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
  48 #include <asm/xen/hypercall.h>
  49 #include <asm/xen/page.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 = 1;
  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                 RING_IDX old_req_cons;
 519                 RING_IDX ring_slots_used;
 520
 521                 queue->last_rx_time = jiffies;
 522
 523                 old_req_cons = queue->rx.req_cons;
 524                 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
 525                 ring_slots_used = queue->rx.req_cons - old_req_cons;
 526
 527                 __skb_queue_tail(&rxq, skb);
 528         }
 529
 530         BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
 531
 532         if (!npo.copy_prod)
 533                 goto done;
 534
 535         BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
 536         gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
 537
 538         while ((skb = __skb_dequeue(&rxq)) != NULL) {
 539
 540                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 541                     queue->vif->gso_prefix_mask) {
 542                         resp = RING_GET_RESPONSE(&queue->rx,
 543                                                  queue->rx.rsp_prod_pvt++);
 544
 545                         resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
 546
 547                         resp->offset = queue->meta[npo.meta_cons].gso_size;
 548                         resp->id = queue->meta[npo.meta_cons].id;
 549                         resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
 550
 551                         npo.meta_cons++;
 552                         XENVIF_RX_CB(skb)->meta_slots_used--;
 553                 }
 554
 555
 556                 queue->stats.tx_bytes += skb->len;
 557                 queue->stats.tx_packets++;
 558
 559                 status = xenvif_check_gop(queue->vif,
 560                                           XENVIF_RX_CB(skb)->meta_slots_used,
 561                                           &npo);
 562
 563                 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
 564                         flags = 0;
 565                 else
 566                         flags = XEN_NETRXF_more_data;
 567
 568                 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
 569                         flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
 570                 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 571                         /* remote but checksummed. */
 572                         flags |= XEN_NETRXF_data_validated;
 573
 574                 offset = 0;
 575                 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
 576                                         status, offset,
 577                                         queue->meta[npo.meta_cons].size,
 578                                         flags);
 579
 580                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 581                     queue->vif->gso_mask) {
 582                         struct xen_netif_extra_info *gso =
 583                                 (struct xen_netif_extra_info *)
 584                                 RING_GET_RESPONSE(&queue->rx,
 585                                                   queue->rx.rsp_prod_pvt++);
 586
 587                         resp->flags |= XEN_NETRXF_extra_info;
 588
 589                         gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
 590                         gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
 591                         gso->u.gso.pad = 0;
 592                         gso->u.gso.features = 0;
 593
 594                         gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
 595                         gso->flags = 0;
 596                 }
 597
 598                 xenvif_add_frag_responses(queue, status,
 599                                           queue->meta + npo.meta_cons + 1,
 600                                           XENVIF_RX_CB(skb)->meta_slots_used);
 601
 602                 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
 603
 604                 need_to_notify |= !!ret;
 605
 606                 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
 607                 dev_kfree_skb(skb);
 608         }
 609
 610 done:
 611         if (need_to_notify)
 612                 notify_remote_via_irq(queue->rx_irq);
 613 }
 614
 615 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 616 {
 617         int more_to_do;
 618
 619         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 620
 621         if (more_to_do)
 622                 napi_schedule(&queue->napi);
 623 }
 624
 625 static void tx_add_credit(struct xenvif_queue *queue)
 626 {
 627         unsigned long max_burst, max_credit;
 628
 629         /*
 630          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 631          * Otherwise the interface can seize up due to insufficient credit.
 632          */
 633         max_burst = RING_GET_REQUEST(&queue->tx, queue->tx.req_cons)->size;
 634         max_burst = min(max_burst, 131072UL);
 635         max_burst = max(max_burst, queue->credit_bytes);
 636
 637         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 638         max_credit = queue->remaining_credit + queue->credit_bytes;
 639         if (max_credit < queue->remaining_credit)
 640                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 641
 642         queue->remaining_credit = min(max_credit, max_burst);
 643 }
 644
 645 static void tx_credit_callback(unsigned long data)
 646 {
 647         struct xenvif_queue *queue = (struct xenvif_queue *)data;
 648         tx_add_credit(queue);
 649         xenvif_napi_schedule_or_enable_events(queue);
 650 }
 651
 652 static void xenvif_tx_err(struct xenvif_queue *queue,
 653                           struct xen_netif_tx_request *txp, RING_IDX end)
 654 {
 655         RING_IDX cons = queue->tx.req_cons;
 656         unsigned long flags;
 657
 658         do {
 659                 spin_lock_irqsave(&queue->response_lock, flags);
 660                 make_tx_response(queue, txp, XEN_NETIF_RSP_ERROR);
 661                 push_tx_responses(queue);
 662                 spin_unlock_irqrestore(&queue->response_lock, flags);
 663                 if (cons == end)
 664                         break;
 665                 txp = RING_GET_REQUEST(&queue->tx, cons++);
 666         } while (1);
 667         queue->tx.req_cons = cons;
 668 }
 669
 670 static void xenvif_fatal_tx_err(struct xenvif *vif)
 671 {
 672         netdev_err(vif->dev, "fatal error; disabling device\n");
 673         vif->disabled = true;
 674         /* Disable the vif from queue 0's kthread */
 675         if (vif->queues)
 676                 xenvif_kick_thread(&vif->queues[0]);
 677 }
 678
 679 static int xenvif_count_requests(struct xenvif_queue *queue,
 680                                  struct xen_netif_tx_request *first,
 681                                  struct xen_netif_tx_request *txp,
 682                                  int work_to_do)
 683 {
 684         RING_IDX cons = queue->tx.req_cons;
 685         int slots = 0;
 686         int drop_err = 0;
 687         int more_data;
 688
 689         if (!(first->flags & XEN_NETTXF_more_data))
 690                 return 0;
 691
 692         do {
 693                 struct xen_netif_tx_request dropped_tx = { 0 };
 694
 695                 if (slots >= work_to_do) {
 696                         netdev_err(queue->vif->dev,
 697                                    "Asked for %d slots but exceeds this limit\n",
 698                                    work_to_do);
 699                         xenvif_fatal_tx_err(queue->vif);
 700                         return -ENODATA;
 701                 }
 702
 703                 /* This guest is really using too many slots and
 704                  * considered malicious.
 705                  */
 706                 if (unlikely(slots >= fatal_skb_slots)) {
 707                         netdev_err(queue->vif->dev,
 708                                    "Malicious frontend using %d slots, threshold %u\n",
 709                                    slots, fatal_skb_slots);
 710                         xenvif_fatal_tx_err(queue->vif);
 711                         return -E2BIG;
 712                 }
 713
 714                 /* Xen network protocol had implicit dependency on
 715                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 716                  * the historical MAX_SKB_FRAGS value 18 to honor the
 717                  * same behavior as before. Any packet using more than
 718                  * 18 slots but less than fatal_skb_slots slots is
 719                  * dropped
 720                  */
 721                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 722                         if (net_ratelimit())
 723                                 netdev_dbg(queue->vif->dev,
 724                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 725                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 726                         drop_err = -E2BIG;
 727                 }
 728
 729                 if (drop_err)
 730                         txp = &dropped_tx;
 731
 732                 memcpy(txp, RING_GET_REQUEST(&queue->tx, cons + slots),
 733                        sizeof(*txp));
 734
 735                 /* If the guest submitted a frame >= 64 KiB then
 736                  * first->size overflowed and following slots will
 737                  * appear to be larger than the frame.
 738                  *
 739                  * This cannot be fatal error as there are buggy
 740                  * frontends that do this.
 741                  *
 742                  * Consume all slots and drop the packet.
 743                  */
 744                 if (!drop_err && txp->size > first->size) {
 745                         if (net_ratelimit())
 746                                 netdev_dbg(queue->vif->dev,
 747                                            "Invalid tx request, slot size %u > remaining size %u\n",
 748                                            txp->size, first->size);
 749                         drop_err = -EIO;
 750                 }
 751
 752                 first->size -= txp->size;
 753                 slots++;
 754
 755                 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
 756                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
 757                                  txp->offset, txp->size);
 758                         xenvif_fatal_tx_err(queue->vif);
 759                         return -EINVAL;
 760                 }
 761
 762                 more_data = txp->flags & XEN_NETTXF_more_data;
 763
 764                 if (!drop_err)
 765                         txp++;
 766
 767         } while (more_data);
 768
 769         if (drop_err) {
 770                 xenvif_tx_err(queue, first, cons + slots);
 771                 return drop_err;
 772         }
 773
 774         return slots;
 775 }
 776
 777
 778 struct xenvif_tx_cb {
 779         u16 pending_idx;
 780 };
 781
 782 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 783
 784 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 785                                           u16 pending_idx,
 786                                           struct xen_netif_tx_request *txp,
 787                                           struct gnttab_map_grant_ref *mop)
 788 {
 789         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 790         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 791                           GNTMAP_host_map | GNTMAP_readonly,
 792                           txp->gref, queue->vif->domid);
 793
 794         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 795                sizeof(*txp));
 796 }
 797
 798 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 799 {
 800         struct sk_buff *skb =
 801                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 802                           GFP_ATOMIC | __GFP_NOWARN);
 803         if (unlikely(skb == NULL))
 804                 return NULL;
 805
 806         /* Packets passed to netif_rx() must have some headroom. */
 807         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 808
 809         /* Initialize it here to avoid later surprises */
 810         skb_shinfo(skb)->destructor_arg = NULL;
 811
 812         return skb;
 813 }
 814
 815 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
 816                                                         struct sk_buff *skb,
 817                                                         struct xen_netif_tx_request *txp,
 818                                                         struct gnttab_map_grant_ref *gop)
 819 {
 820         struct skb_shared_info *shinfo = skb_shinfo(skb);
 821         skb_frag_t *frags = shinfo->frags;
 822         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 823         int start;
 824         pending_ring_idx_t index;
 825         unsigned int nr_slots, frag_overflow = 0;
 826
 827         /* At this point shinfo->nr_frags is in fact the number of
 828          * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
 829          */
 830         if (shinfo->nr_frags > MAX_SKB_FRAGS) {
 831                 frag_overflow = shinfo->nr_frags - MAX_SKB_FRAGS;
 832                 BUG_ON(frag_overflow > MAX_SKB_FRAGS);
 833                 shinfo->nr_frags = MAX_SKB_FRAGS;
 834         }
 835         nr_slots = shinfo->nr_frags;
 836
 837         /* Skip first skb fragment if it is on same page as header fragment. */
 838         start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
 839
 840         for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
 841              shinfo->nr_frags++, txp++, gop++) {
 842                 index = pending_index(queue->pending_cons++);
 843                 pending_idx = queue->pending_ring[index];
 844                 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
 845                 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 846         }
 847
 848         if (frag_overflow) {
 849                 struct sk_buff *nskb = xenvif_alloc_skb(0);
 850                 if (unlikely(nskb == NULL)) {
 851                         if (net_ratelimit())
 852                                 netdev_err(queue->vif->dev,
 853                                            "Can't allocate the frag_list skb.\n");
 854                         return NULL;
 855                 }
 856
 857                 shinfo = skb_shinfo(nskb);
 858                 frags = shinfo->frags;
 859
 860                 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
 861                      shinfo->nr_frags++, txp++, gop++) {
 862                         index = pending_index(queue->pending_cons++);
 863                         pending_idx = queue->pending_ring[index];
 864                         xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
 865                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 866                                              pending_idx);
 867                 }
 868
 869                 skb_shinfo(skb)->frag_list = nskb;
 870         }
 871
 872         return gop;
 873 }
 874
 875 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 876                                            u16 pending_idx,
 877                                            grant_handle_t handle)
 878 {
 879         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 880                      NETBACK_INVALID_HANDLE)) {
 881                 netdev_err(queue->vif->dev,
 882                            "Trying to overwrite active handle! pending_idx: %x\n",
 883                            pending_idx);
 884                 BUG();
 885         }
 886         queue->grant_tx_handle[pending_idx] = handle;
 887 }
 888
 889 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 890                                              u16 pending_idx)
 891 {
 892         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 893                      NETBACK_INVALID_HANDLE)) {
 894                 netdev_err(queue->vif->dev,
 895                            "Trying to unmap invalid handle! pending_idx: %x\n",
 896                            pending_idx);
 897                 BUG();
 898         }
 899         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 900 }
 901
 902 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 903                                struct sk_buff *skb,
 904                                struct gnttab_map_grant_ref **gopp_map,
 905                                struct gnttab_copy **gopp_copy)
 906 {
 907         struct gnttab_map_grant_ref *gop_map = *gopp_map;
 908         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 909         /* This always points to the shinfo of the skb being checked, which
 910          * could be either the first or the one on the frag_list
 911          */
 912         struct skb_shared_info *shinfo = skb_shinfo(skb);
 913         /* If this is non-NULL, we are currently checking the frag_list skb, and
 914          * this points to the shinfo of the first one
 915          */
 916         struct skb_shared_info *first_shinfo = NULL;
 917         int nr_frags = shinfo->nr_frags;
 918         const bool sharedslot = nr_frags &&
 919                                 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
 920         int i, err;
 921
 922         /* Check status of header. */
 923         err = (*gopp_copy)->status;
 924         if (unlikely(err)) {
 925                 if (net_ratelimit())
 926                         netdev_dbg(queue->vif->dev,
 927                                    "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 928                                    (*gopp_copy)->status,
 929                                    pending_idx,
 930                                    (*gopp_copy)->source.u.ref);
 931                 /* The first frag might still have this slot mapped */
 932                 if (!sharedslot)
 933                         xenvif_idx_release(queue, pending_idx,
 934                                            XEN_NETIF_RSP_ERROR);
 935         }
 936         (*gopp_copy)++;
 937
 938 check_frags:
 939         for (i = 0; i < nr_frags; i++, gop_map++) {
 940                 int j, newerr;
 941
 942                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 943
 944                 /* Check error status: if okay then remember grant handle. */
 945                 newerr = gop_map->status;
 946
 947                 if (likely(!newerr)) {
 948                         xenvif_grant_handle_set(queue,
 949                                                 pending_idx,
 950                                                 gop_map->handle);
 951                         /* Had a previous error? Invalidate this fragment. */
 952                         if (unlikely(err)) {
 953                                 xenvif_idx_unmap(queue, pending_idx);
 954                                 /* If the mapping of the first frag was OK, but
 955                                  * the header's copy failed, and they are
 956                                  * sharing a slot, send an error
 957                                  */
 958                                 if (i == 0 && sharedslot)
 959                                         xenvif_idx_release(queue, pending_idx,
 960                                                            XEN_NETIF_RSP_ERROR);
 961                                 else
 962                                         xenvif_idx_release(queue, pending_idx,
 963                                                            XEN_NETIF_RSP_OKAY);
 964                         }
 965                         continue;
 966                 }
 967
 968                 /* Error on this fragment: respond to client with an error. */
 969                 if (net_ratelimit())
 970                         netdev_dbg(queue->vif->dev,
 971                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
 972                                    i,
 973                                    gop_map->status,
 974                                    pending_idx,
 975                                    gop_map->ref);
 976
 977                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
 978
 979                 /* Not the first error? Preceding frags already invalidated. */
 980                 if (err)
 981                         continue;
 982
 983                 /* First error: if the header haven't shared a slot with the
 984                  * first frag, release it as well.
 985                  */
 986                 if (!sharedslot)
 987                         xenvif_idx_release(queue,
 988                                            XENVIF_TX_CB(skb)->pending_idx,
 989                                            XEN_NETIF_RSP_OKAY);
 990
 991                 /* Invalidate preceding fragments of this skb. */
 992                 for (j = 0; j < i; j++) {
 993                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
 994                         xenvif_idx_unmap(queue, pending_idx);
 995                         xenvif_idx_release(queue, pending_idx,
 996                                            XEN_NETIF_RSP_OKAY);
 997                 }
 998
 999                 /* And if we found the error while checking the frag_list, unmap
1000                  * the first skb's frags
1001                  */
1002                 if (first_shinfo) {
1003                         for (j = 0; j < first_shinfo->nr_frags; j++) {
1004                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
1005                                 xenvif_idx_unmap(queue, pending_idx);
1006                                 xenvif_idx_release(queue, pending_idx,
1007                                                    XEN_NETIF_RSP_OKAY);
1008                         }
1009                 }
1010
1011                 /* Remember the error: invalidate all subsequent fragments. */
1012                 err = newerr;
1013         }
1014
1015         if (skb_has_frag_list(skb) && !first_shinfo) {
1016                 first_shinfo = skb_shinfo(skb);
1017                 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1018                 nr_frags = shinfo->nr_frags;
1019
1020                 goto check_frags;
1021         }
1022
1023         *gopp_map = gop_map;
1024         return err;
1025 }
1026
1027 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1028 {
1029         struct skb_shared_info *shinfo = skb_shinfo(skb);
1030         int nr_frags = shinfo->nr_frags;
1031         int i;
1032         u16 prev_pending_idx = INVALID_PENDING_IDX;
1033
1034         for (i = 0; i < nr_frags; i++) {
1035                 skb_frag_t *frag = shinfo->frags + i;
1036                 struct xen_netif_tx_request *txp;
1037                 struct page *page;
1038                 u16 pending_idx;
1039
1040                 pending_idx = frag_get_pending_idx(frag);
1041
1042                 /* If this is not the first frag, chain it to the previous*/
1043                 if (prev_pending_idx == INVALID_PENDING_IDX)
1044                         skb_shinfo(skb)->destructor_arg =
1045                                 &callback_param(queue, pending_idx);
1046                 else
1047                         callback_param(queue, prev_pending_idx).ctx =
1048                                 &callback_param(queue, pending_idx);
1049
1050                 callback_param(queue, pending_idx).ctx = NULL;
1051                 prev_pending_idx = pending_idx;
1052
1053                 txp = &queue->pending_tx_info[pending_idx].req;
1054                 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1055                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1056                 skb->len += txp->size;
1057                 skb->data_len += txp->size;
1058                 skb->truesize += txp->size;
1059
1060                 /* Take an extra reference to offset network stack's put_page */
1061                 get_page(queue->mmap_pages[pending_idx]);
1062         }
1063 }
1064
1065 static int xenvif_get_extras(struct xenvif_queue *queue,
1066                                 struct xen_netif_extra_info *extras,
1067                                 int work_to_do)
1068 {
1069         struct xen_netif_extra_info extra;
1070         RING_IDX cons = queue->tx.req_cons;
1071
1072         do {
1073                 if (unlikely(work_to_do-- <= 0)) {
1074                         netdev_err(queue->vif->dev, "Missing extra info\n");
1075                         xenvif_fatal_tx_err(queue->vif);
1076                         return -EBADR;
1077                 }
1078
1079                 memcpy(&extra, RING_GET_REQUEST(&queue->tx, cons),
1080                        sizeof(extra));
1081                 if (unlikely(!extra.type ||
1082                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1083                         queue->tx.req_cons = ++cons;
1084                         netdev_err(queue->vif->dev,
1085                                    "Invalid extra type: %d\n", extra.type);
1086                         xenvif_fatal_tx_err(queue->vif);
1087                         return -EINVAL;
1088                 }
1089
1090                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1091                 queue->tx.req_cons = ++cons;
1092         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1093
1094         return work_to_do;
1095 }
1096
1097 static int xenvif_set_skb_gso(struct xenvif *vif,
1098                               struct sk_buff *skb,
1099                               struct xen_netif_extra_info *gso)
1100 {
1101         if (!gso->u.gso.size) {
1102                 netdev_err(vif->dev, "GSO size must not be zero.\n");
1103                 xenvif_fatal_tx_err(vif);
1104                 return -EINVAL;
1105         }
1106
1107         switch (gso->u.gso.type) {
1108         case XEN_NETIF_GSO_TYPE_TCPV4:
1109                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1110                 break;
1111         case XEN_NETIF_GSO_TYPE_TCPV6:
1112                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1113                 break;
1114         default:
1115                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1116                 xenvif_fatal_tx_err(vif);
1117                 return -EINVAL;
1118         }
1119
1120         skb_shinfo(skb)->gso_size = gso->u.gso.size;
1121         /* gso_segs will be calculated later */
1122
1123         return 0;
1124 }
1125
1126 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1127 {
1128         bool recalculate_partial_csum = false;
1129
1130         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1131          * peers can fail to set NETRXF_csum_blank when sending a GSO
1132          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1133          * recalculate the partial checksum.
1134          */
1135         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1136                 queue->stats.rx_gso_checksum_fixup++;
1137                 skb->ip_summed = CHECKSUM_PARTIAL;
1138                 recalculate_partial_csum = true;
1139         }
1140
1141         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1142         if (skb->ip_summed != CHECKSUM_PARTIAL)
1143                 return 0;
1144
1145         return skb_checksum_setup(skb, recalculate_partial_csum);
1146 }
1147
1148 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1149 {
1150         u64 now = get_jiffies_64();
1151         u64 next_credit = queue->credit_window_start +
1152                 msecs_to_jiffies(queue->credit_usec / 1000);
1153
1154         /* Timer could already be pending in rare cases. */
1155         if (timer_pending(&queue->credit_timeout))
1156                 return true;
1157
1158         /* Passed the point where we can replenish credit? */
1159         if (time_after_eq64(now, next_credit)) {
1160                 queue->credit_window_start = now;
1161                 tx_add_credit(queue);
1162         }
1163
1164         /* Still too big to send right now? Set a callback. */
1165         if (size > queue->remaining_credit) {
1166                 queue->credit_timeout.data     =
1167                         (unsigned long)queue;
1168                 queue->credit_timeout.function =
1169                         tx_credit_callback;
1170                 mod_timer(&queue->credit_timeout,
1171                           next_credit);
1172                 queue->credit_window_start = next_credit;
1173
1174                 return true;
1175         }
1176
1177         return false;
1178 }
1179
1180 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1181                                      int budget,
1182                                      unsigned *copy_ops,
1183                                      unsigned *map_ops)
1184 {
1185         struct gnttab_map_grant_ref *gop = queue->tx_map_ops, *request_gop;
1186         struct sk_buff *skb;
1187         int ret;
1188
1189         while (skb_queue_len(&queue->tx_queue) < budget) {
1190                 struct xen_netif_tx_request txreq;
1191                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1192                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1193                 u16 pending_idx;
1194                 RING_IDX idx;
1195                 int work_to_do;
1196                 unsigned int data_len;
1197                 pending_ring_idx_t index;
1198
1199                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1200                     XEN_NETIF_TX_RING_SIZE) {
1201                         netdev_err(queue->vif->dev,
1202                                    "Impossible number of requests. "
1203                                    "req_prod %d, req_cons %d, size %ld\n",
1204                                    queue->tx.sring->req_prod, queue->tx.req_cons,
1205                                    XEN_NETIF_TX_RING_SIZE);
1206                         xenvif_fatal_tx_err(queue->vif);
1207                         break;
1208                 }
1209
1210                 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1211                 if (!work_to_do)
1212                         break;
1213
1214                 idx = queue->tx.req_cons;
1215                 rmb(); /* Ensure that we see the request before we copy it. */
1216                 memcpy(&txreq, RING_GET_REQUEST(&queue->tx, idx), sizeof(txreq));
1217
1218                 /* Credit-based scheduling. */
1219                 if (txreq.size > queue->remaining_credit &&
1220                     tx_credit_exceeded(queue, txreq.size))
1221                         break;
1222
1223                 queue->remaining_credit -= txreq.size;
1224
1225                 work_to_do--;
1226                 queue->tx.req_cons = ++idx;
1227
1228                 memset(extras, 0, sizeof(extras));
1229                 if (txreq.flags & XEN_NETTXF_extra_info) {
1230                         work_to_do = xenvif_get_extras(queue, extras,
1231                                                        work_to_do);
1232                         idx = queue->tx.req_cons;
1233                         if (unlikely(work_to_do < 0))
1234                                 break;
1235                 }
1236
1237                 ret = xenvif_count_requests(queue, &txreq, txfrags, work_to_do);
1238                 if (unlikely(ret < 0))
1239                         break;
1240
1241                 idx += ret;
1242
1243                 if (unlikely(txreq.size < ETH_HLEN)) {
1244                         netdev_dbg(queue->vif->dev,
1245                                    "Bad packet size: %d\n", txreq.size);
1246                         xenvif_tx_err(queue, &txreq, idx);
1247                         break;
1248                 }
1249
1250                 /* No crossing a page as the payload mustn't fragment. */
1251                 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1252                         netdev_err(queue->vif->dev,
1253                                    "txreq.offset: %x, size: %u, end: %lu\n",
1254                                    txreq.offset, txreq.size,
1255                                    (txreq.offset&~PAGE_MASK) + txreq.size);
1256                         xenvif_fatal_tx_err(queue->vif);
1257                         break;
1258                 }
1259
1260                 index = pending_index(queue->pending_cons);
1261                 pending_idx = queue->pending_ring[index];
1262
1263                 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
1264                             ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1265                         XEN_NETBACK_TX_COPY_LEN : txreq.size;
1266
1267                 skb = xenvif_alloc_skb(data_len);
1268                 if (unlikely(skb == NULL)) {
1269                         netdev_dbg(queue->vif->dev,
1270                                    "Can't allocate a skb in start_xmit.\n");
1271                         xenvif_tx_err(queue, &txreq, idx);
1272                         break;
1273                 }
1274
1275                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1276                         struct xen_netif_extra_info *gso;
1277                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1278
1279                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1280                                 /* Failure in xenvif_set_skb_gso is fatal. */
1281                                 kfree_skb(skb);
1282                                 break;
1283                         }
1284                 }
1285
1286                 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1287
1288                 __skb_put(skb, data_len);
1289                 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1290                 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1291                 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1292
1293                 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1294                         virt_to_mfn(skb->data);
1295                 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1296                 queue->tx_copy_ops[*copy_ops].dest.offset =
1297                         offset_in_page(skb->data);
1298
1299                 queue->tx_copy_ops[*copy_ops].len = data_len;
1300                 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1301
1302                 (*copy_ops)++;
1303
1304                 skb_shinfo(skb)->nr_frags = ret;
1305                 if (data_len < txreq.size) {
1306                         skb_shinfo(skb)->nr_frags++;
1307                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1308                                              pending_idx);
1309                         xenvif_tx_create_map_op(queue, pending_idx, &txreq, gop);
1310                         gop++;
1311                 } else {
1312                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1313                                              INVALID_PENDING_IDX);
1314                         memcpy(&queue->pending_tx_info[pending_idx].req, &txreq,
1315                                sizeof(txreq));
1316                 }
1317
1318                 queue->pending_cons++;
1319
1320                 request_gop = xenvif_get_requests(queue, skb, txfrags, gop);
1321                 if (request_gop == NULL) {
1322                         kfree_skb(skb);
1323                         xenvif_tx_err(queue, &txreq, idx);
1324                         break;
1325                 }
1326                 gop = request_gop;
1327
1328                 __skb_queue_tail(&queue->tx_queue, skb);
1329
1330                 queue->tx.req_cons = idx;
1331
1332                 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1333                     (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1334                         break;
1335         }
1336
1337         (*map_ops) = gop - queue->tx_map_ops;
1338         return;
1339 }
1340
1341 /* Consolidate skb with a frag_list into a brand new one with local pages on
1342  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1343  */
1344 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1345 {
1346         unsigned int offset = skb_headlen(skb);
1347         skb_frag_t frags[MAX_SKB_FRAGS];
1348         int i, f;
1349         struct ubuf_info *uarg;
1350         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1351
1352         queue->stats.tx_zerocopy_sent += 2;
1353         queue->stats.tx_frag_overflow++;
1354
1355         xenvif_fill_frags(queue, nskb);
1356         /* Subtract frags size, we will correct it later */
1357         skb->truesize -= skb->data_len;
1358         skb->len += nskb->len;
1359         skb->data_len += nskb->len;
1360
1361         /* create a brand new frags array and coalesce there */
1362         for (i = 0; offset < skb->len; i++) {
1363                 struct page *page;
1364                 unsigned int len;
1365
1366                 BUG_ON(i >= MAX_SKB_FRAGS);
1367                 page = alloc_page(GFP_ATOMIC);
1368                 if (!page) {
1369                         int j;
1370                         skb->truesize += skb->data_len;
1371                         for (j = 0; j < i; j++)
1372                                 put_page(frags[j].page.p);
1373                         return -ENOMEM;
1374                 }
1375
1376                 if (offset + PAGE_SIZE < skb->len)
1377                         len = PAGE_SIZE;
1378                 else
1379                         len = skb->len - offset;
1380                 if (skb_copy_bits(skb, offset, page_address(page), len))
1381                         BUG();
1382
1383                 offset += len;
1384                 frags[i].page.p = page;
1385                 frags[i].page_offset = 0;
1386                 skb_frag_size_set(&frags[i], len);
1387         }
1388
1389         /* Copied all the bits from the frag list -- free it. */
1390         skb_frag_list_init(skb);
1391         xenvif_skb_zerocopy_prepare(queue, nskb);
1392         kfree_skb(nskb);
1393
1394         /* Release all the original (foreign) frags. */
1395         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1396                 skb_frag_unref(skb, f);
1397         uarg = skb_shinfo(skb)->destructor_arg;
1398         /* increase inflight counter to offset decrement in callback */
1399         atomic_inc(&queue->inflight_packets);
1400         uarg->callback(uarg, true);
1401         skb_shinfo(skb)->destructor_arg = NULL;
1402
1403         /* Fill the skb with the new (local) frags. */
1404         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1405         skb_shinfo(skb)->nr_frags = i;
1406         skb->truesize += i * PAGE_SIZE;
1407
1408         return 0;
1409 }
1410
1411 static int xenvif_tx_submit(struct xenvif_queue *queue)
1412 {
1413         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1414         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1415         struct sk_buff *skb;
1416         int work_done = 0;
1417
1418         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1419                 struct xen_netif_tx_request *txp;
1420                 u16 pending_idx;
1421                 unsigned data_len;
1422
1423                 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1424                 txp = &queue->pending_tx_info[pending_idx].req;
1425
1426                 /* Check the remap error code. */
1427                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1428                         /* If there was an error, xenvif_tx_check_gop is
1429                          * expected to release all the frags which were mapped,
1430                          * so kfree_skb shouldn't do it again
1431                          */
1432                         skb_shinfo(skb)->nr_frags = 0;
1433                         if (skb_has_frag_list(skb)) {
1434                                 struct sk_buff *nskb =
1435                                                 skb_shinfo(skb)->frag_list;
1436                                 skb_shinfo(nskb)->nr_frags = 0;
1437                         }
1438                         kfree_skb(skb);
1439                         continue;
1440                 }
1441
1442                 data_len = skb->len;
1443                 callback_param(queue, pending_idx).ctx = NULL;
1444                 if (data_len < txp->size) {
1445                         /* Append the packet payload as a fragment. */
1446                         txp->offset += data_len;
1447                         txp->size -= data_len;
1448                 } else {
1449                         /* Schedule a response immediately. */
1450                         xenvif_idx_release(queue, pending_idx,
1451                                            XEN_NETIF_RSP_OKAY);
1452                 }
1453
1454                 if (txp->flags & XEN_NETTXF_csum_blank)
1455                         skb->ip_summed = CHECKSUM_PARTIAL;
1456                 else if (txp->flags & XEN_NETTXF_data_validated)
1457                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1458
1459                 xenvif_fill_frags(queue, skb);
1460
1461                 if (unlikely(skb_has_frag_list(skb))) {
1462                         if (xenvif_handle_frag_list(queue, skb)) {
1463                                 if (net_ratelimit())
1464                                         netdev_err(queue->vif->dev,
1465                                                    "Not enough memory to consolidate frag_list!\n");
1466                                 xenvif_skb_zerocopy_prepare(queue, skb);
1467                                 kfree_skb(skb);
1468                                 continue;
1469                         }
1470                 }
1471
1472                 skb->dev      = queue->vif->dev;
1473                 skb->protocol = eth_type_trans(skb, skb->dev);
1474                 skb_reset_network_header(skb);
1475
1476                 if (checksum_setup(queue, skb)) {
1477                         netdev_dbg(queue->vif->dev,
1478                                    "Can't setup checksum in net_tx_action\n");
1479                         /* We have to set this flag to trigger the callback */
1480                         if (skb_shinfo(skb)->destructor_arg)
1481                                 xenvif_skb_zerocopy_prepare(queue, skb);
1482                         kfree_skb(skb);
1483                         continue;
1484                 }
1485
1486                 skb_probe_transport_header(skb, 0);
1487
1488                 /* If the packet is GSO then we will have just set up the
1489                  * transport header offset in checksum_setup so it's now
1490                  * straightforward to calculate gso_segs.
1491                  */
1492                 if (skb_is_gso(skb)) {
1493                         int mss = skb_shinfo(skb)->gso_size;
1494                         int hdrlen = skb_transport_header(skb) -
1495                                 skb_mac_header(skb) +
1496                                 tcp_hdrlen(skb);
1497
1498                         skb_shinfo(skb)->gso_segs =
1499                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1500                 }
1501
1502                 queue->stats.rx_bytes += skb->len;
1503                 queue->stats.rx_packets++;
1504
1505                 work_done++;
1506
1507                 /* Set this flag right before netif_receive_skb, otherwise
1508                  * someone might think this packet already left netback, and
1509                  * do a skb_copy_ubufs while we are still in control of the
1510                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1511                  */
1512                 if (skb_shinfo(skb)->destructor_arg) {
1513                         xenvif_skb_zerocopy_prepare(queue, skb);
1514                         queue->stats.tx_zerocopy_sent++;
1515                 }
1516
1517                 netif_receive_skb(skb);
1518         }
1519
1520         return work_done;
1521 }
1522
1523 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1524 {
1525         unsigned long flags;
1526         pending_ring_idx_t index;
1527         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1528
1529         /* This is the only place where we grab this lock, to protect callbacks
1530          * from each other.
1531          */
1532         spin_lock_irqsave(&queue->callback_lock, flags);
1533         do {
1534                 u16 pending_idx = ubuf->desc;
1535                 ubuf = (struct ubuf_info *) ubuf->ctx;
1536                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1537                         MAX_PENDING_REQS);
1538                 index = pending_index(queue->dealloc_prod);
1539                 queue->dealloc_ring[index] = pending_idx;
1540                 /* Sync with xenvif_tx_dealloc_action:
1541                  * insert idx then incr producer.
1542                  */
1543                 smp_wmb();
1544                 queue->dealloc_prod++;
1545         } while (ubuf);
1546         wake_up(&queue->dealloc_wq);
1547         spin_unlock_irqrestore(&queue->callback_lock, flags);
1548
1549         if (likely(zerocopy_success))
1550                 queue->stats.tx_zerocopy_success++;
1551         else
1552                 queue->stats.tx_zerocopy_fail++;
1553         xenvif_skb_zerocopy_complete(queue);
1554 }
1555
1556 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1557 {
1558         struct gnttab_unmap_grant_ref *gop;
1559         pending_ring_idx_t dc, dp;
1560         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1561         unsigned int i = 0;
1562
1563         dc = queue->dealloc_cons;
1564         gop = queue->tx_unmap_ops;
1565
1566         /* Free up any grants we have finished using */
1567         do {
1568                 dp = queue->dealloc_prod;
1569
1570                 /* Ensure we see all indices enqueued by all
1571                  * xenvif_zerocopy_callback().
1572                  */
1573                 smp_rmb();
1574
1575                 while (dc != dp) {
1576                         BUG_ON(gop - queue->tx_unmap_ops > MAX_PENDING_REQS);
1577                         pending_idx =
1578                                 queue->dealloc_ring[pending_index(dc++)];
1579
1580                         pending_idx_release[gop-queue->tx_unmap_ops] =
1581                                 pending_idx;
1582                         queue->pages_to_unmap[gop-queue->tx_unmap_ops] =
1583                                 queue->mmap_pages[pending_idx];
1584                         gnttab_set_unmap_op(gop,
1585                                             idx_to_kaddr(queue, pending_idx),
1586                                             GNTMAP_host_map,
1587                                             queue->grant_tx_handle[pending_idx]);
1588                         xenvif_grant_handle_reset(queue, pending_idx);
1589                         ++gop;
1590                 }
1591
1592         } while (dp != queue->dealloc_prod);
1593
1594         queue->dealloc_cons = dc;
1595
1596         if (gop - queue->tx_unmap_ops > 0) {
1597                 int ret;
1598                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1599                                         NULL,
1600                                         queue->pages_to_unmap,
1601                                         gop - queue->tx_unmap_ops);
1602                 if (ret) {
1603                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tx ret %d\n",
1604                                    gop - queue->tx_unmap_ops, ret);
1605                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1606                                 if (gop[i].status != GNTST_okay)
1607                                         netdev_err(queue->vif->dev,
1608                                                    " host_addr: %llx handle: %x status: %d\n",
1609                                                    gop[i].host_addr,
1610                                                    gop[i].handle,
1611                                                    gop[i].status);
1612                         }
1613                         BUG();
1614                 }
1615         }
1616
1617         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1618                 xenvif_idx_release(queue, pending_idx_release[i],
1619                                    XEN_NETIF_RSP_OKAY);
1620 }
1621
1622
1623 /* Called after netfront has transmitted */
1624 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1625 {
1626         unsigned nr_mops, nr_cops = 0;
1627         int work_done, ret;
1628
1629         if (unlikely(!tx_work_todo(queue)))
1630                 return 0;
1631
1632         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1633
1634         if (nr_cops == 0)
1635                 return 0;
1636
1637         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1638         if (nr_mops != 0) {
1639                 ret = gnttab_map_refs(queue->tx_map_ops,
1640                                       NULL,
1641                                       queue->pages_to_map,
1642                                       nr_mops);
1643                 BUG_ON(ret);
1644         }
1645
1646         work_done = xenvif_tx_submit(queue);
1647
1648         return work_done;
1649 }
1650
1651 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1652                                u8 status)
1653 {
1654         struct pending_tx_info *pending_tx_info;
1655         pending_ring_idx_t index;
1656         unsigned long flags;
1657
1658         pending_tx_info = &queue->pending_tx_info[pending_idx];
1659
1660         spin_lock_irqsave(&queue->response_lock, flags);
1661
1662         make_tx_response(queue, &pending_tx_info->req, status);
1663
1664         /* Release the pending index before pusing the Tx response so
1665          * its available before a new Tx request is pushed by the
1666          * frontend.
1667          */
1668         index = pending_index(queue->pending_prod++);
1669         queue->pending_ring[index] = pending_idx;
1670
1671         push_tx_responses(queue);
1672
1673         spin_unlock_irqrestore(&queue->response_lock, flags);
1674 }
1675
1676
1677 static void make_tx_response(struct xenvif_queue *queue,
1678                              struct xen_netif_tx_request *txp,
1679                              s8       st)
1680 {
1681         RING_IDX i = queue->tx.rsp_prod_pvt;
1682         struct xen_netif_tx_response *resp;
1683
1684         resp = RING_GET_RESPONSE(&queue->tx, i);
1685         resp->id     = txp->id;
1686         resp->status = st;
1687
1688         if (txp->flags & XEN_NETTXF_extra_info)
1689                 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1690
1691         queue->tx.rsp_prod_pvt = ++i;
1692 }
1693
1694 static void push_tx_responses(struct xenvif_queue *queue)
1695 {
1696         int notify;
1697
1698         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1699         if (notify)
1700                 notify_remote_via_irq(queue->tx_irq);
1701 }
1702
1703 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1704                                              u16      id,
1705                                              s8       st,
1706                                              u16      offset,
1707                                              u16      size,
1708                                              u16      flags)
1709 {
1710         RING_IDX i = queue->rx.rsp_prod_pvt;
1711         struct xen_netif_rx_response *resp;
1712
1713         resp = RING_GET_RESPONSE(&queue->rx, i);
1714         resp->offset     = offset;
1715         resp->flags      = flags;
1716         resp->id         = id;
1717         resp->status     = (s16)size;
1718         if (st < 0)
1719                 resp->status = (s16)st;
1720
1721         queue->rx.rsp_prod_pvt = ++i;
1722
1723         return resp;
1724 }
1725
1726 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1727 {
1728         int ret;
1729         struct gnttab_unmap_grant_ref tx_unmap_op;
1730
1731         gnttab_set_unmap_op(&tx_unmap_op,
1732                             idx_to_kaddr(queue, pending_idx),
1733                             GNTMAP_host_map,
1734                             queue->grant_tx_handle[pending_idx]);
1735         xenvif_grant_handle_reset(queue, pending_idx);
1736
1737         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1738                                 &queue->mmap_pages[pending_idx], 1);
1739         if (ret) {
1740                 netdev_err(queue->vif->dev,
1741                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: %x status: %d\n",
1742                            ret,
1743                            pending_idx,
1744                            tx_unmap_op.host_addr,
1745                            tx_unmap_op.handle,
1746                            tx_unmap_op.status);
1747                 BUG();
1748         }
1749 }
1750
1751 static inline int tx_work_todo(struct xenvif_queue *queue)
1752 {
1753         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1754                 return 1;
1755
1756         return 0;
1757 }
1758
1759 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1760 {
1761         return queue->dealloc_cons != queue->dealloc_prod;
1762 }
1763
1764 void xenvif_unmap_frontend_rings(struct xenvif_queue *queue)
1765 {
1766         if (queue->tx.sring)
1767                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1768                                         queue->tx.sring);
1769         if (queue->rx.sring)
1770                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1771                                         queue->rx.sring);
1772 }
1773
1774 int xenvif_map_frontend_rings(struct xenvif_queue *queue,
1775                               grant_ref_t tx_ring_ref,
1776                               grant_ref_t rx_ring_ref)
1777 {
1778         void *addr;
1779         struct xen_netif_tx_sring *txs;
1780         struct xen_netif_rx_sring *rxs;
1781
1782         int err = -ENOMEM;
1783
1784         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1785                                      tx_ring_ref, &addr);
1786         if (err)
1787                 goto err;
1788
1789         txs = (struct xen_netif_tx_sring *)addr;
1790         BACK_RING_INIT(&queue->tx, txs, PAGE_SIZE);
1791
1792         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1793                                      rx_ring_ref, &addr);
1794         if (err)
1795                 goto err;
1796
1797         rxs = (struct xen_netif_rx_sring *)addr;
1798         BACK_RING_INIT(&queue->rx, rxs, PAGE_SIZE);
1799
1800         return 0;
1801
1802 err:
1803         xenvif_unmap_frontend_rings(queue);
1804         return err;
1805 }
1806
1807 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
1808 {
1809         struct xenvif *vif = queue->vif;
1810
1811         queue->stalled = true;
1812
1813         /* At least one queue has stalled? Disable the carrier. */
1814         spin_lock(&vif->lock);
1815         if (vif->stalled_queues++ == 0) {
1816                 netdev_info(vif->dev, "Guest Rx stalled");
1817                 netif_carrier_off(vif->dev);
1818         }
1819         spin_unlock(&vif->lock);
1820 }
1821
1822 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
1823 {
1824         struct xenvif *vif = queue->vif;
1825
1826         queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
1827         queue->stalled = false;
1828
1829         /* All queues are ready? Enable the carrier. */
1830         spin_lock(&vif->lock);
1831         if (--vif->stalled_queues == 0) {
1832                 netdev_info(vif->dev, "Guest Rx ready");
1833                 netif_carrier_on(vif->dev);
1834         }
1835         spin_unlock(&vif->lock);
1836 }
1837
1838 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
1839 {
1840         RING_IDX prod, cons;
1841
1842         prod = queue->rx.sring->req_prod;
1843         cons = queue->rx.req_cons;
1844
1845         return !queue->stalled
1846                 && prod - cons < XEN_NETBK_RX_SLOTS_MAX
1847                 && time_after(jiffies,
1848                               queue->last_rx_time + queue->vif->stall_timeout);
1849 }
1850
1851 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
1852 {
1853         RING_IDX prod, cons;
1854
1855         prod = queue->rx.sring->req_prod;
1856         cons = queue->rx.req_cons;
1857
1858         return queue->stalled
1859                 && prod - cons >= XEN_NETBK_RX_SLOTS_MAX;
1860 }
1861
1862 static bool xenvif_have_rx_work(struct xenvif_queue *queue)
1863 {
1864         return (!skb_queue_empty(&queue->rx_queue)
1865                 && xenvif_rx_ring_slots_available(queue, XEN_NETBK_RX_SLOTS_MAX))
1866                 || (queue->vif->stall_timeout &&
1867                     (xenvif_rx_queue_stalled(queue)
1868                      || xenvif_rx_queue_ready(queue)))
1869                 || kthread_should_stop()
1870                 || queue->vif->disabled;
1871 }
1872
1873 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
1874 {
1875         struct sk_buff *skb;
1876         long timeout;
1877
1878         skb = skb_peek(&queue->rx_queue);
1879         if (!skb)
1880                 return MAX_SCHEDULE_TIMEOUT;
1881
1882         timeout = XENVIF_RX_CB(skb)->expires - jiffies;
1883         return timeout < 0 ? 0 : timeout;
1884 }
1885
1886 /* Wait until the guest Rx thread has work.
1887  *
1888  * The timeout needs to be adjusted based on the current head of the
1889  * queue (and not just the head at the beginning).  In particular, if
1890  * the queue is initially empty an infinite timeout is used and this
1891  * needs to be reduced when a skb is queued.
1892  *
1893  * This cannot be done with wait_event_timeout() because it only
1894  * calculates the timeout once.
1895  */
1896 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
1897 {
1898         DEFINE_WAIT(wait);
1899
1900         if (xenvif_have_rx_work(queue))
1901                 return;
1902
1903         for (;;) {
1904                 long ret;
1905
1906                 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
1907                 if (xenvif_have_rx_work(queue))
1908                         break;
1909                 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
1910                 if (!ret)
1911                         break;
1912         }
1913         finish_wait(&queue->wq, &wait);
1914 }
1915
1916 int xenvif_kthread_guest_rx(void *data)
1917 {
1918         struct xenvif_queue *queue = data;
1919         struct xenvif *vif = queue->vif;
1920
1921         if (!vif->stall_timeout)
1922                 xenvif_queue_carrier_on(queue);
1923
1924         for (;;) {
1925                 xenvif_wait_for_rx_work(queue);
1926
1927                 if (kthread_should_stop())
1928                         break;
1929
1930                 /* This frontend is found to be rogue, disable it in
1931                  * kthread context. Currently this is only set when
1932                  * netback finds out frontend sends malformed packet,
1933                  * but we cannot disable the interface in softirq
1934                  * context so we defer it here, if this thread is
1935                  * associated with queue 0.
1936                  */
1937                 if (unlikely(vif->disabled && queue->id == 0)) {
1938                         xenvif_carrier_off(vif);
1939                         break;
1940                 }
1941
1942                 if (!skb_queue_empty(&queue->rx_queue))
1943                         xenvif_rx_action(queue);
1944
1945                 /* If the guest hasn't provided any Rx slots for a
1946                  * while it's probably not responsive, drop the
1947                  * carrier so packets are dropped earlier.
1948                  */
1949                 if (vif->stall_timeout) {
1950                         if (xenvif_rx_queue_stalled(queue))
1951                                 xenvif_queue_carrier_off(queue);
1952                         else if (xenvif_rx_queue_ready(queue))
1953                                 xenvif_queue_carrier_on(queue);
1954                 }
1955
1956                 /* Queued packets may have foreign pages from other
1957                  * domains.  These cannot be queued indefinitely as
1958                  * this would starve guests of grant refs and transmit
1959                  * slots.
1960                  */
1961                 xenvif_rx_queue_drop_expired(queue);
1962
1963                 xenvif_rx_queue_maybe_wake(queue);
1964
1965                 cond_resched();
1966         }
1967
1968         /* Bin any remaining skbs */
1969         xenvif_rx_queue_purge(queue);
1970
1971         return 0;
1972 }
1973
1974 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1975 {
1976         /* Dealloc thread must remain running until all inflight
1977          * packets complete.
1978          */
1979         return kthread_should_stop() &&
1980                 !atomic_read(&queue->inflight_packets);
1981 }
1982
1983 int xenvif_dealloc_kthread(void *data)
1984 {
1985         struct xenvif_queue *queue = data;
1986
1987         for (;;) {
1988                 wait_event_interruptible(queue->dealloc_wq,
1989                                          tx_dealloc_work_todo(queue) ||
1990                                          xenvif_dealloc_kthread_should_stop(queue));
1991                 if (xenvif_dealloc_kthread_should_stop(queue))
1992                         break;
1993
1994                 xenvif_tx_dealloc_action(queue);
1995                 cond_resched();
1996         }
1997
1998         /* Unmap anything remaining*/
1999         if (tx_dealloc_work_todo(queue))
2000                 xenvif_tx_dealloc_action(queue);
2001
2002         return 0;
2003 }
2004
2005 static int __init netback_init(void)
2006 {
2007         int rc = 0;
2008
2009         if (!xen_domain())
2010                 return -ENODEV;
2011
2012         /* Allow as many queues as there are CPUs, by default */
2013         xenvif_max_queues = num_online_cpus();
2014
2015         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2016                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2017                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2018                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2019         }
2020
2021         rc = xenvif_xenbus_init();
2022         if (rc)
2023                 goto failed_init;
2024
2025 #ifdef CONFIG_DEBUG_FS
2026         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2027         if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2028                 pr_warn("Init of debugfs returned %ld!\n",
2029                         PTR_ERR(xen_netback_dbg_root));
2030 #endif /* CONFIG_DEBUG_FS */
2031
2032         return 0;
2033
2034 failed_init:
2035         return rc;
2036 }
2037
2038 module_init(netback_init);
2039
2040 static void __exit netback_fini(void)
2041 {
2042 #ifdef CONFIG_DEBUG_FS
2043         if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2044                 debugfs_remove_recursive(xen_netback_dbg_root);
2045 #endif /* CONFIG_DEBUG_FS */
2046         xenvif_xenbus_fini();
2047 }
2048 module_exit(netback_fini);
2049
2050 MODULE_LICENSE("Dual BSD/GPL");
2051 MODULE_ALIAS("xen-backend:vif");