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                                s8 status);
 101
 102 static void make_tx_response(struct xenvif_queue *queue,
 103                              const struct xen_netif_tx_request *txp,
 104                              unsigned int extra_count,
 105                              s8 status);
 106
 107 static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx);
 108
 109 static inline int tx_work_todo(struct xenvif_queue *queue);
 110
 111 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
 112                                        u16 idx)
 113 {
 114         return page_to_pfn(queue->mmap_pages[idx]);
 115 }
 116
 117 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
 118                                          u16 idx)
 119 {
 120         return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
 121 }
 122
 123 #define callback_param(vif, pending_idx) \
 124         (vif->pending_tx_info[pending_idx].callback_struct)
 125
 126 /* Find the containing VIF's structure from a pointer in pending_tx_info array
 127  */
 128 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
 129 {
 130         u16 pending_idx = ubuf->desc;
 131         struct pending_tx_info *temp =
 132                 container_of(ubuf, struct pending_tx_info, callback_struct);
 133         return container_of(temp - pending_idx,
 134                             struct xenvif_queue,
 135                             pending_tx_info[0]);
 136 }
 137
 138 static u16 frag_get_pending_idx(skb_frag_t *frag)
 139 {
 140         return (u16)frag->page_offset;
 141 }
 142
 143 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 144 {
 145         frag->page_offset = pending_idx;
 146 }
 147
 148 static inline pending_ring_idx_t pending_index(unsigned i)
 149 {
 150         return i & (MAX_PENDING_REQS-1);
 151 }
 152
 153 void xenvif_kick_thread(struct xenvif_queue *queue)
 154 {
 155         wake_up(&queue->wq);
 156 }
 157
 158 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 159 {
 160         int more_to_do;
 161
 162         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 163
 164         if (more_to_do)
 165                 napi_schedule(&queue->napi);
 166         else if (atomic_fetch_andnot(NETBK_TX_EOI | NETBK_COMMON_EOI,
 167                                      &queue->eoi_pending) &
 168                  (NETBK_TX_EOI | NETBK_COMMON_EOI))
 169                 xen_irq_lateeoi(queue->tx_irq, 0);
 170 }
 171
 172 static void tx_add_credit(struct xenvif_queue *queue)
 173 {
 174         unsigned long max_burst, max_credit;
 175
 176         /*
 177          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 178          * Otherwise the interface can seize up due to insufficient credit.
 179          */
 180         max_burst = max(131072UL, queue->credit_bytes);
 181
 182         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 183         max_credit = queue->remaining_credit + queue->credit_bytes;
 184         if (max_credit < queue->remaining_credit)
 185                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 186
 187         queue->remaining_credit = min(max_credit, max_burst);
 188         queue->rate_limited = false;
 189 }
 190
 191 void xenvif_tx_credit_callback(struct timer_list *t)
 192 {
 193         struct xenvif_queue *queue = from_timer(queue, t, credit_timeout);
 194         tx_add_credit(queue);
 195         xenvif_napi_schedule_or_enable_events(queue);
 196 }
 197
 198 static void xenvif_tx_err(struct xenvif_queue *queue,
 199                           struct xen_netif_tx_request *txp,
 200                           unsigned int extra_count, RING_IDX end)
 201 {
 202         RING_IDX cons = queue->tx.req_cons;
 203
 204         do {
 205                 make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
 206                 if (cons == end)
 207                         break;
 208                 RING_COPY_REQUEST(&queue->tx, cons++, txp);
 209                 extra_count = 0; /* only the first frag can have extras */
 210         } while (1);
 211         queue->tx.req_cons = cons;
 212 }
 213
 214 static void xenvif_fatal_tx_err(struct xenvif *vif)
 215 {
 216         netdev_err(vif->dev, "fatal error; disabling device\n");
 217         vif->disabled = true;
 218         /* Disable the vif from queue 0's kthread */
 219         if (vif->num_queues)
 220                 xenvif_kick_thread(&vif->queues[0]);
 221 }
 222
 223 static int xenvif_count_requests(struct xenvif_queue *queue,
 224                                  struct xen_netif_tx_request *first,
 225                                  unsigned int extra_count,
 226                                  struct xen_netif_tx_request *txp,
 227                                  int work_to_do)
 228 {
 229         RING_IDX cons = queue->tx.req_cons;
 230         int slots = 0;
 231         int drop_err = 0;
 232         int more_data;
 233
 234         if (!(first->flags & XEN_NETTXF_more_data))
 235                 return 0;
 236
 237         do {
 238                 struct xen_netif_tx_request dropped_tx = { 0 };
 239
 240                 if (slots >= work_to_do) {
 241                         netdev_err(queue->vif->dev,
 242                                    "Asked for %d slots but exceeds this limit\n",
 243                                    work_to_do);
 244                         xenvif_fatal_tx_err(queue->vif);
 245                         return -ENODATA;
 246                 }
 247
 248                 /* This guest is really using too many slots and
 249                  * considered malicious.
 250                  */
 251                 if (unlikely(slots >= fatal_skb_slots)) {
 252                         netdev_err(queue->vif->dev,
 253                                    "Malicious frontend using %d slots, threshold %u\n",
 254                                    slots, fatal_skb_slots);
 255                         xenvif_fatal_tx_err(queue->vif);
 256                         return -E2BIG;
 257                 }
 258
 259                 /* Xen network protocol had implicit dependency on
 260                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 261                  * the historical MAX_SKB_FRAGS value 18 to honor the
 262                  * same behavior as before. Any packet using more than
 263                  * 18 slots but less than fatal_skb_slots slots is
 264                  * dropped
 265                  */
 266                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 267                         if (net_ratelimit())
 268                                 netdev_dbg(queue->vif->dev,
 269                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 270                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 271                         drop_err = -E2BIG;
 272                 }
 273
 274                 if (drop_err)
 275                         txp = &dropped_tx;
 276
 277                 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
 278
 279                 /* If the guest submitted a frame >= 64 KiB then
 280                  * first->size overflowed and following slots will
 281                  * appear to be larger than the frame.
 282                  *
 283                  * This cannot be fatal error as there are buggy
 284                  * frontends that do this.
 285                  *
 286                  * Consume all slots and drop the packet.
 287                  */
 288                 if (!drop_err && txp->size > first->size) {
 289                         if (net_ratelimit())
 290                                 netdev_dbg(queue->vif->dev,
 291                                            "Invalid tx request, slot size %u > remaining size %u\n",
 292                                            txp->size, first->size);
 293                         drop_err = -EIO;
 294                 }
 295
 296                 first->size -= txp->size;
 297                 slots++;
 298
 299                 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
 300                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 301                                  txp->offset, txp->size);
 302                         xenvif_fatal_tx_err(queue->vif);
 303                         return -EINVAL;
 304                 }
 305
 306                 more_data = txp->flags & XEN_NETTXF_more_data;
 307
 308                 if (!drop_err)
 309                         txp++;
 310
 311         } while (more_data);
 312
 313         if (drop_err) {
 314                 xenvif_tx_err(queue, first, extra_count, cons + slots);
 315                 return drop_err;
 316         }
 317
 318         return slots;
 319 }
 320
 321
 322 struct xenvif_tx_cb {
 323         u16 copy_pending_idx[XEN_NETBK_LEGACY_SLOTS_MAX + 1];
 324         u8 copy_count;
 325         u32 split_mask;
 326 };
 327
 328 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 329 #define copy_pending_idx(skb, i) (XENVIF_TX_CB(skb)->copy_pending_idx[i])
 330 #define copy_count(skb) (XENVIF_TX_CB(skb)->copy_count)
 331
 332 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 333                                            u16 pending_idx,
 334                                            struct xen_netif_tx_request *txp,
 335                                            unsigned int extra_count,
 336                                            struct gnttab_map_grant_ref *mop)
 337 {
 338         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 339         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 340                           GNTMAP_host_map | GNTMAP_readonly,
 341                           txp->gref, queue->vif->domid);
 342
 343         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 344                sizeof(*txp));
 345         queue->pending_tx_info[pending_idx].extra_count = extra_count;
 346 }
 347
 348 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 349 {
 350         struct sk_buff *skb =
 351                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 352                           GFP_ATOMIC | __GFP_NOWARN);
 353
 354         BUILD_BUG_ON(sizeof(*XENVIF_TX_CB(skb)) > sizeof(skb->cb));
 355         if (unlikely(skb == NULL))
 356                 return NULL;
 357
 358         /* Packets passed to netif_rx() must have some headroom. */
 359         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 360
 361         /* Initialize it here to avoid later surprises */
 362         skb_shinfo(skb)->destructor_arg = NULL;
 363
 364         return skb;
 365 }
 366
 367 static void xenvif_get_requests(struct xenvif_queue *queue,
 368                                 struct sk_buff *skb,
 369                                 struct xen_netif_tx_request *first,
 370                                 struct xen_netif_tx_request *txfrags,
 371                                 unsigned *copy_ops,
 372                                 unsigned *map_ops,
 373                                 unsigned int frag_overflow,
 374                                 struct sk_buff *nskb,
 375                                 unsigned int extra_count,
 376                                 unsigned int data_len)
 377 {
 378         struct skb_shared_info *shinfo = skb_shinfo(skb);
 379         skb_frag_t *frags = shinfo->frags;
 380         u16 pending_idx;
 381         pending_ring_idx_t index;
 382         unsigned int nr_slots;
 383         struct gnttab_copy *cop = queue->tx_copy_ops + *copy_ops;
 384         struct gnttab_map_grant_ref *gop = queue->tx_map_ops + *map_ops;
 385         struct xen_netif_tx_request *txp = first;
 386
 387         nr_slots = shinfo->nr_frags + frag_overflow + 1;
 388
 389         copy_count(skb) = 0;
 390         XENVIF_TX_CB(skb)->split_mask = 0;
 391
 392         /* Create copy ops for exactly data_len bytes into the skb head. */
 393         __skb_put(skb, data_len);
 394         while (data_len > 0) {
 395                 int amount = data_len > txp->size ? txp->size : data_len;
 396                 bool split = false;
 397
 398                 cop->source.u.ref = txp->gref;
 399                 cop->source.domid = queue->vif->domid;
 400                 cop->source.offset = txp->offset;
 401
 402                 cop->dest.domid = DOMID_SELF;
 403                 cop->dest.offset = (offset_in_page(skb->data +
 404                                                    skb_headlen(skb) -
 405                                                    data_len)) & ~XEN_PAGE_MASK;
 406                 cop->dest.u.gmfn = virt_to_gfn(skb->data + skb_headlen(skb)
 407                                                - data_len);
 408
 409                 /* Don't cross local page boundary! */
 410                 if (cop->dest.offset + amount > XEN_PAGE_SIZE) {
 411                         amount = XEN_PAGE_SIZE - cop->dest.offset;
 412                         XENVIF_TX_CB(skb)->split_mask |= 1U << copy_count(skb);
 413                         split = true;
 414                 }
 415
 416                 cop->len = amount;
 417                 cop->flags = GNTCOPY_source_gref;
 418
 419                 index = pending_index(queue->pending_cons);
 420                 pending_idx = queue->pending_ring[index];
 421                 callback_param(queue, pending_idx).ctx = NULL;
 422                 copy_pending_idx(skb, copy_count(skb)) = pending_idx;
 423                 if (!split)
 424                         copy_count(skb)++;
 425
 426                 cop++;
 427                 data_len -= amount;
 428
 429                 if (amount == txp->size) {
 430                         /* The copy op covered the full tx_request */
 431
 432                         memcpy(&queue->pending_tx_info[pending_idx].req,
 433                                txp, sizeof(*txp));
 434                         queue->pending_tx_info[pending_idx].extra_count =
 435                                 (txp == first) ? extra_count : 0;
 436
 437                         if (txp == first)
 438                                 txp = txfrags;
 439                         else
 440                                 txp++;
 441                         queue->pending_cons++;
 442                         nr_slots--;
 443                 } else {
 444                         /* The copy op partially covered the tx_request.
 445                          * The remainder will be mapped or copied in the next
 446                          * iteration.
 447                          */
 448                         txp->offset += amount;
 449                         txp->size -= amount;
 450                 }
 451         }
 452
 453         for (shinfo->nr_frags = 0; nr_slots > 0 && shinfo->nr_frags < MAX_SKB_FRAGS;
 454              nr_slots--) {
 455                 if (unlikely(!txp->size)) {
 456                         make_tx_response(queue, txp, 0, XEN_NETIF_RSP_OKAY);
 457                         ++txp;
 458                         continue;
 459                 }
 460
 461                 index = pending_index(queue->pending_cons++);
 462                 pending_idx = queue->pending_ring[index];
 463                 xenvif_tx_create_map_op(queue, pending_idx, txp,
 464                                         txp == first ? extra_count : 0, gop);
 465                 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 466                 ++shinfo->nr_frags;
 467                 ++gop;
 468
 469                 if (txp == first)
 470                         txp = txfrags;
 471                 else
 472                         txp++;
 473         }
 474
 475         if (nr_slots > 0) {
 476
 477                 shinfo = skb_shinfo(nskb);
 478                 frags = shinfo->frags;
 479
 480                 for (shinfo->nr_frags = 0; shinfo->nr_frags < nr_slots; ++txp) {
 481                         if (unlikely(!txp->size)) {
 482                                 make_tx_response(queue, txp, 0,
 483                                                  XEN_NETIF_RSP_OKAY);
 484                                 continue;
 485                         }
 486
 487                         index = pending_index(queue->pending_cons++);
 488                         pending_idx = queue->pending_ring[index];
 489                         xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
 490                                                 gop);
 491                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 492                                              pending_idx);
 493                         ++shinfo->nr_frags;
 494                         ++gop;
 495                 }
 496
 497                 if (shinfo->nr_frags) {
 498                         skb_shinfo(skb)->frag_list = nskb;
 499                         nskb = NULL;
 500                 }
 501         }
 502
 503         if (nskb) {
 504                 /* A frag_list skb was allocated but it is no longer needed
 505                  * because enough slots were converted to copy ops above or some
 506                  * were empty.
 507                  */
 508                 kfree_skb(nskb);
 509         }
 510
 511         (*copy_ops) = cop - queue->tx_copy_ops;
 512         (*map_ops) = gop - queue->tx_map_ops;
 513 }
 514
 515 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 516                                            u16 pending_idx,
 517                                            grant_handle_t handle)
 518 {
 519         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 520                      NETBACK_INVALID_HANDLE)) {
 521                 netdev_err(queue->vif->dev,
 522                            "Trying to overwrite active handle! pending_idx: 0x%x\n",
 523                            pending_idx);
 524                 BUG();
 525         }
 526         queue->grant_tx_handle[pending_idx] = handle;
 527 }
 528
 529 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 530                                              u16 pending_idx)
 531 {
 532         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 533                      NETBACK_INVALID_HANDLE)) {
 534                 netdev_err(queue->vif->dev,
 535                            "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 536                            pending_idx);
 537                 BUG();
 538         }
 539         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 540 }
 541
 542 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 543                                struct sk_buff *skb,
 544                                struct gnttab_map_grant_ref **gopp_map,
 545                                struct gnttab_copy **gopp_copy)
 546 {
 547         struct gnttab_map_grant_ref *gop_map = *gopp_map;
 548         u16 pending_idx;
 549         /* This always points to the shinfo of the skb being checked, which
 550          * could be either the first or the one on the frag_list
 551          */
 552         struct skb_shared_info *shinfo = skb_shinfo(skb);
 553         /* If this is non-NULL, we are currently checking the frag_list skb, and
 554          * this points to the shinfo of the first one
 555          */
 556         struct skb_shared_info *first_shinfo = NULL;
 557         int nr_frags = shinfo->nr_frags;
 558         const bool sharedslot = nr_frags &&
 559                                 frag_get_pending_idx(&shinfo->frags[0]) ==
 560                                     copy_pending_idx(skb, copy_count(skb) - 1);
 561         int i, err = 0;
 562
 563         for (i = 0; i < copy_count(skb); i++) {
 564                 int newerr;
 565
 566                 /* Check status of header. */
 567                 pending_idx = copy_pending_idx(skb, i);
 568
 569                 newerr = (*gopp_copy)->status;
 570
 571                 /* Split copies need to be handled together. */
 572                 if (XENVIF_TX_CB(skb)->split_mask & (1U << i)) {
 573                         (*gopp_copy)++;
 574                         if (!newerr)
 575                                 newerr = (*gopp_copy)->status;
 576                 }
 577                 if (likely(!newerr)) {
 578                         /* The first frag might still have this slot mapped */
 579                         if (i < copy_count(skb) - 1 || !sharedslot)
 580                                 xenvif_idx_release(queue, pending_idx,
 581                                                    XEN_NETIF_RSP_OKAY);
 582                 } else {
 583                         err = newerr;
 584                         if (net_ratelimit())
 585                                 netdev_dbg(queue->vif->dev,
 586                                            "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 587                                            (*gopp_copy)->status,
 588                                            pending_idx,
 589                                            (*gopp_copy)->source.u.ref);
 590                         /* The first frag might still have this slot mapped */
 591                         if (i < copy_count(skb) - 1 || !sharedslot)
 592                                 xenvif_idx_release(queue, pending_idx,
 593                                                    XEN_NETIF_RSP_ERROR);
 594                 }
 595                 (*gopp_copy)++;
 596         }
 597
 598 check_frags:
 599         for (i = 0; i < nr_frags; i++, gop_map++) {
 600                 int j, newerr;
 601
 602                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 603
 604                 /* Check error status: if okay then remember grant handle. */
 605                 newerr = gop_map->status;
 606
 607                 if (likely(!newerr)) {
 608                         xenvif_grant_handle_set(queue,
 609                                                 pending_idx,
 610                                                 gop_map->handle);
 611                         /* Had a previous error? Invalidate this fragment. */
 612                         if (unlikely(err)) {
 613                                 xenvif_idx_unmap(queue, pending_idx);
 614                                 /* If the mapping of the first frag was OK, but
 615                                  * the header's copy failed, and they are
 616                                  * sharing a slot, send an error
 617                                  */
 618                                 if (i == 0 && !first_shinfo && sharedslot)
 619                                         xenvif_idx_release(queue, pending_idx,
 620                                                            XEN_NETIF_RSP_ERROR);
 621                                 else
 622                                         xenvif_idx_release(queue, pending_idx,
 623                                                            XEN_NETIF_RSP_OKAY);
 624                         }
 625                         continue;
 626                 }
 627
 628                 /* Error on this fragment: respond to client with an error. */
 629                 if (net_ratelimit())
 630                         netdev_dbg(queue->vif->dev,
 631                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
 632                                    i,
 633                                    gop_map->status,
 634                                    pending_idx,
 635                                    gop_map->ref);
 636
 637                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
 638
 639                 /* Not the first error? Preceding frags already invalidated. */
 640                 if (err)
 641                         continue;
 642
 643                 /* Invalidate preceding fragments of this skb. */
 644                 for (j = 0; j < i; j++) {
 645                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
 646                         xenvif_idx_unmap(queue, pending_idx);
 647                         xenvif_idx_release(queue, pending_idx,
 648                                            XEN_NETIF_RSP_OKAY);
 649                 }
 650
 651                 /* And if we found the error while checking the frag_list, unmap
 652                  * the first skb's frags
 653                  */
 654                 if (first_shinfo) {
 655                         for (j = 0; j < first_shinfo->nr_frags; j++) {
 656                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
 657                                 xenvif_idx_unmap(queue, pending_idx);
 658                                 xenvif_idx_release(queue, pending_idx,
 659                                                    XEN_NETIF_RSP_OKAY);
 660                         }
 661                 }
 662
 663                 /* Remember the error: invalidate all subsequent fragments. */
 664                 err = newerr;
 665         }
 666
 667         if (skb_has_frag_list(skb) && !first_shinfo) {
 668                 first_shinfo = skb_shinfo(skb);
 669                 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
 670                 nr_frags = shinfo->nr_frags;
 671
 672                 goto check_frags;
 673         }
 674
 675         *gopp_map = gop_map;
 676         return err;
 677 }
 678
 679 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
 680 {
 681         struct skb_shared_info *shinfo = skb_shinfo(skb);
 682         int nr_frags = shinfo->nr_frags;
 683         int i;
 684         u16 prev_pending_idx = INVALID_PENDING_IDX;
 685
 686         for (i = 0; i < nr_frags; i++) {
 687                 skb_frag_t *frag = shinfo->frags + i;
 688                 struct xen_netif_tx_request *txp;
 689                 struct page *page;
 690                 u16 pending_idx;
 691
 692                 pending_idx = frag_get_pending_idx(frag);
 693
 694                 /* If this is not the first frag, chain it to the previous*/
 695                 if (prev_pending_idx == INVALID_PENDING_IDX)
 696                         skb_shinfo(skb)->destructor_arg =
 697                                 &callback_param(queue, pending_idx);
 698                 else
 699                         callback_param(queue, prev_pending_idx).ctx =
 700                                 &callback_param(queue, pending_idx);
 701
 702                 callback_param(queue, pending_idx).ctx = NULL;
 703                 prev_pending_idx = pending_idx;
 704
 705                 txp = &queue->pending_tx_info[pending_idx].req;
 706                 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
 707                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
 708                 skb->len += txp->size;
 709                 skb->data_len += txp->size;
 710                 skb->truesize += txp->size;
 711
 712                 /* Take an extra reference to offset network stack's put_page */
 713                 get_page(queue->mmap_pages[pending_idx]);
 714         }
 715 }
 716
 717 static int xenvif_get_extras(struct xenvif_queue *queue,
 718                              struct xen_netif_extra_info *extras,
 719                              unsigned int *extra_count,
 720                              int work_to_do)
 721 {
 722         struct xen_netif_extra_info extra;
 723         RING_IDX cons = queue->tx.req_cons;
 724
 725         do {
 726                 if (unlikely(work_to_do-- <= 0)) {
 727                         netdev_err(queue->vif->dev, "Missing extra info\n");
 728                         xenvif_fatal_tx_err(queue->vif);
 729                         return -EBADR;
 730                 }
 731
 732                 RING_COPY_REQUEST(&queue->tx, cons, &extra);
 733
 734                 queue->tx.req_cons = ++cons;
 735                 (*extra_count)++;
 736
 737                 if (unlikely(!extra.type ||
 738                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
 739                         netdev_err(queue->vif->dev,
 740                                    "Invalid extra type: %d\n", extra.type);
 741                         xenvif_fatal_tx_err(queue->vif);
 742                         return -EINVAL;
 743                 }
 744
 745                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
 746         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
 747
 748         return work_to_do;
 749 }
 750
 751 static int xenvif_set_skb_gso(struct xenvif *vif,
 752                               struct sk_buff *skb,
 753                               struct xen_netif_extra_info *gso)
 754 {
 755         if (!gso->u.gso.size) {
 756                 netdev_err(vif->dev, "GSO size must not be zero.\n");
 757                 xenvif_fatal_tx_err(vif);
 758                 return -EINVAL;
 759         }
 760
 761         switch (gso->u.gso.type) {
 762         case XEN_NETIF_GSO_TYPE_TCPV4:
 763                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 764                 break;
 765         case XEN_NETIF_GSO_TYPE_TCPV6:
 766                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 767                 break;
 768         default:
 769                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
 770                 xenvif_fatal_tx_err(vif);
 771                 return -EINVAL;
 772         }
 773
 774         skb_shinfo(skb)->gso_size = gso->u.gso.size;
 775         /* gso_segs will be calculated later */
 776
 777         return 0;
 778 }
 779
 780 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
 781 {
 782         bool recalculate_partial_csum = false;
 783
 784         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
 785          * peers can fail to set NETRXF_csum_blank when sending a GSO
 786          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
 787          * recalculate the partial checksum.
 788          */
 789         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
 790                 queue->stats.rx_gso_checksum_fixup++;
 791                 skb->ip_summed = CHECKSUM_PARTIAL;
 792                 recalculate_partial_csum = true;
 793         }
 794
 795         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
 796         if (skb->ip_summed != CHECKSUM_PARTIAL)
 797                 return 0;
 798
 799         return skb_checksum_setup(skb, recalculate_partial_csum);
 800 }
 801
 802 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
 803 {
 804         u64 now = get_jiffies_64();
 805         u64 next_credit = queue->credit_window_start +
 806                 msecs_to_jiffies(queue->credit_usec / 1000);
 807
 808         /* Timer could already be pending in rare cases. */
 809         if (timer_pending(&queue->credit_timeout)) {
 810                 queue->rate_limited = true;
 811                 return true;
 812         }
 813
 814         /* Passed the point where we can replenish credit? */
 815         if (time_after_eq64(now, next_credit)) {
 816                 queue->credit_window_start = now;
 817                 tx_add_credit(queue);
 818         }
 819
 820         /* Still too big to send right now? Set a callback. */
 821         if (size > queue->remaining_credit) {
 822                 mod_timer(&queue->credit_timeout,
 823                           next_credit);
 824                 queue->credit_window_start = next_credit;
 825                 queue->rate_limited = true;
 826
 827                 return true;
 828         }
 829
 830         return false;
 831 }
 832
 833 /* No locking is required in xenvif_mcast_add/del() as they are
 834  * only ever invoked from NAPI poll. An RCU list is used because
 835  * xenvif_mcast_match() is called asynchronously, during start_xmit.
 836  */
 837
 838 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
 839 {
 840         struct xenvif_mcast_addr *mcast;
 841
 842         if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
 843                 if (net_ratelimit())
 844                         netdev_err(vif->dev,
 845                                    "Too many multicast addresses\n");
 846                 return -ENOSPC;
 847         }
 848
 849         mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
 850         if (!mcast)
 851                 return -ENOMEM;
 852
 853         ether_addr_copy(mcast->addr, addr);
 854         list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
 855         vif->fe_mcast_count++;
 856
 857         return 0;
 858 }
 859
 860 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
 861 {
 862         struct xenvif_mcast_addr *mcast;
 863
 864         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 865                 if (ether_addr_equal(addr, mcast->addr)) {
 866                         --vif->fe_mcast_count;
 867                         list_del_rcu(&mcast->entry);
 868                         kfree_rcu(mcast, rcu);
 869                         break;
 870                 }
 871         }
 872 }
 873
 874 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
 875 {
 876         struct xenvif_mcast_addr *mcast;
 877
 878         rcu_read_lock();
 879         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 880                 if (ether_addr_equal(addr, mcast->addr)) {
 881                         rcu_read_unlock();
 882                         return true;
 883                 }
 884         }
 885         rcu_read_unlock();
 886
 887         return false;
 888 }
 889
 890 void xenvif_mcast_addr_list_free(struct xenvif *vif)
 891 {
 892         /* No need for locking or RCU here. NAPI poll and TX queue
 893          * are stopped.
 894          */
 895         while (!list_empty(&vif->fe_mcast_addr)) {
 896                 struct xenvif_mcast_addr *mcast;
 897
 898                 mcast = list_first_entry(&vif->fe_mcast_addr,
 899                                          struct xenvif_mcast_addr,
 900                                          entry);
 901                 --vif->fe_mcast_count;
 902                 list_del(&mcast->entry);
 903                 kfree(mcast);
 904         }
 905 }
 906
 907 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
 908                                      int budget,
 909                                      unsigned *copy_ops,
 910                                      unsigned *map_ops)
 911 {
 912         struct sk_buff *skb, *nskb;
 913         int ret;
 914         unsigned int frag_overflow;
 915
 916         while (skb_queue_len(&queue->tx_queue) < budget) {
 917                 struct xen_netif_tx_request txreq;
 918                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
 919                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
 920                 unsigned int extra_count;
 921                 u16 pending_idx;
 922                 RING_IDX idx;
 923                 int work_to_do;
 924                 unsigned int data_len;
 925                 pending_ring_idx_t index;
 926
 927                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
 928                     XEN_NETIF_TX_RING_SIZE) {
 929                         netdev_err(queue->vif->dev,
 930                                    "Impossible number of requests. "
 931                                    "req_prod %d, req_cons %d, size %ld\n",
 932                                    queue->tx.sring->req_prod, queue->tx.req_cons,
 933                                    XEN_NETIF_TX_RING_SIZE);
 934                         xenvif_fatal_tx_err(queue->vif);
 935                         break;
 936                 }
 937
 938                 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
 939                 if (!work_to_do)
 940                         break;
 941
 942                 idx = queue->tx.req_cons;
 943                 rmb(); /* Ensure that we see the request before we copy it. */
 944                 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
 945
 946                 /* Credit-based scheduling. */
 947                 if (txreq.size > queue->remaining_credit &&
 948                     tx_credit_exceeded(queue, txreq.size))
 949                         break;
 950
 951                 queue->remaining_credit -= txreq.size;
 952
 953                 work_to_do--;
 954                 queue->tx.req_cons = ++idx;
 955
 956                 memset(extras, 0, sizeof(extras));
 957                 extra_count = 0;
 958                 if (txreq.flags & XEN_NETTXF_extra_info) {
 959                         work_to_do = xenvif_get_extras(queue, extras,
 960                                                        &extra_count,
 961                                                        work_to_do);
 962                         idx = queue->tx.req_cons;
 963                         if (unlikely(work_to_do < 0))
 964                                 break;
 965                 }
 966
 967                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
 968                         struct xen_netif_extra_info *extra;
 969
 970                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
 971                         ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
 972
 973                         make_tx_response(queue, &txreq, extra_count,
 974                                          (ret == 0) ?
 975                                          XEN_NETIF_RSP_OKAY :
 976                                          XEN_NETIF_RSP_ERROR);
 977                         continue;
 978                 }
 979
 980                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
 981                         struct xen_netif_extra_info *extra;
 982
 983                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
 984                         xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
 985
 986                         make_tx_response(queue, &txreq, extra_count,
 987                                          XEN_NETIF_RSP_OKAY);
 988                         continue;
 989                 }
 990
 991                 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN) ?
 992                         XEN_NETBACK_TX_COPY_LEN : txreq.size;
 993
 994                 ret = xenvif_count_requests(queue, &txreq, extra_count,
 995                                             txfrags, work_to_do);
 996
 997                 if (unlikely(ret < 0))
 998                         break;
 999
1000                 idx += ret;
1001
1002                 if (unlikely(txreq.size < ETH_HLEN)) {
1003                         netdev_dbg(queue->vif->dev,
1004                                    "Bad packet size: %d\n", txreq.size);
1005                         xenvif_tx_err(queue, &txreq, extra_count, idx);
1006                         break;
1007                 }
1008
1009                 /* No crossing a page as the payload mustn't fragment. */
1010                 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
1011                         netdev_err(queue->vif->dev, "Cross page boundary, txreq.offset: %u, size: %u\n",
1012                                    txreq.offset, txreq.size);
1013                         xenvif_fatal_tx_err(queue->vif);
1014                         break;
1015                 }
1016
1017                 index = pending_index(queue->pending_cons);
1018                 pending_idx = queue->pending_ring[index];
1019
1020                 if (ret >= XEN_NETBK_LEGACY_SLOTS_MAX - 1 && data_len < txreq.size)
1021                         data_len = txreq.size;
1022
1023                 skb = xenvif_alloc_skb(data_len);
1024                 if (unlikely(skb == NULL)) {
1025                         netdev_dbg(queue->vif->dev,
1026                                    "Can't allocate a skb in start_xmit.\n");
1027                         xenvif_tx_err(queue, &txreq, extra_count, idx);
1028                         break;
1029                 }
1030
1031                 skb_shinfo(skb)->nr_frags = ret;
1032                 /* At this point shinfo->nr_frags is in fact the number of
1033                  * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1034                  */
1035                 frag_overflow = 0;
1036                 nskb = NULL;
1037                 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1038                         frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1039                         BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1040                         skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1041                         nskb = xenvif_alloc_skb(0);
1042                         if (unlikely(nskb == NULL)) {
1043                                 skb_shinfo(skb)->nr_frags = 0;
1044                                 kfree_skb(skb);
1045                                 xenvif_tx_err(queue, &txreq, extra_count, idx);
1046                                 if (net_ratelimit())
1047                                         netdev_err(queue->vif->dev,
1048                                                    "Can't allocate the frag_list skb.\n");
1049                                 break;
1050                         }
1051                 }
1052
1053                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1054                         struct xen_netif_extra_info *gso;
1055                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1056
1057                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1058                                 /* Failure in xenvif_set_skb_gso is fatal. */
1059                                 skb_shinfo(skb)->nr_frags = 0;
1060                                 kfree_skb(skb);
1061                                 kfree_skb(nskb);
1062                                 break;
1063                         }
1064                 }
1065
1066                 if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
1067                         struct xen_netif_extra_info *extra;
1068                         enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
1069
1070                         extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
1071
1072                         switch (extra->u.hash.type) {
1073                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
1074                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
1075                                 type = PKT_HASH_TYPE_L3;
1076                                 break;
1077
1078                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
1079                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
1080                                 type = PKT_HASH_TYPE_L4;
1081                                 break;
1082
1083                         default:
1084                                 break;
1085                         }
1086
1087                         if (type != PKT_HASH_TYPE_NONE)
1088                                 skb_set_hash(skb,
1089                                              *(u32 *)extra->u.hash.value,
1090                                              type);
1091                 }
1092
1093                 xenvif_get_requests(queue, skb, &txreq, txfrags, copy_ops,
1094                                     map_ops, frag_overflow, nskb, extra_count,
1095                                     data_len);
1096
1097                 __skb_queue_tail(&queue->tx_queue, skb);
1098
1099                 queue->tx.req_cons = idx;
1100
1101                 if ((*map_ops >= ARRAY_SIZE(queue->tx_map_ops)) ||
1102                     (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1103                         break;
1104         }
1105
1106         return;
1107 }
1108
1109 /* Consolidate skb with a frag_list into a brand new one with local pages on
1110  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1111  */
1112 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1113 {
1114         unsigned int offset = skb_headlen(skb);
1115         skb_frag_t frags[MAX_SKB_FRAGS];
1116         int i, f;
1117         struct ubuf_info *uarg;
1118         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1119
1120         queue->stats.tx_zerocopy_sent += 2;
1121         queue->stats.tx_frag_overflow++;
1122
1123         xenvif_fill_frags(queue, nskb);
1124         /* Subtract frags size, we will correct it later */
1125         skb->truesize -= skb->data_len;
1126         skb->len += nskb->len;
1127         skb->data_len += nskb->len;
1128
1129         /* create a brand new frags array and coalesce there */
1130         for (i = 0; offset < skb->len; i++) {
1131                 struct page *page;
1132                 unsigned int len;
1133
1134                 BUG_ON(i >= MAX_SKB_FRAGS);
1135                 page = alloc_page(GFP_ATOMIC);
1136                 if (!page) {
1137                         int j;
1138                         skb->truesize += skb->data_len;
1139                         for (j = 0; j < i; j++)
1140                                 put_page(frags[j].page.p);
1141                         return -ENOMEM;
1142                 }
1143
1144                 if (offset + PAGE_SIZE < skb->len)
1145                         len = PAGE_SIZE;
1146                 else
1147                         len = skb->len - offset;
1148                 if (skb_copy_bits(skb, offset, page_address(page), len))
1149                         BUG();
1150
1151                 offset += len;
1152                 frags[i].page.p = page;
1153                 frags[i].page_offset = 0;
1154                 skb_frag_size_set(&frags[i], len);
1155         }
1156
1157         /* Release all the original (foreign) frags. */
1158         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1159                 skb_frag_unref(skb, f);
1160         uarg = skb_shinfo(skb)->destructor_arg;
1161         /* increase inflight counter to offset decrement in callback */
1162         atomic_inc(&queue->inflight_packets);
1163         uarg->callback(uarg, true);
1164         skb_shinfo(skb)->destructor_arg = NULL;
1165
1166         /* Fill the skb with the new (local) frags. */
1167         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1168         skb_shinfo(skb)->nr_frags = i;
1169         skb->truesize += i * PAGE_SIZE;
1170
1171         return 0;
1172 }
1173
1174 static int xenvif_tx_submit(struct xenvif_queue *queue)
1175 {
1176         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1177         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1178         struct sk_buff *skb;
1179         int work_done = 0;
1180
1181         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1182                 struct xen_netif_tx_request *txp;
1183                 u16 pending_idx;
1184
1185                 pending_idx = copy_pending_idx(skb, 0);
1186                 txp = &queue->pending_tx_info[pending_idx].req;
1187
1188                 /* Check the remap error code. */
1189                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1190                         /* If there was an error, xenvif_tx_check_gop is
1191                          * expected to release all the frags which were mapped,
1192                          * so kfree_skb shouldn't do it again
1193                          */
1194                         skb_shinfo(skb)->nr_frags = 0;
1195                         if (skb_has_frag_list(skb)) {
1196                                 struct sk_buff *nskb =
1197                                                 skb_shinfo(skb)->frag_list;
1198                                 skb_shinfo(nskb)->nr_frags = 0;
1199                         }
1200                         kfree_skb(skb);
1201                         continue;
1202                 }
1203
1204                 if (txp->flags & XEN_NETTXF_csum_blank)
1205                         skb->ip_summed = CHECKSUM_PARTIAL;
1206                 else if (txp->flags & XEN_NETTXF_data_validated)
1207                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1208
1209                 xenvif_fill_frags(queue, skb);
1210
1211                 if (unlikely(skb_has_frag_list(skb))) {
1212                         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1213                         xenvif_skb_zerocopy_prepare(queue, nskb);
1214                         if (xenvif_handle_frag_list(queue, skb)) {
1215                                 if (net_ratelimit())
1216                                         netdev_err(queue->vif->dev,
1217                                                    "Not enough memory to consolidate frag_list!\n");
1218                                 xenvif_skb_zerocopy_prepare(queue, skb);
1219                                 kfree_skb(skb);
1220                                 continue;
1221                         }
1222                         /* Copied all the bits from the frag list -- free it. */
1223                         skb_frag_list_init(skb);
1224                         kfree_skb(nskb);
1225                 }
1226
1227                 skb->dev      = queue->vif->dev;
1228                 skb->protocol = eth_type_trans(skb, skb->dev);
1229                 skb_reset_network_header(skb);
1230
1231                 if (checksum_setup(queue, skb)) {
1232                         netdev_dbg(queue->vif->dev,
1233                                    "Can't setup checksum in net_tx_action\n");
1234                         /* We have to set this flag to trigger the callback */
1235                         if (skb_shinfo(skb)->destructor_arg)
1236                                 xenvif_skb_zerocopy_prepare(queue, skb);
1237                         kfree_skb(skb);
1238                         continue;
1239                 }
1240
1241                 skb_probe_transport_header(skb, 0);
1242
1243                 /* If the packet is GSO then we will have just set up the
1244                  * transport header offset in checksum_setup so it's now
1245                  * straightforward to calculate gso_segs.
1246                  */
1247                 if (skb_is_gso(skb)) {
1248                         int mss = skb_shinfo(skb)->gso_size;
1249                         int hdrlen = skb_transport_header(skb) -
1250                                 skb_mac_header(skb) +
1251                                 tcp_hdrlen(skb);
1252
1253                         skb_shinfo(skb)->gso_segs =
1254                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1255                 }
1256
1257                 queue->stats.rx_bytes += skb->len;
1258                 queue->stats.rx_packets++;
1259
1260                 work_done++;
1261
1262                 /* Set this flag right before netif_receive_skb, otherwise
1263                  * someone might think this packet already left netback, and
1264                  * do a skb_copy_ubufs while we are still in control of the
1265                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1266                  */
1267                 if (skb_shinfo(skb)->destructor_arg) {
1268                         xenvif_skb_zerocopy_prepare(queue, skb);
1269                         queue->stats.tx_zerocopy_sent++;
1270                 }
1271
1272                 netif_receive_skb(skb);
1273         }
1274
1275         return work_done;
1276 }
1277
1278 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1279 {
1280         unsigned long flags;
1281         pending_ring_idx_t index;
1282         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1283
1284         /* This is the only place where we grab this lock, to protect callbacks
1285          * from each other.
1286          */
1287         spin_lock_irqsave(&queue->callback_lock, flags);
1288         do {
1289                 u16 pending_idx = ubuf->desc;
1290                 ubuf = (struct ubuf_info *) ubuf->ctx;
1291                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1292                         MAX_PENDING_REQS);
1293                 index = pending_index(queue->dealloc_prod);
1294                 queue->dealloc_ring[index] = pending_idx;
1295                 /* Sync with xenvif_tx_dealloc_action:
1296                  * insert idx then incr producer.
1297                  */
1298                 smp_wmb();
1299                 queue->dealloc_prod++;
1300         } while (ubuf);
1301         spin_unlock_irqrestore(&queue->callback_lock, flags);
1302
1303         if (likely(zerocopy_success))
1304                 queue->stats.tx_zerocopy_success++;
1305         else
1306                 queue->stats.tx_zerocopy_fail++;
1307         xenvif_skb_zerocopy_complete(queue);
1308 }
1309
1310 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1311 {
1312         struct gnttab_unmap_grant_ref *gop;
1313         pending_ring_idx_t dc, dp;
1314         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1315         unsigned int i = 0;
1316
1317         dc = queue->dealloc_cons;
1318         gop = queue->tx_unmap_ops;
1319
1320         /* Free up any grants we have finished using */
1321         do {
1322                 dp = queue->dealloc_prod;
1323
1324                 /* Ensure we see all indices enqueued by all
1325                  * xenvif_zerocopy_callback().
1326                  */
1327                 smp_rmb();
1328
1329                 while (dc != dp) {
1330                         BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1331                         pending_idx =
1332                                 queue->dealloc_ring[pending_index(dc++)];
1333
1334                         pending_idx_release[gop - queue->tx_unmap_ops] =
1335                                 pending_idx;
1336                         queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1337                                 queue->mmap_pages[pending_idx];
1338                         gnttab_set_unmap_op(gop,
1339                                             idx_to_kaddr(queue, pending_idx),
1340                                             GNTMAP_host_map,
1341                                             queue->grant_tx_handle[pending_idx]);
1342                         xenvif_grant_handle_reset(queue, pending_idx);
1343                         ++gop;
1344                 }
1345
1346         } while (dp != queue->dealloc_prod);
1347
1348         queue->dealloc_cons = dc;
1349
1350         if (gop - queue->tx_unmap_ops > 0) {
1351                 int ret;
1352                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1353                                         NULL,
1354                                         queue->pages_to_unmap,
1355                                         gop - queue->tx_unmap_ops);
1356                 if (ret) {
1357                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1358                                    gop - queue->tx_unmap_ops, ret);
1359                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1360                                 if (gop[i].status != GNTST_okay)
1361                                         netdev_err(queue->vif->dev,
1362                                                    " host_addr: 0x%llx handle: 0x%x status: %d\n",
1363                                                    gop[i].host_addr,
1364                                                    gop[i].handle,
1365                                                    gop[i].status);
1366                         }
1367                         BUG();
1368                 }
1369         }
1370
1371         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1372                 xenvif_idx_release(queue, pending_idx_release[i],
1373                                    XEN_NETIF_RSP_OKAY);
1374 }
1375
1376
1377 /* Called after netfront has transmitted */
1378 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1379 {
1380         unsigned nr_mops = 0, nr_cops = 0;
1381         int work_done, ret;
1382
1383         if (unlikely(!tx_work_todo(queue)))
1384                 return 0;
1385
1386         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1387
1388         if (nr_cops == 0)
1389                 return 0;
1390
1391         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1392         if (nr_mops != 0) {
1393                 ret = gnttab_map_refs(queue->tx_map_ops,
1394                                       NULL,
1395                                       queue->pages_to_map,
1396                                       nr_mops);
1397                 if (ret) {
1398                         unsigned int i;
1399
1400                         netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1401                                    nr_mops, ret);
1402                         for (i = 0; i < nr_mops; ++i)
1403                                 WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1404                                              GNTST_okay);
1405                 }
1406         }
1407
1408         work_done = xenvif_tx_submit(queue);
1409
1410         return work_done;
1411 }
1412
1413 static void _make_tx_response(struct xenvif_queue *queue,
1414                              const struct xen_netif_tx_request *txp,
1415                              unsigned int extra_count,
1416                              s8 status)
1417 {
1418         RING_IDX i = queue->tx.rsp_prod_pvt;
1419         struct xen_netif_tx_response *resp;
1420
1421         resp = RING_GET_RESPONSE(&queue->tx, i);
1422         resp->id     = txp->id;
1423         resp->status = status;
1424
1425         while (extra_count-- != 0)
1426                 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1427
1428         queue->tx.rsp_prod_pvt = ++i;
1429 }
1430
1431 static void push_tx_responses(struct xenvif_queue *queue)
1432 {
1433         int notify;
1434
1435         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1436         if (notify)
1437                 notify_remote_via_irq(queue->tx_irq);
1438 }
1439
1440 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1441                                s8 status)
1442 {
1443         struct pending_tx_info *pending_tx_info;
1444         pending_ring_idx_t index;
1445         unsigned long flags;
1446
1447         pending_tx_info = &queue->pending_tx_info[pending_idx];
1448
1449         spin_lock_irqsave(&queue->response_lock, flags);
1450
1451         _make_tx_response(queue, &pending_tx_info->req,
1452                           pending_tx_info->extra_count, status);
1453
1454         /* Release the pending index before pusing the Tx response so
1455          * its available before a new Tx request is pushed by the
1456          * frontend.
1457          */
1458         index = pending_index(queue->pending_prod++);
1459         queue->pending_ring[index] = pending_idx;
1460
1461         push_tx_responses(queue);
1462
1463         spin_unlock_irqrestore(&queue->response_lock, flags);
1464 }
1465
1466 static void make_tx_response(struct xenvif_queue *queue,
1467                              const struct xen_netif_tx_request *txp,
1468                              unsigned int extra_count,
1469                              s8 status)
1470 {
1471         unsigned long flags;
1472
1473         spin_lock_irqsave(&queue->response_lock, flags);
1474
1475         _make_tx_response(queue, txp, extra_count, status);
1476         push_tx_responses(queue);
1477
1478         spin_unlock_irqrestore(&queue->response_lock, flags);
1479 }
1480
1481 static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1482 {
1483         int ret;
1484         struct gnttab_unmap_grant_ref tx_unmap_op;
1485
1486         gnttab_set_unmap_op(&tx_unmap_op,
1487                             idx_to_kaddr(queue, pending_idx),
1488                             GNTMAP_host_map,
1489                             queue->grant_tx_handle[pending_idx]);
1490         xenvif_grant_handle_reset(queue, pending_idx);
1491
1492         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1493                                 &queue->mmap_pages[pending_idx], 1);
1494         if (ret) {
1495                 netdev_err(queue->vif->dev,
1496                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1497                            ret,
1498                            pending_idx,
1499                            tx_unmap_op.host_addr,
1500                            tx_unmap_op.handle,
1501                            tx_unmap_op.status);
1502                 BUG();
1503         }
1504 }
1505
1506 static inline int tx_work_todo(struct xenvif_queue *queue)
1507 {
1508         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1509                 return 1;
1510
1511         return 0;
1512 }
1513
1514 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1515 {
1516         return queue->dealloc_cons != queue->dealloc_prod;
1517 }
1518
1519 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1520 {
1521         if (queue->tx.sring)
1522                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1523                                         queue->tx.sring);
1524         if (queue->rx.sring)
1525                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1526                                         queue->rx.sring);
1527 }
1528
1529 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1530                                    grant_ref_t tx_ring_ref,
1531                                    grant_ref_t rx_ring_ref)
1532 {
1533         void *addr;
1534         struct xen_netif_tx_sring *txs;
1535         struct xen_netif_rx_sring *rxs;
1536
1537         int err = -ENOMEM;
1538
1539         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1540                                      &tx_ring_ref, 1, &addr);
1541         if (err)
1542                 goto err;
1543
1544         txs = (struct xen_netif_tx_sring *)addr;
1545         BACK_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE);
1546
1547         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1548                                      &rx_ring_ref, 1, &addr);
1549         if (err)
1550                 goto err;
1551
1552         rxs = (struct xen_netif_rx_sring *)addr;
1553         BACK_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE);
1554
1555         return 0;
1556
1557 err:
1558         xenvif_unmap_frontend_data_rings(queue);
1559         return err;
1560 }
1561
1562 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1563 {
1564         /* Dealloc thread must remain running until all inflight
1565          * packets complete.
1566          */
1567         return kthread_should_stop() &&
1568                 !atomic_read(&queue->inflight_packets);
1569 }
1570
1571 int xenvif_dealloc_kthread(void *data)
1572 {
1573         struct xenvif_queue *queue = data;
1574
1575         for (;;) {
1576                 wait_event_interruptible(queue->dealloc_wq,
1577                                          tx_dealloc_work_todo(queue) ||
1578                                          xenvif_dealloc_kthread_should_stop(queue));
1579                 if (xenvif_dealloc_kthread_should_stop(queue))
1580                         break;
1581
1582                 xenvif_tx_dealloc_action(queue);
1583                 cond_resched();
1584         }
1585
1586         /* Unmap anything remaining*/
1587         if (tx_dealloc_work_todo(queue))
1588                 xenvif_tx_dealloc_action(queue);
1589
1590         return 0;
1591 }
1592
1593 static void make_ctrl_response(struct xenvif *vif,
1594                                const struct xen_netif_ctrl_request *req,
1595                                u32 status, u32 data)
1596 {
1597         RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1598         struct xen_netif_ctrl_response rsp = {
1599                 .id = req->id,
1600                 .type = req->type,
1601                 .status = status,
1602                 .data = data,
1603         };
1604
1605         *RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1606         vif->ctrl.rsp_prod_pvt = ++idx;
1607 }
1608
1609 static void push_ctrl_response(struct xenvif *vif)
1610 {
1611         int notify;
1612
1613         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1614         if (notify)
1615                 notify_remote_via_irq(vif->ctrl_irq);
1616 }
1617
1618 static void process_ctrl_request(struct xenvif *vif,
1619                                  const struct xen_netif_ctrl_request *req)
1620 {
1621         u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1622         u32 data = 0;
1623
1624         switch (req->type) {
1625         case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1626                 status = xenvif_set_hash_alg(vif, req->data[0]);
1627                 break;
1628
1629         case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1630                 status = xenvif_get_hash_flags(vif, &data);
1631                 break;
1632
1633         case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1634                 status = xenvif_set_hash_flags(vif, req->data[0]);
1635                 break;
1636
1637         case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1638                 status = xenvif_set_hash_key(vif, req->data[0],
1639                                              req->data[1]);
1640                 break;
1641
1642         case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1643                 status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1644                 data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1645                 break;
1646
1647         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1648                 status = xenvif_set_hash_mapping_size(vif,
1649                                                       req->data[0]);
1650                 break;
1651
1652         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1653                 status = xenvif_set_hash_mapping(vif, req->data[0],
1654                                                  req->data[1],
1655                                                  req->data[2]);
1656                 break;
1657
1658         default:
1659                 break;
1660         }
1661
1662         make_ctrl_response(vif, req, status, data);
1663         push_ctrl_response(vif);
1664 }
1665
1666 static void xenvif_ctrl_action(struct xenvif *vif)
1667 {
1668         for (;;) {
1669                 RING_IDX req_prod, req_cons;
1670
1671                 req_prod = vif->ctrl.sring->req_prod;
1672                 req_cons = vif->ctrl.req_cons;
1673
1674                 /* Make sure we can see requests before we process them. */
1675                 rmb();
1676
1677                 if (req_cons == req_prod)
1678                         break;
1679
1680                 while (req_cons != req_prod) {
1681                         struct xen_netif_ctrl_request req;
1682
1683                         RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1684                         req_cons++;
1685
1686                         process_ctrl_request(vif, &req);
1687                 }
1688
1689                 vif->ctrl.req_cons = req_cons;
1690                 vif->ctrl.sring->req_event = req_cons + 1;
1691         }
1692 }
1693
1694 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1695 {
1696         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1697                 return true;
1698
1699         return false;
1700 }
1701
1702 irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1703 {
1704         struct xenvif *vif = data;
1705         unsigned int eoi_flag = XEN_EOI_FLAG_SPURIOUS;
1706
1707         while (xenvif_ctrl_work_todo(vif)) {
1708                 xenvif_ctrl_action(vif);
1709                 eoi_flag = 0;
1710         }
1711
1712         xen_irq_lateeoi(irq, eoi_flag);
1713
1714         return IRQ_HANDLED;
1715 }
1716
1717 static int __init netback_init(void)
1718 {
1719         int rc = 0;
1720
1721         if (!xen_domain())
1722                 return -ENODEV;
1723
1724         /* Allow as many queues as there are CPUs but max. 8 if user has not
1725          * specified a value.
1726          */
1727         if (xenvif_max_queues == 0)
1728                 xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1729                                           num_online_cpus());
1730
1731         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1732                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1733                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1734                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1735         }
1736
1737         rc = xenvif_xenbus_init();
1738         if (rc)
1739                 goto failed_init;
1740
1741 #ifdef CONFIG_DEBUG_FS
1742         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1743         if (IS_ERR_OR_NULL(xen_netback_dbg_root))
1744                 pr_warn("Init of debugfs returned %ld!\n",
1745                         PTR_ERR(xen_netback_dbg_root));
1746 #endif /* CONFIG_DEBUG_FS */
1747
1748         return 0;
1749
1750 failed_init:
1751         return rc;
1752 }
1753
1754 module_init(netback_init);
1755
1756 static void __exit netback_fini(void)
1757 {
1758 #ifdef CONFIG_DEBUG_FS
1759         if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
1760                 debugfs_remove_recursive(xen_netback_dbg_root);
1761 #endif /* CONFIG_DEBUG_FS */
1762         xenvif_xenbus_fini();
1763 }
1764 module_exit(netback_fini);
1765
1766 MODULE_LICENSE("Dual BSD/GPL");
1767 MODULE_ALIAS("xen-backend:vif");