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