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