drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <mst@redhat.com>
   3  *
   4  * This work is licensed under the terms of the GNU GPL, version 2.
   5  *
   6  * virtio-net server in host kernel.
   7  */
   8
   9 #include <linux/compat.h>
  10 #include <linux/eventfd.h>
  11 #include <linux/vhost.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/module.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/mutex.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/vmalloc.h>
  21
  22 #include <linux/net.h>
  23 #include <linux/if_packet.h>
  24 #include <linux/if_arp.h>
  25 #include <linux/if_tun.h>
  26 #include <linux/if_macvlan.h>
  27 #include <linux/if_vlan.h>
  28
  29 #include <net/sock.h>
  30
  31 #include "vhost.h"
  32
  33 static int experimental_zcopytx = 1;
  34 module_param(experimental_zcopytx, int, 0444);
  35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  36                                        " 1 -Enable; 0 - Disable");
  37
  38 /* Max number of bytes transferred before requeueing the job.
  39  * Using this limit prevents one virtqueue from starving others. */
  40 #define VHOST_NET_WEIGHT 0x80000
  41
  42 /* MAX number of TX used buffers for outstanding zerocopy */
  43 #define VHOST_MAX_PEND 128
  44 #define VHOST_GOODCOPY_LEN 256
  45
  46 /*
  47  * For transmit, used buffer len is unused; we override it to track buffer
  48  * status internally; used for zerocopy tx only.
  49  */
  50 /* Lower device DMA failed */
  51 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  52 /* Lower device DMA done */
  53 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  54 /* Lower device DMA in progress */
  55 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  56 /* Buffer unused */
  57 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  58
  59 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  60
  61 enum {
  62         VHOST_NET_FEATURES = VHOST_FEATURES |
  63                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  64                          (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  65                          (1ULL << VIRTIO_F_VERSION_1),
  66 };
  67
  68 enum {
  69         VHOST_NET_VQ_RX = 0,
  70         VHOST_NET_VQ_TX = 1,
  71         VHOST_NET_VQ_MAX = 2,
  72 };
  73
  74 struct vhost_net_ubuf_ref {
  75         /* refcount follows semantics similar to kref:
  76          *  0: object is released
  77          *  1: no outstanding ubufs
  78          * >1: outstanding ubufs
  79          */
  80         atomic_t refcount;
  81         wait_queue_head_t wait;
  82         struct vhost_virtqueue *vq;
  83 };
  84
  85 struct vhost_net_virtqueue {
  86         struct vhost_virtqueue vq;
  87         /* hdr is used to store the virtio header.
  88          * Since each iovec has >= 1 byte length, we never need more than
  89          * header length entries to store the header. */
  90         struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
  91         size_t vhost_hlen;
  92         size_t sock_hlen;
  93         /* vhost zerocopy support fields below: */
  94         /* last used idx for outstanding DMA zerocopy buffers */
  95         int upend_idx;
  96         /* first used idx for DMA done zerocopy buffers */
  97         int done_idx;
  98         /* an array of userspace buffers info */
  99         struct ubuf_info *ubuf_info;
 100         /* Reference counting for outstanding ubufs.
 101          * Protected by vq mutex. Writers must also take device mutex. */
 102         struct vhost_net_ubuf_ref *ubufs;
 103 };
 104
 105 struct vhost_net {
 106         struct vhost_dev dev;
 107         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 108         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 109         /* Number of TX recently submitted.
 110          * Protected by tx vq lock. */
 111         unsigned tx_packets;
 112         /* Number of times zerocopy TX recently failed.
 113          * Protected by tx vq lock. */
 114         unsigned tx_zcopy_err;
 115         /* Flush in progress. Protected by tx vq lock. */
 116         bool tx_flush;
 117 };
 118
 119 static unsigned vhost_net_zcopy_mask __read_mostly;
 120
 121 static void vhost_net_enable_zcopy(int vq)
 122 {
 123         vhost_net_zcopy_mask |= 0x1 << vq;
 124 }
 125
 126 static struct vhost_net_ubuf_ref *
 127 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 128 {
 129         struct vhost_net_ubuf_ref *ubufs;
 130         /* No zero copy backend? Nothing to count. */
 131         if (!zcopy)
 132                 return NULL;
 133         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 134         if (!ubufs)
 135                 return ERR_PTR(-ENOMEM);
 136         atomic_set(&ubufs->refcount, 1);
 137         init_waitqueue_head(&ubufs->wait);
 138         ubufs->vq = vq;
 139         return ubufs;
 140 }
 141
 142 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 143 {
 144         int r = atomic_sub_return(1, &ubufs->refcount);
 145         if (unlikely(!r))
 146                 wake_up(&ubufs->wait);
 147         return r;
 148 }
 149
 150 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 151 {
 152         vhost_net_ubuf_put(ubufs);
 153         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 154 }
 155
 156 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 157 {
 158         vhost_net_ubuf_put_and_wait(ubufs);
 159         kfree(ubufs);
 160 }
 161
 162 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 163 {
 164         int i;
 165
 166         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 167                 kfree(n->vqs[i].ubuf_info);
 168                 n->vqs[i].ubuf_info = NULL;
 169         }
 170 }
 171
 172 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 173 {
 174         bool zcopy;
 175         int i;
 176
 177         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 178                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 179                 if (!zcopy)
 180                         continue;
 181                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 182                                               UIO_MAXIOV, GFP_KERNEL);
 183                 if  (!n->vqs[i].ubuf_info)
 184                         goto err;
 185         }
 186         return 0;
 187
 188 err:
 189         vhost_net_clear_ubuf_info(n);
 190         return -ENOMEM;
 191 }
 192
 193 static void vhost_net_vq_reset(struct vhost_net *n)
 194 {
 195         int i;
 196
 197         vhost_net_clear_ubuf_info(n);
 198
 199         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 200                 n->vqs[i].done_idx = 0;
 201                 n->vqs[i].upend_idx = 0;
 202                 n->vqs[i].ubufs = NULL;
 203                 n->vqs[i].vhost_hlen = 0;
 204                 n->vqs[i].sock_hlen = 0;
 205         }
 206
 207 }
 208
 209 static void vhost_net_tx_packet(struct vhost_net *net)
 210 {
 211         ++net->tx_packets;
 212         if (net->tx_packets < 1024)
 213                 return;
 214         net->tx_packets = 0;
 215         net->tx_zcopy_err = 0;
 216 }
 217
 218 static void vhost_net_tx_err(struct vhost_net *net)
 219 {
 220         ++net->tx_zcopy_err;
 221 }
 222
 223 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 224 {
 225         /* TX flush waits for outstanding DMAs to be done.
 226          * Don't start new DMAs.
 227          */
 228         return !net->tx_flush &&
 229                 net->tx_packets / 64 >= net->tx_zcopy_err;
 230 }
 231
 232 static bool vhost_sock_zcopy(struct socket *sock)
 233 {
 234         return unlikely(experimental_zcopytx) &&
 235                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 236 }
 237
 238 /* Pop first len bytes from iovec. Return number of segments used. */
 239 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 240                           size_t len, int iov_count)
 241 {
 242         int seg = 0;
 243         size_t size;
 244
 245         while (len && seg < iov_count) {
 246                 size = min(from->iov_len, len);
 247                 to->iov_base = from->iov_base;
 248                 to->iov_len = size;
 249                 from->iov_len -= size;
 250                 from->iov_base += size;
 251                 len -= size;
 252                 ++from;
 253                 ++to;
 254                 ++seg;
 255         }
 256         return seg;
 257 }
 258 /* Copy iovec entries for len bytes from iovec. */
 259 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 260                            size_t len, int iovcount)
 261 {
 262         int seg = 0;
 263         size_t size;
 264
 265         while (len && seg < iovcount) {
 266                 size = min(from->iov_len, len);
 267                 to->iov_base = from->iov_base;
 268                 to->iov_len = size;
 269                 len -= size;
 270                 ++from;
 271                 ++to;
 272                 ++seg;
 273         }
 274 }
 275
 276 /* In case of DMA done not in order in lower device driver for some reason.
 277  * upend_idx is used to track end of used idx, done_idx is used to track head
 278  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 279  * guest used idx.
 280  */
 281 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 282                                        struct vhost_virtqueue *vq)
 283 {
 284         struct vhost_net_virtqueue *nvq =
 285                 container_of(vq, struct vhost_net_virtqueue, vq);
 286         int i, add;
 287         int j = 0;
 288
 289         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 290                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 291                         vhost_net_tx_err(net);
 292                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 293                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 294                         ++j;
 295                 } else
 296                         break;
 297         }
 298         while (j) {
 299                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 300                 vhost_add_used_and_signal_n(vq->dev, vq,
 301                                             &vq->heads[nvq->done_idx], add);
 302                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 303                 j -= add;
 304         }
 305 }
 306
 307 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 308 {
 309         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 310         struct vhost_virtqueue *vq = ubufs->vq;
 311         int cnt;
 312
 313         rcu_read_lock_bh();
 314
 315         /* set len to mark this desc buffers done DMA */
 316         vq->heads[ubuf->desc].len = success ?
 317                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 318         cnt = vhost_net_ubuf_put(ubufs);
 319
 320         /*
 321          * Trigger polling thread if guest stopped submitting new buffers:
 322          * in this case, the refcount after decrement will eventually reach 1.
 323          * We also trigger polling periodically after each 16 packets
 324          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 325          * less than 10% of times).
 326          */
 327         if (cnt <= 1 || !(cnt % 16))
 328                 vhost_poll_queue(&vq->poll);
 329
 330         rcu_read_unlock_bh();
 331 }
 332
 333 /* Expects to be always run from workqueue - which acts as
 334  * read-size critical section for our kind of RCU. */
 335 static void handle_tx(struct vhost_net *net)
 336 {
 337         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 338         struct vhost_virtqueue *vq = &nvq->vq;
 339         unsigned out, in;
 340         int head;
 341         struct msghdr msg = {
 342                 .msg_name = NULL,
 343                 .msg_namelen = 0,
 344                 .msg_control = NULL,
 345                 .msg_controllen = 0,
 346                 .msg_flags = MSG_DONTWAIT,
 347         };
 348         size_t len, total_len = 0;
 349         int err;
 350         size_t hdr_size;
 351         struct socket *sock;
 352         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 353         bool zcopy, zcopy_used;
 354
 355         mutex_lock(&vq->mutex);
 356         sock = vq->private_data;
 357         if (!sock)
 358                 goto out;
 359
 360         vhost_disable_notify(&net->dev, vq);
 361
 362         hdr_size = nvq->vhost_hlen;
 363         zcopy = nvq->ubufs;
 364
 365         for (;;) {
 366                 /* Release DMAs done buffers first */
 367                 if (zcopy)
 368                         vhost_zerocopy_signal_used(net, vq);
 369
 370                 /* If more outstanding DMAs, queue the work.
 371                  * Handle upend_idx wrap around
 372                  */
 373                 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
 374                               % UIO_MAXIOV == nvq->done_idx))
 375                         break;
 376
 377                 head = vhost_get_vq_desc(vq, vq->iov,
 378                                          ARRAY_SIZE(vq->iov),
 379                                          &out, &in,
 380                                          NULL, NULL);
 381                 /* On error, stop handling until the next kick. */
 382                 if (unlikely(head < 0))
 383                         break;
 384                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 385                 if (head == vq->num) {
 386                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 387                                 vhost_disable_notify(&net->dev, vq);
 388                                 continue;
 389                         }
 390                         break;
 391                 }
 392                 if (in) {
 393                         vq_err(vq, "Unexpected descriptor format for TX: "
 394                                "out %d, int %d\n", out, in);
 395                         break;
 396                 }
 397                 /* Skip header. TODO: support TSO. */
 398                 len = iov_length(vq->iov, out);
 399                 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
 400                 iov_iter_advance(&msg.msg_iter, hdr_size);
 401                 /* Sanity check */
 402                 if (!iov_iter_count(&msg.msg_iter)) {
 403                         vq_err(vq, "Unexpected header len for TX: "
 404                                "%zd expected %zd\n",
 405                                len, hdr_size);
 406                         break;
 407                 }
 408                 len = iov_iter_count(&msg.msg_iter);
 409
 410                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 411                                    && (nvq->upend_idx + 1) % UIO_MAXIOV !=
 412                                       nvq->done_idx
 413                                    && vhost_net_tx_select_zcopy(net);
 414
 415                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 416                 if (zcopy_used) {
 417                         struct ubuf_info *ubuf;
 418                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 419
 420                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 421                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 422                         ubuf->callback = vhost_zerocopy_callback;
 423                         ubuf->ctx = nvq->ubufs;
 424                         ubuf->desc = nvq->upend_idx;
 425                         msg.msg_control = ubuf;
 426                         msg.msg_controllen = sizeof(ubuf);
 427                         ubufs = nvq->ubufs;
 428                         atomic_inc(&ubufs->refcount);
 429                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 430                 } else {
 431                         msg.msg_control = NULL;
 432                         ubufs = NULL;
 433                 }
 434                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 435                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 436                 if (unlikely(err < 0)) {
 437                         if (zcopy_used) {
 438                                 vhost_net_ubuf_put(ubufs);
 439                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 440                                         % UIO_MAXIOV;
 441                         }
 442                         vhost_discard_vq_desc(vq, 1);
 443                         break;
 444                 }
 445                 if (err != len)
 446                         pr_debug("Truncated TX packet: "
 447                                  " len %d != %zd\n", err, len);
 448                 if (!zcopy_used)
 449                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 450                 else
 451                         vhost_zerocopy_signal_used(net, vq);
 452                 total_len += len;
 453                 vhost_net_tx_packet(net);
 454                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 455                         vhost_poll_queue(&vq->poll);
 456                         break;
 457                 }
 458         }
 459 out:
 460         mutex_unlock(&vq->mutex);
 461 }
 462
 463 static int peek_head_len(struct sock *sk)
 464 {
 465         struct sk_buff *head;
 466         int len = 0;
 467         unsigned long flags;
 468
 469         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 470         head = skb_peek(&sk->sk_receive_queue);
 471         if (likely(head)) {
 472                 len = head->len;
 473                 if (skb_vlan_tag_present(head))
 474                         len += VLAN_HLEN;
 475         }
 476
 477         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 478         return len;
 479 }
 480
 481 /* This is a multi-buffer version of vhost_get_desc, that works if
 482  *      vq has read descriptors only.
 483  * @vq          - the relevant virtqueue
 484  * @datalen     - data length we'll be reading
 485  * @iovcount    - returned count of io vectors we fill
 486  * @log         - vhost log
 487  * @log_num     - log offset
 488  * @quota       - headcount quota, 1 for big buffer
 489  *      returns number of buffer heads allocated, negative on error
 490  */
 491 static int get_rx_bufs(struct vhost_virtqueue *vq,
 492                        struct vring_used_elem *heads,
 493                        int datalen,
 494                        unsigned *iovcount,
 495                        struct vhost_log *log,
 496                        unsigned *log_num,
 497                        unsigned int quota)
 498 {
 499         unsigned int out, in;
 500         int seg = 0;
 501         int headcount = 0;
 502         unsigned d;
 503         int r, nlogs = 0;
 504         /* len is always initialized before use since we are always called with
 505          * datalen > 0.
 506          */
 507         u32 uninitialized_var(len);
 508
 509         while (datalen > 0 && headcount < quota) {
 510                 if (unlikely(seg >= UIO_MAXIOV)) {
 511                         r = -ENOBUFS;
 512                         goto err;
 513                 }
 514                 r = vhost_get_vq_desc(vq, vq->iov + seg,
 515                                       ARRAY_SIZE(vq->iov) - seg, &out,
 516                                       &in, log, log_num);
 517                 if (unlikely(r < 0))
 518                         goto err;
 519
 520                 d = r;
 521                 if (d == vq->num) {
 522                         r = 0;
 523                         goto err;
 524                 }
 525                 if (unlikely(out || in <= 0)) {
 526                         vq_err(vq, "unexpected descriptor format for RX: "
 527                                 "out %d, in %d\n", out, in);
 528                         r = -EINVAL;
 529                         goto err;
 530                 }
 531                 if (unlikely(log)) {
 532                         nlogs += *log_num;
 533                         log += *log_num;
 534                 }
 535                 heads[headcount].id = cpu_to_vhost32(vq, d);
 536                 len = iov_length(vq->iov + seg, in);
 537                 heads[headcount].len = cpu_to_vhost32(vq, len);
 538                 datalen -= len;
 539                 ++headcount;
 540                 seg += in;
 541         }
 542         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
 543         *iovcount = seg;
 544         if (unlikely(log))
 545                 *log_num = nlogs;
 546
 547         /* Detect overrun */
 548         if (unlikely(datalen > 0)) {
 549                 r = UIO_MAXIOV + 1;
 550                 goto err;
 551         }
 552         return headcount;
 553 err:
 554         vhost_discard_vq_desc(vq, headcount);
 555         return r;
 556 }
 557
 558 /* Expects to be always run from workqueue - which acts as
 559  * read-size critical section for our kind of RCU. */
 560 static void handle_rx(struct vhost_net *net)
 561 {
 562         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 563         struct vhost_virtqueue *vq = &nvq->vq;
 564         unsigned uninitialized_var(in), log;
 565         struct vhost_log *vq_log;
 566         struct msghdr msg = {
 567                 .msg_name = NULL,
 568                 .msg_namelen = 0,
 569                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 570                 .msg_controllen = 0,
 571                 .msg_flags = MSG_DONTWAIT,
 572         };
 573         struct virtio_net_hdr_mrg_rxbuf hdr = {
 574                 .hdr.flags = 0,
 575                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 576         };
 577         size_t total_len = 0;
 578         int err, mergeable;
 579         s16 headcount;
 580         size_t vhost_hlen, sock_hlen;
 581         size_t vhost_len, sock_len;
 582         struct socket *sock;
 583
 584         mutex_lock(&vq->mutex);
 585         sock = vq->private_data;
 586         if (!sock)
 587                 goto out;
 588         vhost_disable_notify(&net->dev, vq);
 589
 590         vhost_hlen = nvq->vhost_hlen;
 591         sock_hlen = nvq->sock_hlen;
 592
 593         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
 594                 vq->log : NULL;
 595         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
 596
 597         while ((sock_len = peek_head_len(sock->sk))) {
 598                 sock_len += sock_hlen;
 599                 vhost_len = sock_len + vhost_hlen;
 600                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 601                                         &in, vq_log, &log,
 602                                         likely(mergeable) ? UIO_MAXIOV : 1);
 603                 /* On error, stop handling until the next kick. */
 604                 if (unlikely(headcount < 0))
 605                         break;
 606                 /* On overrun, truncate and discard */
 607                 if (unlikely(headcount > UIO_MAXIOV)) {
 608                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
 609                         err = sock->ops->recvmsg(NULL, sock, &msg,
 610                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 611                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 612                         continue;
 613                 }
 614                 /* OK, now we need to know about added descriptors. */
 615                 if (!headcount) {
 616                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 617                                 /* They have slipped one in as we were
 618                                  * doing that: check again. */
 619                                 vhost_disable_notify(&net->dev, vq);
 620                                 continue;
 621                         }
 622                         /* Nothing new?  Wait for eventfd to tell us
 623                          * they refilled. */
 624                         break;
 625                 }
 626                 /* We don't need to be notified again. */
 627                 if (unlikely((vhost_hlen)))
 628                         /* Skip header. TODO: support TSO. */
 629                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
 630                 else
 631                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 632                          * needed because recvmsg can modify msg_iov. */
 633                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
 634                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, sock_len);
 635                 err = sock->ops->recvmsg(NULL, sock, &msg,
 636                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 637                 /* Userspace might have consumed the packet meanwhile:
 638                  * it's not supposed to do this usually, but might be hard
 639                  * to prevent. Discard data we got (if any) and keep going. */
 640                 if (unlikely(err != sock_len)) {
 641                         pr_debug("Discarded rx packet: "
 642                                  " len %d, expected %zd\n", err, sock_len);
 643                         vhost_discard_vq_desc(vq, headcount);
 644                         continue;
 645                 }
 646                 if (unlikely(vhost_hlen) &&
 647                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
 648                                       vhost_hlen)) {
 649                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 650                                vq->iov->iov_base);
 651                         break;
 652                 }
 653                 /* TODO: Should check and handle checksum. */
 654                 if (likely(mergeable) &&
 655                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
 656                                       offsetof(typeof(hdr), num_buffers),
 657                                       sizeof hdr.num_buffers)) {
 658                         vq_err(vq, "Failed num_buffers write");
 659                         vhost_discard_vq_desc(vq, headcount);
 660                         break;
 661                 }
 662                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 663                                             headcount);
 664                 if (unlikely(vq_log))
 665                         vhost_log_write(vq, vq_log, log, vhost_len);
 666                 total_len += vhost_len;
 667                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 668                         vhost_poll_queue(&vq->poll);
 669                         break;
 670                 }
 671         }
 672 out:
 673         mutex_unlock(&vq->mutex);
 674 }
 675
 676 static void handle_tx_kick(struct vhost_work *work)
 677 {
 678         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 679                                                   poll.work);
 680         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 681
 682         handle_tx(net);
 683 }
 684
 685 static void handle_rx_kick(struct vhost_work *work)
 686 {
 687         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 688                                                   poll.work);
 689         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 690
 691         handle_rx(net);
 692 }
 693
 694 static void handle_tx_net(struct vhost_work *work)
 695 {
 696         struct vhost_net *net = container_of(work, struct vhost_net,
 697                                              poll[VHOST_NET_VQ_TX].work);
 698         handle_tx(net);
 699 }
 700
 701 static void handle_rx_net(struct vhost_work *work)
 702 {
 703         struct vhost_net *net = container_of(work, struct vhost_net,
 704                                              poll[VHOST_NET_VQ_RX].work);
 705         handle_rx(net);
 706 }
 707
 708 static int vhost_net_open(struct inode *inode, struct file *f)
 709 {
 710         struct vhost_net *n;
 711         struct vhost_dev *dev;
 712         struct vhost_virtqueue **vqs;
 713         int i;
 714
 715         n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 716         if (!n) {
 717                 n = vmalloc(sizeof *n);
 718                 if (!n)
 719                         return -ENOMEM;
 720         }
 721         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 722         if (!vqs) {
 723                 kvfree(n);
 724                 return -ENOMEM;
 725         }
 726
 727         dev = &n->dev;
 728         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 729         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 730         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 731         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 732         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 733                 n->vqs[i].ubufs = NULL;
 734                 n->vqs[i].ubuf_info = NULL;
 735                 n->vqs[i].upend_idx = 0;
 736                 n->vqs[i].done_idx = 0;
 737                 n->vqs[i].vhost_hlen = 0;
 738                 n->vqs[i].sock_hlen = 0;
 739         }
 740         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 741
 742         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 743         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 744
 745         f->private_data = n;
 746
 747         return 0;
 748 }
 749
 750 static void vhost_net_disable_vq(struct vhost_net *n,
 751                                  struct vhost_virtqueue *vq)
 752 {
 753         struct vhost_net_virtqueue *nvq =
 754                 container_of(vq, struct vhost_net_virtqueue, vq);
 755         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 756         if (!vq->private_data)
 757                 return;
 758         vhost_poll_stop(poll);
 759 }
 760
 761 static int vhost_net_enable_vq(struct vhost_net *n,
 762                                 struct vhost_virtqueue *vq)
 763 {
 764         struct vhost_net_virtqueue *nvq =
 765                 container_of(vq, struct vhost_net_virtqueue, vq);
 766         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 767         struct socket *sock;
 768
 769         sock = vq->private_data;
 770         if (!sock)
 771                 return 0;
 772
 773         return vhost_poll_start(poll, sock->file);
 774 }
 775
 776 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 777                                         struct vhost_virtqueue *vq)
 778 {
 779         struct socket *sock;
 780
 781         mutex_lock(&vq->mutex);
 782         sock = vq->private_data;
 783         vhost_net_disable_vq(n, vq);
 784         vq->private_data = NULL;
 785         mutex_unlock(&vq->mutex);
 786         return sock;
 787 }
 788
 789 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 790                            struct socket **rx_sock)
 791 {
 792         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 793         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 794 }
 795
 796 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 797 {
 798         vhost_poll_flush(n->poll + index);
 799         vhost_poll_flush(&n->vqs[index].vq.poll);
 800 }
 801
 802 static void vhost_net_flush(struct vhost_net *n)
 803 {
 804         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 805         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 806         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 807                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 808                 n->tx_flush = true;
 809                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 810                 /* Wait for all lower device DMAs done. */
 811                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 812                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 813                 n->tx_flush = false;
 814                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
 815                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 816         }
 817 }
 818
 819 static int vhost_net_release(struct inode *inode, struct file *f)
 820 {
 821         struct vhost_net *n = f->private_data;
 822         struct socket *tx_sock;
 823         struct socket *rx_sock;
 824
 825         vhost_net_stop(n, &tx_sock, &rx_sock);
 826         vhost_net_flush(n);
 827         vhost_dev_stop(&n->dev);
 828         vhost_dev_cleanup(&n->dev, false);
 829         vhost_net_vq_reset(n);
 830         if (tx_sock)
 831                 sockfd_put(tx_sock);
 832         if (rx_sock)
 833                 sockfd_put(rx_sock);
 834         /* Make sure no callbacks are outstanding */
 835         synchronize_rcu_bh();
 836         /* We do an extra flush before freeing memory,
 837          * since jobs can re-queue themselves. */
 838         vhost_net_flush(n);
 839         kfree(n->dev.vqs);
 840         kvfree(n);
 841         return 0;
 842 }
 843
 844 static struct socket *get_raw_socket(int fd)
 845 {
 846         struct {
 847                 struct sockaddr_ll sa;
 848                 char  buf[MAX_ADDR_LEN];
 849         } uaddr;
 850         int uaddr_len = sizeof uaddr, r;
 851         struct socket *sock = sockfd_lookup(fd, &r);
 852
 853         if (!sock)
 854                 return ERR_PTR(-ENOTSOCK);
 855
 856         /* Parameter checking */
 857         if (sock->sk->sk_type != SOCK_RAW) {
 858                 r = -ESOCKTNOSUPPORT;
 859                 goto err;
 860         }
 861
 862         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 863                                &uaddr_len, 0);
 864         if (r)
 865                 goto err;
 866
 867         if (uaddr.sa.sll_family != AF_PACKET) {
 868                 r = -EPFNOSUPPORT;
 869                 goto err;
 870         }
 871         return sock;
 872 err:
 873         sockfd_put(sock);
 874         return ERR_PTR(r);
 875 }
 876
 877 static struct socket *get_tap_socket(int fd)
 878 {
 879         struct file *file = fget(fd);
 880         struct socket *sock;
 881
 882         if (!file)
 883                 return ERR_PTR(-EBADF);
 884         sock = tun_get_socket(file);
 885         if (!IS_ERR(sock))
 886                 return sock;
 887         sock = macvtap_get_socket(file);
 888         if (IS_ERR(sock))
 889                 fput(file);
 890         return sock;
 891 }
 892
 893 static struct socket *get_socket(int fd)
 894 {
 895         struct socket *sock;
 896
 897         /* special case to disable backend */
 898         if (fd == -1)
 899                 return NULL;
 900         sock = get_raw_socket(fd);
 901         if (!IS_ERR(sock))
 902                 return sock;
 903         sock = get_tap_socket(fd);
 904         if (!IS_ERR(sock))
 905                 return sock;
 906         return ERR_PTR(-ENOTSOCK);
 907 }
 908
 909 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 910 {
 911         struct socket *sock, *oldsock;
 912         struct vhost_virtqueue *vq;
 913         struct vhost_net_virtqueue *nvq;
 914         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 915         int r;
 916
 917         mutex_lock(&n->dev.mutex);
 918         r = vhost_dev_check_owner(&n->dev);
 919         if (r)
 920                 goto err;
 921
 922         if (index >= VHOST_NET_VQ_MAX) {
 923                 r = -ENOBUFS;
 924                 goto err;
 925         }
 926         vq = &n->vqs[index].vq;
 927         nvq = &n->vqs[index];
 928         mutex_lock(&vq->mutex);
 929
 930         /* Verify that ring has been setup correctly. */
 931         if (!vhost_vq_access_ok(vq)) {
 932                 r = -EFAULT;
 933                 goto err_vq;
 934         }
 935         sock = get_socket(fd);
 936         if (IS_ERR(sock)) {
 937                 r = PTR_ERR(sock);
 938                 goto err_vq;
 939         }
 940
 941         /* start polling new socket */
 942         oldsock = vq->private_data;
 943         if (sock != oldsock) {
 944                 ubufs = vhost_net_ubuf_alloc(vq,
 945                                              sock && vhost_sock_zcopy(sock));
 946                 if (IS_ERR(ubufs)) {
 947                         r = PTR_ERR(ubufs);
 948                         goto err_ubufs;
 949                 }
 950
 951                 vhost_net_disable_vq(n, vq);
 952                 vq->private_data = sock;
 953                 r = vhost_init_used(vq);
 954                 if (r)
 955                         goto err_used;
 956                 r = vhost_net_enable_vq(n, vq);
 957                 if (r)
 958                         goto err_used;
 959
 960                 oldubufs = nvq->ubufs;
 961                 nvq->ubufs = ubufs;
 962
 963                 n->tx_packets = 0;
 964                 n->tx_zcopy_err = 0;
 965                 n->tx_flush = false;
 966         }
 967
 968         mutex_unlock(&vq->mutex);
 969
 970         if (oldubufs) {
 971                 vhost_net_ubuf_put_wait_and_free(oldubufs);
 972                 mutex_lock(&vq->mutex);
 973                 vhost_zerocopy_signal_used(n, vq);
 974                 mutex_unlock(&vq->mutex);
 975         }
 976
 977         if (oldsock) {
 978                 vhost_net_flush_vq(n, index);
 979                 sockfd_put(oldsock);
 980         }
 981
 982         mutex_unlock(&n->dev.mutex);
 983         return 0;
 984
 985 err_used:
 986         vq->private_data = oldsock;
 987         vhost_net_enable_vq(n, vq);
 988         if (ubufs)
 989                 vhost_net_ubuf_put_wait_and_free(ubufs);
 990 err_ubufs:
 991         sockfd_put(sock);
 992 err_vq:
 993         mutex_unlock(&vq->mutex);
 994 err:
 995         mutex_unlock(&n->dev.mutex);
 996         return r;
 997 }
 998
 999 static long vhost_net_reset_owner(struct vhost_net *n)
1000 {
1001         struct socket *tx_sock = NULL;
1002         struct socket *rx_sock = NULL;
1003         long err;
1004         struct vhost_memory *memory;
1005
1006         mutex_lock(&n->dev.mutex);
1007         err = vhost_dev_check_owner(&n->dev);
1008         if (err)
1009                 goto done;
1010         memory = vhost_dev_reset_owner_prepare();
1011         if (!memory) {
1012                 err = -ENOMEM;
1013                 goto done;
1014         }
1015         vhost_net_stop(n, &tx_sock, &rx_sock);
1016         vhost_net_flush(n);
1017         vhost_dev_reset_owner(&n->dev, memory);
1018         vhost_net_vq_reset(n);
1019 done:
1020         mutex_unlock(&n->dev.mutex);
1021         if (tx_sock)
1022                 sockfd_put(tx_sock);
1023         if (rx_sock)
1024                 sockfd_put(rx_sock);
1025         return err;
1026 }
1027
1028 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1029 {
1030         size_t vhost_hlen, sock_hlen, hdr_len;
1031         int i;
1032
1033         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1034                                (1ULL << VIRTIO_F_VERSION_1))) ?
1035                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1036                         sizeof(struct virtio_net_hdr);
1037         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1038                 /* vhost provides vnet_hdr */
1039                 vhost_hlen = hdr_len;
1040                 sock_hlen = 0;
1041         } else {
1042                 /* socket provides vnet_hdr */
1043                 vhost_hlen = 0;
1044                 sock_hlen = hdr_len;
1045         }
1046         mutex_lock(&n->dev.mutex);
1047         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1048             !vhost_log_access_ok(&n->dev)) {
1049                 mutex_unlock(&n->dev.mutex);
1050                 return -EFAULT;
1051         }
1052         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1053                 mutex_lock(&n->vqs[i].vq.mutex);
1054                 n->vqs[i].vq.acked_features = features;
1055                 n->vqs[i].vhost_hlen = vhost_hlen;
1056                 n->vqs[i].sock_hlen = sock_hlen;
1057                 mutex_unlock(&n->vqs[i].vq.mutex);
1058         }
1059         mutex_unlock(&n->dev.mutex);
1060         return 0;
1061 }
1062
1063 static long vhost_net_set_owner(struct vhost_net *n)
1064 {
1065         int r;
1066
1067         mutex_lock(&n->dev.mutex);
1068         if (vhost_dev_has_owner(&n->dev)) {
1069                 r = -EBUSY;
1070                 goto out;
1071         }
1072         r = vhost_net_set_ubuf_info(n);
1073         if (r)
1074                 goto out;
1075         r = vhost_dev_set_owner(&n->dev);
1076         if (r)
1077                 vhost_net_clear_ubuf_info(n);
1078         vhost_net_flush(n);
1079 out:
1080         mutex_unlock(&n->dev.mutex);
1081         return r;
1082 }
1083
1084 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1085                             unsigned long arg)
1086 {
1087         struct vhost_net *n = f->private_data;
1088         void __user *argp = (void __user *)arg;
1089         u64 __user *featurep = argp;
1090         struct vhost_vring_file backend;
1091         u64 features;
1092         int r;
1093
1094         switch (ioctl) {
1095         case VHOST_NET_SET_BACKEND:
1096                 if (copy_from_user(&backend, argp, sizeof backend))
1097                         return -EFAULT;
1098                 return vhost_net_set_backend(n, backend.index, backend.fd);
1099         case VHOST_GET_FEATURES:
1100                 features = VHOST_NET_FEATURES;
1101                 if (copy_to_user(featurep, &features, sizeof features))
1102                         return -EFAULT;
1103                 return 0;
1104         case VHOST_SET_FEATURES:
1105                 if (copy_from_user(&features, featurep, sizeof features))
1106                         return -EFAULT;
1107                 if (features & ~VHOST_NET_FEATURES)
1108                         return -EOPNOTSUPP;
1109                 return vhost_net_set_features(n, features);
1110         case VHOST_RESET_OWNER:
1111                 return vhost_net_reset_owner(n);
1112         case VHOST_SET_OWNER:
1113                 return vhost_net_set_owner(n);
1114         default:
1115                 mutex_lock(&n->dev.mutex);
1116                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1117                 if (r == -ENOIOCTLCMD)
1118                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1119                 else
1120                         vhost_net_flush(n);
1121                 mutex_unlock(&n->dev.mutex);
1122                 return r;
1123         }
1124 }
1125
1126 #ifdef CONFIG_COMPAT
1127 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1128                                    unsigned long arg)
1129 {
1130         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1131 }
1132 #endif
1133
1134 static const struct file_operations vhost_net_fops = {
1135         .owner          = THIS_MODULE,
1136         .release        = vhost_net_release,
1137         .unlocked_ioctl = vhost_net_ioctl,
1138 #ifdef CONFIG_COMPAT
1139         .compat_ioctl   = vhost_net_compat_ioctl,
1140 #endif
1141         .open           = vhost_net_open,
1142         .llseek         = noop_llseek,
1143 };
1144
1145 static struct miscdevice vhost_net_misc = {
1146         .minor = VHOST_NET_MINOR,
1147         .name = "vhost-net",
1148         .fops = &vhost_net_fops,
1149 };
1150
1151 static int vhost_net_init(void)
1152 {
1153         if (experimental_zcopytx)
1154                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1155         return misc_register(&vhost_net_misc);
1156 }
1157 module_init(vhost_net_init);
1158
1159 static void vhost_net_exit(void)
1160 {
1161         misc_deregister(&vhost_net_misc);
1162 }
1163 module_exit(vhost_net_exit);
1164
1165 MODULE_VERSION("0.0.1");
1166 MODULE_LICENSE("GPL v2");
1167 MODULE_AUTHOR("Michael S. Tsirkin");
1168 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1169 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1170 MODULE_ALIAS("devname:vhost-net");