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         /* Number of XDP frames batched */
 120         int batched_xdp;
 121         /* an array of userspace buffers info */
 122         struct ubuf_info *ubuf_info;
 123         /* Reference counting for outstanding ubufs.
 124          * Protected by vq mutex. Writers must also take device mutex. */
 125         struct vhost_net_ubuf_ref *ubufs;
 126         struct ptr_ring *rx_ring;
 127         struct vhost_net_buf rxq;
 128         /* Batched XDP buffs */
 129         struct xdp_buff *xdp;
 130 };
 131
 132 struct vhost_net {
 133         struct vhost_dev dev;
 134         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 135         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 136         /* Number of TX recently submitted.
 137          * Protected by tx vq lock. */
 138         unsigned tx_packets;
 139         /* Number of times zerocopy TX recently failed.
 140          * Protected by tx vq lock. */
 141         unsigned tx_zcopy_err;
 142         /* Flush in progress. Protected by tx vq lock. */
 143         bool tx_flush;
 144 };
 145
 146 static unsigned vhost_net_zcopy_mask __read_mostly;
 147
 148 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 149 {
 150         if (rxq->tail != rxq->head)
 151                 return rxq->queue[rxq->head];
 152         else
 153                 return NULL;
 154 }
 155
 156 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 157 {
 158         return rxq->tail - rxq->head;
 159 }
 160
 161 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 162 {
 163         return rxq->tail == rxq->head;
 164 }
 165
 166 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 167 {
 168         void *ret = vhost_net_buf_get_ptr(rxq);
 169         ++rxq->head;
 170         return ret;
 171 }
 172
 173 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 174 {
 175         struct vhost_net_buf *rxq = &nvq->rxq;
 176
 177         rxq->head = 0;
 178         rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
 179                                               VHOST_NET_BATCH);
 180         return rxq->tail;
 181 }
 182
 183 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 184 {
 185         struct vhost_net_buf *rxq = &nvq->rxq;
 186
 187         if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
 188                 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 189                                    vhost_net_buf_get_size(rxq),
 190                                    tun_ptr_free);
 191                 rxq->head = rxq->tail = 0;
 192         }
 193 }
 194
 195 static int vhost_net_buf_peek_len(void *ptr)
 196 {
 197         if (tun_is_xdp_frame(ptr)) {
 198                 struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
 199
 200                 return xdpf->len;
 201         }
 202
 203         return __skb_array_len_with_tag(ptr);
 204 }
 205
 206 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 207 {
 208         struct vhost_net_buf *rxq = &nvq->rxq;
 209
 210         if (!vhost_net_buf_is_empty(rxq))
 211                 goto out;
 212
 213         if (!vhost_net_buf_produce(nvq))
 214                 return 0;
 215
 216 out:
 217         return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 218 }
 219
 220 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 221 {
 222         rxq->head = rxq->tail = 0;
 223 }
 224
 225 static void vhost_net_enable_zcopy(int vq)
 226 {
 227         vhost_net_zcopy_mask |= 0x1 << vq;
 228 }
 229
 230 static struct vhost_net_ubuf_ref *
 231 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 232 {
 233         struct vhost_net_ubuf_ref *ubufs;
 234         /* No zero copy backend? Nothing to count. */
 235         if (!zcopy)
 236                 return NULL;
 237         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 238         if (!ubufs)
 239                 return ERR_PTR(-ENOMEM);
 240         atomic_set(&ubufs->refcount, 1);
 241         init_waitqueue_head(&ubufs->wait);
 242         ubufs->vq = vq;
 243         return ubufs;
 244 }
 245
 246 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 247 {
 248         int r = atomic_sub_return(1, &ubufs->refcount);
 249         if (unlikely(!r))
 250                 wake_up(&ubufs->wait);
 251         return r;
 252 }
 253
 254 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 255 {
 256         vhost_net_ubuf_put(ubufs);
 257         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 258 }
 259
 260 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 261 {
 262         vhost_net_ubuf_put_and_wait(ubufs);
 263         kfree(ubufs);
 264 }
 265
 266 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 267 {
 268         int i;
 269
 270         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 271                 kfree(n->vqs[i].ubuf_info);
 272                 n->vqs[i].ubuf_info = NULL;
 273         }
 274 }
 275
 276 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 277 {
 278         bool zcopy;
 279         int i;
 280
 281         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 282                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 283                 if (!zcopy)
 284                         continue;
 285                 n->vqs[i].ubuf_info =
 286                         kmalloc_array(UIO_MAXIOV,
 287                                       sizeof(*n->vqs[i].ubuf_info),
 288                                       GFP_KERNEL);
 289                 if  (!n->vqs[i].ubuf_info)
 290                         goto err;
 291         }
 292         return 0;
 293
 294 err:
 295         vhost_net_clear_ubuf_info(n);
 296         return -ENOMEM;
 297 }
 298
 299 static void vhost_net_vq_reset(struct vhost_net *n)
 300 {
 301         int i;
 302
 303         vhost_net_clear_ubuf_info(n);
 304
 305         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 306                 n->vqs[i].done_idx = 0;
 307                 n->vqs[i].upend_idx = 0;
 308                 n->vqs[i].ubufs = NULL;
 309                 n->vqs[i].vhost_hlen = 0;
 310                 n->vqs[i].sock_hlen = 0;
 311                 vhost_net_buf_init(&n->vqs[i].rxq);
 312         }
 313
 314 }
 315
 316 static void vhost_net_tx_packet(struct vhost_net *net)
 317 {
 318         ++net->tx_packets;
 319         if (net->tx_packets < 1024)
 320                 return;
 321         net->tx_packets = 0;
 322         net->tx_zcopy_err = 0;
 323 }
 324
 325 static void vhost_net_tx_err(struct vhost_net *net)
 326 {
 327         ++net->tx_zcopy_err;
 328 }
 329
 330 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 331 {
 332         /* TX flush waits for outstanding DMAs to be done.
 333          * Don't start new DMAs.
 334          */
 335         return !net->tx_flush &&
 336                 net->tx_packets / 64 >= net->tx_zcopy_err;
 337 }
 338
 339 static bool vhost_sock_zcopy(struct socket *sock)
 340 {
 341         return unlikely(experimental_zcopytx) &&
 342                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 343 }
 344
 345 static bool vhost_sock_xdp(struct socket *sock)
 346 {
 347         return sock_flag(sock->sk, SOCK_XDP);
 348 }
 349
 350 /* In case of DMA done not in order in lower device driver for some reason.
 351  * upend_idx is used to track end of used idx, done_idx is used to track head
 352  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 353  * guest used idx.
 354  */
 355 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 356                                        struct vhost_virtqueue *vq)
 357 {
 358         struct vhost_net_virtqueue *nvq =
 359                 container_of(vq, struct vhost_net_virtqueue, vq);
 360         int i, add;
 361         int j = 0;
 362
 363         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 364                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 365                         vhost_net_tx_err(net);
 366                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 367                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 368                         ++j;
 369                 } else
 370                         break;
 371         }
 372         while (j) {
 373                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 374                 vhost_add_used_and_signal_n(vq->dev, vq,
 375                                             &vq->heads[nvq->done_idx], add);
 376                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 377                 j -= add;
 378         }
 379 }
 380
 381 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 382 {
 383         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 384         struct vhost_virtqueue *vq = ubufs->vq;
 385         int cnt;
 386
 387         rcu_read_lock_bh();
 388
 389         /* set len to mark this desc buffers done DMA */
 390         vq->heads[ubuf->desc].len = success ?
 391                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 392         cnt = vhost_net_ubuf_put(ubufs);
 393
 394         /*
 395          * Trigger polling thread if guest stopped submitting new buffers:
 396          * in this case, the refcount after decrement will eventually reach 1.
 397          * We also trigger polling periodically after each 16 packets
 398          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 399          * less than 10% of times).
 400          */
 401         if (cnt <= 1 || !(cnt % 16))
 402                 vhost_poll_queue(&vq->poll);
 403
 404         rcu_read_unlock_bh();
 405 }
 406
 407 static inline unsigned long busy_clock(void)
 408 {
 409         return local_clock() >> 10;
 410 }
 411
 412 static bool vhost_can_busy_poll(unsigned long endtime)
 413 {
 414         return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
 415                       !signal_pending(current));
 416 }
 417
 418 static void vhost_net_disable_vq(struct vhost_net *n,
 419                                  struct vhost_virtqueue *vq)
 420 {
 421         struct vhost_net_virtqueue *nvq =
 422                 container_of(vq, struct vhost_net_virtqueue, vq);
 423         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 424         if (!vq->private_data)
 425                 return;
 426         vhost_poll_stop(poll);
 427 }
 428
 429 static int vhost_net_enable_vq(struct vhost_net *n,
 430                                 struct vhost_virtqueue *vq)
 431 {
 432         struct vhost_net_virtqueue *nvq =
 433                 container_of(vq, struct vhost_net_virtqueue, vq);
 434         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 435         struct socket *sock;
 436
 437         sock = vq->private_data;
 438         if (!sock)
 439                 return 0;
 440
 441         return vhost_poll_start(poll, sock->file);
 442 }
 443
 444 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
 445 {
 446         struct vhost_virtqueue *vq = &nvq->vq;
 447         struct vhost_dev *dev = vq->dev;
 448
 449         if (!nvq->done_idx)
 450                 return;
 451
 452         vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
 453         nvq->done_idx = 0;
 454 }
 455
 456 static void vhost_tx_batch(struct vhost_net *net,
 457                            struct vhost_net_virtqueue *nvq,
 458                            struct socket *sock,
 459                            struct msghdr *msghdr)
 460 {
 461         struct tun_msg_ctl ctl = {
 462                 .type = TUN_MSG_PTR,
 463                 .num = nvq->batched_xdp,
 464                 .ptr = nvq->xdp,
 465         };
 466         int err;
 467
 468         if (nvq->batched_xdp == 0)
 469                 goto signal_used;
 470
 471         msghdr->msg_control = &ctl;
 472         err = sock->ops->sendmsg(sock, msghdr, 0);
 473         if (unlikely(err < 0)) {
 474                 vq_err(&nvq->vq, "Fail to batch sending packets\n");
 475                 return;
 476         }
 477
 478 signal_used:
 479         vhost_net_signal_used(nvq);
 480         nvq->batched_xdp = 0;
 481 }
 482
 483 static int sock_has_rx_data(struct socket *sock)
 484 {
 485         if (unlikely(!sock))
 486                 return 0;
 487
 488         if (sock->ops->peek_len)
 489                 return sock->ops->peek_len(sock);
 490
 491         return skb_queue_empty(&sock->sk->sk_receive_queue);
 492 }
 493
 494 static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
 495                                           struct vhost_virtqueue *vq)
 496 {
 497         if (!vhost_vq_avail_empty(&net->dev, vq)) {
 498                 vhost_poll_queue(&vq->poll);
 499         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 500                 vhost_disable_notify(&net->dev, vq);
 501                 vhost_poll_queue(&vq->poll);
 502         }
 503 }
 504
 505 static void vhost_net_busy_poll(struct vhost_net *net,
 506                                 struct vhost_virtqueue *rvq,
 507                                 struct vhost_virtqueue *tvq,
 508                                 bool *busyloop_intr,
 509                                 bool poll_rx)
 510 {
 511         unsigned long busyloop_timeout;
 512         unsigned long endtime;
 513         struct socket *sock;
 514         struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
 515
 516         /* Try to hold the vq mutex of the paired virtqueue. We can't
 517          * use mutex_lock() here since we could not guarantee a
 518          * consistenet lock ordering.
 519          */
 520         if (!mutex_trylock(&vq->mutex))
 521                 return;
 522
 523         vhost_disable_notify(&net->dev, vq);
 524         sock = rvq->private_data;
 525
 526         busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
 527                                      tvq->busyloop_timeout;
 528
 529         preempt_disable();
 530         endtime = busy_clock() + busyloop_timeout;
 531
 532         while (vhost_can_busy_poll(endtime)) {
 533                 if (vhost_has_work(&net->dev)) {
 534                         *busyloop_intr = true;
 535                         break;
 536                 }
 537
 538                 if ((sock_has_rx_data(sock) &&
 539                      !vhost_vq_avail_empty(&net->dev, rvq)) ||
 540                     !vhost_vq_avail_empty(&net->dev, tvq))
 541                         break;
 542
 543                 cpu_relax();
 544         }
 545
 546         preempt_enable();
 547
 548         if (poll_rx || sock_has_rx_data(sock))
 549                 vhost_net_busy_poll_try_queue(net, vq);
 550         else if (!poll_rx) /* On tx here, sock has no rx data. */
 551                 vhost_enable_notify(&net->dev, rvq);
 552
 553         mutex_unlock(&vq->mutex);
 554 }
 555
 556 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 557                                     struct vhost_net_virtqueue *tnvq,
 558                                     unsigned int *out_num, unsigned int *in_num,
 559                                     struct msghdr *msghdr, bool *busyloop_intr)
 560 {
 561         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 562         struct vhost_virtqueue *rvq = &rnvq->vq;
 563         struct vhost_virtqueue *tvq = &tnvq->vq;
 564
 565         int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 566                                   out_num, in_num, NULL, NULL);
 567
 568         if (r == tvq->num && tvq->busyloop_timeout) {
 569                 /* Flush batched packets first */
 570                 if (!vhost_sock_zcopy(tvq->private_data))
 571                         vhost_tx_batch(net, tnvq, tvq->private_data, msghdr);
 572
 573                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
 574
 575                 r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 576                                       out_num, in_num, NULL, NULL);
 577         }
 578
 579         return r;
 580 }
 581
 582 static bool vhost_exceeds_maxpend(struct vhost_net *net)
 583 {
 584         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 585         struct vhost_virtqueue *vq = &nvq->vq;
 586
 587         return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 588                min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 589 }
 590
 591 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
 592                             size_t hdr_size, int out)
 593 {
 594         /* Skip header. TODO: support TSO. */
 595         size_t len = iov_length(vq->iov, out);
 596
 597         iov_iter_init(iter, WRITE, vq->iov, out, len);
 598         iov_iter_advance(iter, hdr_size);
 599
 600         return iov_iter_count(iter);
 601 }
 602
 603 static bool vhost_exceeds_weight(int pkts, int total_len)
 604 {
 605         return total_len >= VHOST_NET_WEIGHT ||
 606                pkts >= VHOST_NET_PKT_WEIGHT;
 607 }
 608
 609 static int get_tx_bufs(struct vhost_net *net,
 610                        struct vhost_net_virtqueue *nvq,
 611                        struct msghdr *msg,
 612                        unsigned int *out, unsigned int *in,
 613                        size_t *len, bool *busyloop_intr)
 614 {
 615         struct vhost_virtqueue *vq = &nvq->vq;
 616         int ret;
 617
 618         ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
 619
 620         if (ret < 0 || ret == vq->num)
 621                 return ret;
 622
 623         if (*in) {
 624                 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 625                         *out, *in);
 626                 return -EFAULT;
 627         }
 628
 629         /* Sanity check */
 630         *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 631         if (*len == 0) {
 632                 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 633                         *len, nvq->vhost_hlen);
 634                 return -EFAULT;
 635         }
 636
 637         return ret;
 638 }
 639
 640 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 641 {
 642         return total_len < VHOST_NET_WEIGHT &&
 643                !vhost_vq_avail_empty(vq->dev, vq);
 644 }
 645
 646 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 647
 648 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
 649                                struct iov_iter *from)
 650 {
 651         struct vhost_virtqueue *vq = &nvq->vq;
 652         struct socket *sock = vq->private_data;
 653         struct page_frag *alloc_frag = &current->task_frag;
 654         struct virtio_net_hdr *gso;
 655         struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
 656         struct tun_xdp_hdr *hdr;
 657         size_t len = iov_iter_count(from);
 658         int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
 659         int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 660         int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
 661         int sock_hlen = nvq->sock_hlen;
 662         void *buf;
 663         int copied;
 664
 665         if (unlikely(len < nvq->sock_hlen))
 666                 return -EFAULT;
 667
 668         if (SKB_DATA_ALIGN(len + pad) +
 669             SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
 670                 return -ENOSPC;
 671
 672         buflen += SKB_DATA_ALIGN(len + pad);
 673         alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
 674         if (unlikely(!skb_page_frag_refill(buflen, alloc_frag, GFP_KERNEL)))
 675                 return -ENOMEM;
 676
 677         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 678         copied = copy_page_from_iter(alloc_frag->page,
 679                                      alloc_frag->offset +
 680                                      offsetof(struct tun_xdp_hdr, gso),
 681                                      sock_hlen, from);
 682         if (copied != sock_hlen)
 683                 return -EFAULT;
 684
 685         hdr = buf;
 686         gso = &hdr->gso;
 687
 688         if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
 689             vhost16_to_cpu(vq, gso->csum_start) +
 690             vhost16_to_cpu(vq, gso->csum_offset) + 2 >
 691             vhost16_to_cpu(vq, gso->hdr_len)) {
 692                 gso->hdr_len = cpu_to_vhost16(vq,
 693                                vhost16_to_cpu(vq, gso->csum_start) +
 694                                vhost16_to_cpu(vq, gso->csum_offset) + 2);
 695
 696                 if (vhost16_to_cpu(vq, gso->hdr_len) > len)
 697                         return -EINVAL;
 698         }
 699
 700         len -= sock_hlen;
 701         copied = copy_page_from_iter(alloc_frag->page,
 702                                      alloc_frag->offset + pad,
 703                                      len, from);
 704         if (copied != len)
 705                 return -EFAULT;
 706
 707         xdp->data_hard_start = buf;
 708         xdp->data = buf + pad;
 709         xdp->data_end = xdp->data + len;
 710         hdr->buflen = buflen;
 711
 712         get_page(alloc_frag->page);
 713         alloc_frag->offset += buflen;
 714
 715         ++nvq->batched_xdp;
 716
 717         return 0;
 718 }
 719
 720 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 721 {
 722         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 723         struct vhost_virtqueue *vq = &nvq->vq;
 724         unsigned out, in;
 725         int head;
 726         struct msghdr msg = {
 727                 .msg_name = NULL,
 728                 .msg_namelen = 0,
 729                 .msg_control = NULL,
 730                 .msg_controllen = 0,
 731                 .msg_flags = MSG_DONTWAIT,
 732         };
 733         size_t len, total_len = 0;
 734         int err;
 735         int sent_pkts = 0;
 736         bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
 737
 738         for (;;) {
 739                 bool busyloop_intr = false;
 740
 741                 if (nvq->done_idx == VHOST_NET_BATCH)
 742                         vhost_tx_batch(net, nvq, sock, &msg);
 743
 744                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 745                                    &busyloop_intr);
 746                 /* On error, stop handling until the next kick. */
 747                 if (unlikely(head < 0))
 748                         break;
 749                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 750                 if (head == vq->num) {
 751                         if (unlikely(busyloop_intr)) {
 752                                 vhost_poll_queue(&vq->poll);
 753                         } else if (unlikely(vhost_enable_notify(&net->dev,
 754                                                                 vq))) {
 755                                 vhost_disable_notify(&net->dev, vq);
 756                                 continue;
 757                         }
 758                         break;
 759                 }
 760
 761                 total_len += len;
 762
 763                 /* For simplicity, TX batching is only enabled if
 764                  * sndbuf is unlimited.
 765                  */
 766                 if (sock_can_batch) {
 767                         err = vhost_net_build_xdp(nvq, &msg.msg_iter);
 768                         if (!err) {
 769                                 goto done;
 770                         } else if (unlikely(err != -ENOSPC)) {
 771                                 vhost_tx_batch(net, nvq, sock, &msg);
 772                                 vhost_discard_vq_desc(vq, 1);
 773                                 vhost_net_enable_vq(net, vq);
 774                                 break;
 775                         }
 776
 777                         /* We can't build XDP buff, go for single
 778                          * packet path but let's flush batched
 779                          * packets.
 780                          */
 781                         vhost_tx_batch(net, nvq, sock, &msg);
 782                         msg.msg_control = NULL;
 783                 } else {
 784                         if (tx_can_batch(vq, total_len))
 785                                 msg.msg_flags |= MSG_MORE;
 786                         else
 787                                 msg.msg_flags &= ~MSG_MORE;
 788                 }
 789
 790                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 791                 err = sock->ops->sendmsg(sock, &msg, len);
 792                 if (unlikely(err < 0)) {
 793                         vhost_discard_vq_desc(vq, 1);
 794                         vhost_net_enable_vq(net, vq);
 795                         break;
 796                 }
 797                 if (err != len)
 798                         pr_debug("Truncated TX packet: len %d != %zd\n",
 799                                  err, len);
 800 done:
 801                 vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 802                 vq->heads[nvq->done_idx].len = 0;
 803                 ++nvq->done_idx;
 804                 if (vhost_exceeds_weight(++sent_pkts, total_len)) {
 805                         vhost_poll_queue(&vq->poll);
 806                         break;
 807                 }
 808         }
 809
 810         vhost_tx_batch(net, nvq, sock, &msg);
 811 }
 812
 813 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 814 {
 815         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 816         struct vhost_virtqueue *vq = &nvq->vq;
 817         unsigned out, in;
 818         int head;
 819         struct msghdr msg = {
 820                 .msg_name = NULL,
 821                 .msg_namelen = 0,
 822                 .msg_control = NULL,
 823                 .msg_controllen = 0,
 824                 .msg_flags = MSG_DONTWAIT,
 825         };
 826         struct tun_msg_ctl ctl;
 827         size_t len, total_len = 0;
 828         int err;
 829         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 830         bool zcopy_used;
 831         int sent_pkts = 0;
 832
 833         for (;;) {
 834                 bool busyloop_intr;
 835
 836                 /* Release DMAs done buffers first */
 837                 vhost_zerocopy_signal_used(net, vq);
 838
 839                 busyloop_intr = false;
 840                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 841                                    &busyloop_intr);
 842                 /* On error, stop handling until the next kick. */
 843                 if (unlikely(head < 0))
 844                         break;
 845                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 846                 if (head == vq->num) {
 847                         if (unlikely(busyloop_intr)) {
 848                                 vhost_poll_queue(&vq->poll);
 849                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 850                                 vhost_disable_notify(&net->dev, vq);
 851                                 continue;
 852                         }
 853                         break;
 854                 }
 855
 856                 zcopy_used = len >= VHOST_GOODCOPY_LEN
 857                              && !vhost_exceeds_maxpend(net)
 858                              && vhost_net_tx_select_zcopy(net);
 859
 860                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 861                 if (zcopy_used) {
 862                         struct ubuf_info *ubuf;
 863                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 864
 865                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 866                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 867                         ubuf->callback = vhost_zerocopy_callback;
 868                         ubuf->ctx = nvq->ubufs;
 869                         ubuf->desc = nvq->upend_idx;
 870                         refcount_set(&ubuf->refcnt, 1);
 871                         msg.msg_control = &ctl;
 872                         ctl.type = TUN_MSG_UBUF;
 873                         ctl.ptr = ubuf;
 874                         msg.msg_controllen = sizeof(ctl);
 875                         ubufs = nvq->ubufs;
 876                         atomic_inc(&ubufs->refcount);
 877                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 878                 } else {
 879                         msg.msg_control = NULL;
 880                         ubufs = NULL;
 881                 }
 882                 total_len += len;
 883                 if (tx_can_batch(vq, total_len) &&
 884                     likely(!vhost_exceeds_maxpend(net))) {
 885                         msg.msg_flags |= MSG_MORE;
 886                 } else {
 887                         msg.msg_flags &= ~MSG_MORE;
 888                 }
 889
 890                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 891                 err = sock->ops->sendmsg(sock, &msg, len);
 892                 if (unlikely(err < 0)) {
 893                         if (zcopy_used) {
 894                                 vhost_net_ubuf_put(ubufs);
 895                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 896                                         % UIO_MAXIOV;
 897                         }
 898                         vhost_discard_vq_desc(vq, 1);
 899                         vhost_net_enable_vq(net, vq);
 900                         break;
 901                 }
 902                 if (err != len)
 903                         pr_debug("Truncated TX packet: "
 904                                  " len %d != %zd\n", err, len);
 905                 if (!zcopy_used)
 906                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 907                 else
 908                         vhost_zerocopy_signal_used(net, vq);
 909                 vhost_net_tx_packet(net);
 910                 if (unlikely(vhost_exceeds_weight(++sent_pkts, total_len))) {
 911                         vhost_poll_queue(&vq->poll);
 912                         break;
 913                 }
 914         }
 915 }
 916
 917 /* Expects to be always run from workqueue - which acts as
 918  * read-size critical section for our kind of RCU. */
 919 static void handle_tx(struct vhost_net *net)
 920 {
 921         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 922         struct vhost_virtqueue *vq = &nvq->vq;
 923         struct socket *sock;
 924
 925         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
 926         sock = vq->private_data;
 927         if (!sock)
 928                 goto out;
 929
 930         if (!vq_iotlb_prefetch(vq))
 931                 goto out;
 932
 933         vhost_disable_notify(&net->dev, vq);
 934         vhost_net_disable_vq(net, vq);
 935
 936         if (vhost_sock_zcopy(sock))
 937                 handle_tx_zerocopy(net, sock);
 938         else
 939                 handle_tx_copy(net, sock);
 940
 941 out:
 942         mutex_unlock(&vq->mutex);
 943 }
 944
 945 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 946 {
 947         struct sk_buff *head;
 948         int len = 0;
 949         unsigned long flags;
 950
 951         if (rvq->rx_ring)
 952                 return vhost_net_buf_peek(rvq);
 953
 954         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 955         head = skb_peek(&sk->sk_receive_queue);
 956         if (likely(head)) {
 957                 len = head->len;
 958                 if (skb_vlan_tag_present(head))
 959                         len += VLAN_HLEN;
 960         }
 961
 962         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 963         return len;
 964 }
 965
 966 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
 967                                       bool *busyloop_intr)
 968 {
 969         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 970         struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
 971         struct vhost_virtqueue *rvq = &rnvq->vq;
 972         struct vhost_virtqueue *tvq = &tnvq->vq;
 973         int len = peek_head_len(rnvq, sk);
 974
 975         if (!len && rvq->busyloop_timeout) {
 976                 /* Flush batched heads first */
 977                 vhost_net_signal_used(rnvq);
 978                 /* Both tx vq and rx socket were polled here */
 979                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
 980
 981                 len = peek_head_len(rnvq, sk);
 982         }
 983
 984         return len;
 985 }
 986
 987 /* This is a multi-buffer version of vhost_get_desc, that works if
 988  *      vq has read descriptors only.
 989  * @vq          - the relevant virtqueue
 990  * @datalen     - data length we'll be reading
 991  * @iovcount    - returned count of io vectors we fill
 992  * @log         - vhost log
 993  * @log_num     - log offset
 994  * @quota       - headcount quota, 1 for big buffer
 995  *      returns number of buffer heads allocated, negative on error
 996  */
 997 static int get_rx_bufs(struct vhost_virtqueue *vq,
 998                        struct vring_used_elem *heads,
 999                        int datalen,
1000                        unsigned *iovcount,
1001                        struct vhost_log *log,
1002                        unsigned *log_num,
1003                        unsigned int quota)
1004 {
1005         unsigned int out, in;
1006         int seg = 0;
1007         int headcount = 0;
1008         unsigned d;
1009         int r, nlogs = 0;
1010         /* len is always initialized before use since we are always called with
1011          * datalen > 0.
1012          */
1013         u32 uninitialized_var(len);
1014
1015         while (datalen > 0 && headcount < quota) {
1016                 if (unlikely(seg >= UIO_MAXIOV)) {
1017                         r = -ENOBUFS;
1018                         goto err;
1019                 }
1020                 r = vhost_get_vq_desc(vq, vq->iov + seg,
1021                                       ARRAY_SIZE(vq->iov) - seg, &out,
1022                                       &in, log, log_num);
1023                 if (unlikely(r < 0))
1024                         goto err;
1025
1026                 d = r;
1027                 if (d == vq->num) {
1028                         r = 0;
1029                         goto err;
1030                 }
1031                 if (unlikely(out || in <= 0)) {
1032                         vq_err(vq, "unexpected descriptor format for RX: "
1033                                 "out %d, in %d\n", out, in);
1034                         r = -EINVAL;
1035                         goto err;
1036                 }
1037                 if (unlikely(log)) {
1038                         nlogs += *log_num;
1039                         log += *log_num;
1040                 }
1041                 heads[headcount].id = cpu_to_vhost32(vq, d);
1042                 len = iov_length(vq->iov + seg, in);
1043                 heads[headcount].len = cpu_to_vhost32(vq, len);
1044                 datalen -= len;
1045                 ++headcount;
1046                 seg += in;
1047         }
1048         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1049         *iovcount = seg;
1050         if (unlikely(log))
1051                 *log_num = nlogs;
1052
1053         /* Detect overrun */
1054         if (unlikely(datalen > 0)) {
1055                 r = UIO_MAXIOV + 1;
1056                 goto err;
1057         }
1058         return headcount;
1059 err:
1060         vhost_discard_vq_desc(vq, headcount);
1061         return r;
1062 }
1063
1064 /* Expects to be always run from workqueue - which acts as
1065  * read-size critical section for our kind of RCU. */
1066 static void handle_rx(struct vhost_net *net)
1067 {
1068         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1069         struct vhost_virtqueue *vq = &nvq->vq;
1070         unsigned uninitialized_var(in), log;
1071         struct vhost_log *vq_log;
1072         struct msghdr msg = {
1073                 .msg_name = NULL,
1074                 .msg_namelen = 0,
1075                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
1076                 .msg_controllen = 0,
1077                 .msg_flags = MSG_DONTWAIT,
1078         };
1079         struct virtio_net_hdr hdr = {
1080                 .flags = 0,
1081                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
1082         };
1083         size_t total_len = 0;
1084         int err, mergeable;
1085         s16 headcount;
1086         size_t vhost_hlen, sock_hlen;
1087         size_t vhost_len, sock_len;
1088         bool busyloop_intr = false;
1089         struct socket *sock;
1090         struct iov_iter fixup;
1091         __virtio16 num_buffers;
1092         int recv_pkts = 0;
1093
1094         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1095         sock = vq->private_data;
1096         if (!sock)
1097                 goto out;
1098
1099         if (!vq_iotlb_prefetch(vq))
1100                 goto out;
1101
1102         vhost_disable_notify(&net->dev, vq);
1103         vhost_net_disable_vq(net, vq);
1104
1105         vhost_hlen = nvq->vhost_hlen;
1106         sock_hlen = nvq->sock_hlen;
1107
1108         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1109                 vq->log : NULL;
1110         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1111
1112         while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1113                                                       &busyloop_intr))) {
1114                 sock_len += sock_hlen;
1115                 vhost_len = sock_len + vhost_hlen;
1116                 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1117                                         vhost_len, &in, vq_log, &log,
1118                                         likely(mergeable) ? UIO_MAXIOV : 1);
1119                 /* On error, stop handling until the next kick. */
1120                 if (unlikely(headcount < 0))
1121                         goto out;
1122                 /* OK, now we need to know about added descriptors. */
1123                 if (!headcount) {
1124                         if (unlikely(busyloop_intr)) {
1125                                 vhost_poll_queue(&vq->poll);
1126                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1127                                 /* They have slipped one in as we were
1128                                  * doing that: check again. */
1129                                 vhost_disable_notify(&net->dev, vq);
1130                                 continue;
1131                         }
1132                         /* Nothing new?  Wait for eventfd to tell us
1133                          * they refilled. */
1134                         goto out;
1135                 }
1136                 busyloop_intr = false;
1137                 if (nvq->rx_ring)
1138                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1139                 /* On overrun, truncate and discard */
1140                 if (unlikely(headcount > UIO_MAXIOV)) {
1141                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
1142                         err = sock->ops->recvmsg(sock, &msg,
1143                                                  1, MSG_DONTWAIT | MSG_TRUNC);
1144                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
1145                         continue;
1146                 }
1147                 /* We don't need to be notified again. */
1148                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
1149                 fixup = msg.msg_iter;
1150                 if (unlikely((vhost_hlen))) {
1151                         /* We will supply the header ourselves
1152                          * TODO: support TSO.
1153                          */
1154                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
1155                 }
1156                 err = sock->ops->recvmsg(sock, &msg,
1157                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
1158                 /* Userspace might have consumed the packet meanwhile:
1159                  * it's not supposed to do this usually, but might be hard
1160                  * to prevent. Discard data we got (if any) and keep going. */
1161                 if (unlikely(err != sock_len)) {
1162                         pr_debug("Discarded rx packet: "
1163                                  " len %d, expected %zd\n", err, sock_len);
1164                         vhost_discard_vq_desc(vq, headcount);
1165                         continue;
1166                 }
1167                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1168                 if (unlikely(vhost_hlen)) {
1169                         if (copy_to_iter(&hdr, sizeof(hdr),
1170                                          &fixup) != sizeof(hdr)) {
1171                                 vq_err(vq, "Unable to write vnet_hdr "
1172                                        "at addr %p\n", vq->iov->iov_base);
1173                                 goto out;
1174                         }
1175                 } else {
1176                         /* Header came from socket; we'll need to patch
1177                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1178                          */
1179                         iov_iter_advance(&fixup, sizeof(hdr));
1180                 }
1181                 /* TODO: Should check and handle checksum. */
1182
1183                 num_buffers = cpu_to_vhost16(vq, headcount);
1184                 if (likely(mergeable) &&
1185                     copy_to_iter(&num_buffers, sizeof num_buffers,
1186                                  &fixup) != sizeof num_buffers) {
1187                         vq_err(vq, "Failed num_buffers write");
1188                         vhost_discard_vq_desc(vq, headcount);
1189                         goto out;
1190                 }
1191                 nvq->done_idx += headcount;
1192                 if (nvq->done_idx > VHOST_NET_BATCH)
1193                         vhost_net_signal_used(nvq);
1194                 if (unlikely(vq_log))
1195                         vhost_log_write(vq, vq_log, log, vhost_len);
1196                 total_len += vhost_len;
1197                 if (unlikely(vhost_exceeds_weight(++recv_pkts, total_len))) {
1198                         vhost_poll_queue(&vq->poll);
1199                         goto out;
1200                 }
1201         }
1202         if (unlikely(busyloop_intr))
1203                 vhost_poll_queue(&vq->poll);
1204         else
1205                 vhost_net_enable_vq(net, vq);
1206 out:
1207         vhost_net_signal_used(nvq);
1208         mutex_unlock(&vq->mutex);
1209 }
1210
1211 static void handle_tx_kick(struct vhost_work *work)
1212 {
1213         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1214                                                   poll.work);
1215         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1216
1217         handle_tx(net);
1218 }
1219
1220 static void handle_rx_kick(struct vhost_work *work)
1221 {
1222         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1223                                                   poll.work);
1224         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1225
1226         handle_rx(net);
1227 }
1228
1229 static void handle_tx_net(struct vhost_work *work)
1230 {
1231         struct vhost_net *net = container_of(work, struct vhost_net,
1232                                              poll[VHOST_NET_VQ_TX].work);
1233         handle_tx(net);
1234 }
1235
1236 static void handle_rx_net(struct vhost_work *work)
1237 {
1238         struct vhost_net *net = container_of(work, struct vhost_net,
1239                                              poll[VHOST_NET_VQ_RX].work);
1240         handle_rx(net);
1241 }
1242
1243 static int vhost_net_open(struct inode *inode, struct file *f)
1244 {
1245         struct vhost_net *n;
1246         struct vhost_dev *dev;
1247         struct vhost_virtqueue **vqs;
1248         void **queue;
1249         struct xdp_buff *xdp;
1250         int i;
1251
1252         n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1253         if (!n)
1254                 return -ENOMEM;
1255         vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1256         if (!vqs) {
1257                 kvfree(n);
1258                 return -ENOMEM;
1259         }
1260
1261         queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1262                               GFP_KERNEL);
1263         if (!queue) {
1264                 kfree(vqs);
1265                 kvfree(n);
1266                 return -ENOMEM;
1267         }
1268         n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1269
1270         xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1271         if (!xdp) {
1272                 kfree(vqs);
1273                 kvfree(n);
1274                 kfree(queue);
1275                 return -ENOMEM;
1276         }
1277         n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1278
1279         dev = &n->dev;
1280         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1281         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1282         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1283         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1284         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1285                 n->vqs[i].ubufs = NULL;
1286                 n->vqs[i].ubuf_info = NULL;
1287                 n->vqs[i].upend_idx = 0;
1288                 n->vqs[i].done_idx = 0;
1289                 n->vqs[i].batched_xdp = 0;
1290                 n->vqs[i].vhost_hlen = 0;
1291                 n->vqs[i].sock_hlen = 0;
1292                 n->vqs[i].rx_ring = NULL;
1293                 vhost_net_buf_init(&n->vqs[i].rxq);
1294         }
1295         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
1296
1297         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1298         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1299
1300         f->private_data = n;
1301
1302         return 0;
1303 }
1304
1305 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1306                                         struct vhost_virtqueue *vq)
1307 {
1308         struct socket *sock;
1309         struct vhost_net_virtqueue *nvq =
1310                 container_of(vq, struct vhost_net_virtqueue, vq);
1311
1312         mutex_lock(&vq->mutex);
1313         sock = vq->private_data;
1314         vhost_net_disable_vq(n, vq);
1315         vq->private_data = NULL;
1316         vhost_net_buf_unproduce(nvq);
1317         nvq->rx_ring = NULL;
1318         mutex_unlock(&vq->mutex);
1319         return sock;
1320 }
1321
1322 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1323                            struct socket **rx_sock)
1324 {
1325         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1326         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1327 }
1328
1329 static void vhost_net_flush_vq(struct vhost_net *n, int index)
1330 {
1331         vhost_poll_flush(n->poll + index);
1332         vhost_poll_flush(&n->vqs[index].vq.poll);
1333 }
1334
1335 static void vhost_net_flush(struct vhost_net *n)
1336 {
1337         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1338         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1339         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1340                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1341                 n->tx_flush = true;
1342                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1343                 /* Wait for all lower device DMAs done. */
1344                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1345                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1346                 n->tx_flush = false;
1347                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1348                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1349         }
1350 }
1351
1352 static int vhost_net_release(struct inode *inode, struct file *f)
1353 {
1354         struct vhost_net *n = f->private_data;
1355         struct socket *tx_sock;
1356         struct socket *rx_sock;
1357
1358         vhost_net_stop(n, &tx_sock, &rx_sock);
1359         vhost_net_flush(n);
1360         vhost_dev_stop(&n->dev);
1361         vhost_dev_cleanup(&n->dev);
1362         vhost_net_vq_reset(n);
1363         if (tx_sock)
1364                 sockfd_put(tx_sock);
1365         if (rx_sock)
1366                 sockfd_put(rx_sock);
1367         /* Make sure no callbacks are outstanding */
1368         synchronize_rcu();
1369         /* We do an extra flush before freeing memory,
1370          * since jobs can re-queue themselves. */
1371         vhost_net_flush(n);
1372         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1373         kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1374         kfree(n->dev.vqs);
1375         kvfree(n);
1376         return 0;
1377 }
1378
1379 static struct socket *get_raw_socket(int fd)
1380 {
1381         struct {
1382                 struct sockaddr_ll sa;
1383                 char  buf[MAX_ADDR_LEN];
1384         } uaddr;
1385         int r;
1386         struct socket *sock = sockfd_lookup(fd, &r);
1387
1388         if (!sock)
1389                 return ERR_PTR(-ENOTSOCK);
1390
1391         /* Parameter checking */
1392         if (sock->sk->sk_type != SOCK_RAW) {
1393                 r = -ESOCKTNOSUPPORT;
1394                 goto err;
1395         }
1396
1397         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa, 0);
1398         if (r < 0)
1399                 goto err;
1400
1401         if (uaddr.sa.sll_family != AF_PACKET) {
1402                 r = -EPFNOSUPPORT;
1403                 goto err;
1404         }
1405         return sock;
1406 err:
1407         sockfd_put(sock);
1408         return ERR_PTR(r);
1409 }
1410
1411 static struct ptr_ring *get_tap_ptr_ring(int fd)
1412 {
1413         struct ptr_ring *ring;
1414         struct file *file = fget(fd);
1415
1416         if (!file)
1417                 return NULL;
1418         ring = tun_get_tx_ring(file);
1419         if (!IS_ERR(ring))
1420                 goto out;
1421         ring = tap_get_ptr_ring(file);
1422         if (!IS_ERR(ring))
1423                 goto out;
1424         ring = NULL;
1425 out:
1426         fput(file);
1427         return ring;
1428 }
1429
1430 static struct socket *get_tap_socket(int fd)
1431 {
1432         struct file *file = fget(fd);
1433         struct socket *sock;
1434
1435         if (!file)
1436                 return ERR_PTR(-EBADF);
1437         sock = tun_get_socket(file);
1438         if (!IS_ERR(sock))
1439                 return sock;
1440         sock = tap_get_socket(file);
1441         if (IS_ERR(sock))
1442                 fput(file);
1443         return sock;
1444 }
1445
1446 static struct socket *get_socket(int fd)
1447 {
1448         struct socket *sock;
1449
1450         /* special case to disable backend */
1451         if (fd == -1)
1452                 return NULL;
1453         sock = get_raw_socket(fd);
1454         if (!IS_ERR(sock))
1455                 return sock;
1456         sock = get_tap_socket(fd);
1457         if (!IS_ERR(sock))
1458                 return sock;
1459         return ERR_PTR(-ENOTSOCK);
1460 }
1461
1462 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1463 {
1464         struct socket *sock, *oldsock;
1465         struct vhost_virtqueue *vq;
1466         struct vhost_net_virtqueue *nvq;
1467         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1468         int r;
1469
1470         mutex_lock(&n->dev.mutex);
1471         r = vhost_dev_check_owner(&n->dev);
1472         if (r)
1473                 goto err;
1474
1475         if (index >= VHOST_NET_VQ_MAX) {
1476                 r = -ENOBUFS;
1477                 goto err;
1478         }
1479         vq = &n->vqs[index].vq;
1480         nvq = &n->vqs[index];
1481         mutex_lock(&vq->mutex);
1482
1483         /* Verify that ring has been setup correctly. */
1484         if (!vhost_vq_access_ok(vq)) {
1485                 r = -EFAULT;
1486                 goto err_vq;
1487         }
1488         sock = get_socket(fd);
1489         if (IS_ERR(sock)) {
1490                 r = PTR_ERR(sock);
1491                 goto err_vq;
1492         }
1493
1494         /* start polling new socket */
1495         oldsock = vq->private_data;
1496         if (sock != oldsock) {
1497                 ubufs = vhost_net_ubuf_alloc(vq,
1498                                              sock && vhost_sock_zcopy(sock));
1499                 if (IS_ERR(ubufs)) {
1500                         r = PTR_ERR(ubufs);
1501                         goto err_ubufs;
1502                 }
1503
1504                 vhost_net_disable_vq(n, vq);
1505                 vq->private_data = sock;
1506                 vhost_net_buf_unproduce(nvq);
1507                 r = vhost_vq_init_access(vq);
1508                 if (r)
1509                         goto err_used;
1510                 r = vhost_net_enable_vq(n, vq);
1511                 if (r)
1512                         goto err_used;
1513                 if (index == VHOST_NET_VQ_RX)
1514                         nvq->rx_ring = get_tap_ptr_ring(fd);
1515
1516                 oldubufs = nvq->ubufs;
1517                 nvq->ubufs = ubufs;
1518
1519                 n->tx_packets = 0;
1520                 n->tx_zcopy_err = 0;
1521                 n->tx_flush = false;
1522         }
1523
1524         mutex_unlock(&vq->mutex);
1525
1526         if (oldubufs) {
1527                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1528                 mutex_lock(&vq->mutex);
1529                 vhost_zerocopy_signal_used(n, vq);
1530                 mutex_unlock(&vq->mutex);
1531         }
1532
1533         if (oldsock) {
1534                 vhost_net_flush_vq(n, index);
1535                 sockfd_put(oldsock);
1536         }
1537
1538         mutex_unlock(&n->dev.mutex);
1539         return 0;
1540
1541 err_used:
1542         vq->private_data = oldsock;
1543         vhost_net_enable_vq(n, vq);
1544         if (ubufs)
1545                 vhost_net_ubuf_put_wait_and_free(ubufs);
1546 err_ubufs:
1547         if (sock)
1548                 sockfd_put(sock);
1549 err_vq:
1550         mutex_unlock(&vq->mutex);
1551 err:
1552         mutex_unlock(&n->dev.mutex);
1553         return r;
1554 }
1555
1556 static long vhost_net_reset_owner(struct vhost_net *n)
1557 {
1558         struct socket *tx_sock = NULL;
1559         struct socket *rx_sock = NULL;
1560         long err;
1561         struct vhost_umem *umem;
1562
1563         mutex_lock(&n->dev.mutex);
1564         err = vhost_dev_check_owner(&n->dev);
1565         if (err)
1566                 goto done;
1567         umem = vhost_dev_reset_owner_prepare();
1568         if (!umem) {
1569                 err = -ENOMEM;
1570                 goto done;
1571         }
1572         vhost_net_stop(n, &tx_sock, &rx_sock);
1573         vhost_net_flush(n);
1574         vhost_dev_stop(&n->dev);
1575         vhost_dev_reset_owner(&n->dev, umem);
1576         vhost_net_vq_reset(n);
1577 done:
1578         mutex_unlock(&n->dev.mutex);
1579         if (tx_sock)
1580                 sockfd_put(tx_sock);
1581         if (rx_sock)
1582                 sockfd_put(rx_sock);
1583         return err;
1584 }
1585
1586 static int vhost_net_set_backend_features(struct vhost_net *n, u64 features)
1587 {
1588         int i;
1589
1590         mutex_lock(&n->dev.mutex);
1591         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1592                 mutex_lock(&n->vqs[i].vq.mutex);
1593                 n->vqs[i].vq.acked_backend_features = features;
1594                 mutex_unlock(&n->vqs[i].vq.mutex);
1595         }
1596         mutex_unlock(&n->dev.mutex);
1597
1598         return 0;
1599 }
1600
1601 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1602 {
1603         size_t vhost_hlen, sock_hlen, hdr_len;
1604         int i;
1605
1606         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1607                                (1ULL << VIRTIO_F_VERSION_1))) ?
1608                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1609                         sizeof(struct virtio_net_hdr);
1610         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1611                 /* vhost provides vnet_hdr */
1612                 vhost_hlen = hdr_len;
1613                 sock_hlen = 0;
1614         } else {
1615                 /* socket provides vnet_hdr */
1616                 vhost_hlen = 0;
1617                 sock_hlen = hdr_len;
1618         }
1619         mutex_lock(&n->dev.mutex);
1620         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1621             !vhost_log_access_ok(&n->dev))
1622                 goto out_unlock;
1623
1624         if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1625                 if (vhost_init_device_iotlb(&n->dev, true))
1626                         goto out_unlock;
1627         }
1628
1629         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1630                 mutex_lock(&n->vqs[i].vq.mutex);
1631                 n->vqs[i].vq.acked_features = features;
1632                 n->vqs[i].vhost_hlen = vhost_hlen;
1633                 n->vqs[i].sock_hlen = sock_hlen;
1634                 mutex_unlock(&n->vqs[i].vq.mutex);
1635         }
1636         mutex_unlock(&n->dev.mutex);
1637         return 0;
1638
1639 out_unlock:
1640         mutex_unlock(&n->dev.mutex);
1641         return -EFAULT;
1642 }
1643
1644 static long vhost_net_set_owner(struct vhost_net *n)
1645 {
1646         int r;
1647
1648         mutex_lock(&n->dev.mutex);
1649         if (vhost_dev_has_owner(&n->dev)) {
1650                 r = -EBUSY;
1651                 goto out;
1652         }
1653         r = vhost_net_set_ubuf_info(n);
1654         if (r)
1655                 goto out;
1656         r = vhost_dev_set_owner(&n->dev);
1657         if (r)
1658                 vhost_net_clear_ubuf_info(n);
1659         vhost_net_flush(n);
1660 out:
1661         mutex_unlock(&n->dev.mutex);
1662         return r;
1663 }
1664
1665 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1666                             unsigned long arg)
1667 {
1668         struct vhost_net *n = f->private_data;
1669         void __user *argp = (void __user *)arg;
1670         u64 __user *featurep = argp;
1671         struct vhost_vring_file backend;
1672         u64 features;
1673         int r;
1674
1675         switch (ioctl) {
1676         case VHOST_NET_SET_BACKEND:
1677                 if (copy_from_user(&backend, argp, sizeof backend))
1678                         return -EFAULT;
1679                 return vhost_net_set_backend(n, backend.index, backend.fd);
1680         case VHOST_GET_FEATURES:
1681                 features = VHOST_NET_FEATURES;
1682                 if (copy_to_user(featurep, &features, sizeof features))
1683                         return -EFAULT;
1684                 return 0;
1685         case VHOST_SET_FEATURES:
1686                 if (copy_from_user(&features, featurep, sizeof features))
1687                         return -EFAULT;
1688                 if (features & ~VHOST_NET_FEATURES)
1689                         return -EOPNOTSUPP;
1690                 return vhost_net_set_features(n, features);
1691         case VHOST_GET_BACKEND_FEATURES:
1692                 features = VHOST_NET_BACKEND_FEATURES;
1693                 if (copy_to_user(featurep, &features, sizeof(features)))
1694                         return -EFAULT;
1695                 return 0;
1696         case VHOST_SET_BACKEND_FEATURES:
1697                 if (copy_from_user(&features, featurep, sizeof(features)))
1698                         return -EFAULT;
1699                 if (features & ~VHOST_NET_BACKEND_FEATURES)
1700                         return -EOPNOTSUPP;
1701                 return vhost_net_set_backend_features(n, features);
1702         case VHOST_RESET_OWNER:
1703                 return vhost_net_reset_owner(n);
1704         case VHOST_SET_OWNER:
1705                 return vhost_net_set_owner(n);
1706         default:
1707                 mutex_lock(&n->dev.mutex);
1708                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1709                 if (r == -ENOIOCTLCMD)
1710                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1711                 else
1712                         vhost_net_flush(n);
1713                 mutex_unlock(&n->dev.mutex);
1714                 return r;
1715         }
1716 }
1717
1718 #ifdef CONFIG_COMPAT
1719 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1720                                    unsigned long arg)
1721 {
1722         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1723 }
1724 #endif
1725
1726 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1727 {
1728         struct file *file = iocb->ki_filp;
1729         struct vhost_net *n = file->private_data;
1730         struct vhost_dev *dev = &n->dev;
1731         int noblock = file->f_flags & O_NONBLOCK;
1732
1733         return vhost_chr_read_iter(dev, to, noblock);
1734 }
1735
1736 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1737                                         struct iov_iter *from)
1738 {
1739         struct file *file = iocb->ki_filp;
1740         struct vhost_net *n = file->private_data;
1741         struct vhost_dev *dev = &n->dev;
1742
1743         return vhost_chr_write_iter(dev, from);
1744 }
1745
1746 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1747 {
1748         struct vhost_net *n = file->private_data;
1749         struct vhost_dev *dev = &n->dev;
1750
1751         return vhost_chr_poll(file, dev, wait);
1752 }
1753
1754 static const struct file_operations vhost_net_fops = {
1755         .owner          = THIS_MODULE,
1756         .release        = vhost_net_release,
1757         .read_iter      = vhost_net_chr_read_iter,
1758         .write_iter     = vhost_net_chr_write_iter,
1759         .poll           = vhost_net_chr_poll,
1760         .unlocked_ioctl = vhost_net_ioctl,
1761 #ifdef CONFIG_COMPAT
1762         .compat_ioctl   = vhost_net_compat_ioctl,
1763 #endif
1764         .open           = vhost_net_open,
1765         .llseek         = noop_llseek,
1766 };
1767
1768 static struct miscdevice vhost_net_misc = {
1769         .minor = VHOST_NET_MINOR,
1770         .name = "vhost-net",
1771         .fops = &vhost_net_fops,
1772 };
1773
1774 static int vhost_net_init(void)
1775 {
1776         if (experimental_zcopytx)
1777                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1778         return misc_register(&vhost_net_misc);
1779 }
1780 module_init(vhost_net_init);
1781
1782 static void vhost_net_exit(void)
1783 {
1784         misc_deregister(&vhost_net_misc);
1785 }
1786 module_exit(vhost_net_exit);
1787
1788 MODULE_VERSION("0.0.1");
1789 MODULE_LICENSE("GPL v2");
1790 MODULE_AUTHOR("Michael S. Tsirkin");
1791 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1792 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1793 MODULE_ALIAS("devname:vhost-net");