drivers/vhost/net.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (C) 2009 Red Hat, Inc.
   3  * Author: Michael S. Tsirkin <[email protected]>
   4  *
   5  * virtio-net server in host kernel.
   6  */
   7
   8 #include <linux/compat.h>
   9 #include <linux/eventfd.h>
  10 #include <linux/vhost.h>
  11 #include <linux/virtio_net.h>
  12 #include <linux/miscdevice.h>
  13 #include <linux/module.h>
  14 #include <linux/moduleparam.h>
  15 #include <linux/mutex.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19 #include <linux/sched/clock.h>
  20 #include <linux/sched/signal.h>
  21 #include <linux/vmalloc.h>
  22
  23 #include <linux/net.h>
  24 #include <linux/if_packet.h>
  25 #include <linux/if_arp.h>
  26 #include <linux/if_tun.h>
  27 #include <linux/if_macvlan.h>
  28 #include <linux/if_tap.h>
  29 #include <linux/if_vlan.h>
  30 #include <linux/skb_array.h>
  31 #include <linux/skbuff.h>
  32
  33 #include <net/sock.h>
  34 #include <net/xdp.h>
  35
  36 #include "vhost.h"
  37
  38 static int experimental_zcopytx = 0;
  39 module_param(experimental_zcopytx, int, 0444);
  40 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  41                                        " 1 -Enable; 0 - Disable");
  42
  43 /* Max number of bytes transferred before requeueing the job.
  44  * Using this limit prevents one virtqueue from starving others. */
  45 #define VHOST_NET_WEIGHT 0x80000
  46
  47 /* Max number of packets transferred before requeueing the job.
  48  * Using this limit prevents one virtqueue from starving others with small
  49  * pkts.
  50  */
  51 #define VHOST_NET_PKT_WEIGHT 256
  52
  53 /* MAX number of TX used buffers for outstanding zerocopy */
  54 #define VHOST_MAX_PEND 128
  55 #define VHOST_GOODCOPY_LEN 256
  56
  57 /*
  58  * For transmit, used buffer len is unused; we override it to track buffer
  59  * status internally; used for zerocopy tx only.
  60  */
  61 /* Lower device DMA failed */
  62 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  63 /* Lower device DMA done */
  64 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  65 /* Lower device DMA in progress */
  66 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  67 /* Buffer unused */
  68 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  69
  70 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  71
  72 enum {
  73         VHOST_NET_FEATURES = VHOST_FEATURES |
  74                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  75                          (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  76                          (1ULL << VIRTIO_F_ACCESS_PLATFORM) |
  77                          (1ULL << VIRTIO_F_RING_RESET)
  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_msgzc *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         /* Private page frag */
 145         struct page_frag page_frag;
 146         /* Refcount bias of page frag */
 147         int refcnt_bias;
 148 };
 149
 150 static unsigned vhost_net_zcopy_mask __read_mostly;
 151
 152 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 153 {
 154         if (rxq->tail != rxq->head)
 155                 return rxq->queue[rxq->head];
 156         else
 157                 return NULL;
 158 }
 159
 160 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 161 {
 162         return rxq->tail - rxq->head;
 163 }
 164
 165 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 166 {
 167         return rxq->tail == rxq->head;
 168 }
 169
 170 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 171 {
 172         void *ret = vhost_net_buf_get_ptr(rxq);
 173         ++rxq->head;
 174         return ret;
 175 }
 176
 177 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 178 {
 179         struct vhost_net_buf *rxq = &nvq->rxq;
 180
 181         rxq->head = 0;
 182         rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
 183                                               VHOST_NET_BATCH);
 184         return rxq->tail;
 185 }
 186
 187 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 188 {
 189         struct vhost_net_buf *rxq = &nvq->rxq;
 190
 191         if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
 192                 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 193                                    vhost_net_buf_get_size(rxq),
 194                                    tun_ptr_free);
 195                 rxq->head = rxq->tail = 0;
 196         }
 197 }
 198
 199 static int vhost_net_buf_peek_len(void *ptr)
 200 {
 201         if (tun_is_xdp_frame(ptr)) {
 202                 struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
 203
 204                 return xdpf->len;
 205         }
 206
 207         return __skb_array_len_with_tag(ptr);
 208 }
 209
 210 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 211 {
 212         struct vhost_net_buf *rxq = &nvq->rxq;
 213
 214         if (!vhost_net_buf_is_empty(rxq))
 215                 goto out;
 216
 217         if (!vhost_net_buf_produce(nvq))
 218                 return 0;
 219
 220 out:
 221         return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 222 }
 223
 224 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 225 {
 226         rxq->head = rxq->tail = 0;
 227 }
 228
 229 static void vhost_net_enable_zcopy(int vq)
 230 {
 231         vhost_net_zcopy_mask |= 0x1 << vq;
 232 }
 233
 234 static struct vhost_net_ubuf_ref *
 235 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 236 {
 237         struct vhost_net_ubuf_ref *ubufs;
 238         /* No zero copy backend? Nothing to count. */
 239         if (!zcopy)
 240                 return NULL;
 241         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 242         if (!ubufs)
 243                 return ERR_PTR(-ENOMEM);
 244         atomic_set(&ubufs->refcount, 1);
 245         init_waitqueue_head(&ubufs->wait);
 246         ubufs->vq = vq;
 247         return ubufs;
 248 }
 249
 250 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 251 {
 252         int r = atomic_sub_return(1, &ubufs->refcount);
 253         if (unlikely(!r))
 254                 wake_up(&ubufs->wait);
 255         return r;
 256 }
 257
 258 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 259 {
 260         vhost_net_ubuf_put(ubufs);
 261         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 262 }
 263
 264 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 265 {
 266         vhost_net_ubuf_put_and_wait(ubufs);
 267         kfree(ubufs);
 268 }
 269
 270 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 271 {
 272         int i;
 273
 274         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 275                 kfree(n->vqs[i].ubuf_info);
 276                 n->vqs[i].ubuf_info = NULL;
 277         }
 278 }
 279
 280 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 281 {
 282         bool zcopy;
 283         int i;
 284
 285         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 286                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 287                 if (!zcopy)
 288                         continue;
 289                 n->vqs[i].ubuf_info =
 290                         kmalloc_array(UIO_MAXIOV,
 291                                       sizeof(*n->vqs[i].ubuf_info),
 292                                       GFP_KERNEL);
 293                 if  (!n->vqs[i].ubuf_info)
 294                         goto err;
 295         }
 296         return 0;
 297
 298 err:
 299         vhost_net_clear_ubuf_info(n);
 300         return -ENOMEM;
 301 }
 302
 303 static void vhost_net_vq_reset(struct vhost_net *n)
 304 {
 305         int i;
 306
 307         vhost_net_clear_ubuf_info(n);
 308
 309         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 310                 n->vqs[i].done_idx = 0;
 311                 n->vqs[i].upend_idx = 0;
 312                 n->vqs[i].ubufs = NULL;
 313                 n->vqs[i].vhost_hlen = 0;
 314                 n->vqs[i].sock_hlen = 0;
 315                 vhost_net_buf_init(&n->vqs[i].rxq);
 316         }
 317
 318 }
 319
 320 static void vhost_net_tx_packet(struct vhost_net *net)
 321 {
 322         ++net->tx_packets;
 323         if (net->tx_packets < 1024)
 324                 return;
 325         net->tx_packets = 0;
 326         net->tx_zcopy_err = 0;
 327 }
 328
 329 static void vhost_net_tx_err(struct vhost_net *net)
 330 {
 331         ++net->tx_zcopy_err;
 332 }
 333
 334 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 335 {
 336         /* TX flush waits for outstanding DMAs to be done.
 337          * Don't start new DMAs.
 338          */
 339         return !net->tx_flush &&
 340                 net->tx_packets / 64 >= net->tx_zcopy_err;
 341 }
 342
 343 static bool vhost_sock_zcopy(struct socket *sock)
 344 {
 345         return unlikely(experimental_zcopytx) &&
 346                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 347 }
 348
 349 static bool vhost_sock_xdp(struct socket *sock)
 350 {
 351         return sock_flag(sock->sk, SOCK_XDP);
 352 }
 353
 354 /* In case of DMA done not in order in lower device driver for some reason.
 355  * upend_idx is used to track end of used idx, done_idx is used to track head
 356  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 357  * guest used idx.
 358  */
 359 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 360                                        struct vhost_virtqueue *vq)
 361 {
 362         struct vhost_net_virtqueue *nvq =
 363                 container_of(vq, struct vhost_net_virtqueue, vq);
 364         int i, add;
 365         int j = 0;
 366
 367         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 368                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 369                         vhost_net_tx_err(net);
 370                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 371                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 372                         ++j;
 373                 } else
 374                         break;
 375         }
 376         while (j) {
 377                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 378                 vhost_add_used_and_signal_n(vq->dev, vq,
 379                                             &vq->heads[nvq->done_idx], add);
 380                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 381                 j -= add;
 382         }
 383 }
 384
 385 static void vhost_zerocopy_callback(struct sk_buff *skb,
 386                                     struct ubuf_info *ubuf_base, bool success)
 387 {
 388         struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
 389         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 390         struct vhost_virtqueue *vq = ubufs->vq;
 391         int cnt;
 392
 393         rcu_read_lock_bh();
 394
 395         /* set len to mark this desc buffers done DMA */
 396         vq->heads[ubuf->desc].len = success ?
 397                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 398         cnt = vhost_net_ubuf_put(ubufs);
 399
 400         /*
 401          * Trigger polling thread if guest stopped submitting new buffers:
 402          * in this case, the refcount after decrement will eventually reach 1.
 403          * We also trigger polling periodically after each 16 packets
 404          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 405          * less than 10% of times).
 406          */
 407         if (cnt <= 1 || !(cnt % 16))
 408                 vhost_poll_queue(&vq->poll);
 409
 410         rcu_read_unlock_bh();
 411 }
 412
 413 static inline unsigned long busy_clock(void)
 414 {
 415         return local_clock() >> 10;
 416 }
 417
 418 static bool vhost_can_busy_poll(unsigned long endtime)
 419 {
 420         return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
 421                       !signal_pending(current));
 422 }
 423
 424 static void vhost_net_disable_vq(struct vhost_net *n,
 425                                  struct vhost_virtqueue *vq)
 426 {
 427         struct vhost_net_virtqueue *nvq =
 428                 container_of(vq, struct vhost_net_virtqueue, vq);
 429         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 430         if (!vhost_vq_get_backend(vq))
 431                 return;
 432         vhost_poll_stop(poll);
 433 }
 434
 435 static int vhost_net_enable_vq(struct vhost_net *n,
 436                                 struct vhost_virtqueue *vq)
 437 {
 438         struct vhost_net_virtqueue *nvq =
 439                 container_of(vq, struct vhost_net_virtqueue, vq);
 440         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 441         struct socket *sock;
 442
 443         sock = vhost_vq_get_backend(vq);
 444         if (!sock)
 445                 return 0;
 446
 447         return vhost_poll_start(poll, sock->file);
 448 }
 449
 450 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
 451 {
 452         struct vhost_virtqueue *vq = &nvq->vq;
 453         struct vhost_dev *dev = vq->dev;
 454
 455         if (!nvq->done_idx)
 456                 return;
 457
 458         vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
 459         nvq->done_idx = 0;
 460 }
 461
 462 static void vhost_tx_batch(struct vhost_net *net,
 463                            struct vhost_net_virtqueue *nvq,
 464                            struct socket *sock,
 465                            struct msghdr *msghdr)
 466 {
 467         struct tun_msg_ctl ctl = {
 468                 .type = TUN_MSG_PTR,
 469                 .num = nvq->batched_xdp,
 470                 .ptr = nvq->xdp,
 471         };
 472         int i, err;
 473
 474         if (nvq->batched_xdp == 0)
 475                 goto signal_used;
 476
 477         msghdr->msg_control = &ctl;
 478         msghdr->msg_controllen = sizeof(ctl);
 479         err = sock->ops->sendmsg(sock, msghdr, 0);
 480         if (unlikely(err < 0)) {
 481                 vq_err(&nvq->vq, "Fail to batch sending packets\n");
 482
 483                 /* free pages owned by XDP; since this is an unlikely error path,
 484                  * keep it simple and avoid more complex bulk update for the
 485                  * used pages
 486                  */
 487                 for (i = 0; i < nvq->batched_xdp; ++i)
 488                         put_page(virt_to_head_page(nvq->xdp[i].data));
 489                 nvq->batched_xdp = 0;
 490                 nvq->done_idx = 0;
 491                 return;
 492         }
 493
 494 signal_used:
 495         vhost_net_signal_used(nvq);
 496         nvq->batched_xdp = 0;
 497 }
 498
 499 static int sock_has_rx_data(struct socket *sock)
 500 {
 501         if (unlikely(!sock))
 502                 return 0;
 503
 504         if (sock->ops->peek_len)
 505                 return sock->ops->peek_len(sock);
 506
 507         return skb_queue_empty(&sock->sk->sk_receive_queue);
 508 }
 509
 510 static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
 511                                           struct vhost_virtqueue *vq)
 512 {
 513         if (!vhost_vq_avail_empty(&net->dev, vq)) {
 514                 vhost_poll_queue(&vq->poll);
 515         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 516                 vhost_disable_notify(&net->dev, vq);
 517                 vhost_poll_queue(&vq->poll);
 518         }
 519 }
 520
 521 static void vhost_net_busy_poll(struct vhost_net *net,
 522                                 struct vhost_virtqueue *rvq,
 523                                 struct vhost_virtqueue *tvq,
 524                                 bool *busyloop_intr,
 525                                 bool poll_rx)
 526 {
 527         unsigned long busyloop_timeout;
 528         unsigned long endtime;
 529         struct socket *sock;
 530         struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
 531
 532         /* Try to hold the vq mutex of the paired virtqueue. We can't
 533          * use mutex_lock() here since we could not guarantee a
 534          * consistenet lock ordering.
 535          */
 536         if (!mutex_trylock(&vq->mutex))
 537                 return;
 538
 539         vhost_disable_notify(&net->dev, vq);
 540         sock = vhost_vq_get_backend(rvq);
 541
 542         busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
 543                                      tvq->busyloop_timeout;
 544
 545         preempt_disable();
 546         endtime = busy_clock() + busyloop_timeout;
 547
 548         while (vhost_can_busy_poll(endtime)) {
 549                 if (vhost_vq_has_work(vq)) {
 550                         *busyloop_intr = true;
 551                         break;
 552                 }
 553
 554                 if ((sock_has_rx_data(sock) &&
 555                      !vhost_vq_avail_empty(&net->dev, rvq)) ||
 556                     !vhost_vq_avail_empty(&net->dev, tvq))
 557                         break;
 558
 559                 cpu_relax();
 560         }
 561
 562         preempt_enable();
 563
 564         if (poll_rx || sock_has_rx_data(sock))
 565                 vhost_net_busy_poll_try_queue(net, vq);
 566         else if (!poll_rx) /* On tx here, sock has no rx data. */
 567                 vhost_enable_notify(&net->dev, rvq);
 568
 569         mutex_unlock(&vq->mutex);
 570 }
 571
 572 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 573                                     struct vhost_net_virtqueue *tnvq,
 574                                     unsigned int *out_num, unsigned int *in_num,
 575                                     struct msghdr *msghdr, bool *busyloop_intr)
 576 {
 577         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 578         struct vhost_virtqueue *rvq = &rnvq->vq;
 579         struct vhost_virtqueue *tvq = &tnvq->vq;
 580
 581         int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 582                                   out_num, in_num, NULL, NULL);
 583
 584         if (r == tvq->num && tvq->busyloop_timeout) {
 585                 /* Flush batched packets first */
 586                 if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq)))
 587                         vhost_tx_batch(net, tnvq,
 588                                        vhost_vq_get_backend(tvq),
 589                                        msghdr);
 590
 591                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
 592
 593                 r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 594                                       out_num, in_num, NULL, NULL);
 595         }
 596
 597         return r;
 598 }
 599
 600 static bool vhost_exceeds_maxpend(struct vhost_net *net)
 601 {
 602         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 603         struct vhost_virtqueue *vq = &nvq->vq;
 604
 605         return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 606                min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 607 }
 608
 609 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
 610                             size_t hdr_size, int out)
 611 {
 612         /* Skip header. TODO: support TSO. */
 613         size_t len = iov_length(vq->iov, out);
 614
 615         iov_iter_init(iter, ITER_SOURCE, vq->iov, out, len);
 616         iov_iter_advance(iter, hdr_size);
 617
 618         return iov_iter_count(iter);
 619 }
 620
 621 static int get_tx_bufs(struct vhost_net *net,
 622                        struct vhost_net_virtqueue *nvq,
 623                        struct msghdr *msg,
 624                        unsigned int *out, unsigned int *in,
 625                        size_t *len, bool *busyloop_intr)
 626 {
 627         struct vhost_virtqueue *vq = &nvq->vq;
 628         int ret;
 629
 630         ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
 631
 632         if (ret < 0 || ret == vq->num)
 633                 return ret;
 634
 635         if (*in) {
 636                 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 637                         *out, *in);
 638                 return -EFAULT;
 639         }
 640
 641         /* Sanity check */
 642         *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 643         if (*len == 0) {
 644                 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 645                         *len, nvq->vhost_hlen);
 646                 return -EFAULT;
 647         }
 648
 649         return ret;
 650 }
 651
 652 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 653 {
 654         return total_len < VHOST_NET_WEIGHT &&
 655                !vhost_vq_avail_empty(vq->dev, vq);
 656 }
 657
 658 static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
 659                                        struct page_frag *pfrag, gfp_t gfp)
 660 {
 661         if (pfrag->page) {
 662                 if (pfrag->offset + sz <= pfrag->size)
 663                         return true;
 664                 __page_frag_cache_drain(pfrag->page, net->refcnt_bias);
 665         }
 666
 667         pfrag->offset = 0;
 668         net->refcnt_bias = 0;
 669         if (SKB_FRAG_PAGE_ORDER) {
 670                 /* Avoid direct reclaim but allow kswapd to wake */
 671                 pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
 672                                           __GFP_COMP | __GFP_NOWARN |
 673                                           __GFP_NORETRY,
 674                                           SKB_FRAG_PAGE_ORDER);
 675                 if (likely(pfrag->page)) {
 676                         pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
 677                         goto done;
 678                 }
 679         }
 680         pfrag->page = alloc_page(gfp);
 681         if (likely(pfrag->page)) {
 682                 pfrag->size = PAGE_SIZE;
 683                 goto done;
 684         }
 685         return false;
 686
 687 done:
 688         net->refcnt_bias = USHRT_MAX;
 689         page_ref_add(pfrag->page, USHRT_MAX - 1);
 690         return true;
 691 }
 692
 693 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 694
 695 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
 696                                struct iov_iter *from)
 697 {
 698         struct vhost_virtqueue *vq = &nvq->vq;
 699         struct vhost_net *net = container_of(vq->dev, struct vhost_net,
 700                                              dev);
 701         struct socket *sock = vhost_vq_get_backend(vq);
 702         struct page_frag *alloc_frag = &net->page_frag;
 703         struct virtio_net_hdr *gso;
 704         struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
 705         struct tun_xdp_hdr *hdr;
 706         size_t len = iov_iter_count(from);
 707         int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
 708         int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 709         int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
 710         int sock_hlen = nvq->sock_hlen;
 711         void *buf;
 712         int copied;
 713
 714         if (unlikely(len < nvq->sock_hlen))
 715                 return -EFAULT;
 716
 717         if (SKB_DATA_ALIGN(len + pad) +
 718             SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
 719                 return -ENOSPC;
 720
 721         buflen += SKB_DATA_ALIGN(len + pad);
 722         alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
 723         if (unlikely(!vhost_net_page_frag_refill(net, buflen,
 724                                                  alloc_frag, GFP_KERNEL)))
 725                 return -ENOMEM;
 726
 727         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 728         copied = copy_page_from_iter(alloc_frag->page,
 729                                      alloc_frag->offset +
 730                                      offsetof(struct tun_xdp_hdr, gso),
 731                                      sock_hlen, from);
 732         if (copied != sock_hlen)
 733                 return -EFAULT;
 734
 735         hdr = buf;
 736         gso = &hdr->gso;
 737
 738         if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
 739             vhost16_to_cpu(vq, gso->csum_start) +
 740             vhost16_to_cpu(vq, gso->csum_offset) + 2 >
 741             vhost16_to_cpu(vq, gso->hdr_len)) {
 742                 gso->hdr_len = cpu_to_vhost16(vq,
 743                                vhost16_to_cpu(vq, gso->csum_start) +
 744                                vhost16_to_cpu(vq, gso->csum_offset) + 2);
 745
 746                 if (vhost16_to_cpu(vq, gso->hdr_len) > len)
 747                         return -EINVAL;
 748         }
 749
 750         len -= sock_hlen;
 751         copied = copy_page_from_iter(alloc_frag->page,
 752                                      alloc_frag->offset + pad,
 753                                      len, from);
 754         if (copied != len)
 755                 return -EFAULT;
 756
 757         xdp_init_buff(xdp, buflen, NULL);
 758         xdp_prepare_buff(xdp, buf, pad, len, true);
 759         hdr->buflen = buflen;
 760
 761         --net->refcnt_bias;
 762         alloc_frag->offset += buflen;
 763
 764         ++nvq->batched_xdp;
 765
 766         return 0;
 767 }
 768
 769 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 770 {
 771         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 772         struct vhost_virtqueue *vq = &nvq->vq;
 773         unsigned out, in;
 774         int head;
 775         struct msghdr msg = {
 776                 .msg_name = NULL,
 777                 .msg_namelen = 0,
 778                 .msg_control = NULL,
 779                 .msg_controllen = 0,
 780                 .msg_flags = MSG_DONTWAIT,
 781         };
 782         size_t len, total_len = 0;
 783         int err;
 784         int sent_pkts = 0;
 785         bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
 786
 787         do {
 788                 bool busyloop_intr = false;
 789
 790                 if (nvq->done_idx == VHOST_NET_BATCH)
 791                         vhost_tx_batch(net, nvq, sock, &msg);
 792
 793                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 794                                    &busyloop_intr);
 795                 /* On error, stop handling until the next kick. */
 796                 if (unlikely(head < 0))
 797                         break;
 798                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 799                 if (head == vq->num) {
 800                         if (unlikely(busyloop_intr)) {
 801                                 vhost_poll_queue(&vq->poll);
 802                         } else if (unlikely(vhost_enable_notify(&net->dev,
 803                                                                 vq))) {
 804                                 vhost_disable_notify(&net->dev, vq);
 805                                 continue;
 806                         }
 807                         break;
 808                 }
 809
 810                 total_len += len;
 811
 812                 /* For simplicity, TX batching is only enabled if
 813                  * sndbuf is unlimited.
 814                  */
 815                 if (sock_can_batch) {
 816                         err = vhost_net_build_xdp(nvq, &msg.msg_iter);
 817                         if (!err) {
 818                                 goto done;
 819                         } else if (unlikely(err != -ENOSPC)) {
 820                                 vhost_tx_batch(net, nvq, sock, &msg);
 821                                 vhost_discard_vq_desc(vq, 1);
 822                                 vhost_net_enable_vq(net, vq);
 823                                 break;
 824                         }
 825
 826                         /* We can't build XDP buff, go for single
 827                          * packet path but let's flush batched
 828                          * packets.
 829                          */
 830                         vhost_tx_batch(net, nvq, sock, &msg);
 831                         msg.msg_control = NULL;
 832                 } else {
 833                         if (tx_can_batch(vq, total_len))
 834                                 msg.msg_flags |= MSG_MORE;
 835                         else
 836                                 msg.msg_flags &= ~MSG_MORE;
 837                 }
 838
 839                 err = sock->ops->sendmsg(sock, &msg, len);
 840                 if (unlikely(err < 0)) {
 841                         if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) {
 842                                 vhost_discard_vq_desc(vq, 1);
 843                                 vhost_net_enable_vq(net, vq);
 844                                 break;
 845                         }
 846                         pr_debug("Fail to send packet: err %d", err);
 847                 } else if (unlikely(err != len))
 848                         pr_debug("Truncated TX packet: len %d != %zd\n",
 849                                  err, len);
 850 done:
 851                 vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 852                 vq->heads[nvq->done_idx].len = 0;
 853                 ++nvq->done_idx;
 854         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 855
 856         vhost_tx_batch(net, nvq, sock, &msg);
 857 }
 858
 859 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 860 {
 861         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 862         struct vhost_virtqueue *vq = &nvq->vq;
 863         unsigned out, in;
 864         int head;
 865         struct msghdr msg = {
 866                 .msg_name = NULL,
 867                 .msg_namelen = 0,
 868                 .msg_control = NULL,
 869                 .msg_controllen = 0,
 870                 .msg_flags = MSG_DONTWAIT,
 871         };
 872         struct tun_msg_ctl ctl;
 873         size_t len, total_len = 0;
 874         int err;
 875         struct vhost_net_ubuf_ref *ubufs;
 876         struct ubuf_info_msgzc *ubuf;
 877         bool zcopy_used;
 878         int sent_pkts = 0;
 879
 880         do {
 881                 bool busyloop_intr;
 882
 883                 /* Release DMAs done buffers first */
 884                 vhost_zerocopy_signal_used(net, vq);
 885
 886                 busyloop_intr = false;
 887                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 888                                    &busyloop_intr);
 889                 /* On error, stop handling until the next kick. */
 890                 if (unlikely(head < 0))
 891                         break;
 892                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 893                 if (head == vq->num) {
 894                         if (unlikely(busyloop_intr)) {
 895                                 vhost_poll_queue(&vq->poll);
 896                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 897                                 vhost_disable_notify(&net->dev, vq);
 898                                 continue;
 899                         }
 900                         break;
 901                 }
 902
 903                 zcopy_used = len >= VHOST_GOODCOPY_LEN
 904                              && !vhost_exceeds_maxpend(net)
 905                              && vhost_net_tx_select_zcopy(net);
 906
 907                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 908                 if (zcopy_used) {
 909                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 910                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 911                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 912                         ubuf->ctx = nvq->ubufs;
 913                         ubuf->desc = nvq->upend_idx;
 914                         ubuf->ubuf.callback = vhost_zerocopy_callback;
 915                         ubuf->ubuf.flags = SKBFL_ZEROCOPY_FRAG;
 916                         refcount_set(&ubuf->ubuf.refcnt, 1);
 917                         msg.msg_control = &ctl;
 918                         ctl.type = TUN_MSG_UBUF;
 919                         ctl.ptr = &ubuf->ubuf;
 920                         msg.msg_controllen = sizeof(ctl);
 921                         ubufs = nvq->ubufs;
 922                         atomic_inc(&ubufs->refcount);
 923                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 924                 } else {
 925                         msg.msg_control = NULL;
 926                         ubufs = NULL;
 927                 }
 928                 total_len += len;
 929                 if (tx_can_batch(vq, total_len) &&
 930                     likely(!vhost_exceeds_maxpend(net))) {
 931                         msg.msg_flags |= MSG_MORE;
 932                 } else {
 933                         msg.msg_flags &= ~MSG_MORE;
 934                 }
 935
 936                 err = sock->ops->sendmsg(sock, &msg, len);
 937                 if (unlikely(err < 0)) {
 938                         bool retry = err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS;
 939
 940                         if (zcopy_used) {
 941                                 if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
 942                                         vhost_net_ubuf_put(ubufs);
 943                                 if (retry)
 944                                         nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 945                                                 % UIO_MAXIOV;
 946                                 else
 947                                         vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
 948                         }
 949                         if (retry) {
 950                                 vhost_discard_vq_desc(vq, 1);
 951                                 vhost_net_enable_vq(net, vq);
 952                                 break;
 953                         }
 954                         pr_debug("Fail to send packet: err %d", err);
 955                 } else if (unlikely(err != len))
 956                         pr_debug("Truncated TX packet: "
 957                                  " len %d != %zd\n", err, len);
 958                 if (!zcopy_used)
 959                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 960                 else
 961                         vhost_zerocopy_signal_used(net, vq);
 962                 vhost_net_tx_packet(net);
 963         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 964 }
 965
 966 /* Expects to be always run from workqueue - which acts as
 967  * read-size critical section for our kind of RCU. */
 968 static void handle_tx(struct vhost_net *net)
 969 {
 970         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 971         struct vhost_virtqueue *vq = &nvq->vq;
 972         struct socket *sock;
 973
 974         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
 975         sock = vhost_vq_get_backend(vq);
 976         if (!sock)
 977                 goto out;
 978
 979         if (!vq_meta_prefetch(vq))
 980                 goto out;
 981
 982         vhost_disable_notify(&net->dev, vq);
 983         vhost_net_disable_vq(net, vq);
 984
 985         if (vhost_sock_zcopy(sock))
 986                 handle_tx_zerocopy(net, sock);
 987         else
 988                 handle_tx_copy(net, sock);
 989
 990 out:
 991         mutex_unlock(&vq->mutex);
 992 }
 993
 994 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 995 {
 996         struct sk_buff *head;
 997         int len = 0;
 998         unsigned long flags;
 999
1000         if (rvq->rx_ring)
1001                 return vhost_net_buf_peek(rvq);
1002
1003         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
1004         head = skb_peek(&sk->sk_receive_queue);
1005         if (likely(head)) {
1006                 len = head->len;
1007                 if (skb_vlan_tag_present(head))
1008                         len += VLAN_HLEN;
1009         }
1010
1011         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
1012         return len;
1013 }
1014
1015 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
1016                                       bool *busyloop_intr)
1017 {
1018         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1019         struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1020         struct vhost_virtqueue *rvq = &rnvq->vq;
1021         struct vhost_virtqueue *tvq = &tnvq->vq;
1022         int len = peek_head_len(rnvq, sk);
1023
1024         if (!len && rvq->busyloop_timeout) {
1025                 /* Flush batched heads first */
1026                 vhost_net_signal_used(rnvq);
1027                 /* Both tx vq and rx socket were polled here */
1028                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1029
1030                 len = peek_head_len(rnvq, sk);
1031         }
1032
1033         return len;
1034 }
1035
1036 /* This is a multi-buffer version of vhost_get_desc, that works if
1037  *      vq has read descriptors only.
1038  * @vq          - the relevant virtqueue
1039  * @datalen     - data length we'll be reading
1040  * @iovcount    - returned count of io vectors we fill
1041  * @log         - vhost log
1042  * @log_num     - log offset
1043  * @quota       - headcount quota, 1 for big buffer
1044  *      returns number of buffer heads allocated, negative on error
1045  */
1046 static int get_rx_bufs(struct vhost_virtqueue *vq,
1047                        struct vring_used_elem *heads,
1048                        int datalen,
1049                        unsigned *iovcount,
1050                        struct vhost_log *log,
1051                        unsigned *log_num,
1052                        unsigned int quota)
1053 {
1054         unsigned int out, in;
1055         int seg = 0;
1056         int headcount = 0;
1057         unsigned d;
1058         int r, nlogs = 0;
1059         /* len is always initialized before use since we are always called with
1060          * datalen > 0.
1061          */
1062         u32 len;
1063
1064         while (datalen > 0 && headcount < quota) {
1065                 if (unlikely(seg >= UIO_MAXIOV)) {
1066                         r = -ENOBUFS;
1067                         goto err;
1068                 }
1069                 r = vhost_get_vq_desc(vq, vq->iov + seg,
1070                                       ARRAY_SIZE(vq->iov) - seg, &out,
1071                                       &in, log, log_num);
1072                 if (unlikely(r < 0))
1073                         goto err;
1074
1075                 d = r;
1076                 if (d == vq->num) {
1077                         r = 0;
1078                         goto err;
1079                 }
1080                 if (unlikely(out || in <= 0)) {
1081                         vq_err(vq, "unexpected descriptor format for RX: "
1082                                 "out %d, in %d\n", out, in);
1083                         r = -EINVAL;
1084                         goto err;
1085                 }
1086                 if (unlikely(log)) {
1087                         nlogs += *log_num;
1088                         log += *log_num;
1089                 }
1090                 heads[headcount].id = cpu_to_vhost32(vq, d);
1091                 len = iov_length(vq->iov + seg, in);
1092                 heads[headcount].len = cpu_to_vhost32(vq, len);
1093                 datalen -= len;
1094                 ++headcount;
1095                 seg += in;
1096         }
1097         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1098         *iovcount = seg;
1099         if (unlikely(log))
1100                 *log_num = nlogs;
1101
1102         /* Detect overrun */
1103         if (unlikely(datalen > 0)) {
1104                 r = UIO_MAXIOV + 1;
1105                 goto err;
1106         }
1107         return headcount;
1108 err:
1109         vhost_discard_vq_desc(vq, headcount);
1110         return r;
1111 }
1112
1113 /* Expects to be always run from workqueue - which acts as
1114  * read-size critical section for our kind of RCU. */
1115 static void handle_rx(struct vhost_net *net)
1116 {
1117         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1118         struct vhost_virtqueue *vq = &nvq->vq;
1119         unsigned in, log;
1120         struct vhost_log *vq_log;
1121         struct msghdr msg = {
1122                 .msg_name = NULL,
1123                 .msg_namelen = 0,
1124                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
1125                 .msg_controllen = 0,
1126                 .msg_flags = MSG_DONTWAIT,
1127         };
1128         struct virtio_net_hdr hdr = {
1129                 .flags = 0,
1130                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
1131         };
1132         size_t total_len = 0;
1133         int err, mergeable;
1134         s16 headcount;
1135         size_t vhost_hlen, sock_hlen;
1136         size_t vhost_len, sock_len;
1137         bool busyloop_intr = false;
1138         struct socket *sock;
1139         struct iov_iter fixup;
1140         __virtio16 num_buffers;
1141         int recv_pkts = 0;
1142
1143         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1144         sock = vhost_vq_get_backend(vq);
1145         if (!sock)
1146                 goto out;
1147
1148         if (!vq_meta_prefetch(vq))
1149                 goto out;
1150
1151         vhost_disable_notify(&net->dev, vq);
1152         vhost_net_disable_vq(net, vq);
1153
1154         vhost_hlen = nvq->vhost_hlen;
1155         sock_hlen = nvq->sock_hlen;
1156
1157         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1158                 vq->log : NULL;
1159         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1160
1161         do {
1162                 sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1163                                                       &busyloop_intr);
1164                 if (!sock_len)
1165                         break;
1166                 sock_len += sock_hlen;
1167                 vhost_len = sock_len + vhost_hlen;
1168                 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1169                                         vhost_len, &in, vq_log, &log,
1170                                         likely(mergeable) ? UIO_MAXIOV : 1);
1171                 /* On error, stop handling until the next kick. */
1172                 if (unlikely(headcount < 0))
1173                         goto out;
1174                 /* OK, now we need to know about added descriptors. */
1175                 if (!headcount) {
1176                         if (unlikely(busyloop_intr)) {
1177                                 vhost_poll_queue(&vq->poll);
1178                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1179                                 /* They have slipped one in as we were
1180                                  * doing that: check again. */
1181                                 vhost_disable_notify(&net->dev, vq);
1182                                 continue;
1183                         }
1184                         /* Nothing new?  Wait for eventfd to tell us
1185                          * they refilled. */
1186                         goto out;
1187                 }
1188                 busyloop_intr = false;
1189                 if (nvq->rx_ring)
1190                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1191                 /* On overrun, truncate and discard */
1192                 if (unlikely(headcount > UIO_MAXIOV)) {
1193                         iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, 1, 1);
1194                         err = sock->ops->recvmsg(sock, &msg,
1195                                                  1, MSG_DONTWAIT | MSG_TRUNC);
1196                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
1197                         continue;
1198                 }
1199                 /* We don't need to be notified again. */
1200                 iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, in, vhost_len);
1201                 fixup = msg.msg_iter;
1202                 if (unlikely((vhost_hlen))) {
1203                         /* We will supply the header ourselves
1204                          * TODO: support TSO.
1205                          */
1206                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
1207                 }
1208                 err = sock->ops->recvmsg(sock, &msg,
1209                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
1210                 /* Userspace might have consumed the packet meanwhile:
1211                  * it's not supposed to do this usually, but might be hard
1212                  * to prevent. Discard data we got (if any) and keep going. */
1213                 if (unlikely(err != sock_len)) {
1214                         pr_debug("Discarded rx packet: "
1215                                  " len %d, expected %zd\n", err, sock_len);
1216                         vhost_discard_vq_desc(vq, headcount);
1217                         continue;
1218                 }
1219                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1220                 if (unlikely(vhost_hlen)) {
1221                         if (copy_to_iter(&hdr, sizeof(hdr),
1222                                          &fixup) != sizeof(hdr)) {
1223                                 vq_err(vq, "Unable to write vnet_hdr "
1224                                        "at addr %p\n", vq->iov->iov_base);
1225                                 goto out;
1226                         }
1227                 } else {
1228                         /* Header came from socket; we'll need to patch
1229                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1230                          */
1231                         iov_iter_advance(&fixup, sizeof(hdr));
1232                 }
1233                 /* TODO: Should check and handle checksum. */
1234
1235                 num_buffers = cpu_to_vhost16(vq, headcount);
1236                 if (likely(mergeable) &&
1237                     copy_to_iter(&num_buffers, sizeof num_buffers,
1238                                  &fixup) != sizeof num_buffers) {
1239                         vq_err(vq, "Failed num_buffers write");
1240                         vhost_discard_vq_desc(vq, headcount);
1241                         goto out;
1242                 }
1243                 nvq->done_idx += headcount;
1244                 if (nvq->done_idx > VHOST_NET_BATCH)
1245                         vhost_net_signal_used(nvq);
1246                 if (unlikely(vq_log))
1247                         vhost_log_write(vq, vq_log, log, vhost_len,
1248                                         vq->iov, in);
1249                 total_len += vhost_len;
1250         } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1251
1252         if (unlikely(busyloop_intr))
1253                 vhost_poll_queue(&vq->poll);
1254         else if (!sock_len)
1255                 vhost_net_enable_vq(net, vq);
1256 out:
1257         vhost_net_signal_used(nvq);
1258         mutex_unlock(&vq->mutex);
1259 }
1260
1261 static void handle_tx_kick(struct vhost_work *work)
1262 {
1263         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1264                                                   poll.work);
1265         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1266
1267         handle_tx(net);
1268 }
1269
1270 static void handle_rx_kick(struct vhost_work *work)
1271 {
1272         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1273                                                   poll.work);
1274         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1275
1276         handle_rx(net);
1277 }
1278
1279 static void handle_tx_net(struct vhost_work *work)
1280 {
1281         struct vhost_net *net = container_of(work, struct vhost_net,
1282                                              poll[VHOST_NET_VQ_TX].work);
1283         handle_tx(net);
1284 }
1285
1286 static void handle_rx_net(struct vhost_work *work)
1287 {
1288         struct vhost_net *net = container_of(work, struct vhost_net,
1289                                              poll[VHOST_NET_VQ_RX].work);
1290         handle_rx(net);
1291 }
1292
1293 static int vhost_net_open(struct inode *inode, struct file *f)
1294 {
1295         struct vhost_net *n;
1296         struct vhost_dev *dev;
1297         struct vhost_virtqueue **vqs;
1298         void **queue;
1299         struct xdp_buff *xdp;
1300         int i;
1301
1302         n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1303         if (!n)
1304                 return -ENOMEM;
1305         vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1306         if (!vqs) {
1307                 kvfree(n);
1308                 return -ENOMEM;
1309         }
1310
1311         queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1312                               GFP_KERNEL);
1313         if (!queue) {
1314                 kfree(vqs);
1315                 kvfree(n);
1316                 return -ENOMEM;
1317         }
1318         n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1319
1320         xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1321         if (!xdp) {
1322                 kfree(vqs);
1323                 kvfree(n);
1324                 kfree(queue);
1325                 return -ENOMEM;
1326         }
1327         n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1328
1329         dev = &n->dev;
1330         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1331         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1332         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1333         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1334         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1335                 n->vqs[i].ubufs = NULL;
1336                 n->vqs[i].ubuf_info = NULL;
1337                 n->vqs[i].upend_idx = 0;
1338                 n->vqs[i].done_idx = 0;
1339                 n->vqs[i].batched_xdp = 0;
1340                 n->vqs[i].vhost_hlen = 0;
1341                 n->vqs[i].sock_hlen = 0;
1342                 n->vqs[i].rx_ring = NULL;
1343                 vhost_net_buf_init(&n->vqs[i].rxq);
1344         }
1345         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1346                        UIO_MAXIOV + VHOST_NET_BATCH,
1347                        VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true,
1348                        NULL);
1349
1350         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev,
1351                         vqs[VHOST_NET_VQ_TX]);
1352         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev,
1353                         vqs[VHOST_NET_VQ_RX]);
1354
1355         f->private_data = n;
1356         n->page_frag.page = NULL;
1357         n->refcnt_bias = 0;
1358
1359         return 0;
1360 }
1361
1362 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1363                                         struct vhost_virtqueue *vq)
1364 {
1365         struct socket *sock;
1366         struct vhost_net_virtqueue *nvq =
1367                 container_of(vq, struct vhost_net_virtqueue, vq);
1368
1369         mutex_lock(&vq->mutex);
1370         sock = vhost_vq_get_backend(vq);
1371         vhost_net_disable_vq(n, vq);
1372         vhost_vq_set_backend(vq, NULL);
1373         vhost_net_buf_unproduce(nvq);
1374         nvq->rx_ring = NULL;
1375         mutex_unlock(&vq->mutex);
1376         return sock;
1377 }
1378
1379 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1380                            struct socket **rx_sock)
1381 {
1382         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1383         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1384 }
1385
1386 static void vhost_net_flush(struct vhost_net *n)
1387 {
1388         vhost_dev_flush(&n->dev);
1389         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1390                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1391                 n->tx_flush = true;
1392                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1393                 /* Wait for all lower device DMAs done. */
1394                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1395                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1396                 n->tx_flush = false;
1397                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1398                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1399         }
1400 }
1401
1402 static int vhost_net_release(struct inode *inode, struct file *f)
1403 {
1404         struct vhost_net *n = f->private_data;
1405         struct socket *tx_sock;
1406         struct socket *rx_sock;
1407
1408         vhost_net_stop(n, &tx_sock, &rx_sock);
1409         vhost_net_flush(n);
1410         vhost_dev_stop(&n->dev);
1411         vhost_dev_cleanup(&n->dev);
1412         vhost_net_vq_reset(n);
1413         if (tx_sock)
1414                 sockfd_put(tx_sock);
1415         if (rx_sock)
1416                 sockfd_put(rx_sock);
1417         /* Make sure no callbacks are outstanding */
1418         synchronize_rcu();
1419         /* We do an extra flush before freeing memory,
1420          * since jobs can re-queue themselves. */
1421         vhost_net_flush(n);
1422         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1423         kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1424         kfree(n->dev.vqs);
1425         if (n->page_frag.page)
1426                 __page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
1427         kvfree(n);
1428         return 0;
1429 }
1430
1431 static struct socket *get_raw_socket(int fd)
1432 {
1433         int r;
1434         struct socket *sock = sockfd_lookup(fd, &r);
1435
1436         if (!sock)
1437                 return ERR_PTR(-ENOTSOCK);
1438
1439         /* Parameter checking */
1440         if (sock->sk->sk_type != SOCK_RAW) {
1441                 r = -ESOCKTNOSUPPORT;
1442                 goto err;
1443         }
1444
1445         if (sock->sk->sk_family != AF_PACKET) {
1446                 r = -EPFNOSUPPORT;
1447                 goto err;
1448         }
1449         return sock;
1450 err:
1451         sockfd_put(sock);
1452         return ERR_PTR(r);
1453 }
1454
1455 static struct ptr_ring *get_tap_ptr_ring(struct file *file)
1456 {
1457         struct ptr_ring *ring;
1458         ring = tun_get_tx_ring(file);
1459         if (!IS_ERR(ring))
1460                 goto out;
1461         ring = tap_get_ptr_ring(file);
1462         if (!IS_ERR(ring))
1463                 goto out;
1464         ring = NULL;
1465 out:
1466         return ring;
1467 }
1468
1469 static struct socket *get_tap_socket(int fd)
1470 {
1471         struct file *file = fget(fd);
1472         struct socket *sock;
1473
1474         if (!file)
1475                 return ERR_PTR(-EBADF);
1476         sock = tun_get_socket(file);
1477         if (!IS_ERR(sock))
1478                 return sock;
1479         sock = tap_get_socket(file);
1480         if (IS_ERR(sock))
1481                 fput(file);
1482         return sock;
1483 }
1484
1485 static struct socket *get_socket(int fd)
1486 {
1487         struct socket *sock;
1488
1489         /* special case to disable backend */
1490         if (fd == -1)
1491                 return NULL;
1492         sock = get_raw_socket(fd);
1493         if (!IS_ERR(sock))
1494                 return sock;
1495         sock = get_tap_socket(fd);
1496         if (!IS_ERR(sock))
1497                 return sock;
1498         return ERR_PTR(-ENOTSOCK);
1499 }
1500
1501 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1502 {
1503         struct socket *sock, *oldsock;
1504         struct vhost_virtqueue *vq;
1505         struct vhost_net_virtqueue *nvq;
1506         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1507         int r;
1508
1509         mutex_lock(&n->dev.mutex);
1510         r = vhost_dev_check_owner(&n->dev);
1511         if (r)
1512                 goto err;
1513
1514         if (index >= VHOST_NET_VQ_MAX) {
1515                 r = -ENOBUFS;
1516                 goto err;
1517         }
1518         vq = &n->vqs[index].vq;
1519         nvq = &n->vqs[index];
1520         mutex_lock(&vq->mutex);
1521
1522         if (fd == -1)
1523                 vhost_clear_msg(&n->dev);
1524
1525         /* Verify that ring has been setup correctly. */
1526         if (!vhost_vq_access_ok(vq)) {
1527                 r = -EFAULT;
1528                 goto err_vq;
1529         }
1530         sock = get_socket(fd);
1531         if (IS_ERR(sock)) {
1532                 r = PTR_ERR(sock);
1533                 goto err_vq;
1534         }
1535
1536         /* start polling new socket */
1537         oldsock = vhost_vq_get_backend(vq);
1538         if (sock != oldsock) {
1539                 ubufs = vhost_net_ubuf_alloc(vq,
1540                                              sock && vhost_sock_zcopy(sock));
1541                 if (IS_ERR(ubufs)) {
1542                         r = PTR_ERR(ubufs);
1543                         goto err_ubufs;
1544                 }
1545
1546                 vhost_net_disable_vq(n, vq);
1547                 vhost_vq_set_backend(vq, sock);
1548                 vhost_net_buf_unproduce(nvq);
1549                 r = vhost_vq_init_access(vq);
1550                 if (r)
1551                         goto err_used;
1552                 r = vhost_net_enable_vq(n, vq);
1553                 if (r)
1554                         goto err_used;
1555                 if (index == VHOST_NET_VQ_RX) {
1556                         if (sock)
1557                                 nvq->rx_ring = get_tap_ptr_ring(sock->file);
1558                         else
1559                                 nvq->rx_ring = NULL;
1560                 }
1561
1562                 oldubufs = nvq->ubufs;
1563                 nvq->ubufs = ubufs;
1564
1565                 n->tx_packets = 0;
1566                 n->tx_zcopy_err = 0;
1567                 n->tx_flush = false;
1568         }
1569
1570         mutex_unlock(&vq->mutex);
1571
1572         if (oldubufs) {
1573                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1574                 mutex_lock(&vq->mutex);
1575                 vhost_zerocopy_signal_used(n, vq);
1576                 mutex_unlock(&vq->mutex);
1577         }
1578
1579         if (oldsock) {
1580                 vhost_dev_flush(&n->dev);
1581                 sockfd_put(oldsock);
1582         }
1583
1584         mutex_unlock(&n->dev.mutex);
1585         return 0;
1586
1587 err_used:
1588         vhost_vq_set_backend(vq, oldsock);
1589         vhost_net_enable_vq(n, vq);
1590         if (ubufs)
1591                 vhost_net_ubuf_put_wait_and_free(ubufs);
1592 err_ubufs:
1593         if (sock)
1594                 sockfd_put(sock);
1595 err_vq:
1596         mutex_unlock(&vq->mutex);
1597 err:
1598         mutex_unlock(&n->dev.mutex);
1599         return r;
1600 }
1601
1602 static long vhost_net_reset_owner(struct vhost_net *n)
1603 {
1604         struct socket *tx_sock = NULL;
1605         struct socket *rx_sock = NULL;
1606         long err;
1607         struct vhost_iotlb *umem;
1608
1609         mutex_lock(&n->dev.mutex);
1610         err = vhost_dev_check_owner(&n->dev);
1611         if (err)
1612                 goto done;
1613         umem = vhost_dev_reset_owner_prepare();
1614         if (!umem) {
1615                 err = -ENOMEM;
1616                 goto done;
1617         }
1618         vhost_net_stop(n, &tx_sock, &rx_sock);
1619         vhost_net_flush(n);
1620         vhost_dev_stop(&n->dev);
1621         vhost_dev_reset_owner(&n->dev, umem);
1622         vhost_net_vq_reset(n);
1623 done:
1624         mutex_unlock(&n->dev.mutex);
1625         if (tx_sock)
1626                 sockfd_put(tx_sock);
1627         if (rx_sock)
1628                 sockfd_put(rx_sock);
1629         return err;
1630 }
1631
1632 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1633 {
1634         size_t vhost_hlen, sock_hlen, hdr_len;
1635         int i;
1636
1637         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1638                                (1ULL << VIRTIO_F_VERSION_1))) ?
1639                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1640                         sizeof(struct virtio_net_hdr);
1641         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1642                 /* vhost provides vnet_hdr */
1643                 vhost_hlen = hdr_len;
1644                 sock_hlen = 0;
1645         } else {
1646                 /* socket provides vnet_hdr */
1647                 vhost_hlen = 0;
1648                 sock_hlen = hdr_len;
1649         }
1650         mutex_lock(&n->dev.mutex);
1651         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1652             !vhost_log_access_ok(&n->dev))
1653                 goto out_unlock;
1654
1655         if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) {
1656                 if (vhost_init_device_iotlb(&n->dev))
1657                         goto out_unlock;
1658         }
1659
1660         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1661                 mutex_lock(&n->vqs[i].vq.mutex);
1662                 n->vqs[i].vq.acked_features = features;
1663                 n->vqs[i].vhost_hlen = vhost_hlen;
1664                 n->vqs[i].sock_hlen = sock_hlen;
1665                 mutex_unlock(&n->vqs[i].vq.mutex);
1666         }
1667         mutex_unlock(&n->dev.mutex);
1668         return 0;
1669
1670 out_unlock:
1671         mutex_unlock(&n->dev.mutex);
1672         return -EFAULT;
1673 }
1674
1675 static long vhost_net_set_owner(struct vhost_net *n)
1676 {
1677         int r;
1678
1679         mutex_lock(&n->dev.mutex);
1680         if (vhost_dev_has_owner(&n->dev)) {
1681                 r = -EBUSY;
1682                 goto out;
1683         }
1684         r = vhost_net_set_ubuf_info(n);
1685         if (r)
1686                 goto out;
1687         r = vhost_dev_set_owner(&n->dev);
1688         if (r)
1689                 vhost_net_clear_ubuf_info(n);
1690         vhost_net_flush(n);
1691 out:
1692         mutex_unlock(&n->dev.mutex);
1693         return r;
1694 }
1695
1696 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1697                             unsigned long arg)
1698 {
1699         struct vhost_net *n = f->private_data;
1700         void __user *argp = (void __user *)arg;
1701         u64 __user *featurep = argp;
1702         struct vhost_vring_file backend;
1703         u64 features;
1704         int r;
1705
1706         switch (ioctl) {
1707         case VHOST_NET_SET_BACKEND:
1708                 if (copy_from_user(&backend, argp, sizeof backend))
1709                         return -EFAULT;
1710                 return vhost_net_set_backend(n, backend.index, backend.fd);
1711         case VHOST_GET_FEATURES:
1712                 features = VHOST_NET_FEATURES;
1713                 if (copy_to_user(featurep, &features, sizeof features))
1714                         return -EFAULT;
1715                 return 0;
1716         case VHOST_SET_FEATURES:
1717                 if (copy_from_user(&features, featurep, sizeof features))
1718                         return -EFAULT;
1719                 if (features & ~VHOST_NET_FEATURES)
1720                         return -EOPNOTSUPP;
1721                 return vhost_net_set_features(n, features);
1722         case VHOST_GET_BACKEND_FEATURES:
1723                 features = VHOST_NET_BACKEND_FEATURES;
1724                 if (copy_to_user(featurep, &features, sizeof(features)))
1725                         return -EFAULT;
1726                 return 0;
1727         case VHOST_SET_BACKEND_FEATURES:
1728                 if (copy_from_user(&features, featurep, sizeof(features)))
1729                         return -EFAULT;
1730                 if (features & ~VHOST_NET_BACKEND_FEATURES)
1731                         return -EOPNOTSUPP;
1732                 vhost_set_backend_features(&n->dev, features);
1733                 return 0;
1734         case VHOST_RESET_OWNER:
1735                 return vhost_net_reset_owner(n);
1736         case VHOST_SET_OWNER:
1737                 return vhost_net_set_owner(n);
1738         default:
1739                 mutex_lock(&n->dev.mutex);
1740                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1741                 if (r == -ENOIOCTLCMD)
1742                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1743                 else
1744                         vhost_net_flush(n);
1745                 mutex_unlock(&n->dev.mutex);
1746                 return r;
1747         }
1748 }
1749
1750 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1751 {
1752         struct file *file = iocb->ki_filp;
1753         struct vhost_net *n = file->private_data;
1754         struct vhost_dev *dev = &n->dev;
1755         int noblock = file->f_flags & O_NONBLOCK;
1756
1757         return vhost_chr_read_iter(dev, to, noblock);
1758 }
1759
1760 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1761                                         struct iov_iter *from)
1762 {
1763         struct file *file = iocb->ki_filp;
1764         struct vhost_net *n = file->private_data;
1765         struct vhost_dev *dev = &n->dev;
1766
1767         return vhost_chr_write_iter(dev, from);
1768 }
1769
1770 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1771 {
1772         struct vhost_net *n = file->private_data;
1773         struct vhost_dev *dev = &n->dev;
1774
1775         return vhost_chr_poll(file, dev, wait);
1776 }
1777
1778 static const struct file_operations vhost_net_fops = {
1779         .owner          = THIS_MODULE,
1780         .release        = vhost_net_release,
1781         .read_iter      = vhost_net_chr_read_iter,
1782         .write_iter     = vhost_net_chr_write_iter,
1783         .poll           = vhost_net_chr_poll,
1784         .unlocked_ioctl = vhost_net_ioctl,
1785         .compat_ioctl   = compat_ptr_ioctl,
1786         .open           = vhost_net_open,
1787         .llseek         = noop_llseek,
1788 };
1789
1790 static struct miscdevice vhost_net_misc = {
1791         .minor = VHOST_NET_MINOR,
1792         .name = "vhost-net",
1793         .fops = &vhost_net_fops,
1794 };
1795
1796 static int __init vhost_net_init(void)
1797 {
1798         if (experimental_zcopytx)
1799                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1800         return misc_register(&vhost_net_misc);
1801 }
1802 module_init(vhost_net_init);
1803
1804 static void __exit vhost_net_exit(void)
1805 {
1806         misc_deregister(&vhost_net_misc);
1807 }
1808 module_exit(vhost_net_exit);
1809
1810 MODULE_VERSION("0.0.1");
1811 MODULE_LICENSE("GPL v2");
1812 MODULE_AUTHOR("Michael S. Tsirkin");
1813 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1814 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1815 MODULE_ALIAS("devname:vhost-net");