drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <mst@redhat.com>
   3  *
   4  * This work is licensed under the terms of the GNU GPL, version 2.
   5  *
   6  * virtio-net server in host kernel.
   7  */
   8
   9 #include <linux/compat.h>
  10 #include <linux/eventfd.h>
  11 #include <linux/vhost.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/module.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/mutex.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/sched/clock.h>
  21 #include <linux/sched/signal.h>
  22 #include <linux/vmalloc.h>
  23
  24 #include <linux/net.h>
  25 #include <linux/if_packet.h>
  26 #include <linux/if_arp.h>
  27 #include <linux/if_tun.h>
  28 #include <linux/if_macvlan.h>
  29 #include <linux/if_tap.h>
  30 #include <linux/if_vlan.h>
  31 #include <linux/skb_array.h>
  32 #include <linux/skbuff.h>
  33
  34 #include <net/sock.h>
  35 #include <net/xdp.h>
  36
  37 #include "vhost.h"
  38
  39 static int experimental_zcopytx = 1;
  40 module_param(experimental_zcopytx, int, 0444);
  41 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  42                                        " 1 -Enable; 0 - Disable");
  43
  44 /* Max number of bytes transferred before requeueing the job.
  45  * Using this limit prevents one virtqueue from starving others. */
  46 #define VHOST_NET_WEIGHT 0x80000
  47
  48 /* Max number of packets transferred before requeueing the job.
  49  * Using this limit prevents one virtqueue from starving others with small
  50  * pkts.
  51  */
  52 #define VHOST_NET_PKT_WEIGHT 256
  53
  54 /* MAX number of TX used buffers for outstanding zerocopy */
  55 #define VHOST_MAX_PEND 128
  56 #define VHOST_GOODCOPY_LEN 256
  57
  58 /*
  59  * For transmit, used buffer len is unused; we override it to track buffer
  60  * status internally; used for zerocopy tx only.
  61  */
  62 /* Lower device DMA failed */
  63 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  64 /* Lower device DMA done */
  65 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  66 /* Lower device DMA in progress */
  67 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  68 /* Buffer unused */
  69 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  70
  71 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  72
  73 enum {
  74         VHOST_NET_FEATURES = VHOST_FEATURES |
  75                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  76                          (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  77                          (1ULL << VIRTIO_F_IOMMU_PLATFORM)
  78 };
  79
  80 enum {
  81         VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
  82 };
  83
  84 enum {
  85         VHOST_NET_VQ_RX = 0,
  86         VHOST_NET_VQ_TX = 1,
  87         VHOST_NET_VQ_MAX = 2,
  88 };
  89
  90 struct vhost_net_ubuf_ref {
  91         /* refcount follows semantics similar to kref:
  92          *  0: object is released
  93          *  1: no outstanding ubufs
  94          * >1: outstanding ubufs
  95          */
  96         atomic_t refcount;
  97         wait_queue_head_t wait;
  98         struct vhost_virtqueue *vq;
  99 };
 100
 101 #define VHOST_NET_BATCH 64
 102 struct vhost_net_buf {
 103         void **queue;
 104         int tail;
 105         int head;
 106 };
 107
 108 struct vhost_net_virtqueue {
 109         struct vhost_virtqueue vq;
 110         size_t vhost_hlen;
 111         size_t sock_hlen;
 112         /* vhost zerocopy support fields below: */
 113         /* last used idx for outstanding DMA zerocopy buffers */
 114         int upend_idx;
 115         /* For TX, first used idx for DMA done zerocopy buffers
 116          * For RX, number of batched heads
 117          */
 118         int done_idx;
 119         /* Number of XDP frames batched */
 120         int batched_xdp;
 121         /* an array of userspace buffers info */
 122         struct ubuf_info *ubuf_info;
 123         /* Reference counting for outstanding ubufs.
 124          * Protected by vq mutex. Writers must also take device mutex. */
 125         struct vhost_net_ubuf_ref *ubufs;
 126         struct ptr_ring *rx_ring;
 127         struct vhost_net_buf rxq;
 128         /* Batched XDP buffs */
 129         struct xdp_buff *xdp;
 130 };
 131
 132 struct vhost_net {
 133         struct vhost_dev dev;
 134         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 135         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 136         /* Number of TX recently submitted.
 137          * Protected by tx vq lock. */
 138         unsigned tx_packets;
 139         /* Number of times zerocopy TX recently failed.
 140          * Protected by tx vq lock. */
 141         unsigned tx_zcopy_err;
 142         /* Flush in progress. Protected by tx vq lock. */
 143         bool tx_flush;
 144         /* Private page frag */
 145         struct page_frag page_frag;
 146         /* Refcount bias of page frag */
 147         int refcnt_bias;
 148 };
 149
 150 static unsigned vhost_net_zcopy_mask __read_mostly;
 151
 152 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 153 {
 154         if (rxq->tail != rxq->head)
 155                 return rxq->queue[rxq->head];
 156         else
 157                 return NULL;
 158 }
 159
 160 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 161 {
 162         return rxq->tail - rxq->head;
 163 }
 164
 165 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 166 {
 167         return rxq->tail == rxq->head;
 168 }
 169
 170 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 171 {
 172         void *ret = vhost_net_buf_get_ptr(rxq);
 173         ++rxq->head;
 174         return ret;
 175 }
 176
 177 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 178 {
 179         struct vhost_net_buf *rxq = &nvq->rxq;
 180
 181         rxq->head = 0;
 182         rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
 183                                               VHOST_NET_BATCH);
 184         return rxq->tail;
 185 }
 186
 187 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 188 {
 189         struct vhost_net_buf *rxq = &nvq->rxq;
 190
 191         if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
 192                 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 193                                    vhost_net_buf_get_size(rxq),
 194                                    tun_ptr_free);
 195                 rxq->head = rxq->tail = 0;
 196         }
 197 }
 198
 199 static int vhost_net_buf_peek_len(void *ptr)
 200 {
 201         if (tun_is_xdp_frame(ptr)) {
 202                 struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
 203
 204                 return xdpf->len;
 205         }
 206
 207         return __skb_array_len_with_tag(ptr);
 208 }
 209
 210 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 211 {
 212         struct vhost_net_buf *rxq = &nvq->rxq;
 213
 214         if (!vhost_net_buf_is_empty(rxq))
 215                 goto out;
 216
 217         if (!vhost_net_buf_produce(nvq))
 218                 return 0;
 219
 220 out:
 221         return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 222 }
 223
 224 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 225 {
 226         rxq->head = rxq->tail = 0;
 227 }
 228
 229 static void vhost_net_enable_zcopy(int vq)
 230 {
 231         vhost_net_zcopy_mask |= 0x1 << vq;
 232 }
 233
 234 static struct vhost_net_ubuf_ref *
 235 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 236 {
 237         struct vhost_net_ubuf_ref *ubufs;
 238         /* No zero copy backend? Nothing to count. */
 239         if (!zcopy)
 240                 return NULL;
 241         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 242         if (!ubufs)
 243                 return ERR_PTR(-ENOMEM);
 244         atomic_set(&ubufs->refcount, 1);
 245         init_waitqueue_head(&ubufs->wait);
 246         ubufs->vq = vq;
 247         return ubufs;
 248 }
 249
 250 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 251 {
 252         int r = atomic_sub_return(1, &ubufs->refcount);
 253         if (unlikely(!r))
 254                 wake_up(&ubufs->wait);
 255         return r;
 256 }
 257
 258 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 259 {
 260         vhost_net_ubuf_put(ubufs);
 261         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 262 }
 263
 264 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 265 {
 266         vhost_net_ubuf_put_and_wait(ubufs);
 267         kfree(ubufs);
 268 }
 269
 270 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 271 {
 272         int i;
 273
 274         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 275                 kfree(n->vqs[i].ubuf_info);
 276                 n->vqs[i].ubuf_info = NULL;
 277         }
 278 }
 279
 280 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 281 {
 282         bool zcopy;
 283         int i;
 284
 285         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 286                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 287                 if (!zcopy)
 288                         continue;
 289                 n->vqs[i].ubuf_info =
 290                         kmalloc_array(UIO_MAXIOV,
 291                                       sizeof(*n->vqs[i].ubuf_info),
 292                                       GFP_KERNEL);
 293                 if  (!n->vqs[i].ubuf_info)
 294                         goto err;
 295         }
 296         return 0;
 297
 298 err:
 299         vhost_net_clear_ubuf_info(n);
 300         return -ENOMEM;
 301 }
 302
 303 static void vhost_net_vq_reset(struct vhost_net *n)
 304 {
 305         int i;
 306
 307         vhost_net_clear_ubuf_info(n);
 308
 309         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 310                 n->vqs[i].done_idx = 0;
 311                 n->vqs[i].upend_idx = 0;
 312                 n->vqs[i].ubufs = NULL;
 313                 n->vqs[i].vhost_hlen = 0;
 314                 n->vqs[i].sock_hlen = 0;
 315                 vhost_net_buf_init(&n->vqs[i].rxq);
 316         }
 317
 318 }
 319
 320 static void vhost_net_tx_packet(struct vhost_net *net)
 321 {
 322         ++net->tx_packets;
 323         if (net->tx_packets < 1024)
 324                 return;
 325         net->tx_packets = 0;
 326         net->tx_zcopy_err = 0;
 327 }
 328
 329 static void vhost_net_tx_err(struct vhost_net *net)
 330 {
 331         ++net->tx_zcopy_err;
 332 }
 333
 334 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 335 {
 336         /* TX flush waits for outstanding DMAs to be done.
 337          * Don't start new DMAs.
 338          */
 339         return !net->tx_flush &&
 340                 net->tx_packets / 64 >= net->tx_zcopy_err;
 341 }
 342
 343 static bool vhost_sock_zcopy(struct socket *sock)
 344 {
 345         return unlikely(experimental_zcopytx) &&
 346                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 347 }
 348
 349 static bool vhost_sock_xdp(struct socket *sock)
 350 {
 351         return sock_flag(sock->sk, SOCK_XDP);
 352 }
 353
 354 /* In case of DMA done not in order in lower device driver for some reason.
 355  * upend_idx is used to track end of used idx, done_idx is used to track head
 356  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 357  * guest used idx.
 358  */
 359 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 360                                        struct vhost_virtqueue *vq)
 361 {
 362         struct vhost_net_virtqueue *nvq =
 363                 container_of(vq, struct vhost_net_virtqueue, vq);
 364         int i, add;
 365         int j = 0;
 366
 367         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 368                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 369                         vhost_net_tx_err(net);
 370                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 371                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 372                         ++j;
 373                 } else
 374                         break;
 375         }
 376         while (j) {
 377                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 378                 vhost_add_used_and_signal_n(vq->dev, vq,
 379                                             &vq->heads[nvq->done_idx], add);
 380                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 381                 j -= add;
 382         }
 383 }
 384
 385 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 386 {
 387         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 388         struct vhost_virtqueue *vq = ubufs->vq;
 389         int cnt;
 390
 391         rcu_read_lock_bh();
 392
 393         /* set len to mark this desc buffers done DMA */
 394         vq->heads[ubuf->desc].len = success ?
 395                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 396         cnt = vhost_net_ubuf_put(ubufs);
 397
 398         /*
 399          * Trigger polling thread if guest stopped submitting new buffers:
 400          * in this case, the refcount after decrement will eventually reach 1.
 401          * We also trigger polling periodically after each 16 packets
 402          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 403          * less than 10% of times).
 404          */
 405         if (cnt <= 1 || !(cnt % 16))
 406                 vhost_poll_queue(&vq->poll);
 407
 408         rcu_read_unlock_bh();
 409 }
 410
 411 static inline unsigned long busy_clock(void)
 412 {
 413         return local_clock() >> 10;
 414 }
 415
 416 static bool vhost_can_busy_poll(unsigned long endtime)
 417 {
 418         return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
 419                       !signal_pending(current));
 420 }
 421
 422 static void vhost_net_disable_vq(struct vhost_net *n,
 423                                  struct vhost_virtqueue *vq)
 424 {
 425         struct vhost_net_virtqueue *nvq =
 426                 container_of(vq, struct vhost_net_virtqueue, vq);
 427         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 428         if (!vq->private_data)
 429                 return;
 430         vhost_poll_stop(poll);
 431 }
 432
 433 static int vhost_net_enable_vq(struct vhost_net *n,
 434                                 struct vhost_virtqueue *vq)
 435 {
 436         struct vhost_net_virtqueue *nvq =
 437                 container_of(vq, struct vhost_net_virtqueue, vq);
 438         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 439         struct socket *sock;
 440
 441         sock = vq->private_data;
 442         if (!sock)
 443                 return 0;
 444
 445         return vhost_poll_start(poll, sock->file);
 446 }
 447
 448 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
 449 {
 450         struct vhost_virtqueue *vq = &nvq->vq;
 451         struct vhost_dev *dev = vq->dev;
 452
 453         if (!nvq->done_idx)
 454                 return;
 455
 456         vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
 457         nvq->done_idx = 0;
 458 }
 459
 460 static void vhost_tx_batch(struct vhost_net *net,
 461                            struct vhost_net_virtqueue *nvq,
 462                            struct socket *sock,
 463                            struct msghdr *msghdr)
 464 {
 465         struct tun_msg_ctl ctl = {
 466                 .type = TUN_MSG_PTR,
 467                 .num = nvq->batched_xdp,
 468                 .ptr = nvq->xdp,
 469         };
 470         int 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
 482 signal_used:
 483         vhost_net_signal_used(nvq);
 484         nvq->batched_xdp = 0;
 485 }
 486
 487 static 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
 498 static 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
 509 static 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
 560 static 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
 586 static 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
 595 static 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
 607 static bool vhost_exceeds_weight(int pkts, int total_len)
 608 {
 609         return total_len >= VHOST_NET_WEIGHT ||
 610                pkts >= VHOST_NET_PKT_WEIGHT;
 611 }
 612
 613 static int get_tx_bufs(struct vhost_net *net,
 614                        struct vhost_net_virtqueue *nvq,
 615                        struct msghdr *msg,
 616                        unsigned int *out, unsigned int *in,
 617                        size_t *len, bool *busyloop_intr)
 618 {
 619         struct vhost_virtqueue *vq = &nvq->vq;
 620         int ret;
 621
 622         ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
 623
 624         if (ret < 0 || ret == vq->num)
 625                 return ret;
 626
 627         if (*in) {
 628                 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 629                         *out, *in);
 630                 return -EFAULT;
 631         }
 632
 633         /* Sanity check */
 634         *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 635         if (*len == 0) {
 636                 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 637                         *len, nvq->vhost_hlen);
 638                 return -EFAULT;
 639         }
 640
 641         return ret;
 642 }
 643
 644 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 645 {
 646         return total_len < VHOST_NET_WEIGHT &&
 647                !vhost_vq_avail_empty(vq->dev, vq);
 648 }
 649
 650 #define SKB_FRAG_PAGE_ORDER     get_order(32768)
 651
 652 static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
 653                                        struct page_frag *pfrag, gfp_t gfp)
 654 {
 655         if (pfrag->page) {
 656                 if (pfrag->offset + sz <= pfrag->size)
 657                         return true;
 658                 __page_frag_cache_drain(pfrag->page, net->refcnt_bias);
 659         }
 660
 661         pfrag->offset = 0;
 662         net->refcnt_bias = 0;
 663         if (SKB_FRAG_PAGE_ORDER) {
 664                 /* Avoid direct reclaim but allow kswapd to wake */
 665                 pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
 666                                           __GFP_COMP | __GFP_NOWARN |
 667                                           __GFP_NORETRY,
 668                                           SKB_FRAG_PAGE_ORDER);
 669                 if (likely(pfrag->page)) {
 670                         pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
 671                         goto done;
 672                 }
 673         }
 674         pfrag->page = alloc_page(gfp);
 675         if (likely(pfrag->page)) {
 676                 pfrag->size = PAGE_SIZE;
 677                 goto done;
 678         }
 679         return false;
 680
 681 done:
 682         net->refcnt_bias = USHRT_MAX;
 683         page_ref_add(pfrag->page, USHRT_MAX - 1);
 684         return true;
 685 }
 686
 687 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 688
 689 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
 690                                struct iov_iter *from)
 691 {
 692         struct vhost_virtqueue *vq = &nvq->vq;
 693         struct vhost_net *net = container_of(vq->dev, struct vhost_net,
 694                                              dev);
 695         struct socket *sock = vq->private_data;
 696         struct page_frag *alloc_frag = &net->page_frag;
 697         struct virtio_net_hdr *gso;
 698         struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
 699         struct tun_xdp_hdr *hdr;
 700         size_t len = iov_iter_count(from);
 701         int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
 702         int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 703         int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
 704         int sock_hlen = nvq->sock_hlen;
 705         void *buf;
 706         int copied;
 707
 708         if (unlikely(len < nvq->sock_hlen))
 709                 return -EFAULT;
 710
 711         if (SKB_DATA_ALIGN(len + pad) +
 712             SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
 713                 return -ENOSPC;
 714
 715         buflen += SKB_DATA_ALIGN(len + pad);
 716         alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
 717         if (unlikely(!vhost_net_page_frag_refill(net, buflen,
 718                                                  alloc_frag, GFP_KERNEL)))
 719                 return -ENOMEM;
 720
 721         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 722         copied = copy_page_from_iter(alloc_frag->page,
 723                                      alloc_frag->offset +
 724                                      offsetof(struct tun_xdp_hdr, gso),
 725                                      sock_hlen, from);
 726         if (copied != sock_hlen)
 727                 return -EFAULT;
 728
 729         hdr = buf;
 730         gso = &hdr->gso;
 731
 732         if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
 733             vhost16_to_cpu(vq, gso->csum_start) +
 734             vhost16_to_cpu(vq, gso->csum_offset) + 2 >
 735             vhost16_to_cpu(vq, gso->hdr_len)) {
 736                 gso->hdr_len = cpu_to_vhost16(vq,
 737                                vhost16_to_cpu(vq, gso->csum_start) +
 738                                vhost16_to_cpu(vq, gso->csum_offset) + 2);
 739
 740                 if (vhost16_to_cpu(vq, gso->hdr_len) > len)
 741                         return -EINVAL;
 742         }
 743
 744         len -= sock_hlen;
 745         copied = copy_page_from_iter(alloc_frag->page,
 746                                      alloc_frag->offset + pad,
 747                                      len, from);
 748         if (copied != len)
 749                 return -EFAULT;
 750
 751         xdp->data_hard_start = buf;
 752         xdp->data = buf + pad;
 753         xdp->data_end = xdp->data + len;
 754         hdr->buflen = buflen;
 755
 756         --net->refcnt_bias;
 757         alloc_frag->offset += buflen;
 758
 759         ++nvq->batched_xdp;
 760
 761         return 0;
 762 }
 763
 764 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 765 {
 766         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 767         struct vhost_virtqueue *vq = &nvq->vq;
 768         unsigned out, in;
 769         int head;
 770         struct msghdr msg = {
 771                 .msg_name = NULL,
 772                 .msg_namelen = 0,
 773                 .msg_control = NULL,
 774                 .msg_controllen = 0,
 775                 .msg_flags = MSG_DONTWAIT,
 776         };
 777         size_t len, total_len = 0;
 778         int err;
 779         int sent_pkts = 0;
 780         bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
 781
 782         for (;;) {
 783                 bool busyloop_intr = false;
 784
 785                 if (nvq->done_idx == VHOST_NET_BATCH)
 786                         vhost_tx_batch(net, nvq, sock, &msg);
 787
 788                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 789                                    &busyloop_intr);
 790                 /* On error, stop handling until the next kick. */
 791                 if (unlikely(head < 0))
 792                         break;
 793                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 794                 if (head == vq->num) {
 795                         if (unlikely(busyloop_intr)) {
 796                                 vhost_poll_queue(&vq->poll);
 797                         } else if (unlikely(vhost_enable_notify(&net->dev,
 798                                                                 vq))) {
 799                                 vhost_disable_notify(&net->dev, vq);
 800                                 continue;
 801                         }
 802                         break;
 803                 }
 804
 805                 total_len += len;
 806
 807                 /* For simplicity, TX batching is only enabled if
 808                  * sndbuf is unlimited.
 809                  */
 810                 if (sock_can_batch) {
 811                         err = vhost_net_build_xdp(nvq, &msg.msg_iter);
 812                         if (!err) {
 813                                 goto done;
 814                         } else if (unlikely(err != -ENOSPC)) {
 815                                 vhost_tx_batch(net, nvq, sock, &msg);
 816                                 vhost_discard_vq_desc(vq, 1);
 817                                 vhost_net_enable_vq(net, vq);
 818                                 break;
 819                         }
 820
 821                         /* We can't build XDP buff, go for single
 822                          * packet path but let's flush batched
 823                          * packets.
 824                          */
 825                         vhost_tx_batch(net, nvq, sock, &msg);
 826                         msg.msg_control = NULL;
 827                 } else {
 828                         if (tx_can_batch(vq, total_len))
 829                                 msg.msg_flags |= MSG_MORE;
 830                         else
 831                                 msg.msg_flags &= ~MSG_MORE;
 832                 }
 833
 834                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 835                 err = sock->ops->sendmsg(sock, &msg, len);
 836                 if (unlikely(err < 0)) {
 837                         vhost_discard_vq_desc(vq, 1);
 838                         vhost_net_enable_vq(net, vq);
 839                         break;
 840                 }
 841                 if (err != len)
 842                         pr_debug("Truncated TX packet: len %d != %zd\n",
 843                                  err, len);
 844 done:
 845                 vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 846                 vq->heads[nvq->done_idx].len = 0;
 847                 ++nvq->done_idx;
 848                 if (vhost_exceeds_weight(++sent_pkts, total_len)) {
 849                         vhost_poll_queue(&vq->poll);
 850                         break;
 851                 }
 852         }
 853
 854         vhost_tx_batch(net, nvq, sock, &msg);
 855 }
 856
 857 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 858 {
 859         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 860         struct vhost_virtqueue *vq = &nvq->vq;
 861         unsigned out, in;
 862         int head;
 863         struct msghdr msg = {
 864                 .msg_name = NULL,
 865                 .msg_namelen = 0,
 866                 .msg_control = NULL,
 867                 .msg_controllen = 0,
 868                 .msg_flags = MSG_DONTWAIT,
 869         };
 870         struct tun_msg_ctl ctl;
 871         size_t len, total_len = 0;
 872         int err;
 873         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 874         bool zcopy_used;
 875         int sent_pkts = 0;
 876
 877         for (;;) {
 878                 bool busyloop_intr;
 879
 880                 /* Release DMAs done buffers first */
 881                 vhost_zerocopy_signal_used(net, vq);
 882
 883                 busyloop_intr = false;
 884                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 885                                    &busyloop_intr);
 886                 /* On error, stop handling until the next kick. */
 887                 if (unlikely(head < 0))
 888                         break;
 889                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 890                 if (head == vq->num) {
 891                         if (unlikely(busyloop_intr)) {
 892                                 vhost_poll_queue(&vq->poll);
 893                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 894                                 vhost_disable_notify(&net->dev, vq);
 895                                 continue;
 896                         }
 897                         break;
 898                 }
 899
 900                 zcopy_used = len >= VHOST_GOODCOPY_LEN
 901                              && !vhost_exceeds_maxpend(net)
 902                              && vhost_net_tx_select_zcopy(net);
 903
 904                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 905                 if (zcopy_used) {
 906                         struct ubuf_info *ubuf;
 907                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 908
 909                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 910                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 911                         ubuf->callback = vhost_zerocopy_callback;
 912                         ubuf->ctx = nvq->ubufs;
 913                         ubuf->desc = nvq->upend_idx;
 914                         refcount_set(&ubuf->refcnt, 1);
 915                         msg.msg_control = &ctl;
 916                         ctl.type = TUN_MSG_UBUF;
 917                         ctl.ptr = ubuf;
 918                         msg.msg_controllen = sizeof(ctl);
 919                         ubufs = nvq->ubufs;
 920                         atomic_inc(&ubufs->refcount);
 921                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 922                 } else {
 923                         msg.msg_control = NULL;
 924                         ubufs = NULL;
 925                 }
 926                 total_len += len;
 927                 if (tx_can_batch(vq, total_len) &&
 928                     likely(!vhost_exceeds_maxpend(net))) {
 929                         msg.msg_flags |= MSG_MORE;
 930                 } else {
 931                         msg.msg_flags &= ~MSG_MORE;
 932                 }
 933
 934                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 935                 err = sock->ops->sendmsg(sock, &msg, len);
 936                 if (unlikely(err < 0)) {
 937                         if (zcopy_used) {
 938                                 vhost_net_ubuf_put(ubufs);
 939                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 940                                         % UIO_MAXIOV;
 941                         }
 942                         vhost_discard_vq_desc(vq, 1);
 943                         vhost_net_enable_vq(net, vq);
 944                         break;
 945                 }
 946                 if (err != len)
 947                         pr_debug("Truncated TX packet: "
 948                                  " len %d != %zd\n", err, len);
 949                 if (!zcopy_used)
 950                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 951                 else
 952                         vhost_zerocopy_signal_used(net, vq);
 953                 vhost_net_tx_packet(net);
 954                 if (unlikely(vhost_exceeds_weight(++sent_pkts, total_len))) {
 955                         vhost_poll_queue(&vq->poll);
 956                         break;
 957                 }
 958         }
 959 }
 960
 961 /* Expects to be always run from workqueue - which acts as
 962  * read-size critical section for our kind of RCU. */
 963 static void handle_tx(struct vhost_net *net)
 964 {
 965         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 966         struct vhost_virtqueue *vq = &nvq->vq;
 967         struct socket *sock;
 968
 969         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
 970         sock = vq->private_data;
 971         if (!sock)
 972                 goto out;
 973
 974         if (!vq_iotlb_prefetch(vq))
 975                 goto out;
 976
 977         vhost_disable_notify(&net->dev, vq);
 978         vhost_net_disable_vq(net, vq);
 979
 980         if (vhost_sock_zcopy(sock))
 981                 handle_tx_zerocopy(net, sock);
 982         else
 983                 handle_tx_copy(net, sock);
 984
 985 out:
 986         mutex_unlock(&vq->mutex);
 987 }
 988
 989 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 990 {
 991         struct sk_buff *head;
 992         int len = 0;
 993         unsigned long flags;
 994
 995         if (rvq->rx_ring)
 996                 return vhost_net_buf_peek(rvq);
 997
 998         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 999         head = skb_peek(&sk->sk_receive_queue);
1000         if (likely(head)) {
1001                 len = head->len;
1002                 if (skb_vlan_tag_present(head))
1003                         len += VLAN_HLEN;
1004         }
1005
1006         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
1007         return len;
1008 }
1009
1010 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
1011                                       bool *busyloop_intr)
1012 {
1013         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1014         struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1015         struct vhost_virtqueue *rvq = &rnvq->vq;
1016         struct vhost_virtqueue *tvq = &tnvq->vq;
1017         int len = peek_head_len(rnvq, sk);
1018
1019         if (!len && rvq->busyloop_timeout) {
1020                 /* Flush batched heads first */
1021                 vhost_net_signal_used(rnvq);
1022                 /* Both tx vq and rx socket were polled here */
1023                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1024
1025                 len = peek_head_len(rnvq, sk);
1026         }
1027
1028         return len;
1029 }
1030
1031 /* This is a multi-buffer version of vhost_get_desc, that works if
1032  *      vq has read descriptors only.
1033  * @vq          - the relevant virtqueue
1034  * @datalen     - data length we'll be reading
1035  * @iovcount    - returned count of io vectors we fill
1036  * @log         - vhost log
1037  * @log_num     - log offset
1038  * @quota       - headcount quota, 1 for big buffer
1039  *      returns number of buffer heads allocated, negative on error
1040  */
1041 static int get_rx_bufs(struct vhost_virtqueue *vq,
1042                        struct vring_used_elem *heads,
1043                        int datalen,
1044                        unsigned *iovcount,
1045                        struct vhost_log *log,
1046                        unsigned *log_num,
1047                        unsigned int quota)
1048 {
1049         unsigned int out, in;
1050         int seg = 0;
1051         int headcount = 0;
1052         unsigned d;
1053         int r, nlogs = 0;
1054         /* len is always initialized before use since we are always called with
1055          * datalen > 0.
1056          */
1057         u32 uninitialized_var(len);
1058
1059         while (datalen > 0 && headcount < quota) {
1060                 if (unlikely(seg >= UIO_MAXIOV)) {
1061                         r = -ENOBUFS;
1062                         goto err;
1063                 }
1064                 r = vhost_get_vq_desc(vq, vq->iov + seg,
1065                                       ARRAY_SIZE(vq->iov) - seg, &out,
1066                                       &in, log, log_num);
1067                 if (unlikely(r < 0))
1068                         goto err;
1069
1070                 d = r;
1071                 if (d == vq->num) {
1072                         r = 0;
1073                         goto err;
1074                 }
1075                 if (unlikely(out || in <= 0)) {
1076                         vq_err(vq, "unexpected descriptor format for RX: "
1077                                 "out %d, in %d\n", out, in);
1078                         r = -EINVAL;
1079                         goto err;
1080                 }
1081                 if (unlikely(log)) {
1082                         nlogs += *log_num;
1083                         log += *log_num;
1084                 }
1085                 heads[headcount].id = cpu_to_vhost32(vq, d);
1086                 len = iov_length(vq->iov + seg, in);
1087                 heads[headcount].len = cpu_to_vhost32(vq, len);
1088                 datalen -= len;
1089                 ++headcount;
1090                 seg += in;
1091         }
1092         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1093         *iovcount = seg;
1094         if (unlikely(log))
1095                 *log_num = nlogs;
1096
1097         /* Detect overrun */
1098         if (unlikely(datalen > 0)) {
1099                 r = UIO_MAXIOV + 1;
1100                 goto err;
1101         }
1102         return headcount;
1103 err:
1104         vhost_discard_vq_desc(vq, headcount);
1105         return r;
1106 }
1107
1108 /* Expects to be always run from workqueue - which acts as
1109  * read-size critical section for our kind of RCU. */
1110 static void handle_rx(struct vhost_net *net)
1111 {
1112         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1113         struct vhost_virtqueue *vq = &nvq->vq;
1114         unsigned uninitialized_var(in), log;
1115         struct vhost_log *vq_log;
1116         struct msghdr msg = {
1117                 .msg_name = NULL,
1118                 .msg_namelen = 0,
1119                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
1120                 .msg_controllen = 0,
1121                 .msg_flags = MSG_DONTWAIT,
1122         };
1123         struct virtio_net_hdr hdr = {
1124                 .flags = 0,
1125                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
1126         };
1127         size_t total_len = 0;
1128         int err, mergeable;
1129         s16 headcount;
1130         size_t vhost_hlen, sock_hlen;
1131         size_t vhost_len, sock_len;
1132         bool busyloop_intr = false;
1133         struct socket *sock;
1134         struct iov_iter fixup;
1135         __virtio16 num_buffers;
1136         int recv_pkts = 0;
1137
1138         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1139         sock = vq->private_data;
1140         if (!sock)
1141                 goto out;
1142
1143         if (!vq_iotlb_prefetch(vq))
1144                 goto out;
1145
1146         vhost_disable_notify(&net->dev, vq);
1147         vhost_net_disable_vq(net, vq);
1148
1149         vhost_hlen = nvq->vhost_hlen;
1150         sock_hlen = nvq->sock_hlen;
1151
1152         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1153                 vq->log : NULL;
1154         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1155
1156         while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1157                                                       &busyloop_intr))) {
1158                 sock_len += sock_hlen;
1159                 vhost_len = sock_len + vhost_hlen;
1160                 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1161                                         vhost_len, &in, vq_log, &log,
1162                                         likely(mergeable) ? UIO_MAXIOV : 1);
1163                 /* On error, stop handling until the next kick. */
1164                 if (unlikely(headcount < 0))
1165                         goto out;
1166                 /* OK, now we need to know about added descriptors. */
1167                 if (!headcount) {
1168                         if (unlikely(busyloop_intr)) {
1169                                 vhost_poll_queue(&vq->poll);
1170                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1171                                 /* They have slipped one in as we were
1172                                  * doing that: check again. */
1173                                 vhost_disable_notify(&net->dev, vq);
1174                                 continue;
1175                         }
1176                         /* Nothing new?  Wait for eventfd to tell us
1177                          * they refilled. */
1178                         goto out;
1179                 }
1180                 busyloop_intr = false;
1181                 if (nvq->rx_ring)
1182                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1183                 /* On overrun, truncate and discard */
1184                 if (unlikely(headcount > UIO_MAXIOV)) {
1185                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
1186                         err = sock->ops->recvmsg(sock, &msg,
1187                                                  1, MSG_DONTWAIT | MSG_TRUNC);
1188                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
1189                         continue;
1190                 }
1191                 /* We don't need to be notified again. */
1192                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
1193                 fixup = msg.msg_iter;
1194                 if (unlikely((vhost_hlen))) {
1195                         /* We will supply the header ourselves
1196                          * TODO: support TSO.
1197                          */
1198                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
1199                 }
1200                 err = sock->ops->recvmsg(sock, &msg,
1201                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
1202                 /* Userspace might have consumed the packet meanwhile:
1203                  * it's not supposed to do this usually, but might be hard
1204                  * to prevent. Discard data we got (if any) and keep going. */
1205                 if (unlikely(err != sock_len)) {
1206                         pr_debug("Discarded rx packet: "
1207                                  " len %d, expected %zd\n", err, sock_len);
1208                         vhost_discard_vq_desc(vq, headcount);
1209                         continue;
1210                 }
1211                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1212                 if (unlikely(vhost_hlen)) {
1213                         if (copy_to_iter(&hdr, sizeof(hdr),
1214                                          &fixup) != sizeof(hdr)) {
1215                                 vq_err(vq, "Unable to write vnet_hdr "
1216                                        "at addr %p\n", vq->iov->iov_base);
1217                                 goto out;
1218                         }
1219                 } else {
1220                         /* Header came from socket; we'll need to patch
1221                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1222                          */
1223                         iov_iter_advance(&fixup, sizeof(hdr));
1224                 }
1225                 /* TODO: Should check and handle checksum. */
1226
1227                 num_buffers = cpu_to_vhost16(vq, headcount);
1228                 if (likely(mergeable) &&
1229                     copy_to_iter(&num_buffers, sizeof num_buffers,
1230                                  &fixup) != sizeof num_buffers) {
1231                         vq_err(vq, "Failed num_buffers write");
1232                         vhost_discard_vq_desc(vq, headcount);
1233                         goto out;
1234                 }
1235                 nvq->done_idx += headcount;
1236                 if (nvq->done_idx > VHOST_NET_BATCH)
1237                         vhost_net_signal_used(nvq);
1238                 if (unlikely(vq_log))
1239                         vhost_log_write(vq, vq_log, log, vhost_len);
1240                 total_len += vhost_len;
1241                 if (unlikely(vhost_exceeds_weight(++recv_pkts, total_len))) {
1242                         vhost_poll_queue(&vq->poll);
1243                         goto out;
1244                 }
1245         }
1246         if (unlikely(busyloop_intr))
1247                 vhost_poll_queue(&vq->poll);
1248         else
1249                 vhost_net_enable_vq(net, vq);
1250 out:
1251         vhost_net_signal_used(nvq);
1252         mutex_unlock(&vq->mutex);
1253 }
1254
1255 static void handle_tx_kick(struct vhost_work *work)
1256 {
1257         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1258                                                   poll.work);
1259         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1260
1261         handle_tx(net);
1262 }
1263
1264 static void handle_rx_kick(struct vhost_work *work)
1265 {
1266         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1267                                                   poll.work);
1268         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1269
1270         handle_rx(net);
1271 }
1272
1273 static void handle_tx_net(struct vhost_work *work)
1274 {
1275         struct vhost_net *net = container_of(work, struct vhost_net,
1276                                              poll[VHOST_NET_VQ_TX].work);
1277         handle_tx(net);
1278 }
1279
1280 static void handle_rx_net(struct vhost_work *work)
1281 {
1282         struct vhost_net *net = container_of(work, struct vhost_net,
1283                                              poll[VHOST_NET_VQ_RX].work);
1284         handle_rx(net);
1285 }
1286
1287 static int vhost_net_open(struct inode *inode, struct file *f)
1288 {
1289         struct vhost_net *n;
1290         struct vhost_dev *dev;
1291         struct vhost_virtqueue **vqs;
1292         void **queue;
1293         struct xdp_buff *xdp;
1294         int i;
1295
1296         n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1297         if (!n)
1298                 return -ENOMEM;
1299         vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1300         if (!vqs) {
1301                 kvfree(n);
1302                 return -ENOMEM;
1303         }
1304
1305         queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1306                               GFP_KERNEL);
1307         if (!queue) {
1308                 kfree(vqs);
1309                 kvfree(n);
1310                 return -ENOMEM;
1311         }
1312         n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1313
1314         xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1315         if (!xdp) {
1316                 kfree(vqs);
1317                 kvfree(n);
1318                 kfree(queue);
1319                 return -ENOMEM;
1320         }
1321         n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1322
1323         dev = &n->dev;
1324         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1325         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1326         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1327         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1328         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1329                 n->vqs[i].ubufs = NULL;
1330                 n->vqs[i].ubuf_info = NULL;
1331                 n->vqs[i].upend_idx = 0;
1332                 n->vqs[i].done_idx = 0;
1333                 n->vqs[i].batched_xdp = 0;
1334                 n->vqs[i].vhost_hlen = 0;
1335                 n->vqs[i].sock_hlen = 0;
1336                 n->vqs[i].rx_ring = NULL;
1337                 vhost_net_buf_init(&n->vqs[i].rxq);
1338         }
1339         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
1340
1341         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1342         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1343
1344         f->private_data = n;
1345         n->page_frag.page = NULL;
1346         n->refcnt_bias = 0;
1347
1348         return 0;
1349 }
1350
1351 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1352                                         struct vhost_virtqueue *vq)
1353 {
1354         struct socket *sock;
1355         struct vhost_net_virtqueue *nvq =
1356                 container_of(vq, struct vhost_net_virtqueue, vq);
1357
1358         mutex_lock(&vq->mutex);
1359         sock = vq->private_data;
1360         vhost_net_disable_vq(n, vq);
1361         vq->private_data = NULL;
1362         vhost_net_buf_unproduce(nvq);
1363         nvq->rx_ring = NULL;
1364         mutex_unlock(&vq->mutex);
1365         return sock;
1366 }
1367
1368 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1369                            struct socket **rx_sock)
1370 {
1371         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1372         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1373 }
1374
1375 static void vhost_net_flush_vq(struct vhost_net *n, int index)
1376 {
1377         vhost_poll_flush(n->poll + index);
1378         vhost_poll_flush(&n->vqs[index].vq.poll);
1379 }
1380
1381 static void vhost_net_flush(struct vhost_net *n)
1382 {
1383         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1384         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1385         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1386                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1387                 n->tx_flush = true;
1388                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1389                 /* Wait for all lower device DMAs done. */
1390                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1391                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1392                 n->tx_flush = false;
1393                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1394                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1395         }
1396 }
1397
1398 static int vhost_net_release(struct inode *inode, struct file *f)
1399 {
1400         struct vhost_net *n = f->private_data;
1401         struct socket *tx_sock;
1402         struct socket *rx_sock;
1403
1404         vhost_net_stop(n, &tx_sock, &rx_sock);
1405         vhost_net_flush(n);
1406         vhost_dev_stop(&n->dev);
1407         vhost_dev_cleanup(&n->dev);
1408         vhost_net_vq_reset(n);
1409         if (tx_sock)
1410                 sockfd_put(tx_sock);
1411         if (rx_sock)
1412                 sockfd_put(rx_sock);
1413         /* Make sure no callbacks are outstanding */
1414         synchronize_rcu();
1415         /* We do an extra flush before freeing memory,
1416          * since jobs can re-queue themselves. */
1417         vhost_net_flush(n);
1418         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1419         kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1420         kfree(n->dev.vqs);
1421         if (n->page_frag.page)
1422                 __page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
1423         kvfree(n);
1424         return 0;
1425 }
1426
1427 static struct socket *get_raw_socket(int fd)
1428 {
1429         struct {
1430                 struct sockaddr_ll sa;
1431                 char  buf[MAX_ADDR_LEN];
1432         } uaddr;
1433         int r;
1434         struct socket *sock = sockfd_lookup(fd, &r);
1435
1436         if (!sock)
1437                 return ERR_PTR(-ENOTSOCK);
1438
1439         /* Parameter checking */
1440         if (sock->sk->sk_type != SOCK_RAW) {
1441                 r = -ESOCKTNOSUPPORT;
1442                 goto err;
1443         }
1444
1445         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa, 0);
1446         if (r < 0)
1447                 goto err;
1448
1449         if (uaddr.sa.sll_family != AF_PACKET) {
1450                 r = -EPFNOSUPPORT;
1451                 goto err;
1452         }
1453         return sock;
1454 err:
1455         sockfd_put(sock);
1456         return ERR_PTR(r);
1457 }
1458
1459 static struct ptr_ring *get_tap_ptr_ring(int fd)
1460 {
1461         struct ptr_ring *ring;
1462         struct file *file = fget(fd);
1463
1464         if (!file)
1465                 return NULL;
1466         ring = tun_get_tx_ring(file);
1467         if (!IS_ERR(ring))
1468                 goto out;
1469         ring = tap_get_ptr_ring(file);
1470         if (!IS_ERR(ring))
1471                 goto out;
1472         ring = NULL;
1473 out:
1474         fput(file);
1475         return ring;
1476 }
1477
1478 static struct socket *get_tap_socket(int fd)
1479 {
1480         struct file *file = fget(fd);
1481         struct socket *sock;
1482
1483         if (!file)
1484                 return ERR_PTR(-EBADF);
1485         sock = tun_get_socket(file);
1486         if (!IS_ERR(sock))
1487                 return sock;
1488         sock = tap_get_socket(file);
1489         if (IS_ERR(sock))
1490                 fput(file);
1491         return sock;
1492 }
1493
1494 static struct socket *get_socket(int fd)
1495 {
1496         struct socket *sock;
1497
1498         /* special case to disable backend */
1499         if (fd == -1)
1500                 return NULL;
1501         sock = get_raw_socket(fd);
1502         if (!IS_ERR(sock))
1503                 return sock;
1504         sock = get_tap_socket(fd);
1505         if (!IS_ERR(sock))
1506                 return sock;
1507         return ERR_PTR(-ENOTSOCK);
1508 }
1509
1510 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1511 {
1512         struct socket *sock, *oldsock;
1513         struct vhost_virtqueue *vq;
1514         struct vhost_net_virtqueue *nvq;
1515         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1516         int r;
1517
1518         mutex_lock(&n->dev.mutex);
1519         r = vhost_dev_check_owner(&n->dev);
1520         if (r)
1521                 goto err;
1522
1523         if (index >= VHOST_NET_VQ_MAX) {
1524                 r = -ENOBUFS;
1525                 goto err;
1526         }
1527         vq = &n->vqs[index].vq;
1528         nvq = &n->vqs[index];
1529         mutex_lock(&vq->mutex);
1530
1531         /* Verify that ring has been setup correctly. */
1532         if (!vhost_vq_access_ok(vq)) {
1533                 r = -EFAULT;
1534                 goto err_vq;
1535         }
1536         sock = get_socket(fd);
1537         if (IS_ERR(sock)) {
1538                 r = PTR_ERR(sock);
1539                 goto err_vq;
1540         }
1541
1542         /* start polling new socket */
1543         oldsock = vq->private_data;
1544         if (sock != oldsock) {
1545                 ubufs = vhost_net_ubuf_alloc(vq,
1546                                              sock && vhost_sock_zcopy(sock));
1547                 if (IS_ERR(ubufs)) {
1548                         r = PTR_ERR(ubufs);
1549                         goto err_ubufs;
1550                 }
1551
1552                 vhost_net_disable_vq(n, vq);
1553                 vq->private_data = sock;
1554                 vhost_net_buf_unproduce(nvq);
1555                 r = vhost_vq_init_access(vq);
1556                 if (r)
1557                         goto err_used;
1558                 r = vhost_net_enable_vq(n, vq);
1559                 if (r)
1560                         goto err_used;
1561                 if (index == VHOST_NET_VQ_RX)
1562                         nvq->rx_ring = get_tap_ptr_ring(fd);
1563
1564                 oldubufs = nvq->ubufs;
1565                 nvq->ubufs = ubufs;
1566
1567                 n->tx_packets = 0;
1568                 n->tx_zcopy_err = 0;
1569                 n->tx_flush = false;
1570         }
1571
1572         mutex_unlock(&vq->mutex);
1573
1574         if (oldubufs) {
1575                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1576                 mutex_lock(&vq->mutex);
1577                 vhost_zerocopy_signal_used(n, vq);
1578                 mutex_unlock(&vq->mutex);
1579         }
1580
1581         if (oldsock) {
1582                 vhost_net_flush_vq(n, index);
1583                 sockfd_put(oldsock);
1584         }
1585
1586         mutex_unlock(&n->dev.mutex);
1587         return 0;
1588
1589 err_used:
1590         vq->private_data = oldsock;
1591         vhost_net_enable_vq(n, vq);
1592         if (ubufs)
1593                 vhost_net_ubuf_put_wait_and_free(ubufs);
1594 err_ubufs:
1595         if (sock)
1596                 sockfd_put(sock);
1597 err_vq:
1598         mutex_unlock(&vq->mutex);
1599 err:
1600         mutex_unlock(&n->dev.mutex);
1601         return r;
1602 }
1603
1604 static long vhost_net_reset_owner(struct vhost_net *n)
1605 {
1606         struct socket *tx_sock = NULL;
1607         struct socket *rx_sock = NULL;
1608         long err;
1609         struct vhost_umem *umem;
1610
1611         mutex_lock(&n->dev.mutex);
1612         err = vhost_dev_check_owner(&n->dev);
1613         if (err)
1614                 goto done;
1615         umem = vhost_dev_reset_owner_prepare();
1616         if (!umem) {
1617                 err = -ENOMEM;
1618                 goto done;
1619         }
1620         vhost_net_stop(n, &tx_sock, &rx_sock);
1621         vhost_net_flush(n);
1622         vhost_dev_stop(&n->dev);
1623         vhost_dev_reset_owner(&n->dev, umem);
1624         vhost_net_vq_reset(n);
1625 done:
1626         mutex_unlock(&n->dev.mutex);
1627         if (tx_sock)
1628                 sockfd_put(tx_sock);
1629         if (rx_sock)
1630                 sockfd_put(rx_sock);
1631         return err;
1632 }
1633
1634 static int vhost_net_set_backend_features(struct vhost_net *n, u64 features)
1635 {
1636         int i;
1637
1638         mutex_lock(&n->dev.mutex);
1639         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1640                 mutex_lock(&n->vqs[i].vq.mutex);
1641                 n->vqs[i].vq.acked_backend_features = features;
1642                 mutex_unlock(&n->vqs[i].vq.mutex);
1643         }
1644         mutex_unlock(&n->dev.mutex);
1645
1646         return 0;
1647 }
1648
1649 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1650 {
1651         size_t vhost_hlen, sock_hlen, hdr_len;
1652         int i;
1653
1654         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1655                                (1ULL << VIRTIO_F_VERSION_1))) ?
1656                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1657                         sizeof(struct virtio_net_hdr);
1658         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1659                 /* vhost provides vnet_hdr */
1660                 vhost_hlen = hdr_len;
1661                 sock_hlen = 0;
1662         } else {
1663                 /* socket provides vnet_hdr */
1664                 vhost_hlen = 0;
1665                 sock_hlen = hdr_len;
1666         }
1667         mutex_lock(&n->dev.mutex);
1668         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1669             !vhost_log_access_ok(&n->dev))
1670                 goto out_unlock;
1671
1672         if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1673                 if (vhost_init_device_iotlb(&n->dev, true))
1674                         goto out_unlock;
1675         }
1676
1677         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1678                 mutex_lock(&n->vqs[i].vq.mutex);
1679                 n->vqs[i].vq.acked_features = features;
1680                 n->vqs[i].vhost_hlen = vhost_hlen;
1681                 n->vqs[i].sock_hlen = sock_hlen;
1682                 mutex_unlock(&n->vqs[i].vq.mutex);
1683         }
1684         mutex_unlock(&n->dev.mutex);
1685         return 0;
1686
1687 out_unlock:
1688         mutex_unlock(&n->dev.mutex);
1689         return -EFAULT;
1690 }
1691
1692 static long vhost_net_set_owner(struct vhost_net *n)
1693 {
1694         int r;
1695
1696         mutex_lock(&n->dev.mutex);
1697         if (vhost_dev_has_owner(&n->dev)) {
1698                 r = -EBUSY;
1699                 goto out;
1700         }
1701         r = vhost_net_set_ubuf_info(n);
1702         if (r)
1703                 goto out;
1704         r = vhost_dev_set_owner(&n->dev);
1705         if (r)
1706                 vhost_net_clear_ubuf_info(n);
1707         vhost_net_flush(n);
1708 out:
1709         mutex_unlock(&n->dev.mutex);
1710         return r;
1711 }
1712
1713 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1714                             unsigned long arg)
1715 {
1716         struct vhost_net *n = f->private_data;
1717         void __user *argp = (void __user *)arg;
1718         u64 __user *featurep = argp;
1719         struct vhost_vring_file backend;
1720         u64 features;
1721         int r;
1722
1723         switch (ioctl) {
1724         case VHOST_NET_SET_BACKEND:
1725                 if (copy_from_user(&backend, argp, sizeof backend))
1726                         return -EFAULT;
1727                 return vhost_net_set_backend(n, backend.index, backend.fd);
1728         case VHOST_GET_FEATURES:
1729                 features = VHOST_NET_FEATURES;
1730                 if (copy_to_user(featurep, &features, sizeof features))
1731                         return -EFAULT;
1732                 return 0;
1733         case VHOST_SET_FEATURES:
1734                 if (copy_from_user(&features, featurep, sizeof features))
1735                         return -EFAULT;
1736                 if (features & ~VHOST_NET_FEATURES)
1737                         return -EOPNOTSUPP;
1738                 return vhost_net_set_features(n, features);
1739         case VHOST_GET_BACKEND_FEATURES:
1740                 features = VHOST_NET_BACKEND_FEATURES;
1741                 if (copy_to_user(featurep, &features, sizeof(features)))
1742                         return -EFAULT;
1743                 return 0;
1744         case VHOST_SET_BACKEND_FEATURES:
1745                 if (copy_from_user(&features, featurep, sizeof(features)))
1746                         return -EFAULT;
1747                 if (features & ~VHOST_NET_BACKEND_FEATURES)
1748                         return -EOPNOTSUPP;
1749                 return vhost_net_set_backend_features(n, features);
1750         case VHOST_RESET_OWNER:
1751                 return vhost_net_reset_owner(n);
1752         case VHOST_SET_OWNER:
1753                 return vhost_net_set_owner(n);
1754         default:
1755                 mutex_lock(&n->dev.mutex);
1756                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1757                 if (r == -ENOIOCTLCMD)
1758                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1759                 else
1760                         vhost_net_flush(n);
1761                 mutex_unlock(&n->dev.mutex);
1762                 return r;
1763         }
1764 }
1765
1766 #ifdef CONFIG_COMPAT
1767 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1768                                    unsigned long arg)
1769 {
1770         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1771 }
1772 #endif
1773
1774 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1775 {
1776         struct file *file = iocb->ki_filp;
1777         struct vhost_net *n = file->private_data;
1778         struct vhost_dev *dev = &n->dev;
1779         int noblock = file->f_flags & O_NONBLOCK;
1780
1781         return vhost_chr_read_iter(dev, to, noblock);
1782 }
1783
1784 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1785                                         struct iov_iter *from)
1786 {
1787         struct file *file = iocb->ki_filp;
1788         struct vhost_net *n = file->private_data;
1789         struct vhost_dev *dev = &n->dev;
1790
1791         return vhost_chr_write_iter(dev, from);
1792 }
1793
1794 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1795 {
1796         struct vhost_net *n = file->private_data;
1797         struct vhost_dev *dev = &n->dev;
1798
1799         return vhost_chr_poll(file, dev, wait);
1800 }
1801
1802 static const struct file_operations vhost_net_fops = {
1803         .owner          = THIS_MODULE,
1804         .release        = vhost_net_release,
1805         .read_iter      = vhost_net_chr_read_iter,
1806         .write_iter     = vhost_net_chr_write_iter,
1807         .poll           = vhost_net_chr_poll,
1808         .unlocked_ioctl = vhost_net_ioctl,
1809 #ifdef CONFIG_COMPAT
1810         .compat_ioctl   = vhost_net_compat_ioctl,
1811 #endif
1812         .open           = vhost_net_open,
1813         .llseek         = noop_llseek,
1814 };
1815
1816 static struct miscdevice vhost_net_misc = {
1817         .minor = VHOST_NET_MINOR,
1818         .name = "vhost-net",
1819         .fops = &vhost_net_fops,
1820 };
1821
1822 static int vhost_net_init(void)
1823 {
1824         if (experimental_zcopytx)
1825                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1826         return misc_register(&vhost_net_misc);
1827 }
1828 module_init(vhost_net_init);
1829
1830 static void vhost_net_exit(void)
1831 {
1832         misc_deregister(&vhost_net_misc);
1833 }
1834 module_exit(vhost_net_exit);
1835
1836 MODULE_VERSION("0.0.1");
1837 MODULE_LICENSE("GPL v2");
1838 MODULE_AUTHOR("Michael S. Tsirkin");
1839 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1840 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1841 MODULE_ALIAS("devname:vhost-net");