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 static int xenvif_rx_ring_slots_needed(struct xenvif *vif)
 153 {
 154         if (vif->gso_mask)
 155                 return DIV_ROUND_UP(vif->dev->gso_max_size, PAGE_SIZE) + 1;
 156         else
 157                 return DIV_ROUND_UP(vif->dev->mtu, PAGE_SIZE);
 158 }
 159
 160 static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
 161 {
 162         RING_IDX prod, cons;
 163         int needed;
 164
 165         needed = xenvif_rx_ring_slots_needed(queue->vif);
 166
 167         do {
 168                 prod = queue->rx.sring->req_prod;
 169                 cons = queue->rx.req_cons;
 170
 171                 if (prod - cons >= needed)
 172                         return true;
 173
 174                 queue->rx.sring->req_event = prod + 1;
 175
 176                 /* Make sure event is visible before we check prod
 177                  * again.
 178                  */
 179                 mb();
 180         } while (queue->rx.sring->req_prod != prod);
 181
 182         return false;
 183 }
 184
 185 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
 186 {
 187         unsigned long flags;
 188
 189         spin_lock_irqsave(&queue->rx_queue.lock, flags);
 190
 191         __skb_queue_tail(&queue->rx_queue, skb);
 192
 193         queue->rx_queue_len += skb->len;
 194         if (queue->rx_queue_len > queue->rx_queue_max)
 195                 netif_tx_stop_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
 196
 197         spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
 198 }
 199
 200 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
 201 {
 202         struct sk_buff *skb;
 203
 204         spin_lock_irq(&queue->rx_queue.lock);
 205
 206         skb = __skb_dequeue(&queue->rx_queue);
 207         if (skb)
 208                 queue->rx_queue_len -= skb->len;
 209
 210         spin_unlock_irq(&queue->rx_queue.lock);
 211
 212         return skb;
 213 }
 214
 215 static void xenvif_rx_queue_maybe_wake(struct xenvif_queue *queue)
 216 {
 217         spin_lock_irq(&queue->rx_queue.lock);
 218
 219         if (queue->rx_queue_len < queue->rx_queue_max)
 220                 netif_tx_wake_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
 221
 222         spin_unlock_irq(&queue->rx_queue.lock);
 223 }
 224
 225
 226 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
 227 {
 228         struct sk_buff *skb;
 229         while ((skb = xenvif_rx_dequeue(queue)) != NULL)
 230                 kfree_skb(skb);
 231 }
 232
 233 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
 234 {
 235         struct sk_buff *skb;
 236
 237         for(;;) {
 238                 skb = skb_peek(&queue->rx_queue);
 239                 if (!skb)
 240                         break;
 241                 if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
 242                         break;
 243                 xenvif_rx_dequeue(queue);
 244                 kfree_skb(skb);
 245         }
 246 }
 247
 248 struct netrx_pending_operations {
 249         unsigned copy_prod, copy_cons;
 250         unsigned meta_prod, meta_cons;
 251         struct gnttab_copy *copy;
 252         struct xenvif_rx_meta *meta;
 253         int copy_off;
 254         grant_ref_t copy_gref;
 255 };
 256
 257 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue,
 258                                                  struct netrx_pending_operations *npo)
 259 {
 260         struct xenvif_rx_meta *meta;
 261         struct xen_netif_rx_request *req;
 262
 263         req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
 264
 265         meta = npo->meta + npo->meta_prod++;
 266         meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
 267         meta->gso_size = 0;
 268         meta->size = 0;
 269         meta->id = req->id;
 270
 271         npo->copy_off = 0;
 272         npo->copy_gref = req->gref;
 273
 274         return meta;
 275 }
 276
 277 /*
 278  * Set up the grant operations for this fragment. If it's a flipping
 279  * interface, we also set up the unmap request from here.
 280  */
 281 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
 282                                  struct netrx_pending_operations *npo,
 283                                  struct page *page, unsigned long size,
 284                                  unsigned long offset, int *head)
 285 {
 286         struct gnttab_copy *copy_gop;
 287         struct xenvif_rx_meta *meta;
 288         unsigned long bytes;
 289         int gso_type = XEN_NETIF_GSO_TYPE_NONE;
 290
 291         /* Data must not cross a page boundary. */
 292         BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
 293
 294         meta = npo->meta + npo->meta_prod - 1;
 295
 296         /* Skip unused frames from start of page */
 297         page += offset >> PAGE_SHIFT;
 298         offset &= ~PAGE_MASK;
 299
 300         while (size > 0) {
 301                 struct xen_page_foreign *foreign;
 302
 303                 BUG_ON(offset >= PAGE_SIZE);
 304                 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
 305
 306                 if (npo->copy_off == MAX_BUFFER_OFFSET)
 307                         meta = get_next_rx_buffer(queue, npo);
 308
 309                 bytes = PAGE_SIZE - offset;
 310                 if (bytes > size)
 311                         bytes = size;
 312
 313                 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
 314                         bytes = MAX_BUFFER_OFFSET - npo->copy_off;
 315
 316                 copy_gop = npo->copy + npo->copy_prod++;
 317                 copy_gop->flags = GNTCOPY_dest_gref;
 318                 copy_gop->len = bytes;
 319
 320                 foreign = xen_page_foreign(page);
 321                 if (foreign) {
 322                         copy_gop->source.domid = foreign->domid;
 323                         copy_gop->source.u.ref = foreign->gref;
 324                         copy_gop->flags |= GNTCOPY_source_gref;
 325                 } else {
 326                         copy_gop->source.domid = DOMID_SELF;
 327                         copy_gop->source.u.gmfn =
 328                                 virt_to_gfn(page_address(page));
 329                 }
 330                 copy_gop->source.offset = offset;
 331
 332                 copy_gop->dest.domid = queue->vif->domid;
 333                 copy_gop->dest.offset = npo->copy_off;
 334                 copy_gop->dest.u.ref = npo->copy_gref;
 335
 336                 npo->copy_off += bytes;
 337                 meta->size += bytes;
 338
 339                 offset += bytes;
 340                 size -= bytes;
 341
 342                 /* Next frame */
 343                 if (offset == PAGE_SIZE && size) {
 344                         BUG_ON(!PageCompound(page));
 345                         page++;
 346                         offset = 0;
 347                 }
 348
 349                 /* Leave a gap for the GSO descriptor. */
 350                 if (skb_is_gso(skb)) {
 351                         if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 352                                 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
 353                         else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 354                                 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
 355                 }
 356
 357                 if (*head && ((1 << gso_type) & queue->vif->gso_mask))
 358                         queue->rx.req_cons++;
 359
 360                 *head = 0; /* There must be something in this buffer now. */
 361
 362         }
 363 }
 364
 365 /*
 366  * Prepare an SKB to be transmitted to the frontend.
 367  *
 368  * This function is responsible for allocating grant operations, meta
 369  * structures, etc.
 370  *
 371  * It returns the number of meta structures consumed. The number of
 372  * ring slots used is always equal to the number of meta slots used
 373  * plus the number of GSO descriptors used. Currently, we use either
 374  * zero GSO descriptors (for non-GSO packets) or one descriptor (for
 375  * frontend-side LRO).
 376  */
 377 static int xenvif_gop_skb(struct sk_buff *skb,
 378                           struct netrx_pending_operations *npo,
 379                           struct xenvif_queue *queue)
 380 {
 381         struct xenvif *vif = netdev_priv(skb->dev);
 382         int nr_frags = skb_shinfo(skb)->nr_frags;
 383         int i;
 384         struct xen_netif_rx_request *req;
 385         struct xenvif_rx_meta *meta;
 386         unsigned char *data;
 387         int head = 1;
 388         int old_meta_prod;
 389         int gso_type;
 390
 391         old_meta_prod = npo->meta_prod;
 392
 393         gso_type = XEN_NETIF_GSO_TYPE_NONE;
 394         if (skb_is_gso(skb)) {
 395                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 396                         gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
 397                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 398                         gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
 399         }
 400
 401         /* Set up a GSO prefix descriptor, if necessary */
 402         if ((1 << gso_type) & vif->gso_prefix_mask) {
 403                 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
 404                 meta = npo->meta + npo->meta_prod++;
 405                 meta->gso_type = gso_type;
 406                 meta->gso_size = skb_shinfo(skb)->gso_size;
 407                 meta->size = 0;
 408                 meta->id = req->id;
 409         }
 410
 411         req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
 412         meta = npo->meta + npo->meta_prod++;
 413
 414         if ((1 << gso_type) & vif->gso_mask) {
 415                 meta->gso_type = gso_type;
 416                 meta->gso_size = skb_shinfo(skb)->gso_size;
 417         } else {
 418                 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
 419                 meta->gso_size = 0;
 420         }
 421
 422         meta->size = 0;
 423         meta->id = req->id;
 424         npo->copy_off = 0;
 425         npo->copy_gref = req->gref;
 426
 427         data = skb->data;
 428         while (data < skb_tail_pointer(skb)) {
 429                 unsigned int offset = offset_in_page(data);
 430                 unsigned int len = PAGE_SIZE - offset;
 431
 432                 if (data + len > skb_tail_pointer(skb))
 433                         len = skb_tail_pointer(skb) - data;
 434
 435                 xenvif_gop_frag_copy(queue, skb, npo,
 436                                      virt_to_page(data), len, offset, &head);
 437                 data += len;
 438         }
 439
 440         for (i = 0; i < nr_frags; i++) {
 441                 xenvif_gop_frag_copy(queue, skb, npo,
 442                                      skb_frag_page(&skb_shinfo(skb)->frags[i]),
 443                                      skb_frag_size(&skb_shinfo(skb)->frags[i]),
 444                                      skb_shinfo(skb)->frags[i].page_offset,
 445                                      &head);
 446         }
 447
 448         return npo->meta_prod - old_meta_prod;
 449 }
 450
 451 /*
 452  * This is a twin to xenvif_gop_skb.  Assume that xenvif_gop_skb was
 453  * used to set up the operations on the top of
 454  * netrx_pending_operations, which have since been done.  Check that
 455  * they didn't give any errors and advance over them.
 456  */
 457 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
 458                             struct netrx_pending_operations *npo)
 459 {
 460         struct gnttab_copy     *copy_op;
 461         int status = XEN_NETIF_RSP_OKAY;
 462         int i;
 463
 464         for (i = 0; i < nr_meta_slots; i++) {
 465                 copy_op = npo->copy + npo->copy_cons++;
 466                 if (copy_op->status != GNTST_okay) {
 467                         netdev_dbg(vif->dev,
 468                                    "Bad status %d from copy to DOM%d.\n",
 469                                    copy_op->status, vif->domid);
 470                         status = XEN_NETIF_RSP_ERROR;
 471                 }
 472         }
 473
 474         return status;
 475 }
 476
 477 static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status,
 478                                       struct xenvif_rx_meta *meta,
 479                                       int nr_meta_slots)
 480 {
 481         int i;
 482         unsigned long offset;
 483
 484         /* No fragments used */
 485         if (nr_meta_slots <= 1)
 486                 return;
 487
 488         nr_meta_slots--;
 489
 490         for (i = 0; i < nr_meta_slots; i++) {
 491                 int flags;
 492                 if (i == nr_meta_slots - 1)
 493                         flags = 0;
 494                 else
 495                         flags = XEN_NETRXF_more_data;
 496
 497                 offset = 0;
 498                 make_rx_response(queue, meta[i].id, status, offset,
 499                                  meta[i].size, flags);
 500         }
 501 }
 502
 503 void xenvif_kick_thread(struct xenvif_queue *queue)
 504 {
 505         wake_up(&queue->wq);
 506 }
 507
 508 static void xenvif_rx_action(struct xenvif_queue *queue)
 509 {
 510         s8 status;
 511         u16 flags;
 512         struct xen_netif_rx_response *resp;
 513         struct sk_buff_head rxq;
 514         struct sk_buff *skb;
 515         LIST_HEAD(notify);
 516         int ret;
 517         unsigned long offset;
 518         bool need_to_notify = false;
 519
 520         struct netrx_pending_operations npo = {
 521                 .copy  = queue->grant_copy_op,
 522                 .meta  = queue->meta,
 523         };
 524
 525         skb_queue_head_init(&rxq);
 526
 527         while (xenvif_rx_ring_slots_available(queue)
 528                && (skb = xenvif_rx_dequeue(queue)) != NULL) {
 529                 queue->last_rx_time = jiffies;
 530
 531                 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
 532
 533                 __skb_queue_tail(&rxq, skb);
 534         }
 535
 536         BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
 537
 538         if (!npo.copy_prod)
 539                 goto done;
 540
 541         BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
 542         gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
 543
 544         while ((skb = __skb_dequeue(&rxq)) != NULL) {
 545
 546                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 547                     queue->vif->gso_prefix_mask) {
 548                         resp = RING_GET_RESPONSE(&queue->rx,
 549                                                  queue->rx.rsp_prod_pvt++);
 550
 551                         resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
 552
 553                         resp->offset = queue->meta[npo.meta_cons].gso_size;
 554                         resp->id = queue->meta[npo.meta_cons].id;
 555                         resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
 556
 557                         npo.meta_cons++;
 558                         XENVIF_RX_CB(skb)->meta_slots_used--;
 559                 }
 560
 561
 562                 queue->stats.tx_bytes += skb->len;
 563                 queue->stats.tx_packets++;
 564
 565                 status = xenvif_check_gop(queue->vif,
 566                                           XENVIF_RX_CB(skb)->meta_slots_used,
 567                                           &npo);
 568
 569                 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
 570                         flags = 0;
 571                 else
 572                         flags = XEN_NETRXF_more_data;
 573
 574                 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
 575                         flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
 576                 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 577                         /* remote but checksummed. */
 578                         flags |= XEN_NETRXF_data_validated;
 579
 580                 offset = 0;
 581                 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
 582                                         status, offset,
 583                                         queue->meta[npo.meta_cons].size,
 584                                         flags);
 585
 586                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 587                     queue->vif->gso_mask) {
 588                         struct xen_netif_extra_info *gso =
 589                                 (struct xen_netif_extra_info *)
 590                                 RING_GET_RESPONSE(&queue->rx,
 591                                                   queue->rx.rsp_prod_pvt++);
 592
 593                         resp->flags |= XEN_NETRXF_extra_info;
 594
 595                         gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
 596                         gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
 597                         gso->u.gso.pad = 0;
 598                         gso->u.gso.features = 0;
 599
 600                         gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
 601                         gso->flags = 0;
 602                 }
 603
 604                 xenvif_add_frag_responses(queue, status,
 605                                           queue->meta + npo.meta_cons + 1,
 606                                           XENVIF_RX_CB(skb)->meta_slots_used);
 607
 608                 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
 609
 610                 need_to_notify |= !!ret;
 611
 612                 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
 613                 dev_kfree_skb(skb);
 614         }
 615
 616 done:
 617         if (need_to_notify)
 618                 notify_remote_via_irq(queue->rx_irq);
 619 }
 620
 621 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 622 {
 623         int more_to_do;
 624
 625         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 626
 627         if (more_to_do)
 628                 napi_schedule(&queue->napi);
 629 }
 630
 631 static void tx_add_credit(struct xenvif_queue *queue)
 632 {
 633         unsigned long max_burst, max_credit;
 634
 635         /*
 636          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 637          * Otherwise the interface can seize up due to insufficient credit.
 638          */
 639         max_burst = RING_GET_REQUEST(&queue->tx, queue->tx.req_cons)->size;
 640         max_burst = min(max_burst, 131072UL);
 641         max_burst = max(max_burst, queue->credit_bytes);
 642
 643         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 644         max_credit = queue->remaining_credit + queue->credit_bytes;
 645         if (max_credit < queue->remaining_credit)
 646                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 647
 648         queue->remaining_credit = min(max_credit, max_burst);
 649 }
 650
 651 void xenvif_tx_credit_callback(unsigned long data)
 652 {
 653         struct xenvif_queue *queue = (struct xenvif_queue *)data;
 654         tx_add_credit(queue);
 655         xenvif_napi_schedule_or_enable_events(queue);
 656 }
 657
 658 static void xenvif_tx_err(struct xenvif_queue *queue,
 659                           struct xen_netif_tx_request *txp, RING_IDX end)
 660 {
 661         RING_IDX cons = queue->tx.req_cons;
 662         unsigned long flags;
 663
 664         do {
 665                 spin_lock_irqsave(&queue->response_lock, flags);
 666                 make_tx_response(queue, txp, XEN_NETIF_RSP_ERROR);
 667                 push_tx_responses(queue);
 668                 spin_unlock_irqrestore(&queue->response_lock, flags);
 669                 if (cons == end)
 670                         break;
 671                 txp = RING_GET_REQUEST(&queue->tx, cons++);
 672         } while (1);
 673         queue->tx.req_cons = cons;
 674 }
 675
 676 static void xenvif_fatal_tx_err(struct xenvif *vif)
 677 {
 678         netdev_err(vif->dev, "fatal error; disabling device\n");
 679         vif->disabled = true;
 680         /* Disable the vif from queue 0's kthread */
 681         if (vif->queues)
 682                 xenvif_kick_thread(&vif->queues[0]);
 683 }
 684
 685 static int xenvif_count_requests(struct xenvif_queue *queue,
 686                                  struct xen_netif_tx_request *first,
 687                                  struct xen_netif_tx_request *txp,
 688                                  int work_to_do)
 689 {
 690         RING_IDX cons = queue->tx.req_cons;
 691         int slots = 0;
 692         int drop_err = 0;
 693         int more_data;
 694
 695         if (!(first->flags & XEN_NETTXF_more_data))
 696                 return 0;
 697
 698         do {
 699                 struct xen_netif_tx_request dropped_tx = { 0 };
 700
 701                 if (slots >= work_to_do) {
 702                         netdev_err(queue->vif->dev,
 703                                    "Asked for %d slots but exceeds this limit\n",
 704                                    work_to_do);
 705                         xenvif_fatal_tx_err(queue->vif);
 706                         return -ENODATA;
 707                 }
 708
 709                 /* This guest is really using too many slots and
 710                  * considered malicious.
 711                  */
 712                 if (unlikely(slots >= fatal_skb_slots)) {
 713                         netdev_err(queue->vif->dev,
 714                                    "Malicious frontend using %d slots, threshold %u\n",
 715                                    slots, fatal_skb_slots);
 716                         xenvif_fatal_tx_err(queue->vif);
 717                         return -E2BIG;
 718                 }
 719
 720                 /* Xen network protocol had implicit dependency on
 721                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 722                  * the historical MAX_SKB_FRAGS value 18 to honor the
 723                  * same behavior as before. Any packet using more than
 724                  * 18 slots but less than fatal_skb_slots slots is
 725                  * dropped
 726                  */
 727                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 728                         if (net_ratelimit())
 729                                 netdev_dbg(queue->vif->dev,
 730                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 731                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 732                         drop_err = -E2BIG;
 733                 }
 734
 735                 if (drop_err)
 736                         txp = &dropped_tx;
 737
 738                 memcpy(txp, RING_GET_REQUEST(&queue->tx, cons + slots),
 739                        sizeof(*txp));
 740
 741                 /* If the guest submitted a frame >= 64 KiB then
 742                  * first->size overflowed and following slots will
 743                  * appear to be larger than the frame.
 744                  *
 745                  * This cannot be fatal error as there are buggy
 746                  * frontends that do this.
 747                  *
 748                  * Consume all slots and drop the packet.
 749                  */
 750                 if (!drop_err && txp->size > first->size) {
 751                         if (net_ratelimit())
 752                                 netdev_dbg(queue->vif->dev,
 753                                            "Invalid tx request, slot size %u > remaining size %u\n",
 754                                            txp->size, first->size);
 755                         drop_err = -EIO;
 756                 }
 757
 758                 first->size -= txp->size;
 759                 slots++;
 760
 761                 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
 762                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 763                                  txp->offset, txp->size);
 764                         xenvif_fatal_tx_err(queue->vif);
 765                         return -EINVAL;
 766                 }
 767
 768                 more_data = txp->flags & XEN_NETTXF_more_data;
 769
 770                 if (!drop_err)
 771                         txp++;
 772
 773         } while (more_data);
 774
 775         if (drop_err) {
 776                 xenvif_tx_err(queue, first, cons + slots);
 777                 return drop_err;
 778         }
 779
 780         return slots;
 781 }
 782
 783
 784 struct xenvif_tx_cb {
 785         u16 pending_idx;
 786 };
 787
 788 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 789
 790 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 791                                           u16 pending_idx,
 792                                           struct xen_netif_tx_request *txp,
 793                                           struct gnttab_map_grant_ref *mop)
 794 {
 795         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 796         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 797                           GNTMAP_host_map | GNTMAP_readonly,
 798                           txp->gref, queue->vif->domid);
 799
 800         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 801                sizeof(*txp));
 802 }
 803
 804 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 805 {
 806         struct sk_buff *skb =
 807                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 808                           GFP_ATOMIC | __GFP_NOWARN);
 809         if (unlikely(skb == NULL))
 810                 return NULL;
 811
 812         /* Packets passed to netif_rx() must have some headroom. */
 813         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 814
 815         /* Initialize it here to avoid later surprises */
 816         skb_shinfo(skb)->destructor_arg = NULL;
 817
 818         return skb;
 819 }
 820
 821 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
 822                                                         struct sk_buff *skb,
 823                                                         struct xen_netif_tx_request *txp,
 824                                                         struct gnttab_map_grant_ref *gop,
 825                                                         unsigned int frag_overflow,
 826                                                         struct sk_buff *nskb)
 827 {
 828         struct skb_shared_info *shinfo = skb_shinfo(skb);
 829         skb_frag_t *frags = shinfo->frags;
 830         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 831         int start;
 832         pending_ring_idx_t index;
 833         unsigned int nr_slots;
 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
 850                 shinfo = skb_shinfo(nskb);
 851                 frags = shinfo->frags;
 852
 853                 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
 854                      shinfo->nr_frags++, txp++, gop++) {
 855                         index = pending_index(queue->pending_cons++);
 856                         pending_idx = queue->pending_ring[index];
 857                         xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
 858                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 859                                              pending_idx);
 860                 }
 861
 862                 skb_shinfo(skb)->frag_list = nskb;
 863         }
 864
 865         return gop;
 866 }
 867
 868 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 869                                            u16 pending_idx,
 870                                            grant_handle_t handle)
 871 {
 872         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 873                      NETBACK_INVALID_HANDLE)) {
 874                 netdev_err(queue->vif->dev,
 875                            "Trying to overwrite active handle! pending_idx: 0x%x\n",
 876                            pending_idx);
 877                 BUG();
 878         }
 879         queue->grant_tx_handle[pending_idx] = handle;
 880 }
 881
 882 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 883                                              u16 pending_idx)
 884 {
 885         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 886                      NETBACK_INVALID_HANDLE)) {
 887                 netdev_err(queue->vif->dev,
 888                            "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 889                            pending_idx);
 890                 BUG();
 891         }
 892         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 893 }
 894
 895 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 896                                struct sk_buff *skb,
 897                                struct gnttab_map_grant_ref **gopp_map,
 898                                struct gnttab_copy **gopp_copy)
 899 {
 900         struct gnttab_map_grant_ref *gop_map = *gopp_map;
 901         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 902         /* This always points to the shinfo of the skb being checked, which
 903          * could be either the first or the one on the frag_list
 904          */
 905         struct skb_shared_info *shinfo = skb_shinfo(skb);
 906         /* If this is non-NULL, we are currently checking the frag_list skb, and
 907          * this points to the shinfo of the first one
 908          */
 909         struct skb_shared_info *first_shinfo = NULL;
 910         int nr_frags = shinfo->nr_frags;
 911         const bool sharedslot = nr_frags &&
 912                                 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
 913         int i, err;
 914
 915         /* Check status of header. */
 916         err = (*gopp_copy)->status;
 917         if (unlikely(err)) {
 918                 if (net_ratelimit())
 919                         netdev_dbg(queue->vif->dev,
 920                                    "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 921                                    (*gopp_copy)->status,
 922                                    pending_idx,
 923                                    (*gopp_copy)->source.u.ref);
 924                 /* The first frag might still have this slot mapped */
 925                 if (!sharedslot)
 926                         xenvif_idx_release(queue, pending_idx,
 927                                            XEN_NETIF_RSP_ERROR);
 928         }
 929         (*gopp_copy)++;
 930
 931 check_frags:
 932         for (i = 0; i < nr_frags; i++, gop_map++) {
 933                 int j, newerr;
 934
 935                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 936
 937                 /* Check error status: if okay then remember grant handle. */
 938                 newerr = gop_map->status;
 939
 940                 if (likely(!newerr)) {
 941                         xenvif_grant_handle_set(queue,
 942                                                 pending_idx,
 943                                                 gop_map->handle);
 944                         /* Had a previous error? Invalidate this fragment. */
 945                         if (unlikely(err)) {
 946                                 xenvif_idx_unmap(queue, pending_idx);
 947                                 /* If the mapping of the first frag was OK, but
 948                                  * the header's copy failed, and they are
 949                                  * sharing a slot, send an error
 950                                  */
 951                                 if (i == 0 && sharedslot)
 952                                         xenvif_idx_release(queue, pending_idx,
 953                                                            XEN_NETIF_RSP_ERROR);
 954                                 else
 955                                         xenvif_idx_release(queue, pending_idx,
 956                                                            XEN_NETIF_RSP_OKAY);
 957                         }
 958                         continue;
 959                 }
 960
 961                 /* Error on this fragment: respond to client with an error. */
 962                 if (net_ratelimit())
 963                         netdev_dbg(queue->vif->dev,
 964                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
 965                                    i,
 966                                    gop_map->status,
 967                                    pending_idx,
 968                                    gop_map->ref);
 969
 970                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
 971
 972                 /* Not the first error? Preceding frags already invalidated. */
 973                 if (err)
 974                         continue;
 975
 976                 /* First error: if the header haven't shared a slot with the
 977                  * first frag, release it as well.
 978                  */
 979                 if (!sharedslot)
 980                         xenvif_idx_release(queue,
 981                                            XENVIF_TX_CB(skb)->pending_idx,
 982                                            XEN_NETIF_RSP_OKAY);
 983
 984                 /* Invalidate preceding fragments of this skb. */
 985                 for (j = 0; j < i; j++) {
 986                         pending_idx = frag_get_pending_idx(&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                 /* And if we found the error while checking the frag_list, unmap
 993                  * the first skb's frags
 994                  */
 995                 if (first_shinfo) {
 996                         for (j = 0; j < first_shinfo->nr_frags; j++) {
 997                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
 998                                 xenvif_idx_unmap(queue, pending_idx);
 999                                 xenvif_idx_release(queue, pending_idx,
1000                                                    XEN_NETIF_RSP_OKAY);
1001                         }
1002                 }
1003
1004                 /* Remember the error: invalidate all subsequent fragments. */
1005                 err = newerr;
1006         }
1007
1008         if (skb_has_frag_list(skb) && !first_shinfo) {
1009                 first_shinfo = skb_shinfo(skb);
1010                 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1011                 nr_frags = shinfo->nr_frags;
1012
1013                 goto check_frags;
1014         }
1015
1016         *gopp_map = gop_map;
1017         return err;
1018 }
1019
1020 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1021 {
1022         struct skb_shared_info *shinfo = skb_shinfo(skb);
1023         int nr_frags = shinfo->nr_frags;
1024         int i;
1025         u16 prev_pending_idx = INVALID_PENDING_IDX;
1026
1027         for (i = 0; i < nr_frags; i++) {
1028                 skb_frag_t *frag = shinfo->frags + i;
1029                 struct xen_netif_tx_request *txp;
1030                 struct page *page;
1031                 u16 pending_idx;
1032
1033                 pending_idx = frag_get_pending_idx(frag);
1034
1035                 /* If this is not the first frag, chain it to the previous*/
1036                 if (prev_pending_idx == INVALID_PENDING_IDX)
1037                         skb_shinfo(skb)->destructor_arg =
1038                                 &callback_param(queue, pending_idx);
1039                 else
1040                         callback_param(queue, prev_pending_idx).ctx =
1041                                 &callback_param(queue, pending_idx);
1042
1043                 callback_param(queue, pending_idx).ctx = NULL;
1044                 prev_pending_idx = pending_idx;
1045
1046                 txp = &queue->pending_tx_info[pending_idx].req;
1047                 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1048                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1049                 skb->len += txp->size;
1050                 skb->data_len += txp->size;
1051                 skb->truesize += txp->size;
1052
1053                 /* Take an extra reference to offset network stack's put_page */
1054                 get_page(queue->mmap_pages[pending_idx]);
1055         }
1056 }
1057
1058 static int xenvif_get_extras(struct xenvif_queue *queue,
1059                                 struct xen_netif_extra_info *extras,
1060                                 int work_to_do)
1061 {
1062         struct xen_netif_extra_info extra;
1063         RING_IDX cons = queue->tx.req_cons;
1064
1065         do {
1066                 if (unlikely(work_to_do-- <= 0)) {
1067                         netdev_err(queue->vif->dev, "Missing extra info\n");
1068                         xenvif_fatal_tx_err(queue->vif);
1069                         return -EBADR;
1070                 }
1071
1072                 memcpy(&extra, RING_GET_REQUEST(&queue->tx, cons),
1073                        sizeof(extra));
1074                 if (unlikely(!extra.type ||
1075                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1076                         queue->tx.req_cons = ++cons;
1077                         netdev_err(queue->vif->dev,
1078                                    "Invalid extra type: %d\n", extra.type);
1079                         xenvif_fatal_tx_err(queue->vif);
1080                         return -EINVAL;
1081                 }
1082
1083                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1084                 queue->tx.req_cons = ++cons;
1085         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1086
1087         return work_to_do;
1088 }
1089
1090 static int xenvif_set_skb_gso(struct xenvif *vif,
1091                               struct sk_buff *skb,
1092                               struct xen_netif_extra_info *gso)
1093 {
1094         if (!gso->u.gso.size) {
1095                 netdev_err(vif->dev, "GSO size must not be zero.\n");
1096                 xenvif_fatal_tx_err(vif);
1097                 return -EINVAL;
1098         }
1099
1100         switch (gso->u.gso.type) {
1101         case XEN_NETIF_GSO_TYPE_TCPV4:
1102                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1103                 break;
1104         case XEN_NETIF_GSO_TYPE_TCPV6:
1105                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1106                 break;
1107         default:
1108                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1109                 xenvif_fatal_tx_err(vif);
1110                 return -EINVAL;
1111         }
1112
1113         skb_shinfo(skb)->gso_size = gso->u.gso.size;
1114         /* gso_segs will be calculated later */
1115
1116         return 0;
1117 }
1118
1119 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1120 {
1121         bool recalculate_partial_csum = false;
1122
1123         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1124          * peers can fail to set NETRXF_csum_blank when sending a GSO
1125          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1126          * recalculate the partial checksum.
1127          */
1128         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1129                 queue->stats.rx_gso_checksum_fixup++;
1130                 skb->ip_summed = CHECKSUM_PARTIAL;
1131                 recalculate_partial_csum = true;
1132         }
1133
1134         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1135         if (skb->ip_summed != CHECKSUM_PARTIAL)
1136                 return 0;
1137
1138         return skb_checksum_setup(skb, recalculate_partial_csum);
1139 }
1140
1141 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1142 {
1143         u64 now = get_jiffies_64();
1144         u64 next_credit = queue->credit_window_start +
1145                 msecs_to_jiffies(queue->credit_usec / 1000);
1146
1147         /* Timer could already be pending in rare cases. */
1148         if (timer_pending(&queue->credit_timeout))
1149                 return true;
1150
1151         /* Passed the point where we can replenish credit? */
1152         if (time_after_eq64(now, next_credit)) {
1153                 queue->credit_window_start = now;
1154                 tx_add_credit(queue);
1155         }
1156
1157         /* Still too big to send right now? Set a callback. */
1158         if (size > queue->remaining_credit) {
1159                 queue->credit_timeout.data     =
1160                         (unsigned long)queue;
1161                 mod_timer(&queue->credit_timeout,
1162                           next_credit);
1163                 queue->credit_window_start = next_credit;
1164
1165                 return true;
1166         }
1167
1168         return false;
1169 }
1170
1171 /* No locking is required in xenvif_mcast_add/del() as they are
1172  * only ever invoked from NAPI poll. An RCU list is used because
1173  * xenvif_mcast_match() is called asynchronously, during start_xmit.
1174  */
1175
1176 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
1177 {
1178         struct xenvif_mcast_addr *mcast;
1179
1180         if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
1181                 if (net_ratelimit())
1182                         netdev_err(vif->dev,
1183                                    "Too many multicast addresses\n");
1184                 return -ENOSPC;
1185         }
1186
1187         mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
1188         if (!mcast)
1189                 return -ENOMEM;
1190
1191         ether_addr_copy(mcast->addr, addr);
1192         list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
1193         vif->fe_mcast_count++;
1194
1195         return 0;
1196 }
1197
1198 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
1199 {
1200         struct xenvif_mcast_addr *mcast;
1201
1202         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1203                 if (ether_addr_equal(addr, mcast->addr)) {
1204                         --vif->fe_mcast_count;
1205                         list_del_rcu(&mcast->entry);
1206                         kfree_rcu(mcast, rcu);
1207                         break;
1208                 }
1209         }
1210 }
1211
1212 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
1213 {
1214         struct xenvif_mcast_addr *mcast;
1215
1216         rcu_read_lock();
1217         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1218                 if (ether_addr_equal(addr, mcast->addr)) {
1219                         rcu_read_unlock();
1220                         return true;
1221                 }
1222         }
1223         rcu_read_unlock();
1224
1225         return false;
1226 }
1227
1228 void xenvif_mcast_addr_list_free(struct xenvif *vif)
1229 {
1230         /* No need for locking or RCU here. NAPI poll and TX queue
1231          * are stopped.
1232          */
1233         while (!list_empty(&vif->fe_mcast_addr)) {
1234                 struct xenvif_mcast_addr *mcast;
1235
1236                 mcast = list_first_entry(&vif->fe_mcast_addr,
1237                                          struct xenvif_mcast_addr,
1238                                          entry);
1239                 --vif->fe_mcast_count;
1240                 list_del(&mcast->entry);
1241                 kfree(mcast);
1242         }
1243 }
1244
1245 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1246                                      int budget,
1247                                      unsigned *copy_ops,
1248                                      unsigned *map_ops)
1249 {
1250         struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
1251         struct sk_buff *skb, *nskb;
1252         int ret;
1253         unsigned int frag_overflow;
1254
1255         while (skb_queue_len(&queue->tx_queue) < budget) {
1256                 struct xen_netif_tx_request txreq;
1257                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1258                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1259                 u16 pending_idx;
1260                 RING_IDX idx;
1261                 int work_to_do;
1262                 unsigned int data_len;
1263                 pending_ring_idx_t index;
1264
1265                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1266                     XEN_NETIF_TX_RING_SIZE) {
1267                         netdev_err(queue->vif->dev,
1268                                    "Impossible number of requests. "
1269                                    "req_prod %d, req_cons %d, size %ld\n",
1270                                    queue->tx.sring->req_prod, queue->tx.req_cons,
1271                                    XEN_NETIF_TX_RING_SIZE);
1272                         xenvif_fatal_tx_err(queue->vif);
1273                         break;
1274                 }
1275
1276                 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1277                 if (!work_to_do)
1278                         break;
1279
1280                 idx = queue->tx.req_cons;
1281                 rmb(); /* Ensure that we see the request before we copy it. */
1282                 memcpy(&txreq, RING_GET_REQUEST(&queue->tx, idx), sizeof(txreq));
1283
1284                 /* Credit-based scheduling. */
1285                 if (txreq.size > queue->remaining_credit &&
1286                     tx_credit_exceeded(queue, txreq.size))
1287                         break;
1288
1289                 queue->remaining_credit -= txreq.size;
1290
1291                 work_to_do--;
1292                 queue->tx.req_cons = ++idx;
1293
1294                 memset(extras, 0, sizeof(extras));
1295                 if (txreq.flags & XEN_NETTXF_extra_info) {
1296                         work_to_do = xenvif_get_extras(queue, extras,
1297                                                        work_to_do);
1298                         idx = queue->tx.req_cons;
1299                         if (unlikely(work_to_do < 0))
1300                                 break;
1301                 }
1302
1303                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
1304                         struct xen_netif_extra_info *extra;
1305
1306                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
1307                         ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
1308
1309                         make_tx_response(queue, &txreq,
1310                                          (ret == 0) ?
1311                                          XEN_NETIF_RSP_OKAY :
1312                                          XEN_NETIF_RSP_ERROR);
1313                         push_tx_responses(queue);
1314                         continue;
1315                 }
1316
1317                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
1318                         struct xen_netif_extra_info *extra;
1319
1320                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
1321                         xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
1322
1323                         make_tx_response(queue, &txreq, XEN_NETIF_RSP_OKAY);
1324                         push_tx_responses(queue);
1325                         continue;
1326                 }
1327
1328                 ret = xenvif_count_requests(queue, &txreq, txfrags, work_to_do);
1329                 if (unlikely(ret < 0))
1330                         break;
1331
1332                 idx += ret;
1333
1334                 if (unlikely(txreq.size < ETH_HLEN)) {
1335                         netdev_dbg(queue->vif->dev,
1336                                    "Bad packet size: %d\n", txreq.size);
1337                         xenvif_tx_err(queue, &txreq, idx);
1338                         break;
1339                 }
1340
1341                 /* No crossing a page as the payload mustn't fragment. */
1342                 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1343                         netdev_err(queue->vif->dev,
1344                                    "txreq.offset: %u, size: %u, end: %lu\n",
1345                                    txreq.offset, txreq.size,
1346                                    (unsigned long)(txreq.offset&~PAGE_MASK) + txreq.size);
1347                         xenvif_fatal_tx_err(queue->vif);
1348                         break;
1349                 }
1350
1351                 index = pending_index(queue->pending_cons);
1352                 pending_idx = queue->pending_ring[index];
1353
1354                 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
1355                             ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1356                         XEN_NETBACK_TX_COPY_LEN : txreq.size;
1357
1358                 skb = xenvif_alloc_skb(data_len);
1359                 if (unlikely(skb == NULL)) {
1360                         netdev_dbg(queue->vif->dev,
1361                                    "Can't allocate a skb in start_xmit.\n");
1362                         xenvif_tx_err(queue, &txreq, idx);
1363                         break;
1364                 }
1365
1366                 skb_shinfo(skb)->nr_frags = ret;
1367                 if (data_len < txreq.size)
1368                         skb_shinfo(skb)->nr_frags++;
1369                 /* At this point shinfo->nr_frags is in fact the number of
1370                  * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1371                  */
1372                 frag_overflow = 0;
1373                 nskb = NULL;
1374                 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1375                         frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1376                         BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1377                         skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1378                         nskb = xenvif_alloc_skb(0);
1379                         if (unlikely(nskb == NULL)) {
1380                                 kfree_skb(skb);
1381                                 xenvif_tx_err(queue, &txreq, idx);
1382                                 if (net_ratelimit())
1383                                         netdev_err(queue->vif->dev,
1384                                                    "Can't allocate the frag_list skb.\n");
1385                                 break;
1386                         }
1387                 }
1388
1389                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1390                         struct xen_netif_extra_info *gso;
1391                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1392
1393                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1394                                 /* Failure in xenvif_set_skb_gso is fatal. */
1395                                 kfree_skb(skb);
1396                                 kfree_skb(nskb);
1397                                 break;
1398                         }
1399                 }
1400
1401                 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1402
1403                 __skb_put(skb, data_len);
1404                 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1405                 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1406                 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1407
1408                 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1409                         virt_to_gfn(skb->data);
1410                 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1411                 queue->tx_copy_ops[*copy_ops].dest.offset =
1412                         offset_in_page(skb->data);
1413
1414                 queue->tx_copy_ops[*copy_ops].len = data_len;
1415                 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1416
1417                 (*copy_ops)++;
1418
1419                 if (data_len < txreq.size) {
1420                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1421                                              pending_idx);
1422                         xenvif_tx_create_map_op(queue, pending_idx, &txreq, gop);
1423                         gop++;
1424                 } else {
1425                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1426                                              INVALID_PENDING_IDX);
1427                         memcpy(&queue->pending_tx_info[pending_idx].req, &txreq,
1428                                sizeof(txreq));
1429                 }
1430
1431                 queue->pending_cons++;
1432
1433                 gop = xenvif_get_requests(queue, skb, txfrags, gop,
1434                                           frag_overflow, nskb);
1435
1436                 __skb_queue_tail(&queue->tx_queue, skb);
1437
1438                 queue->tx.req_cons = idx;
1439
1440                 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1441                     (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1442                         break;
1443         }
1444
1445         (*map_ops) = gop - queue->tx_map_ops;
1446         return;
1447 }
1448
1449 /* Consolidate skb with a frag_list into a brand new one with local pages on
1450  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1451  */
1452 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1453 {
1454         unsigned int offset = skb_headlen(skb);
1455         skb_frag_t frags[MAX_SKB_FRAGS];
1456         int i, f;
1457         struct ubuf_info *uarg;
1458         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1459
1460         queue->stats.tx_zerocopy_sent += 2;
1461         queue->stats.tx_frag_overflow++;
1462
1463         xenvif_fill_frags(queue, nskb);
1464         /* Subtract frags size, we will correct it later */
1465         skb->truesize -= skb->data_len;
1466         skb->len += nskb->len;
1467         skb->data_len += nskb->len;
1468
1469         /* create a brand new frags array and coalesce there */
1470         for (i = 0; offset < skb->len; i++) {
1471                 struct page *page;
1472                 unsigned int len;
1473
1474                 BUG_ON(i >= MAX_SKB_FRAGS);
1475                 page = alloc_page(GFP_ATOMIC);
1476                 if (!page) {
1477                         int j;
1478                         skb->truesize += skb->data_len;
1479                         for (j = 0; j < i; j++)
1480                                 put_page(frags[j].page.p);
1481                         return -ENOMEM;
1482                 }
1483
1484                 if (offset + PAGE_SIZE < skb->len)
1485                         len = PAGE_SIZE;
1486                 else
1487                         len = skb->len - offset;
1488                 if (skb_copy_bits(skb, offset, page_address(page), len))
1489                         BUG();
1490
1491                 offset += len;
1492                 frags[i].page.p = page;
1493                 frags[i].page_offset = 0;
1494                 skb_frag_size_set(&frags[i], len);
1495         }
1496
1497         /* Copied all the bits from the frag list -- free it. */
1498         skb_frag_list_init(skb);
1499         xenvif_skb_zerocopy_prepare(queue, nskb);
1500         kfree_skb(nskb);
1501
1502         /* Release all the original (foreign) frags. */
1503         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1504                 skb_frag_unref(skb, f);
1505         uarg = skb_shinfo(skb)->destructor_arg;
1506         /* increase inflight counter to offset decrement in callback */
1507         atomic_inc(&queue->inflight_packets);
1508         uarg->callback(uarg, true);
1509         skb_shinfo(skb)->destructor_arg = NULL;
1510
1511         /* Fill the skb with the new (local) frags. */
1512         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1513         skb_shinfo(skb)->nr_frags = i;
1514         skb->truesize += i * PAGE_SIZE;
1515
1516         return 0;
1517 }
1518
1519 static int xenvif_tx_submit(struct xenvif_queue *queue)
1520 {
1521         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1522         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1523         struct sk_buff *skb;
1524         int work_done = 0;
1525
1526         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1527                 struct xen_netif_tx_request *txp;
1528                 u16 pending_idx;
1529                 unsigned data_len;
1530
1531                 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1532                 txp = &queue->pending_tx_info[pending_idx].req;
1533
1534                 /* Check the remap error code. */
1535                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1536                         /* If there was an error, xenvif_tx_check_gop is
1537                          * expected to release all the frags which were mapped,
1538                          * so kfree_skb shouldn't do it again
1539                          */
1540                         skb_shinfo(skb)->nr_frags = 0;
1541                         if (skb_has_frag_list(skb)) {
1542                                 struct sk_buff *nskb =
1543                                                 skb_shinfo(skb)->frag_list;
1544                                 skb_shinfo(nskb)->nr_frags = 0;
1545                         }
1546                         kfree_skb(skb);
1547                         continue;
1548                 }
1549
1550                 data_len = skb->len;
1551                 callback_param(queue, pending_idx).ctx = NULL;
1552                 if (data_len < txp->size) {
1553                         /* Append the packet payload as a fragment. */
1554                         txp->offset += data_len;
1555                         txp->size -= data_len;
1556                 } else {
1557                         /* Schedule a response immediately. */
1558                         xenvif_idx_release(queue, pending_idx,
1559                                            XEN_NETIF_RSP_OKAY);
1560                 }
1561
1562                 if (txp->flags & XEN_NETTXF_csum_blank)
1563                         skb->ip_summed = CHECKSUM_PARTIAL;
1564                 else if (txp->flags & XEN_NETTXF_data_validated)
1565                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1566
1567                 xenvif_fill_frags(queue, skb);
1568
1569                 if (unlikely(skb_has_frag_list(skb))) {
1570                         if (xenvif_handle_frag_list(queue, skb)) {
1571                                 if (net_ratelimit())
1572                                         netdev_err(queue->vif->dev,
1573                                                    "Not enough memory to consolidate frag_list!\n");
1574                                 xenvif_skb_zerocopy_prepare(queue, skb);
1575                                 kfree_skb(skb);
1576                                 continue;
1577                         }
1578                 }
1579
1580                 skb->dev      = queue->vif->dev;
1581                 skb->protocol = eth_type_trans(skb, skb->dev);
1582                 skb_reset_network_header(skb);
1583
1584                 if (checksum_setup(queue, skb)) {
1585                         netdev_dbg(queue->vif->dev,
1586                                    "Can't setup checksum in net_tx_action\n");
1587                         /* We have to set this flag to trigger the callback */
1588                         if (skb_shinfo(skb)->destructor_arg)
1589                                 xenvif_skb_zerocopy_prepare(queue, skb);
1590                         kfree_skb(skb);
1591                         continue;
1592                 }
1593
1594                 skb_probe_transport_header(skb, 0);
1595
1596                 /* If the packet is GSO then we will have just set up the
1597                  * transport header offset in checksum_setup so it's now
1598                  * straightforward to calculate gso_segs.
1599                  */
1600                 if (skb_is_gso(skb)) {
1601                         int mss = skb_shinfo(skb)->gso_size;
1602                         int hdrlen = skb_transport_header(skb) -
1603                                 skb_mac_header(skb) +
1604                                 tcp_hdrlen(skb);
1605
1606                         skb_shinfo(skb)->gso_segs =
1607                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1608                 }
1609
1610                 queue->stats.rx_bytes += skb->len;
1611                 queue->stats.rx_packets++;
1612
1613                 work_done++;
1614
1615                 /* Set this flag right before netif_receive_skb, otherwise
1616                  * someone might think this packet already left netback, and
1617                  * do a skb_copy_ubufs while we are still in control of the
1618                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1619                  */
1620                 if (skb_shinfo(skb)->destructor_arg) {
1621                         xenvif_skb_zerocopy_prepare(queue, skb);
1622                         queue->stats.tx_zerocopy_sent++;
1623                 }
1624
1625                 netif_receive_skb(skb);
1626         }
1627
1628         return work_done;
1629 }
1630
1631 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1632 {
1633         unsigned long flags;
1634         pending_ring_idx_t index;
1635         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1636
1637         /* This is the only place where we grab this lock, to protect callbacks
1638          * from each other.
1639          */
1640         spin_lock_irqsave(&queue->callback_lock, flags);
1641         do {
1642                 u16 pending_idx = ubuf->desc;
1643                 ubuf = (struct ubuf_info *) ubuf->ctx;
1644                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1645                         MAX_PENDING_REQS);
1646                 index = pending_index(queue->dealloc_prod);
1647                 queue->dealloc_ring[index] = pending_idx;
1648                 /* Sync with xenvif_tx_dealloc_action:
1649                  * insert idx then incr producer.
1650                  */
1651                 smp_wmb();
1652                 queue->dealloc_prod++;
1653         } while (ubuf);
1654         spin_unlock_irqrestore(&queue->callback_lock, flags);
1655
1656         if (likely(zerocopy_success))
1657                 queue->stats.tx_zerocopy_success++;
1658         else
1659                 queue->stats.tx_zerocopy_fail++;
1660         xenvif_skb_zerocopy_complete(queue);
1661 }
1662
1663 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1664 {
1665         struct gnttab_unmap_grant_ref *gop;
1666         pending_ring_idx_t dc, dp;
1667         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1668         unsigned int i = 0;
1669
1670         dc = queue->dealloc_cons;
1671         gop = queue->tx_unmap_ops;
1672
1673         /* Free up any grants we have finished using */
1674         do {
1675                 dp = queue->dealloc_prod;
1676
1677                 /* Ensure we see all indices enqueued by all
1678                  * xenvif_zerocopy_callback().
1679                  */
1680                 smp_rmb();
1681
1682                 while (dc != dp) {
1683                         BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1684                         pending_idx =
1685                                 queue->dealloc_ring[pending_index(dc++)];
1686
1687                         pending_idx_release[gop - queue->tx_unmap_ops] =
1688                                 pending_idx;
1689                         queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1690                                 queue->mmap_pages[pending_idx];
1691                         gnttab_set_unmap_op(gop,
1692                                             idx_to_kaddr(queue, pending_idx),
1693                                             GNTMAP_host_map,
1694                                             queue->grant_tx_handle[pending_idx]);
1695                         xenvif_grant_handle_reset(queue, pending_idx);
1696                         ++gop;
1697                 }
1698
1699         } while (dp != queue->dealloc_prod);
1700
1701         queue->dealloc_cons = dc;
1702
1703         if (gop - queue->tx_unmap_ops > 0) {
1704                 int ret;
1705                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1706                                         NULL,
1707                                         queue->pages_to_unmap,
1708                                         gop - queue->tx_unmap_ops);
1709                 if (ret) {
1710                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1711                                    gop - queue->tx_unmap_ops, ret);
1712                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1713                                 if (gop[i].status != GNTST_okay)
1714                                         netdev_err(queue->vif->dev,
1715                                                    " host_addr: 0x%llx handle: 0x%x status: %d\n",
1716                                                    gop[i].host_addr,
1717                                                    gop[i].handle,
1718                                                    gop[i].status);
1719                         }
1720                         BUG();
1721                 }
1722         }
1723
1724         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1725                 xenvif_idx_release(queue, pending_idx_release[i],
1726                                    XEN_NETIF_RSP_OKAY);
1727 }
1728
1729
1730 /* Called after netfront has transmitted */
1731 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1732 {
1733         unsigned nr_mops, nr_cops = 0;
1734         int work_done, ret;
1735
1736         if (unlikely(!tx_work_todo(queue)))
1737                 return 0;
1738
1739         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1740
1741         if (nr_cops == 0)
1742                 return 0;
1743
1744         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1745         if (nr_mops != 0) {
1746                 ret = gnttab_map_refs(queue->tx_map_ops,
1747                                       NULL,
1748                                       queue->pages_to_map,
1749                                       nr_mops);
1750                 BUG_ON(ret);
1751         }
1752
1753         work_done = xenvif_tx_submit(queue);
1754
1755         return work_done;
1756 }
1757
1758 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1759                                u8 status)
1760 {
1761         struct pending_tx_info *pending_tx_info;
1762         pending_ring_idx_t index;
1763         unsigned long flags;
1764
1765         pending_tx_info = &queue->pending_tx_info[pending_idx];
1766
1767         spin_lock_irqsave(&queue->response_lock, flags);
1768
1769         make_tx_response(queue, &pending_tx_info->req, status);
1770
1771         /* Release the pending index before pusing the Tx response so
1772          * its available before a new Tx request is pushed by the
1773          * frontend.
1774          */
1775         index = pending_index(queue->pending_prod++);
1776         queue->pending_ring[index] = pending_idx;
1777
1778         push_tx_responses(queue);
1779
1780         spin_unlock_irqrestore(&queue->response_lock, flags);
1781 }
1782
1783
1784 static void make_tx_response(struct xenvif_queue *queue,
1785                              struct xen_netif_tx_request *txp,
1786                              s8       st)
1787 {
1788         RING_IDX i = queue->tx.rsp_prod_pvt;
1789         struct xen_netif_tx_response *resp;
1790
1791         resp = RING_GET_RESPONSE(&queue->tx, i);
1792         resp->id     = txp->id;
1793         resp->status = st;
1794
1795         if (txp->flags & XEN_NETTXF_extra_info)
1796                 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1797
1798         queue->tx.rsp_prod_pvt = ++i;
1799 }
1800
1801 static void push_tx_responses(struct xenvif_queue *queue)
1802 {
1803         int notify;
1804
1805         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1806         if (notify)
1807                 notify_remote_via_irq(queue->tx_irq);
1808 }
1809
1810 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1811                                              u16      id,
1812                                              s8       st,
1813                                              u16      offset,
1814                                              u16      size,
1815                                              u16      flags)
1816 {
1817         RING_IDX i = queue->rx.rsp_prod_pvt;
1818         struct xen_netif_rx_response *resp;
1819
1820         resp = RING_GET_RESPONSE(&queue->rx, i);
1821         resp->offset     = offset;
1822         resp->flags      = flags;
1823         resp->id         = id;
1824         resp->status     = (s16)size;
1825         if (st < 0)
1826                 resp->status = (s16)st;
1827
1828         queue->rx.rsp_prod_pvt = ++i;
1829
1830         return resp;
1831 }
1832
1833 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1834 {
1835         int ret;
1836         struct gnttab_unmap_grant_ref tx_unmap_op;
1837
1838         gnttab_set_unmap_op(&tx_unmap_op,
1839                             idx_to_kaddr(queue, pending_idx),
1840                             GNTMAP_host_map,
1841                             queue->grant_tx_handle[pending_idx]);
1842         xenvif_grant_handle_reset(queue, pending_idx);
1843
1844         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1845                                 &queue->mmap_pages[pending_idx], 1);
1846         if (ret) {
1847                 netdev_err(queue->vif->dev,
1848                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1849                            ret,
1850                            pending_idx,
1851                            tx_unmap_op.host_addr,
1852                            tx_unmap_op.handle,
1853                            tx_unmap_op.status);
1854                 BUG();
1855         }
1856 }
1857
1858 static inline int tx_work_todo(struct xenvif_queue *queue)
1859 {
1860         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1861                 return 1;
1862
1863         return 0;
1864 }
1865
1866 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1867 {
1868         return queue->dealloc_cons != queue->dealloc_prod;
1869 }
1870
1871 void xenvif_unmap_frontend_rings(struct xenvif_queue *queue)
1872 {
1873         if (queue->tx.sring)
1874                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1875                                         queue->tx.sring);
1876         if (queue->rx.sring)
1877                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1878                                         queue->rx.sring);
1879 }
1880
1881 int xenvif_map_frontend_rings(struct xenvif_queue *queue,
1882                               grant_ref_t tx_ring_ref,
1883                               grant_ref_t rx_ring_ref)
1884 {
1885         void *addr;
1886         struct xen_netif_tx_sring *txs;
1887         struct xen_netif_rx_sring *rxs;
1888
1889         int err = -ENOMEM;
1890
1891         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1892                                      &tx_ring_ref, 1, &addr);
1893         if (err)
1894                 goto err;
1895
1896         txs = (struct xen_netif_tx_sring *)addr;
1897         BACK_RING_INIT(&queue->tx, txs, PAGE_SIZE);
1898
1899         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1900                                      &rx_ring_ref, 1, &addr);
1901         if (err)
1902                 goto err;
1903
1904         rxs = (struct xen_netif_rx_sring *)addr;
1905         BACK_RING_INIT(&queue->rx, rxs, PAGE_SIZE);
1906
1907         return 0;
1908
1909 err:
1910         xenvif_unmap_frontend_rings(queue);
1911         return err;
1912 }
1913
1914 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
1915 {
1916         struct xenvif *vif = queue->vif;
1917
1918         queue->stalled = true;
1919
1920         /* At least one queue has stalled? Disable the carrier. */
1921         spin_lock(&vif->lock);
1922         if (vif->stalled_queues++ == 0) {
1923                 netdev_info(vif->dev, "Guest Rx stalled");
1924                 netif_carrier_off(vif->dev);
1925         }
1926         spin_unlock(&vif->lock);
1927 }
1928
1929 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
1930 {
1931         struct xenvif *vif = queue->vif;
1932
1933         queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
1934         queue->stalled = false;
1935
1936         /* All queues are ready? Enable the carrier. */
1937         spin_lock(&vif->lock);
1938         if (--vif->stalled_queues == 0) {
1939                 netdev_info(vif->dev, "Guest Rx ready");
1940                 netif_carrier_on(vif->dev);
1941         }
1942         spin_unlock(&vif->lock);
1943 }
1944
1945 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
1946 {
1947         RING_IDX prod, cons;
1948
1949         prod = queue->rx.sring->req_prod;
1950         cons = queue->rx.req_cons;
1951
1952         return !queue->stalled && prod - cons < 1
1953                 && time_after(jiffies,
1954                               queue->last_rx_time + queue->vif->stall_timeout);
1955 }
1956
1957 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
1958 {
1959         RING_IDX prod, cons;
1960
1961         prod = queue->rx.sring->req_prod;
1962         cons = queue->rx.req_cons;
1963
1964         return queue->stalled && prod - cons >= 1;
1965 }
1966
1967 static bool xenvif_have_rx_work(struct xenvif_queue *queue)
1968 {
1969         return (!skb_queue_empty(&queue->rx_queue)
1970                 && xenvif_rx_ring_slots_available(queue))
1971                 || (queue->vif->stall_timeout &&
1972                     (xenvif_rx_queue_stalled(queue)
1973                      || xenvif_rx_queue_ready(queue)))
1974                 || kthread_should_stop()
1975                 || queue->vif->disabled;
1976 }
1977
1978 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
1979 {
1980         struct sk_buff *skb;
1981         long timeout;
1982
1983         skb = skb_peek(&queue->rx_queue);
1984         if (!skb)
1985                 return MAX_SCHEDULE_TIMEOUT;
1986
1987         timeout = XENVIF_RX_CB(skb)->expires - jiffies;
1988         return timeout < 0 ? 0 : timeout;
1989 }
1990
1991 /* Wait until the guest Rx thread has work.
1992  *
1993  * The timeout needs to be adjusted based on the current head of the
1994  * queue (and not just the head at the beginning).  In particular, if
1995  * the queue is initially empty an infinite timeout is used and this
1996  * needs to be reduced when a skb is queued.
1997  *
1998  * This cannot be done with wait_event_timeout() because it only
1999  * calculates the timeout once.
2000  */
2001 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
2002 {
2003         DEFINE_WAIT(wait);
2004
2005         if (xenvif_have_rx_work(queue))
2006                 return;
2007
2008         for (;;) {
2009                 long ret;
2010
2011                 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
2012                 if (xenvif_have_rx_work(queue))
2013                         break;
2014                 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
2015                 if (!ret)
2016                         break;
2017         }
2018         finish_wait(&queue->wq, &wait);
2019 }
2020
2021 int xenvif_kthread_guest_rx(void *data)
2022 {
2023         struct xenvif_queue *queue = data;
2024         struct xenvif *vif = queue->vif;
2025
2026         if (!vif->stall_timeout)
2027                 xenvif_queue_carrier_on(queue);
2028
2029         for (;;) {
2030                 xenvif_wait_for_rx_work(queue);
2031
2032                 if (kthread_should_stop())
2033                         break;
2034
2035                 /* This frontend is found to be rogue, disable it in
2036                  * kthread context. Currently this is only set when
2037                  * netback finds out frontend sends malformed packet,
2038                  * but we cannot disable the interface in softirq
2039                  * context so we defer it here, if this thread is
2040                  * associated with queue 0.
2041                  */
2042                 if (unlikely(vif->disabled && queue->id == 0)) {
2043                         xenvif_carrier_off(vif);
2044                         break;
2045                 }
2046
2047                 if (!skb_queue_empty(&queue->rx_queue))
2048                         xenvif_rx_action(queue);
2049
2050                 /* If the guest hasn't provided any Rx slots for a
2051                  * while it's probably not responsive, drop the
2052                  * carrier so packets are dropped earlier.
2053                  */
2054                 if (vif->stall_timeout) {
2055                         if (xenvif_rx_queue_stalled(queue))
2056                                 xenvif_queue_carrier_off(queue);
2057                         else if (xenvif_rx_queue_ready(queue))
2058                                 xenvif_queue_carrier_on(queue);
2059                 }
2060
2061                 /* Queued packets may have foreign pages from other
2062                  * domains.  These cannot be queued indefinitely as
2063                  * this would starve guests of grant refs and transmit
2064                  * slots.
2065                  */
2066                 xenvif_rx_queue_drop_expired(queue);
2067
2068                 xenvif_rx_queue_maybe_wake(queue);
2069
2070                 cond_resched();
2071         }
2072
2073         /* Bin any remaining skbs */
2074         xenvif_rx_queue_purge(queue);
2075
2076         return 0;
2077 }
2078
2079 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
2080 {
2081         /* Dealloc thread must remain running until all inflight
2082          * packets complete.
2083          */
2084         return kthread_should_stop() &&
2085                 !atomic_read(&queue->inflight_packets);
2086 }
2087
2088 int xenvif_dealloc_kthread(void *data)
2089 {
2090         struct xenvif_queue *queue = data;
2091
2092         for (;;) {
2093                 wait_event_interruptible(queue->dealloc_wq,
2094                                          tx_dealloc_work_todo(queue) ||
2095                                          xenvif_dealloc_kthread_should_stop(queue));
2096                 if (xenvif_dealloc_kthread_should_stop(queue))
2097                         break;
2098
2099                 xenvif_tx_dealloc_action(queue);
2100                 cond_resched();
2101         }
2102
2103         /* Unmap anything remaining*/
2104         if (tx_dealloc_work_todo(queue))
2105                 xenvif_tx_dealloc_action(queue);
2106
2107         return 0;
2108 }
2109
2110 static int __init netback_init(void)
2111 {
2112         int rc = 0;
2113
2114         if (!xen_domain())
2115                 return -ENODEV;
2116
2117         /* Allow as many queues as there are CPUs if user has not
2118          * specified a value.
2119          */
2120         if (xenvif_max_queues == 0)
2121                 xenvif_max_queues = num_online_cpus();
2122
2123         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2124                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2125                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2126                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2127         }
2128
2129         rc = xenvif_xenbus_init();
2130         if (rc)
2131                 goto failed_init;
2132
2133 #ifdef CONFIG_DEBUG_FS
2134         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2135         if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2136                 pr_warn("Init of debugfs returned %ld!\n",
2137                         PTR_ERR(xen_netback_dbg_root));
2138 #endif /* CONFIG_DEBUG_FS */
2139
2140         return 0;
2141
2142 failed_init:
2143         return rc;
2144 }
2145
2146 module_init(netback_init);
2147
2148 static void __exit netback_fini(void)
2149 {
2150 #ifdef CONFIG_DEBUG_FS
2151         if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2152                 debugfs_remove_recursive(xen_netback_dbg_root);
2153 #endif /* CONFIG_DEBUG_FS */
2154         xenvif_xenbus_fini();
2155 }
2156 module_exit(netback_fini);
2157
2158 MODULE_LICENSE("Dual BSD/GPL");
2159 MODULE_ALIAS("xen-backend:vif");