drivers/net/xen-netback/netback.c

   1 /*
   2  * Back-end of the driver for virtual network devices. This portion of the
   3  * driver exports a 'unified' network-device interface that can be accessed
   4  * by any operating system that implements a compatible front end. A
   5  * reference front-end implementation can be found in:
   6  *  drivers/net/xen-netfront.c
   7  *
   8  * Copyright (c) 2002-2005, K A Fraser
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public License version 2
  12  * as published by the Free Software Foundation; or, when distributed
  13  * separately from the Linux kernel or incorporated into other
  14  * software packages, subject to the following license:
  15  *
  16  * Permission is hereby granted, free of charge, to any person obtaining a copy
  17  * of this source file (the "Software"), to deal in the Software without
  18  * restriction, including without limitation the rights to use, copy, modify,
  19  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  20  * and to permit persons to whom the Software is furnished to do so, subject to
  21  * the following conditions:
  22  *
  23  * The above copyright notice and this permission notice shall be included in
  24  * all copies or substantial portions of the Software.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  27  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  28  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  29  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  30  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  31  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  32  * IN THE SOFTWARE.
  33  */
  34
  35 #include "common.h"
  36
  37 #include <linux/kthread.h>
  38 #include <linux/if_vlan.h>
  39 #include <linux/udp.h>
  40 #include <linux/highmem.h>
  41
  42 #include <net/tcp.h>
  43
  44 #include <xen/xen.h>
  45 #include <xen/events.h>
  46 #include <xen/interface/memory.h>
  47 #include <xen/page.h>
  48
  49 #include <asm/xen/hypercall.h>
  50
  51 /* Provide an option to disable split event channels at load time as
  52  * event channels are limited resource. Split event channels are
  53  * enabled by default.
  54  */
  55 bool separate_tx_rx_irq = true;
  56 module_param(separate_tx_rx_irq, bool, 0644);
  57
  58 /* The time that packets can stay on the guest Rx internal queue
  59  * before they are dropped.
  60  */
  61 unsigned int rx_drain_timeout_msecs = 10000;
  62 module_param(rx_drain_timeout_msecs, uint, 0444);
  63
  64 /* The length of time before the frontend is considered unresponsive
  65  * because it isn't providing Rx slots.
  66  */
  67 unsigned int rx_stall_timeout_msecs = 60000;
  68 module_param(rx_stall_timeout_msecs, uint, 0444);
  69
  70 #define MAX_QUEUES_DEFAULT 8
  71 unsigned int xenvif_max_queues;
  72 module_param_named(max_queues, xenvif_max_queues, uint, 0644);
  73 MODULE_PARM_DESC(max_queues,
  74                  "Maximum number of queues per virtual interface");
  75
  76 /*
  77  * This is the maximum slots a skb can have. If a guest sends a skb
  78  * which exceeds this limit it is considered malicious.
  79  */
  80 #define FATAL_SKB_SLOTS_DEFAULT 20
  81 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
  82 module_param(fatal_skb_slots, uint, 0444);
  83
  84 /* The amount to copy out of the first guest Tx slot into the skb's
  85  * linear area.  If the first slot has more data, it will be mapped
  86  * and put into the first frag.
  87  *
  88  * This is sized to avoid pulling headers from the frags for most
  89  * TCP/IP packets.
  90  */
  91 #define XEN_NETBACK_TX_COPY_LEN 128
  92
  93 /* This is the maximum number of flows in the hash cache. */
  94 #define XENVIF_HASH_CACHE_SIZE_DEFAULT 64
  95 unsigned int xenvif_hash_cache_size = XENVIF_HASH_CACHE_SIZE_DEFAULT;
  96 module_param_named(hash_cache_size, xenvif_hash_cache_size, uint, 0644);
  97 MODULE_PARM_DESC(hash_cache_size, "Number of flows in the hash cache");
  98
  99 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
 100                                u8 status);
 101
 102 static void make_tx_response(struct xenvif_queue *queue,
 103                              struct xen_netif_tx_request *txp,
 104                              unsigned int extra_count,
 105                              s8       st);
 106 static void push_tx_responses(struct xenvif_queue *queue);
 107
 108 static inline int tx_work_todo(struct xenvif_queue *queue);
 109
 110 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
 111                                        u16 idx)
 112 {
 113         return page_to_pfn(queue->mmap_pages[idx]);
 114 }
 115
 116 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
 117                                          u16 idx)
 118 {
 119         return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
 120 }
 121
 122 #define callback_param(vif, pending_idx) \
 123         (vif->pending_tx_info[pending_idx].callback_struct)
 124
 125 /* Find the containing VIF's structure from a pointer in pending_tx_info array
 126  */
 127 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
 128 {
 129         u16 pending_idx = ubuf->desc;
 130         struct pending_tx_info *temp =
 131                 container_of(ubuf, struct pending_tx_info, callback_struct);
 132         return container_of(temp - pending_idx,
 133                             struct xenvif_queue,
 134                             pending_tx_info[0]);
 135 }
 136
 137 static u16 frag_get_pending_idx(skb_frag_t *frag)
 138 {
 139         return (u16)frag->page_offset;
 140 }
 141
 142 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 143 {
 144         frag->page_offset = pending_idx;
 145 }
 146
 147 static inline pending_ring_idx_t pending_index(unsigned i)
 148 {
 149         return i & (MAX_PENDING_REQS-1);
 150 }
 151
 152 void xenvif_kick_thread(struct xenvif_queue *queue)
 153 {
 154         wake_up(&queue->wq);
 155 }
 156
 157 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 158 {
 159         int more_to_do;
 160
 161         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 162
 163         if (more_to_do)
 164                 napi_schedule(&queue->napi);
 165         else if (atomic_fetch_andnot(NETBK_TX_EOI | NETBK_COMMON_EOI,
 166                                      &queue->eoi_pending) &
 167                  (NETBK_TX_EOI | NETBK_COMMON_EOI))
 168                 xen_irq_lateeoi(queue->tx_irq, 0);
 169 }
 170
 171 static void tx_add_credit(struct xenvif_queue *queue)
 172 {
 173         unsigned long max_burst, max_credit;
 174
 175         /*
 176          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 177          * Otherwise the interface can seize up due to insufficient credit.
 178          */
 179         max_burst = max(131072UL, queue->credit_bytes);
 180
 181         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 182         max_credit = queue->remaining_credit + queue->credit_bytes;
 183         if (max_credit < queue->remaining_credit)
 184                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 185
 186         queue->remaining_credit = min(max_credit, max_burst);
 187         queue->rate_limited = false;
 188 }
 189
 190 void xenvif_tx_credit_callback(unsigned long data)
 191 {
 192         struct xenvif_queue *queue = (struct xenvif_queue *)data;
 193         tx_add_credit(queue);
 194         xenvif_napi_schedule_or_enable_events(queue);
 195 }
 196
 197 static void xenvif_tx_err(struct xenvif_queue *queue,
 198                           struct xen_netif_tx_request *txp,
 199                           unsigned int extra_count, RING_IDX end)
 200 {
 201         RING_IDX cons = queue->tx.req_cons;
 202         unsigned long flags;
 203
 204         do {
 205                 spin_lock_irqsave(&queue->response_lock, flags);
 206                 make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
 207                 push_tx_responses(queue);
 208                 spin_unlock_irqrestore(&queue->response_lock, flags);
 209                 if (cons == end)
 210                         break;
 211                 RING_COPY_REQUEST(&queue->tx, cons++, txp);
 212                 extra_count = 0; /* only the first frag can have extras */
 213         } while (1);
 214         queue->tx.req_cons = cons;
 215 }
 216
 217 static void xenvif_fatal_tx_err(struct xenvif *vif)
 218 {
 219         netdev_err(vif->dev, "fatal error; disabling device\n");
 220         vif->disabled = true;
 221         /* Disable the vif from queue 0's kthread */
 222         if (vif->queues)
 223                 xenvif_kick_thread(&vif->queues[0]);
 224 }
 225
 226 static int xenvif_count_requests(struct xenvif_queue *queue,
 227                                  struct xen_netif_tx_request *first,
 228                                  unsigned int extra_count,
 229                                  struct xen_netif_tx_request *txp,
 230                                  int work_to_do)
 231 {
 232         RING_IDX cons = queue->tx.req_cons;
 233         int slots = 0;
 234         int drop_err = 0;
 235         int more_data;
 236
 237         if (!(first->flags & XEN_NETTXF_more_data))
 238                 return 0;
 239
 240         do {
 241                 struct xen_netif_tx_request dropped_tx = { 0 };
 242
 243                 if (slots >= work_to_do) {
 244                         netdev_err(queue->vif->dev,
 245                                    "Asked for %d slots but exceeds this limit\n",
 246                                    work_to_do);
 247                         xenvif_fatal_tx_err(queue->vif);
 248                         return -ENODATA;
 249                 }
 250
 251                 /* This guest is really using too many slots and
 252                  * considered malicious.
 253                  */
 254                 if (unlikely(slots >= fatal_skb_slots)) {
 255                         netdev_err(queue->vif->dev,
 256                                    "Malicious frontend using %d slots, threshold %u\n",
 257                                    slots, fatal_skb_slots);
 258                         xenvif_fatal_tx_err(queue->vif);
 259                         return -E2BIG;
 260                 }
 261
 262                 /* Xen network protocol had implicit dependency on
 263                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 264                  * the historical MAX_SKB_FRAGS value 18 to honor the
 265                  * same behavior as before. Any packet using more than
 266                  * 18 slots but less than fatal_skb_slots slots is
 267                  * dropped
 268                  */
 269                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 270                         if (net_ratelimit())
 271                                 netdev_dbg(queue->vif->dev,
 272                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 273                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 274                         drop_err = -E2BIG;
 275                 }
 276
 277                 if (drop_err)
 278                         txp = &dropped_tx;
 279
 280                 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
 281
 282                 /* If the guest submitted a frame >= 64 KiB then
 283                  * first->size overflowed and following slots will
 284                  * appear to be larger than the frame.
 285                  *
 286                  * This cannot be fatal error as there are buggy
 287                  * frontends that do this.
 288                  *
 289                  * Consume all slots and drop the packet.
 290                  */
 291                 if (!drop_err && txp->size > first->size) {
 292                         if (net_ratelimit())
 293                                 netdev_dbg(queue->vif->dev,
 294                                            "Invalid tx request, slot size %u > remaining size %u\n",
 295                                            txp->size, first->size);
 296                         drop_err = -EIO;
 297                 }
 298
 299                 first->size -= txp->size;
 300                 slots++;
 301
 302                 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
 303                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 304                                  txp->offset, txp->size);
 305                         xenvif_fatal_tx_err(queue->vif);
 306                         return -EINVAL;
 307                 }
 308
 309                 more_data = txp->flags & XEN_NETTXF_more_data;
 310
 311                 if (!drop_err)
 312                         txp++;
 313
 314         } while (more_data);
 315
 316         if (drop_err) {
 317                 xenvif_tx_err(queue, first, extra_count, cons + slots);
 318                 return drop_err;
 319         }
 320
 321         return slots;
 322 }
 323
 324
 325 struct xenvif_tx_cb {
 326         u16 pending_idx;
 327 };
 328
 329 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 330
 331 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 332                                            u16 pending_idx,
 333                                            struct xen_netif_tx_request *txp,
 334                                            unsigned int extra_count,
 335                                            struct gnttab_map_grant_ref *mop)
 336 {
 337         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 338         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 339                           GNTMAP_host_map | GNTMAP_readonly,
 340                           txp->gref, queue->vif->domid);
 341
 342         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 343                sizeof(*txp));
 344         queue->pending_tx_info[pending_idx].extra_count = extra_count;
 345 }
 346
 347 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 348 {
 349         struct sk_buff *skb =
 350                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 351                           GFP_ATOMIC | __GFP_NOWARN);
 352         if (unlikely(skb == NULL))
 353                 return NULL;
 354
 355         /* Packets passed to netif_rx() must have some headroom. */
 356         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 357
 358         /* Initialize it here to avoid later surprises */
 359         skb_shinfo(skb)->destructor_arg = NULL;
 360
 361         return skb;
 362 }
 363
 364 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
 365                                                         struct sk_buff *skb,
 366                                                         struct xen_netif_tx_request *txp,
 367                                                         struct gnttab_map_grant_ref *gop,
 368                                                         unsigned int frag_overflow,
 369                                                         struct sk_buff *nskb)
 370 {
 371         struct skb_shared_info *shinfo = skb_shinfo(skb);
 372         skb_frag_t *frags = shinfo->frags;
 373         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 374         int start;
 375         pending_ring_idx_t index;
 376         unsigned int nr_slots;
 377
 378         nr_slots = shinfo->nr_frags;
 379
 380         /* Skip first skb fragment if it is on same page as header fragment. */
 381         start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
 382
 383         for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
 384              shinfo->nr_frags++, txp++, gop++) {
 385                 index = pending_index(queue->pending_cons++);
 386                 pending_idx = queue->pending_ring[index];
 387                 xenvif_tx_create_map_op(queue, pending_idx, txp, 0, gop);
 388                 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 389         }
 390
 391         if (frag_overflow) {
 392
 393                 shinfo = skb_shinfo(nskb);
 394                 frags = shinfo->frags;
 395
 396                 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
 397                      shinfo->nr_frags++, txp++, gop++) {
 398                         index = pending_index(queue->pending_cons++);
 399                         pending_idx = queue->pending_ring[index];
 400                         xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
 401                                                 gop);
 402                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 403                                              pending_idx);
 404                 }
 405
 406                 skb_shinfo(skb)->frag_list = nskb;
 407         }
 408
 409         return gop;
 410 }
 411
 412 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 413                                            u16 pending_idx,
 414                                            grant_handle_t handle)
 415 {
 416         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 417                      NETBACK_INVALID_HANDLE)) {
 418                 netdev_err(queue->vif->dev,
 419                            "Trying to overwrite active handle! pending_idx: 0x%x\n",
 420                            pending_idx);
 421                 BUG();
 422         }
 423         queue->grant_tx_handle[pending_idx] = handle;
 424 }
 425
 426 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 427                                              u16 pending_idx)
 428 {
 429         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 430                      NETBACK_INVALID_HANDLE)) {
 431                 netdev_err(queue->vif->dev,
 432                            "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 433                            pending_idx);
 434                 BUG();
 435         }
 436         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 437 }
 438
 439 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 440                                struct sk_buff *skb,
 441                                struct gnttab_map_grant_ref **gopp_map,
 442                                struct gnttab_copy **gopp_copy)
 443 {
 444         struct gnttab_map_grant_ref *gop_map = *gopp_map;
 445         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 446         /* This always points to the shinfo of the skb being checked, which
 447          * could be either the first or the one on the frag_list
 448          */
 449         struct skb_shared_info *shinfo = skb_shinfo(skb);
 450         /* If this is non-NULL, we are currently checking the frag_list skb, and
 451          * this points to the shinfo of the first one
 452          */
 453         struct skb_shared_info *first_shinfo = NULL;
 454         int nr_frags = shinfo->nr_frags;
 455         const bool sharedslot = nr_frags &&
 456                                 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
 457         int i, err;
 458
 459         /* Check status of header. */
 460         err = (*gopp_copy)->status;
 461         if (unlikely(err)) {
 462                 if (net_ratelimit())
 463                         netdev_dbg(queue->vif->dev,
 464                                    "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 465                                    (*gopp_copy)->status,
 466                                    pending_idx,
 467                                    (*gopp_copy)->source.u.ref);
 468                 /* The first frag might still have this slot mapped */
 469                 if (!sharedslot)
 470                         xenvif_idx_release(queue, pending_idx,
 471                                            XEN_NETIF_RSP_ERROR);
 472         }
 473         (*gopp_copy)++;
 474
 475 check_frags:
 476         for (i = 0; i < nr_frags; i++, gop_map++) {
 477                 int j, newerr;
 478
 479                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 480
 481                 /* Check error status: if okay then remember grant handle. */
 482                 newerr = gop_map->status;
 483
 484                 if (likely(!newerr)) {
 485                         xenvif_grant_handle_set(queue,
 486                                                 pending_idx,
 487                                                 gop_map->handle);
 488                         /* Had a previous error? Invalidate this fragment. */
 489                         if (unlikely(err)) {
 490                                 xenvif_idx_unmap(queue, pending_idx);
 491                                 /* If the mapping of the first frag was OK, but
 492                                  * the header's copy failed, and they are
 493                                  * sharing a slot, send an error
 494                                  */
 495                                 if (i == 0 && !first_shinfo && sharedslot)
 496                                         xenvif_idx_release(queue, pending_idx,
 497                                                            XEN_NETIF_RSP_ERROR);
 498                                 else
 499                                         xenvif_idx_release(queue, pending_idx,
 500                                                            XEN_NETIF_RSP_OKAY);
 501                         }
 502                         continue;
 503                 }
 504
 505                 /* Error on this fragment: respond to client with an error. */
 506                 if (net_ratelimit())
 507                         netdev_dbg(queue->vif->dev,
 508                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
 509                                    i,
 510                                    gop_map->status,
 511                                    pending_idx,
 512                                    gop_map->ref);
 513
 514                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
 515
 516                 /* Not the first error? Preceding frags already invalidated. */
 517                 if (err)
 518                         continue;
 519
 520                 /* First error: if the header haven't shared a slot with the
 521                  * first frag, release it as well.
 522                  */
 523                 if (!sharedslot)
 524                         xenvif_idx_release(queue,
 525                                            XENVIF_TX_CB(skb)->pending_idx,
 526                                            XEN_NETIF_RSP_OKAY);
 527
 528                 /* Invalidate preceding fragments of this skb. */
 529                 for (j = 0; j < i; j++) {
 530                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
 531                         xenvif_idx_unmap(queue, pending_idx);
 532                         xenvif_idx_release(queue, pending_idx,
 533                                            XEN_NETIF_RSP_OKAY);
 534                 }
 535
 536                 /* And if we found the error while checking the frag_list, unmap
 537                  * the first skb's frags
 538                  */
 539                 if (first_shinfo) {
 540                         for (j = 0; j < first_shinfo->nr_frags; j++) {
 541                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
 542                                 xenvif_idx_unmap(queue, pending_idx);
 543                                 xenvif_idx_release(queue, pending_idx,
 544                                                    XEN_NETIF_RSP_OKAY);
 545                         }
 546                 }
 547
 548                 /* Remember the error: invalidate all subsequent fragments. */
 549                 err = newerr;
 550         }
 551
 552         if (skb_has_frag_list(skb) && !first_shinfo) {
 553                 first_shinfo = skb_shinfo(skb);
 554                 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
 555                 nr_frags = shinfo->nr_frags;
 556
 557                 goto check_frags;
 558         }
 559
 560         *gopp_map = gop_map;
 561         return err;
 562 }
 563
 564 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
 565 {
 566         struct skb_shared_info *shinfo = skb_shinfo(skb);
 567         int nr_frags = shinfo->nr_frags;
 568         int i;
 569         u16 prev_pending_idx = INVALID_PENDING_IDX;
 570
 571         for (i = 0; i < nr_frags; i++) {
 572                 skb_frag_t *frag = shinfo->frags + i;
 573                 struct xen_netif_tx_request *txp;
 574                 struct page *page;
 575                 u16 pending_idx;
 576
 577                 pending_idx = frag_get_pending_idx(frag);
 578
 579                 /* If this is not the first frag, chain it to the previous*/
 580                 if (prev_pending_idx == INVALID_PENDING_IDX)
 581                         skb_shinfo(skb)->destructor_arg =
 582                                 &callback_param(queue, pending_idx);
 583                 else
 584                         callback_param(queue, prev_pending_idx).ctx =
 585                                 &callback_param(queue, pending_idx);
 586
 587                 callback_param(queue, pending_idx).ctx = NULL;
 588                 prev_pending_idx = pending_idx;
 589
 590                 txp = &queue->pending_tx_info[pending_idx].req;
 591                 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
 592                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
 593                 skb->len += txp->size;
 594                 skb->data_len += txp->size;
 595                 skb->truesize += txp->size;
 596
 597                 /* Take an extra reference to offset network stack's put_page */
 598                 get_page(queue->mmap_pages[pending_idx]);
 599         }
 600 }
 601
 602 static int xenvif_get_extras(struct xenvif_queue *queue,
 603                              struct xen_netif_extra_info *extras,
 604                              unsigned int *extra_count,
 605                              int work_to_do)
 606 {
 607         struct xen_netif_extra_info extra;
 608         RING_IDX cons = queue->tx.req_cons;
 609
 610         do {
 611                 if (unlikely(work_to_do-- <= 0)) {
 612                         netdev_err(queue->vif->dev, "Missing extra info\n");
 613                         xenvif_fatal_tx_err(queue->vif);
 614                         return -EBADR;
 615                 }
 616
 617                 RING_COPY_REQUEST(&queue->tx, cons, &extra);
 618
 619                 queue->tx.req_cons = ++cons;
 620                 (*extra_count)++;
 621
 622                 if (unlikely(!extra.type ||
 623                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
 624                         netdev_err(queue->vif->dev,
 625                                    "Invalid extra type: %d\n", extra.type);
 626                         xenvif_fatal_tx_err(queue->vif);
 627                         return -EINVAL;
 628                 }
 629
 630                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
 631         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
 632
 633         return work_to_do;
 634 }
 635
 636 static int xenvif_set_skb_gso(struct xenvif *vif,
 637                               struct sk_buff *skb,
 638                               struct xen_netif_extra_info *gso)
 639 {
 640         if (!gso->u.gso.size) {
 641                 netdev_err(vif->dev, "GSO size must not be zero.\n");
 642                 xenvif_fatal_tx_err(vif);
 643                 return -EINVAL;
 644         }
 645
 646         switch (gso->u.gso.type) {
 647         case XEN_NETIF_GSO_TYPE_TCPV4:
 648                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 649                 break;
 650         case XEN_NETIF_GSO_TYPE_TCPV6:
 651                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 652                 break;
 653         default:
 654                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
 655                 xenvif_fatal_tx_err(vif);
 656                 return -EINVAL;
 657         }
 658
 659         skb_shinfo(skb)->gso_size = gso->u.gso.size;
 660         /* gso_segs will be calculated later */
 661
 662         return 0;
 663 }
 664
 665 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
 666 {
 667         bool recalculate_partial_csum = false;
 668
 669         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
 670          * peers can fail to set NETRXF_csum_blank when sending a GSO
 671          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
 672          * recalculate the partial checksum.
 673          */
 674         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
 675                 queue->stats.rx_gso_checksum_fixup++;
 676                 skb->ip_summed = CHECKSUM_PARTIAL;
 677                 recalculate_partial_csum = true;
 678         }
 679
 680         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
 681         if (skb->ip_summed != CHECKSUM_PARTIAL)
 682                 return 0;
 683
 684         return skb_checksum_setup(skb, recalculate_partial_csum);
 685 }
 686
 687 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
 688 {
 689         u64 now = get_jiffies_64();
 690         u64 next_credit = queue->credit_window_start +
 691                 msecs_to_jiffies(queue->credit_usec / 1000);
 692
 693         /* Timer could already be pending in rare cases. */
 694         if (timer_pending(&queue->credit_timeout)) {
 695                 queue->rate_limited = true;
 696                 return true;
 697         }
 698
 699         /* Passed the point where we can replenish credit? */
 700         if (time_after_eq64(now, next_credit)) {
 701                 queue->credit_window_start = now;
 702                 tx_add_credit(queue);
 703         }
 704
 705         /* Still too big to send right now? Set a callback. */
 706         if (size > queue->remaining_credit) {
 707                 queue->credit_timeout.data     =
 708                         (unsigned long)queue;
 709                 mod_timer(&queue->credit_timeout,
 710                           next_credit);
 711                 queue->credit_window_start = next_credit;
 712                 queue->rate_limited = true;
 713
 714                 return true;
 715         }
 716
 717         return false;
 718 }
 719
 720 /* No locking is required in xenvif_mcast_add/del() as they are
 721  * only ever invoked from NAPI poll. An RCU list is used because
 722  * xenvif_mcast_match() is called asynchronously, during start_xmit.
 723  */
 724
 725 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
 726 {
 727         struct xenvif_mcast_addr *mcast;
 728
 729         if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
 730                 if (net_ratelimit())
 731                         netdev_err(vif->dev,
 732                                    "Too many multicast addresses\n");
 733                 return -ENOSPC;
 734         }
 735
 736         mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
 737         if (!mcast)
 738                 return -ENOMEM;
 739
 740         ether_addr_copy(mcast->addr, addr);
 741         list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
 742         vif->fe_mcast_count++;
 743
 744         return 0;
 745 }
 746
 747 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
 748 {
 749         struct xenvif_mcast_addr *mcast;
 750
 751         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 752                 if (ether_addr_equal(addr, mcast->addr)) {
 753                         --vif->fe_mcast_count;
 754                         list_del_rcu(&mcast->entry);
 755                         kfree_rcu(mcast, rcu);
 756                         break;
 757                 }
 758         }
 759 }
 760
 761 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
 762 {
 763         struct xenvif_mcast_addr *mcast;
 764
 765         rcu_read_lock();
 766         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 767                 if (ether_addr_equal(addr, mcast->addr)) {
 768                         rcu_read_unlock();
 769                         return true;
 770                 }
 771         }
 772         rcu_read_unlock();
 773
 774         return false;
 775 }
 776
 777 void xenvif_mcast_addr_list_free(struct xenvif *vif)
 778 {
 779         /* No need for locking or RCU here. NAPI poll and TX queue
 780          * are stopped.
 781          */
 782         while (!list_empty(&vif->fe_mcast_addr)) {
 783                 struct xenvif_mcast_addr *mcast;
 784
 785                 mcast = list_first_entry(&vif->fe_mcast_addr,
 786                                          struct xenvif_mcast_addr,
 787                                          entry);
 788                 --vif->fe_mcast_count;
 789                 list_del(&mcast->entry);
 790                 kfree(mcast);
 791         }
 792 }
 793
 794 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
 795                                      int budget,
 796                                      unsigned *copy_ops,
 797                                      unsigned *map_ops)
 798 {
 799         struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
 800         struct sk_buff *skb, *nskb;
 801         int ret;
 802         unsigned int frag_overflow;
 803
 804         while (skb_queue_len(&queue->tx_queue) < budget) {
 805                 struct xen_netif_tx_request txreq;
 806                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
 807                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
 808                 unsigned int extra_count;
 809                 u16 pending_idx;
 810                 RING_IDX idx;
 811                 int work_to_do;
 812                 unsigned int data_len;
 813                 pending_ring_idx_t index;
 814
 815                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
 816                     XEN_NETIF_TX_RING_SIZE) {
 817                         netdev_err(queue->vif->dev,
 818                                    "Impossible number of requests. "
 819                                    "req_prod %d, req_cons %d, size %ld\n",
 820                                    queue->tx.sring->req_prod, queue->tx.req_cons,
 821                                    XEN_NETIF_TX_RING_SIZE);
 822                         xenvif_fatal_tx_err(queue->vif);
 823                         break;
 824                 }
 825
 826                 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
 827                 if (!work_to_do)
 828                         break;
 829
 830                 idx = queue->tx.req_cons;
 831                 rmb(); /* Ensure that we see the request before we copy it. */
 832                 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
 833
 834                 /* Credit-based scheduling. */
 835                 if (txreq.size > queue->remaining_credit &&
 836                     tx_credit_exceeded(queue, txreq.size))
 837                         break;
 838
 839                 queue->remaining_credit -= txreq.size;
 840
 841                 work_to_do--;
 842                 queue->tx.req_cons = ++idx;
 843
 844                 memset(extras, 0, sizeof(extras));
 845                 extra_count = 0;
 846                 if (txreq.flags & XEN_NETTXF_extra_info) {
 847                         work_to_do = xenvif_get_extras(queue, extras,
 848                                                        &extra_count,
 849                                                        work_to_do);
 850                         idx = queue->tx.req_cons;
 851                         if (unlikely(work_to_do < 0))
 852                                 break;
 853                 }
 854
 855                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
 856                         struct xen_netif_extra_info *extra;
 857
 858                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
 859                         ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
 860
 861                         make_tx_response(queue, &txreq, extra_count,
 862                                          (ret == 0) ?
 863                                          XEN_NETIF_RSP_OKAY :
 864                                          XEN_NETIF_RSP_ERROR);
 865                         push_tx_responses(queue);
 866                         continue;
 867                 }
 868
 869                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
 870                         struct xen_netif_extra_info *extra;
 871
 872                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
 873                         xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
 874
 875                         make_tx_response(queue, &txreq, extra_count,
 876                                          XEN_NETIF_RSP_OKAY);
 877                         push_tx_responses(queue);
 878                         continue;
 879                 }
 880
 881                 ret = xenvif_count_requests(queue, &txreq, extra_count,
 882                                             txfrags, work_to_do);
 883                 if (unlikely(ret < 0))
 884                         break;
 885
 886                 idx += ret;
 887
 888                 if (unlikely(txreq.size < ETH_HLEN)) {
 889                         netdev_dbg(queue->vif->dev,
 890                                    "Bad packet size: %d\n", txreq.size);
 891                         xenvif_tx_err(queue, &txreq, extra_count, idx);
 892                         break;
 893                 }
 894
 895                 /* No crossing a page as the payload mustn't fragment. */
 896                 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
 897                         netdev_err(queue->vif->dev,
 898                                    "txreq.offset: %u, size: %u, end: %lu\n",
 899                                    txreq.offset, txreq.size,
 900                                    (unsigned long)(txreq.offset&~XEN_PAGE_MASK) + txreq.size);
 901                         xenvif_fatal_tx_err(queue->vif);
 902                         break;
 903                 }
 904
 905                 index = pending_index(queue->pending_cons);
 906                 pending_idx = queue->pending_ring[index];
 907
 908                 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
 909                             ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
 910                         XEN_NETBACK_TX_COPY_LEN : txreq.size;
 911
 912                 skb = xenvif_alloc_skb(data_len);
 913                 if (unlikely(skb == NULL)) {
 914                         netdev_dbg(queue->vif->dev,
 915                                    "Can't allocate a skb in start_xmit.\n");
 916                         xenvif_tx_err(queue, &txreq, extra_count, idx);
 917                         break;
 918                 }
 919
 920                 skb_shinfo(skb)->nr_frags = ret;
 921                 if (data_len < txreq.size)
 922                         skb_shinfo(skb)->nr_frags++;
 923                 /* At this point shinfo->nr_frags is in fact the number of
 924                  * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
 925                  */
 926                 frag_overflow = 0;
 927                 nskb = NULL;
 928                 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
 929                         frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
 930                         BUG_ON(frag_overflow > MAX_SKB_FRAGS);
 931                         skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
 932                         nskb = xenvif_alloc_skb(0);
 933                         if (unlikely(nskb == NULL)) {
 934                                 skb_shinfo(skb)->nr_frags = 0;
 935                                 kfree_skb(skb);
 936                                 xenvif_tx_err(queue, &txreq, extra_count, idx);
 937                                 if (net_ratelimit())
 938                                         netdev_err(queue->vif->dev,
 939                                                    "Can't allocate the frag_list skb.\n");
 940                                 break;
 941                         }
 942                 }
 943
 944                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
 945                         struct xen_netif_extra_info *gso;
 946                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
 947
 948                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
 949                                 /* Failure in xenvif_set_skb_gso is fatal. */
 950                                 skb_shinfo(skb)->nr_frags = 0;
 951                                 kfree_skb(skb);
 952                                 kfree_skb(nskb);
 953                                 break;
 954                         }
 955                 }
 956
 957                 if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
 958                         struct xen_netif_extra_info *extra;
 959                         enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
 960
 961                         extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
 962
 963                         switch (extra->u.hash.type) {
 964                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
 965                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
 966                                 type = PKT_HASH_TYPE_L3;
 967                                 break;
 968
 969                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
 970                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
 971                                 type = PKT_HASH_TYPE_L4;
 972                                 break;
 973
 974                         default:
 975                                 break;
 976                         }
 977
 978                         if (type != PKT_HASH_TYPE_NONE)
 979                                 skb_set_hash(skb,
 980                                              *(u32 *)extra->u.hash.value,
 981                                              type);
 982                 }
 983
 984                 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
 985
 986                 __skb_put(skb, data_len);
 987                 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
 988                 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
 989                 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
 990
 991                 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
 992                         virt_to_gfn(skb->data);
 993                 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
 994                 queue->tx_copy_ops[*copy_ops].dest.offset =
 995                         offset_in_page(skb->data) & ~XEN_PAGE_MASK;
 996
 997                 queue->tx_copy_ops[*copy_ops].len = data_len;
 998                 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
 999
1000                 (*copy_ops)++;
1001
1002                 if (data_len < txreq.size) {
1003                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1004                                              pending_idx);
1005                         xenvif_tx_create_map_op(queue, pending_idx, &txreq,
1006                                                 extra_count, gop);
1007                         gop++;
1008                 } else {
1009                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1010                                              INVALID_PENDING_IDX);
1011                         memcpy(&queue->pending_tx_info[pending_idx].req,
1012                                &txreq, sizeof(txreq));
1013                         queue->pending_tx_info[pending_idx].extra_count =
1014                                 extra_count;
1015                 }
1016
1017                 queue->pending_cons++;
1018
1019                 gop = xenvif_get_requests(queue, skb, txfrags, gop,
1020                                           frag_overflow, nskb);
1021
1022                 __skb_queue_tail(&queue->tx_queue, skb);
1023
1024                 queue->tx.req_cons = idx;
1025
1026                 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1027                     (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1028                         break;
1029         }
1030
1031         (*map_ops) = gop - queue->tx_map_ops;
1032         return;
1033 }
1034
1035 /* Consolidate skb with a frag_list into a brand new one with local pages on
1036  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1037  */
1038 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1039 {
1040         unsigned int offset = skb_headlen(skb);
1041         skb_frag_t frags[MAX_SKB_FRAGS];
1042         int i, f;
1043         struct ubuf_info *uarg;
1044         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1045
1046         queue->stats.tx_zerocopy_sent += 2;
1047         queue->stats.tx_frag_overflow++;
1048
1049         xenvif_fill_frags(queue, nskb);
1050         /* Subtract frags size, we will correct it later */
1051         skb->truesize -= skb->data_len;
1052         skb->len += nskb->len;
1053         skb->data_len += nskb->len;
1054
1055         /* create a brand new frags array and coalesce there */
1056         for (i = 0; offset < skb->len; i++) {
1057                 struct page *page;
1058                 unsigned int len;
1059
1060                 BUG_ON(i >= MAX_SKB_FRAGS);
1061                 page = alloc_page(GFP_ATOMIC);
1062                 if (!page) {
1063                         int j;
1064                         skb->truesize += skb->data_len;
1065                         for (j = 0; j < i; j++)
1066                                 put_page(frags[j].page.p);
1067                         return -ENOMEM;
1068                 }
1069
1070                 if (offset + PAGE_SIZE < skb->len)
1071                         len = PAGE_SIZE;
1072                 else
1073                         len = skb->len - offset;
1074                 if (skb_copy_bits(skb, offset, page_address(page), len))
1075                         BUG();
1076
1077                 offset += len;
1078                 frags[i].page.p = page;
1079                 frags[i].page_offset = 0;
1080                 skb_frag_size_set(&frags[i], len);
1081         }
1082
1083         /* Release all the original (foreign) frags. */
1084         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1085                 skb_frag_unref(skb, f);
1086         uarg = skb_shinfo(skb)->destructor_arg;
1087         /* increase inflight counter to offset decrement in callback */
1088         atomic_inc(&queue->inflight_packets);
1089         uarg->callback(uarg, true);
1090         skb_shinfo(skb)->destructor_arg = NULL;
1091
1092         /* Fill the skb with the new (local) frags. */
1093         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1094         skb_shinfo(skb)->nr_frags = i;
1095         skb->truesize += i * PAGE_SIZE;
1096
1097         return 0;
1098 }
1099
1100 static int xenvif_tx_submit(struct xenvif_queue *queue)
1101 {
1102         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1103         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1104         struct sk_buff *skb;
1105         int work_done = 0;
1106
1107         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1108                 struct xen_netif_tx_request *txp;
1109                 u16 pending_idx;
1110                 unsigned data_len;
1111
1112                 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1113                 txp = &queue->pending_tx_info[pending_idx].req;
1114
1115                 /* Check the remap error code. */
1116                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1117                         /* If there was an error, xenvif_tx_check_gop is
1118                          * expected to release all the frags which were mapped,
1119                          * so kfree_skb shouldn't do it again
1120                          */
1121                         skb_shinfo(skb)->nr_frags = 0;
1122                         if (skb_has_frag_list(skb)) {
1123                                 struct sk_buff *nskb =
1124                                                 skb_shinfo(skb)->frag_list;
1125                                 skb_shinfo(nskb)->nr_frags = 0;
1126                         }
1127                         kfree_skb(skb);
1128                         continue;
1129                 }
1130
1131                 data_len = skb->len;
1132                 callback_param(queue, pending_idx).ctx = NULL;
1133                 if (data_len < txp->size) {
1134                         /* Append the packet payload as a fragment. */
1135                         txp->offset += data_len;
1136                         txp->size -= data_len;
1137                 } else {
1138                         /* Schedule a response immediately. */
1139                         xenvif_idx_release(queue, pending_idx,
1140                                            XEN_NETIF_RSP_OKAY);
1141                 }
1142
1143                 if (txp->flags & XEN_NETTXF_csum_blank)
1144                         skb->ip_summed = CHECKSUM_PARTIAL;
1145                 else if (txp->flags & XEN_NETTXF_data_validated)
1146                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1147
1148                 xenvif_fill_frags(queue, skb);
1149
1150                 if (unlikely(skb_has_frag_list(skb))) {
1151                         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1152                         xenvif_skb_zerocopy_prepare(queue, nskb);
1153                         if (xenvif_handle_frag_list(queue, skb)) {
1154                                 if (net_ratelimit())
1155                                         netdev_err(queue->vif->dev,
1156                                                    "Not enough memory to consolidate frag_list!\n");
1157                                 xenvif_skb_zerocopy_prepare(queue, skb);
1158                                 kfree_skb(skb);
1159                                 continue;
1160                         }
1161                         /* Copied all the bits from the frag list -- free it. */
1162                         skb_frag_list_init(skb);
1163                         kfree_skb(nskb);
1164                 }
1165
1166                 skb->dev      = queue->vif->dev;
1167                 skb->protocol = eth_type_trans(skb, skb->dev);
1168                 skb_reset_network_header(skb);
1169
1170                 if (checksum_setup(queue, skb)) {
1171                         netdev_dbg(queue->vif->dev,
1172                                    "Can't setup checksum in net_tx_action\n");
1173                         /* We have to set this flag to trigger the callback */
1174                         if (skb_shinfo(skb)->destructor_arg)
1175                                 xenvif_skb_zerocopy_prepare(queue, skb);
1176                         kfree_skb(skb);
1177                         continue;
1178                 }
1179
1180                 skb_probe_transport_header(skb, 0);
1181
1182                 /* If the packet is GSO then we will have just set up the
1183                  * transport header offset in checksum_setup so it's now
1184                  * straightforward to calculate gso_segs.
1185                  */
1186                 if (skb_is_gso(skb)) {
1187                         int mss = skb_shinfo(skb)->gso_size;
1188                         int hdrlen = skb_transport_header(skb) -
1189                                 skb_mac_header(skb) +
1190                                 tcp_hdrlen(skb);
1191
1192                         skb_shinfo(skb)->gso_segs =
1193                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1194                 }
1195
1196                 queue->stats.rx_bytes += skb->len;
1197                 queue->stats.rx_packets++;
1198
1199                 work_done++;
1200
1201                 /* Set this flag right before netif_receive_skb, otherwise
1202                  * someone might think this packet already left netback, and
1203                  * do a skb_copy_ubufs while we are still in control of the
1204                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1205                  */
1206                 if (skb_shinfo(skb)->destructor_arg) {
1207                         xenvif_skb_zerocopy_prepare(queue, skb);
1208                         queue->stats.tx_zerocopy_sent++;
1209                 }
1210
1211                 netif_receive_skb(skb);
1212         }
1213
1214         return work_done;
1215 }
1216
1217 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1218 {
1219         unsigned long flags;
1220         pending_ring_idx_t index;
1221         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1222
1223         /* This is the only place where we grab this lock, to protect callbacks
1224          * from each other.
1225          */
1226         spin_lock_irqsave(&queue->callback_lock, flags);
1227         do {
1228                 u16 pending_idx = ubuf->desc;
1229                 ubuf = (struct ubuf_info *) ubuf->ctx;
1230                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1231                         MAX_PENDING_REQS);
1232                 index = pending_index(queue->dealloc_prod);
1233                 queue->dealloc_ring[index] = pending_idx;
1234                 /* Sync with xenvif_tx_dealloc_action:
1235                  * insert idx then incr producer.
1236                  */
1237                 smp_wmb();
1238                 queue->dealloc_prod++;
1239         } while (ubuf);
1240         spin_unlock_irqrestore(&queue->callback_lock, flags);
1241
1242         if (likely(zerocopy_success))
1243                 queue->stats.tx_zerocopy_success++;
1244         else
1245                 queue->stats.tx_zerocopy_fail++;
1246         xenvif_skb_zerocopy_complete(queue);
1247 }
1248
1249 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1250 {
1251         struct gnttab_unmap_grant_ref *gop;
1252         pending_ring_idx_t dc, dp;
1253         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1254         unsigned int i = 0;
1255
1256         dc = queue->dealloc_cons;
1257         gop = queue->tx_unmap_ops;
1258
1259         /* Free up any grants we have finished using */
1260         do {
1261                 dp = queue->dealloc_prod;
1262
1263                 /* Ensure we see all indices enqueued by all
1264                  * xenvif_zerocopy_callback().
1265                  */
1266                 smp_rmb();
1267
1268                 while (dc != dp) {
1269                         BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1270                         pending_idx =
1271                                 queue->dealloc_ring[pending_index(dc++)];
1272
1273                         pending_idx_release[gop - queue->tx_unmap_ops] =
1274                                 pending_idx;
1275                         queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1276                                 queue->mmap_pages[pending_idx];
1277                         gnttab_set_unmap_op(gop,
1278                                             idx_to_kaddr(queue, pending_idx),
1279                                             GNTMAP_host_map,
1280                                             queue->grant_tx_handle[pending_idx]);
1281                         xenvif_grant_handle_reset(queue, pending_idx);
1282                         ++gop;
1283                 }
1284
1285         } while (dp != queue->dealloc_prod);
1286
1287         queue->dealloc_cons = dc;
1288
1289         if (gop - queue->tx_unmap_ops > 0) {
1290                 int ret;
1291                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1292                                         NULL,
1293                                         queue->pages_to_unmap,
1294                                         gop - queue->tx_unmap_ops);
1295                 if (ret) {
1296                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1297                                    gop - queue->tx_unmap_ops, ret);
1298                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1299                                 if (gop[i].status != GNTST_okay)
1300                                         netdev_err(queue->vif->dev,
1301                                                    " host_addr: 0x%llx handle: 0x%x status: %d\n",
1302                                                    gop[i].host_addr,
1303                                                    gop[i].handle,
1304                                                    gop[i].status);
1305                         }
1306                         BUG();
1307                 }
1308         }
1309
1310         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1311                 xenvif_idx_release(queue, pending_idx_release[i],
1312                                    XEN_NETIF_RSP_OKAY);
1313 }
1314
1315
1316 /* Called after netfront has transmitted */
1317 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1318 {
1319         unsigned nr_mops, nr_cops = 0;
1320         int work_done, ret;
1321
1322         if (unlikely(!tx_work_todo(queue)))
1323                 return 0;
1324
1325         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1326
1327         if (nr_cops == 0)
1328                 return 0;
1329
1330         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1331         if (nr_mops != 0) {
1332                 ret = gnttab_map_refs(queue->tx_map_ops,
1333                                       NULL,
1334                                       queue->pages_to_map,
1335                                       nr_mops);
1336                 if (ret) {
1337                         unsigned int i;
1338
1339                         netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1340                                    nr_mops, ret);
1341                         for (i = 0; i < nr_mops; ++i)
1342                                 WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1343                                              GNTST_okay);
1344                 }
1345         }
1346
1347         work_done = xenvif_tx_submit(queue);
1348
1349         return work_done;
1350 }
1351
1352 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1353                                u8 status)
1354 {
1355         struct pending_tx_info *pending_tx_info;
1356         pending_ring_idx_t index;
1357         unsigned long flags;
1358
1359         pending_tx_info = &queue->pending_tx_info[pending_idx];
1360
1361         spin_lock_irqsave(&queue->response_lock, flags);
1362
1363         make_tx_response(queue, &pending_tx_info->req,
1364                          pending_tx_info->extra_count, status);
1365
1366         /* Release the pending index before pusing the Tx response so
1367          * its available before a new Tx request is pushed by the
1368          * frontend.
1369          */
1370         index = pending_index(queue->pending_prod++);
1371         queue->pending_ring[index] = pending_idx;
1372
1373         push_tx_responses(queue);
1374
1375         spin_unlock_irqrestore(&queue->response_lock, flags);
1376 }
1377
1378
1379 static void make_tx_response(struct xenvif_queue *queue,
1380                              struct xen_netif_tx_request *txp,
1381                              unsigned int extra_count,
1382                              s8       st)
1383 {
1384         RING_IDX i = queue->tx.rsp_prod_pvt;
1385         struct xen_netif_tx_response *resp;
1386
1387         resp = RING_GET_RESPONSE(&queue->tx, i);
1388         resp->id     = txp->id;
1389         resp->status = st;
1390
1391         while (extra_count-- != 0)
1392                 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1393
1394         queue->tx.rsp_prod_pvt = ++i;
1395 }
1396
1397 static void push_tx_responses(struct xenvif_queue *queue)
1398 {
1399         int notify;
1400
1401         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1402         if (notify)
1403                 notify_remote_via_irq(queue->tx_irq);
1404 }
1405
1406 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1407 {
1408         int ret;
1409         struct gnttab_unmap_grant_ref tx_unmap_op;
1410
1411         gnttab_set_unmap_op(&tx_unmap_op,
1412                             idx_to_kaddr(queue, pending_idx),
1413                             GNTMAP_host_map,
1414                             queue->grant_tx_handle[pending_idx]);
1415         xenvif_grant_handle_reset(queue, pending_idx);
1416
1417         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1418                                 &queue->mmap_pages[pending_idx], 1);
1419         if (ret) {
1420                 netdev_err(queue->vif->dev,
1421                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1422                            ret,
1423                            pending_idx,
1424                            tx_unmap_op.host_addr,
1425                            tx_unmap_op.handle,
1426                            tx_unmap_op.status);
1427                 BUG();
1428         }
1429 }
1430
1431 static inline int tx_work_todo(struct xenvif_queue *queue)
1432 {
1433         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1434                 return 1;
1435
1436         return 0;
1437 }
1438
1439 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1440 {
1441         return queue->dealloc_cons != queue->dealloc_prod;
1442 }
1443
1444 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1445 {
1446         if (queue->tx.sring)
1447                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1448                                         queue->tx.sring);
1449         if (queue->rx.sring)
1450                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1451                                         queue->rx.sring);
1452 }
1453
1454 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1455                                    grant_ref_t tx_ring_ref,
1456                                    grant_ref_t rx_ring_ref)
1457 {
1458         void *addr;
1459         struct xen_netif_tx_sring *txs;
1460         struct xen_netif_rx_sring *rxs;
1461
1462         int err = -ENOMEM;
1463
1464         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1465                                      &tx_ring_ref, 1, &addr);
1466         if (err)
1467                 goto err;
1468
1469         txs = (struct xen_netif_tx_sring *)addr;
1470         BACK_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE);
1471
1472         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1473                                      &rx_ring_ref, 1, &addr);
1474         if (err)
1475                 goto err;
1476
1477         rxs = (struct xen_netif_rx_sring *)addr;
1478         BACK_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE);
1479
1480         return 0;
1481
1482 err:
1483         xenvif_unmap_frontend_data_rings(queue);
1484         return err;
1485 }
1486
1487 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1488 {
1489         /* Dealloc thread must remain running until all inflight
1490          * packets complete.
1491          */
1492         return kthread_should_stop() &&
1493                 !atomic_read(&queue->inflight_packets);
1494 }
1495
1496 int xenvif_dealloc_kthread(void *data)
1497 {
1498         struct xenvif_queue *queue = data;
1499
1500         for (;;) {
1501                 wait_event_interruptible(queue->dealloc_wq,
1502                                          tx_dealloc_work_todo(queue) ||
1503                                          xenvif_dealloc_kthread_should_stop(queue));
1504                 if (xenvif_dealloc_kthread_should_stop(queue))
1505                         break;
1506
1507                 xenvif_tx_dealloc_action(queue);
1508                 cond_resched();
1509         }
1510
1511         /* Unmap anything remaining*/
1512         if (tx_dealloc_work_todo(queue))
1513                 xenvif_tx_dealloc_action(queue);
1514
1515         return 0;
1516 }
1517
1518 static void make_ctrl_response(struct xenvif *vif,
1519                                const struct xen_netif_ctrl_request *req,
1520                                u32 status, u32 data)
1521 {
1522         RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1523         struct xen_netif_ctrl_response rsp = {
1524                 .id = req->id,
1525                 .type = req->type,
1526                 .status = status,
1527                 .data = data,
1528         };
1529
1530         *RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1531         vif->ctrl.rsp_prod_pvt = ++idx;
1532 }
1533
1534 static void push_ctrl_response(struct xenvif *vif)
1535 {
1536         int notify;
1537
1538         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1539         if (notify)
1540                 notify_remote_via_irq(vif->ctrl_irq);
1541 }
1542
1543 static void process_ctrl_request(struct xenvif *vif,
1544                                  const struct xen_netif_ctrl_request *req)
1545 {
1546         u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1547         u32 data = 0;
1548
1549         switch (req->type) {
1550         case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1551                 status = xenvif_set_hash_alg(vif, req->data[0]);
1552                 break;
1553
1554         case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1555                 status = xenvif_get_hash_flags(vif, &data);
1556                 break;
1557
1558         case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1559                 status = xenvif_set_hash_flags(vif, req->data[0]);
1560                 break;
1561
1562         case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1563                 status = xenvif_set_hash_key(vif, req->data[0],
1564                                              req->data[1]);
1565                 break;
1566
1567         case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1568                 status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1569                 data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1570                 break;
1571
1572         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1573                 status = xenvif_set_hash_mapping_size(vif,
1574                                                       req->data[0]);
1575                 break;
1576
1577         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1578                 status = xenvif_set_hash_mapping(vif, req->data[0],
1579                                                  req->data[1],
1580                                                  req->data[2]);
1581                 break;
1582
1583         default:
1584                 break;
1585         }
1586
1587         make_ctrl_response(vif, req, status, data);
1588         push_ctrl_response(vif);
1589 }
1590
1591 static void xenvif_ctrl_action(struct xenvif *vif)
1592 {
1593         for (;;) {
1594                 RING_IDX req_prod, req_cons;
1595
1596                 req_prod = vif->ctrl.sring->req_prod;
1597                 req_cons = vif->ctrl.req_cons;
1598
1599                 /* Make sure we can see requests before we process them. */
1600                 rmb();
1601
1602                 if (req_cons == req_prod)
1603                         break;
1604
1605                 while (req_cons != req_prod) {
1606                         struct xen_netif_ctrl_request req;
1607
1608                         RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1609                         req_cons++;
1610
1611                         process_ctrl_request(vif, &req);
1612                 }
1613
1614                 vif->ctrl.req_cons = req_cons;
1615                 vif->ctrl.sring->req_event = req_cons + 1;
1616         }
1617 }
1618
1619 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1620 {
1621         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1622                 return 1;
1623
1624         return 0;
1625 }
1626
1627 irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1628 {
1629         struct xenvif *vif = data;
1630         unsigned int eoi_flag = XEN_EOI_FLAG_SPURIOUS;
1631
1632         while (xenvif_ctrl_work_todo(vif)) {
1633                 xenvif_ctrl_action(vif);
1634                 eoi_flag = 0;
1635         }
1636
1637         xen_irq_lateeoi(irq, eoi_flag);
1638
1639         return IRQ_HANDLED;
1640 }
1641
1642 static int __init netback_init(void)
1643 {
1644         int rc = 0;
1645
1646         if (!xen_domain())
1647                 return -ENODEV;
1648
1649         /* Allow as many queues as there are CPUs but max. 8 if user has not
1650          * specified a value.
1651          */
1652         if (xenvif_max_queues == 0)
1653                 xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1654                                           num_online_cpus());
1655
1656         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1657                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1658                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1659                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1660         }
1661
1662         rc = xenvif_xenbus_init();
1663         if (rc)
1664                 goto failed_init;
1665
1666 #ifdef CONFIG_DEBUG_FS
1667         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1668         if (IS_ERR_OR_NULL(xen_netback_dbg_root))
1669                 pr_warn("Init of debugfs returned %ld!\n",
1670                         PTR_ERR(xen_netback_dbg_root));
1671 #endif /* CONFIG_DEBUG_FS */
1672
1673         return 0;
1674
1675 failed_init:
1676         return rc;
1677 }
1678
1679 module_init(netback_init);
1680
1681 static void __exit netback_fini(void)
1682 {
1683 #ifdef CONFIG_DEBUG_FS
1684         if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
1685                 debugfs_remove_recursive(xen_netback_dbg_root);
1686 #endif /* CONFIG_DEBUG_FS */
1687         xenvif_xenbus_fini();
1688 }
1689 module_exit(netback_fini);
1690
1691 MODULE_LICENSE("Dual BSD/GPL");
1692 MODULE_ALIAS("xen-backend:vif");