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_IOMMU_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 ubuf_info *ubuf, bool success)
 385 {
 386         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 387         struct vhost_virtqueue *vq = ubufs->vq;
 388         int cnt;
 389
 390         rcu_read_lock_bh();
 391
 392         /* set len to mark this desc buffers done DMA */
 393         vq->heads[ubuf->desc].len = success ?
 394                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 395         cnt = vhost_net_ubuf_put(ubufs);
 396
 397         /*
 398          * Trigger polling thread if guest stopped submitting new buffers:
 399          * in this case, the refcount after decrement will eventually reach 1.
 400          * We also trigger polling periodically after each 16 packets
 401          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 402          * less than 10% of times).
 403          */
 404         if (cnt <= 1 || !(cnt % 16))
 405                 vhost_poll_queue(&vq->poll);
 406
 407         rcu_read_unlock_bh();
 408 }
 409
 410 static inline unsigned long busy_clock(void)
 411 {
 412         return local_clock() >> 10;
 413 }
 414
 415 static bool vhost_can_busy_poll(unsigned long endtime)
 416 {
 417         return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
 418                       !signal_pending(current));
 419 }
 420
 421 static void vhost_net_disable_vq(struct vhost_net *n,
 422                                  struct vhost_virtqueue *vq)
 423 {
 424         struct vhost_net_virtqueue *nvq =
 425                 container_of(vq, struct vhost_net_virtqueue, vq);
 426         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 427         if (!vhost_vq_get_backend(vq))
 428                 return;
 429         vhost_poll_stop(poll);
 430 }
 431
 432 static int vhost_net_enable_vq(struct vhost_net *n,
 433                                 struct vhost_virtqueue *vq)
 434 {
 435         struct vhost_net_virtqueue *nvq =
 436                 container_of(vq, struct vhost_net_virtqueue, vq);
 437         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 438         struct socket *sock;
 439
 440         sock = vhost_vq_get_backend(vq);
 441         if (!sock)
 442                 return 0;
 443
 444         return vhost_poll_start(poll, sock->file);
 445 }
 446
 447 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
 448 {
 449         struct vhost_virtqueue *vq = &nvq->vq;
 450         struct vhost_dev *dev = vq->dev;
 451
 452         if (!nvq->done_idx)
 453                 return;
 454
 455         vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
 456         nvq->done_idx = 0;
 457 }
 458
 459 static void vhost_tx_batch(struct vhost_net *net,
 460                            struct vhost_net_virtqueue *nvq,
 461                            struct socket *sock,
 462                            struct msghdr *msghdr)
 463 {
 464         struct tun_msg_ctl ctl = {
 465                 .type = TUN_MSG_PTR,
 466                 .num = nvq->batched_xdp,
 467                 .ptr = nvq->xdp,
 468         };
 469         int err;
 470
 471         if (nvq->batched_xdp == 0)
 472                 goto signal_used;
 473
 474         msghdr->msg_control = &ctl;
 475         err = sock->ops->sendmsg(sock, msghdr, 0);
 476         if (unlikely(err < 0)) {
 477                 vq_err(&nvq->vq, "Fail to batch sending packets\n");
 478                 return;
 479         }
 480
 481 signal_used:
 482         vhost_net_signal_used(nvq);
 483         nvq->batched_xdp = 0;
 484 }
 485
 486 static int sock_has_rx_data(struct socket *sock)
 487 {
 488         if (unlikely(!sock))
 489                 return 0;
 490
 491         if (sock->ops->peek_len)
 492                 return sock->ops->peek_len(sock);
 493
 494         return skb_queue_empty(&sock->sk->sk_receive_queue);
 495 }
 496
 497 static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
 498                                           struct vhost_virtqueue *vq)
 499 {
 500         if (!vhost_vq_avail_empty(&net->dev, vq)) {
 501                 vhost_poll_queue(&vq->poll);
 502         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 503                 vhost_disable_notify(&net->dev, vq);
 504                 vhost_poll_queue(&vq->poll);
 505         }
 506 }
 507
 508 static void vhost_net_busy_poll(struct vhost_net *net,
 509                                 struct vhost_virtqueue *rvq,
 510                                 struct vhost_virtqueue *tvq,
 511                                 bool *busyloop_intr,
 512                                 bool poll_rx)
 513 {
 514         unsigned long busyloop_timeout;
 515         unsigned long endtime;
 516         struct socket *sock;
 517         struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
 518
 519         /* Try to hold the vq mutex of the paired virtqueue. We can't
 520          * use mutex_lock() here since we could not guarantee a
 521          * consistenet lock ordering.
 522          */
 523         if (!mutex_trylock(&vq->mutex))
 524                 return;
 525
 526         vhost_disable_notify(&net->dev, vq);
 527         sock = vhost_vq_get_backend(rvq);
 528
 529         busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
 530                                      tvq->busyloop_timeout;
 531
 532         preempt_disable();
 533         endtime = busy_clock() + busyloop_timeout;
 534
 535         while (vhost_can_busy_poll(endtime)) {
 536                 if (vhost_has_work(&net->dev)) {
 537                         *busyloop_intr = true;
 538                         break;
 539                 }
 540
 541                 if ((sock_has_rx_data(sock) &&
 542                      !vhost_vq_avail_empty(&net->dev, rvq)) ||
 543                     !vhost_vq_avail_empty(&net->dev, tvq))
 544                         break;
 545
 546                 cpu_relax();
 547         }
 548
 549         preempt_enable();
 550
 551         if (poll_rx || sock_has_rx_data(sock))
 552                 vhost_net_busy_poll_try_queue(net, vq);
 553         else if (!poll_rx) /* On tx here, sock has no rx data. */
 554                 vhost_enable_notify(&net->dev, rvq);
 555
 556         mutex_unlock(&vq->mutex);
 557 }
 558
 559 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 560                                     struct vhost_net_virtqueue *tnvq,
 561                                     unsigned int *out_num, unsigned int *in_num,
 562                                     struct msghdr *msghdr, bool *busyloop_intr)
 563 {
 564         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 565         struct vhost_virtqueue *rvq = &rnvq->vq;
 566         struct vhost_virtqueue *tvq = &tnvq->vq;
 567
 568         int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 569                                   out_num, in_num, NULL, NULL);
 570
 571         if (r == tvq->num && tvq->busyloop_timeout) {
 572                 /* Flush batched packets first */
 573                 if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq)))
 574                         vhost_tx_batch(net, tnvq,
 575                                        vhost_vq_get_backend(tvq),
 576                                        msghdr);
 577
 578                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
 579
 580                 r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 581                                       out_num, in_num, NULL, NULL);
 582         }
 583
 584         return r;
 585 }
 586
 587 static bool vhost_exceeds_maxpend(struct vhost_net *net)
 588 {
 589         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 590         struct vhost_virtqueue *vq = &nvq->vq;
 591
 592         return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 593                min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 594 }
 595
 596 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
 597                             size_t hdr_size, int out)
 598 {
 599         /* Skip header. TODO: support TSO. */
 600         size_t len = iov_length(vq->iov, out);
 601
 602         iov_iter_init(iter, WRITE, vq->iov, out, len);
 603         iov_iter_advance(iter, hdr_size);
 604
 605         return iov_iter_count(iter);
 606 }
 607
 608 static int get_tx_bufs(struct vhost_net *net,
 609                        struct vhost_net_virtqueue *nvq,
 610                        struct msghdr *msg,
 611                        unsigned int *out, unsigned int *in,
 612                        size_t *len, bool *busyloop_intr)
 613 {
 614         struct vhost_virtqueue *vq = &nvq->vq;
 615         int ret;
 616
 617         ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
 618
 619         if (ret < 0 || ret == vq->num)
 620                 return ret;
 621
 622         if (*in) {
 623                 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 624                         *out, *in);
 625                 return -EFAULT;
 626         }
 627
 628         /* Sanity check */
 629         *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 630         if (*len == 0) {
 631                 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 632                         *len, nvq->vhost_hlen);
 633                 return -EFAULT;
 634         }
 635
 636         return ret;
 637 }
 638
 639 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 640 {
 641         return total_len < VHOST_NET_WEIGHT &&
 642                !vhost_vq_avail_empty(vq->dev, vq);
 643 }
 644
 645 #define SKB_FRAG_PAGE_ORDER     get_order(32768)
 646
 647 static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
 648                                        struct page_frag *pfrag, gfp_t gfp)
 649 {
 650         if (pfrag->page) {
 651                 if (pfrag->offset + sz <= pfrag->size)
 652                         return true;
 653                 __page_frag_cache_drain(pfrag->page, net->refcnt_bias);
 654         }
 655
 656         pfrag->offset = 0;
 657         net->refcnt_bias = 0;
 658         if (SKB_FRAG_PAGE_ORDER) {
 659                 /* Avoid direct reclaim but allow kswapd to wake */
 660                 pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
 661                                           __GFP_COMP | __GFP_NOWARN |
 662                                           __GFP_NORETRY,
 663                                           SKB_FRAG_PAGE_ORDER);
 664                 if (likely(pfrag->page)) {
 665                         pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
 666                         goto done;
 667                 }
 668         }
 669         pfrag->page = alloc_page(gfp);
 670         if (likely(pfrag->page)) {
 671                 pfrag->size = PAGE_SIZE;
 672                 goto done;
 673         }
 674         return false;
 675
 676 done:
 677         net->refcnt_bias = USHRT_MAX;
 678         page_ref_add(pfrag->page, USHRT_MAX - 1);
 679         return true;
 680 }
 681
 682 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 683
 684 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
 685                                struct iov_iter *from)
 686 {
 687         struct vhost_virtqueue *vq = &nvq->vq;
 688         struct vhost_net *net = container_of(vq->dev, struct vhost_net,
 689                                              dev);
 690         struct socket *sock = vhost_vq_get_backend(vq);
 691         struct page_frag *alloc_frag = &net->page_frag;
 692         struct virtio_net_hdr *gso;
 693         struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
 694         struct tun_xdp_hdr *hdr;
 695         size_t len = iov_iter_count(from);
 696         int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
 697         int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 698         int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
 699         int sock_hlen = nvq->sock_hlen;
 700         void *buf;
 701         int copied;
 702
 703         if (unlikely(len < nvq->sock_hlen))
 704                 return -EFAULT;
 705
 706         if (SKB_DATA_ALIGN(len + pad) +
 707             SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
 708                 return -ENOSPC;
 709
 710         buflen += SKB_DATA_ALIGN(len + pad);
 711         alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
 712         if (unlikely(!vhost_net_page_frag_refill(net, buflen,
 713                                                  alloc_frag, GFP_KERNEL)))
 714                 return -ENOMEM;
 715
 716         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 717         copied = copy_page_from_iter(alloc_frag->page,
 718                                      alloc_frag->offset +
 719                                      offsetof(struct tun_xdp_hdr, gso),
 720                                      sock_hlen, from);
 721         if (copied != sock_hlen)
 722                 return -EFAULT;
 723
 724         hdr = buf;
 725         gso = &hdr->gso;
 726
 727         if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
 728             vhost16_to_cpu(vq, gso->csum_start) +
 729             vhost16_to_cpu(vq, gso->csum_offset) + 2 >
 730             vhost16_to_cpu(vq, gso->hdr_len)) {
 731                 gso->hdr_len = cpu_to_vhost16(vq,
 732                                vhost16_to_cpu(vq, gso->csum_start) +
 733                                vhost16_to_cpu(vq, gso->csum_offset) + 2);
 734
 735                 if (vhost16_to_cpu(vq, gso->hdr_len) > len)
 736                         return -EINVAL;
 737         }
 738
 739         len -= sock_hlen;
 740         copied = copy_page_from_iter(alloc_frag->page,
 741                                      alloc_frag->offset + pad,
 742                                      len, from);
 743         if (copied != len)
 744                 return -EFAULT;
 745
 746         xdp->data_hard_start = buf;
 747         xdp->data = buf + pad;
 748         xdp->data_end = xdp->data + len;
 749         hdr->buflen = buflen;
 750
 751         --net->refcnt_bias;
 752         alloc_frag->offset += buflen;
 753
 754         ++nvq->batched_xdp;
 755
 756         return 0;
 757 }
 758
 759 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 760 {
 761         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 762         struct vhost_virtqueue *vq = &nvq->vq;
 763         unsigned out, in;
 764         int head;
 765         struct msghdr msg = {
 766                 .msg_name = NULL,
 767                 .msg_namelen = 0,
 768                 .msg_control = NULL,
 769                 .msg_controllen = 0,
 770                 .msg_flags = MSG_DONTWAIT,
 771         };
 772         size_t len, total_len = 0;
 773         int err;
 774         int sent_pkts = 0;
 775         bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
 776
 777         do {
 778                 bool busyloop_intr = false;
 779
 780                 if (nvq->done_idx == VHOST_NET_BATCH)
 781                         vhost_tx_batch(net, nvq, sock, &msg);
 782
 783                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 784                                    &busyloop_intr);
 785                 /* On error, stop handling until the next kick. */
 786                 if (unlikely(head < 0))
 787                         break;
 788                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 789                 if (head == vq->num) {
 790                         if (unlikely(busyloop_intr)) {
 791                                 vhost_poll_queue(&vq->poll);
 792                         } else if (unlikely(vhost_enable_notify(&net->dev,
 793                                                                 vq))) {
 794                                 vhost_disable_notify(&net->dev, vq);
 795                                 continue;
 796                         }
 797                         break;
 798                 }
 799
 800                 total_len += len;
 801
 802                 /* For simplicity, TX batching is only enabled if
 803                  * sndbuf is unlimited.
 804                  */
 805                 if (sock_can_batch) {
 806                         err = vhost_net_build_xdp(nvq, &msg.msg_iter);
 807                         if (!err) {
 808                                 goto done;
 809                         } else if (unlikely(err != -ENOSPC)) {
 810                                 vhost_tx_batch(net, nvq, sock, &msg);
 811                                 vhost_discard_vq_desc(vq, 1);
 812                                 vhost_net_enable_vq(net, vq);
 813                                 break;
 814                         }
 815
 816                         /* We can't build XDP buff, go for single
 817                          * packet path but let's flush batched
 818                          * packets.
 819                          */
 820                         vhost_tx_batch(net, nvq, sock, &msg);
 821                         msg.msg_control = NULL;
 822                 } else {
 823                         if (tx_can_batch(vq, total_len))
 824                                 msg.msg_flags |= MSG_MORE;
 825                         else
 826                                 msg.msg_flags &= ~MSG_MORE;
 827                 }
 828
 829                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 830                 err = sock->ops->sendmsg(sock, &msg, len);
 831                 if (unlikely(err < 0)) {
 832                         vhost_discard_vq_desc(vq, 1);
 833                         vhost_net_enable_vq(net, vq);
 834                         break;
 835                 }
 836                 if (err != len)
 837                         pr_debug("Truncated TX packet: len %d != %zd\n",
 838                                  err, len);
 839 done:
 840                 vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 841                 vq->heads[nvq->done_idx].len = 0;
 842                 ++nvq->done_idx;
 843         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 844
 845         vhost_tx_batch(net, nvq, sock, &msg);
 846 }
 847
 848 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 849 {
 850         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 851         struct vhost_virtqueue *vq = &nvq->vq;
 852         unsigned out, in;
 853         int head;
 854         struct msghdr msg = {
 855                 .msg_name = NULL,
 856                 .msg_namelen = 0,
 857                 .msg_control = NULL,
 858                 .msg_controllen = 0,
 859                 .msg_flags = MSG_DONTWAIT,
 860         };
 861         struct tun_msg_ctl ctl;
 862         size_t len, total_len = 0;
 863         int err;
 864         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 865         bool zcopy_used;
 866         int sent_pkts = 0;
 867
 868         do {
 869                 bool busyloop_intr;
 870
 871                 /* Release DMAs done buffers first */
 872                 vhost_zerocopy_signal_used(net, vq);
 873
 874                 busyloop_intr = false;
 875                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 876                                    &busyloop_intr);
 877                 /* On error, stop handling until the next kick. */
 878                 if (unlikely(head < 0))
 879                         break;
 880                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 881                 if (head == vq->num) {
 882                         if (unlikely(busyloop_intr)) {
 883                                 vhost_poll_queue(&vq->poll);
 884                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 885                                 vhost_disable_notify(&net->dev, vq);
 886                                 continue;
 887                         }
 888                         break;
 889                 }
 890
 891                 zcopy_used = len >= VHOST_GOODCOPY_LEN
 892                              && !vhost_exceeds_maxpend(net)
 893                              && vhost_net_tx_select_zcopy(net);
 894
 895                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 896                 if (zcopy_used) {
 897                         struct ubuf_info *ubuf;
 898                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 899
 900                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 901                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 902                         ubuf->callback = vhost_zerocopy_callback;
 903                         ubuf->ctx = nvq->ubufs;
 904                         ubuf->desc = nvq->upend_idx;
 905                         refcount_set(&ubuf->refcnt, 1);
 906                         msg.msg_control = &ctl;
 907                         ctl.type = TUN_MSG_UBUF;
 908                         ctl.ptr = ubuf;
 909                         msg.msg_controllen = sizeof(ctl);
 910                         ubufs = nvq->ubufs;
 911                         atomic_inc(&ubufs->refcount);
 912                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 913                 } else {
 914                         msg.msg_control = NULL;
 915                         ubufs = NULL;
 916                 }
 917                 total_len += len;
 918                 if (tx_can_batch(vq, total_len) &&
 919                     likely(!vhost_exceeds_maxpend(net))) {
 920                         msg.msg_flags |= MSG_MORE;
 921                 } else {
 922                         msg.msg_flags &= ~MSG_MORE;
 923                 }
 924
 925                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 926                 err = sock->ops->sendmsg(sock, &msg, len);
 927                 if (unlikely(err < 0)) {
 928                         if (zcopy_used) {
 929                                 vhost_net_ubuf_put(ubufs);
 930                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 931                                         % UIO_MAXIOV;
 932                         }
 933                         vhost_discard_vq_desc(vq, 1);
 934                         vhost_net_enable_vq(net, vq);
 935                         break;
 936                 }
 937                 if (err != len)
 938                         pr_debug("Truncated TX packet: "
 939                                  " len %d != %zd\n", err, len);
 940                 if (!zcopy_used)
 941                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 942                 else
 943                         vhost_zerocopy_signal_used(net, vq);
 944                 vhost_net_tx_packet(net);
 945         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 946 }
 947
 948 /* Expects to be always run from workqueue - which acts as
 949  * read-size critical section for our kind of RCU. */
 950 static void handle_tx(struct vhost_net *net)
 951 {
 952         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 953         struct vhost_virtqueue *vq = &nvq->vq;
 954         struct socket *sock;
 955
 956         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
 957         sock = vhost_vq_get_backend(vq);
 958         if (!sock)
 959                 goto out;
 960
 961         if (!vq_meta_prefetch(vq))
 962                 goto out;
 963
 964         vhost_disable_notify(&net->dev, vq);
 965         vhost_net_disable_vq(net, vq);
 966
 967         if (vhost_sock_zcopy(sock))
 968                 handle_tx_zerocopy(net, sock);
 969         else
 970                 handle_tx_copy(net, sock);
 971
 972 out:
 973         mutex_unlock(&vq->mutex);
 974 }
 975
 976 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 977 {
 978         struct sk_buff *head;
 979         int len = 0;
 980         unsigned long flags;
 981
 982         if (rvq->rx_ring)
 983                 return vhost_net_buf_peek(rvq);
 984
 985         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 986         head = skb_peek(&sk->sk_receive_queue);
 987         if (likely(head)) {
 988                 len = head->len;
 989                 if (skb_vlan_tag_present(head))
 990                         len += VLAN_HLEN;
 991         }
 992
 993         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 994         return len;
 995 }
 996
 997 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
 998                                       bool *busyloop_intr)
 999 {
1000         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1001         struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1002         struct vhost_virtqueue *rvq = &rnvq->vq;
1003         struct vhost_virtqueue *tvq = &tnvq->vq;
1004         int len = peek_head_len(rnvq, sk);
1005
1006         if (!len && rvq->busyloop_timeout) {
1007                 /* Flush batched heads first */
1008                 vhost_net_signal_used(rnvq);
1009                 /* Both tx vq and rx socket were polled here */
1010                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1011
1012                 len = peek_head_len(rnvq, sk);
1013         }
1014
1015         return len;
1016 }
1017
1018 /* This is a multi-buffer version of vhost_get_desc, that works if
1019  *      vq has read descriptors only.
1020  * @vq          - the relevant virtqueue
1021  * @datalen     - data length we'll be reading
1022  * @iovcount    - returned count of io vectors we fill
1023  * @log         - vhost log
1024  * @log_num     - log offset
1025  * @quota       - headcount quota, 1 for big buffer
1026  *      returns number of buffer heads allocated, negative on error
1027  */
1028 static int get_rx_bufs(struct vhost_virtqueue *vq,
1029                        struct vring_used_elem *heads,
1030                        int datalen,
1031                        unsigned *iovcount,
1032                        struct vhost_log *log,
1033                        unsigned *log_num,
1034                        unsigned int quota)
1035 {
1036         unsigned int out, in;
1037         int seg = 0;
1038         int headcount = 0;
1039         unsigned d;
1040         int r, nlogs = 0;
1041         /* len is always initialized before use since we are always called with
1042          * datalen > 0.
1043          */
1044         u32 uninitialized_var(len);
1045
1046         while (datalen > 0 && headcount < quota) {
1047                 if (unlikely(seg >= UIO_MAXIOV)) {
1048                         r = -ENOBUFS;
1049                         goto err;
1050                 }
1051                 r = vhost_get_vq_desc(vq, vq->iov + seg,
1052                                       ARRAY_SIZE(vq->iov) - seg, &out,
1053                                       &in, log, log_num);
1054                 if (unlikely(r < 0))
1055                         goto err;
1056
1057                 d = r;
1058                 if (d == vq->num) {
1059                         r = 0;
1060                         goto err;
1061                 }
1062                 if (unlikely(out || in <= 0)) {
1063                         vq_err(vq, "unexpected descriptor format for RX: "
1064                                 "out %d, in %d\n", out, in);
1065                         r = -EINVAL;
1066                         goto err;
1067                 }
1068                 if (unlikely(log)) {
1069                         nlogs += *log_num;
1070                         log += *log_num;
1071                 }
1072                 heads[headcount].id = cpu_to_vhost32(vq, d);
1073                 len = iov_length(vq->iov + seg, in);
1074                 heads[headcount].len = cpu_to_vhost32(vq, len);
1075                 datalen -= len;
1076                 ++headcount;
1077                 seg += in;
1078         }
1079         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1080         *iovcount = seg;
1081         if (unlikely(log))
1082                 *log_num = nlogs;
1083
1084         /* Detect overrun */
1085         if (unlikely(datalen > 0)) {
1086                 r = UIO_MAXIOV + 1;
1087                 goto err;
1088         }
1089         return headcount;
1090 err:
1091         vhost_discard_vq_desc(vq, headcount);
1092         return r;
1093 }
1094
1095 /* Expects to be always run from workqueue - which acts as
1096  * read-size critical section for our kind of RCU. */
1097 static void handle_rx(struct vhost_net *net)
1098 {
1099         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1100         struct vhost_virtqueue *vq = &nvq->vq;
1101         unsigned uninitialized_var(in), log;
1102         struct vhost_log *vq_log;
1103         struct msghdr msg = {
1104                 .msg_name = NULL,
1105                 .msg_namelen = 0,
1106                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
1107                 .msg_controllen = 0,
1108                 .msg_flags = MSG_DONTWAIT,
1109         };
1110         struct virtio_net_hdr hdr = {
1111                 .flags = 0,
1112                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
1113         };
1114         size_t total_len = 0;
1115         int err, mergeable;
1116         s16 headcount;
1117         size_t vhost_hlen, sock_hlen;
1118         size_t vhost_len, sock_len;
1119         bool busyloop_intr = false;
1120         struct socket *sock;
1121         struct iov_iter fixup;
1122         __virtio16 num_buffers;
1123         int recv_pkts = 0;
1124
1125         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1126         sock = vhost_vq_get_backend(vq);
1127         if (!sock)
1128                 goto out;
1129
1130         if (!vq_meta_prefetch(vq))
1131                 goto out;
1132
1133         vhost_disable_notify(&net->dev, vq);
1134         vhost_net_disable_vq(net, vq);
1135
1136         vhost_hlen = nvq->vhost_hlen;
1137         sock_hlen = nvq->sock_hlen;
1138
1139         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1140                 vq->log : NULL;
1141         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1142
1143         do {
1144                 sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1145                                                       &busyloop_intr);
1146                 if (!sock_len)
1147                         break;
1148                 sock_len += sock_hlen;
1149                 vhost_len = sock_len + vhost_hlen;
1150                 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1151                                         vhost_len, &in, vq_log, &log,
1152                                         likely(mergeable) ? UIO_MAXIOV : 1);
1153                 /* On error, stop handling until the next kick. */
1154                 if (unlikely(headcount < 0))
1155                         goto out;
1156                 /* OK, now we need to know about added descriptors. */
1157                 if (!headcount) {
1158                         if (unlikely(busyloop_intr)) {
1159                                 vhost_poll_queue(&vq->poll);
1160                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1161                                 /* They have slipped one in as we were
1162                                  * doing that: check again. */
1163                                 vhost_disable_notify(&net->dev, vq);
1164                                 continue;
1165                         }
1166                         /* Nothing new?  Wait for eventfd to tell us
1167                          * they refilled. */
1168                         goto out;
1169                 }
1170                 busyloop_intr = false;
1171                 if (nvq->rx_ring)
1172                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1173                 /* On overrun, truncate and discard */
1174                 if (unlikely(headcount > UIO_MAXIOV)) {
1175                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
1176                         err = sock->ops->recvmsg(sock, &msg,
1177                                                  1, MSG_DONTWAIT | MSG_TRUNC);
1178                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
1179                         continue;
1180                 }
1181                 /* We don't need to be notified again. */
1182                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
1183                 fixup = msg.msg_iter;
1184                 if (unlikely((vhost_hlen))) {
1185                         /* We will supply the header ourselves
1186                          * TODO: support TSO.
1187                          */
1188                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
1189                 }
1190                 err = sock->ops->recvmsg(sock, &msg,
1191                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
1192                 /* Userspace might have consumed the packet meanwhile:
1193                  * it's not supposed to do this usually, but might be hard
1194                  * to prevent. Discard data we got (if any) and keep going. */
1195                 if (unlikely(err != sock_len)) {
1196                         pr_debug("Discarded rx packet: "
1197                                  " len %d, expected %zd\n", err, sock_len);
1198                         vhost_discard_vq_desc(vq, headcount);
1199                         continue;
1200                 }
1201                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1202                 if (unlikely(vhost_hlen)) {
1203                         if (copy_to_iter(&hdr, sizeof(hdr),
1204                                          &fixup) != sizeof(hdr)) {
1205                                 vq_err(vq, "Unable to write vnet_hdr "
1206                                        "at addr %p\n", vq->iov->iov_base);
1207                                 goto out;
1208                         }
1209                 } else {
1210                         /* Header came from socket; we'll need to patch
1211                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1212                          */
1213                         iov_iter_advance(&fixup, sizeof(hdr));
1214                 }
1215                 /* TODO: Should check and handle checksum. */
1216
1217                 num_buffers = cpu_to_vhost16(vq, headcount);
1218                 if (likely(mergeable) &&
1219                     copy_to_iter(&num_buffers, sizeof num_buffers,
1220                                  &fixup) != sizeof num_buffers) {
1221                         vq_err(vq, "Failed num_buffers write");
1222                         vhost_discard_vq_desc(vq, headcount);
1223                         goto out;
1224                 }
1225                 nvq->done_idx += headcount;
1226                 if (nvq->done_idx > VHOST_NET_BATCH)
1227                         vhost_net_signal_used(nvq);
1228                 if (unlikely(vq_log))
1229                         vhost_log_write(vq, vq_log, log, vhost_len,
1230                                         vq->iov, in);
1231                 total_len += vhost_len;
1232         } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1233
1234         if (unlikely(busyloop_intr))
1235                 vhost_poll_queue(&vq->poll);
1236         else if (!sock_len)
1237                 vhost_net_enable_vq(net, vq);
1238 out:
1239         vhost_net_signal_used(nvq);
1240         mutex_unlock(&vq->mutex);
1241 }
1242
1243 static void handle_tx_kick(struct vhost_work *work)
1244 {
1245         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1246                                                   poll.work);
1247         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1248
1249         handle_tx(net);
1250 }
1251
1252 static void handle_rx_kick(struct vhost_work *work)
1253 {
1254         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1255                                                   poll.work);
1256         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1257
1258         handle_rx(net);
1259 }
1260
1261 static void handle_tx_net(struct vhost_work *work)
1262 {
1263         struct vhost_net *net = container_of(work, struct vhost_net,
1264                                              poll[VHOST_NET_VQ_TX].work);
1265         handle_tx(net);
1266 }
1267
1268 static void handle_rx_net(struct vhost_work *work)
1269 {
1270         struct vhost_net *net = container_of(work, struct vhost_net,
1271                                              poll[VHOST_NET_VQ_RX].work);
1272         handle_rx(net);
1273 }
1274
1275 static int vhost_net_open(struct inode *inode, struct file *f)
1276 {
1277         struct vhost_net *n;
1278         struct vhost_dev *dev;
1279         struct vhost_virtqueue **vqs;
1280         void **queue;
1281         struct xdp_buff *xdp;
1282         int i;
1283
1284         n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1285         if (!n)
1286                 return -ENOMEM;
1287         vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1288         if (!vqs) {
1289                 kvfree(n);
1290                 return -ENOMEM;
1291         }
1292
1293         queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1294                               GFP_KERNEL);
1295         if (!queue) {
1296                 kfree(vqs);
1297                 kvfree(n);
1298                 return -ENOMEM;
1299         }
1300         n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1301
1302         xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1303         if (!xdp) {
1304                 kfree(vqs);
1305                 kvfree(n);
1306                 kfree(queue);
1307                 return -ENOMEM;
1308         }
1309         n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1310
1311         dev = &n->dev;
1312         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1313         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1314         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1315         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1316         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1317                 n->vqs[i].ubufs = NULL;
1318                 n->vqs[i].ubuf_info = NULL;
1319                 n->vqs[i].upend_idx = 0;
1320                 n->vqs[i].done_idx = 0;
1321                 n->vqs[i].batched_xdp = 0;
1322                 n->vqs[i].vhost_hlen = 0;
1323                 n->vqs[i].sock_hlen = 0;
1324                 n->vqs[i].rx_ring = NULL;
1325                 vhost_net_buf_init(&n->vqs[i].rxq);
1326         }
1327         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1328                        UIO_MAXIOV + VHOST_NET_BATCH,
1329                        VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT,
1330                        NULL);
1331
1332         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1333         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1334
1335         f->private_data = n;
1336         n->page_frag.page = NULL;
1337         n->refcnt_bias = 0;
1338
1339         return 0;
1340 }
1341
1342 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1343                                         struct vhost_virtqueue *vq)
1344 {
1345         struct socket *sock;
1346         struct vhost_net_virtqueue *nvq =
1347                 container_of(vq, struct vhost_net_virtqueue, vq);
1348
1349         mutex_lock(&vq->mutex);
1350         sock = vhost_vq_get_backend(vq);
1351         vhost_net_disable_vq(n, vq);
1352         vhost_vq_set_backend(vq, NULL);
1353         vhost_net_buf_unproduce(nvq);
1354         nvq->rx_ring = NULL;
1355         mutex_unlock(&vq->mutex);
1356         return sock;
1357 }
1358
1359 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1360                            struct socket **rx_sock)
1361 {
1362         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1363         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1364 }
1365
1366 static void vhost_net_flush_vq(struct vhost_net *n, int index)
1367 {
1368         vhost_poll_flush(n->poll + index);
1369         vhost_poll_flush(&n->vqs[index].vq.poll);
1370 }
1371
1372 static void vhost_net_flush(struct vhost_net *n)
1373 {
1374         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1375         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1376         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1377                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1378                 n->tx_flush = true;
1379                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1380                 /* Wait for all lower device DMAs done. */
1381                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1382                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1383                 n->tx_flush = false;
1384                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1385                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1386         }
1387 }
1388
1389 static int vhost_net_release(struct inode *inode, struct file *f)
1390 {
1391         struct vhost_net *n = f->private_data;
1392         struct socket *tx_sock;
1393         struct socket *rx_sock;
1394
1395         vhost_net_stop(n, &tx_sock, &rx_sock);
1396         vhost_net_flush(n);
1397         vhost_dev_stop(&n->dev);
1398         vhost_dev_cleanup(&n->dev);
1399         vhost_net_vq_reset(n);
1400         if (tx_sock)
1401                 sockfd_put(tx_sock);
1402         if (rx_sock)
1403                 sockfd_put(rx_sock);
1404         /* Make sure no callbacks are outstanding */
1405         synchronize_rcu();
1406         /* We do an extra flush before freeing memory,
1407          * since jobs can re-queue themselves. */
1408         vhost_net_flush(n);
1409         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1410         kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1411         kfree(n->dev.vqs);
1412         if (n->page_frag.page)
1413                 __page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
1414         kvfree(n);
1415         return 0;
1416 }
1417
1418 static struct socket *get_raw_socket(int fd)
1419 {
1420         int r;
1421         struct socket *sock = sockfd_lookup(fd, &r);
1422
1423         if (!sock)
1424                 return ERR_PTR(-ENOTSOCK);
1425
1426         /* Parameter checking */
1427         if (sock->sk->sk_type != SOCK_RAW) {
1428                 r = -ESOCKTNOSUPPORT;
1429                 goto err;
1430         }
1431
1432         if (sock->sk->sk_family != AF_PACKET) {
1433                 r = -EPFNOSUPPORT;
1434                 goto err;
1435         }
1436         return sock;
1437 err:
1438         sockfd_put(sock);
1439         return ERR_PTR(r);
1440 }
1441
1442 static struct ptr_ring *get_tap_ptr_ring(int fd)
1443 {
1444         struct ptr_ring *ring;
1445         struct file *file = fget(fd);
1446
1447         if (!file)
1448                 return NULL;
1449         ring = tun_get_tx_ring(file);
1450         if (!IS_ERR(ring))
1451                 goto out;
1452         ring = tap_get_ptr_ring(file);
1453         if (!IS_ERR(ring))
1454                 goto out;
1455         ring = NULL;
1456 out:
1457         fput(file);
1458         return ring;
1459 }
1460
1461 static struct socket *get_tap_socket(int fd)
1462 {
1463         struct file *file = fget(fd);
1464         struct socket *sock;
1465
1466         if (!file)
1467                 return ERR_PTR(-EBADF);
1468         sock = tun_get_socket(file);
1469         if (!IS_ERR(sock))
1470                 return sock;
1471         sock = tap_get_socket(file);
1472         if (IS_ERR(sock))
1473                 fput(file);
1474         return sock;
1475 }
1476
1477 static struct socket *get_socket(int fd)
1478 {
1479         struct socket *sock;
1480
1481         /* special case to disable backend */
1482         if (fd == -1)
1483                 return NULL;
1484         sock = get_raw_socket(fd);
1485         if (!IS_ERR(sock))
1486                 return sock;
1487         sock = get_tap_socket(fd);
1488         if (!IS_ERR(sock))
1489                 return sock;
1490         return ERR_PTR(-ENOTSOCK);
1491 }
1492
1493 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1494 {
1495         struct socket *sock, *oldsock;
1496         struct vhost_virtqueue *vq;
1497         struct vhost_net_virtqueue *nvq;
1498         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1499         int r;
1500
1501         mutex_lock(&n->dev.mutex);
1502         r = vhost_dev_check_owner(&n->dev);
1503         if (r)
1504                 goto err;
1505
1506         if (index >= VHOST_NET_VQ_MAX) {
1507                 r = -ENOBUFS;
1508                 goto err;
1509         }
1510         vq = &n->vqs[index].vq;
1511         nvq = &n->vqs[index];
1512         mutex_lock(&vq->mutex);
1513
1514         /* Verify that ring has been setup correctly. */
1515         if (!vhost_vq_access_ok(vq)) {
1516                 r = -EFAULT;
1517                 goto err_vq;
1518         }
1519         sock = get_socket(fd);
1520         if (IS_ERR(sock)) {
1521                 r = PTR_ERR(sock);
1522                 goto err_vq;
1523         }
1524
1525         /* start polling new socket */
1526         oldsock = vhost_vq_get_backend(vq);
1527         if (sock != oldsock) {
1528                 ubufs = vhost_net_ubuf_alloc(vq,
1529                                              sock && vhost_sock_zcopy(sock));
1530                 if (IS_ERR(ubufs)) {
1531                         r = PTR_ERR(ubufs);
1532                         goto err_ubufs;
1533                 }
1534
1535                 vhost_net_disable_vq(n, vq);
1536                 vhost_vq_set_backend(vq, sock);
1537                 vhost_net_buf_unproduce(nvq);
1538                 r = vhost_vq_init_access(vq);
1539                 if (r)
1540                         goto err_used;
1541                 r = vhost_net_enable_vq(n, vq);
1542                 if (r)
1543                         goto err_used;
1544                 if (index == VHOST_NET_VQ_RX)
1545                         nvq->rx_ring = get_tap_ptr_ring(fd);
1546
1547                 oldubufs = nvq->ubufs;
1548                 nvq->ubufs = ubufs;
1549
1550                 n->tx_packets = 0;
1551                 n->tx_zcopy_err = 0;
1552                 n->tx_flush = false;
1553         }
1554
1555         mutex_unlock(&vq->mutex);
1556
1557         if (oldubufs) {
1558                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1559                 mutex_lock(&vq->mutex);
1560                 vhost_zerocopy_signal_used(n, vq);
1561                 mutex_unlock(&vq->mutex);
1562         }
1563
1564         if (oldsock) {
1565                 vhost_net_flush_vq(n, index);
1566                 sockfd_put(oldsock);
1567         }
1568
1569         mutex_unlock(&n->dev.mutex);
1570         return 0;
1571
1572 err_used:
1573         vhost_vq_set_backend(vq, oldsock);
1574         vhost_net_enable_vq(n, vq);
1575         if (ubufs)
1576                 vhost_net_ubuf_put_wait_and_free(ubufs);
1577 err_ubufs:
1578         if (sock)
1579                 sockfd_put(sock);
1580 err_vq:
1581         mutex_unlock(&vq->mutex);
1582 err:
1583         mutex_unlock(&n->dev.mutex);
1584         return r;
1585 }
1586
1587 static long vhost_net_reset_owner(struct vhost_net *n)
1588 {
1589         struct socket *tx_sock = NULL;
1590         struct socket *rx_sock = NULL;
1591         long err;
1592         struct vhost_iotlb *umem;
1593
1594         mutex_lock(&n->dev.mutex);
1595         err = vhost_dev_check_owner(&n->dev);
1596         if (err)
1597                 goto done;
1598         umem = vhost_dev_reset_owner_prepare();
1599         if (!umem) {
1600                 err = -ENOMEM;
1601                 goto done;
1602         }
1603         vhost_net_stop(n, &tx_sock, &rx_sock);
1604         vhost_net_flush(n);
1605         vhost_dev_stop(&n->dev);
1606         vhost_dev_reset_owner(&n->dev, umem);
1607         vhost_net_vq_reset(n);
1608 done:
1609         mutex_unlock(&n->dev.mutex);
1610         if (tx_sock)
1611                 sockfd_put(tx_sock);
1612         if (rx_sock)
1613                 sockfd_put(rx_sock);
1614         return err;
1615 }
1616
1617 static int vhost_net_set_backend_features(struct vhost_net *n, u64 features)
1618 {
1619         int i;
1620
1621         mutex_lock(&n->dev.mutex);
1622         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1623                 mutex_lock(&n->vqs[i].vq.mutex);
1624                 n->vqs[i].vq.acked_backend_features = features;
1625                 mutex_unlock(&n->vqs[i].vq.mutex);
1626         }
1627         mutex_unlock(&n->dev.mutex);
1628
1629         return 0;
1630 }
1631
1632 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1633 {
1634         size_t vhost_hlen, sock_hlen, hdr_len;
1635         int i;
1636
1637         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1638                                (1ULL << VIRTIO_F_VERSION_1))) ?
1639                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1640                         sizeof(struct virtio_net_hdr);
1641         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1642                 /* vhost provides vnet_hdr */
1643                 vhost_hlen = hdr_len;
1644                 sock_hlen = 0;
1645         } else {
1646                 /* socket provides vnet_hdr */
1647                 vhost_hlen = 0;
1648                 sock_hlen = hdr_len;
1649         }
1650         mutex_lock(&n->dev.mutex);
1651         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1652             !vhost_log_access_ok(&n->dev))
1653                 goto out_unlock;
1654
1655         if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1656                 if (vhost_init_device_iotlb(&n->dev, true))
1657                         goto out_unlock;
1658         }
1659
1660         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1661                 mutex_lock(&n->vqs[i].vq.mutex);
1662                 n->vqs[i].vq.acked_features = features;
1663                 n->vqs[i].vhost_hlen = vhost_hlen;
1664                 n->vqs[i].sock_hlen = sock_hlen;
1665                 mutex_unlock(&n->vqs[i].vq.mutex);
1666         }
1667         mutex_unlock(&n->dev.mutex);
1668         return 0;
1669
1670 out_unlock:
1671         mutex_unlock(&n->dev.mutex);
1672         return -EFAULT;
1673 }
1674
1675 static long vhost_net_set_owner(struct vhost_net *n)
1676 {
1677         int r;
1678
1679         mutex_lock(&n->dev.mutex);
1680         if (vhost_dev_has_owner(&n->dev)) {
1681                 r = -EBUSY;
1682                 goto out;
1683         }
1684         r = vhost_net_set_ubuf_info(n);
1685         if (r)
1686                 goto out;
1687         r = vhost_dev_set_owner(&n->dev);
1688         if (r)
1689                 vhost_net_clear_ubuf_info(n);
1690         vhost_net_flush(n);
1691 out:
1692         mutex_unlock(&n->dev.mutex);
1693         return r;
1694 }
1695
1696 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1697                             unsigned long arg)
1698 {
1699         struct vhost_net *n = f->private_data;
1700         void __user *argp = (void __user *)arg;
1701         u64 __user *featurep = argp;
1702         struct vhost_vring_file backend;
1703         u64 features;
1704         int r;
1705
1706         switch (ioctl) {
1707         case VHOST_NET_SET_BACKEND:
1708                 if (copy_from_user(&backend, argp, sizeof backend))
1709                         return -EFAULT;
1710                 return vhost_net_set_backend(n, backend.index, backend.fd);
1711         case VHOST_GET_FEATURES:
1712                 features = VHOST_NET_FEATURES;
1713                 if (copy_to_user(featurep, &features, sizeof features))
1714                         return -EFAULT;
1715                 return 0;
1716         case VHOST_SET_FEATURES:
1717                 if (copy_from_user(&features, featurep, sizeof features))
1718                         return -EFAULT;
1719                 if (features & ~VHOST_NET_FEATURES)
1720                         return -EOPNOTSUPP;
1721                 return vhost_net_set_features(n, features);
1722         case VHOST_GET_BACKEND_FEATURES:
1723                 features = VHOST_NET_BACKEND_FEATURES;
1724                 if (copy_to_user(featurep, &features, sizeof(features)))
1725                         return -EFAULT;
1726                 return 0;
1727         case VHOST_SET_BACKEND_FEATURES:
1728                 if (copy_from_user(&features, featurep, sizeof(features)))
1729                         return -EFAULT;
1730                 if (features & ~VHOST_NET_BACKEND_FEATURES)
1731                         return -EOPNOTSUPP;
1732                 return vhost_net_set_backend_features(n, features);
1733         case VHOST_RESET_OWNER:
1734                 return vhost_net_reset_owner(n);
1735         case VHOST_SET_OWNER:
1736                 return vhost_net_set_owner(n);
1737         default:
1738                 mutex_lock(&n->dev.mutex);
1739                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1740                 if (r == -ENOIOCTLCMD)
1741                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1742                 else
1743                         vhost_net_flush(n);
1744                 mutex_unlock(&n->dev.mutex);
1745                 return r;
1746         }
1747 }
1748
1749 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1750 {
1751         struct file *file = iocb->ki_filp;
1752         struct vhost_net *n = file->private_data;
1753         struct vhost_dev *dev = &n->dev;
1754         int noblock = file->f_flags & O_NONBLOCK;
1755
1756         return vhost_chr_read_iter(dev, to, noblock);
1757 }
1758
1759 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1760                                         struct iov_iter *from)
1761 {
1762         struct file *file = iocb->ki_filp;
1763         struct vhost_net *n = file->private_data;
1764         struct vhost_dev *dev = &n->dev;
1765
1766         return vhost_chr_write_iter(dev, from);
1767 }
1768
1769 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1770 {
1771         struct vhost_net *n = file->private_data;
1772         struct vhost_dev *dev = &n->dev;
1773
1774         return vhost_chr_poll(file, dev, wait);
1775 }
1776
1777 static const struct file_operations vhost_net_fops = {
1778         .owner          = THIS_MODULE,
1779         .release        = vhost_net_release,
1780         .read_iter      = vhost_net_chr_read_iter,
1781         .write_iter     = vhost_net_chr_write_iter,
1782         .poll           = vhost_net_chr_poll,
1783         .unlocked_ioctl = vhost_net_ioctl,
1784         .compat_ioctl   = compat_ptr_ioctl,
1785         .open           = vhost_net_open,
1786         .llseek         = noop_llseek,
1787 };
1788
1789 static struct miscdevice vhost_net_misc = {
1790         .minor = VHOST_NET_MINOR,
1791         .name = "vhost-net",
1792         .fops = &vhost_net_fops,
1793 };
1794
1795 static int vhost_net_init(void)
1796 {
1797         if (experimental_zcopytx)
1798                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1799         return misc_register(&vhost_net_misc);
1800 }
1801 module_init(vhost_net_init);
1802
1803 static void vhost_net_exit(void)
1804 {
1805         misc_deregister(&vhost_net_misc);
1806 }
1807 module_exit(vhost_net_exit);
1808
1809 MODULE_VERSION("0.0.1");
1810 MODULE_LICENSE("GPL v2");
1811 MODULE_AUTHOR("Michael S. Tsirkin");
1812 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1813 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1814 MODULE_ALIAS("devname:vhost-net");