drivers/vhost/net.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (C) 2009 Red Hat, Inc.
   3  * Author: Michael S. Tsirkin <mst@redhat.com>
   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 };
  78
  79 enum {
  80         VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
  81 };
  82
  83 enum {
  84         VHOST_NET_VQ_RX = 0,
  85         VHOST_NET_VQ_TX = 1,
  86         VHOST_NET_VQ_MAX = 2,
  87 };
  88
  89 struct vhost_net_ubuf_ref {
  90         /* refcount follows semantics similar to kref:
  91          *  0: object is released
  92          *  1: no outstanding ubufs
  93          * >1: outstanding ubufs
  94          */
  95         atomic_t refcount;
  96         wait_queue_head_t wait;
  97         struct vhost_virtqueue *vq;
  98 };
  99
 100 #define VHOST_NET_BATCH 64
 101 struct vhost_net_buf {
 102         void **queue;
 103         int tail;
 104         int head;
 105 };
 106
 107 struct vhost_net_virtqueue {
 108         struct vhost_virtqueue vq;
 109         size_t vhost_hlen;
 110         size_t sock_hlen;
 111         /* vhost zerocopy support fields below: */
 112         /* last used idx for outstanding DMA zerocopy buffers */
 113         int upend_idx;
 114         /* For TX, first used idx for DMA done zerocopy buffers
 115          * For RX, number of batched heads
 116          */
 117         int done_idx;
 118         /* Number of XDP frames batched */
 119         int batched_xdp;
 120         /* an array of userspace buffers info */
 121         struct ubuf_info *ubuf_info;
 122         /* Reference counting for outstanding ubufs.
 123          * Protected by vq mutex. Writers must also take device mutex. */
 124         struct vhost_net_ubuf_ref *ubufs;
 125         struct ptr_ring *rx_ring;
 126         struct vhost_net_buf rxq;
 127         /* Batched XDP buffs */
 128         struct xdp_buff *xdp;
 129 };
 130
 131 struct vhost_net {
 132         struct vhost_dev dev;
 133         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 134         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 135         /* Number of TX recently submitted.
 136          * Protected by tx vq lock. */
 137         unsigned tx_packets;
 138         /* Number of times zerocopy TX recently failed.
 139          * Protected by tx vq lock. */
 140         unsigned tx_zcopy_err;
 141         /* Flush in progress. Protected by tx vq lock. */
 142         bool tx_flush;
 143         /* Private page frag */
 144         struct page_frag page_frag;
 145         /* Refcount bias of page frag */
 146         int refcnt_bias;
 147 };
 148
 149 static unsigned vhost_net_zcopy_mask __read_mostly;
 150
 151 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 152 {
 153         if (rxq->tail != rxq->head)
 154                 return rxq->queue[rxq->head];
 155         else
 156                 return NULL;
 157 }
 158
 159 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 160 {
 161         return rxq->tail - rxq->head;
 162 }
 163
 164 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 165 {
 166         return rxq->tail == rxq->head;
 167 }
 168
 169 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 170 {
 171         void *ret = vhost_net_buf_get_ptr(rxq);
 172         ++rxq->head;
 173         return ret;
 174 }
 175
 176 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 177 {
 178         struct vhost_net_buf *rxq = &nvq->rxq;
 179
 180         rxq->head = 0;
 181         rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
 182                                               VHOST_NET_BATCH);
 183         return rxq->tail;
 184 }
 185
 186 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 187 {
 188         struct vhost_net_buf *rxq = &nvq->rxq;
 189
 190         if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
 191                 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 192                                    vhost_net_buf_get_size(rxq),
 193                                    tun_ptr_free);
 194                 rxq->head = rxq->tail = 0;
 195         }
 196 }
 197
 198 static int vhost_net_buf_peek_len(void *ptr)
 199 {
 200         if (tun_is_xdp_frame(ptr)) {
 201                 struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
 202
 203                 return xdpf->len;
 204         }
 205
 206         return __skb_array_len_with_tag(ptr);
 207 }
 208
 209 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 210 {
 211         struct vhost_net_buf *rxq = &nvq->rxq;
 212
 213         if (!vhost_net_buf_is_empty(rxq))
 214                 goto out;
 215
 216         if (!vhost_net_buf_produce(nvq))
 217                 return 0;
 218
 219 out:
 220         return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 221 }
 222
 223 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 224 {
 225         rxq->head = rxq->tail = 0;
 226 }
 227
 228 static void vhost_net_enable_zcopy(int vq)
 229 {
 230         vhost_net_zcopy_mask |= 0x1 << vq;
 231 }
 232
 233 static struct vhost_net_ubuf_ref *
 234 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 235 {
 236         struct vhost_net_ubuf_ref *ubufs;
 237         /* No zero copy backend? Nothing to count. */
 238         if (!zcopy)
 239                 return NULL;
 240         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 241         if (!ubufs)
 242                 return ERR_PTR(-ENOMEM);
 243         atomic_set(&ubufs->refcount, 1);
 244         init_waitqueue_head(&ubufs->wait);
 245         ubufs->vq = vq;
 246         return ubufs;
 247 }
 248
 249 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 250 {
 251         int r = atomic_sub_return(1, &ubufs->refcount);
 252         if (unlikely(!r))
 253                 wake_up(&ubufs->wait);
 254         return r;
 255 }
 256
 257 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 258 {
 259         vhost_net_ubuf_put(ubufs);
 260         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 261 }
 262
 263 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 264 {
 265         vhost_net_ubuf_put_and_wait(ubufs);
 266         kfree(ubufs);
 267 }
 268
 269 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 270 {
 271         int i;
 272
 273         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 274                 kfree(n->vqs[i].ubuf_info);
 275                 n->vqs[i].ubuf_info = NULL;
 276         }
 277 }
 278
 279 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 280 {
 281         bool zcopy;
 282         int i;
 283
 284         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 285                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 286                 if (!zcopy)
 287                         continue;
 288                 n->vqs[i].ubuf_info =
 289                         kmalloc_array(UIO_MAXIOV,
 290                                       sizeof(*n->vqs[i].ubuf_info),
 291                                       GFP_KERNEL);
 292                 if  (!n->vqs[i].ubuf_info)
 293                         goto err;
 294         }
 295         return 0;
 296
 297 err:
 298         vhost_net_clear_ubuf_info(n);
 299         return -ENOMEM;
 300 }
 301
 302 static void vhost_net_vq_reset(struct vhost_net *n)
 303 {
 304         int i;
 305
 306         vhost_net_clear_ubuf_info(n);
 307
 308         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 309                 n->vqs[i].done_idx = 0;
 310                 n->vqs[i].upend_idx = 0;
 311                 n->vqs[i].ubufs = NULL;
 312                 n->vqs[i].vhost_hlen = 0;
 313                 n->vqs[i].sock_hlen = 0;
 314                 vhost_net_buf_init(&n->vqs[i].rxq);
 315         }
 316
 317 }
 318
 319 static void vhost_net_tx_packet(struct vhost_net *net)
 320 {
 321         ++net->tx_packets;
 322         if (net->tx_packets < 1024)
 323                 return;
 324         net->tx_packets = 0;
 325         net->tx_zcopy_err = 0;
 326 }
 327
 328 static void vhost_net_tx_err(struct vhost_net *net)
 329 {
 330         ++net->tx_zcopy_err;
 331 }
 332
 333 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 334 {
 335         /* TX flush waits for outstanding DMAs to be done.
 336          * Don't start new DMAs.
 337          */
 338         return !net->tx_flush &&
 339                 net->tx_packets / 64 >= net->tx_zcopy_err;
 340 }
 341
 342 static bool vhost_sock_zcopy(struct socket *sock)
 343 {
 344         return unlikely(experimental_zcopytx) &&
 345                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 346 }
 347
 348 static bool vhost_sock_xdp(struct socket *sock)
 349 {
 350         return sock_flag(sock->sk, SOCK_XDP);
 351 }
 352
 353 /* In case of DMA done not in order in lower device driver for some reason.
 354  * upend_idx is used to track end of used idx, done_idx is used to track head
 355  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 356  * guest used idx.
 357  */
 358 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 359                                        struct vhost_virtqueue *vq)
 360 {
 361         struct vhost_net_virtqueue *nvq =
 362                 container_of(vq, struct vhost_net_virtqueue, vq);
 363         int i, add;
 364         int j = 0;
 365
 366         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 367                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 368                         vhost_net_tx_err(net);
 369                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 370                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 371                         ++j;
 372                 } else
 373                         break;
 374         }
 375         while (j) {
 376                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 377                 vhost_add_used_and_signal_n(vq->dev, vq,
 378                                             &vq->heads[nvq->done_idx], add);
 379                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 380                 j -= add;
 381         }
 382 }
 383
 384 static void vhost_zerocopy_callback(struct sk_buff *skb,
 385                                     struct ubuf_info *ubuf, bool success)
 386 {
 387         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 388         struct vhost_virtqueue *vq = ubufs->vq;
 389         int cnt;
 390
 391         rcu_read_lock_bh();
 392
 393         /* set len to mark this desc buffers done DMA */
 394         vq->heads[ubuf->desc].len = success ?
 395                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 396         cnt = vhost_net_ubuf_put(ubufs);
 397
 398         /*
 399          * Trigger polling thread if guest stopped submitting new buffers:
 400          * in this case, the refcount after decrement will eventually reach 1.
 401          * We also trigger polling periodically after each 16 packets
 402          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 403          * less than 10% of times).
 404          */
 405         if (cnt <= 1 || !(cnt % 16))
 406                 vhost_poll_queue(&vq->poll);
 407
 408         rcu_read_unlock_bh();
 409 }
 410
 411 static inline unsigned long busy_clock(void)
 412 {
 413         return local_clock() >> 10;
 414 }
 415
 416 static bool vhost_can_busy_poll(unsigned long endtime)
 417 {
 418         return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
 419                       !signal_pending(current));
 420 }
 421
 422 static void vhost_net_disable_vq(struct vhost_net *n,
 423                                  struct vhost_virtqueue *vq)
 424 {
 425         struct vhost_net_virtqueue *nvq =
 426                 container_of(vq, struct vhost_net_virtqueue, vq);
 427         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 428         if (!vhost_vq_get_backend(vq))
 429                 return;
 430         vhost_poll_stop(poll);
 431 }
 432
 433 static int vhost_net_enable_vq(struct vhost_net *n,
 434                                 struct vhost_virtqueue *vq)
 435 {
 436         struct vhost_net_virtqueue *nvq =
 437                 container_of(vq, struct vhost_net_virtqueue, vq);
 438         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 439         struct socket *sock;
 440
 441         sock = vhost_vq_get_backend(vq);
 442         if (!sock)
 443                 return 0;
 444
 445         return vhost_poll_start(poll, sock->file);
 446 }
 447
 448 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
 449 {
 450         struct vhost_virtqueue *vq = &nvq->vq;
 451         struct vhost_dev *dev = vq->dev;
 452
 453         if (!nvq->done_idx)
 454                 return;
 455
 456         vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
 457         nvq->done_idx = 0;
 458 }
 459
 460 static void vhost_tx_batch(struct vhost_net *net,
 461                            struct vhost_net_virtqueue *nvq,
 462                            struct socket *sock,
 463                            struct msghdr *msghdr)
 464 {
 465         struct tun_msg_ctl ctl = {
 466                 .type = TUN_MSG_PTR,
 467                 .num = nvq->batched_xdp,
 468                 .ptr = nvq->xdp,
 469         };
 470         int i, err;
 471
 472         if (nvq->batched_xdp == 0)
 473                 goto signal_used;
 474
 475         msghdr->msg_control = &ctl;
 476         err = sock->ops->sendmsg(sock, msghdr, 0);
 477         if (unlikely(err < 0)) {
 478                 vq_err(&nvq->vq, "Fail to batch sending packets\n");
 479
 480                 /* free pages owned by XDP; since this is an unlikely error path,
 481                  * keep it simple and avoid more complex bulk update for the
 482                  * used pages
 483                  */
 484                 for (i = 0; i < nvq->batched_xdp; ++i)
 485                         put_page(virt_to_head_page(nvq->xdp[i].data));
 486                 nvq->batched_xdp = 0;
 487                 nvq->done_idx = 0;
 488                 return;
 489         }
 490
 491 signal_used:
 492         vhost_net_signal_used(nvq);
 493         nvq->batched_xdp = 0;
 494 }
 495
 496 static int sock_has_rx_data(struct socket *sock)
 497 {
 498         if (unlikely(!sock))
 499                 return 0;
 500
 501         if (sock->ops->peek_len)
 502                 return sock->ops->peek_len(sock);
 503
 504         return skb_queue_empty(&sock->sk->sk_receive_queue);
 505 }
 506
 507 static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
 508                                           struct vhost_virtqueue *vq)
 509 {
 510         if (!vhost_vq_avail_empty(&net->dev, vq)) {
 511                 vhost_poll_queue(&vq->poll);
 512         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 513                 vhost_disable_notify(&net->dev, vq);
 514                 vhost_poll_queue(&vq->poll);
 515         }
 516 }
 517
 518 static void vhost_net_busy_poll(struct vhost_net *net,
 519                                 struct vhost_virtqueue *rvq,
 520                                 struct vhost_virtqueue *tvq,
 521                                 bool *busyloop_intr,
 522                                 bool poll_rx)
 523 {
 524         unsigned long busyloop_timeout;
 525         unsigned long endtime;
 526         struct socket *sock;
 527         struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
 528
 529         /* Try to hold the vq mutex of the paired virtqueue. We can't
 530          * use mutex_lock() here since we could not guarantee a
 531          * consistenet lock ordering.
 532          */
 533         if (!mutex_trylock(&vq->mutex))
 534                 return;
 535
 536         vhost_disable_notify(&net->dev, vq);
 537         sock = vhost_vq_get_backend(rvq);
 538
 539         busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
 540                                      tvq->busyloop_timeout;
 541
 542         preempt_disable();
 543         endtime = busy_clock() + busyloop_timeout;
 544
 545         while (vhost_can_busy_poll(endtime)) {
 546                 if (vhost_has_work(&net->dev)) {
 547                         *busyloop_intr = true;
 548                         break;
 549                 }
 550
 551                 if ((sock_has_rx_data(sock) &&
 552                      !vhost_vq_avail_empty(&net->dev, rvq)) ||
 553                     !vhost_vq_avail_empty(&net->dev, tvq))
 554                         break;
 555
 556                 cpu_relax();
 557         }
 558
 559         preempt_enable();
 560
 561         if (poll_rx || sock_has_rx_data(sock))
 562                 vhost_net_busy_poll_try_queue(net, vq);
 563         else if (!poll_rx) /* On tx here, sock has no rx data. */
 564                 vhost_enable_notify(&net->dev, rvq);
 565
 566         mutex_unlock(&vq->mutex);
 567 }
 568
 569 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 570                                     struct vhost_net_virtqueue *tnvq,
 571                                     unsigned int *out_num, unsigned int *in_num,
 572                                     struct msghdr *msghdr, bool *busyloop_intr)
 573 {
 574         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 575         struct vhost_virtqueue *rvq = &rnvq->vq;
 576         struct vhost_virtqueue *tvq = &tnvq->vq;
 577
 578         int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 579                                   out_num, in_num, NULL, NULL);
 580
 581         if (r == tvq->num && tvq->busyloop_timeout) {
 582                 /* Flush batched packets first */
 583                 if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq)))
 584                         vhost_tx_batch(net, tnvq,
 585                                        vhost_vq_get_backend(tvq),
 586                                        msghdr);
 587
 588                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
 589
 590                 r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 591                                       out_num, in_num, NULL, NULL);
 592         }
 593
 594         return r;
 595 }
 596
 597 static bool vhost_exceeds_maxpend(struct vhost_net *net)
 598 {
 599         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 600         struct vhost_virtqueue *vq = &nvq->vq;
 601
 602         return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 603                min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 604 }
 605
 606 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
 607                             size_t hdr_size, int out)
 608 {
 609         /* Skip header. TODO: support TSO. */
 610         size_t len = iov_length(vq->iov, out);
 611
 612         iov_iter_init(iter, WRITE, vq->iov, out, len);
 613         iov_iter_advance(iter, hdr_size);
 614
 615         return iov_iter_count(iter);
 616 }
 617
 618 static int get_tx_bufs(struct vhost_net *net,
 619                        struct vhost_net_virtqueue *nvq,
 620                        struct msghdr *msg,
 621                        unsigned int *out, unsigned int *in,
 622                        size_t *len, bool *busyloop_intr)
 623 {
 624         struct vhost_virtqueue *vq = &nvq->vq;
 625         int ret;
 626
 627         ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
 628
 629         if (ret < 0 || ret == vq->num)
 630                 return ret;
 631
 632         if (*in) {
 633                 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 634                         *out, *in);
 635                 return -EFAULT;
 636         }
 637
 638         /* Sanity check */
 639         *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 640         if (*len == 0) {
 641                 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 642                         *len, nvq->vhost_hlen);
 643                 return -EFAULT;
 644         }
 645
 646         return ret;
 647 }
 648
 649 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 650 {
 651         return total_len < VHOST_NET_WEIGHT &&
 652                !vhost_vq_avail_empty(vq->dev, vq);
 653 }
 654
 655 static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
 656                                        struct page_frag *pfrag, gfp_t gfp)
 657 {
 658         if (pfrag->page) {
 659                 if (pfrag->offset + sz <= pfrag->size)
 660                         return true;
 661                 __page_frag_cache_drain(pfrag->page, net->refcnt_bias);
 662         }
 663
 664         pfrag->offset = 0;
 665         net->refcnt_bias = 0;
 666         if (SKB_FRAG_PAGE_ORDER) {
 667                 /* Avoid direct reclaim but allow kswapd to wake */
 668                 pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
 669                                           __GFP_COMP | __GFP_NOWARN |
 670                                           __GFP_NORETRY,
 671                                           SKB_FRAG_PAGE_ORDER);
 672                 if (likely(pfrag->page)) {
 673                         pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
 674                         goto done;
 675                 }
 676         }
 677         pfrag->page = alloc_page(gfp);
 678         if (likely(pfrag->page)) {
 679                 pfrag->size = PAGE_SIZE;
 680                 goto done;
 681         }
 682         return false;
 683
 684 done:
 685         net->refcnt_bias = USHRT_MAX;
 686         page_ref_add(pfrag->page, USHRT_MAX - 1);
 687         return true;
 688 }
 689
 690 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 691
 692 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
 693                                struct iov_iter *from)
 694 {
 695         struct vhost_virtqueue *vq = &nvq->vq;
 696         struct vhost_net *net = container_of(vq->dev, struct vhost_net,
 697                                              dev);
 698         struct socket *sock = vhost_vq_get_backend(vq);
 699         struct page_frag *alloc_frag = &net->page_frag;
 700         struct virtio_net_hdr *gso;
 701         struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
 702         struct tun_xdp_hdr *hdr;
 703         size_t len = iov_iter_count(from);
 704         int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
 705         int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 706         int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
 707         int sock_hlen = nvq->sock_hlen;
 708         void *buf;
 709         int copied;
 710
 711         if (unlikely(len < nvq->sock_hlen))
 712                 return -EFAULT;
 713
 714         if (SKB_DATA_ALIGN(len + pad) +
 715             SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
 716                 return -ENOSPC;
 717
 718         buflen += SKB_DATA_ALIGN(len + pad);
 719         alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
 720         if (unlikely(!vhost_net_page_frag_refill(net, buflen,
 721                                                  alloc_frag, GFP_KERNEL)))
 722                 return -ENOMEM;
 723
 724         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 725         copied = copy_page_from_iter(alloc_frag->page,
 726                                      alloc_frag->offset +
 727                                      offsetof(struct tun_xdp_hdr, gso),
 728                                      sock_hlen, from);
 729         if (copied != sock_hlen)
 730                 return -EFAULT;
 731
 732         hdr = buf;
 733         gso = &hdr->gso;
 734
 735         if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
 736             vhost16_to_cpu(vq, gso->csum_start) +
 737             vhost16_to_cpu(vq, gso->csum_offset) + 2 >
 738             vhost16_to_cpu(vq, gso->hdr_len)) {
 739                 gso->hdr_len = cpu_to_vhost16(vq,
 740                                vhost16_to_cpu(vq, gso->csum_start) +
 741                                vhost16_to_cpu(vq, gso->csum_offset) + 2);
 742
 743                 if (vhost16_to_cpu(vq, gso->hdr_len) > len)
 744                         return -EINVAL;
 745         }
 746
 747         len -= sock_hlen;
 748         copied = copy_page_from_iter(alloc_frag->page,
 749                                      alloc_frag->offset + pad,
 750                                      len, from);
 751         if (copied != len)
 752                 return -EFAULT;
 753
 754         xdp_init_buff(xdp, buflen, NULL);
 755         xdp_prepare_buff(xdp, buf, pad, len, true);
 756         hdr->buflen = buflen;
 757
 758         --net->refcnt_bias;
 759         alloc_frag->offset += buflen;
 760
 761         ++nvq->batched_xdp;
 762
 763         return 0;
 764 }
 765
 766 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 767 {
 768         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 769         struct vhost_virtqueue *vq = &nvq->vq;
 770         unsigned out, in;
 771         int head;
 772         struct msghdr msg = {
 773                 .msg_name = NULL,
 774                 .msg_namelen = 0,
 775                 .msg_control = NULL,
 776                 .msg_controllen = 0,
 777                 .msg_flags = MSG_DONTWAIT,
 778         };
 779         size_t len, total_len = 0;
 780         int err;
 781         int sent_pkts = 0;
 782         bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
 783
 784         do {
 785                 bool busyloop_intr = false;
 786
 787                 if (nvq->done_idx == VHOST_NET_BATCH)
 788                         vhost_tx_batch(net, nvq, sock, &msg);
 789
 790                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 791                                    &busyloop_intr);
 792                 /* On error, stop handling until the next kick. */
 793                 if (unlikely(head < 0))
 794                         break;
 795                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 796                 if (head == vq->num) {
 797                         if (unlikely(busyloop_intr)) {
 798                                 vhost_poll_queue(&vq->poll);
 799                         } else if (unlikely(vhost_enable_notify(&net->dev,
 800                                                                 vq))) {
 801                                 vhost_disable_notify(&net->dev, vq);
 802                                 continue;
 803                         }
 804                         break;
 805                 }
 806
 807                 total_len += len;
 808
 809                 /* For simplicity, TX batching is only enabled if
 810                  * sndbuf is unlimited.
 811                  */
 812                 if (sock_can_batch) {
 813                         err = vhost_net_build_xdp(nvq, &msg.msg_iter);
 814                         if (!err) {
 815                                 goto done;
 816                         } else if (unlikely(err != -ENOSPC)) {
 817                                 vhost_tx_batch(net, nvq, sock, &msg);
 818                                 vhost_discard_vq_desc(vq, 1);
 819                                 vhost_net_enable_vq(net, vq);
 820                                 break;
 821                         }
 822
 823                         /* We can't build XDP buff, go for single
 824                          * packet path but let's flush batched
 825                          * packets.
 826                          */
 827                         vhost_tx_batch(net, nvq, sock, &msg);
 828                         msg.msg_control = NULL;
 829                 } else {
 830                         if (tx_can_batch(vq, total_len))
 831                                 msg.msg_flags |= MSG_MORE;
 832                         else
 833                                 msg.msg_flags &= ~MSG_MORE;
 834                 }
 835
 836                 err = sock->ops->sendmsg(sock, &msg, len);
 837                 if (unlikely(err < 0)) {
 838                         if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) {
 839                                 vhost_discard_vq_desc(vq, 1);
 840                                 vhost_net_enable_vq(net, vq);
 841                                 break;
 842                         }
 843                         pr_debug("Fail to send packet: err %d", err);
 844                 } else if (unlikely(err != len))
 845                         pr_debug("Truncated TX packet: len %d != %zd\n",
 846                                  err, len);
 847 done:
 848                 vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 849                 vq->heads[nvq->done_idx].len = 0;
 850                 ++nvq->done_idx;
 851         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 852
 853         vhost_tx_batch(net, nvq, sock, &msg);
 854 }
 855
 856 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 857 {
 858         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 859         struct vhost_virtqueue *vq = &nvq->vq;
 860         unsigned out, in;
 861         int head;
 862         struct msghdr msg = {
 863                 .msg_name = NULL,
 864                 .msg_namelen = 0,
 865                 .msg_control = NULL,
 866                 .msg_controllen = 0,
 867                 .msg_flags = MSG_DONTWAIT,
 868         };
 869         struct tun_msg_ctl ctl;
 870         size_t len, total_len = 0;
 871         int err;
 872         struct vhost_net_ubuf_ref *ubufs;
 873         struct ubuf_info *ubuf;
 874         bool zcopy_used;
 875         int sent_pkts = 0;
 876
 877         do {
 878                 bool busyloop_intr;
 879
 880                 /* Release DMAs done buffers first */
 881                 vhost_zerocopy_signal_used(net, vq);
 882
 883                 busyloop_intr = false;
 884                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 885                                    &busyloop_intr);
 886                 /* On error, stop handling until the next kick. */
 887                 if (unlikely(head < 0))
 888                         break;
 889                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 890                 if (head == vq->num) {
 891                         if (unlikely(busyloop_intr)) {
 892                                 vhost_poll_queue(&vq->poll);
 893                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 894                                 vhost_disable_notify(&net->dev, vq);
 895                                 continue;
 896                         }
 897                         break;
 898                 }
 899
 900                 zcopy_used = len >= VHOST_GOODCOPY_LEN
 901                              && !vhost_exceeds_maxpend(net)
 902                              && vhost_net_tx_select_zcopy(net);
 903
 904                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 905                 if (zcopy_used) {
 906                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 907                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 908                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 909                         ubuf->callback = vhost_zerocopy_callback;
 910                         ubuf->ctx = nvq->ubufs;
 911                         ubuf->desc = nvq->upend_idx;
 912                         ubuf->flags = SKBFL_ZEROCOPY_FRAG;
 913                         refcount_set(&ubuf->refcnt, 1);
 914                         msg.msg_control = &ctl;
 915                         ctl.type = TUN_MSG_UBUF;
 916                         ctl.ptr = ubuf;
 917                         msg.msg_controllen = sizeof(ctl);
 918                         ubufs = nvq->ubufs;
 919                         atomic_inc(&ubufs->refcount);
 920                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 921                 } else {
 922                         msg.msg_control = NULL;
 923                         ubufs = NULL;
 924                 }
 925                 total_len += len;
 926                 if (tx_can_batch(vq, total_len) &&
 927                     likely(!vhost_exceeds_maxpend(net))) {
 928                         msg.msg_flags |= MSG_MORE;
 929                 } else {
 930                         msg.msg_flags &= ~MSG_MORE;
 931                 }
 932
 933                 err = sock->ops->sendmsg(sock, &msg, len);
 934                 if (unlikely(err < 0)) {
 935                         if (zcopy_used) {
 936                                 if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
 937                                         vhost_net_ubuf_put(ubufs);
 938                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 939                                         % UIO_MAXIOV;
 940                         }
 941                         if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) {
 942                                 vhost_discard_vq_desc(vq, 1);
 943                                 vhost_net_enable_vq(net, vq);
 944                                 break;
 945                         }
 946                         pr_debug("Fail to send packet: err %d", err);
 947                 } else if (unlikely(err != len))
 948                         pr_debug("Truncated TX packet: "
 949                                  " len %d != %zd\n", err, len);
 950                 if (!zcopy_used)
 951                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 952                 else
 953                         vhost_zerocopy_signal_used(net, vq);
 954                 vhost_net_tx_packet(net);
 955         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 956 }
 957
 958 /* Expects to be always run from workqueue - which acts as
 959  * read-size critical section for our kind of RCU. */
 960 static void handle_tx(struct vhost_net *net)
 961 {
 962         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 963         struct vhost_virtqueue *vq = &nvq->vq;
 964         struct socket *sock;
 965
 966         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
 967         sock = vhost_vq_get_backend(vq);
 968         if (!sock)
 969                 goto out;
 970
 971         if (!vq_meta_prefetch(vq))
 972                 goto out;
 973
 974         vhost_disable_notify(&net->dev, vq);
 975         vhost_net_disable_vq(net, vq);
 976
 977         if (vhost_sock_zcopy(sock))
 978                 handle_tx_zerocopy(net, sock);
 979         else
 980                 handle_tx_copy(net, sock);
 981
 982 out:
 983         mutex_unlock(&vq->mutex);
 984 }
 985
 986 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 987 {
 988         struct sk_buff *head;
 989         int len = 0;
 990         unsigned long flags;
 991
 992         if (rvq->rx_ring)
 993                 return vhost_net_buf_peek(rvq);
 994
 995         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 996         head = skb_peek(&sk->sk_receive_queue);
 997         if (likely(head)) {
 998                 len = head->len;
 999                 if (skb_vlan_tag_present(head))
1000                         len += VLAN_HLEN;
1001         }
1002
1003         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
1004         return len;
1005 }
1006
1007 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
1008                                       bool *busyloop_intr)
1009 {
1010         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1011         struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1012         struct vhost_virtqueue *rvq = &rnvq->vq;
1013         struct vhost_virtqueue *tvq = &tnvq->vq;
1014         int len = peek_head_len(rnvq, sk);
1015
1016         if (!len && rvq->busyloop_timeout) {
1017                 /* Flush batched heads first */
1018                 vhost_net_signal_used(rnvq);
1019                 /* Both tx vq and rx socket were polled here */
1020                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1021
1022                 len = peek_head_len(rnvq, sk);
1023         }
1024
1025         return len;
1026 }
1027
1028 /* This is a multi-buffer version of vhost_get_desc, that works if
1029  *      vq has read descriptors only.
1030  * @vq          - the relevant virtqueue
1031  * @datalen     - data length we'll be reading
1032  * @iovcount    - returned count of io vectors we fill
1033  * @log         - vhost log
1034  * @log_num     - log offset
1035  * @quota       - headcount quota, 1 for big buffer
1036  *      returns number of buffer heads allocated, negative on error
1037  */
1038 static int get_rx_bufs(struct vhost_virtqueue *vq,
1039                        struct vring_used_elem *heads,
1040                        int datalen,
1041                        unsigned *iovcount,
1042                        struct vhost_log *log,
1043                        unsigned *log_num,
1044                        unsigned int quota)
1045 {
1046         unsigned int out, in;
1047         int seg = 0;
1048         int headcount = 0;
1049         unsigned d;
1050         int r, nlogs = 0;
1051         /* len is always initialized before use since we are always called with
1052          * datalen > 0.
1053          */
1054         u32 len;
1055
1056         while (datalen > 0 && headcount < quota) {
1057                 if (unlikely(seg >= UIO_MAXIOV)) {
1058                         r = -ENOBUFS;
1059                         goto err;
1060                 }
1061                 r = vhost_get_vq_desc(vq, vq->iov + seg,
1062                                       ARRAY_SIZE(vq->iov) - seg, &out,
1063                                       &in, log, log_num);
1064                 if (unlikely(r < 0))
1065                         goto err;
1066
1067                 d = r;
1068                 if (d == vq->num) {
1069                         r = 0;
1070                         goto err;
1071                 }
1072                 if (unlikely(out || in <= 0)) {
1073                         vq_err(vq, "unexpected descriptor format for RX: "
1074                                 "out %d, in %d\n", out, in);
1075                         r = -EINVAL;
1076                         goto err;
1077                 }
1078                 if (unlikely(log)) {
1079                         nlogs += *log_num;
1080                         log += *log_num;
1081                 }
1082                 heads[headcount].id = cpu_to_vhost32(vq, d);
1083                 len = iov_length(vq->iov + seg, in);
1084                 heads[headcount].len = cpu_to_vhost32(vq, len);
1085                 datalen -= len;
1086                 ++headcount;
1087                 seg += in;
1088         }
1089         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1090         *iovcount = seg;
1091         if (unlikely(log))
1092                 *log_num = nlogs;
1093
1094         /* Detect overrun */
1095         if (unlikely(datalen > 0)) {
1096                 r = UIO_MAXIOV + 1;
1097                 goto err;
1098         }
1099         return headcount;
1100 err:
1101         vhost_discard_vq_desc(vq, headcount);
1102         return r;
1103 }
1104
1105 /* Expects to be always run from workqueue - which acts as
1106  * read-size critical section for our kind of RCU. */
1107 static void handle_rx(struct vhost_net *net)
1108 {
1109         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1110         struct vhost_virtqueue *vq = &nvq->vq;
1111         unsigned in, log;
1112         struct vhost_log *vq_log;
1113         struct msghdr msg = {
1114                 .msg_name = NULL,
1115                 .msg_namelen = 0,
1116                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
1117                 .msg_controllen = 0,
1118                 .msg_flags = MSG_DONTWAIT,
1119         };
1120         struct virtio_net_hdr hdr = {
1121                 .flags = 0,
1122                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
1123         };
1124         size_t total_len = 0;
1125         int err, mergeable;
1126         s16 headcount;
1127         size_t vhost_hlen, sock_hlen;
1128         size_t vhost_len, sock_len;
1129         bool busyloop_intr = false;
1130         struct socket *sock;
1131         struct iov_iter fixup;
1132         __virtio16 num_buffers;
1133         int recv_pkts = 0;
1134
1135         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1136         sock = vhost_vq_get_backend(vq);
1137         if (!sock)
1138                 goto out;
1139
1140         if (!vq_meta_prefetch(vq))
1141                 goto out;
1142
1143         vhost_disable_notify(&net->dev, vq);
1144         vhost_net_disable_vq(net, vq);
1145
1146         vhost_hlen = nvq->vhost_hlen;
1147         sock_hlen = nvq->sock_hlen;
1148
1149         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1150                 vq->log : NULL;
1151         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1152
1153         do {
1154                 sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1155                                                       &busyloop_intr);
1156                 if (!sock_len)
1157                         break;
1158                 sock_len += sock_hlen;
1159                 vhost_len = sock_len + vhost_hlen;
1160                 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1161                                         vhost_len, &in, vq_log, &log,
1162                                         likely(mergeable) ? UIO_MAXIOV : 1);
1163                 /* On error, stop handling until the next kick. */
1164                 if (unlikely(headcount < 0))
1165                         goto out;
1166                 /* OK, now we need to know about added descriptors. */
1167                 if (!headcount) {
1168                         if (unlikely(busyloop_intr)) {
1169                                 vhost_poll_queue(&vq->poll);
1170                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1171                                 /* They have slipped one in as we were
1172                                  * doing that: check again. */
1173                                 vhost_disable_notify(&net->dev, vq);
1174                                 continue;
1175                         }
1176                         /* Nothing new?  Wait for eventfd to tell us
1177                          * they refilled. */
1178                         goto out;
1179                 }
1180                 busyloop_intr = false;
1181                 if (nvq->rx_ring)
1182                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1183                 /* On overrun, truncate and discard */
1184                 if (unlikely(headcount > UIO_MAXIOV)) {
1185                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
1186                         err = sock->ops->recvmsg(sock, &msg,
1187                                                  1, MSG_DONTWAIT | MSG_TRUNC);
1188                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
1189                         continue;
1190                 }
1191                 /* We don't need to be notified again. */
1192                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
1193                 fixup = msg.msg_iter;
1194                 if (unlikely((vhost_hlen))) {
1195                         /* We will supply the header ourselves
1196                          * TODO: support TSO.
1197                          */
1198                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
1199                 }
1200                 err = sock->ops->recvmsg(sock, &msg,
1201                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
1202                 /* Userspace might have consumed the packet meanwhile:
1203                  * it's not supposed to do this usually, but might be hard
1204                  * to prevent. Discard data we got (if any) and keep going. */
1205                 if (unlikely(err != sock_len)) {
1206                         pr_debug("Discarded rx packet: "
1207                                  " len %d, expected %zd\n", err, sock_len);
1208                         vhost_discard_vq_desc(vq, headcount);
1209                         continue;
1210                 }
1211                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1212                 if (unlikely(vhost_hlen)) {
1213                         if (copy_to_iter(&hdr, sizeof(hdr),
1214                                          &fixup) != sizeof(hdr)) {
1215                                 vq_err(vq, "Unable to write vnet_hdr "
1216                                        "at addr %p\n", vq->iov->iov_base);
1217                                 goto out;
1218                         }
1219                 } else {
1220                         /* Header came from socket; we'll need to patch
1221                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1222                          */
1223                         iov_iter_advance(&fixup, sizeof(hdr));
1224                 }
1225                 /* TODO: Should check and handle checksum. */
1226
1227                 num_buffers = cpu_to_vhost16(vq, headcount);
1228                 if (likely(mergeable) &&
1229                     copy_to_iter(&num_buffers, sizeof num_buffers,
1230                                  &fixup) != sizeof num_buffers) {
1231                         vq_err(vq, "Failed num_buffers write");
1232                         vhost_discard_vq_desc(vq, headcount);
1233                         goto out;
1234                 }
1235                 nvq->done_idx += headcount;
1236                 if (nvq->done_idx > VHOST_NET_BATCH)
1237                         vhost_net_signal_used(nvq);
1238                 if (unlikely(vq_log))
1239                         vhost_log_write(vq, vq_log, log, vhost_len,
1240                                         vq->iov, in);
1241                 total_len += vhost_len;
1242         } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1243
1244         if (unlikely(busyloop_intr))
1245                 vhost_poll_queue(&vq->poll);
1246         else if (!sock_len)
1247                 vhost_net_enable_vq(net, vq);
1248 out:
1249         vhost_net_signal_used(nvq);
1250         mutex_unlock(&vq->mutex);
1251 }
1252
1253 static void handle_tx_kick(struct vhost_work *work)
1254 {
1255         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1256                                                   poll.work);
1257         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1258
1259         handle_tx(net);
1260 }
1261
1262 static void handle_rx_kick(struct vhost_work *work)
1263 {
1264         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1265                                                   poll.work);
1266         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1267
1268         handle_rx(net);
1269 }
1270
1271 static void handle_tx_net(struct vhost_work *work)
1272 {
1273         struct vhost_net *net = container_of(work, struct vhost_net,
1274                                              poll[VHOST_NET_VQ_TX].work);
1275         handle_tx(net);
1276 }
1277
1278 static void handle_rx_net(struct vhost_work *work)
1279 {
1280         struct vhost_net *net = container_of(work, struct vhost_net,
1281                                              poll[VHOST_NET_VQ_RX].work);
1282         handle_rx(net);
1283 }
1284
1285 static int vhost_net_open(struct inode *inode, struct file *f)
1286 {
1287         struct vhost_net *n;
1288         struct vhost_dev *dev;
1289         struct vhost_virtqueue **vqs;
1290         void **queue;
1291         struct xdp_buff *xdp;
1292         int i;
1293
1294         n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1295         if (!n)
1296                 return -ENOMEM;
1297         vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1298         if (!vqs) {
1299                 kvfree(n);
1300                 return -ENOMEM;
1301         }
1302
1303         queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1304                               GFP_KERNEL);
1305         if (!queue) {
1306                 kfree(vqs);
1307                 kvfree(n);
1308                 return -ENOMEM;
1309         }
1310         n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1311
1312         xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1313         if (!xdp) {
1314                 kfree(vqs);
1315                 kvfree(n);
1316                 kfree(queue);
1317                 return -ENOMEM;
1318         }
1319         n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1320
1321         dev = &n->dev;
1322         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1323         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1324         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1325         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1326         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1327                 n->vqs[i].ubufs = NULL;
1328                 n->vqs[i].ubuf_info = NULL;
1329                 n->vqs[i].upend_idx = 0;
1330                 n->vqs[i].done_idx = 0;
1331                 n->vqs[i].batched_xdp = 0;
1332                 n->vqs[i].vhost_hlen = 0;
1333                 n->vqs[i].sock_hlen = 0;
1334                 n->vqs[i].rx_ring = NULL;
1335                 vhost_net_buf_init(&n->vqs[i].rxq);
1336         }
1337         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1338                        UIO_MAXIOV + VHOST_NET_BATCH,
1339                        VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true,
1340                        NULL);
1341
1342         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1343         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1344
1345         f->private_data = n;
1346         n->page_frag.page = NULL;
1347         n->refcnt_bias = 0;
1348
1349         return 0;
1350 }
1351
1352 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1353                                         struct vhost_virtqueue *vq)
1354 {
1355         struct socket *sock;
1356         struct vhost_net_virtqueue *nvq =
1357                 container_of(vq, struct vhost_net_virtqueue, vq);
1358
1359         mutex_lock(&vq->mutex);
1360         sock = vhost_vq_get_backend(vq);
1361         vhost_net_disable_vq(n, vq);
1362         vhost_vq_set_backend(vq, NULL);
1363         vhost_net_buf_unproduce(nvq);
1364         nvq->rx_ring = NULL;
1365         mutex_unlock(&vq->mutex);
1366         return sock;
1367 }
1368
1369 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1370                            struct socket **rx_sock)
1371 {
1372         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1373         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1374 }
1375
1376 static void vhost_net_flush_vq(struct vhost_net *n, int index)
1377 {
1378         vhost_poll_flush(n->poll + index);
1379         vhost_poll_flush(&n->vqs[index].vq.poll);
1380 }
1381
1382 static void vhost_net_flush(struct vhost_net *n)
1383 {
1384         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1385         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1386         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1387                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1388                 n->tx_flush = true;
1389                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1390                 /* Wait for all lower device DMAs done. */
1391                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1392                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1393                 n->tx_flush = false;
1394                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1395                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1396         }
1397 }
1398
1399 static int vhost_net_release(struct inode *inode, struct file *f)
1400 {
1401         struct vhost_net *n = f->private_data;
1402         struct socket *tx_sock;
1403         struct socket *rx_sock;
1404
1405         vhost_net_stop(n, &tx_sock, &rx_sock);
1406         vhost_net_flush(n);
1407         vhost_dev_stop(&n->dev);
1408         vhost_dev_cleanup(&n->dev);
1409         vhost_net_vq_reset(n);
1410         if (tx_sock)
1411                 sockfd_put(tx_sock);
1412         if (rx_sock)
1413                 sockfd_put(rx_sock);
1414         /* Make sure no callbacks are outstanding */
1415         synchronize_rcu();
1416         /* We do an extra flush before freeing memory,
1417          * since jobs can re-queue themselves. */
1418         vhost_net_flush(n);
1419         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1420         kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1421         kfree(n->dev.vqs);
1422         if (n->page_frag.page)
1423                 __page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
1424         kvfree(n);
1425         return 0;
1426 }
1427
1428 static struct socket *get_raw_socket(int fd)
1429 {
1430         int r;
1431         struct socket *sock = sockfd_lookup(fd, &r);
1432
1433         if (!sock)
1434                 return ERR_PTR(-ENOTSOCK);
1435
1436         /* Parameter checking */
1437         if (sock->sk->sk_type != SOCK_RAW) {
1438                 r = -ESOCKTNOSUPPORT;
1439                 goto err;
1440         }
1441
1442         if (sock->sk->sk_family != AF_PACKET) {
1443                 r = -EPFNOSUPPORT;
1444                 goto err;
1445         }
1446         return sock;
1447 err:
1448         sockfd_put(sock);
1449         return ERR_PTR(r);
1450 }
1451
1452 static struct ptr_ring *get_tap_ptr_ring(int fd)
1453 {
1454         struct ptr_ring *ring;
1455         struct file *file = fget(fd);
1456
1457         if (!file)
1458                 return NULL;
1459         ring = tun_get_tx_ring(file);
1460         if (!IS_ERR(ring))
1461                 goto out;
1462         ring = tap_get_ptr_ring(file);
1463         if (!IS_ERR(ring))
1464                 goto out;
1465         ring = NULL;
1466 out:
1467         fput(file);
1468         return ring;
1469 }
1470
1471 static struct socket *get_tap_socket(int fd)
1472 {
1473         struct file *file = fget(fd);
1474         struct socket *sock;
1475
1476         if (!file)
1477                 return ERR_PTR(-EBADF);
1478         sock = tun_get_socket(file);
1479         if (!IS_ERR(sock))
1480                 return sock;
1481         sock = tap_get_socket(file);
1482         if (IS_ERR(sock))
1483                 fput(file);
1484         return sock;
1485 }
1486
1487 static struct socket *get_socket(int fd)
1488 {
1489         struct socket *sock;
1490
1491         /* special case to disable backend */
1492         if (fd == -1)
1493                 return NULL;
1494         sock = get_raw_socket(fd);
1495         if (!IS_ERR(sock))
1496                 return sock;
1497         sock = get_tap_socket(fd);
1498         if (!IS_ERR(sock))
1499                 return sock;
1500         return ERR_PTR(-ENOTSOCK);
1501 }
1502
1503 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1504 {
1505         struct socket *sock, *oldsock;
1506         struct vhost_virtqueue *vq;
1507         struct vhost_net_virtqueue *nvq;
1508         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1509         int r;
1510
1511         mutex_lock(&n->dev.mutex);
1512         r = vhost_dev_check_owner(&n->dev);
1513         if (r)
1514                 goto err;
1515
1516         if (index >= VHOST_NET_VQ_MAX) {
1517                 r = -ENOBUFS;
1518                 goto err;
1519         }
1520         vq = &n->vqs[index].vq;
1521         nvq = &n->vqs[index];
1522         mutex_lock(&vq->mutex);
1523
1524         /* Verify that ring has been setup correctly. */
1525         if (!vhost_vq_access_ok(vq)) {
1526                 r = -EFAULT;
1527                 goto err_vq;
1528         }
1529         sock = get_socket(fd);
1530         if (IS_ERR(sock)) {
1531                 r = PTR_ERR(sock);
1532                 goto err_vq;
1533         }
1534
1535         /* start polling new socket */
1536         oldsock = vhost_vq_get_backend(vq);
1537         if (sock != oldsock) {
1538                 ubufs = vhost_net_ubuf_alloc(vq,
1539                                              sock && vhost_sock_zcopy(sock));
1540                 if (IS_ERR(ubufs)) {
1541                         r = PTR_ERR(ubufs);
1542                         goto err_ubufs;
1543                 }
1544
1545                 vhost_net_disable_vq(n, vq);
1546                 vhost_vq_set_backend(vq, sock);
1547                 vhost_net_buf_unproduce(nvq);
1548                 r = vhost_vq_init_access(vq);
1549                 if (r)
1550                         goto err_used;
1551                 r = vhost_net_enable_vq(n, vq);
1552                 if (r)
1553                         goto err_used;
1554                 if (index == VHOST_NET_VQ_RX)
1555                         nvq->rx_ring = get_tap_ptr_ring(fd);
1556
1557                 oldubufs = nvq->ubufs;
1558                 nvq->ubufs = ubufs;
1559
1560                 n->tx_packets = 0;
1561                 n->tx_zcopy_err = 0;
1562                 n->tx_flush = false;
1563         }
1564
1565         mutex_unlock(&vq->mutex);
1566
1567         if (oldubufs) {
1568                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1569                 mutex_lock(&vq->mutex);
1570                 vhost_zerocopy_signal_used(n, vq);
1571                 mutex_unlock(&vq->mutex);
1572         }
1573
1574         if (oldsock) {
1575                 vhost_net_flush_vq(n, index);
1576                 sockfd_put(oldsock);
1577         }
1578
1579         mutex_unlock(&n->dev.mutex);
1580         return 0;
1581
1582 err_used:
1583         vhost_vq_set_backend(vq, oldsock);
1584         vhost_net_enable_vq(n, vq);
1585         if (ubufs)
1586                 vhost_net_ubuf_put_wait_and_free(ubufs);
1587 err_ubufs:
1588         if (sock)
1589                 sockfd_put(sock);
1590 err_vq:
1591         mutex_unlock(&vq->mutex);
1592 err:
1593         mutex_unlock(&n->dev.mutex);
1594         return r;
1595 }
1596
1597 static long vhost_net_reset_owner(struct vhost_net *n)
1598 {
1599         struct socket *tx_sock = NULL;
1600         struct socket *rx_sock = NULL;
1601         long err;
1602         struct vhost_iotlb *umem;
1603
1604         mutex_lock(&n->dev.mutex);
1605         err = vhost_dev_check_owner(&n->dev);
1606         if (err)
1607                 goto done;
1608         umem = vhost_dev_reset_owner_prepare();
1609         if (!umem) {
1610                 err = -ENOMEM;
1611                 goto done;
1612         }
1613         vhost_net_stop(n, &tx_sock, &rx_sock);
1614         vhost_net_flush(n);
1615         vhost_dev_stop(&n->dev);
1616         vhost_dev_reset_owner(&n->dev, umem);
1617         vhost_net_vq_reset(n);
1618 done:
1619         mutex_unlock(&n->dev.mutex);
1620         if (tx_sock)
1621                 sockfd_put(tx_sock);
1622         if (rx_sock)
1623                 sockfd_put(rx_sock);
1624         return err;
1625 }
1626
1627 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1628 {
1629         size_t vhost_hlen, sock_hlen, hdr_len;
1630         int i;
1631
1632         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1633                                (1ULL << VIRTIO_F_VERSION_1))) ?
1634                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1635                         sizeof(struct virtio_net_hdr);
1636         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1637                 /* vhost provides vnet_hdr */
1638                 vhost_hlen = hdr_len;
1639                 sock_hlen = 0;
1640         } else {
1641                 /* socket provides vnet_hdr */
1642                 vhost_hlen = 0;
1643                 sock_hlen = hdr_len;
1644         }
1645         mutex_lock(&n->dev.mutex);
1646         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1647             !vhost_log_access_ok(&n->dev))
1648                 goto out_unlock;
1649
1650         if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) {
1651                 if (vhost_init_device_iotlb(&n->dev, true))
1652                         goto out_unlock;
1653         }
1654
1655         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1656                 mutex_lock(&n->vqs[i].vq.mutex);
1657                 n->vqs[i].vq.acked_features = features;
1658                 n->vqs[i].vhost_hlen = vhost_hlen;
1659                 n->vqs[i].sock_hlen = sock_hlen;
1660                 mutex_unlock(&n->vqs[i].vq.mutex);
1661         }
1662         mutex_unlock(&n->dev.mutex);
1663         return 0;
1664
1665 out_unlock:
1666         mutex_unlock(&n->dev.mutex);
1667         return -EFAULT;
1668 }
1669
1670 static long vhost_net_set_owner(struct vhost_net *n)
1671 {
1672         int r;
1673
1674         mutex_lock(&n->dev.mutex);
1675         if (vhost_dev_has_owner(&n->dev)) {
1676                 r = -EBUSY;
1677                 goto out;
1678         }
1679         r = vhost_net_set_ubuf_info(n);
1680         if (r)
1681                 goto out;
1682         r = vhost_dev_set_owner(&n->dev);
1683         if (r)
1684                 vhost_net_clear_ubuf_info(n);
1685         vhost_net_flush(n);
1686 out:
1687         mutex_unlock(&n->dev.mutex);
1688         return r;
1689 }
1690
1691 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1692                             unsigned long arg)
1693 {
1694         struct vhost_net *n = f->private_data;
1695         void __user *argp = (void __user *)arg;
1696         u64 __user *featurep = argp;
1697         struct vhost_vring_file backend;
1698         u64 features;
1699         int r;
1700
1701         switch (ioctl) {
1702         case VHOST_NET_SET_BACKEND:
1703                 if (copy_from_user(&backend, argp, sizeof backend))
1704                         return -EFAULT;
1705                 return vhost_net_set_backend(n, backend.index, backend.fd);
1706         case VHOST_GET_FEATURES:
1707                 features = VHOST_NET_FEATURES;
1708                 if (copy_to_user(featurep, &features, sizeof features))
1709                         return -EFAULT;
1710                 return 0;
1711         case VHOST_SET_FEATURES:
1712                 if (copy_from_user(&features, featurep, sizeof features))
1713                         return -EFAULT;
1714                 if (features & ~VHOST_NET_FEATURES)
1715                         return -EOPNOTSUPP;
1716                 return vhost_net_set_features(n, features);
1717         case VHOST_GET_BACKEND_FEATURES:
1718                 features = VHOST_NET_BACKEND_FEATURES;
1719                 if (copy_to_user(featurep, &features, sizeof(features)))
1720                         return -EFAULT;
1721                 return 0;
1722         case VHOST_SET_BACKEND_FEATURES:
1723                 if (copy_from_user(&features, featurep, sizeof(features)))
1724                         return -EFAULT;
1725                 if (features & ~VHOST_NET_BACKEND_FEATURES)
1726                         return -EOPNOTSUPP;
1727                 vhost_set_backend_features(&n->dev, features);
1728                 return 0;
1729         case VHOST_RESET_OWNER:
1730                 return vhost_net_reset_owner(n);
1731         case VHOST_SET_OWNER:
1732                 return vhost_net_set_owner(n);
1733         default:
1734                 mutex_lock(&n->dev.mutex);
1735                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1736                 if (r == -ENOIOCTLCMD)
1737                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1738                 else
1739                         vhost_net_flush(n);
1740                 mutex_unlock(&n->dev.mutex);
1741                 return r;
1742         }
1743 }
1744
1745 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1746 {
1747         struct file *file = iocb->ki_filp;
1748         struct vhost_net *n = file->private_data;
1749         struct vhost_dev *dev = &n->dev;
1750         int noblock = file->f_flags & O_NONBLOCK;
1751
1752         return vhost_chr_read_iter(dev, to, noblock);
1753 }
1754
1755 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1756                                         struct iov_iter *from)
1757 {
1758         struct file *file = iocb->ki_filp;
1759         struct vhost_net *n = file->private_data;
1760         struct vhost_dev *dev = &n->dev;
1761
1762         return vhost_chr_write_iter(dev, from);
1763 }
1764
1765 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1766 {
1767         struct vhost_net *n = file->private_data;
1768         struct vhost_dev *dev = &n->dev;
1769
1770         return vhost_chr_poll(file, dev, wait);
1771 }
1772
1773 static const struct file_operations vhost_net_fops = {
1774         .owner          = THIS_MODULE,
1775         .release        = vhost_net_release,
1776         .read_iter      = vhost_net_chr_read_iter,
1777         .write_iter     = vhost_net_chr_write_iter,
1778         .poll           = vhost_net_chr_poll,
1779         .unlocked_ioctl = vhost_net_ioctl,
1780         .compat_ioctl   = compat_ptr_ioctl,
1781         .open           = vhost_net_open,
1782         .llseek         = noop_llseek,
1783 };
1784
1785 static struct miscdevice vhost_net_misc = {
1786         .minor = VHOST_NET_MINOR,
1787         .name = "vhost-net",
1788         .fops = &vhost_net_fops,
1789 };
1790
1791 static int vhost_net_init(void)
1792 {
1793         if (experimental_zcopytx)
1794                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1795         return misc_register(&vhost_net_misc);
1796 }
1797 module_init(vhost_net_init);
1798
1799 static void vhost_net_exit(void)
1800 {
1801         misc_deregister(&vhost_net_misc);
1802 }
1803 module_exit(vhost_net_exit);
1804
1805 MODULE_VERSION("0.0.1");
1806 MODULE_LICENSE("GPL v2");
1807 MODULE_AUTHOR("Michael S. Tsirkin");
1808 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1809 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1810 MODULE_ALIAS("devname:vhost-net");