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_tap.h> 28 #include <linux/if_vlan.h> 29 30 #include <net/sock.h> 31 32 #include "vhost.h" 33 34 static int experimental_zcopytx = 1; 35 module_param(experimental_zcopytx, int, 0444); 36 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;" 37 " 1 -Enable; 0 - Disable"); 38 39 /* Max number of bytes transferred before requeueing the job. 40 * Using this limit prevents one virtqueue from starving others. */ 41 #define VHOST_NET_WEIGHT 0x80000 42 43 /* MAX number of TX used buffers for outstanding zerocopy */ 44 #define VHOST_MAX_PEND 128 45 #define VHOST_GOODCOPY_LEN 256 46 47 /* 48 * For transmit, used buffer len is unused; we override it to track buffer 49 * status internally; used for zerocopy tx only. 50 */ 51 /* Lower device DMA failed */ 52 #define VHOST_DMA_FAILED_LEN ((__force __virtio32)3) 53 /* Lower device DMA done */ 54 #define VHOST_DMA_DONE_LEN ((__force __virtio32)2) 55 /* Lower device DMA in progress */ 56 #define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1) 57 /* Buffer unused */ 58 #define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0) 59 60 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN) 61 62 enum { 63 VHOST_NET_FEATURES = VHOST_FEATURES | 64 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) | 65 (1ULL << VIRTIO_NET_F_MRG_RXBUF) | 66 (1ULL << VIRTIO_F_IOMMU_PLATFORM) 67 }; 68 69 enum { 70 VHOST_NET_VQ_RX = 0, 71 VHOST_NET_VQ_TX = 1, 72 VHOST_NET_VQ_MAX = 2, 73 }; 74 75 struct vhost_net_ubuf_ref { 76 /* refcount follows semantics similar to kref: 77 * 0: object is released 78 * 1: no outstanding ubufs 79 * >1: outstanding ubufs 80 */ 81 atomic_t refcount; 82 wait_queue_head_t wait; 83 struct vhost_virtqueue *vq; 84 }; 85 86 struct vhost_net_virtqueue { 87 struct vhost_virtqueue vq; 88 size_t vhost_hlen; 89 size_t sock_hlen; 90 /* vhost zerocopy support fields below: */ 91 /* last used idx for outstanding DMA zerocopy buffers */ 92 int upend_idx; 93 /* first used idx for DMA done zerocopy buffers */ 94 int done_idx; 95 /* an array of userspace buffers info */ 96 struct ubuf_info *ubuf_info; 97 /* Reference counting for outstanding ubufs. 98 * Protected by vq mutex. Writers must also take device mutex. */ 99 struct vhost_net_ubuf_ref *ubufs; 100 }; 101 102 struct vhost_net { 103 struct vhost_dev dev; 104 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX]; 105 struct vhost_poll poll[VHOST_NET_VQ_MAX]; 106 /* Number of TX recently submitted. 107 * Protected by tx vq lock. */ 108 unsigned tx_packets; 109 /* Number of times zerocopy TX recently failed. 110 * Protected by tx vq lock. */ 111 unsigned tx_zcopy_err; 112 /* Flush in progress. Protected by tx vq lock. */ 113 bool tx_flush; 114 }; 115 116 static unsigned vhost_net_zcopy_mask __read_mostly; 117 118 static void vhost_net_enable_zcopy(int vq) 119 { 120 vhost_net_zcopy_mask |= 0x1 << vq; 121 } 122 123 static struct vhost_net_ubuf_ref * 124 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy) 125 { 126 struct vhost_net_ubuf_ref *ubufs; 127 /* No zero copy backend? Nothing to count. */ 128 if (!zcopy) 129 return NULL; 130 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL); 131 if (!ubufs) 132 return ERR_PTR(-ENOMEM); 133 atomic_set(&ubufs->refcount, 1); 134 init_waitqueue_head(&ubufs->wait); 135 ubufs->vq = vq; 136 return ubufs; 137 } 138 139 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs) 140 { 141 int r = atomic_sub_return(1, &ubufs->refcount); 142 if (unlikely(!r)) 143 wake_up(&ubufs->wait); 144 return r; 145 } 146 147 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs) 148 { 149 vhost_net_ubuf_put(ubufs); 150 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount)); 151 } 152 153 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs) 154 { 155 vhost_net_ubuf_put_and_wait(ubufs); 156 kfree(ubufs); 157 } 158 159 static void vhost_net_clear_ubuf_info(struct vhost_net *n) 160 { 161 int i; 162 163 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 164 kfree(n->vqs[i].ubuf_info); 165 n->vqs[i].ubuf_info = NULL; 166 } 167 } 168 169 static int vhost_net_set_ubuf_info(struct vhost_net *n) 170 { 171 bool zcopy; 172 int i; 173 174 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 175 zcopy = vhost_net_zcopy_mask & (0x1 << i); 176 if (!zcopy) 177 continue; 178 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) * 179 UIO_MAXIOV, GFP_KERNEL); 180 if (!n->vqs[i].ubuf_info) 181 goto err; 182 } 183 return 0; 184 185 err: 186 vhost_net_clear_ubuf_info(n); 187 return -ENOMEM; 188 } 189 190 static void vhost_net_vq_reset(struct vhost_net *n) 191 { 192 int i; 193 194 vhost_net_clear_ubuf_info(n); 195 196 for (i = 0; i < VHOST_NET_VQ_MAX; i++) { 197 n->vqs[i].done_idx = 0; 198 n->vqs[i].upend_idx = 0; 199 n->vqs[i].ubufs = NULL; 200 n->vqs[i].vhost_hlen = 0; 201 n->vqs[i].sock_hlen = 0; 202 } 203 204 } 205 206 static void vhost_net_tx_packet(struct vhost_net *net) 207 { 208 ++net->tx_packets; 209 if (net->tx_packets < 1024) 210 return; 211 net->tx_packets = 0; 212 net->tx_zcopy_err = 0; 213 } 214 215 static void vhost_net_tx_err(struct vhost_net *net) 216 { 217 ++net->tx_zcopy_err; 218 } 219 220 static bool vhost_net_tx_select_zcopy(struct vhost_net *net) 221 { 222 /* TX flush waits for outstanding DMAs to be done. 223 * Don't start new DMAs. 224 */ 225 return !net->tx_flush && 226 net->tx_packets / 64 >= net->tx_zcopy_err; 227 } 228 229 static bool vhost_sock_zcopy(struct socket *sock) 230 { 231 return unlikely(experimental_zcopytx) && 232 sock_flag(sock->sk, SOCK_ZEROCOPY); 233 } 234 235 /* In case of DMA done not in order in lower device driver for some reason. 236 * upend_idx is used to track end of used idx, done_idx is used to track head 237 * of used idx. Once lower device DMA done contiguously, we will signal KVM 238 * guest used idx. 239 */ 240 static void vhost_zerocopy_signal_used(struct vhost_net *net, 241 struct vhost_virtqueue *vq) 242 { 243 struct vhost_net_virtqueue *nvq = 244 container_of(vq, struct vhost_net_virtqueue, vq); 245 int i, add; 246 int j = 0; 247 248 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) { 249 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN) 250 vhost_net_tx_err(net); 251 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) { 252 vq->heads[i].len = VHOST_DMA_CLEAR_LEN; 253 ++j; 254 } else 255 break; 256 } 257 while (j) { 258 add = min(UIO_MAXIOV - nvq->done_idx, j); 259 vhost_add_used_and_signal_n(vq->dev, vq, 260 &vq->heads[nvq->done_idx], add); 261 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV; 262 j -= add; 263 } 264 } 265 266 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success) 267 { 268 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx; 269 struct vhost_virtqueue *vq = ubufs->vq; 270 int cnt; 271 272 rcu_read_lock_bh(); 273 274 /* set len to mark this desc buffers done DMA */ 275 vq->heads[ubuf->desc].len = success ? 276 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN; 277 cnt = vhost_net_ubuf_put(ubufs); 278 279 /* 280 * Trigger polling thread if guest stopped submitting new buffers: 281 * in this case, the refcount after decrement will eventually reach 1. 282 * We also trigger polling periodically after each 16 packets 283 * (the value 16 here is more or less arbitrary, it's tuned to trigger 284 * less than 10% of times). 285 */ 286 if (cnt <= 1 || !(cnt % 16)) 287 vhost_poll_queue(&vq->poll); 288 289 rcu_read_unlock_bh(); 290 } 291 292 static inline unsigned long busy_clock(void) 293 { 294 return local_clock() >> 10; 295 } 296 297 static bool vhost_can_busy_poll(struct vhost_dev *dev, 298 unsigned long endtime) 299 { 300 return likely(!need_resched()) && 301 likely(!time_after(busy_clock(), endtime)) && 302 likely(!signal_pending(current)) && 303 !vhost_has_work(dev); 304 } 305 306 static void vhost_net_disable_vq(struct vhost_net *n, 307 struct vhost_virtqueue *vq) 308 { 309 struct vhost_net_virtqueue *nvq = 310 container_of(vq, struct vhost_net_virtqueue, vq); 311 struct vhost_poll *poll = n->poll + (nvq - n->vqs); 312 if (!vq->private_data) 313 return; 314 vhost_poll_stop(poll); 315 } 316 317 static int vhost_net_enable_vq(struct vhost_net *n, 318 struct vhost_virtqueue *vq) 319 { 320 struct vhost_net_virtqueue *nvq = 321 container_of(vq, struct vhost_net_virtqueue, vq); 322 struct vhost_poll *poll = n->poll + (nvq - n->vqs); 323 struct socket *sock; 324 325 sock = vq->private_data; 326 if (!sock) 327 return 0; 328 329 return vhost_poll_start(poll, sock->file); 330 } 331 332 static int vhost_net_tx_get_vq_desc(struct vhost_net *net, 333 struct vhost_virtqueue *vq, 334 struct iovec iov[], unsigned int iov_size, 335 unsigned int *out_num, unsigned int *in_num) 336 { 337 unsigned long uninitialized_var(endtime); 338 int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov), 339 out_num, in_num, NULL, NULL); 340 341 if (r == vq->num && vq->busyloop_timeout) { 342 preempt_disable(); 343 endtime = busy_clock() + vq->busyloop_timeout; 344 while (vhost_can_busy_poll(vq->dev, endtime) && 345 vhost_vq_avail_empty(vq->dev, vq)) 346 cpu_relax(); 347 preempt_enable(); 348 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov), 349 out_num, in_num, NULL, NULL); 350 } 351 352 return r; 353 } 354 355 static bool vhost_exceeds_maxpend(struct vhost_net *net) 356 { 357 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 358 struct vhost_virtqueue *vq = &nvq->vq; 359 360 return (nvq->upend_idx + vq->num - VHOST_MAX_PEND) % UIO_MAXIOV 361 == nvq->done_idx; 362 } 363 364 /* Expects to be always run from workqueue - which acts as 365 * read-size critical section for our kind of RCU. */ 366 static void handle_tx(struct vhost_net *net) 367 { 368 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 369 struct vhost_virtqueue *vq = &nvq->vq; 370 unsigned out, in; 371 int head; 372 struct msghdr msg = { 373 .msg_name = NULL, 374 .msg_namelen = 0, 375 .msg_control = NULL, 376 .msg_controllen = 0, 377 .msg_flags = MSG_DONTWAIT, 378 }; 379 size_t len, total_len = 0; 380 int err; 381 size_t hdr_size; 382 struct socket *sock; 383 struct vhost_net_ubuf_ref *uninitialized_var(ubufs); 384 bool zcopy, zcopy_used; 385 386 mutex_lock(&vq->mutex); 387 sock = vq->private_data; 388 if (!sock) 389 goto out; 390 391 if (!vq_iotlb_prefetch(vq)) 392 goto out; 393 394 vhost_disable_notify(&net->dev, vq); 395 396 hdr_size = nvq->vhost_hlen; 397 zcopy = nvq->ubufs; 398 399 for (;;) { 400 /* Release DMAs done buffers first */ 401 if (zcopy) 402 vhost_zerocopy_signal_used(net, vq); 403 404 /* If more outstanding DMAs, queue the work. 405 * Handle upend_idx wrap around 406 */ 407 if (unlikely(vhost_exceeds_maxpend(net))) 408 break; 409 410 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov, 411 ARRAY_SIZE(vq->iov), 412 &out, &in); 413 /* On error, stop handling until the next kick. */ 414 if (unlikely(head < 0)) 415 break; 416 /* Nothing new? Wait for eventfd to tell us they refilled. */ 417 if (head == vq->num) { 418 if (unlikely(vhost_enable_notify(&net->dev, vq))) { 419 vhost_disable_notify(&net->dev, vq); 420 continue; 421 } 422 break; 423 } 424 if (in) { 425 vq_err(vq, "Unexpected descriptor format for TX: " 426 "out %d, int %d\n", out, in); 427 break; 428 } 429 /* Skip header. TODO: support TSO. */ 430 len = iov_length(vq->iov, out); 431 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len); 432 iov_iter_advance(&msg.msg_iter, hdr_size); 433 /* Sanity check */ 434 if (!msg_data_left(&msg)) { 435 vq_err(vq, "Unexpected header len for TX: " 436 "%zd expected %zd\n", 437 len, hdr_size); 438 break; 439 } 440 len = msg_data_left(&msg); 441 442 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN 443 && (nvq->upend_idx + 1) % UIO_MAXIOV != 444 nvq->done_idx 445 && vhost_net_tx_select_zcopy(net); 446 447 /* use msg_control to pass vhost zerocopy ubuf info to skb */ 448 if (zcopy_used) { 449 struct ubuf_info *ubuf; 450 ubuf = nvq->ubuf_info + nvq->upend_idx; 451 452 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head); 453 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS; 454 ubuf->callback = vhost_zerocopy_callback; 455 ubuf->ctx = nvq->ubufs; 456 ubuf->desc = nvq->upend_idx; 457 msg.msg_control = ubuf; 458 msg.msg_controllen = sizeof(ubuf); 459 ubufs = nvq->ubufs; 460 atomic_inc(&ubufs->refcount); 461 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV; 462 } else { 463 msg.msg_control = NULL; 464 ubufs = NULL; 465 } 466 467 total_len += len; 468 if (total_len < VHOST_NET_WEIGHT && 469 !vhost_vq_avail_empty(&net->dev, vq) && 470 likely(!vhost_exceeds_maxpend(net))) { 471 msg.msg_flags |= MSG_MORE; 472 } else { 473 msg.msg_flags &= ~MSG_MORE; 474 } 475 476 /* TODO: Check specific error and bomb out unless ENOBUFS? */ 477 err = sock->ops->sendmsg(sock, &msg, len); 478 if (unlikely(err < 0)) { 479 if (zcopy_used) { 480 vhost_net_ubuf_put(ubufs); 481 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1) 482 % UIO_MAXIOV; 483 } 484 vhost_discard_vq_desc(vq, 1); 485 break; 486 } 487 if (err != len) 488 pr_debug("Truncated TX packet: " 489 " len %d != %zd\n", err, len); 490 if (!zcopy_used) 491 vhost_add_used_and_signal(&net->dev, vq, head, 0); 492 else 493 vhost_zerocopy_signal_used(net, vq); 494 vhost_net_tx_packet(net); 495 if (unlikely(total_len >= VHOST_NET_WEIGHT)) { 496 vhost_poll_queue(&vq->poll); 497 break; 498 } 499 } 500 out: 501 mutex_unlock(&vq->mutex); 502 } 503 504 static int peek_head_len(struct sock *sk) 505 { 506 struct socket *sock = sk->sk_socket; 507 struct sk_buff *head; 508 int len = 0; 509 unsigned long flags; 510 511 if (sock->ops->peek_len) 512 return sock->ops->peek_len(sock); 513 514 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags); 515 head = skb_peek(&sk->sk_receive_queue); 516 if (likely(head)) { 517 len = head->len; 518 if (skb_vlan_tag_present(head)) 519 len += VLAN_HLEN; 520 } 521 522 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags); 523 return len; 524 } 525 526 static int sk_has_rx_data(struct sock *sk) 527 { 528 struct socket *sock = sk->sk_socket; 529 530 if (sock->ops->peek_len) 531 return sock->ops->peek_len(sock); 532 533 return skb_queue_empty(&sk->sk_receive_queue); 534 } 535 536 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk) 537 { 538 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 539 struct vhost_virtqueue *vq = &nvq->vq; 540 unsigned long uninitialized_var(endtime); 541 int len = peek_head_len(sk); 542 543 if (!len && vq->busyloop_timeout) { 544 /* Both tx vq and rx socket were polled here */ 545 mutex_lock(&vq->mutex); 546 vhost_disable_notify(&net->dev, vq); 547 548 preempt_disable(); 549 endtime = busy_clock() + vq->busyloop_timeout; 550 551 while (vhost_can_busy_poll(&net->dev, endtime) && 552 !sk_has_rx_data(sk) && 553 vhost_vq_avail_empty(&net->dev, vq)) 554 cpu_relax(); 555 556 preempt_enable(); 557 558 if (vhost_enable_notify(&net->dev, vq)) 559 vhost_poll_queue(&vq->poll); 560 mutex_unlock(&vq->mutex); 561 562 len = peek_head_len(sk); 563 } 564 565 return len; 566 } 567 568 /* This is a multi-buffer version of vhost_get_desc, that works if 569 * vq has read descriptors only. 570 * @vq - the relevant virtqueue 571 * @datalen - data length we'll be reading 572 * @iovcount - returned count of io vectors we fill 573 * @log - vhost log 574 * @log_num - log offset 575 * @quota - headcount quota, 1 for big buffer 576 * returns number of buffer heads allocated, negative on error 577 */ 578 static int get_rx_bufs(struct vhost_virtqueue *vq, 579 struct vring_used_elem *heads, 580 int datalen, 581 unsigned *iovcount, 582 struct vhost_log *log, 583 unsigned *log_num, 584 unsigned int quota) 585 { 586 unsigned int out, in; 587 int seg = 0; 588 int headcount = 0; 589 unsigned d; 590 int r, nlogs = 0; 591 /* len is always initialized before use since we are always called with 592 * datalen > 0. 593 */ 594 u32 uninitialized_var(len); 595 596 while (datalen > 0 && headcount < quota) { 597 if (unlikely(seg >= UIO_MAXIOV)) { 598 r = -ENOBUFS; 599 goto err; 600 } 601 r = vhost_get_vq_desc(vq, vq->iov + seg, 602 ARRAY_SIZE(vq->iov) - seg, &out, 603 &in, log, log_num); 604 if (unlikely(r < 0)) 605 goto err; 606 607 d = r; 608 if (d == vq->num) { 609 r = 0; 610 goto err; 611 } 612 if (unlikely(out || in <= 0)) { 613 vq_err(vq, "unexpected descriptor format for RX: " 614 "out %d, in %d\n", out, in); 615 r = -EINVAL; 616 goto err; 617 } 618 if (unlikely(log)) { 619 nlogs += *log_num; 620 log += *log_num; 621 } 622 heads[headcount].id = cpu_to_vhost32(vq, d); 623 len = iov_length(vq->iov + seg, in); 624 heads[headcount].len = cpu_to_vhost32(vq, len); 625 datalen -= len; 626 ++headcount; 627 seg += in; 628 } 629 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen); 630 *iovcount = seg; 631 if (unlikely(log)) 632 *log_num = nlogs; 633 634 /* Detect overrun */ 635 if (unlikely(datalen > 0)) { 636 r = UIO_MAXIOV + 1; 637 goto err; 638 } 639 return headcount; 640 err: 641 vhost_discard_vq_desc(vq, headcount); 642 return r; 643 } 644 645 /* Expects to be always run from workqueue - which acts as 646 * read-size critical section for our kind of RCU. */ 647 static void handle_rx(struct vhost_net *net) 648 { 649 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX]; 650 struct vhost_virtqueue *vq = &nvq->vq; 651 unsigned uninitialized_var(in), log; 652 struct vhost_log *vq_log; 653 struct msghdr msg = { 654 .msg_name = NULL, 655 .msg_namelen = 0, 656 .msg_control = NULL, /* FIXME: get and handle RX aux data. */ 657 .msg_controllen = 0, 658 .msg_flags = MSG_DONTWAIT, 659 }; 660 struct virtio_net_hdr hdr = { 661 .flags = 0, 662 .gso_type = VIRTIO_NET_HDR_GSO_NONE 663 }; 664 size_t total_len = 0; 665 int err, mergeable; 666 s16 headcount; 667 size_t vhost_hlen, sock_hlen; 668 size_t vhost_len, sock_len; 669 struct socket *sock; 670 struct iov_iter fixup; 671 __virtio16 num_buffers; 672 673 mutex_lock(&vq->mutex); 674 sock = vq->private_data; 675 if (!sock) 676 goto out; 677 678 if (!vq_iotlb_prefetch(vq)) 679 goto out; 680 681 vhost_disable_notify(&net->dev, vq); 682 vhost_net_disable_vq(net, vq); 683 684 vhost_hlen = nvq->vhost_hlen; 685 sock_hlen = nvq->sock_hlen; 686 687 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ? 688 vq->log : NULL; 689 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF); 690 691 while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) { 692 sock_len += sock_hlen; 693 vhost_len = sock_len + vhost_hlen; 694 headcount = get_rx_bufs(vq, vq->heads, vhost_len, 695 &in, vq_log, &log, 696 likely(mergeable) ? UIO_MAXIOV : 1); 697 /* On error, stop handling until the next kick. */ 698 if (unlikely(headcount < 0)) 699 goto out; 700 /* On overrun, truncate and discard */ 701 if (unlikely(headcount > UIO_MAXIOV)) { 702 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1); 703 err = sock->ops->recvmsg(sock, &msg, 704 1, MSG_DONTWAIT | MSG_TRUNC); 705 pr_debug("Discarded rx packet: len %zd\n", sock_len); 706 continue; 707 } 708 /* OK, now we need to know about added descriptors. */ 709 if (!headcount) { 710 if (unlikely(vhost_enable_notify(&net->dev, vq))) { 711 /* They have slipped one in as we were 712 * doing that: check again. */ 713 vhost_disable_notify(&net->dev, vq); 714 continue; 715 } 716 /* Nothing new? Wait for eventfd to tell us 717 * they refilled. */ 718 goto out; 719 } 720 /* We don't need to be notified again. */ 721 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len); 722 fixup = msg.msg_iter; 723 if (unlikely((vhost_hlen))) { 724 /* We will supply the header ourselves 725 * TODO: support TSO. 726 */ 727 iov_iter_advance(&msg.msg_iter, vhost_hlen); 728 } 729 err = sock->ops->recvmsg(sock, &msg, 730 sock_len, MSG_DONTWAIT | MSG_TRUNC); 731 /* Userspace might have consumed the packet meanwhile: 732 * it's not supposed to do this usually, but might be hard 733 * to prevent. Discard data we got (if any) and keep going. */ 734 if (unlikely(err != sock_len)) { 735 pr_debug("Discarded rx packet: " 736 " len %d, expected %zd\n", err, sock_len); 737 vhost_discard_vq_desc(vq, headcount); 738 continue; 739 } 740 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */ 741 if (unlikely(vhost_hlen)) { 742 if (copy_to_iter(&hdr, sizeof(hdr), 743 &fixup) != sizeof(hdr)) { 744 vq_err(vq, "Unable to write vnet_hdr " 745 "at addr %p\n", vq->iov->iov_base); 746 goto out; 747 } 748 } else { 749 /* Header came from socket; we'll need to patch 750 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF 751 */ 752 iov_iter_advance(&fixup, sizeof(hdr)); 753 } 754 /* TODO: Should check and handle checksum. */ 755 756 num_buffers = cpu_to_vhost16(vq, headcount); 757 if (likely(mergeable) && 758 copy_to_iter(&num_buffers, sizeof num_buffers, 759 &fixup) != sizeof num_buffers) { 760 vq_err(vq, "Failed num_buffers write"); 761 vhost_discard_vq_desc(vq, headcount); 762 goto out; 763 } 764 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads, 765 headcount); 766 if (unlikely(vq_log)) 767 vhost_log_write(vq, vq_log, log, vhost_len); 768 total_len += vhost_len; 769 if (unlikely(total_len >= VHOST_NET_WEIGHT)) { 770 vhost_poll_queue(&vq->poll); 771 goto out; 772 } 773 } 774 vhost_net_enable_vq(net, vq); 775 out: 776 mutex_unlock(&vq->mutex); 777 } 778 779 static void handle_tx_kick(struct vhost_work *work) 780 { 781 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, 782 poll.work); 783 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); 784 785 handle_tx(net); 786 } 787 788 static void handle_rx_kick(struct vhost_work *work) 789 { 790 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, 791 poll.work); 792 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); 793 794 handle_rx(net); 795 } 796 797 static void handle_tx_net(struct vhost_work *work) 798 { 799 struct vhost_net *net = container_of(work, struct vhost_net, 800 poll[VHOST_NET_VQ_TX].work); 801 handle_tx(net); 802 } 803 804 static void handle_rx_net(struct vhost_work *work) 805 { 806 struct vhost_net *net = container_of(work, struct vhost_net, 807 poll[VHOST_NET_VQ_RX].work); 808 handle_rx(net); 809 } 810 811 static int vhost_net_open(struct inode *inode, struct file *f) 812 { 813 struct vhost_net *n; 814 struct vhost_dev *dev; 815 struct vhost_virtqueue **vqs; 816 int i; 817 818 n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT); 819 if (!n) { 820 n = vmalloc(sizeof *n); 821 if (!n) 822 return -ENOMEM; 823 } 824 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL); 825 if (!vqs) { 826 kvfree(n); 827 return -ENOMEM; 828 } 829 830 dev = &n->dev; 831 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq; 832 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq; 833 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick; 834 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick; 835 for (i = 0; i < VHOST_NET_VQ_MAX; i++) { 836 n->vqs[i].ubufs = NULL; 837 n->vqs[i].ubuf_info = NULL; 838 n->vqs[i].upend_idx = 0; 839 n->vqs[i].done_idx = 0; 840 n->vqs[i].vhost_hlen = 0; 841 n->vqs[i].sock_hlen = 0; 842 } 843 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX); 844 845 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev); 846 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev); 847 848 f->private_data = n; 849 850 return 0; 851 } 852 853 static struct socket *vhost_net_stop_vq(struct vhost_net *n, 854 struct vhost_virtqueue *vq) 855 { 856 struct socket *sock; 857 858 mutex_lock(&vq->mutex); 859 sock = vq->private_data; 860 vhost_net_disable_vq(n, vq); 861 vq->private_data = NULL; 862 mutex_unlock(&vq->mutex); 863 return sock; 864 } 865 866 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock, 867 struct socket **rx_sock) 868 { 869 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq); 870 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq); 871 } 872 873 static void vhost_net_flush_vq(struct vhost_net *n, int index) 874 { 875 vhost_poll_flush(n->poll + index); 876 vhost_poll_flush(&n->vqs[index].vq.poll); 877 } 878 879 static void vhost_net_flush(struct vhost_net *n) 880 { 881 vhost_net_flush_vq(n, VHOST_NET_VQ_TX); 882 vhost_net_flush_vq(n, VHOST_NET_VQ_RX); 883 if (n->vqs[VHOST_NET_VQ_TX].ubufs) { 884 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 885 n->tx_flush = true; 886 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 887 /* Wait for all lower device DMAs done. */ 888 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs); 889 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 890 n->tx_flush = false; 891 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1); 892 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 893 } 894 } 895 896 static int vhost_net_release(struct inode *inode, struct file *f) 897 { 898 struct vhost_net *n = f->private_data; 899 struct socket *tx_sock; 900 struct socket *rx_sock; 901 902 vhost_net_stop(n, &tx_sock, &rx_sock); 903 vhost_net_flush(n); 904 vhost_dev_stop(&n->dev); 905 vhost_dev_cleanup(&n->dev, false); 906 vhost_net_vq_reset(n); 907 if (tx_sock) 908 sockfd_put(tx_sock); 909 if (rx_sock) 910 sockfd_put(rx_sock); 911 /* Make sure no callbacks are outstanding */ 912 synchronize_rcu_bh(); 913 /* We do an extra flush before freeing memory, 914 * since jobs can re-queue themselves. */ 915 vhost_net_flush(n); 916 kfree(n->dev.vqs); 917 kvfree(n); 918 return 0; 919 } 920 921 static struct socket *get_raw_socket(int fd) 922 { 923 struct { 924 struct sockaddr_ll sa; 925 char buf[MAX_ADDR_LEN]; 926 } uaddr; 927 int uaddr_len = sizeof uaddr, r; 928 struct socket *sock = sockfd_lookup(fd, &r); 929 930 if (!sock) 931 return ERR_PTR(-ENOTSOCK); 932 933 /* Parameter checking */ 934 if (sock->sk->sk_type != SOCK_RAW) { 935 r = -ESOCKTNOSUPPORT; 936 goto err; 937 } 938 939 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa, 940 &uaddr_len, 0); 941 if (r) 942 goto err; 943 944 if (uaddr.sa.sll_family != AF_PACKET) { 945 r = -EPFNOSUPPORT; 946 goto err; 947 } 948 return sock; 949 err: 950 sockfd_put(sock); 951 return ERR_PTR(r); 952 } 953 954 static struct socket *get_tap_socket(int fd) 955 { 956 struct file *file = fget(fd); 957 struct socket *sock; 958 959 if (!file) 960 return ERR_PTR(-EBADF); 961 sock = tun_get_socket(file); 962 if (!IS_ERR(sock)) 963 return sock; 964 sock = tap_get_socket(file); 965 if (IS_ERR(sock)) 966 fput(file); 967 return sock; 968 } 969 970 static struct socket *get_socket(int fd) 971 { 972 struct socket *sock; 973 974 /* special case to disable backend */ 975 if (fd == -1) 976 return NULL; 977 sock = get_raw_socket(fd); 978 if (!IS_ERR(sock)) 979 return sock; 980 sock = get_tap_socket(fd); 981 if (!IS_ERR(sock)) 982 return sock; 983 return ERR_PTR(-ENOTSOCK); 984 } 985 986 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd) 987 { 988 struct socket *sock, *oldsock; 989 struct vhost_virtqueue *vq; 990 struct vhost_net_virtqueue *nvq; 991 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL; 992 int r; 993 994 mutex_lock(&n->dev.mutex); 995 r = vhost_dev_check_owner(&n->dev); 996 if (r) 997 goto err; 998 999 if (index >= VHOST_NET_VQ_MAX) { 1000 r = -ENOBUFS; 1001 goto err; 1002 } 1003 vq = &n->vqs[index].vq; 1004 nvq = &n->vqs[index]; 1005 mutex_lock(&vq->mutex); 1006 1007 /* Verify that ring has been setup correctly. */ 1008 if (!vhost_vq_access_ok(vq)) { 1009 r = -EFAULT; 1010 goto err_vq; 1011 } 1012 sock = get_socket(fd); 1013 if (IS_ERR(sock)) { 1014 r = PTR_ERR(sock); 1015 goto err_vq; 1016 } 1017 1018 /* start polling new socket */ 1019 oldsock = vq->private_data; 1020 if (sock != oldsock) { 1021 ubufs = vhost_net_ubuf_alloc(vq, 1022 sock && vhost_sock_zcopy(sock)); 1023 if (IS_ERR(ubufs)) { 1024 r = PTR_ERR(ubufs); 1025 goto err_ubufs; 1026 } 1027 1028 vhost_net_disable_vq(n, vq); 1029 vq->private_data = sock; 1030 r = vhost_vq_init_access(vq); 1031 if (r) 1032 goto err_used; 1033 r = vhost_net_enable_vq(n, vq); 1034 if (r) 1035 goto err_used; 1036 1037 oldubufs = nvq->ubufs; 1038 nvq->ubufs = ubufs; 1039 1040 n->tx_packets = 0; 1041 n->tx_zcopy_err = 0; 1042 n->tx_flush = false; 1043 } 1044 1045 mutex_unlock(&vq->mutex); 1046 1047 if (oldubufs) { 1048 vhost_net_ubuf_put_wait_and_free(oldubufs); 1049 mutex_lock(&vq->mutex); 1050 vhost_zerocopy_signal_used(n, vq); 1051 mutex_unlock(&vq->mutex); 1052 } 1053 1054 if (oldsock) { 1055 vhost_net_flush_vq(n, index); 1056 sockfd_put(oldsock); 1057 } 1058 1059 mutex_unlock(&n->dev.mutex); 1060 return 0; 1061 1062 err_used: 1063 vq->private_data = oldsock; 1064 vhost_net_enable_vq(n, vq); 1065 if (ubufs) 1066 vhost_net_ubuf_put_wait_and_free(ubufs); 1067 err_ubufs: 1068 sockfd_put(sock); 1069 err_vq: 1070 mutex_unlock(&vq->mutex); 1071 err: 1072 mutex_unlock(&n->dev.mutex); 1073 return r; 1074 } 1075 1076 static long vhost_net_reset_owner(struct vhost_net *n) 1077 { 1078 struct socket *tx_sock = NULL; 1079 struct socket *rx_sock = NULL; 1080 long err; 1081 struct vhost_umem *umem; 1082 1083 mutex_lock(&n->dev.mutex); 1084 err = vhost_dev_check_owner(&n->dev); 1085 if (err) 1086 goto done; 1087 umem = vhost_dev_reset_owner_prepare(); 1088 if (!umem) { 1089 err = -ENOMEM; 1090 goto done; 1091 } 1092 vhost_net_stop(n, &tx_sock, &rx_sock); 1093 vhost_net_flush(n); 1094 vhost_dev_reset_owner(&n->dev, umem); 1095 vhost_net_vq_reset(n); 1096 done: 1097 mutex_unlock(&n->dev.mutex); 1098 if (tx_sock) 1099 sockfd_put(tx_sock); 1100 if (rx_sock) 1101 sockfd_put(rx_sock); 1102 return err; 1103 } 1104 1105 static int vhost_net_set_features(struct vhost_net *n, u64 features) 1106 { 1107 size_t vhost_hlen, sock_hlen, hdr_len; 1108 int i; 1109 1110 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) | 1111 (1ULL << VIRTIO_F_VERSION_1))) ? 1112 sizeof(struct virtio_net_hdr_mrg_rxbuf) : 1113 sizeof(struct virtio_net_hdr); 1114 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) { 1115 /* vhost provides vnet_hdr */ 1116 vhost_hlen = hdr_len; 1117 sock_hlen = 0; 1118 } else { 1119 /* socket provides vnet_hdr */ 1120 vhost_hlen = 0; 1121 sock_hlen = hdr_len; 1122 } 1123 mutex_lock(&n->dev.mutex); 1124 if ((features & (1 << VHOST_F_LOG_ALL)) && 1125 !vhost_log_access_ok(&n->dev)) 1126 goto out_unlock; 1127 1128 if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) { 1129 if (vhost_init_device_iotlb(&n->dev, true)) 1130 goto out_unlock; 1131 } 1132 1133 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 1134 mutex_lock(&n->vqs[i].vq.mutex); 1135 n->vqs[i].vq.acked_features = features; 1136 n->vqs[i].vhost_hlen = vhost_hlen; 1137 n->vqs[i].sock_hlen = sock_hlen; 1138 mutex_unlock(&n->vqs[i].vq.mutex); 1139 } 1140 mutex_unlock(&n->dev.mutex); 1141 return 0; 1142 1143 out_unlock: 1144 mutex_unlock(&n->dev.mutex); 1145 return -EFAULT; 1146 } 1147 1148 static long vhost_net_set_owner(struct vhost_net *n) 1149 { 1150 int r; 1151 1152 mutex_lock(&n->dev.mutex); 1153 if (vhost_dev_has_owner(&n->dev)) { 1154 r = -EBUSY; 1155 goto out; 1156 } 1157 r = vhost_net_set_ubuf_info(n); 1158 if (r) 1159 goto out; 1160 r = vhost_dev_set_owner(&n->dev); 1161 if (r) 1162 vhost_net_clear_ubuf_info(n); 1163 vhost_net_flush(n); 1164 out: 1165 mutex_unlock(&n->dev.mutex); 1166 return r; 1167 } 1168 1169 static long vhost_net_ioctl(struct file *f, unsigned int ioctl, 1170 unsigned long arg) 1171 { 1172 struct vhost_net *n = f->private_data; 1173 void __user *argp = (void __user *)arg; 1174 u64 __user *featurep = argp; 1175 struct vhost_vring_file backend; 1176 u64 features; 1177 int r; 1178 1179 switch (ioctl) { 1180 case VHOST_NET_SET_BACKEND: 1181 if (copy_from_user(&backend, argp, sizeof backend)) 1182 return -EFAULT; 1183 return vhost_net_set_backend(n, backend.index, backend.fd); 1184 case VHOST_GET_FEATURES: 1185 features = VHOST_NET_FEATURES; 1186 if (copy_to_user(featurep, &features, sizeof features)) 1187 return -EFAULT; 1188 return 0; 1189 case VHOST_SET_FEATURES: 1190 if (copy_from_user(&features, featurep, sizeof features)) 1191 return -EFAULT; 1192 if (features & ~VHOST_NET_FEATURES) 1193 return -EOPNOTSUPP; 1194 return vhost_net_set_features(n, features); 1195 case VHOST_RESET_OWNER: 1196 return vhost_net_reset_owner(n); 1197 case VHOST_SET_OWNER: 1198 return vhost_net_set_owner(n); 1199 default: 1200 mutex_lock(&n->dev.mutex); 1201 r = vhost_dev_ioctl(&n->dev, ioctl, argp); 1202 if (r == -ENOIOCTLCMD) 1203 r = vhost_vring_ioctl(&n->dev, ioctl, argp); 1204 else 1205 vhost_net_flush(n); 1206 mutex_unlock(&n->dev.mutex); 1207 return r; 1208 } 1209 } 1210 1211 #ifdef CONFIG_COMPAT 1212 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl, 1213 unsigned long arg) 1214 { 1215 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg)); 1216 } 1217 #endif 1218 1219 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to) 1220 { 1221 struct file *file = iocb->ki_filp; 1222 struct vhost_net *n = file->private_data; 1223 struct vhost_dev *dev = &n->dev; 1224 int noblock = file->f_flags & O_NONBLOCK; 1225 1226 return vhost_chr_read_iter(dev, to, noblock); 1227 } 1228 1229 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb, 1230 struct iov_iter *from) 1231 { 1232 struct file *file = iocb->ki_filp; 1233 struct vhost_net *n = file->private_data; 1234 struct vhost_dev *dev = &n->dev; 1235 1236 return vhost_chr_write_iter(dev, from); 1237 } 1238 1239 static unsigned int vhost_net_chr_poll(struct file *file, poll_table *wait) 1240 { 1241 struct vhost_net *n = file->private_data; 1242 struct vhost_dev *dev = &n->dev; 1243 1244 return vhost_chr_poll(file, dev, wait); 1245 } 1246 1247 static const struct file_operations vhost_net_fops = { 1248 .owner = THIS_MODULE, 1249 .release = vhost_net_release, 1250 .read_iter = vhost_net_chr_read_iter, 1251 .write_iter = vhost_net_chr_write_iter, 1252 .poll = vhost_net_chr_poll, 1253 .unlocked_ioctl = vhost_net_ioctl, 1254 #ifdef CONFIG_COMPAT 1255 .compat_ioctl = vhost_net_compat_ioctl, 1256 #endif 1257 .open = vhost_net_open, 1258 .llseek = noop_llseek, 1259 }; 1260 1261 static struct miscdevice vhost_net_misc = { 1262 .minor = VHOST_NET_MINOR, 1263 .name = "vhost-net", 1264 .fops = &vhost_net_fops, 1265 }; 1266 1267 static int vhost_net_init(void) 1268 { 1269 if (experimental_zcopytx) 1270 vhost_net_enable_zcopy(VHOST_NET_VQ_TX); 1271 return misc_register(&vhost_net_misc); 1272 } 1273 module_init(vhost_net_init); 1274 1275 static void vhost_net_exit(void) 1276 { 1277 misc_deregister(&vhost_net_misc); 1278 } 1279 module_exit(vhost_net_exit); 1280 1281 MODULE_VERSION("0.0.1"); 1282 MODULE_LICENSE("GPL v2"); 1283 MODULE_AUTHOR("Michael S. Tsirkin"); 1284 MODULE_DESCRIPTION("Host kernel accelerator for virtio net"); 1285 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR); 1286 MODULE_ALIAS("devname:vhost-net"); 1287