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                              unsigned int extra_count,
  99                              s8       st);
 100 static void push_tx_responses(struct xenvif_queue *queue);
 101
 102 static inline int tx_work_todo(struct xenvif_queue *queue);
 103
 104 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
 105                                              u16      id,
 106                                              s8       st,
 107                                              u16      offset,
 108                                              u16      size,
 109                                              u16      flags);
 110
 111 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
 112                                        u16 idx)
 113 {
 114         return page_to_pfn(queue->mmap_pages[idx]);
 115 }
 116
 117 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
 118                                          u16 idx)
 119 {
 120         return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
 121 }
 122
 123 #define callback_param(vif, pending_idx) \
 124         (vif->pending_tx_info[pending_idx].callback_struct)
 125
 126 /* Find the containing VIF's structure from a pointer in pending_tx_info array
 127  */
 128 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
 129 {
 130         u16 pending_idx = ubuf->desc;
 131         struct pending_tx_info *temp =
 132                 container_of(ubuf, struct pending_tx_info, callback_struct);
 133         return container_of(temp - pending_idx,
 134                             struct xenvif_queue,
 135                             pending_tx_info[0]);
 136 }
 137
 138 static u16 frag_get_pending_idx(skb_frag_t *frag)
 139 {
 140         return (u16)frag->page_offset;
 141 }
 142
 143 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 144 {
 145         frag->page_offset = pending_idx;
 146 }
 147
 148 static inline pending_ring_idx_t pending_index(unsigned i)
 149 {
 150         return i & (MAX_PENDING_REQS-1);
 151 }
 152
 153 static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
 154 {
 155         RING_IDX prod, cons;
 156         struct sk_buff *skb;
 157         int needed;
 158
 159         skb = skb_peek(&queue->rx_queue);
 160         if (!skb)
 161                 return false;
 162
 163         needed = DIV_ROUND_UP(skb->len, XEN_PAGE_SIZE);
 164         if (skb_is_gso(skb))
 165                 needed++;
 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         RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
 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 struct gop_frag_copy {
 278         struct xenvif_queue *queue;
 279         struct netrx_pending_operations *npo;
 280         struct xenvif_rx_meta *meta;
 281         int head;
 282         int gso_type;
 283
 284         struct page *page;
 285 };
 286
 287 static void xenvif_setup_copy_gop(unsigned long gfn,
 288                                   unsigned int offset,
 289                                   unsigned int *len,
 290                                   struct gop_frag_copy *info)
 291 {
 292         struct gnttab_copy *copy_gop;
 293         struct xen_page_foreign *foreign;
 294         /* Convenient aliases */
 295         struct xenvif_queue *queue = info->queue;
 296         struct netrx_pending_operations *npo = info->npo;
 297         struct page *page = info->page;
 298
 299         BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
 300
 301         if (npo->copy_off == MAX_BUFFER_OFFSET)
 302                 info->meta = get_next_rx_buffer(queue, npo);
 303
 304         if (npo->copy_off + *len > MAX_BUFFER_OFFSET)
 305                 *len = MAX_BUFFER_OFFSET - npo->copy_off;
 306
 307         copy_gop = npo->copy + npo->copy_prod++;
 308         copy_gop->flags = GNTCOPY_dest_gref;
 309         copy_gop->len = *len;
 310
 311         foreign = xen_page_foreign(page);
 312         if (foreign) {
 313                 copy_gop->source.domid = foreign->domid;
 314                 copy_gop->source.u.ref = foreign->gref;
 315                 copy_gop->flags |= GNTCOPY_source_gref;
 316         } else {
 317                 copy_gop->source.domid = DOMID_SELF;
 318                 copy_gop->source.u.gmfn = gfn;
 319         }
 320         copy_gop->source.offset = offset;
 321
 322         copy_gop->dest.domid = queue->vif->domid;
 323         copy_gop->dest.offset = npo->copy_off;
 324         copy_gop->dest.u.ref = npo->copy_gref;
 325
 326         npo->copy_off += *len;
 327         info->meta->size += *len;
 328
 329         /* Leave a gap for the GSO descriptor. */
 330         if (info->head && ((1 << info->gso_type) & queue->vif->gso_mask))
 331                 queue->rx.req_cons++;
 332
 333         info->head = 0; /* There must be something in this buffer now */
 334 }
 335
 336 static void xenvif_gop_frag_copy_grant(unsigned long gfn,
 337                                        unsigned offset,
 338                                        unsigned int len,
 339                                        void *data)
 340 {
 341         unsigned int bytes;
 342
 343         while (len) {
 344                 bytes = len;
 345                 xenvif_setup_copy_gop(gfn, offset, &bytes, data);
 346                 offset += bytes;
 347                 len -= bytes;
 348         }
 349 }
 350
 351 /*
 352  * Set up the grant operations for this fragment. If it's a flipping
 353  * interface, we also set up the unmap request from here.
 354  */
 355 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
 356                                  struct netrx_pending_operations *npo,
 357                                  struct page *page, unsigned long size,
 358                                  unsigned long offset, int *head)
 359 {
 360         struct gop_frag_copy info = {
 361                 .queue = queue,
 362                 .npo = npo,
 363                 .head = *head,
 364                 .gso_type = XEN_NETIF_GSO_TYPE_NONE,
 365         };
 366         unsigned long bytes;
 367
 368         if (skb_is_gso(skb)) {
 369                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 370                         info.gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
 371                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 372                         info.gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
 373         }
 374
 375         /* Data must not cross a page boundary. */
 376         BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
 377
 378         info.meta = npo->meta + npo->meta_prod - 1;
 379
 380         /* Skip unused frames from start of page */
 381         page += offset >> PAGE_SHIFT;
 382         offset &= ~PAGE_MASK;
 383
 384         while (size > 0) {
 385                 BUG_ON(offset >= PAGE_SIZE);
 386
 387                 bytes = PAGE_SIZE - offset;
 388                 if (bytes > size)
 389                         bytes = size;
 390
 391                 info.page = page;
 392                 gnttab_foreach_grant_in_range(page, offset, bytes,
 393                                               xenvif_gop_frag_copy_grant,
 394                                               &info);
 395                 size -= bytes;
 396                 offset = 0;
 397
 398                 /* Next page */
 399                 if (size) {
 400                         BUG_ON(!PageCompound(page));
 401                         page++;
 402                 }
 403         }
 404
 405         *head = info.head;
 406 }
 407
 408 /*
 409  * Prepare an SKB to be transmitted to the frontend.
 410  *
 411  * This function is responsible for allocating grant operations, meta
 412  * structures, etc.
 413  *
 414  * It returns the number of meta structures consumed. The number of
 415  * ring slots used is always equal to the number of meta slots used
 416  * plus the number of GSO descriptors used. Currently, we use either
 417  * zero GSO descriptors (for non-GSO packets) or one descriptor (for
 418  * frontend-side LRO).
 419  */
 420 static int xenvif_gop_skb(struct sk_buff *skb,
 421                           struct netrx_pending_operations *npo,
 422                           struct xenvif_queue *queue)
 423 {
 424         struct xenvif *vif = netdev_priv(skb->dev);
 425         int nr_frags = skb_shinfo(skb)->nr_frags;
 426         int i;
 427         struct xen_netif_rx_request req;
 428         struct xenvif_rx_meta *meta;
 429         unsigned char *data;
 430         int head = 1;
 431         int old_meta_prod;
 432         int gso_type;
 433
 434         old_meta_prod = npo->meta_prod;
 435
 436         gso_type = XEN_NETIF_GSO_TYPE_NONE;
 437         if (skb_is_gso(skb)) {
 438                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 439                         gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
 440                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 441                         gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
 442         }
 443
 444         /* Set up a GSO prefix descriptor, if necessary */
 445         if ((1 << gso_type) & vif->gso_prefix_mask) {
 446                 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
 447                 meta = npo->meta + npo->meta_prod++;
 448                 meta->gso_type = gso_type;
 449                 meta->gso_size = skb_shinfo(skb)->gso_size;
 450                 meta->size = 0;
 451                 meta->id = req.id;
 452         }
 453
 454         RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
 455         meta = npo->meta + npo->meta_prod++;
 456
 457         if ((1 << gso_type) & vif->gso_mask) {
 458                 meta->gso_type = gso_type;
 459                 meta->gso_size = skb_shinfo(skb)->gso_size;
 460         } else {
 461                 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
 462                 meta->gso_size = 0;
 463         }
 464
 465         meta->size = 0;
 466         meta->id = req.id;
 467         npo->copy_off = 0;
 468         npo->copy_gref = req.gref;
 469
 470         data = skb->data;
 471         while (data < skb_tail_pointer(skb)) {
 472                 unsigned int offset = offset_in_page(data);
 473                 unsigned int len = PAGE_SIZE - offset;
 474
 475                 if (data + len > skb_tail_pointer(skb))
 476                         len = skb_tail_pointer(skb) - data;
 477
 478                 xenvif_gop_frag_copy(queue, skb, npo,
 479                                      virt_to_page(data), len, offset, &head);
 480                 data += len;
 481         }
 482
 483         for (i = 0; i < nr_frags; i++) {
 484                 xenvif_gop_frag_copy(queue, skb, npo,
 485                                      skb_frag_page(&skb_shinfo(skb)->frags[i]),
 486                                      skb_frag_size(&skb_shinfo(skb)->frags[i]),
 487                                      skb_shinfo(skb)->frags[i].page_offset,
 488                                      &head);
 489         }
 490
 491         return npo->meta_prod - old_meta_prod;
 492 }
 493
 494 /*
 495  * This is a twin to xenvif_gop_skb.  Assume that xenvif_gop_skb was
 496  * used to set up the operations on the top of
 497  * netrx_pending_operations, which have since been done.  Check that
 498  * they didn't give any errors and advance over them.
 499  */
 500 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
 501                             struct netrx_pending_operations *npo)
 502 {
 503         struct gnttab_copy     *copy_op;
 504         int status = XEN_NETIF_RSP_OKAY;
 505         int i;
 506
 507         for (i = 0; i < nr_meta_slots; i++) {
 508                 copy_op = npo->copy + npo->copy_cons++;
 509                 if (copy_op->status != GNTST_okay) {
 510                         netdev_dbg(vif->dev,
 511                                    "Bad status %d from copy to DOM%d.\n",
 512                                    copy_op->status, vif->domid);
 513                         status = XEN_NETIF_RSP_ERROR;
 514                 }
 515         }
 516
 517         return status;
 518 }
 519
 520 static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status,
 521                                       struct xenvif_rx_meta *meta,
 522                                       int nr_meta_slots)
 523 {
 524         int i;
 525         unsigned long offset;
 526
 527         /* No fragments used */
 528         if (nr_meta_slots <= 1)
 529                 return;
 530
 531         nr_meta_slots--;
 532
 533         for (i = 0; i < nr_meta_slots; i++) {
 534                 int flags;
 535                 if (i == nr_meta_slots - 1)
 536                         flags = 0;
 537                 else
 538                         flags = XEN_NETRXF_more_data;
 539
 540                 offset = 0;
 541                 make_rx_response(queue, meta[i].id, status, offset,
 542                                  meta[i].size, flags);
 543         }
 544 }
 545
 546 void xenvif_kick_thread(struct xenvif_queue *queue)
 547 {
 548         wake_up(&queue->wq);
 549 }
 550
 551 static void xenvif_rx_action(struct xenvif_queue *queue)
 552 {
 553         s8 status;
 554         u16 flags;
 555         struct xen_netif_rx_response *resp;
 556         struct sk_buff_head rxq;
 557         struct sk_buff *skb;
 558         LIST_HEAD(notify);
 559         int ret;
 560         unsigned long offset;
 561         bool need_to_notify = false;
 562
 563         struct netrx_pending_operations npo = {
 564                 .copy  = queue->grant_copy_op,
 565                 .meta  = queue->meta,
 566         };
 567
 568         skb_queue_head_init(&rxq);
 569
 570         while (xenvif_rx_ring_slots_available(queue)
 571                && (skb = xenvif_rx_dequeue(queue)) != NULL) {
 572                 queue->last_rx_time = jiffies;
 573
 574                 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
 575
 576                 __skb_queue_tail(&rxq, skb);
 577         }
 578
 579         BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
 580
 581         if (!npo.copy_prod)
 582                 goto done;
 583
 584         BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
 585         gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
 586
 587         while ((skb = __skb_dequeue(&rxq)) != NULL) {
 588
 589                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 590                     queue->vif->gso_prefix_mask) {
 591                         resp = RING_GET_RESPONSE(&queue->rx,
 592                                                  queue->rx.rsp_prod_pvt++);
 593
 594                         resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
 595
 596                         resp->offset = queue->meta[npo.meta_cons].gso_size;
 597                         resp->id = queue->meta[npo.meta_cons].id;
 598                         resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
 599
 600                         npo.meta_cons++;
 601                         XENVIF_RX_CB(skb)->meta_slots_used--;
 602                 }
 603
 604
 605                 queue->stats.tx_bytes += skb->len;
 606                 queue->stats.tx_packets++;
 607
 608                 status = xenvif_check_gop(queue->vif,
 609                                           XENVIF_RX_CB(skb)->meta_slots_used,
 610                                           &npo);
 611
 612                 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
 613                         flags = 0;
 614                 else
 615                         flags = XEN_NETRXF_more_data;
 616
 617                 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
 618                         flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
 619                 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 620                         /* remote but checksummed. */
 621                         flags |= XEN_NETRXF_data_validated;
 622
 623                 offset = 0;
 624                 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
 625                                         status, offset,
 626                                         queue->meta[npo.meta_cons].size,
 627                                         flags);
 628
 629                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 630                     queue->vif->gso_mask) {
 631                         struct xen_netif_extra_info *gso =
 632                                 (struct xen_netif_extra_info *)
 633                                 RING_GET_RESPONSE(&queue->rx,
 634                                                   queue->rx.rsp_prod_pvt++);
 635
 636                         resp->flags |= XEN_NETRXF_extra_info;
 637
 638                         gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
 639                         gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
 640                         gso->u.gso.pad = 0;
 641                         gso->u.gso.features = 0;
 642
 643                         gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
 644                         gso->flags = 0;
 645                 }
 646
 647                 xenvif_add_frag_responses(queue, status,
 648                                           queue->meta + npo.meta_cons + 1,
 649                                           XENVIF_RX_CB(skb)->meta_slots_used);
 650
 651                 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
 652
 653                 need_to_notify |= !!ret;
 654
 655                 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
 656                 dev_kfree_skb(skb);
 657         }
 658
 659 done:
 660         if (need_to_notify)
 661                 notify_remote_via_irq(queue->rx_irq);
 662 }
 663
 664 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 665 {
 666         int more_to_do;
 667
 668         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 669
 670         if (more_to_do)
 671                 napi_schedule(&queue->napi);
 672 }
 673
 674 static void tx_add_credit(struct xenvif_queue *queue)
 675 {
 676         unsigned long max_burst, max_credit;
 677
 678         /*
 679          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 680          * Otherwise the interface can seize up due to insufficient credit.
 681          */
 682         max_burst = max(131072UL, queue->credit_bytes);
 683
 684         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 685         max_credit = queue->remaining_credit + queue->credit_bytes;
 686         if (max_credit < queue->remaining_credit)
 687                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 688
 689         queue->remaining_credit = min(max_credit, max_burst);
 690 }
 691
 692 void xenvif_tx_credit_callback(unsigned long data)
 693 {
 694         struct xenvif_queue *queue = (struct xenvif_queue *)data;
 695         tx_add_credit(queue);
 696         xenvif_napi_schedule_or_enable_events(queue);
 697 }
 698
 699 static void xenvif_tx_err(struct xenvif_queue *queue,
 700                           struct xen_netif_tx_request *txp,
 701                           unsigned int extra_count, RING_IDX end)
 702 {
 703         RING_IDX cons = queue->tx.req_cons;
 704         unsigned long flags;
 705
 706         do {
 707                 spin_lock_irqsave(&queue->response_lock, flags);
 708                 make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
 709                 push_tx_responses(queue);
 710                 spin_unlock_irqrestore(&queue->response_lock, flags);
 711                 if (cons == end)
 712                         break;
 713                 RING_COPY_REQUEST(&queue->tx, cons++, txp);
 714                 extra_count = 0; /* only the first frag can have extras */
 715         } while (1);
 716         queue->tx.req_cons = cons;
 717 }
 718
 719 static void xenvif_fatal_tx_err(struct xenvif *vif)
 720 {
 721         netdev_err(vif->dev, "fatal error; disabling device\n");
 722         vif->disabled = true;
 723         /* Disable the vif from queue 0's kthread */
 724         if (vif->queues)
 725                 xenvif_kick_thread(&vif->queues[0]);
 726 }
 727
 728 static int xenvif_count_requests(struct xenvif_queue *queue,
 729                                  struct xen_netif_tx_request *first,
 730                                  unsigned int extra_count,
 731                                  struct xen_netif_tx_request *txp,
 732                                  int work_to_do)
 733 {
 734         RING_IDX cons = queue->tx.req_cons;
 735         int slots = 0;
 736         int drop_err = 0;
 737         int more_data;
 738
 739         if (!(first->flags & XEN_NETTXF_more_data))
 740                 return 0;
 741
 742         do {
 743                 struct xen_netif_tx_request dropped_tx = { 0 };
 744
 745                 if (slots >= work_to_do) {
 746                         netdev_err(queue->vif->dev,
 747                                    "Asked for %d slots but exceeds this limit\n",
 748                                    work_to_do);
 749                         xenvif_fatal_tx_err(queue->vif);
 750                         return -ENODATA;
 751                 }
 752
 753                 /* This guest is really using too many slots and
 754                  * considered malicious.
 755                  */
 756                 if (unlikely(slots >= fatal_skb_slots)) {
 757                         netdev_err(queue->vif->dev,
 758                                    "Malicious frontend using %d slots, threshold %u\n",
 759                                    slots, fatal_skb_slots);
 760                         xenvif_fatal_tx_err(queue->vif);
 761                         return -E2BIG;
 762                 }
 763
 764                 /* Xen network protocol had implicit dependency on
 765                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 766                  * the historical MAX_SKB_FRAGS value 18 to honor the
 767                  * same behavior as before. Any packet using more than
 768                  * 18 slots but less than fatal_skb_slots slots is
 769                  * dropped
 770                  */
 771                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 772                         if (net_ratelimit())
 773                                 netdev_dbg(queue->vif->dev,
 774                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 775                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 776                         drop_err = -E2BIG;
 777                 }
 778
 779                 if (drop_err)
 780                         txp = &dropped_tx;
 781
 782                 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
 783
 784                 /* If the guest submitted a frame >= 64 KiB then
 785                  * first->size overflowed and following slots will
 786                  * appear to be larger than the frame.
 787                  *
 788                  * This cannot be fatal error as there are buggy
 789                  * frontends that do this.
 790                  *
 791                  * Consume all slots and drop the packet.
 792                  */
 793                 if (!drop_err && txp->size > first->size) {
 794                         if (net_ratelimit())
 795                                 netdev_dbg(queue->vif->dev,
 796                                            "Invalid tx request, slot size %u > remaining size %u\n",
 797                                            txp->size, first->size);
 798                         drop_err = -EIO;
 799                 }
 800
 801                 first->size -= txp->size;
 802                 slots++;
 803
 804                 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
 805                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 806                                  txp->offset, txp->size);
 807                         xenvif_fatal_tx_err(queue->vif);
 808                         return -EINVAL;
 809                 }
 810
 811                 more_data = txp->flags & XEN_NETTXF_more_data;
 812
 813                 if (!drop_err)
 814                         txp++;
 815
 816         } while (more_data);
 817
 818         if (drop_err) {
 819                 xenvif_tx_err(queue, first, extra_count, cons + slots);
 820                 return drop_err;
 821         }
 822
 823         return slots;
 824 }
 825
 826
 827 struct xenvif_tx_cb {
 828         u16 pending_idx;
 829 };
 830
 831 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 832
 833 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 834                                            u16 pending_idx,
 835                                            struct xen_netif_tx_request *txp,
 836                                            unsigned int extra_count,
 837                                            struct gnttab_map_grant_ref *mop)
 838 {
 839         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 840         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 841                           GNTMAP_host_map | GNTMAP_readonly,
 842                           txp->gref, queue->vif->domid);
 843
 844         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 845                sizeof(*txp));
 846         queue->pending_tx_info[pending_idx].extra_count = extra_count;
 847 }
 848
 849 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 850 {
 851         struct sk_buff *skb =
 852                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 853                           GFP_ATOMIC | __GFP_NOWARN);
 854         if (unlikely(skb == NULL))
 855                 return NULL;
 856
 857         /* Packets passed to netif_rx() must have some headroom. */
 858         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 859
 860         /* Initialize it here to avoid later surprises */
 861         skb_shinfo(skb)->destructor_arg = NULL;
 862
 863         return skb;
 864 }
 865
 866 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
 867                                                         struct sk_buff *skb,
 868                                                         struct xen_netif_tx_request *txp,
 869                                                         struct gnttab_map_grant_ref *gop,
 870                                                         unsigned int frag_overflow,
 871                                                         struct sk_buff *nskb)
 872 {
 873         struct skb_shared_info *shinfo = skb_shinfo(skb);
 874         skb_frag_t *frags = shinfo->frags;
 875         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 876         int start;
 877         pending_ring_idx_t index;
 878         unsigned int nr_slots;
 879
 880         nr_slots = shinfo->nr_frags;
 881
 882         /* Skip first skb fragment if it is on same page as header fragment. */
 883         start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
 884
 885         for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
 886              shinfo->nr_frags++, txp++, gop++) {
 887                 index = pending_index(queue->pending_cons++);
 888                 pending_idx = queue->pending_ring[index];
 889                 xenvif_tx_create_map_op(queue, pending_idx, txp, 0, gop);
 890                 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 891         }
 892
 893         if (frag_overflow) {
 894
 895                 shinfo = skb_shinfo(nskb);
 896                 frags = shinfo->frags;
 897
 898                 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
 899                      shinfo->nr_frags++, txp++, gop++) {
 900                         index = pending_index(queue->pending_cons++);
 901                         pending_idx = queue->pending_ring[index];
 902                         xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
 903                                                 gop);
 904                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 905                                              pending_idx);
 906                 }
 907
 908                 skb_shinfo(skb)->frag_list = nskb;
 909         }
 910
 911         return gop;
 912 }
 913
 914 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 915                                            u16 pending_idx,
 916                                            grant_handle_t handle)
 917 {
 918         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 919                      NETBACK_INVALID_HANDLE)) {
 920                 netdev_err(queue->vif->dev,
 921                            "Trying to overwrite active handle! pending_idx: 0x%x\n",
 922                            pending_idx);
 923                 BUG();
 924         }
 925         queue->grant_tx_handle[pending_idx] = handle;
 926 }
 927
 928 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 929                                              u16 pending_idx)
 930 {
 931         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 932                      NETBACK_INVALID_HANDLE)) {
 933                 netdev_err(queue->vif->dev,
 934                            "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 935                            pending_idx);
 936                 BUG();
 937         }
 938         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 939 }
 940
 941 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 942                                struct sk_buff *skb,
 943                                struct gnttab_map_grant_ref **gopp_map,
 944                                struct gnttab_copy **gopp_copy)
 945 {
 946         struct gnttab_map_grant_ref *gop_map = *gopp_map;
 947         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 948         /* This always points to the shinfo of the skb being checked, which
 949          * could be either the first or the one on the frag_list
 950          */
 951         struct skb_shared_info *shinfo = skb_shinfo(skb);
 952         /* If this is non-NULL, we are currently checking the frag_list skb, and
 953          * this points to the shinfo of the first one
 954          */
 955         struct skb_shared_info *first_shinfo = NULL;
 956         int nr_frags = shinfo->nr_frags;
 957         const bool sharedslot = nr_frags &&
 958                                 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
 959         int i, err;
 960
 961         /* Check status of header. */
 962         err = (*gopp_copy)->status;
 963         if (unlikely(err)) {
 964                 if (net_ratelimit())
 965                         netdev_dbg(queue->vif->dev,
 966                                    "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 967                                    (*gopp_copy)->status,
 968                                    pending_idx,
 969                                    (*gopp_copy)->source.u.ref);
 970                 /* The first frag might still have this slot mapped */
 971                 if (!sharedslot)
 972                         xenvif_idx_release(queue, pending_idx,
 973                                            XEN_NETIF_RSP_ERROR);
 974         }
 975         (*gopp_copy)++;
 976
 977 check_frags:
 978         for (i = 0; i < nr_frags; i++, gop_map++) {
 979                 int j, newerr;
 980
 981                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 982
 983                 /* Check error status: if okay then remember grant handle. */
 984                 newerr = gop_map->status;
 985
 986                 if (likely(!newerr)) {
 987                         xenvif_grant_handle_set(queue,
 988                                                 pending_idx,
 989                                                 gop_map->handle);
 990                         /* Had a previous error? Invalidate this fragment. */
 991                         if (unlikely(err)) {
 992                                 xenvif_idx_unmap(queue, pending_idx);
 993                                 /* If the mapping of the first frag was OK, but
 994                                  * the header's copy failed, and they are
 995                                  * sharing a slot, send an error
 996                                  */
 997                                 if (i == 0 && sharedslot)
 998                                         xenvif_idx_release(queue, pending_idx,
 999                                                            XEN_NETIF_RSP_ERROR);
1000                                 else
1001                                         xenvif_idx_release(queue, pending_idx,
1002                                                            XEN_NETIF_RSP_OKAY);
1003                         }
1004                         continue;
1005                 }
1006
1007                 /* Error on this fragment: respond to client with an error. */
1008                 if (net_ratelimit())
1009                         netdev_dbg(queue->vif->dev,
1010                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
1011                                    i,
1012                                    gop_map->status,
1013                                    pending_idx,
1014                                    gop_map->ref);
1015
1016                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
1017
1018                 /* Not the first error? Preceding frags already invalidated. */
1019                 if (err)
1020                         continue;
1021
1022                 /* First error: if the header haven't shared a slot with the
1023                  * first frag, release it as well.
1024                  */
1025                 if (!sharedslot)
1026                         xenvif_idx_release(queue,
1027                                            XENVIF_TX_CB(skb)->pending_idx,
1028                                            XEN_NETIF_RSP_OKAY);
1029
1030                 /* Invalidate preceding fragments of this skb. */
1031                 for (j = 0; j < i; j++) {
1032                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1033                         xenvif_idx_unmap(queue, pending_idx);
1034                         xenvif_idx_release(queue, pending_idx,
1035                                            XEN_NETIF_RSP_OKAY);
1036                 }
1037
1038                 /* And if we found the error while checking the frag_list, unmap
1039                  * the first skb's frags
1040                  */
1041                 if (first_shinfo) {
1042                         for (j = 0; j < first_shinfo->nr_frags; j++) {
1043                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
1044                                 xenvif_idx_unmap(queue, pending_idx);
1045                                 xenvif_idx_release(queue, pending_idx,
1046                                                    XEN_NETIF_RSP_OKAY);
1047                         }
1048                 }
1049
1050                 /* Remember the error: invalidate all subsequent fragments. */
1051                 err = newerr;
1052         }
1053
1054         if (skb_has_frag_list(skb) && !first_shinfo) {
1055                 first_shinfo = skb_shinfo(skb);
1056                 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1057                 nr_frags = shinfo->nr_frags;
1058
1059                 goto check_frags;
1060         }
1061
1062         *gopp_map = gop_map;
1063         return err;
1064 }
1065
1066 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1067 {
1068         struct skb_shared_info *shinfo = skb_shinfo(skb);
1069         int nr_frags = shinfo->nr_frags;
1070         int i;
1071         u16 prev_pending_idx = INVALID_PENDING_IDX;
1072
1073         for (i = 0; i < nr_frags; i++) {
1074                 skb_frag_t *frag = shinfo->frags + i;
1075                 struct xen_netif_tx_request *txp;
1076                 struct page *page;
1077                 u16 pending_idx;
1078
1079                 pending_idx = frag_get_pending_idx(frag);
1080
1081                 /* If this is not the first frag, chain it to the previous*/
1082                 if (prev_pending_idx == INVALID_PENDING_IDX)
1083                         skb_shinfo(skb)->destructor_arg =
1084                                 &callback_param(queue, pending_idx);
1085                 else
1086                         callback_param(queue, prev_pending_idx).ctx =
1087                                 &callback_param(queue, pending_idx);
1088
1089                 callback_param(queue, pending_idx).ctx = NULL;
1090                 prev_pending_idx = pending_idx;
1091
1092                 txp = &queue->pending_tx_info[pending_idx].req;
1093                 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1094                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1095                 skb->len += txp->size;
1096                 skb->data_len += txp->size;
1097                 skb->truesize += txp->size;
1098
1099                 /* Take an extra reference to offset network stack's put_page */
1100                 get_page(queue->mmap_pages[pending_idx]);
1101         }
1102 }
1103
1104 static int xenvif_get_extras(struct xenvif_queue *queue,
1105                              struct xen_netif_extra_info *extras,
1106                              unsigned int *extra_count,
1107                              int work_to_do)
1108 {
1109         struct xen_netif_extra_info extra;
1110         RING_IDX cons = queue->tx.req_cons;
1111
1112         do {
1113                 if (unlikely(work_to_do-- <= 0)) {
1114                         netdev_err(queue->vif->dev, "Missing extra info\n");
1115                         xenvif_fatal_tx_err(queue->vif);
1116                         return -EBADR;
1117                 }
1118
1119                 RING_COPY_REQUEST(&queue->tx, cons, &extra);
1120
1121                 queue->tx.req_cons = ++cons;
1122                 (*extra_count)++;
1123
1124                 if (unlikely(!extra.type ||
1125                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1126                         netdev_err(queue->vif->dev,
1127                                    "Invalid extra type: %d\n", extra.type);
1128                         xenvif_fatal_tx_err(queue->vif);
1129                         return -EINVAL;
1130                 }
1131
1132                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1133         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1134
1135         return work_to_do;
1136 }
1137
1138 static int xenvif_set_skb_gso(struct xenvif *vif,
1139                               struct sk_buff *skb,
1140                               struct xen_netif_extra_info *gso)
1141 {
1142         if (!gso->u.gso.size) {
1143                 netdev_err(vif->dev, "GSO size must not be zero.\n");
1144                 xenvif_fatal_tx_err(vif);
1145                 return -EINVAL;
1146         }
1147
1148         switch (gso->u.gso.type) {
1149         case XEN_NETIF_GSO_TYPE_TCPV4:
1150                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1151                 break;
1152         case XEN_NETIF_GSO_TYPE_TCPV6:
1153                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1154                 break;
1155         default:
1156                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1157                 xenvif_fatal_tx_err(vif);
1158                 return -EINVAL;
1159         }
1160
1161         skb_shinfo(skb)->gso_size = gso->u.gso.size;
1162         /* gso_segs will be calculated later */
1163
1164         return 0;
1165 }
1166
1167 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1168 {
1169         bool recalculate_partial_csum = false;
1170
1171         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1172          * peers can fail to set NETRXF_csum_blank when sending a GSO
1173          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1174          * recalculate the partial checksum.
1175          */
1176         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1177                 queue->stats.rx_gso_checksum_fixup++;
1178                 skb->ip_summed = CHECKSUM_PARTIAL;
1179                 recalculate_partial_csum = true;
1180         }
1181
1182         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1183         if (skb->ip_summed != CHECKSUM_PARTIAL)
1184                 return 0;
1185
1186         return skb_checksum_setup(skb, recalculate_partial_csum);
1187 }
1188
1189 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1190 {
1191         u64 now = get_jiffies_64();
1192         u64 next_credit = queue->credit_window_start +
1193                 msecs_to_jiffies(queue->credit_usec / 1000);
1194
1195         /* Timer could already be pending in rare cases. */
1196         if (timer_pending(&queue->credit_timeout))
1197                 return true;
1198
1199         /* Passed the point where we can replenish credit? */
1200         if (time_after_eq64(now, next_credit)) {
1201                 queue->credit_window_start = now;
1202                 tx_add_credit(queue);
1203         }
1204
1205         /* Still too big to send right now? Set a callback. */
1206         if (size > queue->remaining_credit) {
1207                 queue->credit_timeout.data     =
1208                         (unsigned long)queue;
1209                 mod_timer(&queue->credit_timeout,
1210                           next_credit);
1211                 queue->credit_window_start = next_credit;
1212
1213                 return true;
1214         }
1215
1216         return false;
1217 }
1218
1219 /* No locking is required in xenvif_mcast_add/del() as they are
1220  * only ever invoked from NAPI poll. An RCU list is used because
1221  * xenvif_mcast_match() is called asynchronously, during start_xmit.
1222  */
1223
1224 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
1225 {
1226         struct xenvif_mcast_addr *mcast;
1227
1228         if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
1229                 if (net_ratelimit())
1230                         netdev_err(vif->dev,
1231                                    "Too many multicast addresses\n");
1232                 return -ENOSPC;
1233         }
1234
1235         mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
1236         if (!mcast)
1237                 return -ENOMEM;
1238
1239         ether_addr_copy(mcast->addr, addr);
1240         list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
1241         vif->fe_mcast_count++;
1242
1243         return 0;
1244 }
1245
1246 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
1247 {
1248         struct xenvif_mcast_addr *mcast;
1249
1250         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1251                 if (ether_addr_equal(addr, mcast->addr)) {
1252                         --vif->fe_mcast_count;
1253                         list_del_rcu(&mcast->entry);
1254                         kfree_rcu(mcast, rcu);
1255                         break;
1256                 }
1257         }
1258 }
1259
1260 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
1261 {
1262         struct xenvif_mcast_addr *mcast;
1263
1264         rcu_read_lock();
1265         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1266                 if (ether_addr_equal(addr, mcast->addr)) {
1267                         rcu_read_unlock();
1268                         return true;
1269                 }
1270         }
1271         rcu_read_unlock();
1272
1273         return false;
1274 }
1275
1276 void xenvif_mcast_addr_list_free(struct xenvif *vif)
1277 {
1278         /* No need for locking or RCU here. NAPI poll and TX queue
1279          * are stopped.
1280          */
1281         while (!list_empty(&vif->fe_mcast_addr)) {
1282                 struct xenvif_mcast_addr *mcast;
1283
1284                 mcast = list_first_entry(&vif->fe_mcast_addr,
1285                                          struct xenvif_mcast_addr,
1286                                          entry);
1287                 --vif->fe_mcast_count;
1288                 list_del(&mcast->entry);
1289                 kfree(mcast);
1290         }
1291 }
1292
1293 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1294                                      int budget,
1295                                      unsigned *copy_ops,
1296                                      unsigned *map_ops)
1297 {
1298         struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
1299         struct sk_buff *skb, *nskb;
1300         int ret;
1301         unsigned int frag_overflow;
1302
1303         while (skb_queue_len(&queue->tx_queue) < budget) {
1304                 struct xen_netif_tx_request txreq;
1305                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1306                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1307                 unsigned int extra_count;
1308                 u16 pending_idx;
1309                 RING_IDX idx;
1310                 int work_to_do;
1311                 unsigned int data_len;
1312                 pending_ring_idx_t index;
1313
1314                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1315                     XEN_NETIF_TX_RING_SIZE) {
1316                         netdev_err(queue->vif->dev,
1317                                    "Impossible number of requests. "
1318                                    "req_prod %d, req_cons %d, size %ld\n",
1319                                    queue->tx.sring->req_prod, queue->tx.req_cons,
1320                                    XEN_NETIF_TX_RING_SIZE);
1321                         xenvif_fatal_tx_err(queue->vif);
1322                         break;
1323                 }
1324
1325                 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1326                 if (!work_to_do)
1327                         break;
1328
1329                 idx = queue->tx.req_cons;
1330                 rmb(); /* Ensure that we see the request before we copy it. */
1331                 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
1332
1333                 /* Credit-based scheduling. */
1334                 if (txreq.size > queue->remaining_credit &&
1335                     tx_credit_exceeded(queue, txreq.size))
1336                         break;
1337
1338                 queue->remaining_credit -= txreq.size;
1339
1340                 work_to_do--;
1341                 queue->tx.req_cons = ++idx;
1342
1343                 memset(extras, 0, sizeof(extras));
1344                 extra_count = 0;
1345                 if (txreq.flags & XEN_NETTXF_extra_info) {
1346                         work_to_do = xenvif_get_extras(queue, extras,
1347                                                        &extra_count,
1348                                                        work_to_do);
1349                         idx = queue->tx.req_cons;
1350                         if (unlikely(work_to_do < 0))
1351                                 break;
1352                 }
1353
1354                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
1355                         struct xen_netif_extra_info *extra;
1356
1357                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
1358                         ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
1359
1360                         make_tx_response(queue, &txreq, extra_count,
1361                                          (ret == 0) ?
1362                                          XEN_NETIF_RSP_OKAY :
1363                                          XEN_NETIF_RSP_ERROR);
1364                         push_tx_responses(queue);
1365                         continue;
1366                 }
1367
1368                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
1369                         struct xen_netif_extra_info *extra;
1370
1371                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
1372                         xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
1373
1374                         make_tx_response(queue, &txreq, extra_count,
1375                                          XEN_NETIF_RSP_OKAY);
1376                         push_tx_responses(queue);
1377                         continue;
1378                 }
1379
1380                 ret = xenvif_count_requests(queue, &txreq, extra_count,
1381                                             txfrags, work_to_do);
1382                 if (unlikely(ret < 0))
1383                         break;
1384
1385                 idx += ret;
1386
1387                 if (unlikely(txreq.size < ETH_HLEN)) {
1388                         netdev_dbg(queue->vif->dev,
1389                                    "Bad packet size: %d\n", txreq.size);
1390                         xenvif_tx_err(queue, &txreq, extra_count, idx);
1391                         break;
1392                 }
1393
1394                 /* No crossing a page as the payload mustn't fragment. */
1395                 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
1396                         netdev_err(queue->vif->dev,
1397                                    "txreq.offset: %u, size: %u, end: %lu\n",
1398                                    txreq.offset, txreq.size,
1399                                    (unsigned long)(txreq.offset&~XEN_PAGE_MASK) + txreq.size);
1400                         xenvif_fatal_tx_err(queue->vif);
1401                         break;
1402                 }
1403
1404                 index = pending_index(queue->pending_cons);
1405                 pending_idx = queue->pending_ring[index];
1406
1407                 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
1408                             ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1409                         XEN_NETBACK_TX_COPY_LEN : txreq.size;
1410
1411                 skb = xenvif_alloc_skb(data_len);
1412                 if (unlikely(skb == NULL)) {
1413                         netdev_dbg(queue->vif->dev,
1414                                    "Can't allocate a skb in start_xmit.\n");
1415                         xenvif_tx_err(queue, &txreq, extra_count, idx);
1416                         break;
1417                 }
1418
1419                 skb_shinfo(skb)->nr_frags = ret;
1420                 if (data_len < txreq.size)
1421                         skb_shinfo(skb)->nr_frags++;
1422                 /* At this point shinfo->nr_frags is in fact the number of
1423                  * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1424                  */
1425                 frag_overflow = 0;
1426                 nskb = NULL;
1427                 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1428                         frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1429                         BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1430                         skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1431                         nskb = xenvif_alloc_skb(0);
1432                         if (unlikely(nskb == NULL)) {
1433                                 kfree_skb(skb);
1434                                 xenvif_tx_err(queue, &txreq, extra_count, idx);
1435                                 if (net_ratelimit())
1436                                         netdev_err(queue->vif->dev,
1437                                                    "Can't allocate the frag_list skb.\n");
1438                                 break;
1439                         }
1440                 }
1441
1442                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1443                         struct xen_netif_extra_info *gso;
1444                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1445
1446                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1447                                 /* Failure in xenvif_set_skb_gso is fatal. */
1448                                 kfree_skb(skb);
1449                                 kfree_skb(nskb);
1450                                 break;
1451                         }
1452                 }
1453
1454                 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1455
1456                 __skb_put(skb, data_len);
1457                 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1458                 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1459                 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1460
1461                 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1462                         virt_to_gfn(skb->data);
1463                 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1464                 queue->tx_copy_ops[*copy_ops].dest.offset =
1465                         offset_in_page(skb->data) & ~XEN_PAGE_MASK;
1466
1467                 queue->tx_copy_ops[*copy_ops].len = data_len;
1468                 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1469
1470                 (*copy_ops)++;
1471
1472                 if (data_len < txreq.size) {
1473                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1474                                              pending_idx);
1475                         xenvif_tx_create_map_op(queue, pending_idx, &txreq,
1476                                                 extra_count, gop);
1477                         gop++;
1478                 } else {
1479                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1480                                              INVALID_PENDING_IDX);
1481                         memcpy(&queue->pending_tx_info[pending_idx].req,
1482                                &txreq, sizeof(txreq));
1483                         queue->pending_tx_info[pending_idx].extra_count =
1484                                 extra_count;
1485                 }
1486
1487                 queue->pending_cons++;
1488
1489                 gop = xenvif_get_requests(queue, skb, txfrags, gop,
1490                                           frag_overflow, nskb);
1491
1492                 __skb_queue_tail(&queue->tx_queue, skb);
1493
1494                 queue->tx.req_cons = idx;
1495
1496                 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1497                     (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1498                         break;
1499         }
1500
1501         (*map_ops) = gop - queue->tx_map_ops;
1502         return;
1503 }
1504
1505 /* Consolidate skb with a frag_list into a brand new one with local pages on
1506  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1507  */
1508 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1509 {
1510         unsigned int offset = skb_headlen(skb);
1511         skb_frag_t frags[MAX_SKB_FRAGS];
1512         int i, f;
1513         struct ubuf_info *uarg;
1514         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1515
1516         queue->stats.tx_zerocopy_sent += 2;
1517         queue->stats.tx_frag_overflow++;
1518
1519         xenvif_fill_frags(queue, nskb);
1520         /* Subtract frags size, we will correct it later */
1521         skb->truesize -= skb->data_len;
1522         skb->len += nskb->len;
1523         skb->data_len += nskb->len;
1524
1525         /* create a brand new frags array and coalesce there */
1526         for (i = 0; offset < skb->len; i++) {
1527                 struct page *page;
1528                 unsigned int len;
1529
1530                 BUG_ON(i >= MAX_SKB_FRAGS);
1531                 page = alloc_page(GFP_ATOMIC);
1532                 if (!page) {
1533                         int j;
1534                         skb->truesize += skb->data_len;
1535                         for (j = 0; j < i; j++)
1536                                 put_page(frags[j].page.p);
1537                         return -ENOMEM;
1538                 }
1539
1540                 if (offset + PAGE_SIZE < skb->len)
1541                         len = PAGE_SIZE;
1542                 else
1543                         len = skb->len - offset;
1544                 if (skb_copy_bits(skb, offset, page_address(page), len))
1545                         BUG();
1546
1547                 offset += len;
1548                 frags[i].page.p = page;
1549                 frags[i].page_offset = 0;
1550                 skb_frag_size_set(&frags[i], len);
1551         }
1552
1553         /* Copied all the bits from the frag list -- free it. */
1554         skb_frag_list_init(skb);
1555         xenvif_skb_zerocopy_prepare(queue, nskb);
1556         kfree_skb(nskb);
1557
1558         /* Release all the original (foreign) frags. */
1559         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1560                 skb_frag_unref(skb, f);
1561         uarg = skb_shinfo(skb)->destructor_arg;
1562         /* increase inflight counter to offset decrement in callback */
1563         atomic_inc(&queue->inflight_packets);
1564         uarg->callback(uarg, true);
1565         skb_shinfo(skb)->destructor_arg = NULL;
1566
1567         /* Fill the skb with the new (local) frags. */
1568         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1569         skb_shinfo(skb)->nr_frags = i;
1570         skb->truesize += i * PAGE_SIZE;
1571
1572         return 0;
1573 }
1574
1575 static int xenvif_tx_submit(struct xenvif_queue *queue)
1576 {
1577         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1578         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1579         struct sk_buff *skb;
1580         int work_done = 0;
1581
1582         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1583                 struct xen_netif_tx_request *txp;
1584                 u16 pending_idx;
1585                 unsigned data_len;
1586
1587                 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1588                 txp = &queue->pending_tx_info[pending_idx].req;
1589
1590                 /* Check the remap error code. */
1591                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1592                         /* If there was an error, xenvif_tx_check_gop is
1593                          * expected to release all the frags which were mapped,
1594                          * so kfree_skb shouldn't do it again
1595                          */
1596                         skb_shinfo(skb)->nr_frags = 0;
1597                         if (skb_has_frag_list(skb)) {
1598                                 struct sk_buff *nskb =
1599                                                 skb_shinfo(skb)->frag_list;
1600                                 skb_shinfo(nskb)->nr_frags = 0;
1601                         }
1602                         kfree_skb(skb);
1603                         continue;
1604                 }
1605
1606                 data_len = skb->len;
1607                 callback_param(queue, pending_idx).ctx = NULL;
1608                 if (data_len < txp->size) {
1609                         /* Append the packet payload as a fragment. */
1610                         txp->offset += data_len;
1611                         txp->size -= data_len;
1612                 } else {
1613                         /* Schedule a response immediately. */
1614                         xenvif_idx_release(queue, pending_idx,
1615                                            XEN_NETIF_RSP_OKAY);
1616                 }
1617
1618                 if (txp->flags & XEN_NETTXF_csum_blank)
1619                         skb->ip_summed = CHECKSUM_PARTIAL;
1620                 else if (txp->flags & XEN_NETTXF_data_validated)
1621                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1622
1623                 xenvif_fill_frags(queue, skb);
1624
1625                 if (unlikely(skb_has_frag_list(skb))) {
1626                         if (xenvif_handle_frag_list(queue, skb)) {
1627                                 if (net_ratelimit())
1628                                         netdev_err(queue->vif->dev,
1629                                                    "Not enough memory to consolidate frag_list!\n");
1630                                 xenvif_skb_zerocopy_prepare(queue, skb);
1631                                 kfree_skb(skb);
1632                                 continue;
1633                         }
1634                 }
1635
1636                 skb->dev      = queue->vif->dev;
1637                 skb->protocol = eth_type_trans(skb, skb->dev);
1638                 skb_reset_network_header(skb);
1639
1640                 if (checksum_setup(queue, skb)) {
1641                         netdev_dbg(queue->vif->dev,
1642                                    "Can't setup checksum in net_tx_action\n");
1643                         /* We have to set this flag to trigger the callback */
1644                         if (skb_shinfo(skb)->destructor_arg)
1645                                 xenvif_skb_zerocopy_prepare(queue, skb);
1646                         kfree_skb(skb);
1647                         continue;
1648                 }
1649
1650                 skb_probe_transport_header(skb, 0);
1651
1652                 /* If the packet is GSO then we will have just set up the
1653                  * transport header offset in checksum_setup so it's now
1654                  * straightforward to calculate gso_segs.
1655                  */
1656                 if (skb_is_gso(skb)) {
1657                         int mss = skb_shinfo(skb)->gso_size;
1658                         int hdrlen = skb_transport_header(skb) -
1659                                 skb_mac_header(skb) +
1660                                 tcp_hdrlen(skb);
1661
1662                         skb_shinfo(skb)->gso_segs =
1663                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1664                 }
1665
1666                 queue->stats.rx_bytes += skb->len;
1667                 queue->stats.rx_packets++;
1668
1669                 work_done++;
1670
1671                 /* Set this flag right before netif_receive_skb, otherwise
1672                  * someone might think this packet already left netback, and
1673                  * do a skb_copy_ubufs while we are still in control of the
1674                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1675                  */
1676                 if (skb_shinfo(skb)->destructor_arg) {
1677                         xenvif_skb_zerocopy_prepare(queue, skb);
1678                         queue->stats.tx_zerocopy_sent++;
1679                 }
1680
1681                 netif_receive_skb(skb);
1682         }
1683
1684         return work_done;
1685 }
1686
1687 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1688 {
1689         unsigned long flags;
1690         pending_ring_idx_t index;
1691         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1692
1693         /* This is the only place where we grab this lock, to protect callbacks
1694          * from each other.
1695          */
1696         spin_lock_irqsave(&queue->callback_lock, flags);
1697         do {
1698                 u16 pending_idx = ubuf->desc;
1699                 ubuf = (struct ubuf_info *) ubuf->ctx;
1700                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1701                         MAX_PENDING_REQS);
1702                 index = pending_index(queue->dealloc_prod);
1703                 queue->dealloc_ring[index] = pending_idx;
1704                 /* Sync with xenvif_tx_dealloc_action:
1705                  * insert idx then incr producer.
1706                  */
1707                 smp_wmb();
1708                 queue->dealloc_prod++;
1709         } while (ubuf);
1710         spin_unlock_irqrestore(&queue->callback_lock, flags);
1711
1712         if (likely(zerocopy_success))
1713                 queue->stats.tx_zerocopy_success++;
1714         else
1715                 queue->stats.tx_zerocopy_fail++;
1716         xenvif_skb_zerocopy_complete(queue);
1717 }
1718
1719 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1720 {
1721         struct gnttab_unmap_grant_ref *gop;
1722         pending_ring_idx_t dc, dp;
1723         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1724         unsigned int i = 0;
1725
1726         dc = queue->dealloc_cons;
1727         gop = queue->tx_unmap_ops;
1728
1729         /* Free up any grants we have finished using */
1730         do {
1731                 dp = queue->dealloc_prod;
1732
1733                 /* Ensure we see all indices enqueued by all
1734                  * xenvif_zerocopy_callback().
1735                  */
1736                 smp_rmb();
1737
1738                 while (dc != dp) {
1739                         BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1740                         pending_idx =
1741                                 queue->dealloc_ring[pending_index(dc++)];
1742
1743                         pending_idx_release[gop - queue->tx_unmap_ops] =
1744                                 pending_idx;
1745                         queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1746                                 queue->mmap_pages[pending_idx];
1747                         gnttab_set_unmap_op(gop,
1748                                             idx_to_kaddr(queue, pending_idx),
1749                                             GNTMAP_host_map,
1750                                             queue->grant_tx_handle[pending_idx]);
1751                         xenvif_grant_handle_reset(queue, pending_idx);
1752                         ++gop;
1753                 }
1754
1755         } while (dp != queue->dealloc_prod);
1756
1757         queue->dealloc_cons = dc;
1758
1759         if (gop - queue->tx_unmap_ops > 0) {
1760                 int ret;
1761                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1762                                         NULL,
1763                                         queue->pages_to_unmap,
1764                                         gop - queue->tx_unmap_ops);
1765                 if (ret) {
1766                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1767                                    gop - queue->tx_unmap_ops, ret);
1768                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1769                                 if (gop[i].status != GNTST_okay)
1770                                         netdev_err(queue->vif->dev,
1771                                                    " host_addr: 0x%llx handle: 0x%x status: %d\n",
1772                                                    gop[i].host_addr,
1773                                                    gop[i].handle,
1774                                                    gop[i].status);
1775                         }
1776                         BUG();
1777                 }
1778         }
1779
1780         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1781                 xenvif_idx_release(queue, pending_idx_release[i],
1782                                    XEN_NETIF_RSP_OKAY);
1783 }
1784
1785
1786 /* Called after netfront has transmitted */
1787 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1788 {
1789         unsigned nr_mops, nr_cops = 0;
1790         int work_done, ret;
1791
1792         if (unlikely(!tx_work_todo(queue)))
1793                 return 0;
1794
1795         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1796
1797         if (nr_cops == 0)
1798                 return 0;
1799
1800         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1801         if (nr_mops != 0) {
1802                 ret = gnttab_map_refs(queue->tx_map_ops,
1803                                       NULL,
1804                                       queue->pages_to_map,
1805                                       nr_mops);
1806                 BUG_ON(ret);
1807         }
1808
1809         work_done = xenvif_tx_submit(queue);
1810
1811         return work_done;
1812 }
1813
1814 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1815                                u8 status)
1816 {
1817         struct pending_tx_info *pending_tx_info;
1818         pending_ring_idx_t index;
1819         unsigned long flags;
1820
1821         pending_tx_info = &queue->pending_tx_info[pending_idx];
1822
1823         spin_lock_irqsave(&queue->response_lock, flags);
1824
1825         make_tx_response(queue, &pending_tx_info->req,
1826                          pending_tx_info->extra_count, status);
1827
1828         /* Release the pending index before pusing the Tx response so
1829          * its available before a new Tx request is pushed by the
1830          * frontend.
1831          */
1832         index = pending_index(queue->pending_prod++);
1833         queue->pending_ring[index] = pending_idx;
1834
1835         push_tx_responses(queue);
1836
1837         spin_unlock_irqrestore(&queue->response_lock, flags);
1838 }
1839
1840
1841 static void make_tx_response(struct xenvif_queue *queue,
1842                              struct xen_netif_tx_request *txp,
1843                              unsigned int extra_count,
1844                              s8       st)
1845 {
1846         RING_IDX i = queue->tx.rsp_prod_pvt;
1847         struct xen_netif_tx_response *resp;
1848
1849         resp = RING_GET_RESPONSE(&queue->tx, i);
1850         resp->id     = txp->id;
1851         resp->status = st;
1852
1853         while (extra_count-- != 0)
1854                 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1855
1856         queue->tx.rsp_prod_pvt = ++i;
1857 }
1858
1859 static void push_tx_responses(struct xenvif_queue *queue)
1860 {
1861         int notify;
1862
1863         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1864         if (notify)
1865                 notify_remote_via_irq(queue->tx_irq);
1866 }
1867
1868 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1869                                              u16      id,
1870                                              s8       st,
1871                                              u16      offset,
1872                                              u16      size,
1873                                              u16      flags)
1874 {
1875         RING_IDX i = queue->rx.rsp_prod_pvt;
1876         struct xen_netif_rx_response *resp;
1877
1878         resp = RING_GET_RESPONSE(&queue->rx, i);
1879         resp->offset     = offset;
1880         resp->flags      = flags;
1881         resp->id         = id;
1882         resp->status     = (s16)size;
1883         if (st < 0)
1884                 resp->status = (s16)st;
1885
1886         queue->rx.rsp_prod_pvt = ++i;
1887
1888         return resp;
1889 }
1890
1891 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1892 {
1893         int ret;
1894         struct gnttab_unmap_grant_ref tx_unmap_op;
1895
1896         gnttab_set_unmap_op(&tx_unmap_op,
1897                             idx_to_kaddr(queue, pending_idx),
1898                             GNTMAP_host_map,
1899                             queue->grant_tx_handle[pending_idx]);
1900         xenvif_grant_handle_reset(queue, pending_idx);
1901
1902         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1903                                 &queue->mmap_pages[pending_idx], 1);
1904         if (ret) {
1905                 netdev_err(queue->vif->dev,
1906                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1907                            ret,
1908                            pending_idx,
1909                            tx_unmap_op.host_addr,
1910                            tx_unmap_op.handle,
1911                            tx_unmap_op.status);
1912                 BUG();
1913         }
1914 }
1915
1916 static inline int tx_work_todo(struct xenvif_queue *queue)
1917 {
1918         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1919                 return 1;
1920
1921         return 0;
1922 }
1923
1924 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1925 {
1926         return queue->dealloc_cons != queue->dealloc_prod;
1927 }
1928
1929 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1930 {
1931         if (queue->tx.sring)
1932                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1933                                         queue->tx.sring);
1934         if (queue->rx.sring)
1935                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1936                                         queue->rx.sring);
1937 }
1938
1939 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1940                                    grant_ref_t tx_ring_ref,
1941                                    grant_ref_t rx_ring_ref)
1942 {
1943         void *addr;
1944         struct xen_netif_tx_sring *txs;
1945         struct xen_netif_rx_sring *rxs;
1946
1947         int err = -ENOMEM;
1948
1949         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1950                                      &tx_ring_ref, 1, &addr);
1951         if (err)
1952                 goto err;
1953
1954         txs = (struct xen_netif_tx_sring *)addr;
1955         BACK_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE);
1956
1957         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1958                                      &rx_ring_ref, 1, &addr);
1959         if (err)
1960                 goto err;
1961
1962         rxs = (struct xen_netif_rx_sring *)addr;
1963         BACK_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE);
1964
1965         return 0;
1966
1967 err:
1968         xenvif_unmap_frontend_data_rings(queue);
1969         return err;
1970 }
1971
1972 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
1973 {
1974         struct xenvif *vif = queue->vif;
1975
1976         queue->stalled = true;
1977
1978         /* At least one queue has stalled? Disable the carrier. */
1979         spin_lock(&vif->lock);
1980         if (vif->stalled_queues++ == 0) {
1981                 netdev_info(vif->dev, "Guest Rx stalled");
1982                 netif_carrier_off(vif->dev);
1983         }
1984         spin_unlock(&vif->lock);
1985 }
1986
1987 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
1988 {
1989         struct xenvif *vif = queue->vif;
1990
1991         queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
1992         queue->stalled = false;
1993
1994         /* All queues are ready? Enable the carrier. */
1995         spin_lock(&vif->lock);
1996         if (--vif->stalled_queues == 0) {
1997                 netdev_info(vif->dev, "Guest Rx ready");
1998                 netif_carrier_on(vif->dev);
1999         }
2000         spin_unlock(&vif->lock);
2001 }
2002
2003 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
2004 {
2005         RING_IDX prod, cons;
2006
2007         prod = queue->rx.sring->req_prod;
2008         cons = queue->rx.req_cons;
2009
2010         return !queue->stalled && prod - cons < 1
2011                 && time_after(jiffies,
2012                               queue->last_rx_time + queue->vif->stall_timeout);
2013 }
2014
2015 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
2016 {
2017         RING_IDX prod, cons;
2018
2019         prod = queue->rx.sring->req_prod;
2020         cons = queue->rx.req_cons;
2021
2022         return queue->stalled && prod - cons >= 1;
2023 }
2024
2025 static bool xenvif_have_rx_work(struct xenvif_queue *queue)
2026 {
2027         return xenvif_rx_ring_slots_available(queue)
2028                 || (queue->vif->stall_timeout &&
2029                     (xenvif_rx_queue_stalled(queue)
2030                      || xenvif_rx_queue_ready(queue)))
2031                 || kthread_should_stop()
2032                 || queue->vif->disabled;
2033 }
2034
2035 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
2036 {
2037         struct sk_buff *skb;
2038         long timeout;
2039
2040         skb = skb_peek(&queue->rx_queue);
2041         if (!skb)
2042                 return MAX_SCHEDULE_TIMEOUT;
2043
2044         timeout = XENVIF_RX_CB(skb)->expires - jiffies;
2045         return timeout < 0 ? 0 : timeout;
2046 }
2047
2048 /* Wait until the guest Rx thread has work.
2049  *
2050  * The timeout needs to be adjusted based on the current head of the
2051  * queue (and not just the head at the beginning).  In particular, if
2052  * the queue is initially empty an infinite timeout is used and this
2053  * needs to be reduced when a skb is queued.
2054  *
2055  * This cannot be done with wait_event_timeout() because it only
2056  * calculates the timeout once.
2057  */
2058 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
2059 {
2060         DEFINE_WAIT(wait);
2061
2062         if (xenvif_have_rx_work(queue))
2063                 return;
2064
2065         for (;;) {
2066                 long ret;
2067
2068                 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
2069                 if (xenvif_have_rx_work(queue))
2070                         break;
2071                 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
2072                 if (!ret)
2073                         break;
2074         }
2075         finish_wait(&queue->wq, &wait);
2076 }
2077
2078 int xenvif_kthread_guest_rx(void *data)
2079 {
2080         struct xenvif_queue *queue = data;
2081         struct xenvif *vif = queue->vif;
2082
2083         if (!vif->stall_timeout)
2084                 xenvif_queue_carrier_on(queue);
2085
2086         for (;;) {
2087                 xenvif_wait_for_rx_work(queue);
2088
2089                 if (kthread_should_stop())
2090                         break;
2091
2092                 /* This frontend is found to be rogue, disable it in
2093                  * kthread context. Currently this is only set when
2094                  * netback finds out frontend sends malformed packet,
2095                  * but we cannot disable the interface in softirq
2096                  * context so we defer it here, if this thread is
2097                  * associated with queue 0.
2098                  */
2099                 if (unlikely(vif->disabled && queue->id == 0)) {
2100                         xenvif_carrier_off(vif);
2101                         break;
2102                 }
2103
2104                 if (!skb_queue_empty(&queue->rx_queue))
2105                         xenvif_rx_action(queue);
2106
2107                 /* If the guest hasn't provided any Rx slots for a
2108                  * while it's probably not responsive, drop the
2109                  * carrier so packets are dropped earlier.
2110                  */
2111                 if (vif->stall_timeout) {
2112                         if (xenvif_rx_queue_stalled(queue))
2113                                 xenvif_queue_carrier_off(queue);
2114                         else if (xenvif_rx_queue_ready(queue))
2115                                 xenvif_queue_carrier_on(queue);
2116                 }
2117
2118                 /* Queued packets may have foreign pages from other
2119                  * domains.  These cannot be queued indefinitely as
2120                  * this would starve guests of grant refs and transmit
2121                  * slots.
2122                  */
2123                 xenvif_rx_queue_drop_expired(queue);
2124
2125                 xenvif_rx_queue_maybe_wake(queue);
2126
2127                 cond_resched();
2128         }
2129
2130         /* Bin any remaining skbs */
2131         xenvif_rx_queue_purge(queue);
2132
2133         return 0;
2134 }
2135
2136 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
2137 {
2138         /* Dealloc thread must remain running until all inflight
2139          * packets complete.
2140          */
2141         return kthread_should_stop() &&
2142                 !atomic_read(&queue->inflight_packets);
2143 }
2144
2145 int xenvif_dealloc_kthread(void *data)
2146 {
2147         struct xenvif_queue *queue = data;
2148
2149         for (;;) {
2150                 wait_event_interruptible(queue->dealloc_wq,
2151                                          tx_dealloc_work_todo(queue) ||
2152                                          xenvif_dealloc_kthread_should_stop(queue));
2153                 if (xenvif_dealloc_kthread_should_stop(queue))
2154                         break;
2155
2156                 xenvif_tx_dealloc_action(queue);
2157                 cond_resched();
2158         }
2159
2160         /* Unmap anything remaining*/
2161         if (tx_dealloc_work_todo(queue))
2162                 xenvif_tx_dealloc_action(queue);
2163
2164         return 0;
2165 }
2166
2167 static void make_ctrl_response(struct xenvif *vif,
2168                                const struct xen_netif_ctrl_request *req,
2169                                u32 status, u32 data)
2170 {
2171         RING_IDX idx = vif->ctrl.rsp_prod_pvt;
2172         struct xen_netif_ctrl_response rsp = {
2173                 .id = req->id,
2174                 .type = req->type,
2175                 .status = status,
2176                 .data = data,
2177         };
2178
2179         *RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
2180         vif->ctrl.rsp_prod_pvt = ++idx;
2181 }
2182
2183 static void push_ctrl_response(struct xenvif *vif)
2184 {
2185         int notify;
2186
2187         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
2188         if (notify)
2189                 notify_remote_via_irq(vif->ctrl_irq);
2190 }
2191
2192 static void process_ctrl_request(struct xenvif *vif,
2193                                  const struct xen_netif_ctrl_request *req)
2194 {
2195         make_ctrl_response(vif, req, XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED,
2196                            0);
2197         push_ctrl_response(vif);
2198 }
2199
2200 static void xenvif_ctrl_action(struct xenvif *vif)
2201 {
2202         for (;;) {
2203                 RING_IDX req_prod, req_cons;
2204
2205                 req_prod = vif->ctrl.sring->req_prod;
2206                 req_cons = vif->ctrl.req_cons;
2207
2208                 /* Make sure we can see requests before we process them. */
2209                 rmb();
2210
2211                 if (req_cons == req_prod)
2212                         break;
2213
2214                 while (req_cons != req_prod) {
2215                         struct xen_netif_ctrl_request req;
2216
2217                         RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
2218                         req_cons++;
2219
2220                         process_ctrl_request(vif, &req);
2221                 }
2222
2223                 vif->ctrl.req_cons = req_cons;
2224                 vif->ctrl.sring->req_event = req_cons + 1;
2225         }
2226 }
2227
2228 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
2229 {
2230         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
2231                 return 1;
2232
2233         return 0;
2234 }
2235
2236 int xenvif_ctrl_kthread(void *data)
2237 {
2238         struct xenvif *vif = data;
2239
2240         for (;;) {
2241                 wait_event_interruptible(vif->ctrl_wq,
2242                                          xenvif_ctrl_work_todo(vif) ||
2243                                          kthread_should_stop());
2244                 if (kthread_should_stop())
2245                         break;
2246
2247                 while (xenvif_ctrl_work_todo(vif))
2248                         xenvif_ctrl_action(vif);
2249
2250                 cond_resched();
2251         }
2252
2253         return 0;
2254 }
2255
2256 static int __init netback_init(void)
2257 {
2258         int rc = 0;
2259
2260         if (!xen_domain())
2261                 return -ENODEV;
2262
2263         /* Allow as many queues as there are CPUs if user has not
2264          * specified a value.
2265          */
2266         if (xenvif_max_queues == 0)
2267                 xenvif_max_queues = num_online_cpus();
2268
2269         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2270                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2271                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2272                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2273         }
2274
2275         rc = xenvif_xenbus_init();
2276         if (rc)
2277                 goto failed_init;
2278
2279 #ifdef CONFIG_DEBUG_FS
2280         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2281         if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2282                 pr_warn("Init of debugfs returned %ld!\n",
2283                         PTR_ERR(xen_netback_dbg_root));
2284 #endif /* CONFIG_DEBUG_FS */
2285
2286         return 0;
2287
2288 failed_init:
2289         return rc;
2290 }
2291
2292 module_init(netback_init);
2293
2294 static void __exit netback_fini(void)
2295 {
2296 #ifdef CONFIG_DEBUG_FS
2297         if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2298                 debugfs_remove_recursive(xen_netback_dbg_root);
2299 #endif /* CONFIG_DEBUG_FS */
2300         xenvif_xenbus_fini();
2301 }
2302 module_exit(netback_fini);
2303
2304 MODULE_LICENSE("Dual BSD/GPL");
2305 MODULE_ALIAS("xen-backend:vif");