Pileus Git - ~andy/linux/blob - drivers/net/xen-netback/netback.c

   1 /*
   2  * Back-end of the driver for virtual network devices. This portion of the
   3  * driver exports a 'unified' network-device interface that can be accessed
   4  * by any operating system that implements a compatible front end. A
   5  * reference front-end implementation can be found in:
   6  *  drivers/net/xen-netfront.c
   7  *
   8  * Copyright (c) 2002-2005, K A Fraser
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public License version 2
  12  * as published by the Free Software Foundation; or, when distributed
  13  * separately from the Linux kernel or incorporated into other
  14  * software packages, subject to the following license:
  15  *
  16  * Permission is hereby granted, free of charge, to any person obtaining a copy
  17  * of this source file (the "Software"), to deal in the Software without
  18  * restriction, including without limitation the rights to use, copy, modify,
  19  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  20  * and to permit persons to whom the Software is furnished to do so, subject to
  21  * the following conditions:
  22  *
  23  * The above copyright notice and this permission notice shall be included in
  24  * all copies or substantial portions of the Software.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  27  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  28  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  29  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  30  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  31  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  32  * IN THE SOFTWARE.
  33  */
  34
  35 #include "common.h"
  36
  37 #include <linux/kthread.h>
  38 #include <linux/if_vlan.h>
  39 #include <linux/udp.h>
  40
  41 #include <net/tcp.h>
  42
  43 #include <xen/xen.h>
  44 #include <xen/events.h>
  45 #include <xen/interface/memory.h>
  46
  47 #include <asm/xen/hypercall.h>
  48 #include <asm/xen/page.h>
  49
  50 /*
  51  * This is the maximum slots a skb can have. If a guest sends a skb
  52  * which exceeds this limit it is considered malicious.
  53  */
  54 #define MAX_SKB_SLOTS_DEFAULT 20
  55 static unsigned int max_skb_slots = MAX_SKB_SLOTS_DEFAULT;
  56 module_param(max_skb_slots, uint, 0444);
  57
  58 typedef unsigned int pending_ring_idx_t;
  59 #define INVALID_PENDING_RING_IDX (~0U)
  60
  61 struct pending_tx_info {
  62         struct xen_netif_tx_request req; /* coalesced tx request */
  63         struct xenvif *vif;
  64         pending_ring_idx_t head; /* head != INVALID_PENDING_RING_IDX
  65                                   * if it is head of one or more tx
  66                                   * reqs
  67                                   */
  68 };
  69
  70 struct netbk_rx_meta {
  71         int id;
  72         int size;
  73         int gso_size;
  74 };
  75
  76 #define MAX_PENDING_REQS 256
  77
  78 /* Discriminate from any valid pending_idx value. */
  79 #define INVALID_PENDING_IDX 0xFFFF
  80
  81 #define MAX_BUFFER_OFFSET PAGE_SIZE
  82
  83 /* extra field used in struct page */
  84 union page_ext {
  85         struct {
  86 #if BITS_PER_LONG < 64
  87 #define IDX_WIDTH   8
  88 #define GROUP_WIDTH (BITS_PER_LONG - IDX_WIDTH)
  89                 unsigned int group:GROUP_WIDTH;
  90                 unsigned int idx:IDX_WIDTH;
  91 #else
  92                 unsigned int group, idx;
  93 #endif
  94         } e;
  95         void *mapping;
  96 };
  97
  98 struct xen_netbk {
  99         wait_queue_head_t wq;
 100         struct task_struct *task;
 101
 102         struct sk_buff_head rx_queue;
 103         struct sk_buff_head tx_queue;
 104
 105         struct timer_list net_timer;
 106
 107         struct page *mmap_pages[MAX_PENDING_REQS];
 108
 109         pending_ring_idx_t pending_prod;
 110         pending_ring_idx_t pending_cons;
 111         struct list_head net_schedule_list;
 112
 113         /* Protect the net_schedule_list in netif. */
 114         spinlock_t net_schedule_list_lock;
 115
 116         atomic_t netfront_count;
 117
 118         struct pending_tx_info pending_tx_info[MAX_PENDING_REQS];
 119         /* Coalescing tx requests before copying makes number of grant
 120          * copy ops greater or equal to number of slots required. In
 121          * worst case a tx request consumes 2 gnttab_copy.
 122          */
 123         struct gnttab_copy tx_copy_ops[2*MAX_PENDING_REQS];
 124
 125         u16 pending_ring[MAX_PENDING_REQS];
 126
 127         /*
 128          * Given MAX_BUFFER_OFFSET of 4096 the worst case is that each
 129          * head/fragment page uses 2 copy operations because it
 130          * straddles two buffers in the frontend.
 131          */
 132         struct gnttab_copy grant_copy_op[2*XEN_NETIF_RX_RING_SIZE];
 133         struct netbk_rx_meta meta[2*XEN_NETIF_RX_RING_SIZE];
 134 };
 135
 136 static struct xen_netbk *xen_netbk;
 137 static int xen_netbk_group_nr;
 138
 139 /*
 140  * If head != INVALID_PENDING_RING_IDX, it means this tx request is head of
 141  * one or more merged tx requests, otherwise it is the continuation of
 142  * previous tx request.
 143  */
 144 static inline int pending_tx_is_head(struct xen_netbk *netbk, RING_IDX idx)
 145 {
 146         return netbk->pending_tx_info[idx].head != INVALID_PENDING_RING_IDX;
 147 }
 148
 149 void xen_netbk_add_xenvif(struct xenvif *vif)
 150 {
 151         int i;
 152         int min_netfront_count;
 153         int min_group = 0;
 154         struct xen_netbk *netbk;
 155
 156         min_netfront_count = atomic_read(&xen_netbk[0].netfront_count);
 157         for (i = 0; i < xen_netbk_group_nr; i++) {
 158                 int netfront_count = atomic_read(&xen_netbk[i].netfront_count);
 159                 if (netfront_count < min_netfront_count) {
 160                         min_group = i;
 161                         min_netfront_count = netfront_count;
 162                 }
 163         }
 164
 165         netbk = &xen_netbk[min_group];
 166
 167         vif->netbk = netbk;
 168         atomic_inc(&netbk->netfront_count);
 169 }
 170
 171 void xen_netbk_remove_xenvif(struct xenvif *vif)
 172 {
 173         struct xen_netbk *netbk = vif->netbk;
 174         vif->netbk = NULL;
 175         atomic_dec(&netbk->netfront_count);
 176 }
 177
 178 static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx,
 179                                   u8 status);
 180 static void make_tx_response(struct xenvif *vif,
 181                              struct xen_netif_tx_request *txp,
 182                              s8       st);
 183 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
 184                                              u16      id,
 185                                              s8       st,
 186                                              u16      offset,
 187                                              u16      size,
 188                                              u16      flags);
 189
 190 static inline unsigned long idx_to_pfn(struct xen_netbk *netbk,
 191                                        u16 idx)
 192 {
 193         return page_to_pfn(netbk->mmap_pages[idx]);
 194 }
 195
 196 static inline unsigned long idx_to_kaddr(struct xen_netbk *netbk,
 197                                          u16 idx)
 198 {
 199         return (unsigned long)pfn_to_kaddr(idx_to_pfn(netbk, idx));
 200 }
 201
 202 /* extra field used in struct page */
 203 static inline void set_page_ext(struct page *pg, struct xen_netbk *netbk,
 204                                 unsigned int idx)
 205 {
 206         unsigned int group = netbk - xen_netbk;
 207         union page_ext ext = { .e = { .group = group + 1, .idx = idx } };
 208
 209         BUILD_BUG_ON(sizeof(ext) > sizeof(ext.mapping));
 210         pg->mapping = ext.mapping;
 211 }
 212
 213 static int get_page_ext(struct page *pg,
 214                         unsigned int *pgroup, unsigned int *pidx)
 215 {
 216         union page_ext ext = { .mapping = pg->mapping };
 217         struct xen_netbk *netbk;
 218         unsigned int group, idx;
 219
 220         group = ext.e.group - 1;
 221
 222         if (group < 0 || group >= xen_netbk_group_nr)
 223                 return 0;
 224
 225         netbk = &xen_netbk[group];
 226
 227         idx = ext.e.idx;
 228
 229         if ((idx < 0) || (idx >= MAX_PENDING_REQS))
 230                 return 0;
 231
 232         if (netbk->mmap_pages[idx] != pg)
 233                 return 0;
 234
 235         *pgroup = group;
 236         *pidx = idx;
 237
 238         return 1;
 239 }
 240
 241 /*
 242  * This is the amount of packet we copy rather than map, so that the
 243  * guest can't fiddle with the contents of the headers while we do
 244  * packet processing on them (netfilter, routing, etc).
 245  */
 246 #define PKT_PROT_LEN    (ETH_HLEN + \
 247                          VLAN_HLEN + \
 248                          sizeof(struct iphdr) + MAX_IPOPTLEN + \
 249                          sizeof(struct tcphdr) + MAX_TCP_OPTION_SPACE)
 250
 251 static u16 frag_get_pending_idx(skb_frag_t *frag)
 252 {
 253         return (u16)frag->page_offset;
 254 }
 255
 256 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 257 {
 258         frag->page_offset = pending_idx;
 259 }
 260
 261 static inline pending_ring_idx_t pending_index(unsigned i)
 262 {
 263         return i & (MAX_PENDING_REQS-1);
 264 }
 265
 266 static inline pending_ring_idx_t nr_pending_reqs(struct xen_netbk *netbk)
 267 {
 268         return MAX_PENDING_REQS -
 269                 netbk->pending_prod + netbk->pending_cons;
 270 }
 271
 272 static void xen_netbk_kick_thread(struct xen_netbk *netbk)
 273 {
 274         wake_up(&netbk->wq);
 275 }
 276
 277 static int max_required_rx_slots(struct xenvif *vif)
 278 {
 279         int max = DIV_ROUND_UP(vif->dev->mtu, PAGE_SIZE);
 280
 281         /* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
 282         if (vif->can_sg || vif->gso || vif->gso_prefix)
 283                 max += MAX_SKB_FRAGS + 1; /* extra_info + frags */
 284
 285         return max;
 286 }
 287
 288 int xen_netbk_rx_ring_full(struct xenvif *vif)
 289 {
 290         RING_IDX peek   = vif->rx_req_cons_peek;
 291         RING_IDX needed = max_required_rx_slots(vif);
 292
 293         return ((vif->rx.sring->req_prod - peek) < needed) ||
 294                ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
 295 }
 296
 297 int xen_netbk_must_stop_queue(struct xenvif *vif)
 298 {
 299         if (!xen_netbk_rx_ring_full(vif))
 300                 return 0;
 301
 302         vif->rx.sring->req_event = vif->rx_req_cons_peek +
 303                 max_required_rx_slots(vif);
 304         mb(); /* request notification /then/ check the queue */
 305
 306         return xen_netbk_rx_ring_full(vif);
 307 }
 308
 309 /*
 310  * Returns true if we should start a new receive buffer instead of
 311  * adding 'size' bytes to a buffer which currently contains 'offset'
 312  * bytes.
 313  */
 314 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
 315 {
 316         /* simple case: we have completely filled the current buffer. */
 317         if (offset == MAX_BUFFER_OFFSET)
 318                 return true;
 319
 320         /*
 321          * complex case: start a fresh buffer if the current frag
 322          * would overflow the current buffer but only if:
 323          *     (i)   this frag would fit completely in the next buffer
 324          * and (ii)  there is already some data in the current buffer
 325          * and (iii) this is not the head buffer.
 326          *
 327          * Where:
 328          * - (i) stops us splitting a frag into two copies
 329          *   unless the frag is too large for a single buffer.
 330          * - (ii) stops us from leaving a buffer pointlessly empty.
 331          * - (iii) stops us leaving the first buffer
 332          *   empty. Strictly speaking this is already covered
 333          *   by (ii) but is explicitly checked because
 334          *   netfront relies on the first buffer being
 335          *   non-empty and can crash otherwise.
 336          *
 337          * This means we will effectively linearise small
 338          * frags but do not needlessly split large buffers
 339          * into multiple copies tend to give large frags their
 340          * own buffers as before.
 341          */
 342         if ((offset + size > MAX_BUFFER_OFFSET) &&
 343             (size <= MAX_BUFFER_OFFSET) && offset && !head)
 344                 return true;
 345
 346         return false;
 347 }
 348
 349 /*
 350  * Figure out how many ring slots we're going to need to send @skb to
 351  * the guest. This function is essentially a dry run of
 352  * netbk_gop_frag_copy.
 353  */
 354 unsigned int xen_netbk_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
 355 {
 356         unsigned int count;
 357         int i, copy_off;
 358
 359         count = DIV_ROUND_UP(skb_headlen(skb), PAGE_SIZE);
 360
 361         copy_off = skb_headlen(skb) % PAGE_SIZE;
 362
 363         if (skb_shinfo(skb)->gso_size)
 364                 count++;
 365
 366         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 367                 unsigned long size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
 368                 unsigned long offset = skb_shinfo(skb)->frags[i].page_offset;
 369                 unsigned long bytes;
 370
 371                 offset &= ~PAGE_MASK;
 372
 373                 while (size > 0) {
 374                         BUG_ON(offset >= PAGE_SIZE);
 375                         BUG_ON(copy_off > MAX_BUFFER_OFFSET);
 376
 377                         bytes = PAGE_SIZE - offset;
 378
 379                         if (bytes > size)
 380                                 bytes = size;
 381
 382                         if (start_new_rx_buffer(copy_off, bytes, 0)) {
 383                                 count++;
 384                                 copy_off = 0;
 385                         }
 386
 387                         if (copy_off + bytes > MAX_BUFFER_OFFSET)
 388                                 bytes = MAX_BUFFER_OFFSET - copy_off;
 389
 390                         copy_off += bytes;
 391
 392                         offset += bytes;
 393                         size -= bytes;
 394
 395                         if (offset == PAGE_SIZE)
 396                                 offset = 0;
 397                 }
 398         }
 399         return count;
 400 }
 401
 402 struct netrx_pending_operations {
 403         unsigned copy_prod, copy_cons;
 404         unsigned meta_prod, meta_cons;
 405         struct gnttab_copy *copy;
 406         struct netbk_rx_meta *meta;
 407         int copy_off;
 408         grant_ref_t copy_gref;
 409 };
 410
 411 static struct netbk_rx_meta *get_next_rx_buffer(struct xenvif *vif,
 412                                                 struct netrx_pending_operations *npo)
 413 {
 414         struct netbk_rx_meta *meta;
 415         struct xen_netif_rx_request *req;
 416
 417         req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
 418
 419         meta = npo->meta + npo->meta_prod++;
 420         meta->gso_size = 0;
 421         meta->size = 0;
 422         meta->id = req->id;
 423
 424         npo->copy_off = 0;
 425         npo->copy_gref = req->gref;
 426
 427         return meta;
 428 }
 429
 430 /*
 431  * Set up the grant operations for this fragment. If it's a flipping
 432  * interface, we also set up the unmap request from here.
 433  */
 434 static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
 435                                 struct netrx_pending_operations *npo,
 436                                 struct page *page, unsigned long size,
 437                                 unsigned long offset, int *head)
 438 {
 439         struct gnttab_copy *copy_gop;
 440         struct netbk_rx_meta *meta;
 441         /*
 442          * These variables are used iff get_page_ext returns true,
 443          * in which case they are guaranteed to be initialized.
 444          */
 445         unsigned int uninitialized_var(group), uninitialized_var(idx);
 446         int foreign = get_page_ext(page, &group, &idx);
 447         unsigned long bytes;
 448
 449         /* Data must not cross a page boundary. */
 450         BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
 451
 452         meta = npo->meta + npo->meta_prod - 1;
 453
 454         /* Skip unused frames from start of page */
 455         page += offset >> PAGE_SHIFT;
 456         offset &= ~PAGE_MASK;
 457
 458         while (size > 0) {
 459                 BUG_ON(offset >= PAGE_SIZE);
 460                 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
 461
 462                 bytes = PAGE_SIZE - offset;
 463
 464                 if (bytes > size)
 465                         bytes = size;
 466
 467                 if (start_new_rx_buffer(npo->copy_off, bytes, *head)) {
 468                         /*
 469                          * Netfront requires there to be some data in the head
 470                          * buffer.
 471                          */
 472                         BUG_ON(*head);
 473
 474                         meta = get_next_rx_buffer(vif, npo);
 475                 }
 476
 477                 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
 478                         bytes = MAX_BUFFER_OFFSET - npo->copy_off;
 479
 480                 copy_gop = npo->copy + npo->copy_prod++;
 481                 copy_gop->flags = GNTCOPY_dest_gref;
 482                 if (foreign) {
 483                         struct xen_netbk *netbk = &xen_netbk[group];
 484                         struct pending_tx_info *src_pend;
 485
 486                         src_pend = &netbk->pending_tx_info[idx];
 487
 488                         copy_gop->source.domid = src_pend->vif->domid;
 489                         copy_gop->source.u.ref = src_pend->req.gref;
 490                         copy_gop->flags |= GNTCOPY_source_gref;
 491                 } else {
 492                         void *vaddr = page_address(page);
 493                         copy_gop->source.domid = DOMID_SELF;
 494                         copy_gop->source.u.gmfn = virt_to_mfn(vaddr);
 495                 }
 496                 copy_gop->source.offset = offset;
 497                 copy_gop->dest.domid = vif->domid;
 498
 499                 copy_gop->dest.offset = npo->copy_off;
 500                 copy_gop->dest.u.ref = npo->copy_gref;
 501                 copy_gop->len = bytes;
 502
 503                 npo->copy_off += bytes;
 504                 meta->size += bytes;
 505
 506                 offset += bytes;
 507                 size -= bytes;
 508
 509                 /* Next frame */
 510                 if (offset == PAGE_SIZE && size) {
 511                         BUG_ON(!PageCompound(page));
 512                         page++;
 513                         offset = 0;
 514                 }
 515
 516                 /* Leave a gap for the GSO descriptor. */
 517                 if (*head && skb_shinfo(skb)->gso_size && !vif->gso_prefix)
 518                         vif->rx.req_cons++;
 519
 520                 *head = 0; /* There must be something in this buffer now. */
 521
 522         }
 523 }
 524
 525 /*
 526  * Prepare an SKB to be transmitted to the frontend.
 527  *
 528  * This function is responsible for allocating grant operations, meta
 529  * structures, etc.
 530  *
 531  * It returns the number of meta structures consumed. The number of
 532  * ring slots used is always equal to the number of meta slots used
 533  * plus the number of GSO descriptors used. Currently, we use either
 534  * zero GSO descriptors (for non-GSO packets) or one descriptor (for
 535  * frontend-side LRO).
 536  */
 537 static int netbk_gop_skb(struct sk_buff *skb,
 538                          struct netrx_pending_operations *npo)
 539 {
 540         struct xenvif *vif = netdev_priv(skb->dev);
 541         int nr_frags = skb_shinfo(skb)->nr_frags;
 542         int i;
 543         struct xen_netif_rx_request *req;
 544         struct netbk_rx_meta *meta;
 545         unsigned char *data;
 546         int head = 1;
 547         int old_meta_prod;
 548
 549         old_meta_prod = npo->meta_prod;
 550
 551         /* Set up a GSO prefix descriptor, if necessary */
 552         if (skb_shinfo(skb)->gso_size && vif->gso_prefix) {
 553                 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
 554                 meta = npo->meta + npo->meta_prod++;
 555                 meta->gso_size = skb_shinfo(skb)->gso_size;
 556                 meta->size = 0;
 557                 meta->id = req->id;
 558         }
 559
 560         req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
 561         meta = npo->meta + npo->meta_prod++;
 562
 563         if (!vif->gso_prefix)
 564                 meta->gso_size = skb_shinfo(skb)->gso_size;
 565         else
 566                 meta->gso_size = 0;
 567
 568         meta->size = 0;
 569         meta->id = req->id;
 570         npo->copy_off = 0;
 571         npo->copy_gref = req->gref;
 572
 573         data = skb->data;
 574         while (data < skb_tail_pointer(skb)) {
 575                 unsigned int offset = offset_in_page(data);
 576                 unsigned int len = PAGE_SIZE - offset;
 577
 578                 if (data + len > skb_tail_pointer(skb))
 579                         len = skb_tail_pointer(skb) - data;
 580
 581                 netbk_gop_frag_copy(vif, skb, npo,
 582                                     virt_to_page(data), len, offset, &head);
 583                 data += len;
 584         }
 585
 586         for (i = 0; i < nr_frags; i++) {
 587                 netbk_gop_frag_copy(vif, skb, npo,
 588                                     skb_frag_page(&skb_shinfo(skb)->frags[i]),
 589                                     skb_frag_size(&skb_shinfo(skb)->frags[i]),
 590                                     skb_shinfo(skb)->frags[i].page_offset,
 591                                     &head);
 592         }
 593
 594         return npo->meta_prod - old_meta_prod;
 595 }
 596
 597 /*
 598  * This is a twin to netbk_gop_skb.  Assume that netbk_gop_skb was
 599  * used to set up the operations on the top of
 600  * netrx_pending_operations, which have since been done.  Check that
 601  * they didn't give any errors and advance over them.
 602  */
 603 static int netbk_check_gop(struct xenvif *vif, int nr_meta_slots,
 604                            struct netrx_pending_operations *npo)
 605 {
 606         struct gnttab_copy     *copy_op;
 607         int status = XEN_NETIF_RSP_OKAY;
 608         int i;
 609
 610         for (i = 0; i < nr_meta_slots; i++) {
 611                 copy_op = npo->copy + npo->copy_cons++;
 612                 if (copy_op->status != GNTST_okay) {
 613                         netdev_dbg(vif->dev,
 614                                    "Bad status %d from copy to DOM%d.\n",
 615                                    copy_op->status, vif->domid);
 616                         status = XEN_NETIF_RSP_ERROR;
 617                 }
 618         }
 619
 620         return status;
 621 }
 622
 623 static void netbk_add_frag_responses(struct xenvif *vif, int status,
 624                                      struct netbk_rx_meta *meta,
 625                                      int nr_meta_slots)
 626 {
 627         int i;
 628         unsigned long offset;
 629
 630         /* No fragments used */
 631         if (nr_meta_slots <= 1)
 632                 return;
 633
 634         nr_meta_slots--;
 635
 636         for (i = 0; i < nr_meta_slots; i++) {
 637                 int flags;
 638                 if (i == nr_meta_slots - 1)
 639                         flags = 0;
 640                 else
 641                         flags = XEN_NETRXF_more_data;
 642
 643                 offset = 0;
 644                 make_rx_response(vif, meta[i].id, status, offset,
 645                                  meta[i].size, flags);
 646         }
 647 }
 648
 649 struct skb_cb_overlay {
 650         int meta_slots_used;
 651 };
 652
 653 static void xen_netbk_rx_action(struct xen_netbk *netbk)
 654 {
 655         struct xenvif *vif = NULL, *tmp;
 656         s8 status;
 657         u16 irq, flags;
 658         struct xen_netif_rx_response *resp;
 659         struct sk_buff_head rxq;
 660         struct sk_buff *skb;
 661         LIST_HEAD(notify);
 662         int ret;
 663         int nr_frags;
 664         int count;
 665         unsigned long offset;
 666         struct skb_cb_overlay *sco;
 667
 668         struct netrx_pending_operations npo = {
 669                 .copy  = netbk->grant_copy_op,
 670                 .meta  = netbk->meta,
 671         };
 672
 673         skb_queue_head_init(&rxq);
 674
 675         count = 0;
 676
 677         while ((skb = skb_dequeue(&netbk->rx_queue)) != NULL) {
 678                 vif = netdev_priv(skb->dev);
 679                 nr_frags = skb_shinfo(skb)->nr_frags;
 680
 681                 sco = (struct skb_cb_overlay *)skb->cb;
 682                 sco->meta_slots_used = netbk_gop_skb(skb, &npo);
 683
 684                 count += nr_frags + 1;
 685
 686                 __skb_queue_tail(&rxq, skb);
 687
 688                 /* Filled the batch queue? */
 689                 /* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
 690                 if (count + MAX_SKB_FRAGS >= XEN_NETIF_RX_RING_SIZE)
 691                         break;
 692         }
 693
 694         BUG_ON(npo.meta_prod > ARRAY_SIZE(netbk->meta));
 695
 696         if (!npo.copy_prod)
 697                 return;
 698
 699         BUG_ON(npo.copy_prod > ARRAY_SIZE(netbk->grant_copy_op));
 700         gnttab_batch_copy(netbk->grant_copy_op, npo.copy_prod);
 701
 702         while ((skb = __skb_dequeue(&rxq)) != NULL) {
 703                 sco = (struct skb_cb_overlay *)skb->cb;
 704
 705                 vif = netdev_priv(skb->dev);
 706
 707                 if (netbk->meta[npo.meta_cons].gso_size && vif->gso_prefix) {
 708                         resp = RING_GET_RESPONSE(&vif->rx,
 709                                                 vif->rx.rsp_prod_pvt++);
 710
 711                         resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
 712
 713                         resp->offset = netbk->meta[npo.meta_cons].gso_size;
 714                         resp->id = netbk->meta[npo.meta_cons].id;
 715                         resp->status = sco->meta_slots_used;
 716
 717                         npo.meta_cons++;
 718                         sco->meta_slots_used--;
 719                 }
 720
 721
 722                 vif->dev->stats.tx_bytes += skb->len;
 723                 vif->dev->stats.tx_packets++;
 724
 725                 status = netbk_check_gop(vif, sco->meta_slots_used, &npo);
 726
 727                 if (sco->meta_slots_used == 1)
 728                         flags = 0;
 729                 else
 730                         flags = XEN_NETRXF_more_data;
 731
 732                 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
 733                         flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
 734                 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 735                         /* remote but checksummed. */
 736                         flags |= XEN_NETRXF_data_validated;
 737
 738                 offset = 0;
 739                 resp = make_rx_response(vif, netbk->meta[npo.meta_cons].id,
 740                                         status, offset,
 741                                         netbk->meta[npo.meta_cons].size,
 742                                         flags);
 743
 744                 if (netbk->meta[npo.meta_cons].gso_size && !vif->gso_prefix) {
 745                         struct xen_netif_extra_info *gso =
 746                                 (struct xen_netif_extra_info *)
 747                                 RING_GET_RESPONSE(&vif->rx,
 748                                                   vif->rx.rsp_prod_pvt++);
 749
 750                         resp->flags |= XEN_NETRXF_extra_info;
 751
 752                         gso->u.gso.size = netbk->meta[npo.meta_cons].gso_size;
 753                         gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
 754                         gso->u.gso.pad = 0;
 755                         gso->u.gso.features = 0;
 756
 757                         gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
 758                         gso->flags = 0;
 759                 }
 760
 761                 netbk_add_frag_responses(vif, status,
 762                                          netbk->meta + npo.meta_cons + 1,
 763                                          sco->meta_slots_used);
 764
 765                 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
 766                 irq = vif->irq;
 767                 if (ret && list_empty(&vif->notify_list))
 768                         list_add_tail(&vif->notify_list, &notify);
 769
 770                 xenvif_notify_tx_completion(vif);
 771
 772                 xenvif_put(vif);
 773                 npo.meta_cons += sco->meta_slots_used;
 774                 dev_kfree_skb(skb);
 775         }
 776
 777         list_for_each_entry_safe(vif, tmp, &notify, notify_list) {
 778                 notify_remote_via_irq(vif->irq);
 779                 list_del_init(&vif->notify_list);
 780         }
 781
 782         /* More work to do? */
 783         if (!skb_queue_empty(&netbk->rx_queue) &&
 784                         !timer_pending(&netbk->net_timer))
 785                 xen_netbk_kick_thread(netbk);
 786 }
 787
 788 void xen_netbk_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb)
 789 {
 790         struct xen_netbk *netbk = vif->netbk;
 791
 792         skb_queue_tail(&netbk->rx_queue, skb);
 793
 794         xen_netbk_kick_thread(netbk);
 795 }
 796
 797 static void xen_netbk_alarm(unsigned long data)
 798 {
 799         struct xen_netbk *netbk = (struct xen_netbk *)data;
 800         xen_netbk_kick_thread(netbk);
 801 }
 802
 803 static int __on_net_schedule_list(struct xenvif *vif)
 804 {
 805         return !list_empty(&vif->schedule_list);
 806 }
 807
 808 /* Must be called with net_schedule_list_lock held */
 809 static void remove_from_net_schedule_list(struct xenvif *vif)
 810 {
 811         if (likely(__on_net_schedule_list(vif))) {
 812                 list_del_init(&vif->schedule_list);
 813                 xenvif_put(vif);
 814         }
 815 }
 816
 817 static struct xenvif *poll_net_schedule_list(struct xen_netbk *netbk)
 818 {
 819         struct xenvif *vif = NULL;
 820
 821         spin_lock_irq(&netbk->net_schedule_list_lock);
 822         if (list_empty(&netbk->net_schedule_list))
 823                 goto out;
 824
 825         vif = list_first_entry(&netbk->net_schedule_list,
 826                                struct xenvif, schedule_list);
 827         if (!vif)
 828                 goto out;
 829
 830         xenvif_get(vif);
 831
 832         remove_from_net_schedule_list(vif);
 833 out:
 834         spin_unlock_irq(&netbk->net_schedule_list_lock);
 835         return vif;
 836 }
 837
 838 void xen_netbk_schedule_xenvif(struct xenvif *vif)
 839 {
 840         unsigned long flags;
 841         struct xen_netbk *netbk = vif->netbk;
 842
 843         if (__on_net_schedule_list(vif))
 844                 goto kick;
 845
 846         spin_lock_irqsave(&netbk->net_schedule_list_lock, flags);
 847         if (!__on_net_schedule_list(vif) &&
 848             likely(xenvif_schedulable(vif))) {
 849                 list_add_tail(&vif->schedule_list, &netbk->net_schedule_list);
 850                 xenvif_get(vif);
 851         }
 852         spin_unlock_irqrestore(&netbk->net_schedule_list_lock, flags);
 853
 854 kick:
 855         smp_mb();
 856         if ((nr_pending_reqs(netbk) < (MAX_PENDING_REQS/2)) &&
 857             !list_empty(&netbk->net_schedule_list))
 858                 xen_netbk_kick_thread(netbk);
 859 }
 860
 861 void xen_netbk_deschedule_xenvif(struct xenvif *vif)
 862 {
 863         struct xen_netbk *netbk = vif->netbk;
 864         spin_lock_irq(&netbk->net_schedule_list_lock);
 865         remove_from_net_schedule_list(vif);
 866         spin_unlock_irq(&netbk->net_schedule_list_lock);
 867 }
 868
 869 void xen_netbk_check_rx_xenvif(struct xenvif *vif)
 870 {
 871         int more_to_do;
 872
 873         RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
 874
 875         if (more_to_do)
 876                 xen_netbk_schedule_xenvif(vif);
 877 }
 878
 879 static void tx_add_credit(struct xenvif *vif)
 880 {
 881         unsigned long max_burst, max_credit;
 882
 883         /*
 884          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 885          * Otherwise the interface can seize up due to insufficient credit.
 886          */
 887         max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
 888         max_burst = min(max_burst, 131072UL);
 889         max_burst = max(max_burst, vif->credit_bytes);
 890
 891         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 892         max_credit = vif->remaining_credit + vif->credit_bytes;
 893         if (max_credit < vif->remaining_credit)
 894                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 895
 896         vif->remaining_credit = min(max_credit, max_burst);
 897 }
 898
 899 static void tx_credit_callback(unsigned long data)
 900 {
 901         struct xenvif *vif = (struct xenvif *)data;
 902         tx_add_credit(vif);
 903         xen_netbk_check_rx_xenvif(vif);
 904 }
 905
 906 static void netbk_tx_err(struct xenvif *vif,
 907                          struct xen_netif_tx_request *txp, RING_IDX end)
 908 {
 909         RING_IDX cons = vif->tx.req_cons;
 910
 911         do {
 912                 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
 913                 if (cons == end)
 914                         break;
 915                 txp = RING_GET_REQUEST(&vif->tx, cons++);
 916         } while (1);
 917         vif->tx.req_cons = cons;
 918         xen_netbk_check_rx_xenvif(vif);
 919         xenvif_put(vif);
 920 }
 921
 922 static void netbk_fatal_tx_err(struct xenvif *vif)
 923 {
 924         netdev_err(vif->dev, "fatal error; disabling device\n");
 925         xenvif_carrier_off(vif);
 926         xenvif_put(vif);
 927 }
 928
 929 static int netbk_count_requests(struct xenvif *vif,
 930                                 struct xen_netif_tx_request *first,
 931                                 RING_IDX first_idx,
 932                                 struct xen_netif_tx_request *txp,
 933                                 int work_to_do)
 934 {
 935         RING_IDX cons = vif->tx.req_cons;
 936         int slots = 0;
 937         int drop_err = 0;
 938
 939         if (!(first->flags & XEN_NETTXF_more_data))
 940                 return 0;
 941
 942         do {
 943                 if (slots >= work_to_do) {
 944                         netdev_err(vif->dev,
 945                                    "Asked for %d slots but exceeds this limit\n",
 946                                    work_to_do);
 947                         netbk_fatal_tx_err(vif);
 948                         return -ENODATA;
 949                 }
 950
 951                 /* This guest is really using too many slots and
 952                  * considered malicious.
 953                  */
 954                 if (unlikely(slots >= max_skb_slots)) {
 955                         netdev_err(vif->dev,
 956                                    "Malicious frontend using %d slots, threshold %u\n",
 957                                    slots, max_skb_slots);
 958                         netbk_fatal_tx_err(vif);
 959                         return -E2BIG;
 960                 }
 961
 962                 /* Xen network protocol had implicit dependency on
 963                  * MAX_SKB_FRAGS. XEN_NETIF_NR_SLOTS_MIN is set to the
 964                  * historical MAX_SKB_FRAGS value 18 to honor the same
 965                  * behavior as before. Any packet using more than 18
 966                  * slots but less than max_skb_slots slots is dropped
 967                  */
 968                 if (!drop_err && slots >= XEN_NETIF_NR_SLOTS_MIN) {
 969                         if (net_ratelimit())
 970                                 netdev_dbg(vif->dev,
 971                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 972                                            slots, XEN_NETIF_NR_SLOTS_MIN);
 973                         drop_err = -E2BIG;
 974                 }
 975
 976                 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + slots),
 977                        sizeof(*txp));
 978                 if (txp->size > first->size) {
 979                         netdev_err(vif->dev,
 980                                    "Invalid tx request, slot size %u > remaining size %u\n",
 981                                    txp->size, first->size);
 982                         netbk_fatal_tx_err(vif);
 983                         return -EIO;
 984                 }
 985
 986                 first->size -= txp->size;
 987                 slots++;
 988
 989                 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
 990                         netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
 991                                  txp->offset, txp->size);
 992                         netbk_fatal_tx_err(vif);
 993                         return -EINVAL;
 994                 }
 995         } while ((txp++)->flags & XEN_NETTXF_more_data);
 996
 997         if (drop_err) {
 998                 netbk_tx_err(vif, first, first_idx + slots);
 999                 return drop_err;
1000         }
1001
1002         return slots;
1003 }
1004
1005 static struct page *xen_netbk_alloc_page(struct xen_netbk *netbk,
1006                                          u16 pending_idx)
1007 {
1008         struct page *page;
1009         page = alloc_page(GFP_KERNEL|__GFP_COLD);
1010         if (!page)
1011                 return NULL;
1012         set_page_ext(page, netbk, pending_idx);
1013         netbk->mmap_pages[pending_idx] = page;
1014         return page;
1015 }
1016
1017 static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
1018                                                   struct xenvif *vif,
1019                                                   struct sk_buff *skb,
1020                                                   struct xen_netif_tx_request *txp,
1021                                                   struct gnttab_copy *gop)
1022 {
1023         struct skb_shared_info *shinfo = skb_shinfo(skb);
1024         skb_frag_t *frags = shinfo->frags;
1025         u16 pending_idx = *((u16 *)skb->data);
1026         u16 head_idx = 0;
1027         int slot, start;
1028         struct page *page;
1029         pending_ring_idx_t index, start_idx = 0;
1030         uint16_t dst_offset;
1031         unsigned int nr_slots;
1032         struct pending_tx_info *first = NULL;
1033
1034         /* At this point shinfo->nr_frags is in fact the number of
1035          * slots, which can be as large as XEN_NETIF_NR_SLOTS_MIN.
1036          */
1037         nr_slots = shinfo->nr_frags;
1038
1039         /* Skip first skb fragment if it is on same page as header fragment. */
1040         start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
1041
1042         /* Coalesce tx requests, at this point the packet passed in
1043          * should be <= 64K. Any packets larger than 64K have been
1044          * handled in netbk_count_requests().
1045          */
1046         for (shinfo->nr_frags = slot = start; slot < nr_slots;
1047              shinfo->nr_frags++) {
1048                 struct pending_tx_info *pending_tx_info =
1049                         netbk->pending_tx_info;
1050
1051                 page = alloc_page(GFP_KERNEL|__GFP_COLD);
1052                 if (!page)
1053                         goto err;
1054
1055                 dst_offset = 0;
1056                 first = NULL;
1057                 while (dst_offset < PAGE_SIZE && slot < nr_slots) {
1058                         gop->flags = GNTCOPY_source_gref;
1059
1060                         gop->source.u.ref = txp->gref;
1061                         gop->source.domid = vif->domid;
1062                         gop->source.offset = txp->offset;
1063
1064                         gop->dest.domid = DOMID_SELF;
1065
1066                         gop->dest.offset = dst_offset;
1067                         gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1068
1069                         if (dst_offset + txp->size > PAGE_SIZE) {
1070                                 /* This page can only merge a portion
1071                                  * of tx request. Do not increment any
1072                                  * pointer / counter here. The txp
1073                                  * will be dealt with in future
1074                                  * rounds, eventually hitting the
1075                                  * `else` branch.
1076                                  */
1077                                 gop->len = PAGE_SIZE - dst_offset;
1078                                 txp->offset += gop->len;
1079                                 txp->size -= gop->len;
1080                                 dst_offset += gop->len; /* quit loop */
1081                         } else {
1082                                 /* This tx request can be merged in the page */
1083                                 gop->len = txp->size;
1084                                 dst_offset += gop->len;
1085
1086                                 index = pending_index(netbk->pending_cons++);
1087
1088                                 pending_idx = netbk->pending_ring[index];
1089
1090                                 memcpy(&pending_tx_info[pending_idx].req, txp,
1091                                        sizeof(*txp));
1092                                 xenvif_get(vif);
1093
1094                                 pending_tx_info[pending_idx].vif = vif;
1095
1096                                 /* Poison these fields, corresponding
1097                                  * fields for head tx req will be set
1098                                  * to correct values after the loop.
1099                                  */
1100                                 netbk->mmap_pages[pending_idx] = (void *)(~0UL);
1101                                 pending_tx_info[pending_idx].head =
1102                                         INVALID_PENDING_RING_IDX;
1103
1104                                 if (!first) {
1105                                         first = &pending_tx_info[pending_idx];
1106                                         start_idx = index;
1107                                         head_idx = pending_idx;
1108                                 }
1109
1110                                 txp++;
1111                                 slot++;
1112                         }
1113
1114                         gop++;
1115                 }
1116
1117                 first->req.offset = 0;
1118                 first->req.size = dst_offset;
1119                 first->head = start_idx;
1120                 set_page_ext(page, netbk, head_idx);
1121                 netbk->mmap_pages[head_idx] = page;
1122                 frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
1123         }
1124
1125         BUG_ON(shinfo->nr_frags > MAX_SKB_FRAGS);
1126
1127         return gop;
1128 err:
1129         /* Unwind, freeing all pages and sending error responses. */
1130         while (shinfo->nr_frags-- > start) {
1131                 xen_netbk_idx_release(netbk,
1132                                 frag_get_pending_idx(&frags[shinfo->nr_frags]),
1133                                 XEN_NETIF_RSP_ERROR);
1134         }
1135         /* The head too, if necessary. */
1136         if (start)
1137                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
1138
1139         return NULL;
1140 }
1141
1142 static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
1143                                   struct sk_buff *skb,
1144                                   struct gnttab_copy **gopp)
1145 {
1146         struct gnttab_copy *gop = *gopp;
1147         u16 pending_idx = *((u16 *)skb->data);
1148         struct skb_shared_info *shinfo = skb_shinfo(skb);
1149         struct pending_tx_info *tx_info;
1150         int nr_frags = shinfo->nr_frags;
1151         int i, err, start;
1152         u16 peek; /* peek into next tx request */
1153
1154         /* Check status of header. */
1155         err = gop->status;
1156         if (unlikely(err))
1157                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
1158
1159         /* Skip first skb fragment if it is on same page as header fragment. */
1160         start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
1161
1162         for (i = start; i < nr_frags; i++) {
1163                 int j, newerr;
1164                 pending_ring_idx_t head;
1165
1166                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
1167                 tx_info = &netbk->pending_tx_info[pending_idx];
1168                 head = tx_info->head;
1169
1170                 /* Check error status: if okay then remember grant handle. */
1171                 do {
1172                         newerr = (++gop)->status;
1173                         if (newerr)
1174                                 break;
1175                         peek = netbk->pending_ring[pending_index(++head)];
1176                 } while (!pending_tx_is_head(netbk, peek));
1177
1178                 if (likely(!newerr)) {
1179                         /* Had a previous error? Invalidate this fragment. */
1180                         if (unlikely(err))
1181                                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1182                         continue;
1183                 }
1184
1185                 /* Error on this fragment: respond to client with an error. */
1186                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
1187
1188                 /* Not the first error? Preceding frags already invalidated. */
1189                 if (err)
1190                         continue;
1191
1192                 /* First error: invalidate header and preceding fragments. */
1193                 pending_idx = *((u16 *)skb->data);
1194                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1195                 for (j = start; j < i; j++) {
1196                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1197                         xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1198                 }
1199
1200                 /* Remember the error: invalidate all subsequent fragments. */
1201                 err = newerr;
1202         }
1203
1204         *gopp = gop + 1;
1205         return err;
1206 }
1207
1208 static void xen_netbk_fill_frags(struct xen_netbk *netbk, struct sk_buff *skb)
1209 {
1210         struct skb_shared_info *shinfo = skb_shinfo(skb);
1211         int nr_frags = shinfo->nr_frags;
1212         int i;
1213
1214         for (i = 0; i < nr_frags; i++) {
1215                 skb_frag_t *frag = shinfo->frags + i;
1216                 struct xen_netif_tx_request *txp;
1217                 struct page *page;
1218                 u16 pending_idx;
1219
1220                 pending_idx = frag_get_pending_idx(frag);
1221
1222                 txp = &netbk->pending_tx_info[pending_idx].req;
1223                 page = virt_to_page(idx_to_kaddr(netbk, pending_idx));
1224                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1225                 skb->len += txp->size;
1226                 skb->data_len += txp->size;
1227                 skb->truesize += txp->size;
1228
1229                 /* Take an extra reference to offset xen_netbk_idx_release */
1230                 get_page(netbk->mmap_pages[pending_idx]);
1231                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1232         }
1233 }
1234
1235 static int xen_netbk_get_extras(struct xenvif *vif,
1236                                 struct xen_netif_extra_info *extras,
1237                                 int work_to_do)
1238 {
1239         struct xen_netif_extra_info extra;
1240         RING_IDX cons = vif->tx.req_cons;
1241
1242         do {
1243                 if (unlikely(work_to_do-- <= 0)) {
1244                         netdev_err(vif->dev, "Missing extra info\n");
1245                         netbk_fatal_tx_err(vif);
1246                         return -EBADR;
1247                 }
1248
1249                 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1250                        sizeof(extra));
1251                 if (unlikely(!extra.type ||
1252                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1253                         vif->tx.req_cons = ++cons;
1254                         netdev_err(vif->dev,
1255                                    "Invalid extra type: %d\n", extra.type);
1256                         netbk_fatal_tx_err(vif);
1257                         return -EINVAL;
1258                 }
1259
1260                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1261                 vif->tx.req_cons = ++cons;
1262         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1263
1264         return work_to_do;
1265 }
1266
1267 static int netbk_set_skb_gso(struct xenvif *vif,
1268                              struct sk_buff *skb,
1269                              struct xen_netif_extra_info *gso)
1270 {
1271         if (!gso->u.gso.size) {
1272                 netdev_err(vif->dev, "GSO size must not be zero.\n");
1273                 netbk_fatal_tx_err(vif);
1274                 return -EINVAL;
1275         }
1276
1277         /* Currently only TCPv4 S.O. is supported. */
1278         if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
1279                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1280                 netbk_fatal_tx_err(vif);
1281                 return -EINVAL;
1282         }
1283
1284         skb_shinfo(skb)->gso_size = gso->u.gso.size;
1285         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1286
1287         /* Header must be checked, and gso_segs computed. */
1288         skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1289         skb_shinfo(skb)->gso_segs = 0;
1290
1291         return 0;
1292 }
1293
1294 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1295 {
1296         struct iphdr *iph;
1297         int err = -EPROTO;
1298         int recalculate_partial_csum = 0;
1299
1300         /*
1301          * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1302          * peers can fail to set NETRXF_csum_blank when sending a GSO
1303          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1304          * recalculate the partial checksum.
1305          */
1306         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1307                 vif->rx_gso_checksum_fixup++;
1308                 skb->ip_summed = CHECKSUM_PARTIAL;
1309                 recalculate_partial_csum = 1;
1310         }
1311
1312         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1313         if (skb->ip_summed != CHECKSUM_PARTIAL)
1314                 return 0;
1315
1316         if (skb->protocol != htons(ETH_P_IP))
1317                 goto out;
1318
1319         iph = (void *)skb->data;
1320         switch (iph->protocol) {
1321         case IPPROTO_TCP:
1322                 if (!skb_partial_csum_set(skb, 4 * iph->ihl,
1323                                           offsetof(struct tcphdr, check)))
1324                         goto out;
1325
1326                 if (recalculate_partial_csum) {
1327                         struct tcphdr *tcph = tcp_hdr(skb);
1328                         tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1329                                                          skb->len - iph->ihl*4,
1330                                                          IPPROTO_TCP, 0);
1331                 }
1332                 break;
1333         case IPPROTO_UDP:
1334                 if (!skb_partial_csum_set(skb, 4 * iph->ihl,
1335                                           offsetof(struct udphdr, check)))
1336                         goto out;
1337
1338                 if (recalculate_partial_csum) {
1339                         struct udphdr *udph = udp_hdr(skb);
1340                         udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1341                                                          skb->len - iph->ihl*4,
1342                                                          IPPROTO_UDP, 0);
1343                 }
1344                 break;
1345         default:
1346                 if (net_ratelimit())
1347                         netdev_err(vif->dev,
1348                                    "Attempting to checksum a non-TCP/UDP packet, dropping a protocol %d packet\n",
1349                                    iph->protocol);
1350                 goto out;
1351         }
1352
1353         err = 0;
1354
1355 out:
1356         return err;
1357 }
1358
1359 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1360 {
1361         unsigned long now = jiffies;
1362         unsigned long next_credit =
1363                 vif->credit_timeout.expires +
1364                 msecs_to_jiffies(vif->credit_usec / 1000);
1365
1366         /* Timer could already be pending in rare cases. */
1367         if (timer_pending(&vif->credit_timeout))
1368                 return true;
1369
1370         /* Passed the point where we can replenish credit? */
1371         if (time_after_eq(now, next_credit)) {
1372                 vif->credit_timeout.expires = now;
1373                 tx_add_credit(vif);
1374         }
1375
1376         /* Still too big to send right now? Set a callback. */
1377         if (size > vif->remaining_credit) {
1378                 vif->credit_timeout.data     =
1379                         (unsigned long)vif;
1380                 vif->credit_timeout.function =
1381                         tx_credit_callback;
1382                 mod_timer(&vif->credit_timeout,
1383                           next_credit);
1384
1385                 return true;
1386         }
1387
1388         return false;
1389 }
1390
1391 static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
1392 {
1393         struct gnttab_copy *gop = netbk->tx_copy_ops, *request_gop;
1394         struct sk_buff *skb;
1395         int ret;
1396
1397         while ((nr_pending_reqs(netbk) + XEN_NETIF_NR_SLOTS_MIN
1398                 < MAX_PENDING_REQS) &&
1399                 !list_empty(&netbk->net_schedule_list)) {
1400                 struct xenvif *vif;
1401                 struct xen_netif_tx_request txreq;
1402                 struct xen_netif_tx_request txfrags[max_skb_slots];
1403                 struct page *page;
1404                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1405                 u16 pending_idx;
1406                 RING_IDX idx;
1407                 int work_to_do;
1408                 unsigned int data_len;
1409                 pending_ring_idx_t index;
1410
1411                 /* Get a netif from the list with work to do. */
1412                 vif = poll_net_schedule_list(netbk);
1413                 /* This can sometimes happen because the test of
1414                  * list_empty(net_schedule_list) at the top of the
1415                  * loop is unlocked.  Just go back and have another
1416                  * look.
1417                  */
1418                 if (!vif)
1419                         continue;
1420
1421                 if (vif->tx.sring->req_prod - vif->tx.req_cons >
1422                     XEN_NETIF_TX_RING_SIZE) {
1423                         netdev_err(vif->dev,
1424                                    "Impossible number of requests. "
1425                                    "req_prod %d, req_cons %d, size %ld\n",
1426                                    vif->tx.sring->req_prod, vif->tx.req_cons,
1427                                    XEN_NETIF_TX_RING_SIZE);
1428                         netbk_fatal_tx_err(vif);
1429                         continue;
1430                 }
1431
1432                 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, work_to_do);
1433                 if (!work_to_do) {
1434                         xenvif_put(vif);
1435                         continue;
1436                 }
1437
1438                 idx = vif->tx.req_cons;
1439                 rmb(); /* Ensure that we see the request before we copy it. */
1440                 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1441
1442                 /* Credit-based scheduling. */
1443                 if (txreq.size > vif->remaining_credit &&
1444                     tx_credit_exceeded(vif, txreq.size)) {
1445                         xenvif_put(vif);
1446                         continue;
1447                 }
1448
1449                 vif->remaining_credit -= txreq.size;
1450
1451                 work_to_do--;
1452                 vif->tx.req_cons = ++idx;
1453
1454                 memset(extras, 0, sizeof(extras));
1455                 if (txreq.flags & XEN_NETTXF_extra_info) {
1456                         work_to_do = xen_netbk_get_extras(vif, extras,
1457                                                           work_to_do);
1458                         idx = vif->tx.req_cons;
1459                         if (unlikely(work_to_do < 0))
1460                                 continue;
1461                 }
1462
1463                 ret = netbk_count_requests(vif, &txreq, idx,
1464                                            txfrags, work_to_do);
1465                 if (unlikely(ret < 0))
1466                         continue;
1467
1468                 idx += ret;
1469
1470                 if (unlikely(txreq.size < ETH_HLEN)) {
1471                         netdev_dbg(vif->dev,
1472                                    "Bad packet size: %d\n", txreq.size);
1473                         netbk_tx_err(vif, &txreq, idx);
1474                         continue;
1475                 }
1476
1477                 /* No crossing a page as the payload mustn't fragment. */
1478                 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1479                         netdev_err(vif->dev,
1480                                    "txreq.offset: %x, size: %u, end: %lu\n",
1481                                    txreq.offset, txreq.size,
1482                                    (txreq.offset&~PAGE_MASK) + txreq.size);
1483                         netbk_fatal_tx_err(vif);
1484                         continue;
1485                 }
1486
1487                 index = pending_index(netbk->pending_cons);
1488                 pending_idx = netbk->pending_ring[index];
1489
1490                 data_len = (txreq.size > PKT_PROT_LEN &&
1491                             ret < XEN_NETIF_NR_SLOTS_MIN) ?
1492                         PKT_PROT_LEN : txreq.size;
1493
1494                 skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1495                                 GFP_ATOMIC | __GFP_NOWARN);
1496                 if (unlikely(skb == NULL)) {
1497                         netdev_dbg(vif->dev,
1498                                    "Can't allocate a skb in start_xmit.\n");
1499                         netbk_tx_err(vif, &txreq, idx);
1500                         break;
1501                 }
1502
1503                 /* Packets passed to netif_rx() must have some headroom. */
1504                 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1505
1506                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1507                         struct xen_netif_extra_info *gso;
1508                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1509
1510                         if (netbk_set_skb_gso(vif, skb, gso)) {
1511                                 /* Failure in netbk_set_skb_gso is fatal. */
1512                                 kfree_skb(skb);
1513                                 continue;
1514                         }
1515                 }
1516
1517                 /* XXX could copy straight to head */
1518                 page = xen_netbk_alloc_page(netbk, pending_idx);
1519                 if (!page) {
1520                         kfree_skb(skb);
1521                         netbk_tx_err(vif, &txreq, idx);
1522                         continue;
1523                 }
1524
1525                 gop->source.u.ref = txreq.gref;
1526                 gop->source.domid = vif->domid;
1527                 gop->source.offset = txreq.offset;
1528
1529                 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1530                 gop->dest.domid = DOMID_SELF;
1531                 gop->dest.offset = txreq.offset;
1532
1533                 gop->len = txreq.size;
1534                 gop->flags = GNTCOPY_source_gref;
1535
1536                 gop++;
1537
1538                 memcpy(&netbk->pending_tx_info[pending_idx].req,
1539                        &txreq, sizeof(txreq));
1540                 netbk->pending_tx_info[pending_idx].vif = vif;
1541                 netbk->pending_tx_info[pending_idx].head = index;
1542                 *((u16 *)skb->data) = pending_idx;
1543
1544                 __skb_put(skb, data_len);
1545
1546                 skb_shinfo(skb)->nr_frags = ret;
1547                 if (data_len < txreq.size) {
1548                         skb_shinfo(skb)->nr_frags++;
1549                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1550                                              pending_idx);
1551                 } else {
1552                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1553                                              INVALID_PENDING_IDX);
1554                 }
1555
1556                 netbk->pending_cons++;
1557
1558                 request_gop = xen_netbk_get_requests(netbk, vif,
1559                                                      skb, txfrags, gop);
1560                 if (request_gop == NULL) {
1561                         kfree_skb(skb);
1562                         netbk_tx_err(vif, &txreq, idx);
1563                         continue;
1564                 }
1565                 gop = request_gop;
1566
1567                 __skb_queue_tail(&netbk->tx_queue, skb);
1568
1569                 vif->tx.req_cons = idx;
1570                 xen_netbk_check_rx_xenvif(vif);
1571
1572                 if ((gop-netbk->tx_copy_ops) >= ARRAY_SIZE(netbk->tx_copy_ops))
1573                         break;
1574         }
1575
1576         return gop - netbk->tx_copy_ops;
1577 }
1578
1579 static void xen_netbk_tx_submit(struct xen_netbk *netbk)
1580 {
1581         struct gnttab_copy *gop = netbk->tx_copy_ops;
1582         struct sk_buff *skb;
1583
1584         while ((skb = __skb_dequeue(&netbk->tx_queue)) != NULL) {
1585                 struct xen_netif_tx_request *txp;
1586                 struct xenvif *vif;
1587                 u16 pending_idx;
1588                 unsigned data_len;
1589
1590                 pending_idx = *((u16 *)skb->data);
1591                 vif = netbk->pending_tx_info[pending_idx].vif;
1592                 txp = &netbk->pending_tx_info[pending_idx].req;
1593
1594                 /* Check the remap error code. */
1595                 if (unlikely(xen_netbk_tx_check_gop(netbk, skb, &gop))) {
1596                         netdev_dbg(vif->dev, "netback grant failed.\n");
1597                         skb_shinfo(skb)->nr_frags = 0;
1598                         kfree_skb(skb);
1599                         continue;
1600                 }
1601
1602                 data_len = skb->len;
1603                 memcpy(skb->data,
1604                        (void *)(idx_to_kaddr(netbk, pending_idx)|txp->offset),
1605                        data_len);
1606                 if (data_len < txp->size) {
1607                         /* Append the packet payload as a fragment. */
1608                         txp->offset += data_len;
1609                         txp->size -= data_len;
1610                 } else {
1611                         /* Schedule a response immediately. */
1612                         xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1613                 }
1614
1615                 if (txp->flags & XEN_NETTXF_csum_blank)
1616                         skb->ip_summed = CHECKSUM_PARTIAL;
1617                 else if (txp->flags & XEN_NETTXF_data_validated)
1618                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1619
1620                 xen_netbk_fill_frags(netbk, skb);
1621
1622                 /*
1623                  * If the initial fragment was < PKT_PROT_LEN then
1624                  * pull through some bytes from the other fragments to
1625                  * increase the linear region to PKT_PROT_LEN bytes.
1626                  */
1627                 if (skb_headlen(skb) < PKT_PROT_LEN && skb_is_nonlinear(skb)) {
1628                         int target = min_t(int, skb->len, PKT_PROT_LEN);
1629                         __pskb_pull_tail(skb, target - skb_headlen(skb));
1630                 }
1631
1632                 skb->dev      = vif->dev;
1633                 skb->protocol = eth_type_trans(skb, skb->dev);
1634                 skb_reset_network_header(skb);
1635
1636                 if (checksum_setup(vif, skb)) {
1637                         netdev_dbg(vif->dev,
1638                                    "Can't setup checksum in net_tx_action\n");
1639                         kfree_skb(skb);
1640                         continue;
1641                 }
1642
1643                 skb_probe_transport_header(skb, 0);
1644
1645                 vif->dev->stats.rx_bytes += skb->len;
1646                 vif->dev->stats.rx_packets++;
1647
1648                 xenvif_receive_skb(vif, skb);
1649         }
1650 }
1651
1652 /* Called after netfront has transmitted */
1653 static void xen_netbk_tx_action(struct xen_netbk *netbk)
1654 {
1655         unsigned nr_gops;
1656
1657         nr_gops = xen_netbk_tx_build_gops(netbk);
1658
1659         if (nr_gops == 0)
1660                 return;
1661
1662         gnttab_batch_copy(netbk->tx_copy_ops, nr_gops);
1663
1664         xen_netbk_tx_submit(netbk);
1665 }
1666
1667 static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx,
1668                                   u8 status)
1669 {
1670         struct xenvif *vif;
1671         struct pending_tx_info *pending_tx_info;
1672         pending_ring_idx_t head;
1673         u16 peek; /* peek into next tx request */
1674
1675         BUG_ON(netbk->mmap_pages[pending_idx] == (void *)(~0UL));
1676
1677         /* Already complete? */
1678         if (netbk->mmap_pages[pending_idx] == NULL)
1679                 return;
1680
1681         pending_tx_info = &netbk->pending_tx_info[pending_idx];
1682
1683         vif = pending_tx_info->vif;
1684         head = pending_tx_info->head;
1685
1686         BUG_ON(!pending_tx_is_head(netbk, head));
1687         BUG_ON(netbk->pending_ring[pending_index(head)] != pending_idx);
1688
1689         do {
1690                 pending_ring_idx_t index;
1691                 pending_ring_idx_t idx = pending_index(head);
1692                 u16 info_idx = netbk->pending_ring[idx];
1693
1694                 pending_tx_info = &netbk->pending_tx_info[info_idx];
1695                 make_tx_response(vif, &pending_tx_info->req, status);
1696
1697                 /* Setting any number other than
1698                  * INVALID_PENDING_RING_IDX indicates this slot is
1699                  * starting a new packet / ending a previous packet.
1700                  */
1701                 pending_tx_info->head = 0;
1702
1703                 index = pending_index(netbk->pending_prod++);
1704                 netbk->pending_ring[index] = netbk->pending_ring[info_idx];
1705
1706                 xenvif_put(vif);
1707
1708                 peek = netbk->pending_ring[pending_index(++head)];
1709
1710         } while (!pending_tx_is_head(netbk, peek));
1711
1712         netbk->mmap_pages[pending_idx]->mapping = 0;
1713         put_page(netbk->mmap_pages[pending_idx]);
1714         netbk->mmap_pages[pending_idx] = NULL;
1715 }
1716
1717
1718 static void make_tx_response(struct xenvif *vif,
1719                              struct xen_netif_tx_request *txp,
1720                              s8       st)
1721 {
1722         RING_IDX i = vif->tx.rsp_prod_pvt;
1723         struct xen_netif_tx_response *resp;
1724         int notify;
1725
1726         resp = RING_GET_RESPONSE(&vif->tx, i);
1727         resp->id     = txp->id;
1728         resp->status = st;
1729
1730         if (txp->flags & XEN_NETTXF_extra_info)
1731                 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1732
1733         vif->tx.rsp_prod_pvt = ++i;
1734         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1735         if (notify)
1736                 notify_remote_via_irq(vif->irq);
1737 }
1738
1739 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1740                                              u16      id,
1741                                              s8       st,
1742                                              u16      offset,
1743                                              u16      size,
1744                                              u16      flags)
1745 {
1746         RING_IDX i = vif->rx.rsp_prod_pvt;
1747         struct xen_netif_rx_response *resp;
1748
1749         resp = RING_GET_RESPONSE(&vif->rx, i);
1750         resp->offset     = offset;
1751         resp->flags      = flags;
1752         resp->id         = id;
1753         resp->status     = (s16)size;
1754         if (st < 0)
1755                 resp->status = (s16)st;
1756
1757         vif->rx.rsp_prod_pvt = ++i;
1758
1759         return resp;
1760 }
1761
1762 static inline int rx_work_todo(struct xen_netbk *netbk)
1763 {
1764         return !skb_queue_empty(&netbk->rx_queue);
1765 }
1766
1767 static inline int tx_work_todo(struct xen_netbk *netbk)
1768 {
1769
1770         if ((nr_pending_reqs(netbk) + XEN_NETIF_NR_SLOTS_MIN
1771              < MAX_PENDING_REQS) &&
1772              !list_empty(&netbk->net_schedule_list))
1773                 return 1;
1774
1775         return 0;
1776 }
1777
1778 static int xen_netbk_kthread(void *data)
1779 {
1780         struct xen_netbk *netbk = data;
1781         while (!kthread_should_stop()) {
1782                 wait_event_interruptible(netbk->wq,
1783                                 rx_work_todo(netbk) ||
1784                                 tx_work_todo(netbk) ||
1785                                 kthread_should_stop());
1786                 cond_resched();
1787
1788                 if (kthread_should_stop())
1789                         break;
1790
1791                 if (rx_work_todo(netbk))
1792                         xen_netbk_rx_action(netbk);
1793
1794                 if (tx_work_todo(netbk))
1795                         xen_netbk_tx_action(netbk);
1796         }
1797
1798         return 0;
1799 }
1800
1801 void xen_netbk_unmap_frontend_rings(struct xenvif *vif)
1802 {
1803         if (vif->tx.sring)
1804                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1805                                         vif->tx.sring);
1806         if (vif->rx.sring)
1807                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1808                                         vif->rx.sring);
1809 }
1810
1811 int xen_netbk_map_frontend_rings(struct xenvif *vif,
1812                                  grant_ref_t tx_ring_ref,
1813                                  grant_ref_t rx_ring_ref)
1814 {
1815         void *addr;
1816         struct xen_netif_tx_sring *txs;
1817         struct xen_netif_rx_sring *rxs;
1818
1819         int err = -ENOMEM;
1820
1821         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1822                                      tx_ring_ref, &addr);
1823         if (err)
1824                 goto err;
1825
1826         txs = (struct xen_netif_tx_sring *)addr;
1827         BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1828
1829         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1830                                      rx_ring_ref, &addr);
1831         if (err)
1832                 goto err;
1833
1834         rxs = (struct xen_netif_rx_sring *)addr;
1835         BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1836
1837         vif->rx_req_cons_peek = 0;
1838
1839         return 0;
1840
1841 err:
1842         xen_netbk_unmap_frontend_rings(vif);
1843         return err;
1844 }
1845
1846 static int __init netback_init(void)
1847 {
1848         int i;
1849         int rc = 0;
1850         int group;
1851
1852         if (!xen_domain())
1853                 return -ENODEV;
1854
1855         if (max_skb_slots < XEN_NETIF_NR_SLOTS_MIN) {
1856                 printk(KERN_INFO
1857                        "xen-netback: max_skb_slots too small (%d), bump it to XEN_NETIF_NR_SLOTS_MIN (%d)\n",
1858                        max_skb_slots, XEN_NETIF_NR_SLOTS_MIN);
1859                 max_skb_slots = XEN_NETIF_NR_SLOTS_MIN;
1860         }
1861
1862         xen_netbk_group_nr = num_online_cpus();
1863         xen_netbk = vzalloc(sizeof(struct xen_netbk) * xen_netbk_group_nr);
1864         if (!xen_netbk)
1865                 return -ENOMEM;
1866
1867         for (group = 0; group < xen_netbk_group_nr; group++) {
1868                 struct xen_netbk *netbk = &xen_netbk[group];
1869                 skb_queue_head_init(&netbk->rx_queue);
1870                 skb_queue_head_init(&netbk->tx_queue);
1871
1872                 init_timer(&netbk->net_timer);
1873                 netbk->net_timer.data = (unsigned long)netbk;
1874                 netbk->net_timer.function = xen_netbk_alarm;
1875
1876                 netbk->pending_cons = 0;
1877                 netbk->pending_prod = MAX_PENDING_REQS;
1878                 for (i = 0; i < MAX_PENDING_REQS; i++)
1879                         netbk->pending_ring[i] = i;
1880
1881                 init_waitqueue_head(&netbk->wq);
1882                 netbk->task = kthread_create(xen_netbk_kthread,
1883                                              (void *)netbk,
1884                                              "netback/%u", group);
1885
1886                 if (IS_ERR(netbk->task)) {
1887                         printk(KERN_ALERT "kthread_create() fails at netback\n");
1888                         del_timer(&netbk->net_timer);
1889                         rc = PTR_ERR(netbk->task);
1890                         goto failed_init;
1891                 }
1892
1893                 kthread_bind(netbk->task, group);
1894
1895                 INIT_LIST_HEAD(&netbk->net_schedule_list);
1896
1897                 spin_lock_init(&netbk->net_schedule_list_lock);
1898
1899                 atomic_set(&netbk->netfront_count, 0);
1900
1901                 wake_up_process(netbk->task);
1902         }
1903
1904         rc = xenvif_xenbus_init();
1905         if (rc)
1906                 goto failed_init;
1907
1908         return 0;
1909
1910 failed_init:
1911         while (--group >= 0) {
1912                 struct xen_netbk *netbk = &xen_netbk[group];
1913                 for (i = 0; i < MAX_PENDING_REQS; i++) {
1914                         if (netbk->mmap_pages[i])
1915                                 __free_page(netbk->mmap_pages[i]);
1916                 }
1917                 del_timer(&netbk->net_timer);
1918                 kthread_stop(netbk->task);
1919         }
1920         vfree(xen_netbk);
1921         return rc;
1922
1923 }
1924
1925 module_init(netback_init);
1926
1927 MODULE_LICENSE("Dual BSD/GPL");
1928 MODULE_ALIAS("xen-backend:vif");