drivers/net/xen-netfront.c

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