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_IOMMU_PLATFORM) 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 size_t vhost_hlen; 88 size_t sock_hlen; 89 /* vhost zerocopy support fields below: */ 90 /* last used idx for outstanding DMA zerocopy buffers */ 91 int upend_idx; 92 /* first used idx for DMA done zerocopy buffers */ 93 int done_idx; 94 /* an array of userspace buffers info */ 95 struct ubuf_info *ubuf_info; 96 /* Reference counting for outstanding ubufs. 97 * Protected by vq mutex. Writers must also take device mutex. */ 98 struct vhost_net_ubuf_ref *ubufs; 99 }; 100 101 struct vhost_net { 102 struct vhost_dev dev; 103 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX]; 104 struct vhost_poll poll[VHOST_NET_VQ_MAX]; 105 /* Number of TX recently submitted. 106 * Protected by tx vq lock. */ 107 unsigned tx_packets; 108 /* Number of times zerocopy TX recently failed. 109 * Protected by tx vq lock. */ 110 unsigned tx_zcopy_err; 111 /* Flush in progress. Protected by tx vq lock. */ 112 bool tx_flush; 113 }; 114 115 static unsigned vhost_net_zcopy_mask __read_mostly; 116 117 static void vhost_net_enable_zcopy(int vq) 118 { 119 vhost_net_zcopy_mask |= 0x1 << vq; 120 } 121 122 static struct vhost_net_ubuf_ref * 123 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy) 124 { 125 struct vhost_net_ubuf_ref *ubufs; 126 /* No zero copy backend? Nothing to count. */ 127 if (!zcopy) 128 return NULL; 129 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL); 130 if (!ubufs) 131 return ERR_PTR(-ENOMEM); 132 atomic_set(&ubufs->refcount, 1); 133 init_waitqueue_head(&ubufs->wait); 134 ubufs->vq = vq; 135 return ubufs; 136 } 137 138 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs) 139 { 140 int r = atomic_sub_return(1, &ubufs->refcount); 141 if (unlikely(!r)) 142 wake_up(&ubufs->wait); 143 return r; 144 } 145 146 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs) 147 { 148 vhost_net_ubuf_put(ubufs); 149 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount)); 150 } 151 152 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs) 153 { 154 vhost_net_ubuf_put_and_wait(ubufs); 155 kfree(ubufs); 156 } 157 158 static void vhost_net_clear_ubuf_info(struct vhost_net *n) 159 { 160 int i; 161 162 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 163 kfree(n->vqs[i].ubuf_info); 164 n->vqs[i].ubuf_info = NULL; 165 } 166 } 167 168 static int vhost_net_set_ubuf_info(struct vhost_net *n) 169 { 170 bool zcopy; 171 int i; 172 173 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 174 zcopy = vhost_net_zcopy_mask & (0x1 << i); 175 if (!zcopy) 176 continue; 177 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) * 178 UIO_MAXIOV, GFP_KERNEL); 179 if (!n->vqs[i].ubuf_info) 180 goto err; 181 } 182 return 0; 183 184 err: 185 vhost_net_clear_ubuf_info(n); 186 return -ENOMEM; 187 } 188 189 static void vhost_net_vq_reset(struct vhost_net *n) 190 { 191 int i; 192 193 vhost_net_clear_ubuf_info(n); 194 195 for (i = 0; i < VHOST_NET_VQ_MAX; i++) { 196 n->vqs[i].done_idx = 0; 197 n->vqs[i].upend_idx = 0; 198 n->vqs[i].ubufs = NULL; 199 n->vqs[i].vhost_hlen = 0; 200 n->vqs[i].sock_hlen = 0; 201 } 202 203 } 204 205 static void vhost_net_tx_packet(struct vhost_net *net) 206 { 207 ++net->tx_packets; 208 if (net->tx_packets < 1024) 209 return; 210 net->tx_packets = 0; 211 net->tx_zcopy_err = 0; 212 } 213 214 static void vhost_net_tx_err(struct vhost_net *net) 215 { 216 ++net->tx_zcopy_err; 217 } 218 219 static bool vhost_net_tx_select_zcopy(struct vhost_net *net) 220 { 221 /* TX flush waits for outstanding DMAs to be done. 222 * Don't start new DMAs. 223 */ 224 return !net->tx_flush && 225 net->tx_packets / 64 >= net->tx_zcopy_err; 226 } 227 228 static bool vhost_sock_zcopy(struct socket *sock) 229 { 230 return unlikely(experimental_zcopytx) && 231 sock_flag(sock->sk, SOCK_ZEROCOPY); 232 } 233 234 /* In case of DMA done not in order in lower device driver for some reason. 235 * upend_idx is used to track end of used idx, done_idx is used to track head 236 * of used idx. Once lower device DMA done contiguously, we will signal KVM 237 * guest used idx. 238 */ 239 static void vhost_zerocopy_signal_used(struct vhost_net *net, 240 struct vhost_virtqueue *vq) 241 { 242 struct vhost_net_virtqueue *nvq = 243 container_of(vq, struct vhost_net_virtqueue, vq); 244 int i, add; 245 int j = 0; 246 247 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) { 248 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN) 249 vhost_net_tx_err(net); 250 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) { 251 vq->heads[i].len = VHOST_DMA_CLEAR_LEN; 252 ++j; 253 } else 254 break; 255 } 256 while (j) { 257 add = min(UIO_MAXIOV - nvq->done_idx, j); 258 vhost_add_used_and_signal_n(vq->dev, vq, 259 &vq->heads[nvq->done_idx], add); 260 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV; 261 j -= add; 262 } 263 } 264 265 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success) 266 { 267 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx; 268 struct vhost_virtqueue *vq = ubufs->vq; 269 int cnt; 270 271 rcu_read_lock_bh(); 272 273 /* set len to mark this desc buffers done DMA */ 274 vq->heads[ubuf->desc].len = success ? 275 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN; 276 cnt = vhost_net_ubuf_put(ubufs); 277 278 /* 279 * Trigger polling thread if guest stopped submitting new buffers: 280 * in this case, the refcount after decrement will eventually reach 1. 281 * We also trigger polling periodically after each 16 packets 282 * (the value 16 here is more or less arbitrary, it's tuned to trigger 283 * less than 10% of times). 284 */ 285 if (cnt <= 1 || !(cnt % 16)) 286 vhost_poll_queue(&vq->poll); 287 288 rcu_read_unlock_bh(); 289 } 290 291 static inline unsigned long busy_clock(void) 292 { 293 return local_clock() >> 10; 294 } 295 296 static bool vhost_can_busy_poll(struct vhost_dev *dev, 297 unsigned long endtime) 298 { 299 return likely(!need_resched()) && 300 likely(!time_after(busy_clock(), endtime)) && 301 likely(!signal_pending(current)) && 302 !vhost_has_work(dev); 303 } 304 305 static void vhost_net_disable_vq(struct vhost_net *n, 306 struct vhost_virtqueue *vq) 307 { 308 struct vhost_net_virtqueue *nvq = 309 container_of(vq, struct vhost_net_virtqueue, vq); 310 struct vhost_poll *poll = n->poll + (nvq - n->vqs); 311 if (!vq->private_data) 312 return; 313 vhost_poll_stop(poll); 314 } 315 316 static int vhost_net_enable_vq(struct vhost_net *n, 317 struct vhost_virtqueue *vq) 318 { 319 struct vhost_net_virtqueue *nvq = 320 container_of(vq, struct vhost_net_virtqueue, vq); 321 struct vhost_poll *poll = n->poll + (nvq - n->vqs); 322 struct socket *sock; 323 324 sock = vq->private_data; 325 if (!sock) 326 return 0; 327 328 return vhost_poll_start(poll, sock->file); 329 } 330 331 static int vhost_net_tx_get_vq_desc(struct vhost_net *net, 332 struct vhost_virtqueue *vq, 333 struct iovec iov[], unsigned int iov_size, 334 unsigned int *out_num, unsigned int *in_num) 335 { 336 unsigned long uninitialized_var(endtime); 337 int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov), 338 out_num, in_num, NULL, NULL); 339 340 if (r == vq->num && vq->busyloop_timeout) { 341 preempt_disable(); 342 endtime = busy_clock() + vq->busyloop_timeout; 343 while (vhost_can_busy_poll(vq->dev, endtime) && 344 vhost_vq_avail_empty(vq->dev, vq)) 345 cpu_relax(); 346 preempt_enable(); 347 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov), 348 out_num, in_num, NULL, NULL); 349 } 350 351 return r; 352 } 353 354 static bool vhost_exceeds_maxpend(struct vhost_net *net) 355 { 356 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 357 struct vhost_virtqueue *vq = &nvq->vq; 358 359 return (nvq->upend_idx + vq->num - VHOST_MAX_PEND) % UIO_MAXIOV 360 == nvq->done_idx; 361 } 362 363 /* Expects to be always run from workqueue - which acts as 364 * read-size critical section for our kind of RCU. */ 365 static void handle_tx(struct vhost_net *net) 366 { 367 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 368 struct vhost_virtqueue *vq = &nvq->vq; 369 unsigned out, in; 370 int head; 371 struct msghdr msg = { 372 .msg_name = NULL, 373 .msg_namelen = 0, 374 .msg_control = NULL, 375 .msg_controllen = 0, 376 .msg_flags = MSG_DONTWAIT, 377 }; 378 size_t len, total_len = 0; 379 int err; 380 size_t hdr_size; 381 struct socket *sock; 382 struct vhost_net_ubuf_ref *uninitialized_var(ubufs); 383 bool zcopy, zcopy_used; 384 385 mutex_lock(&vq->mutex); 386 sock = vq->private_data; 387 if (!sock) 388 goto out; 389 390 if (!vq_iotlb_prefetch(vq)) 391 goto out; 392 393 vhost_disable_notify(&net->dev, vq); 394 395 hdr_size = nvq->vhost_hlen; 396 zcopy = nvq->ubufs; 397 398 for (;;) { 399 /* Release DMAs done buffers first */ 400 if (zcopy) 401 vhost_zerocopy_signal_used(net, vq); 402 403 /* If more outstanding DMAs, queue the work. 404 * Handle upend_idx wrap around 405 */ 406 if (unlikely(vhost_exceeds_maxpend(net))) 407 break; 408 409 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov, 410 ARRAY_SIZE(vq->iov), 411 &out, &in); 412 /* On error, stop handling until the next kick. */ 413 if (unlikely(head < 0)) 414 break; 415 /* Nothing new? Wait for eventfd to tell us they refilled. */ 416 if (head == vq->num) { 417 if (unlikely(vhost_enable_notify(&net->dev, vq))) { 418 vhost_disable_notify(&net->dev, vq); 419 continue; 420 } 421 break; 422 } 423 if (in) { 424 vq_err(vq, "Unexpected descriptor format for TX: " 425 "out %d, int %d\n", out, in); 426 break; 427 } 428 /* Skip header. TODO: support TSO. */ 429 len = iov_length(vq->iov, out); 430 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len); 431 iov_iter_advance(&msg.msg_iter, hdr_size); 432 /* Sanity check */ 433 if (!msg_data_left(&msg)) { 434 vq_err(vq, "Unexpected header len for TX: " 435 "%zd expected %zd\n", 436 len, hdr_size); 437 break; 438 } 439 len = msg_data_left(&msg); 440 441 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN 442 && (nvq->upend_idx + 1) % UIO_MAXIOV != 443 nvq->done_idx 444 && vhost_net_tx_select_zcopy(net); 445 446 /* use msg_control to pass vhost zerocopy ubuf info to skb */ 447 if (zcopy_used) { 448 struct ubuf_info *ubuf; 449 ubuf = nvq->ubuf_info + nvq->upend_idx; 450 451 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head); 452 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS; 453 ubuf->callback = vhost_zerocopy_callback; 454 ubuf->ctx = nvq->ubufs; 455 ubuf->desc = nvq->upend_idx; 456 msg.msg_control = ubuf; 457 msg.msg_controllen = sizeof(ubuf); 458 ubufs = nvq->ubufs; 459 atomic_inc(&ubufs->refcount); 460 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV; 461 } else { 462 msg.msg_control = NULL; 463 ubufs = NULL; 464 } 465 466 total_len += len; 467 if (total_len < VHOST_NET_WEIGHT && 468 !vhost_vq_avail_empty(&net->dev, vq) && 469 likely(!vhost_exceeds_maxpend(net))) { 470 msg.msg_flags |= MSG_MORE; 471 } else { 472 msg.msg_flags &= ~MSG_MORE; 473 } 474 475 /* TODO: Check specific error and bomb out unless ENOBUFS? */ 476 err = sock->ops->sendmsg(sock, &msg, len); 477 if (unlikely(err < 0)) { 478 if (zcopy_used) { 479 vhost_net_ubuf_put(ubufs); 480 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1) 481 % UIO_MAXIOV; 482 } 483 vhost_discard_vq_desc(vq, 1); 484 break; 485 } 486 if (err != len) 487 pr_debug("Truncated TX packet: " 488 " len %d != %zd\n", err, len); 489 if (!zcopy_used) 490 vhost_add_used_and_signal(&net->dev, vq, head, 0); 491 else 492 vhost_zerocopy_signal_used(net, vq); 493 vhost_net_tx_packet(net); 494 if (unlikely(total_len >= VHOST_NET_WEIGHT)) { 495 vhost_poll_queue(&vq->poll); 496 break; 497 } 498 } 499 out: 500 mutex_unlock(&vq->mutex); 501 } 502 503 static int peek_head_len(struct sock *sk) 504 { 505 struct socket *sock = sk->sk_socket; 506 struct sk_buff *head; 507 int len = 0; 508 unsigned long flags; 509 510 if (sock->ops->peek_len) 511 return sock->ops->peek_len(sock); 512 513 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags); 514 head = skb_peek(&sk->sk_receive_queue); 515 if (likely(head)) { 516 len = head->len; 517 if (skb_vlan_tag_present(head)) 518 len += VLAN_HLEN; 519 } 520 521 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags); 522 return len; 523 } 524 525 static int sk_has_rx_data(struct sock *sk) 526 { 527 struct socket *sock = sk->sk_socket; 528 529 if (sock->ops->peek_len) 530 return sock->ops->peek_len(sock); 531 532 return skb_queue_empty(&sk->sk_receive_queue); 533 } 534 535 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk) 536 { 537 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 538 struct vhost_virtqueue *vq = &nvq->vq; 539 unsigned long uninitialized_var(endtime); 540 int len = peek_head_len(sk); 541 542 if (!len && vq->busyloop_timeout) { 543 /* Both tx vq and rx socket were polled here */ 544 mutex_lock(&vq->mutex); 545 vhost_disable_notify(&net->dev, vq); 546 547 preempt_disable(); 548 endtime = busy_clock() + vq->busyloop_timeout; 549 550 while (vhost_can_busy_poll(&net->dev, endtime) && 551 !sk_has_rx_data(sk) && 552 vhost_vq_avail_empty(&net->dev, vq)) 553 cpu_relax(); 554 555 preempt_enable(); 556 557 if (vhost_enable_notify(&net->dev, vq)) 558 vhost_poll_queue(&vq->poll); 559 mutex_unlock(&vq->mutex); 560 561 len = peek_head_len(sk); 562 } 563 564 return len; 565 } 566 567 /* This is a multi-buffer version of vhost_get_desc, that works if 568 * vq has read descriptors only. 569 * @vq - the relevant virtqueue 570 * @datalen - data length we'll be reading 571 * @iovcount - returned count of io vectors we fill 572 * @log - vhost log 573 * @log_num - log offset 574 * @quota - headcount quota, 1 for big buffer 575 * returns number of buffer heads allocated, negative on error 576 */ 577 static int get_rx_bufs(struct vhost_virtqueue *vq, 578 struct vring_used_elem *heads, 579 int datalen, 580 unsigned *iovcount, 581 struct vhost_log *log, 582 unsigned *log_num, 583 unsigned int quota) 584 { 585 unsigned int out, in; 586 int seg = 0; 587 int headcount = 0; 588 unsigned d; 589 int r, nlogs = 0; 590 /* len is always initialized before use since we are always called with 591 * datalen > 0. 592 */ 593 u32 uninitialized_var(len); 594 595 while (datalen > 0 && headcount < quota) { 596 if (unlikely(seg >= UIO_MAXIOV)) { 597 r = -ENOBUFS; 598 goto err; 599 } 600 r = vhost_get_vq_desc(vq, vq->iov + seg, 601 ARRAY_SIZE(vq->iov) - seg, &out, 602 &in, log, log_num); 603 if (unlikely(r < 0)) 604 goto err; 605 606 d = r; 607 if (d == vq->num) { 608 r = 0; 609 goto err; 610 } 611 if (unlikely(out || in <= 0)) { 612 vq_err(vq, "unexpected descriptor format for RX: " 613 "out %d, in %d\n", out, in); 614 r = -EINVAL; 615 goto err; 616 } 617 if (unlikely(log)) { 618 nlogs += *log_num; 619 log += *log_num; 620 } 621 heads[headcount].id = cpu_to_vhost32(vq, d); 622 len = iov_length(vq->iov + seg, in); 623 heads[headcount].len = cpu_to_vhost32(vq, len); 624 datalen -= len; 625 ++headcount; 626 seg += in; 627 } 628 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen); 629 *iovcount = seg; 630 if (unlikely(log)) 631 *log_num = nlogs; 632 633 /* Detect overrun */ 634 if (unlikely(datalen > 0)) { 635 r = UIO_MAXIOV + 1; 636 goto err; 637 } 638 return headcount; 639 err: 640 vhost_discard_vq_desc(vq, headcount); 641 return r; 642 } 643 644 /* Expects to be always run from workqueue - which acts as 645 * read-size critical section for our kind of RCU. */ 646 static void handle_rx(struct vhost_net *net) 647 { 648 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX]; 649 struct vhost_virtqueue *vq = &nvq->vq; 650 unsigned uninitialized_var(in), log; 651 struct vhost_log *vq_log; 652 struct msghdr msg = { 653 .msg_name = NULL, 654 .msg_namelen = 0, 655 .msg_control = NULL, /* FIXME: get and handle RX aux data. */ 656 .msg_controllen = 0, 657 .msg_flags = MSG_DONTWAIT, 658 }; 659 struct virtio_net_hdr hdr = { 660 .flags = 0, 661 .gso_type = VIRTIO_NET_HDR_GSO_NONE 662 }; 663 size_t total_len = 0; 664 int err, mergeable; 665 s16 headcount; 666 size_t vhost_hlen, sock_hlen; 667 size_t vhost_len, sock_len; 668 struct socket *sock; 669 struct iov_iter fixup; 670 __virtio16 num_buffers; 671 672 mutex_lock(&vq->mutex); 673 sock = vq->private_data; 674 if (!sock) 675 goto out; 676 677 if (!vq_iotlb_prefetch(vq)) 678 goto out; 679 680 vhost_disable_notify(&net->dev, vq); 681 vhost_net_disable_vq(net, vq); 682 683 vhost_hlen = nvq->vhost_hlen; 684 sock_hlen = nvq->sock_hlen; 685 686 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ? 687 vq->log : NULL; 688 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF); 689 690 while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) { 691 sock_len += sock_hlen; 692 vhost_len = sock_len + vhost_hlen; 693 headcount = get_rx_bufs(vq, vq->heads, vhost_len, 694 &in, vq_log, &log, 695 likely(mergeable) ? UIO_MAXIOV : 1); 696 /* On error, stop handling until the next kick. */ 697 if (unlikely(headcount < 0)) 698 goto out; 699 /* On overrun, truncate and discard */ 700 if (unlikely(headcount > UIO_MAXIOV)) { 701 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1); 702 err = sock->ops->recvmsg(sock, &msg, 703 1, MSG_DONTWAIT | MSG_TRUNC); 704 pr_debug("Discarded rx packet: len %zd\n", sock_len); 705 continue; 706 } 707 /* OK, now we need to know about added descriptors. */ 708 if (!headcount) { 709 if (unlikely(vhost_enable_notify(&net->dev, vq))) { 710 /* They have slipped one in as we were 711 * doing that: check again. */ 712 vhost_disable_notify(&net->dev, vq); 713 continue; 714 } 715 /* Nothing new? Wait for eventfd to tell us 716 * they refilled. */ 717 goto out; 718 } 719 /* We don't need to be notified again. */ 720 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len); 721 fixup = msg.msg_iter; 722 if (unlikely((vhost_hlen))) { 723 /* We will supply the header ourselves 724 * TODO: support TSO. 725 */ 726 iov_iter_advance(&msg.msg_iter, vhost_hlen); 727 } 728 err = sock->ops->recvmsg(sock, &msg, 729 sock_len, MSG_DONTWAIT | MSG_TRUNC); 730 /* Userspace might have consumed the packet meanwhile: 731 * it's not supposed to do this usually, but might be hard 732 * to prevent. Discard data we got (if any) and keep going. */ 733 if (unlikely(err != sock_len)) { 734 pr_debug("Discarded rx packet: " 735 " len %d, expected %zd\n", err, sock_len); 736 vhost_discard_vq_desc(vq, headcount); 737 continue; 738 } 739 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */ 740 if (unlikely(vhost_hlen)) { 741 if (copy_to_iter(&hdr, sizeof(hdr), 742 &fixup) != sizeof(hdr)) { 743 vq_err(vq, "Unable to write vnet_hdr " 744 "at addr %p\n", vq->iov->iov_base); 745 goto out; 746 } 747 } else { 748 /* Header came from socket; we'll need to patch 749 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF 750 */ 751 iov_iter_advance(&fixup, sizeof(hdr)); 752 } 753 /* TODO: Should check and handle checksum. */ 754 755 num_buffers = cpu_to_vhost16(vq, headcount); 756 if (likely(mergeable) && 757 copy_to_iter(&num_buffers, sizeof num_buffers, 758 &fixup) != sizeof num_buffers) { 759 vq_err(vq, "Failed num_buffers write"); 760 vhost_discard_vq_desc(vq, headcount); 761 goto out; 762 } 763 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads, 764 headcount); 765 if (unlikely(vq_log)) 766 vhost_log_write(vq, vq_log, log, vhost_len); 767 total_len += vhost_len; 768 if (unlikely(total_len >= VHOST_NET_WEIGHT)) { 769 vhost_poll_queue(&vq->poll); 770 goto out; 771 } 772 } 773 vhost_net_enable_vq(net, vq); 774 out: 775 mutex_unlock(&vq->mutex); 776 } 777 778 static void handle_tx_kick(struct vhost_work *work) 779 { 780 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, 781 poll.work); 782 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); 783 784 handle_tx(net); 785 } 786 787 static void handle_rx_kick(struct vhost_work *work) 788 { 789 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, 790 poll.work); 791 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); 792 793 handle_rx(net); 794 } 795 796 static void handle_tx_net(struct vhost_work *work) 797 { 798 struct vhost_net *net = container_of(work, struct vhost_net, 799 poll[VHOST_NET_VQ_TX].work); 800 handle_tx(net); 801 } 802 803 static void handle_rx_net(struct vhost_work *work) 804 { 805 struct vhost_net *net = container_of(work, struct vhost_net, 806 poll[VHOST_NET_VQ_RX].work); 807 handle_rx(net); 808 } 809 810 static int vhost_net_open(struct inode *inode, struct file *f) 811 { 812 struct vhost_net *n; 813 struct vhost_dev *dev; 814 struct vhost_virtqueue **vqs; 815 int i; 816 817 n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT); 818 if (!n) { 819 n = vmalloc(sizeof *n); 820 if (!n) 821 return -ENOMEM; 822 } 823 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL); 824 if (!vqs) { 825 kvfree(n); 826 return -ENOMEM; 827 } 828 829 dev = &n->dev; 830 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq; 831 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq; 832 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick; 833 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick; 834 for (i = 0; i < VHOST_NET_VQ_MAX; i++) { 835 n->vqs[i].ubufs = NULL; 836 n->vqs[i].ubuf_info = NULL; 837 n->vqs[i].upend_idx = 0; 838 n->vqs[i].done_idx = 0; 839 n->vqs[i].vhost_hlen = 0; 840 n->vqs[i].sock_hlen = 0; 841 } 842 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX); 843 844 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev); 845 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev); 846 847 f->private_data = n; 848 849 return 0; 850 } 851 852 static struct socket *vhost_net_stop_vq(struct vhost_net *n, 853 struct vhost_virtqueue *vq) 854 { 855 struct socket *sock; 856 857 mutex_lock(&vq->mutex); 858 sock = vq->private_data; 859 vhost_net_disable_vq(n, vq); 860 vq->private_data = NULL; 861 mutex_unlock(&vq->mutex); 862 return sock; 863 } 864 865 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock, 866 struct socket **rx_sock) 867 { 868 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq); 869 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq); 870 } 871 872 static void vhost_net_flush_vq(struct vhost_net *n, int index) 873 { 874 vhost_poll_flush(n->poll + index); 875 vhost_poll_flush(&n->vqs[index].vq.poll); 876 } 877 878 static void vhost_net_flush(struct vhost_net *n) 879 { 880 vhost_net_flush_vq(n, VHOST_NET_VQ_TX); 881 vhost_net_flush_vq(n, VHOST_NET_VQ_RX); 882 if (n->vqs[VHOST_NET_VQ_TX].ubufs) { 883 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 884 n->tx_flush = true; 885 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 886 /* Wait for all lower device DMAs done. */ 887 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs); 888 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 889 n->tx_flush = false; 890 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1); 891 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 892 } 893 } 894 895 static int vhost_net_release(struct inode *inode, struct file *f) 896 { 897 struct vhost_net *n = f->private_data; 898 struct socket *tx_sock; 899 struct socket *rx_sock; 900 901 vhost_net_stop(n, &tx_sock, &rx_sock); 902 vhost_net_flush(n); 903 vhost_dev_stop(&n->dev); 904 vhost_dev_cleanup(&n->dev, false); 905 vhost_net_vq_reset(n); 906 if (tx_sock) 907 sockfd_put(tx_sock); 908 if (rx_sock) 909 sockfd_put(rx_sock); 910 /* Make sure no callbacks are outstanding */ 911 synchronize_rcu_bh(); 912 /* We do an extra flush before freeing memory, 913 * since jobs can re-queue themselves. */ 914 vhost_net_flush(n); 915 kfree(n->dev.vqs); 916 kvfree(n); 917 return 0; 918 } 919 920 static struct socket *get_raw_socket(int fd) 921 { 922 struct { 923 struct sockaddr_ll sa; 924 char buf[MAX_ADDR_LEN]; 925 } uaddr; 926 int uaddr_len = sizeof uaddr, r; 927 struct socket *sock = sockfd_lookup(fd, &r); 928 929 if (!sock) 930 return ERR_PTR(-ENOTSOCK); 931 932 /* Parameter checking */ 933 if (sock->sk->sk_type != SOCK_RAW) { 934 r = -ESOCKTNOSUPPORT; 935 goto err; 936 } 937 938 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa, 939 &uaddr_len, 0); 940 if (r) 941 goto err; 942 943 if (uaddr.sa.sll_family != AF_PACKET) { 944 r = -EPFNOSUPPORT; 945 goto err; 946 } 947 return sock; 948 err: 949 sockfd_put(sock); 950 return ERR_PTR(r); 951 } 952 953 static struct socket *get_tap_socket(int fd) 954 { 955 struct file *file = fget(fd); 956 struct socket *sock; 957 958 if (!file) 959 return ERR_PTR(-EBADF); 960 sock = tun_get_socket(file); 961 if (!IS_ERR(sock)) 962 return sock; 963 sock = macvtap_get_socket(file); 964 if (IS_ERR(sock)) 965 fput(file); 966 return sock; 967 } 968 969 static struct socket *get_socket(int fd) 970 { 971 struct socket *sock; 972 973 /* special case to disable backend */ 974 if (fd == -1) 975 return NULL; 976 sock = get_raw_socket(fd); 977 if (!IS_ERR(sock)) 978 return sock; 979 sock = get_tap_socket(fd); 980 if (!IS_ERR(sock)) 981 return sock; 982 return ERR_PTR(-ENOTSOCK); 983 } 984 985 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd) 986 { 987 struct socket *sock, *oldsock; 988 struct vhost_virtqueue *vq; 989 struct vhost_net_virtqueue *nvq; 990 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL; 991 int r; 992 993 mutex_lock(&n->dev.mutex); 994 r = vhost_dev_check_owner(&n->dev); 995 if (r) 996 goto err; 997 998 if (index >= VHOST_NET_VQ_MAX) { 999 r = -ENOBUFS; 1000 goto err; 1001 } 1002 vq = &n->vqs[index].vq; 1003 nvq = &n->vqs[index]; 1004 mutex_lock(&vq->mutex); 1005 1006 /* Verify that ring has been setup correctly. */ 1007 if (!vhost_vq_access_ok(vq)) { 1008 r = -EFAULT; 1009 goto err_vq; 1010 } 1011 sock = get_socket(fd); 1012 if (IS_ERR(sock)) { 1013 r = PTR_ERR(sock); 1014 goto err_vq; 1015 } 1016 1017 /* start polling new socket */ 1018 oldsock = vq->private_data; 1019 if (sock != oldsock) { 1020 ubufs = vhost_net_ubuf_alloc(vq, 1021 sock && vhost_sock_zcopy(sock)); 1022 if (IS_ERR(ubufs)) { 1023 r = PTR_ERR(ubufs); 1024 goto err_ubufs; 1025 } 1026 1027 vhost_net_disable_vq(n, vq); 1028 vq->private_data = sock; 1029 r = vhost_vq_init_access(vq); 1030 if (r) 1031 goto err_used; 1032 r = vhost_net_enable_vq(n, vq); 1033 if (r) 1034 goto err_used; 1035 1036 oldubufs = nvq->ubufs; 1037 nvq->ubufs = ubufs; 1038 1039 n->tx_packets = 0; 1040 n->tx_zcopy_err = 0; 1041 n->tx_flush = false; 1042 } 1043 1044 mutex_unlock(&vq->mutex); 1045 1046 if (oldubufs) { 1047 vhost_net_ubuf_put_wait_and_free(oldubufs); 1048 mutex_lock(&vq->mutex); 1049 vhost_zerocopy_signal_used(n, vq); 1050 mutex_unlock(&vq->mutex); 1051 } 1052 1053 if (oldsock) { 1054 vhost_net_flush_vq(n, index); 1055 sockfd_put(oldsock); 1056 } 1057 1058 mutex_unlock(&n->dev.mutex); 1059 return 0; 1060 1061 err_used: 1062 vq->private_data = oldsock; 1063 vhost_net_enable_vq(n, vq); 1064 if (ubufs) 1065 vhost_net_ubuf_put_wait_and_free(ubufs); 1066 err_ubufs: 1067 sockfd_put(sock); 1068 err_vq: 1069 mutex_unlock(&vq->mutex); 1070 err: 1071 mutex_unlock(&n->dev.mutex); 1072 return r; 1073 } 1074 1075 static long vhost_net_reset_owner(struct vhost_net *n) 1076 { 1077 struct socket *tx_sock = NULL; 1078 struct socket *rx_sock = NULL; 1079 long err; 1080 struct vhost_umem *umem; 1081 1082 mutex_lock(&n->dev.mutex); 1083 err = vhost_dev_check_owner(&n->dev); 1084 if (err) 1085 goto done; 1086 umem = vhost_dev_reset_owner_prepare(); 1087 if (!umem) { 1088 err = -ENOMEM; 1089 goto done; 1090 } 1091 vhost_net_stop(n, &tx_sock, &rx_sock); 1092 vhost_net_flush(n); 1093 vhost_dev_reset_owner(&n->dev, umem); 1094 vhost_net_vq_reset(n); 1095 done: 1096 mutex_unlock(&n->dev.mutex); 1097 if (tx_sock) 1098 sockfd_put(tx_sock); 1099 if (rx_sock) 1100 sockfd_put(rx_sock); 1101 return err; 1102 } 1103 1104 static int vhost_net_set_features(struct vhost_net *n, u64 features) 1105 { 1106 size_t vhost_hlen, sock_hlen, hdr_len; 1107 int i; 1108 1109 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) | 1110 (1ULL << VIRTIO_F_VERSION_1))) ? 1111 sizeof(struct virtio_net_hdr_mrg_rxbuf) : 1112 sizeof(struct virtio_net_hdr); 1113 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) { 1114 /* vhost provides vnet_hdr */ 1115 vhost_hlen = hdr_len; 1116 sock_hlen = 0; 1117 } else { 1118 /* socket provides vnet_hdr */ 1119 vhost_hlen = 0; 1120 sock_hlen = hdr_len; 1121 } 1122 mutex_lock(&n->dev.mutex); 1123 if ((features & (1 << VHOST_F_LOG_ALL)) && 1124 !vhost_log_access_ok(&n->dev)) 1125 goto out_unlock; 1126 1127 if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) { 1128 if (vhost_init_device_iotlb(&n->dev, true)) 1129 goto out_unlock; 1130 } 1131 1132 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 1133 mutex_lock(&n->vqs[i].vq.mutex); 1134 n->vqs[i].vq.acked_features = features; 1135 n->vqs[i].vhost_hlen = vhost_hlen; 1136 n->vqs[i].sock_hlen = sock_hlen; 1137 mutex_unlock(&n->vqs[i].vq.mutex); 1138 } 1139 mutex_unlock(&n->dev.mutex); 1140 return 0; 1141 1142 out_unlock: 1143 mutex_unlock(&n->dev.mutex); 1144 return -EFAULT; 1145 } 1146 1147 static long vhost_net_set_owner(struct vhost_net *n) 1148 { 1149 int r; 1150 1151 mutex_lock(&n->dev.mutex); 1152 if (vhost_dev_has_owner(&n->dev)) { 1153 r = -EBUSY; 1154 goto out; 1155 } 1156 r = vhost_net_set_ubuf_info(n); 1157 if (r) 1158 goto out; 1159 r = vhost_dev_set_owner(&n->dev); 1160 if (r) 1161 vhost_net_clear_ubuf_info(n); 1162 vhost_net_flush(n); 1163 out: 1164 mutex_unlock(&n->dev.mutex); 1165 return r; 1166 } 1167 1168 static long vhost_net_ioctl(struct file *f, unsigned int ioctl, 1169 unsigned long arg) 1170 { 1171 struct vhost_net *n = f->private_data; 1172 void __user *argp = (void __user *)arg; 1173 u64 __user *featurep = argp; 1174 struct vhost_vring_file backend; 1175 u64 features; 1176 int r; 1177 1178 switch (ioctl) { 1179 case VHOST_NET_SET_BACKEND: 1180 if (copy_from_user(&backend, argp, sizeof backend)) 1181 return -EFAULT; 1182 return vhost_net_set_backend(n, backend.index, backend.fd); 1183 case VHOST_GET_FEATURES: 1184 features = VHOST_NET_FEATURES; 1185 if (copy_to_user(featurep, &features, sizeof features)) 1186 return -EFAULT; 1187 return 0; 1188 case VHOST_SET_FEATURES: 1189 if (copy_from_user(&features, featurep, sizeof features)) 1190 return -EFAULT; 1191 if (features & ~VHOST_NET_FEATURES) 1192 return -EOPNOTSUPP; 1193 return vhost_net_set_features(n, features); 1194 case VHOST_RESET_OWNER: 1195 return vhost_net_reset_owner(n); 1196 case VHOST_SET_OWNER: 1197 return vhost_net_set_owner(n); 1198 default: 1199 mutex_lock(&n->dev.mutex); 1200 r = vhost_dev_ioctl(&n->dev, ioctl, argp); 1201 if (r == -ENOIOCTLCMD) 1202 r = vhost_vring_ioctl(&n->dev, ioctl, argp); 1203 else 1204 vhost_net_flush(n); 1205 mutex_unlock(&n->dev.mutex); 1206 return r; 1207 } 1208 } 1209 1210 #ifdef CONFIG_COMPAT 1211 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl, 1212 unsigned long arg) 1213 { 1214 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg)); 1215 } 1216 #endif 1217 1218 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to) 1219 { 1220 struct file *file = iocb->ki_filp; 1221 struct vhost_net *n = file->private_data; 1222 struct vhost_dev *dev = &n->dev; 1223 int noblock = file->f_flags & O_NONBLOCK; 1224 1225 return vhost_chr_read_iter(dev, to, noblock); 1226 } 1227 1228 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb, 1229 struct iov_iter *from) 1230 { 1231 struct file *file = iocb->ki_filp; 1232 struct vhost_net *n = file->private_data; 1233 struct vhost_dev *dev = &n->dev; 1234 1235 return vhost_chr_write_iter(dev, from); 1236 } 1237 1238 static unsigned int vhost_net_chr_poll(struct file *file, poll_table *wait) 1239 { 1240 struct vhost_net *n = file->private_data; 1241 struct vhost_dev *dev = &n->dev; 1242 1243 return vhost_chr_poll(file, dev, wait); 1244 } 1245 1246 static const struct file_operations vhost_net_fops = { 1247 .owner = THIS_MODULE, 1248 .release = vhost_net_release, 1249 .read_iter = vhost_net_chr_read_iter, 1250 .write_iter = vhost_net_chr_write_iter, 1251 .poll = vhost_net_chr_poll, 1252 .unlocked_ioctl = vhost_net_ioctl, 1253 #ifdef CONFIG_COMPAT 1254 .compat_ioctl = vhost_net_compat_ioctl, 1255 #endif 1256 .open = vhost_net_open, 1257 .llseek = noop_llseek, 1258 }; 1259 1260 static struct miscdevice vhost_net_misc = { 1261 .minor = VHOST_NET_MINOR, 1262 .name = "vhost-net", 1263 .fops = &vhost_net_fops, 1264 }; 1265 1266 static int vhost_net_init(void) 1267 { 1268 if (experimental_zcopytx) 1269 vhost_net_enable_zcopy(VHOST_NET_VQ_TX); 1270 return misc_register(&vhost_net_misc); 1271 } 1272 module_init(vhost_net_init); 1273 1274 static void vhost_net_exit(void) 1275 { 1276 misc_deregister(&vhost_net_misc); 1277 } 1278 module_exit(vhost_net_exit); 1279 1280 MODULE_VERSION("0.0.1"); 1281 MODULE_LICENSE("GPL v2"); 1282 MODULE_AUTHOR("Michael S. Tsirkin"); 1283 MODULE_DESCRIPTION("Host kernel accelerator for virtio net"); 1284 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR); 1285 MODULE_ALIAS("devname:vhost-net"); 1286