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 FATAL_SKB_SLOTS_DEFAULT 20
  55 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
  56 module_param(fatal_skb_slots, uint, 0444);
  57
  58 /*
  59  * To avoid confusion, we define XEN_NETBK_LEGACY_SLOTS_MAX indicating
  60  * the maximum slots a valid packet can use. Now this value is defined
  61  * to be XEN_NETIF_NR_SLOTS_MIN, which is supposed to be supported by
  62  * all backend.
  63  */
  64 #define XEN_NETBK_LEGACY_SLOTS_MAX XEN_NETIF_NR_SLOTS_MIN
  65
  66 typedef unsigned int pending_ring_idx_t;
  67 #define INVALID_PENDING_RING_IDX (~0U)
  68
  69 struct pending_tx_info {
  70         struct xen_netif_tx_request req; /* coalesced tx request */
  71         struct xenvif *vif;
  72         pending_ring_idx_t head; /* head != INVALID_PENDING_RING_IDX
  73                                   * if it is head of one or more tx
  74                                   * reqs
  75                                   */
  76 };
  77
  78 struct netbk_rx_meta {
  79         int id;
  80         int size;
  81         int gso_size;
  82 };
  83
  84 #define MAX_PENDING_REQS 256
  85
  86 /* Discriminate from any valid pending_idx value. */
  87 #define INVALID_PENDING_IDX 0xFFFF
  88
  89 #define MAX_BUFFER_OFFSET PAGE_SIZE
  90
  91 /* extra field used in struct page */
  92 union page_ext {
  93         struct {
  94 #if BITS_PER_LONG < 64
  95 #define IDX_WIDTH   8
  96 #define GROUP_WIDTH (BITS_PER_LONG - IDX_WIDTH)
  97                 unsigned int group:GROUP_WIDTH;
  98                 unsigned int idx:IDX_WIDTH;
  99 #else
 100                 unsigned int group, idx;
 101 #endif
 102         } e;
 103         void *mapping;
 104 };
 105
 106 struct xen_netbk {
 107         wait_queue_head_t wq;
 108         struct task_struct *task;
 109
 110         struct sk_buff_head rx_queue;
 111         struct sk_buff_head tx_queue;
 112
 113         struct timer_list net_timer;
 114
 115         struct page *mmap_pages[MAX_PENDING_REQS];
 116
 117         pending_ring_idx_t pending_prod;
 118         pending_ring_idx_t pending_cons;
 119         struct list_head net_schedule_list;
 120
 121         /* Protect the net_schedule_list in netif. */
 122         spinlock_t net_schedule_list_lock;
 123
 124         atomic_t netfront_count;
 125
 126         struct pending_tx_info pending_tx_info[MAX_PENDING_REQS];
 127         /* Coalescing tx requests before copying makes number of grant
 128          * copy ops greater or equal to number of slots required. In
 129          * worst case a tx request consumes 2 gnttab_copy.
 130          */
 131         struct gnttab_copy tx_copy_ops[2*MAX_PENDING_REQS];
 132
 133         u16 pending_ring[MAX_PENDING_REQS];
 134
 135         /*
 136          * Given MAX_BUFFER_OFFSET of 4096 the worst case is that each
 137          * head/fragment page uses 2 copy operations because it
 138          * straddles two buffers in the frontend.
 139          */
 140         struct gnttab_copy grant_copy_op[2*XEN_NETIF_RX_RING_SIZE];
 141         struct netbk_rx_meta meta[2*XEN_NETIF_RX_RING_SIZE];
 142 };
 143
 144 static struct xen_netbk *xen_netbk;
 145 static int xen_netbk_group_nr;
 146
 147 /*
 148  * If head != INVALID_PENDING_RING_IDX, it means this tx request is head of
 149  * one or more merged tx requests, otherwise it is the continuation of
 150  * previous tx request.
 151  */
 152 static inline int pending_tx_is_head(struct xen_netbk *netbk, RING_IDX idx)
 153 {
 154         return netbk->pending_tx_info[idx].head != INVALID_PENDING_RING_IDX;
 155 }
 156
 157 void xen_netbk_add_xenvif(struct xenvif *vif)
 158 {
 159         int i;
 160         int min_netfront_count;
 161         int min_group = 0;
 162         struct xen_netbk *netbk;
 163
 164         min_netfront_count = atomic_read(&xen_netbk[0].netfront_count);
 165         for (i = 0; i < xen_netbk_group_nr; i++) {
 166                 int netfront_count = atomic_read(&xen_netbk[i].netfront_count);
 167                 if (netfront_count < min_netfront_count) {
 168                         min_group = i;
 169                         min_netfront_count = netfront_count;
 170                 }
 171         }
 172
 173         netbk = &xen_netbk[min_group];
 174
 175         vif->netbk = netbk;
 176         atomic_inc(&netbk->netfront_count);
 177 }
 178
 179 void xen_netbk_remove_xenvif(struct xenvif *vif)
 180 {
 181         struct xen_netbk *netbk = vif->netbk;
 182         vif->netbk = NULL;
 183         atomic_dec(&netbk->netfront_count);
 184 }
 185
 186 static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx,
 187                                   u8 status);
 188 static void make_tx_response(struct xenvif *vif,
 189                              struct xen_netif_tx_request *txp,
 190                              s8       st);
 191 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
 192                                              u16      id,
 193                                              s8       st,
 194                                              u16      offset,
 195                                              u16      size,
 196                                              u16      flags);
 197
 198 static inline unsigned long idx_to_pfn(struct xen_netbk *netbk,
 199                                        u16 idx)
 200 {
 201         return page_to_pfn(netbk->mmap_pages[idx]);
 202 }
 203
 204 static inline unsigned long idx_to_kaddr(struct xen_netbk *netbk,
 205                                          u16 idx)
 206 {
 207         return (unsigned long)pfn_to_kaddr(idx_to_pfn(netbk, idx));
 208 }
 209
 210 /* extra field used in struct page */
 211 static inline void set_page_ext(struct page *pg, struct xen_netbk *netbk,
 212                                 unsigned int idx)
 213 {
 214         unsigned int group = netbk - xen_netbk;
 215         union page_ext ext = { .e = { .group = group + 1, .idx = idx } };
 216
 217         BUILD_BUG_ON(sizeof(ext) > sizeof(ext.mapping));
 218         pg->mapping = ext.mapping;
 219 }
 220
 221 static int get_page_ext(struct page *pg,
 222                         unsigned int *pgroup, unsigned int *pidx)
 223 {
 224         union page_ext ext = { .mapping = pg->mapping };
 225         struct xen_netbk *netbk;
 226         unsigned int group, idx;
 227
 228         group = ext.e.group - 1;
 229
 230         if (group < 0 || group >= xen_netbk_group_nr)
 231                 return 0;
 232
 233         netbk = &xen_netbk[group];
 234
 235         idx = ext.e.idx;
 236
 237         if ((idx < 0) || (idx >= MAX_PENDING_REQS))
 238                 return 0;
 239
 240         if (netbk->mmap_pages[idx] != pg)
 241                 return 0;
 242
 243         *pgroup = group;
 244         *pidx = idx;
 245
 246         return 1;
 247 }
 248
 249 /*
 250  * This is the amount of packet we copy rather than map, so that the
 251  * guest can't fiddle with the contents of the headers while we do
 252  * packet processing on them (netfilter, routing, etc).
 253  */
 254 #define PKT_PROT_LEN    (ETH_HLEN + \
 255                          VLAN_HLEN + \
 256                          sizeof(struct iphdr) + MAX_IPOPTLEN + \
 257                          sizeof(struct tcphdr) + MAX_TCP_OPTION_SPACE)
 258
 259 static u16 frag_get_pending_idx(skb_frag_t *frag)
 260 {
 261         return (u16)frag->page_offset;
 262 }
 263
 264 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 265 {
 266         frag->page_offset = pending_idx;
 267 }
 268
 269 static inline pending_ring_idx_t pending_index(unsigned i)
 270 {
 271         return i & (MAX_PENDING_REQS-1);
 272 }
 273
 274 static inline pending_ring_idx_t nr_pending_reqs(struct xen_netbk *netbk)
 275 {
 276         return MAX_PENDING_REQS -
 277                 netbk->pending_prod + netbk->pending_cons;
 278 }
 279
 280 static void xen_netbk_kick_thread(struct xen_netbk *netbk)
 281 {
 282         wake_up(&netbk->wq);
 283 }
 284
 285 static int max_required_rx_slots(struct xenvif *vif)
 286 {
 287         int max = DIV_ROUND_UP(vif->dev->mtu, PAGE_SIZE);
 288
 289         /* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
 290         if (vif->can_sg || vif->gso || vif->gso_prefix)
 291                 max += MAX_SKB_FRAGS + 1; /* extra_info + frags */
 292
 293         return max;
 294 }
 295
 296 int xen_netbk_rx_ring_full(struct xenvif *vif)
 297 {
 298         RING_IDX peek   = vif->rx_req_cons_peek;
 299         RING_IDX needed = max_required_rx_slots(vif);
 300
 301         return ((vif->rx.sring->req_prod - peek) < needed) ||
 302                ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
 303 }
 304
 305 int xen_netbk_must_stop_queue(struct xenvif *vif)
 306 {
 307         if (!xen_netbk_rx_ring_full(vif))
 308                 return 0;
 309
 310         vif->rx.sring->req_event = vif->rx_req_cons_peek +
 311                 max_required_rx_slots(vif);
 312         mb(); /* request notification /then/ check the queue */
 313
 314         return xen_netbk_rx_ring_full(vif);
 315 }
 316
 317 /*
 318  * Returns true if we should start a new receive buffer instead of
 319  * adding 'size' bytes to a buffer which currently contains 'offset'
 320  * bytes.
 321  */
 322 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
 323 {
 324         /* simple case: we have completely filled the current buffer. */
 325         if (offset == MAX_BUFFER_OFFSET)
 326                 return true;
 327
 328         /*
 329          * complex case: start a fresh buffer if the current frag
 330          * would overflow the current buffer but only if:
 331          *     (i)   this frag would fit completely in the next buffer
 332          * and (ii)  there is already some data in the current buffer
 333          * and (iii) this is not the head buffer.
 334          *
 335          * Where:
 336          * - (i) stops us splitting a frag into two copies
 337          *   unless the frag is too large for a single buffer.
 338          * - (ii) stops us from leaving a buffer pointlessly empty.
 339          * - (iii) stops us leaving the first buffer
 340          *   empty. Strictly speaking this is already covered
 341          *   by (ii) but is explicitly checked because
 342          *   netfront relies on the first buffer being
 343          *   non-empty and can crash otherwise.
 344          *
 345          * This means we will effectively linearise small
 346          * frags but do not needlessly split large buffers
 347          * into multiple copies tend to give large frags their
 348          * own buffers as before.
 349          */
 350         if ((offset + size > MAX_BUFFER_OFFSET) &&
 351             (size <= MAX_BUFFER_OFFSET) && offset && !head)
 352                 return true;
 353
 354         return false;
 355 }
 356
 357 /*
 358  * Figure out how many ring slots we're going to need to send @skb to
 359  * the guest. This function is essentially a dry run of
 360  * netbk_gop_frag_copy.
 361  */
 362 unsigned int xen_netbk_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
 363 {
 364         unsigned int count;
 365         int i, copy_off;
 366
 367         count = DIV_ROUND_UP(skb_headlen(skb), PAGE_SIZE);
 368
 369         copy_off = skb_headlen(skb) % PAGE_SIZE;
 370
 371         if (skb_shinfo(skb)->gso_size)
 372                 count++;
 373
 374         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 375                 unsigned long size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
 376                 unsigned long offset = skb_shinfo(skb)->frags[i].page_offset;
 377                 unsigned long bytes;
 378
 379                 offset &= ~PAGE_MASK;
 380
 381                 while (size > 0) {
 382                         BUG_ON(offset >= PAGE_SIZE);
 383                         BUG_ON(copy_off > MAX_BUFFER_OFFSET);
 384
 385                         bytes = PAGE_SIZE - offset;
 386
 387                         if (bytes > size)
 388                                 bytes = size;
 389
 390                         if (start_new_rx_buffer(copy_off, bytes, 0)) {
 391                                 count++;
 392                                 copy_off = 0;
 393                         }
 394
 395                         if (copy_off + bytes > MAX_BUFFER_OFFSET)
 396                                 bytes = MAX_BUFFER_OFFSET - copy_off;
 397
 398                         copy_off += bytes;
 399
 400                         offset += bytes;
 401                         size -= bytes;
 402
 403                         if (offset == PAGE_SIZE)
 404                                 offset = 0;
 405                 }
 406         }
 407         return count;
 408 }
 409
 410 struct netrx_pending_operations {
 411         unsigned copy_prod, copy_cons;
 412         unsigned meta_prod, meta_cons;
 413         struct gnttab_copy *copy;
 414         struct netbk_rx_meta *meta;
 415         int copy_off;
 416         grant_ref_t copy_gref;
 417 };
 418
 419 static struct netbk_rx_meta *get_next_rx_buffer(struct xenvif *vif,
 420                                                 struct netrx_pending_operations *npo)
 421 {
 422         struct netbk_rx_meta *meta;
 423         struct xen_netif_rx_request *req;
 424
 425         req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
 426
 427         meta = npo->meta + npo->meta_prod++;
 428         meta->gso_size = 0;
 429         meta->size = 0;
 430         meta->id = req->id;
 431
 432         npo->copy_off = 0;
 433         npo->copy_gref = req->gref;
 434
 435         return meta;
 436 }
 437
 438 /*
 439  * Set up the grant operations for this fragment. If it's a flipping
 440  * interface, we also set up the unmap request from here.
 441  */
 442 static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
 443                                 struct netrx_pending_operations *npo,
 444                                 struct page *page, unsigned long size,
 445                                 unsigned long offset, int *head)
 446 {
 447         struct gnttab_copy *copy_gop;
 448         struct netbk_rx_meta *meta;
 449         /*
 450          * These variables are used iff get_page_ext returns true,
 451          * in which case they are guaranteed to be initialized.
 452          */
 453         unsigned int uninitialized_var(group), uninitialized_var(idx);
 454         int foreign = get_page_ext(page, &group, &idx);
 455         unsigned long bytes;
 456
 457         /* Data must not cross a page boundary. */
 458         BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
 459
 460         meta = npo->meta + npo->meta_prod - 1;
 461
 462         /* Skip unused frames from start of page */
 463         page += offset >> PAGE_SHIFT;
 464         offset &= ~PAGE_MASK;
 465
 466         while (size > 0) {
 467                 BUG_ON(offset >= PAGE_SIZE);
 468                 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
 469
 470                 bytes = PAGE_SIZE - offset;
 471
 472                 if (bytes > size)
 473                         bytes = size;
 474
 475                 if (start_new_rx_buffer(npo->copy_off, bytes, *head)) {
 476                         /*
 477                          * Netfront requires there to be some data in the head
 478                          * buffer.
 479                          */
 480                         BUG_ON(*head);
 481
 482                         meta = get_next_rx_buffer(vif, npo);
 483                 }
 484
 485                 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
 486                         bytes = MAX_BUFFER_OFFSET - npo->copy_off;
 487
 488                 copy_gop = npo->copy + npo->copy_prod++;
 489                 copy_gop->flags = GNTCOPY_dest_gref;
 490                 if (foreign) {
 491                         struct xen_netbk *netbk = &xen_netbk[group];
 492                         struct pending_tx_info *src_pend;
 493
 494                         src_pend = &netbk->pending_tx_info[idx];
 495
 496                         copy_gop->source.domid = src_pend->vif->domid;
 497                         copy_gop->source.u.ref = src_pend->req.gref;
 498                         copy_gop->flags |= GNTCOPY_source_gref;
 499                 } else {
 500                         void *vaddr = page_address(page);
 501                         copy_gop->source.domid = DOMID_SELF;
 502                         copy_gop->source.u.gmfn = virt_to_mfn(vaddr);
 503                 }
 504                 copy_gop->source.offset = offset;
 505                 copy_gop->dest.domid = vif->domid;
 506
 507                 copy_gop->dest.offset = npo->copy_off;
 508                 copy_gop->dest.u.ref = npo->copy_gref;
 509                 copy_gop->len = bytes;
 510
 511                 npo->copy_off += bytes;
 512                 meta->size += bytes;
 513
 514                 offset += bytes;
 515                 size -= bytes;
 516
 517                 /* Next frame */
 518                 if (offset == PAGE_SIZE && size) {
 519                         BUG_ON(!PageCompound(page));
 520                         page++;
 521                         offset = 0;
 522                 }
 523
 524                 /* Leave a gap for the GSO descriptor. */
 525                 if (*head && skb_shinfo(skb)->gso_size && !vif->gso_prefix)
 526                         vif->rx.req_cons++;
 527
 528                 *head = 0; /* There must be something in this buffer now. */
 529
 530         }
 531 }
 532
 533 /*
 534  * Prepare an SKB to be transmitted to the frontend.
 535  *
 536  * This function is responsible for allocating grant operations, meta
 537  * structures, etc.
 538  *
 539  * It returns the number of meta structures consumed. The number of
 540  * ring slots used is always equal to the number of meta slots used
 541  * plus the number of GSO descriptors used. Currently, we use either
 542  * zero GSO descriptors (for non-GSO packets) or one descriptor (for
 543  * frontend-side LRO).
 544  */
 545 static int netbk_gop_skb(struct sk_buff *skb,
 546                          struct netrx_pending_operations *npo)
 547 {
 548         struct xenvif *vif = netdev_priv(skb->dev);
 549         int nr_frags = skb_shinfo(skb)->nr_frags;
 550         int i;
 551         struct xen_netif_rx_request *req;
 552         struct netbk_rx_meta *meta;
 553         unsigned char *data;
 554         int head = 1;
 555         int old_meta_prod;
 556
 557         old_meta_prod = npo->meta_prod;
 558
 559         /* Set up a GSO prefix descriptor, if necessary */
 560         if (skb_shinfo(skb)->gso_size && vif->gso_prefix) {
 561                 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
 562                 meta = npo->meta + npo->meta_prod++;
 563                 meta->gso_size = skb_shinfo(skb)->gso_size;
 564                 meta->size = 0;
 565                 meta->id = req->id;
 566         }
 567
 568         req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
 569         meta = npo->meta + npo->meta_prod++;
 570
 571         if (!vif->gso_prefix)
 572                 meta->gso_size = skb_shinfo(skb)->gso_size;
 573         else
 574                 meta->gso_size = 0;
 575
 576         meta->size = 0;
 577         meta->id = req->id;
 578         npo->copy_off = 0;
 579         npo->copy_gref = req->gref;
 580
 581         data = skb->data;
 582         while (data < skb_tail_pointer(skb)) {
 583                 unsigned int offset = offset_in_page(data);
 584                 unsigned int len = PAGE_SIZE - offset;
 585
 586                 if (data + len > skb_tail_pointer(skb))
 587                         len = skb_tail_pointer(skb) - data;
 588
 589                 netbk_gop_frag_copy(vif, skb, npo,
 590                                     virt_to_page(data), len, offset, &head);
 591                 data += len;
 592         }
 593
 594         for (i = 0; i < nr_frags; i++) {
 595                 netbk_gop_frag_copy(vif, skb, npo,
 596                                     skb_frag_page(&skb_shinfo(skb)->frags[i]),
 597                                     skb_frag_size(&skb_shinfo(skb)->frags[i]),
 598                                     skb_shinfo(skb)->frags[i].page_offset,
 599                                     &head);
 600         }
 601
 602         return npo->meta_prod - old_meta_prod;
 603 }
 604
 605 /*
 606  * This is a twin to netbk_gop_skb.  Assume that netbk_gop_skb was
 607  * used to set up the operations on the top of
 608  * netrx_pending_operations, which have since been done.  Check that
 609  * they didn't give any errors and advance over them.
 610  */
 611 static int netbk_check_gop(struct xenvif *vif, int nr_meta_slots,
 612                            struct netrx_pending_operations *npo)
 613 {
 614         struct gnttab_copy     *copy_op;
 615         int status = XEN_NETIF_RSP_OKAY;
 616         int i;
 617
 618         for (i = 0; i < nr_meta_slots; i++) {
 619                 copy_op = npo->copy + npo->copy_cons++;
 620                 if (copy_op->status != GNTST_okay) {
 621                         netdev_dbg(vif->dev,
 622                                    "Bad status %d from copy to DOM%d.\n",
 623                                    copy_op->status, vif->domid);
 624                         status = XEN_NETIF_RSP_ERROR;
 625                 }
 626         }
 627
 628         return status;
 629 }
 630
 631 static void netbk_add_frag_responses(struct xenvif *vif, int status,
 632                                      struct netbk_rx_meta *meta,
 633                                      int nr_meta_slots)
 634 {
 635         int i;
 636         unsigned long offset;
 637
 638         /* No fragments used */
 639         if (nr_meta_slots <= 1)
 640                 return;
 641
 642         nr_meta_slots--;
 643
 644         for (i = 0; i < nr_meta_slots; i++) {
 645                 int flags;
 646                 if (i == nr_meta_slots - 1)
 647                         flags = 0;
 648                 else
 649                         flags = XEN_NETRXF_more_data;
 650
 651                 offset = 0;
 652                 make_rx_response(vif, meta[i].id, status, offset,
 653                                  meta[i].size, flags);
 654         }
 655 }
 656
 657 struct skb_cb_overlay {
 658         int meta_slots_used;
 659 };
 660
 661 static void xen_netbk_rx_action(struct xen_netbk *netbk)
 662 {
 663         struct xenvif *vif = NULL, *tmp;
 664         s8 status;
 665         u16 irq, flags;
 666         struct xen_netif_rx_response *resp;
 667         struct sk_buff_head rxq;
 668         struct sk_buff *skb;
 669         LIST_HEAD(notify);
 670         int ret;
 671         int nr_frags;
 672         int count;
 673         unsigned long offset;
 674         struct skb_cb_overlay *sco;
 675
 676         struct netrx_pending_operations npo = {
 677                 .copy  = netbk->grant_copy_op,
 678                 .meta  = netbk->meta,
 679         };
 680
 681         skb_queue_head_init(&rxq);
 682
 683         count = 0;
 684
 685         while ((skb = skb_dequeue(&netbk->rx_queue)) != NULL) {
 686                 vif = netdev_priv(skb->dev);
 687                 nr_frags = skb_shinfo(skb)->nr_frags;
 688
 689                 sco = (struct skb_cb_overlay *)skb->cb;
 690                 sco->meta_slots_used = netbk_gop_skb(skb, &npo);
 691
 692                 count += nr_frags + 1;
 693
 694                 __skb_queue_tail(&rxq, skb);
 695
 696                 /* Filled the batch queue? */
 697                 /* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
 698                 if (count + MAX_SKB_FRAGS >= XEN_NETIF_RX_RING_SIZE)
 699                         break;
 700         }
 701
 702         BUG_ON(npo.meta_prod > ARRAY_SIZE(netbk->meta));
 703
 704         if (!npo.copy_prod)
 705                 return;
 706
 707         BUG_ON(npo.copy_prod > ARRAY_SIZE(netbk->grant_copy_op));
 708         gnttab_batch_copy(netbk->grant_copy_op, npo.copy_prod);
 709
 710         while ((skb = __skb_dequeue(&rxq)) != NULL) {
 711                 sco = (struct skb_cb_overlay *)skb->cb;
 712
 713                 vif = netdev_priv(skb->dev);
 714
 715                 if (netbk->meta[npo.meta_cons].gso_size && vif->gso_prefix) {
 716                         resp = RING_GET_RESPONSE(&vif->rx,
 717                                                 vif->rx.rsp_prod_pvt++);
 718
 719                         resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
 720
 721                         resp->offset = netbk->meta[npo.meta_cons].gso_size;
 722                         resp->id = netbk->meta[npo.meta_cons].id;
 723                         resp->status = sco->meta_slots_used;
 724
 725                         npo.meta_cons++;
 726                         sco->meta_slots_used--;
 727                 }
 728
 729
 730                 vif->dev->stats.tx_bytes += skb->len;
 731                 vif->dev->stats.tx_packets++;
 732
 733                 status = netbk_check_gop(vif, sco->meta_slots_used, &npo);
 734
 735                 if (sco->meta_slots_used == 1)
 736                         flags = 0;
 737                 else
 738                         flags = XEN_NETRXF_more_data;
 739
 740                 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
 741                         flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
 742                 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 743                         /* remote but checksummed. */
 744                         flags |= XEN_NETRXF_data_validated;
 745
 746                 offset = 0;
 747                 resp = make_rx_response(vif, netbk->meta[npo.meta_cons].id,
 748                                         status, offset,
 749                                         netbk->meta[npo.meta_cons].size,
 750                                         flags);
 751
 752                 if (netbk->meta[npo.meta_cons].gso_size && !vif->gso_prefix) {
 753                         struct xen_netif_extra_info *gso =
 754                                 (struct xen_netif_extra_info *)
 755                                 RING_GET_RESPONSE(&vif->rx,
 756                                                   vif->rx.rsp_prod_pvt++);
 757
 758                         resp->flags |= XEN_NETRXF_extra_info;
 759
 760                         gso->u.gso.size = netbk->meta[npo.meta_cons].gso_size;
 761                         gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
 762                         gso->u.gso.pad = 0;
 763                         gso->u.gso.features = 0;
 764
 765                         gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
 766                         gso->flags = 0;
 767                 }
 768
 769                 netbk_add_frag_responses(vif, status,
 770                                          netbk->meta + npo.meta_cons + 1,
 771                                          sco->meta_slots_used);
 772
 773                 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
 774                 irq = vif->irq;
 775                 if (ret && list_empty(&vif->notify_list))
 776                         list_add_tail(&vif->notify_list, &notify);
 777
 778                 xenvif_notify_tx_completion(vif);
 779
 780                 xenvif_put(vif);
 781                 npo.meta_cons += sco->meta_slots_used;
 782                 dev_kfree_skb(skb);
 783         }
 784
 785         list_for_each_entry_safe(vif, tmp, &notify, notify_list) {
 786                 notify_remote_via_irq(vif->irq);
 787                 list_del_init(&vif->notify_list);
 788         }
 789
 790         /* More work to do? */
 791         if (!skb_queue_empty(&netbk->rx_queue) &&
 792                         !timer_pending(&netbk->net_timer))
 793                 xen_netbk_kick_thread(netbk);
 794 }
 795
 796 void xen_netbk_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb)
 797 {
 798         struct xen_netbk *netbk = vif->netbk;
 799
 800         skb_queue_tail(&netbk->rx_queue, skb);
 801
 802         xen_netbk_kick_thread(netbk);
 803 }
 804
 805 static void xen_netbk_alarm(unsigned long data)
 806 {
 807         struct xen_netbk *netbk = (struct xen_netbk *)data;
 808         xen_netbk_kick_thread(netbk);
 809 }
 810
 811 static int __on_net_schedule_list(struct xenvif *vif)
 812 {
 813         return !list_empty(&vif->schedule_list);
 814 }
 815
 816 /* Must be called with net_schedule_list_lock held */
 817 static void remove_from_net_schedule_list(struct xenvif *vif)
 818 {
 819         if (likely(__on_net_schedule_list(vif))) {
 820                 list_del_init(&vif->schedule_list);
 821                 xenvif_put(vif);
 822         }
 823 }
 824
 825 static struct xenvif *poll_net_schedule_list(struct xen_netbk *netbk)
 826 {
 827         struct xenvif *vif = NULL;
 828
 829         spin_lock_irq(&netbk->net_schedule_list_lock);
 830         if (list_empty(&netbk->net_schedule_list))
 831                 goto out;
 832
 833         vif = list_first_entry(&netbk->net_schedule_list,
 834                                struct xenvif, schedule_list);
 835         if (!vif)
 836                 goto out;
 837
 838         xenvif_get(vif);
 839
 840         remove_from_net_schedule_list(vif);
 841 out:
 842         spin_unlock_irq(&netbk->net_schedule_list_lock);
 843         return vif;
 844 }
 845
 846 void xen_netbk_schedule_xenvif(struct xenvif *vif)
 847 {
 848         unsigned long flags;
 849         struct xen_netbk *netbk = vif->netbk;
 850
 851         if (__on_net_schedule_list(vif))
 852                 goto kick;
 853
 854         spin_lock_irqsave(&netbk->net_schedule_list_lock, flags);
 855         if (!__on_net_schedule_list(vif) &&
 856             likely(xenvif_schedulable(vif))) {
 857                 list_add_tail(&vif->schedule_list, &netbk->net_schedule_list);
 858                 xenvif_get(vif);
 859         }
 860         spin_unlock_irqrestore(&netbk->net_schedule_list_lock, flags);
 861
 862 kick:
 863         smp_mb();
 864         if ((nr_pending_reqs(netbk) < (MAX_PENDING_REQS/2)) &&
 865             !list_empty(&netbk->net_schedule_list))
 866                 xen_netbk_kick_thread(netbk);
 867 }
 868
 869 void xen_netbk_deschedule_xenvif(struct xenvif *vif)
 870 {
 871         struct xen_netbk *netbk = vif->netbk;
 872         spin_lock_irq(&netbk->net_schedule_list_lock);
 873         remove_from_net_schedule_list(vif);
 874         spin_unlock_irq(&netbk->net_schedule_list_lock);
 875 }
 876
 877 void xen_netbk_check_rx_xenvif(struct xenvif *vif)
 878 {
 879         int more_to_do;
 880
 881         RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
 882
 883         if (more_to_do)
 884                 xen_netbk_schedule_xenvif(vif);
 885 }
 886
 887 static void tx_add_credit(struct xenvif *vif)
 888 {
 889         unsigned long max_burst, max_credit;
 890
 891         /*
 892          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 893          * Otherwise the interface can seize up due to insufficient credit.
 894          */
 895         max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
 896         max_burst = min(max_burst, 131072UL);
 897         max_burst = max(max_burst, vif->credit_bytes);
 898
 899         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 900         max_credit = vif->remaining_credit + vif->credit_bytes;
 901         if (max_credit < vif->remaining_credit)
 902                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 903
 904         vif->remaining_credit = min(max_credit, max_burst);
 905 }
 906
 907 static void tx_credit_callback(unsigned long data)
 908 {
 909         struct xenvif *vif = (struct xenvif *)data;
 910         tx_add_credit(vif);
 911         xen_netbk_check_rx_xenvif(vif);
 912 }
 913
 914 static void netbk_tx_err(struct xenvif *vif,
 915                          struct xen_netif_tx_request *txp, RING_IDX end)
 916 {
 917         RING_IDX cons = vif->tx.req_cons;
 918
 919         do {
 920                 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
 921                 if (cons == end)
 922                         break;
 923                 txp = RING_GET_REQUEST(&vif->tx, cons++);
 924         } while (1);
 925         vif->tx.req_cons = cons;
 926         xen_netbk_check_rx_xenvif(vif);
 927         xenvif_put(vif);
 928 }
 929
 930 static void netbk_fatal_tx_err(struct xenvif *vif)
 931 {
 932         netdev_err(vif->dev, "fatal error; disabling device\n");
 933         xenvif_carrier_off(vif);
 934         xenvif_put(vif);
 935 }
 936
 937 static int netbk_count_requests(struct xenvif *vif,
 938                                 struct xen_netif_tx_request *first,
 939                                 struct xen_netif_tx_request *txp,
 940                                 int work_to_do)
 941 {
 942         RING_IDX cons = vif->tx.req_cons;
 943         int slots = 0;
 944         int drop_err = 0;
 945         int more_data;
 946
 947         if (!(first->flags & XEN_NETTXF_more_data))
 948                 return 0;
 949
 950         do {
 951                 struct xen_netif_tx_request dropped_tx = { 0 };
 952
 953                 if (slots >= work_to_do) {
 954                         netdev_err(vif->dev,
 955                                    "Asked for %d slots but exceeds this limit\n",
 956                                    work_to_do);
 957                         netbk_fatal_tx_err(vif);
 958                         return -ENODATA;
 959                 }
 960
 961                 /* This guest is really using too many slots and
 962                  * considered malicious.
 963                  */
 964                 if (unlikely(slots >= fatal_skb_slots)) {
 965                         netdev_err(vif->dev,
 966                                    "Malicious frontend using %d slots, threshold %u\n",
 967                                    slots, fatal_skb_slots);
 968                         netbk_fatal_tx_err(vif);
 969                         return -E2BIG;
 970                 }
 971
 972                 /* Xen network protocol had implicit dependency on
 973                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 974                  * the historical MAX_SKB_FRAGS value 18 to honor the
 975                  * same behavior as before. Any packet using more than
 976                  * 18 slots but less than fatal_skb_slots slots is
 977                  * dropped
 978                  */
 979                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 980                         if (net_ratelimit())
 981                                 netdev_dbg(vif->dev,
 982                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 983                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 984                         drop_err = -E2BIG;
 985                 }
 986
 987                 if (drop_err)
 988                         txp = &dropped_tx;
 989
 990                 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + slots),
 991                        sizeof(*txp));
 992
 993                 /* If the guest submitted a frame >= 64 KiB then
 994                  * first->size overflowed and following slots will
 995                  * appear to be larger than the frame.
 996                  *
 997                  * This cannot be fatal error as there are buggy
 998                  * frontends that do this.
 999                  *
1000                  * Consume all slots and drop the packet.
1001                  */
1002                 if (!drop_err && txp->size > first->size) {
1003                         if (net_ratelimit())
1004                                 netdev_dbg(vif->dev,
1005                                            "Invalid tx request, slot size %u > remaining size %u\n",
1006                                            txp->size, first->size);
1007                         drop_err = -EIO;
1008                 }
1009
1010                 first->size -= txp->size;
1011                 slots++;
1012
1013                 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
1014                         netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
1015                                  txp->offset, txp->size);
1016                         netbk_fatal_tx_err(vif);
1017                         return -EINVAL;
1018                 }
1019
1020                 more_data = txp->flags & XEN_NETTXF_more_data;
1021
1022                 if (!drop_err)
1023                         txp++;
1024
1025         } while (more_data);
1026
1027         if (drop_err) {
1028                 netbk_tx_err(vif, first, cons + slots);
1029                 return drop_err;
1030         }
1031
1032         return slots;
1033 }
1034
1035 static struct page *xen_netbk_alloc_page(struct xen_netbk *netbk,
1036                                          u16 pending_idx)
1037 {
1038         struct page *page;
1039         page = alloc_page(GFP_KERNEL|__GFP_COLD);
1040         if (!page)
1041                 return NULL;
1042         set_page_ext(page, netbk, pending_idx);
1043         netbk->mmap_pages[pending_idx] = page;
1044         return page;
1045 }
1046
1047 static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
1048                                                   struct xenvif *vif,
1049                                                   struct sk_buff *skb,
1050                                                   struct xen_netif_tx_request *txp,
1051                                                   struct gnttab_copy *gop)
1052 {
1053         struct skb_shared_info *shinfo = skb_shinfo(skb);
1054         skb_frag_t *frags = shinfo->frags;
1055         u16 pending_idx = *((u16 *)skb->data);
1056         u16 head_idx = 0;
1057         int slot, start;
1058         struct page *page;
1059         pending_ring_idx_t index, start_idx = 0;
1060         uint16_t dst_offset;
1061         unsigned int nr_slots;
1062         struct pending_tx_info *first = NULL;
1063
1064         /* At this point shinfo->nr_frags is in fact the number of
1065          * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1066          */
1067         nr_slots = shinfo->nr_frags;
1068
1069         /* Skip first skb fragment if it is on same page as header fragment. */
1070         start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
1071
1072         /* Coalesce tx requests, at this point the packet passed in
1073          * should be <= 64K. Any packets larger than 64K have been
1074          * handled in netbk_count_requests().
1075          */
1076         for (shinfo->nr_frags = slot = start; slot < nr_slots;
1077              shinfo->nr_frags++) {
1078                 struct pending_tx_info *pending_tx_info =
1079                         netbk->pending_tx_info;
1080
1081                 page = alloc_page(GFP_KERNEL|__GFP_COLD);
1082                 if (!page)
1083                         goto err;
1084
1085                 dst_offset = 0;
1086                 first = NULL;
1087                 while (dst_offset < PAGE_SIZE && slot < nr_slots) {
1088                         gop->flags = GNTCOPY_source_gref;
1089
1090                         gop->source.u.ref = txp->gref;
1091                         gop->source.domid = vif->domid;
1092                         gop->source.offset = txp->offset;
1093
1094                         gop->dest.domid = DOMID_SELF;
1095
1096                         gop->dest.offset = dst_offset;
1097                         gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1098
1099                         if (dst_offset + txp->size > PAGE_SIZE) {
1100                                 /* This page can only merge a portion
1101                                  * of tx request. Do not increment any
1102                                  * pointer / counter here. The txp
1103                                  * will be dealt with in future
1104                                  * rounds, eventually hitting the
1105                                  * `else` branch.
1106                                  */
1107                                 gop->len = PAGE_SIZE - dst_offset;
1108                                 txp->offset += gop->len;
1109                                 txp->size -= gop->len;
1110                                 dst_offset += gop->len; /* quit loop */
1111                         } else {
1112                                 /* This tx request can be merged in the page */
1113                                 gop->len = txp->size;
1114                                 dst_offset += gop->len;
1115
1116                                 index = pending_index(netbk->pending_cons++);
1117
1118                                 pending_idx = netbk->pending_ring[index];
1119
1120                                 memcpy(&pending_tx_info[pending_idx].req, txp,
1121                                        sizeof(*txp));
1122                                 xenvif_get(vif);
1123
1124                                 pending_tx_info[pending_idx].vif = vif;
1125
1126                                 /* Poison these fields, corresponding
1127                                  * fields for head tx req will be set
1128                                  * to correct values after the loop.
1129                                  */
1130                                 netbk->mmap_pages[pending_idx] = (void *)(~0UL);
1131                                 pending_tx_info[pending_idx].head =
1132                                         INVALID_PENDING_RING_IDX;
1133
1134                                 if (!first) {
1135                                         first = &pending_tx_info[pending_idx];
1136                                         start_idx = index;
1137                                         head_idx = pending_idx;
1138                                 }
1139
1140                                 txp++;
1141                                 slot++;
1142                         }
1143
1144                         gop++;
1145                 }
1146
1147                 first->req.offset = 0;
1148                 first->req.size = dst_offset;
1149                 first->head = start_idx;
1150                 set_page_ext(page, netbk, head_idx);
1151                 netbk->mmap_pages[head_idx] = page;
1152                 frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
1153         }
1154
1155         BUG_ON(shinfo->nr_frags > MAX_SKB_FRAGS);
1156
1157         return gop;
1158 err:
1159         /* Unwind, freeing all pages and sending error responses. */
1160         while (shinfo->nr_frags-- > start) {
1161                 xen_netbk_idx_release(netbk,
1162                                 frag_get_pending_idx(&frags[shinfo->nr_frags]),
1163                                 XEN_NETIF_RSP_ERROR);
1164         }
1165         /* The head too, if necessary. */
1166         if (start)
1167                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
1168
1169         return NULL;
1170 }
1171
1172 static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
1173                                   struct sk_buff *skb,
1174                                   struct gnttab_copy **gopp)
1175 {
1176         struct gnttab_copy *gop = *gopp;
1177         u16 pending_idx = *((u16 *)skb->data);
1178         struct skb_shared_info *shinfo = skb_shinfo(skb);
1179         struct pending_tx_info *tx_info;
1180         int nr_frags = shinfo->nr_frags;
1181         int i, err, start;
1182         u16 peek; /* peek into next tx request */
1183
1184         /* Check status of header. */
1185         err = gop->status;
1186         if (unlikely(err))
1187                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
1188
1189         /* Skip first skb fragment if it is on same page as header fragment. */
1190         start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
1191
1192         for (i = start; i < nr_frags; i++) {
1193                 int j, newerr;
1194                 pending_ring_idx_t head;
1195
1196                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
1197                 tx_info = &netbk->pending_tx_info[pending_idx];
1198                 head = tx_info->head;
1199
1200                 /* Check error status: if okay then remember grant handle. */
1201                 do {
1202                         newerr = (++gop)->status;
1203                         if (newerr)
1204                                 break;
1205                         peek = netbk->pending_ring[pending_index(++head)];
1206                 } while (!pending_tx_is_head(netbk, peek));
1207
1208                 if (likely(!newerr)) {
1209                         /* Had a previous error? Invalidate this fragment. */
1210                         if (unlikely(err))
1211                                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1212                         continue;
1213                 }
1214
1215                 /* Error on this fragment: respond to client with an error. */
1216                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
1217
1218                 /* Not the first error? Preceding frags already invalidated. */
1219                 if (err)
1220                         continue;
1221
1222                 /* First error: invalidate header and preceding fragments. */
1223                 pending_idx = *((u16 *)skb->data);
1224                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1225                 for (j = start; j < i; j++) {
1226                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1227                         xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1228                 }
1229
1230                 /* Remember the error: invalidate all subsequent fragments. */
1231                 err = newerr;
1232         }
1233
1234         *gopp = gop + 1;
1235         return err;
1236 }
1237
1238 static void xen_netbk_fill_frags(struct xen_netbk *netbk, struct sk_buff *skb)
1239 {
1240         struct skb_shared_info *shinfo = skb_shinfo(skb);
1241         int nr_frags = shinfo->nr_frags;
1242         int i;
1243
1244         for (i = 0; i < nr_frags; i++) {
1245                 skb_frag_t *frag = shinfo->frags + i;
1246                 struct xen_netif_tx_request *txp;
1247                 struct page *page;
1248                 u16 pending_idx;
1249
1250                 pending_idx = frag_get_pending_idx(frag);
1251
1252                 txp = &netbk->pending_tx_info[pending_idx].req;
1253                 page = virt_to_page(idx_to_kaddr(netbk, pending_idx));
1254                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1255                 skb->len += txp->size;
1256                 skb->data_len += txp->size;
1257                 skb->truesize += txp->size;
1258
1259                 /* Take an extra reference to offset xen_netbk_idx_release */
1260                 get_page(netbk->mmap_pages[pending_idx]);
1261                 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1262         }
1263 }
1264
1265 static int xen_netbk_get_extras(struct xenvif *vif,
1266                                 struct xen_netif_extra_info *extras,
1267                                 int work_to_do)
1268 {
1269         struct xen_netif_extra_info extra;
1270         RING_IDX cons = vif->tx.req_cons;
1271
1272         do {
1273                 if (unlikely(work_to_do-- <= 0)) {
1274                         netdev_err(vif->dev, "Missing extra info\n");
1275                         netbk_fatal_tx_err(vif);
1276                         return -EBADR;
1277                 }
1278
1279                 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1280                        sizeof(extra));
1281                 if (unlikely(!extra.type ||
1282                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1283                         vif->tx.req_cons = ++cons;
1284                         netdev_err(vif->dev,
1285                                    "Invalid extra type: %d\n", extra.type);
1286                         netbk_fatal_tx_err(vif);
1287                         return -EINVAL;
1288                 }
1289
1290                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1291                 vif->tx.req_cons = ++cons;
1292         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1293
1294         return work_to_do;
1295 }
1296
1297 static int netbk_set_skb_gso(struct xenvif *vif,
1298                              struct sk_buff *skb,
1299                              struct xen_netif_extra_info *gso)
1300 {
1301         if (!gso->u.gso.size) {
1302                 netdev_err(vif->dev, "GSO size must not be zero.\n");
1303                 netbk_fatal_tx_err(vif);
1304                 return -EINVAL;
1305         }
1306
1307         /* Currently only TCPv4 S.O. is supported. */
1308         if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
1309                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1310                 netbk_fatal_tx_err(vif);
1311                 return -EINVAL;
1312         }
1313
1314         skb_shinfo(skb)->gso_size = gso->u.gso.size;
1315         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1316
1317         /* Header must be checked, and gso_segs computed. */
1318         skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1319         skb_shinfo(skb)->gso_segs = 0;
1320
1321         return 0;
1322 }
1323
1324 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1325 {
1326         struct iphdr *iph;
1327         int err = -EPROTO;
1328         int recalculate_partial_csum = 0;
1329
1330         /*
1331          * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1332          * peers can fail to set NETRXF_csum_blank when sending a GSO
1333          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1334          * recalculate the partial checksum.
1335          */
1336         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1337                 vif->rx_gso_checksum_fixup++;
1338                 skb->ip_summed = CHECKSUM_PARTIAL;
1339                 recalculate_partial_csum = 1;
1340         }
1341
1342         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1343         if (skb->ip_summed != CHECKSUM_PARTIAL)
1344                 return 0;
1345
1346         if (skb->protocol != htons(ETH_P_IP))
1347                 goto out;
1348
1349         iph = (void *)skb->data;
1350         switch (iph->protocol) {
1351         case IPPROTO_TCP:
1352                 if (!skb_partial_csum_set(skb, 4 * iph->ihl,
1353                                           offsetof(struct tcphdr, check)))
1354                         goto out;
1355
1356                 if (recalculate_partial_csum) {
1357                         struct tcphdr *tcph = tcp_hdr(skb);
1358                         tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1359                                                          skb->len - iph->ihl*4,
1360                                                          IPPROTO_TCP, 0);
1361                 }
1362                 break;
1363         case IPPROTO_UDP:
1364                 if (!skb_partial_csum_set(skb, 4 * iph->ihl,
1365                                           offsetof(struct udphdr, check)))
1366                         goto out;
1367
1368                 if (recalculate_partial_csum) {
1369                         struct udphdr *udph = udp_hdr(skb);
1370                         udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1371                                                          skb->len - iph->ihl*4,
1372                                                          IPPROTO_UDP, 0);
1373                 }
1374                 break;
1375         default:
1376                 if (net_ratelimit())
1377                         netdev_err(vif->dev,
1378                                    "Attempting to checksum a non-TCP/UDP packet, dropping a protocol %d packet\n",
1379                                    iph->protocol);
1380                 goto out;
1381         }
1382
1383         err = 0;
1384
1385 out:
1386         return err;
1387 }
1388
1389 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1390 {
1391         unsigned long now = jiffies;
1392         unsigned long next_credit =
1393                 vif->credit_timeout.expires +
1394                 msecs_to_jiffies(vif->credit_usec / 1000);
1395
1396         /* Timer could already be pending in rare cases. */
1397         if (timer_pending(&vif->credit_timeout))
1398                 return true;
1399
1400         /* Passed the point where we can replenish credit? */
1401         if (time_after_eq(now, next_credit)) {
1402                 vif->credit_timeout.expires = now;
1403                 tx_add_credit(vif);
1404         }
1405
1406         /* Still too big to send right now? Set a callback. */
1407         if (size > vif->remaining_credit) {
1408                 vif->credit_timeout.data     =
1409                         (unsigned long)vif;
1410                 vif->credit_timeout.function =
1411                         tx_credit_callback;
1412                 mod_timer(&vif->credit_timeout,
1413                           next_credit);
1414
1415                 return true;
1416         }
1417
1418         return false;
1419 }
1420
1421 static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
1422 {
1423         struct gnttab_copy *gop = netbk->tx_copy_ops, *request_gop;
1424         struct sk_buff *skb;
1425         int ret;
1426
1427         while ((nr_pending_reqs(netbk) + XEN_NETBK_LEGACY_SLOTS_MAX
1428                 < MAX_PENDING_REQS) &&
1429                 !list_empty(&netbk->net_schedule_list)) {
1430                 struct xenvif *vif;
1431                 struct xen_netif_tx_request txreq;
1432                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1433                 struct page *page;
1434                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1435                 u16 pending_idx;
1436                 RING_IDX idx;
1437                 int work_to_do;
1438                 unsigned int data_len;
1439                 pending_ring_idx_t index;
1440
1441                 /* Get a netif from the list with work to do. */
1442                 vif = poll_net_schedule_list(netbk);
1443                 /* This can sometimes happen because the test of
1444                  * list_empty(net_schedule_list) at the top of the
1445                  * loop is unlocked.  Just go back and have another
1446                  * look.
1447                  */
1448                 if (!vif)
1449                         continue;
1450
1451                 if (vif->tx.sring->req_prod - vif->tx.req_cons >
1452                     XEN_NETIF_TX_RING_SIZE) {
1453                         netdev_err(vif->dev,
1454                                    "Impossible number of requests. "
1455                                    "req_prod %d, req_cons %d, size %ld\n",
1456                                    vif->tx.sring->req_prod, vif->tx.req_cons,
1457                                    XEN_NETIF_TX_RING_SIZE);
1458                         netbk_fatal_tx_err(vif);
1459                         continue;
1460                 }
1461
1462                 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, work_to_do);
1463                 if (!work_to_do) {
1464                         xenvif_put(vif);
1465                         continue;
1466                 }
1467
1468                 idx = vif->tx.req_cons;
1469                 rmb(); /* Ensure that we see the request before we copy it. */
1470                 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1471
1472                 /* Credit-based scheduling. */
1473                 if (txreq.size > vif->remaining_credit &&
1474                     tx_credit_exceeded(vif, txreq.size)) {
1475                         xenvif_put(vif);
1476                         continue;
1477                 }
1478
1479                 vif->remaining_credit -= txreq.size;
1480
1481                 work_to_do--;
1482                 vif->tx.req_cons = ++idx;
1483
1484                 memset(extras, 0, sizeof(extras));
1485                 if (txreq.flags & XEN_NETTXF_extra_info) {
1486                         work_to_do = xen_netbk_get_extras(vif, extras,
1487                                                           work_to_do);
1488                         idx = vif->tx.req_cons;
1489                         if (unlikely(work_to_do < 0))
1490                                 continue;
1491                 }
1492
1493                 ret = netbk_count_requests(vif, &txreq, txfrags, work_to_do);
1494                 if (unlikely(ret < 0))
1495                         continue;
1496
1497                 idx += ret;
1498
1499                 if (unlikely(txreq.size < ETH_HLEN)) {
1500                         netdev_dbg(vif->dev,
1501                                    "Bad packet size: %d\n", txreq.size);
1502                         netbk_tx_err(vif, &txreq, idx);
1503                         continue;
1504                 }
1505
1506                 /* No crossing a page as the payload mustn't fragment. */
1507                 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1508                         netdev_err(vif->dev,
1509                                    "txreq.offset: %x, size: %u, end: %lu\n",
1510                                    txreq.offset, txreq.size,
1511                                    (txreq.offset&~PAGE_MASK) + txreq.size);
1512                         netbk_fatal_tx_err(vif);
1513                         continue;
1514                 }
1515
1516                 index = pending_index(netbk->pending_cons);
1517                 pending_idx = netbk->pending_ring[index];
1518
1519                 data_len = (txreq.size > PKT_PROT_LEN &&
1520                             ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1521                         PKT_PROT_LEN : txreq.size;
1522
1523                 skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1524                                 GFP_ATOMIC | __GFP_NOWARN);
1525                 if (unlikely(skb == NULL)) {
1526                         netdev_dbg(vif->dev,
1527                                    "Can't allocate a skb in start_xmit.\n");
1528                         netbk_tx_err(vif, &txreq, idx);
1529                         break;
1530                 }
1531
1532                 /* Packets passed to netif_rx() must have some headroom. */
1533                 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1534
1535                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1536                         struct xen_netif_extra_info *gso;
1537                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1538
1539                         if (netbk_set_skb_gso(vif, skb, gso)) {
1540                                 /* Failure in netbk_set_skb_gso is fatal. */
1541                                 kfree_skb(skb);
1542                                 continue;
1543                         }
1544                 }
1545
1546                 /* XXX could copy straight to head */
1547                 page = xen_netbk_alloc_page(netbk, pending_idx);
1548                 if (!page) {
1549                         kfree_skb(skb);
1550                         netbk_tx_err(vif, &txreq, idx);
1551                         continue;
1552                 }
1553
1554                 gop->source.u.ref = txreq.gref;
1555                 gop->source.domid = vif->domid;
1556                 gop->source.offset = txreq.offset;
1557
1558                 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1559                 gop->dest.domid = DOMID_SELF;
1560                 gop->dest.offset = txreq.offset;
1561
1562                 gop->len = txreq.size;
1563                 gop->flags = GNTCOPY_source_gref;
1564
1565                 gop++;
1566
1567                 memcpy(&netbk->pending_tx_info[pending_idx].req,
1568                        &txreq, sizeof(txreq));
1569                 netbk->pending_tx_info[pending_idx].vif = vif;
1570                 netbk->pending_tx_info[pending_idx].head = index;
1571                 *((u16 *)skb->data) = pending_idx;
1572
1573                 __skb_put(skb, data_len);
1574
1575                 skb_shinfo(skb)->nr_frags = ret;
1576                 if (data_len < txreq.size) {
1577                         skb_shinfo(skb)->nr_frags++;
1578                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1579                                              pending_idx);
1580                 } else {
1581                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1582                                              INVALID_PENDING_IDX);
1583                 }
1584
1585                 netbk->pending_cons++;
1586
1587                 request_gop = xen_netbk_get_requests(netbk, vif,
1588                                                      skb, txfrags, gop);
1589                 if (request_gop == NULL) {
1590                         kfree_skb(skb);
1591                         netbk_tx_err(vif, &txreq, idx);
1592                         continue;
1593                 }
1594                 gop = request_gop;
1595
1596                 __skb_queue_tail(&netbk->tx_queue, skb);
1597
1598                 vif->tx.req_cons = idx;
1599                 xen_netbk_check_rx_xenvif(vif);
1600
1601                 if ((gop-netbk->tx_copy_ops) >= ARRAY_SIZE(netbk->tx_copy_ops))
1602                         break;
1603         }
1604
1605         return gop - netbk->tx_copy_ops;
1606 }
1607
1608 static void xen_netbk_tx_submit(struct xen_netbk *netbk)
1609 {
1610         struct gnttab_copy *gop = netbk->tx_copy_ops;
1611         struct sk_buff *skb;
1612
1613         while ((skb = __skb_dequeue(&netbk->tx_queue)) != NULL) {
1614                 struct xen_netif_tx_request *txp;
1615                 struct xenvif *vif;
1616                 u16 pending_idx;
1617                 unsigned data_len;
1618
1619                 pending_idx = *((u16 *)skb->data);
1620                 vif = netbk->pending_tx_info[pending_idx].vif;
1621                 txp = &netbk->pending_tx_info[pending_idx].req;
1622
1623                 /* Check the remap error code. */
1624                 if (unlikely(xen_netbk_tx_check_gop(netbk, skb, &gop))) {
1625                         netdev_dbg(vif->dev, "netback grant failed.\n");
1626                         skb_shinfo(skb)->nr_frags = 0;
1627                         kfree_skb(skb);
1628                         continue;
1629                 }
1630
1631                 data_len = skb->len;
1632                 memcpy(skb->data,
1633                        (void *)(idx_to_kaddr(netbk, pending_idx)|txp->offset),
1634                        data_len);
1635                 if (data_len < txp->size) {
1636                         /* Append the packet payload as a fragment. */
1637                         txp->offset += data_len;
1638                         txp->size -= data_len;
1639                 } else {
1640                         /* Schedule a response immediately. */
1641                         xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1642                 }
1643
1644                 if (txp->flags & XEN_NETTXF_csum_blank)
1645                         skb->ip_summed = CHECKSUM_PARTIAL;
1646                 else if (txp->flags & XEN_NETTXF_data_validated)
1647                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1648
1649                 xen_netbk_fill_frags(netbk, skb);
1650
1651                 /*
1652                  * If the initial fragment was < PKT_PROT_LEN then
1653                  * pull through some bytes from the other fragments to
1654                  * increase the linear region to PKT_PROT_LEN bytes.
1655                  */
1656                 if (skb_headlen(skb) < PKT_PROT_LEN && skb_is_nonlinear(skb)) {
1657                         int target = min_t(int, skb->len, PKT_PROT_LEN);
1658                         __pskb_pull_tail(skb, target - skb_headlen(skb));
1659                 }
1660
1661                 skb->dev      = vif->dev;
1662                 skb->protocol = eth_type_trans(skb, skb->dev);
1663                 skb_reset_network_header(skb);
1664
1665                 if (checksum_setup(vif, skb)) {
1666                         netdev_dbg(vif->dev,
1667                                    "Can't setup checksum in net_tx_action\n");
1668                         kfree_skb(skb);
1669                         continue;
1670                 }
1671
1672                 skb_probe_transport_header(skb, 0);
1673
1674                 vif->dev->stats.rx_bytes += skb->len;
1675                 vif->dev->stats.rx_packets++;
1676
1677                 xenvif_receive_skb(vif, skb);
1678         }
1679 }
1680
1681 /* Called after netfront has transmitted */
1682 static void xen_netbk_tx_action(struct xen_netbk *netbk)
1683 {
1684         unsigned nr_gops;
1685
1686         nr_gops = xen_netbk_tx_build_gops(netbk);
1687
1688         if (nr_gops == 0)
1689                 return;
1690
1691         gnttab_batch_copy(netbk->tx_copy_ops, nr_gops);
1692
1693         xen_netbk_tx_submit(netbk);
1694 }
1695
1696 static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx,
1697                                   u8 status)
1698 {
1699         struct xenvif *vif;
1700         struct pending_tx_info *pending_tx_info;
1701         pending_ring_idx_t head;
1702         u16 peek; /* peek into next tx request */
1703
1704         BUG_ON(netbk->mmap_pages[pending_idx] == (void *)(~0UL));
1705
1706         /* Already complete? */
1707         if (netbk->mmap_pages[pending_idx] == NULL)
1708                 return;
1709
1710         pending_tx_info = &netbk->pending_tx_info[pending_idx];
1711
1712         vif = pending_tx_info->vif;
1713         head = pending_tx_info->head;
1714
1715         BUG_ON(!pending_tx_is_head(netbk, head));
1716         BUG_ON(netbk->pending_ring[pending_index(head)] != pending_idx);
1717
1718         do {
1719                 pending_ring_idx_t index;
1720                 pending_ring_idx_t idx = pending_index(head);
1721                 u16 info_idx = netbk->pending_ring[idx];
1722
1723                 pending_tx_info = &netbk->pending_tx_info[info_idx];
1724                 make_tx_response(vif, &pending_tx_info->req, status);
1725
1726                 /* Setting any number other than
1727                  * INVALID_PENDING_RING_IDX indicates this slot is
1728                  * starting a new packet / ending a previous packet.
1729                  */
1730                 pending_tx_info->head = 0;
1731
1732                 index = pending_index(netbk->pending_prod++);
1733                 netbk->pending_ring[index] = netbk->pending_ring[info_idx];
1734
1735                 xenvif_put(vif);
1736
1737                 peek = netbk->pending_ring[pending_index(++head)];
1738
1739         } while (!pending_tx_is_head(netbk, peek));
1740
1741         netbk->mmap_pages[pending_idx]->mapping = 0;
1742         put_page(netbk->mmap_pages[pending_idx]);
1743         netbk->mmap_pages[pending_idx] = NULL;
1744 }
1745
1746
1747 static void make_tx_response(struct xenvif *vif,
1748                              struct xen_netif_tx_request *txp,
1749                              s8       st)
1750 {
1751         RING_IDX i = vif->tx.rsp_prod_pvt;
1752         struct xen_netif_tx_response *resp;
1753         int notify;
1754
1755         resp = RING_GET_RESPONSE(&vif->tx, i);
1756         resp->id     = txp->id;
1757         resp->status = st;
1758
1759         if (txp->flags & XEN_NETTXF_extra_info)
1760                 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1761
1762         vif->tx.rsp_prod_pvt = ++i;
1763         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1764         if (notify)
1765                 notify_remote_via_irq(vif->irq);
1766 }
1767
1768 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1769                                              u16      id,
1770                                              s8       st,
1771                                              u16      offset,
1772                                              u16      size,
1773                                              u16      flags)
1774 {
1775         RING_IDX i = vif->rx.rsp_prod_pvt;
1776         struct xen_netif_rx_response *resp;
1777
1778         resp = RING_GET_RESPONSE(&vif->rx, i);
1779         resp->offset     = offset;
1780         resp->flags      = flags;
1781         resp->id         = id;
1782         resp->status     = (s16)size;
1783         if (st < 0)
1784                 resp->status = (s16)st;
1785
1786         vif->rx.rsp_prod_pvt = ++i;
1787
1788         return resp;
1789 }
1790
1791 static inline int rx_work_todo(struct xen_netbk *netbk)
1792 {
1793         return !skb_queue_empty(&netbk->rx_queue);
1794 }
1795
1796 static inline int tx_work_todo(struct xen_netbk *netbk)
1797 {
1798
1799         if ((nr_pending_reqs(netbk) + XEN_NETBK_LEGACY_SLOTS_MAX
1800              < MAX_PENDING_REQS) &&
1801              !list_empty(&netbk->net_schedule_list))
1802                 return 1;
1803
1804         return 0;
1805 }
1806
1807 static int xen_netbk_kthread(void *data)
1808 {
1809         struct xen_netbk *netbk = data;
1810         while (!kthread_should_stop()) {
1811                 wait_event_interruptible(netbk->wq,
1812                                 rx_work_todo(netbk) ||
1813                                 tx_work_todo(netbk) ||
1814                                 kthread_should_stop());
1815                 cond_resched();
1816
1817                 if (kthread_should_stop())
1818                         break;
1819
1820                 if (rx_work_todo(netbk))
1821                         xen_netbk_rx_action(netbk);
1822
1823                 if (tx_work_todo(netbk))
1824                         xen_netbk_tx_action(netbk);
1825         }
1826
1827         return 0;
1828 }
1829
1830 void xen_netbk_unmap_frontend_rings(struct xenvif *vif)
1831 {
1832         if (vif->tx.sring)
1833                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1834                                         vif->tx.sring);
1835         if (vif->rx.sring)
1836                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1837                                         vif->rx.sring);
1838 }
1839
1840 int xen_netbk_map_frontend_rings(struct xenvif *vif,
1841                                  grant_ref_t tx_ring_ref,
1842                                  grant_ref_t rx_ring_ref)
1843 {
1844         void *addr;
1845         struct xen_netif_tx_sring *txs;
1846         struct xen_netif_rx_sring *rxs;
1847
1848         int err = -ENOMEM;
1849
1850         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1851                                      tx_ring_ref, &addr);
1852         if (err)
1853                 goto err;
1854
1855         txs = (struct xen_netif_tx_sring *)addr;
1856         BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1857
1858         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1859                                      rx_ring_ref, &addr);
1860         if (err)
1861                 goto err;
1862
1863         rxs = (struct xen_netif_rx_sring *)addr;
1864         BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1865
1866         vif->rx_req_cons_peek = 0;
1867
1868         return 0;
1869
1870 err:
1871         xen_netbk_unmap_frontend_rings(vif);
1872         return err;
1873 }
1874
1875 static int __init netback_init(void)
1876 {
1877         int i;
1878         int rc = 0;
1879         int group;
1880
1881         if (!xen_domain())
1882                 return -ENODEV;
1883
1884         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1885                 printk(KERN_INFO
1886                        "xen-netback: fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1887                        fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1888                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1889         }
1890
1891         xen_netbk_group_nr = num_online_cpus();
1892         xen_netbk = vzalloc(sizeof(struct xen_netbk) * xen_netbk_group_nr);
1893         if (!xen_netbk)
1894                 return -ENOMEM;
1895
1896         for (group = 0; group < xen_netbk_group_nr; group++) {
1897                 struct xen_netbk *netbk = &xen_netbk[group];
1898                 skb_queue_head_init(&netbk->rx_queue);
1899                 skb_queue_head_init(&netbk->tx_queue);
1900
1901                 init_timer(&netbk->net_timer);
1902                 netbk->net_timer.data = (unsigned long)netbk;
1903                 netbk->net_timer.function = xen_netbk_alarm;
1904
1905                 netbk->pending_cons = 0;
1906                 netbk->pending_prod = MAX_PENDING_REQS;
1907                 for (i = 0; i < MAX_PENDING_REQS; i++)
1908                         netbk->pending_ring[i] = i;
1909
1910                 init_waitqueue_head(&netbk->wq);
1911                 netbk->task = kthread_create(xen_netbk_kthread,
1912                                              (void *)netbk,
1913                                              "netback/%u", group);
1914
1915                 if (IS_ERR(netbk->task)) {
1916                         printk(KERN_ALERT "kthread_create() fails at netback\n");
1917                         del_timer(&netbk->net_timer);
1918                         rc = PTR_ERR(netbk->task);
1919                         goto failed_init;
1920                 }
1921
1922                 kthread_bind(netbk->task, group);
1923
1924                 INIT_LIST_HEAD(&netbk->net_schedule_list);
1925
1926                 spin_lock_init(&netbk->net_schedule_list_lock);
1927
1928                 atomic_set(&netbk->netfront_count, 0);
1929
1930                 wake_up_process(netbk->task);
1931         }
1932
1933         rc = xenvif_xenbus_init();
1934         if (rc)
1935                 goto failed_init;
1936
1937         return 0;
1938
1939 failed_init:
1940         while (--group >= 0) {
1941                 struct xen_netbk *netbk = &xen_netbk[group];
1942                 del_timer(&netbk->net_timer);
1943                 kthread_stop(netbk->task);
1944         }
1945         vfree(xen_netbk);
1946         return rc;
1947
1948 }
1949
1950 module_init(netback_init);
1951
1952 MODULE_LICENSE("Dual BSD/GPL");
1953 MODULE_ALIAS("xen-backend:vif");