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 #define MAX_QUEUES_DEFAULT 8
  71 unsigned int xenvif_max_queues;
  72 module_param_named(max_queues, xenvif_max_queues, uint, 0644);
  73 MODULE_PARM_DESC(max_queues,
  74                  "Maximum number of queues per virtual interface");
  75
  76 /*
  77  * This is the maximum slots a skb can have. If a guest sends a skb
  78  * which exceeds this limit it is considered malicious.
  79  */
  80 #define FATAL_SKB_SLOTS_DEFAULT 20
  81 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
  82 module_param(fatal_skb_slots, uint, 0444);
  83
  84 /* The amount to copy out of the first guest Tx slot into the skb's
  85  * linear area.  If the first slot has more data, it will be mapped
  86  * and put into the first frag.
  87  *
  88  * This is sized to avoid pulling headers from the frags for most
  89  * TCP/IP packets.
  90  */
  91 #define XEN_NETBACK_TX_COPY_LEN 128
  92
  93
  94 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
  95                                u8 status);
  96
  97 static void make_tx_response(struct xenvif_queue *queue,
  98                              struct xen_netif_tx_request *txp,
  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 int xenvif_rx_ring_slots_needed(struct xenvif *vif)
 154 {
 155         if (vif->gso_mask)
 156                 return DIV_ROUND_UP(vif->dev->gso_max_size, XEN_PAGE_SIZE) + 1;
 157         else
 158                 return DIV_ROUND_UP(vif->dev->mtu, XEN_PAGE_SIZE);
 159 }
 160
 161 static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
 162 {
 163         RING_IDX prod, cons;
 164         int needed;
 165
 166         needed = xenvif_rx_ring_slots_needed(queue->vif);
 167
 168         do {
 169                 prod = queue->rx.sring->req_prod;
 170                 cons = queue->rx.req_cons;
 171
 172                 if (prod - cons >= needed)
 173                         return true;
 174
 175                 queue->rx.sring->req_event = prod + 1;
 176
 177                 /* Make sure event is visible before we check prod
 178                  * again.
 179                  */
 180                 mb();
 181         } while (queue->rx.sring->req_prod != prod);
 182
 183         return false;
 184 }
 185
 186 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
 187 {
 188         unsigned long flags;
 189
 190         spin_lock_irqsave(&queue->rx_queue.lock, flags);
 191
 192         if (queue->rx_queue_len >= queue->rx_queue_max) {
 193                 netif_tx_stop_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
 194                 kfree_skb(skb);
 195                 queue->vif->dev->stats.rx_dropped++;
 196         } else {
 197                 __skb_queue_tail(&queue->rx_queue, skb);
 198
 199                 queue->rx_queue_len += skb->len;
 200         }
 201
 202         spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
 203 }
 204
 205 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
 206 {
 207         struct sk_buff *skb;
 208
 209         spin_lock_irq(&queue->rx_queue.lock);
 210
 211         skb = __skb_dequeue(&queue->rx_queue);
 212         if (skb)
 213                 queue->rx_queue_len -= skb->len;
 214
 215         spin_unlock_irq(&queue->rx_queue.lock);
 216
 217         return skb;
 218 }
 219
 220 static void xenvif_rx_queue_maybe_wake(struct xenvif_queue *queue)
 221 {
 222         spin_lock_irq(&queue->rx_queue.lock);
 223
 224         if (queue->rx_queue_len < queue->rx_queue_max)
 225                 netif_tx_wake_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
 226
 227         spin_unlock_irq(&queue->rx_queue.lock);
 228 }
 229
 230
 231 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
 232 {
 233         struct sk_buff *skb;
 234         while ((skb = xenvif_rx_dequeue(queue)) != NULL)
 235                 kfree_skb(skb);
 236 }
 237
 238 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
 239 {
 240         struct sk_buff *skb;
 241
 242         for(;;) {
 243                 skb = skb_peek(&queue->rx_queue);
 244                 if (!skb)
 245                         break;
 246                 if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
 247                         break;
 248                 xenvif_rx_dequeue(queue);
 249                 kfree_skb(skb);
 250                 queue->vif->dev->stats.rx_dropped++;
 251         }
 252 }
 253
 254 struct netrx_pending_operations {
 255         unsigned copy_prod, copy_cons;
 256         unsigned meta_prod, meta_cons;
 257         struct gnttab_copy *copy;
 258         struct xenvif_rx_meta *meta;
 259         int copy_off;
 260         grant_ref_t copy_gref;
 261 };
 262
 263 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue,
 264                                                  struct netrx_pending_operations *npo)
 265 {
 266         struct xenvif_rx_meta *meta;
 267         struct xen_netif_rx_request req;
 268
 269         RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
 270
 271         meta = npo->meta + npo->meta_prod++;
 272         meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
 273         meta->gso_size = 0;
 274         meta->size = 0;
 275         meta->id = req.id;
 276
 277         npo->copy_off = 0;
 278         npo->copy_gref = req.gref;
 279
 280         return meta;
 281 }
 282
 283 struct gop_frag_copy {
 284         struct xenvif_queue *queue;
 285         struct netrx_pending_operations *npo;
 286         struct xenvif_rx_meta *meta;
 287         int head;
 288         int gso_type;
 289
 290         struct page *page;
 291 };
 292
 293 static void xenvif_setup_copy_gop(unsigned long gfn,
 294                                   unsigned int offset,
 295                                   unsigned int *len,
 296                                   struct gop_frag_copy *info)
 297 {
 298         struct gnttab_copy *copy_gop;
 299         struct xen_page_foreign *foreign;
 300         /* Convenient aliases */
 301         struct xenvif_queue *queue = info->queue;
 302         struct netrx_pending_operations *npo = info->npo;
 303         struct page *page = info->page;
 304
 305         BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
 306
 307         if (npo->copy_off == MAX_BUFFER_OFFSET)
 308                 info->meta = get_next_rx_buffer(queue, npo);
 309
 310         if (npo->copy_off + *len > MAX_BUFFER_OFFSET)
 311                 *len = MAX_BUFFER_OFFSET - npo->copy_off;
 312
 313         copy_gop = npo->copy + npo->copy_prod++;
 314         copy_gop->flags = GNTCOPY_dest_gref;
 315         copy_gop->len = *len;
 316
 317         foreign = xen_page_foreign(page);
 318         if (foreign) {
 319                 copy_gop->source.domid = foreign->domid;
 320                 copy_gop->source.u.ref = foreign->gref;
 321                 copy_gop->flags |= GNTCOPY_source_gref;
 322         } else {
 323                 copy_gop->source.domid = DOMID_SELF;
 324                 copy_gop->source.u.gmfn = gfn;
 325         }
 326         copy_gop->source.offset = offset;
 327
 328         copy_gop->dest.domid = queue->vif->domid;
 329         copy_gop->dest.offset = npo->copy_off;
 330         copy_gop->dest.u.ref = npo->copy_gref;
 331
 332         npo->copy_off += *len;
 333         info->meta->size += *len;
 334
 335         /* Leave a gap for the GSO descriptor. */
 336         if (info->head && ((1 << info->gso_type) & queue->vif->gso_mask))
 337                 queue->rx.req_cons++;
 338
 339         info->head = 0; /* There must be something in this buffer now */
 340 }
 341
 342 static void xenvif_gop_frag_copy_grant(unsigned long gfn,
 343                                        unsigned offset,
 344                                        unsigned int len,
 345                                        void *data)
 346 {
 347         unsigned int bytes;
 348
 349         while (len) {
 350                 bytes = len;
 351                 xenvif_setup_copy_gop(gfn, offset, &bytes, data);
 352                 offset += bytes;
 353                 len -= bytes;
 354         }
 355 }
 356
 357 /*
 358  * Set up the grant operations for this fragment. If it's a flipping
 359  * interface, we also set up the unmap request from here.
 360  */
 361 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
 362                                  struct netrx_pending_operations *npo,
 363                                  struct page *page, unsigned long size,
 364                                  unsigned long offset, int *head)
 365 {
 366         struct gop_frag_copy info = {
 367                 .queue = queue,
 368                 .npo = npo,
 369                 .head = *head,
 370                 .gso_type = XEN_NETIF_GSO_TYPE_NONE,
 371         };
 372         unsigned long bytes;
 373
 374         if (skb_is_gso(skb)) {
 375                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 376                         info.gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
 377                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 378                         info.gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
 379         }
 380
 381         /* Data must not cross a page boundary. */
 382         BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
 383
 384         info.meta = npo->meta + npo->meta_prod - 1;
 385
 386         /* Skip unused frames from start of page */
 387         page += offset >> PAGE_SHIFT;
 388         offset &= ~PAGE_MASK;
 389
 390         while (size > 0) {
 391                 BUG_ON(offset >= PAGE_SIZE);
 392
 393                 bytes = PAGE_SIZE - offset;
 394                 if (bytes > size)
 395                         bytes = size;
 396
 397                 info.page = page;
 398                 gnttab_foreach_grant_in_range(page, offset, bytes,
 399                                               xenvif_gop_frag_copy_grant,
 400                                               &info);
 401                 size -= bytes;
 402                 offset = 0;
 403
 404                 /* Next page */
 405                 if (size) {
 406                         BUG_ON(!PageCompound(page));
 407                         page++;
 408                 }
 409         }
 410
 411         *head = info.head;
 412 }
 413
 414 /*
 415  * Prepare an SKB to be transmitted to the frontend.
 416  *
 417  * This function is responsible for allocating grant operations, meta
 418  * structures, etc.
 419  *
 420  * It returns the number of meta structures consumed. The number of
 421  * ring slots used is always equal to the number of meta slots used
 422  * plus the number of GSO descriptors used. Currently, we use either
 423  * zero GSO descriptors (for non-GSO packets) or one descriptor (for
 424  * frontend-side LRO).
 425  */
 426 static int xenvif_gop_skb(struct sk_buff *skb,
 427                           struct netrx_pending_operations *npo,
 428                           struct xenvif_queue *queue)
 429 {
 430         struct xenvif *vif = netdev_priv(skb->dev);
 431         int nr_frags = skb_shinfo(skb)->nr_frags;
 432         int i;
 433         struct xen_netif_rx_request req;
 434         struct xenvif_rx_meta *meta;
 435         unsigned char *data;
 436         int head = 1;
 437         int old_meta_prod;
 438         int gso_type;
 439
 440         old_meta_prod = npo->meta_prod;
 441
 442         gso_type = XEN_NETIF_GSO_TYPE_NONE;
 443         if (skb_is_gso(skb)) {
 444                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 445                         gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
 446                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 447                         gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
 448         }
 449
 450         /* Set up a GSO prefix descriptor, if necessary */
 451         if ((1 << gso_type) & vif->gso_prefix_mask) {
 452                 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
 453                 meta = npo->meta + npo->meta_prod++;
 454                 meta->gso_type = gso_type;
 455                 meta->gso_size = skb_shinfo(skb)->gso_size;
 456                 meta->size = 0;
 457                 meta->id = req.id;
 458         }
 459
 460         RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
 461         meta = npo->meta + npo->meta_prod++;
 462
 463         if ((1 << gso_type) & vif->gso_mask) {
 464                 meta->gso_type = gso_type;
 465                 meta->gso_size = skb_shinfo(skb)->gso_size;
 466         } else {
 467                 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
 468                 meta->gso_size = 0;
 469         }
 470
 471         meta->size = 0;
 472         meta->id = req.id;
 473         npo->copy_off = 0;
 474         npo->copy_gref = req.gref;
 475
 476         data = skb->data;
 477         while (data < skb_tail_pointer(skb)) {
 478                 unsigned int offset = offset_in_page(data);
 479                 unsigned int len = PAGE_SIZE - offset;
 480
 481                 if (data + len > skb_tail_pointer(skb))
 482                         len = skb_tail_pointer(skb) - data;
 483
 484                 xenvif_gop_frag_copy(queue, skb, npo,
 485                                      virt_to_page(data), len, offset, &head);
 486                 data += len;
 487         }
 488
 489         for (i = 0; i < nr_frags; i++) {
 490                 xenvif_gop_frag_copy(queue, skb, npo,
 491                                      skb_frag_page(&skb_shinfo(skb)->frags[i]),
 492                                      skb_frag_size(&skb_shinfo(skb)->frags[i]),
 493                                      skb_shinfo(skb)->frags[i].page_offset,
 494                                      &head);
 495         }
 496
 497         return npo->meta_prod - old_meta_prod;
 498 }
 499
 500 /*
 501  * This is a twin to xenvif_gop_skb.  Assume that xenvif_gop_skb was
 502  * used to set up the operations on the top of
 503  * netrx_pending_operations, which have since been done.  Check that
 504  * they didn't give any errors and advance over them.
 505  */
 506 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
 507                             struct netrx_pending_operations *npo)
 508 {
 509         struct gnttab_copy     *copy_op;
 510         int status = XEN_NETIF_RSP_OKAY;
 511         int i;
 512
 513         for (i = 0; i < nr_meta_slots; i++) {
 514                 copy_op = npo->copy + npo->copy_cons++;
 515                 if (copy_op->status != GNTST_okay) {
 516                         netdev_dbg(vif->dev,
 517                                    "Bad status %d from copy to DOM%d.\n",
 518                                    copy_op->status, vif->domid);
 519                         status = XEN_NETIF_RSP_ERROR;
 520                 }
 521         }
 522
 523         return status;
 524 }
 525
 526 static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status,
 527                                       struct xenvif_rx_meta *meta,
 528                                       int nr_meta_slots)
 529 {
 530         int i;
 531         unsigned long offset;
 532
 533         /* No fragments used */
 534         if (nr_meta_slots <= 1)
 535                 return;
 536
 537         nr_meta_slots--;
 538
 539         for (i = 0; i < nr_meta_slots; i++) {
 540                 int flags;
 541                 if (i == nr_meta_slots - 1)
 542                         flags = 0;
 543                 else
 544                         flags = XEN_NETRXF_more_data;
 545
 546                 offset = 0;
 547                 make_rx_response(queue, meta[i].id, status, offset,
 548                                  meta[i].size, flags);
 549         }
 550 }
 551
 552 void xenvif_kick_thread(struct xenvif_queue *queue)
 553 {
 554         wake_up(&queue->wq);
 555 }
 556
 557 static void xenvif_rx_action(struct xenvif_queue *queue)
 558 {
 559         s8 status;
 560         u16 flags;
 561         struct xen_netif_rx_response *resp;
 562         struct sk_buff_head rxq;
 563         struct sk_buff *skb;
 564         LIST_HEAD(notify);
 565         int ret;
 566         unsigned long offset;
 567         bool need_to_notify = false;
 568
 569         struct netrx_pending_operations npo = {
 570                 .copy  = queue->grant_copy_op,
 571                 .meta  = queue->meta,
 572         };
 573
 574         skb_queue_head_init(&rxq);
 575
 576         while (xenvif_rx_ring_slots_available(queue)
 577                && (skb = xenvif_rx_dequeue(queue)) != NULL) {
 578                 queue->last_rx_time = jiffies;
 579
 580                 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
 581
 582                 __skb_queue_tail(&rxq, skb);
 583         }
 584
 585         BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
 586
 587         if (!npo.copy_prod)
 588                 goto done;
 589
 590         BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
 591         gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
 592
 593         while ((skb = __skb_dequeue(&rxq)) != NULL) {
 594
 595                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 596                     queue->vif->gso_prefix_mask) {
 597                         resp = RING_GET_RESPONSE(&queue->rx,
 598                                                  queue->rx.rsp_prod_pvt++);
 599
 600                         resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
 601
 602                         resp->offset = queue->meta[npo.meta_cons].gso_size;
 603                         resp->id = queue->meta[npo.meta_cons].id;
 604                         resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
 605
 606                         npo.meta_cons++;
 607                         XENVIF_RX_CB(skb)->meta_slots_used--;
 608                 }
 609
 610
 611                 queue->stats.tx_bytes += skb->len;
 612                 queue->stats.tx_packets++;
 613
 614                 status = xenvif_check_gop(queue->vif,
 615                                           XENVIF_RX_CB(skb)->meta_slots_used,
 616                                           &npo);
 617
 618                 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
 619                         flags = 0;
 620                 else
 621                         flags = XEN_NETRXF_more_data;
 622
 623                 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
 624                         flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
 625                 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 626                         /* remote but checksummed. */
 627                         flags |= XEN_NETRXF_data_validated;
 628
 629                 offset = 0;
 630                 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
 631                                         status, offset,
 632                                         queue->meta[npo.meta_cons].size,
 633                                         flags);
 634
 635                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 636                     queue->vif->gso_mask) {
 637                         struct xen_netif_extra_info *gso =
 638                                 (struct xen_netif_extra_info *)
 639                                 RING_GET_RESPONSE(&queue->rx,
 640                                                   queue->rx.rsp_prod_pvt++);
 641
 642                         resp->flags |= XEN_NETRXF_extra_info;
 643
 644                         gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
 645                         gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
 646                         gso->u.gso.pad = 0;
 647                         gso->u.gso.features = 0;
 648
 649                         gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
 650                         gso->flags = 0;
 651                 }
 652
 653                 xenvif_add_frag_responses(queue, status,
 654                                           queue->meta + npo.meta_cons + 1,
 655                                           XENVIF_RX_CB(skb)->meta_slots_used);
 656
 657                 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
 658
 659                 need_to_notify |= !!ret;
 660
 661                 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
 662                 dev_kfree_skb(skb);
 663         }
 664
 665 done:
 666         if (need_to_notify)
 667                 notify_remote_via_irq(queue->rx_irq);
 668 }
 669
 670 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 671 {
 672         int more_to_do;
 673
 674         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 675
 676         if (more_to_do)
 677                 napi_schedule(&queue->napi);
 678         else if (xenvif_atomic_fetch_andnot(NETBK_TX_EOI | NETBK_COMMON_EOI,
 679                                      &queue->eoi_pending) &
 680                  (NETBK_TX_EOI | NETBK_COMMON_EOI))
 681                 xen_irq_lateeoi(queue->tx_irq, 0);
 682 }
 683
 684 static void tx_add_credit(struct xenvif_queue *queue)
 685 {
 686         unsigned long max_burst, max_credit;
 687
 688         /*
 689          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 690          * Otherwise the interface can seize up due to insufficient credit.
 691          */
 692         max_burst = max(131072UL, queue->credit_bytes);
 693
 694         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 695         max_credit = queue->remaining_credit + queue->credit_bytes;
 696         if (max_credit < queue->remaining_credit)
 697                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 698
 699         queue->remaining_credit = min(max_credit, max_burst);
 700         queue->rate_limited = false;
 701 }
 702
 703 void xenvif_tx_credit_callback(unsigned long data)
 704 {
 705         struct xenvif_queue *queue = (struct xenvif_queue *)data;
 706         tx_add_credit(queue);
 707         xenvif_napi_schedule_or_enable_events(queue);
 708 }
 709
 710 static void xenvif_tx_err(struct xenvif_queue *queue,
 711                           struct xen_netif_tx_request *txp, RING_IDX end)
 712 {
 713         RING_IDX cons = queue->tx.req_cons;
 714         unsigned long flags;
 715
 716         do {
 717                 spin_lock_irqsave(&queue->response_lock, flags);
 718                 make_tx_response(queue, txp, XEN_NETIF_RSP_ERROR);
 719                 push_tx_responses(queue);
 720                 spin_unlock_irqrestore(&queue->response_lock, flags);
 721                 if (cons == end)
 722                         break;
 723                 RING_COPY_REQUEST(&queue->tx, cons++, txp);
 724         } while (1);
 725         queue->tx.req_cons = cons;
 726 }
 727
 728 static void xenvif_fatal_tx_err(struct xenvif *vif)
 729 {
 730         netdev_err(vif->dev, "fatal error; disabling device\n");
 731         vif->disabled = true;
 732         /* Disable the vif from queue 0's kthread */
 733         if (vif->queues)
 734                 xenvif_kick_thread(&vif->queues[0]);
 735 }
 736
 737 static int xenvif_count_requests(struct xenvif_queue *queue,
 738                                  struct xen_netif_tx_request *first,
 739                                  struct xen_netif_tx_request *txp,
 740                                  int work_to_do)
 741 {
 742         RING_IDX cons = queue->tx.req_cons;
 743         int slots = 0;
 744         int drop_err = 0;
 745         int more_data;
 746
 747         if (!(first->flags & XEN_NETTXF_more_data))
 748                 return 0;
 749
 750         do {
 751                 struct xen_netif_tx_request dropped_tx = { 0 };
 752
 753                 if (slots >= work_to_do) {
 754                         netdev_err(queue->vif->dev,
 755                                    "Asked for %d slots but exceeds this limit\n",
 756                                    work_to_do);
 757                         xenvif_fatal_tx_err(queue->vif);
 758                         return -ENODATA;
 759                 }
 760
 761                 /* This guest is really using too many slots and
 762                  * considered malicious.
 763                  */
 764                 if (unlikely(slots >= fatal_skb_slots)) {
 765                         netdev_err(queue->vif->dev,
 766                                    "Malicious frontend using %d slots, threshold %u\n",
 767                                    slots, fatal_skb_slots);
 768                         xenvif_fatal_tx_err(queue->vif);
 769                         return -E2BIG;
 770                 }
 771
 772                 /* Xen network protocol had implicit dependency on
 773                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 774                  * the historical MAX_SKB_FRAGS value 18 to honor the
 775                  * same behavior as before. Any packet using more than
 776                  * 18 slots but less than fatal_skb_slots slots is
 777                  * dropped
 778                  */
 779                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 780                         if (net_ratelimit())
 781                                 netdev_dbg(queue->vif->dev,
 782                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 783                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 784                         drop_err = -E2BIG;
 785                 }
 786
 787                 if (drop_err)
 788                         txp = &dropped_tx;
 789
 790                 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
 791
 792                 /* If the guest submitted a frame >= 64 KiB then
 793                  * first->size overflowed and following slots will
 794                  * appear to be larger than the frame.
 795                  *
 796                  * This cannot be fatal error as there are buggy
 797                  * frontends that do this.
 798                  *
 799                  * Consume all slots and drop the packet.
 800                  */
 801                 if (!drop_err && txp->size > first->size) {
 802                         if (net_ratelimit())
 803                                 netdev_dbg(queue->vif->dev,
 804                                            "Invalid tx request, slot size %u > remaining size %u\n",
 805                                            txp->size, first->size);
 806                         drop_err = -EIO;
 807                 }
 808
 809                 first->size -= txp->size;
 810                 slots++;
 811
 812                 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
 813                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 814                                  txp->offset, txp->size);
 815                         xenvif_fatal_tx_err(queue->vif);
 816                         return -EINVAL;
 817                 }
 818
 819                 more_data = txp->flags & XEN_NETTXF_more_data;
 820
 821                 if (!drop_err)
 822                         txp++;
 823
 824         } while (more_data);
 825
 826         if (drop_err) {
 827                 xenvif_tx_err(queue, first, cons + slots);
 828                 return drop_err;
 829         }
 830
 831         return slots;
 832 }
 833
 834
 835 struct xenvif_tx_cb {
 836         u16 pending_idx;
 837 };
 838
 839 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 840
 841 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 842                                           u16 pending_idx,
 843                                           struct xen_netif_tx_request *txp,
 844                                           struct gnttab_map_grant_ref *mop)
 845 {
 846         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 847         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 848                           GNTMAP_host_map | GNTMAP_readonly,
 849                           txp->gref, queue->vif->domid);
 850
 851         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 852                sizeof(*txp));
 853 }
 854
 855 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 856 {
 857         struct sk_buff *skb =
 858                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 859                           GFP_ATOMIC | __GFP_NOWARN);
 860         if (unlikely(skb == NULL))
 861                 return NULL;
 862
 863         /* Packets passed to netif_rx() must have some headroom. */
 864         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 865
 866         /* Initialize it here to avoid later surprises */
 867         skb_shinfo(skb)->destructor_arg = NULL;
 868
 869         return skb;
 870 }
 871
 872 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
 873                                                         struct sk_buff *skb,
 874                                                         struct xen_netif_tx_request *txp,
 875                                                         struct gnttab_map_grant_ref *gop,
 876                                                         unsigned int frag_overflow,
 877                                                         struct sk_buff *nskb)
 878 {
 879         struct skb_shared_info *shinfo = skb_shinfo(skb);
 880         skb_frag_t *frags = shinfo->frags;
 881         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 882         int start;
 883         pending_ring_idx_t index;
 884         unsigned int nr_slots;
 885
 886         nr_slots = shinfo->nr_frags;
 887
 888         /* Skip first skb fragment if it is on same page as header fragment. */
 889         start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
 890
 891         for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
 892              shinfo->nr_frags++, txp++, gop++) {
 893                 index = pending_index(queue->pending_cons++);
 894                 pending_idx = queue->pending_ring[index];
 895                 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
 896                 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 897         }
 898
 899         if (frag_overflow) {
 900
 901                 shinfo = skb_shinfo(nskb);
 902                 frags = shinfo->frags;
 903
 904                 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
 905                      shinfo->nr_frags++, txp++, gop++) {
 906                         index = pending_index(queue->pending_cons++);
 907                         pending_idx = queue->pending_ring[index];
 908                         xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
 909                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 910                                              pending_idx);
 911                 }
 912
 913                 skb_shinfo(skb)->frag_list = nskb;
 914         }
 915
 916         return gop;
 917 }
 918
 919 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 920                                            u16 pending_idx,
 921                                            grant_handle_t handle)
 922 {
 923         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 924                      NETBACK_INVALID_HANDLE)) {
 925                 netdev_err(queue->vif->dev,
 926                            "Trying to overwrite active handle! pending_idx: 0x%x\n",
 927                            pending_idx);
 928                 BUG();
 929         }
 930         queue->grant_tx_handle[pending_idx] = handle;
 931 }
 932
 933 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 934                                              u16 pending_idx)
 935 {
 936         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 937                      NETBACK_INVALID_HANDLE)) {
 938                 netdev_err(queue->vif->dev,
 939                            "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 940                            pending_idx);
 941                 BUG();
 942         }
 943         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 944 }
 945
 946 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 947                                struct sk_buff *skb,
 948                                struct gnttab_map_grant_ref **gopp_map,
 949                                struct gnttab_copy **gopp_copy)
 950 {
 951         struct gnttab_map_grant_ref *gop_map = *gopp_map;
 952         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 953         /* This always points to the shinfo of the skb being checked, which
 954          * could be either the first or the one on the frag_list
 955          */
 956         struct skb_shared_info *shinfo = skb_shinfo(skb);
 957         /* If this is non-NULL, we are currently checking the frag_list skb, and
 958          * this points to the shinfo of the first one
 959          */
 960         struct skb_shared_info *first_shinfo = NULL;
 961         int nr_frags = shinfo->nr_frags;
 962         const bool sharedslot = nr_frags &&
 963                                 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
 964         int i, err;
 965
 966         /* Check status of header. */
 967         err = (*gopp_copy)->status;
 968         if (unlikely(err)) {
 969                 if (net_ratelimit())
 970                         netdev_dbg(queue->vif->dev,
 971                                    "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 972                                    (*gopp_copy)->status,
 973                                    pending_idx,
 974                                    (*gopp_copy)->source.u.ref);
 975                 /* The first frag might still have this slot mapped */
 976                 if (!sharedslot)
 977                         xenvif_idx_release(queue, pending_idx,
 978                                            XEN_NETIF_RSP_ERROR);
 979         }
 980         (*gopp_copy)++;
 981
 982 check_frags:
 983         for (i = 0; i < nr_frags; i++, gop_map++) {
 984                 int j, newerr;
 985
 986                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 987
 988                 /* Check error status: if okay then remember grant handle. */
 989                 newerr = gop_map->status;
 990
 991                 if (likely(!newerr)) {
 992                         xenvif_grant_handle_set(queue,
 993                                                 pending_idx,
 994                                                 gop_map->handle);
 995                         /* Had a previous error? Invalidate this fragment. */
 996                         if (unlikely(err)) {
 997                                 xenvif_idx_unmap(queue, pending_idx);
 998                                 /* If the mapping of the first frag was OK, but
 999                                  * the header's copy failed, and they are
1000                                  * sharing a slot, send an error
1001                                  */
1002                                 if (i == 0 && !first_shinfo && sharedslot)
1003                                         xenvif_idx_release(queue, pending_idx,
1004                                                            XEN_NETIF_RSP_ERROR);
1005                                 else
1006                                         xenvif_idx_release(queue, pending_idx,
1007                                                            XEN_NETIF_RSP_OKAY);
1008                         }
1009                         continue;
1010                 }
1011
1012                 /* Error on this fragment: respond to client with an error. */
1013                 if (net_ratelimit())
1014                         netdev_dbg(queue->vif->dev,
1015                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
1016                                    i,
1017                                    gop_map->status,
1018                                    pending_idx,
1019                                    gop_map->ref);
1020
1021                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
1022
1023                 /* Not the first error? Preceding frags already invalidated. */
1024                 if (err)
1025                         continue;
1026
1027                 /* First error: if the header haven't shared a slot with the
1028                  * first frag, release it as well.
1029                  */
1030                 if (!sharedslot)
1031                         xenvif_idx_release(queue,
1032                                            XENVIF_TX_CB(skb)->pending_idx,
1033                                            XEN_NETIF_RSP_OKAY);
1034
1035                 /* Invalidate preceding fragments of this skb. */
1036                 for (j = 0; j < i; j++) {
1037                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1038                         xenvif_idx_unmap(queue, pending_idx);
1039                         xenvif_idx_release(queue, pending_idx,
1040                                            XEN_NETIF_RSP_OKAY);
1041                 }
1042
1043                 /* And if we found the error while checking the frag_list, unmap
1044                  * the first skb's frags
1045                  */
1046                 if (first_shinfo) {
1047                         for (j = 0; j < first_shinfo->nr_frags; j++) {
1048                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
1049                                 xenvif_idx_unmap(queue, pending_idx);
1050                                 xenvif_idx_release(queue, pending_idx,
1051                                                    XEN_NETIF_RSP_OKAY);
1052                         }
1053                 }
1054
1055                 /* Remember the error: invalidate all subsequent fragments. */
1056                 err = newerr;
1057         }
1058
1059         if (skb_has_frag_list(skb) && !first_shinfo) {
1060                 first_shinfo = skb_shinfo(skb);
1061                 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1062                 nr_frags = shinfo->nr_frags;
1063
1064                 goto check_frags;
1065         }
1066
1067         *gopp_map = gop_map;
1068         return err;
1069 }
1070
1071 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1072 {
1073         struct skb_shared_info *shinfo = skb_shinfo(skb);
1074         int nr_frags = shinfo->nr_frags;
1075         int i;
1076         u16 prev_pending_idx = INVALID_PENDING_IDX;
1077
1078         for (i = 0; i < nr_frags; i++) {
1079                 skb_frag_t *frag = shinfo->frags + i;
1080                 struct xen_netif_tx_request *txp;
1081                 struct page *page;
1082                 u16 pending_idx;
1083
1084                 pending_idx = frag_get_pending_idx(frag);
1085
1086                 /* If this is not the first frag, chain it to the previous*/
1087                 if (prev_pending_idx == INVALID_PENDING_IDX)
1088                         skb_shinfo(skb)->destructor_arg =
1089                                 &callback_param(queue, pending_idx);
1090                 else
1091                         callback_param(queue, prev_pending_idx).ctx =
1092                                 &callback_param(queue, pending_idx);
1093
1094                 callback_param(queue, pending_idx).ctx = NULL;
1095                 prev_pending_idx = pending_idx;
1096
1097                 txp = &queue->pending_tx_info[pending_idx].req;
1098                 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1099                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1100                 skb->len += txp->size;
1101                 skb->data_len += txp->size;
1102                 skb->truesize += txp->size;
1103
1104                 /* Take an extra reference to offset network stack's put_page */
1105                 get_page(queue->mmap_pages[pending_idx]);
1106         }
1107 }
1108
1109 static int xenvif_get_extras(struct xenvif_queue *queue,
1110                                 struct xen_netif_extra_info *extras,
1111                                 int work_to_do)
1112 {
1113         struct xen_netif_extra_info extra;
1114         RING_IDX cons = queue->tx.req_cons;
1115
1116         do {
1117                 if (unlikely(work_to_do-- <= 0)) {
1118                         netdev_err(queue->vif->dev, "Missing extra info\n");
1119                         xenvif_fatal_tx_err(queue->vif);
1120                         return -EBADR;
1121                 }
1122
1123                 RING_COPY_REQUEST(&queue->tx, cons, &extra);
1124                 if (unlikely(!extra.type ||
1125                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1126                         queue->tx.req_cons = ++cons;
1127                         netdev_err(queue->vif->dev,
1128                                    "Invalid extra type: %d\n", extra.type);
1129                         xenvif_fatal_tx_err(queue->vif);
1130                         return -EINVAL;
1131                 }
1132
1133                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1134                 queue->tx.req_cons = ++cons;
1135         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1136
1137         return work_to_do;
1138 }
1139
1140 static int xenvif_set_skb_gso(struct xenvif *vif,
1141                               struct sk_buff *skb,
1142                               struct xen_netif_extra_info *gso)
1143 {
1144         if (!gso->u.gso.size) {
1145                 netdev_err(vif->dev, "GSO size must not be zero.\n");
1146                 xenvif_fatal_tx_err(vif);
1147                 return -EINVAL;
1148         }
1149
1150         switch (gso->u.gso.type) {
1151         case XEN_NETIF_GSO_TYPE_TCPV4:
1152                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1153                 break;
1154         case XEN_NETIF_GSO_TYPE_TCPV6:
1155                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1156                 break;
1157         default:
1158                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1159                 xenvif_fatal_tx_err(vif);
1160                 return -EINVAL;
1161         }
1162
1163         skb_shinfo(skb)->gso_size = gso->u.gso.size;
1164         /* gso_segs will be calculated later */
1165
1166         return 0;
1167 }
1168
1169 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1170 {
1171         bool recalculate_partial_csum = false;
1172
1173         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1174          * peers can fail to set NETRXF_csum_blank when sending a GSO
1175          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1176          * recalculate the partial checksum.
1177          */
1178         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1179                 queue->stats.rx_gso_checksum_fixup++;
1180                 skb->ip_summed = CHECKSUM_PARTIAL;
1181                 recalculate_partial_csum = true;
1182         }
1183
1184         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1185         if (skb->ip_summed != CHECKSUM_PARTIAL)
1186                 return 0;
1187
1188         return skb_checksum_setup(skb, recalculate_partial_csum);
1189 }
1190
1191 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1192 {
1193         u64 now = get_jiffies_64();
1194         u64 next_credit = queue->credit_window_start +
1195                 msecs_to_jiffies(queue->credit_usec / 1000);
1196
1197         /* Timer could already be pending in rare cases. */
1198         if (timer_pending(&queue->credit_timeout)) {
1199                 queue->rate_limited = true;
1200                 return true;
1201         }
1202
1203         /* Passed the point where we can replenish credit? */
1204         if (time_after_eq64(now, next_credit)) {
1205                 queue->credit_window_start = now;
1206                 tx_add_credit(queue);
1207         }
1208
1209         /* Still too big to send right now? Set a callback. */
1210         if (size > queue->remaining_credit) {
1211                 queue->credit_timeout.data     =
1212                         (unsigned long)queue;
1213                 mod_timer(&queue->credit_timeout,
1214                           next_credit);
1215                 queue->credit_window_start = next_credit;
1216                 queue->rate_limited = true;
1217
1218                 return true;
1219         }
1220
1221         return false;
1222 }
1223
1224 /* No locking is required in xenvif_mcast_add/del() as they are
1225  * only ever invoked from NAPI poll. An RCU list is used because
1226  * xenvif_mcast_match() is called asynchronously, during start_xmit.
1227  */
1228
1229 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
1230 {
1231         struct xenvif_mcast_addr *mcast;
1232
1233         if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
1234                 if (net_ratelimit())
1235                         netdev_err(vif->dev,
1236                                    "Too many multicast addresses\n");
1237                 return -ENOSPC;
1238         }
1239
1240         mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
1241         if (!mcast)
1242                 return -ENOMEM;
1243
1244         ether_addr_copy(mcast->addr, addr);
1245         list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
1246         vif->fe_mcast_count++;
1247
1248         return 0;
1249 }
1250
1251 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
1252 {
1253         struct xenvif_mcast_addr *mcast;
1254
1255         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1256                 if (ether_addr_equal(addr, mcast->addr)) {
1257                         --vif->fe_mcast_count;
1258                         list_del_rcu(&mcast->entry);
1259                         kfree_rcu(mcast, rcu);
1260                         break;
1261                 }
1262         }
1263 }
1264
1265 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
1266 {
1267         struct xenvif_mcast_addr *mcast;
1268
1269         rcu_read_lock();
1270         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1271                 if (ether_addr_equal(addr, mcast->addr)) {
1272                         rcu_read_unlock();
1273                         return true;
1274                 }
1275         }
1276         rcu_read_unlock();
1277
1278         return false;
1279 }
1280
1281 void xenvif_mcast_addr_list_free(struct xenvif *vif)
1282 {
1283         /* No need for locking or RCU here. NAPI poll and TX queue
1284          * are stopped.
1285          */
1286         while (!list_empty(&vif->fe_mcast_addr)) {
1287                 struct xenvif_mcast_addr *mcast;
1288
1289                 mcast = list_first_entry(&vif->fe_mcast_addr,
1290                                          struct xenvif_mcast_addr,
1291                                          entry);
1292                 --vif->fe_mcast_count;
1293                 list_del(&mcast->entry);
1294                 kfree(mcast);
1295         }
1296 }
1297
1298 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1299                                      int budget,
1300                                      unsigned *copy_ops,
1301                                      unsigned *map_ops)
1302 {
1303         struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
1304         struct sk_buff *skb, *nskb;
1305         int ret;
1306         unsigned int frag_overflow;
1307
1308         while (skb_queue_len(&queue->tx_queue) < budget) {
1309                 struct xen_netif_tx_request txreq;
1310                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1311                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1312                 u16 pending_idx;
1313                 RING_IDX idx;
1314                 int work_to_do;
1315                 unsigned int data_len;
1316                 pending_ring_idx_t index;
1317
1318                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1319                     XEN_NETIF_TX_RING_SIZE) {
1320                         netdev_err(queue->vif->dev,
1321                                    "Impossible number of requests. "
1322                                    "req_prod %d, req_cons %d, size %ld\n",
1323                                    queue->tx.sring->req_prod, queue->tx.req_cons,
1324                                    XEN_NETIF_TX_RING_SIZE);
1325                         xenvif_fatal_tx_err(queue->vif);
1326                         break;
1327                 }
1328
1329                 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1330                 if (!work_to_do)
1331                         break;
1332
1333                 idx = queue->tx.req_cons;
1334                 rmb(); /* Ensure that we see the request before we copy it. */
1335                 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
1336
1337                 /* Credit-based scheduling. */
1338                 if (txreq.size > queue->remaining_credit &&
1339                     tx_credit_exceeded(queue, txreq.size))
1340                         break;
1341
1342                 queue->remaining_credit -= txreq.size;
1343
1344                 work_to_do--;
1345                 queue->tx.req_cons = ++idx;
1346
1347                 memset(extras, 0, sizeof(extras));
1348                 if (txreq.flags & XEN_NETTXF_extra_info) {
1349                         work_to_do = xenvif_get_extras(queue, extras,
1350                                                        work_to_do);
1351                         idx = queue->tx.req_cons;
1352                         if (unlikely(work_to_do < 0))
1353                                 break;
1354                 }
1355
1356                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
1357                         struct xen_netif_extra_info *extra;
1358
1359                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
1360                         ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
1361
1362                         make_tx_response(queue, &txreq,
1363                                          (ret == 0) ?
1364                                          XEN_NETIF_RSP_OKAY :
1365                                          XEN_NETIF_RSP_ERROR);
1366                         push_tx_responses(queue);
1367                         continue;
1368                 }
1369
1370                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
1371                         struct xen_netif_extra_info *extra;
1372
1373                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
1374                         xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
1375
1376                         make_tx_response(queue, &txreq, XEN_NETIF_RSP_OKAY);
1377                         push_tx_responses(queue);
1378                         continue;
1379                 }
1380
1381                 ret = xenvif_count_requests(queue, &txreq, 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, 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, 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                                 skb_shinfo(skb)->nr_frags = 0;
1434                                 kfree_skb(skb);
1435                                 xenvif_tx_err(queue, &txreq, idx);
1436                                 if (net_ratelimit())
1437                                         netdev_err(queue->vif->dev,
1438                                                    "Can't allocate the frag_list skb.\n");
1439                                 break;
1440                         }
1441                 }
1442
1443                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1444                         struct xen_netif_extra_info *gso;
1445                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1446
1447                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1448                                 /* Failure in xenvif_set_skb_gso is fatal. */
1449                                 skb_shinfo(skb)->nr_frags = 0;
1450                                 kfree_skb(skb);
1451                                 kfree_skb(nskb);
1452                                 break;
1453                         }
1454                 }
1455
1456                 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1457
1458                 __skb_put(skb, data_len);
1459                 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1460                 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1461                 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1462
1463                 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1464                         virt_to_gfn(skb->data);
1465                 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1466                 queue->tx_copy_ops[*copy_ops].dest.offset =
1467                         offset_in_page(skb->data) & ~XEN_PAGE_MASK;
1468
1469                 queue->tx_copy_ops[*copy_ops].len = data_len;
1470                 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1471
1472                 (*copy_ops)++;
1473
1474                 if (data_len < txreq.size) {
1475                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1476                                              pending_idx);
1477                         xenvif_tx_create_map_op(queue, pending_idx, &txreq, gop);
1478                         gop++;
1479                 } else {
1480                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1481                                              INVALID_PENDING_IDX);
1482                         memcpy(&queue->pending_tx_info[pending_idx].req, &txreq,
1483                                sizeof(txreq));
1484                 }
1485
1486                 queue->pending_cons++;
1487
1488                 gop = xenvif_get_requests(queue, skb, txfrags, gop,
1489                                           frag_overflow, nskb);
1490
1491                 __skb_queue_tail(&queue->tx_queue, skb);
1492
1493                 queue->tx.req_cons = idx;
1494
1495                 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1496                     (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1497                         break;
1498         }
1499
1500         (*map_ops) = gop - queue->tx_map_ops;
1501         return;
1502 }
1503
1504 /* Consolidate skb with a frag_list into a brand new one with local pages on
1505  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1506  */
1507 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1508 {
1509         unsigned int offset = skb_headlen(skb);
1510         skb_frag_t frags[MAX_SKB_FRAGS];
1511         int i, f;
1512         struct ubuf_info *uarg;
1513         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1514
1515         queue->stats.tx_zerocopy_sent += 2;
1516         queue->stats.tx_frag_overflow++;
1517
1518         xenvif_fill_frags(queue, nskb);
1519         /* Subtract frags size, we will correct it later */
1520         skb->truesize -= skb->data_len;
1521         skb->len += nskb->len;
1522         skb->data_len += nskb->len;
1523
1524         /* create a brand new frags array and coalesce there */
1525         for (i = 0; offset < skb->len; i++) {
1526                 struct page *page;
1527                 unsigned int len;
1528
1529                 BUG_ON(i >= MAX_SKB_FRAGS);
1530                 page = alloc_page(GFP_ATOMIC);
1531                 if (!page) {
1532                         int j;
1533                         skb->truesize += skb->data_len;
1534                         for (j = 0; j < i; j++)
1535                                 put_page(frags[j].page.p);
1536                         return -ENOMEM;
1537                 }
1538
1539                 if (offset + PAGE_SIZE < skb->len)
1540                         len = PAGE_SIZE;
1541                 else
1542                         len = skb->len - offset;
1543                 if (skb_copy_bits(skb, offset, page_address(page), len))
1544                         BUG();
1545
1546                 offset += len;
1547                 frags[i].page.p = page;
1548                 frags[i].page_offset = 0;
1549                 skb_frag_size_set(&frags[i], len);
1550         }
1551
1552         /* Release all the original (foreign) frags. */
1553         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1554                 skb_frag_unref(skb, f);
1555         uarg = skb_shinfo(skb)->destructor_arg;
1556         /* increase inflight counter to offset decrement in callback */
1557         atomic_inc(&queue->inflight_packets);
1558         uarg->callback(uarg, true);
1559         skb_shinfo(skb)->destructor_arg = NULL;
1560
1561         /* Fill the skb with the new (local) frags. */
1562         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1563         skb_shinfo(skb)->nr_frags = i;
1564         skb->truesize += i * PAGE_SIZE;
1565
1566         return 0;
1567 }
1568
1569 static int xenvif_tx_submit(struct xenvif_queue *queue)
1570 {
1571         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1572         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1573         struct sk_buff *skb;
1574         int work_done = 0;
1575
1576         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1577                 struct xen_netif_tx_request *txp;
1578                 u16 pending_idx;
1579                 unsigned data_len;
1580
1581                 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1582                 txp = &queue->pending_tx_info[pending_idx].req;
1583
1584                 /* Check the remap error code. */
1585                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1586                         /* If there was an error, xenvif_tx_check_gop is
1587                          * expected to release all the frags which were mapped,
1588                          * so kfree_skb shouldn't do it again
1589                          */
1590                         skb_shinfo(skb)->nr_frags = 0;
1591                         if (skb_has_frag_list(skb)) {
1592                                 struct sk_buff *nskb =
1593                                                 skb_shinfo(skb)->frag_list;
1594                                 skb_shinfo(nskb)->nr_frags = 0;
1595                         }
1596                         kfree_skb(skb);
1597                         continue;
1598                 }
1599
1600                 data_len = skb->len;
1601                 callback_param(queue, pending_idx).ctx = NULL;
1602                 if (data_len < txp->size) {
1603                         /* Append the packet payload as a fragment. */
1604                         txp->offset += data_len;
1605                         txp->size -= data_len;
1606                 } else {
1607                         /* Schedule a response immediately. */
1608                         xenvif_idx_release(queue, pending_idx,
1609                                            XEN_NETIF_RSP_OKAY);
1610                 }
1611
1612                 if (txp->flags & XEN_NETTXF_csum_blank)
1613                         skb->ip_summed = CHECKSUM_PARTIAL;
1614                 else if (txp->flags & XEN_NETTXF_data_validated)
1615                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1616
1617                 xenvif_fill_frags(queue, skb);
1618
1619                 if (unlikely(skb_has_frag_list(skb))) {
1620                         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1621                         xenvif_skb_zerocopy_prepare(queue, nskb);
1622                         if (xenvif_handle_frag_list(queue, skb)) {
1623                                 if (net_ratelimit())
1624                                         netdev_err(queue->vif->dev,
1625                                                    "Not enough memory to consolidate frag_list!\n");
1626                                 xenvif_skb_zerocopy_prepare(queue, skb);
1627                                 kfree_skb(skb);
1628                                 continue;
1629                         }
1630                         /* Copied all the bits from the frag list -- free it. */
1631                         skb_frag_list_init(skb);
1632                         kfree_skb(nskb);
1633                 }
1634
1635                 skb->dev      = queue->vif->dev;
1636                 skb->protocol = eth_type_trans(skb, skb->dev);
1637                 skb_reset_network_header(skb);
1638
1639                 if (checksum_setup(queue, skb)) {
1640                         netdev_dbg(queue->vif->dev,
1641                                    "Can't setup checksum in net_tx_action\n");
1642                         /* We have to set this flag to trigger the callback */
1643                         if (skb_shinfo(skb)->destructor_arg)
1644                                 xenvif_skb_zerocopy_prepare(queue, skb);
1645                         kfree_skb(skb);
1646                         continue;
1647                 }
1648
1649                 skb_probe_transport_header(skb, 0);
1650
1651                 /* If the packet is GSO then we will have just set up the
1652                  * transport header offset in checksum_setup so it's now
1653                  * straightforward to calculate gso_segs.
1654                  */
1655                 if (skb_is_gso(skb)) {
1656                         int mss = skb_shinfo(skb)->gso_size;
1657                         int hdrlen = skb_transport_header(skb) -
1658                                 skb_mac_header(skb) +
1659                                 tcp_hdrlen(skb);
1660
1661                         skb_shinfo(skb)->gso_segs =
1662                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1663                 }
1664
1665                 queue->stats.rx_bytes += skb->len;
1666                 queue->stats.rx_packets++;
1667
1668                 work_done++;
1669
1670                 /* Set this flag right before netif_receive_skb, otherwise
1671                  * someone might think this packet already left netback, and
1672                  * do a skb_copy_ubufs while we are still in control of the
1673                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1674                  */
1675                 if (skb_shinfo(skb)->destructor_arg) {
1676                         xenvif_skb_zerocopy_prepare(queue, skb);
1677                         queue->stats.tx_zerocopy_sent++;
1678                 }
1679
1680                 netif_receive_skb(skb);
1681         }
1682
1683         return work_done;
1684 }
1685
1686 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1687 {
1688         unsigned long flags;
1689         pending_ring_idx_t index;
1690         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1691
1692         /* This is the only place where we grab this lock, to protect callbacks
1693          * from each other.
1694          */
1695         spin_lock_irqsave(&queue->callback_lock, flags);
1696         do {
1697                 u16 pending_idx = ubuf->desc;
1698                 ubuf = (struct ubuf_info *) ubuf->ctx;
1699                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1700                         MAX_PENDING_REQS);
1701                 index = pending_index(queue->dealloc_prod);
1702                 queue->dealloc_ring[index] = pending_idx;
1703                 /* Sync with xenvif_tx_dealloc_action:
1704                  * insert idx then incr producer.
1705                  */
1706                 smp_wmb();
1707                 queue->dealloc_prod++;
1708         } while (ubuf);
1709         spin_unlock_irqrestore(&queue->callback_lock, flags);
1710
1711         if (likely(zerocopy_success))
1712                 queue->stats.tx_zerocopy_success++;
1713         else
1714                 queue->stats.tx_zerocopy_fail++;
1715         xenvif_skb_zerocopy_complete(queue);
1716 }
1717
1718 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1719 {
1720         struct gnttab_unmap_grant_ref *gop;
1721         pending_ring_idx_t dc, dp;
1722         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1723         unsigned int i = 0;
1724
1725         dc = queue->dealloc_cons;
1726         gop = queue->tx_unmap_ops;
1727
1728         /* Free up any grants we have finished using */
1729         do {
1730                 dp = queue->dealloc_prod;
1731
1732                 /* Ensure we see all indices enqueued by all
1733                  * xenvif_zerocopy_callback().
1734                  */
1735                 smp_rmb();
1736
1737                 while (dc != dp) {
1738                         BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1739                         pending_idx =
1740                                 queue->dealloc_ring[pending_index(dc++)];
1741
1742                         pending_idx_release[gop - queue->tx_unmap_ops] =
1743                                 pending_idx;
1744                         queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1745                                 queue->mmap_pages[pending_idx];
1746                         gnttab_set_unmap_op(gop,
1747                                             idx_to_kaddr(queue, pending_idx),
1748                                             GNTMAP_host_map,
1749                                             queue->grant_tx_handle[pending_idx]);
1750                         xenvif_grant_handle_reset(queue, pending_idx);
1751                         ++gop;
1752                 }
1753
1754         } while (dp != queue->dealloc_prod);
1755
1756         queue->dealloc_cons = dc;
1757
1758         if (gop - queue->tx_unmap_ops > 0) {
1759                 int ret;
1760                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1761                                         NULL,
1762                                         queue->pages_to_unmap,
1763                                         gop - queue->tx_unmap_ops);
1764                 if (ret) {
1765                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1766                                    gop - queue->tx_unmap_ops, ret);
1767                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1768                                 if (gop[i].status != GNTST_okay)
1769                                         netdev_err(queue->vif->dev,
1770                                                    " host_addr: 0x%llx handle: 0x%x status: %d\n",
1771                                                    gop[i].host_addr,
1772                                                    gop[i].handle,
1773                                                    gop[i].status);
1774                         }
1775                         BUG();
1776                 }
1777         }
1778
1779         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1780                 xenvif_idx_release(queue, pending_idx_release[i],
1781                                    XEN_NETIF_RSP_OKAY);
1782 }
1783
1784
1785 /* Called after netfront has transmitted */
1786 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1787 {
1788         unsigned nr_mops, nr_cops = 0;
1789         int work_done, ret;
1790
1791         if (unlikely(!tx_work_todo(queue)))
1792                 return 0;
1793
1794         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1795
1796         if (nr_cops == 0)
1797                 return 0;
1798
1799         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1800         if (nr_mops != 0) {
1801                 ret = gnttab_map_refs(queue->tx_map_ops,
1802                                       NULL,
1803                                       queue->pages_to_map,
1804                                       nr_mops);
1805                 if (ret) {
1806                         unsigned int i;
1807
1808                         netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1809                                    nr_mops, ret);
1810                         for (i = 0; i < nr_mops; ++i)
1811                                 WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1812                                              GNTST_okay);
1813                 }
1814         }
1815
1816         work_done = xenvif_tx_submit(queue);
1817
1818         return work_done;
1819 }
1820
1821 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1822                                u8 status)
1823 {
1824         struct pending_tx_info *pending_tx_info;
1825         pending_ring_idx_t index;
1826         unsigned long flags;
1827
1828         pending_tx_info = &queue->pending_tx_info[pending_idx];
1829
1830         spin_lock_irqsave(&queue->response_lock, flags);
1831
1832         make_tx_response(queue, &pending_tx_info->req, status);
1833
1834         /* Release the pending index before pusing the Tx response so
1835          * its available before a new Tx request is pushed by the
1836          * frontend.
1837          */
1838         index = pending_index(queue->pending_prod++);
1839         queue->pending_ring[index] = pending_idx;
1840
1841         push_tx_responses(queue);
1842
1843         spin_unlock_irqrestore(&queue->response_lock, flags);
1844 }
1845
1846
1847 static void make_tx_response(struct xenvif_queue *queue,
1848                              struct xen_netif_tx_request *txp,
1849                              s8       st)
1850 {
1851         RING_IDX i = queue->tx.rsp_prod_pvt;
1852         struct xen_netif_tx_response *resp;
1853
1854         resp = RING_GET_RESPONSE(&queue->tx, i);
1855         resp->id     = txp->id;
1856         resp->status = st;
1857
1858         if (txp->flags & XEN_NETTXF_extra_info)
1859                 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1860
1861         queue->tx.rsp_prod_pvt = ++i;
1862 }
1863
1864 static void push_tx_responses(struct xenvif_queue *queue)
1865 {
1866         int notify;
1867
1868         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1869         if (notify)
1870                 notify_remote_via_irq(queue->tx_irq);
1871 }
1872
1873 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1874                                              u16      id,
1875                                              s8       st,
1876                                              u16      offset,
1877                                              u16      size,
1878                                              u16      flags)
1879 {
1880         RING_IDX i = queue->rx.rsp_prod_pvt;
1881         struct xen_netif_rx_response *resp;
1882
1883         resp = RING_GET_RESPONSE(&queue->rx, i);
1884         resp->offset     = offset;
1885         resp->flags      = flags;
1886         resp->id         = id;
1887         resp->status     = (s16)size;
1888         if (st < 0)
1889                 resp->status = (s16)st;
1890
1891         queue->rx.rsp_prod_pvt = ++i;
1892
1893         return resp;
1894 }
1895
1896 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1897 {
1898         int ret;
1899         struct gnttab_unmap_grant_ref tx_unmap_op;
1900
1901         gnttab_set_unmap_op(&tx_unmap_op,
1902                             idx_to_kaddr(queue, pending_idx),
1903                             GNTMAP_host_map,
1904                             queue->grant_tx_handle[pending_idx]);
1905         xenvif_grant_handle_reset(queue, pending_idx);
1906
1907         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1908                                 &queue->mmap_pages[pending_idx], 1);
1909         if (ret) {
1910                 netdev_err(queue->vif->dev,
1911                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1912                            ret,
1913                            pending_idx,
1914                            tx_unmap_op.host_addr,
1915                            tx_unmap_op.handle,
1916                            tx_unmap_op.status);
1917                 BUG();
1918         }
1919 }
1920
1921 static inline int tx_work_todo(struct xenvif_queue *queue)
1922 {
1923         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1924                 return 1;
1925
1926         return 0;
1927 }
1928
1929 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1930 {
1931         return queue->dealloc_cons != queue->dealloc_prod;
1932 }
1933
1934 void xenvif_unmap_frontend_rings(struct xenvif_queue *queue)
1935 {
1936         if (queue->tx.sring)
1937                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1938                                         queue->tx.sring);
1939         if (queue->rx.sring)
1940                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1941                                         queue->rx.sring);
1942 }
1943
1944 int xenvif_map_frontend_rings(struct xenvif_queue *queue,
1945                               grant_ref_t tx_ring_ref,
1946                               grant_ref_t rx_ring_ref)
1947 {
1948         void *addr;
1949         struct xen_netif_tx_sring *txs;
1950         struct xen_netif_rx_sring *rxs;
1951
1952         int err = -ENOMEM;
1953
1954         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1955                                      &tx_ring_ref, 1, &addr);
1956         if (err)
1957                 goto err;
1958
1959         txs = (struct xen_netif_tx_sring *)addr;
1960         BACK_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE);
1961
1962         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1963                                      &rx_ring_ref, 1, &addr);
1964         if (err)
1965                 goto err;
1966
1967         rxs = (struct xen_netif_rx_sring *)addr;
1968         BACK_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE);
1969
1970         return 0;
1971
1972 err:
1973         xenvif_unmap_frontend_rings(queue);
1974         return err;
1975 }
1976
1977 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
1978 {
1979         struct xenvif *vif = queue->vif;
1980
1981         queue->stalled = true;
1982
1983         /* At least one queue has stalled? Disable the carrier. */
1984         spin_lock(&vif->lock);
1985         if (vif->stalled_queues++ == 0) {
1986                 netdev_info(vif->dev, "Guest Rx stalled");
1987                 netif_carrier_off(vif->dev);
1988         }
1989         spin_unlock(&vif->lock);
1990 }
1991
1992 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
1993 {
1994         struct xenvif *vif = queue->vif;
1995
1996         queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
1997         queue->stalled = false;
1998
1999         /* All queues are ready? Enable the carrier. */
2000         spin_lock(&vif->lock);
2001         if (--vif->stalled_queues == 0) {
2002                 netdev_info(vif->dev, "Guest Rx ready");
2003                 netif_carrier_on(vif->dev);
2004         }
2005         spin_unlock(&vif->lock);
2006 }
2007
2008 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
2009 {
2010         RING_IDX prod, cons;
2011
2012         prod = queue->rx.sring->req_prod;
2013         cons = queue->rx.req_cons;
2014
2015         return !queue->stalled && prod - cons < 1
2016                 && time_after(jiffies,
2017                               queue->last_rx_time + queue->vif->stall_timeout);
2018 }
2019
2020 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
2021 {
2022         RING_IDX prod, cons;
2023
2024         prod = queue->rx.sring->req_prod;
2025         cons = queue->rx.req_cons;
2026
2027         return queue->stalled && prod - cons >= 1;
2028 }
2029
2030 bool xenvif_have_rx_work(struct xenvif_queue *queue, bool test_kthread)
2031 {
2032         return (!skb_queue_empty(&queue->rx_queue)
2033                 && xenvif_rx_ring_slots_available(queue))
2034                 || (queue->vif->stall_timeout &&
2035                     (xenvif_rx_queue_stalled(queue)
2036                      || xenvif_rx_queue_ready(queue)))
2037                 || (test_kthread && kthread_should_stop())
2038                 || queue->vif->disabled;
2039 }
2040
2041 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
2042 {
2043         struct sk_buff *skb;
2044         long timeout;
2045
2046         skb = skb_peek(&queue->rx_queue);
2047         if (!skb)
2048                 return MAX_SCHEDULE_TIMEOUT;
2049
2050         timeout = XENVIF_RX_CB(skb)->expires - jiffies;
2051         return timeout < 0 ? 0 : timeout;
2052 }
2053
2054 /* Wait until the guest Rx thread has work.
2055  *
2056  * The timeout needs to be adjusted based on the current head of the
2057  * queue (and not just the head at the beginning).  In particular, if
2058  * the queue is initially empty an infinite timeout is used and this
2059  * needs to be reduced when a skb is queued.
2060  *
2061  * This cannot be done with wait_event_timeout() because it only
2062  * calculates the timeout once.
2063  */
2064 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
2065 {
2066         DEFINE_WAIT(wait);
2067
2068         if (xenvif_have_rx_work(queue, true))
2069                 return;
2070
2071         for (;;) {
2072                 long ret;
2073
2074                 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
2075                 if (xenvif_have_rx_work(queue, true))
2076                         break;
2077                 if (xenvif_atomic_fetch_andnot(NETBK_RX_EOI | NETBK_COMMON_EOI,
2078                                         &queue->eoi_pending) &
2079                     (NETBK_RX_EOI | NETBK_COMMON_EOI))
2080                         xen_irq_lateeoi(queue->rx_irq, 0);
2081
2082                 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
2083                 if (!ret)
2084                         break;
2085         }
2086         finish_wait(&queue->wq, &wait);
2087 }
2088
2089 int xenvif_kthread_guest_rx(void *data)
2090 {
2091         struct xenvif_queue *queue = data;
2092         struct xenvif *vif = queue->vif;
2093
2094         if (!vif->stall_timeout)
2095                 xenvif_queue_carrier_on(queue);
2096
2097         for (;;) {
2098                 xenvif_wait_for_rx_work(queue);
2099
2100                 if (kthread_should_stop())
2101                         break;
2102
2103                 /* This frontend is found to be rogue, disable it in
2104                  * kthread context. Currently this is only set when
2105                  * netback finds out frontend sends malformed packet,
2106                  * but we cannot disable the interface in softirq
2107                  * context so we defer it here, if this thread is
2108                  * associated with queue 0.
2109                  */
2110                 if (unlikely(vif->disabled && queue->id == 0)) {
2111                         xenvif_carrier_off(vif);
2112                         break;
2113                 }
2114
2115                 if (!skb_queue_empty(&queue->rx_queue))
2116                         xenvif_rx_action(queue);
2117
2118                 /* If the guest hasn't provided any Rx slots for a
2119                  * while it's probably not responsive, drop the
2120                  * carrier so packets are dropped earlier.
2121                  */
2122                 if (vif->stall_timeout) {
2123                         if (xenvif_rx_queue_stalled(queue))
2124                                 xenvif_queue_carrier_off(queue);
2125                         else if (xenvif_rx_queue_ready(queue))
2126                                 xenvif_queue_carrier_on(queue);
2127                 }
2128
2129                 /* Queued packets may have foreign pages from other
2130                  * domains.  These cannot be queued indefinitely as
2131                  * this would starve guests of grant refs and transmit
2132                  * slots.
2133                  */
2134                 xenvif_rx_queue_drop_expired(queue);
2135
2136                 xenvif_rx_queue_maybe_wake(queue);
2137
2138                 cond_resched();
2139         }
2140
2141         /* Bin any remaining skbs */
2142         xenvif_rx_queue_purge(queue);
2143
2144         return 0;
2145 }
2146
2147 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
2148 {
2149         /* Dealloc thread must remain running until all inflight
2150          * packets complete.
2151          */
2152         return kthread_should_stop() &&
2153                 !atomic_read(&queue->inflight_packets);
2154 }
2155
2156 int xenvif_dealloc_kthread(void *data)
2157 {
2158         struct xenvif_queue *queue = data;
2159
2160         for (;;) {
2161                 wait_event_interruptible(queue->dealloc_wq,
2162                                          tx_dealloc_work_todo(queue) ||
2163                                          xenvif_dealloc_kthread_should_stop(queue));
2164                 if (xenvif_dealloc_kthread_should_stop(queue))
2165                         break;
2166
2167                 xenvif_tx_dealloc_action(queue);
2168                 cond_resched();
2169         }
2170
2171         /* Unmap anything remaining*/
2172         if (tx_dealloc_work_todo(queue))
2173                 xenvif_tx_dealloc_action(queue);
2174
2175         return 0;
2176 }
2177
2178 static int __init netback_init(void)
2179 {
2180         int rc = 0;
2181
2182         if (!xen_domain())
2183                 return -ENODEV;
2184
2185         /* Allow as many queues as there are CPUs but max. 8 if user has not
2186          * specified a value.
2187          */
2188         if (xenvif_max_queues == 0)
2189                 xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
2190                                           num_online_cpus());
2191
2192         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2193                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2194                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2195                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2196         }
2197
2198         rc = xenvif_xenbus_init();
2199         if (rc)
2200                 goto failed_init;
2201
2202 #ifdef CONFIG_DEBUG_FS
2203         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2204         if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2205                 pr_warn("Init of debugfs returned %ld!\n",
2206                         PTR_ERR(xen_netback_dbg_root));
2207 #endif /* CONFIG_DEBUG_FS */
2208
2209         return 0;
2210
2211 failed_init:
2212         return rc;
2213 }
2214
2215 module_init(netback_init);
2216
2217 static void __exit netback_fini(void)
2218 {
2219 #ifdef CONFIG_DEBUG_FS
2220         if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2221                 debugfs_remove_recursive(xen_netback_dbg_root);
2222 #endif /* CONFIG_DEBUG_FS */
2223         xenvif_xenbus_fini();
2224 }
2225 module_exit(netback_fini);
2226
2227 MODULE_LICENSE("Dual BSD/GPL");
2228 MODULE_ALIAS("xen-backend:vif");