linux/drivers/vhost/net.c
<<
>>
Prefs
   1/* Copyright (C) 2009 Red Hat, Inc.
   2 * Author: Michael S. Tsirkin <mst@redhat.com>
   3 *
   4 * This work is licensed under the terms of the GNU GPL, version 2.
   5 *
   6 * virtio-net server in host kernel.
   7 */
   8
   9#include <linux/compat.h>
  10#include <linux/eventfd.h>
  11#include <linux/vhost.h>
  12#include <linux/virtio_net.h>
  13#include <linux/miscdevice.h>
  14#include <linux/module.h>
  15#include <linux/moduleparam.h>
  16#include <linux/mutex.h>
  17#include <linux/workqueue.h>
  18#include <linux/file.h>
  19#include <linux/slab.h>
  20#include <linux/sched/clock.h>
  21#include <linux/sched/signal.h>
  22#include <linux/vmalloc.h>
  23
  24#include <linux/net.h>
  25#include <linux/if_packet.h>
  26#include <linux/if_arp.h>
  27#include <linux/if_tun.h>
  28#include <linux/if_macvlan.h>
  29#include <linux/if_tap.h>
  30#include <linux/if_vlan.h>
  31#include <linux/skb_array.h>
  32#include <linux/skbuff.h>
  33
  34#include <net/sock.h>
  35#include <net/xdp.h>
  36
  37#include "vhost.h"
  38
  39static int experimental_zcopytx;
  40module_param(experimental_zcopytx, int, 0444);
  41MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  42                                       " 1 -Enable; 0 - Disable");
  43
  44/* Max number of bytes transferred before requeueing the job.
  45 * Using this limit prevents one virtqueue from starving others. */
  46#define VHOST_NET_WEIGHT 0x80000
  47
  48/* Max number of packets transferred before requeueing the job.
  49 * Using this limit prevents one virtqueue from starving others with small
  50 * pkts.
  51 */
  52#define VHOST_NET_PKT_WEIGHT 256
  53
  54/* MAX number of TX used buffers for outstanding zerocopy */
  55#define VHOST_MAX_PEND 128
  56#define VHOST_GOODCOPY_LEN 256
  57
  58/*
  59 * For transmit, used buffer len is unused; we override it to track buffer
  60 * status internally; used for zerocopy tx only.
  61 */
  62/* Lower device DMA failed */
  63#define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  64/* Lower device DMA done */
  65#define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  66/* Lower device DMA in progress */
  67#define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  68/* Buffer unused */
  69#define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  70
  71#define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  72
  73enum {
  74        VHOST_NET_FEATURES = VHOST_FEATURES |
  75                         (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  76                         (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  77                         (1ULL << VIRTIO_F_IOMMU_PLATFORM)
  78};
  79
  80enum {
  81        VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
  82};
  83
  84enum {
  85        VHOST_NET_VQ_RX = 0,
  86        VHOST_NET_VQ_TX = 1,
  87        VHOST_NET_VQ_MAX = 2,
  88};
  89
  90struct 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
 102struct vhost_net_buf {
 103        void **queue;
 104        int tail;
 105        int head;
 106};
 107
 108struct vhost_net_virtqueue {
 109        struct vhost_virtqueue vq;
 110        size_t vhost_hlen;
 111        size_t sock_hlen;
 112        /* vhost zerocopy support fields below: */
 113        /* last used idx for outstanding DMA zerocopy buffers */
 114        int upend_idx;
 115        /* For TX, first used idx for DMA done zerocopy buffers
 116         * For RX, number of batched heads
 117         */
 118        int done_idx;
 119        /* Number of XDP frames batched */
 120        int batched_xdp;
 121        /* an array of userspace buffers info */
 122        struct ubuf_info *ubuf_info;
 123        /* Reference counting for outstanding ubufs.
 124         * Protected by vq mutex. Writers must also take device mutex. */
 125        struct vhost_net_ubuf_ref *ubufs;
 126        struct ptr_ring *rx_ring;
 127        struct vhost_net_buf rxq;
 128        /* Batched XDP buffs */
 129        struct xdp_buff *xdp;
 130};
 131
 132struct 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
 150static unsigned vhost_net_zcopy_mask __read_mostly;
 151
 152static 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
 160static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 161{
 162        return rxq->tail - rxq->head;
 163}
 164
 165static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 166{
 167        return rxq->tail == rxq->head;
 168}
 169
 170static 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
 177static 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
 187static 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
 199static 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
 210static 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
 220out:
 221        return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 222}
 223
 224static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 225{
 226        rxq->head = rxq->tail = 0;
 227}
 228
 229static void vhost_net_enable_zcopy(int vq)
 230{
 231        vhost_net_zcopy_mask |= 0x1 << vq;
 232}
 233
 234static struct vhost_net_ubuf_ref *
 235vhost_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
 250static 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
 258static 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
 264static 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
 270static 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
 280static 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
 298err:
 299        vhost_net_clear_ubuf_info(n);
 300        return -ENOMEM;
 301}
 302
 303static 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
 320static 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
 329static void vhost_net_tx_err(struct vhost_net *net)
 330{
 331        ++net->tx_zcopy_err;
 332}
 333
 334static 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
 343static bool vhost_sock_zcopy(struct socket *sock)
 344{
 345        return unlikely(experimental_zcopytx) &&
 346                sock_flag(sock->sk, SOCK_ZEROCOPY);
 347}
 348
 349static 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 */
 359static 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
 385static void vhost_zerocopy_callback(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
 411static inline unsigned long busy_clock(void)
 412{
 413        return local_clock() >> 10;
 414}
 415
 416static 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
 422static 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 (!vq->private_data)
 429                return;
 430        vhost_poll_stop(poll);
 431}
 432
 433static 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 = vq->private_data;
 442        if (!sock)
 443                return 0;
 444
 445        return vhost_poll_start(poll, sock->file);
 446}
 447
 448static 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
 460static 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 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                return;
 480        }
 481
 482signal_used:
 483        vhost_net_signal_used(nvq);
 484        nvq->batched_xdp = 0;
 485}
 486
 487static int sock_has_rx_data(struct socket *sock)
 488{
 489        if (unlikely(!sock))
 490                return 0;
 491
 492        if (sock->ops->peek_len)
 493                return sock->ops->peek_len(sock);
 494
 495        return skb_queue_empty(&sock->sk->sk_receive_queue);
 496}
 497
 498static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
 499                                          struct vhost_virtqueue *vq)
 500{
 501        if (!vhost_vq_avail_empty(&net->dev, vq)) {
 502                vhost_poll_queue(&vq->poll);
 503        } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 504                vhost_disable_notify(&net->dev, vq);
 505                vhost_poll_queue(&vq->poll);
 506        }
 507}
 508
 509static void vhost_net_busy_poll(struct vhost_net *net,
 510                                struct vhost_virtqueue *rvq,
 511                                struct vhost_virtqueue *tvq,
 512                                bool *busyloop_intr,
 513                                bool poll_rx)
 514{
 515        unsigned long busyloop_timeout;
 516        unsigned long endtime;
 517        struct socket *sock;
 518        struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
 519
 520        /* Try to hold the vq mutex of the paired virtqueue. We can't
 521         * use mutex_lock() here since we could not guarantee a
 522         * consistenet lock ordering.
 523         */
 524        if (!mutex_trylock(&vq->mutex))
 525                return;
 526
 527        vhost_disable_notify(&net->dev, vq);
 528        sock = rvq->private_data;
 529
 530        busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
 531                                     tvq->busyloop_timeout;
 532
 533        preempt_disable();
 534        endtime = busy_clock() + busyloop_timeout;
 535
 536        while (vhost_can_busy_poll(endtime)) {
 537                if (vhost_has_work(&net->dev)) {
 538                        *busyloop_intr = true;
 539                        break;
 540                }
 541
 542                if ((sock_has_rx_data(sock) &&
 543                     !vhost_vq_avail_empty(&net->dev, rvq)) ||
 544                    !vhost_vq_avail_empty(&net->dev, tvq))
 545                        break;
 546
 547                cpu_relax();
 548        }
 549
 550        preempt_enable();
 551
 552        if (poll_rx || sock_has_rx_data(sock))
 553                vhost_net_busy_poll_try_queue(net, vq);
 554        else if (!poll_rx) /* On tx here, sock has no rx data. */
 555                vhost_enable_notify(&net->dev, rvq);
 556
 557        mutex_unlock(&vq->mutex);
 558}
 559
 560static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 561                                    struct vhost_net_virtqueue *tnvq,
 562                                    unsigned int *out_num, unsigned int *in_num,
 563                                    struct msghdr *msghdr, bool *busyloop_intr)
 564{
 565        struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 566        struct vhost_virtqueue *rvq = &rnvq->vq;
 567        struct vhost_virtqueue *tvq = &tnvq->vq;
 568
 569        int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 570                                  out_num, in_num, NULL, NULL);
 571
 572        if (r == tvq->num && tvq->busyloop_timeout) {
 573                /* Flush batched packets first */
 574                if (!vhost_sock_zcopy(tvq->private_data))
 575                        vhost_tx_batch(net, tnvq, tvq->private_data, msghdr);
 576
 577                vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
 578
 579                r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 580                                      out_num, in_num, NULL, NULL);
 581        }
 582
 583        return r;
 584}
 585
 586static bool vhost_exceeds_maxpend(struct vhost_net *net)
 587{
 588        struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 589        struct vhost_virtqueue *vq = &nvq->vq;
 590
 591        return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 592               min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 593}
 594
 595static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
 596                            size_t hdr_size, int out)
 597{
 598        /* Skip header. TODO: support TSO. */
 599        size_t len = iov_length(vq->iov, out);
 600
 601        iov_iter_init(iter, WRITE, vq->iov, out, len);
 602        iov_iter_advance(iter, hdr_size);
 603
 604        return iov_iter_count(iter);
 605}
 606
 607static int get_tx_bufs(struct vhost_net *net,
 608                       struct vhost_net_virtqueue *nvq,
 609                       struct msghdr *msg,
 610                       unsigned int *out, unsigned int *in,
 611                       size_t *len, bool *busyloop_intr)
 612{
 613        struct vhost_virtqueue *vq = &nvq->vq;
 614        int ret;
 615
 616        ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
 617
 618        if (ret < 0 || ret == vq->num)
 619                return ret;
 620
 621        if (*in) {
 622                vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 623                        *out, *in);
 624                return -EFAULT;
 625        }
 626
 627        /* Sanity check */
 628        *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 629        if (*len == 0) {
 630                vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 631                        *len, nvq->vhost_hlen);
 632                return -EFAULT;
 633        }
 634
 635        return ret;
 636}
 637
 638static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 639{
 640        return total_len < VHOST_NET_WEIGHT &&
 641               !vhost_vq_avail_empty(vq->dev, vq);
 642}
 643
 644#define SKB_FRAG_PAGE_ORDER     get_order(32768)
 645
 646static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
 647                                       struct page_frag *pfrag, gfp_t gfp)
 648{
 649        if (pfrag->page) {
 650                if (pfrag->offset + sz <= pfrag->size)
 651                        return true;
 652                __page_frag_cache_drain(pfrag->page, net->refcnt_bias);
 653        }
 654
 655        pfrag->offset = 0;
 656        net->refcnt_bias = 0;
 657        if (SKB_FRAG_PAGE_ORDER) {
 658                /* Avoid direct reclaim but allow kswapd to wake */
 659                pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
 660                                          __GFP_COMP | __GFP_NOWARN |
 661                                          __GFP_NORETRY,
 662                                          SKB_FRAG_PAGE_ORDER);
 663                if (likely(pfrag->page)) {
 664                        pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
 665                        goto done;
 666                }
 667        }
 668        pfrag->page = alloc_page(gfp);
 669        if (likely(pfrag->page)) {
 670                pfrag->size = PAGE_SIZE;
 671                goto done;
 672        }
 673        return false;
 674
 675done:
 676        net->refcnt_bias = USHRT_MAX;
 677        page_ref_add(pfrag->page, USHRT_MAX - 1);
 678        return true;
 679}
 680
 681#define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 682
 683static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
 684                               struct iov_iter *from)
 685{
 686        struct vhost_virtqueue *vq = &nvq->vq;
 687        struct vhost_net *net = container_of(vq->dev, struct vhost_net,
 688                                             dev);
 689        struct socket *sock = vq->private_data;
 690        struct page_frag *alloc_frag = &net->page_frag;
 691        struct virtio_net_hdr *gso;
 692        struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
 693        struct tun_xdp_hdr *hdr;
 694        size_t len = iov_iter_count(from);
 695        int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
 696        int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 697        int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
 698        int sock_hlen = nvq->sock_hlen;
 699        void *buf;
 700        int copied;
 701
 702        if (unlikely(len < nvq->sock_hlen))
 703                return -EFAULT;
 704
 705        if (SKB_DATA_ALIGN(len + pad) +
 706            SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
 707                return -ENOSPC;
 708
 709        buflen += SKB_DATA_ALIGN(len + pad);
 710        alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
 711        if (unlikely(!vhost_net_page_frag_refill(net, buflen,
 712                                                 alloc_frag, GFP_KERNEL)))
 713                return -ENOMEM;
 714
 715        buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 716        copied = copy_page_from_iter(alloc_frag->page,
 717                                     alloc_frag->offset +
 718                                     offsetof(struct tun_xdp_hdr, gso),
 719                                     sock_hlen, from);
 720        if (copied != sock_hlen)
 721                return -EFAULT;
 722
 723        hdr = buf;
 724        gso = &hdr->gso;
 725
 726        if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
 727            vhost16_to_cpu(vq, gso->csum_start) +
 728            vhost16_to_cpu(vq, gso->csum_offset) + 2 >
 729            vhost16_to_cpu(vq, gso->hdr_len)) {
 730                gso->hdr_len = cpu_to_vhost16(vq,
 731                               vhost16_to_cpu(vq, gso->csum_start) +
 732                               vhost16_to_cpu(vq, gso->csum_offset) + 2);
 733
 734                if (vhost16_to_cpu(vq, gso->hdr_len) > len)
 735                        return -EINVAL;
 736        }
 737
 738        len -= sock_hlen;
 739        copied = copy_page_from_iter(alloc_frag->page,
 740                                     alloc_frag->offset + pad,
 741                                     len, from);
 742        if (copied != len)
 743                return -EFAULT;
 744
 745        xdp->data_hard_start = buf;
 746        xdp->data = buf + pad;
 747        xdp->data_end = xdp->data + len;
 748        hdr->buflen = buflen;
 749
 750        --net->refcnt_bias;
 751        alloc_frag->offset += buflen;
 752
 753        ++nvq->batched_xdp;
 754
 755        return 0;
 756}
 757
 758static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 759{
 760        struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 761        struct vhost_virtqueue *vq = &nvq->vq;
 762        unsigned out, in;
 763        int head;
 764        struct msghdr msg = {
 765                .msg_name = NULL,
 766                .msg_namelen = 0,
 767                .msg_control = NULL,
 768                .msg_controllen = 0,
 769                .msg_flags = MSG_DONTWAIT,
 770        };
 771        size_t len, total_len = 0;
 772        int err;
 773        int sent_pkts = 0;
 774        bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
 775
 776        do {
 777                bool busyloop_intr = false;
 778
 779                if (nvq->done_idx == VHOST_NET_BATCH)
 780                        vhost_tx_batch(net, nvq, sock, &msg);
 781
 782                head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 783                                   &busyloop_intr);
 784                /* On error, stop handling until the next kick. */
 785                if (unlikely(head < 0))
 786                        break;
 787                /* Nothing new?  Wait for eventfd to tell us they refilled. */
 788                if (head == vq->num) {
 789                        if (unlikely(busyloop_intr)) {
 790                                vhost_poll_queue(&vq->poll);
 791                        } else if (unlikely(vhost_enable_notify(&net->dev,
 792                                                                vq))) {
 793                                vhost_disable_notify(&net->dev, vq);
 794                                continue;
 795                        }
 796                        break;
 797                }
 798
 799                total_len += len;
 800
 801                /* For simplicity, TX batching is only enabled if
 802                 * sndbuf is unlimited.
 803                 */
 804                if (sock_can_batch) {
 805                        err = vhost_net_build_xdp(nvq, &msg.msg_iter);
 806                        if (!err) {
 807                                goto done;
 808                        } else if (unlikely(err != -ENOSPC)) {
 809                                vhost_tx_batch(net, nvq, sock, &msg);
 810                                vhost_discard_vq_desc(vq, 1);
 811                                vhost_net_enable_vq(net, vq);
 812                                break;
 813                        }
 814
 815                        /* We can't build XDP buff, go for single
 816                         * packet path but let's flush batched
 817                         * packets.
 818                         */
 819                        vhost_tx_batch(net, nvq, sock, &msg);
 820                        msg.msg_control = NULL;
 821                } else {
 822                        if (tx_can_batch(vq, total_len))
 823                                msg.msg_flags |= MSG_MORE;
 824                        else
 825                                msg.msg_flags &= ~MSG_MORE;
 826                }
 827
 828                /* TODO: Check specific error and bomb out unless ENOBUFS? */
 829                err = sock->ops->sendmsg(sock, &msg, len);
 830                if (unlikely(err < 0)) {
 831                        vhost_discard_vq_desc(vq, 1);
 832                        vhost_net_enable_vq(net, vq);
 833                        break;
 834                }
 835                if (err != len)
 836                        pr_debug("Truncated TX packet: len %d != %zd\n",
 837                                 err, len);
 838done:
 839                vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 840                vq->heads[nvq->done_idx].len = 0;
 841                ++nvq->done_idx;
 842        } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 843
 844        vhost_tx_batch(net, nvq, sock, &msg);
 845}
 846
 847static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 848{
 849        struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 850        struct vhost_virtqueue *vq = &nvq->vq;
 851        unsigned out, in;
 852        int head;
 853        struct msghdr msg = {
 854                .msg_name = NULL,
 855                .msg_namelen = 0,
 856                .msg_control = NULL,
 857                .msg_controllen = 0,
 858                .msg_flags = MSG_DONTWAIT,
 859        };
 860        struct tun_msg_ctl ctl;
 861        size_t len, total_len = 0;
 862        int err;
 863        struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 864        bool zcopy_used;
 865        int sent_pkts = 0;
 866
 867        do {
 868                bool busyloop_intr;
 869
 870                /* Release DMAs done buffers first */
 871                vhost_zerocopy_signal_used(net, vq);
 872
 873                busyloop_intr = false;
 874                head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 875                                   &busyloop_intr);
 876                /* On error, stop handling until the next kick. */
 877                if (unlikely(head < 0))
 878                        break;
 879                /* Nothing new?  Wait for eventfd to tell us they refilled. */
 880                if (head == vq->num) {
 881                        if (unlikely(busyloop_intr)) {
 882                                vhost_poll_queue(&vq->poll);
 883                        } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 884                                vhost_disable_notify(&net->dev, vq);
 885                                continue;
 886                        }
 887                        break;
 888                }
 889
 890                zcopy_used = len >= VHOST_GOODCOPY_LEN
 891                             && !vhost_exceeds_maxpend(net)
 892                             && vhost_net_tx_select_zcopy(net);
 893
 894                /* use msg_control to pass vhost zerocopy ubuf info to skb */
 895                if (zcopy_used) {
 896                        struct ubuf_info *ubuf;
 897                        ubuf = nvq->ubuf_info + nvq->upend_idx;
 898
 899                        vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 900                        vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 901                        ubuf->callback = vhost_zerocopy_callback;
 902                        ubuf->ctx = nvq->ubufs;
 903                        ubuf->desc = nvq->upend_idx;
 904                        refcount_set(&ubuf->refcnt, 1);
 905                        msg.msg_control = &ctl;
 906                        ctl.type = TUN_MSG_UBUF;
 907                        ctl.ptr = ubuf;
 908                        msg.msg_controllen = sizeof(ctl);
 909                        ubufs = nvq->ubufs;
 910                        atomic_inc(&ubufs->refcount);
 911                        nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 912                } else {
 913                        msg.msg_control = NULL;
 914                        ubufs = NULL;
 915                }
 916                total_len += len;
 917                if (tx_can_batch(vq, total_len) &&
 918                    likely(!vhost_exceeds_maxpend(net))) {
 919                        msg.msg_flags |= MSG_MORE;
 920                } else {
 921                        msg.msg_flags &= ~MSG_MORE;
 922                }
 923
 924                /* TODO: Check specific error and bomb out unless ENOBUFS? */
 925                err = sock->ops->sendmsg(sock, &msg, len);
 926                if (unlikely(err < 0)) {
 927                        if (zcopy_used) {
 928                                vhost_net_ubuf_put(ubufs);
 929                                nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 930                                        % UIO_MAXIOV;
 931                        }
 932                        vhost_discard_vq_desc(vq, 1);
 933                        vhost_net_enable_vq(net, vq);
 934                        break;
 935                }
 936                if (err != len)
 937                        pr_debug("Truncated TX packet: "
 938                                 " len %d != %zd\n", err, len);
 939                if (!zcopy_used)
 940                        vhost_add_used_and_signal(&net->dev, vq, head, 0);
 941                else
 942                        vhost_zerocopy_signal_used(net, vq);
 943                vhost_net_tx_packet(net);
 944        } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 945}
 946
 947/* Expects to be always run from workqueue - which acts as
 948 * read-size critical section for our kind of RCU. */
 949static void handle_tx(struct vhost_net *net)
 950{
 951        struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 952        struct vhost_virtqueue *vq = &nvq->vq;
 953        struct socket *sock;
 954
 955        mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
 956        sock = vq->private_data;
 957        if (!sock)
 958                goto out;
 959
 960        if (!vq_iotlb_prefetch(vq))
 961                goto out;
 962
 963        vhost_disable_notify(&net->dev, vq);
 964        vhost_net_disable_vq(net, vq);
 965
 966        if (vhost_sock_zcopy(sock))
 967                handle_tx_zerocopy(net, sock);
 968        else
 969                handle_tx_copy(net, sock);
 970
 971out:
 972        mutex_unlock(&vq->mutex);
 973}
 974
 975static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 976{
 977        struct sk_buff *head;
 978        int len = 0;
 979        unsigned long flags;
 980
 981        if (rvq->rx_ring)
 982                return vhost_net_buf_peek(rvq);
 983
 984        spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 985        head = skb_peek(&sk->sk_receive_queue);
 986        if (likely(head)) {
 987                len = head->len;
 988                if (skb_vlan_tag_present(head))
 989                        len += VLAN_HLEN;
 990        }
 991
 992        spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 993        return len;
 994}
 995
 996static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
 997                                      bool *busyloop_intr)
 998{
 999        struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1000        struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1001        struct vhost_virtqueue *rvq = &rnvq->vq;
1002        struct vhost_virtqueue *tvq = &tnvq->vq;
1003        int len = peek_head_len(rnvq, sk);
1004
1005        if (!len && rvq->busyloop_timeout) {
1006                /* Flush batched heads first */
1007                vhost_net_signal_used(rnvq);
1008                /* Both tx vq and rx socket were polled here */
1009                vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1010
1011                len = peek_head_len(rnvq, sk);
1012        }
1013
1014        return len;
1015}
1016
1017/* This is a multi-buffer version of vhost_get_desc, that works if
1018 *      vq has read descriptors only.
1019 * @vq          - the relevant virtqueue
1020 * @datalen     - data length we'll be reading
1021 * @iovcount    - returned count of io vectors we fill
1022 * @log         - vhost log
1023 * @log_num     - log offset
1024 * @quota       - headcount quota, 1 for big buffer
1025 *      returns number of buffer heads allocated, negative on error
1026 */
1027static int get_rx_bufs(struct vhost_virtqueue *vq,
1028                       struct vring_used_elem *heads,
1029                       int datalen,
1030                       unsigned *iovcount,
1031                       struct vhost_log *log,
1032                       unsigned *log_num,
1033                       unsigned int quota)
1034{
1035        unsigned int out, in;
1036        int seg = 0;
1037        int headcount = 0;
1038        unsigned d;
1039        int r, nlogs = 0;
1040        /* len is always initialized before use since we are always called with
1041         * datalen > 0.
1042         */
1043        u32 uninitialized_var(len);
1044
1045        while (datalen > 0 && headcount < quota) {
1046                if (unlikely(seg >= UIO_MAXIOV)) {
1047                        r = -ENOBUFS;
1048                        goto err;
1049                }
1050                r = vhost_get_vq_desc(vq, vq->iov + seg,
1051                                      ARRAY_SIZE(vq->iov) - seg, &out,
1052                                      &in, log, log_num);
1053                if (unlikely(r < 0))
1054                        goto err;
1055
1056                d = r;
1057                if (d == vq->num) {
1058                        r = 0;
1059                        goto err;
1060                }
1061                if (unlikely(out || in <= 0)) {
1062                        vq_err(vq, "unexpected descriptor format for RX: "
1063                                "out %d, in %d\n", out, in);
1064                        r = -EINVAL;
1065                        goto err;
1066                }
1067                if (unlikely(log)) {
1068                        nlogs += *log_num;
1069                        log += *log_num;
1070                }
1071                heads[headcount].id = cpu_to_vhost32(vq, d);
1072                len = iov_length(vq->iov + seg, in);
1073                heads[headcount].len = cpu_to_vhost32(vq, len);
1074                datalen -= len;
1075                ++headcount;
1076                seg += in;
1077        }
1078        heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1079        *iovcount = seg;
1080        if (unlikely(log))
1081                *log_num = nlogs;
1082
1083        /* Detect overrun */
1084        if (unlikely(datalen > 0)) {
1085                r = UIO_MAXIOV + 1;
1086                goto err;
1087        }
1088        return headcount;
1089err:
1090        vhost_discard_vq_desc(vq, headcount);
1091        return r;
1092}
1093
1094/* Expects to be always run from workqueue - which acts as
1095 * read-size critical section for our kind of RCU. */
1096static void handle_rx(struct vhost_net *net)
1097{
1098        struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1099        struct vhost_virtqueue *vq = &nvq->vq;
1100        unsigned uninitialized_var(in), log;
1101        struct vhost_log *vq_log;
1102        struct msghdr msg = {
1103                .msg_name = NULL,
1104                .msg_namelen = 0,
1105                .msg_control = NULL, /* FIXME: get and handle RX aux data. */
1106                .msg_controllen = 0,
1107                .msg_flags = MSG_DONTWAIT,
1108        };
1109        struct virtio_net_hdr hdr = {
1110                .flags = 0,
1111                .gso_type = VIRTIO_NET_HDR_GSO_NONE
1112        };
1113        size_t total_len = 0;
1114        int err, mergeable;
1115        s16 headcount;
1116        size_t vhost_hlen, sock_hlen;
1117        size_t vhost_len, sock_len;
1118        bool busyloop_intr = false;
1119        struct socket *sock;
1120        struct iov_iter fixup;
1121        __virtio16 num_buffers;
1122        int recv_pkts = 0;
1123
1124        mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1125        sock = vq->private_data;
1126        if (!sock)
1127                goto out;
1128
1129        if (!vq_iotlb_prefetch(vq))
1130                goto out;
1131
1132        vhost_disable_notify(&net->dev, vq);
1133        vhost_net_disable_vq(net, vq);
1134
1135        vhost_hlen = nvq->vhost_hlen;
1136        sock_hlen = nvq->sock_hlen;
1137
1138        vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1139                vq->log : NULL;
1140        mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1141
1142        do {
1143                sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1144                                                      &busyloop_intr);
1145                if (!sock_len)
1146                        break;
1147                sock_len += sock_hlen;
1148                vhost_len = sock_len + vhost_hlen;
1149                headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1150                                        vhost_len, &in, vq_log, &log,
1151                                        likely(mergeable) ? UIO_MAXIOV : 1);
1152                /* On error, stop handling until the next kick. */
1153                if (unlikely(headcount < 0))
1154                        goto out;
1155                /* OK, now we need to know about added descriptors. */
1156                if (!headcount) {
1157                        if (unlikely(busyloop_intr)) {
1158                                vhost_poll_queue(&vq->poll);
1159                        } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1160                                /* They have slipped one in as we were
1161                                 * doing that: check again. */
1162                                vhost_disable_notify(&net->dev, vq);
1163                                continue;
1164                        }
1165                        /* Nothing new?  Wait for eventfd to tell us
1166                         * they refilled. */
1167                        goto out;
1168                }
1169                busyloop_intr = false;
1170                if (nvq->rx_ring)
1171                        msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1172                /* On overrun, truncate and discard */
1173                if (unlikely(headcount > UIO_MAXIOV)) {
1174                        iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
1175                        err = sock->ops->recvmsg(sock, &msg,
1176                                                 1, MSG_DONTWAIT | MSG_TRUNC);
1177                        pr_debug("Discarded rx packet: len %zd\n", sock_len);
1178                        continue;
1179                }
1180                /* We don't need to be notified again. */
1181                iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
1182                fixup = msg.msg_iter;
1183                if (unlikely((vhost_hlen))) {
1184                        /* We will supply the header ourselves
1185                         * TODO: support TSO.
1186                         */
1187                        iov_iter_advance(&msg.msg_iter, vhost_hlen);
1188                }
1189                err = sock->ops->recvmsg(sock, &msg,
1190                                         sock_len, MSG_DONTWAIT | MSG_TRUNC);
1191                /* Userspace might have consumed the packet meanwhile:
1192                 * it's not supposed to do this usually, but might be hard
1193                 * to prevent. Discard data we got (if any) and keep going. */
1194                if (unlikely(err != sock_len)) {
1195                        pr_debug("Discarded rx packet: "
1196                                 " len %d, expected %zd\n", err, sock_len);
1197                        vhost_discard_vq_desc(vq, headcount);
1198                        continue;
1199                }
1200                /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1201                if (unlikely(vhost_hlen)) {
1202                        if (copy_to_iter(&hdr, sizeof(hdr),
1203                                         &fixup) != sizeof(hdr)) {
1204                                vq_err(vq, "Unable to write vnet_hdr "
1205                                       "at addr %p\n", vq->iov->iov_base);
1206                                goto out;
1207                        }
1208                } else {
1209                        /* Header came from socket; we'll need to patch
1210                         * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1211                         */
1212                        iov_iter_advance(&fixup, sizeof(hdr));
1213                }
1214                /* TODO: Should check and handle checksum. */
1215
1216                num_buffers = cpu_to_vhost16(vq, headcount);
1217                if (likely(mergeable) &&
1218                    copy_to_iter(&num_buffers, sizeof num_buffers,
1219                                 &fixup) != sizeof num_buffers) {
1220                        vq_err(vq, "Failed num_buffers write");
1221                        vhost_discard_vq_desc(vq, headcount);
1222                        goto out;
1223                }
1224                nvq->done_idx += headcount;
1225                if (nvq->done_idx > VHOST_NET_BATCH)
1226                        vhost_net_signal_used(nvq);
1227                if (unlikely(vq_log))
1228                        vhost_log_write(vq, vq_log, log, vhost_len,
1229                                        vq->iov, in);
1230                total_len += vhost_len;
1231        } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1232
1233        if (unlikely(busyloop_intr))
1234                vhost_poll_queue(&vq->poll);
1235        else if (!sock_len)
1236                vhost_net_enable_vq(net, vq);
1237out:
1238        vhost_net_signal_used(nvq);
1239        mutex_unlock(&vq->mutex);
1240}
1241
1242static void handle_tx_kick(struct vhost_work *work)
1243{
1244        struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1245                                                  poll.work);
1246        struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1247
1248        handle_tx(net);
1249}
1250
1251static void handle_rx_kick(struct vhost_work *work)
1252{
1253        struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1254                                                  poll.work);
1255        struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1256
1257        handle_rx(net);
1258}
1259
1260static void handle_tx_net(struct vhost_work *work)
1261{
1262        struct vhost_net *net = container_of(work, struct vhost_net,
1263                                             poll[VHOST_NET_VQ_TX].work);
1264        handle_tx(net);
1265}
1266
1267static void handle_rx_net(struct vhost_work *work)
1268{
1269        struct vhost_net *net = container_of(work, struct vhost_net,
1270                                             poll[VHOST_NET_VQ_RX].work);
1271        handle_rx(net);
1272}
1273
1274static int vhost_net_open(struct inode *inode, struct file *f)
1275{
1276        struct vhost_net *n;
1277        struct vhost_dev *dev;
1278        struct vhost_virtqueue **vqs;
1279        void **queue;
1280        struct xdp_buff *xdp;
1281        int i;
1282
1283        n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1284        if (!n)
1285                return -ENOMEM;
1286        vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1287        if (!vqs) {
1288                kvfree(n);
1289                return -ENOMEM;
1290        }
1291
1292        queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1293                              GFP_KERNEL);
1294        if (!queue) {
1295                kfree(vqs);
1296                kvfree(n);
1297                return -ENOMEM;
1298        }
1299        n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1300
1301        xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1302        if (!xdp) {
1303                kfree(vqs);
1304                kvfree(n);
1305                kfree(queue);
1306                return -ENOMEM;
1307        }
1308        n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1309
1310        dev = &n->dev;
1311        vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1312        vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1313        n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1314        n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1315        for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1316                n->vqs[i].ubufs = NULL;
1317                n->vqs[i].ubuf_info = NULL;
1318                n->vqs[i].upend_idx = 0;
1319                n->vqs[i].done_idx = 0;
1320                n->vqs[i].batched_xdp = 0;
1321                n->vqs[i].vhost_hlen = 0;
1322                n->vqs[i].sock_hlen = 0;
1323                n->vqs[i].rx_ring = NULL;
1324                vhost_net_buf_init(&n->vqs[i].rxq);
1325        }
1326        vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1327                       UIO_MAXIOV + VHOST_NET_BATCH,
1328                       VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT);
1329
1330        vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1331        vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1332
1333        f->private_data = n;
1334        n->page_frag.page = NULL;
1335        n->refcnt_bias = 0;
1336
1337        return 0;
1338}
1339
1340static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1341                                        struct vhost_virtqueue *vq)
1342{
1343        struct socket *sock;
1344        struct vhost_net_virtqueue *nvq =
1345                container_of(vq, struct vhost_net_virtqueue, vq);
1346
1347        mutex_lock(&vq->mutex);
1348        sock = vq->private_data;
1349        vhost_net_disable_vq(n, vq);
1350        vq->private_data = NULL;
1351        vhost_net_buf_unproduce(nvq);
1352        nvq->rx_ring = NULL;
1353        mutex_unlock(&vq->mutex);
1354        return sock;
1355}
1356
1357static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1358                           struct socket **rx_sock)
1359{
1360        *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1361        *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1362}
1363
1364static void vhost_net_flush_vq(struct vhost_net *n, int index)
1365{
1366        vhost_poll_flush(n->poll + index);
1367        vhost_poll_flush(&n->vqs[index].vq.poll);
1368}
1369
1370static void vhost_net_flush(struct vhost_net *n)
1371{
1372        vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1373        vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1374        if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1375                mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1376                n->tx_flush = true;
1377                mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1378                /* Wait for all lower device DMAs done. */
1379                vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1380                mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1381                n->tx_flush = false;
1382                atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1383                mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1384        }
1385}
1386
1387static int vhost_net_release(struct inode *inode, struct file *f)
1388{
1389        struct vhost_net *n = f->private_data;
1390        struct socket *tx_sock;
1391        struct socket *rx_sock;
1392
1393        vhost_net_stop(n, &tx_sock, &rx_sock);
1394        vhost_net_flush(n);
1395        vhost_dev_stop(&n->dev);
1396        vhost_dev_cleanup(&n->dev);
1397        vhost_net_vq_reset(n);
1398        if (tx_sock)
1399                sockfd_put(tx_sock);
1400        if (rx_sock)
1401                sockfd_put(rx_sock);
1402        /* Make sure no callbacks are outstanding */
1403        synchronize_rcu_bh();
1404        /* We do an extra flush before freeing memory,
1405         * since jobs can re-queue themselves. */
1406        vhost_net_flush(n);
1407        kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1408        kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1409        kfree(n->dev.vqs);
1410        if (n->page_frag.page)
1411                __page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
1412        kvfree(n);
1413        return 0;
1414}
1415
1416static struct socket *get_raw_socket(int fd)
1417{
1418        struct {
1419                struct sockaddr_ll sa;
1420                char  buf[MAX_ADDR_LEN];
1421        } uaddr;
1422        int r;
1423        struct socket *sock = sockfd_lookup(fd, &r);
1424
1425        if (!sock)
1426                return ERR_PTR(-ENOTSOCK);
1427
1428        /* Parameter checking */
1429        if (sock->sk->sk_type != SOCK_RAW) {
1430                r = -ESOCKTNOSUPPORT;
1431                goto err;
1432        }
1433
1434        r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa, 0);
1435        if (r < 0)
1436                goto err;
1437
1438        if (uaddr.sa.sll_family != AF_PACKET) {
1439                r = -EPFNOSUPPORT;
1440                goto err;
1441        }
1442        return sock;
1443err:
1444        sockfd_put(sock);
1445        return ERR_PTR(r);
1446}
1447
1448static struct ptr_ring *get_tap_ptr_ring(int fd)
1449{
1450        struct ptr_ring *ring;
1451        struct file *file = fget(fd);
1452
1453        if (!file)
1454                return NULL;
1455        ring = tun_get_tx_ring(file);
1456        if (!IS_ERR(ring))
1457                goto out;
1458        ring = tap_get_ptr_ring(file);
1459        if (!IS_ERR(ring))
1460                goto out;
1461        ring = NULL;
1462out:
1463        fput(file);
1464        return ring;
1465}
1466
1467static struct socket *get_tap_socket(int fd)
1468{
1469        struct file *file = fget(fd);
1470        struct socket *sock;
1471
1472        if (!file)
1473                return ERR_PTR(-EBADF);
1474        sock = tun_get_socket(file);
1475        if (!IS_ERR(sock))
1476                return sock;
1477        sock = tap_get_socket(file);
1478        if (IS_ERR(sock))
1479                fput(file);
1480        return sock;
1481}
1482
1483static struct socket *get_socket(int fd)
1484{
1485        struct socket *sock;
1486
1487        /* special case to disable backend */
1488        if (fd == -1)
1489                return NULL;
1490        sock = get_raw_socket(fd);
1491        if (!IS_ERR(sock))
1492                return sock;
1493        sock = get_tap_socket(fd);
1494        if (!IS_ERR(sock))
1495                return sock;
1496        return ERR_PTR(-ENOTSOCK);
1497}
1498
1499static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1500{
1501        struct socket *sock, *oldsock;
1502        struct vhost_virtqueue *vq;
1503        struct vhost_net_virtqueue *nvq;
1504        struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1505        int r;
1506
1507        mutex_lock(&n->dev.mutex);
1508        r = vhost_dev_check_owner(&n->dev);
1509        if (r)
1510                goto err;
1511
1512        if (index >= VHOST_NET_VQ_MAX) {
1513                r = -ENOBUFS;
1514                goto err;
1515        }
1516        vq = &n->vqs[index].vq;
1517        nvq = &n->vqs[index];
1518        mutex_lock(&vq->mutex);
1519
1520        /* Verify that ring has been setup correctly. */
1521        if (!vhost_vq_access_ok(vq)) {
1522                r = -EFAULT;
1523                goto err_vq;
1524        }
1525        sock = get_socket(fd);
1526        if (IS_ERR(sock)) {
1527                r = PTR_ERR(sock);
1528                goto err_vq;
1529        }
1530
1531        /* start polling new socket */
1532        oldsock = vq->private_data;
1533        if (sock != oldsock) {
1534                ubufs = vhost_net_ubuf_alloc(vq,
1535                                             sock && vhost_sock_zcopy(sock));
1536                if (IS_ERR(ubufs)) {
1537                        r = PTR_ERR(ubufs);
1538                        goto err_ubufs;
1539                }
1540
1541                vhost_net_disable_vq(n, vq);
1542                vq->private_data = sock;
1543                vhost_net_buf_unproduce(nvq);
1544                r = vhost_vq_init_access(vq);
1545                if (r)
1546                        goto err_used;
1547                r = vhost_net_enable_vq(n, vq);
1548                if (r)
1549                        goto err_used;
1550                if (index == VHOST_NET_VQ_RX)
1551                        nvq->rx_ring = get_tap_ptr_ring(fd);
1552
1553                oldubufs = nvq->ubufs;
1554                nvq->ubufs = ubufs;
1555
1556                n->tx_packets = 0;
1557                n->tx_zcopy_err = 0;
1558                n->tx_flush = false;
1559        }
1560
1561        mutex_unlock(&vq->mutex);
1562
1563        if (oldubufs) {
1564                vhost_net_ubuf_put_wait_and_free(oldubufs);
1565                mutex_lock(&vq->mutex);
1566                vhost_zerocopy_signal_used(n, vq);
1567                mutex_unlock(&vq->mutex);
1568        }
1569
1570        if (oldsock) {
1571                vhost_net_flush_vq(n, index);
1572                sockfd_put(oldsock);
1573        }
1574
1575        mutex_unlock(&n->dev.mutex);
1576        return 0;
1577
1578err_used:
1579        vq->private_data = oldsock;
1580        vhost_net_enable_vq(n, vq);
1581        if (ubufs)
1582                vhost_net_ubuf_put_wait_and_free(ubufs);
1583err_ubufs:
1584        if (sock)
1585                sockfd_put(sock);
1586err_vq:
1587        mutex_unlock(&vq->mutex);
1588err:
1589        mutex_unlock(&n->dev.mutex);
1590        return r;
1591}
1592
1593static long vhost_net_reset_owner(struct vhost_net *n)
1594{
1595        struct socket *tx_sock = NULL;
1596        struct socket *rx_sock = NULL;
1597        long err;
1598        struct vhost_umem *umem;
1599
1600        mutex_lock(&n->dev.mutex);
1601        err = vhost_dev_check_owner(&n->dev);
1602        if (err)
1603                goto done;
1604        umem = vhost_dev_reset_owner_prepare();
1605        if (!umem) {
1606                err = -ENOMEM;
1607                goto done;
1608        }
1609        vhost_net_stop(n, &tx_sock, &rx_sock);
1610        vhost_net_flush(n);
1611        vhost_dev_stop(&n->dev);
1612        vhost_dev_reset_owner(&n->dev, umem);
1613        vhost_net_vq_reset(n);
1614done:
1615        mutex_unlock(&n->dev.mutex);
1616        if (tx_sock)
1617                sockfd_put(tx_sock);
1618        if (rx_sock)
1619                sockfd_put(rx_sock);
1620        return err;
1621}
1622
1623static int vhost_net_set_backend_features(struct vhost_net *n, u64 features)
1624{
1625        int i;
1626
1627        mutex_lock(&n->dev.mutex);
1628        for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1629                mutex_lock(&n->vqs[i].vq.mutex);
1630                n->vqs[i].vq.acked_backend_features = features;
1631                mutex_unlock(&n->vqs[i].vq.mutex);
1632        }
1633        mutex_unlock(&n->dev.mutex);
1634
1635        return 0;
1636}
1637
1638static int vhost_net_set_features(struct vhost_net *n, u64 features)
1639{
1640        size_t vhost_hlen, sock_hlen, hdr_len;
1641        int i;
1642
1643        hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1644                               (1ULL << VIRTIO_F_VERSION_1))) ?
1645                        sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1646                        sizeof(struct virtio_net_hdr);
1647        if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1648                /* vhost provides vnet_hdr */
1649                vhost_hlen = hdr_len;
1650                sock_hlen = 0;
1651        } else {
1652                /* socket provides vnet_hdr */
1653                vhost_hlen = 0;
1654                sock_hlen = hdr_len;
1655        }
1656        mutex_lock(&n->dev.mutex);
1657        if ((features & (1 << VHOST_F_LOG_ALL)) &&
1658            !vhost_log_access_ok(&n->dev))
1659                goto out_unlock;
1660
1661        if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1662                if (vhost_init_device_iotlb(&n->dev, true))
1663                        goto out_unlock;
1664        }
1665
1666        for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1667                mutex_lock(&n->vqs[i].vq.mutex);
1668                n->vqs[i].vq.acked_features = features;
1669                n->vqs[i].vhost_hlen = vhost_hlen;
1670                n->vqs[i].sock_hlen = sock_hlen;
1671                mutex_unlock(&n->vqs[i].vq.mutex);
1672        }
1673        mutex_unlock(&n->dev.mutex);
1674        return 0;
1675
1676out_unlock:
1677        mutex_unlock(&n->dev.mutex);
1678        return -EFAULT;
1679}
1680
1681static long vhost_net_set_owner(struct vhost_net *n)
1682{
1683        int r;
1684
1685        mutex_lock(&n->dev.mutex);
1686        if (vhost_dev_has_owner(&n->dev)) {
1687                r = -EBUSY;
1688                goto out;
1689        }
1690        r = vhost_net_set_ubuf_info(n);
1691        if (r)
1692                goto out;
1693        r = vhost_dev_set_owner(&n->dev);
1694        if (r)
1695                vhost_net_clear_ubuf_info(n);
1696        vhost_net_flush(n);
1697out:
1698        mutex_unlock(&n->dev.mutex);
1699        return r;
1700}
1701
1702static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1703                            unsigned long arg)
1704{
1705        struct vhost_net *n = f->private_data;
1706        void __user *argp = (void __user *)arg;
1707        u64 __user *featurep = argp;
1708        struct vhost_vring_file backend;
1709        u64 features;
1710        int r;
1711
1712        switch (ioctl) {
1713        case VHOST_NET_SET_BACKEND:
1714                if (copy_from_user(&backend, argp, sizeof backend))
1715                        return -EFAULT;
1716                return vhost_net_set_backend(n, backend.index, backend.fd);
1717        case VHOST_GET_FEATURES:
1718                features = VHOST_NET_FEATURES;
1719                if (copy_to_user(featurep, &features, sizeof features))
1720                        return -EFAULT;
1721                return 0;
1722        case VHOST_SET_FEATURES:
1723                if (copy_from_user(&features, featurep, sizeof features))
1724                        return -EFAULT;
1725                if (features & ~VHOST_NET_FEATURES)
1726                        return -EOPNOTSUPP;
1727                return vhost_net_set_features(n, features);
1728        case VHOST_GET_BACKEND_FEATURES:
1729                features = VHOST_NET_BACKEND_FEATURES;
1730                if (copy_to_user(featurep, &features, sizeof(features)))
1731                        return -EFAULT;
1732                return 0;
1733        case VHOST_SET_BACKEND_FEATURES:
1734                if (copy_from_user(&features, featurep, sizeof(features)))
1735                        return -EFAULT;
1736                if (features & ~VHOST_NET_BACKEND_FEATURES)
1737                        return -EOPNOTSUPP;
1738                return vhost_net_set_backend_features(n, features);
1739        case VHOST_RESET_OWNER:
1740                return vhost_net_reset_owner(n);
1741        case VHOST_SET_OWNER:
1742                return vhost_net_set_owner(n);
1743        default:
1744                mutex_lock(&n->dev.mutex);
1745                r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1746                if (r == -ENOIOCTLCMD)
1747                        r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1748                else
1749                        vhost_net_flush(n);
1750                mutex_unlock(&n->dev.mutex);
1751                return r;
1752        }
1753}
1754
1755#ifdef CONFIG_COMPAT
1756static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1757                                   unsigned long arg)
1758{
1759        return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1760}
1761#endif
1762
1763static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1764{
1765        struct file *file = iocb->ki_filp;
1766        struct vhost_net *n = file->private_data;
1767        struct vhost_dev *dev = &n->dev;
1768        int noblock = file->f_flags & O_NONBLOCK;
1769
1770        return vhost_chr_read_iter(dev, to, noblock);
1771}
1772
1773static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1774                                        struct iov_iter *from)
1775{
1776        struct file *file = iocb->ki_filp;
1777        struct vhost_net *n = file->private_data;
1778        struct vhost_dev *dev = &n->dev;
1779
1780        return vhost_chr_write_iter(dev, from);
1781}
1782
1783static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1784{
1785        struct vhost_net *n = file->private_data;
1786        struct vhost_dev *dev = &n->dev;
1787
1788        return vhost_chr_poll(file, dev, wait);
1789}
1790
1791static const struct file_operations vhost_net_fops = {
1792        .owner          = THIS_MODULE,
1793        .release        = vhost_net_release,
1794        .read_iter      = vhost_net_chr_read_iter,
1795        .write_iter     = vhost_net_chr_write_iter,
1796        .poll           = vhost_net_chr_poll,
1797        .unlocked_ioctl = vhost_net_ioctl,
1798#ifdef CONFIG_COMPAT
1799        .compat_ioctl   = vhost_net_compat_ioctl,
1800#endif
1801        .open           = vhost_net_open,
1802        .llseek         = noop_llseek,
1803};
1804
1805static struct miscdevice vhost_net_misc = {
1806        .minor = VHOST_NET_MINOR,
1807        .name = "vhost-net",
1808        .fops = &vhost_net_fops,
1809};
1810
1811static int vhost_net_init(void)
1812{
1813        if (experimental_zcopytx)
1814                vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1815        return misc_register(&vhost_net_misc);
1816}
1817module_init(vhost_net_init);
1818
1819static void vhost_net_exit(void)
1820{
1821        misc_deregister(&vhost_net_misc);
1822}
1823module_exit(vhost_net_exit);
1824
1825MODULE_VERSION("0.0.1");
1826MODULE_LICENSE("GPL v2");
1827MODULE_AUTHOR("Michael S. Tsirkin");
1828MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1829MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1830MODULE_ALIAS("devname:vhost-net");
1831