drivers/vhost/net.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (C) 2009 Red Hat, Inc.
   3  * Author: Michael S. Tsirkin <mst@redhat.com>
   4  *
   5  * virtio-net server in host kernel.
   6  */
   7
   8 #include <linux/compat.h>
   9 #include <linux/eventfd.h>
  10 #include <linux/vhost.h>
  11 #include <linux/virtio_net.h>
  12 #include <linux/miscdevice.h>
  13 #include <linux/module.h>
  14 #include <linux/moduleparam.h>
  15 #include <linux/mutex.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19 #include <linux/sched/clock.h>
  20 #include <linux/sched/signal.h>
  21 #include <linux/vmalloc.h>
  22
  23 #include <linux/net.h>
  24 #include <linux/if_packet.h>
  25 #include <linux/if_arp.h>
  26 #include <linux/if_tun.h>
  27 #include <linux/if_macvlan.h>
  28 #include <linux/if_tap.h>
  29 #include <linux/if_vlan.h>
  30 #include <linux/skb_array.h>
  31 #include <linux/skbuff.h>
  32
  33 #include <net/sock.h>
  34 #include <net/xdp.h>
  35
  36 #include "vhost.h"
  37
  38 static int experimental_zcopytx = 0;
  39 module_param(experimental_zcopytx, int, 0444);
  40 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  41                                        " 1 -Enable; 0 - Disable");
  42
  43 /* Max number of bytes transferred before requeueing the job.
  44  * Using this limit prevents one virtqueue from starving others. */
  45 #define VHOST_NET_WEIGHT 0x80000
  46
  47 /* Max number of packets transferred before requeueing the job.
  48  * Using this limit prevents one virtqueue from starving others with small
  49  * pkts.
  50  */
  51 #define VHOST_NET_PKT_WEIGHT 256
  52
  53 /* MAX number of TX used buffers for outstanding zerocopy */
  54 #define VHOST_MAX_PEND 128
  55 #define VHOST_GOODCOPY_LEN 256
  56
  57 /*
  58  * For transmit, used buffer len is unused; we override it to track buffer
  59  * status internally; used for zerocopy tx only.
  60  */
  61 /* Lower device DMA failed */
  62 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  63 /* Lower device DMA done */
  64 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  65 /* Lower device DMA in progress */
  66 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  67 /* Buffer unused */
  68 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  69
  70 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  71
  72 enum {
  73         VHOST_NET_FEATURES = VHOST_FEATURES |
  74                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  75                          (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  76                          (1ULL << VIRTIO_F_ACCESS_PLATFORM)
  77 };
  78
  79 enum {
  80         VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
  81 };
  82
  83 enum {
  84         VHOST_NET_VQ_RX = 0,
  85         VHOST_NET_VQ_TX = 1,
  86         VHOST_NET_VQ_MAX = 2,
  87 };
  88
  89 struct vhost_net_ubuf_ref {
  90         /* refcount follows semantics similar to kref:
  91          *  0: object is released
  92          *  1: no outstanding ubufs
  93          * >1: outstanding ubufs
  94          */
  95         atomic_t refcount;
  96         wait_queue_head_t wait;
  97         struct vhost_virtqueue *vq;
  98 };
  99
 100 #define VHOST_NET_BATCH 64
 101 struct vhost_net_buf {
 102         void **queue;
 103         int tail;
 104         int head;
 105 };
 106
 107 struct vhost_net_virtqueue {
 108         struct vhost_virtqueue vq;
 109         size_t vhost_hlen;
 110         size_t sock_hlen;
 111         /* vhost zerocopy support fields below: */
 112         /* last used idx for outstanding DMA zerocopy buffers */
 113         int upend_idx;
 114         /* For TX, first used idx for DMA done zerocopy buffers
 115          * For RX, number of batched heads
 116          */
 117         int done_idx;
 118         /* Number of XDP frames batched */
 119         int batched_xdp;
 120         /* an array of userspace buffers info */
 121         struct ubuf_info_msgzc *ubuf_info;
 122         /* Reference counting for outstanding ubufs.
 123          * Protected by vq mutex. Writers must also take device mutex. */
 124         struct vhost_net_ubuf_ref *ubufs;
 125         struct ptr_ring *rx_ring;
 126         struct vhost_net_buf rxq;
 127         /* Batched XDP buffs */
 128         struct xdp_buff *xdp;
 129 };
 130
 131 struct vhost_net {
 132         struct vhost_dev dev;
 133         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 134         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 135         /* Number of TX recently submitted.
 136          * Protected by tx vq lock. */
 137         unsigned tx_packets;
 138         /* Number of times zerocopy TX recently failed.
 139          * Protected by tx vq lock. */
 140         unsigned tx_zcopy_err;
 141         /* Flush in progress. Protected by tx vq lock. */
 142         bool tx_flush;
 143         /* Private page frag */
 144         struct page_frag page_frag;
 145         /* Refcount bias of page frag */
 146         int refcnt_bias;
 147 };
 148
 149 static unsigned vhost_net_zcopy_mask __read_mostly;
 150
 151 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 152 {
 153         if (rxq->tail != rxq->head)
 154                 return rxq->queue[rxq->head];
 155         else
 156                 return NULL;
 157 }
 158
 159 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 160 {
 161         return rxq->tail - rxq->head;
 162 }
 163
 164 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 165 {
 166         return rxq->tail == rxq->head;
 167 }
 168
 169 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 170 {
 171         void *ret = vhost_net_buf_get_ptr(rxq);
 172         ++rxq->head;
 173         return ret;
 174 }
 175
 176 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 177 {
 178         struct vhost_net_buf *rxq = &nvq->rxq;
 179
 180         rxq->head = 0;
 181         rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
 182                                               VHOST_NET_BATCH);
 183         return rxq->tail;
 184 }
 185
 186 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 187 {
 188         struct vhost_net_buf *rxq = &nvq->rxq;
 189
 190         if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
 191                 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 192                                    vhost_net_buf_get_size(rxq),
 193                                    tun_ptr_free);
 194                 rxq->head = rxq->tail = 0;
 195         }
 196 }
 197
 198 static int vhost_net_buf_peek_len(void *ptr)
 199 {
 200         if (tun_is_xdp_frame(ptr)) {
 201                 struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
 202
 203                 return xdpf->len;
 204         }
 205
 206         return __skb_array_len_with_tag(ptr);
 207 }
 208
 209 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 210 {
 211         struct vhost_net_buf *rxq = &nvq->rxq;
 212
 213         if (!vhost_net_buf_is_empty(rxq))
 214                 goto out;
 215
 216         if (!vhost_net_buf_produce(nvq))
 217                 return 0;
 218
 219 out:
 220         return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 221 }
 222
 223 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 224 {
 225         rxq->head = rxq->tail = 0;
 226 }
 227
 228 static void vhost_net_enable_zcopy(int vq)
 229 {
 230         vhost_net_zcopy_mask |= 0x1 << vq;
 231 }
 232
 233 static struct vhost_net_ubuf_ref *
 234 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 235 {
 236         struct vhost_net_ubuf_ref *ubufs;
 237         /* No zero copy backend? Nothing to count. */
 238         if (!zcopy)
 239                 return NULL;
 240         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 241         if (!ubufs)
 242                 return ERR_PTR(-ENOMEM);
 243         atomic_set(&ubufs->refcount, 1);
 244         init_waitqueue_head(&ubufs->wait);
 245         ubufs->vq = vq;
 246         return ubufs;
 247 }
 248
 249 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 250 {
 251         int r = atomic_sub_return(1, &ubufs->refcount);
 252         if (unlikely(!r))
 253                 wake_up(&ubufs->wait);
 254         return r;
 255 }
 256
 257 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 258 {
 259         vhost_net_ubuf_put(ubufs);
 260         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 261 }
 262
 263 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 264 {
 265         vhost_net_ubuf_put_and_wait(ubufs);
 266         kfree(ubufs);
 267 }
 268
 269 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 270 {
 271         int i;
 272
 273         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 274                 kfree(n->vqs[i].ubuf_info);
 275                 n->vqs[i].ubuf_info = NULL;
 276         }
 277 }
 278
 279 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 280 {
 281         bool zcopy;
 282         int i;
 283
 284         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 285                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 286                 if (!zcopy)
 287                         continue;
 288                 n->vqs[i].ubuf_info =
 289                         kmalloc_array(UIO_MAXIOV,
 290                                       sizeof(*n->vqs[i].ubuf_info),
 291                                       GFP_KERNEL);
 292                 if  (!n->vqs[i].ubuf_info)
 293                         goto err;
 294         }
 295         return 0;
 296
 297 err:
 298         vhost_net_clear_ubuf_info(n);
 299         return -ENOMEM;
 300 }
 301
 302 static void vhost_net_vq_reset(struct vhost_net *n)
 303 {
 304         int i;
 305
 306         vhost_net_clear_ubuf_info(n);
 307
 308         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 309                 n->vqs[i].done_idx = 0;
 310                 n->vqs[i].upend_idx = 0;
 311                 n->vqs[i].ubufs = NULL;
 312                 n->vqs[i].vhost_hlen = 0;
 313                 n->vqs[i].sock_hlen = 0;
 314                 vhost_net_buf_init(&n->vqs[i].rxq);
 315         }
 316
 317 }
 318
 319 static void vhost_net_tx_packet(struct vhost_net *net)
 320 {
 321         ++net->tx_packets;
 322         if (net->tx_packets < 1024)
 323                 return;
 324         net->tx_packets = 0;
 325         net->tx_zcopy_err = 0;
 326 }
 327
 328 static void vhost_net_tx_err(struct vhost_net *net)
 329 {
 330         ++net->tx_zcopy_err;
 331 }
 332
 333 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 334 {
 335         /* TX flush waits for outstanding DMAs to be done.
 336          * Don't start new DMAs.
 337          */
 338         return !net->tx_flush &&
 339                 net->tx_packets / 64 >= net->tx_zcopy_err;
 340 }
 341
 342 static bool vhost_sock_zcopy(struct socket *sock)
 343 {
 344         return unlikely(experimental_zcopytx) &&
 345                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 346 }
 347
 348 static bool vhost_sock_xdp(struct socket *sock)
 349 {
 350         return sock_flag(sock->sk, SOCK_XDP);
 351 }
 352
 353 /* In case of DMA done not in order in lower device driver for some reason.
 354  * upend_idx is used to track end of used idx, done_idx is used to track head
 355  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 356  * guest used idx.
 357  */
 358 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 359                                        struct vhost_virtqueue *vq)
 360 {
 361         struct vhost_net_virtqueue *nvq =
 362                 container_of(vq, struct vhost_net_virtqueue, vq);
 363         int i, add;
 364         int j = 0;
 365
 366         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 367                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 368                         vhost_net_tx_err(net);
 369                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 370                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 371                         ++j;
 372                 } else
 373                         break;
 374         }
 375         while (j) {
 376                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 377                 vhost_add_used_and_signal_n(vq->dev, vq,
 378                                             &vq->heads[nvq->done_idx], add);
 379                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 380                 j -= add;
 381         }
 382 }
 383
 384 static void vhost_zerocopy_callback(struct sk_buff *skb,
 385                                     struct ubuf_info *ubuf_base, bool success)
 386 {
 387         struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
 388         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 389         struct vhost_virtqueue *vq = ubufs->vq;
 390         int cnt;
 391
 392         rcu_read_lock_bh();
 393
 394         /* set len to mark this desc buffers done DMA */
 395         vq->heads[ubuf->desc].len = success ?
 396                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 397         cnt = vhost_net_ubuf_put(ubufs);
 398
 399         /*
 400          * Trigger polling thread if guest stopped submitting new buffers:
 401          * in this case, the refcount after decrement will eventually reach 1.
 402          * We also trigger polling periodically after each 16 packets
 403          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 404          * less than 10% of times).
 405          */
 406         if (cnt <= 1 || !(cnt % 16))
 407                 vhost_poll_queue(&vq->poll);
 408
 409         rcu_read_unlock_bh();
 410 }
 411
 412 static inline unsigned long busy_clock(void)
 413 {
 414         return local_clock() >> 10;
 415 }
 416
 417 static bool vhost_can_busy_poll(unsigned long endtime)
 418 {
 419         return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
 420                       !signal_pending(current));
 421 }
 422
 423 static void vhost_net_disable_vq(struct vhost_net *n,
 424                                  struct vhost_virtqueue *vq)
 425 {
 426         struct vhost_net_virtqueue *nvq =
 427                 container_of(vq, struct vhost_net_virtqueue, vq);
 428         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 429         if (!vhost_vq_get_backend(vq))
 430                 return;
 431         vhost_poll_stop(poll);
 432 }
 433
 434 static int vhost_net_enable_vq(struct vhost_net *n,
 435                                 struct vhost_virtqueue *vq)
 436 {
 437         struct vhost_net_virtqueue *nvq =
 438                 container_of(vq, struct vhost_net_virtqueue, vq);
 439         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 440         struct socket *sock;
 441
 442         sock = vhost_vq_get_backend(vq);
 443         if (!sock)
 444                 return 0;
 445
 446         return vhost_poll_start(poll, sock->file);
 447 }
 448
 449 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
 450 {
 451         struct vhost_virtqueue *vq = &nvq->vq;
 452         struct vhost_dev *dev = vq->dev;
 453
 454         if (!nvq->done_idx)
 455                 return;
 456
 457         vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
 458         nvq->done_idx = 0;
 459 }
 460
 461 static void vhost_tx_batch(struct vhost_net *net,
 462                            struct vhost_net_virtqueue *nvq,
 463                            struct socket *sock,
 464                            struct msghdr *msghdr)
 465 {
 466         struct tun_msg_ctl ctl = {
 467                 .type = TUN_MSG_PTR,
 468                 .num = nvq->batched_xdp,
 469                 .ptr = nvq->xdp,
 470         };
 471         int i, err;
 472
 473         if (nvq->batched_xdp == 0)
 474                 goto signal_used;
 475
 476         msghdr->msg_control = &ctl;
 477         msghdr->msg_controllen = sizeof(ctl);
 478         err = sock->ops->sendmsg(sock, msghdr, 0);
 479         if (unlikely(err < 0)) {
 480                 vq_err(&nvq->vq, "Fail to batch sending packets\n");
 481
 482                 /* free pages owned by XDP; since this is an unlikely error path,
 483                  * keep it simple and avoid more complex bulk update for the
 484                  * used pages
 485                  */
 486                 for (i = 0; i < nvq->batched_xdp; ++i)
 487                         put_page(virt_to_head_page(nvq->xdp[i].data));
 488                 nvq->batched_xdp = 0;
 489                 nvq->done_idx = 0;
 490                 return;
 491         }
 492
 493 signal_used:
 494         vhost_net_signal_used(nvq);
 495         nvq->batched_xdp = 0;
 496 }
 497
 498 static int sock_has_rx_data(struct socket *sock)
 499 {
 500         if (unlikely(!sock))
 501                 return 0;
 502
 503         if (sock->ops->peek_len)
 504                 return sock->ops->peek_len(sock);
 505
 506         return skb_queue_empty(&sock->sk->sk_receive_queue);
 507 }
 508
 509 static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
 510                                           struct vhost_virtqueue *vq)
 511 {
 512         if (!vhost_vq_avail_empty(&net->dev, vq)) {
 513                 vhost_poll_queue(&vq->poll);
 514         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 515                 vhost_disable_notify(&net->dev, vq);
 516                 vhost_poll_queue(&vq->poll);
 517         }
 518 }
 519
 520 static void vhost_net_busy_poll(struct vhost_net *net,
 521                                 struct vhost_virtqueue *rvq,
 522                                 struct vhost_virtqueue *tvq,
 523                                 bool *busyloop_intr,
 524                                 bool poll_rx)
 525 {
 526         unsigned long busyloop_timeout;
 527         unsigned long endtime;
 528         struct socket *sock;
 529         struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
 530
 531         /* Try to hold the vq mutex of the paired virtqueue. We can't
 532          * use mutex_lock() here since we could not guarantee a
 533          * consistenet lock ordering.
 534          */
 535         if (!mutex_trylock(&vq->mutex))
 536                 return;
 537
 538         vhost_disable_notify(&net->dev, vq);
 539         sock = vhost_vq_get_backend(rvq);
 540
 541         busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
 542                                      tvq->busyloop_timeout;
 543
 544         preempt_disable();
 545         endtime = busy_clock() + busyloop_timeout;
 546
 547         while (vhost_can_busy_poll(endtime)) {
 548                 if (vhost_has_work(&net->dev)) {
 549                         *busyloop_intr = true;
 550                         break;
 551                 }
 552
 553                 if ((sock_has_rx_data(sock) &&
 554                      !vhost_vq_avail_empty(&net->dev, rvq)) ||
 555                     !vhost_vq_avail_empty(&net->dev, tvq))
 556                         break;
 557
 558                 cpu_relax();
 559         }
 560
 561         preempt_enable();
 562
 563         if (poll_rx || sock_has_rx_data(sock))
 564                 vhost_net_busy_poll_try_queue(net, vq);
 565         else if (!poll_rx) /* On tx here, sock has no rx data. */
 566                 vhost_enable_notify(&net->dev, rvq);
 567
 568         mutex_unlock(&vq->mutex);
 569 }
 570
 571 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 572                                     struct vhost_net_virtqueue *tnvq,
 573                                     unsigned int *out_num, unsigned int *in_num,
 574                                     struct msghdr *msghdr, bool *busyloop_intr)
 575 {
 576         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 577         struct vhost_virtqueue *rvq = &rnvq->vq;
 578         struct vhost_virtqueue *tvq = &tnvq->vq;
 579
 580         int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 581                                   out_num, in_num, NULL, NULL);
 582
 583         if (r == tvq->num && tvq->busyloop_timeout) {
 584                 /* Flush batched packets first */
 585                 if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq)))
 586                         vhost_tx_batch(net, tnvq,
 587                                        vhost_vq_get_backend(tvq),
 588                                        msghdr);
 589
 590                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
 591
 592                 r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 593                                       out_num, in_num, NULL, NULL);
 594         }
 595
 596         return r;
 597 }
 598
 599 static bool vhost_exceeds_maxpend(struct vhost_net *net)
 600 {
 601         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 602         struct vhost_virtqueue *vq = &nvq->vq;
 603
 604         return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 605                min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 606 }
 607
 608 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
 609                             size_t hdr_size, int out)
 610 {
 611         /* Skip header. TODO: support TSO. */
 612         size_t len = iov_length(vq->iov, out);
 613
 614         iov_iter_init(iter, WRITE, vq->iov, out, len);
 615         iov_iter_advance(iter, hdr_size);
 616
 617         return iov_iter_count(iter);
 618 }
 619
 620 static int get_tx_bufs(struct vhost_net *net,
 621                        struct vhost_net_virtqueue *nvq,
 622                        struct msghdr *msg,
 623                        unsigned int *out, unsigned int *in,
 624                        size_t *len, bool *busyloop_intr)
 625 {
 626         struct vhost_virtqueue *vq = &nvq->vq;
 627         int ret;
 628
 629         ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
 630
 631         if (ret < 0 || ret == vq->num)
 632                 return ret;
 633
 634         if (*in) {
 635                 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 636                         *out, *in);
 637                 return -EFAULT;
 638         }
 639
 640         /* Sanity check */
 641         *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 642         if (*len == 0) {
 643                 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 644                         *len, nvq->vhost_hlen);
 645                 return -EFAULT;
 646         }
 647
 648         return ret;
 649 }
 650
 651 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 652 {
 653         return total_len < VHOST_NET_WEIGHT &&
 654                !vhost_vq_avail_empty(vq->dev, vq);
 655 }
 656
 657 static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
 658                                        struct page_frag *pfrag, gfp_t gfp)
 659 {
 660         if (pfrag->page) {
 661                 if (pfrag->offset + sz <= pfrag->size)
 662                         return true;
 663                 __page_frag_cache_drain(pfrag->page, net->refcnt_bias);
 664         }
 665
 666         pfrag->offset = 0;
 667         net->refcnt_bias = 0;
 668         if (SKB_FRAG_PAGE_ORDER) {
 669                 /* Avoid direct reclaim but allow kswapd to wake */
 670                 pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
 671                                           __GFP_COMP | __GFP_NOWARN |
 672                                           __GFP_NORETRY,
 673                                           SKB_FRAG_PAGE_ORDER);
 674                 if (likely(pfrag->page)) {
 675                         pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
 676                         goto done;
 677                 }
 678         }
 679         pfrag->page = alloc_page(gfp);
 680         if (likely(pfrag->page)) {
 681                 pfrag->size = PAGE_SIZE;
 682                 goto done;
 683         }
 684         return false;
 685
 686 done:
 687         net->refcnt_bias = USHRT_MAX;
 688         page_ref_add(pfrag->page, USHRT_MAX - 1);
 689         return true;
 690 }
 691
 692 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 693
 694 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
 695                                struct iov_iter *from)
 696 {
 697         struct vhost_virtqueue *vq = &nvq->vq;
 698         struct vhost_net *net = container_of(vq->dev, struct vhost_net,
 699                                              dev);
 700         struct socket *sock = vhost_vq_get_backend(vq);
 701         struct page_frag *alloc_frag = &net->page_frag;
 702         struct virtio_net_hdr *gso;
 703         struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
 704         struct tun_xdp_hdr *hdr;
 705         size_t len = iov_iter_count(from);
 706         int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
 707         int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 708         int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
 709         int sock_hlen = nvq->sock_hlen;
 710         void *buf;
 711         int copied;
 712
 713         if (unlikely(len < nvq->sock_hlen))
 714                 return -EFAULT;
 715
 716         if (SKB_DATA_ALIGN(len + pad) +
 717             SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
 718                 return -ENOSPC;
 719
 720         buflen += SKB_DATA_ALIGN(len + pad);
 721         alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
 722         if (unlikely(!vhost_net_page_frag_refill(net, buflen,
 723                                                  alloc_frag, GFP_KERNEL)))
 724                 return -ENOMEM;
 725
 726         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 727         copied = copy_page_from_iter(alloc_frag->page,
 728                                      alloc_frag->offset +
 729                                      offsetof(struct tun_xdp_hdr, gso),
 730                                      sock_hlen, from);
 731         if (copied != sock_hlen)
 732                 return -EFAULT;
 733
 734         hdr = buf;
 735         gso = &hdr->gso;
 736
 737         if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
 738             vhost16_to_cpu(vq, gso->csum_start) +
 739             vhost16_to_cpu(vq, gso->csum_offset) + 2 >
 740             vhost16_to_cpu(vq, gso->hdr_len)) {
 741                 gso->hdr_len = cpu_to_vhost16(vq,
 742                                vhost16_to_cpu(vq, gso->csum_start) +
 743                                vhost16_to_cpu(vq, gso->csum_offset) + 2);
 744
 745                 if (vhost16_to_cpu(vq, gso->hdr_len) > len)
 746                         return -EINVAL;
 747         }
 748
 749         len -= sock_hlen;
 750         copied = copy_page_from_iter(alloc_frag->page,
 751                                      alloc_frag->offset + pad,
 752                                      len, from);
 753         if (copied != len)
 754                 return -EFAULT;
 755
 756         xdp_init_buff(xdp, buflen, NULL);
 757         xdp_prepare_buff(xdp, buf, pad, len, true);
 758         hdr->buflen = buflen;
 759
 760         --net->refcnt_bias;
 761         alloc_frag->offset += buflen;
 762
 763         ++nvq->batched_xdp;
 764
 765         return 0;
 766 }
 767
 768 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 769 {
 770         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 771         struct vhost_virtqueue *vq = &nvq->vq;
 772         unsigned out, in;
 773         int head;
 774         struct msghdr msg = {
 775                 .msg_name = NULL,
 776                 .msg_namelen = 0,
 777                 .msg_control = NULL,
 778                 .msg_controllen = 0,
 779                 .msg_flags = MSG_DONTWAIT,
 780         };
 781         size_t len, total_len = 0;
 782         int err;
 783         int sent_pkts = 0;
 784         bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
 785
 786         do {
 787                 bool busyloop_intr = false;
 788
 789                 if (nvq->done_idx == VHOST_NET_BATCH)
 790                         vhost_tx_batch(net, nvq, sock, &msg);
 791
 792                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 793                                    &busyloop_intr);
 794                 /* On error, stop handling until the next kick. */
 795                 if (unlikely(head < 0))
 796                         break;
 797                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 798                 if (head == vq->num) {
 799                         if (unlikely(busyloop_intr)) {
 800                                 vhost_poll_queue(&vq->poll);
 801                         } else if (unlikely(vhost_enable_notify(&net->dev,
 802                                                                 vq))) {
 803                                 vhost_disable_notify(&net->dev, vq);
 804                                 continue;
 805                         }
 806                         break;
 807                 }
 808
 809                 total_len += len;
 810
 811                 /* For simplicity, TX batching is only enabled if
 812                  * sndbuf is unlimited.
 813                  */
 814                 if (sock_can_batch) {
 815                         err = vhost_net_build_xdp(nvq, &msg.msg_iter);
 816                         if (!err) {
 817                                 goto done;
 818                         } else if (unlikely(err != -ENOSPC)) {
 819                                 vhost_tx_batch(net, nvq, sock, &msg);
 820                                 vhost_discard_vq_desc(vq, 1);
 821                                 vhost_net_enable_vq(net, vq);
 822                                 break;
 823                         }
 824
 825                         /* We can't build XDP buff, go for single
 826                          * packet path but let's flush batched
 827                          * packets.
 828                          */
 829                         vhost_tx_batch(net, nvq, sock, &msg);
 830                         msg.msg_control = NULL;
 831                 } else {
 832                         if (tx_can_batch(vq, total_len))
 833                                 msg.msg_flags |= MSG_MORE;
 834                         else
 835                                 msg.msg_flags &= ~MSG_MORE;
 836                 }
 837
 838                 err = sock->ops->sendmsg(sock, &msg, len);
 839                 if (unlikely(err < 0)) {
 840                         if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) {
 841                                 vhost_discard_vq_desc(vq, 1);
 842                                 vhost_net_enable_vq(net, vq);
 843                                 break;
 844                         }
 845                         pr_debug("Fail to send packet: err %d", err);
 846                 } else if (unlikely(err != len))
 847                         pr_debug("Truncated TX packet: len %d != %zd\n",
 848                                  err, len);
 849 done:
 850                 vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 851                 vq->heads[nvq->done_idx].len = 0;
 852                 ++nvq->done_idx;
 853         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 854
 855         vhost_tx_batch(net, nvq, sock, &msg);
 856 }
 857
 858 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 859 {
 860         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 861         struct vhost_virtqueue *vq = &nvq->vq;
 862         unsigned out, in;
 863         int head;
 864         struct msghdr msg = {
 865                 .msg_name = NULL,
 866                 .msg_namelen = 0,
 867                 .msg_control = NULL,
 868                 .msg_controllen = 0,
 869                 .msg_flags = MSG_DONTWAIT,
 870         };
 871         struct tun_msg_ctl ctl;
 872         size_t len, total_len = 0;
 873         int err;
 874         struct vhost_net_ubuf_ref *ubufs;
 875         struct ubuf_info_msgzc *ubuf;
 876         bool zcopy_used;
 877         int sent_pkts = 0;
 878
 879         do {
 880                 bool busyloop_intr;
 881
 882                 /* Release DMAs done buffers first */
 883                 vhost_zerocopy_signal_used(net, vq);
 884
 885                 busyloop_intr = false;
 886                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 887                                    &busyloop_intr);
 888                 /* On error, stop handling until the next kick. */
 889                 if (unlikely(head < 0))
 890                         break;
 891                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 892                 if (head == vq->num) {
 893                         if (unlikely(busyloop_intr)) {
 894                                 vhost_poll_queue(&vq->poll);
 895                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 896                                 vhost_disable_notify(&net->dev, vq);
 897                                 continue;
 898                         }
 899                         break;
 900                 }
 901
 902                 zcopy_used = len >= VHOST_GOODCOPY_LEN
 903                              && !vhost_exceeds_maxpend(net)
 904                              && vhost_net_tx_select_zcopy(net);
 905
 906                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 907                 if (zcopy_used) {
 908                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 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->ctx = nvq->ubufs;
 912                         ubuf->desc = nvq->upend_idx;
 913                         ubuf->ubuf.callback = vhost_zerocopy_callback;
 914                         ubuf->ubuf.flags = SKBFL_ZEROCOPY_FRAG;
 915                         refcount_set(&ubuf->ubuf.refcnt, 1);
 916                         msg.msg_control = &ctl;
 917                         ctl.type = TUN_MSG_UBUF;
 918                         ctl.ptr = &ubuf->ubuf;
 919                         msg.msg_controllen = sizeof(ctl);
 920                         ubufs = nvq->ubufs;
 921                         atomic_inc(&ubufs->refcount);
 922                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 923                 } else {
 924                         msg.msg_control = NULL;
 925                         ubufs = NULL;
 926                 }
 927                 total_len += len;
 928                 if (tx_can_batch(vq, total_len) &&
 929                     likely(!vhost_exceeds_maxpend(net))) {
 930                         msg.msg_flags |= MSG_MORE;
 931                 } else {
 932                         msg.msg_flags &= ~MSG_MORE;
 933                 }
 934
 935                 err = sock->ops->sendmsg(sock, &msg, len);
 936                 if (unlikely(err < 0)) {
 937                         if (zcopy_used) {
 938                                 if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
 939                                         vhost_net_ubuf_put(ubufs);
 940                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 941                                         % UIO_MAXIOV;
 942                         }
 943                         if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) {
 944                                 vhost_discard_vq_desc(vq, 1);
 945                                 vhost_net_enable_vq(net, vq);
 946                                 break;
 947                         }
 948                         pr_debug("Fail to send packet: err %d", err);
 949                 } else if (unlikely(err != len))
 950                         pr_debug("Truncated TX packet: "
 951                                  " len %d != %zd\n", err, len);
 952                 if (!zcopy_used)
 953                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 954                 else
 955                         vhost_zerocopy_signal_used(net, vq);
 956                 vhost_net_tx_packet(net);
 957         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 958 }
 959
 960 /* Expects to be always run from workqueue - which acts as
 961  * read-size critical section for our kind of RCU. */
 962 static void handle_tx(struct vhost_net *net)
 963 {
 964         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 965         struct vhost_virtqueue *vq = &nvq->vq;
 966         struct socket *sock;
 967
 968         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
 969         sock = vhost_vq_get_backend(vq);
 970         if (!sock)
 971                 goto out;
 972
 973         if (!vq_meta_prefetch(vq))
 974                 goto out;
 975
 976         vhost_disable_notify(&net->dev, vq);
 977         vhost_net_disable_vq(net, vq);
 978
 979         if (vhost_sock_zcopy(sock))
 980                 handle_tx_zerocopy(net, sock);
 981         else
 982                 handle_tx_copy(net, sock);
 983
 984 out:
 985         mutex_unlock(&vq->mutex);
 986 }
 987
 988 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 989 {
 990         struct sk_buff *head;
 991         int len = 0;
 992         unsigned long flags;
 993
 994         if (rvq->rx_ring)
 995                 return vhost_net_buf_peek(rvq);
 996
 997         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 998         head = skb_peek(&sk->sk_receive_queue);
 999         if (likely(head)) {
1000                 len = head->len;
1001                 if (skb_vlan_tag_present(head))
1002                         len += VLAN_HLEN;
1003         }
1004
1005         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
1006         return len;
1007 }
1008
1009 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
1010                                       bool *busyloop_intr)
1011 {
1012         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1013         struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1014         struct vhost_virtqueue *rvq = &rnvq->vq;
1015         struct vhost_virtqueue *tvq = &tnvq->vq;
1016         int len = peek_head_len(rnvq, sk);
1017
1018         if (!len && rvq->busyloop_timeout) {
1019                 /* Flush batched heads first */
1020                 vhost_net_signal_used(rnvq);
1021                 /* Both tx vq and rx socket were polled here */
1022                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1023
1024                 len = peek_head_len(rnvq, sk);
1025         }
1026
1027         return len;
1028 }
1029
1030 /* This is a multi-buffer version of vhost_get_desc, that works if
1031  *      vq has read descriptors only.
1032  * @vq          - the relevant virtqueue
1033  * @datalen     - data length we'll be reading
1034  * @iovcount    - returned count of io vectors we fill
1035  * @log         - vhost log
1036  * @log_num     - log offset
1037  * @quota       - headcount quota, 1 for big buffer
1038  *      returns number of buffer heads allocated, negative on error
1039  */
1040 static int get_rx_bufs(struct vhost_virtqueue *vq,
1041                        struct vring_used_elem *heads,
1042                        int datalen,
1043                        unsigned *iovcount,
1044                        struct vhost_log *log,
1045                        unsigned *log_num,
1046                        unsigned int quota)
1047 {
1048         unsigned int out, in;
1049         int seg = 0;
1050         int headcount = 0;
1051         unsigned d;
1052         int r, nlogs = 0;
1053         /* len is always initialized before use since we are always called with
1054          * datalen > 0.
1055          */
1056         u32 len;
1057
1058         while (datalen > 0 && headcount < quota) {
1059                 if (unlikely(seg >= UIO_MAXIOV)) {
1060                         r = -ENOBUFS;
1061                         goto err;
1062                 }
1063                 r = vhost_get_vq_desc(vq, vq->iov + seg,
1064                                       ARRAY_SIZE(vq->iov) - seg, &out,
1065                                       &in, log, log_num);
1066                 if (unlikely(r < 0))
1067                         goto err;
1068
1069                 d = r;
1070                 if (d == vq->num) {
1071                         r = 0;
1072                         goto err;
1073                 }
1074                 if (unlikely(out || in <= 0)) {
1075                         vq_err(vq, "unexpected descriptor format for RX: "
1076                                 "out %d, in %d\n", out, in);
1077                         r = -EINVAL;
1078                         goto err;
1079                 }
1080                 if (unlikely(log)) {
1081                         nlogs += *log_num;
1082                         log += *log_num;
1083                 }
1084                 heads[headcount].id = cpu_to_vhost32(vq, d);
1085                 len = iov_length(vq->iov + seg, in);
1086                 heads[headcount].len = cpu_to_vhost32(vq, len);
1087                 datalen -= len;
1088                 ++headcount;
1089                 seg += in;
1090         }
1091         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1092         *iovcount = seg;
1093         if (unlikely(log))
1094                 *log_num = nlogs;
1095
1096         /* Detect overrun */
1097         if (unlikely(datalen > 0)) {
1098                 r = UIO_MAXIOV + 1;
1099                 goto err;
1100         }
1101         return headcount;
1102 err:
1103         vhost_discard_vq_desc(vq, headcount);
1104         return r;
1105 }
1106
1107 /* Expects to be always run from workqueue - which acts as
1108  * read-size critical section for our kind of RCU. */
1109 static void handle_rx(struct vhost_net *net)
1110 {
1111         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1112         struct vhost_virtqueue *vq = &nvq->vq;
1113         unsigned in, log;
1114         struct vhost_log *vq_log;
1115         struct msghdr msg = {
1116                 .msg_name = NULL,
1117                 .msg_namelen = 0,
1118                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
1119                 .msg_controllen = 0,
1120                 .msg_flags = MSG_DONTWAIT,
1121         };
1122         struct virtio_net_hdr hdr = {
1123                 .flags = 0,
1124                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
1125         };
1126         size_t total_len = 0;
1127         int err, mergeable;
1128         s16 headcount;
1129         size_t vhost_hlen, sock_hlen;
1130         size_t vhost_len, sock_len;
1131         bool busyloop_intr = false;
1132         struct socket *sock;
1133         struct iov_iter fixup;
1134         __virtio16 num_buffers;
1135         int recv_pkts = 0;
1136
1137         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1138         sock = vhost_vq_get_backend(vq);
1139         if (!sock)
1140                 goto out;
1141
1142         if (!vq_meta_prefetch(vq))
1143                 goto out;
1144
1145         vhost_disable_notify(&net->dev, vq);
1146         vhost_net_disable_vq(net, vq);
1147
1148         vhost_hlen = nvq->vhost_hlen;
1149         sock_hlen = nvq->sock_hlen;
1150
1151         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1152                 vq->log : NULL;
1153         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1154
1155         do {
1156                 sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1157                                                       &busyloop_intr);
1158                 if (!sock_len)
1159                         break;
1160                 sock_len += sock_hlen;
1161                 vhost_len = sock_len + vhost_hlen;
1162                 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1163                                         vhost_len, &in, vq_log, &log,
1164                                         likely(mergeable) ? UIO_MAXIOV : 1);
1165                 /* On error, stop handling until the next kick. */
1166                 if (unlikely(headcount < 0))
1167                         goto out;
1168                 /* OK, now we need to know about added descriptors. */
1169                 if (!headcount) {
1170                         if (unlikely(busyloop_intr)) {
1171                                 vhost_poll_queue(&vq->poll);
1172                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1173                                 /* They have slipped one in as we were
1174                                  * doing that: check again. */
1175                                 vhost_disable_notify(&net->dev, vq);
1176                                 continue;
1177                         }
1178                         /* Nothing new?  Wait for eventfd to tell us
1179                          * they refilled. */
1180                         goto out;
1181                 }
1182                 busyloop_intr = false;
1183                 if (nvq->rx_ring)
1184                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1185                 /* On overrun, truncate and discard */
1186                 if (unlikely(headcount > UIO_MAXIOV)) {
1187                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
1188                         err = sock->ops->recvmsg(sock, &msg,
1189                                                  1, MSG_DONTWAIT | MSG_TRUNC);
1190                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
1191                         continue;
1192                 }
1193                 /* We don't need to be notified again. */
1194                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
1195                 fixup = msg.msg_iter;
1196                 if (unlikely((vhost_hlen))) {
1197                         /* We will supply the header ourselves
1198                          * TODO: support TSO.
1199                          */
1200                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
1201                 }
1202                 err = sock->ops->recvmsg(sock, &msg,
1203                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
1204                 /* Userspace might have consumed the packet meanwhile:
1205                  * it's not supposed to do this usually, but might be hard
1206                  * to prevent. Discard data we got (if any) and keep going. */
1207                 if (unlikely(err != sock_len)) {
1208                         pr_debug("Discarded rx packet: "
1209                                  " len %d, expected %zd\n", err, sock_len);
1210                         vhost_discard_vq_desc(vq, headcount);
1211                         continue;
1212                 }
1213                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1214                 if (unlikely(vhost_hlen)) {
1215                         if (copy_to_iter(&hdr, sizeof(hdr),
1216                                          &fixup) != sizeof(hdr)) {
1217                                 vq_err(vq, "Unable to write vnet_hdr "
1218                                        "at addr %p\n", vq->iov->iov_base);
1219                                 goto out;
1220                         }
1221                 } else {
1222                         /* Header came from socket; we'll need to patch
1223                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1224                          */
1225                         iov_iter_advance(&fixup, sizeof(hdr));
1226                 }
1227                 /* TODO: Should check and handle checksum. */
1228
1229                 num_buffers = cpu_to_vhost16(vq, headcount);
1230                 if (likely(mergeable) &&
1231                     copy_to_iter(&num_buffers, sizeof num_buffers,
1232                                  &fixup) != sizeof num_buffers) {
1233                         vq_err(vq, "Failed num_buffers write");
1234                         vhost_discard_vq_desc(vq, headcount);
1235                         goto out;
1236                 }
1237                 nvq->done_idx += headcount;
1238                 if (nvq->done_idx > VHOST_NET_BATCH)
1239                         vhost_net_signal_used(nvq);
1240                 if (unlikely(vq_log))
1241                         vhost_log_write(vq, vq_log, log, vhost_len,
1242                                         vq->iov, in);
1243                 total_len += vhost_len;
1244         } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1245
1246         if (unlikely(busyloop_intr))
1247                 vhost_poll_queue(&vq->poll);
1248         else if (!sock_len)
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                        UIO_MAXIOV + VHOST_NET_BATCH,
1341                        VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true,
1342                        NULL);
1343
1344         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1345         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1346
1347         f->private_data = n;
1348         n->page_frag.page = NULL;
1349         n->refcnt_bias = 0;
1350
1351         return 0;
1352 }
1353
1354 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1355                                         struct vhost_virtqueue *vq)
1356 {
1357         struct socket *sock;
1358         struct vhost_net_virtqueue *nvq =
1359                 container_of(vq, struct vhost_net_virtqueue, vq);
1360
1361         mutex_lock(&vq->mutex);
1362         sock = vhost_vq_get_backend(vq);
1363         vhost_net_disable_vq(n, vq);
1364         vhost_vq_set_backend(vq, NULL);
1365         vhost_net_buf_unproduce(nvq);
1366         nvq->rx_ring = NULL;
1367         mutex_unlock(&vq->mutex);
1368         return sock;
1369 }
1370
1371 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1372                            struct socket **rx_sock)
1373 {
1374         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1375         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1376 }
1377
1378 static void vhost_net_flush(struct vhost_net *n)
1379 {
1380         vhost_dev_flush(&n->dev);
1381         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1382                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1383                 n->tx_flush = true;
1384                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1385                 /* Wait for all lower device DMAs done. */
1386                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1387                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1388                 n->tx_flush = false;
1389                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1390                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1391         }
1392 }
1393
1394 static int vhost_net_release(struct inode *inode, struct file *f)
1395 {
1396         struct vhost_net *n = f->private_data;
1397         struct socket *tx_sock;
1398         struct socket *rx_sock;
1399
1400         vhost_net_stop(n, &tx_sock, &rx_sock);
1401         vhost_net_flush(n);
1402         vhost_dev_stop(&n->dev);
1403         vhost_dev_cleanup(&n->dev);
1404         vhost_net_vq_reset(n);
1405         if (tx_sock)
1406                 sockfd_put(tx_sock);
1407         if (rx_sock)
1408                 sockfd_put(rx_sock);
1409         /* Make sure no callbacks are outstanding */
1410         synchronize_rcu();
1411         /* We do an extra flush before freeing memory,
1412          * since jobs can re-queue themselves. */
1413         vhost_net_flush(n);
1414         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1415         kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1416         kfree(n->dev.vqs);
1417         if (n->page_frag.page)
1418                 __page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
1419         kvfree(n);
1420         return 0;
1421 }
1422
1423 static struct socket *get_raw_socket(int fd)
1424 {
1425         int r;
1426         struct socket *sock = sockfd_lookup(fd, &r);
1427
1428         if (!sock)
1429                 return ERR_PTR(-ENOTSOCK);
1430
1431         /* Parameter checking */
1432         if (sock->sk->sk_type != SOCK_RAW) {
1433                 r = -ESOCKTNOSUPPORT;
1434                 goto err;
1435         }
1436
1437         if (sock->sk->sk_family != AF_PACKET) {
1438                 r = -EPFNOSUPPORT;
1439                 goto err;
1440         }
1441         return sock;
1442 err:
1443         sockfd_put(sock);
1444         return ERR_PTR(r);
1445 }
1446
1447 static struct ptr_ring *get_tap_ptr_ring(struct file *file)
1448 {
1449         struct ptr_ring *ring;
1450         ring = tun_get_tx_ring(file);
1451         if (!IS_ERR(ring))
1452                 goto out;
1453         ring = tap_get_ptr_ring(file);
1454         if (!IS_ERR(ring))
1455                 goto out;
1456         ring = NULL;
1457 out:
1458         return ring;
1459 }
1460
1461 static struct socket *get_tap_socket(int fd)
1462 {
1463         struct file *file = fget(fd);
1464         struct socket *sock;
1465
1466         if (!file)
1467                 return ERR_PTR(-EBADF);
1468         sock = tun_get_socket(file);
1469         if (!IS_ERR(sock))
1470                 return sock;
1471         sock = tap_get_socket(file);
1472         if (IS_ERR(sock))
1473                 fput(file);
1474         return sock;
1475 }
1476
1477 static struct socket *get_socket(int fd)
1478 {
1479         struct socket *sock;
1480
1481         /* special case to disable backend */
1482         if (fd == -1)
1483                 return NULL;
1484         sock = get_raw_socket(fd);
1485         if (!IS_ERR(sock))
1486                 return sock;
1487         sock = get_tap_socket(fd);
1488         if (!IS_ERR(sock))
1489                 return sock;
1490         return ERR_PTR(-ENOTSOCK);
1491 }
1492
1493 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1494 {
1495         struct socket *sock, *oldsock;
1496         struct vhost_virtqueue *vq;
1497         struct vhost_net_virtqueue *nvq;
1498         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1499         int r;
1500
1501         mutex_lock(&n->dev.mutex);
1502         r = vhost_dev_check_owner(&n->dev);
1503         if (r)
1504                 goto err;
1505
1506         if (index >= VHOST_NET_VQ_MAX) {
1507                 r = -ENOBUFS;
1508                 goto err;
1509         }
1510         vq = &n->vqs[index].vq;
1511         nvq = &n->vqs[index];
1512         mutex_lock(&vq->mutex);
1513
1514         /* Verify that ring has been setup correctly. */
1515         if (!vhost_vq_access_ok(vq)) {
1516                 r = -EFAULT;
1517                 goto err_vq;
1518         }
1519         sock = get_socket(fd);
1520         if (IS_ERR(sock)) {
1521                 r = PTR_ERR(sock);
1522                 goto err_vq;
1523         }
1524
1525         /* start polling new socket */
1526         oldsock = vhost_vq_get_backend(vq);
1527         if (sock != oldsock) {
1528                 ubufs = vhost_net_ubuf_alloc(vq,
1529                                              sock && vhost_sock_zcopy(sock));
1530                 if (IS_ERR(ubufs)) {
1531                         r = PTR_ERR(ubufs);
1532                         goto err_ubufs;
1533                 }
1534
1535                 vhost_net_disable_vq(n, vq);
1536                 vhost_vq_set_backend(vq, sock);
1537                 vhost_net_buf_unproduce(nvq);
1538                 r = vhost_vq_init_access(vq);
1539                 if (r)
1540                         goto err_used;
1541                 r = vhost_net_enable_vq(n, vq);
1542                 if (r)
1543                         goto err_used;
1544                 if (index == VHOST_NET_VQ_RX) {
1545                         if (sock)
1546                                 nvq->rx_ring = get_tap_ptr_ring(sock->file);
1547                         else
1548                                 nvq->rx_ring = NULL;
1549                 }
1550
1551                 oldubufs = nvq->ubufs;
1552                 nvq->ubufs = ubufs;
1553
1554                 n->tx_packets = 0;
1555                 n->tx_zcopy_err = 0;
1556                 n->tx_flush = false;
1557         }
1558
1559         mutex_unlock(&vq->mutex);
1560
1561         if (oldubufs) {
1562                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1563                 mutex_lock(&vq->mutex);
1564                 vhost_zerocopy_signal_used(n, vq);
1565                 mutex_unlock(&vq->mutex);
1566         }
1567
1568         if (oldsock) {
1569                 vhost_dev_flush(&n->dev);
1570                 sockfd_put(oldsock);
1571         }
1572
1573         mutex_unlock(&n->dev.mutex);
1574         return 0;
1575
1576 err_used:
1577         vhost_vq_set_backend(vq, oldsock);
1578         vhost_net_enable_vq(n, vq);
1579         if (ubufs)
1580                 vhost_net_ubuf_put_wait_and_free(ubufs);
1581 err_ubufs:
1582         if (sock)
1583                 sockfd_put(sock);
1584 err_vq:
1585         mutex_unlock(&vq->mutex);
1586 err:
1587         mutex_unlock(&n->dev.mutex);
1588         return r;
1589 }
1590
1591 static long vhost_net_reset_owner(struct vhost_net *n)
1592 {
1593         struct socket *tx_sock = NULL;
1594         struct socket *rx_sock = NULL;
1595         long err;
1596         struct vhost_iotlb *umem;
1597
1598         mutex_lock(&n->dev.mutex);
1599         err = vhost_dev_check_owner(&n->dev);
1600         if (err)
1601                 goto done;
1602         umem = vhost_dev_reset_owner_prepare();
1603         if (!umem) {
1604                 err = -ENOMEM;
1605                 goto done;
1606         }
1607         vhost_net_stop(n, &tx_sock, &rx_sock);
1608         vhost_net_flush(n);
1609         vhost_dev_stop(&n->dev);
1610         vhost_dev_reset_owner(&n->dev, umem);
1611         vhost_net_vq_reset(n);
1612 done:
1613         mutex_unlock(&n->dev.mutex);
1614         if (tx_sock)
1615                 sockfd_put(tx_sock);
1616         if (rx_sock)
1617                 sockfd_put(rx_sock);
1618         return err;
1619 }
1620
1621 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1622 {
1623         size_t vhost_hlen, sock_hlen, hdr_len;
1624         int i;
1625
1626         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1627                                (1ULL << VIRTIO_F_VERSION_1))) ?
1628                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1629                         sizeof(struct virtio_net_hdr);
1630         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1631                 /* vhost provides vnet_hdr */
1632                 vhost_hlen = hdr_len;
1633                 sock_hlen = 0;
1634         } else {
1635                 /* socket provides vnet_hdr */
1636                 vhost_hlen = 0;
1637                 sock_hlen = hdr_len;
1638         }
1639         mutex_lock(&n->dev.mutex);
1640         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1641             !vhost_log_access_ok(&n->dev))
1642                 goto out_unlock;
1643
1644         if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) {
1645                 if (vhost_init_device_iotlb(&n->dev, true))
1646                         goto out_unlock;
1647         }
1648
1649         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1650                 mutex_lock(&n->vqs[i].vq.mutex);
1651                 n->vqs[i].vq.acked_features = features;
1652                 n->vqs[i].vhost_hlen = vhost_hlen;
1653                 n->vqs[i].sock_hlen = sock_hlen;
1654                 mutex_unlock(&n->vqs[i].vq.mutex);
1655         }
1656         mutex_unlock(&n->dev.mutex);
1657         return 0;
1658
1659 out_unlock:
1660         mutex_unlock(&n->dev.mutex);
1661         return -EFAULT;
1662 }
1663
1664 static long vhost_net_set_owner(struct vhost_net *n)
1665 {
1666         int r;
1667
1668         mutex_lock(&n->dev.mutex);
1669         if (vhost_dev_has_owner(&n->dev)) {
1670                 r = -EBUSY;
1671                 goto out;
1672         }
1673         r = vhost_net_set_ubuf_info(n);
1674         if (r)
1675                 goto out;
1676         r = vhost_dev_set_owner(&n->dev);
1677         if (r)
1678                 vhost_net_clear_ubuf_info(n);
1679         vhost_net_flush(n);
1680 out:
1681         mutex_unlock(&n->dev.mutex);
1682         return r;
1683 }
1684
1685 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1686                             unsigned long arg)
1687 {
1688         struct vhost_net *n = f->private_data;
1689         void __user *argp = (void __user *)arg;
1690         u64 __user *featurep = argp;
1691         struct vhost_vring_file backend;
1692         u64 features;
1693         int r;
1694
1695         switch (ioctl) {
1696         case VHOST_NET_SET_BACKEND:
1697                 if (copy_from_user(&backend, argp, sizeof backend))
1698                         return -EFAULT;
1699                 return vhost_net_set_backend(n, backend.index, backend.fd);
1700         case VHOST_GET_FEATURES:
1701                 features = VHOST_NET_FEATURES;
1702                 if (copy_to_user(featurep, &features, sizeof features))
1703                         return -EFAULT;
1704                 return 0;
1705         case VHOST_SET_FEATURES:
1706                 if (copy_from_user(&features, featurep, sizeof features))
1707                         return -EFAULT;
1708                 if (features & ~VHOST_NET_FEATURES)
1709                         return -EOPNOTSUPP;
1710                 return vhost_net_set_features(n, features);
1711         case VHOST_GET_BACKEND_FEATURES:
1712                 features = VHOST_NET_BACKEND_FEATURES;
1713                 if (copy_to_user(featurep, &features, sizeof(features)))
1714                         return -EFAULT;
1715                 return 0;
1716         case VHOST_SET_BACKEND_FEATURES:
1717                 if (copy_from_user(&features, featurep, sizeof(features)))
1718                         return -EFAULT;
1719                 if (features & ~VHOST_NET_BACKEND_FEATURES)
1720                         return -EOPNOTSUPP;
1721                 vhost_set_backend_features(&n->dev, features);
1722                 return 0;
1723         case VHOST_RESET_OWNER:
1724                 return vhost_net_reset_owner(n);
1725         case VHOST_SET_OWNER:
1726                 return vhost_net_set_owner(n);
1727         default:
1728                 mutex_lock(&n->dev.mutex);
1729                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1730                 if (r == -ENOIOCTLCMD)
1731                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1732                 else
1733                         vhost_net_flush(n);
1734                 mutex_unlock(&n->dev.mutex);
1735                 return r;
1736         }
1737 }
1738
1739 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1740 {
1741         struct file *file = iocb->ki_filp;
1742         struct vhost_net *n = file->private_data;
1743         struct vhost_dev *dev = &n->dev;
1744         int noblock = file->f_flags & O_NONBLOCK;
1745
1746         return vhost_chr_read_iter(dev, to, noblock);
1747 }
1748
1749 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1750                                         struct iov_iter *from)
1751 {
1752         struct file *file = iocb->ki_filp;
1753         struct vhost_net *n = file->private_data;
1754         struct vhost_dev *dev = &n->dev;
1755
1756         return vhost_chr_write_iter(dev, from);
1757 }
1758
1759 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1760 {
1761         struct vhost_net *n = file->private_data;
1762         struct vhost_dev *dev = &n->dev;
1763
1764         return vhost_chr_poll(file, dev, wait);
1765 }
1766
1767 static const struct file_operations vhost_net_fops = {
1768         .owner          = THIS_MODULE,
1769         .release        = vhost_net_release,
1770         .read_iter      = vhost_net_chr_read_iter,
1771         .write_iter     = vhost_net_chr_write_iter,
1772         .poll           = vhost_net_chr_poll,
1773         .unlocked_ioctl = vhost_net_ioctl,
1774         .compat_ioctl   = compat_ptr_ioctl,
1775         .open           = vhost_net_open,
1776         .llseek         = noop_llseek,
1777 };
1778
1779 static struct miscdevice vhost_net_misc = {
1780         .minor = VHOST_NET_MINOR,
1781         .name = "vhost-net",
1782         .fops = &vhost_net_fops,
1783 };
1784
1785 static int __init vhost_net_init(void)
1786 {
1787         if (experimental_zcopytx)
1788                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1789         return misc_register(&vhost_net_misc);
1790 }
1791 module_init(vhost_net_init);
1792
1793 static void __exit vhost_net_exit(void)
1794 {
1795         misc_deregister(&vhost_net_misc);
1796 }
1797 module_exit(vhost_net_exit);
1798
1799 MODULE_VERSION("0.0.1");
1800 MODULE_LICENSE("GPL v2");
1801 MODULE_AUTHOR("Michael S. Tsirkin");
1802 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1803 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1804 MODULE_ALIAS("devname:vhost-net");