drivers/vhost/net.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (C) 2009 Red Hat, Inc.
   3  * Author: Michael S. Tsirkin <mst@redhat.com>
   4  *
   5  * virtio-net server in host kernel.
   6  */
   7
   8 #include <linux/compat.h>
   9 #include <linux/eventfd.h>
  10 #include <linux/vhost.h>
  11 #include <linux/virtio_net.h>
  12 #include <linux/miscdevice.h>
  13 #include <linux/module.h>
  14 #include <linux/moduleparam.h>
  15 #include <linux/mutex.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/file.h>
  18 #include <linux/slab.h>
  19 #include <linux/sched/clock.h>
  20 #include <linux/sched/signal.h>
  21 #include <linux/vmalloc.h>
  22
  23 #include <linux/net.h>
  24 #include <linux/if_packet.h>
  25 #include <linux/if_arp.h>
  26 #include <linux/if_tun.h>
  27 #include <linux/if_macvlan.h>
  28 #include <linux/if_tap.h>
  29 #include <linux/if_vlan.h>
  30 #include <linux/skb_array.h>
  31 #include <linux/skbuff.h>
  32
  33 #include <net/sock.h>
  34 #include <net/xdp.h>
  35
  36 #include "vhost.h"
  37
  38 static int experimental_zcopytx = 0;
  39 module_param(experimental_zcopytx, int, 0444);
  40 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  41                                        " 1 -Enable; 0 - Disable");
  42
  43 /* Max number of bytes transferred before requeueing the job.
  44  * Using this limit prevents one virtqueue from starving others. */
  45 #define VHOST_NET_WEIGHT 0x80000
  46
  47 /* Max number of packets transferred before requeueing the job.
  48  * Using this limit prevents one virtqueue from starving others with small
  49  * pkts.
  50  */
  51 #define VHOST_NET_PKT_WEIGHT 256
  52
  53 /* MAX number of TX used buffers for outstanding zerocopy */
  54 #define VHOST_MAX_PEND 128
  55 #define VHOST_GOODCOPY_LEN 256
  56
  57 /*
  58  * For transmit, used buffer len is unused; we override it to track buffer
  59  * status internally; used for zerocopy tx only.
  60  */
  61 /* Lower device DMA failed */
  62 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  63 /* Lower device DMA done */
  64 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  65 /* Lower device DMA in progress */
  66 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  67 /* Buffer unused */
  68 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  69
  70 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  71
  72 enum {
  73         VHOST_NET_FEATURES = VHOST_FEATURES |
  74                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  75                          (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  76                          (1ULL << VIRTIO_F_IOMMU_PLATFORM)
  77 };
  78
  79 enum {
  80         VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
  81 };
  82
  83 enum {
  84         VHOST_NET_VQ_RX = 0,
  85         VHOST_NET_VQ_TX = 1,
  86         VHOST_NET_VQ_MAX = 2,
  87 };
  88
  89 struct vhost_net_ubuf_ref {
  90         /* refcount follows semantics similar to kref:
  91          *  0: object is released
  92          *  1: no outstanding ubufs
  93          * >1: outstanding ubufs
  94          */
  95         atomic_t refcount;
  96         wait_queue_head_t wait;
  97         struct vhost_virtqueue *vq;
  98 };
  99
 100 #define VHOST_NET_BATCH 64
 101 struct vhost_net_buf {
 102         void **queue;
 103         int tail;
 104         int head;
 105 };
 106
 107 struct vhost_net_virtqueue {
 108         struct vhost_virtqueue vq;
 109         size_t vhost_hlen;
 110         size_t sock_hlen;
 111         /* vhost zerocopy support fields below: */
 112         /* last used idx for outstanding DMA zerocopy buffers */
 113         int upend_idx;
 114         /* For TX, first used idx for DMA done zerocopy buffers
 115          * For RX, number of batched heads
 116          */
 117         int done_idx;
 118         /* Number of XDP frames batched */
 119         int batched_xdp;
 120         /* an array of userspace buffers info */
 121         struct ubuf_info *ubuf_info;
 122         /* Reference counting for outstanding ubufs.
 123          * Protected by vq mutex. Writers must also take device mutex. */
 124         struct vhost_net_ubuf_ref *ubufs;
 125         struct ptr_ring *rx_ring;
 126         struct vhost_net_buf rxq;
 127         /* Batched XDP buffs */
 128         struct xdp_buff *xdp;
 129 };
 130
 131 struct vhost_net {
 132         struct vhost_dev dev;
 133         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 134         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 135         /* Number of TX recently submitted.
 136          * Protected by tx vq lock. */
 137         unsigned tx_packets;
 138         /* Number of times zerocopy TX recently failed.
 139          * Protected by tx vq lock. */
 140         unsigned tx_zcopy_err;
 141         /* Flush in progress. Protected by tx vq lock. */
 142         bool tx_flush;
 143         /* Private page frag */
 144         struct page_frag page_frag;
 145         /* Refcount bias of page frag */
 146         int refcnt_bias;
 147 };
 148
 149 static unsigned vhost_net_zcopy_mask __read_mostly;
 150
 151 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 152 {
 153         if (rxq->tail != rxq->head)
 154                 return rxq->queue[rxq->head];
 155         else
 156                 return NULL;
 157 }
 158
 159 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 160 {
 161         return rxq->tail - rxq->head;
 162 }
 163
 164 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 165 {
 166         return rxq->tail == rxq->head;
 167 }
 168
 169 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 170 {
 171         void *ret = vhost_net_buf_get_ptr(rxq);
 172         ++rxq->head;
 173         return ret;
 174 }
 175
 176 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 177 {
 178         struct vhost_net_buf *rxq = &nvq->rxq;
 179
 180         rxq->head = 0;
 181         rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
 182                                               VHOST_NET_BATCH);
 183         return rxq->tail;
 184 }
 185
 186 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 187 {
 188         struct vhost_net_buf *rxq = &nvq->rxq;
 189
 190         if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
 191                 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 192                                    vhost_net_buf_get_size(rxq),
 193                                    tun_ptr_free);
 194                 rxq->head = rxq->tail = 0;
 195         }
 196 }
 197
 198 static int vhost_net_buf_peek_len(void *ptr)
 199 {
 200         if (tun_is_xdp_frame(ptr)) {
 201                 struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
 202
 203                 return xdpf->len;
 204         }
 205
 206         return __skb_array_len_with_tag(ptr);
 207 }
 208
 209 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 210 {
 211         struct vhost_net_buf *rxq = &nvq->rxq;
 212
 213         if (!vhost_net_buf_is_empty(rxq))
 214                 goto out;
 215
 216         if (!vhost_net_buf_produce(nvq))
 217                 return 0;
 218
 219 out:
 220         return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 221 }
 222
 223 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 224 {
 225         rxq->head = rxq->tail = 0;
 226 }
 227
 228 static void vhost_net_enable_zcopy(int vq)
 229 {
 230         vhost_net_zcopy_mask |= 0x1 << vq;
 231 }
 232
 233 static struct vhost_net_ubuf_ref *
 234 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 235 {
 236         struct vhost_net_ubuf_ref *ubufs;
 237         /* No zero copy backend? Nothing to count. */
 238         if (!zcopy)
 239                 return NULL;
 240         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 241         if (!ubufs)
 242                 return ERR_PTR(-ENOMEM);
 243         atomic_set(&ubufs->refcount, 1);
 244         init_waitqueue_head(&ubufs->wait);
 245         ubufs->vq = vq;
 246         return ubufs;
 247 }
 248
 249 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 250 {
 251         int r = atomic_sub_return(1, &ubufs->refcount);
 252         if (unlikely(!r))
 253                 wake_up(&ubufs->wait);
 254         return r;
 255 }
 256
 257 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 258 {
 259         vhost_net_ubuf_put(ubufs);
 260         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 261 }
 262
 263 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 264 {
 265         vhost_net_ubuf_put_and_wait(ubufs);
 266         kfree(ubufs);
 267 }
 268
 269 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 270 {
 271         int i;
 272
 273         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 274                 kfree(n->vqs[i].ubuf_info);
 275                 n->vqs[i].ubuf_info = NULL;
 276         }
 277 }
 278
 279 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 280 {
 281         bool zcopy;
 282         int i;
 283
 284         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 285                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 286                 if (!zcopy)
 287                         continue;
 288                 n->vqs[i].ubuf_info =
 289                         kmalloc_array(UIO_MAXIOV,
 290                                       sizeof(*n->vqs[i].ubuf_info),
 291                                       GFP_KERNEL);
 292                 if  (!n->vqs[i].ubuf_info)
 293                         goto err;
 294         }
 295         return 0;
 296
 297 err:
 298         vhost_net_clear_ubuf_info(n);
 299         return -ENOMEM;
 300 }
 301
 302 static void vhost_net_vq_reset(struct vhost_net *n)
 303 {
 304         int i;
 305
 306         vhost_net_clear_ubuf_info(n);
 307
 308         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 309                 n->vqs[i].done_idx = 0;
 310                 n->vqs[i].upend_idx = 0;
 311                 n->vqs[i].ubufs = NULL;
 312                 n->vqs[i].vhost_hlen = 0;
 313                 n->vqs[i].sock_hlen = 0;
 314                 vhost_net_buf_init(&n->vqs[i].rxq);
 315         }
 316
 317 }
 318
 319 static void vhost_net_tx_packet(struct vhost_net *net)
 320 {
 321         ++net->tx_packets;
 322         if (net->tx_packets < 1024)
 323                 return;
 324         net->tx_packets = 0;
 325         net->tx_zcopy_err = 0;
 326 }
 327
 328 static void vhost_net_tx_err(struct vhost_net *net)
 329 {
 330         ++net->tx_zcopy_err;
 331 }
 332
 333 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 334 {
 335         /* TX flush waits for outstanding DMAs to be done.
 336          * Don't start new DMAs.
 337          */
 338         return !net->tx_flush &&
 339                 net->tx_packets / 64 >= net->tx_zcopy_err;
 340 }
 341
 342 static bool vhost_sock_zcopy(struct socket *sock)
 343 {
 344         return unlikely(experimental_zcopytx) &&
 345                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 346 }
 347
 348 static bool vhost_sock_xdp(struct socket *sock)
 349 {
 350         return sock_flag(sock->sk, SOCK_XDP);
 351 }
 352
 353 /* In case of DMA done not in order in lower device driver for some reason.
 354  * upend_idx is used to track end of used idx, done_idx is used to track head
 355  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 356  * guest used idx.
 357  */
 358 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 359                                        struct vhost_virtqueue *vq)
 360 {
 361         struct vhost_net_virtqueue *nvq =
 362                 container_of(vq, struct vhost_net_virtqueue, vq);
 363         int i, add;
 364         int j = 0;
 365
 366         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 367                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 368                         vhost_net_tx_err(net);
 369                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 370                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 371                         ++j;
 372                 } else
 373                         break;
 374         }
 375         while (j) {
 376                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 377                 vhost_add_used_and_signal_n(vq->dev, vq,
 378                                             &vq->heads[nvq->done_idx], add);
 379                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 380                 j -= add;
 381         }
 382 }
 383
 384 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 385 {
 386         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 387         struct vhost_virtqueue *vq = ubufs->vq;
 388         int cnt;
 389
 390         rcu_read_lock_bh();
 391
 392         /* set len to mark this desc buffers done DMA */
 393         vq->heads[ubuf->desc].len = success ?
 394                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 395         cnt = vhost_net_ubuf_put(ubufs);
 396
 397         /*
 398          * Trigger polling thread if guest stopped submitting new buffers:
 399          * in this case, the refcount after decrement will eventually reach 1.
 400          * We also trigger polling periodically after each 16 packets
 401          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 402          * less than 10% of times).
 403          */
 404         if (cnt <= 1 || !(cnt % 16))
 405                 vhost_poll_queue(&vq->poll);
 406
 407         rcu_read_unlock_bh();
 408 }
 409
 410 static inline unsigned long busy_clock(void)
 411 {
 412         return local_clock() >> 10;
 413 }
 414
 415 static bool vhost_can_busy_poll(unsigned long endtime)
 416 {
 417         return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
 418                       !signal_pending(current));
 419 }
 420
 421 static void vhost_net_disable_vq(struct vhost_net *n,
 422                                  struct vhost_virtqueue *vq)
 423 {
 424         struct vhost_net_virtqueue *nvq =
 425                 container_of(vq, struct vhost_net_virtqueue, vq);
 426         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 427         if (!vq->private_data)
 428                 return;
 429         vhost_poll_stop(poll);
 430 }
 431
 432 static int vhost_net_enable_vq(struct vhost_net *n,
 433                                 struct vhost_virtqueue *vq)
 434 {
 435         struct vhost_net_virtqueue *nvq =
 436                 container_of(vq, struct vhost_net_virtqueue, vq);
 437         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 438         struct socket *sock;
 439
 440         sock = vq->private_data;
 441         if (!sock)
 442                 return 0;
 443
 444         return vhost_poll_start(poll, sock->file);
 445 }
 446
 447 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
 448 {
 449         struct vhost_virtqueue *vq = &nvq->vq;
 450         struct vhost_dev *dev = vq->dev;
 451
 452         if (!nvq->done_idx)
 453                 return;
 454
 455         vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
 456         nvq->done_idx = 0;
 457 }
 458
 459 static void vhost_tx_batch(struct vhost_net *net,
 460                            struct vhost_net_virtqueue *nvq,
 461                            struct socket *sock,
 462                            struct msghdr *msghdr)
 463 {
 464         struct tun_msg_ctl ctl = {
 465                 .type = TUN_MSG_PTR,
 466                 .num = nvq->batched_xdp,
 467                 .ptr = nvq->xdp,
 468         };
 469         int err;
 470
 471         if (nvq->batched_xdp == 0)
 472                 goto signal_used;
 473
 474         msghdr->msg_control = &ctl;
 475         err = sock->ops->sendmsg(sock, msghdr, 0);
 476         if (unlikely(err < 0)) {
 477                 vq_err(&nvq->vq, "Fail to batch sending packets\n");
 478                 return;
 479         }
 480
 481 signal_used:
 482         vhost_net_signal_used(nvq);
 483         nvq->batched_xdp = 0;
 484 }
 485
 486 static int sock_has_rx_data(struct socket *sock)
 487 {
 488         if (unlikely(!sock))
 489                 return 0;
 490
 491         if (sock->ops->peek_len)
 492                 return sock->ops->peek_len(sock);
 493
 494         return skb_queue_empty(&sock->sk->sk_receive_queue);
 495 }
 496
 497 static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
 498                                           struct vhost_virtqueue *vq)
 499 {
 500         if (!vhost_vq_avail_empty(&net->dev, vq)) {
 501                 vhost_poll_queue(&vq->poll);
 502         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 503                 vhost_disable_notify(&net->dev, vq);
 504                 vhost_poll_queue(&vq->poll);
 505         }
 506 }
 507
 508 static void vhost_net_busy_poll(struct vhost_net *net,
 509                                 struct vhost_virtqueue *rvq,
 510                                 struct vhost_virtqueue *tvq,
 511                                 bool *busyloop_intr,
 512                                 bool poll_rx)
 513 {
 514         unsigned long busyloop_timeout;
 515         unsigned long endtime;
 516         struct socket *sock;
 517         struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
 518
 519         /* Try to hold the vq mutex of the paired virtqueue. We can't
 520          * use mutex_lock() here since we could not guarantee a
 521          * consistenet lock ordering.
 522          */
 523         if (!mutex_trylock(&vq->mutex))
 524                 return;
 525
 526         vhost_disable_notify(&net->dev, vq);
 527         sock = rvq->private_data;
 528
 529         busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
 530                                      tvq->busyloop_timeout;
 531
 532         preempt_disable();
 533         endtime = busy_clock() + busyloop_timeout;
 534
 535         while (vhost_can_busy_poll(endtime)) {
 536                 if (vhost_has_work(&net->dev)) {
 537                         *busyloop_intr = true;
 538                         break;
 539                 }
 540
 541                 if ((sock_has_rx_data(sock) &&
 542                      !vhost_vq_avail_empty(&net->dev, rvq)) ||
 543                     !vhost_vq_avail_empty(&net->dev, tvq))
 544                         break;
 545
 546                 cpu_relax();
 547         }
 548
 549         preempt_enable();
 550
 551         if (poll_rx || sock_has_rx_data(sock))
 552                 vhost_net_busy_poll_try_queue(net, vq);
 553         else if (!poll_rx) /* On tx here, sock has no rx data. */
 554                 vhost_enable_notify(&net->dev, rvq);
 555
 556         mutex_unlock(&vq->mutex);
 557 }
 558
 559 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 560                                     struct vhost_net_virtqueue *tnvq,
 561                                     unsigned int *out_num, unsigned int *in_num,
 562                                     struct msghdr *msghdr, bool *busyloop_intr)
 563 {
 564         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 565         struct vhost_virtqueue *rvq = &rnvq->vq;
 566         struct vhost_virtqueue *tvq = &tnvq->vq;
 567
 568         int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 569                                   out_num, in_num, NULL, NULL);
 570
 571         if (r == tvq->num && tvq->busyloop_timeout) {
 572                 /* Flush batched packets first */
 573                 if (!vhost_sock_zcopy(tvq->private_data))
 574                         vhost_tx_batch(net, tnvq, tvq->private_data, msghdr);
 575
 576                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
 577
 578                 r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 579                                       out_num, in_num, NULL, NULL);
 580         }
 581
 582         return r;
 583 }
 584
 585 static bool vhost_exceeds_maxpend(struct vhost_net *net)
 586 {
 587         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 588         struct vhost_virtqueue *vq = &nvq->vq;
 589
 590         return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 591                min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 592 }
 593
 594 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
 595                             size_t hdr_size, int out)
 596 {
 597         /* Skip header. TODO: support TSO. */
 598         size_t len = iov_length(vq->iov, out);
 599
 600         iov_iter_init(iter, WRITE, vq->iov, out, len);
 601         iov_iter_advance(iter, hdr_size);
 602
 603         return iov_iter_count(iter);
 604 }
 605
 606 static int get_tx_bufs(struct vhost_net *net,
 607                        struct vhost_net_virtqueue *nvq,
 608                        struct msghdr *msg,
 609                        unsigned int *out, unsigned int *in,
 610                        size_t *len, bool *busyloop_intr)
 611 {
 612         struct vhost_virtqueue *vq = &nvq->vq;
 613         int ret;
 614
 615         ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
 616
 617         if (ret < 0 || ret == vq->num)
 618                 return ret;
 619
 620         if (*in) {
 621                 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 622                         *out, *in);
 623                 return -EFAULT;
 624         }
 625
 626         /* Sanity check */
 627         *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 628         if (*len == 0) {
 629                 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 630                         *len, nvq->vhost_hlen);
 631                 return -EFAULT;
 632         }
 633
 634         return ret;
 635 }
 636
 637 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 638 {
 639         return total_len < VHOST_NET_WEIGHT &&
 640                !vhost_vq_avail_empty(vq->dev, vq);
 641 }
 642
 643 #define SKB_FRAG_PAGE_ORDER     get_order(32768)
 644
 645 static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
 646                                        struct page_frag *pfrag, gfp_t gfp)
 647 {
 648         if (pfrag->page) {
 649                 if (pfrag->offset + sz <= pfrag->size)
 650                         return true;
 651                 __page_frag_cache_drain(pfrag->page, net->refcnt_bias);
 652         }
 653
 654         pfrag->offset = 0;
 655         net->refcnt_bias = 0;
 656         if (SKB_FRAG_PAGE_ORDER) {
 657                 /* Avoid direct reclaim but allow kswapd to wake */
 658                 pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
 659                                           __GFP_COMP | __GFP_NOWARN |
 660                                           __GFP_NORETRY,
 661                                           SKB_FRAG_PAGE_ORDER);
 662                 if (likely(pfrag->page)) {
 663                         pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
 664                         goto done;
 665                 }
 666         }
 667         pfrag->page = alloc_page(gfp);
 668         if (likely(pfrag->page)) {
 669                 pfrag->size = PAGE_SIZE;
 670                 goto done;
 671         }
 672         return false;
 673
 674 done:
 675         net->refcnt_bias = USHRT_MAX;
 676         page_ref_add(pfrag->page, USHRT_MAX - 1);
 677         return true;
 678 }
 679
 680 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 681
 682 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
 683                                struct iov_iter *from)
 684 {
 685         struct vhost_virtqueue *vq = &nvq->vq;
 686         struct vhost_net *net = container_of(vq->dev, struct vhost_net,
 687                                              dev);
 688         struct socket *sock = vq->private_data;
 689         struct page_frag *alloc_frag = &net->page_frag;
 690         struct virtio_net_hdr *gso;
 691         struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
 692         struct tun_xdp_hdr *hdr;
 693         size_t len = iov_iter_count(from);
 694         int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
 695         int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 696         int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
 697         int sock_hlen = nvq->sock_hlen;
 698         void *buf;
 699         int copied;
 700
 701         if (unlikely(len < nvq->sock_hlen))
 702                 return -EFAULT;
 703
 704         if (SKB_DATA_ALIGN(len + pad) +
 705             SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
 706                 return -ENOSPC;
 707
 708         buflen += SKB_DATA_ALIGN(len + pad);
 709         alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
 710         if (unlikely(!vhost_net_page_frag_refill(net, buflen,
 711                                                  alloc_frag, GFP_KERNEL)))
 712                 return -ENOMEM;
 713
 714         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 715         copied = copy_page_from_iter(alloc_frag->page,
 716                                      alloc_frag->offset +
 717                                      offsetof(struct tun_xdp_hdr, gso),
 718                                      sock_hlen, from);
 719         if (copied != sock_hlen)
 720                 return -EFAULT;
 721
 722         hdr = buf;
 723         gso = &hdr->gso;
 724
 725         if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
 726             vhost16_to_cpu(vq, gso->csum_start) +
 727             vhost16_to_cpu(vq, gso->csum_offset) + 2 >
 728             vhost16_to_cpu(vq, gso->hdr_len)) {
 729                 gso->hdr_len = cpu_to_vhost16(vq,
 730                                vhost16_to_cpu(vq, gso->csum_start) +
 731                                vhost16_to_cpu(vq, gso->csum_offset) + 2);
 732
 733                 if (vhost16_to_cpu(vq, gso->hdr_len) > len)
 734                         return -EINVAL;
 735         }
 736
 737         len -= sock_hlen;
 738         copied = copy_page_from_iter(alloc_frag->page,
 739                                      alloc_frag->offset + pad,
 740                                      len, from);
 741         if (copied != len)
 742                 return -EFAULT;
 743
 744         xdp->data_hard_start = buf;
 745         xdp->data = buf + pad;
 746         xdp->data_end = xdp->data + len;
 747         hdr->buflen = buflen;
 748
 749         --net->refcnt_bias;
 750         alloc_frag->offset += buflen;
 751
 752         ++nvq->batched_xdp;
 753
 754         return 0;
 755 }
 756
 757 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 758 {
 759         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 760         struct vhost_virtqueue *vq = &nvq->vq;
 761         unsigned out, in;
 762         int head;
 763         struct msghdr msg = {
 764                 .msg_name = NULL,
 765                 .msg_namelen = 0,
 766                 .msg_control = NULL,
 767                 .msg_controllen = 0,
 768                 .msg_flags = MSG_DONTWAIT,
 769         };
 770         size_t len, total_len = 0;
 771         int err;
 772         int sent_pkts = 0;
 773         bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
 774
 775         do {
 776                 bool busyloop_intr = false;
 777
 778                 if (nvq->done_idx == VHOST_NET_BATCH)
 779                         vhost_tx_batch(net, nvq, sock, &msg);
 780
 781                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 782                                    &busyloop_intr);
 783                 /* On error, stop handling until the next kick. */
 784                 if (unlikely(head < 0))
 785                         break;
 786                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 787                 if (head == vq->num) {
 788                         if (unlikely(busyloop_intr)) {
 789                                 vhost_poll_queue(&vq->poll);
 790                         } else if (unlikely(vhost_enable_notify(&net->dev,
 791                                                                 vq))) {
 792                                 vhost_disable_notify(&net->dev, vq);
 793                                 continue;
 794                         }
 795                         break;
 796                 }
 797
 798                 total_len += len;
 799
 800                 /* For simplicity, TX batching is only enabled if
 801                  * sndbuf is unlimited.
 802                  */
 803                 if (sock_can_batch) {
 804                         err = vhost_net_build_xdp(nvq, &msg.msg_iter);
 805                         if (!err) {
 806                                 goto done;
 807                         } else if (unlikely(err != -ENOSPC)) {
 808                                 vhost_tx_batch(net, nvq, sock, &msg);
 809                                 vhost_discard_vq_desc(vq, 1);
 810                                 vhost_net_enable_vq(net, vq);
 811                                 break;
 812                         }
 813
 814                         /* We can't build XDP buff, go for single
 815                          * packet path but let's flush batched
 816                          * packets.
 817                          */
 818                         vhost_tx_batch(net, nvq, sock, &msg);
 819                         msg.msg_control = NULL;
 820                 } else {
 821                         if (tx_can_batch(vq, total_len))
 822                                 msg.msg_flags |= MSG_MORE;
 823                         else
 824                                 msg.msg_flags &= ~MSG_MORE;
 825                 }
 826
 827                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 828                 err = sock->ops->sendmsg(sock, &msg, len);
 829                 if (unlikely(err < 0)) {
 830                         vhost_discard_vq_desc(vq, 1);
 831                         vhost_net_enable_vq(net, vq);
 832                         break;
 833                 }
 834                 if (err != len)
 835                         pr_debug("Truncated TX packet: len %d != %zd\n",
 836                                  err, len);
 837 done:
 838                 vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 839                 vq->heads[nvq->done_idx].len = 0;
 840                 ++nvq->done_idx;
 841         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 842
 843         vhost_tx_batch(net, nvq, sock, &msg);
 844 }
 845
 846 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 847 {
 848         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 849         struct vhost_virtqueue *vq = &nvq->vq;
 850         unsigned out, in;
 851         int head;
 852         struct msghdr msg = {
 853                 .msg_name = NULL,
 854                 .msg_namelen = 0,
 855                 .msg_control = NULL,
 856                 .msg_controllen = 0,
 857                 .msg_flags = MSG_DONTWAIT,
 858         };
 859         struct tun_msg_ctl ctl;
 860         size_t len, total_len = 0;
 861         int err;
 862         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 863         bool zcopy_used;
 864         int sent_pkts = 0;
 865
 866         do {
 867                 bool busyloop_intr;
 868
 869                 /* Release DMAs done buffers first */
 870                 vhost_zerocopy_signal_used(net, vq);
 871
 872                 busyloop_intr = false;
 873                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 874                                    &busyloop_intr);
 875                 /* On error, stop handling until the next kick. */
 876                 if (unlikely(head < 0))
 877                         break;
 878                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 879                 if (head == vq->num) {
 880                         if (unlikely(busyloop_intr)) {
 881                                 vhost_poll_queue(&vq->poll);
 882                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 883                                 vhost_disable_notify(&net->dev, vq);
 884                                 continue;
 885                         }
 886                         break;
 887                 }
 888
 889                 zcopy_used = len >= VHOST_GOODCOPY_LEN
 890                              && !vhost_exceeds_maxpend(net)
 891                              && vhost_net_tx_select_zcopy(net);
 892
 893                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 894                 if (zcopy_used) {
 895                         struct ubuf_info *ubuf;
 896                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 897
 898                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 899                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 900                         ubuf->callback = vhost_zerocopy_callback;
 901                         ubuf->ctx = nvq->ubufs;
 902                         ubuf->desc = nvq->upend_idx;
 903                         refcount_set(&ubuf->refcnt, 1);
 904                         msg.msg_control = &ctl;
 905                         ctl.type = TUN_MSG_UBUF;
 906                         ctl.ptr = ubuf;
 907                         msg.msg_controllen = sizeof(ctl);
 908                         ubufs = nvq->ubufs;
 909                         atomic_inc(&ubufs->refcount);
 910                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 911                 } else {
 912                         msg.msg_control = NULL;
 913                         ubufs = NULL;
 914                 }
 915                 total_len += len;
 916                 if (tx_can_batch(vq, total_len) &&
 917                     likely(!vhost_exceeds_maxpend(net))) {
 918                         msg.msg_flags |= MSG_MORE;
 919                 } else {
 920                         msg.msg_flags &= ~MSG_MORE;
 921                 }
 922
 923                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 924                 err = sock->ops->sendmsg(sock, &msg, len);
 925                 if (unlikely(err < 0)) {
 926                         if (zcopy_used) {
 927                                 vhost_net_ubuf_put(ubufs);
 928                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 929                                         % UIO_MAXIOV;
 930                         }
 931                         vhost_discard_vq_desc(vq, 1);
 932                         vhost_net_enable_vq(net, vq);
 933                         break;
 934                 }
 935                 if (err != len)
 936                         pr_debug("Truncated TX packet: "
 937                                  " len %d != %zd\n", err, len);
 938                 if (!zcopy_used)
 939                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 940                 else
 941                         vhost_zerocopy_signal_used(net, vq);
 942                 vhost_net_tx_packet(net);
 943         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 944 }
 945
 946 /* Expects to be always run from workqueue - which acts as
 947  * read-size critical section for our kind of RCU. */
 948 static void handle_tx(struct vhost_net *net)
 949 {
 950         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 951         struct vhost_virtqueue *vq = &nvq->vq;
 952         struct socket *sock;
 953
 954         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
 955         sock = vq->private_data;
 956         if (!sock)
 957                 goto out;
 958
 959         if (!vq_meta_prefetch(vq))
 960                 goto out;
 961
 962         vhost_disable_notify(&net->dev, vq);
 963         vhost_net_disable_vq(net, vq);
 964
 965         if (vhost_sock_zcopy(sock))
 966                 handle_tx_zerocopy(net, sock);
 967         else
 968                 handle_tx_copy(net, sock);
 969
 970 out:
 971         mutex_unlock(&vq->mutex);
 972 }
 973
 974 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 975 {
 976         struct sk_buff *head;
 977         int len = 0;
 978         unsigned long flags;
 979
 980         if (rvq->rx_ring)
 981                 return vhost_net_buf_peek(rvq);
 982
 983         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 984         head = skb_peek(&sk->sk_receive_queue);
 985         if (likely(head)) {
 986                 len = head->len;
 987                 if (skb_vlan_tag_present(head))
 988                         len += VLAN_HLEN;
 989         }
 990
 991         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 992         return len;
 993 }
 994
 995 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
 996                                       bool *busyloop_intr)
 997 {
 998         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 999         struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1000         struct vhost_virtqueue *rvq = &rnvq->vq;
1001         struct vhost_virtqueue *tvq = &tnvq->vq;
1002         int len = peek_head_len(rnvq, sk);
1003
1004         if (!len && rvq->busyloop_timeout) {
1005                 /* Flush batched heads first */
1006                 vhost_net_signal_used(rnvq);
1007                 /* Both tx vq and rx socket were polled here */
1008                 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1009
1010                 len = peek_head_len(rnvq, sk);
1011         }
1012
1013         return len;
1014 }
1015
1016 /* This is a multi-buffer version of vhost_get_desc, that works if
1017  *      vq has read descriptors only.
1018  * @vq          - the relevant virtqueue
1019  * @datalen     - data length we'll be reading
1020  * @iovcount    - returned count of io vectors we fill
1021  * @log         - vhost log
1022  * @log_num     - log offset
1023  * @quota       - headcount quota, 1 for big buffer
1024  *      returns number of buffer heads allocated, negative on error
1025  */
1026 static int get_rx_bufs(struct vhost_virtqueue *vq,
1027                        struct vring_used_elem *heads,
1028                        int datalen,
1029                        unsigned *iovcount,
1030                        struct vhost_log *log,
1031                        unsigned *log_num,
1032                        unsigned int quota)
1033 {
1034         unsigned int out, in;
1035         int seg = 0;
1036         int headcount = 0;
1037         unsigned d;
1038         int r, nlogs = 0;
1039         /* len is always initialized before use since we are always called with
1040          * datalen > 0.
1041          */
1042         u32 uninitialized_var(len);
1043
1044         while (datalen > 0 && headcount < quota) {
1045                 if (unlikely(seg >= UIO_MAXIOV)) {
1046                         r = -ENOBUFS;
1047                         goto err;
1048                 }
1049                 r = vhost_get_vq_desc(vq, vq->iov + seg,
1050                                       ARRAY_SIZE(vq->iov) - seg, &out,
1051                                       &in, log, log_num);
1052                 if (unlikely(r < 0))
1053                         goto err;
1054
1055                 d = r;
1056                 if (d == vq->num) {
1057                         r = 0;
1058                         goto err;
1059                 }
1060                 if (unlikely(out || in <= 0)) {
1061                         vq_err(vq, "unexpected descriptor format for RX: "
1062                                 "out %d, in %d\n", out, in);
1063                         r = -EINVAL;
1064                         goto err;
1065                 }
1066                 if (unlikely(log)) {
1067                         nlogs += *log_num;
1068                         log += *log_num;
1069                 }
1070                 heads[headcount].id = cpu_to_vhost32(vq, d);
1071                 len = iov_length(vq->iov + seg, in);
1072                 heads[headcount].len = cpu_to_vhost32(vq, len);
1073                 datalen -= len;
1074                 ++headcount;
1075                 seg += in;
1076         }
1077         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1078         *iovcount = seg;
1079         if (unlikely(log))
1080                 *log_num = nlogs;
1081
1082         /* Detect overrun */
1083         if (unlikely(datalen > 0)) {
1084                 r = UIO_MAXIOV + 1;
1085                 goto err;
1086         }
1087         return headcount;
1088 err:
1089         vhost_discard_vq_desc(vq, headcount);
1090         return r;
1091 }
1092
1093 /* Expects to be always run from workqueue - which acts as
1094  * read-size critical section for our kind of RCU. */
1095 static void handle_rx(struct vhost_net *net)
1096 {
1097         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1098         struct vhost_virtqueue *vq = &nvq->vq;
1099         unsigned uninitialized_var(in), log;
1100         struct vhost_log *vq_log;
1101         struct msghdr msg = {
1102                 .msg_name = NULL,
1103                 .msg_namelen = 0,
1104                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
1105                 .msg_controllen = 0,
1106                 .msg_flags = MSG_DONTWAIT,
1107         };
1108         struct virtio_net_hdr hdr = {
1109                 .flags = 0,
1110                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
1111         };
1112         size_t total_len = 0;
1113         int err, mergeable;
1114         s16 headcount;
1115         size_t vhost_hlen, sock_hlen;
1116         size_t vhost_len, sock_len;
1117         bool busyloop_intr = false;
1118         struct socket *sock;
1119         struct iov_iter fixup;
1120         __virtio16 num_buffers;
1121         int recv_pkts = 0;
1122
1123         mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1124         sock = vq->private_data;
1125         if (!sock)
1126                 goto out;
1127
1128         if (!vq_meta_prefetch(vq))
1129                 goto out;
1130
1131         vhost_disable_notify(&net->dev, vq);
1132         vhost_net_disable_vq(net, vq);
1133
1134         vhost_hlen = nvq->vhost_hlen;
1135         sock_hlen = nvq->sock_hlen;
1136
1137         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1138                 vq->log : NULL;
1139         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1140
1141         do {
1142                 sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1143                                                       &busyloop_intr);
1144                 if (!sock_len)
1145                         break;
1146                 sock_len += sock_hlen;
1147                 vhost_len = sock_len + vhost_hlen;
1148                 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1149                                         vhost_len, &in, vq_log, &log,
1150                                         likely(mergeable) ? UIO_MAXIOV : 1);
1151                 /* On error, stop handling until the next kick. */
1152                 if (unlikely(headcount < 0))
1153                         goto out;
1154                 /* OK, now we need to know about added descriptors. */
1155                 if (!headcount) {
1156                         if (unlikely(busyloop_intr)) {
1157                                 vhost_poll_queue(&vq->poll);
1158                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1159                                 /* They have slipped one in as we were
1160                                  * doing that: check again. */
1161                                 vhost_disable_notify(&net->dev, vq);
1162                                 continue;
1163                         }
1164                         /* Nothing new?  Wait for eventfd to tell us
1165                          * they refilled. */
1166                         goto out;
1167                 }
1168                 busyloop_intr = false;
1169                 if (nvq->rx_ring)
1170                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1171                 /* On overrun, truncate and discard */
1172                 if (unlikely(headcount > UIO_MAXIOV)) {
1173                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
1174                         err = sock->ops->recvmsg(sock, &msg,
1175                                                  1, MSG_DONTWAIT | MSG_TRUNC);
1176                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
1177                         continue;
1178                 }
1179                 /* We don't need to be notified again. */
1180                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
1181                 fixup = msg.msg_iter;
1182                 if (unlikely((vhost_hlen))) {
1183                         /* We will supply the header ourselves
1184                          * TODO: support TSO.
1185                          */
1186                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
1187                 }
1188                 err = sock->ops->recvmsg(sock, &msg,
1189                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
1190                 /* Userspace might have consumed the packet meanwhile:
1191                  * it's not supposed to do this usually, but might be hard
1192                  * to prevent. Discard data we got (if any) and keep going. */
1193                 if (unlikely(err != sock_len)) {
1194                         pr_debug("Discarded rx packet: "
1195                                  " len %d, expected %zd\n", err, sock_len);
1196                         vhost_discard_vq_desc(vq, headcount);
1197                         continue;
1198                 }
1199                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1200                 if (unlikely(vhost_hlen)) {
1201                         if (copy_to_iter(&hdr, sizeof(hdr),
1202                                          &fixup) != sizeof(hdr)) {
1203                                 vq_err(vq, "Unable to write vnet_hdr "
1204                                        "at addr %p\n", vq->iov->iov_base);
1205                                 goto out;
1206                         }
1207                 } else {
1208                         /* Header came from socket; we'll need to patch
1209                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1210                          */
1211                         iov_iter_advance(&fixup, sizeof(hdr));
1212                 }
1213                 /* TODO: Should check and handle checksum. */
1214
1215                 num_buffers = cpu_to_vhost16(vq, headcount);
1216                 if (likely(mergeable) &&
1217                     copy_to_iter(&num_buffers, sizeof num_buffers,
1218                                  &fixup) != sizeof num_buffers) {
1219                         vq_err(vq, "Failed num_buffers write");
1220                         vhost_discard_vq_desc(vq, headcount);
1221                         goto out;
1222                 }
1223                 nvq->done_idx += headcount;
1224                 if (nvq->done_idx > VHOST_NET_BATCH)
1225                         vhost_net_signal_used(nvq);
1226                 if (unlikely(vq_log))
1227                         vhost_log_write(vq, vq_log, log, vhost_len,
1228                                         vq->iov, in);
1229                 total_len += vhost_len;
1230         } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1231
1232         if (unlikely(busyloop_intr))
1233                 vhost_poll_queue(&vq->poll);
1234         else if (!sock_len)
1235                 vhost_net_enable_vq(net, vq);
1236 out:
1237         vhost_net_signal_used(nvq);
1238         mutex_unlock(&vq->mutex);
1239 }
1240
1241 static void handle_tx_kick(struct vhost_work *work)
1242 {
1243         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1244                                                   poll.work);
1245         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1246
1247         handle_tx(net);
1248 }
1249
1250 static void handle_rx_kick(struct vhost_work *work)
1251 {
1252         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1253                                                   poll.work);
1254         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1255
1256         handle_rx(net);
1257 }
1258
1259 static void handle_tx_net(struct vhost_work *work)
1260 {
1261         struct vhost_net *net = container_of(work, struct vhost_net,
1262                                              poll[VHOST_NET_VQ_TX].work);
1263         handle_tx(net);
1264 }
1265
1266 static void handle_rx_net(struct vhost_work *work)
1267 {
1268         struct vhost_net *net = container_of(work, struct vhost_net,
1269                                              poll[VHOST_NET_VQ_RX].work);
1270         handle_rx(net);
1271 }
1272
1273 static int vhost_net_open(struct inode *inode, struct file *f)
1274 {
1275         struct vhost_net *n;
1276         struct vhost_dev *dev;
1277         struct vhost_virtqueue **vqs;
1278         void **queue;
1279         struct xdp_buff *xdp;
1280         int i;
1281
1282         n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1283         if (!n)
1284                 return -ENOMEM;
1285         vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1286         if (!vqs) {
1287                 kvfree(n);
1288                 return -ENOMEM;
1289         }
1290
1291         queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1292                               GFP_KERNEL);
1293         if (!queue) {
1294                 kfree(vqs);
1295                 kvfree(n);
1296                 return -ENOMEM;
1297         }
1298         n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1299
1300         xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1301         if (!xdp) {
1302                 kfree(vqs);
1303                 kvfree(n);
1304                 kfree(queue);
1305                 return -ENOMEM;
1306         }
1307         n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1308
1309         dev = &n->dev;
1310         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1311         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1312         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1313         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1314         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1315                 n->vqs[i].ubufs = NULL;
1316                 n->vqs[i].ubuf_info = NULL;
1317                 n->vqs[i].upend_idx = 0;
1318                 n->vqs[i].done_idx = 0;
1319                 n->vqs[i].batched_xdp = 0;
1320                 n->vqs[i].vhost_hlen = 0;
1321                 n->vqs[i].sock_hlen = 0;
1322                 n->vqs[i].rx_ring = NULL;
1323                 vhost_net_buf_init(&n->vqs[i].rxq);
1324         }
1325         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1326                        UIO_MAXIOV + VHOST_NET_BATCH,
1327                        VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT);
1328
1329         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1330         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1331
1332         f->private_data = n;
1333         n->page_frag.page = NULL;
1334         n->refcnt_bias = 0;
1335
1336         return 0;
1337 }
1338
1339 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1340                                         struct vhost_virtqueue *vq)
1341 {
1342         struct socket *sock;
1343         struct vhost_net_virtqueue *nvq =
1344                 container_of(vq, struct vhost_net_virtqueue, vq);
1345
1346         mutex_lock(&vq->mutex);
1347         sock = vq->private_data;
1348         vhost_net_disable_vq(n, vq);
1349         vq->private_data = NULL;
1350         vhost_net_buf_unproduce(nvq);
1351         nvq->rx_ring = NULL;
1352         mutex_unlock(&vq->mutex);
1353         return sock;
1354 }
1355
1356 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1357                            struct socket **rx_sock)
1358 {
1359         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1360         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1361 }
1362
1363 static void vhost_net_flush_vq(struct vhost_net *n, int index)
1364 {
1365         vhost_poll_flush(n->poll + index);
1366         vhost_poll_flush(&n->vqs[index].vq.poll);
1367 }
1368
1369 static void vhost_net_flush(struct vhost_net *n)
1370 {
1371         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1372         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1373         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1374                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1375                 n->tx_flush = true;
1376                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1377                 /* Wait for all lower device DMAs done. */
1378                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1379                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1380                 n->tx_flush = false;
1381                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1382                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1383         }
1384 }
1385
1386 static int vhost_net_release(struct inode *inode, struct file *f)
1387 {
1388         struct vhost_net *n = f->private_data;
1389         struct socket *tx_sock;
1390         struct socket *rx_sock;
1391
1392         vhost_net_stop(n, &tx_sock, &rx_sock);
1393         vhost_net_flush(n);
1394         vhost_dev_stop(&n->dev);
1395         vhost_dev_cleanup(&n->dev);
1396         vhost_net_vq_reset(n);
1397         if (tx_sock)
1398                 sockfd_put(tx_sock);
1399         if (rx_sock)
1400                 sockfd_put(rx_sock);
1401         /* Make sure no callbacks are outstanding */
1402         synchronize_rcu();
1403         /* We do an extra flush before freeing memory,
1404          * since jobs can re-queue themselves. */
1405         vhost_net_flush(n);
1406         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1407         kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1408         kfree(n->dev.vqs);
1409         if (n->page_frag.page)
1410                 __page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
1411         kvfree(n);
1412         return 0;
1413 }
1414
1415 static struct socket *get_raw_socket(int fd)
1416 {
1417         int r;
1418         struct socket *sock = sockfd_lookup(fd, &r);
1419
1420         if (!sock)
1421                 return ERR_PTR(-ENOTSOCK);
1422
1423         /* Parameter checking */
1424         if (sock->sk->sk_type != SOCK_RAW) {
1425                 r = -ESOCKTNOSUPPORT;
1426                 goto err;
1427         }
1428
1429         if (sock->sk->sk_family != AF_PACKET) {
1430                 r = -EPFNOSUPPORT;
1431                 goto err;
1432         }
1433         return sock;
1434 err:
1435         sockfd_put(sock);
1436         return ERR_PTR(r);
1437 }
1438
1439 static struct ptr_ring *get_tap_ptr_ring(int fd)
1440 {
1441         struct ptr_ring *ring;
1442         struct file *file = fget(fd);
1443
1444         if (!file)
1445                 return NULL;
1446         ring = tun_get_tx_ring(file);
1447         if (!IS_ERR(ring))
1448                 goto out;
1449         ring = tap_get_ptr_ring(file);
1450         if (!IS_ERR(ring))
1451                 goto out;
1452         ring = NULL;
1453 out:
1454         fput(file);
1455         return ring;
1456 }
1457
1458 static struct socket *get_tap_socket(int fd)
1459 {
1460         struct file *file = fget(fd);
1461         struct socket *sock;
1462
1463         if (!file)
1464                 return ERR_PTR(-EBADF);
1465         sock = tun_get_socket(file);
1466         if (!IS_ERR(sock))
1467                 return sock;
1468         sock = tap_get_socket(file);
1469         if (IS_ERR(sock))
1470                 fput(file);
1471         return sock;
1472 }
1473
1474 static struct socket *get_socket(int fd)
1475 {
1476         struct socket *sock;
1477
1478         /* special case to disable backend */
1479         if (fd == -1)
1480                 return NULL;
1481         sock = get_raw_socket(fd);
1482         if (!IS_ERR(sock))
1483                 return sock;
1484         sock = get_tap_socket(fd);
1485         if (!IS_ERR(sock))
1486                 return sock;
1487         return ERR_PTR(-ENOTSOCK);
1488 }
1489
1490 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1491 {
1492         struct socket *sock, *oldsock;
1493         struct vhost_virtqueue *vq;
1494         struct vhost_net_virtqueue *nvq;
1495         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1496         int r;
1497
1498         mutex_lock(&n->dev.mutex);
1499         r = vhost_dev_check_owner(&n->dev);
1500         if (r)
1501                 goto err;
1502
1503         if (index >= VHOST_NET_VQ_MAX) {
1504                 r = -ENOBUFS;
1505                 goto err;
1506         }
1507         vq = &n->vqs[index].vq;
1508         nvq = &n->vqs[index];
1509         mutex_lock(&vq->mutex);
1510
1511         /* Verify that ring has been setup correctly. */
1512         if (!vhost_vq_access_ok(vq)) {
1513                 r = -EFAULT;
1514                 goto err_vq;
1515         }
1516         sock = get_socket(fd);
1517         if (IS_ERR(sock)) {
1518                 r = PTR_ERR(sock);
1519                 goto err_vq;
1520         }
1521
1522         /* start polling new socket */
1523         oldsock = vq->private_data;
1524         if (sock != oldsock) {
1525                 ubufs = vhost_net_ubuf_alloc(vq,
1526                                              sock && vhost_sock_zcopy(sock));
1527                 if (IS_ERR(ubufs)) {
1528                         r = PTR_ERR(ubufs);
1529                         goto err_ubufs;
1530                 }
1531
1532                 vhost_net_disable_vq(n, vq);
1533                 vq->private_data = sock;
1534                 vhost_net_buf_unproduce(nvq);
1535                 r = vhost_vq_init_access(vq);
1536                 if (r)
1537                         goto err_used;
1538                 r = vhost_net_enable_vq(n, vq);
1539                 if (r)
1540                         goto err_used;
1541                 if (index == VHOST_NET_VQ_RX)
1542                         nvq->rx_ring = get_tap_ptr_ring(fd);
1543
1544                 oldubufs = nvq->ubufs;
1545                 nvq->ubufs = ubufs;
1546
1547                 n->tx_packets = 0;
1548                 n->tx_zcopy_err = 0;
1549                 n->tx_flush = false;
1550         }
1551
1552         mutex_unlock(&vq->mutex);
1553
1554         if (oldubufs) {
1555                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1556                 mutex_lock(&vq->mutex);
1557                 vhost_zerocopy_signal_used(n, vq);
1558                 mutex_unlock(&vq->mutex);
1559         }
1560
1561         if (oldsock) {
1562                 vhost_net_flush_vq(n, index);
1563                 sockfd_put(oldsock);
1564         }
1565
1566         mutex_unlock(&n->dev.mutex);
1567         return 0;
1568
1569 err_used:
1570         vq->private_data = oldsock;
1571         vhost_net_enable_vq(n, vq);
1572         if (ubufs)
1573                 vhost_net_ubuf_put_wait_and_free(ubufs);
1574 err_ubufs:
1575         if (sock)
1576                 sockfd_put(sock);
1577 err_vq:
1578         mutex_unlock(&vq->mutex);
1579 err:
1580         mutex_unlock(&n->dev.mutex);
1581         return r;
1582 }
1583
1584 static long vhost_net_reset_owner(struct vhost_net *n)
1585 {
1586         struct socket *tx_sock = NULL;
1587         struct socket *rx_sock = NULL;
1588         long err;
1589         struct vhost_umem *umem;
1590
1591         mutex_lock(&n->dev.mutex);
1592         err = vhost_dev_check_owner(&n->dev);
1593         if (err)
1594                 goto done;
1595         umem = vhost_dev_reset_owner_prepare();
1596         if (!umem) {
1597                 err = -ENOMEM;
1598                 goto done;
1599         }
1600         vhost_net_stop(n, &tx_sock, &rx_sock);
1601         vhost_net_flush(n);
1602         vhost_dev_stop(&n->dev);
1603         vhost_dev_reset_owner(&n->dev, umem);
1604         vhost_net_vq_reset(n);
1605 done:
1606         mutex_unlock(&n->dev.mutex);
1607         if (tx_sock)
1608                 sockfd_put(tx_sock);
1609         if (rx_sock)
1610                 sockfd_put(rx_sock);
1611         return err;
1612 }
1613
1614 static int vhost_net_set_backend_features(struct vhost_net *n, u64 features)
1615 {
1616         int i;
1617
1618         mutex_lock(&n->dev.mutex);
1619         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1620                 mutex_lock(&n->vqs[i].vq.mutex);
1621                 n->vqs[i].vq.acked_backend_features = features;
1622                 mutex_unlock(&n->vqs[i].vq.mutex);
1623         }
1624         mutex_unlock(&n->dev.mutex);
1625
1626         return 0;
1627 }
1628
1629 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1630 {
1631         size_t vhost_hlen, sock_hlen, hdr_len;
1632         int i;
1633
1634         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1635                                (1ULL << VIRTIO_F_VERSION_1))) ?
1636                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1637                         sizeof(struct virtio_net_hdr);
1638         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1639                 /* vhost provides vnet_hdr */
1640                 vhost_hlen = hdr_len;
1641                 sock_hlen = 0;
1642         } else {
1643                 /* socket provides vnet_hdr */
1644                 vhost_hlen = 0;
1645                 sock_hlen = hdr_len;
1646         }
1647         mutex_lock(&n->dev.mutex);
1648         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1649             !vhost_log_access_ok(&n->dev))
1650                 goto out_unlock;
1651
1652         if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1653                 if (vhost_init_device_iotlb(&n->dev, true))
1654                         goto out_unlock;
1655         }
1656
1657         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1658                 mutex_lock(&n->vqs[i].vq.mutex);
1659                 n->vqs[i].vq.acked_features = features;
1660                 n->vqs[i].vhost_hlen = vhost_hlen;
1661                 n->vqs[i].sock_hlen = sock_hlen;
1662                 mutex_unlock(&n->vqs[i].vq.mutex);
1663         }
1664         mutex_unlock(&n->dev.mutex);
1665         return 0;
1666
1667 out_unlock:
1668         mutex_unlock(&n->dev.mutex);
1669         return -EFAULT;
1670 }
1671
1672 static long vhost_net_set_owner(struct vhost_net *n)
1673 {
1674         int r;
1675
1676         mutex_lock(&n->dev.mutex);
1677         if (vhost_dev_has_owner(&n->dev)) {
1678                 r = -EBUSY;
1679                 goto out;
1680         }
1681         r = vhost_net_set_ubuf_info(n);
1682         if (r)
1683                 goto out;
1684         r = vhost_dev_set_owner(&n->dev);
1685         if (r)
1686                 vhost_net_clear_ubuf_info(n);
1687         vhost_net_flush(n);
1688 out:
1689         mutex_unlock(&n->dev.mutex);
1690         return r;
1691 }
1692
1693 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1694                             unsigned long arg)
1695 {
1696         struct vhost_net *n = f->private_data;
1697         void __user *argp = (void __user *)arg;
1698         u64 __user *featurep = argp;
1699         struct vhost_vring_file backend;
1700         u64 features;
1701         int r;
1702
1703         switch (ioctl) {
1704         case VHOST_NET_SET_BACKEND:
1705                 if (copy_from_user(&backend, argp, sizeof backend))
1706                         return -EFAULT;
1707                 return vhost_net_set_backend(n, backend.index, backend.fd);
1708         case VHOST_GET_FEATURES:
1709                 features = VHOST_NET_FEATURES;
1710                 if (copy_to_user(featurep, &features, sizeof features))
1711                         return -EFAULT;
1712                 return 0;
1713         case VHOST_SET_FEATURES:
1714                 if (copy_from_user(&features, featurep, sizeof features))
1715                         return -EFAULT;
1716                 if (features & ~VHOST_NET_FEATURES)
1717                         return -EOPNOTSUPP;
1718                 return vhost_net_set_features(n, features);
1719         case VHOST_GET_BACKEND_FEATURES:
1720                 features = VHOST_NET_BACKEND_FEATURES;
1721                 if (copy_to_user(featurep, &features, sizeof(features)))
1722                         return -EFAULT;
1723                 return 0;
1724         case VHOST_SET_BACKEND_FEATURES:
1725                 if (copy_from_user(&features, featurep, sizeof(features)))
1726                         return -EFAULT;
1727                 if (features & ~VHOST_NET_BACKEND_FEATURES)
1728                         return -EOPNOTSUPP;
1729                 return vhost_net_set_backend_features(n, features);
1730         case VHOST_RESET_OWNER:
1731                 return vhost_net_reset_owner(n);
1732         case VHOST_SET_OWNER:
1733                 return vhost_net_set_owner(n);
1734         default:
1735                 mutex_lock(&n->dev.mutex);
1736                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1737                 if (r == -ENOIOCTLCMD)
1738                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1739                 else
1740                         vhost_net_flush(n);
1741                 mutex_unlock(&n->dev.mutex);
1742                 return r;
1743         }
1744 }
1745
1746 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1747 {
1748         struct file *file = iocb->ki_filp;
1749         struct vhost_net *n = file->private_data;
1750         struct vhost_dev *dev = &n->dev;
1751         int noblock = file->f_flags & O_NONBLOCK;
1752
1753         return vhost_chr_read_iter(dev, to, noblock);
1754 }
1755
1756 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1757                                         struct iov_iter *from)
1758 {
1759         struct file *file = iocb->ki_filp;
1760         struct vhost_net *n = file->private_data;
1761         struct vhost_dev *dev = &n->dev;
1762
1763         return vhost_chr_write_iter(dev, from);
1764 }
1765
1766 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1767 {
1768         struct vhost_net *n = file->private_data;
1769         struct vhost_dev *dev = &n->dev;
1770
1771         return vhost_chr_poll(file, dev, wait);
1772 }
1773
1774 static const struct file_operations vhost_net_fops = {
1775         .owner          = THIS_MODULE,
1776         .release        = vhost_net_release,
1777         .read_iter      = vhost_net_chr_read_iter,
1778         .write_iter     = vhost_net_chr_write_iter,
1779         .poll           = vhost_net_chr_poll,
1780         .unlocked_ioctl = vhost_net_ioctl,
1781         .compat_ioctl   = compat_ptr_ioctl,
1782         .open           = vhost_net_open,
1783         .llseek         = noop_llseek,
1784 };
1785
1786 static struct miscdevice vhost_net_misc = {
1787         .minor = VHOST_NET_MINOR,
1788         .name = "vhost-net",
1789         .fops = &vhost_net_fops,
1790 };
1791
1792 static int vhost_net_init(void)
1793 {
1794         if (experimental_zcopytx)
1795                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1796         return misc_register(&vhost_net_misc);
1797 }
1798 module_init(vhost_net_init);
1799
1800 static void vhost_net_exit(void)
1801 {
1802         misc_deregister(&vhost_net_misc);
1803 }
1804 module_exit(vhost_net_exit);
1805
1806 MODULE_VERSION("0.0.1");
1807 MODULE_LICENSE("GPL v2");
1808 MODULE_AUTHOR("Michael S. Tsirkin");
1809 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1810 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1811 MODULE_ALIAS("devname:vhost-net");