drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <mst@redhat.com>
   3  *
   4  * This work is licensed under the terms of the GNU GPL, version 2.
   5  *
   6  * virtio-net server in host kernel.
   7  */
   8
   9 #include <linux/compat.h>
  10 #include <linux/eventfd.h>
  11 #include <linux/vhost.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/module.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/mutex.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/sched/clock.h>
  21 #include <linux/sched/signal.h>
  22 #include <linux/vmalloc.h>
  23
  24 #include <linux/net.h>
  25 #include <linux/if_packet.h>
  26 #include <linux/if_arp.h>
  27 #include <linux/if_tun.h>
  28 #include <linux/if_macvlan.h>
  29 #include <linux/if_tap.h>
  30 #include <linux/if_vlan.h>
  31 #include <linux/skb_array.h>
  32 #include <linux/skbuff.h>
  33
  34 #include <net/sock.h>
  35 #include <net/xdp.h>
  36
  37 #include "vhost.h"
  38
  39 static int experimental_zcopytx = 1;
  40 module_param(experimental_zcopytx, int, 0444);
  41 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  42                                        " 1 -Enable; 0 - Disable");
  43
  44 /* Max number of bytes transferred before requeueing the job.
  45  * Using this limit prevents one virtqueue from starving others. */
  46 #define VHOST_NET_WEIGHT 0x80000
  47
  48 /* Max number of packets transferred before requeueing the job.
  49  * Using this limit prevents one virtqueue from starving others with small
  50  * pkts.
  51  */
  52 #define VHOST_NET_PKT_WEIGHT 256
  53
  54 /* MAX number of TX used buffers for outstanding zerocopy */
  55 #define VHOST_MAX_PEND 128
  56 #define VHOST_GOODCOPY_LEN 256
  57
  58 /*
  59  * For transmit, used buffer len is unused; we override it to track buffer
  60  * status internally; used for zerocopy tx only.
  61  */
  62 /* Lower device DMA failed */
  63 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  64 /* Lower device DMA done */
  65 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  66 /* Lower device DMA in progress */
  67 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  68 /* Buffer unused */
  69 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  70
  71 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  72
  73 enum {
  74         VHOST_NET_FEATURES = VHOST_FEATURES |
  75                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  76                          (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  77                          (1ULL << VIRTIO_F_IOMMU_PLATFORM)
  78 };
  79
  80 enum {
  81         VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
  82 };
  83
  84 enum {
  85         VHOST_NET_VQ_RX = 0,
  86         VHOST_NET_VQ_TX = 1,
  87         VHOST_NET_VQ_MAX = 2,
  88 };
  89
  90 struct vhost_net_ubuf_ref {
  91         /* refcount follows semantics similar to kref:
  92          *  0: object is released
  93          *  1: no outstanding ubufs
  94          * >1: outstanding ubufs
  95          */
  96         atomic_t refcount;
  97         wait_queue_head_t wait;
  98         struct vhost_virtqueue *vq;
  99 };
 100
 101 #define VHOST_NET_BATCH 64
 102 struct vhost_net_buf {
 103         void **queue;
 104         int tail;
 105         int head;
 106 };
 107
 108 struct vhost_net_virtqueue {
 109         struct vhost_virtqueue vq;
 110         size_t vhost_hlen;
 111         size_t sock_hlen;
 112         /* vhost zerocopy support fields below: */
 113         /* last used idx for outstanding DMA zerocopy buffers */
 114         int upend_idx;
 115         /* For TX, first used idx for DMA done zerocopy buffers
 116          * For RX, number of batched heads
 117          */
 118         int done_idx;
 119         /* an array of userspace buffers info */
 120         struct ubuf_info *ubuf_info;
 121         /* Reference counting for outstanding ubufs.
 122          * Protected by vq mutex. Writers must also take device mutex. */
 123         struct vhost_net_ubuf_ref *ubufs;
 124         struct ptr_ring *rx_ring;
 125         struct vhost_net_buf rxq;
 126 };
 127
 128 struct vhost_net {
 129         struct vhost_dev dev;
 130         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 131         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 132         /* Number of TX recently submitted.
 133          * Protected by tx vq lock. */
 134         unsigned tx_packets;
 135         /* Number of times zerocopy TX recently failed.
 136          * Protected by tx vq lock. */
 137         unsigned tx_zcopy_err;
 138         /* Flush in progress. Protected by tx vq lock. */
 139         bool tx_flush;
 140 };
 141
 142 static unsigned vhost_net_zcopy_mask __read_mostly;
 143
 144 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 145 {
 146         if (rxq->tail != rxq->head)
 147                 return rxq->queue[rxq->head];
 148         else
 149                 return NULL;
 150 }
 151
 152 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 153 {
 154         return rxq->tail - rxq->head;
 155 }
 156
 157 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 158 {
 159         return rxq->tail == rxq->head;
 160 }
 161
 162 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 163 {
 164         void *ret = vhost_net_buf_get_ptr(rxq);
 165         ++rxq->head;
 166         return ret;
 167 }
 168
 169 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 170 {
 171         struct vhost_net_buf *rxq = &nvq->rxq;
 172
 173         rxq->head = 0;
 174         rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
 175                                               VHOST_NET_BATCH);
 176         return rxq->tail;
 177 }
 178
 179 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 180 {
 181         struct vhost_net_buf *rxq = &nvq->rxq;
 182
 183         if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
 184                 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 185                                    vhost_net_buf_get_size(rxq),
 186                                    tun_ptr_free);
 187                 rxq->head = rxq->tail = 0;
 188         }
 189 }
 190
 191 static int vhost_net_buf_peek_len(void *ptr)
 192 {
 193         if (tun_is_xdp_frame(ptr)) {
 194                 struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
 195
 196                 return xdpf->len;
 197         }
 198
 199         return __skb_array_len_with_tag(ptr);
 200 }
 201
 202 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 203 {
 204         struct vhost_net_buf *rxq = &nvq->rxq;
 205
 206         if (!vhost_net_buf_is_empty(rxq))
 207                 goto out;
 208
 209         if (!vhost_net_buf_produce(nvq))
 210                 return 0;
 211
 212 out:
 213         return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 214 }
 215
 216 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 217 {
 218         rxq->head = rxq->tail = 0;
 219 }
 220
 221 static void vhost_net_enable_zcopy(int vq)
 222 {
 223         vhost_net_zcopy_mask |= 0x1 << vq;
 224 }
 225
 226 static struct vhost_net_ubuf_ref *
 227 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 228 {
 229         struct vhost_net_ubuf_ref *ubufs;
 230         /* No zero copy backend? Nothing to count. */
 231         if (!zcopy)
 232                 return NULL;
 233         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 234         if (!ubufs)
 235                 return ERR_PTR(-ENOMEM);
 236         atomic_set(&ubufs->refcount, 1);
 237         init_waitqueue_head(&ubufs->wait);
 238         ubufs->vq = vq;
 239         return ubufs;
 240 }
 241
 242 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 243 {
 244         int r = atomic_sub_return(1, &ubufs->refcount);
 245         if (unlikely(!r))
 246                 wake_up(&ubufs->wait);
 247         return r;
 248 }
 249
 250 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 251 {
 252         vhost_net_ubuf_put(ubufs);
 253         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 254 }
 255
 256 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 257 {
 258         vhost_net_ubuf_put_and_wait(ubufs);
 259         kfree(ubufs);
 260 }
 261
 262 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 263 {
 264         int i;
 265
 266         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 267                 kfree(n->vqs[i].ubuf_info);
 268                 n->vqs[i].ubuf_info = NULL;
 269         }
 270 }
 271
 272 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 273 {
 274         bool zcopy;
 275         int i;
 276
 277         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 278                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 279                 if (!zcopy)
 280                         continue;
 281                 n->vqs[i].ubuf_info =
 282                         kmalloc_array(UIO_MAXIOV,
 283                                       sizeof(*n->vqs[i].ubuf_info),
 284                                       GFP_KERNEL);
 285                 if  (!n->vqs[i].ubuf_info)
 286                         goto err;
 287         }
 288         return 0;
 289
 290 err:
 291         vhost_net_clear_ubuf_info(n);
 292         return -ENOMEM;
 293 }
 294
 295 static void vhost_net_vq_reset(struct vhost_net *n)
 296 {
 297         int i;
 298
 299         vhost_net_clear_ubuf_info(n);
 300
 301         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 302                 n->vqs[i].done_idx = 0;
 303                 n->vqs[i].upend_idx = 0;
 304                 n->vqs[i].ubufs = NULL;
 305                 n->vqs[i].vhost_hlen = 0;
 306                 n->vqs[i].sock_hlen = 0;
 307                 vhost_net_buf_init(&n->vqs[i].rxq);
 308         }
 309
 310 }
 311
 312 static void vhost_net_tx_packet(struct vhost_net *net)
 313 {
 314         ++net->tx_packets;
 315         if (net->tx_packets < 1024)
 316                 return;
 317         net->tx_packets = 0;
 318         net->tx_zcopy_err = 0;
 319 }
 320
 321 static void vhost_net_tx_err(struct vhost_net *net)
 322 {
 323         ++net->tx_zcopy_err;
 324 }
 325
 326 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 327 {
 328         /* TX flush waits for outstanding DMAs to be done.
 329          * Don't start new DMAs.
 330          */
 331         return !net->tx_flush &&
 332                 net->tx_packets / 64 >= net->tx_zcopy_err;
 333 }
 334
 335 static bool vhost_sock_zcopy(struct socket *sock)
 336 {
 337         return unlikely(experimental_zcopytx) &&
 338                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 339 }
 340
 341 /* In case of DMA done not in order in lower device driver for some reason.
 342  * upend_idx is used to track end of used idx, done_idx is used to track head
 343  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 344  * guest used idx.
 345  */
 346 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 347                                        struct vhost_virtqueue *vq)
 348 {
 349         struct vhost_net_virtqueue *nvq =
 350                 container_of(vq, struct vhost_net_virtqueue, vq);
 351         int i, add;
 352         int j = 0;
 353
 354         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 355                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 356                         vhost_net_tx_err(net);
 357                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 358                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 359                         ++j;
 360                 } else
 361                         break;
 362         }
 363         while (j) {
 364                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 365                 vhost_add_used_and_signal_n(vq->dev, vq,
 366                                             &vq->heads[nvq->done_idx], add);
 367                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 368                 j -= add;
 369         }
 370 }
 371
 372 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 373 {
 374         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 375         struct vhost_virtqueue *vq = ubufs->vq;
 376         int cnt;
 377
 378         rcu_read_lock_bh();
 379
 380         /* set len to mark this desc buffers done DMA */
 381         vq->heads[ubuf->desc].len = success ?
 382                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 383         cnt = vhost_net_ubuf_put(ubufs);
 384
 385         /*
 386          * Trigger polling thread if guest stopped submitting new buffers:
 387          * in this case, the refcount after decrement will eventually reach 1.
 388          * We also trigger polling periodically after each 16 packets
 389          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 390          * less than 10% of times).
 391          */
 392         if (cnt <= 1 || !(cnt % 16))
 393                 vhost_poll_queue(&vq->poll);
 394
 395         rcu_read_unlock_bh();
 396 }
 397
 398 static inline unsigned long busy_clock(void)
 399 {
 400         return local_clock() >> 10;
 401 }
 402
 403 static bool vhost_can_busy_poll(unsigned long endtime)
 404 {
 405         return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
 406                       !signal_pending(current));
 407 }
 408
 409 static void vhost_net_disable_vq(struct vhost_net *n,
 410                                  struct vhost_virtqueue *vq)
 411 {
 412         struct vhost_net_virtqueue *nvq =
 413                 container_of(vq, struct vhost_net_virtqueue, vq);
 414         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 415         if (!vq->private_data)
 416                 return;
 417         vhost_poll_stop(poll);
 418 }
 419
 420 static int vhost_net_enable_vq(struct vhost_net *n,
 421                                 struct vhost_virtqueue *vq)
 422 {
 423         struct vhost_net_virtqueue *nvq =
 424                 container_of(vq, struct vhost_net_virtqueue, vq);
 425         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 426         struct socket *sock;
 427
 428         sock = vq->private_data;
 429         if (!sock)
 430                 return 0;
 431
 432         return vhost_poll_start(poll, sock->file);
 433 }
 434
 435 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
 436 {
 437         struct vhost_virtqueue *vq = &nvq->vq;
 438         struct vhost_dev *dev = vq->dev;
 439
 440         if (!nvq->done_idx)
 441                 return;
 442
 443         vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
 444         nvq->done_idx = 0;
 445 }
 446
 447 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 448                                     struct vhost_net_virtqueue *nvq,
 449                                     unsigned int *out_num, unsigned int *in_num,
 450                                     bool *busyloop_intr)
 451 {
 452         struct vhost_virtqueue *vq = &nvq->vq;
 453         unsigned long uninitialized_var(endtime);
 454         int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 455                                   out_num, in_num, NULL, NULL);
 456
 457         if (r == vq->num && vq->busyloop_timeout) {
 458                 if (!vhost_sock_zcopy(vq->private_data))
 459                         vhost_net_signal_used(nvq);
 460                 preempt_disable();
 461                 endtime = busy_clock() + vq->busyloop_timeout;
 462                 while (vhost_can_busy_poll(endtime)) {
 463                         if (vhost_has_work(vq->dev)) {
 464                                 *busyloop_intr = true;
 465                                 break;
 466                         }
 467                         if (!vhost_vq_avail_empty(vq->dev, vq))
 468                                 break;
 469                         cpu_relax();
 470                 }
 471                 preempt_enable();
 472                 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 473                                       out_num, in_num, NULL, NULL);
 474         }
 475
 476         return r;
 477 }
 478
 479 static bool vhost_exceeds_maxpend(struct vhost_net *net)
 480 {
 481         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 482         struct vhost_virtqueue *vq = &nvq->vq;
 483
 484         return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 485                min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 486 }
 487
 488 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
 489                             size_t hdr_size, int out)
 490 {
 491         /* Skip header. TODO: support TSO. */
 492         size_t len = iov_length(vq->iov, out);
 493
 494         iov_iter_init(iter, WRITE, vq->iov, out, len);
 495         iov_iter_advance(iter, hdr_size);
 496
 497         return iov_iter_count(iter);
 498 }
 499
 500 static bool vhost_exceeds_weight(int pkts, int total_len)
 501 {
 502         return total_len >= VHOST_NET_WEIGHT ||
 503                pkts >= VHOST_NET_PKT_WEIGHT;
 504 }
 505
 506 static int get_tx_bufs(struct vhost_net *net,
 507                        struct vhost_net_virtqueue *nvq,
 508                        struct msghdr *msg,
 509                        unsigned int *out, unsigned int *in,
 510                        size_t *len, bool *busyloop_intr)
 511 {
 512         struct vhost_virtqueue *vq = &nvq->vq;
 513         int ret;
 514
 515         ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, busyloop_intr);
 516
 517         if (ret < 0 || ret == vq->num)
 518                 return ret;
 519
 520         if (*in) {
 521                 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 522                         *out, *in);
 523                 return -EFAULT;
 524         }
 525
 526         /* Sanity check */
 527         *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 528         if (*len == 0) {
 529                 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 530                         *len, nvq->vhost_hlen);
 531                 return -EFAULT;
 532         }
 533
 534         return ret;
 535 }
 536
 537 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 538 {
 539         return total_len < VHOST_NET_WEIGHT &&
 540                !vhost_vq_avail_empty(vq->dev, vq);
 541 }
 542
 543 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 544 {
 545         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 546         struct vhost_virtqueue *vq = &nvq->vq;
 547         unsigned out, in;
 548         int head;
 549         struct msghdr msg = {
 550                 .msg_name = NULL,
 551                 .msg_namelen = 0,
 552                 .msg_control = NULL,
 553                 .msg_controllen = 0,
 554                 .msg_flags = MSG_DONTWAIT,
 555         };
 556         size_t len, total_len = 0;
 557         int err;
 558         int sent_pkts = 0;
 559
 560         for (;;) {
 561                 bool busyloop_intr = false;
 562
 563                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 564                                    &busyloop_intr);
 565                 /* On error, stop handling until the next kick. */
 566                 if (unlikely(head < 0))
 567                         break;
 568                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 569                 if (head == vq->num) {
 570                         if (unlikely(busyloop_intr)) {
 571                                 vhost_poll_queue(&vq->poll);
 572                         } else if (unlikely(vhost_enable_notify(&net->dev,
 573                                                                 vq))) {
 574                                 vhost_disable_notify(&net->dev, vq);
 575                                 continue;
 576                         }
 577                         break;
 578                 }
 579
 580                 vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 581                 vq->heads[nvq->done_idx].len = 0;
 582
 583                 total_len += len;
 584                 if (tx_can_batch(vq, total_len))
 585                         msg.msg_flags |= MSG_MORE;
 586                 else
 587                         msg.msg_flags &= ~MSG_MORE;
 588
 589                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 590                 err = sock->ops->sendmsg(sock, &msg, len);
 591                 if (unlikely(err < 0)) {
 592                         vhost_discard_vq_desc(vq, 1);
 593                         vhost_net_enable_vq(net, vq);
 594                         break;
 595                 }
 596                 if (err != len)
 597                         pr_debug("Truncated TX packet: len %d != %zd\n",
 598                                  err, len);
 599                 if (++nvq->done_idx >= VHOST_NET_BATCH)
 600                         vhost_net_signal_used(nvq);
 601                 if (vhost_exceeds_weight(++sent_pkts, total_len)) {
 602                         vhost_poll_queue(&vq->poll);
 603                         break;
 604                 }
 605         }
 606
 607         vhost_net_signal_used(nvq);
 608 }
 609
 610 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 611 {
 612         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 613         struct vhost_virtqueue *vq = &nvq->vq;
 614         unsigned out, in;
 615         int head;
 616         struct msghdr msg = {
 617                 .msg_name = NULL,
 618                 .msg_namelen = 0,
 619                 .msg_control = NULL,
 620                 .msg_controllen = 0,
 621                 .msg_flags = MSG_DONTWAIT,
 622         };
 623         size_t len, total_len = 0;
 624         int err;
 625         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 626         bool zcopy_used;
 627         int sent_pkts = 0;
 628
 629         for (;;) {
 630                 bool busyloop_intr;
 631
 632                 /* Release DMAs done buffers first */
 633                 vhost_zerocopy_signal_used(net, vq);
 634
 635                 busyloop_intr = false;
 636                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 637                                    &busyloop_intr);
 638                 /* On error, stop handling until the next kick. */
 639                 if (unlikely(head < 0))
 640                         break;
 641                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 642                 if (head == vq->num) {
 643                         if (unlikely(busyloop_intr)) {
 644                                 vhost_poll_queue(&vq->poll);
 645                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 646                                 vhost_disable_notify(&net->dev, vq);
 647                                 continue;
 648                         }
 649                         break;
 650                 }
 651
 652                 zcopy_used = len >= VHOST_GOODCOPY_LEN
 653                              && !vhost_exceeds_maxpend(net)
 654                              && vhost_net_tx_select_zcopy(net);
 655
 656                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 657                 if (zcopy_used) {
 658                         struct ubuf_info *ubuf;
 659                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 660
 661                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 662                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 663                         ubuf->callback = vhost_zerocopy_callback;
 664                         ubuf->ctx = nvq->ubufs;
 665                         ubuf->desc = nvq->upend_idx;
 666                         refcount_set(&ubuf->refcnt, 1);
 667                         msg.msg_control = ubuf;
 668                         msg.msg_controllen = sizeof(ubuf);
 669                         ubufs = nvq->ubufs;
 670                         atomic_inc(&ubufs->refcount);
 671                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 672                 } else {
 673                         msg.msg_control = NULL;
 674                         ubufs = NULL;
 675                 }
 676                 total_len += len;
 677                 if (tx_can_batch(vq, total_len) &&
 678                     likely(!vhost_exceeds_maxpend(net))) {
 679                         msg.msg_flags |= MSG_MORE;
 680                 } else {
 681                         msg.msg_flags &= ~MSG_MORE;
 682                 }
 683
 684                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 685                 err = sock->ops->sendmsg(sock, &msg, len);
 686                 if (unlikely(err < 0)) {
 687                         if (zcopy_used) {
 688                                 vhost_net_ubuf_put(ubufs);
 689                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 690                                         % UIO_MAXIOV;
 691                         }
 692                         vhost_discard_vq_desc(vq, 1);
 693                         vhost_net_enable_vq(net, vq);
 694                         break;
 695                 }
 696                 if (err != len)
 697                         pr_debug("Truncated TX packet: "
 698                                  " len %d != %zd\n", err, len);
 699                 if (!zcopy_used)
 700                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 701                 else
 702                         vhost_zerocopy_signal_used(net, vq);
 703                 vhost_net_tx_packet(net);
 704                 if (unlikely(vhost_exceeds_weight(++sent_pkts, total_len))) {
 705                         vhost_poll_queue(&vq->poll);
 706                         break;
 707                 }
 708         }
 709 }
 710
 711 /* Expects to be always run from workqueue - which acts as
 712  * read-size critical section for our kind of RCU. */
 713 static void handle_tx(struct vhost_net *net)
 714 {
 715         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 716         struct vhost_virtqueue *vq = &nvq->vq;
 717         struct socket *sock;
 718
 719         mutex_lock(&vq->mutex);
 720         sock = vq->private_data;
 721         if (!sock)
 722                 goto out;
 723
 724         if (!vq_iotlb_prefetch(vq))
 725                 goto out;
 726
 727         vhost_disable_notify(&net->dev, vq);
 728         vhost_net_disable_vq(net, vq);
 729
 730         if (vhost_sock_zcopy(sock))
 731                 handle_tx_zerocopy(net, sock);
 732         else
 733                 handle_tx_copy(net, sock);
 734
 735 out:
 736         mutex_unlock(&vq->mutex);
 737 }
 738
 739 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 740 {
 741         struct sk_buff *head;
 742         int len = 0;
 743         unsigned long flags;
 744
 745         if (rvq->rx_ring)
 746                 return vhost_net_buf_peek(rvq);
 747
 748         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 749         head = skb_peek(&sk->sk_receive_queue);
 750         if (likely(head)) {
 751                 len = head->len;
 752                 if (skb_vlan_tag_present(head))
 753                         len += VLAN_HLEN;
 754         }
 755
 756         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 757         return len;
 758 }
 759
 760 static int sk_has_rx_data(struct sock *sk)
 761 {
 762         struct socket *sock = sk->sk_socket;
 763
 764         if (sock->ops->peek_len)
 765                 return sock->ops->peek_len(sock);
 766
 767         return skb_queue_empty(&sk->sk_receive_queue);
 768 }
 769
 770 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
 771                                       bool *busyloop_intr)
 772 {
 773         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 774         struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
 775         struct vhost_virtqueue *rvq = &rnvq->vq;
 776         struct vhost_virtqueue *tvq = &tnvq->vq;
 777         unsigned long uninitialized_var(endtime);
 778         int len = peek_head_len(rnvq, sk);
 779
 780         if (!len && tvq->busyloop_timeout) {
 781                 /* Flush batched heads first */
 782                 vhost_net_signal_used(rnvq);
 783                 /* Both tx vq and rx socket were polled here */
 784                 mutex_lock_nested(&tvq->mutex, 1);
 785                 vhost_disable_notify(&net->dev, tvq);
 786
 787                 preempt_disable();
 788                 endtime = busy_clock() + tvq->busyloop_timeout;
 789
 790                 while (vhost_can_busy_poll(endtime)) {
 791                         if (vhost_has_work(&net->dev)) {
 792                                 *busyloop_intr = true;
 793                                 break;
 794                         }
 795                         if ((sk_has_rx_data(sk) &&
 796                              !vhost_vq_avail_empty(&net->dev, rvq)) ||
 797                             !vhost_vq_avail_empty(&net->dev, tvq))
 798                                 break;
 799                         cpu_relax();
 800                 }
 801
 802                 preempt_enable();
 803
 804                 if (!vhost_vq_avail_empty(&net->dev, tvq)) {
 805                         vhost_poll_queue(&tvq->poll);
 806                 } else if (unlikely(vhost_enable_notify(&net->dev, tvq))) {
 807                         vhost_disable_notify(&net->dev, tvq);
 808                         vhost_poll_queue(&tvq->poll);
 809                 }
 810
 811                 mutex_unlock(&tvq->mutex);
 812
 813                 len = peek_head_len(rnvq, sk);
 814         }
 815
 816         return len;
 817 }
 818
 819 /* This is a multi-buffer version of vhost_get_desc, that works if
 820  *      vq has read descriptors only.
 821  * @vq          - the relevant virtqueue
 822  * @datalen     - data length we'll be reading
 823  * @iovcount    - returned count of io vectors we fill
 824  * @log         - vhost log
 825  * @log_num     - log offset
 826  * @quota       - headcount quota, 1 for big buffer
 827  *      returns number of buffer heads allocated, negative on error
 828  */
 829 static int get_rx_bufs(struct vhost_virtqueue *vq,
 830                        struct vring_used_elem *heads,
 831                        int datalen,
 832                        unsigned *iovcount,
 833                        struct vhost_log *log,
 834                        unsigned *log_num,
 835                        unsigned int quota)
 836 {
 837         unsigned int out, in;
 838         int seg = 0;
 839         int headcount = 0;
 840         unsigned d;
 841         int r, nlogs = 0;
 842         /* len is always initialized before use since we are always called with
 843          * datalen > 0.
 844          */
 845         u32 uninitialized_var(len);
 846
 847         while (datalen > 0 && headcount < quota) {
 848                 if (unlikely(seg >= UIO_MAXIOV)) {
 849                         r = -ENOBUFS;
 850                         goto err;
 851                 }
 852                 r = vhost_get_vq_desc(vq, vq->iov + seg,
 853                                       ARRAY_SIZE(vq->iov) - seg, &out,
 854                                       &in, log, log_num);
 855                 if (unlikely(r < 0))
 856                         goto err;
 857
 858                 d = r;
 859                 if (d == vq->num) {
 860                         r = 0;
 861                         goto err;
 862                 }
 863                 if (unlikely(out || in <= 0)) {
 864                         vq_err(vq, "unexpected descriptor format for RX: "
 865                                 "out %d, in %d\n", out, in);
 866                         r = -EINVAL;
 867                         goto err;
 868                 }
 869                 if (unlikely(log)) {
 870                         nlogs += *log_num;
 871                         log += *log_num;
 872                 }
 873                 heads[headcount].id = cpu_to_vhost32(vq, d);
 874                 len = iov_length(vq->iov + seg, in);
 875                 heads[headcount].len = cpu_to_vhost32(vq, len);
 876                 datalen -= len;
 877                 ++headcount;
 878                 seg += in;
 879         }
 880         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
 881         *iovcount = seg;
 882         if (unlikely(log))
 883                 *log_num = nlogs;
 884
 885         /* Detect overrun */
 886         if (unlikely(datalen > 0)) {
 887                 r = UIO_MAXIOV + 1;
 888                 goto err;
 889         }
 890         return headcount;
 891 err:
 892         vhost_discard_vq_desc(vq, headcount);
 893         return r;
 894 }
 895
 896 /* Expects to be always run from workqueue - which acts as
 897  * read-size critical section for our kind of RCU. */
 898 static void handle_rx(struct vhost_net *net)
 899 {
 900         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 901         struct vhost_virtqueue *vq = &nvq->vq;
 902         unsigned uninitialized_var(in), log;
 903         struct vhost_log *vq_log;
 904         struct msghdr msg = {
 905                 .msg_name = NULL,
 906                 .msg_namelen = 0,
 907                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 908                 .msg_controllen = 0,
 909                 .msg_flags = MSG_DONTWAIT,
 910         };
 911         struct virtio_net_hdr hdr = {
 912                 .flags = 0,
 913                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
 914         };
 915         size_t total_len = 0;
 916         int err, mergeable;
 917         s16 headcount;
 918         size_t vhost_hlen, sock_hlen;
 919         size_t vhost_len, sock_len;
 920         bool busyloop_intr = false;
 921         struct socket *sock;
 922         struct iov_iter fixup;
 923         __virtio16 num_buffers;
 924         int recv_pkts = 0;
 925
 926         mutex_lock_nested(&vq->mutex, 0);
 927         sock = vq->private_data;
 928         if (!sock)
 929                 goto out;
 930
 931         if (!vq_iotlb_prefetch(vq))
 932                 goto out;
 933
 934         vhost_disable_notify(&net->dev, vq);
 935         vhost_net_disable_vq(net, vq);
 936
 937         vhost_hlen = nvq->vhost_hlen;
 938         sock_hlen = nvq->sock_hlen;
 939
 940         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
 941                 vq->log : NULL;
 942         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
 943
 944         while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
 945                                                       &busyloop_intr))) {
 946                 sock_len += sock_hlen;
 947                 vhost_len = sock_len + vhost_hlen;
 948                 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
 949                                         vhost_len, &in, vq_log, &log,
 950                                         likely(mergeable) ? UIO_MAXIOV : 1);
 951                 /* On error, stop handling until the next kick. */
 952                 if (unlikely(headcount < 0))
 953                         goto out;
 954                 /* OK, now we need to know about added descriptors. */
 955                 if (!headcount) {
 956                         if (unlikely(busyloop_intr)) {
 957                                 vhost_poll_queue(&vq->poll);
 958                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 959                                 /* They have slipped one in as we were
 960                                  * doing that: check again. */
 961                                 vhost_disable_notify(&net->dev, vq);
 962                                 continue;
 963                         }
 964                         /* Nothing new?  Wait for eventfd to tell us
 965                          * they refilled. */
 966                         goto out;
 967                 }
 968                 busyloop_intr = false;
 969                 if (nvq->rx_ring)
 970                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
 971                 /* On overrun, truncate and discard */
 972                 if (unlikely(headcount > UIO_MAXIOV)) {
 973                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
 974                         err = sock->ops->recvmsg(sock, &msg,
 975                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 976                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 977                         continue;
 978                 }
 979                 /* We don't need to be notified again. */
 980                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
 981                 fixup = msg.msg_iter;
 982                 if (unlikely((vhost_hlen))) {
 983                         /* We will supply the header ourselves
 984                          * TODO: support TSO.
 985                          */
 986                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
 987                 }
 988                 err = sock->ops->recvmsg(sock, &msg,
 989                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 990                 /* Userspace might have consumed the packet meanwhile:
 991                  * it's not supposed to do this usually, but might be hard
 992                  * to prevent. Discard data we got (if any) and keep going. */
 993                 if (unlikely(err != sock_len)) {
 994                         pr_debug("Discarded rx packet: "
 995                                  " len %d, expected %zd\n", err, sock_len);
 996                         vhost_discard_vq_desc(vq, headcount);
 997                         continue;
 998                 }
 999                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1000                 if (unlikely(vhost_hlen)) {
1001                         if (copy_to_iter(&hdr, sizeof(hdr),
1002                                          &fixup) != sizeof(hdr)) {
1003                                 vq_err(vq, "Unable to write vnet_hdr "
1004                                        "at addr %p\n", vq->iov->iov_base);
1005                                 goto out;
1006                         }
1007                 } else {
1008                         /* Header came from socket; we'll need to patch
1009                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1010                          */
1011                         iov_iter_advance(&fixup, sizeof(hdr));
1012                 }
1013                 /* TODO: Should check and handle checksum. */
1014
1015                 num_buffers = cpu_to_vhost16(vq, headcount);
1016                 if (likely(mergeable) &&
1017                     copy_to_iter(&num_buffers, sizeof num_buffers,
1018                                  &fixup) != sizeof num_buffers) {
1019                         vq_err(vq, "Failed num_buffers write");
1020                         vhost_discard_vq_desc(vq, headcount);
1021                         goto out;
1022                 }
1023                 nvq->done_idx += headcount;
1024                 if (nvq->done_idx > VHOST_NET_BATCH)
1025                         vhost_net_signal_used(nvq);
1026                 if (unlikely(vq_log))
1027                         vhost_log_write(vq, vq_log, log, vhost_len);
1028                 total_len += vhost_len;
1029                 if (unlikely(vhost_exceeds_weight(++recv_pkts, total_len))) {
1030                         vhost_poll_queue(&vq->poll);
1031                         goto out;
1032                 }
1033         }
1034         if (unlikely(busyloop_intr))
1035                 vhost_poll_queue(&vq->poll);
1036         else
1037                 vhost_net_enable_vq(net, vq);
1038 out:
1039         vhost_net_signal_used(nvq);
1040         mutex_unlock(&vq->mutex);
1041 }
1042
1043 static void handle_tx_kick(struct vhost_work *work)
1044 {
1045         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1046                                                   poll.work);
1047         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1048
1049         handle_tx(net);
1050 }
1051
1052 static void handle_rx_kick(struct vhost_work *work)
1053 {
1054         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1055                                                   poll.work);
1056         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1057
1058         handle_rx(net);
1059 }
1060
1061 static void handle_tx_net(struct vhost_work *work)
1062 {
1063         struct vhost_net *net = container_of(work, struct vhost_net,
1064                                              poll[VHOST_NET_VQ_TX].work);
1065         handle_tx(net);
1066 }
1067
1068 static void handle_rx_net(struct vhost_work *work)
1069 {
1070         struct vhost_net *net = container_of(work, struct vhost_net,
1071                                              poll[VHOST_NET_VQ_RX].work);
1072         handle_rx(net);
1073 }
1074
1075 static int vhost_net_open(struct inode *inode, struct file *f)
1076 {
1077         struct vhost_net *n;
1078         struct vhost_dev *dev;
1079         struct vhost_virtqueue **vqs;
1080         void **queue;
1081         int i;
1082
1083         n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1084         if (!n)
1085                 return -ENOMEM;
1086         vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1087         if (!vqs) {
1088                 kvfree(n);
1089                 return -ENOMEM;
1090         }
1091
1092         queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1093                               GFP_KERNEL);
1094         if (!queue) {
1095                 kfree(vqs);
1096                 kvfree(n);
1097                 return -ENOMEM;
1098         }
1099         n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1100
1101         dev = &n->dev;
1102         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1103         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1104         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1105         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1106         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1107                 n->vqs[i].ubufs = NULL;
1108                 n->vqs[i].ubuf_info = NULL;
1109                 n->vqs[i].upend_idx = 0;
1110                 n->vqs[i].done_idx = 0;
1111                 n->vqs[i].vhost_hlen = 0;
1112                 n->vqs[i].sock_hlen = 0;
1113                 n->vqs[i].rx_ring = NULL;
1114                 vhost_net_buf_init(&n->vqs[i].rxq);
1115         }
1116         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
1117
1118         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1119         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1120
1121         f->private_data = n;
1122
1123         return 0;
1124 }
1125
1126 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1127                                         struct vhost_virtqueue *vq)
1128 {
1129         struct socket *sock;
1130         struct vhost_net_virtqueue *nvq =
1131                 container_of(vq, struct vhost_net_virtqueue, vq);
1132
1133         mutex_lock(&vq->mutex);
1134         sock = vq->private_data;
1135         vhost_net_disable_vq(n, vq);
1136         vq->private_data = NULL;
1137         vhost_net_buf_unproduce(nvq);
1138         nvq->rx_ring = NULL;
1139         mutex_unlock(&vq->mutex);
1140         return sock;
1141 }
1142
1143 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1144                            struct socket **rx_sock)
1145 {
1146         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1147         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1148 }
1149
1150 static void vhost_net_flush_vq(struct vhost_net *n, int index)
1151 {
1152         vhost_poll_flush(n->poll + index);
1153         vhost_poll_flush(&n->vqs[index].vq.poll);
1154 }
1155
1156 static void vhost_net_flush(struct vhost_net *n)
1157 {
1158         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1159         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1160         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1161                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1162                 n->tx_flush = true;
1163                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1164                 /* Wait for all lower device DMAs done. */
1165                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1166                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1167                 n->tx_flush = false;
1168                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1169                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1170         }
1171 }
1172
1173 static int vhost_net_release(struct inode *inode, struct file *f)
1174 {
1175         struct vhost_net *n = f->private_data;
1176         struct socket *tx_sock;
1177         struct socket *rx_sock;
1178
1179         vhost_net_stop(n, &tx_sock, &rx_sock);
1180         vhost_net_flush(n);
1181         vhost_dev_stop(&n->dev);
1182         vhost_dev_cleanup(&n->dev);
1183         vhost_net_vq_reset(n);
1184         if (tx_sock)
1185                 sockfd_put(tx_sock);
1186         if (rx_sock)
1187                 sockfd_put(rx_sock);
1188         /* Make sure no callbacks are outstanding */
1189         synchronize_rcu_bh();
1190         /* We do an extra flush before freeing memory,
1191          * since jobs can re-queue themselves. */
1192         vhost_net_flush(n);
1193         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1194         kfree(n->dev.vqs);
1195         kvfree(n);
1196         return 0;
1197 }
1198
1199 static struct socket *get_raw_socket(int fd)
1200 {
1201         struct {
1202                 struct sockaddr_ll sa;
1203                 char  buf[MAX_ADDR_LEN];
1204         } uaddr;
1205         int r;
1206         struct socket *sock = sockfd_lookup(fd, &r);
1207
1208         if (!sock)
1209                 return ERR_PTR(-ENOTSOCK);
1210
1211         /* Parameter checking */
1212         if (sock->sk->sk_type != SOCK_RAW) {
1213                 r = -ESOCKTNOSUPPORT;
1214                 goto err;
1215         }
1216
1217         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa, 0);
1218         if (r < 0)
1219                 goto err;
1220
1221         if (uaddr.sa.sll_family != AF_PACKET) {
1222                 r = -EPFNOSUPPORT;
1223                 goto err;
1224         }
1225         return sock;
1226 err:
1227         sockfd_put(sock);
1228         return ERR_PTR(r);
1229 }
1230
1231 static struct ptr_ring *get_tap_ptr_ring(int fd)
1232 {
1233         struct ptr_ring *ring;
1234         struct file *file = fget(fd);
1235
1236         if (!file)
1237                 return NULL;
1238         ring = tun_get_tx_ring(file);
1239         if (!IS_ERR(ring))
1240                 goto out;
1241         ring = tap_get_ptr_ring(file);
1242         if (!IS_ERR(ring))
1243                 goto out;
1244         ring = NULL;
1245 out:
1246         fput(file);
1247         return ring;
1248 }
1249
1250 static struct socket *get_tap_socket(int fd)
1251 {
1252         struct file *file = fget(fd);
1253         struct socket *sock;
1254
1255         if (!file)
1256                 return ERR_PTR(-EBADF);
1257         sock = tun_get_socket(file);
1258         if (!IS_ERR(sock))
1259                 return sock;
1260         sock = tap_get_socket(file);
1261         if (IS_ERR(sock))
1262                 fput(file);
1263         return sock;
1264 }
1265
1266 static struct socket *get_socket(int fd)
1267 {
1268         struct socket *sock;
1269
1270         /* special case to disable backend */
1271         if (fd == -1)
1272                 return NULL;
1273         sock = get_raw_socket(fd);
1274         if (!IS_ERR(sock))
1275                 return sock;
1276         sock = get_tap_socket(fd);
1277         if (!IS_ERR(sock))
1278                 return sock;
1279         return ERR_PTR(-ENOTSOCK);
1280 }
1281
1282 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1283 {
1284         struct socket *sock, *oldsock;
1285         struct vhost_virtqueue *vq;
1286         struct vhost_net_virtqueue *nvq;
1287         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1288         int r;
1289
1290         mutex_lock(&n->dev.mutex);
1291         r = vhost_dev_check_owner(&n->dev);
1292         if (r)
1293                 goto err;
1294
1295         if (index >= VHOST_NET_VQ_MAX) {
1296                 r = -ENOBUFS;
1297                 goto err;
1298         }
1299         vq = &n->vqs[index].vq;
1300         nvq = &n->vqs[index];
1301         mutex_lock(&vq->mutex);
1302
1303         /* Verify that ring has been setup correctly. */
1304         if (!vhost_vq_access_ok(vq)) {
1305                 r = -EFAULT;
1306                 goto err_vq;
1307         }
1308         sock = get_socket(fd);
1309         if (IS_ERR(sock)) {
1310                 r = PTR_ERR(sock);
1311                 goto err_vq;
1312         }
1313
1314         /* start polling new socket */
1315         oldsock = vq->private_data;
1316         if (sock != oldsock) {
1317                 ubufs = vhost_net_ubuf_alloc(vq,
1318                                              sock && vhost_sock_zcopy(sock));
1319                 if (IS_ERR(ubufs)) {
1320                         r = PTR_ERR(ubufs);
1321                         goto err_ubufs;
1322                 }
1323
1324                 vhost_net_disable_vq(n, vq);
1325                 vq->private_data = sock;
1326                 vhost_net_buf_unproduce(nvq);
1327                 r = vhost_vq_init_access(vq);
1328                 if (r)
1329                         goto err_used;
1330                 r = vhost_net_enable_vq(n, vq);
1331                 if (r)
1332                         goto err_used;
1333                 if (index == VHOST_NET_VQ_RX)
1334                         nvq->rx_ring = get_tap_ptr_ring(fd);
1335
1336                 oldubufs = nvq->ubufs;
1337                 nvq->ubufs = ubufs;
1338
1339                 n->tx_packets = 0;
1340                 n->tx_zcopy_err = 0;
1341                 n->tx_flush = false;
1342         }
1343
1344         mutex_unlock(&vq->mutex);
1345
1346         if (oldubufs) {
1347                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1348                 mutex_lock(&vq->mutex);
1349                 vhost_zerocopy_signal_used(n, vq);
1350                 mutex_unlock(&vq->mutex);
1351         }
1352
1353         if (oldsock) {
1354                 vhost_net_flush_vq(n, index);
1355                 sockfd_put(oldsock);
1356         }
1357
1358         mutex_unlock(&n->dev.mutex);
1359         return 0;
1360
1361 err_used:
1362         vq->private_data = oldsock;
1363         vhost_net_enable_vq(n, vq);
1364         if (ubufs)
1365                 vhost_net_ubuf_put_wait_and_free(ubufs);
1366 err_ubufs:
1367         if (sock)
1368                 sockfd_put(sock);
1369 err_vq:
1370         mutex_unlock(&vq->mutex);
1371 err:
1372         mutex_unlock(&n->dev.mutex);
1373         return r;
1374 }
1375
1376 static long vhost_net_reset_owner(struct vhost_net *n)
1377 {
1378         struct socket *tx_sock = NULL;
1379         struct socket *rx_sock = NULL;
1380         long err;
1381         struct vhost_umem *umem;
1382
1383         mutex_lock(&n->dev.mutex);
1384         err = vhost_dev_check_owner(&n->dev);
1385         if (err)
1386                 goto done;
1387         umem = vhost_dev_reset_owner_prepare();
1388         if (!umem) {
1389                 err = -ENOMEM;
1390                 goto done;
1391         }
1392         vhost_net_stop(n, &tx_sock, &rx_sock);
1393         vhost_net_flush(n);
1394         vhost_dev_stop(&n->dev);
1395         vhost_dev_reset_owner(&n->dev, umem);
1396         vhost_net_vq_reset(n);
1397 done:
1398         mutex_unlock(&n->dev.mutex);
1399         if (tx_sock)
1400                 sockfd_put(tx_sock);
1401         if (rx_sock)
1402                 sockfd_put(rx_sock);
1403         return err;
1404 }
1405
1406 static int vhost_net_set_backend_features(struct vhost_net *n, u64 features)
1407 {
1408         int i;
1409
1410         mutex_lock(&n->dev.mutex);
1411         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1412                 mutex_lock(&n->vqs[i].vq.mutex);
1413                 n->vqs[i].vq.acked_backend_features = features;
1414                 mutex_unlock(&n->vqs[i].vq.mutex);
1415         }
1416         mutex_unlock(&n->dev.mutex);
1417
1418         return 0;
1419 }
1420
1421 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1422 {
1423         size_t vhost_hlen, sock_hlen, hdr_len;
1424         int i;
1425
1426         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1427                                (1ULL << VIRTIO_F_VERSION_1))) ?
1428                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1429                         sizeof(struct virtio_net_hdr);
1430         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1431                 /* vhost provides vnet_hdr */
1432                 vhost_hlen = hdr_len;
1433                 sock_hlen = 0;
1434         } else {
1435                 /* socket provides vnet_hdr */
1436                 vhost_hlen = 0;
1437                 sock_hlen = hdr_len;
1438         }
1439         mutex_lock(&n->dev.mutex);
1440         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1441             !vhost_log_access_ok(&n->dev))
1442                 goto out_unlock;
1443
1444         if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1445                 if (vhost_init_device_iotlb(&n->dev, true))
1446                         goto out_unlock;
1447         }
1448
1449         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1450                 mutex_lock(&n->vqs[i].vq.mutex);
1451                 n->vqs[i].vq.acked_features = features;
1452                 n->vqs[i].vhost_hlen = vhost_hlen;
1453                 n->vqs[i].sock_hlen = sock_hlen;
1454                 mutex_unlock(&n->vqs[i].vq.mutex);
1455         }
1456         mutex_unlock(&n->dev.mutex);
1457         return 0;
1458
1459 out_unlock:
1460         mutex_unlock(&n->dev.mutex);
1461         return -EFAULT;
1462 }
1463
1464 static long vhost_net_set_owner(struct vhost_net *n)
1465 {
1466         int r;
1467
1468         mutex_lock(&n->dev.mutex);
1469         if (vhost_dev_has_owner(&n->dev)) {
1470                 r = -EBUSY;
1471                 goto out;
1472         }
1473         r = vhost_net_set_ubuf_info(n);
1474         if (r)
1475                 goto out;
1476         r = vhost_dev_set_owner(&n->dev);
1477         if (r)
1478                 vhost_net_clear_ubuf_info(n);
1479         vhost_net_flush(n);
1480 out:
1481         mutex_unlock(&n->dev.mutex);
1482         return r;
1483 }
1484
1485 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1486                             unsigned long arg)
1487 {
1488         struct vhost_net *n = f->private_data;
1489         void __user *argp = (void __user *)arg;
1490         u64 __user *featurep = argp;
1491         struct vhost_vring_file backend;
1492         u64 features;
1493         int r;
1494
1495         switch (ioctl) {
1496         case VHOST_NET_SET_BACKEND:
1497                 if (copy_from_user(&backend, argp, sizeof backend))
1498                         return -EFAULT;
1499                 return vhost_net_set_backend(n, backend.index, backend.fd);
1500         case VHOST_GET_FEATURES:
1501                 features = VHOST_NET_FEATURES;
1502                 if (copy_to_user(featurep, &features, sizeof features))
1503                         return -EFAULT;
1504                 return 0;
1505         case VHOST_SET_FEATURES:
1506                 if (copy_from_user(&features, featurep, sizeof features))
1507                         return -EFAULT;
1508                 if (features & ~VHOST_NET_FEATURES)
1509                         return -EOPNOTSUPP;
1510                 return vhost_net_set_features(n, features);
1511         case VHOST_GET_BACKEND_FEATURES:
1512                 features = VHOST_NET_BACKEND_FEATURES;
1513                 if (copy_to_user(featurep, &features, sizeof(features)))
1514                         return -EFAULT;
1515                 return 0;
1516         case VHOST_SET_BACKEND_FEATURES:
1517                 if (copy_from_user(&features, featurep, sizeof(features)))
1518                         return -EFAULT;
1519                 if (features & ~VHOST_NET_BACKEND_FEATURES)
1520                         return -EOPNOTSUPP;
1521                 return vhost_net_set_backend_features(n, features);
1522         case VHOST_RESET_OWNER:
1523                 return vhost_net_reset_owner(n);
1524         case VHOST_SET_OWNER:
1525                 return vhost_net_set_owner(n);
1526         default:
1527                 mutex_lock(&n->dev.mutex);
1528                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1529                 if (r == -ENOIOCTLCMD)
1530                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1531                 else
1532                         vhost_net_flush(n);
1533                 mutex_unlock(&n->dev.mutex);
1534                 return r;
1535         }
1536 }
1537
1538 #ifdef CONFIG_COMPAT
1539 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1540                                    unsigned long arg)
1541 {
1542         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1543 }
1544 #endif
1545
1546 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1547 {
1548         struct file *file = iocb->ki_filp;
1549         struct vhost_net *n = file->private_data;
1550         struct vhost_dev *dev = &n->dev;
1551         int noblock = file->f_flags & O_NONBLOCK;
1552
1553         return vhost_chr_read_iter(dev, to, noblock);
1554 }
1555
1556 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1557                                         struct iov_iter *from)
1558 {
1559         struct file *file = iocb->ki_filp;
1560         struct vhost_net *n = file->private_data;
1561         struct vhost_dev *dev = &n->dev;
1562
1563         return vhost_chr_write_iter(dev, from);
1564 }
1565
1566 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1567 {
1568         struct vhost_net *n = file->private_data;
1569         struct vhost_dev *dev = &n->dev;
1570
1571         return vhost_chr_poll(file, dev, wait);
1572 }
1573
1574 static const struct file_operations vhost_net_fops = {
1575         .owner          = THIS_MODULE,
1576         .release        = vhost_net_release,
1577         .read_iter      = vhost_net_chr_read_iter,
1578         .write_iter     = vhost_net_chr_write_iter,
1579         .poll           = vhost_net_chr_poll,
1580         .unlocked_ioctl = vhost_net_ioctl,
1581 #ifdef CONFIG_COMPAT
1582         .compat_ioctl   = vhost_net_compat_ioctl,
1583 #endif
1584         .open           = vhost_net_open,
1585         .llseek         = noop_llseek,
1586 };
1587
1588 static struct miscdevice vhost_net_misc = {
1589         .minor = VHOST_NET_MINOR,
1590         .name = "vhost-net",
1591         .fops = &vhost_net_fops,
1592 };
1593
1594 static int vhost_net_init(void)
1595 {
1596         if (experimental_zcopytx)
1597                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1598         return misc_register(&vhost_net_misc);
1599 }
1600 module_init(vhost_net_init);
1601
1602 static void vhost_net_exit(void)
1603 {
1604         misc_deregister(&vhost_net_misc);
1605 }
1606 module_exit(vhost_net_exit);
1607
1608 MODULE_VERSION("0.0.1");
1609 MODULE_LICENSE("GPL v2");
1610 MODULE_AUTHOR("Michael S. Tsirkin");
1611 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1612 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1613 MODULE_ALIAS("devname:vhost-net");