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 /* The module parameter tells that we have to put data
 100  * for xen-netfront with the XDP_PACKET_HEADROOM offset
 101  * needed for XDP processing
 102  */
 103 bool provides_xdp_headroom = true;
 104 module_param(provides_xdp_headroom, bool, 0644);
 105
 106 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
 107                                u8 status);
 108
 109 static void make_tx_response(struct xenvif_queue *queue,
 110                              struct xen_netif_tx_request *txp,
 111                              unsigned int extra_count,
 112                              s8       st);
 113 static void push_tx_responses(struct xenvif_queue *queue);
 114
 115 static inline int tx_work_todo(struct xenvif_queue *queue);
 116
 117 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
 118                                        u16 idx)
 119 {
 120         return page_to_pfn(queue->mmap_pages[idx]);
 121 }
 122
 123 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
 124                                          u16 idx)
 125 {
 126         return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
 127 }
 128
 129 #define callback_param(vif, pending_idx) \
 130         (vif->pending_tx_info[pending_idx].callback_struct)
 131
 132 /* Find the containing VIF's structure from a pointer in pending_tx_info array
 133  */
 134 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
 135 {
 136         u16 pending_idx = ubuf->desc;
 137         struct pending_tx_info *temp =
 138                 container_of(ubuf, struct pending_tx_info, callback_struct);
 139         return container_of(temp - pending_idx,
 140                             struct xenvif_queue,
 141                             pending_tx_info[0]);
 142 }
 143
 144 static u16 frag_get_pending_idx(skb_frag_t *frag)
 145 {
 146         return (u16)skb_frag_off(frag);
 147 }
 148
 149 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 150 {
 151         skb_frag_off_set(frag, pending_idx);
 152 }
 153
 154 static inline pending_ring_idx_t pending_index(unsigned i)
 155 {
 156         return i & (MAX_PENDING_REQS-1);
 157 }
 158
 159 void xenvif_kick_thread(struct xenvif_queue *queue)
 160 {
 161         wake_up(&queue->wq);
 162 }
 163
 164 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 165 {
 166         int more_to_do;
 167
 168         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 169
 170         if (more_to_do)
 171                 napi_schedule(&queue->napi);
 172         else if (atomic_fetch_andnot(NETBK_TX_EOI | NETBK_COMMON_EOI,
 173                                      &queue->eoi_pending) &
 174                  (NETBK_TX_EOI | NETBK_COMMON_EOI))
 175                 xen_irq_lateeoi(queue->tx_irq, 0);
 176 }
 177
 178 static void tx_add_credit(struct xenvif_queue *queue)
 179 {
 180         unsigned long max_burst, max_credit;
 181
 182         /*
 183          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 184          * Otherwise the interface can seize up due to insufficient credit.
 185          */
 186         max_burst = max(131072UL, queue->credit_bytes);
 187
 188         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 189         max_credit = queue->remaining_credit + queue->credit_bytes;
 190         if (max_credit < queue->remaining_credit)
 191                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 192
 193         queue->remaining_credit = min(max_credit, max_burst);
 194         queue->rate_limited = false;
 195 }
 196
 197 void xenvif_tx_credit_callback(struct timer_list *t)
 198 {
 199         struct xenvif_queue *queue = from_timer(queue, t, credit_timeout);
 200         tx_add_credit(queue);
 201         xenvif_napi_schedule_or_enable_events(queue);
 202 }
 203
 204 static void xenvif_tx_err(struct xenvif_queue *queue,
 205                           struct xen_netif_tx_request *txp,
 206                           unsigned int extra_count, RING_IDX end)
 207 {
 208         RING_IDX cons = queue->tx.req_cons;
 209         unsigned long flags;
 210
 211         do {
 212                 spin_lock_irqsave(&queue->response_lock, flags);
 213                 make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
 214                 push_tx_responses(queue);
 215                 spin_unlock_irqrestore(&queue->response_lock, flags);
 216                 if (cons == end)
 217                         break;
 218                 RING_COPY_REQUEST(&queue->tx, cons++, txp);
 219                 extra_count = 0; /* only the first frag can have extras */
 220         } while (1);
 221         queue->tx.req_cons = cons;
 222 }
 223
 224 static void xenvif_fatal_tx_err(struct xenvif *vif)
 225 {
 226         netdev_err(vif->dev, "fatal error; disabling device\n");
 227         vif->disabled = true;
 228         /* Disable the vif from queue 0's kthread */
 229         if (vif->num_queues)
 230                 xenvif_kick_thread(&vif->queues[0]);
 231 }
 232
 233 static int xenvif_count_requests(struct xenvif_queue *queue,
 234                                  struct xen_netif_tx_request *first,
 235                                  unsigned int extra_count,
 236                                  struct xen_netif_tx_request *txp,
 237                                  int work_to_do)
 238 {
 239         RING_IDX cons = queue->tx.req_cons;
 240         int slots = 0;
 241         int drop_err = 0;
 242         int more_data;
 243
 244         if (!(first->flags & XEN_NETTXF_more_data))
 245                 return 0;
 246
 247         do {
 248                 struct xen_netif_tx_request dropped_tx = { 0 };
 249
 250                 if (slots >= work_to_do) {
 251                         netdev_err(queue->vif->dev,
 252                                    "Asked for %d slots but exceeds this limit\n",
 253                                    work_to_do);
 254                         xenvif_fatal_tx_err(queue->vif);
 255                         return -ENODATA;
 256                 }
 257
 258                 /* This guest is really using too many slots and
 259                  * considered malicious.
 260                  */
 261                 if (unlikely(slots >= fatal_skb_slots)) {
 262                         netdev_err(queue->vif->dev,
 263                                    "Malicious frontend using %d slots, threshold %u\n",
 264                                    slots, fatal_skb_slots);
 265                         xenvif_fatal_tx_err(queue->vif);
 266                         return -E2BIG;
 267                 }
 268
 269                 /* Xen network protocol had implicit dependency on
 270                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 271                  * the historical MAX_SKB_FRAGS value 18 to honor the
 272                  * same behavior as before. Any packet using more than
 273                  * 18 slots but less than fatal_skb_slots slots is
 274                  * dropped
 275                  */
 276                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 277                         if (net_ratelimit())
 278                                 netdev_dbg(queue->vif->dev,
 279                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 280                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 281                         drop_err = -E2BIG;
 282                 }
 283
 284                 if (drop_err)
 285                         txp = &dropped_tx;
 286
 287                 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
 288
 289                 /* If the guest submitted a frame >= 64 KiB then
 290                  * first->size overflowed and following slots will
 291                  * appear to be larger than the frame.
 292                  *
 293                  * This cannot be fatal error as there are buggy
 294                  * frontends that do this.
 295                  *
 296                  * Consume all slots and drop the packet.
 297                  */
 298                 if (!drop_err && txp->size > first->size) {
 299                         if (net_ratelimit())
 300                                 netdev_dbg(queue->vif->dev,
 301                                            "Invalid tx request, slot size %u > remaining size %u\n",
 302                                            txp->size, first->size);
 303                         drop_err = -EIO;
 304                 }
 305
 306                 first->size -= txp->size;
 307                 slots++;
 308
 309                 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
 310                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 311                                  txp->offset, txp->size);
 312                         xenvif_fatal_tx_err(queue->vif);
 313                         return -EINVAL;
 314                 }
 315
 316                 more_data = txp->flags & XEN_NETTXF_more_data;
 317
 318                 if (!drop_err)
 319                         txp++;
 320
 321         } while (more_data);
 322
 323         if (drop_err) {
 324                 xenvif_tx_err(queue, first, extra_count, cons + slots);
 325                 return drop_err;
 326         }
 327
 328         return slots;
 329 }
 330
 331
 332 struct xenvif_tx_cb {
 333         u16 pending_idx;
 334 };
 335
 336 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 337
 338 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 339                                            u16 pending_idx,
 340                                            struct xen_netif_tx_request *txp,
 341                                            unsigned int extra_count,
 342                                            struct gnttab_map_grant_ref *mop)
 343 {
 344         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 345         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 346                           GNTMAP_host_map | GNTMAP_readonly,
 347                           txp->gref, queue->vif->domid);
 348
 349         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 350                sizeof(*txp));
 351         queue->pending_tx_info[pending_idx].extra_count = extra_count;
 352 }
 353
 354 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 355 {
 356         struct sk_buff *skb =
 357                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 358                           GFP_ATOMIC | __GFP_NOWARN);
 359         if (unlikely(skb == NULL))
 360                 return NULL;
 361
 362         /* Packets passed to netif_rx() must have some headroom. */
 363         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 364
 365         /* Initialize it here to avoid later surprises */
 366         skb_shinfo(skb)->destructor_arg = NULL;
 367
 368         return skb;
 369 }
 370
 371 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
 372                                                         struct sk_buff *skb,
 373                                                         struct xen_netif_tx_request *txp,
 374                                                         struct gnttab_map_grant_ref *gop,
 375                                                         unsigned int frag_overflow,
 376                                                         struct sk_buff *nskb)
 377 {
 378         struct skb_shared_info *shinfo = skb_shinfo(skb);
 379         skb_frag_t *frags = shinfo->frags;
 380         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 381         int start;
 382         pending_ring_idx_t index;
 383         unsigned int nr_slots;
 384
 385         nr_slots = shinfo->nr_frags;
 386
 387         /* Skip first skb fragment if it is on same page as header fragment. */
 388         start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
 389
 390         for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
 391              shinfo->nr_frags++, txp++, gop++) {
 392                 index = pending_index(queue->pending_cons++);
 393                 pending_idx = queue->pending_ring[index];
 394                 xenvif_tx_create_map_op(queue, pending_idx, txp, 0, gop);
 395                 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 396         }
 397
 398         if (frag_overflow) {
 399
 400                 shinfo = skb_shinfo(nskb);
 401                 frags = shinfo->frags;
 402
 403                 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
 404                      shinfo->nr_frags++, txp++, gop++) {
 405                         index = pending_index(queue->pending_cons++);
 406                         pending_idx = queue->pending_ring[index];
 407                         xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
 408                                                 gop);
 409                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 410                                              pending_idx);
 411                 }
 412
 413                 skb_shinfo(skb)->frag_list = nskb;
 414         }
 415
 416         return gop;
 417 }
 418
 419 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 420                                            u16 pending_idx,
 421                                            grant_handle_t handle)
 422 {
 423         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 424                      NETBACK_INVALID_HANDLE)) {
 425                 netdev_err(queue->vif->dev,
 426                            "Trying to overwrite active handle! pending_idx: 0x%x\n",
 427                            pending_idx);
 428                 BUG();
 429         }
 430         queue->grant_tx_handle[pending_idx] = handle;
 431 }
 432
 433 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 434                                              u16 pending_idx)
 435 {
 436         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 437                      NETBACK_INVALID_HANDLE)) {
 438                 netdev_err(queue->vif->dev,
 439                            "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 440                            pending_idx);
 441                 BUG();
 442         }
 443         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 444 }
 445
 446 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 447                                struct sk_buff *skb,
 448                                struct gnttab_map_grant_ref **gopp_map,
 449                                struct gnttab_copy **gopp_copy)
 450 {
 451         struct gnttab_map_grant_ref *gop_map = *gopp_map;
 452         u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
 453         /* This always points to the shinfo of the skb being checked, which
 454          * could be either the first or the one on the frag_list
 455          */
 456         struct skb_shared_info *shinfo = skb_shinfo(skb);
 457         /* If this is non-NULL, we are currently checking the frag_list skb, and
 458          * this points to the shinfo of the first one
 459          */
 460         struct skb_shared_info *first_shinfo = NULL;
 461         int nr_frags = shinfo->nr_frags;
 462         const bool sharedslot = nr_frags &&
 463                                 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
 464         int i, err;
 465
 466         /* Check status of header. */
 467         err = (*gopp_copy)->status;
 468         if (unlikely(err)) {
 469                 if (net_ratelimit())
 470                         netdev_dbg(queue->vif->dev,
 471                                    "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 472                                    (*gopp_copy)->status,
 473                                    pending_idx,
 474                                    (*gopp_copy)->source.u.ref);
 475                 /* The first frag might still have this slot mapped */
 476                 if (!sharedslot)
 477                         xenvif_idx_release(queue, pending_idx,
 478                                            XEN_NETIF_RSP_ERROR);
 479         }
 480         (*gopp_copy)++;
 481
 482 check_frags:
 483         for (i = 0; i < nr_frags; i++, gop_map++) {
 484                 int j, newerr;
 485
 486                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 487
 488                 /* Check error status: if okay then remember grant handle. */
 489                 newerr = gop_map->status;
 490
 491                 if (likely(!newerr)) {
 492                         xenvif_grant_handle_set(queue,
 493                                                 pending_idx,
 494                                                 gop_map->handle);
 495                         /* Had a previous error? Invalidate this fragment. */
 496                         if (unlikely(err)) {
 497                                 xenvif_idx_unmap(queue, pending_idx);
 498                                 /* If the mapping of the first frag was OK, but
 499                                  * the header's copy failed, and they are
 500                                  * sharing a slot, send an error
 501                                  */
 502                                 if (i == 0 && sharedslot)
 503                                         xenvif_idx_release(queue, pending_idx,
 504                                                            XEN_NETIF_RSP_ERROR);
 505                                 else
 506                                         xenvif_idx_release(queue, pending_idx,
 507                                                            XEN_NETIF_RSP_OKAY);
 508                         }
 509                         continue;
 510                 }
 511
 512                 /* Error on this fragment: respond to client with an error. */
 513                 if (net_ratelimit())
 514                         netdev_dbg(queue->vif->dev,
 515                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
 516                                    i,
 517                                    gop_map->status,
 518                                    pending_idx,
 519                                    gop_map->ref);
 520
 521                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
 522
 523                 /* Not the first error? Preceding frags already invalidated. */
 524                 if (err)
 525                         continue;
 526
 527                 /* First error: if the header haven't shared a slot with the
 528                  * first frag, release it as well.
 529                  */
 530                 if (!sharedslot)
 531                         xenvif_idx_release(queue,
 532                                            XENVIF_TX_CB(skb)->pending_idx,
 533                                            XEN_NETIF_RSP_OKAY);
 534
 535                 /* Invalidate preceding fragments of this skb. */
 536                 for (j = 0; j < i; j++) {
 537                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
 538                         xenvif_idx_unmap(queue, pending_idx);
 539                         xenvif_idx_release(queue, pending_idx,
 540                                            XEN_NETIF_RSP_OKAY);
 541                 }
 542
 543                 /* And if we found the error while checking the frag_list, unmap
 544                  * the first skb's frags
 545                  */
 546                 if (first_shinfo) {
 547                         for (j = 0; j < first_shinfo->nr_frags; j++) {
 548                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
 549                                 xenvif_idx_unmap(queue, pending_idx);
 550                                 xenvif_idx_release(queue, pending_idx,
 551                                                    XEN_NETIF_RSP_OKAY);
 552                         }
 553                 }
 554
 555                 /* Remember the error: invalidate all subsequent fragments. */
 556                 err = newerr;
 557         }
 558
 559         if (skb_has_frag_list(skb) && !first_shinfo) {
 560                 first_shinfo = shinfo;
 561                 shinfo = skb_shinfo(shinfo->frag_list);
 562                 nr_frags = shinfo->nr_frags;
 563
 564                 goto check_frags;
 565         }
 566
 567         *gopp_map = gop_map;
 568         return err;
 569 }
 570
 571 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
 572 {
 573         struct skb_shared_info *shinfo = skb_shinfo(skb);
 574         int nr_frags = shinfo->nr_frags;
 575         int i;
 576         u16 prev_pending_idx = INVALID_PENDING_IDX;
 577
 578         for (i = 0; i < nr_frags; i++) {
 579                 skb_frag_t *frag = shinfo->frags + i;
 580                 struct xen_netif_tx_request *txp;
 581                 struct page *page;
 582                 u16 pending_idx;
 583
 584                 pending_idx = frag_get_pending_idx(frag);
 585
 586                 /* If this is not the first frag, chain it to the previous*/
 587                 if (prev_pending_idx == INVALID_PENDING_IDX)
 588                         skb_shinfo(skb)->destructor_arg =
 589                                 &callback_param(queue, pending_idx);
 590                 else
 591                         callback_param(queue, prev_pending_idx).ctx =
 592                                 &callback_param(queue, pending_idx);
 593
 594                 callback_param(queue, pending_idx).ctx = NULL;
 595                 prev_pending_idx = pending_idx;
 596
 597                 txp = &queue->pending_tx_info[pending_idx].req;
 598                 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
 599                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
 600                 skb->len += txp->size;
 601                 skb->data_len += txp->size;
 602                 skb->truesize += txp->size;
 603
 604                 /* Take an extra reference to offset network stack's put_page */
 605                 get_page(queue->mmap_pages[pending_idx]);
 606         }
 607 }
 608
 609 static int xenvif_get_extras(struct xenvif_queue *queue,
 610                              struct xen_netif_extra_info *extras,
 611                              unsigned int *extra_count,
 612                              int work_to_do)
 613 {
 614         struct xen_netif_extra_info extra;
 615         RING_IDX cons = queue->tx.req_cons;
 616
 617         do {
 618                 if (unlikely(work_to_do-- <= 0)) {
 619                         netdev_err(queue->vif->dev, "Missing extra info\n");
 620                         xenvif_fatal_tx_err(queue->vif);
 621                         return -EBADR;
 622                 }
 623
 624                 RING_COPY_REQUEST(&queue->tx, cons, &extra);
 625
 626                 queue->tx.req_cons = ++cons;
 627                 (*extra_count)++;
 628
 629                 if (unlikely(!extra.type ||
 630                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
 631                         netdev_err(queue->vif->dev,
 632                                    "Invalid extra type: %d\n", extra.type);
 633                         xenvif_fatal_tx_err(queue->vif);
 634                         return -EINVAL;
 635                 }
 636
 637                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
 638         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
 639
 640         return work_to_do;
 641 }
 642
 643 static int xenvif_set_skb_gso(struct xenvif *vif,
 644                               struct sk_buff *skb,
 645                               struct xen_netif_extra_info *gso)
 646 {
 647         if (!gso->u.gso.size) {
 648                 netdev_err(vif->dev, "GSO size must not be zero.\n");
 649                 xenvif_fatal_tx_err(vif);
 650                 return -EINVAL;
 651         }
 652
 653         switch (gso->u.gso.type) {
 654         case XEN_NETIF_GSO_TYPE_TCPV4:
 655                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 656                 break;
 657         case XEN_NETIF_GSO_TYPE_TCPV6:
 658                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 659                 break;
 660         default:
 661                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
 662                 xenvif_fatal_tx_err(vif);
 663                 return -EINVAL;
 664         }
 665
 666         skb_shinfo(skb)->gso_size = gso->u.gso.size;
 667         /* gso_segs will be calculated later */
 668
 669         return 0;
 670 }
 671
 672 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
 673 {
 674         bool recalculate_partial_csum = false;
 675
 676         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
 677          * peers can fail to set NETRXF_csum_blank when sending a GSO
 678          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
 679          * recalculate the partial checksum.
 680          */
 681         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
 682                 queue->stats.rx_gso_checksum_fixup++;
 683                 skb->ip_summed = CHECKSUM_PARTIAL;
 684                 recalculate_partial_csum = true;
 685         }
 686
 687         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
 688         if (skb->ip_summed != CHECKSUM_PARTIAL)
 689                 return 0;
 690
 691         return skb_checksum_setup(skb, recalculate_partial_csum);
 692 }
 693
 694 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
 695 {
 696         u64 now = get_jiffies_64();
 697         u64 next_credit = queue->credit_window_start +
 698                 msecs_to_jiffies(queue->credit_usec / 1000);
 699
 700         /* Timer could already be pending in rare cases. */
 701         if (timer_pending(&queue->credit_timeout)) {
 702                 queue->rate_limited = true;
 703                 return true;
 704         }
 705
 706         /* Passed the point where we can replenish credit? */
 707         if (time_after_eq64(now, next_credit)) {
 708                 queue->credit_window_start = now;
 709                 tx_add_credit(queue);
 710         }
 711
 712         /* Still too big to send right now? Set a callback. */
 713         if (size > queue->remaining_credit) {
 714                 mod_timer(&queue->credit_timeout,
 715                           next_credit);
 716                 queue->credit_window_start = next_credit;
 717                 queue->rate_limited = true;
 718
 719                 return true;
 720         }
 721
 722         return false;
 723 }
 724
 725 /* No locking is required in xenvif_mcast_add/del() as they are
 726  * only ever invoked from NAPI poll. An RCU list is used because
 727  * xenvif_mcast_match() is called asynchronously, during start_xmit.
 728  */
 729
 730 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
 731 {
 732         struct xenvif_mcast_addr *mcast;
 733
 734         if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
 735                 if (net_ratelimit())
 736                         netdev_err(vif->dev,
 737                                    "Too many multicast addresses\n");
 738                 return -ENOSPC;
 739         }
 740
 741         mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
 742         if (!mcast)
 743                 return -ENOMEM;
 744
 745         ether_addr_copy(mcast->addr, addr);
 746         list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
 747         vif->fe_mcast_count++;
 748
 749         return 0;
 750 }
 751
 752 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
 753 {
 754         struct xenvif_mcast_addr *mcast;
 755
 756         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 757                 if (ether_addr_equal(addr, mcast->addr)) {
 758                         --vif->fe_mcast_count;
 759                         list_del_rcu(&mcast->entry);
 760                         kfree_rcu(mcast, rcu);
 761                         break;
 762                 }
 763         }
 764 }
 765
 766 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
 767 {
 768         struct xenvif_mcast_addr *mcast;
 769
 770         rcu_read_lock();
 771         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 772                 if (ether_addr_equal(addr, mcast->addr)) {
 773                         rcu_read_unlock();
 774                         return true;
 775                 }
 776         }
 777         rcu_read_unlock();
 778
 779         return false;
 780 }
 781
 782 void xenvif_mcast_addr_list_free(struct xenvif *vif)
 783 {
 784         /* No need for locking or RCU here. NAPI poll and TX queue
 785          * are stopped.
 786          */
 787         while (!list_empty(&vif->fe_mcast_addr)) {
 788                 struct xenvif_mcast_addr *mcast;
 789
 790                 mcast = list_first_entry(&vif->fe_mcast_addr,
 791                                          struct xenvif_mcast_addr,
 792                                          entry);
 793                 --vif->fe_mcast_count;
 794                 list_del(&mcast->entry);
 795                 kfree(mcast);
 796         }
 797 }
 798
 799 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
 800                                      int budget,
 801                                      unsigned *copy_ops,
 802                                      unsigned *map_ops)
 803 {
 804         struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
 805         struct sk_buff *skb, *nskb;
 806         int ret;
 807         unsigned int frag_overflow;
 808
 809         while (skb_queue_len(&queue->tx_queue) < budget) {
 810                 struct xen_netif_tx_request txreq;
 811                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
 812                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
 813                 unsigned int extra_count;
 814                 u16 pending_idx;
 815                 RING_IDX idx;
 816                 int work_to_do;
 817                 unsigned int data_len;
 818                 pending_ring_idx_t index;
 819
 820                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
 821                     XEN_NETIF_TX_RING_SIZE) {
 822                         netdev_err(queue->vif->dev,
 823                                    "Impossible number of requests. "
 824                                    "req_prod %d, req_cons %d, size %ld\n",
 825                                    queue->tx.sring->req_prod, queue->tx.req_cons,
 826                                    XEN_NETIF_TX_RING_SIZE);
 827                         xenvif_fatal_tx_err(queue->vif);
 828                         break;
 829                 }
 830
 831                 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
 832                 if (!work_to_do)
 833                         break;
 834
 835                 idx = queue->tx.req_cons;
 836                 rmb(); /* Ensure that we see the request before we copy it. */
 837                 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
 838
 839                 /* Credit-based scheduling. */
 840                 if (txreq.size > queue->remaining_credit &&
 841                     tx_credit_exceeded(queue, txreq.size))
 842                         break;
 843
 844                 queue->remaining_credit -= txreq.size;
 845
 846                 work_to_do--;
 847                 queue->tx.req_cons = ++idx;
 848
 849                 memset(extras, 0, sizeof(extras));
 850                 extra_count = 0;
 851                 if (txreq.flags & XEN_NETTXF_extra_info) {
 852                         work_to_do = xenvif_get_extras(queue, extras,
 853                                                        &extra_count,
 854                                                        work_to_do);
 855                         idx = queue->tx.req_cons;
 856                         if (unlikely(work_to_do < 0))
 857                                 break;
 858                 }
 859
 860                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
 861                         struct xen_netif_extra_info *extra;
 862
 863                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
 864                         ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
 865
 866                         make_tx_response(queue, &txreq, extra_count,
 867                                          (ret == 0) ?
 868                                          XEN_NETIF_RSP_OKAY :
 869                                          XEN_NETIF_RSP_ERROR);
 870                         push_tx_responses(queue);
 871                         continue;
 872                 }
 873
 874                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
 875                         struct xen_netif_extra_info *extra;
 876
 877                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
 878                         xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
 879
 880                         make_tx_response(queue, &txreq, extra_count,
 881                                          XEN_NETIF_RSP_OKAY);
 882                         push_tx_responses(queue);
 883                         continue;
 884                 }
 885
 886                 ret = xenvif_count_requests(queue, &txreq, extra_count,
 887                                             txfrags, work_to_do);
 888                 if (unlikely(ret < 0))
 889                         break;
 890
 891                 idx += ret;
 892
 893                 if (unlikely(txreq.size < ETH_HLEN)) {
 894                         netdev_dbg(queue->vif->dev,
 895                                    "Bad packet size: %d\n", txreq.size);
 896                         xenvif_tx_err(queue, &txreq, extra_count, idx);
 897                         break;
 898                 }
 899
 900                 /* No crossing a page as the payload mustn't fragment. */
 901                 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
 902                         netdev_err(queue->vif->dev,
 903                                    "txreq.offset: %u, size: %u, end: %lu\n",
 904                                    txreq.offset, txreq.size,
 905                                    (unsigned long)(txreq.offset&~XEN_PAGE_MASK) + txreq.size);
 906                         xenvif_fatal_tx_err(queue->vif);
 907                         break;
 908                 }
 909
 910                 index = pending_index(queue->pending_cons);
 911                 pending_idx = queue->pending_ring[index];
 912
 913                 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
 914                             ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
 915                         XEN_NETBACK_TX_COPY_LEN : txreq.size;
 916
 917                 skb = xenvif_alloc_skb(data_len);
 918                 if (unlikely(skb == NULL)) {
 919                         netdev_dbg(queue->vif->dev,
 920                                    "Can't allocate a skb in start_xmit.\n");
 921                         xenvif_tx_err(queue, &txreq, extra_count, idx);
 922                         break;
 923                 }
 924
 925                 skb_shinfo(skb)->nr_frags = ret;
 926                 if (data_len < txreq.size)
 927                         skb_shinfo(skb)->nr_frags++;
 928                 /* At this point shinfo->nr_frags is in fact the number of
 929                  * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
 930                  */
 931                 frag_overflow = 0;
 932                 nskb = NULL;
 933                 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
 934                         frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
 935                         BUG_ON(frag_overflow > MAX_SKB_FRAGS);
 936                         skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
 937                         nskb = xenvif_alloc_skb(0);
 938                         if (unlikely(nskb == NULL)) {
 939                                 skb_shinfo(skb)->nr_frags = 0;
 940                                 kfree_skb(skb);
 941                                 xenvif_tx_err(queue, &txreq, extra_count, idx);
 942                                 if (net_ratelimit())
 943                                         netdev_err(queue->vif->dev,
 944                                                    "Can't allocate the frag_list skb.\n");
 945                                 break;
 946                         }
 947                 }
 948
 949                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
 950                         struct xen_netif_extra_info *gso;
 951                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
 952
 953                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
 954                                 /* Failure in xenvif_set_skb_gso is fatal. */
 955                                 skb_shinfo(skb)->nr_frags = 0;
 956                                 kfree_skb(skb);
 957                                 kfree_skb(nskb);
 958                                 break;
 959                         }
 960                 }
 961
 962                 if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
 963                         struct xen_netif_extra_info *extra;
 964                         enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
 965
 966                         extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
 967
 968                         switch (extra->u.hash.type) {
 969                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
 970                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
 971                                 type = PKT_HASH_TYPE_L3;
 972                                 break;
 973
 974                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
 975                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
 976                                 type = PKT_HASH_TYPE_L4;
 977                                 break;
 978
 979                         default:
 980                                 break;
 981                         }
 982
 983                         if (type != PKT_HASH_TYPE_NONE)
 984                                 skb_set_hash(skb,
 985                                              *(u32 *)extra->u.hash.value,
 986                                              type);
 987                 }
 988
 989                 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
 990
 991                 __skb_put(skb, data_len);
 992                 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
 993                 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
 994                 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
 995
 996                 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
 997                         virt_to_gfn(skb->data);
 998                 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
 999                 queue->tx_copy_ops[*copy_ops].dest.offset =
1000                         offset_in_page(skb->data) & ~XEN_PAGE_MASK;
1001
1002                 queue->tx_copy_ops[*copy_ops].len = data_len;
1003                 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1004
1005                 (*copy_ops)++;
1006
1007                 if (data_len < txreq.size) {
1008                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1009                                              pending_idx);
1010                         xenvif_tx_create_map_op(queue, pending_idx, &txreq,
1011                                                 extra_count, gop);
1012                         gop++;
1013                 } else {
1014                         frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1015                                              INVALID_PENDING_IDX);
1016                         memcpy(&queue->pending_tx_info[pending_idx].req,
1017                                &txreq, sizeof(txreq));
1018                         queue->pending_tx_info[pending_idx].extra_count =
1019                                 extra_count;
1020                 }
1021
1022                 queue->pending_cons++;
1023
1024                 gop = xenvif_get_requests(queue, skb, txfrags, gop,
1025                                           frag_overflow, nskb);
1026
1027                 __skb_queue_tail(&queue->tx_queue, skb);
1028
1029                 queue->tx.req_cons = idx;
1030
1031                 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1032                     (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1033                         break;
1034         }
1035
1036         (*map_ops) = gop - queue->tx_map_ops;
1037         return;
1038 }
1039
1040 /* Consolidate skb with a frag_list into a brand new one with local pages on
1041  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1042  */
1043 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1044 {
1045         unsigned int offset = skb_headlen(skb);
1046         skb_frag_t frags[MAX_SKB_FRAGS];
1047         int i, f;
1048         struct ubuf_info *uarg;
1049         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1050
1051         queue->stats.tx_zerocopy_sent += 2;
1052         queue->stats.tx_frag_overflow++;
1053
1054         xenvif_fill_frags(queue, nskb);
1055         /* Subtract frags size, we will correct it later */
1056         skb->truesize -= skb->data_len;
1057         skb->len += nskb->len;
1058         skb->data_len += nskb->len;
1059
1060         /* create a brand new frags array and coalesce there */
1061         for (i = 0; offset < skb->len; i++) {
1062                 struct page *page;
1063                 unsigned int len;
1064
1065                 BUG_ON(i >= MAX_SKB_FRAGS);
1066                 page = alloc_page(GFP_ATOMIC);
1067                 if (!page) {
1068                         int j;
1069                         skb->truesize += skb->data_len;
1070                         for (j = 0; j < i; j++)
1071                                 put_page(skb_frag_page(&frags[j]));
1072                         return -ENOMEM;
1073                 }
1074
1075                 if (offset + PAGE_SIZE < skb->len)
1076                         len = PAGE_SIZE;
1077                 else
1078                         len = skb->len - offset;
1079                 if (skb_copy_bits(skb, offset, page_address(page), len))
1080                         BUG();
1081
1082                 offset += len;
1083                 __skb_frag_set_page(&frags[i], page);
1084                 skb_frag_off_set(&frags[i], 0);
1085                 skb_frag_size_set(&frags[i], len);
1086         }
1087
1088         /* Release all the original (foreign) frags. */
1089         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1090                 skb_frag_unref(skb, f);
1091         uarg = skb_shinfo(skb)->destructor_arg;
1092         /* increase inflight counter to offset decrement in callback */
1093         atomic_inc(&queue->inflight_packets);
1094         uarg->callback(NULL, uarg, true);
1095         skb_shinfo(skb)->destructor_arg = NULL;
1096
1097         /* Fill the skb with the new (local) frags. */
1098         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1099         skb_shinfo(skb)->nr_frags = i;
1100         skb->truesize += i * PAGE_SIZE;
1101
1102         return 0;
1103 }
1104
1105 static int xenvif_tx_submit(struct xenvif_queue *queue)
1106 {
1107         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1108         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1109         struct sk_buff *skb;
1110         int work_done = 0;
1111
1112         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1113                 struct xen_netif_tx_request *txp;
1114                 u16 pending_idx;
1115                 unsigned data_len;
1116
1117                 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1118                 txp = &queue->pending_tx_info[pending_idx].req;
1119
1120                 /* Check the remap error code. */
1121                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1122                         /* If there was an error, xenvif_tx_check_gop is
1123                          * expected to release all the frags which were mapped,
1124                          * so kfree_skb shouldn't do it again
1125                          */
1126                         skb_shinfo(skb)->nr_frags = 0;
1127                         if (skb_has_frag_list(skb)) {
1128                                 struct sk_buff *nskb =
1129                                                 skb_shinfo(skb)->frag_list;
1130                                 skb_shinfo(nskb)->nr_frags = 0;
1131                         }
1132                         kfree_skb(skb);
1133                         continue;
1134                 }
1135
1136                 data_len = skb->len;
1137                 callback_param(queue, pending_idx).ctx = NULL;
1138                 if (data_len < txp->size) {
1139                         /* Append the packet payload as a fragment. */
1140                         txp->offset += data_len;
1141                         txp->size -= data_len;
1142                 } else {
1143                         /* Schedule a response immediately. */
1144                         xenvif_idx_release(queue, pending_idx,
1145                                            XEN_NETIF_RSP_OKAY);
1146                 }
1147
1148                 if (txp->flags & XEN_NETTXF_csum_blank)
1149                         skb->ip_summed = CHECKSUM_PARTIAL;
1150                 else if (txp->flags & XEN_NETTXF_data_validated)
1151                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1152
1153                 xenvif_fill_frags(queue, skb);
1154
1155                 if (unlikely(skb_has_frag_list(skb))) {
1156                         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1157                         xenvif_skb_zerocopy_prepare(queue, nskb);
1158                         if (xenvif_handle_frag_list(queue, skb)) {
1159                                 if (net_ratelimit())
1160                                         netdev_err(queue->vif->dev,
1161                                                    "Not enough memory to consolidate frag_list!\n");
1162                                 xenvif_skb_zerocopy_prepare(queue, skb);
1163                                 kfree_skb(skb);
1164                                 continue;
1165                         }
1166                         /* Copied all the bits from the frag list -- free it. */
1167                         skb_frag_list_init(skb);
1168                         kfree_skb(nskb);
1169                 }
1170
1171                 skb->dev      = queue->vif->dev;
1172                 skb->protocol = eth_type_trans(skb, skb->dev);
1173                 skb_reset_network_header(skb);
1174
1175                 if (checksum_setup(queue, skb)) {
1176                         netdev_dbg(queue->vif->dev,
1177                                    "Can't setup checksum in net_tx_action\n");
1178                         /* We have to set this flag to trigger the callback */
1179                         if (skb_shinfo(skb)->destructor_arg)
1180                                 xenvif_skb_zerocopy_prepare(queue, skb);
1181                         kfree_skb(skb);
1182                         continue;
1183                 }
1184
1185                 skb_probe_transport_header(skb);
1186
1187                 /* If the packet is GSO then we will have just set up the
1188                  * transport header offset in checksum_setup so it's now
1189                  * straightforward to calculate gso_segs.
1190                  */
1191                 if (skb_is_gso(skb)) {
1192                         int mss, hdrlen;
1193
1194                         /* GSO implies having the L4 header. */
1195                         WARN_ON_ONCE(!skb_transport_header_was_set(skb));
1196                         if (unlikely(!skb_transport_header_was_set(skb))) {
1197                                 kfree_skb(skb);
1198                                 continue;
1199                         }
1200
1201                         mss = skb_shinfo(skb)->gso_size;
1202                         hdrlen = skb_transport_header(skb) -
1203                                 skb_mac_header(skb) +
1204                                 tcp_hdrlen(skb);
1205
1206                         skb_shinfo(skb)->gso_segs =
1207                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1208                 }
1209
1210                 queue->stats.rx_bytes += skb->len;
1211                 queue->stats.rx_packets++;
1212
1213                 work_done++;
1214
1215                 /* Set this flag right before netif_receive_skb, otherwise
1216                  * someone might think this packet already left netback, and
1217                  * do a skb_copy_ubufs while we are still in control of the
1218                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1219                  */
1220                 if (skb_shinfo(skb)->destructor_arg) {
1221                         xenvif_skb_zerocopy_prepare(queue, skb);
1222                         queue->stats.tx_zerocopy_sent++;
1223                 }
1224
1225                 netif_receive_skb(skb);
1226         }
1227
1228         return work_done;
1229 }
1230
1231 void xenvif_zerocopy_callback(struct sk_buff *skb, struct ubuf_info *ubuf,
1232                               bool zerocopy_success)
1233 {
1234         unsigned long flags;
1235         pending_ring_idx_t index;
1236         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1237
1238         /* This is the only place where we grab this lock, to protect callbacks
1239          * from each other.
1240          */
1241         spin_lock_irqsave(&queue->callback_lock, flags);
1242         do {
1243                 u16 pending_idx = ubuf->desc;
1244                 ubuf = (struct ubuf_info *) ubuf->ctx;
1245                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1246                         MAX_PENDING_REQS);
1247                 index = pending_index(queue->dealloc_prod);
1248                 queue->dealloc_ring[index] = pending_idx;
1249                 /* Sync with xenvif_tx_dealloc_action:
1250                  * insert idx then incr producer.
1251                  */
1252                 smp_wmb();
1253                 queue->dealloc_prod++;
1254         } while (ubuf);
1255         spin_unlock_irqrestore(&queue->callback_lock, flags);
1256
1257         if (likely(zerocopy_success))
1258                 queue->stats.tx_zerocopy_success++;
1259         else
1260                 queue->stats.tx_zerocopy_fail++;
1261         xenvif_skb_zerocopy_complete(queue);
1262 }
1263
1264 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1265 {
1266         struct gnttab_unmap_grant_ref *gop;
1267         pending_ring_idx_t dc, dp;
1268         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1269         unsigned int i = 0;
1270
1271         dc = queue->dealloc_cons;
1272         gop = queue->tx_unmap_ops;
1273
1274         /* Free up any grants we have finished using */
1275         do {
1276                 dp = queue->dealloc_prod;
1277
1278                 /* Ensure we see all indices enqueued by all
1279                  * xenvif_zerocopy_callback().
1280                  */
1281                 smp_rmb();
1282
1283                 while (dc != dp) {
1284                         BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1285                         pending_idx =
1286                                 queue->dealloc_ring[pending_index(dc++)];
1287
1288                         pending_idx_release[gop - queue->tx_unmap_ops] =
1289                                 pending_idx;
1290                         queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1291                                 queue->mmap_pages[pending_idx];
1292                         gnttab_set_unmap_op(gop,
1293                                             idx_to_kaddr(queue, pending_idx),
1294                                             GNTMAP_host_map,
1295                                             queue->grant_tx_handle[pending_idx]);
1296                         xenvif_grant_handle_reset(queue, pending_idx);
1297                         ++gop;
1298                 }
1299
1300         } while (dp != queue->dealloc_prod);
1301
1302         queue->dealloc_cons = dc;
1303
1304         if (gop - queue->tx_unmap_ops > 0) {
1305                 int ret;
1306                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1307                                         NULL,
1308                                         queue->pages_to_unmap,
1309                                         gop - queue->tx_unmap_ops);
1310                 if (ret) {
1311                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1312                                    gop - queue->tx_unmap_ops, ret);
1313                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1314                                 if (gop[i].status != GNTST_okay)
1315                                         netdev_err(queue->vif->dev,
1316                                                    " host_addr: 0x%llx handle: 0x%x status: %d\n",
1317                                                    gop[i].host_addr,
1318                                                    gop[i].handle,
1319                                                    gop[i].status);
1320                         }
1321                         BUG();
1322                 }
1323         }
1324
1325         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1326                 xenvif_idx_release(queue, pending_idx_release[i],
1327                                    XEN_NETIF_RSP_OKAY);
1328 }
1329
1330
1331 /* Called after netfront has transmitted */
1332 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1333 {
1334         unsigned nr_mops, nr_cops = 0;
1335         int work_done, ret;
1336
1337         if (unlikely(!tx_work_todo(queue)))
1338                 return 0;
1339
1340         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1341
1342         if (nr_cops == 0)
1343                 return 0;
1344
1345         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1346         if (nr_mops != 0) {
1347                 ret = gnttab_map_refs(queue->tx_map_ops,
1348                                       NULL,
1349                                       queue->pages_to_map,
1350                                       nr_mops);
1351                 if (ret) {
1352                         unsigned int i;
1353
1354                         netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1355                                    nr_mops, ret);
1356                         for (i = 0; i < nr_mops; ++i)
1357                                 WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1358                                              GNTST_okay);
1359                 }
1360         }
1361
1362         work_done = xenvif_tx_submit(queue);
1363
1364         return work_done;
1365 }
1366
1367 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1368                                u8 status)
1369 {
1370         struct pending_tx_info *pending_tx_info;
1371         pending_ring_idx_t index;
1372         unsigned long flags;
1373
1374         pending_tx_info = &queue->pending_tx_info[pending_idx];
1375
1376         spin_lock_irqsave(&queue->response_lock, flags);
1377
1378         make_tx_response(queue, &pending_tx_info->req,
1379                          pending_tx_info->extra_count, status);
1380
1381         /* Release the pending index before pusing the Tx response so
1382          * its available before a new Tx request is pushed by the
1383          * frontend.
1384          */
1385         index = pending_index(queue->pending_prod++);
1386         queue->pending_ring[index] = pending_idx;
1387
1388         push_tx_responses(queue);
1389
1390         spin_unlock_irqrestore(&queue->response_lock, flags);
1391 }
1392
1393
1394 static void make_tx_response(struct xenvif_queue *queue,
1395                              struct xen_netif_tx_request *txp,
1396                              unsigned int extra_count,
1397                              s8       st)
1398 {
1399         RING_IDX i = queue->tx.rsp_prod_pvt;
1400         struct xen_netif_tx_response *resp;
1401
1402         resp = RING_GET_RESPONSE(&queue->tx, i);
1403         resp->id     = txp->id;
1404         resp->status = st;
1405
1406         while (extra_count-- != 0)
1407                 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1408
1409         queue->tx.rsp_prod_pvt = ++i;
1410 }
1411
1412 static void push_tx_responses(struct xenvif_queue *queue)
1413 {
1414         int notify;
1415
1416         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1417         if (notify)
1418                 notify_remote_via_irq(queue->tx_irq);
1419 }
1420
1421 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1422 {
1423         int ret;
1424         struct gnttab_unmap_grant_ref tx_unmap_op;
1425
1426         gnttab_set_unmap_op(&tx_unmap_op,
1427                             idx_to_kaddr(queue, pending_idx),
1428                             GNTMAP_host_map,
1429                             queue->grant_tx_handle[pending_idx]);
1430         xenvif_grant_handle_reset(queue, pending_idx);
1431
1432         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1433                                 &queue->mmap_pages[pending_idx], 1);
1434         if (ret) {
1435                 netdev_err(queue->vif->dev,
1436                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1437                            ret,
1438                            pending_idx,
1439                            tx_unmap_op.host_addr,
1440                            tx_unmap_op.handle,
1441                            tx_unmap_op.status);
1442                 BUG();
1443         }
1444 }
1445
1446 static inline int tx_work_todo(struct xenvif_queue *queue)
1447 {
1448         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1449                 return 1;
1450
1451         return 0;
1452 }
1453
1454 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1455 {
1456         return queue->dealloc_cons != queue->dealloc_prod;
1457 }
1458
1459 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1460 {
1461         if (queue->tx.sring)
1462                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1463                                         queue->tx.sring);
1464         if (queue->rx.sring)
1465                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1466                                         queue->rx.sring);
1467 }
1468
1469 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1470                                    grant_ref_t tx_ring_ref,
1471                                    grant_ref_t rx_ring_ref)
1472 {
1473         void *addr;
1474         struct xen_netif_tx_sring *txs;
1475         struct xen_netif_rx_sring *rxs;
1476         RING_IDX rsp_prod, req_prod;
1477         int err = -ENOMEM;
1478
1479         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1480                                      &tx_ring_ref, 1, &addr);
1481         if (err)
1482                 goto err;
1483
1484         txs = (struct xen_netif_tx_sring *)addr;
1485         rsp_prod = READ_ONCE(txs->rsp_prod);
1486         req_prod = READ_ONCE(txs->req_prod);
1487
1488         BACK_RING_ATTACH(&queue->tx, txs, rsp_prod, XEN_PAGE_SIZE);
1489
1490         err = -EIO;
1491         if (req_prod - rsp_prod > RING_SIZE(&queue->tx))
1492                 goto err;
1493
1494         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1495                                      &rx_ring_ref, 1, &addr);
1496         if (err)
1497                 goto err;
1498
1499         rxs = (struct xen_netif_rx_sring *)addr;
1500         rsp_prod = READ_ONCE(rxs->rsp_prod);
1501         req_prod = READ_ONCE(rxs->req_prod);
1502
1503         BACK_RING_ATTACH(&queue->rx, rxs, rsp_prod, XEN_PAGE_SIZE);
1504
1505         err = -EIO;
1506         if (req_prod - rsp_prod > RING_SIZE(&queue->rx))
1507                 goto err;
1508
1509         return 0;
1510
1511 err:
1512         xenvif_unmap_frontend_data_rings(queue);
1513         return err;
1514 }
1515
1516 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1517 {
1518         /* Dealloc thread must remain running until all inflight
1519          * packets complete.
1520          */
1521         return kthread_should_stop() &&
1522                 !atomic_read(&queue->inflight_packets);
1523 }
1524
1525 int xenvif_dealloc_kthread(void *data)
1526 {
1527         struct xenvif_queue *queue = data;
1528
1529         for (;;) {
1530                 wait_event_interruptible(queue->dealloc_wq,
1531                                          tx_dealloc_work_todo(queue) ||
1532                                          xenvif_dealloc_kthread_should_stop(queue));
1533                 if (xenvif_dealloc_kthread_should_stop(queue))
1534                         break;
1535
1536                 xenvif_tx_dealloc_action(queue);
1537                 cond_resched();
1538         }
1539
1540         /* Unmap anything remaining*/
1541         if (tx_dealloc_work_todo(queue))
1542                 xenvif_tx_dealloc_action(queue);
1543
1544         return 0;
1545 }
1546
1547 static void make_ctrl_response(struct xenvif *vif,
1548                                const struct xen_netif_ctrl_request *req,
1549                                u32 status, u32 data)
1550 {
1551         RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1552         struct xen_netif_ctrl_response rsp = {
1553                 .id = req->id,
1554                 .type = req->type,
1555                 .status = status,
1556                 .data = data,
1557         };
1558
1559         *RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1560         vif->ctrl.rsp_prod_pvt = ++idx;
1561 }
1562
1563 static void push_ctrl_response(struct xenvif *vif)
1564 {
1565         int notify;
1566
1567         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1568         if (notify)
1569                 notify_remote_via_irq(vif->ctrl_irq);
1570 }
1571
1572 static void process_ctrl_request(struct xenvif *vif,
1573                                  const struct xen_netif_ctrl_request *req)
1574 {
1575         u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1576         u32 data = 0;
1577
1578         switch (req->type) {
1579         case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1580                 status = xenvif_set_hash_alg(vif, req->data[0]);
1581                 break;
1582
1583         case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1584                 status = xenvif_get_hash_flags(vif, &data);
1585                 break;
1586
1587         case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1588                 status = xenvif_set_hash_flags(vif, req->data[0]);
1589                 break;
1590
1591         case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1592                 status = xenvif_set_hash_key(vif, req->data[0],
1593                                              req->data[1]);
1594                 break;
1595
1596         case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1597                 status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1598                 data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1599                 break;
1600
1601         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1602                 status = xenvif_set_hash_mapping_size(vif,
1603                                                       req->data[0]);
1604                 break;
1605
1606         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1607                 status = xenvif_set_hash_mapping(vif, req->data[0],
1608                                                  req->data[1],
1609                                                  req->data[2]);
1610                 break;
1611
1612         default:
1613                 break;
1614         }
1615
1616         make_ctrl_response(vif, req, status, data);
1617         push_ctrl_response(vif);
1618 }
1619
1620 static void xenvif_ctrl_action(struct xenvif *vif)
1621 {
1622         for (;;) {
1623                 RING_IDX req_prod, req_cons;
1624
1625                 req_prod = vif->ctrl.sring->req_prod;
1626                 req_cons = vif->ctrl.req_cons;
1627
1628                 /* Make sure we can see requests before we process them. */
1629                 rmb();
1630
1631                 if (req_cons == req_prod)
1632                         break;
1633
1634                 while (req_cons != req_prod) {
1635                         struct xen_netif_ctrl_request req;
1636
1637                         RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1638                         req_cons++;
1639
1640                         process_ctrl_request(vif, &req);
1641                 }
1642
1643                 vif->ctrl.req_cons = req_cons;
1644                 vif->ctrl.sring->req_event = req_cons + 1;
1645         }
1646 }
1647
1648 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1649 {
1650         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1651                 return true;
1652
1653         return false;
1654 }
1655
1656 irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1657 {
1658         struct xenvif *vif = data;
1659         unsigned int eoi_flag = XEN_EOI_FLAG_SPURIOUS;
1660
1661         while (xenvif_ctrl_work_todo(vif)) {
1662                 xenvif_ctrl_action(vif);
1663                 eoi_flag = 0;
1664         }
1665
1666         xen_irq_lateeoi(irq, eoi_flag);
1667
1668         return IRQ_HANDLED;
1669 }
1670
1671 static int __init netback_init(void)
1672 {
1673         int rc = 0;
1674
1675         if (!xen_domain())
1676                 return -ENODEV;
1677
1678         /* Allow as many queues as there are CPUs but max. 8 if user has not
1679          * specified a value.
1680          */
1681         if (xenvif_max_queues == 0)
1682                 xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1683                                           num_online_cpus());
1684
1685         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1686                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1687                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1688                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1689         }
1690
1691         rc = xenvif_xenbus_init();
1692         if (rc)
1693                 goto failed_init;
1694
1695 #ifdef CONFIG_DEBUG_FS
1696         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1697 #endif /* CONFIG_DEBUG_FS */
1698
1699         return 0;
1700
1701 failed_init:
1702         return rc;
1703 }
1704
1705 module_init(netback_init);
1706
1707 static void __exit netback_fini(void)
1708 {
1709 #ifdef CONFIG_DEBUG_FS
1710         debugfs_remove_recursive(xen_netback_dbg_root);
1711 #endif /* CONFIG_DEBUG_FS */
1712         xenvif_xenbus_fini();
1713 }
1714 module_exit(netback_fini);
1715
1716 MODULE_LICENSE("Dual BSD/GPL");
1717 MODULE_ALIAS("xen-backend:vif");