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 /* This is the maximum number of flows in the hash cache. */
  93 #define XENVIF_HASH_CACHE_SIZE_DEFAULT 64
  94 unsigned int xenvif_hash_cache_size = XENVIF_HASH_CACHE_SIZE_DEFAULT;
  95 module_param_named(hash_cache_size, xenvif_hash_cache_size, uint, 0644);
  96 MODULE_PARM_DESC(hash_cache_size, "Number of flows in the hash cache");
  97
  98 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
  99                                u8 status);
 100
 101 static void make_tx_response(struct xenvif_queue *queue,
 102                              struct xen_netif_tx_request *txp,
 103                              unsigned int extra_count,
 104                              s8       st);
 105 static void push_tx_responses(struct xenvif_queue *queue);
 106
 107 static inline int tx_work_todo(struct xenvif_queue *queue);
 108
 109 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
 110                                              u16      id,
 111                                              s8       st,
 112                                              u16      offset,
 113                                              u16      size,
 114                                              u16      flags);
 115
 116 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
 117                                        u16 idx)
 118 {
 119         return page_to_pfn(queue->mmap_pages[idx]);
 120 }
 121
 122 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
 123                                          u16 idx)
 124 {
 125         return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
 126 }
 127
 128 #define callback_param(vif, pending_idx) \
 129         (vif->pending_tx_info[pending_idx].callback_struct)
 130
 131 /* Find the containing VIF's structure from a pointer in pending_tx_info array
 132  */
 133 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
 134 {
 135         u16 pending_idx = ubuf->desc;
 136         struct pending_tx_info *temp =
 137                 container_of(ubuf, struct pending_tx_info, callback_struct);
 138         return container_of(temp - pending_idx,
 139                             struct xenvif_queue,
 140                             pending_tx_info[0]);
 141 }
 142
 143 static u16 frag_get_pending_idx(skb_frag_t *frag)
 144 {
 145         return (u16)frag->page_offset;
 146 }
 147
 148 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 149 {
 150         frag->page_offset = pending_idx;
 151 }
 152
 153 static inline pending_ring_idx_t pending_index(unsigned i)
 154 {
 155         return i & (MAX_PENDING_REQS-1);
 156 }
 157
 158 static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
 159 {
 160         RING_IDX prod, cons;
 161         struct sk_buff *skb;
 162         int needed;
 163
 164         skb = skb_peek(&queue->rx_queue);
 165         if (!skb)
 166                 return false;
 167
 168         needed = DIV_ROUND_UP(skb->len, XEN_PAGE_SIZE);
 169         if (skb_is_gso(skb))
 170                 needed++;
 171         if (skb->sw_hash)
 172                 needed++;
 173
 174         do {
 175                 prod = queue->rx.sring->req_prod;
 176                 cons = queue->rx.req_cons;
 177
 178                 if (prod - cons >= needed)
 179                         return true;
 180
 181                 queue->rx.sring->req_event = prod + 1;
 182
 183                 /* Make sure event is visible before we check prod
 184                  * again.
 185                  */
 186                 mb();
 187         } while (queue->rx.sring->req_prod != prod);
 188
 189         return false;
 190 }
 191
 192 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
 193 {
 194         unsigned long flags;
 195
 196         spin_lock_irqsave(&queue->rx_queue.lock, flags);
 197
 198         __skb_queue_tail(&queue->rx_queue, skb);
 199
 200         queue->rx_queue_len += skb->len;
 201         if (queue->rx_queue_len > queue->rx_queue_max)
 202                 netif_tx_stop_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
 203
 204         spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
 205 }
 206
 207 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
 208 {
 209         struct sk_buff *skb;
 210
 211         spin_lock_irq(&queue->rx_queue.lock);
 212
 213         skb = __skb_dequeue(&queue->rx_queue);
 214         if (skb)
 215                 queue->rx_queue_len -= skb->len;
 216
 217         spin_unlock_irq(&queue->rx_queue.lock);
 218
 219         return skb;
 220 }
 221
 222 static void xenvif_rx_queue_maybe_wake(struct xenvif_queue *queue)
 223 {
 224         spin_lock_irq(&queue->rx_queue.lock);
 225
 226         if (queue->rx_queue_len < queue->rx_queue_max)
 227                 netif_tx_wake_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
 228
 229         spin_unlock_irq(&queue->rx_queue.lock);
 230 }
 231
 232
 233 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
 234 {
 235         struct sk_buff *skb;
 236         while ((skb = xenvif_rx_dequeue(queue)) != NULL)
 237                 kfree_skb(skb);
 238 }
 239
 240 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
 241 {
 242         struct sk_buff *skb;
 243
 244         for(;;) {
 245                 skb = skb_peek(&queue->rx_queue);
 246                 if (!skb)
 247                         break;
 248                 if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
 249                         break;
 250                 xenvif_rx_dequeue(queue);
 251                 kfree_skb(skb);
 252         }
 253 }
 254
 255 struct netrx_pending_operations {
 256         unsigned copy_prod, copy_cons;
 257         unsigned meta_prod, meta_cons;
 258         struct gnttab_copy *copy;
 259         struct xenvif_rx_meta *meta;
 260         int copy_off;
 261         grant_ref_t copy_gref;
 262 };
 263
 264 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue,
 265                                                  struct netrx_pending_operations *npo)
 266 {
 267         struct xenvif_rx_meta *meta;
 268         struct xen_netif_rx_request req;
 269
 270         RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
 271
 272         meta = npo->meta + npo->meta_prod++;
 273         meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
 274         meta->gso_size = 0;
 275         meta->size = 0;
 276         meta->id = req.id;
 277
 278         npo->copy_off = 0;
 279         npo->copy_gref = req.gref;
 280
 281         return meta;
 282 }
 283
 284 struct gop_frag_copy {
 285         struct xenvif_queue *queue;
 286         struct netrx_pending_operations *npo;
 287         struct xenvif_rx_meta *meta;
 288         int head;
 289         int gso_type;
 290         int protocol;
 291         int hash_present;
 292
 293         struct page *page;
 294 };
 295
 296 static void xenvif_setup_copy_gop(unsigned long gfn,
 297                                   unsigned int offset,
 298                                   unsigned int *len,
 299                                   struct gop_frag_copy *info)
 300 {
 301         struct gnttab_copy *copy_gop;
 302         struct xen_page_foreign *foreign;
 303         /* Convenient aliases */
 304         struct xenvif_queue *queue = info->queue;
 305         struct netrx_pending_operations *npo = info->npo;
 306         struct page *page = info->page;
 307
 308         BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
 309
 310         if (npo->copy_off == MAX_BUFFER_OFFSET)
 311                 info->meta = get_next_rx_buffer(queue, npo);
 312
 313         if (npo->copy_off + *len > MAX_BUFFER_OFFSET)
 314                 *len = MAX_BUFFER_OFFSET - npo->copy_off;
 315
 316         copy_gop = npo->copy + npo->copy_prod++;
 317         copy_gop->flags = GNTCOPY_dest_gref;
 318         copy_gop->len = *len;
 319
 320         foreign = xen_page_foreign(page);
 321         if (foreign) {
 322                 copy_gop->source.domid = foreign->domid;
 323                 copy_gop->source.u.ref = foreign->gref;
 324                 copy_gop->flags |= GNTCOPY_source_gref;
 325         } else {
 326                 copy_gop->source.domid = DOMID_SELF;
 327                 copy_gop->source.u.gmfn = gfn;
 328         }
 329         copy_gop->source.offset = offset;
 330
 331         copy_gop->dest.domid = queue->vif->domid;
 332         copy_gop->dest.offset = npo->copy_off;
 333         copy_gop->dest.u.ref = npo->copy_gref;
 334
 335         npo->copy_off += *len;
 336         info->meta->size += *len;
 337
 338         if (!info->head)
 339                 return;
 340
 341         /* Leave a gap for the GSO descriptor. */
 342         if ((1 << info->gso_type) & queue->vif->gso_mask)
 343                 queue->rx.req_cons++;
 344
 345         /* Leave a gap for the hash extra segment. */
 346         if (info->hash_present)
 347                 queue->rx.req_cons++;
 348
 349         info->head = 0; /* There must be something in this buffer now */
 350 }
 351
 352 static void xenvif_gop_frag_copy_grant(unsigned long gfn,
 353                                        unsigned offset,
 354                                        unsigned int len,
 355                                        void *data)
 356 {
 357         unsigned int bytes;
 358
 359         while (len) {
 360                 bytes = len;
 361                 xenvif_setup_copy_gop(gfn, offset, &bytes, data);
 362                 offset += bytes;
 363                 len -= bytes;
 364         }
 365 }
 366
 367 /*
 368  * Set up the grant operations for this fragment. If it's a flipping
 369  * interface, we also set up the unmap request from here.
 370  */
 371 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
 372                                  struct netrx_pending_operations *npo,
 373                                  struct page *page, unsigned long size,
 374                                  unsigned long offset, int *head)
 375 {
 376         struct gop_frag_copy info = {
 377                 .queue = queue,
 378                 .npo = npo,
 379                 .head = *head,
 380                 .gso_type = XEN_NETIF_GSO_TYPE_NONE,
 381                 /* xenvif_set_skb_hash() will have either set a s/w
 382                  * hash or cleared the hash depending on
 383                  * whether the the frontend wants a hash for this skb.
 384                  */
 385                 .hash_present = skb->sw_hash,
 386         };
 387         unsigned long bytes;
 388
 389         if (skb_is_gso(skb)) {
 390                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 391                         info.gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
 392                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 393                         info.gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
 394         }
 395
 396         /* Data must not cross a page boundary. */
 397         BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
 398
 399         info.meta = npo->meta + npo->meta_prod - 1;
 400
 401         /* Skip unused frames from start of page */
 402         page += offset >> PAGE_SHIFT;
 403         offset &= ~PAGE_MASK;
 404
 405         while (size > 0) {
 406                 BUG_ON(offset >= PAGE_SIZE);
 407
 408                 bytes = PAGE_SIZE - offset;
 409                 if (bytes > size)
 410                         bytes = size;
 411
 412                 info.page = page;
 413                 gnttab_foreach_grant_in_range(page, offset, bytes,
 414                                               xenvif_gop_frag_copy_grant,
 415                                               &info);
 416                 size -= bytes;
 417                 offset = 0;
 418
 419                 /* Next page */
 420                 if (size) {
 421                         BUG_ON(!PageCompound(page));
 422                         page++;
 423                 }
 424         }
 425
 426         *head = info.head;
 427 }
 428
 429 /*
 430  * Prepare an SKB to be transmitted to the frontend.
 431  *
 432  * This function is responsible for allocating grant operations, meta
 433  * structures, etc.
 434  *
 435  * It returns the number of meta structures consumed. The number of
 436  * ring slots used is always equal to the number of meta slots used
 437  * plus the number of GSO descriptors used. Currently, we use either
 438  * zero GSO descriptors (for non-GSO packets) or one descriptor (for
 439  * frontend-side LRO).
 440  */
 441 static int xenvif_gop_skb(struct sk_buff *skb,
 442                           struct netrx_pending_operations *npo,
 443                           struct xenvif_queue *queue)
 444 {
 445         struct xenvif *vif = netdev_priv(skb->dev);
 446         int nr_frags = skb_shinfo(skb)->nr_frags;
 447         int i;
 448         struct xen_netif_rx_request req;
 449         struct xenvif_rx_meta *meta;
 450         unsigned char *data;
 451         int head = 1;
 452         int old_meta_prod;
 453         int gso_type;
 454
 455         old_meta_prod = npo->meta_prod;
 456
 457         gso_type = XEN_NETIF_GSO_TYPE_NONE;
 458         if (skb_is_gso(skb)) {
 459                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 460                         gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
 461                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 462                         gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
 463         }
 464
 465         /* Set up a GSO prefix descriptor, if necessary */
 466         if ((1 << gso_type) & vif->gso_prefix_mask) {
 467                 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
 468                 meta = npo->meta + npo->meta_prod++;
 469                 meta->gso_type = gso_type;
 470                 meta->gso_size = skb_shinfo(skb)->gso_size;
 471                 meta->size = 0;
 472                 meta->id = req.id;
 473         }
 474
 475         RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
 476         meta = npo->meta + npo->meta_prod++;
 477
 478         if ((1 << gso_type) & vif->gso_mask) {
 479                 meta->gso_type = gso_type;
 480                 meta->gso_size = skb_shinfo(skb)->gso_size;
 481         } else {
 482                 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
 483                 meta->gso_size = 0;
 484         }
 485
 486         meta->size = 0;
 487         meta->id = req.id;
 488         npo->copy_off = 0;
 489         npo->copy_gref = req.gref;
 490
 491         data = skb->data;
 492         while (data < skb_tail_pointer(skb)) {
 493                 unsigned int offset = offset_in_page(data);
 494                 unsigned int len = PAGE_SIZE - offset;
 495
 496                 if (data + len > skb_tail_pointer(skb))
 497                         len = skb_tail_pointer(skb) - data;
 498
 499                 xenvif_gop_frag_copy(queue, skb, npo,
 500                                      virt_to_page(data), len, offset, &head);
 501                 data += len;
 502         }
 503
 504         for (i = 0; i < nr_frags; i++) {
 505                 xenvif_gop_frag_copy(queue, skb, npo,
 506                                      skb_frag_page(&skb_shinfo(skb)->frags[i]),
 507                                      skb_frag_size(&skb_shinfo(skb)->frags[i]),
 508                                      skb_shinfo(skb)->frags[i].page_offset,
 509                                      &head);
 510         }
 511
 512         return npo->meta_prod - old_meta_prod;
 513 }
 514
 515 /*
 516  * This is a twin to xenvif_gop_skb.  Assume that xenvif_gop_skb was
 517  * used to set up the operations on the top of
 518  * netrx_pending_operations, which have since been done.  Check that
 519  * they didn't give any errors and advance over them.
 520  */
 521 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
 522                             struct netrx_pending_operations *npo)
 523 {
 524         struct gnttab_copy     *copy_op;
 525         int status = XEN_NETIF_RSP_OKAY;
 526         int i;
 527
 528         for (i = 0; i < nr_meta_slots; i++) {
 529                 copy_op = npo->copy + npo->copy_cons++;
 530                 if (copy_op->status != GNTST_okay) {
 531                         netdev_dbg(vif->dev,
 532                                    "Bad status %d from copy to DOM%d.\n",
 533                                    copy_op->status, vif->domid);
 534                         status = XEN_NETIF_RSP_ERROR;
 535                 }
 536         }
 537
 538         return status;
 539 }
 540
 541 static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status,
 542                                       struct xenvif_rx_meta *meta,
 543                                       int nr_meta_slots)
 544 {
 545         int i;
 546         unsigned long offset;
 547
 548         /* No fragments used */
 549         if (nr_meta_slots <= 1)
 550                 return;
 551
 552         nr_meta_slots--;
 553
 554         for (i = 0; i < nr_meta_slots; i++) {
 555                 int flags;
 556                 if (i == nr_meta_slots - 1)
 557                         flags = 0;
 558                 else
 559                         flags = XEN_NETRXF_more_data;
 560
 561                 offset = 0;
 562                 make_rx_response(queue, meta[i].id, status, offset,
 563                                  meta[i].size, flags);
 564         }
 565 }
 566
 567 void xenvif_kick_thread(struct xenvif_queue *queue)
 568 {
 569         wake_up(&queue->wq);
 570 }
 571
 572 static void xenvif_rx_action(struct xenvif_queue *queue)
 573 {
 574         struct xenvif *vif = queue->vif;
 575         s8 status;
 576         u16 flags;
 577         struct xen_netif_rx_response *resp;
 578         struct sk_buff_head rxq;
 579         struct sk_buff *skb;
 580         LIST_HEAD(notify);
 581         int ret;
 582         unsigned long offset;
 583         bool need_to_notify = false;
 584
 585         struct netrx_pending_operations npo = {
 586                 .copy  = queue->grant_copy_op,
 587                 .meta  = queue->meta,
 588         };
 589
 590         skb_queue_head_init(&rxq);
 591
 592         while (xenvif_rx_ring_slots_available(queue)
 593                && (skb = xenvif_rx_dequeue(queue)) != NULL) {
 594                 queue->last_rx_time = jiffies;
 595
 596                 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
 597
 598                 __skb_queue_tail(&rxq, skb);
 599         }
 600
 601         BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
 602
 603         if (!npo.copy_prod)
 604                 goto done;
 605
 606         BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
 607         gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
 608
 609         while ((skb = __skb_dequeue(&rxq)) != NULL) {
 610                 struct xen_netif_extra_info *extra = NULL;
 611
 612                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 613                     vif->gso_prefix_mask) {
 614                         resp = RING_GET_RESPONSE(&queue->rx,
 615                                                  queue->rx.rsp_prod_pvt++);
 616
 617                         resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
 618
 619                         resp->offset = queue->meta[npo.meta_cons].gso_size;
 620                         resp->id = queue->meta[npo.meta_cons].id;
 621                         resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
 622
 623                         npo.meta_cons++;
 624                         XENVIF_RX_CB(skb)->meta_slots_used--;
 625                 }
 626
 627
 628                 queue->stats.tx_bytes += skb->len;
 629                 queue->stats.tx_packets++;
 630
 631                 status = xenvif_check_gop(vif,
 632                                           XENVIF_RX_CB(skb)->meta_slots_used,
 633                                           &npo);
 634
 635                 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
 636                         flags = 0;
 637                 else
 638                         flags = XEN_NETRXF_more_data;
 639
 640                 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
 641                         flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
 642                 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 643                         /* remote but checksummed. */
 644                         flags |= XEN_NETRXF_data_validated;
 645
 646                 offset = 0;
 647                 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
 648                                         status, offset,
 649                                         queue->meta[npo.meta_cons].size,
 650                                         flags);
 651
 652                 if ((1 << queue->meta[npo.meta_cons].gso_type) &
 653                     vif->gso_mask) {
 654                         extra = (struct xen_netif_extra_info *)
 655                                 RING_GET_RESPONSE(&queue->rx,
 656                                                   queue->rx.rsp_prod_pvt++);
 657
 658                         resp->flags |= XEN_NETRXF_extra_info;
 659
 660                         extra->u.gso.type = queue->meta[npo.meta_cons].gso_type;
 661                         extra->u.gso.size = queue->meta[npo.meta_cons].gso_size;
 662                         extra->u.gso.pad = 0;
 663                         extra->u.gso.features = 0;
 664
 665                         extra->type = XEN_NETIF_EXTRA_TYPE_GSO;
 666                         extra->flags = 0;
 667                 }
 668
 669                 if (skb->sw_hash) {
 670                         /* Since the skb got here via xenvif_select_queue()
 671                          * we know that the hash has been re-calculated
 672                          * according to a configuration set by the frontend
 673                          * and therefore we know that it is legitimate to
 674                          * pass it to the frontend.
 675                          */
 676                         if (resp->flags & XEN_NETRXF_extra_info)
 677                                 extra->flags |= XEN_NETIF_EXTRA_FLAG_MORE;
 678                         else
 679                                 resp->flags |= XEN_NETRXF_extra_info;
 680
 681                         extra = (struct xen_netif_extra_info *)
 682                                 RING_GET_RESPONSE(&queue->rx,
 683                                                   queue->rx.rsp_prod_pvt++);
 684
 685                         extra->u.hash.algorithm =
 686                                 XEN_NETIF_CTRL_HASH_ALGORITHM_TOEPLITZ;
 687
 688                         if (skb->l4_hash)
 689                                 extra->u.hash.type =
 690                                         skb->protocol == htons(ETH_P_IP) ?
 691                                         _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP :
 692                                         _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP;
 693                         else
 694                                 extra->u.hash.type =
 695                                         skb->protocol == htons(ETH_P_IP) ?
 696                                         _XEN_NETIF_CTRL_HASH_TYPE_IPV4 :
 697                                         _XEN_NETIF_CTRL_HASH_TYPE_IPV6;
 698
 699                         *(uint32_t *)extra->u.hash.value =
 700                                 skb_get_hash_raw(skb);
 701
 702                         extra->type = XEN_NETIF_EXTRA_TYPE_HASH;
 703                         extra->flags = 0;
 704                 }
 705
 706                 xenvif_add_frag_responses(queue, status,
 707                                           queue->meta + npo.meta_cons + 1,
 708                                           XENVIF_RX_CB(skb)->meta_slots_used);
 709
 710                 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
 711
 712                 need_to_notify |= !!ret;
 713
 714                 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
 715                 dev_kfree_skb(skb);
 716         }
 717
 718 done:
 719         if (need_to_notify)
 720                 notify_remote_via_irq(queue->rx_irq);
 721 }
 722
 723 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 724 {
 725         int more_to_do;
 726
 727         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 728
 729         if (more_to_do)
 730                 napi_schedule(&queue->napi);
 731 }
 732
 733 static void tx_add_credit(struct xenvif_queue *queue)
 734 {
 735         unsigned long max_burst, max_credit;
 736
 737         /*
 738          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 739          * Otherwise the interface can seize up due to insufficient credit.
 740          */
 741         max_burst = max(131072UL, queue->credit_bytes);
 742
 743         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 744         max_credit = queue->remaining_credit + queue->credit_bytes;
 745         if (max_credit < queue->remaining_credit)
 746                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 747
 748         queue->remaining_credit = min(max_credit, max_burst);
 749 }
 750
 751 void xenvif_tx_credit_callback(unsigned long data)
 752 {
 753         struct xenvif_queue *queue = (struct xenvif_queue *)data;
 754         tx_add_credit(queue);
 755         xenvif_napi_schedule_or_enable_events(queue);
 756 }
 757
 758 static void xenvif_tx_err(struct xenvif_queue *queue,
 759                           struct xen_netif_tx_request *txp,
 760                           unsigned int extra_count, RING_IDX end)
 761 {
 762         RING_IDX cons = queue->tx.req_cons;
 763         unsigned long flags;
 764
 765         do {
 766                 spin_lock_irqsave(&queue->response_lock, flags);
 767                 make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
 768                 push_tx_responses(queue);
 769                 spin_unlock_irqrestore(&queue->response_lock, flags);
 770                 if (cons == end)
 771                         break;
 772                 RING_COPY_REQUEST(&queue->tx, cons++, txp);
 773                 extra_count = 0; /* only the first frag can have extras */
 774         } while (1);
 775         queue->tx.req_cons = cons;
 776 }
 777
 778 static void xenvif_fatal_tx_err(struct xenvif *vif)
 779 {
 780         netdev_err(vif->dev, "fatal error; disabling device\n");
 781         vif->disabled = true;
 782         /* Disable the vif from queue 0's kthread */
 783         if (vif->queues)
 784                 xenvif_kick_thread(&vif->queues[0]);
 785 }
 786
 787 static int xenvif_count_requests(struct xenvif_queue *queue,
 788                                  struct xen_netif_tx_request *first,
 789                                  unsigned int extra_count,
 790                                  struct xen_netif_tx_request *txp,
 791                                  int work_to_do)
 792 {
 793         RING_IDX cons = queue->tx.req_cons;
 794         int slots = 0;
 795         int drop_err = 0;
 796         int more_data;
 797
 798         if (!(first->flags & XEN_NETTXF_more_data))
 799                 return 0;
 800
 801         do {
 802                 struct xen_netif_tx_request dropped_tx = { 0 };
 803
 804                 if (slots >= work_to_do) {
 805                         netdev_err(queue->vif->dev,
 806                                    "Asked for %d slots but exceeds this limit\n",
 807                                    work_to_do);
 808                         xenvif_fatal_tx_err(queue->vif);
 809                         return -ENODATA;
 810                 }
 811
 812                 /* This guest is really using too many slots and
 813                  * considered malicious.
 814                  */
 815                 if (unlikely(slots >= fatal_skb_slots)) {
 816                         netdev_err(queue->vif->dev,
 817                                    "Malicious frontend using %d slots, threshold %u\n",
 818                                    slots, fatal_skb_slots);
 819                         xenvif_fatal_tx_err(queue->vif);
 820                         return -E2BIG;
 821                 }
 822
 823                 /* Xen network protocol had implicit dependency on
 824                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 825                  * the historical MAX_SKB_FRAGS value 18 to honor the
 826                  * same behavior as before. Any packet using more than
 827                  * 18 slots but less than fatal_skb_slots slots is
 828                  * dropped
 829                  */
 830                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 831                         if (net_ratelimit())
 832                                 netdev_dbg(queue->vif->dev,
 833                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 834                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 835                         drop_err = -E2BIG;
 836                 }
 837
 838                 if (drop_err)
 839                         txp = &dropped_tx;
 840
 841                 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
 842
 843                 /* If the guest submitted a frame >= 64 KiB then
 844                  * first->size overflowed and following slots will
 845                  * appear to be larger than the frame.
 846                  *
 847                  * This cannot be fatal error as there are buggy
 848                  * frontends that do this.
 849                  *
 850                  * Consume all slots and drop the packet.
 851                  */
 852                 if (!drop_err && txp->size > first->size) {
 853                         if (net_ratelimit())
 854                                 netdev_dbg(queue->vif->dev,
 855                                            "Invalid tx request, slot size %u > remaining size %u\n",
 856                                            txp->size, first->size);
 857                         drop_err = -EIO;
 858                 }
 859
 860                 first->size -= txp->size;
 861                 slots++;
 862
 863                 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
 864                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 865                                  txp->offset, txp->size);
 866                         xenvif_fatal_tx_err(queue->vif);
 867                         return -EINVAL;
 868                 }
 869
 870                 more_data = txp->flags & XEN_NETTXF_more_data;
 871
 872                 if (!drop_err)
 873                         txp++;
 874
 875         } while (more_data);
 876
 877         if (drop_err) {
 878                 xenvif_tx_err(queue, first, extra_count, cons + slots);
 879                 return drop_err;
 880         }
 881
 882         return slots;
 883 }
 884
 885
 886 struct xenvif_tx_cb {
 887         u16 pending_idx;
 888 };
 889
 890 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 891
 892 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 893                                            u16 pending_idx,
 894                                            struct xen_netif_tx_request *txp,
 895                                            unsigned int extra_count,
 896                                            struct gnttab_map_grant_ref *mop)
 897 {
 898         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 899         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 900                           GNTMAP_host_map | GNTMAP_readonly,
 901                           txp->gref, queue->vif->domid);
 902
 903         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 904                sizeof(*txp));
 905         queue->pending_tx_info[pending_idx].extra_count = extra_count;
 906 }
 907
 908 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 909 {
 910         struct sk_buff *skb =
 911                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 912                           GFP_ATOMIC | __GFP_NOWARN);
 913         if (unlikely(skb == NULL))
 914                 return NULL;
 915
 916         /* Packets passed to netif_rx() must have some headroom. */
 917         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 918
 919         /* Initialize it here to avoid later surprises */
 920         skb_shinfo(skb)->destructor_arg = NULL;
 921
 922         return skb;
 923 }
 924
 925 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
 926                                                         struct sk_buff *skb,
 927                                                         struct xen_netif_tx_request *txp,
 928                                                         struct gnttab_map_grant_ref *gop,
 929                                                         unsigned int frag_overflow,
 930                                                         struct sk_buff *nskb)
 931 {
 932         struct skb_shared_info *shinfo = skb_shinfo(skb);
 933         skb_frag_t *frags = shinfo->frags;
 934         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 935         int start;
 936         pending_ring_idx_t index;
 937         unsigned int nr_slots;
 938
 939         nr_slots = shinfo->nr_frags;
 940
 941         /* Skip first skb fragment if it is on same page as header fragment. */
 942         start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
 943
 944         for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
 945              shinfo->nr_frags++, txp++, gop++) {
 946                 index = pending_index(queue->pending_cons++);
 947                 pending_idx = queue->pending_ring[index];
 948                 xenvif_tx_create_map_op(queue, pending_idx, txp, 0, gop);
 949                 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 950         }
 951
 952         if (frag_overflow) {
 953
 954                 shinfo = skb_shinfo(nskb);
 955                 frags = shinfo->frags;
 956
 957                 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
 958                      shinfo->nr_frags++, txp++, gop++) {
 959                         index = pending_index(queue->pending_cons++);
 960                         pending_idx = queue->pending_ring[index];
 961                         xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
 962                                                 gop);
 963                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 964                                              pending_idx);
 965                 }
 966
 967                 skb_shinfo(skb)->frag_list = nskb;
 968         }
 969
 970         return gop;
 971 }
 972
 973 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 974                                            u16 pending_idx,
 975                                            grant_handle_t handle)
 976 {
 977         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 978                      NETBACK_INVALID_HANDLE)) {
 979                 netdev_err(queue->vif->dev,
 980                            "Trying to overwrite active handle! pending_idx: 0x%x\n",
 981                            pending_idx);
 982                 BUG();
 983         }
 984         queue->grant_tx_handle[pending_idx] = handle;
 985 }
 986
 987 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 988                                              u16 pending_idx)
 989 {
 990         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 991                      NETBACK_INVALID_HANDLE)) {
 992                 netdev_err(queue->vif->dev,
 993                            "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 994                            pending_idx);
 995                 BUG();
 996         }
 997         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 998 }
 999
1000 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
1001                                struct sk_buff *skb,
1002                                struct gnttab_map_grant_ref **gopp_map,
1003                                struct gnttab_copy **gopp_copy)
1004 {
1005         struct gnttab_map_grant_ref *gop_map = *gopp_map;
1006         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1007         /* This always points to the shinfo of the skb being checked, which
1008          * could be either the first or the one on the frag_list
1009          */
1010         struct skb_shared_info *shinfo = skb_shinfo(skb);
1011         /* If this is non-NULL, we are currently checking the frag_list skb, and
1012          * this points to the shinfo of the first one
1013          */
1014         struct skb_shared_info *first_shinfo = NULL;
1015         int nr_frags = shinfo->nr_frags;
1016         const bool sharedslot = nr_frags &&
1017                                 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
1018         int i, err;
1019
1020         /* Check status of header. */
1021         err = (*gopp_copy)->status;
1022         if (unlikely(err)) {
1023                 if (net_ratelimit())
1024                         netdev_dbg(queue->vif->dev,
1025                                    "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
1026                                    (*gopp_copy)->status,
1027                                    pending_idx,
1028                                    (*gopp_copy)->source.u.ref);
1029                 /* The first frag might still have this slot mapped */
1030                 if (!sharedslot)
1031                         xenvif_idx_release(queue, pending_idx,
1032                                            XEN_NETIF_RSP_ERROR);
1033         }
1034         (*gopp_copy)++;
1035
1036 check_frags:
1037         for (i = 0; i < nr_frags; i++, gop_map++) {
1038                 int j, newerr;
1039
1040                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
1041
1042                 /* Check error status: if okay then remember grant handle. */
1043                 newerr = gop_map->status;
1044
1045                 if (likely(!newerr)) {
1046                         xenvif_grant_handle_set(queue,
1047                                                 pending_idx,
1048                                                 gop_map->handle);
1049                         /* Had a previous error? Invalidate this fragment. */
1050                         if (unlikely(err)) {
1051                                 xenvif_idx_unmap(queue, pending_idx);
1052                                 /* If the mapping of the first frag was OK, but
1053                                  * the header's copy failed, and they are
1054                                  * sharing a slot, send an error
1055                                  */
1056                                 if (i == 0 && sharedslot)
1057                                         xenvif_idx_release(queue, pending_idx,
1058                                                            XEN_NETIF_RSP_ERROR);
1059                                 else
1060                                         xenvif_idx_release(queue, pending_idx,
1061                                                            XEN_NETIF_RSP_OKAY);
1062                         }
1063                         continue;
1064                 }
1065
1066                 /* Error on this fragment: respond to client with an error. */
1067                 if (net_ratelimit())
1068                         netdev_dbg(queue->vif->dev,
1069                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
1070                                    i,
1071                                    gop_map->status,
1072                                    pending_idx,
1073                                    gop_map->ref);
1074
1075                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
1076
1077                 /* Not the first error? Preceding frags already invalidated. */
1078                 if (err)
1079                         continue;
1080
1081                 /* First error: if the header haven't shared a slot with the
1082                  * first frag, release it as well.
1083                  */
1084                 if (!sharedslot)
1085                         xenvif_idx_release(queue,
1086                                            XENVIF_TX_CB(skb)->pending_idx,
1087                                            XEN_NETIF_RSP_OKAY);
1088
1089                 /* Invalidate preceding fragments of this skb. */
1090                 for (j = 0; j < i; j++) {
1091                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1092                         xenvif_idx_unmap(queue, pending_idx);
1093                         xenvif_idx_release(queue, pending_idx,
1094                                            XEN_NETIF_RSP_OKAY);
1095                 }
1096
1097                 /* And if we found the error while checking the frag_list, unmap
1098                  * the first skb's frags
1099                  */
1100                 if (first_shinfo) {
1101                         for (j = 0; j < first_shinfo->nr_frags; j++) {
1102                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
1103                                 xenvif_idx_unmap(queue, pending_idx);
1104                                 xenvif_idx_release(queue, pending_idx,
1105                                                    XEN_NETIF_RSP_OKAY);
1106                         }
1107                 }
1108
1109                 /* Remember the error: invalidate all subsequent fragments. */
1110                 err = newerr;
1111         }
1112
1113         if (skb_has_frag_list(skb) && !first_shinfo) {
1114                 first_shinfo = skb_shinfo(skb);
1115                 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1116                 nr_frags = shinfo->nr_frags;
1117
1118                 goto check_frags;
1119         }
1120
1121         *gopp_map = gop_map;
1122         return err;
1123 }
1124
1125 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1126 {
1127         struct skb_shared_info *shinfo = skb_shinfo(skb);
1128         int nr_frags = shinfo->nr_frags;
1129         int i;
1130         u16 prev_pending_idx = INVALID_PENDING_IDX;
1131
1132         for (i = 0; i < nr_frags; i++) {
1133                 skb_frag_t *frag = shinfo->frags + i;
1134                 struct xen_netif_tx_request *txp;
1135                 struct page *page;
1136                 u16 pending_idx;
1137
1138                 pending_idx = frag_get_pending_idx(frag);
1139
1140                 /* If this is not the first frag, chain it to the previous*/
1141                 if (prev_pending_idx == INVALID_PENDING_IDX)
1142                         skb_shinfo(skb)->destructor_arg =
1143                                 &callback_param(queue, pending_idx);
1144                 else
1145                         callback_param(queue, prev_pending_idx).ctx =
1146                                 &callback_param(queue, pending_idx);
1147
1148                 callback_param(queue, pending_idx).ctx = NULL;
1149                 prev_pending_idx = pending_idx;
1150
1151                 txp = &queue->pending_tx_info[pending_idx].req;
1152                 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1153                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1154                 skb->len += txp->size;
1155                 skb->data_len += txp->size;
1156                 skb->truesize += txp->size;
1157
1158                 /* Take an extra reference to offset network stack's put_page */
1159                 get_page(queue->mmap_pages[pending_idx]);
1160         }
1161 }
1162
1163 static int xenvif_get_extras(struct xenvif_queue *queue,
1164                              struct xen_netif_extra_info *extras,
1165                              unsigned int *extra_count,
1166                              int work_to_do)
1167 {
1168         struct xen_netif_extra_info extra;
1169         RING_IDX cons = queue->tx.req_cons;
1170
1171         do {
1172                 if (unlikely(work_to_do-- <= 0)) {
1173                         netdev_err(queue->vif->dev, "Missing extra info\n");
1174                         xenvif_fatal_tx_err(queue->vif);
1175                         return -EBADR;
1176                 }
1177
1178                 RING_COPY_REQUEST(&queue->tx, cons, &extra);
1179
1180                 queue->tx.req_cons = ++cons;
1181                 (*extra_count)++;
1182
1183                 if (unlikely(!extra.type ||
1184                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1185                         netdev_err(queue->vif->dev,
1186                                    "Invalid extra type: %d\n", extra.type);
1187                         xenvif_fatal_tx_err(queue->vif);
1188                         return -EINVAL;
1189                 }
1190
1191                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1192         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1193
1194         return work_to_do;
1195 }
1196
1197 static int xenvif_set_skb_gso(struct xenvif *vif,
1198                               struct sk_buff *skb,
1199                               struct xen_netif_extra_info *gso)
1200 {
1201         if (!gso->u.gso.size) {
1202                 netdev_err(vif->dev, "GSO size must not be zero.\n");
1203                 xenvif_fatal_tx_err(vif);
1204                 return -EINVAL;
1205         }
1206
1207         switch (gso->u.gso.type) {
1208         case XEN_NETIF_GSO_TYPE_TCPV4:
1209                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1210                 break;
1211         case XEN_NETIF_GSO_TYPE_TCPV6:
1212                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1213                 break;
1214         default:
1215                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1216                 xenvif_fatal_tx_err(vif);
1217                 return -EINVAL;
1218         }
1219
1220         skb_shinfo(skb)->gso_size = gso->u.gso.size;
1221         /* gso_segs will be calculated later */
1222
1223         return 0;
1224 }
1225
1226 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1227 {
1228         bool recalculate_partial_csum = false;
1229
1230         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1231          * peers can fail to set NETRXF_csum_blank when sending a GSO
1232          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1233          * recalculate the partial checksum.
1234          */
1235         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1236                 queue->stats.rx_gso_checksum_fixup++;
1237                 skb->ip_summed = CHECKSUM_PARTIAL;
1238                 recalculate_partial_csum = true;
1239         }
1240
1241         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1242         if (skb->ip_summed != CHECKSUM_PARTIAL)
1243                 return 0;
1244
1245         return skb_checksum_setup(skb, recalculate_partial_csum);
1246 }
1247
1248 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1249 {
1250         u64 now = get_jiffies_64();
1251         u64 next_credit = queue->credit_window_start +
1252                 msecs_to_jiffies(queue->credit_usec / 1000);
1253
1254         /* Timer could already be pending in rare cases. */
1255         if (timer_pending(&queue->credit_timeout))
1256                 return true;
1257
1258         /* Passed the point where we can replenish credit? */
1259         if (time_after_eq64(now, next_credit)) {
1260                 queue->credit_window_start = now;
1261                 tx_add_credit(queue);
1262         }
1263
1264         /* Still too big to send right now? Set a callback. */
1265         if (size > queue->remaining_credit) {
1266                 queue->credit_timeout.data     =
1267                         (unsigned long)queue;
1268                 mod_timer(&queue->credit_timeout,
1269                           next_credit);
1270                 queue->credit_window_start = next_credit;
1271
1272                 return true;
1273         }
1274
1275         return false;
1276 }
1277
1278 /* No locking is required in xenvif_mcast_add/del() as they are
1279  * only ever invoked from NAPI poll. An RCU list is used because
1280  * xenvif_mcast_match() is called asynchronously, during start_xmit.
1281  */
1282
1283 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
1284 {
1285         struct xenvif_mcast_addr *mcast;
1286
1287         if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
1288                 if (net_ratelimit())
1289                         netdev_err(vif->dev,
1290                                    "Too many multicast addresses\n");
1291                 return -ENOSPC;
1292         }
1293
1294         mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
1295         if (!mcast)
1296                 return -ENOMEM;
1297
1298         ether_addr_copy(mcast->addr, addr);
1299         list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
1300         vif->fe_mcast_count++;
1301
1302         return 0;
1303 }
1304
1305 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
1306 {
1307         struct xenvif_mcast_addr *mcast;
1308
1309         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1310                 if (ether_addr_equal(addr, mcast->addr)) {
1311                         --vif->fe_mcast_count;
1312                         list_del_rcu(&mcast->entry);
1313                         kfree_rcu(mcast, rcu);
1314                         break;
1315                 }
1316         }
1317 }
1318
1319 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
1320 {
1321         struct xenvif_mcast_addr *mcast;
1322
1323         rcu_read_lock();
1324         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1325                 if (ether_addr_equal(addr, mcast->addr)) {
1326                         rcu_read_unlock();
1327                         return true;
1328                 }
1329         }
1330         rcu_read_unlock();
1331
1332         return false;
1333 }
1334
1335 void xenvif_mcast_addr_list_free(struct xenvif *vif)
1336 {
1337         /* No need for locking or RCU here. NAPI poll and TX queue
1338          * are stopped.
1339          */
1340         while (!list_empty(&vif->fe_mcast_addr)) {
1341                 struct xenvif_mcast_addr *mcast;
1342
1343                 mcast = list_first_entry(&vif->fe_mcast_addr,
1344                                          struct xenvif_mcast_addr,
1345                                          entry);
1346                 --vif->fe_mcast_count;
1347                 list_del(&mcast->entry);
1348                 kfree(mcast);
1349         }
1350 }
1351
1352 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1353                                      int budget,
1354                                      unsigned *copy_ops,
1355                                      unsigned *map_ops)
1356 {
1357         struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
1358         struct sk_buff *skb, *nskb;
1359         int ret;
1360         unsigned int frag_overflow;
1361
1362         while (skb_queue_len(&queue->tx_queue) < budget) {
1363                 struct xen_netif_tx_request txreq;
1364                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1365                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1366                 unsigned int extra_count;
1367                 u16 pending_idx;
1368                 RING_IDX idx;
1369                 int work_to_do;
1370                 unsigned int data_len;
1371                 pending_ring_idx_t index;
1372
1373                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1374                     XEN_NETIF_TX_RING_SIZE) {
1375                         netdev_err(queue->vif->dev,
1376                                    "Impossible number of requests. "
1377                                    "req_prod %d, req_cons %d, size %ld\n",
1378                                    queue->tx.sring->req_prod, queue->tx.req_cons,
1379                                    XEN_NETIF_TX_RING_SIZE);
1380                         xenvif_fatal_tx_err(queue->vif);
1381                         break;
1382                 }
1383
1384                 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1385                 if (!work_to_do)
1386                         break;
1387
1388                 idx = queue->tx.req_cons;
1389                 rmb(); /* Ensure that we see the request before we copy it. */
1390                 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
1391
1392                 /* Credit-based scheduling. */
1393                 if (txreq.size > queue->remaining_credit &&
1394                     tx_credit_exceeded(queue, txreq.size))
1395                         break;
1396
1397                 queue->remaining_credit -= txreq.size;
1398
1399                 work_to_do--;
1400                 queue->tx.req_cons = ++idx;
1401
1402                 memset(extras, 0, sizeof(extras));
1403                 extra_count = 0;
1404                 if (txreq.flags & XEN_NETTXF_extra_info) {
1405                         work_to_do = xenvif_get_extras(queue, extras,
1406                                                        &extra_count,
1407                                                        work_to_do);
1408                         idx = queue->tx.req_cons;
1409                         if (unlikely(work_to_do < 0))
1410                                 break;
1411                 }
1412
1413                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
1414                         struct xen_netif_extra_info *extra;
1415
1416                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
1417                         ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
1418
1419                         make_tx_response(queue, &txreq, extra_count,
1420                                          (ret == 0) ?
1421                                          XEN_NETIF_RSP_OKAY :
1422                                          XEN_NETIF_RSP_ERROR);
1423                         push_tx_responses(queue);
1424                         continue;
1425                 }
1426
1427                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
1428                         struct xen_netif_extra_info *extra;
1429
1430                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
1431                         xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
1432
1433                         make_tx_response(queue, &txreq, extra_count,
1434                                          XEN_NETIF_RSP_OKAY);
1435                         push_tx_responses(queue);
1436                         continue;
1437                 }
1438
1439                 ret = xenvif_count_requests(queue, &txreq, extra_count,
1440                                             txfrags, work_to_do);
1441                 if (unlikely(ret < 0))
1442                         break;
1443
1444                 idx += ret;
1445
1446                 if (unlikely(txreq.size < ETH_HLEN)) {
1447                         netdev_dbg(queue->vif->dev,
1448                                    "Bad packet size: %d\n", txreq.size);
1449                         xenvif_tx_err(queue, &txreq, extra_count, idx);
1450                         break;
1451                 }
1452
1453                 /* No crossing a page as the payload mustn't fragment. */
1454                 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
1455                         netdev_err(queue->vif->dev,
1456                                    "txreq.offset: %u, size: %u, end: %lu\n",
1457                                    txreq.offset, txreq.size,
1458                                    (unsigned long)(txreq.offset&~XEN_PAGE_MASK) + txreq.size);
1459                         xenvif_fatal_tx_err(queue->vif);
1460                         break;
1461                 }
1462
1463                 index = pending_index(queue->pending_cons);
1464                 pending_idx = queue->pending_ring[index];
1465
1466                 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
1467                             ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1468                         XEN_NETBACK_TX_COPY_LEN : txreq.size;
1469
1470                 skb = xenvif_alloc_skb(data_len);
1471                 if (unlikely(skb == NULL)) {
1472                         netdev_dbg(queue->vif->dev,
1473                                    "Can't allocate a skb in start_xmit.\n");
1474                         xenvif_tx_err(queue, &txreq, extra_count, idx);
1475                         break;
1476                 }
1477
1478                 skb_shinfo(skb)->nr_frags = ret;
1479                 if (data_len < txreq.size)
1480                         skb_shinfo(skb)->nr_frags++;
1481                 /* At this point shinfo->nr_frags is in fact the number of
1482                  * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1483                  */
1484                 frag_overflow = 0;
1485                 nskb = NULL;
1486                 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1487                         frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1488                         BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1489                         skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1490                         nskb = xenvif_alloc_skb(0);
1491                         if (unlikely(nskb == NULL)) {
1492                                 kfree_skb(skb);
1493                                 xenvif_tx_err(queue, &txreq, extra_count, idx);
1494                                 if (net_ratelimit())
1495                                         netdev_err(queue->vif->dev,
1496                                                    "Can't allocate the frag_list skb.\n");
1497                                 break;
1498                         }
1499                 }
1500
1501                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1502                         struct xen_netif_extra_info *gso;
1503                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1504
1505                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1506                                 /* Failure in xenvif_set_skb_gso is fatal. */
1507                                 kfree_skb(skb);
1508                                 kfree_skb(nskb);
1509                                 break;
1510                         }
1511                 }
1512
1513                 if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
1514                         struct xen_netif_extra_info *extra;
1515                         enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
1516
1517                         extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
1518
1519                         switch (extra->u.hash.type) {
1520                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
1521                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
1522                                 type = PKT_HASH_TYPE_L3;
1523                                 break;
1524
1525                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
1526                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
1527                                 type = PKT_HASH_TYPE_L4;
1528                                 break;
1529
1530                         default:
1531                                 break;
1532                         }
1533
1534                         if (type != PKT_HASH_TYPE_NONE)
1535                                 skb_set_hash(skb,
1536                                              *(u32 *)extra->u.hash.value,
1537                                              type);
1538                 }
1539
1540                 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1541
1542                 __skb_put(skb, data_len);
1543                 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1544                 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1545                 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1546
1547                 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1548                         virt_to_gfn(skb->data);
1549                 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1550                 queue->tx_copy_ops[*copy_ops].dest.offset =
1551                         offset_in_page(skb->data) & ~XEN_PAGE_MASK;
1552
1553                 queue->tx_copy_ops[*copy_ops].len = data_len;
1554                 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1555
1556                 (*copy_ops)++;
1557
1558                 if (data_len < txreq.size) {
1559                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1560                                              pending_idx);
1561                         xenvif_tx_create_map_op(queue, pending_idx, &txreq,
1562                                                 extra_count, gop);
1563                         gop++;
1564                 } else {
1565                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1566                                              INVALID_PENDING_IDX);
1567                         memcpy(&queue->pending_tx_info[pending_idx].req,
1568                                &txreq, sizeof(txreq));
1569                         queue->pending_tx_info[pending_idx].extra_count =
1570                                 extra_count;
1571                 }
1572
1573                 queue->pending_cons++;
1574
1575                 gop = xenvif_get_requests(queue, skb, txfrags, gop,
1576                                           frag_overflow, nskb);
1577
1578                 __skb_queue_tail(&queue->tx_queue, skb);
1579
1580                 queue->tx.req_cons = idx;
1581
1582                 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1583                     (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1584                         break;
1585         }
1586
1587         (*map_ops) = gop - queue->tx_map_ops;
1588         return;
1589 }
1590
1591 /* Consolidate skb with a frag_list into a brand new one with local pages on
1592  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1593  */
1594 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1595 {
1596         unsigned int offset = skb_headlen(skb);
1597         skb_frag_t frags[MAX_SKB_FRAGS];
1598         int i, f;
1599         struct ubuf_info *uarg;
1600         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1601
1602         queue->stats.tx_zerocopy_sent += 2;
1603         queue->stats.tx_frag_overflow++;
1604
1605         xenvif_fill_frags(queue, nskb);
1606         /* Subtract frags size, we will correct it later */
1607         skb->truesize -= skb->data_len;
1608         skb->len += nskb->len;
1609         skb->data_len += nskb->len;
1610
1611         /* create a brand new frags array and coalesce there */
1612         for (i = 0; offset < skb->len; i++) {
1613                 struct page *page;
1614                 unsigned int len;
1615
1616                 BUG_ON(i >= MAX_SKB_FRAGS);
1617                 page = alloc_page(GFP_ATOMIC);
1618                 if (!page) {
1619                         int j;
1620                         skb->truesize += skb->data_len;
1621                         for (j = 0; j < i; j++)
1622                                 put_page(frags[j].page.p);
1623                         return -ENOMEM;
1624                 }
1625
1626                 if (offset + PAGE_SIZE < skb->len)
1627                         len = PAGE_SIZE;
1628                 else
1629                         len = skb->len - offset;
1630                 if (skb_copy_bits(skb, offset, page_address(page), len))
1631                         BUG();
1632
1633                 offset += len;
1634                 frags[i].page.p = page;
1635                 frags[i].page_offset = 0;
1636                 skb_frag_size_set(&frags[i], len);
1637         }
1638
1639         /* Copied all the bits from the frag list -- free it. */
1640         skb_frag_list_init(skb);
1641         xenvif_skb_zerocopy_prepare(queue, nskb);
1642         kfree_skb(nskb);
1643
1644         /* Release all the original (foreign) frags. */
1645         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1646                 skb_frag_unref(skb, f);
1647         uarg = skb_shinfo(skb)->destructor_arg;
1648         /* increase inflight counter to offset decrement in callback */
1649         atomic_inc(&queue->inflight_packets);
1650         uarg->callback(uarg, true);
1651         skb_shinfo(skb)->destructor_arg = NULL;
1652
1653         /* Fill the skb with the new (local) frags. */
1654         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1655         skb_shinfo(skb)->nr_frags = i;
1656         skb->truesize += i * PAGE_SIZE;
1657
1658         return 0;
1659 }
1660
1661 static int xenvif_tx_submit(struct xenvif_queue *queue)
1662 {
1663         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1664         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1665         struct sk_buff *skb;
1666         int work_done = 0;
1667
1668         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1669                 struct xen_netif_tx_request *txp;
1670                 u16 pending_idx;
1671                 unsigned data_len;
1672
1673                 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1674                 txp = &queue->pending_tx_info[pending_idx].req;
1675
1676                 /* Check the remap error code. */
1677                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1678                         /* If there was an error, xenvif_tx_check_gop is
1679                          * expected to release all the frags which were mapped,
1680                          * so kfree_skb shouldn't do it again
1681                          */
1682                         skb_shinfo(skb)->nr_frags = 0;
1683                         if (skb_has_frag_list(skb)) {
1684                                 struct sk_buff *nskb =
1685                                                 skb_shinfo(skb)->frag_list;
1686                                 skb_shinfo(nskb)->nr_frags = 0;
1687                         }
1688                         kfree_skb(skb);
1689                         continue;
1690                 }
1691
1692                 data_len = skb->len;
1693                 callback_param(queue, pending_idx).ctx = NULL;
1694                 if (data_len < txp->size) {
1695                         /* Append the packet payload as a fragment. */
1696                         txp->offset += data_len;
1697                         txp->size -= data_len;
1698                 } else {
1699                         /* Schedule a response immediately. */
1700                         xenvif_idx_release(queue, pending_idx,
1701                                            XEN_NETIF_RSP_OKAY);
1702                 }
1703
1704                 if (txp->flags & XEN_NETTXF_csum_blank)
1705                         skb->ip_summed = CHECKSUM_PARTIAL;
1706                 else if (txp->flags & XEN_NETTXF_data_validated)
1707                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1708
1709                 xenvif_fill_frags(queue, skb);
1710
1711                 if (unlikely(skb_has_frag_list(skb))) {
1712                         if (xenvif_handle_frag_list(queue, skb)) {
1713                                 if (net_ratelimit())
1714                                         netdev_err(queue->vif->dev,
1715                                                    "Not enough memory to consolidate frag_list!\n");
1716                                 xenvif_skb_zerocopy_prepare(queue, skb);
1717                                 kfree_skb(skb);
1718                                 continue;
1719                         }
1720                 }
1721
1722                 skb->dev      = queue->vif->dev;
1723                 skb->protocol = eth_type_trans(skb, skb->dev);
1724                 skb_reset_network_header(skb);
1725
1726                 if (checksum_setup(queue, skb)) {
1727                         netdev_dbg(queue->vif->dev,
1728                                    "Can't setup checksum in net_tx_action\n");
1729                         /* We have to set this flag to trigger the callback */
1730                         if (skb_shinfo(skb)->destructor_arg)
1731                                 xenvif_skb_zerocopy_prepare(queue, skb);
1732                         kfree_skb(skb);
1733                         continue;
1734                 }
1735
1736                 skb_probe_transport_header(skb, 0);
1737
1738                 /* If the packet is GSO then we will have just set up the
1739                  * transport header offset in checksum_setup so it's now
1740                  * straightforward to calculate gso_segs.
1741                  */
1742                 if (skb_is_gso(skb)) {
1743                         int mss = skb_shinfo(skb)->gso_size;
1744                         int hdrlen = skb_transport_header(skb) -
1745                                 skb_mac_header(skb) +
1746                                 tcp_hdrlen(skb);
1747
1748                         skb_shinfo(skb)->gso_segs =
1749                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1750                 }
1751
1752                 queue->stats.rx_bytes += skb->len;
1753                 queue->stats.rx_packets++;
1754
1755                 work_done++;
1756
1757                 /* Set this flag right before netif_receive_skb, otherwise
1758                  * someone might think this packet already left netback, and
1759                  * do a skb_copy_ubufs while we are still in control of the
1760                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1761                  */
1762                 if (skb_shinfo(skb)->destructor_arg) {
1763                         xenvif_skb_zerocopy_prepare(queue, skb);
1764                         queue->stats.tx_zerocopy_sent++;
1765                 }
1766
1767                 netif_receive_skb(skb);
1768         }
1769
1770         return work_done;
1771 }
1772
1773 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1774 {
1775         unsigned long flags;
1776         pending_ring_idx_t index;
1777         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1778
1779         /* This is the only place where we grab this lock, to protect callbacks
1780          * from each other.
1781          */
1782         spin_lock_irqsave(&queue->callback_lock, flags);
1783         do {
1784                 u16 pending_idx = ubuf->desc;
1785                 ubuf = (struct ubuf_info *) ubuf->ctx;
1786                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1787                         MAX_PENDING_REQS);
1788                 index = pending_index(queue->dealloc_prod);
1789                 queue->dealloc_ring[index] = pending_idx;
1790                 /* Sync with xenvif_tx_dealloc_action:
1791                  * insert idx then incr producer.
1792                  */
1793                 smp_wmb();
1794                 queue->dealloc_prod++;
1795         } while (ubuf);
1796         spin_unlock_irqrestore(&queue->callback_lock, flags);
1797
1798         if (likely(zerocopy_success))
1799                 queue->stats.tx_zerocopy_success++;
1800         else
1801                 queue->stats.tx_zerocopy_fail++;
1802         xenvif_skb_zerocopy_complete(queue);
1803 }
1804
1805 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1806 {
1807         struct gnttab_unmap_grant_ref *gop;
1808         pending_ring_idx_t dc, dp;
1809         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1810         unsigned int i = 0;
1811
1812         dc = queue->dealloc_cons;
1813         gop = queue->tx_unmap_ops;
1814
1815         /* Free up any grants we have finished using */
1816         do {
1817                 dp = queue->dealloc_prod;
1818
1819                 /* Ensure we see all indices enqueued by all
1820                  * xenvif_zerocopy_callback().
1821                  */
1822                 smp_rmb();
1823
1824                 while (dc != dp) {
1825                         BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1826                         pending_idx =
1827                                 queue->dealloc_ring[pending_index(dc++)];
1828
1829                         pending_idx_release[gop - queue->tx_unmap_ops] =
1830                                 pending_idx;
1831                         queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1832                                 queue->mmap_pages[pending_idx];
1833                         gnttab_set_unmap_op(gop,
1834                                             idx_to_kaddr(queue, pending_idx),
1835                                             GNTMAP_host_map,
1836                                             queue->grant_tx_handle[pending_idx]);
1837                         xenvif_grant_handle_reset(queue, pending_idx);
1838                         ++gop;
1839                 }
1840
1841         } while (dp != queue->dealloc_prod);
1842
1843         queue->dealloc_cons = dc;
1844
1845         if (gop - queue->tx_unmap_ops > 0) {
1846                 int ret;
1847                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1848                                         NULL,
1849                                         queue->pages_to_unmap,
1850                                         gop - queue->tx_unmap_ops);
1851                 if (ret) {
1852                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1853                                    gop - queue->tx_unmap_ops, ret);
1854                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1855                                 if (gop[i].status != GNTST_okay)
1856                                         netdev_err(queue->vif->dev,
1857                                                    " host_addr: 0x%llx handle: 0x%x status: %d\n",
1858                                                    gop[i].host_addr,
1859                                                    gop[i].handle,
1860                                                    gop[i].status);
1861                         }
1862                         BUG();
1863                 }
1864         }
1865
1866         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1867                 xenvif_idx_release(queue, pending_idx_release[i],
1868                                    XEN_NETIF_RSP_OKAY);
1869 }
1870
1871
1872 /* Called after netfront has transmitted */
1873 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1874 {
1875         unsigned nr_mops, nr_cops = 0;
1876         int work_done, ret;
1877
1878         if (unlikely(!tx_work_todo(queue)))
1879                 return 0;
1880
1881         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1882
1883         if (nr_cops == 0)
1884                 return 0;
1885
1886         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1887         if (nr_mops != 0) {
1888                 ret = gnttab_map_refs(queue->tx_map_ops,
1889                                       NULL,
1890                                       queue->pages_to_map,
1891                                       nr_mops);
1892                 BUG_ON(ret);
1893         }
1894
1895         work_done = xenvif_tx_submit(queue);
1896
1897         return work_done;
1898 }
1899
1900 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1901                                u8 status)
1902 {
1903         struct pending_tx_info *pending_tx_info;
1904         pending_ring_idx_t index;
1905         unsigned long flags;
1906
1907         pending_tx_info = &queue->pending_tx_info[pending_idx];
1908
1909         spin_lock_irqsave(&queue->response_lock, flags);
1910
1911         make_tx_response(queue, &pending_tx_info->req,
1912                          pending_tx_info->extra_count, status);
1913
1914         /* Release the pending index before pusing the Tx response so
1915          * its available before a new Tx request is pushed by the
1916          * frontend.
1917          */
1918         index = pending_index(queue->pending_prod++);
1919         queue->pending_ring[index] = pending_idx;
1920
1921         push_tx_responses(queue);
1922
1923         spin_unlock_irqrestore(&queue->response_lock, flags);
1924 }
1925
1926
1927 static void make_tx_response(struct xenvif_queue *queue,
1928                              struct xen_netif_tx_request *txp,
1929                              unsigned int extra_count,
1930                              s8       st)
1931 {
1932         RING_IDX i = queue->tx.rsp_prod_pvt;
1933         struct xen_netif_tx_response *resp;
1934
1935         resp = RING_GET_RESPONSE(&queue->tx, i);
1936         resp->id     = txp->id;
1937         resp->status = st;
1938
1939         while (extra_count-- != 0)
1940                 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1941
1942         queue->tx.rsp_prod_pvt = ++i;
1943 }
1944
1945 static void push_tx_responses(struct xenvif_queue *queue)
1946 {
1947         int notify;
1948
1949         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1950         if (notify)
1951                 notify_remote_via_irq(queue->tx_irq);
1952 }
1953
1954 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1955                                              u16      id,
1956                                              s8       st,
1957                                              u16      offset,
1958                                              u16      size,
1959                                              u16      flags)
1960 {
1961         RING_IDX i = queue->rx.rsp_prod_pvt;
1962         struct xen_netif_rx_response *resp;
1963
1964         resp = RING_GET_RESPONSE(&queue->rx, i);
1965         resp->offset     = offset;
1966         resp->flags      = flags;
1967         resp->id         = id;
1968         resp->status     = (s16)size;
1969         if (st < 0)
1970                 resp->status = (s16)st;
1971
1972         queue->rx.rsp_prod_pvt = ++i;
1973
1974         return resp;
1975 }
1976
1977 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1978 {
1979         int ret;
1980         struct gnttab_unmap_grant_ref tx_unmap_op;
1981
1982         gnttab_set_unmap_op(&tx_unmap_op,
1983                             idx_to_kaddr(queue, pending_idx),
1984                             GNTMAP_host_map,
1985                             queue->grant_tx_handle[pending_idx]);
1986         xenvif_grant_handle_reset(queue, pending_idx);
1987
1988         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1989                                 &queue->mmap_pages[pending_idx], 1);
1990         if (ret) {
1991                 netdev_err(queue->vif->dev,
1992                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1993                            ret,
1994                            pending_idx,
1995                            tx_unmap_op.host_addr,
1996                            tx_unmap_op.handle,
1997                            tx_unmap_op.status);
1998                 BUG();
1999         }
2000 }
2001
2002 static inline int tx_work_todo(struct xenvif_queue *queue)
2003 {
2004         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
2005                 return 1;
2006
2007         return 0;
2008 }
2009
2010 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
2011 {
2012         return queue->dealloc_cons != queue->dealloc_prod;
2013 }
2014
2015 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
2016 {
2017         if (queue->tx.sring)
2018                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
2019                                         queue->tx.sring);
2020         if (queue->rx.sring)
2021                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
2022                                         queue->rx.sring);
2023 }
2024
2025 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
2026                                    grant_ref_t tx_ring_ref,
2027                                    grant_ref_t rx_ring_ref)
2028 {
2029         void *addr;
2030         struct xen_netif_tx_sring *txs;
2031         struct xen_netif_rx_sring *rxs;
2032
2033         int err = -ENOMEM;
2034
2035         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
2036                                      &tx_ring_ref, 1, &addr);
2037         if (err)
2038                 goto err;
2039
2040         txs = (struct xen_netif_tx_sring *)addr;
2041         BACK_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE);
2042
2043         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
2044                                      &rx_ring_ref, 1, &addr);
2045         if (err)
2046                 goto err;
2047
2048         rxs = (struct xen_netif_rx_sring *)addr;
2049         BACK_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE);
2050
2051         return 0;
2052
2053 err:
2054         xenvif_unmap_frontend_data_rings(queue);
2055         return err;
2056 }
2057
2058 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
2059 {
2060         struct xenvif *vif = queue->vif;
2061
2062         queue->stalled = true;
2063
2064         /* At least one queue has stalled? Disable the carrier. */
2065         spin_lock(&vif->lock);
2066         if (vif->stalled_queues++ == 0) {
2067                 netdev_info(vif->dev, "Guest Rx stalled");
2068                 netif_carrier_off(vif->dev);
2069         }
2070         spin_unlock(&vif->lock);
2071 }
2072
2073 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
2074 {
2075         struct xenvif *vif = queue->vif;
2076
2077         queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
2078         queue->stalled = false;
2079
2080         /* All queues are ready? Enable the carrier. */
2081         spin_lock(&vif->lock);
2082         if (--vif->stalled_queues == 0) {
2083                 netdev_info(vif->dev, "Guest Rx ready");
2084                 netif_carrier_on(vif->dev);
2085         }
2086         spin_unlock(&vif->lock);
2087 }
2088
2089 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
2090 {
2091         RING_IDX prod, cons;
2092
2093         prod = queue->rx.sring->req_prod;
2094         cons = queue->rx.req_cons;
2095
2096         return !queue->stalled && prod - cons < 1
2097                 && time_after(jiffies,
2098                               queue->last_rx_time + queue->vif->stall_timeout);
2099 }
2100
2101 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
2102 {
2103         RING_IDX prod, cons;
2104
2105         prod = queue->rx.sring->req_prod;
2106         cons = queue->rx.req_cons;
2107
2108         return queue->stalled && prod - cons >= 1;
2109 }
2110
2111 static bool xenvif_have_rx_work(struct xenvif_queue *queue)
2112 {
2113         return xenvif_rx_ring_slots_available(queue)
2114                 || (queue->vif->stall_timeout &&
2115                     (xenvif_rx_queue_stalled(queue)
2116                      || xenvif_rx_queue_ready(queue)))
2117                 || kthread_should_stop()
2118                 || queue->vif->disabled;
2119 }
2120
2121 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
2122 {
2123         struct sk_buff *skb;
2124         long timeout;
2125
2126         skb = skb_peek(&queue->rx_queue);
2127         if (!skb)
2128                 return MAX_SCHEDULE_TIMEOUT;
2129
2130         timeout = XENVIF_RX_CB(skb)->expires - jiffies;
2131         return timeout < 0 ? 0 : timeout;
2132 }
2133
2134 /* Wait until the guest Rx thread has work.
2135  *
2136  * The timeout needs to be adjusted based on the current head of the
2137  * queue (and not just the head at the beginning).  In particular, if
2138  * the queue is initially empty an infinite timeout is used and this
2139  * needs to be reduced when a skb is queued.
2140  *
2141  * This cannot be done with wait_event_timeout() because it only
2142  * calculates the timeout once.
2143  */
2144 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
2145 {
2146         DEFINE_WAIT(wait);
2147
2148         if (xenvif_have_rx_work(queue))
2149                 return;
2150
2151         for (;;) {
2152                 long ret;
2153
2154                 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
2155                 if (xenvif_have_rx_work(queue))
2156                         break;
2157                 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
2158                 if (!ret)
2159                         break;
2160         }
2161         finish_wait(&queue->wq, &wait);
2162 }
2163
2164 int xenvif_kthread_guest_rx(void *data)
2165 {
2166         struct xenvif_queue *queue = data;
2167         struct xenvif *vif = queue->vif;
2168
2169         if (!vif->stall_timeout)
2170                 xenvif_queue_carrier_on(queue);
2171
2172         for (;;) {
2173                 xenvif_wait_for_rx_work(queue);
2174
2175                 if (kthread_should_stop())
2176                         break;
2177
2178                 /* This frontend is found to be rogue, disable it in
2179                  * kthread context. Currently this is only set when
2180                  * netback finds out frontend sends malformed packet,
2181                  * but we cannot disable the interface in softirq
2182                  * context so we defer it here, if this thread is
2183                  * associated with queue 0.
2184                  */
2185                 if (unlikely(vif->disabled && queue->id == 0)) {
2186                         xenvif_carrier_off(vif);
2187                         break;
2188                 }
2189
2190                 if (!skb_queue_empty(&queue->rx_queue))
2191                         xenvif_rx_action(queue);
2192
2193                 /* If the guest hasn't provided any Rx slots for a
2194                  * while it's probably not responsive, drop the
2195                  * carrier so packets are dropped earlier.
2196                  */
2197                 if (vif->stall_timeout) {
2198                         if (xenvif_rx_queue_stalled(queue))
2199                                 xenvif_queue_carrier_off(queue);
2200                         else if (xenvif_rx_queue_ready(queue))
2201                                 xenvif_queue_carrier_on(queue);
2202                 }
2203
2204                 /* Queued packets may have foreign pages from other
2205                  * domains.  These cannot be queued indefinitely as
2206                  * this would starve guests of grant refs and transmit
2207                  * slots.
2208                  */
2209                 xenvif_rx_queue_drop_expired(queue);
2210
2211                 xenvif_rx_queue_maybe_wake(queue);
2212
2213                 cond_resched();
2214         }
2215
2216         /* Bin any remaining skbs */
2217         xenvif_rx_queue_purge(queue);
2218
2219         return 0;
2220 }
2221
2222 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
2223 {
2224         /* Dealloc thread must remain running until all inflight
2225          * packets complete.
2226          */
2227         return kthread_should_stop() &&
2228                 !atomic_read(&queue->inflight_packets);
2229 }
2230
2231 int xenvif_dealloc_kthread(void *data)
2232 {
2233         struct xenvif_queue *queue = data;
2234
2235         for (;;) {
2236                 wait_event_interruptible(queue->dealloc_wq,
2237                                          tx_dealloc_work_todo(queue) ||
2238                                          xenvif_dealloc_kthread_should_stop(queue));
2239                 if (xenvif_dealloc_kthread_should_stop(queue))
2240                         break;
2241
2242                 xenvif_tx_dealloc_action(queue);
2243                 cond_resched();
2244         }
2245
2246         /* Unmap anything remaining*/
2247         if (tx_dealloc_work_todo(queue))
2248                 xenvif_tx_dealloc_action(queue);
2249
2250         return 0;
2251 }
2252
2253 static void make_ctrl_response(struct xenvif *vif,
2254                                const struct xen_netif_ctrl_request *req,
2255                                u32 status, u32 data)
2256 {
2257         RING_IDX idx = vif->ctrl.rsp_prod_pvt;
2258         struct xen_netif_ctrl_response rsp = {
2259                 .id = req->id,
2260                 .type = req->type,
2261                 .status = status,
2262                 .data = data,
2263         };
2264
2265         *RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
2266         vif->ctrl.rsp_prod_pvt = ++idx;
2267 }
2268
2269 static void push_ctrl_response(struct xenvif *vif)
2270 {
2271         int notify;
2272
2273         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
2274         if (notify)
2275                 notify_remote_via_irq(vif->ctrl_irq);
2276 }
2277
2278 static void process_ctrl_request(struct xenvif *vif,
2279                                  const struct xen_netif_ctrl_request *req)
2280 {
2281         u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
2282         u32 data = 0;
2283
2284         switch (req->type) {
2285         case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
2286                 status = xenvif_set_hash_alg(vif, req->data[0]);
2287                 break;
2288
2289         case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
2290                 status = xenvif_get_hash_flags(vif, &data);
2291                 break;
2292
2293         case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
2294                 status = xenvif_set_hash_flags(vif, req->data[0]);
2295                 break;
2296
2297         case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
2298                 status = xenvif_set_hash_key(vif, req->data[0],
2299                                              req->data[1]);
2300                 break;
2301
2302         case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
2303                 status = XEN_NETIF_CTRL_STATUS_SUCCESS;
2304                 data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
2305                 break;
2306
2307         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
2308                 status = xenvif_set_hash_mapping_size(vif,
2309                                                       req->data[0]);
2310                 break;
2311
2312         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
2313                 status = xenvif_set_hash_mapping(vif, req->data[0],
2314                                                  req->data[1],
2315                                                  req->data[2]);
2316                 break;
2317
2318         default:
2319                 break;
2320         }
2321
2322         make_ctrl_response(vif, req, status, data);
2323         push_ctrl_response(vif);
2324 }
2325
2326 static void xenvif_ctrl_action(struct xenvif *vif)
2327 {
2328         for (;;) {
2329                 RING_IDX req_prod, req_cons;
2330
2331                 req_prod = vif->ctrl.sring->req_prod;
2332                 req_cons = vif->ctrl.req_cons;
2333
2334                 /* Make sure we can see requests before we process them. */
2335                 rmb();
2336
2337                 if (req_cons == req_prod)
2338                         break;
2339
2340                 while (req_cons != req_prod) {
2341                         struct xen_netif_ctrl_request req;
2342
2343                         RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
2344                         req_cons++;
2345
2346                         process_ctrl_request(vif, &req);
2347                 }
2348
2349                 vif->ctrl.req_cons = req_cons;
2350                 vif->ctrl.sring->req_event = req_cons + 1;
2351         }
2352 }
2353
2354 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
2355 {
2356         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
2357                 return 1;
2358
2359         return 0;
2360 }
2361
2362 irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
2363 {
2364         struct xenvif *vif = data;
2365
2366         while (xenvif_ctrl_work_todo(vif))
2367                 xenvif_ctrl_action(vif);
2368
2369         return IRQ_HANDLED;
2370 }
2371
2372 static int __init netback_init(void)
2373 {
2374         int rc = 0;
2375
2376         if (!xen_domain())
2377                 return -ENODEV;
2378
2379         /* Allow as many queues as there are CPUs if user has not
2380          * specified a value.
2381          */
2382         if (xenvif_max_queues == 0)
2383                 xenvif_max_queues = num_online_cpus();
2384
2385         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2386                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2387                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2388                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2389         }
2390
2391         rc = xenvif_xenbus_init();
2392         if (rc)
2393                 goto failed_init;
2394
2395 #ifdef CONFIG_DEBUG_FS
2396         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2397         if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2398                 pr_warn("Init of debugfs returned %ld!\n",
2399                         PTR_ERR(xen_netback_dbg_root));
2400 #endif /* CONFIG_DEBUG_FS */
2401
2402         return 0;
2403
2404 failed_init:
2405         return rc;
2406 }
2407
2408 module_init(netback_init);
2409
2410 static void __exit netback_fini(void)
2411 {
2412 #ifdef CONFIG_DEBUG_FS
2413         if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2414                 debugfs_remove_recursive(xen_netback_dbg_root);
2415 #endif /* CONFIG_DEBUG_FS */
2416         xenvif_xenbus_fini();
2417 }
2418 module_exit(netback_fini);
2419
2420 MODULE_LICENSE("Dual BSD/GPL");
2421 MODULE_ALIAS("xen-backend:vif");