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/sched/clock.h> 21 #include <linux/sched/signal.h> 22 #include <linux/vmalloc.h> 23 24 #include <linux/net.h> 25 #include <linux/if_packet.h> 26 #include <linux/if_arp.h> 27 #include <linux/if_tun.h> 28 #include <linux/if_macvlan.h> 29 #include <linux/if_tap.h> 30 #include <linux/if_vlan.h> 31 32 #include <net/sock.h> 33 34 #include "vhost.h" 35 36 static int experimental_zcopytx = 1; 37 module_param(experimental_zcopytx, int, 0444); 38 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;" 39 " 1 -Enable; 0 - Disable"); 40 41 /* Max number of bytes transferred before requeueing the job. 42 * Using this limit prevents one virtqueue from starving others. */ 43 #define VHOST_NET_WEIGHT 0x80000 44 45 /* MAX number of TX used buffers for outstanding zerocopy */ 46 #define VHOST_MAX_PEND 128 47 #define VHOST_GOODCOPY_LEN 256 48 49 /* 50 * For transmit, used buffer len is unused; we override it to track buffer 51 * status internally; used for zerocopy tx only. 52 */ 53 /* Lower device DMA failed */ 54 #define VHOST_DMA_FAILED_LEN ((__force __virtio32)3) 55 /* Lower device DMA done */ 56 #define VHOST_DMA_DONE_LEN ((__force __virtio32)2) 57 /* Lower device DMA in progress */ 58 #define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1) 59 /* Buffer unused */ 60 #define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0) 61 62 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN) 63 64 enum { 65 VHOST_NET_FEATURES = VHOST_FEATURES | 66 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) | 67 (1ULL << VIRTIO_NET_F_MRG_RXBUF) | 68 (1ULL << VIRTIO_F_IOMMU_PLATFORM) 69 }; 70 71 enum { 72 VHOST_NET_VQ_RX = 0, 73 VHOST_NET_VQ_TX = 1, 74 VHOST_NET_VQ_MAX = 2, 75 }; 76 77 struct vhost_net_ubuf_ref { 78 /* refcount follows semantics similar to kref: 79 * 0: object is released 80 * 1: no outstanding ubufs 81 * >1: outstanding ubufs 82 */ 83 atomic_t refcount; 84 wait_queue_head_t wait; 85 struct vhost_virtqueue *vq; 86 }; 87 88 struct vhost_net_virtqueue { 89 struct vhost_virtqueue vq; 90 size_t vhost_hlen; 91 size_t sock_hlen; 92 /* vhost zerocopy support fields below: */ 93 /* last used idx for outstanding DMA zerocopy buffers */ 94 int upend_idx; 95 /* first used idx for DMA done zerocopy buffers */ 96 int done_idx; 97 /* an array of userspace buffers info */ 98 struct ubuf_info *ubuf_info; 99 /* Reference counting for outstanding ubufs. 100 * Protected by vq mutex. Writers must also take device mutex. */ 101 struct vhost_net_ubuf_ref *ubufs; 102 }; 103 104 struct vhost_net { 105 struct vhost_dev dev; 106 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX]; 107 struct vhost_poll poll[VHOST_NET_VQ_MAX]; 108 /* Number of TX recently submitted. 109 * Protected by tx vq lock. */ 110 unsigned tx_packets; 111 /* Number of times zerocopy TX recently failed. 112 * Protected by tx vq lock. */ 113 unsigned tx_zcopy_err; 114 /* Flush in progress. Protected by tx vq lock. */ 115 bool tx_flush; 116 }; 117 118 static unsigned vhost_net_zcopy_mask __read_mostly; 119 120 static void vhost_net_enable_zcopy(int vq) 121 { 122 vhost_net_zcopy_mask |= 0x1 << vq; 123 } 124 125 static struct vhost_net_ubuf_ref * 126 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy) 127 { 128 struct vhost_net_ubuf_ref *ubufs; 129 /* No zero copy backend? Nothing to count. */ 130 if (!zcopy) 131 return NULL; 132 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL); 133 if (!ubufs) 134 return ERR_PTR(-ENOMEM); 135 atomic_set(&ubufs->refcount, 1); 136 init_waitqueue_head(&ubufs->wait); 137 ubufs->vq = vq; 138 return ubufs; 139 } 140 141 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs) 142 { 143 int r = atomic_sub_return(1, &ubufs->refcount); 144 if (unlikely(!r)) 145 wake_up(&ubufs->wait); 146 return r; 147 } 148 149 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs) 150 { 151 vhost_net_ubuf_put(ubufs); 152 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount)); 153 } 154 155 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs) 156 { 157 vhost_net_ubuf_put_and_wait(ubufs); 158 kfree(ubufs); 159 } 160 161 static void vhost_net_clear_ubuf_info(struct vhost_net *n) 162 { 163 int i; 164 165 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 166 kfree(n->vqs[i].ubuf_info); 167 n->vqs[i].ubuf_info = NULL; 168 } 169 } 170 171 static int vhost_net_set_ubuf_info(struct vhost_net *n) 172 { 173 bool zcopy; 174 int i; 175 176 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 177 zcopy = vhost_net_zcopy_mask & (0x1 << i); 178 if (!zcopy) 179 continue; 180 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) * 181 UIO_MAXIOV, GFP_KERNEL); 182 if (!n->vqs[i].ubuf_info) 183 goto err; 184 } 185 return 0; 186 187 err: 188 vhost_net_clear_ubuf_info(n); 189 return -ENOMEM; 190 } 191 192 static void vhost_net_vq_reset(struct vhost_net *n) 193 { 194 int i; 195 196 vhost_net_clear_ubuf_info(n); 197 198 for (i = 0; i < VHOST_NET_VQ_MAX; i++) { 199 n->vqs[i].done_idx = 0; 200 n->vqs[i].upend_idx = 0; 201 n->vqs[i].ubufs = NULL; 202 n->vqs[i].vhost_hlen = 0; 203 n->vqs[i].sock_hlen = 0; 204 } 205 206 } 207 208 static void vhost_net_tx_packet(struct vhost_net *net) 209 { 210 ++net->tx_packets; 211 if (net->tx_packets < 1024) 212 return; 213 net->tx_packets = 0; 214 net->tx_zcopy_err = 0; 215 } 216 217 static void vhost_net_tx_err(struct vhost_net *net) 218 { 219 ++net->tx_zcopy_err; 220 } 221 222 static bool vhost_net_tx_select_zcopy(struct vhost_net *net) 223 { 224 /* TX flush waits for outstanding DMAs to be done. 225 * Don't start new DMAs. 226 */ 227 return !net->tx_flush && 228 net->tx_packets / 64 >= net->tx_zcopy_err; 229 } 230 231 static bool vhost_sock_zcopy(struct socket *sock) 232 { 233 return unlikely(experimental_zcopytx) && 234 sock_flag(sock->sk, SOCK_ZEROCOPY); 235 } 236 237 /* In case of DMA done not in order in lower device driver for some reason. 238 * upend_idx is used to track end of used idx, done_idx is used to track head 239 * of used idx. Once lower device DMA done contiguously, we will signal KVM 240 * guest used idx. 241 */ 242 static void vhost_zerocopy_signal_used(struct vhost_net *net, 243 struct vhost_virtqueue *vq) 244 { 245 struct vhost_net_virtqueue *nvq = 246 container_of(vq, struct vhost_net_virtqueue, vq); 247 int i, add; 248 int j = 0; 249 250 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) { 251 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN) 252 vhost_net_tx_err(net); 253 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) { 254 vq->heads[i].len = VHOST_DMA_CLEAR_LEN; 255 ++j; 256 } else 257 break; 258 } 259 while (j) { 260 add = min(UIO_MAXIOV - nvq->done_idx, j); 261 vhost_add_used_and_signal_n(vq->dev, vq, 262 &vq->heads[nvq->done_idx], add); 263 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV; 264 j -= add; 265 } 266 } 267 268 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success) 269 { 270 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx; 271 struct vhost_virtqueue *vq = ubufs->vq; 272 int cnt; 273 274 rcu_read_lock_bh(); 275 276 /* set len to mark this desc buffers done DMA */ 277 vq->heads[ubuf->desc].len = success ? 278 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN; 279 cnt = vhost_net_ubuf_put(ubufs); 280 281 /* 282 * Trigger polling thread if guest stopped submitting new buffers: 283 * in this case, the refcount after decrement will eventually reach 1. 284 * We also trigger polling periodically after each 16 packets 285 * (the value 16 here is more or less arbitrary, it's tuned to trigger 286 * less than 10% of times). 287 */ 288 if (cnt <= 1 || !(cnt % 16)) 289 vhost_poll_queue(&vq->poll); 290 291 rcu_read_unlock_bh(); 292 } 293 294 static inline unsigned long busy_clock(void) 295 { 296 return local_clock() >> 10; 297 } 298 299 static bool vhost_can_busy_poll(struct vhost_dev *dev, 300 unsigned long endtime) 301 { 302 return likely(!need_resched()) && 303 likely(!time_after(busy_clock(), endtime)) && 304 likely(!signal_pending(current)) && 305 !vhost_has_work(dev); 306 } 307 308 static void vhost_net_disable_vq(struct vhost_net *n, 309 struct vhost_virtqueue *vq) 310 { 311 struct vhost_net_virtqueue *nvq = 312 container_of(vq, struct vhost_net_virtqueue, vq); 313 struct vhost_poll *poll = n->poll + (nvq - n->vqs); 314 if (!vq->private_data) 315 return; 316 vhost_poll_stop(poll); 317 } 318 319 static int vhost_net_enable_vq(struct vhost_net *n, 320 struct vhost_virtqueue *vq) 321 { 322 struct vhost_net_virtqueue *nvq = 323 container_of(vq, struct vhost_net_virtqueue, vq); 324 struct vhost_poll *poll = n->poll + (nvq - n->vqs); 325 struct socket *sock; 326 327 sock = vq->private_data; 328 if (!sock) 329 return 0; 330 331 return vhost_poll_start(poll, sock->file); 332 } 333 334 static int vhost_net_tx_get_vq_desc(struct vhost_net *net, 335 struct vhost_virtqueue *vq, 336 struct iovec iov[], unsigned int iov_size, 337 unsigned int *out_num, unsigned int *in_num) 338 { 339 unsigned long uninitialized_var(endtime); 340 int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov), 341 out_num, in_num, NULL, NULL); 342 343 if (r == vq->num && vq->busyloop_timeout) { 344 preempt_disable(); 345 endtime = busy_clock() + vq->busyloop_timeout; 346 while (vhost_can_busy_poll(vq->dev, endtime) && 347 vhost_vq_avail_empty(vq->dev, vq)) 348 cpu_relax(); 349 preempt_enable(); 350 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov), 351 out_num, in_num, NULL, NULL); 352 } 353 354 return r; 355 } 356 357 static bool vhost_exceeds_maxpend(struct vhost_net *net) 358 { 359 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 360 struct vhost_virtqueue *vq = &nvq->vq; 361 362 return (nvq->upend_idx + vq->num - VHOST_MAX_PEND) % UIO_MAXIOV 363 == nvq->done_idx; 364 } 365 366 /* Expects to be always run from workqueue - which acts as 367 * read-size critical section for our kind of RCU. */ 368 static void handle_tx(struct vhost_net *net) 369 { 370 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 371 struct vhost_virtqueue *vq = &nvq->vq; 372 unsigned out, in; 373 int head; 374 struct msghdr msg = { 375 .msg_name = NULL, 376 .msg_namelen = 0, 377 .msg_control = NULL, 378 .msg_controllen = 0, 379 .msg_flags = MSG_DONTWAIT, 380 }; 381 size_t len, total_len = 0; 382 int err; 383 size_t hdr_size; 384 struct socket *sock; 385 struct vhost_net_ubuf_ref *uninitialized_var(ubufs); 386 bool zcopy, zcopy_used; 387 388 mutex_lock(&vq->mutex); 389 sock = vq->private_data; 390 if (!sock) 391 goto out; 392 393 if (!vq_iotlb_prefetch(vq)) 394 goto out; 395 396 vhost_disable_notify(&net->dev, vq); 397 398 hdr_size = nvq->vhost_hlen; 399 zcopy = nvq->ubufs; 400 401 for (;;) { 402 /* Release DMAs done buffers first */ 403 if (zcopy) 404 vhost_zerocopy_signal_used(net, vq); 405 406 /* If more outstanding DMAs, queue the work. 407 * Handle upend_idx wrap around 408 */ 409 if (unlikely(vhost_exceeds_maxpend(net))) 410 break; 411 412 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov, 413 ARRAY_SIZE(vq->iov), 414 &out, &in); 415 /* On error, stop handling until the next kick. */ 416 if (unlikely(head < 0)) 417 break; 418 /* Nothing new? Wait for eventfd to tell us they refilled. */ 419 if (head == vq->num) { 420 if (unlikely(vhost_enable_notify(&net->dev, vq))) { 421 vhost_disable_notify(&net->dev, vq); 422 continue; 423 } 424 break; 425 } 426 if (in) { 427 vq_err(vq, "Unexpected descriptor format for TX: " 428 "out %d, int %d\n", out, in); 429 break; 430 } 431 /* Skip header. TODO: support TSO. */ 432 len = iov_length(vq->iov, out); 433 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len); 434 iov_iter_advance(&msg.msg_iter, hdr_size); 435 /* Sanity check */ 436 if (!msg_data_left(&msg)) { 437 vq_err(vq, "Unexpected header len for TX: " 438 "%zd expected %zd\n", 439 len, hdr_size); 440 break; 441 } 442 len = msg_data_left(&msg); 443 444 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN 445 && (nvq->upend_idx + 1) % UIO_MAXIOV != 446 nvq->done_idx 447 && vhost_net_tx_select_zcopy(net); 448 449 /* use msg_control to pass vhost zerocopy ubuf info to skb */ 450 if (zcopy_used) { 451 struct ubuf_info *ubuf; 452 ubuf = nvq->ubuf_info + nvq->upend_idx; 453 454 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head); 455 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS; 456 ubuf->callback = vhost_zerocopy_callback; 457 ubuf->ctx = nvq->ubufs; 458 ubuf->desc = nvq->upend_idx; 459 msg.msg_control = ubuf; 460 msg.msg_controllen = sizeof(ubuf); 461 ubufs = nvq->ubufs; 462 atomic_inc(&ubufs->refcount); 463 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV; 464 } else { 465 msg.msg_control = NULL; 466 ubufs = NULL; 467 } 468 469 total_len += len; 470 if (total_len < VHOST_NET_WEIGHT && 471 !vhost_vq_avail_empty(&net->dev, vq) && 472 likely(!vhost_exceeds_maxpend(net))) { 473 msg.msg_flags |= MSG_MORE; 474 } else { 475 msg.msg_flags &= ~MSG_MORE; 476 } 477 478 /* TODO: Check specific error and bomb out unless ENOBUFS? */ 479 err = sock->ops->sendmsg(sock, &msg, len); 480 if (unlikely(err < 0)) { 481 if (zcopy_used) { 482 vhost_net_ubuf_put(ubufs); 483 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1) 484 % UIO_MAXIOV; 485 } 486 vhost_discard_vq_desc(vq, 1); 487 break; 488 } 489 if (err != len) 490 pr_debug("Truncated TX packet: " 491 " len %d != %zd\n", err, len); 492 if (!zcopy_used) 493 vhost_add_used_and_signal(&net->dev, vq, head, 0); 494 else 495 vhost_zerocopy_signal_used(net, vq); 496 vhost_net_tx_packet(net); 497 if (unlikely(total_len >= VHOST_NET_WEIGHT)) { 498 vhost_poll_queue(&vq->poll); 499 break; 500 } 501 } 502 out: 503 mutex_unlock(&vq->mutex); 504 } 505 506 static int peek_head_len(struct sock *sk) 507 { 508 struct socket *sock = sk->sk_socket; 509 struct sk_buff *head; 510 int len = 0; 511 unsigned long flags; 512 513 if (sock->ops->peek_len) 514 return sock->ops->peek_len(sock); 515 516 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags); 517 head = skb_peek(&sk->sk_receive_queue); 518 if (likely(head)) { 519 len = head->len; 520 if (skb_vlan_tag_present(head)) 521 len += VLAN_HLEN; 522 } 523 524 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags); 525 return len; 526 } 527 528 static int sk_has_rx_data(struct sock *sk) 529 { 530 struct socket *sock = sk->sk_socket; 531 532 if (sock->ops->peek_len) 533 return sock->ops->peek_len(sock); 534 535 return skb_queue_empty(&sk->sk_receive_queue); 536 } 537 538 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk) 539 { 540 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 541 struct vhost_virtqueue *vq = &nvq->vq; 542 unsigned long uninitialized_var(endtime); 543 int len = peek_head_len(sk); 544 545 if (!len && vq->busyloop_timeout) { 546 /* Both tx vq and rx socket were polled here */ 547 mutex_lock(&vq->mutex); 548 vhost_disable_notify(&net->dev, vq); 549 550 preempt_disable(); 551 endtime = busy_clock() + vq->busyloop_timeout; 552 553 while (vhost_can_busy_poll(&net->dev, endtime) && 554 !sk_has_rx_data(sk) && 555 vhost_vq_avail_empty(&net->dev, vq)) 556 cpu_relax(); 557 558 preempt_enable(); 559 560 if (vhost_enable_notify(&net->dev, vq)) 561 vhost_poll_queue(&vq->poll); 562 mutex_unlock(&vq->mutex); 563 564 len = peek_head_len(sk); 565 } 566 567 return len; 568 } 569 570 /* This is a multi-buffer version of vhost_get_desc, that works if 571 * vq has read descriptors only. 572 * @vq - the relevant virtqueue 573 * @datalen - data length we'll be reading 574 * @iovcount - returned count of io vectors we fill 575 * @log - vhost log 576 * @log_num - log offset 577 * @quota - headcount quota, 1 for big buffer 578 * returns number of buffer heads allocated, negative on error 579 */ 580 static int get_rx_bufs(struct vhost_virtqueue *vq, 581 struct vring_used_elem *heads, 582 int datalen, 583 unsigned *iovcount, 584 struct vhost_log *log, 585 unsigned *log_num, 586 unsigned int quota) 587 { 588 unsigned int out, in; 589 int seg = 0; 590 int headcount = 0; 591 unsigned d; 592 int r, nlogs = 0; 593 /* len is always initialized before use since we are always called with 594 * datalen > 0. 595 */ 596 u32 uninitialized_var(len); 597 598 while (datalen > 0 && headcount < quota) { 599 if (unlikely(seg >= UIO_MAXIOV)) { 600 r = -ENOBUFS; 601 goto err; 602 } 603 r = vhost_get_vq_desc(vq, vq->iov + seg, 604 ARRAY_SIZE(vq->iov) - seg, &out, 605 &in, log, log_num); 606 if (unlikely(r < 0)) 607 goto err; 608 609 d = r; 610 if (d == vq->num) { 611 r = 0; 612 goto err; 613 } 614 if (unlikely(out || in <= 0)) { 615 vq_err(vq, "unexpected descriptor format for RX: " 616 "out %d, in %d\n", out, in); 617 r = -EINVAL; 618 goto err; 619 } 620 if (unlikely(log)) { 621 nlogs += *log_num; 622 log += *log_num; 623 } 624 heads[headcount].id = cpu_to_vhost32(vq, d); 625 len = iov_length(vq->iov + seg, in); 626 heads[headcount].len = cpu_to_vhost32(vq, len); 627 datalen -= len; 628 ++headcount; 629 seg += in; 630 } 631 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen); 632 *iovcount = seg; 633 if (unlikely(log)) 634 *log_num = nlogs; 635 636 /* Detect overrun */ 637 if (unlikely(datalen > 0)) { 638 r = UIO_MAXIOV + 1; 639 goto err; 640 } 641 return headcount; 642 err: 643 vhost_discard_vq_desc(vq, headcount); 644 return r; 645 } 646 647 /* Expects to be always run from workqueue - which acts as 648 * read-size critical section for our kind of RCU. */ 649 static void handle_rx(struct vhost_net *net) 650 { 651 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX]; 652 struct vhost_virtqueue *vq = &nvq->vq; 653 unsigned uninitialized_var(in), log; 654 struct vhost_log *vq_log; 655 struct msghdr msg = { 656 .msg_name = NULL, 657 .msg_namelen = 0, 658 .msg_control = NULL, /* FIXME: get and handle RX aux data. */ 659 .msg_controllen = 0, 660 .msg_flags = MSG_DONTWAIT, 661 }; 662 struct virtio_net_hdr hdr = { 663 .flags = 0, 664 .gso_type = VIRTIO_NET_HDR_GSO_NONE 665 }; 666 size_t total_len = 0; 667 int err, mergeable; 668 s16 headcount; 669 size_t vhost_hlen, sock_hlen; 670 size_t vhost_len, sock_len; 671 struct socket *sock; 672 struct iov_iter fixup; 673 __virtio16 num_buffers; 674 675 mutex_lock(&vq->mutex); 676 sock = vq->private_data; 677 if (!sock) 678 goto out; 679 680 if (!vq_iotlb_prefetch(vq)) 681 goto out; 682 683 vhost_disable_notify(&net->dev, vq); 684 vhost_net_disable_vq(net, vq); 685 686 vhost_hlen = nvq->vhost_hlen; 687 sock_hlen = nvq->sock_hlen; 688 689 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ? 690 vq->log : NULL; 691 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF); 692 693 while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) { 694 sock_len += sock_hlen; 695 vhost_len = sock_len + vhost_hlen; 696 headcount = get_rx_bufs(vq, vq->heads, vhost_len, 697 &in, vq_log, &log, 698 likely(mergeable) ? UIO_MAXIOV : 1); 699 /* On error, stop handling until the next kick. */ 700 if (unlikely(headcount < 0)) 701 goto out; 702 /* On overrun, truncate and discard */ 703 if (unlikely(headcount > UIO_MAXIOV)) { 704 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1); 705 err = sock->ops->recvmsg(sock, &msg, 706 1, MSG_DONTWAIT | MSG_TRUNC); 707 pr_debug("Discarded rx packet: len %zd\n", sock_len); 708 continue; 709 } 710 /* OK, now we need to know about added descriptors. */ 711 if (!headcount) { 712 if (unlikely(vhost_enable_notify(&net->dev, vq))) { 713 /* They have slipped one in as we were 714 * doing that: check again. */ 715 vhost_disable_notify(&net->dev, vq); 716 continue; 717 } 718 /* Nothing new? Wait for eventfd to tell us 719 * they refilled. */ 720 goto out; 721 } 722 /* We don't need to be notified again. */ 723 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len); 724 fixup = msg.msg_iter; 725 if (unlikely((vhost_hlen))) { 726 /* We will supply the header ourselves 727 * TODO: support TSO. 728 */ 729 iov_iter_advance(&msg.msg_iter, vhost_hlen); 730 } 731 err = sock->ops->recvmsg(sock, &msg, 732 sock_len, MSG_DONTWAIT | MSG_TRUNC); 733 /* Userspace might have consumed the packet meanwhile: 734 * it's not supposed to do this usually, but might be hard 735 * to prevent. Discard data we got (if any) and keep going. */ 736 if (unlikely(err != sock_len)) { 737 pr_debug("Discarded rx packet: " 738 " len %d, expected %zd\n", err, sock_len); 739 vhost_discard_vq_desc(vq, headcount); 740 continue; 741 } 742 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */ 743 if (unlikely(vhost_hlen)) { 744 if (copy_to_iter(&hdr, sizeof(hdr), 745 &fixup) != sizeof(hdr)) { 746 vq_err(vq, "Unable to write vnet_hdr " 747 "at addr %p\n", vq->iov->iov_base); 748 goto out; 749 } 750 } else { 751 /* Header came from socket; we'll need to patch 752 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF 753 */ 754 iov_iter_advance(&fixup, sizeof(hdr)); 755 } 756 /* TODO: Should check and handle checksum. */ 757 758 num_buffers = cpu_to_vhost16(vq, headcount); 759 if (likely(mergeable) && 760 copy_to_iter(&num_buffers, sizeof num_buffers, 761 &fixup) != sizeof num_buffers) { 762 vq_err(vq, "Failed num_buffers write"); 763 vhost_discard_vq_desc(vq, headcount); 764 goto out; 765 } 766 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads, 767 headcount); 768 if (unlikely(vq_log)) 769 vhost_log_write(vq, vq_log, log, vhost_len); 770 total_len += vhost_len; 771 if (unlikely(total_len >= VHOST_NET_WEIGHT)) { 772 vhost_poll_queue(&vq->poll); 773 goto out; 774 } 775 } 776 vhost_net_enable_vq(net, vq); 777 out: 778 mutex_unlock(&vq->mutex); 779 } 780 781 static void handle_tx_kick(struct vhost_work *work) 782 { 783 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, 784 poll.work); 785 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); 786 787 handle_tx(net); 788 } 789 790 static void handle_rx_kick(struct vhost_work *work) 791 { 792 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, 793 poll.work); 794 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); 795 796 handle_rx(net); 797 } 798 799 static void handle_tx_net(struct vhost_work *work) 800 { 801 struct vhost_net *net = container_of(work, struct vhost_net, 802 poll[VHOST_NET_VQ_TX].work); 803 handle_tx(net); 804 } 805 806 static void handle_rx_net(struct vhost_work *work) 807 { 808 struct vhost_net *net = container_of(work, struct vhost_net, 809 poll[VHOST_NET_VQ_RX].work); 810 handle_rx(net); 811 } 812 813 static int vhost_net_open(struct inode *inode, struct file *f) 814 { 815 struct vhost_net *n; 816 struct vhost_dev *dev; 817 struct vhost_virtqueue **vqs; 818 int i; 819 820 n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT); 821 if (!n) { 822 n = vmalloc(sizeof *n); 823 if (!n) 824 return -ENOMEM; 825 } 826 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL); 827 if (!vqs) { 828 kvfree(n); 829 return -ENOMEM; 830 } 831 832 dev = &n->dev; 833 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq; 834 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq; 835 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick; 836 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick; 837 for (i = 0; i < VHOST_NET_VQ_MAX; i++) { 838 n->vqs[i].ubufs = NULL; 839 n->vqs[i].ubuf_info = NULL; 840 n->vqs[i].upend_idx = 0; 841 n->vqs[i].done_idx = 0; 842 n->vqs[i].vhost_hlen = 0; 843 n->vqs[i].sock_hlen = 0; 844 } 845 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX); 846 847 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev); 848 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev); 849 850 f->private_data = n; 851 852 return 0; 853 } 854 855 static struct socket *vhost_net_stop_vq(struct vhost_net *n, 856 struct vhost_virtqueue *vq) 857 { 858 struct socket *sock; 859 860 mutex_lock(&vq->mutex); 861 sock = vq->private_data; 862 vhost_net_disable_vq(n, vq); 863 vq->private_data = NULL; 864 mutex_unlock(&vq->mutex); 865 return sock; 866 } 867 868 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock, 869 struct socket **rx_sock) 870 { 871 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq); 872 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq); 873 } 874 875 static void vhost_net_flush_vq(struct vhost_net *n, int index) 876 { 877 vhost_poll_flush(n->poll + index); 878 vhost_poll_flush(&n->vqs[index].vq.poll); 879 } 880 881 static void vhost_net_flush(struct vhost_net *n) 882 { 883 vhost_net_flush_vq(n, VHOST_NET_VQ_TX); 884 vhost_net_flush_vq(n, VHOST_NET_VQ_RX); 885 if (n->vqs[VHOST_NET_VQ_TX].ubufs) { 886 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 887 n->tx_flush = true; 888 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 889 /* Wait for all lower device DMAs done. */ 890 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs); 891 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 892 n->tx_flush = false; 893 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1); 894 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 895 } 896 } 897 898 static int vhost_net_release(struct inode *inode, struct file *f) 899 { 900 struct vhost_net *n = f->private_data; 901 struct socket *tx_sock; 902 struct socket *rx_sock; 903 904 vhost_net_stop(n, &tx_sock, &rx_sock); 905 vhost_net_flush(n); 906 vhost_dev_stop(&n->dev); 907 vhost_dev_cleanup(&n->dev, false); 908 vhost_net_vq_reset(n); 909 if (tx_sock) 910 sockfd_put(tx_sock); 911 if (rx_sock) 912 sockfd_put(rx_sock); 913 /* Make sure no callbacks are outstanding */ 914 synchronize_rcu_bh(); 915 /* We do an extra flush before freeing memory, 916 * since jobs can re-queue themselves. */ 917 vhost_net_flush(n); 918 kfree(n->dev.vqs); 919 kvfree(n); 920 return 0; 921 } 922 923 static struct socket *get_raw_socket(int fd) 924 { 925 struct { 926 struct sockaddr_ll sa; 927 char buf[MAX_ADDR_LEN]; 928 } uaddr; 929 int uaddr_len = sizeof uaddr, r; 930 struct socket *sock = sockfd_lookup(fd, &r); 931 932 if (!sock) 933 return ERR_PTR(-ENOTSOCK); 934 935 /* Parameter checking */ 936 if (sock->sk->sk_type != SOCK_RAW) { 937 r = -ESOCKTNOSUPPORT; 938 goto err; 939 } 940 941 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa, 942 &uaddr_len, 0); 943 if (r) 944 goto err; 945 946 if (uaddr.sa.sll_family != AF_PACKET) { 947 r = -EPFNOSUPPORT; 948 goto err; 949 } 950 return sock; 951 err: 952 sockfd_put(sock); 953 return ERR_PTR(r); 954 } 955 956 static struct socket *get_tap_socket(int fd) 957 { 958 struct file *file = fget(fd); 959 struct socket *sock; 960 961 if (!file) 962 return ERR_PTR(-EBADF); 963 sock = tun_get_socket(file); 964 if (!IS_ERR(sock)) 965 return sock; 966 sock = tap_get_socket(file); 967 if (IS_ERR(sock)) 968 fput(file); 969 return sock; 970 } 971 972 static struct socket *get_socket(int fd) 973 { 974 struct socket *sock; 975 976 /* special case to disable backend */ 977 if (fd == -1) 978 return NULL; 979 sock = get_raw_socket(fd); 980 if (!IS_ERR(sock)) 981 return sock; 982 sock = get_tap_socket(fd); 983 if (!IS_ERR(sock)) 984 return sock; 985 return ERR_PTR(-ENOTSOCK); 986 } 987 988 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd) 989 { 990 struct socket *sock, *oldsock; 991 struct vhost_virtqueue *vq; 992 struct vhost_net_virtqueue *nvq; 993 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL; 994 int r; 995 996 mutex_lock(&n->dev.mutex); 997 r = vhost_dev_check_owner(&n->dev); 998 if (r) 999 goto err; 1000 1001 if (index >= VHOST_NET_VQ_MAX) { 1002 r = -ENOBUFS; 1003 goto err; 1004 } 1005 vq = &n->vqs[index].vq; 1006 nvq = &n->vqs[index]; 1007 mutex_lock(&vq->mutex); 1008 1009 /* Verify that ring has been setup correctly. */ 1010 if (!vhost_vq_access_ok(vq)) { 1011 r = -EFAULT; 1012 goto err_vq; 1013 } 1014 sock = get_socket(fd); 1015 if (IS_ERR(sock)) { 1016 r = PTR_ERR(sock); 1017 goto err_vq; 1018 } 1019 1020 /* start polling new socket */ 1021 oldsock = vq->private_data; 1022 if (sock != oldsock) { 1023 ubufs = vhost_net_ubuf_alloc(vq, 1024 sock && vhost_sock_zcopy(sock)); 1025 if (IS_ERR(ubufs)) { 1026 r = PTR_ERR(ubufs); 1027 goto err_ubufs; 1028 } 1029 1030 vhost_net_disable_vq(n, vq); 1031 vq->private_data = sock; 1032 r = vhost_vq_init_access(vq); 1033 if (r) 1034 goto err_used; 1035 r = vhost_net_enable_vq(n, vq); 1036 if (r) 1037 goto err_used; 1038 1039 oldubufs = nvq->ubufs; 1040 nvq->ubufs = ubufs; 1041 1042 n->tx_packets = 0; 1043 n->tx_zcopy_err = 0; 1044 n->tx_flush = false; 1045 } 1046 1047 mutex_unlock(&vq->mutex); 1048 1049 if (oldubufs) { 1050 vhost_net_ubuf_put_wait_and_free(oldubufs); 1051 mutex_lock(&vq->mutex); 1052 vhost_zerocopy_signal_used(n, vq); 1053 mutex_unlock(&vq->mutex); 1054 } 1055 1056 if (oldsock) { 1057 vhost_net_flush_vq(n, index); 1058 sockfd_put(oldsock); 1059 } 1060 1061 mutex_unlock(&n->dev.mutex); 1062 return 0; 1063 1064 err_used: 1065 vq->private_data = oldsock; 1066 vhost_net_enable_vq(n, vq); 1067 if (ubufs) 1068 vhost_net_ubuf_put_wait_and_free(ubufs); 1069 err_ubufs: 1070 sockfd_put(sock); 1071 err_vq: 1072 mutex_unlock(&vq->mutex); 1073 err: 1074 mutex_unlock(&n->dev.mutex); 1075 return r; 1076 } 1077 1078 static long vhost_net_reset_owner(struct vhost_net *n) 1079 { 1080 struct socket *tx_sock = NULL; 1081 struct socket *rx_sock = NULL; 1082 long err; 1083 struct vhost_umem *umem; 1084 1085 mutex_lock(&n->dev.mutex); 1086 err = vhost_dev_check_owner(&n->dev); 1087 if (err) 1088 goto done; 1089 umem = vhost_dev_reset_owner_prepare(); 1090 if (!umem) { 1091 err = -ENOMEM; 1092 goto done; 1093 } 1094 vhost_net_stop(n, &tx_sock, &rx_sock); 1095 vhost_net_flush(n); 1096 vhost_dev_reset_owner(&n->dev, umem); 1097 vhost_net_vq_reset(n); 1098 done: 1099 mutex_unlock(&n->dev.mutex); 1100 if (tx_sock) 1101 sockfd_put(tx_sock); 1102 if (rx_sock) 1103 sockfd_put(rx_sock); 1104 return err; 1105 } 1106 1107 static int vhost_net_set_features(struct vhost_net *n, u64 features) 1108 { 1109 size_t vhost_hlen, sock_hlen, hdr_len; 1110 int i; 1111 1112 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) | 1113 (1ULL << VIRTIO_F_VERSION_1))) ? 1114 sizeof(struct virtio_net_hdr_mrg_rxbuf) : 1115 sizeof(struct virtio_net_hdr); 1116 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) { 1117 /* vhost provides vnet_hdr */ 1118 vhost_hlen = hdr_len; 1119 sock_hlen = 0; 1120 } else { 1121 /* socket provides vnet_hdr */ 1122 vhost_hlen = 0; 1123 sock_hlen = hdr_len; 1124 } 1125 mutex_lock(&n->dev.mutex); 1126 if ((features & (1 << VHOST_F_LOG_ALL)) && 1127 !vhost_log_access_ok(&n->dev)) 1128 goto out_unlock; 1129 1130 if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) { 1131 if (vhost_init_device_iotlb(&n->dev, true)) 1132 goto out_unlock; 1133 } 1134 1135 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 1136 mutex_lock(&n->vqs[i].vq.mutex); 1137 n->vqs[i].vq.acked_features = features; 1138 n->vqs[i].vhost_hlen = vhost_hlen; 1139 n->vqs[i].sock_hlen = sock_hlen; 1140 mutex_unlock(&n->vqs[i].vq.mutex); 1141 } 1142 mutex_unlock(&n->dev.mutex); 1143 return 0; 1144 1145 out_unlock: 1146 mutex_unlock(&n->dev.mutex); 1147 return -EFAULT; 1148 } 1149 1150 static long vhost_net_set_owner(struct vhost_net *n) 1151 { 1152 int r; 1153 1154 mutex_lock(&n->dev.mutex); 1155 if (vhost_dev_has_owner(&n->dev)) { 1156 r = -EBUSY; 1157 goto out; 1158 } 1159 r = vhost_net_set_ubuf_info(n); 1160 if (r) 1161 goto out; 1162 r = vhost_dev_set_owner(&n->dev); 1163 if (r) 1164 vhost_net_clear_ubuf_info(n); 1165 vhost_net_flush(n); 1166 out: 1167 mutex_unlock(&n->dev.mutex); 1168 return r; 1169 } 1170 1171 static long vhost_net_ioctl(struct file *f, unsigned int ioctl, 1172 unsigned long arg) 1173 { 1174 struct vhost_net *n = f->private_data; 1175 void __user *argp = (void __user *)arg; 1176 u64 __user *featurep = argp; 1177 struct vhost_vring_file backend; 1178 u64 features; 1179 int r; 1180 1181 switch (ioctl) { 1182 case VHOST_NET_SET_BACKEND: 1183 if (copy_from_user(&backend, argp, sizeof backend)) 1184 return -EFAULT; 1185 return vhost_net_set_backend(n, backend.index, backend.fd); 1186 case VHOST_GET_FEATURES: 1187 features = VHOST_NET_FEATURES; 1188 if (copy_to_user(featurep, &features, sizeof features)) 1189 return -EFAULT; 1190 return 0; 1191 case VHOST_SET_FEATURES: 1192 if (copy_from_user(&features, featurep, sizeof features)) 1193 return -EFAULT; 1194 if (features & ~VHOST_NET_FEATURES) 1195 return -EOPNOTSUPP; 1196 return vhost_net_set_features(n, features); 1197 case VHOST_RESET_OWNER: 1198 return vhost_net_reset_owner(n); 1199 case VHOST_SET_OWNER: 1200 return vhost_net_set_owner(n); 1201 default: 1202 mutex_lock(&n->dev.mutex); 1203 r = vhost_dev_ioctl(&n->dev, ioctl, argp); 1204 if (r == -ENOIOCTLCMD) 1205 r = vhost_vring_ioctl(&n->dev, ioctl, argp); 1206 else 1207 vhost_net_flush(n); 1208 mutex_unlock(&n->dev.mutex); 1209 return r; 1210 } 1211 } 1212 1213 #ifdef CONFIG_COMPAT 1214 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl, 1215 unsigned long arg) 1216 { 1217 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg)); 1218 } 1219 #endif 1220 1221 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to) 1222 { 1223 struct file *file = iocb->ki_filp; 1224 struct vhost_net *n = file->private_data; 1225 struct vhost_dev *dev = &n->dev; 1226 int noblock = file->f_flags & O_NONBLOCK; 1227 1228 return vhost_chr_read_iter(dev, to, noblock); 1229 } 1230 1231 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb, 1232 struct iov_iter *from) 1233 { 1234 struct file *file = iocb->ki_filp; 1235 struct vhost_net *n = file->private_data; 1236 struct vhost_dev *dev = &n->dev; 1237 1238 return vhost_chr_write_iter(dev, from); 1239 } 1240 1241 static unsigned int vhost_net_chr_poll(struct file *file, poll_table *wait) 1242 { 1243 struct vhost_net *n = file->private_data; 1244 struct vhost_dev *dev = &n->dev; 1245 1246 return vhost_chr_poll(file, dev, wait); 1247 } 1248 1249 static const struct file_operations vhost_net_fops = { 1250 .owner = THIS_MODULE, 1251 .release = vhost_net_release, 1252 .read_iter = vhost_net_chr_read_iter, 1253 .write_iter = vhost_net_chr_write_iter, 1254 .poll = vhost_net_chr_poll, 1255 .unlocked_ioctl = vhost_net_ioctl, 1256 #ifdef CONFIG_COMPAT 1257 .compat_ioctl = vhost_net_compat_ioctl, 1258 #endif 1259 .open = vhost_net_open, 1260 .llseek = noop_llseek, 1261 }; 1262 1263 static struct miscdevice vhost_net_misc = { 1264 .minor = VHOST_NET_MINOR, 1265 .name = "vhost-net", 1266 .fops = &vhost_net_fops, 1267 }; 1268 1269 static int vhost_net_init(void) 1270 { 1271 if (experimental_zcopytx) 1272 vhost_net_enable_zcopy(VHOST_NET_VQ_TX); 1273 return misc_register(&vhost_net_misc); 1274 } 1275 module_init(vhost_net_init); 1276 1277 static void vhost_net_exit(void) 1278 { 1279 misc_deregister(&vhost_net_misc); 1280 } 1281 module_exit(vhost_net_exit); 1282 1283 MODULE_VERSION("0.0.1"); 1284 MODULE_LICENSE("GPL v2"); 1285 MODULE_AUTHOR("Michael S. Tsirkin"); 1286 MODULE_DESCRIPTION("Host kernel accelerator for virtio net"); 1287 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR); 1288 MODULE_ALIAS("devname:vhost-net"); 1289