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                                s8 status);
 108
 109 static void make_tx_response(struct xenvif_queue *queue,
 110                              const struct xen_netif_tx_request *txp,
 111                              unsigned int extra_count,
 112                              s8 status);
 113
 114 static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx);
 115
 116 static inline int tx_work_todo(struct xenvif_queue *queue);
 117
 118 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
 119                                        u16 idx)
 120 {
 121         return page_to_pfn(queue->mmap_pages[idx]);
 122 }
 123
 124 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
 125                                          u16 idx)
 126 {
 127         return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
 128 }
 129
 130 #define callback_param(vif, pending_idx) \
 131         (vif->pending_tx_info[pending_idx].callback_struct)
 132
 133 /* Find the containing VIF's structure from a pointer in pending_tx_info array
 134  */
 135 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info_msgzc *ubuf)
 136 {
 137         u16 pending_idx = ubuf->desc;
 138         struct pending_tx_info *temp =
 139                 container_of(ubuf, struct pending_tx_info, callback_struct);
 140         return container_of(temp - pending_idx,
 141                             struct xenvif_queue,
 142                             pending_tx_info[0]);
 143 }
 144
 145 static u16 frag_get_pending_idx(skb_frag_t *frag)
 146 {
 147         return (u16)skb_frag_off(frag);
 148 }
 149
 150 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 151 {
 152         skb_frag_off_set(frag, pending_idx);
 153 }
 154
 155 static inline pending_ring_idx_t pending_index(unsigned i)
 156 {
 157         return i & (MAX_PENDING_REQS-1);
 158 }
 159
 160 void xenvif_kick_thread(struct xenvif_queue *queue)
 161 {
 162         wake_up(&queue->wq);
 163 }
 164
 165 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 166 {
 167         int more_to_do;
 168
 169         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 170
 171         if (more_to_do)
 172                 napi_schedule(&queue->napi);
 173         else if (atomic_fetch_andnot(NETBK_TX_EOI | NETBK_COMMON_EOI,
 174                                      &queue->eoi_pending) &
 175                  (NETBK_TX_EOI | NETBK_COMMON_EOI))
 176                 xen_irq_lateeoi(queue->tx_irq, 0);
 177 }
 178
 179 static void tx_add_credit(struct xenvif_queue *queue)
 180 {
 181         unsigned long max_burst, max_credit;
 182
 183         /*
 184          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 185          * Otherwise the interface can seize up due to insufficient credit.
 186          */
 187         max_burst = max(131072UL, queue->credit_bytes);
 188
 189         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 190         max_credit = queue->remaining_credit + queue->credit_bytes;
 191         if (max_credit < queue->remaining_credit)
 192                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 193
 194         queue->remaining_credit = min(max_credit, max_burst);
 195         queue->rate_limited = false;
 196 }
 197
 198 void xenvif_tx_credit_callback(struct timer_list *t)
 199 {
 200         struct xenvif_queue *queue = from_timer(queue, t, credit_timeout);
 201         tx_add_credit(queue);
 202         xenvif_napi_schedule_or_enable_events(queue);
 203 }
 204
 205 static void xenvif_tx_err(struct xenvif_queue *queue,
 206                           struct xen_netif_tx_request *txp,
 207                           unsigned int extra_count, RING_IDX end)
 208 {
 209         RING_IDX cons = queue->tx.req_cons;
 210
 211         do {
 212                 make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
 213                 if (cons == end)
 214                         break;
 215                 RING_COPY_REQUEST(&queue->tx, cons++, txp);
 216                 extra_count = 0; /* only the first frag can have extras */
 217         } while (1);
 218         queue->tx.req_cons = cons;
 219 }
 220
 221 static void xenvif_fatal_tx_err(struct xenvif *vif)
 222 {
 223         netdev_err(vif->dev, "fatal error; disabling device\n");
 224         vif->disabled = true;
 225         /* Disable the vif from queue 0's kthread */
 226         if (vif->num_queues)
 227                 xenvif_kick_thread(&vif->queues[0]);
 228 }
 229
 230 static int xenvif_count_requests(struct xenvif_queue *queue,
 231                                  struct xen_netif_tx_request *first,
 232                                  unsigned int extra_count,
 233                                  struct xen_netif_tx_request *txp,
 234                                  int work_to_do)
 235 {
 236         RING_IDX cons = queue->tx.req_cons;
 237         int slots = 0;
 238         int drop_err = 0;
 239         int more_data;
 240
 241         if (!(first->flags & XEN_NETTXF_more_data))
 242                 return 0;
 243
 244         do {
 245                 struct xen_netif_tx_request dropped_tx = { 0 };
 246
 247                 if (slots >= work_to_do) {
 248                         netdev_err(queue->vif->dev,
 249                                    "Asked for %d slots but exceeds this limit\n",
 250                                    work_to_do);
 251                         xenvif_fatal_tx_err(queue->vif);
 252                         return -ENODATA;
 253                 }
 254
 255                 /* This guest is really using too many slots and
 256                  * considered malicious.
 257                  */
 258                 if (unlikely(slots >= fatal_skb_slots)) {
 259                         netdev_err(queue->vif->dev,
 260                                    "Malicious frontend using %d slots, threshold %u\n",
 261                                    slots, fatal_skb_slots);
 262                         xenvif_fatal_tx_err(queue->vif);
 263                         return -E2BIG;
 264                 }
 265
 266                 /* Xen network protocol had implicit dependency on
 267                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 268                  * the historical MAX_SKB_FRAGS value 18 to honor the
 269                  * same behavior as before. Any packet using more than
 270                  * 18 slots but less than fatal_skb_slots slots is
 271                  * dropped
 272                  */
 273                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 274                         if (net_ratelimit())
 275                                 netdev_dbg(queue->vif->dev,
 276                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 277                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 278                         drop_err = -E2BIG;
 279                 }
 280
 281                 if (drop_err)
 282                         txp = &dropped_tx;
 283
 284                 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
 285
 286                 /* If the guest submitted a frame >= 64 KiB then
 287                  * first->size overflowed and following slots will
 288                  * appear to be larger than the frame.
 289                  *
 290                  * This cannot be fatal error as there are buggy
 291                  * frontends that do this.
 292                  *
 293                  * Consume all slots and drop the packet.
 294                  */
 295                 if (!drop_err && txp->size > first->size) {
 296                         if (net_ratelimit())
 297                                 netdev_dbg(queue->vif->dev,
 298                                            "Invalid tx request, slot size %u > remaining size %u\n",
 299                                            txp->size, first->size);
 300                         drop_err = -EIO;
 301                 }
 302
 303                 first->size -= txp->size;
 304                 slots++;
 305
 306                 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
 307                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 308                                  txp->offset, txp->size);
 309                         xenvif_fatal_tx_err(queue->vif);
 310                         return -EINVAL;
 311                 }
 312
 313                 more_data = txp->flags & XEN_NETTXF_more_data;
 314
 315                 if (!drop_err)
 316                         txp++;
 317
 318         } while (more_data);
 319
 320         if (drop_err) {
 321                 xenvif_tx_err(queue, first, extra_count, cons + slots);
 322                 return drop_err;
 323         }
 324
 325         return slots;
 326 }
 327
 328
 329 struct xenvif_tx_cb {
 330         u16 copy_pending_idx[XEN_NETBK_LEGACY_SLOTS_MAX + 1];
 331         u8 copy_count;
 332         u32 split_mask;
 333 };
 334
 335 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 336 #define copy_pending_idx(skb, i) (XENVIF_TX_CB(skb)->copy_pending_idx[i])
 337 #define copy_count(skb) (XENVIF_TX_CB(skb)->copy_count)
 338
 339 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 340                                            u16 pending_idx,
 341                                            struct xen_netif_tx_request *txp,
 342                                            unsigned int extra_count,
 343                                            struct gnttab_map_grant_ref *mop)
 344 {
 345         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 346         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 347                           GNTMAP_host_map | GNTMAP_readonly,
 348                           txp->gref, queue->vif->domid);
 349
 350         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 351                sizeof(*txp));
 352         queue->pending_tx_info[pending_idx].extra_count = extra_count;
 353 }
 354
 355 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 356 {
 357         struct sk_buff *skb =
 358                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 359                           GFP_ATOMIC | __GFP_NOWARN);
 360
 361         BUILD_BUG_ON(sizeof(*XENVIF_TX_CB(skb)) > sizeof(skb->cb));
 362         if (unlikely(skb == NULL))
 363                 return NULL;
 364
 365         /* Packets passed to netif_rx() must have some headroom. */
 366         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 367
 368         /* Initialize it here to avoid later surprises */
 369         skb_shinfo(skb)->destructor_arg = NULL;
 370
 371         return skb;
 372 }
 373
 374 static void xenvif_get_requests(struct xenvif_queue *queue,
 375                                 struct sk_buff *skb,
 376                                 struct xen_netif_tx_request *first,
 377                                 struct xen_netif_tx_request *txfrags,
 378                                 unsigned *copy_ops,
 379                                 unsigned *map_ops,
 380                                 unsigned int frag_overflow,
 381                                 struct sk_buff *nskb,
 382                                 unsigned int extra_count,
 383                                 unsigned int data_len)
 384 {
 385         struct skb_shared_info *shinfo = skb_shinfo(skb);
 386         skb_frag_t *frags = shinfo->frags;
 387         u16 pending_idx;
 388         pending_ring_idx_t index;
 389         unsigned int nr_slots;
 390         struct gnttab_copy *cop = queue->tx_copy_ops + *copy_ops;
 391         struct gnttab_map_grant_ref *gop = queue->tx_map_ops + *map_ops;
 392         struct xen_netif_tx_request *txp = first;
 393
 394         nr_slots = shinfo->nr_frags + frag_overflow + 1;
 395
 396         copy_count(skb) = 0;
 397         XENVIF_TX_CB(skb)->split_mask = 0;
 398
 399         /* Create copy ops for exactly data_len bytes into the skb head. */
 400         __skb_put(skb, data_len);
 401         while (data_len > 0) {
 402                 int amount = data_len > txp->size ? txp->size : data_len;
 403                 bool split = false;
 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                 /* Don't cross local page boundary! */
 417                 if (cop->dest.offset + amount > XEN_PAGE_SIZE) {
 418                         amount = XEN_PAGE_SIZE - cop->dest.offset;
 419                         XENVIF_TX_CB(skb)->split_mask |= 1U << copy_count(skb);
 420                         split = true;
 421                 }
 422
 423                 cop->len = amount;
 424                 cop->flags = GNTCOPY_source_gref;
 425
 426                 index = pending_index(queue->pending_cons);
 427                 pending_idx = queue->pending_ring[index];
 428                 callback_param(queue, pending_idx).ctx = NULL;
 429                 copy_pending_idx(skb, copy_count(skb)) = pending_idx;
 430                 if (!split)
 431                         copy_count(skb)++;
 432
 433                 cop++;
 434                 data_len -= amount;
 435
 436                 if (amount == txp->size) {
 437                         /* The copy op covered the full tx_request */
 438
 439                         memcpy(&queue->pending_tx_info[pending_idx].req,
 440                                txp, sizeof(*txp));
 441                         queue->pending_tx_info[pending_idx].extra_count =
 442                                 (txp == first) ? extra_count : 0;
 443
 444                         if (txp == first)
 445                                 txp = txfrags;
 446                         else
 447                                 txp++;
 448                         queue->pending_cons++;
 449                         nr_slots--;
 450                 } else {
 451                         /* The copy op partially covered the tx_request.
 452                          * The remainder will be mapped or copied in the next
 453                          * iteration.
 454                          */
 455                         txp->offset += amount;
 456                         txp->size -= amount;
 457                 }
 458         }
 459
 460         for (shinfo->nr_frags = 0; nr_slots > 0 && shinfo->nr_frags < MAX_SKB_FRAGS;
 461              nr_slots--) {
 462                 if (unlikely(!txp->size)) {
 463                         make_tx_response(queue, txp, 0, XEN_NETIF_RSP_OKAY);
 464                         ++txp;
 465                         continue;
 466                 }
 467
 468                 index = pending_index(queue->pending_cons++);
 469                 pending_idx = queue->pending_ring[index];
 470                 xenvif_tx_create_map_op(queue, pending_idx, txp,
 471                                         txp == first ? extra_count : 0, gop);
 472                 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 473                 ++shinfo->nr_frags;
 474                 ++gop;
 475
 476                 if (txp == first)
 477                         txp = txfrags;
 478                 else
 479                         txp++;
 480         }
 481
 482         if (nr_slots > 0) {
 483
 484                 shinfo = skb_shinfo(nskb);
 485                 frags = shinfo->frags;
 486
 487                 for (shinfo->nr_frags = 0; shinfo->nr_frags < nr_slots; ++txp) {
 488                         if (unlikely(!txp->size)) {
 489                                 make_tx_response(queue, txp, 0,
 490                                                  XEN_NETIF_RSP_OKAY);
 491                                 continue;
 492                         }
 493
 494                         index = pending_index(queue->pending_cons++);
 495                         pending_idx = queue->pending_ring[index];
 496                         xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
 497                                                 gop);
 498                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 499                                              pending_idx);
 500                         ++shinfo->nr_frags;
 501                         ++gop;
 502                 }
 503
 504                 if (shinfo->nr_frags) {
 505                         skb_shinfo(skb)->frag_list = nskb;
 506                         nskb = NULL;
 507                 }
 508         }
 509
 510         if (nskb) {
 511                 /* A frag_list skb was allocated but it is no longer needed
 512                  * because enough slots were converted to copy ops above or some
 513                  * were empty.
 514                  */
 515                 kfree_skb(nskb);
 516         }
 517
 518         (*copy_ops) = cop - queue->tx_copy_ops;
 519         (*map_ops) = gop - queue->tx_map_ops;
 520 }
 521
 522 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 523                                            u16 pending_idx,
 524                                            grant_handle_t handle)
 525 {
 526         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 527                      NETBACK_INVALID_HANDLE)) {
 528                 netdev_err(queue->vif->dev,
 529                            "Trying to overwrite active handle! pending_idx: 0x%x\n",
 530                            pending_idx);
 531                 BUG();
 532         }
 533         queue->grant_tx_handle[pending_idx] = handle;
 534 }
 535
 536 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 537                                              u16 pending_idx)
 538 {
 539         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 540                      NETBACK_INVALID_HANDLE)) {
 541                 netdev_err(queue->vif->dev,
 542                            "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 543                            pending_idx);
 544                 BUG();
 545         }
 546         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 547 }
 548
 549 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 550                                struct sk_buff *skb,
 551                                struct gnttab_map_grant_ref **gopp_map,
 552                                struct gnttab_copy **gopp_copy)
 553 {
 554         struct gnttab_map_grant_ref *gop_map = *gopp_map;
 555         u16 pending_idx;
 556         /* This always points to the shinfo of the skb being checked, which
 557          * could be either the first or the one on the frag_list
 558          */
 559         struct skb_shared_info *shinfo = skb_shinfo(skb);
 560         /* If this is non-NULL, we are currently checking the frag_list skb, and
 561          * this points to the shinfo of the first one
 562          */
 563         struct skb_shared_info *first_shinfo = NULL;
 564         int nr_frags = shinfo->nr_frags;
 565         const bool sharedslot = nr_frags &&
 566                                 frag_get_pending_idx(&shinfo->frags[0]) ==
 567                                     copy_pending_idx(skb, copy_count(skb) - 1);
 568         int i, err = 0;
 569
 570         for (i = 0; i < copy_count(skb); i++) {
 571                 int newerr;
 572
 573                 /* Check status of header. */
 574                 pending_idx = copy_pending_idx(skb, i);
 575
 576                 newerr = (*gopp_copy)->status;
 577
 578                 /* Split copies need to be handled together. */
 579                 if (XENVIF_TX_CB(skb)->split_mask & (1U << i)) {
 580                         (*gopp_copy)++;
 581                         if (!newerr)
 582                                 newerr = (*gopp_copy)->status;
 583                 }
 584                 if (likely(!newerr)) {
 585                         /* The first frag might still have this slot mapped */
 586                         if (i < copy_count(skb) - 1 || !sharedslot)
 587                                 xenvif_idx_release(queue, pending_idx,
 588                                                    XEN_NETIF_RSP_OKAY);
 589                 } else {
 590                         err = newerr;
 591                         if (net_ratelimit())
 592                                 netdev_dbg(queue->vif->dev,
 593                                            "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 594                                            (*gopp_copy)->status,
 595                                            pending_idx,
 596                                            (*gopp_copy)->source.u.ref);
 597                         /* The first frag might still have this slot mapped */
 598                         if (i < copy_count(skb) - 1 || !sharedslot)
 599                                 xenvif_idx_release(queue, pending_idx,
 600                                                    XEN_NETIF_RSP_ERROR);
 601                 }
 602                 (*gopp_copy)++;
 603         }
 604
 605 check_frags:
 606         for (i = 0; i < nr_frags; i++, gop_map++) {
 607                 int j, newerr;
 608
 609                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 610
 611                 /* Check error status: if okay then remember grant handle. */
 612                 newerr = gop_map->status;
 613
 614                 if (likely(!newerr)) {
 615                         xenvif_grant_handle_set(queue,
 616                                                 pending_idx,
 617                                                 gop_map->handle);
 618                         /* Had a previous error? Invalidate this fragment. */
 619                         if (unlikely(err)) {
 620                                 xenvif_idx_unmap(queue, pending_idx);
 621                                 /* If the mapping of the first frag was OK, but
 622                                  * the header's copy failed, and they are
 623                                  * sharing a slot, send an error
 624                                  */
 625                                 if (i == 0 && !first_shinfo && sharedslot)
 626                                         xenvif_idx_release(queue, pending_idx,
 627                                                            XEN_NETIF_RSP_ERROR);
 628                                 else
 629                                         xenvif_idx_release(queue, pending_idx,
 630                                                            XEN_NETIF_RSP_OKAY);
 631                         }
 632                         continue;
 633                 }
 634
 635                 /* Error on this fragment: respond to client with an error. */
 636                 if (net_ratelimit())
 637                         netdev_dbg(queue->vif->dev,
 638                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
 639                                    i,
 640                                    gop_map->status,
 641                                    pending_idx,
 642                                    gop_map->ref);
 643
 644                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
 645
 646                 /* Not the first error? Preceding frags already invalidated. */
 647                 if (err)
 648                         continue;
 649
 650                 /* Invalidate preceding fragments of this skb. */
 651                 for (j = 0; j < i; j++) {
 652                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
 653                         xenvif_idx_unmap(queue, pending_idx);
 654                         xenvif_idx_release(queue, pending_idx,
 655                                            XEN_NETIF_RSP_OKAY);
 656                 }
 657
 658                 /* And if we found the error while checking the frag_list, unmap
 659                  * the first skb's frags
 660                  */
 661                 if (first_shinfo) {
 662                         for (j = 0; j < first_shinfo->nr_frags; j++) {
 663                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
 664                                 xenvif_idx_unmap(queue, pending_idx);
 665                                 xenvif_idx_release(queue, pending_idx,
 666                                                    XEN_NETIF_RSP_OKAY);
 667                         }
 668                 }
 669
 670                 /* Remember the error: invalidate all subsequent fragments. */
 671                 err = newerr;
 672         }
 673
 674         if (skb_has_frag_list(skb) && !first_shinfo) {
 675                 first_shinfo = shinfo;
 676                 shinfo = skb_shinfo(shinfo->frag_list);
 677                 nr_frags = shinfo->nr_frags;
 678
 679                 goto check_frags;
 680         }
 681
 682         *gopp_map = gop_map;
 683         return err;
 684 }
 685
 686 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
 687 {
 688         struct skb_shared_info *shinfo = skb_shinfo(skb);
 689         int nr_frags = shinfo->nr_frags;
 690         int i;
 691         u16 prev_pending_idx = INVALID_PENDING_IDX;
 692
 693         for (i = 0; i < nr_frags; i++) {
 694                 skb_frag_t *frag = shinfo->frags + i;
 695                 struct xen_netif_tx_request *txp;
 696                 struct page *page;
 697                 u16 pending_idx;
 698
 699                 pending_idx = frag_get_pending_idx(frag);
 700
 701                 /* If this is not the first frag, chain it to the previous*/
 702                 if (prev_pending_idx == INVALID_PENDING_IDX)
 703                         skb_shinfo(skb)->destructor_arg =
 704                                 &callback_param(queue, pending_idx);
 705                 else
 706                         callback_param(queue, prev_pending_idx).ctx =
 707                                 &callback_param(queue, pending_idx);
 708
 709                 callback_param(queue, pending_idx).ctx = NULL;
 710                 prev_pending_idx = pending_idx;
 711
 712                 txp = &queue->pending_tx_info[pending_idx].req;
 713                 page = virt_to_page((void *)idx_to_kaddr(queue, pending_idx));
 714                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
 715                 skb->len += txp->size;
 716                 skb->data_len += txp->size;
 717                 skb->truesize += txp->size;
 718
 719                 /* Take an extra reference to offset network stack's put_page */
 720                 get_page(queue->mmap_pages[pending_idx]);
 721         }
 722 }
 723
 724 static int xenvif_get_extras(struct xenvif_queue *queue,
 725                              struct xen_netif_extra_info *extras,
 726                              unsigned int *extra_count,
 727                              int work_to_do)
 728 {
 729         struct xen_netif_extra_info extra;
 730         RING_IDX cons = queue->tx.req_cons;
 731
 732         do {
 733                 if (unlikely(work_to_do-- <= 0)) {
 734                         netdev_err(queue->vif->dev, "Missing extra info\n");
 735                         xenvif_fatal_tx_err(queue->vif);
 736                         return -EBADR;
 737                 }
 738
 739                 RING_COPY_REQUEST(&queue->tx, cons, &extra);
 740
 741                 queue->tx.req_cons = ++cons;
 742                 (*extra_count)++;
 743
 744                 if (unlikely(!extra.type ||
 745                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
 746                         netdev_err(queue->vif->dev,
 747                                    "Invalid extra type: %d\n", extra.type);
 748                         xenvif_fatal_tx_err(queue->vif);
 749                         return -EINVAL;
 750                 }
 751
 752                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
 753         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
 754
 755         return work_to_do;
 756 }
 757
 758 static int xenvif_set_skb_gso(struct xenvif *vif,
 759                               struct sk_buff *skb,
 760                               struct xen_netif_extra_info *gso)
 761 {
 762         if (!gso->u.gso.size) {
 763                 netdev_err(vif->dev, "GSO size must not be zero.\n");
 764                 xenvif_fatal_tx_err(vif);
 765                 return -EINVAL;
 766         }
 767
 768         switch (gso->u.gso.type) {
 769         case XEN_NETIF_GSO_TYPE_TCPV4:
 770                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 771                 break;
 772         case XEN_NETIF_GSO_TYPE_TCPV6:
 773                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 774                 break;
 775         default:
 776                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
 777                 xenvif_fatal_tx_err(vif);
 778                 return -EINVAL;
 779         }
 780
 781         skb_shinfo(skb)->gso_size = gso->u.gso.size;
 782         /* gso_segs will be calculated later */
 783
 784         return 0;
 785 }
 786
 787 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
 788 {
 789         bool recalculate_partial_csum = false;
 790
 791         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
 792          * peers can fail to set NETRXF_csum_blank when sending a GSO
 793          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
 794          * recalculate the partial checksum.
 795          */
 796         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
 797                 queue->stats.rx_gso_checksum_fixup++;
 798                 skb->ip_summed = CHECKSUM_PARTIAL;
 799                 recalculate_partial_csum = true;
 800         }
 801
 802         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
 803         if (skb->ip_summed != CHECKSUM_PARTIAL)
 804                 return 0;
 805
 806         return skb_checksum_setup(skb, recalculate_partial_csum);
 807 }
 808
 809 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
 810 {
 811         u64 now = get_jiffies_64();
 812         u64 next_credit = queue->credit_window_start +
 813                 msecs_to_jiffies(queue->credit_usec / 1000);
 814
 815         /* Timer could already be pending in rare cases. */
 816         if (timer_pending(&queue->credit_timeout)) {
 817                 queue->rate_limited = true;
 818                 return true;
 819         }
 820
 821         /* Passed the point where we can replenish credit? */
 822         if (time_after_eq64(now, next_credit)) {
 823                 queue->credit_window_start = now;
 824                 tx_add_credit(queue);
 825         }
 826
 827         /* Still too big to send right now? Set a callback. */
 828         if (size > queue->remaining_credit) {
 829                 mod_timer(&queue->credit_timeout,
 830                           next_credit);
 831                 queue->credit_window_start = next_credit;
 832                 queue->rate_limited = true;
 833
 834                 return true;
 835         }
 836
 837         return false;
 838 }
 839
 840 /* No locking is required in xenvif_mcast_add/del() as they are
 841  * only ever invoked from NAPI poll. An RCU list is used because
 842  * xenvif_mcast_match() is called asynchronously, during start_xmit.
 843  */
 844
 845 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
 846 {
 847         struct xenvif_mcast_addr *mcast;
 848
 849         if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
 850                 if (net_ratelimit())
 851                         netdev_err(vif->dev,
 852                                    "Too many multicast addresses\n");
 853                 return -ENOSPC;
 854         }
 855
 856         mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
 857         if (!mcast)
 858                 return -ENOMEM;
 859
 860         ether_addr_copy(mcast->addr, addr);
 861         list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
 862         vif->fe_mcast_count++;
 863
 864         return 0;
 865 }
 866
 867 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
 868 {
 869         struct xenvif_mcast_addr *mcast;
 870
 871         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 872                 if (ether_addr_equal(addr, mcast->addr)) {
 873                         --vif->fe_mcast_count;
 874                         list_del_rcu(&mcast->entry);
 875                         kfree_rcu(mcast, rcu);
 876                         break;
 877                 }
 878         }
 879 }
 880
 881 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
 882 {
 883         struct xenvif_mcast_addr *mcast;
 884
 885         rcu_read_lock();
 886         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 887                 if (ether_addr_equal(addr, mcast->addr)) {
 888                         rcu_read_unlock();
 889                         return true;
 890                 }
 891         }
 892         rcu_read_unlock();
 893
 894         return false;
 895 }
 896
 897 void xenvif_mcast_addr_list_free(struct xenvif *vif)
 898 {
 899         /* No need for locking or RCU here. NAPI poll and TX queue
 900          * are stopped.
 901          */
 902         while (!list_empty(&vif->fe_mcast_addr)) {
 903                 struct xenvif_mcast_addr *mcast;
 904
 905                 mcast = list_first_entry(&vif->fe_mcast_addr,
 906                                          struct xenvif_mcast_addr,
 907                                          entry);
 908                 --vif->fe_mcast_count;
 909                 list_del(&mcast->entry);
 910                 kfree(mcast);
 911         }
 912 }
 913
 914 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
 915                                      int budget,
 916                                      unsigned *copy_ops,
 917                                      unsigned *map_ops)
 918 {
 919         struct sk_buff *skb, *nskb;
 920         int ret;
 921         unsigned int frag_overflow;
 922
 923         while (skb_queue_len(&queue->tx_queue) < budget) {
 924                 struct xen_netif_tx_request txreq;
 925                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
 926                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
 927                 unsigned int extra_count;
 928                 RING_IDX idx;
 929                 int work_to_do;
 930                 unsigned int data_len;
 931
 932                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
 933                     XEN_NETIF_TX_RING_SIZE) {
 934                         netdev_err(queue->vif->dev,
 935                                    "Impossible number of requests. "
 936                                    "req_prod %d, req_cons %d, size %ld\n",
 937                                    queue->tx.sring->req_prod, queue->tx.req_cons,
 938                                    XEN_NETIF_TX_RING_SIZE);
 939                         xenvif_fatal_tx_err(queue->vif);
 940                         break;
 941                 }
 942
 943                 work_to_do = XEN_RING_NR_UNCONSUMED_REQUESTS(&queue->tx);
 944                 if (!work_to_do)
 945                         break;
 946
 947                 idx = queue->tx.req_cons;
 948                 rmb(); /* Ensure that we see the request before we copy it. */
 949                 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
 950
 951                 /* Credit-based scheduling. */
 952                 if (txreq.size > queue->remaining_credit &&
 953                     tx_credit_exceeded(queue, txreq.size))
 954                         break;
 955
 956                 queue->remaining_credit -= txreq.size;
 957
 958                 work_to_do--;
 959                 queue->tx.req_cons = ++idx;
 960
 961                 memset(extras, 0, sizeof(extras));
 962                 extra_count = 0;
 963                 if (txreq.flags & XEN_NETTXF_extra_info) {
 964                         work_to_do = xenvif_get_extras(queue, extras,
 965                                                        &extra_count,
 966                                                        work_to_do);
 967                         idx = queue->tx.req_cons;
 968                         if (unlikely(work_to_do < 0))
 969                                 break;
 970                 }
 971
 972                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
 973                         struct xen_netif_extra_info *extra;
 974
 975                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
 976                         ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
 977
 978                         make_tx_response(queue, &txreq, extra_count,
 979                                          (ret == 0) ?
 980                                          XEN_NETIF_RSP_OKAY :
 981                                          XEN_NETIF_RSP_ERROR);
 982                         continue;
 983                 }
 984
 985                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
 986                         struct xen_netif_extra_info *extra;
 987
 988                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
 989                         xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
 990
 991                         make_tx_response(queue, &txreq, extra_count,
 992                                          XEN_NETIF_RSP_OKAY);
 993                         continue;
 994                 }
 995
 996                 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN) ?
 997                         XEN_NETBACK_TX_COPY_LEN : txreq.size;
 998
 999                 ret = xenvif_count_requests(queue, &txreq, extra_count,
1000                                             txfrags, work_to_do);
1001
1002                 if (unlikely(ret < 0))
1003                         break;
1004
1005                 idx += ret;
1006
1007                 if (unlikely(txreq.size < ETH_HLEN)) {
1008                         netdev_dbg(queue->vif->dev,
1009                                    "Bad packet size: %d\n", txreq.size);
1010                         xenvif_tx_err(queue, &txreq, extra_count, idx);
1011                         break;
1012                 }
1013
1014                 /* No crossing a page as the payload mustn't fragment. */
1015                 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
1016                         netdev_err(queue->vif->dev, "Cross page boundary, txreq.offset: %u, size: %u\n",
1017                                    txreq.offset, txreq.size);
1018                         xenvif_fatal_tx_err(queue->vif);
1019                         break;
1020                 }
1021
1022                 if (ret >= XEN_NETBK_LEGACY_SLOTS_MAX - 1 && data_len < txreq.size)
1023                         data_len = txreq.size;
1024
1025                 skb = xenvif_alloc_skb(data_len);
1026                 if (unlikely(skb == NULL)) {
1027                         netdev_dbg(queue->vif->dev,
1028                                    "Can't allocate a skb in start_xmit.\n");
1029                         xenvif_tx_err(queue, &txreq, extra_count, idx);
1030                         break;
1031                 }
1032
1033                 skb_shinfo(skb)->nr_frags = ret;
1034                 /* At this point shinfo->nr_frags is in fact the number of
1035                  * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1036                  */
1037                 frag_overflow = 0;
1038                 nskb = NULL;
1039                 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1040                         frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1041                         BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1042                         skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1043                         nskb = xenvif_alloc_skb(0);
1044                         if (unlikely(nskb == NULL)) {
1045                                 skb_shinfo(skb)->nr_frags = 0;
1046                                 kfree_skb(skb);
1047                                 xenvif_tx_err(queue, &txreq, extra_count, idx);
1048                                 if (net_ratelimit())
1049                                         netdev_err(queue->vif->dev,
1050                                                    "Can't allocate the frag_list skb.\n");
1051                                 break;
1052                         }
1053                 }
1054
1055                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1056                         struct xen_netif_extra_info *gso;
1057                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1058
1059                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1060                                 /* Failure in xenvif_set_skb_gso is fatal. */
1061                                 skb_shinfo(skb)->nr_frags = 0;
1062                                 kfree_skb(skb);
1063                                 kfree_skb(nskb);
1064                                 break;
1065                         }
1066                 }
1067
1068                 if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
1069                         struct xen_netif_extra_info *extra;
1070                         enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
1071
1072                         extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
1073
1074                         switch (extra->u.hash.type) {
1075                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
1076                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
1077                                 type = PKT_HASH_TYPE_L3;
1078                                 break;
1079
1080                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
1081                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
1082                                 type = PKT_HASH_TYPE_L4;
1083                                 break;
1084
1085                         default:
1086                                 break;
1087                         }
1088
1089                         if (type != PKT_HASH_TYPE_NONE)
1090                                 skb_set_hash(skb,
1091                                              *(u32 *)extra->u.hash.value,
1092                                              type);
1093                 }
1094
1095                 xenvif_get_requests(queue, skb, &txreq, txfrags, copy_ops,
1096                                     map_ops, frag_overflow, nskb, extra_count,
1097                                     data_len);
1098
1099                 __skb_queue_tail(&queue->tx_queue, skb);
1100
1101                 queue->tx.req_cons = idx;
1102         }
1103
1104         return;
1105 }
1106
1107 /* Consolidate skb with a frag_list into a brand new one with local pages on
1108  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1109  */
1110 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1111 {
1112         unsigned int offset = skb_headlen(skb);
1113         skb_frag_t frags[MAX_SKB_FRAGS];
1114         int i, f;
1115         struct ubuf_info *uarg;
1116         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1117
1118         queue->stats.tx_zerocopy_sent += 2;
1119         queue->stats.tx_frag_overflow++;
1120
1121         xenvif_fill_frags(queue, nskb);
1122         /* Subtract frags size, we will correct it later */
1123         skb->truesize -= skb->data_len;
1124         skb->len += nskb->len;
1125         skb->data_len += nskb->len;
1126
1127         /* create a brand new frags array and coalesce there */
1128         for (i = 0; offset < skb->len; i++) {
1129                 struct page *page;
1130                 unsigned int len;
1131
1132                 BUG_ON(i >= MAX_SKB_FRAGS);
1133                 page = alloc_page(GFP_ATOMIC);
1134                 if (!page) {
1135                         int j;
1136                         skb->truesize += skb->data_len;
1137                         for (j = 0; j < i; j++)
1138                                 put_page(skb_frag_page(&frags[j]));
1139                         return -ENOMEM;
1140                 }
1141
1142                 if (offset + PAGE_SIZE < skb->len)
1143                         len = PAGE_SIZE;
1144                 else
1145                         len = skb->len - offset;
1146                 if (skb_copy_bits(skb, offset, page_address(page), len))
1147                         BUG();
1148
1149                 offset += len;
1150                 skb_frag_fill_page_desc(&frags[i], page, 0, len);
1151         }
1152
1153         /* Release all the original (foreign) frags. */
1154         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1155                 skb_frag_unref(skb, f);
1156         uarg = skb_shinfo(skb)->destructor_arg;
1157         /* increase inflight counter to offset decrement in callback */
1158         atomic_inc(&queue->inflight_packets);
1159         uarg->callback(NULL, uarg, true);
1160         skb_shinfo(skb)->destructor_arg = NULL;
1161
1162         /* Fill the skb with the new (local) frags. */
1163         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1164         skb_shinfo(skb)->nr_frags = i;
1165         skb->truesize += i * PAGE_SIZE;
1166
1167         return 0;
1168 }
1169
1170 static int xenvif_tx_submit(struct xenvif_queue *queue)
1171 {
1172         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1173         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1174         struct sk_buff *skb;
1175         int work_done = 0;
1176
1177         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1178                 struct xen_netif_tx_request *txp;
1179                 u16 pending_idx;
1180
1181                 pending_idx = copy_pending_idx(skb, 0);
1182                 txp = &queue->pending_tx_info[pending_idx].req;
1183
1184                 /* Check the remap error code. */
1185                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1186                         /* If there was an error, xenvif_tx_check_gop is
1187                          * expected to release all the frags which were mapped,
1188                          * so kfree_skb shouldn't do it again
1189                          */
1190                         skb_shinfo(skb)->nr_frags = 0;
1191                         if (skb_has_frag_list(skb)) {
1192                                 struct sk_buff *nskb =
1193                                                 skb_shinfo(skb)->frag_list;
1194                                 skb_shinfo(nskb)->nr_frags = 0;
1195                         }
1196                         kfree_skb(skb);
1197                         continue;
1198                 }
1199
1200                 if (txp->flags & XEN_NETTXF_csum_blank)
1201                         skb->ip_summed = CHECKSUM_PARTIAL;
1202                 else if (txp->flags & XEN_NETTXF_data_validated)
1203                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1204
1205                 xenvif_fill_frags(queue, skb);
1206
1207                 if (unlikely(skb_has_frag_list(skb))) {
1208                         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1209                         xenvif_skb_zerocopy_prepare(queue, nskb);
1210                         if (xenvif_handle_frag_list(queue, skb)) {
1211                                 if (net_ratelimit())
1212                                         netdev_err(queue->vif->dev,
1213                                                    "Not enough memory to consolidate frag_list!\n");
1214                                 xenvif_skb_zerocopy_prepare(queue, skb);
1215                                 kfree_skb(skb);
1216                                 continue;
1217                         }
1218                         /* Copied all the bits from the frag list -- free it. */
1219                         skb_frag_list_init(skb);
1220                         kfree_skb(nskb);
1221                 }
1222
1223                 skb->dev      = queue->vif->dev;
1224                 skb->protocol = eth_type_trans(skb, skb->dev);
1225                 skb_reset_network_header(skb);
1226
1227                 if (checksum_setup(queue, skb)) {
1228                         netdev_dbg(queue->vif->dev,
1229                                    "Can't setup checksum in net_tx_action\n");
1230                         /* We have to set this flag to trigger the callback */
1231                         if (skb_shinfo(skb)->destructor_arg)
1232                                 xenvif_skb_zerocopy_prepare(queue, skb);
1233                         kfree_skb(skb);
1234                         continue;
1235                 }
1236
1237                 skb_probe_transport_header(skb);
1238
1239                 /* If the packet is GSO then we will have just set up the
1240                  * transport header offset in checksum_setup so it's now
1241                  * straightforward to calculate gso_segs.
1242                  */
1243                 if (skb_is_gso(skb)) {
1244                         int mss, hdrlen;
1245
1246                         /* GSO implies having the L4 header. */
1247                         WARN_ON_ONCE(!skb_transport_header_was_set(skb));
1248                         if (unlikely(!skb_transport_header_was_set(skb))) {
1249                                 kfree_skb(skb);
1250                                 continue;
1251                         }
1252
1253                         mss = skb_shinfo(skb)->gso_size;
1254                         hdrlen = skb_tcp_all_headers(skb);
1255
1256                         skb_shinfo(skb)->gso_segs =
1257                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1258                 }
1259
1260                 queue->stats.rx_bytes += skb->len;
1261                 queue->stats.rx_packets++;
1262
1263                 work_done++;
1264
1265                 /* Set this flag right before netif_receive_skb, otherwise
1266                  * someone might think this packet already left netback, and
1267                  * do a skb_copy_ubufs while we are still in control of the
1268                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1269                  */
1270                 if (skb_shinfo(skb)->destructor_arg) {
1271                         xenvif_skb_zerocopy_prepare(queue, skb);
1272                         queue->stats.tx_zerocopy_sent++;
1273                 }
1274
1275                 netif_receive_skb(skb);
1276         }
1277
1278         return work_done;
1279 }
1280
1281 void xenvif_zerocopy_callback(struct sk_buff *skb, struct ubuf_info *ubuf_base,
1282                               bool zerocopy_success)
1283 {
1284         unsigned long flags;
1285         pending_ring_idx_t index;
1286         struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
1287         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1288
1289         /* This is the only place where we grab this lock, to protect callbacks
1290          * from each other.
1291          */
1292         spin_lock_irqsave(&queue->callback_lock, flags);
1293         do {
1294                 u16 pending_idx = ubuf->desc;
1295                 ubuf = (struct ubuf_info_msgzc *) ubuf->ctx;
1296                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1297                         MAX_PENDING_REQS);
1298                 index = pending_index(queue->dealloc_prod);
1299                 queue->dealloc_ring[index] = pending_idx;
1300                 /* Sync with xenvif_tx_dealloc_action:
1301                  * insert idx then incr producer.
1302                  */
1303                 smp_wmb();
1304                 queue->dealloc_prod++;
1305         } while (ubuf);
1306         spin_unlock_irqrestore(&queue->callback_lock, flags);
1307
1308         if (likely(zerocopy_success))
1309                 queue->stats.tx_zerocopy_success++;
1310         else
1311                 queue->stats.tx_zerocopy_fail++;
1312         xenvif_skb_zerocopy_complete(queue);
1313 }
1314
1315 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1316 {
1317         struct gnttab_unmap_grant_ref *gop;
1318         pending_ring_idx_t dc, dp;
1319         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1320         unsigned int i = 0;
1321
1322         dc = queue->dealloc_cons;
1323         gop = queue->tx_unmap_ops;
1324
1325         /* Free up any grants we have finished using */
1326         do {
1327                 dp = queue->dealloc_prod;
1328
1329                 /* Ensure we see all indices enqueued by all
1330                  * xenvif_zerocopy_callback().
1331                  */
1332                 smp_rmb();
1333
1334                 while (dc != dp) {
1335                         BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1336                         pending_idx =
1337                                 queue->dealloc_ring[pending_index(dc++)];
1338
1339                         pending_idx_release[gop - queue->tx_unmap_ops] =
1340                                 pending_idx;
1341                         queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1342                                 queue->mmap_pages[pending_idx];
1343                         gnttab_set_unmap_op(gop,
1344                                             idx_to_kaddr(queue, pending_idx),
1345                                             GNTMAP_host_map,
1346                                             queue->grant_tx_handle[pending_idx]);
1347                         xenvif_grant_handle_reset(queue, pending_idx);
1348                         ++gop;
1349                 }
1350
1351         } while (dp != queue->dealloc_prod);
1352
1353         queue->dealloc_cons = dc;
1354
1355         if (gop - queue->tx_unmap_ops > 0) {
1356                 int ret;
1357                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1358                                         NULL,
1359                                         queue->pages_to_unmap,
1360                                         gop - queue->tx_unmap_ops);
1361                 if (ret) {
1362                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1363                                    gop - queue->tx_unmap_ops, ret);
1364                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1365                                 if (gop[i].status != GNTST_okay)
1366                                         netdev_err(queue->vif->dev,
1367                                                    " host_addr: 0x%llx handle: 0x%x status: %d\n",
1368                                                    gop[i].host_addr,
1369                                                    gop[i].handle,
1370                                                    gop[i].status);
1371                         }
1372                         BUG();
1373                 }
1374         }
1375
1376         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1377                 xenvif_idx_release(queue, pending_idx_release[i],
1378                                    XEN_NETIF_RSP_OKAY);
1379 }
1380
1381
1382 /* Called after netfront has transmitted */
1383 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1384 {
1385         unsigned nr_mops = 0, nr_cops = 0;
1386         int work_done, ret;
1387
1388         if (unlikely(!tx_work_todo(queue)))
1389                 return 0;
1390
1391         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1392
1393         if (nr_cops == 0)
1394                 return 0;
1395
1396         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1397         if (nr_mops != 0) {
1398                 ret = gnttab_map_refs(queue->tx_map_ops,
1399                                       NULL,
1400                                       queue->pages_to_map,
1401                                       nr_mops);
1402                 if (ret) {
1403                         unsigned int i;
1404
1405                         netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1406                                    nr_mops, ret);
1407                         for (i = 0; i < nr_mops; ++i)
1408                                 WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1409                                              GNTST_okay);
1410                 }
1411         }
1412
1413         work_done = xenvif_tx_submit(queue);
1414
1415         return work_done;
1416 }
1417
1418 static void _make_tx_response(struct xenvif_queue *queue,
1419                              const struct xen_netif_tx_request *txp,
1420                              unsigned int extra_count,
1421                              s8 status)
1422 {
1423         RING_IDX i = queue->tx.rsp_prod_pvt;
1424         struct xen_netif_tx_response *resp;
1425
1426         resp = RING_GET_RESPONSE(&queue->tx, i);
1427         resp->id     = txp->id;
1428         resp->status = status;
1429
1430         while (extra_count-- != 0)
1431                 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1432
1433         queue->tx.rsp_prod_pvt = ++i;
1434 }
1435
1436 static void push_tx_responses(struct xenvif_queue *queue)
1437 {
1438         int notify;
1439
1440         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1441         if (notify)
1442                 notify_remote_via_irq(queue->tx_irq);
1443 }
1444
1445 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1446                                s8 status)
1447 {
1448         struct pending_tx_info *pending_tx_info;
1449         pending_ring_idx_t index;
1450         unsigned long flags;
1451
1452         pending_tx_info = &queue->pending_tx_info[pending_idx];
1453
1454         spin_lock_irqsave(&queue->response_lock, flags);
1455
1456         _make_tx_response(queue, &pending_tx_info->req,
1457                           pending_tx_info->extra_count, status);
1458
1459         /* Release the pending index before pusing the Tx response so
1460          * its available before a new Tx request is pushed by the
1461          * frontend.
1462          */
1463         index = pending_index(queue->pending_prod++);
1464         queue->pending_ring[index] = pending_idx;
1465
1466         push_tx_responses(queue);
1467
1468         spin_unlock_irqrestore(&queue->response_lock, flags);
1469 }
1470
1471 static void make_tx_response(struct xenvif_queue *queue,
1472                              const struct xen_netif_tx_request *txp,
1473                              unsigned int extra_count,
1474                              s8 status)
1475 {
1476         unsigned long flags;
1477
1478         spin_lock_irqsave(&queue->response_lock, flags);
1479
1480         _make_tx_response(queue, txp, extra_count, status);
1481         push_tx_responses(queue);
1482
1483         spin_unlock_irqrestore(&queue->response_lock, flags);
1484 }
1485
1486 static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1487 {
1488         int ret;
1489         struct gnttab_unmap_grant_ref tx_unmap_op;
1490
1491         gnttab_set_unmap_op(&tx_unmap_op,
1492                             idx_to_kaddr(queue, pending_idx),
1493                             GNTMAP_host_map,
1494                             queue->grant_tx_handle[pending_idx]);
1495         xenvif_grant_handle_reset(queue, pending_idx);
1496
1497         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1498                                 &queue->mmap_pages[pending_idx], 1);
1499         if (ret) {
1500                 netdev_err(queue->vif->dev,
1501                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1502                            ret,
1503                            pending_idx,
1504                            tx_unmap_op.host_addr,
1505                            tx_unmap_op.handle,
1506                            tx_unmap_op.status);
1507                 BUG();
1508         }
1509 }
1510
1511 static inline int tx_work_todo(struct xenvif_queue *queue)
1512 {
1513         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1514                 return 1;
1515
1516         return 0;
1517 }
1518
1519 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1520 {
1521         return queue->dealloc_cons != queue->dealloc_prod;
1522 }
1523
1524 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1525 {
1526         if (queue->tx.sring)
1527                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1528                                         queue->tx.sring);
1529         if (queue->rx.sring)
1530                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1531                                         queue->rx.sring);
1532 }
1533
1534 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1535                                    grant_ref_t tx_ring_ref,
1536                                    grant_ref_t rx_ring_ref)
1537 {
1538         void *addr;
1539         struct xen_netif_tx_sring *txs;
1540         struct xen_netif_rx_sring *rxs;
1541         RING_IDX rsp_prod, req_prod;
1542         int err;
1543
1544         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1545                                      &tx_ring_ref, 1, &addr);
1546         if (err)
1547                 goto err;
1548
1549         txs = (struct xen_netif_tx_sring *)addr;
1550         rsp_prod = READ_ONCE(txs->rsp_prod);
1551         req_prod = READ_ONCE(txs->req_prod);
1552
1553         BACK_RING_ATTACH(&queue->tx, txs, rsp_prod, XEN_PAGE_SIZE);
1554
1555         err = -EIO;
1556         if (req_prod - rsp_prod > RING_SIZE(&queue->tx))
1557                 goto err;
1558
1559         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1560                                      &rx_ring_ref, 1, &addr);
1561         if (err)
1562                 goto err;
1563
1564         rxs = (struct xen_netif_rx_sring *)addr;
1565         rsp_prod = READ_ONCE(rxs->rsp_prod);
1566         req_prod = READ_ONCE(rxs->req_prod);
1567
1568         BACK_RING_ATTACH(&queue->rx, rxs, rsp_prod, XEN_PAGE_SIZE);
1569
1570         err = -EIO;
1571         if (req_prod - rsp_prod > RING_SIZE(&queue->rx))
1572                 goto err;
1573
1574         return 0;
1575
1576 err:
1577         xenvif_unmap_frontend_data_rings(queue);
1578         return err;
1579 }
1580
1581 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1582 {
1583         /* Dealloc thread must remain running until all inflight
1584          * packets complete.
1585          */
1586         return kthread_should_stop() &&
1587                 !atomic_read(&queue->inflight_packets);
1588 }
1589
1590 int xenvif_dealloc_kthread(void *data)
1591 {
1592         struct xenvif_queue *queue = data;
1593
1594         for (;;) {
1595                 wait_event_interruptible(queue->dealloc_wq,
1596                                          tx_dealloc_work_todo(queue) ||
1597                                          xenvif_dealloc_kthread_should_stop(queue));
1598                 if (xenvif_dealloc_kthread_should_stop(queue))
1599                         break;
1600
1601                 xenvif_tx_dealloc_action(queue);
1602                 cond_resched();
1603         }
1604
1605         /* Unmap anything remaining*/
1606         if (tx_dealloc_work_todo(queue))
1607                 xenvif_tx_dealloc_action(queue);
1608
1609         return 0;
1610 }
1611
1612 static void make_ctrl_response(struct xenvif *vif,
1613                                const struct xen_netif_ctrl_request *req,
1614                                u32 status, u32 data)
1615 {
1616         RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1617         struct xen_netif_ctrl_response rsp = {
1618                 .id = req->id,
1619                 .type = req->type,
1620                 .status = status,
1621                 .data = data,
1622         };
1623
1624         *RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1625         vif->ctrl.rsp_prod_pvt = ++idx;
1626 }
1627
1628 static void push_ctrl_response(struct xenvif *vif)
1629 {
1630         int notify;
1631
1632         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1633         if (notify)
1634                 notify_remote_via_irq(vif->ctrl_irq);
1635 }
1636
1637 static void process_ctrl_request(struct xenvif *vif,
1638                                  const struct xen_netif_ctrl_request *req)
1639 {
1640         u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1641         u32 data = 0;
1642
1643         switch (req->type) {
1644         case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1645                 status = xenvif_set_hash_alg(vif, req->data[0]);
1646                 break;
1647
1648         case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1649                 status = xenvif_get_hash_flags(vif, &data);
1650                 break;
1651
1652         case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1653                 status = xenvif_set_hash_flags(vif, req->data[0]);
1654                 break;
1655
1656         case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1657                 status = xenvif_set_hash_key(vif, req->data[0],
1658                                              req->data[1]);
1659                 break;
1660
1661         case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1662                 status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1663                 data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1664                 break;
1665
1666         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1667                 status = xenvif_set_hash_mapping_size(vif,
1668                                                       req->data[0]);
1669                 break;
1670
1671         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1672                 status = xenvif_set_hash_mapping(vif, req->data[0],
1673                                                  req->data[1],
1674                                                  req->data[2]);
1675                 break;
1676
1677         default:
1678                 break;
1679         }
1680
1681         make_ctrl_response(vif, req, status, data);
1682         push_ctrl_response(vif);
1683 }
1684
1685 static void xenvif_ctrl_action(struct xenvif *vif)
1686 {
1687         for (;;) {
1688                 RING_IDX req_prod, req_cons;
1689
1690                 req_prod = vif->ctrl.sring->req_prod;
1691                 req_cons = vif->ctrl.req_cons;
1692
1693                 /* Make sure we can see requests before we process them. */
1694                 rmb();
1695
1696                 if (req_cons == req_prod)
1697                         break;
1698
1699                 while (req_cons != req_prod) {
1700                         struct xen_netif_ctrl_request req;
1701
1702                         RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1703                         req_cons++;
1704
1705                         process_ctrl_request(vif, &req);
1706                 }
1707
1708                 vif->ctrl.req_cons = req_cons;
1709                 vif->ctrl.sring->req_event = req_cons + 1;
1710         }
1711 }
1712
1713 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1714 {
1715         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1716                 return true;
1717
1718         return false;
1719 }
1720
1721 irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1722 {
1723         struct xenvif *vif = data;
1724         unsigned int eoi_flag = XEN_EOI_FLAG_SPURIOUS;
1725
1726         while (xenvif_ctrl_work_todo(vif)) {
1727                 xenvif_ctrl_action(vif);
1728                 eoi_flag = 0;
1729         }
1730
1731         xen_irq_lateeoi(irq, eoi_flag);
1732
1733         return IRQ_HANDLED;
1734 }
1735
1736 static int __init netback_init(void)
1737 {
1738         int rc = 0;
1739
1740         if (!xen_domain())
1741                 return -ENODEV;
1742
1743         /* Allow as many queues as there are CPUs but max. 8 if user has not
1744          * specified a value.
1745          */
1746         if (xenvif_max_queues == 0)
1747                 xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1748                                           num_online_cpus());
1749
1750         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1751                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1752                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1753                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1754         }
1755
1756         rc = xenvif_xenbus_init();
1757         if (rc)
1758                 goto failed_init;
1759
1760 #ifdef CONFIG_DEBUG_FS
1761         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1762 #endif /* CONFIG_DEBUG_FS */
1763
1764         return 0;
1765
1766 failed_init:
1767         return rc;
1768 }
1769
1770 module_init(netback_init);
1771
1772 static void __exit netback_fini(void)
1773 {
1774 #ifdef CONFIG_DEBUG_FS
1775         debugfs_remove_recursive(xen_netback_dbg_root);
1776 #endif /* CONFIG_DEBUG_FS */
1777         xenvif_xenbus_fini();
1778 }
1779 module_exit(netback_fini);
1780
1781 MODULE_DESCRIPTION("Xen backend network device module");
1782 MODULE_LICENSE("Dual BSD/GPL");
1783 MODULE_ALIAS("xen-backend:vif");