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