; 673 RING_IDX cons = np->rx.rsp_cons; 674 int err = 0; 675 676 do { 677 struct sk_buff *skb; 678 grant_ref_t ref; 679 680 if (unlikely(cons + 1 == rp)) { 681 if (net_ratelimit()) 682 dev_warn(dev, "Missing extra info\n"); 683 err = -EBADR; 684 break; 685 } 686 687 extra = (struct xen_netif_extra_info *) 688 RING_GET_RESPONSE(&np->rx, ++cons); 689 690 if (unlikely(!extra->type || 691 extra->type >= XEN_NETIF_EXTRA_TYPE_MAX)) { 692 if (net_ratelimit()) 693 dev_warn(dev, "Invalid extra type: %d\n", 694 extra->type); 695 err = -EINVAL; 696 } else { 697 memcpy(&extras[extra->type - 1], extra, 698 sizeof(*extra)); 699 } 700 701 skb = xennet_get_rx_skb(np, cons); 702 ref = xennet_get_rx_ref(np, cons); 703 xennet_move_rx_slot(np, skb, ref); 704 } while (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE); 705 706 np->rx.rsp_cons = cons; 707 return err; 708 } 709 710 static int xennet_get_responses(struct netfront_info *np, 711 struct netfront_rx_info *rinfo, RING_IDX rp, 712 struct sk_buff_head *list) 713 { 714 struct xen_netif_rx_response *rx = &rinfo->rx; 715 struct xen_netif_extra_info *extras = rinfo->extras; 716 struct device *dev = &np->netdev->dev; 717 RING_IDX cons = np->rx.rsp_cons; 718 struct sk_buff *skb = xennet_get_rx_skb(np, cons); 719 grant_ref_t ref = xennet_get_rx_ref(np, cons); 720 int max = MAX_SKB_FRAGS + (rx->status <= RX_COPY_THRESHOLD); 721 int frags = 1; 722 int err = 0; 723 unsigned long ret; 724 725 if (rx->flags & XEN_NETRXF_extra_info) { 726 err = xennet_get_extras(np, extras, rp); 727 cons = np->rx.rsp_cons; 728 } 729 730 for (;;) { 731 if (unlikely(rx->status < 0 || 732 rx->offset + rx->status > PAGE_SIZE)) { 733 if (net_ratelimit()) 734 dev_warn(dev, "rx->offset: %x, size: %u\n", 735 rx->offset, rx->status); 736 xennet_move_rx_slot(np, skb, ref); 737 err = -EINVAL; 738 goto next; 739 } 740 741 /* 742 * This definitely indicates a bug, either in this driver or in 743 * the backend driver. In future this should flag the bad 744 * situation to the system controller to reboot the backed. 745 */ 746 if (ref == GRANT_INVALID_REF) { 747 if (net_ratelimit()) 748 dev_warn(dev, "Bad rx response id %d.\n", 749 rx->id); 750 err = -EINVAL; 751 goto next; 752 } 753 754 ret = gnttab_end_foreign_access_ref(ref, 0); 755 BUG_ON(!ret); 756 757 gnttab_release_grant_reference(&np->gref_rx_head, ref); 758 759 __skb_queue_tail(list, skb); 760 761 next: 762 if (!(rx->flags & XEN_NETRXF_more_data)) 763 break; 764 765 if (cons + frags == rp) { 766 if (net_ratelimit()) 767 dev_warn(dev, "Need more frags\n"); 768 err = -ENOENT; 769 break; 770 } 771 772 rx = RING_GET_RESPONSE(&np->rx, cons + frags); 773 skb = xennet_get_rx_skb(np, cons + frags); 774 ref = xennet_get_rx_ref(np, cons + frags); 775 frags++; 776 } 777 778 if (unlikely(frags > max)) { 779 if (net_ratelimit()) 780 dev_warn(dev, "Too many frags\n"); 781 err = -E2BIG; 782 } 783 784 if (unlikely(err)) 785 np->rx.rsp_cons = cons + frags; 786 787 return err; 788 } 789 790 static int xennet_set_skb_gso(struct sk_buff *skb, 791 struct xen_netif_extra_info *gso) 792 { 793 if (!gso->u.gso.size) { 794 if (net_ratelimit()) 795 printk(KERN_WARNING "GSO size must not be zero.\n"); 796 return -EINVAL; 797 } 798 799 /* Currently only TCPv4 S.O. is supported. */ 800 if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) { 801 if (net_ratelimit()) 802 printk(KERN_WARNING "Bad GSO type %d.\n", gso->u.gso.type); 803 return -EINVAL; 804 } 805 806 skb_shinfo(skb)->gso_size = gso->u.gso.size; 807 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4; 808 809 /* Header must be checked, and gso_segs computed. */ 810 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY; 811 skb_shinfo(skb)->gso_segs = 0; 812 813 return 0; 814 } 815 816 static RING_IDX xennet_fill_frags(struct netfront_info *np, 817 struct sk_buff *skb, 818 struct sk_buff_head *list) 819 { 820 struct skb_shared_info *shinfo = skb_shinfo(skb); 821 int nr_frags = shinfo->nr_frags; 822 RING_IDX cons = np->rx.rsp_cons; 823 struct sk_buff *nskb; 824 825 while ((nskb = __skb_dequeue(list))) { 826 struct xen_netif_rx_response *rx = 827 RING_GET_RESPONSE(&np->rx, ++cons); 828 skb_frag_t *nfrag = &skb_shinfo(nskb)->frags[0]; 829 830 __skb_fill_page_desc(skb, nr_frags, 831 skb_frag_page(nfrag), 832 rx->offset, rx->status); 833 834 skb->data_len += rx->status; 835 836 skb_shinfo(nskb)->nr_frags = 0; 837 kfree_skb(nskb); 838 839 nr_frags++; 840 } 841 842 shinfo->nr_frags = nr_frags; 843 return cons; 844 } 845 846 static int checksum_setup(struct net_device *dev, struct sk_buff *skb) 847 { 848 struct iphdr *iph; 849 unsigned char *th; 850 int err = -EPROTO; 851 int recalculate_partial_csum = 0; 852 853 /* 854 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy 855 * peers can fail to set NETRXF_csum_blank when sending a GSO 856 * frame. In this case force the SKB to CHECKSUM_PARTIAL and 857 * recalculate the partial checksum. 858 */ 859 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) { 860 struct netfront_info *np = netdev_priv(dev); 861 np->rx_gso_checksum_fixup++; 862 skb->ip_summed = CHECKSUM_PARTIAL; 863 recalculate_partial_csum = 1; 864 } 865 866 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */ 867 if (skb->ip_summed != CHECKSUM_PARTIAL) 868 return 0; 869 870 if (skb->protocol != htons(ETH_P_IP)) 871 goto out; 872 873 iph = (void *)skb->data; 874 th = skb->data + 4 * iph->ihl; 875 if (th >= skb_tail_pointer(skb)) 876 goto out; 877 878 skb->csum_start = th - skb->head; 879 switch (iph->protocol) { 880 case IPPROTO_TCP: 881 skb->csum_offset = offsetof(struct tcphdr, check); 882 883 if (recalculate_partial_csum) { 884 struct tcphdr *tcph = (struct tcphdr *)th; 885 tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr, 886 skb->len - iph->ihl*4, 887 IPPROTO_TCP, 0); 888 } 889 break; 890 case IPPROTO_UDP: 891 skb->csum_offset = offsetof(struct udphdr, check); 892 893 if (recalculate_partial_csum) { 894 struct udphdr *udph = (struct udphdr *)th; 895 udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr, 896 skb->len - iph->ihl*4, 897 IPPROTO_UDP, 0); 898 } 899 break; 900 default: 901 if (net_ratelimit()) 902 printk(KERN_ERR "Attempting to checksum a non-" 903 "TCP/UDP packet, dropping a protocol" 904 " %d packet", iph->protocol); 905 goto out; 906 } 907 908 if ((th + skb->csum_offset + 2) > skb_tail_pointer(skb)) 909 goto out; 910 911 err = 0; 912 913 out: 914 return err; 915 } 916 917 static int handle_incoming_queue(struct net_device *dev, 918 struct sk_buff_head *rxq) 919 { 920 struct netfront_info *np = netdev_priv(dev); 921 struct netfront_stats *stats = this_cpu_ptr(np->stats); 922 int packets_dropped = 0; 923 struct sk_buff *skb; 924 925 while ((skb = __skb_dequeue(rxq)) != NULL) { 926 int pull_to = NETFRONT_SKB_CB(skb)->pull_to; 927 928 __pskb_pull_tail(skb, pull_to - skb_headlen(skb)); 929 930 /* Ethernet work: Delayed to here as it peeks the header. */ 931 skb->protocol = eth_type_trans(skb, dev); 932 933 if (checksum_setup(dev, skb)) { 934 kfree_skb(skb); 935 packets_dropped++; 936 dev->stats.rx_errors++; 937 continue; 938 } 939 940 u64_stats_update_begin(&stats->syncp); 941 stats->rx_packets++; 942 stats->rx_bytes += skb->len; 943 u64_stats_update_end(&stats->syncp); 944 945 /* Pass it up. */ 946 netif_receive_skb(skb); 947 } 948 949 return packets_dropped; 950 } 951 952 static int xennet_poll(struct napi_struct *napi, int budget) 953 { 954 struct netfront_info *np = container_of(napi, struct netfront_info, napi); 955 struct net_device *dev = np->netdev; 956 struct sk_buff *skb; 957 struct netfront_rx_info rinfo; 958 struct xen_netif_rx_response *rx = &rinfo.rx; 959 struct xen_netif_extra_info *extras = rinfo.extras; 960 RING_IDX i, rp; 961 int work_done; 962 struct sk_buff_head rxq; 963 struct sk_buff_head errq; 964 struct sk_buff_head tmpq; 965 unsigned long flags; 966 int err; 967 968 spin_lock(&np->rx_lock); 969 970 skb_queue_head_init(&rxq); 971 skb_queue_head_init(&errq); 972 skb_queue_head_init(&tmpq); 973 974 rp = np->rx.sring->rsp_prod; 975 rmb(); /* Ensure we see queued responses up to 'rp'. */ 976 977 i = np->rx.rsp_cons; 978 work_done = 0; 979 while ((i != rp) && (work_done < budget)) { 980 memcpy(rx, RING_GET_RESPONSE(&np->rx, i), sizeof(*rx)); 981 memset(extras, 0, sizeof(rinfo.extras)); 982 983 err = xennet_get_responses(np, &rinfo, rp, &tmpq); 984 985 if (unlikely(err)) { 986 err: 987 while ((skb = __skb_dequeue(&tmpq))) 988 __skb_queue_tail(&errq, skb); 989 dev->stats.rx_errors++; 990 i = np->rx.rsp_cons; 991 continue; 992 } 993 994 skb = __skb_dequeue(&tmpq); 995 996 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) { 997 struct xen_netif_extra_info *gso; 998 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1]; 999 1000 if (unlikely(xennet_set_skb_gso(skb, gso))) { 1001 __skb_queue_head(&tmpq, skb); 1002 np->rx.rsp_cons += skb_queue_len(&tmpq); 1003 goto err; 1004 } 1005 } 1006 1007 NETFRONT_SKB_CB(skb)->pull_to = rx->status; 1008 if (NETFRONT_SKB_CB(skb)->pull_to > RX_COPY_THRESHOLD) 1009 NETFRONT_SKB_CB(skb)->pull_to = RX_COPY_THRESHOLD; 1010 1011 skb_shinfo(skb)->frags[0].page_offset = rx->offset; 1012 skb_frag_size_set(&skb_shinfo(skb)->frags[0], rx->status); 1013 skb->data_len = rx->status; 1014 1015 i = xennet_fill_frags(np, skb, &tmpq); 1016 1017 /* 1018 * Truesize approximates the size of true data plus 1019 * any supervisor overheads. Adding hypervisor 1020 * overheads has been shown to significantly reduce 1021 * achievable bandwidth with the default receive 1022 * buffer size. It is therefore not wise to account 1023 * for it here. 1024 * 1025 * After alloc_skb(RX_COPY_THRESHOLD), truesize is set 1026 * to RX_COPY_THRESHOLD + the supervisor 1027 * overheads. Here, we add the size of the data pulled 1028 * in xennet_fill_frags(). 1029 * 1030 * We also adjust for any unused space in the main 1031 * data area by subtracting (RX_COPY_THRESHOLD - 1032 * len). This is especially important with drivers 1033 * which split incoming packets into header and data, 1034 * using only 66 bytes of the main data area (see the 1035 * e1000 driver for example.) On such systems, 1036 * without this last adjustement, our achievable 1037 * receive throughout using the standard receive 1038 * buffer size was cut by 25%(!!!). 1039 */ 1040 skb->truesize += skb->data_len - RX_COPY_THRESHOLD; 1041 skb->len += skb->data_len; 1042 1043 if (rx->flags & XEN_NETRXF_csum_blank) 1044 skb->ip_summed = CHECKSUM_PARTIAL; 1045 else if (rx->flags & XEN_NETRXF_data_validated) 1046 skb->ip_summed = CHECKSUM_UNNECESSARY; 1047 1048 __skb_queue_tail(&rxq, skb); 1049 1050 np->rx.rsp_cons = ++i; 1051 work_done++; 1052 } 1053 1054 __skb_queue_purge(&errq); 1055 1056 work_done -= handle_incoming_queue(dev, &rxq); 1057 1058 /* If we get a callback with very few responses, reduce fill target. */ 1059 /* NB. Note exponential increase, linear decrease. */ 1060 if (((np->rx.req_prod_pvt - np->rx.sring->rsp_prod) > 1061 ((3*np->rx_target) / 4)) && 1062 (--np->rx_target < np->rx_min_target)) 1063 np->rx_target = np->rx_min_target; 1064 1065 xennet_alloc_rx_buffers(dev); 1066 1067 if (work_done < budget) { 1068 int more_to_do = 0; 1069 1070 local_irq_save(flags); 1071 1072 RING_FINAL_CHECK_FOR_RESPONSES(&np->rx, more_to_do); 1073 if (!more_to_do) 1074 __napi_complete(napi); 1075 1076 local_irq_restore(flags); 1077 } 1078 1079 spin_unlock(&np->rx_lock); 1080 1081 return work_done; 1082 } 1083 1084 static int xennet_change_mtu(struct net_device *dev, int mtu) 1085 { 1086 int max = xennet_can_sg(dev) ? 65535 - ETH_HLEN : ETH_DATA_LEN; 1087 1088 if (mtu > max) 1089 return -EINVAL; 1090 dev->mtu = mtu; 1091 return 0; 1092 } 1093 1094 static struct rtnl_link_stats64 *xennet_get_stats64(struct net_device *dev, 1095 struct rtnl_link_stats64 *tot) 1096 { 1097 struct netfront_info *np = netdev_priv(dev); 1098 int cpu; 1099 1100 for_each_possible_cpu(cpu) { 1101 struct netfront_stats *stats = per_cpu_ptr(np->stats, cpu); 1102 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 1103 unsigned int start; 1104 1105 do { 1106 start = u64_stats_fetch_begin_bh(&stats->syncp); 1107 1108 rx_packets = stats->rx_packets; 1109 tx_packets = stats->tx_packets; 1110 rx_bytes = stats->rx_bytes; 1111 tx_bytes = stats->tx_bytes; 1112 } while (u64_stats_fetch_retry_bh(&stats->syncp, start)); 1113 1114 tot->rx_packets += rx_packets; 1115 tot->tx_packets += tx_packets; 1116 tot->rx_bytes += rx_bytes; 1117 tot->tx_bytes += tx_bytes; 1118 } 1119 1120 tot->rx_errors = dev->stats.rx_errors; 1121 tot->tx_dropped = dev->stats.tx_dropped; 1122 1123 return tot; 1124 } 1125 1126 static void xennet_release_tx_bufs(struct netfront_info *np) 1127 { 1128 struct sk_buff *skb; 1129 int i; 1130 1131 for (i = 0; i < NET_TX_RING_SIZE; i++) { 1132 /* Skip over entries which are actually freelist references */ 1133 if (skb_entry_is_link(&np->tx_skbs[i])) 1134 continue; 1135 1136 skb = np->tx_skbs[i].skb; 1137 gnttab_end_foreign_access_ref(np->grant_tx_ref[i], 1138 GNTMAP_readonly); 1139 gnttab_release_grant_reference(&np->gref_tx_head, 1140 np->grant_tx_ref[i]); 1141 np->grant_tx_ref[i] = GRANT_INVALID_REF; 1142 add_id_to_freelist(&np->tx_skb_freelist, np->tx_skbs, i); 1143 dev_kfree_skb_irq(skb); 1144 } 1145 } 1146 1147 static void xennet_release_rx_bufs(struct netfront_info *np) 1148 { 1149 struct mmu_update *mmu = np->rx_mmu; 1150 struct multicall_entry *mcl = np->rx_mcl; 1151 struct sk_buff_head free_list; 1152 struct sk_buff *skb; 1153 unsigned long mfn; 1154 int xfer = 0, noxfer = 0, unused = 0; 1155 int id, ref; 1156 1157 dev_warn(&np->netdev->dev, "%s: fix me for copying receiver.\n", 1158 __func__); 1159 return; 1160 1161 skb_queue_head_init(&free_list); 1162 1163 spin_lock_bh(&np->rx_lock); 1164 1165 for (id = 0; id < NET_RX_RING_SIZE; id++) { 1166 ref = np->grant_rx_ref[id]; 1167 if (ref == GRANT_INVALID_REF) { 1168 unused++; 1169 continue; 1170 } 1171 1172 skb = np->rx_skbs[id]; 1173 mfn = gnttab_end_foreign_transfer_ref(ref); 1174 gnttab_release_grant_reference(&np->gref_rx_head, ref); 1175 np->grant_rx_ref[id] = GRANT_INVALID_REF; 1176 1177 if (0 == mfn) { 1178 skb_shinfo(skb)->nr_frags = 0; 1179 dev_kfree_skb(skb); 1180 noxfer++; 1181 continue; 1182 } 1183 1184 if (!xen_feature(XENFEAT_auto_translated_physmap)) { 1185 /* Remap the page. */ 1186 const struct page *page = 1187 skb_frag_page(&skb_shinfo(skb)->frags[0]); 1188 unsigned long pfn = page_to_pfn(page); 1189 void *vaddr = page_address(page); 1190 1191 MULTI_update_va_mapping(mcl, (unsigned long)vaddr, 1192 mfn_pte(mfn, PAGE_KERNEL), 1193 0); 1194 mcl++; 1195 mmu->ptr = ((u64)mfn << PAGE_SHIFT) 1196 | MMU_MACHPHYS_UPDATE; 1197 mmu->val = pfn; 1198 mmu++; 1199 1200 set_phys_to_machine(pfn, mfn); 1201 } 1202 __skb_queue_tail(&free_list, skb); 1203 xfer++; 1204 } 1205 1206 dev_info(&np->netdev->dev, "%s: %d xfer, %d noxfer, %d unused\n", 1207 __func__, xfer, noxfer, unused); 1208 1209 if (xfer) { 1210 if (!xen_feature(XENFEAT_auto_translated_physmap)) { 1211 /* Do all the remapping work and M2P updates. */ 1212 MULTI_mmu_update(mcl, np->rx_mmu, mmu - np->rx_mmu, 1213 NULL, DOMID_SELF); 1214 mcl++; 1215 HYPERVISOR_multicall(np->rx_mcl, mcl - np->rx_mcl); 1216 } 1217 } 1218 1219 __skb_queue_purge(&free_list); 1220 1221 spin_unlock_bh(&np->rx_lock); 1222 } 1223 1224 static void xennet_uninit(struct net_device *dev) 1225 { 1226 struct netfront_info *np = netdev_priv(dev); 1227 xennet_release_tx_bufs(np); 1228 xennet_release_rx_bufs(np); 1229 gnttab_free_grant_references(np->gref_tx_head); 1230 gnttab_free_grant_references(np->gref_rx_head); 1231 } 1232 1233 static netdev_features_t xennet_fix_features(struct net_device *dev, 1234 netdev_features_t features) 1235 { 1236 struct netfront_info *np = netdev_priv(dev); 1237 int val; 1238 1239 if (features & NETIF_F_SG) { 1240 if (xenbus_scanf(XBT_NIL, np->xbdev->otherend, "feature-sg", 1241 "%d", &val) < 0) 1242 val = 0; 1243 1244 if (!val) 1245 features &= ~NETIF_F_SG; 1246 } 1247 1248 if (features & NETIF_F_TSO) { 1249 if (xenbus_scanf(XBT_NIL, np->xbdev->otherend, 1250 "feature-gso-tcpv4", "%d", &val) < 0) 1251 val = 0; 1252 1253 if (!val) 1254 features &= ~NETIF_F_TSO; 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_interrupt(int irq, void *dev_id) 1272 { 1273 struct net_device *dev = dev_id; 1274 struct netfront_info *np = netdev_priv(dev); 1275 unsigned long flags; 1276 1277 spin_lock_irqsave(&np->tx_lock, flags); 1278 1279 if (likely(netif_carrier_ok(dev))) { 1280 xennet_tx_buf_gc(dev); 1281 /* Under tx_lock: protects access to rx shared-ring indexes. */ 1282 if (RING_HAS_UNCONSUMED_RESPONSES(&np->rx)) 1283 napi_schedule(&np->napi); 1284 } 1285 1286 spin_unlock_irqrestore(&np->tx_lock, flags); 1287 1288 return IRQ_HANDLED; 1289 } 1290 1291 #ifdef CONFIG_NET_POLL_CONTROLLER 1292 static void xennet_poll_controller(struct net_device *dev) 1293 { 1294 xennet_interrupt(0, dev); 1295 } 1296 #endif 1297 1298 static const struct net_device_ops xennet_netdev_ops = { 1299 .ndo_open = xennet_open, 1300 .ndo_uninit = xennet_uninit, 1301 .ndo_stop = xennet_close, 1302 .ndo_start_xmit = xennet_start_xmit, 1303 .ndo_change_mtu = xennet_change_mtu, 1304 .ndo_get_stats64 = xennet_get_stats64, 1305 .ndo_set_mac_address = eth_mac_addr, 1306 .ndo_validate_addr = eth_validate_addr, 1307 .ndo_fix_features = xennet_fix_features, 1308 .ndo_set_features = xennet_set_features, 1309 #ifdef CONFIG_NET_POLL_CONTROLLER 1310 .ndo_poll_controller = xennet_poll_controller, 1311 #endif 1312 }; 1313 1314 static struct net_device *xennet_create_dev(struct xenbus_device *dev) 1315 { 1316 int i, err; 1317 struct net_device *netdev; 1318 struct netfront_info *np; 1319 1320 netdev = alloc_etherdev(sizeof(struct netfront_info)); 1321 if (!netdev) 1322 return ERR_PTR(-ENOMEM); 1323 1324 np = netdev_priv(netdev); 1325 np->xbdev = dev; 1326 1327 spin_lock_init(&np->tx_lock); 1328 spin_lock_init(&np->rx_lock); 1329 1330 skb_queue_head_init(&np->rx_batch); 1331 np->rx_target = RX_DFL_MIN_TARGET; 1332 np->rx_min_target = RX_DFL_MIN_TARGET; 1333 np->rx_max_target = RX_MAX_TARGET; 1334 1335 init_timer(&np->rx_refill_timer); 1336 np->rx_refill_timer.data = (unsigned long)netdev; 1337 np->rx_refill_timer.function = rx_refill_timeout; 1338 1339 err = -ENOMEM; 1340 np->stats = alloc_percpu(struct netfront_stats); 1341 if (np->stats == NULL) 1342 goto exit; 1343 1344 /* Initialise tx_skbs as a free chain containing every entry. */ 1345 np->tx_skb_freelist = 0; 1346 for (i = 0; i < NET_TX_RING_SIZE; i++) { 1347 skb_entry_set_link(&np->tx_skbs[i], i+1); 1348 np->grant_tx_ref[i] = GRANT_INVALID_REF; 1349 } 1350 1351 /* Clear out rx_skbs */ 1352 for (i = 0; i < NET_RX_RING_SIZE; i++) { 1353 np->rx_skbs[i] = NULL; 1354 np->grant_rx_ref[i] = GRANT_INVALID_REF; 1355 } 1356 1357 /* A grant for every tx ring slot */ 1358 if (gnttab_alloc_grant_references(TX_MAX_TARGET, 1359 &np->gref_tx_head) < 0) { 1360 printk(KERN_ALERT "#### netfront can't alloc tx grant refs\n"); 1361 err = -ENOMEM; 1362 goto exit_free_stats; 1363 } 1364 /* A grant for every rx ring slot */ 1365 if (gnttab_alloc_grant_references(RX_MAX_TARGET, 1366 &np->gref_rx_head) < 0) { 1367 printk(KERN_ALERT "#### netfront can't alloc rx grant refs\n"); 1368 err = -ENOMEM; 1369 goto exit_free_tx; 1370 } 1371 1372 netdev->netdev_ops = &xennet_netdev_ops; 1373 1374 netif_napi_add(netdev, &np->napi, xennet_poll, 64); 1375 netdev->features = NETIF_F_IP_CSUM | NETIF_F_RXCSUM | 1376 NETIF_F_GSO_ROBUST; 1377 netdev->hw_features = NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO; 1378 1379 /* 1380 * Assume that all hw features are available for now. This set 1381 * will be adjusted by the call to netdev_update_features() in 1382 * xennet_connect() which is the earliest point where we can 1383 * negotiate with the backend regarding supported features. 1384 */ 1385 netdev->features |= netdev->hw_features; 1386 1387 SET_ETHTOOL_OPS(netdev, &xennet_ethtool_ops); 1388 SET_NETDEV_DEV(netdev, &dev->dev); 1389 1390 np->netdev = netdev; 1391 1392 netif_carrier_off(netdev); 1393 1394 return netdev; 1395 1396 exit_free_tx: 1397 gnttab_free_grant_references(np->gref_tx_head); 1398 exit_free_stats: 1399 free_percpu(np->stats); 1400 exit: 1401 free_netdev(netdev); 1402 return ERR_PTR(err); 1403 } 1404 1405 /** 1406 * Entry point to this code when a new device is created. Allocate the basic 1407 * structures and the ring buffers for communication with the backend, and 1408 * inform the backend of the appropriate details for those. 1409 */ 1410 static int netfront_probe(struct xenbus_device *dev, 1411 const struct xenbus_device_id *id) 1412 { 1413 int err; 1414 struct net_device *netdev; 1415 struct netfront_info *info; 1416 1417 netdev = xennet_create_dev(dev); 1418 if (IS_ERR(netdev)) { 1419 err = PTR_ERR(netdev); 1420 xenbus_dev_fatal(dev, err, "creating netdev"); 1421 return err; 1422 } 1423 1424 info = netdev_priv(netdev); 1425 dev_set_drvdata(&dev->dev, info); 1426 1427 err = register_netdev(info->netdev); 1428 if (err) { 1429 printk(KERN_WARNING "%s: register_netdev err=%d\n", 1430 __func__, err); 1431 goto fail; 1432 } 1433 1434 err = xennet_sysfs_addif(info->netdev); 1435 if (err) { 1436 unregister_netdev(info->netdev); 1437 printk(KERN_WARNING "%s: add sysfs failed err=%d\n", 1438 __func__, err); 1439 goto fail; 1440 } 1441 1442 return 0; 1443 1444 fail: 1445 free_netdev(netdev); 1446 dev_set_drvdata(&dev->dev, NULL); 1447 return err; 1448 } 1449 1450 static void xennet_end_access(int ref, void *page) 1451 { 1452 /* This frees the page as a side-effect */ 1453 if (ref != GRANT_INVALID_REF) 1454 gnttab_end_foreign_access(ref, 0, (unsigned long)page); 1455 } 1456 1457 static void xennet_disconnect_backend(struct netfront_info *info) 1458 { 1459 /* Stop old i/f to prevent errors whilst we rebuild the state. */ 1460 spin_lock_bh(&info->rx_lock); 1461 spin_lock_irq(&info->tx_lock); 1462 netif_carrier_off(info->netdev); 1463 spin_unlock_irq(&info->tx_lock); 1464 spin_unlock_bh(&info->rx_lock); 1465 1466 if (info->netdev->irq) 1467 unbind_from_irqhandler(info->netdev->irq, info->netdev); 1468 info->evtchn = info->netdev->irq = 0; 1469 1470 /* End access and free the pages */ 1471 xennet_end_access(info->tx_ring_ref, info->tx.sring); 1472 xennet_end_access(info->rx_ring_ref, info->rx.sring); 1473 1474 info->tx_ring_ref = GRANT_INVALID_REF; 1475 info->rx_ring_ref = GRANT_INVALID_REF; 1476 info->tx.sring = NULL; 1477 info->rx.sring = NULL; 1478 } 1479 1480 /** 1481 * We are reconnecting to the backend, due to a suspend/resume, or a backend 1482 * driver restart. We tear down our netif structure and recreate it, but 1483 * leave the device-layer structures intact so that this is transparent to the 1484 * rest of the kernel. 1485 */ 1486 static int netfront_resume(struct xenbus_device *dev) 1487 { 1488 struct netfront_info *info = dev_get_drvdata(&dev->dev); 1489 1490 dev_dbg(&dev->dev, "%s\n", dev->nodename); 1491 1492 xennet_disconnect_backend(info); 1493 return 0; 1494 } 1495 1496 static int xen_net_read_mac(struct xenbus_device *dev, u8 mac[]) 1497 { 1498 char *s, *e, *macstr; 1499 int i; 1500 1501 macstr = s = xenbus_read(XBT_NIL, dev->nodename, "mac", NULL); 1502 if (IS_ERR(macstr)) 1503 return PTR_ERR(macstr); 1504 1505 for (i = 0; i < ETH_ALEN; i++) { 1506 mac[i] = simple_strtoul(s, &e, 16); 1507 if ((s == e) || (*e != ((i == ETH_ALEN-1) ? '\0' : ':'))) { 1508 kfree(macstr); 1509 return -ENOENT; 1510 } 1511 s = e+1; 1512 } 1513 1514 kfree(macstr); 1515 return 0; 1516 } 1517 1518 static int setup_netfront(struct xenbus_device *dev, struct netfront_info *info) 1519 { 1520 struct xen_netif_tx_sring *txs; 1521 struct xen_netif_rx_sring *rxs; 1522 int err; 1523 struct net_device *netdev = info->netdev; 1524 1525 info->tx_ring_ref = GRANT_INVALID_REF; 1526 info->rx_ring_ref = GRANT_INVALID_REF; 1527 info->rx.sring = NULL; 1528 info->tx.sring = NULL; 1529 netdev->irq = 0; 1530 1531 err = xen_net_read_mac(dev, netdev->dev_addr); 1532 if (err) { 1533 xenbus_dev_fatal(dev, err, "parsing %s/mac", dev->nodename); 1534 goto fail; 1535 } 1536 1537 txs = (struct xen_netif_tx_sring *)get_zeroed_page(GFP_NOIO | __GFP_HIGH); 1538 if (!txs) { 1539 err = -ENOMEM; 1540 xenbus_dev_fatal(dev, err, "allocating tx ring page"); 1541 goto fail; 1542 } 1543 SHARED_RING_INIT(txs); 1544 FRONT_RING_INIT(&info->tx, txs, PAGE_SIZE); 1545 1546 err = xenbus_grant_ring(dev, virt_to_mfn(txs)); 1547 if (err < 0) { 1548 free_page((unsigned long)txs); 1549 goto fail; 1550 } 1551 1552 info->tx_ring_ref = err; 1553 rxs = (struct xen_netif_rx_sring *)get_zeroed_page(GFP_NOIO | __GFP_HIGH); 1554 if (!rxs) { 1555 err = -ENOMEM; 1556 xenbus_dev_fatal(dev, err, "allocating rx ring page"); 1557 goto fail; 1558 } 1559 SHARED_RING_INIT(rxs); 1560 FRONT_RING_INIT(&info->rx, rxs, PAGE_SIZE); 1561 1562 err = xenbus_grant_ring(dev, virt_to_mfn(rxs)); 1563 if (err < 0) { 1564 free_page((unsigned long)rxs); 1565 goto fail; 1566 } 1567 info->rx_ring_ref = err; 1568 1569 err = xenbus_alloc_evtchn(dev, &info->evtchn); 1570 if (err) 1571 goto fail; 1572 1573 err = bind_evtchn_to_irqhandler(info->evtchn, xennet_interrupt, 1574 0, netdev->name, netdev); 1575 if (err < 0) 1576 goto fail; 1577 netdev->irq = err; 1578 return 0; 1579 1580 fail: 1581 return err; 1582 } 1583 1584 /* Common code used when first setting up, and when resuming. */ 1585 static int talk_to_netback(struct xenbus_device *dev, 1586 struct netfront_info *info) 1587 { 1588 const char *message; 1589 struct xenbus_transaction xbt; 1590 int err; 1591 1592 /* Create shared ring, alloc event channel. */ 1593 err = setup_netfront(dev, info); 1594 if (err) 1595 goto out; 1596 1597 again: 1598 err = xenbus_transaction_start(&xbt); 1599 if (err) { 1600 xenbus_dev_fatal(dev, err, "starting transaction"); 1601 goto destroy_ring; 1602 } 1603 1604 err = xenbus_printf(xbt, dev->nodename, "tx-ring-ref", "%u", 1605 info->tx_ring_ref); 1606 if (err) { 1607 message = "writing tx ring-ref"; 1608 goto abort_transaction; 1609 } 1610 err = xenbus_printf(xbt, dev->nodename, "rx-ring-ref", "%u", 1611 info->rx_ring_ref); 1612 if (err) { 1613 message = "writing rx ring-ref"; 1614 goto abort_transaction; 1615 } 1616 err = xenbus_printf(xbt, dev->nodename, 1617 "event-channel", "%u", info->evtchn); 1618 if (err) { 1619 message = "writing event-channel"; 1620 goto abort_transaction; 1621 } 1622 1623 err = xenbus_printf(xbt, dev->nodename, "request-rx-copy", "%u", 1624 1); 1625 if (err) { 1626 message = "writing request-rx-copy"; 1627 goto abort_transaction; 1628 } 1629 1630 err = xenbus_printf(xbt, dev->nodename, "feature-rx-notify", "%d", 1); 1631 if (err) { 1632 message = "writing feature-rx-notify"; 1633 goto abort_transaction; 1634 } 1635 1636 err = xenbus_printf(xbt, dev->nodename, "feature-sg", "%d", 1); 1637 if (err) { 1638 message = "writing feature-sg"; 1639 goto abort_transaction; 1640 } 1641 1642 err = xenbus_printf(xbt, dev->nodename, "feature-gso-tcpv4", "%d", 1); 1643 if (err) { 1644 message = "writing feature-gso-tcpv4"; 1645 goto abort_transaction; 1646 } 1647 1648 err = xenbus_transaction_end(xbt, 0); 1649 if (err) { 1650 if (err == -EAGAIN) 1651 goto again; 1652 xenbus_dev_fatal(dev, err, "completing transaction"); 1653 goto destroy_ring; 1654 } 1655 1656 return 0; 1657 1658 abort_transaction: 1659 xenbus_transaction_end(xbt, 1); 1660 xenbus_dev_fatal(dev, err, "%s", message); 1661 destroy_ring: 1662 xennet_disconnect_backend(info); 1663 out: 1664 return err; 1665 } 1666 1667 static int xennet_connect(struct net_device *dev) 1668 { 1669 struct netfront_info *np = netdev_priv(dev); 1670 int i, requeue_idx, err; 1671 struct sk_buff *skb; 1672 grant_ref_t ref; 1673 struct xen_netif_rx_request *req; 1674 unsigned int feature_rx_copy; 1675 1676 err = xenbus_scanf(XBT_NIL, np->xbdev->otherend, 1677 "feature-rx-copy", "%u", &feature_rx_copy); 1678 if (err != 1) 1679 feature_rx_copy = 0; 1680 1681 if (!feature_rx_copy) { 1682 dev_info(&dev->dev, 1683 "backend does not support copying receive path\n"); 1684 return -ENODEV; 1685 } 1686 1687 err = talk_to_netback(np->xbdev, np); 1688 if (err) 1689 return err; 1690 1691 rtnl_lock(); 1692 netdev_update_features(dev); 1693 rtnl_unlock(); 1694 1695 spin_lock_bh(&np->rx_lock); 1696 spin_lock_irq(&np->tx_lock); 1697 1698 /* Step 1: Discard all pending TX packet fragments. */ 1699 xennet_release_tx_bufs(np); 1700 1701 /* Step 2: Rebuild the RX buffer freelist and the RX ring itself. */ 1702 for (requeue_idx = 0, i = 0; i < NET_RX_RING_SIZE; i++) { 1703 skb_frag_t *frag; 1704 const struct page *page; 1705 if (!np->rx_skbs[i]) 1706 continue; 1707 1708 skb = np->rx_skbs[requeue_idx] = xennet_get_rx_skb(np, i); 1709 ref = np->grant_rx_ref[requeue_idx] = xennet_get_rx_ref(np, i); 1710 req = RING_GET_REQUEST(&np->rx, requeue_idx); 1711 1712 frag = &skb_shinfo(skb)->frags[0]; 1713 page = skb_frag_page(frag); 1714 gnttab_grant_foreign_access_ref( 1715 ref, np->xbdev->otherend_id, 1716 pfn_to_mfn(page_to_pfn(page)), 1717 0); 1718 req->gref = ref; 1719 req->id = requeue_idx; 1720 1721 requeue_idx++; 1722 } 1723 1724 np->rx.req_prod_pvt = requeue_idx; 1725 1726 /* 1727 * Step 3: All public and private state should now be sane. Get 1728 * ready to start sending and receiving packets and give the driver 1729 * domain a kick because we've probably just requeued some 1730 * packets. 1731 */ 1732 netif_carrier_on(np->netdev); 1733 notify_remote_via_irq(np->netdev->irq); 1734 xennet_tx_buf_gc(dev); 1735 xennet_alloc_rx_buffers(dev); 1736 1737 spin_unlock_irq(&np->tx_lock); 1738 spin_unlock_bh(&np->rx_lock); 1739 1740 return 0; 1741 } 1742 1743 /** 1744 * Callback received when the backend's state changes. 1745 */ 1746 static void netback_changed(struct xenbus_device *dev, 1747 enum xenbus_state backend_state) 1748 { 1749 struct netfront_info *np = dev_get_drvdata(&dev->dev); 1750 struct net_device *netdev = np->netdev; 1751 1752 dev_dbg(&dev->dev, "%s\n", xenbus_strstate(backend_state)); 1753 1754 switch (backend_state) { 1755 case XenbusStateInitialising: 1756 case XenbusStateInitialised: 1757 case XenbusStateReconfiguring: 1758 case XenbusStateReconfigured: 1759 case XenbusStateUnknown: 1760 case XenbusStateClosed: 1761 break; 1762 1763 case XenbusStateInitWait: 1764 if (dev->state != XenbusStateInitialising) 1765 break; 1766 if (xennet_connect(netdev) != 0) 1767 break; 1768 xenbus_switch_state(dev, XenbusStateConnected); 1769 break; 1770 1771 case XenbusStateConnected: 1772 netdev_notify_peers(netdev); 1773 break; 1774 1775 case XenbusStateClosing: 1776 xenbus_frontend_closed(dev); 1777 break; 1778 } 1779 } 1780 1781 static const struct xennet_stat { 1782 char name[ETH_GSTRING_LEN]; 1783 u16 offset; 1784 } xennet_stats[] = { 1785 { 1786 "rx_gso_checksum_fixup", 1787 offsetof(struct netfront_info, rx_gso_checksum_fixup) 1788 }, 1789 }; 1790 1791 static int xennet_get_sset_count(struct net_device *dev, int string_set) 1792 { 1793 switch (string_set) { 1794 case ETH_SS_STATS: 1795 return ARRAY_SIZE(xennet_stats); 1796 default: 1797 return -EINVAL; 1798 } 1799 } 1800 1801 static void xennet_get_ethtool_stats(struct net_device *dev, 1802 struct ethtool_stats *stats, u64 * data) 1803 { 1804 void *np = netdev_priv(dev); 1805 int i; 1806 1807 for (i = 0; i < ARRAY_SIZE(xennet_stats); i++) 1808 data[i] = *(unsigned long *)(np + xennet_stats[i].offset); 1809 } 1810 1811 static void xennet_get_strings(struct net_device *dev, u32 stringset, u8 * data) 1812 { 1813 int i; 1814 1815 switch (stringset) { 1816 case ETH_SS_STATS: 1817 for (i = 0; i < ARRAY_SIZE(xennet_stats); i++) 1818 memcpy(data + i * ETH_GSTRING_LEN, 1819 xennet_stats[i].name, ETH_GSTRING_LEN); 1820 break; 1821 } 1822 } 1823 1824 static const struct ethtool_ops xennet_ethtool_ops = 1825 { 1826 .get_link = ethtool_op_get_link, 1827 1828 .get_sset_count = xennet_get_sset_count, 1829 .get_ethtool_stats = xennet_get_ethtool_stats, 1830 .get_strings = xennet_get_strings, 1831 }; 1832 1833 #ifdef CONFIG_SYSFS 1834 static ssize_t show_rxbuf_min(struct device *dev, 1835 struct device_attribute *attr, char *buf) 1836 { 1837 struct net_device *netdev = to_net_dev(dev); 1838 struct netfront_info *info = netdev_priv(netdev); 1839 1840 return sprintf(buf, "%u\n", info->rx_min_target); 1841 } 1842 1843 static ssize_t store_rxbuf_min(struct device *dev, 1844 struct device_attribute *attr, 1845 const char *buf, size_t len) 1846 { 1847 struct net_device *netdev = to_net_dev(dev); 1848 struct netfront_info *np = netdev_priv(netdev); 1849 char *endp; 1850 unsigned long target; 1851 1852 if (!capable(CAP_NET_ADMIN)) 1853 return -EPERM; 1854 1855 target = simple_strtoul(buf, &endp, 0); 1856 if (endp == buf) 1857 return -EBADMSG; 1858 1859 if (target < RX_MIN_TARGET) 1860 target = RX_MIN_TARGET; 1861 if (target > RX_MAX_TARGET) 1862 target = RX_MAX_TARGET; 1863 1864 spin_lock_bh(&np->rx_lock); 1865 if (target > np->rx_max_target) 1866 np->rx_max_target = target; 1867 np->rx_min_target = target; 1868 if (target > np->rx_target) 1869 np->rx_target = target; 1870 1871 xennet_alloc_rx_buffers(netdev); 1872 1873 spin_unlock_bh(&np->rx_lock); 1874 return len; 1875 } 1876 1877 static ssize_t show_rxbuf_max(struct device *dev, 1878 struct device_attribute *attr, char *buf) 1879 { 1880 struct net_device *netdev = to_net_dev(dev); 1881 struct netfront_info *info = netdev_priv(netdev); 1882 1883 return sprintf(buf, "%u\n", info->rx_max_target); 1884 } 1885 1886 static ssize_t store_rxbuf_max(struct device *dev, 1887 struct device_attribute *attr, 1888 const char *buf, size_t len) 1889 { 1890 struct net_device *netdev = to_net_dev(dev); 1891 struct netfront_info *np = netdev_priv(netdev); 1892 char *endp; 1893 unsigned long target; 1894 1895 if (!capable(CAP_NET_ADMIN)) 1896 return -EPERM; 1897 1898 target = simple_strtoul(buf, &endp, 0); 1899 if (endp == buf) 1900 return -EBADMSG; 1901 1902 if (target < RX_MIN_TARGET) 1903 target = RX_MIN_TARGET; 1904 if (target > RX_MAX_TARGET) 1905 target = RX_MAX_TARGET; 1906 1907 spin_lock_bh(&np->rx_lock); 1908 if (target < np->rx_min_target) 1909 np->rx_min_target = target; 1910 np->rx_max_target = target; 1911 if (target < np->rx_target) 1912 np->rx_target = target; 1913 1914 xennet_alloc_rx_buffers(netdev); 1915 1916 spin_unlock_bh(&np->rx_lock); 1917 return len; 1918 } 1919 1920 static ssize_t show_rxbuf_cur(struct device *dev, 1921 struct device_attribute *attr, char *buf) 1922 { 1923 struct net_device *netdev = to_net_dev(dev); 1924 struct netfront_info *info = netdev_priv(netdev); 1925 1926 return sprintf(buf, "%u\n", info->rx_target); 1927 } 1928 1929 static struct device_attribute xennet_attrs[] = { 1930 __ATTR(rxbuf_min, S_IRUGO|S_IWUSR, show_rxbuf_min, store_rxbuf_min), 1931 __ATTR(rxbuf_max, S_IRUGO|S_IWUSR, show_rxbuf_max, store_rxbuf_max), 1932 __ATTR(rxbuf_cur, S_IRUGO, show_rxbuf_cur, NULL), 1933 }; 1934 1935 static int xennet_sysfs_addif(struct net_device *netdev) 1936 { 1937 int i; 1938 int err; 1939 1940 for (i = 0; i < ARRAY_SIZE(xennet_attrs); i++) { 1941 err = device_create_file(&netdev->dev, 1942 &xennet_attrs[i]); 1943 if (err) 1944 goto fail; 1945 } 1946 return 0; 1947 1948 fail: 1949 while (--i >= 0) 1950 device_remove_file(&netdev->dev, &xennet_attrs[i]); 1951 return err; 1952 } 1953 1954 static void xennet_sysfs_delif(struct net_device *netdev) 1955 { 1956 int i; 1957 1958 for (i = 0; i < ARRAY_SIZE(xennet_attrs); i++) 1959 device_remove_file(&netdev->dev, &xennet_attrs[i]); 1960 } 1961 1962 #endif /* CONFIG_SYSFS */ 1963 1964 static const struct xenbus_device_id netfront_ids[] = { 1965 { "vif" }, 1966 { "" } 1967 }; 1968 1969 1970 static int xennet_remove(struct xenbus_device *dev) 1971 { 1972 struct netfront_info *info = dev_get_drvdata(&dev->dev); 1973 1974 dev_dbg(&dev->dev, "%s\n", dev->nodename); 1975 1976 xennet_disconnect_backend(info); 1977 1978 xennet_sysfs_delif(info->netdev); 1979 1980 unregister_netdev(info->netdev); 1981 1982 del_timer_sync(&info->rx_refill_timer); 1983 1984 free_percpu(info->stats); 1985 1986 free_netdev(info->netdev); 1987 1988 return 0; 1989 } 1990 1991 static DEFINE_XENBUS_DRIVER(netfront, , 1992 .probe = netfront_probe, 1993 .remove = xennet_remove, 1994 .resume = netfront_resume, 1995 .otherend_changed = netback_changed, 1996 ); 1997 1998 static int __init netif_init(void) 1999 { 2000 if (!xen_domain()) 2001 return -ENODEV; 2002 2003 if (xen_hvm_domain() && !xen_platform_pci_unplug) 2004 return -ENODEV; 2005 2006 printk(KERN_INFO "Initialising Xen virtual ethernet driver.\n"); 2007 2008 return xenbus_register_frontend(&netfront_driver); 2009 } 2010 module_init(netif_init); 2011 2012 2013 static void __exit netif_exit(void) 2014 { 2015 xenbus_unregister_driver(&netfront_driver); 2016 } 2017 module_exit(netif_exit); 2018 2019 MODULE_DESCRIPTION("Xen virtual network device frontend"); 2020 MODULE_LICENSE("GPL"); 2021 MODULE_ALIAS("xen:vif"); 2022 MODULE_ALIAS("xennet"); 2023