drivers/net/xen-netback/netback.c

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