drivers/net/xen-netback/netback.c

   1 /*
   2  * Back-end of the driver for virtual network devices. This portion of the
   3  * driver exports a 'unified' network-device interface that can be accessed
   4  * by any operating system that implements a compatible front end. A
   5  * reference front-end implementation can be found in:
   6  *  drivers/net/xen-netfront.c
   7  *
   8  * Copyright (c) 2002-2005, K A Fraser
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public License version 2
  12  * as published by the Free Software Foundation; or, when distributed
  13  * separately from the Linux kernel or incorporated into other
  14  * software packages, subject to the following license:
  15  *
  16  * Permission is hereby granted, free of charge, to any person obtaining a copy
  17  * of this source file (the "Software"), to deal in the Software without
  18  * restriction, including without limitation the rights to use, copy, modify,
  19  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  20  * and to permit persons to whom the Software is furnished to do so, subject to
  21  * the following conditions:
  22  *
  23  * The above copyright notice and this permission notice shall be included in
  24  * all copies or substantial portions of the Software.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  27  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  28  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  29  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  30  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  31  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  32  * IN THE SOFTWARE.
  33  */
  34
  35 #include "common.h"
  36
  37 #include <linux/kthread.h>
  38 #include <linux/if_vlan.h>
  39 #include <linux/udp.h>
  40 #include <linux/highmem.h>
  41
  42 #include <net/tcp.h>
  43
  44 #include <xen/xen.h>
  45 #include <xen/events.h>
  46 #include <xen/interface/memory.h>
  47 #include <xen/page.h>
  48
  49 #include <asm/xen/hypercall.h>
  50
  51 /* Provide an option to disable split event channels at load time as
  52  * event channels are limited resource. Split event channels are
  53  * enabled by default.
  54  */
  55 bool separate_tx_rx_irq = true;
  56 module_param(separate_tx_rx_irq, bool, 0644);
  57
  58 /* The time that packets can stay on the guest Rx internal queue
  59  * before they are dropped.
  60  */
  61 unsigned int rx_drain_timeout_msecs = 10000;
  62 module_param(rx_drain_timeout_msecs, uint, 0444);
  63
  64 /* The length of time before the frontend is considered unresponsive
  65  * because it isn't providing Rx slots.
  66  */
  67 unsigned int rx_stall_timeout_msecs = 60000;
  68 module_param(rx_stall_timeout_msecs, uint, 0444);
  69
  70 #define MAX_QUEUES_DEFAULT 8
  71 unsigned int xenvif_max_queues;
  72 module_param_named(max_queues, xenvif_max_queues, uint, 0644);
  73 MODULE_PARM_DESC(max_queues,
  74                  "Maximum number of queues per virtual interface");
  75
  76 /*
  77  * This is the maximum slots a skb can have. If a guest sends a skb
  78  * which exceeds this limit it is considered malicious.
  79  */
  80 #define FATAL_SKB_SLOTS_DEFAULT 20
  81 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
  82 module_param(fatal_skb_slots, uint, 0444);
  83
  84 /* The amount to copy out of the first guest Tx slot into the skb's
  85  * linear area.  If the first slot has more data, it will be mapped
  86  * and put into the first frag.
  87  *
  88  * This is sized to avoid pulling headers from the frags for most
  89  * TCP/IP packets.
  90  */
  91 #define XEN_NETBACK_TX_COPY_LEN 128
  92
  93 /* This is the maximum number of flows in the hash cache. */
  94 #define XENVIF_HASH_CACHE_SIZE_DEFAULT 64
  95 unsigned int xenvif_hash_cache_size = XENVIF_HASH_CACHE_SIZE_DEFAULT;
  96 module_param_named(hash_cache_size, xenvif_hash_cache_size, uint, 0644);
  97 MODULE_PARM_DESC(hash_cache_size, "Number of flows in the hash cache");
  98
  99 /* The module parameter tells that we have to put data
 100  * for xen-netfront with the XDP_PACKET_HEADROOM offset
 101  * needed for XDP processing
 102  */
 103 bool provides_xdp_headroom = true;
 104 module_param(provides_xdp_headroom, bool, 0644);
 105
 106 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
 107                                u8 status);
 108
 109 static void make_tx_response(struct xenvif_queue *queue,
 110                              struct xen_netif_tx_request *txp,
 111                              unsigned int extra_count,
 112                              s8       st);
 113 static void push_tx_responses(struct xenvif_queue *queue);
 114
 115 static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx);
 116
 117 static inline int tx_work_todo(struct xenvif_queue *queue);
 118
 119 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
 120                                        u16 idx)
 121 {
 122         return page_to_pfn(queue->mmap_pages[idx]);
 123 }
 124
 125 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
 126                                          u16 idx)
 127 {
 128         return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
 129 }
 130
 131 #define callback_param(vif, pending_idx) \
 132         (vif->pending_tx_info[pending_idx].callback_struct)
 133
 134 /* Find the containing VIF's structure from a pointer in pending_tx_info array
 135  */
 136 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info_msgzc *ubuf)
 137 {
 138         u16 pending_idx = ubuf->desc;
 139         struct pending_tx_info *temp =
 140                 container_of(ubuf, struct pending_tx_info, callback_struct);
 141         return container_of(temp - pending_idx,
 142                             struct xenvif_queue,
 143                             pending_tx_info[0]);
 144 }
 145
 146 static u16 frag_get_pending_idx(skb_frag_t *frag)
 147 {
 148         return (u16)skb_frag_off(frag);
 149 }
 150
 151 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 152 {
 153         skb_frag_off_set(frag, pending_idx);
 154 }
 155
 156 static inline pending_ring_idx_t pending_index(unsigned i)
 157 {
 158         return i & (MAX_PENDING_REQS-1);
 159 }
 160
 161 void xenvif_kick_thread(struct xenvif_queue *queue)
 162 {
 163         wake_up(&queue->wq);
 164 }
 165
 166 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 167 {
 168         int more_to_do;
 169
 170         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 171
 172         if (more_to_do)
 173                 napi_schedule(&queue->napi);
 174         else if (atomic_fetch_andnot(NETBK_TX_EOI | NETBK_COMMON_EOI,
 175                                      &queue->eoi_pending) &
 176                  (NETBK_TX_EOI | NETBK_COMMON_EOI))
 177                 xen_irq_lateeoi(queue->tx_irq, 0);
 178 }
 179
 180 static void tx_add_credit(struct xenvif_queue *queue)
 181 {
 182         unsigned long max_burst, max_credit;
 183
 184         /*
 185          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 186          * Otherwise the interface can seize up due to insufficient credit.
 187          */
 188         max_burst = max(131072UL, queue->credit_bytes);
 189
 190         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 191         max_credit = queue->remaining_credit + queue->credit_bytes;
 192         if (max_credit < queue->remaining_credit)
 193                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 194
 195         queue->remaining_credit = min(max_credit, max_burst);
 196         queue->rate_limited = false;
 197 }
 198
 199 void xenvif_tx_credit_callback(struct timer_list *t)
 200 {
 201         struct xenvif_queue *queue = from_timer(queue, t, credit_timeout);
 202         tx_add_credit(queue);
 203         xenvif_napi_schedule_or_enable_events(queue);
 204 }
 205
 206 static void xenvif_tx_err(struct xenvif_queue *queue,
 207                           struct xen_netif_tx_request *txp,
 208                           unsigned int extra_count, RING_IDX end)
 209 {
 210         RING_IDX cons = queue->tx.req_cons;
 211         unsigned long flags;
 212
 213         do {
 214                 spin_lock_irqsave(&queue->response_lock, flags);
 215                 make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
 216                 push_tx_responses(queue);
 217                 spin_unlock_irqrestore(&queue->response_lock, flags);
 218                 if (cons == end)
 219                         break;
 220                 RING_COPY_REQUEST(&queue->tx, cons++, txp);
 221                 extra_count = 0; /* only the first frag can have extras */
 222         } while (1);
 223         queue->tx.req_cons = cons;
 224 }
 225
 226 static void xenvif_fatal_tx_err(struct xenvif *vif)
 227 {
 228         netdev_err(vif->dev, "fatal error; disabling device\n");
 229         vif->disabled = true;
 230         /* Disable the vif from queue 0's kthread */
 231         if (vif->num_queues)
 232                 xenvif_kick_thread(&vif->queues[0]);
 233 }
 234
 235 static int xenvif_count_requests(struct xenvif_queue *queue,
 236                                  struct xen_netif_tx_request *first,
 237                                  unsigned int extra_count,
 238                                  struct xen_netif_tx_request *txp,
 239                                  int work_to_do)
 240 {
 241         RING_IDX cons = queue->tx.req_cons;
 242         int slots = 0;
 243         int drop_err = 0;
 244         int more_data;
 245
 246         if (!(first->flags & XEN_NETTXF_more_data))
 247                 return 0;
 248
 249         do {
 250                 struct xen_netif_tx_request dropped_tx = { 0 };
 251
 252                 if (slots >= work_to_do) {
 253                         netdev_err(queue->vif->dev,
 254                                    "Asked for %d slots but exceeds this limit\n",
 255                                    work_to_do);
 256                         xenvif_fatal_tx_err(queue->vif);
 257                         return -ENODATA;
 258                 }
 259
 260                 /* This guest is really using too many slots and
 261                  * considered malicious.
 262                  */
 263                 if (unlikely(slots >= fatal_skb_slots)) {
 264                         netdev_err(queue->vif->dev,
 265                                    "Malicious frontend using %d slots, threshold %u\n",
 266                                    slots, fatal_skb_slots);
 267                         xenvif_fatal_tx_err(queue->vif);
 268                         return -E2BIG;
 269                 }
 270
 271                 /* Xen network protocol had implicit dependency on
 272                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 273                  * the historical MAX_SKB_FRAGS value 18 to honor the
 274                  * same behavior as before. Any packet using more than
 275                  * 18 slots but less than fatal_skb_slots slots is
 276                  * dropped
 277                  */
 278                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 279                         if (net_ratelimit())
 280                                 netdev_dbg(queue->vif->dev,
 281                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 282                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 283                         drop_err = -E2BIG;
 284                 }
 285
 286                 if (drop_err)
 287                         txp = &dropped_tx;
 288
 289                 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
 290
 291                 /* If the guest submitted a frame >= 64 KiB then
 292                  * first->size overflowed and following slots will
 293                  * appear to be larger than the frame.
 294                  *
 295                  * This cannot be fatal error as there are buggy
 296                  * frontends that do this.
 297                  *
 298                  * Consume all slots and drop the packet.
 299                  */
 300                 if (!drop_err && txp->size > first->size) {
 301                         if (net_ratelimit())
 302                                 netdev_dbg(queue->vif->dev,
 303                                            "Invalid tx request, slot size %u > remaining size %u\n",
 304                                            txp->size, first->size);
 305                         drop_err = -EIO;
 306                 }
 307
 308                 first->size -= txp->size;
 309                 slots++;
 310
 311                 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
 312                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 313                                  txp->offset, txp->size);
 314                         xenvif_fatal_tx_err(queue->vif);
 315                         return -EINVAL;
 316                 }
 317
 318                 more_data = txp->flags & XEN_NETTXF_more_data;
 319
 320                 if (!drop_err)
 321                         txp++;
 322
 323         } while (more_data);
 324
 325         if (drop_err) {
 326                 xenvif_tx_err(queue, first, extra_count, cons + slots);
 327                 return drop_err;
 328         }
 329
 330         return slots;
 331 }
 332
 333
 334 struct xenvif_tx_cb {
 335         u16 copy_pending_idx[XEN_NETBK_LEGACY_SLOTS_MAX + 1];
 336         u8 copy_count;
 337 };
 338
 339 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 340 #define copy_pending_idx(skb, i) (XENVIF_TX_CB(skb)->copy_pending_idx[i])
 341 #define copy_count(skb) (XENVIF_TX_CB(skb)->copy_count)
 342
 343 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 344                                            u16 pending_idx,
 345                                            struct xen_netif_tx_request *txp,
 346                                            unsigned int extra_count,
 347                                            struct gnttab_map_grant_ref *mop)
 348 {
 349         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 350         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 351                           GNTMAP_host_map | GNTMAP_readonly,
 352                           txp->gref, queue->vif->domid);
 353
 354         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 355                sizeof(*txp));
 356         queue->pending_tx_info[pending_idx].extra_count = extra_count;
 357 }
 358
 359 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 360 {
 361         struct sk_buff *skb =
 362                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 363                           GFP_ATOMIC | __GFP_NOWARN);
 364         if (unlikely(skb == NULL))
 365                 return NULL;
 366
 367         /* Packets passed to netif_rx() must have some headroom. */
 368         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 369
 370         /* Initialize it here to avoid later surprises */
 371         skb_shinfo(skb)->destructor_arg = NULL;
 372
 373         return skb;
 374 }
 375
 376 static void xenvif_get_requests(struct xenvif_queue *queue,
 377                                 struct sk_buff *skb,
 378                                 struct xen_netif_tx_request *first,
 379                                 struct xen_netif_tx_request *txfrags,
 380                                 unsigned *copy_ops,
 381                                 unsigned *map_ops,
 382                                 unsigned int frag_overflow,
 383                                 struct sk_buff *nskb,
 384                                 unsigned int extra_count,
 385                                 unsigned int data_len)
 386 {
 387         struct skb_shared_info *shinfo = skb_shinfo(skb);
 388         skb_frag_t *frags = shinfo->frags;
 389         u16 pending_idx;
 390         pending_ring_idx_t index;
 391         unsigned int nr_slots;
 392         struct gnttab_copy *cop = queue->tx_copy_ops + *copy_ops;
 393         struct gnttab_map_grant_ref *gop = queue->tx_map_ops + *map_ops;
 394         struct xen_netif_tx_request *txp = first;
 395
 396         nr_slots = shinfo->nr_frags + 1;
 397
 398         copy_count(skb) = 0;
 399
 400         /* Create copy ops for exactly data_len bytes into the skb head. */
 401         __skb_put(skb, data_len);
 402         while (data_len > 0) {
 403                 int amount = data_len > txp->size ? txp->size : data_len;
 404
 405                 cop->source.u.ref = txp->gref;
 406                 cop->source.domid = queue->vif->domid;
 407                 cop->source.offset = txp->offset;
 408
 409                 cop->dest.domid = DOMID_SELF;
 410                 cop->dest.offset = (offset_in_page(skb->data +
 411                                                    skb_headlen(skb) -
 412                                                    data_len)) & ~XEN_PAGE_MASK;
 413                 cop->dest.u.gmfn = virt_to_gfn(skb->data + skb_headlen(skb)
 414                                                - data_len);
 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                 copy_count(skb)++;
 424
 425                 cop++;
 426                 data_len -= amount;
 427
 428                 if (amount == txp->size) {
 429                         /* The copy op covered the full tx_request */
 430
 431                         memcpy(&queue->pending_tx_info[pending_idx].req,
 432                                txp, sizeof(*txp));
 433                         queue->pending_tx_info[pending_idx].extra_count =
 434                                 (txp == first) ? extra_count : 0;
 435
 436                         if (txp == first)
 437                                 txp = txfrags;
 438                         else
 439                                 txp++;
 440                         queue->pending_cons++;
 441                         nr_slots--;
 442                 } else {
 443                         /* The copy op partially covered the tx_request.
 444                          * The remainder will be mapped.
 445                          */
 446                         txp->offset += amount;
 447                         txp->size -= amount;
 448                 }
 449         }
 450
 451         for (shinfo->nr_frags = 0; shinfo->nr_frags < nr_slots;
 452              shinfo->nr_frags++, gop++) {
 453                 index = pending_index(queue->pending_cons++);
 454                 pending_idx = queue->pending_ring[index];
 455                 xenvif_tx_create_map_op(queue, pending_idx, txp,
 456                                         txp == first ? extra_count : 0, gop);
 457                 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 458
 459                 if (txp == first)
 460                         txp = txfrags;
 461                 else
 462                         txp++;
 463         }
 464
 465         if (frag_overflow) {
 466
 467                 shinfo = skb_shinfo(nskb);
 468                 frags = shinfo->frags;
 469
 470                 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
 471                      shinfo->nr_frags++, txp++, gop++) {
 472                         index = pending_index(queue->pending_cons++);
 473                         pending_idx = queue->pending_ring[index];
 474                         xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
 475                                                 gop);
 476                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 477                                              pending_idx);
 478                 }
 479
 480                 skb_shinfo(skb)->frag_list = nskb;
 481         }
 482
 483         (*copy_ops) = cop - queue->tx_copy_ops;
 484         (*map_ops) = gop - queue->tx_map_ops;
 485 }
 486
 487 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 488                                            u16 pending_idx,
 489                                            grant_handle_t handle)
 490 {
 491         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 492                      NETBACK_INVALID_HANDLE)) {
 493                 netdev_err(queue->vif->dev,
 494                            "Trying to overwrite active handle! pending_idx: 0x%x\n",
 495                            pending_idx);
 496                 BUG();
 497         }
 498         queue->grant_tx_handle[pending_idx] = handle;
 499 }
 500
 501 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 502                                              u16 pending_idx)
 503 {
 504         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 505                      NETBACK_INVALID_HANDLE)) {
 506                 netdev_err(queue->vif->dev,
 507                            "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 508                            pending_idx);
 509                 BUG();
 510         }
 511         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 512 }
 513
 514 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 515                                struct sk_buff *skb,
 516                                struct gnttab_map_grant_ref **gopp_map,
 517                                struct gnttab_copy **gopp_copy)
 518 {
 519         struct gnttab_map_grant_ref *gop_map = *gopp_map;
 520         u16 pending_idx;
 521         /* This always points to the shinfo of the skb being checked, which
 522          * could be either the first or the one on the frag_list
 523          */
 524         struct skb_shared_info *shinfo = skb_shinfo(skb);
 525         /* If this is non-NULL, we are currently checking the frag_list skb, and
 526          * this points to the shinfo of the first one
 527          */
 528         struct skb_shared_info *first_shinfo = NULL;
 529         int nr_frags = shinfo->nr_frags;
 530         const bool sharedslot = nr_frags &&
 531                                 frag_get_pending_idx(&shinfo->frags[0]) ==
 532                                     copy_pending_idx(skb, copy_count(skb) - 1);
 533         int i, err = 0;
 534
 535         for (i = 0; i < copy_count(skb); i++) {
 536                 int newerr;
 537
 538                 /* Check status of header. */
 539                 pending_idx = copy_pending_idx(skb, i);
 540
 541                 newerr = (*gopp_copy)->status;
 542                 if (likely(!newerr)) {
 543                         /* The first frag might still have this slot mapped */
 544                         if (i < copy_count(skb) - 1 || !sharedslot)
 545                                 xenvif_idx_release(queue, pending_idx,
 546                                                    XEN_NETIF_RSP_OKAY);
 547                 } else {
 548                         err = newerr;
 549                         if (net_ratelimit())
 550                                 netdev_dbg(queue->vif->dev,
 551                                            "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 552                                            (*gopp_copy)->status,
 553                                            pending_idx,
 554                                            (*gopp_copy)->source.u.ref);
 555                         /* The first frag might still have this slot mapped */
 556                         if (i < copy_count(skb) - 1 || !sharedslot)
 557                                 xenvif_idx_release(queue, pending_idx,
 558                                                    XEN_NETIF_RSP_ERROR);
 559                 }
 560                 (*gopp_copy)++;
 561         }
 562
 563 check_frags:
 564         for (i = 0; i < nr_frags; i++, gop_map++) {
 565                 int j, newerr;
 566
 567                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 568
 569                 /* Check error status: if okay then remember grant handle. */
 570                 newerr = gop_map->status;
 571
 572                 if (likely(!newerr)) {
 573                         xenvif_grant_handle_set(queue,
 574                                                 pending_idx,
 575                                                 gop_map->handle);
 576                         /* Had a previous error? Invalidate this fragment. */
 577                         if (unlikely(err)) {
 578                                 xenvif_idx_unmap(queue, pending_idx);
 579                                 /* If the mapping of the first frag was OK, but
 580                                  * the header's copy failed, and they are
 581                                  * sharing a slot, send an error
 582                                  */
 583                                 if (i == 0 && !first_shinfo && sharedslot)
 584                                         xenvif_idx_release(queue, pending_idx,
 585                                                            XEN_NETIF_RSP_ERROR);
 586                                 else
 587                                         xenvif_idx_release(queue, pending_idx,
 588                                                            XEN_NETIF_RSP_OKAY);
 589                         }
 590                         continue;
 591                 }
 592
 593                 /* Error on this fragment: respond to client with an error. */
 594                 if (net_ratelimit())
 595                         netdev_dbg(queue->vif->dev,
 596                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
 597                                    i,
 598                                    gop_map->status,
 599                                    pending_idx,
 600                                    gop_map->ref);
 601
 602                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
 603
 604                 /* Not the first error? Preceding frags already invalidated. */
 605                 if (err)
 606                         continue;
 607
 608                 /* Invalidate preceding fragments of this skb. */
 609                 for (j = 0; j < i; j++) {
 610                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
 611                         xenvif_idx_unmap(queue, pending_idx);
 612                         xenvif_idx_release(queue, pending_idx,
 613                                            XEN_NETIF_RSP_OKAY);
 614                 }
 615
 616                 /* And if we found the error while checking the frag_list, unmap
 617                  * the first skb's frags
 618                  */
 619                 if (first_shinfo) {
 620                         for (j = 0; j < first_shinfo->nr_frags; j++) {
 621                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
 622                                 xenvif_idx_unmap(queue, pending_idx);
 623                                 xenvif_idx_release(queue, pending_idx,
 624                                                    XEN_NETIF_RSP_OKAY);
 625                         }
 626                 }
 627
 628                 /* Remember the error: invalidate all subsequent fragments. */
 629                 err = newerr;
 630         }
 631
 632         if (skb_has_frag_list(skb) && !first_shinfo) {
 633                 first_shinfo = shinfo;
 634                 shinfo = skb_shinfo(shinfo->frag_list);
 635                 nr_frags = shinfo->nr_frags;
 636
 637                 goto check_frags;
 638         }
 639
 640         *gopp_map = gop_map;
 641         return err;
 642 }
 643
 644 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
 645 {
 646         struct skb_shared_info *shinfo = skb_shinfo(skb);
 647         int nr_frags = shinfo->nr_frags;
 648         int i;
 649         u16 prev_pending_idx = INVALID_PENDING_IDX;
 650
 651         for (i = 0; i < nr_frags; i++) {
 652                 skb_frag_t *frag = shinfo->frags + i;
 653                 struct xen_netif_tx_request *txp;
 654                 struct page *page;
 655                 u16 pending_idx;
 656
 657                 pending_idx = frag_get_pending_idx(frag);
 658
 659                 /* If this is not the first frag, chain it to the previous*/
 660                 if (prev_pending_idx == INVALID_PENDING_IDX)
 661                         skb_shinfo(skb)->destructor_arg =
 662                                 &callback_param(queue, pending_idx);
 663                 else
 664                         callback_param(queue, prev_pending_idx).ctx =
 665                                 &callback_param(queue, pending_idx);
 666
 667                 callback_param(queue, pending_idx).ctx = NULL;
 668                 prev_pending_idx = pending_idx;
 669
 670                 txp = &queue->pending_tx_info[pending_idx].req;
 671                 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
 672                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
 673                 skb->len += txp->size;
 674                 skb->data_len += txp->size;
 675                 skb->truesize += txp->size;
 676
 677                 /* Take an extra reference to offset network stack's put_page */
 678                 get_page(queue->mmap_pages[pending_idx]);
 679         }
 680 }
 681
 682 static int xenvif_get_extras(struct xenvif_queue *queue,
 683                              struct xen_netif_extra_info *extras,
 684                              unsigned int *extra_count,
 685                              int work_to_do)
 686 {
 687         struct xen_netif_extra_info extra;
 688         RING_IDX cons = queue->tx.req_cons;
 689
 690         do {
 691                 if (unlikely(work_to_do-- <= 0)) {
 692                         netdev_err(queue->vif->dev, "Missing extra info\n");
 693                         xenvif_fatal_tx_err(queue->vif);
 694                         return -EBADR;
 695                 }
 696
 697                 RING_COPY_REQUEST(&queue->tx, cons, &extra);
 698
 699                 queue->tx.req_cons = ++cons;
 700                 (*extra_count)++;
 701
 702                 if (unlikely(!extra.type ||
 703                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
 704                         netdev_err(queue->vif->dev,
 705                                    "Invalid extra type: %d\n", extra.type);
 706                         xenvif_fatal_tx_err(queue->vif);
 707                         return -EINVAL;
 708                 }
 709
 710                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
 711         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
 712
 713         return work_to_do;
 714 }
 715
 716 static int xenvif_set_skb_gso(struct xenvif *vif,
 717                               struct sk_buff *skb,
 718                               struct xen_netif_extra_info *gso)
 719 {
 720         if (!gso->u.gso.size) {
 721                 netdev_err(vif->dev, "GSO size must not be zero.\n");
 722                 xenvif_fatal_tx_err(vif);
 723                 return -EINVAL;
 724         }
 725
 726         switch (gso->u.gso.type) {
 727         case XEN_NETIF_GSO_TYPE_TCPV4:
 728                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 729                 break;
 730         case XEN_NETIF_GSO_TYPE_TCPV6:
 731                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 732                 break;
 733         default:
 734                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
 735                 xenvif_fatal_tx_err(vif);
 736                 return -EINVAL;
 737         }
 738
 739         skb_shinfo(skb)->gso_size = gso->u.gso.size;
 740         /* gso_segs will be calculated later */
 741
 742         return 0;
 743 }
 744
 745 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
 746 {
 747         bool recalculate_partial_csum = false;
 748
 749         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
 750          * peers can fail to set NETRXF_csum_blank when sending a GSO
 751          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
 752          * recalculate the partial checksum.
 753          */
 754         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
 755                 queue->stats.rx_gso_checksum_fixup++;
 756                 skb->ip_summed = CHECKSUM_PARTIAL;
 757                 recalculate_partial_csum = true;
 758         }
 759
 760         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
 761         if (skb->ip_summed != CHECKSUM_PARTIAL)
 762                 return 0;
 763
 764         return skb_checksum_setup(skb, recalculate_partial_csum);
 765 }
 766
 767 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
 768 {
 769         u64 now = get_jiffies_64();
 770         u64 next_credit = queue->credit_window_start +
 771                 msecs_to_jiffies(queue->credit_usec / 1000);
 772
 773         /* Timer could already be pending in rare cases. */
 774         if (timer_pending(&queue->credit_timeout)) {
 775                 queue->rate_limited = true;
 776                 return true;
 777         }
 778
 779         /* Passed the point where we can replenish credit? */
 780         if (time_after_eq64(now, next_credit)) {
 781                 queue->credit_window_start = now;
 782                 tx_add_credit(queue);
 783         }
 784
 785         /* Still too big to send right now? Set a callback. */
 786         if (size > queue->remaining_credit) {
 787                 mod_timer(&queue->credit_timeout,
 788                           next_credit);
 789                 queue->credit_window_start = next_credit;
 790                 queue->rate_limited = true;
 791
 792                 return true;
 793         }
 794
 795         return false;
 796 }
 797
 798 /* No locking is required in xenvif_mcast_add/del() as they are
 799  * only ever invoked from NAPI poll. An RCU list is used because
 800  * xenvif_mcast_match() is called asynchronously, during start_xmit.
 801  */
 802
 803 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
 804 {
 805         struct xenvif_mcast_addr *mcast;
 806
 807         if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
 808                 if (net_ratelimit())
 809                         netdev_err(vif->dev,
 810                                    "Too many multicast addresses\n");
 811                 return -ENOSPC;
 812         }
 813
 814         mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
 815         if (!mcast)
 816                 return -ENOMEM;
 817
 818         ether_addr_copy(mcast->addr, addr);
 819         list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
 820         vif->fe_mcast_count++;
 821
 822         return 0;
 823 }
 824
 825 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
 826 {
 827         struct xenvif_mcast_addr *mcast;
 828
 829         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 830                 if (ether_addr_equal(addr, mcast->addr)) {
 831                         --vif->fe_mcast_count;
 832                         list_del_rcu(&mcast->entry);
 833                         kfree_rcu(mcast, rcu);
 834                         break;
 835                 }
 836         }
 837 }
 838
 839 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
 840 {
 841         struct xenvif_mcast_addr *mcast;
 842
 843         rcu_read_lock();
 844         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 845                 if (ether_addr_equal(addr, mcast->addr)) {
 846                         rcu_read_unlock();
 847                         return true;
 848                 }
 849         }
 850         rcu_read_unlock();
 851
 852         return false;
 853 }
 854
 855 void xenvif_mcast_addr_list_free(struct xenvif *vif)
 856 {
 857         /* No need for locking or RCU here. NAPI poll and TX queue
 858          * are stopped.
 859          */
 860         while (!list_empty(&vif->fe_mcast_addr)) {
 861                 struct xenvif_mcast_addr *mcast;
 862
 863                 mcast = list_first_entry(&vif->fe_mcast_addr,
 864                                          struct xenvif_mcast_addr,
 865                                          entry);
 866                 --vif->fe_mcast_count;
 867                 list_del(&mcast->entry);
 868                 kfree(mcast);
 869         }
 870 }
 871
 872 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
 873                                      int budget,
 874                                      unsigned *copy_ops,
 875                                      unsigned *map_ops)
 876 {
 877         struct sk_buff *skb, *nskb;
 878         int ret;
 879         unsigned int frag_overflow;
 880
 881         while (skb_queue_len(&queue->tx_queue) < budget) {
 882                 struct xen_netif_tx_request txreq;
 883                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
 884                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
 885                 unsigned int extra_count;
 886                 RING_IDX idx;
 887                 int work_to_do;
 888                 unsigned int data_len;
 889
 890                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
 891                     XEN_NETIF_TX_RING_SIZE) {
 892                         netdev_err(queue->vif->dev,
 893                                    "Impossible number of requests. "
 894                                    "req_prod %d, req_cons %d, size %ld\n",
 895                                    queue->tx.sring->req_prod, queue->tx.req_cons,
 896                                    XEN_NETIF_TX_RING_SIZE);
 897                         xenvif_fatal_tx_err(queue->vif);
 898                         break;
 899                 }
 900
 901                 work_to_do = XEN_RING_NR_UNCONSUMED_REQUESTS(&queue->tx);
 902                 if (!work_to_do)
 903                         break;
 904
 905                 idx = queue->tx.req_cons;
 906                 rmb(); /* Ensure that we see the request before we copy it. */
 907                 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
 908
 909                 /* Credit-based scheduling. */
 910                 if (txreq.size > queue->remaining_credit &&
 911                     tx_credit_exceeded(queue, txreq.size))
 912                         break;
 913
 914                 queue->remaining_credit -= txreq.size;
 915
 916                 work_to_do--;
 917                 queue->tx.req_cons = ++idx;
 918
 919                 memset(extras, 0, sizeof(extras));
 920                 extra_count = 0;
 921                 if (txreq.flags & XEN_NETTXF_extra_info) {
 922                         work_to_do = xenvif_get_extras(queue, extras,
 923                                                        &extra_count,
 924                                                        work_to_do);
 925                         idx = queue->tx.req_cons;
 926                         if (unlikely(work_to_do < 0))
 927                                 break;
 928                 }
 929
 930                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
 931                         struct xen_netif_extra_info *extra;
 932
 933                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
 934                         ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
 935
 936                         make_tx_response(queue, &txreq, extra_count,
 937                                          (ret == 0) ?
 938                                          XEN_NETIF_RSP_OKAY :
 939                                          XEN_NETIF_RSP_ERROR);
 940                         push_tx_responses(queue);
 941                         continue;
 942                 }
 943
 944                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
 945                         struct xen_netif_extra_info *extra;
 946
 947                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
 948                         xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
 949
 950                         make_tx_response(queue, &txreq, extra_count,
 951                                          XEN_NETIF_RSP_OKAY);
 952                         push_tx_responses(queue);
 953                         continue;
 954                 }
 955
 956                 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN) ?
 957                         XEN_NETBACK_TX_COPY_LEN : txreq.size;
 958
 959                 ret = xenvif_count_requests(queue, &txreq, extra_count,
 960                                             txfrags, work_to_do);
 961
 962                 if (unlikely(ret < 0))
 963                         break;
 964
 965                 idx += ret;
 966
 967                 if (unlikely(txreq.size < ETH_HLEN)) {
 968                         netdev_dbg(queue->vif->dev,
 969                                    "Bad packet size: %d\n", txreq.size);
 970                         xenvif_tx_err(queue, &txreq, extra_count, idx);
 971                         break;
 972                 }
 973
 974                 /* No crossing a page as the payload mustn't fragment. */
 975                 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
 976                         netdev_err(queue->vif->dev,
 977                                    "txreq.offset: %u, size: %u, end: %lu\n",
 978                                    txreq.offset, txreq.size,
 979                                    (unsigned long)(txreq.offset&~XEN_PAGE_MASK) + txreq.size);
 980                         xenvif_fatal_tx_err(queue->vif);
 981                         break;
 982                 }
 983
 984                 if (ret >= XEN_NETBK_LEGACY_SLOTS_MAX - 1 && data_len < txreq.size)
 985                         data_len = txreq.size;
 986
 987                 skb = xenvif_alloc_skb(data_len);
 988                 if (unlikely(skb == NULL)) {
 989                         netdev_dbg(queue->vif->dev,
 990                                    "Can't allocate a skb in start_xmit.\n");
 991                         xenvif_tx_err(queue, &txreq, extra_count, idx);
 992                         break;
 993                 }
 994
 995                 skb_shinfo(skb)->nr_frags = ret;
 996                 /* At this point shinfo->nr_frags is in fact the number of
 997                  * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
 998                  */
 999                 frag_overflow = 0;
1000                 nskb = NULL;
1001                 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1002                         frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1003                         BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1004                         skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1005                         nskb = xenvif_alloc_skb(0);
1006                         if (unlikely(nskb == NULL)) {
1007                                 skb_shinfo(skb)->nr_frags = 0;
1008                                 kfree_skb(skb);
1009                                 xenvif_tx_err(queue, &txreq, extra_count, idx);
1010                                 if (net_ratelimit())
1011                                         netdev_err(queue->vif->dev,
1012                                                    "Can't allocate the frag_list skb.\n");
1013                                 break;
1014                         }
1015                 }
1016
1017                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1018                         struct xen_netif_extra_info *gso;
1019                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1020
1021                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1022                                 /* Failure in xenvif_set_skb_gso is fatal. */
1023                                 skb_shinfo(skb)->nr_frags = 0;
1024                                 kfree_skb(skb);
1025                                 kfree_skb(nskb);
1026                                 break;
1027                         }
1028                 }
1029
1030                 if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
1031                         struct xen_netif_extra_info *extra;
1032                         enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
1033
1034                         extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
1035
1036                         switch (extra->u.hash.type) {
1037                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
1038                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
1039                                 type = PKT_HASH_TYPE_L3;
1040                                 break;
1041
1042                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
1043                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
1044                                 type = PKT_HASH_TYPE_L4;
1045                                 break;
1046
1047                         default:
1048                                 break;
1049                         }
1050
1051                         if (type != PKT_HASH_TYPE_NONE)
1052                                 skb_set_hash(skb,
1053                                              *(u32 *)extra->u.hash.value,
1054                                              type);
1055                 }
1056
1057                 xenvif_get_requests(queue, skb, &txreq, txfrags, copy_ops,
1058                                     map_ops, frag_overflow, nskb, extra_count,
1059                                     data_len);
1060
1061                 __skb_queue_tail(&queue->tx_queue, skb);
1062
1063                 queue->tx.req_cons = idx;
1064
1065                 if ((*map_ops >= ARRAY_SIZE(queue->tx_map_ops)) ||
1066                     (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1067                         break;
1068         }
1069
1070         return;
1071 }
1072
1073 /* Consolidate skb with a frag_list into a brand new one with local pages on
1074  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1075  */
1076 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1077 {
1078         unsigned int offset = skb_headlen(skb);
1079         skb_frag_t frags[MAX_SKB_FRAGS];
1080         int i, f;
1081         struct ubuf_info *uarg;
1082         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1083
1084         queue->stats.tx_zerocopy_sent += 2;
1085         queue->stats.tx_frag_overflow++;
1086
1087         xenvif_fill_frags(queue, nskb);
1088         /* Subtract frags size, we will correct it later */
1089         skb->truesize -= skb->data_len;
1090         skb->len += nskb->len;
1091         skb->data_len += nskb->len;
1092
1093         /* create a brand new frags array and coalesce there */
1094         for (i = 0; offset < skb->len; i++) {
1095                 struct page *page;
1096                 unsigned int len;
1097
1098                 BUG_ON(i >= MAX_SKB_FRAGS);
1099                 page = alloc_page(GFP_ATOMIC);
1100                 if (!page) {
1101                         int j;
1102                         skb->truesize += skb->data_len;
1103                         for (j = 0; j < i; j++)
1104                                 put_page(skb_frag_page(&frags[j]));
1105                         return -ENOMEM;
1106                 }
1107
1108                 if (offset + PAGE_SIZE < skb->len)
1109                         len = PAGE_SIZE;
1110                 else
1111                         len = skb->len - offset;
1112                 if (skb_copy_bits(skb, offset, page_address(page), len))
1113                         BUG();
1114
1115                 offset += len;
1116                 __skb_frag_set_page(&frags[i], page);
1117                 skb_frag_off_set(&frags[i], 0);
1118                 skb_frag_size_set(&frags[i], len);
1119         }
1120
1121         /* Release all the original (foreign) frags. */
1122         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1123                 skb_frag_unref(skb, f);
1124         uarg = skb_shinfo(skb)->destructor_arg;
1125         /* increase inflight counter to offset decrement in callback */
1126         atomic_inc(&queue->inflight_packets);
1127         uarg->callback(NULL, uarg, true);
1128         skb_shinfo(skb)->destructor_arg = NULL;
1129
1130         /* Fill the skb with the new (local) frags. */
1131         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1132         skb_shinfo(skb)->nr_frags = i;
1133         skb->truesize += i * PAGE_SIZE;
1134
1135         return 0;
1136 }
1137
1138 static int xenvif_tx_submit(struct xenvif_queue *queue)
1139 {
1140         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1141         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1142         struct sk_buff *skb;
1143         int work_done = 0;
1144
1145         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1146                 struct xen_netif_tx_request *txp;
1147                 u16 pending_idx;
1148
1149                 pending_idx = copy_pending_idx(skb, 0);
1150                 txp = &queue->pending_tx_info[pending_idx].req;
1151
1152                 /* Check the remap error code. */
1153                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1154                         /* If there was an error, xenvif_tx_check_gop is
1155                          * expected to release all the frags which were mapped,
1156                          * so kfree_skb shouldn't do it again
1157                          */
1158                         skb_shinfo(skb)->nr_frags = 0;
1159                         if (skb_has_frag_list(skb)) {
1160                                 struct sk_buff *nskb =
1161                                                 skb_shinfo(skb)->frag_list;
1162                                 skb_shinfo(nskb)->nr_frags = 0;
1163                         }
1164                         kfree_skb(skb);
1165                         continue;
1166                 }
1167
1168                 if (txp->flags & XEN_NETTXF_csum_blank)
1169                         skb->ip_summed = CHECKSUM_PARTIAL;
1170                 else if (txp->flags & XEN_NETTXF_data_validated)
1171                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1172
1173                 xenvif_fill_frags(queue, skb);
1174
1175                 if (unlikely(skb_has_frag_list(skb))) {
1176                         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1177                         xenvif_skb_zerocopy_prepare(queue, nskb);
1178                         if (xenvif_handle_frag_list(queue, skb)) {
1179                                 if (net_ratelimit())
1180                                         netdev_err(queue->vif->dev,
1181                                                    "Not enough memory to consolidate frag_list!\n");
1182                                 xenvif_skb_zerocopy_prepare(queue, skb);
1183                                 kfree_skb(skb);
1184                                 continue;
1185                         }
1186                         /* Copied all the bits from the frag list -- free it. */
1187                         skb_frag_list_init(skb);
1188                         kfree_skb(nskb);
1189                 }
1190
1191                 skb->dev      = queue->vif->dev;
1192                 skb->protocol = eth_type_trans(skb, skb->dev);
1193                 skb_reset_network_header(skb);
1194
1195                 if (checksum_setup(queue, skb)) {
1196                         netdev_dbg(queue->vif->dev,
1197                                    "Can't setup checksum in net_tx_action\n");
1198                         /* We have to set this flag to trigger the callback */
1199                         if (skb_shinfo(skb)->destructor_arg)
1200                                 xenvif_skb_zerocopy_prepare(queue, skb);
1201                         kfree_skb(skb);
1202                         continue;
1203                 }
1204
1205                 skb_probe_transport_header(skb);
1206
1207                 /* If the packet is GSO then we will have just set up the
1208                  * transport header offset in checksum_setup so it's now
1209                  * straightforward to calculate gso_segs.
1210                  */
1211                 if (skb_is_gso(skb)) {
1212                         int mss, hdrlen;
1213
1214                         /* GSO implies having the L4 header. */
1215                         WARN_ON_ONCE(!skb_transport_header_was_set(skb));
1216                         if (unlikely(!skb_transport_header_was_set(skb))) {
1217                                 kfree_skb(skb);
1218                                 continue;
1219                         }
1220
1221                         mss = skb_shinfo(skb)->gso_size;
1222                         hdrlen = skb_tcp_all_headers(skb);
1223
1224                         skb_shinfo(skb)->gso_segs =
1225                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1226                 }
1227
1228                 queue->stats.rx_bytes += skb->len;
1229                 queue->stats.rx_packets++;
1230
1231                 work_done++;
1232
1233                 /* Set this flag right before netif_receive_skb, otherwise
1234                  * someone might think this packet already left netback, and
1235                  * do a skb_copy_ubufs while we are still in control of the
1236                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1237                  */
1238                 if (skb_shinfo(skb)->destructor_arg) {
1239                         xenvif_skb_zerocopy_prepare(queue, skb);
1240                         queue->stats.tx_zerocopy_sent++;
1241                 }
1242
1243                 netif_receive_skb(skb);
1244         }
1245
1246         return work_done;
1247 }
1248
1249 void xenvif_zerocopy_callback(struct sk_buff *skb, struct ubuf_info *ubuf_base,
1250                               bool zerocopy_success)
1251 {
1252         unsigned long flags;
1253         pending_ring_idx_t index;
1254         struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
1255         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1256
1257         /* This is the only place where we grab this lock, to protect callbacks
1258          * from each other.
1259          */
1260         spin_lock_irqsave(&queue->callback_lock, flags);
1261         do {
1262                 u16 pending_idx = ubuf->desc;
1263                 ubuf = (struct ubuf_info_msgzc *) ubuf->ctx;
1264                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1265                         MAX_PENDING_REQS);
1266                 index = pending_index(queue->dealloc_prod);
1267                 queue->dealloc_ring[index] = pending_idx;
1268                 /* Sync with xenvif_tx_dealloc_action:
1269                  * insert idx then incr producer.
1270                  */
1271                 smp_wmb();
1272                 queue->dealloc_prod++;
1273         } while (ubuf);
1274         spin_unlock_irqrestore(&queue->callback_lock, flags);
1275
1276         if (likely(zerocopy_success))
1277                 queue->stats.tx_zerocopy_success++;
1278         else
1279                 queue->stats.tx_zerocopy_fail++;
1280         xenvif_skb_zerocopy_complete(queue);
1281 }
1282
1283 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1284 {
1285         struct gnttab_unmap_grant_ref *gop;
1286         pending_ring_idx_t dc, dp;
1287         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1288         unsigned int i = 0;
1289
1290         dc = queue->dealloc_cons;
1291         gop = queue->tx_unmap_ops;
1292
1293         /* Free up any grants we have finished using */
1294         do {
1295                 dp = queue->dealloc_prod;
1296
1297                 /* Ensure we see all indices enqueued by all
1298                  * xenvif_zerocopy_callback().
1299                  */
1300                 smp_rmb();
1301
1302                 while (dc != dp) {
1303                         BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1304                         pending_idx =
1305                                 queue->dealloc_ring[pending_index(dc++)];
1306
1307                         pending_idx_release[gop - queue->tx_unmap_ops] =
1308                                 pending_idx;
1309                         queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1310                                 queue->mmap_pages[pending_idx];
1311                         gnttab_set_unmap_op(gop,
1312                                             idx_to_kaddr(queue, pending_idx),
1313                                             GNTMAP_host_map,
1314                                             queue->grant_tx_handle[pending_idx]);
1315                         xenvif_grant_handle_reset(queue, pending_idx);
1316                         ++gop;
1317                 }
1318
1319         } while (dp != queue->dealloc_prod);
1320
1321         queue->dealloc_cons = dc;
1322
1323         if (gop - queue->tx_unmap_ops > 0) {
1324                 int ret;
1325                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1326                                         NULL,
1327                                         queue->pages_to_unmap,
1328                                         gop - queue->tx_unmap_ops);
1329                 if (ret) {
1330                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1331                                    gop - queue->tx_unmap_ops, ret);
1332                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1333                                 if (gop[i].status != GNTST_okay)
1334                                         netdev_err(queue->vif->dev,
1335                                                    " host_addr: 0x%llx handle: 0x%x status: %d\n",
1336                                                    gop[i].host_addr,
1337                                                    gop[i].handle,
1338                                                    gop[i].status);
1339                         }
1340                         BUG();
1341                 }
1342         }
1343
1344         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1345                 xenvif_idx_release(queue, pending_idx_release[i],
1346                                    XEN_NETIF_RSP_OKAY);
1347 }
1348
1349
1350 /* Called after netfront has transmitted */
1351 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1352 {
1353         unsigned nr_mops = 0, nr_cops = 0;
1354         int work_done, ret;
1355
1356         if (unlikely(!tx_work_todo(queue)))
1357                 return 0;
1358
1359         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1360
1361         if (nr_cops == 0)
1362                 return 0;
1363
1364         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1365         if (nr_mops != 0) {
1366                 ret = gnttab_map_refs(queue->tx_map_ops,
1367                                       NULL,
1368                                       queue->pages_to_map,
1369                                       nr_mops);
1370                 if (ret) {
1371                         unsigned int i;
1372
1373                         netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1374                                    nr_mops, ret);
1375                         for (i = 0; i < nr_mops; ++i)
1376                                 WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1377                                              GNTST_okay);
1378                 }
1379         }
1380
1381         work_done = xenvif_tx_submit(queue);
1382
1383         return work_done;
1384 }
1385
1386 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1387                                u8 status)
1388 {
1389         struct pending_tx_info *pending_tx_info;
1390         pending_ring_idx_t index;
1391         unsigned long flags;
1392
1393         pending_tx_info = &queue->pending_tx_info[pending_idx];
1394
1395         spin_lock_irqsave(&queue->response_lock, flags);
1396
1397         make_tx_response(queue, &pending_tx_info->req,
1398                          pending_tx_info->extra_count, status);
1399
1400         /* Release the pending index before pusing the Tx response so
1401          * its available before a new Tx request is pushed by the
1402          * frontend.
1403          */
1404         index = pending_index(queue->pending_prod++);
1405         queue->pending_ring[index] = pending_idx;
1406
1407         push_tx_responses(queue);
1408
1409         spin_unlock_irqrestore(&queue->response_lock, flags);
1410 }
1411
1412
1413 static void make_tx_response(struct xenvif_queue *queue,
1414                              struct xen_netif_tx_request *txp,
1415                              unsigned int extra_count,
1416                              s8       st)
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 = st;
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_unmap(struct xenvif_queue *queue, u16 pending_idx)
1441 {
1442         int ret;
1443         struct gnttab_unmap_grant_ref tx_unmap_op;
1444
1445         gnttab_set_unmap_op(&tx_unmap_op,
1446                             idx_to_kaddr(queue, pending_idx),
1447                             GNTMAP_host_map,
1448                             queue->grant_tx_handle[pending_idx]);
1449         xenvif_grant_handle_reset(queue, pending_idx);
1450
1451         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1452                                 &queue->mmap_pages[pending_idx], 1);
1453         if (ret) {
1454                 netdev_err(queue->vif->dev,
1455                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1456                            ret,
1457                            pending_idx,
1458                            tx_unmap_op.host_addr,
1459                            tx_unmap_op.handle,
1460                            tx_unmap_op.status);
1461                 BUG();
1462         }
1463 }
1464
1465 static inline int tx_work_todo(struct xenvif_queue *queue)
1466 {
1467         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1468                 return 1;
1469
1470         return 0;
1471 }
1472
1473 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1474 {
1475         return queue->dealloc_cons != queue->dealloc_prod;
1476 }
1477
1478 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1479 {
1480         if (queue->tx.sring)
1481                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1482                                         queue->tx.sring);
1483         if (queue->rx.sring)
1484                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1485                                         queue->rx.sring);
1486 }
1487
1488 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1489                                    grant_ref_t tx_ring_ref,
1490                                    grant_ref_t rx_ring_ref)
1491 {
1492         void *addr;
1493         struct xen_netif_tx_sring *txs;
1494         struct xen_netif_rx_sring *rxs;
1495         RING_IDX rsp_prod, req_prod;
1496         int err;
1497
1498         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1499                                      &tx_ring_ref, 1, &addr);
1500         if (err)
1501                 goto err;
1502
1503         txs = (struct xen_netif_tx_sring *)addr;
1504         rsp_prod = READ_ONCE(txs->rsp_prod);
1505         req_prod = READ_ONCE(txs->req_prod);
1506
1507         BACK_RING_ATTACH(&queue->tx, txs, rsp_prod, XEN_PAGE_SIZE);
1508
1509         err = -EIO;
1510         if (req_prod - rsp_prod > RING_SIZE(&queue->tx))
1511                 goto err;
1512
1513         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1514                                      &rx_ring_ref, 1, &addr);
1515         if (err)
1516                 goto err;
1517
1518         rxs = (struct xen_netif_rx_sring *)addr;
1519         rsp_prod = READ_ONCE(rxs->rsp_prod);
1520         req_prod = READ_ONCE(rxs->req_prod);
1521
1522         BACK_RING_ATTACH(&queue->rx, rxs, rsp_prod, XEN_PAGE_SIZE);
1523
1524         err = -EIO;
1525         if (req_prod - rsp_prod > RING_SIZE(&queue->rx))
1526                 goto err;
1527
1528         return 0;
1529
1530 err:
1531         xenvif_unmap_frontend_data_rings(queue);
1532         return err;
1533 }
1534
1535 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1536 {
1537         /* Dealloc thread must remain running until all inflight
1538          * packets complete.
1539          */
1540         return kthread_should_stop() &&
1541                 !atomic_read(&queue->inflight_packets);
1542 }
1543
1544 int xenvif_dealloc_kthread(void *data)
1545 {
1546         struct xenvif_queue *queue = data;
1547
1548         for (;;) {
1549                 wait_event_interruptible(queue->dealloc_wq,
1550                                          tx_dealloc_work_todo(queue) ||
1551                                          xenvif_dealloc_kthread_should_stop(queue));
1552                 if (xenvif_dealloc_kthread_should_stop(queue))
1553                         break;
1554
1555                 xenvif_tx_dealloc_action(queue);
1556                 cond_resched();
1557         }
1558
1559         /* Unmap anything remaining*/
1560         if (tx_dealloc_work_todo(queue))
1561                 xenvif_tx_dealloc_action(queue);
1562
1563         return 0;
1564 }
1565
1566 static void make_ctrl_response(struct xenvif *vif,
1567                                const struct xen_netif_ctrl_request *req,
1568                                u32 status, u32 data)
1569 {
1570         RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1571         struct xen_netif_ctrl_response rsp = {
1572                 .id = req->id,
1573                 .type = req->type,
1574                 .status = status,
1575                 .data = data,
1576         };
1577
1578         *RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1579         vif->ctrl.rsp_prod_pvt = ++idx;
1580 }
1581
1582 static void push_ctrl_response(struct xenvif *vif)
1583 {
1584         int notify;
1585
1586         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1587         if (notify)
1588                 notify_remote_via_irq(vif->ctrl_irq);
1589 }
1590
1591 static void process_ctrl_request(struct xenvif *vif,
1592                                  const struct xen_netif_ctrl_request *req)
1593 {
1594         u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1595         u32 data = 0;
1596
1597         switch (req->type) {
1598         case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1599                 status = xenvif_set_hash_alg(vif, req->data[0]);
1600                 break;
1601
1602         case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1603                 status = xenvif_get_hash_flags(vif, &data);
1604                 break;
1605
1606         case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1607                 status = xenvif_set_hash_flags(vif, req->data[0]);
1608                 break;
1609
1610         case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1611                 status = xenvif_set_hash_key(vif, req->data[0],
1612                                              req->data[1]);
1613                 break;
1614
1615         case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1616                 status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1617                 data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1618                 break;
1619
1620         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1621                 status = xenvif_set_hash_mapping_size(vif,
1622                                                       req->data[0]);
1623                 break;
1624
1625         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1626                 status = xenvif_set_hash_mapping(vif, req->data[0],
1627                                                  req->data[1],
1628                                                  req->data[2]);
1629                 break;
1630
1631         default:
1632                 break;
1633         }
1634
1635         make_ctrl_response(vif, req, status, data);
1636         push_ctrl_response(vif);
1637 }
1638
1639 static void xenvif_ctrl_action(struct xenvif *vif)
1640 {
1641         for (;;) {
1642                 RING_IDX req_prod, req_cons;
1643
1644                 req_prod = vif->ctrl.sring->req_prod;
1645                 req_cons = vif->ctrl.req_cons;
1646
1647                 /* Make sure we can see requests before we process them. */
1648                 rmb();
1649
1650                 if (req_cons == req_prod)
1651                         break;
1652
1653                 while (req_cons != req_prod) {
1654                         struct xen_netif_ctrl_request req;
1655
1656                         RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1657                         req_cons++;
1658
1659                         process_ctrl_request(vif, &req);
1660                 }
1661
1662                 vif->ctrl.req_cons = req_cons;
1663                 vif->ctrl.sring->req_event = req_cons + 1;
1664         }
1665 }
1666
1667 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1668 {
1669         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1670                 return true;
1671
1672         return false;
1673 }
1674
1675 irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1676 {
1677         struct xenvif *vif = data;
1678         unsigned int eoi_flag = XEN_EOI_FLAG_SPURIOUS;
1679
1680         while (xenvif_ctrl_work_todo(vif)) {
1681                 xenvif_ctrl_action(vif);
1682                 eoi_flag = 0;
1683         }
1684
1685         xen_irq_lateeoi(irq, eoi_flag);
1686
1687         return IRQ_HANDLED;
1688 }
1689
1690 static int __init netback_init(void)
1691 {
1692         int rc = 0;
1693
1694         if (!xen_domain())
1695                 return -ENODEV;
1696
1697         /* Allow as many queues as there are CPUs but max. 8 if user has not
1698          * specified a value.
1699          */
1700         if (xenvif_max_queues == 0)
1701                 xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1702                                           num_online_cpus());
1703
1704         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1705                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1706                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1707                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1708         }
1709
1710         rc = xenvif_xenbus_init();
1711         if (rc)
1712                 goto failed_init;
1713
1714 #ifdef CONFIG_DEBUG_FS
1715         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1716 #endif /* CONFIG_DEBUG_FS */
1717
1718         return 0;
1719
1720 failed_init:
1721         return rc;
1722 }
1723
1724 module_init(netback_init);
1725
1726 static void __exit netback_fini(void)
1727 {
1728 #ifdef CONFIG_DEBUG_FS
1729         debugfs_remove_recursive(xen_netback_dbg_root);
1730 #endif /* CONFIG_DEBUG_FS */
1731         xenvif_xenbus_fini();
1732 }
1733 module_exit(netback_fini);
1734
1735 MODULE_LICENSE("Dual BSD/GPL");
1736 MODULE_ALIAS("xen-backend:vif");