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 #include <linux/highmem.h>
  41
  42 #include <net/tcp.h>
  43
  44 #include <xen/xen.h>
  45 #include <xen/events.h>
  46 #include <xen/interface/memory.h>
  47 #include <xen/page.h>
  48
  49 #include <asm/xen/hypercall.h>
  50
  51 /* Provide an option to disable split event channels at load time as
  52  * event channels are limited resource. Split event channels are
  53  * enabled by default.
  54  */
  55 bool separate_tx_rx_irq = true;
  56 module_param(separate_tx_rx_irq, bool, 0644);
  57
  58 /* The time that packets can stay on the guest Rx internal queue
  59  * before they are dropped.
  60  */
  61 unsigned int rx_drain_timeout_msecs = 10000;
  62 module_param(rx_drain_timeout_msecs, uint, 0444);
  63
  64 /* The length of time before the frontend is considered unresponsive
  65  * because it isn't providing Rx slots.
  66  */
  67 unsigned int rx_stall_timeout_msecs = 60000;
  68 module_param(rx_stall_timeout_msecs, uint, 0444);
  69
  70 unsigned int xenvif_max_queues;
  71 module_param_named(max_queues, xenvif_max_queues, uint, 0644);
  72 MODULE_PARM_DESC(max_queues,
  73                  "Maximum number of queues per virtual interface");
  74
  75 /*
  76  * This is the maximum slots a skb can have. If a guest sends a skb
  77  * which exceeds this limit it is considered malicious.
  78  */
  79 #define FATAL_SKB_SLOTS_DEFAULT 20
  80 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
  81 module_param(fatal_skb_slots, uint, 0444);
  82
  83 /* The amount to copy out of the first guest Tx slot into the skb's
  84  * linear area.  If the first slot has more data, it will be mapped
  85  * and put into the first frag.
  86  *
  87  * This is sized to avoid pulling headers from the frags for most
  88  * TCP/IP packets.
  89  */
  90 #define XEN_NETBACK_TX_COPY_LEN 128
  91
  92
  93 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
  94                                u8 status);
  95
  96 static void make_tx_response(struct xenvif_queue *queue,
  97                              struct xen_netif_tx_request *txp,
  98                              s8       st);
  99 static void push_tx_responses(struct xenvif_queue *queue);
 100
 101 static inline int tx_work_todo(struct xenvif_queue *queue);
 102
 103 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
 104                                              u16      id,
 105                                              s8       st,
 106                                              u16      offset,
 107                                              u16      size,
 108                                              u16      flags);
 109
 110 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
 111                                        u16 idx)
 112 {
 113         return page_to_pfn(queue->mmap_pages[idx]);
 114 }
 115
 116 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
 117                                          u16 idx)
 118 {
 119         return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
 120 }
 121
 122 #define callback_param(vif, pending_idx) \
 123         (vif->pending_tx_info[pending_idx].callback_struct)
 124
 125 /* Find the containing VIF's structure from a pointer in pending_tx_info array
 126  */
 127 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
 128 {
 129         u16 pending_idx = ubuf->desc;
 130         struct pending_tx_info *temp =
 131                 container_of(ubuf, struct pending_tx_info, callback_struct);
 132         return container_of(temp - pending_idx,
 133                             struct xenvif_queue,
 134                             pending_tx_info[0]);
 135 }
 136
 137 static u16 frag_get_pending_idx(skb_frag_t *frag)
 138 {
 139         return (u16)frag->page_offset;
 140 }
 141
 142 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 143 {
 144         frag->page_offset = pending_idx;
 145 }
 146
 147 static inline pending_ring_idx_t pending_index(unsigned i)
 148 {
 149         return i & (MAX_PENDING_REQS-1);
 150 }
 151
 152 bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue, int needed)
 153 {
 154         RING_IDX prod, cons;
 155
 156         do {
 157                 prod = queue->rx.sring->req_prod;
 158                 cons = queue->rx.req_cons;
 159
 160                 if (prod - cons >= needed)
 161                         return true;
 162
 163                 queue->rx.sring->req_event = prod + 1;
 164
 165                 /* Make sure event is visible before we check prod
 166                  * again.
 167                  */
 168                 mb();
 169         } while (queue->rx.sring->req_prod != prod);
 170
 171         return false;
 172 }
 173
 174 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
 175 {
 176         unsigned long flags;
 177
 178         spin_lock_irqsave(&queue->rx_queue.lock, flags);
 179
 180         __skb_queue_tail(&queue->rx_queue, skb);
 181
 182         queue->rx_queue_len += skb->len;
 183         if (queue->rx_queue_len > queue->rx_queue_max)
 184                 netif_tx_stop_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
 185
 186         spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
 187 }
 188
 189 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
 190 {
 191         struct sk_buff *skb;
 192
 193         spin_lock_irq(&queue->rx_queue.lock);
 194
 195         skb = __skb_dequeue(&queue->rx_queue);
 196         if (skb)
 197                 queue->rx_queue_len -= skb->len;
 198
 199         spin_unlock_irq(&queue->rx_queue.lock);
 200
 201         return skb;
 202 }
 203
 204 static void xenvif_rx_queue_maybe_wake(struct xenvif_queue *queue)
 205 {
 206         spin_lock_irq(&queue->rx_queue.lock);
 207
 208         if (queue->rx_queue_len < queue->rx_queue_max)
 209                 netif_tx_wake_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
 210
 211         spin_unlock_irq(&queue->rx_queue.lock);
 212 }
 213
 214
 215 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
 216 {
 217         struct sk_buff *skb;
 218         while ((skb = xenvif_rx_dequeue(queue)) != NULL)
 219                 kfree_skb(skb);
 220 }
 221
 222 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
 223 {
 224         struct sk_buff *skb;
 225
 226         for(;;) {
 227                 skb = skb_peek(&queue->rx_queue);
 228                 if (!skb)
 229                         break;
 230                 if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
 231                         break;
 232                 xenvif_rx_dequeue(queue);
 233                 kfree_skb(skb);
 234         }
 235 }
 236
 237 struct netrx_pending_operations {
 238         unsigned copy_prod, copy_cons;
 239         unsigned meta_prod, meta_cons;
 240         struct gnttab_copy *copy;
 241         struct xenvif_rx_meta *meta;
 242         int copy_off;
 243         grant_ref_t copy_gref;
 244 };
 245
 246 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue,
 247                                                  struct netrx_pending_operations *npo)
 248 {
 249         struct xenvif_rx_meta *meta;
 250         struct xen_netif_rx_request *req;
 251
 252         req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
 253
 254         meta = npo->meta + npo->meta_prod++;
 255         meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
 256         meta->gso_size = 0;
 257         meta->size = 0;
 258         meta->id = req->id;
 259
 260         npo->copy_off = 0;
 261         npo->copy_gref = req->gref;
 262
 263         return meta;
 264 }
 265
 266 /*
 267  * Set up the grant operations for this fragment. If it's a flipping
 268  * interface, we also set up the unmap request from here.
 269  */
 270 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
 271                                  struct netrx_pending_operations *npo,
 272                                  struct page *page, unsigned long size,
 273                                  unsigned long offset, int *head)
 274 {
 275         struct gnttab_copy *copy_gop;
 276         struct xenvif_rx_meta *meta;
 277         unsigned long bytes;
 278         int gso_type = XEN_NETIF_GSO_TYPE_NONE;
 279
 280         /* Data must not cross a page boundary. */
 281         BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
 282
 283         meta = npo->meta + npo->meta_prod - 1;
 284
 285         /* Skip unused frames from start of page */
 286         page += offset >> PAGE_SHIFT;
 287         offset &= ~PAGE_MASK;
 288
 289         while (size > 0) {
 290                 struct xen_page_foreign *foreign;
 291
 292                 BUG_ON(offset >= PAGE_SIZE);
 293                 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
 294
 295                 if (npo->copy_off == MAX_BUFFER_OFFSET)
 296                         meta = get_next_rx_buffer(queue, npo);
 297
 298                 bytes = PAGE_SIZE - offset;
 299                 if (bytes > size)
 300                         bytes = size;
 301
 302                 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
 303                         bytes = MAX_BUFFER_OFFSET - npo->copy_off;
 304
 305                 copy_gop = npo->copy + npo->copy_prod++;
 306                 copy_gop->flags = GNTCOPY_dest_gref;
 307                 copy_gop->len = bytes;
 308
 309                 foreign = xen_page_foreign(page);
 310                 if (foreign) {
 311                         copy_gop->source.domid = foreign->domid;
 312                         copy_gop->source.u.ref = foreign->gref;
 313                         copy_gop->flags |= GNTCOPY_source_gref;
 314                 } else {
 315                         copy_gop->source.domid = DOMID_SELF;
 316                         copy_gop->source.u.gmfn =
 317                                 virt_to_mfn(page_address(page));
 318                 }
 319                 copy_gop->source.offset = offset;
 320
 321                 copy_gop->dest.domid = queue->vif->domid;
 322                 copy_gop->dest.offset = npo->copy_off;
 323                 copy_gop->dest.u.ref = npo->copy_gref;
 324
 325                 npo->copy_off += bytes;
 326                 meta->size += bytes;
 327
 328                 offset += bytes;
 329                 size -= bytes;
 330
 331                 /* Next frame */
 332                 if (offset == PAGE_SIZE && size) {
 333                         BUG_ON(!PageCompound(page));
 334                         page++;
 335                         offset = 0;
 336                 }
 337
 338                 /* Leave a gap for the GSO descriptor. */
 339                 if (skb_is_gso(skb)) {
 340                         if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 341                                 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
 342                         else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 343                                 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
 344                 }
 345
 346                 if (*head && ((1 << gso_type) & queue->vif->gso_mask))
 347                         queue->rx.req_cons++;
 348
 349                 *head = 0; /* There must be something in this buffer now. */
 350
 351         }
 352 }
 353
 354 /*
 355  * Prepare an SKB to be transmitted to the frontend.
 356  *
 357  * This function is responsible for allocating grant operations, meta
 358  * structures, etc.
 359  *
 360  * It returns the number of meta structures consumed. The number of
 361  * ring slots used is always equal to the number of meta slots used
 362  * plus the number of GSO descriptors used. Currently, we use either
 363  * zero GSO descriptors (for non-GSO packets) or one descriptor (for
 364  * frontend-side LRO).
 365  */
 366 static int xenvif_gop_skb(struct sk_buff *skb,
 367                           struct netrx_pending_operations *npo,
 368                           struct xenvif_queue *queue)
 369 {
 370         struct xenvif *vif = netdev_priv(skb->dev);
 371         int nr_frags = skb_shinfo(skb)->nr_frags;
 372         int i;
 373         struct xen_netif_rx_request *req;
 374         struct xenvif_rx_meta *meta;
 375         unsigned char *data;
 376         int head = 1;
 377         int old_meta_prod;
 378         int gso_type;
 379
 380         old_meta_prod = npo->meta_prod;
 381
 382         gso_type = XEN_NETIF_GSO_TYPE_NONE;
 383         if (skb_is_gso(skb)) {
 384                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 385                         gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
 386                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 387                         gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
 388         }
 389
 390         /* Set up a GSO prefix descriptor, if necessary */
 391         if ((1 << gso_type) & vif->gso_prefix_mask) {
 392                 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
 393                 meta = npo->meta + npo->meta_prod++;
 394                 meta->gso_type = gso_type;
 395                 meta->gso_size = skb_shinfo(skb)->gso_size;
 396                 meta->size = 0;
 397                 meta->id = req->id;
 398         }
 399
 400         req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
 401         meta = npo->meta + npo->meta_prod++;
 402
 403         if ((1 << gso_type) & vif->gso_mask) {
 404                 meta->gso_type = gso_type;
 405                 meta->gso_size = skb_shinfo(skb)->gso_size;
 406         } else {
 407                 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
 408                 meta->gso_size = 0;
 409         }
 410
 411         meta->size = 0;
 412         meta->id = req->id;
 413         npo->copy_off = 0;
 414         npo->copy_gref = req->gref;
 415
 416         data = skb->data;
 417         while (data < skb_tail_pointer(skb)) {
 418                 unsigned int offset = offset_in_page(data);
 419                 unsigned int len = PAGE_SIZE - offset;
 420
 421                 if (data + len > skb_tail_pointer(skb))
 422                         len = skb_tail_pointer(skb) - data;
 423
 424                 xenvif_gop_frag_copy(queue, skb, npo,
 425                                      virt_to_page(data), len, offset, &head);
 426                 data += len;
 427         }
 428
 429         for (i = 0; i < nr_frags; i++) {
 430                 xenvif_gop_frag_copy(queue, skb, npo,
 431                                      skb_frag_page(&skb_shinfo(skb)->frags[i]),
 432                                      skb_frag_size(&skb_shinfo(skb)->frags[i]),
 433                                      skb_shinfo(skb)->frags[i].page_offset,
 434                                      &head);
 435         }
 436
 437         return npo->meta_prod - old_meta_prod;
 438 }
 439
 440 /*
 441  * This is a twin to xenvif_gop_skb.  Assume that xenvif_gop_skb was
 442  * used to set up the operations on the top of
 443  * netrx_pending_operations, which have since been done.  Check that
 444  * they didn't give any errors and advance over them.
 445  */
 446 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
 447                             struct netrx_pending_operations *npo)
 448 {
 449         struct gnttab_copy     *copy_op;
 450         int status = XEN_NETIF_RSP_OKAY;
 451         int i;
 452
 453         for (i = 0; i < nr_meta_slots; i++) {
 454                 copy_op = npo->copy + npo->copy_cons++;
 455                 if (copy_op->status != GNTST_okay) {
 456                         netdev_dbg(vif->dev,
 457                                    "Bad status %d from copy to DOM%d.\n",
 458                                    copy_op->status, vif->domid);
 459                         status = XEN_NETIF_RSP_ERROR;
 460                 }
 461         }
 462
 463         return status;
 464 }
 465
 466 static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status,
 467                                       struct xenvif_rx_meta *meta,
 468                                       int nr_meta_slots)
 469 {
 470         int i;
 471         unsigned long offset;
 472
 473         /* No fragments used */
 474         if (nr_meta_slots <= 1)
 475                 return;
 476
 477         nr_meta_slots--;
 478
 479         for (i = 0; i < nr_meta_slots; i++) {
 480                 int flags;
 481                 if (i == nr_meta_slots - 1)
 482                         flags = 0;
 483                 else
 484                         flags = XEN_NETRXF_more_data;
 485
 486                 offset = 0;
 487                 make_rx_response(queue, meta[i].id, status, offset,
 488                                  meta[i].size, flags);
 489         }
 490 }
 491
 492 void xenvif_kick_thread(struct xenvif_queue *queue)
 493 {
 494         wake_up(&queue->wq);
 495 }
 496
 497 static void xenvif_rx_action(struct xenvif_queue *queue)
 498 {
 499         s8 status;
 500         u16 flags;
 501         struct xen_netif_rx_response *resp;
 502         struct sk_buff_head rxq;
 503         struct sk_buff *skb;
 504         LIST_HEAD(notify);
 505         int ret;
 506         unsigned long offset;
 507         bool need_to_notify = false;
 508
 509         struct netrx_pending_operations npo = {
 510                 .copy  = queue->grant_copy_op,
 511                 .meta  = queue->meta,
 512         };
 513
 514         skb_queue_head_init(&rxq);
 515
 516         while (xenvif_rx_ring_slots_available(queue, XEN_NETBK_RX_SLOTS_MAX)
 517                && (skb = xenvif_rx_dequeue(queue)) != NULL) {
 518                 queue->last_rx_time = jiffies;
 519
 520                 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
 521
 522                 __skb_queue_tail(&rxq, skb);
 523         }
 524
 525         BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
 526
 527         if (!npo.copy_prod)
 528                 goto done;
 529
 530         BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
 531         gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
 532
 533         while ((skb = __skb_dequeue(&rxq)) != NULL) {
 534
 535                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 536                     queue->vif->gso_prefix_mask) {
 537                         resp = RING_GET_RESPONSE(&queue->rx,
 538                                                  queue->rx.rsp_prod_pvt++);
 539
 540                         resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
 541
 542                         resp->offset = queue->meta[npo.meta_cons].gso_size;
 543                         resp->id = queue->meta[npo.meta_cons].id;
 544                         resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
 545
 546                         npo.meta_cons++;
 547                         XENVIF_RX_CB(skb)->meta_slots_used--;
 548                 }
 549
 550
 551                 queue->stats.tx_bytes += skb->len;
 552                 queue->stats.tx_packets++;
 553
 554                 status = xenvif_check_gop(queue->vif,
 555                                           XENVIF_RX_CB(skb)->meta_slots_used,
 556                                           &npo);
 557
 558                 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
 559                         flags = 0;
 560                 else
 561                         flags = XEN_NETRXF_more_data;
 562
 563                 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
 564                         flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
 565                 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 566                         /* remote but checksummed. */
 567                         flags |= XEN_NETRXF_data_validated;
 568
 569                 offset = 0;
 570                 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
 571                                         status, offset,
 572                                         queue->meta[npo.meta_cons].size,
 573                                         flags);
 574
 575                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 576                     queue->vif->gso_mask) {
 577                         struct xen_netif_extra_info *gso =
 578                                 (struct xen_netif_extra_info *)
 579                                 RING_GET_RESPONSE(&queue->rx,
 580                                                   queue->rx.rsp_prod_pvt++);
 581
 582                         resp->flags |= XEN_NETRXF_extra_info;
 583
 584                         gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
 585                         gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
 586                         gso->u.gso.pad = 0;
 587                         gso->u.gso.features = 0;
 588
 589                         gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
 590                         gso->flags = 0;
 591                 }
 592
 593                 xenvif_add_frag_responses(queue, status,
 594                                           queue->meta + npo.meta_cons + 1,
 595                                           XENVIF_RX_CB(skb)->meta_slots_used);
 596
 597                 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
 598
 599                 need_to_notify |= !!ret;
 600
 601                 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
 602                 dev_kfree_skb(skb);
 603         }
 604
 605 done:
 606         if (need_to_notify)
 607                 notify_remote_via_irq(queue->rx_irq);
 608 }
 609
 610 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 611 {
 612         int more_to_do;
 613
 614         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 615
 616         if (more_to_do)
 617                 napi_schedule(&queue->napi);
 618 }
 619
 620 static void tx_add_credit(struct xenvif_queue *queue)
 621 {
 622         unsigned long max_burst, max_credit;
 623
 624         /*
 625          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 626          * Otherwise the interface can seize up due to insufficient credit.
 627          */
 628         max_burst = RING_GET_REQUEST(&queue->tx, queue->tx.req_cons)->size;
 629         max_burst = min(max_burst, 131072UL);
 630         max_burst = max(max_burst, queue->credit_bytes);
 631
 632         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 633         max_credit = queue->remaining_credit + queue->credit_bytes;
 634         if (max_credit < queue->remaining_credit)
 635                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 636
 637         queue->remaining_credit = min(max_credit, max_burst);
 638 }
 639
 640 void xenvif_tx_credit_callback(unsigned long data)
 641 {
 642         struct xenvif_queue *queue = (struct xenvif_queue *)data;
 643         tx_add_credit(queue);
 644         xenvif_napi_schedule_or_enable_events(queue);
 645 }
 646
 647 static void xenvif_tx_err(struct xenvif_queue *queue,
 648                           struct xen_netif_tx_request *txp, RING_IDX end)
 649 {
 650         RING_IDX cons = queue->tx.req_cons;
 651         unsigned long flags;
 652
 653         do {
 654                 spin_lock_irqsave(&queue->response_lock, flags);
 655                 make_tx_response(queue, txp, XEN_NETIF_RSP_ERROR);
 656                 push_tx_responses(queue);
 657                 spin_unlock_irqrestore(&queue->response_lock, flags);
 658                 if (cons == end)
 659                         break;
 660                 txp = RING_GET_REQUEST(&queue->tx, cons++);
 661         } while (1);
 662         queue->tx.req_cons = cons;
 663 }
 664
 665 static void xenvif_fatal_tx_err(struct xenvif *vif)
 666 {
 667         netdev_err(vif->dev, "fatal error; disabling device\n");
 668         vif->disabled = true;
 669         /* Disable the vif from queue 0's kthread */
 670         if (vif->queues)
 671                 xenvif_kick_thread(&vif->queues[0]);
 672 }
 673
 674 static int xenvif_count_requests(struct xenvif_queue *queue,
 675                                  struct xen_netif_tx_request *first,
 676                                  struct xen_netif_tx_request *txp,
 677                                  int work_to_do)
 678 {
 679         RING_IDX cons = queue->tx.req_cons;
 680         int slots = 0;
 681         int drop_err = 0;
 682         int more_data;
 683
 684         if (!(first->flags & XEN_NETTXF_more_data))
 685                 return 0;
 686
 687         do {
 688                 struct xen_netif_tx_request dropped_tx = { 0 };
 689
 690                 if (slots >= work_to_do) {
 691                         netdev_err(queue->vif->dev,
 692                                    "Asked for %d slots but exceeds this limit\n",
 693                                    work_to_do);
 694                         xenvif_fatal_tx_err(queue->vif);
 695                         return -ENODATA;
 696                 }
 697
 698                 /* This guest is really using too many slots and
 699                  * considered malicious.
 700                  */
 701                 if (unlikely(slots >= fatal_skb_slots)) {
 702                         netdev_err(queue->vif->dev,
 703                                    "Malicious frontend using %d slots, threshold %u\n",
 704                                    slots, fatal_skb_slots);
 705                         xenvif_fatal_tx_err(queue->vif);
 706                         return -E2BIG;
 707                 }
 708
 709                 /* Xen network protocol had implicit dependency on
 710                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 711                  * the historical MAX_SKB_FRAGS value 18 to honor the
 712                  * same behavior as before. Any packet using more than
 713                  * 18 slots but less than fatal_skb_slots slots is
 714                  * dropped
 715                  */
 716                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 717                         if (net_ratelimit())
 718                                 netdev_dbg(queue->vif->dev,
 719                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 720                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 721                         drop_err = -E2BIG;
 722                 }
 723
 724                 if (drop_err)
 725                         txp = &dropped_tx;
 726
 727                 memcpy(txp, RING_GET_REQUEST(&queue->tx, cons + slots),
 728                        sizeof(*txp));
 729
 730                 /* If the guest submitted a frame >= 64 KiB then
 731                  * first->size overflowed and following slots will
 732                  * appear to be larger than the frame.
 733                  *
 734                  * This cannot be fatal error as there are buggy
 735                  * frontends that do this.
 736                  *
 737                  * Consume all slots and drop the packet.
 738                  */
 739                 if (!drop_err && txp->size > first->size) {
 740                         if (net_ratelimit())
 741                                 netdev_dbg(queue->vif->dev,
 742                                            "Invalid tx request, slot size %u > remaining size %u\n",
 743                                            txp->size, first->size);
 744                         drop_err = -EIO;
 745                 }
 746
 747                 first->size -= txp->size;
 748                 slots++;
 749
 750                 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
 751                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 752                                  txp->offset, txp->size);
 753                         xenvif_fatal_tx_err(queue->vif);
 754                         return -EINVAL;
 755                 }
 756
 757                 more_data = txp->flags & XEN_NETTXF_more_data;
 758
 759                 if (!drop_err)
 760                         txp++;
 761
 762         } while (more_data);
 763
 764         if (drop_err) {
 765                 xenvif_tx_err(queue, first, cons + slots);
 766                 return drop_err;
 767         }
 768
 769         return slots;
 770 }
 771
 772
 773 struct xenvif_tx_cb {
 774         u16 pending_idx;
 775 };
 776
 777 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 778
 779 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 780                                           u16 pending_idx,
 781                                           struct xen_netif_tx_request *txp,
 782                                           struct gnttab_map_grant_ref *mop)
 783 {
 784         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 785         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 786                           GNTMAP_host_map | GNTMAP_readonly,
 787                           txp->gref, queue->vif->domid);
 788
 789         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 790                sizeof(*txp));
 791 }
 792
 793 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 794 {
 795         struct sk_buff *skb =
 796                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 797                           GFP_ATOMIC | __GFP_NOWARN);
 798         if (unlikely(skb == NULL))
 799                 return NULL;
 800
 801         /* Packets passed to netif_rx() must have some headroom. */
 802         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 803
 804         /* Initialize it here to avoid later surprises */
 805         skb_shinfo(skb)->destructor_arg = NULL;
 806
 807         return skb;
 808 }
 809
 810 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
 811                                                         struct sk_buff *skb,
 812                                                         struct xen_netif_tx_request *txp,
 813                                                         struct gnttab_map_grant_ref *gop,
 814                                                         unsigned int frag_overflow,
 815                                                         struct sk_buff *nskb)
 816 {
 817         struct skb_shared_info *shinfo = skb_shinfo(skb);
 818         skb_frag_t *frags = shinfo->frags;
 819         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 820         int start;
 821         pending_ring_idx_t index;
 822         unsigned int nr_slots;
 823
 824         nr_slots = shinfo->nr_frags;
 825
 826         /* Skip first skb fragment if it is on same page as header fragment. */
 827         start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
 828
 829         for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
 830              shinfo->nr_frags++, txp++, gop++) {
 831                 index = pending_index(queue->pending_cons++);
 832                 pending_idx = queue->pending_ring[index];
 833                 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
 834                 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 835         }
 836
 837         if (frag_overflow) {
 838
 839                 shinfo = skb_shinfo(nskb);
 840                 frags = shinfo->frags;
 841
 842                 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
 843                      shinfo->nr_frags++, txp++, gop++) {
 844                         index = pending_index(queue->pending_cons++);
 845                         pending_idx = queue->pending_ring[index];
 846                         xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
 847                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 848                                              pending_idx);
 849                 }
 850
 851                 skb_shinfo(skb)->frag_list = nskb;
 852         }
 853
 854         return gop;
 855 }
 856
 857 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 858                                            u16 pending_idx,
 859                                            grant_handle_t handle)
 860 {
 861         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 862                      NETBACK_INVALID_HANDLE)) {
 863                 netdev_err(queue->vif->dev,
 864                            "Trying to overwrite active handle! pending_idx: 0x%x\n",
 865                            pending_idx);
 866                 BUG();
 867         }
 868         queue->grant_tx_handle[pending_idx] = handle;
 869 }
 870
 871 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 872                                              u16 pending_idx)
 873 {
 874         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 875                      NETBACK_INVALID_HANDLE)) {
 876                 netdev_err(queue->vif->dev,
 877                            "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 878                            pending_idx);
 879                 BUG();
 880         }
 881         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 882 }
 883
 884 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 885                                struct sk_buff *skb,
 886                                struct gnttab_map_grant_ref **gopp_map,
 887                                struct gnttab_copy **gopp_copy)
 888 {
 889         struct gnttab_map_grant_ref *gop_map = *gopp_map;
 890         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 891         /* This always points to the shinfo of the skb being checked, which
 892          * could be either the first or the one on the frag_list
 893          */
 894         struct skb_shared_info *shinfo = skb_shinfo(skb);
 895         /* If this is non-NULL, we are currently checking the frag_list skb, and
 896          * this points to the shinfo of the first one
 897          */
 898         struct skb_shared_info *first_shinfo = NULL;
 899         int nr_frags = shinfo->nr_frags;
 900         const bool sharedslot = nr_frags &&
 901                                 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
 902         int i, err;
 903
 904         /* Check status of header. */
 905         err = (*gopp_copy)->status;
 906         if (unlikely(err)) {
 907                 if (net_ratelimit())
 908                         netdev_dbg(queue->vif->dev,
 909                                    "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 910                                    (*gopp_copy)->status,
 911                                    pending_idx,
 912                                    (*gopp_copy)->source.u.ref);
 913                 /* The first frag might still have this slot mapped */
 914                 if (!sharedslot)
 915                         xenvif_idx_release(queue, pending_idx,
 916                                            XEN_NETIF_RSP_ERROR);
 917         }
 918         (*gopp_copy)++;
 919
 920 check_frags:
 921         for (i = 0; i < nr_frags; i++, gop_map++) {
 922                 int j, newerr;
 923
 924                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 925
 926                 /* Check error status: if okay then remember grant handle. */
 927                 newerr = gop_map->status;
 928
 929                 if (likely(!newerr)) {
 930                         xenvif_grant_handle_set(queue,
 931                                                 pending_idx,
 932                                                 gop_map->handle);
 933                         /* Had a previous error? Invalidate this fragment. */
 934                         if (unlikely(err)) {
 935                                 xenvif_idx_unmap(queue, pending_idx);
 936                                 /* If the mapping of the first frag was OK, but
 937                                  * the header's copy failed, and they are
 938                                  * sharing a slot, send an error
 939                                  */
 940                                 if (i == 0 && sharedslot)
 941                                         xenvif_idx_release(queue, pending_idx,
 942                                                            XEN_NETIF_RSP_ERROR);
 943                                 else
 944                                         xenvif_idx_release(queue, pending_idx,
 945                                                            XEN_NETIF_RSP_OKAY);
 946                         }
 947                         continue;
 948                 }
 949
 950                 /* Error on this fragment: respond to client with an error. */
 951                 if (net_ratelimit())
 952                         netdev_dbg(queue->vif->dev,
 953                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
 954                                    i,
 955                                    gop_map->status,
 956                                    pending_idx,
 957                                    gop_map->ref);
 958
 959                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
 960
 961                 /* Not the first error? Preceding frags already invalidated. */
 962                 if (err)
 963                         continue;
 964
 965                 /* First error: if the header haven't shared a slot with the
 966                  * first frag, release it as well.
 967                  */
 968                 if (!sharedslot)
 969                         xenvif_idx_release(queue,
 970                                            XENVIF_TX_CB(skb)->pending_idx,
 971                                            XEN_NETIF_RSP_OKAY);
 972
 973                 /* Invalidate preceding fragments of this skb. */
 974                 for (j = 0; j < i; j++) {
 975                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
 976                         xenvif_idx_unmap(queue, pending_idx);
 977                         xenvif_idx_release(queue, pending_idx,
 978                                            XEN_NETIF_RSP_OKAY);
 979                 }
 980
 981                 /* And if we found the error while checking the frag_list, unmap
 982                  * the first skb's frags
 983                  */
 984                 if (first_shinfo) {
 985                         for (j = 0; j < first_shinfo->nr_frags; j++) {
 986                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
 987                                 xenvif_idx_unmap(queue, pending_idx);
 988                                 xenvif_idx_release(queue, pending_idx,
 989                                                    XEN_NETIF_RSP_OKAY);
 990                         }
 991                 }
 992
 993                 /* Remember the error: invalidate all subsequent fragments. */
 994                 err = newerr;
 995         }
 996
 997         if (skb_has_frag_list(skb) && !first_shinfo) {
 998                 first_shinfo = skb_shinfo(skb);
 999                 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1000                 nr_frags = shinfo->nr_frags;
1001
1002                 goto check_frags;
1003         }
1004
1005         *gopp_map = gop_map;
1006         return err;
1007 }
1008
1009 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1010 {
1011         struct skb_shared_info *shinfo = skb_shinfo(skb);
1012         int nr_frags = shinfo->nr_frags;
1013         int i;
1014         u16 prev_pending_idx = INVALID_PENDING_IDX;
1015
1016         for (i = 0; i < nr_frags; i++) {
1017                 skb_frag_t *frag = shinfo->frags + i;
1018                 struct xen_netif_tx_request *txp;
1019                 struct page *page;
1020                 u16 pending_idx;
1021
1022                 pending_idx = frag_get_pending_idx(frag);
1023
1024                 /* If this is not the first frag, chain it to the previous*/
1025                 if (prev_pending_idx == INVALID_PENDING_IDX)
1026                         skb_shinfo(skb)->destructor_arg =
1027                                 &callback_param(queue, pending_idx);
1028                 else
1029                         callback_param(queue, prev_pending_idx).ctx =
1030                                 &callback_param(queue, pending_idx);
1031
1032                 callback_param(queue, pending_idx).ctx = NULL;
1033                 prev_pending_idx = pending_idx;
1034
1035                 txp = &queue->pending_tx_info[pending_idx].req;
1036                 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1037                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1038                 skb->len += txp->size;
1039                 skb->data_len += txp->size;
1040                 skb->truesize += txp->size;
1041
1042                 /* Take an extra reference to offset network stack's put_page */
1043                 get_page(queue->mmap_pages[pending_idx]);
1044         }
1045 }
1046
1047 static int xenvif_get_extras(struct xenvif_queue *queue,
1048                                 struct xen_netif_extra_info *extras,
1049                                 int work_to_do)
1050 {
1051         struct xen_netif_extra_info extra;
1052         RING_IDX cons = queue->tx.req_cons;
1053
1054         do {
1055                 if (unlikely(work_to_do-- <= 0)) {
1056                         netdev_err(queue->vif->dev, "Missing extra info\n");
1057                         xenvif_fatal_tx_err(queue->vif);
1058                         return -EBADR;
1059                 }
1060
1061                 memcpy(&extra, RING_GET_REQUEST(&queue->tx, cons),
1062                        sizeof(extra));
1063                 if (unlikely(!extra.type ||
1064                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1065                         queue->tx.req_cons = ++cons;
1066                         netdev_err(queue->vif->dev,
1067                                    "Invalid extra type: %d\n", extra.type);
1068                         xenvif_fatal_tx_err(queue->vif);
1069                         return -EINVAL;
1070                 }
1071
1072                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1073                 queue->tx.req_cons = ++cons;
1074         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1075
1076         return work_to_do;
1077 }
1078
1079 static int xenvif_set_skb_gso(struct xenvif *vif,
1080                               struct sk_buff *skb,
1081                               struct xen_netif_extra_info *gso)
1082 {
1083         if (!gso->u.gso.size) {
1084                 netdev_err(vif->dev, "GSO size must not be zero.\n");
1085                 xenvif_fatal_tx_err(vif);
1086                 return -EINVAL;
1087         }
1088
1089         switch (gso->u.gso.type) {
1090         case XEN_NETIF_GSO_TYPE_TCPV4:
1091                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1092                 break;
1093         case XEN_NETIF_GSO_TYPE_TCPV6:
1094                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1095                 break;
1096         default:
1097                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1098                 xenvif_fatal_tx_err(vif);
1099                 return -EINVAL;
1100         }
1101
1102         skb_shinfo(skb)->gso_size = gso->u.gso.size;
1103         /* gso_segs will be calculated later */
1104
1105         return 0;
1106 }
1107
1108 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1109 {
1110         bool recalculate_partial_csum = false;
1111
1112         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1113          * peers can fail to set NETRXF_csum_blank when sending a GSO
1114          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1115          * recalculate the partial checksum.
1116          */
1117         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1118                 queue->stats.rx_gso_checksum_fixup++;
1119                 skb->ip_summed = CHECKSUM_PARTIAL;
1120                 recalculate_partial_csum = true;
1121         }
1122
1123         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1124         if (skb->ip_summed != CHECKSUM_PARTIAL)
1125                 return 0;
1126
1127         return skb_checksum_setup(skb, recalculate_partial_csum);
1128 }
1129
1130 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1131 {
1132         u64 now = get_jiffies_64();
1133         u64 next_credit = queue->credit_window_start +
1134                 msecs_to_jiffies(queue->credit_usec / 1000);
1135
1136         /* Timer could already be pending in rare cases. */
1137         if (timer_pending(&queue->credit_timeout))
1138                 return true;
1139
1140         /* Passed the point where we can replenish credit? */
1141         if (time_after_eq64(now, next_credit)) {
1142                 queue->credit_window_start = now;
1143                 tx_add_credit(queue);
1144         }
1145
1146         /* Still too big to send right now? Set a callback. */
1147         if (size > queue->remaining_credit) {
1148                 queue->credit_timeout.data     =
1149                         (unsigned long)queue;
1150                 mod_timer(&queue->credit_timeout,
1151                           next_credit);
1152                 queue->credit_window_start = next_credit;
1153
1154                 return true;
1155         }
1156
1157         return false;
1158 }
1159
1160 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1161                                      int budget,
1162                                      unsigned *copy_ops,
1163                                      unsigned *map_ops)
1164 {
1165         struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
1166         struct sk_buff *skb, *nskb;
1167         int ret;
1168         unsigned int frag_overflow;
1169
1170         while (skb_queue_len(&queue->tx_queue) < budget) {
1171                 struct xen_netif_tx_request txreq;
1172                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1173                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1174                 u16 pending_idx;
1175                 RING_IDX idx;
1176                 int work_to_do;
1177                 unsigned int data_len;
1178                 pending_ring_idx_t index;
1179
1180                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1181                     XEN_NETIF_TX_RING_SIZE) {
1182                         netdev_err(queue->vif->dev,
1183                                    "Impossible number of requests. "
1184                                    "req_prod %d, req_cons %d, size %ld\n",
1185                                    queue->tx.sring->req_prod, queue->tx.req_cons,
1186                                    XEN_NETIF_TX_RING_SIZE);
1187                         xenvif_fatal_tx_err(queue->vif);
1188                         break;
1189                 }
1190
1191                 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1192                 if (!work_to_do)
1193                         break;
1194
1195                 idx = queue->tx.req_cons;
1196                 rmb(); /* Ensure that we see the request before we copy it. */
1197                 memcpy(&txreq, RING_GET_REQUEST(&queue->tx, idx), sizeof(txreq));
1198
1199                 /* Credit-based scheduling. */
1200                 if (txreq.size > queue->remaining_credit &&
1201                     tx_credit_exceeded(queue, txreq.size))
1202                         break;
1203
1204                 queue->remaining_credit -= txreq.size;
1205
1206                 work_to_do--;
1207                 queue->tx.req_cons = ++idx;
1208
1209                 memset(extras, 0, sizeof(extras));
1210                 if (txreq.flags & XEN_NETTXF_extra_info) {
1211                         work_to_do = xenvif_get_extras(queue, extras,
1212                                                        work_to_do);
1213                         idx = queue->tx.req_cons;
1214                         if (unlikely(work_to_do < 0))
1215                                 break;
1216                 }
1217
1218                 ret = xenvif_count_requests(queue, &txreq, txfrags, work_to_do);
1219                 if (unlikely(ret < 0))
1220                         break;
1221
1222                 idx += ret;
1223
1224                 if (unlikely(txreq.size < ETH_HLEN)) {
1225                         netdev_dbg(queue->vif->dev,
1226                                    "Bad packet size: %d\n", txreq.size);
1227                         xenvif_tx_err(queue, &txreq, idx);
1228                         break;
1229                 }
1230
1231                 /* No crossing a page as the payload mustn't fragment. */
1232                 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1233                         netdev_err(queue->vif->dev,
1234                                    "txreq.offset: %u, size: %u, end: %lu\n",
1235                                    txreq.offset, txreq.size,
1236                                    (unsigned long)(txreq.offset&~PAGE_MASK) + txreq.size);
1237                         xenvif_fatal_tx_err(queue->vif);
1238                         break;
1239                 }
1240
1241                 index = pending_index(queue->pending_cons);
1242                 pending_idx = queue->pending_ring[index];
1243
1244                 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
1245                             ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1246                         XEN_NETBACK_TX_COPY_LEN : txreq.size;
1247
1248                 skb = xenvif_alloc_skb(data_len);
1249                 if (unlikely(skb == NULL)) {
1250                         netdev_dbg(queue->vif->dev,
1251                                    "Can't allocate a skb in start_xmit.\n");
1252                         xenvif_tx_err(queue, &txreq, idx);
1253                         break;
1254                 }
1255
1256                 skb_shinfo(skb)->nr_frags = ret;
1257                 if (data_len < txreq.size)
1258                         skb_shinfo(skb)->nr_frags++;
1259                 /* At this point shinfo->nr_frags is in fact the number of
1260                  * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1261                  */
1262                 frag_overflow = 0;
1263                 nskb = NULL;
1264                 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1265                         frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1266                         BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1267                         skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1268                         nskb = xenvif_alloc_skb(0);
1269                         if (unlikely(nskb == NULL)) {
1270                                 kfree_skb(skb);
1271                                 xenvif_tx_err(queue, &txreq, idx);
1272                                 if (net_ratelimit())
1273                                         netdev_err(queue->vif->dev,
1274                                                    "Can't allocate the frag_list skb.\n");
1275                                 break;
1276                         }
1277                 }
1278
1279                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1280                         struct xen_netif_extra_info *gso;
1281                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1282
1283                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1284                                 /* Failure in xenvif_set_skb_gso is fatal. */
1285                                 kfree_skb(skb);
1286                                 kfree_skb(nskb);
1287                                 break;
1288                         }
1289                 }
1290
1291                 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1292
1293                 __skb_put(skb, data_len);
1294                 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1295                 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1296                 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1297
1298                 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1299                         virt_to_mfn(skb->data);
1300                 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1301                 queue->tx_copy_ops[*copy_ops].dest.offset =
1302                         offset_in_page(skb->data);
1303
1304                 queue->tx_copy_ops[*copy_ops].len = data_len;
1305                 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1306
1307                 (*copy_ops)++;
1308
1309                 if (data_len < txreq.size) {
1310                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1311                                              pending_idx);
1312                         xenvif_tx_create_map_op(queue, pending_idx, &txreq, gop);
1313                         gop++;
1314                 } else {
1315                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1316                                              INVALID_PENDING_IDX);
1317                         memcpy(&queue->pending_tx_info[pending_idx].req, &txreq,
1318                                sizeof(txreq));
1319                 }
1320
1321                 queue->pending_cons++;
1322
1323                 gop = xenvif_get_requests(queue, skb, txfrags, gop,
1324                                           frag_overflow, nskb);
1325
1326                 __skb_queue_tail(&queue->tx_queue, skb);
1327
1328                 queue->tx.req_cons = idx;
1329
1330                 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1331                     (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1332                         break;
1333         }
1334
1335         (*map_ops) = gop - queue->tx_map_ops;
1336         return;
1337 }
1338
1339 /* Consolidate skb with a frag_list into a brand new one with local pages on
1340  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1341  */
1342 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1343 {
1344         unsigned int offset = skb_headlen(skb);
1345         skb_frag_t frags[MAX_SKB_FRAGS];
1346         int i, f;
1347         struct ubuf_info *uarg;
1348         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1349
1350         queue->stats.tx_zerocopy_sent += 2;
1351         queue->stats.tx_frag_overflow++;
1352
1353         xenvif_fill_frags(queue, nskb);
1354         /* Subtract frags size, we will correct it later */
1355         skb->truesize -= skb->data_len;
1356         skb->len += nskb->len;
1357         skb->data_len += nskb->len;
1358
1359         /* create a brand new frags array and coalesce there */
1360         for (i = 0; offset < skb->len; i++) {
1361                 struct page *page;
1362                 unsigned int len;
1363
1364                 BUG_ON(i >= MAX_SKB_FRAGS);
1365                 page = alloc_page(GFP_ATOMIC);
1366                 if (!page) {
1367                         int j;
1368                         skb->truesize += skb->data_len;
1369                         for (j = 0; j < i; j++)
1370                                 put_page(frags[j].page.p);
1371                         return -ENOMEM;
1372                 }
1373
1374                 if (offset + PAGE_SIZE < skb->len)
1375                         len = PAGE_SIZE;
1376                 else
1377                         len = skb->len - offset;
1378                 if (skb_copy_bits(skb, offset, page_address(page), len))
1379                         BUG();
1380
1381                 offset += len;
1382                 frags[i].page.p = page;
1383                 frags[i].page_offset = 0;
1384                 skb_frag_size_set(&frags[i], len);
1385         }
1386
1387         /* Copied all the bits from the frag list -- free it. */
1388         skb_frag_list_init(skb);
1389         xenvif_skb_zerocopy_prepare(queue, nskb);
1390         kfree_skb(nskb);
1391
1392         /* Release all the original (foreign) frags. */
1393         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1394                 skb_frag_unref(skb, f);
1395         uarg = skb_shinfo(skb)->destructor_arg;
1396         /* increase inflight counter to offset decrement in callback */
1397         atomic_inc(&queue->inflight_packets);
1398         uarg->callback(uarg, true);
1399         skb_shinfo(skb)->destructor_arg = NULL;
1400
1401         /* Fill the skb with the new (local) frags. */
1402         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1403         skb_shinfo(skb)->nr_frags = i;
1404         skb->truesize += i * PAGE_SIZE;
1405
1406         return 0;
1407 }
1408
1409 static int xenvif_tx_submit(struct xenvif_queue *queue)
1410 {
1411         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1412         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1413         struct sk_buff *skb;
1414         int work_done = 0;
1415
1416         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1417                 struct xen_netif_tx_request *txp;
1418                 u16 pending_idx;
1419                 unsigned data_len;
1420
1421                 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1422                 txp = &queue->pending_tx_info[pending_idx].req;
1423
1424                 /* Check the remap error code. */
1425                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1426                         /* If there was an error, xenvif_tx_check_gop is
1427                          * expected to release all the frags which were mapped,
1428                          * so kfree_skb shouldn't do it again
1429                          */
1430                         skb_shinfo(skb)->nr_frags = 0;
1431                         if (skb_has_frag_list(skb)) {
1432                                 struct sk_buff *nskb =
1433                                                 skb_shinfo(skb)->frag_list;
1434                                 skb_shinfo(nskb)->nr_frags = 0;
1435                         }
1436                         kfree_skb(skb);
1437                         continue;
1438                 }
1439
1440                 data_len = skb->len;
1441                 callback_param(queue, pending_idx).ctx = NULL;
1442                 if (data_len < txp->size) {
1443                         /* Append the packet payload as a fragment. */
1444                         txp->offset += data_len;
1445                         txp->size -= data_len;
1446                 } else {
1447                         /* Schedule a response immediately. */
1448                         xenvif_idx_release(queue, pending_idx,
1449                                            XEN_NETIF_RSP_OKAY);
1450                 }
1451
1452                 if (txp->flags & XEN_NETTXF_csum_blank)
1453                         skb->ip_summed = CHECKSUM_PARTIAL;
1454                 else if (txp->flags & XEN_NETTXF_data_validated)
1455                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1456
1457                 xenvif_fill_frags(queue, skb);
1458
1459                 if (unlikely(skb_has_frag_list(skb))) {
1460                         if (xenvif_handle_frag_list(queue, skb)) {
1461                                 if (net_ratelimit())
1462                                         netdev_err(queue->vif->dev,
1463                                                    "Not enough memory to consolidate frag_list!\n");
1464                                 xenvif_skb_zerocopy_prepare(queue, skb);
1465                                 kfree_skb(skb);
1466                                 continue;
1467                         }
1468                 }
1469
1470                 skb->dev      = queue->vif->dev;
1471                 skb->protocol = eth_type_trans(skb, skb->dev);
1472                 skb_reset_network_header(skb);
1473
1474                 if (checksum_setup(queue, skb)) {
1475                         netdev_dbg(queue->vif->dev,
1476                                    "Can't setup checksum in net_tx_action\n");
1477                         /* We have to set this flag to trigger the callback */
1478                         if (skb_shinfo(skb)->destructor_arg)
1479                                 xenvif_skb_zerocopy_prepare(queue, skb);
1480                         kfree_skb(skb);
1481                         continue;
1482                 }
1483
1484                 skb_probe_transport_header(skb, 0);
1485
1486                 /* If the packet is GSO then we will have just set up the
1487                  * transport header offset in checksum_setup so it's now
1488                  * straightforward to calculate gso_segs.
1489                  */
1490                 if (skb_is_gso(skb)) {
1491                         int mss = skb_shinfo(skb)->gso_size;
1492                         int hdrlen = skb_transport_header(skb) -
1493                                 skb_mac_header(skb) +
1494                                 tcp_hdrlen(skb);
1495
1496                         skb_shinfo(skb)->gso_segs =
1497                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1498                 }
1499
1500                 queue->stats.rx_bytes += skb->len;
1501                 queue->stats.rx_packets++;
1502
1503                 work_done++;
1504
1505                 /* Set this flag right before netif_receive_skb, otherwise
1506                  * someone might think this packet already left netback, and
1507                  * do a skb_copy_ubufs while we are still in control of the
1508                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1509                  */
1510                 if (skb_shinfo(skb)->destructor_arg) {
1511                         xenvif_skb_zerocopy_prepare(queue, skb);
1512                         queue->stats.tx_zerocopy_sent++;
1513                 }
1514
1515                 netif_receive_skb(skb);
1516         }
1517
1518         return work_done;
1519 }
1520
1521 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1522 {
1523         unsigned long flags;
1524         pending_ring_idx_t index;
1525         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1526
1527         /* This is the only place where we grab this lock, to protect callbacks
1528          * from each other.
1529          */
1530         spin_lock_irqsave(&queue->callback_lock, flags);
1531         do {
1532                 u16 pending_idx = ubuf->desc;
1533                 ubuf = (struct ubuf_info *) ubuf->ctx;
1534                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1535                         MAX_PENDING_REQS);
1536                 index = pending_index(queue->dealloc_prod);
1537                 queue->dealloc_ring[index] = pending_idx;
1538                 /* Sync with xenvif_tx_dealloc_action:
1539                  * insert idx then incr producer.
1540                  */
1541                 smp_wmb();
1542                 queue->dealloc_prod++;
1543         } while (ubuf);
1544         spin_unlock_irqrestore(&queue->callback_lock, flags);
1545
1546         if (likely(zerocopy_success))
1547                 queue->stats.tx_zerocopy_success++;
1548         else
1549                 queue->stats.tx_zerocopy_fail++;
1550         xenvif_skb_zerocopy_complete(queue);
1551 }
1552
1553 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1554 {
1555         struct gnttab_unmap_grant_ref *gop;
1556         pending_ring_idx_t dc, dp;
1557         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1558         unsigned int i = 0;
1559
1560         dc = queue->dealloc_cons;
1561         gop = queue->tx_unmap_ops;
1562
1563         /* Free up any grants we have finished using */
1564         do {
1565                 dp = queue->dealloc_prod;
1566
1567                 /* Ensure we see all indices enqueued by all
1568                  * xenvif_zerocopy_callback().
1569                  */
1570                 smp_rmb();
1571
1572                 while (dc != dp) {
1573                         BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1574                         pending_idx =
1575                                 queue->dealloc_ring[pending_index(dc++)];
1576
1577                         pending_idx_release[gop - queue->tx_unmap_ops] =
1578                                 pending_idx;
1579                         queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1580                                 queue->mmap_pages[pending_idx];
1581                         gnttab_set_unmap_op(gop,
1582                                             idx_to_kaddr(queue, pending_idx),
1583                                             GNTMAP_host_map,
1584                                             queue->grant_tx_handle[pending_idx]);
1585                         xenvif_grant_handle_reset(queue, pending_idx);
1586                         ++gop;
1587                 }
1588
1589         } while (dp != queue->dealloc_prod);
1590
1591         queue->dealloc_cons = dc;
1592
1593         if (gop - queue->tx_unmap_ops > 0) {
1594                 int ret;
1595                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1596                                         NULL,
1597                                         queue->pages_to_unmap,
1598                                         gop - queue->tx_unmap_ops);
1599                 if (ret) {
1600                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1601                                    gop - queue->tx_unmap_ops, ret);
1602                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1603                                 if (gop[i].status != GNTST_okay)
1604                                         netdev_err(queue->vif->dev,
1605                                                    " host_addr: 0x%llx handle: 0x%x status: %d\n",
1606                                                    gop[i].host_addr,
1607                                                    gop[i].handle,
1608                                                    gop[i].status);
1609                         }
1610                         BUG();
1611                 }
1612         }
1613
1614         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1615                 xenvif_idx_release(queue, pending_idx_release[i],
1616                                    XEN_NETIF_RSP_OKAY);
1617 }
1618
1619
1620 /* Called after netfront has transmitted */
1621 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1622 {
1623         unsigned nr_mops, nr_cops = 0;
1624         int work_done, ret;
1625
1626         if (unlikely(!tx_work_todo(queue)))
1627                 return 0;
1628
1629         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1630
1631         if (nr_cops == 0)
1632                 return 0;
1633
1634         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1635         if (nr_mops != 0) {
1636                 ret = gnttab_map_refs(queue->tx_map_ops,
1637                                       NULL,
1638                                       queue->pages_to_map,
1639                                       nr_mops);
1640                 BUG_ON(ret);
1641         }
1642
1643         work_done = xenvif_tx_submit(queue);
1644
1645         return work_done;
1646 }
1647
1648 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1649                                u8 status)
1650 {
1651         struct pending_tx_info *pending_tx_info;
1652         pending_ring_idx_t index;
1653         unsigned long flags;
1654
1655         pending_tx_info = &queue->pending_tx_info[pending_idx];
1656
1657         spin_lock_irqsave(&queue->response_lock, flags);
1658
1659         make_tx_response(queue, &pending_tx_info->req, status);
1660
1661         /* Release the pending index before pusing the Tx response so
1662          * its available before a new Tx request is pushed by the
1663          * frontend.
1664          */
1665         index = pending_index(queue->pending_prod++);
1666         queue->pending_ring[index] = pending_idx;
1667
1668         push_tx_responses(queue);
1669
1670         spin_unlock_irqrestore(&queue->response_lock, flags);
1671 }
1672
1673
1674 static void make_tx_response(struct xenvif_queue *queue,
1675                              struct xen_netif_tx_request *txp,
1676                              s8       st)
1677 {
1678         RING_IDX i = queue->tx.rsp_prod_pvt;
1679         struct xen_netif_tx_response *resp;
1680
1681         resp = RING_GET_RESPONSE(&queue->tx, i);
1682         resp->id     = txp->id;
1683         resp->status = st;
1684
1685         if (txp->flags & XEN_NETTXF_extra_info)
1686                 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1687
1688         queue->tx.rsp_prod_pvt = ++i;
1689 }
1690
1691 static void push_tx_responses(struct xenvif_queue *queue)
1692 {
1693         int notify;
1694
1695         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1696         if (notify)
1697                 notify_remote_via_irq(queue->tx_irq);
1698 }
1699
1700 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1701                                              u16      id,
1702                                              s8       st,
1703                                              u16      offset,
1704                                              u16      size,
1705                                              u16      flags)
1706 {
1707         RING_IDX i = queue->rx.rsp_prod_pvt;
1708         struct xen_netif_rx_response *resp;
1709
1710         resp = RING_GET_RESPONSE(&queue->rx, i);
1711         resp->offset     = offset;
1712         resp->flags      = flags;
1713         resp->id         = id;
1714         resp->status     = (s16)size;
1715         if (st < 0)
1716                 resp->status = (s16)st;
1717
1718         queue->rx.rsp_prod_pvt = ++i;
1719
1720         return resp;
1721 }
1722
1723 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1724 {
1725         int ret;
1726         struct gnttab_unmap_grant_ref tx_unmap_op;
1727
1728         gnttab_set_unmap_op(&tx_unmap_op,
1729                             idx_to_kaddr(queue, pending_idx),
1730                             GNTMAP_host_map,
1731                             queue->grant_tx_handle[pending_idx]);
1732         xenvif_grant_handle_reset(queue, pending_idx);
1733
1734         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1735                                 &queue->mmap_pages[pending_idx], 1);
1736         if (ret) {
1737                 netdev_err(queue->vif->dev,
1738                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1739                            ret,
1740                            pending_idx,
1741                            tx_unmap_op.host_addr,
1742                            tx_unmap_op.handle,
1743                            tx_unmap_op.status);
1744                 BUG();
1745         }
1746 }
1747
1748 static inline int tx_work_todo(struct xenvif_queue *queue)
1749 {
1750         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1751                 return 1;
1752
1753         return 0;
1754 }
1755
1756 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1757 {
1758         return queue->dealloc_cons != queue->dealloc_prod;
1759 }
1760
1761 void xenvif_unmap_frontend_rings(struct xenvif_queue *queue)
1762 {
1763         if (queue->tx.sring)
1764                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1765                                         queue->tx.sring);
1766         if (queue->rx.sring)
1767                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1768                                         queue->rx.sring);
1769 }
1770
1771 int xenvif_map_frontend_rings(struct xenvif_queue *queue,
1772                               grant_ref_t tx_ring_ref,
1773                               grant_ref_t rx_ring_ref)
1774 {
1775         void *addr;
1776         struct xen_netif_tx_sring *txs;
1777         struct xen_netif_rx_sring *rxs;
1778
1779         int err = -ENOMEM;
1780
1781         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1782                                      &tx_ring_ref, 1, &addr);
1783         if (err)
1784                 goto err;
1785
1786         txs = (struct xen_netif_tx_sring *)addr;
1787         BACK_RING_INIT(&queue->tx, txs, PAGE_SIZE);
1788
1789         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1790                                      &rx_ring_ref, 1, &addr);
1791         if (err)
1792                 goto err;
1793
1794         rxs = (struct xen_netif_rx_sring *)addr;
1795         BACK_RING_INIT(&queue->rx, rxs, PAGE_SIZE);
1796
1797         return 0;
1798
1799 err:
1800         xenvif_unmap_frontend_rings(queue);
1801         return err;
1802 }
1803
1804 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
1805 {
1806         struct xenvif *vif = queue->vif;
1807
1808         queue->stalled = true;
1809
1810         /* At least one queue has stalled? Disable the carrier. */
1811         spin_lock(&vif->lock);
1812         if (vif->stalled_queues++ == 0) {
1813                 netdev_info(vif->dev, "Guest Rx stalled");
1814                 netif_carrier_off(vif->dev);
1815         }
1816         spin_unlock(&vif->lock);
1817 }
1818
1819 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
1820 {
1821         struct xenvif *vif = queue->vif;
1822
1823         queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
1824         queue->stalled = false;
1825
1826         /* All queues are ready? Enable the carrier. */
1827         spin_lock(&vif->lock);
1828         if (--vif->stalled_queues == 0) {
1829                 netdev_info(vif->dev, "Guest Rx ready");
1830                 netif_carrier_on(vif->dev);
1831         }
1832         spin_unlock(&vif->lock);
1833 }
1834
1835 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
1836 {
1837         RING_IDX prod, cons;
1838
1839         prod = queue->rx.sring->req_prod;
1840         cons = queue->rx.req_cons;
1841
1842         return !queue->stalled
1843                 && prod - cons < XEN_NETBK_RX_SLOTS_MAX
1844                 && time_after(jiffies,
1845                               queue->last_rx_time + queue->vif->stall_timeout);
1846 }
1847
1848 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
1849 {
1850         RING_IDX prod, cons;
1851
1852         prod = queue->rx.sring->req_prod;
1853         cons = queue->rx.req_cons;
1854
1855         return queue->stalled
1856                 && prod - cons >= XEN_NETBK_RX_SLOTS_MAX;
1857 }
1858
1859 static bool xenvif_have_rx_work(struct xenvif_queue *queue)
1860 {
1861         return (!skb_queue_empty(&queue->rx_queue)
1862                 && xenvif_rx_ring_slots_available(queue, XEN_NETBK_RX_SLOTS_MAX))
1863                 || (queue->vif->stall_timeout &&
1864                     (xenvif_rx_queue_stalled(queue)
1865                      || xenvif_rx_queue_ready(queue)))
1866                 || kthread_should_stop()
1867                 || queue->vif->disabled;
1868 }
1869
1870 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
1871 {
1872         struct sk_buff *skb;
1873         long timeout;
1874
1875         skb = skb_peek(&queue->rx_queue);
1876         if (!skb)
1877                 return MAX_SCHEDULE_TIMEOUT;
1878
1879         timeout = XENVIF_RX_CB(skb)->expires - jiffies;
1880         return timeout < 0 ? 0 : timeout;
1881 }
1882
1883 /* Wait until the guest Rx thread has work.
1884  *
1885  * The timeout needs to be adjusted based on the current head of the
1886  * queue (and not just the head at the beginning).  In particular, if
1887  * the queue is initially empty an infinite timeout is used and this
1888  * needs to be reduced when a skb is queued.
1889  *
1890  * This cannot be done with wait_event_timeout() because it only
1891  * calculates the timeout once.
1892  */
1893 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
1894 {
1895         DEFINE_WAIT(wait);
1896
1897         if (xenvif_have_rx_work(queue))
1898                 return;
1899
1900         for (;;) {
1901                 long ret;
1902
1903                 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
1904                 if (xenvif_have_rx_work(queue))
1905                         break;
1906                 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
1907                 if (!ret)
1908                         break;
1909         }
1910         finish_wait(&queue->wq, &wait);
1911 }
1912
1913 int xenvif_kthread_guest_rx(void *data)
1914 {
1915         struct xenvif_queue *queue = data;
1916         struct xenvif *vif = queue->vif;
1917
1918         if (!vif->stall_timeout)
1919                 xenvif_queue_carrier_on(queue);
1920
1921         for (;;) {
1922                 xenvif_wait_for_rx_work(queue);
1923
1924                 if (kthread_should_stop())
1925                         break;
1926
1927                 /* This frontend is found to be rogue, disable it in
1928                  * kthread context. Currently this is only set when
1929                  * netback finds out frontend sends malformed packet,
1930                  * but we cannot disable the interface in softirq
1931                  * context so we defer it here, if this thread is
1932                  * associated with queue 0.
1933                  */
1934                 if (unlikely(vif->disabled && queue->id == 0)) {
1935                         xenvif_carrier_off(vif);
1936                         break;
1937                 }
1938
1939                 if (!skb_queue_empty(&queue->rx_queue))
1940                         xenvif_rx_action(queue);
1941
1942                 /* If the guest hasn't provided any Rx slots for a
1943                  * while it's probably not responsive, drop the
1944                  * carrier so packets are dropped earlier.
1945                  */
1946                 if (vif->stall_timeout) {
1947                         if (xenvif_rx_queue_stalled(queue))
1948                                 xenvif_queue_carrier_off(queue);
1949                         else if (xenvif_rx_queue_ready(queue))
1950                                 xenvif_queue_carrier_on(queue);
1951                 }
1952
1953                 /* Queued packets may have foreign pages from other
1954                  * domains.  These cannot be queued indefinitely as
1955                  * this would starve guests of grant refs and transmit
1956                  * slots.
1957                  */
1958                 xenvif_rx_queue_drop_expired(queue);
1959
1960                 xenvif_rx_queue_maybe_wake(queue);
1961
1962                 cond_resched();
1963         }
1964
1965         /* Bin any remaining skbs */
1966         xenvif_rx_queue_purge(queue);
1967
1968         return 0;
1969 }
1970
1971 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1972 {
1973         /* Dealloc thread must remain running until all inflight
1974          * packets complete.
1975          */
1976         return kthread_should_stop() &&
1977                 !atomic_read(&queue->inflight_packets);
1978 }
1979
1980 int xenvif_dealloc_kthread(void *data)
1981 {
1982         struct xenvif_queue *queue = data;
1983
1984         for (;;) {
1985                 wait_event_interruptible(queue->dealloc_wq,
1986                                          tx_dealloc_work_todo(queue) ||
1987                                          xenvif_dealloc_kthread_should_stop(queue));
1988                 if (xenvif_dealloc_kthread_should_stop(queue))
1989                         break;
1990
1991                 xenvif_tx_dealloc_action(queue);
1992                 cond_resched();
1993         }
1994
1995         /* Unmap anything remaining*/
1996         if (tx_dealloc_work_todo(queue))
1997                 xenvif_tx_dealloc_action(queue);
1998
1999         return 0;
2000 }
2001
2002 static int __init netback_init(void)
2003 {
2004         int rc = 0;
2005
2006         if (!xen_domain())
2007                 return -ENODEV;
2008
2009         /* Allow as many queues as there are CPUs, by default */
2010         xenvif_max_queues = num_online_cpus();
2011
2012         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2013                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2014                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2015                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2016         }
2017
2018         rc = xenvif_xenbus_init();
2019         if (rc)
2020                 goto failed_init;
2021
2022 #ifdef CONFIG_DEBUG_FS
2023         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2024         if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2025                 pr_warn("Init of debugfs returned %ld!\n",
2026                         PTR_ERR(xen_netback_dbg_root));
2027 #endif /* CONFIG_DEBUG_FS */
2028
2029         return 0;
2030
2031 failed_init:
2032         return rc;
2033 }
2034
2035 module_init(netback_init);
2036
2037 static void __exit netback_fini(void)
2038 {
2039 #ifdef CONFIG_DEBUG_FS
2040         if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2041                 debugfs_remove_recursive(xen_netback_dbg_root);
2042 #endif /* CONFIG_DEBUG_FS */
2043         xenvif_xenbus_fini();
2044 }
2045 module_exit(netback_fini);
2046
2047 MODULE_LICENSE("Dual BSD/GPL");
2048 MODULE_ALIAS("xen-backend:vif");