1 // SPDX-License-Identifier: GPL-2.0-only 2 /* Copyright (C) 2009 Red Hat, Inc. 3 * Author: Michael S. Tsirkin <mst@redhat.com> 4 * 5 * virtio-net server in host kernel. 6 */ 7 8 #include <linux/compat.h> 9 #include <linux/eventfd.h> 10 #include <linux/vhost.h> 11 #include <linux/virtio_net.h> 12 #include <linux/miscdevice.h> 13 #include <linux/module.h> 14 #include <linux/moduleparam.h> 15 #include <linux/mutex.h> 16 #include <linux/workqueue.h> 17 #include <linux/file.h> 18 #include <linux/slab.h> 19 #include <linux/sched/clock.h> 20 #include <linux/sched/signal.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 #include <linux/skb_array.h> 31 #include <linux/skbuff.h> 32 33 #include <net/sock.h> 34 #include <net/xdp.h> 35 36 #include "vhost.h" 37 38 static int experimental_zcopytx = 0; 39 module_param(experimental_zcopytx, int, 0444); 40 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;" 41 " 1 -Enable; 0 - Disable"); 42 43 /* Max number of bytes transferred before requeueing the job. 44 * Using this limit prevents one virtqueue from starving others. */ 45 #define VHOST_NET_WEIGHT 0x80000 46 47 /* Max number of packets transferred before requeueing the job. 48 * Using this limit prevents one virtqueue from starving others with small 49 * pkts. 50 */ 51 #define VHOST_NET_PKT_WEIGHT 256 52 53 /* MAX number of TX used buffers for outstanding zerocopy */ 54 #define VHOST_MAX_PEND 128 55 #define VHOST_GOODCOPY_LEN 256 56 57 /* 58 * For transmit, used buffer len is unused; we override it to track buffer 59 * status internally; used for zerocopy tx only. 60 */ 61 /* Lower device DMA failed */ 62 #define VHOST_DMA_FAILED_LEN ((__force __virtio32)3) 63 /* Lower device DMA done */ 64 #define VHOST_DMA_DONE_LEN ((__force __virtio32)2) 65 /* Lower device DMA in progress */ 66 #define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1) 67 /* Buffer unused */ 68 #define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0) 69 70 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN) 71 72 enum { 73 VHOST_NET_FEATURES = VHOST_FEATURES | 74 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) | 75 (1ULL << VIRTIO_NET_F_MRG_RXBUF) | 76 (1ULL << VIRTIO_F_ACCESS_PLATFORM) 77 }; 78 79 enum { 80 VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2) 81 }; 82 83 enum { 84 VHOST_NET_VQ_RX = 0, 85 VHOST_NET_VQ_TX = 1, 86 VHOST_NET_VQ_MAX = 2, 87 }; 88 89 struct vhost_net_ubuf_ref { 90 /* refcount follows semantics similar to kref: 91 * 0: object is released 92 * 1: no outstanding ubufs 93 * >1: outstanding ubufs 94 */ 95 atomic_t refcount; 96 wait_queue_head_t wait; 97 struct vhost_virtqueue *vq; 98 }; 99 100 #define VHOST_NET_BATCH 64 101 struct vhost_net_buf { 102 void **queue; 103 int tail; 104 int head; 105 }; 106 107 struct vhost_net_virtqueue { 108 struct vhost_virtqueue vq; 109 size_t vhost_hlen; 110 size_t sock_hlen; 111 /* vhost zerocopy support fields below: */ 112 /* last used idx for outstanding DMA zerocopy buffers */ 113 int upend_idx; 114 /* For TX, first used idx for DMA done zerocopy buffers 115 * For RX, number of batched heads 116 */ 117 int done_idx; 118 /* Number of XDP frames batched */ 119 int batched_xdp; 120 /* an array of userspace buffers info */ 121 struct ubuf_info *ubuf_info; 122 /* Reference counting for outstanding ubufs. 123 * Protected by vq mutex. Writers must also take device mutex. */ 124 struct vhost_net_ubuf_ref *ubufs; 125 struct ptr_ring *rx_ring; 126 struct vhost_net_buf rxq; 127 /* Batched XDP buffs */ 128 struct xdp_buff *xdp; 129 }; 130 131 struct vhost_net { 132 struct vhost_dev dev; 133 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX]; 134 struct vhost_poll poll[VHOST_NET_VQ_MAX]; 135 /* Number of TX recently submitted. 136 * Protected by tx vq lock. */ 137 unsigned tx_packets; 138 /* Number of times zerocopy TX recently failed. 139 * Protected by tx vq lock. */ 140 unsigned tx_zcopy_err; 141 /* Flush in progress. Protected by tx vq lock. */ 142 bool tx_flush; 143 /* Private page frag */ 144 struct page_frag page_frag; 145 /* Refcount bias of page frag */ 146 int refcnt_bias; 147 }; 148 149 static unsigned vhost_net_zcopy_mask __read_mostly; 150 151 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq) 152 { 153 if (rxq->tail != rxq->head) 154 return rxq->queue[rxq->head]; 155 else 156 return NULL; 157 } 158 159 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq) 160 { 161 return rxq->tail - rxq->head; 162 } 163 164 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq) 165 { 166 return rxq->tail == rxq->head; 167 } 168 169 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq) 170 { 171 void *ret = vhost_net_buf_get_ptr(rxq); 172 ++rxq->head; 173 return ret; 174 } 175 176 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq) 177 { 178 struct vhost_net_buf *rxq = &nvq->rxq; 179 180 rxq->head = 0; 181 rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue, 182 VHOST_NET_BATCH); 183 return rxq->tail; 184 } 185 186 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq) 187 { 188 struct vhost_net_buf *rxq = &nvq->rxq; 189 190 if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) { 191 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head, 192 vhost_net_buf_get_size(rxq), 193 tun_ptr_free); 194 rxq->head = rxq->tail = 0; 195 } 196 } 197 198 static int vhost_net_buf_peek_len(void *ptr) 199 { 200 if (tun_is_xdp_frame(ptr)) { 201 struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr); 202 203 return xdpf->len; 204 } 205 206 return __skb_array_len_with_tag(ptr); 207 } 208 209 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq) 210 { 211 struct vhost_net_buf *rxq = &nvq->rxq; 212 213 if (!vhost_net_buf_is_empty(rxq)) 214 goto out; 215 216 if (!vhost_net_buf_produce(nvq)) 217 return 0; 218 219 out: 220 return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq)); 221 } 222 223 static void vhost_net_buf_init(struct vhost_net_buf *rxq) 224 { 225 rxq->head = rxq->tail = 0; 226 } 227 228 static void vhost_net_enable_zcopy(int vq) 229 { 230 vhost_net_zcopy_mask |= 0x1 << vq; 231 } 232 233 static struct vhost_net_ubuf_ref * 234 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy) 235 { 236 struct vhost_net_ubuf_ref *ubufs; 237 /* No zero copy backend? Nothing to count. */ 238 if (!zcopy) 239 return NULL; 240 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL); 241 if (!ubufs) 242 return ERR_PTR(-ENOMEM); 243 atomic_set(&ubufs->refcount, 1); 244 init_waitqueue_head(&ubufs->wait); 245 ubufs->vq = vq; 246 return ubufs; 247 } 248 249 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs) 250 { 251 int r = atomic_sub_return(1, &ubufs->refcount); 252 if (unlikely(!r)) 253 wake_up(&ubufs->wait); 254 return r; 255 } 256 257 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs) 258 { 259 vhost_net_ubuf_put(ubufs); 260 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount)); 261 } 262 263 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs) 264 { 265 vhost_net_ubuf_put_and_wait(ubufs); 266 kfree(ubufs); 267 } 268 269 static void vhost_net_clear_ubuf_info(struct vhost_net *n) 270 { 271 int i; 272 273 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 274 kfree(n->vqs[i].ubuf_info); 275 n->vqs[i].ubuf_info = NULL; 276 } 277 } 278 279 static int vhost_net_set_ubuf_info(struct vhost_net *n) 280 { 281 bool zcopy; 282 int i; 283 284 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 285 zcopy = vhost_net_zcopy_mask & (0x1 << i); 286 if (!zcopy) 287 continue; 288 n->vqs[i].ubuf_info = 289 kmalloc_array(UIO_MAXIOV, 290 sizeof(*n->vqs[i].ubuf_info), 291 GFP_KERNEL); 292 if (!n->vqs[i].ubuf_info) 293 goto err; 294 } 295 return 0; 296 297 err: 298 vhost_net_clear_ubuf_info(n); 299 return -ENOMEM; 300 } 301 302 static void vhost_net_vq_reset(struct vhost_net *n) 303 { 304 int i; 305 306 vhost_net_clear_ubuf_info(n); 307 308 for (i = 0; i < VHOST_NET_VQ_MAX; i++) { 309 n->vqs[i].done_idx = 0; 310 n->vqs[i].upend_idx = 0; 311 n->vqs[i].ubufs = NULL; 312 n->vqs[i].vhost_hlen = 0; 313 n->vqs[i].sock_hlen = 0; 314 vhost_net_buf_init(&n->vqs[i].rxq); 315 } 316 317 } 318 319 static void vhost_net_tx_packet(struct vhost_net *net) 320 { 321 ++net->tx_packets; 322 if (net->tx_packets < 1024) 323 return; 324 net->tx_packets = 0; 325 net->tx_zcopy_err = 0; 326 } 327 328 static void vhost_net_tx_err(struct vhost_net *net) 329 { 330 ++net->tx_zcopy_err; 331 } 332 333 static bool vhost_net_tx_select_zcopy(struct vhost_net *net) 334 { 335 /* TX flush waits for outstanding DMAs to be done. 336 * Don't start new DMAs. 337 */ 338 return !net->tx_flush && 339 net->tx_packets / 64 >= net->tx_zcopy_err; 340 } 341 342 static bool vhost_sock_zcopy(struct socket *sock) 343 { 344 return unlikely(experimental_zcopytx) && 345 sock_flag(sock->sk, SOCK_ZEROCOPY); 346 } 347 348 static bool vhost_sock_xdp(struct socket *sock) 349 { 350 return sock_flag(sock->sk, SOCK_XDP); 351 } 352 353 /* In case of DMA done not in order in lower device driver for some reason. 354 * upend_idx is used to track end of used idx, done_idx is used to track head 355 * of used idx. Once lower device DMA done contiguously, we will signal KVM 356 * guest used idx. 357 */ 358 static void vhost_zerocopy_signal_used(struct vhost_net *net, 359 struct vhost_virtqueue *vq) 360 { 361 struct vhost_net_virtqueue *nvq = 362 container_of(vq, struct vhost_net_virtqueue, vq); 363 int i, add; 364 int j = 0; 365 366 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) { 367 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN) 368 vhost_net_tx_err(net); 369 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) { 370 vq->heads[i].len = VHOST_DMA_CLEAR_LEN; 371 ++j; 372 } else 373 break; 374 } 375 while (j) { 376 add = min(UIO_MAXIOV - nvq->done_idx, j); 377 vhost_add_used_and_signal_n(vq->dev, vq, 378 &vq->heads[nvq->done_idx], add); 379 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV; 380 j -= add; 381 } 382 } 383 384 static void vhost_zerocopy_callback(struct sk_buff *skb, 385 struct ubuf_info *ubuf, bool success) 386 { 387 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx; 388 struct vhost_virtqueue *vq = ubufs->vq; 389 int cnt; 390 391 rcu_read_lock_bh(); 392 393 /* set len to mark this desc buffers done DMA */ 394 vq->heads[ubuf->desc].len = success ? 395 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN; 396 cnt = vhost_net_ubuf_put(ubufs); 397 398 /* 399 * Trigger polling thread if guest stopped submitting new buffers: 400 * in this case, the refcount after decrement will eventually reach 1. 401 * We also trigger polling periodically after each 16 packets 402 * (the value 16 here is more or less arbitrary, it's tuned to trigger 403 * less than 10% of times). 404 */ 405 if (cnt <= 1 || !(cnt % 16)) 406 vhost_poll_queue(&vq->poll); 407 408 rcu_read_unlock_bh(); 409 } 410 411 static inline unsigned long busy_clock(void) 412 { 413 return local_clock() >> 10; 414 } 415 416 static bool vhost_can_busy_poll(unsigned long endtime) 417 { 418 return likely(!need_resched() && !time_after(busy_clock(), endtime) && 419 !signal_pending(current)); 420 } 421 422 static void vhost_net_disable_vq(struct vhost_net *n, 423 struct vhost_virtqueue *vq) 424 { 425 struct vhost_net_virtqueue *nvq = 426 container_of(vq, struct vhost_net_virtqueue, vq); 427 struct vhost_poll *poll = n->poll + (nvq - n->vqs); 428 if (!vhost_vq_get_backend(vq)) 429 return; 430 vhost_poll_stop(poll); 431 } 432 433 static int vhost_net_enable_vq(struct vhost_net *n, 434 struct vhost_virtqueue *vq) 435 { 436 struct vhost_net_virtqueue *nvq = 437 container_of(vq, struct vhost_net_virtqueue, vq); 438 struct vhost_poll *poll = n->poll + (nvq - n->vqs); 439 struct socket *sock; 440 441 sock = vhost_vq_get_backend(vq); 442 if (!sock) 443 return 0; 444 445 return vhost_poll_start(poll, sock->file); 446 } 447 448 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq) 449 { 450 struct vhost_virtqueue *vq = &nvq->vq; 451 struct vhost_dev *dev = vq->dev; 452 453 if (!nvq->done_idx) 454 return; 455 456 vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx); 457 nvq->done_idx = 0; 458 } 459 460 static void vhost_tx_batch(struct vhost_net *net, 461 struct vhost_net_virtqueue *nvq, 462 struct socket *sock, 463 struct msghdr *msghdr) 464 { 465 struct tun_msg_ctl ctl = { 466 .type = TUN_MSG_PTR, 467 .num = nvq->batched_xdp, 468 .ptr = nvq->xdp, 469 }; 470 int i, err; 471 472 if (nvq->batched_xdp == 0) 473 goto signal_used; 474 475 msghdr->msg_control = &ctl; 476 msghdr->msg_controllen = sizeof(ctl); 477 err = sock->ops->sendmsg(sock, msghdr, 0); 478 if (unlikely(err < 0)) { 479 vq_err(&nvq->vq, "Fail to batch sending packets\n"); 480 481 /* free pages owned by XDP; since this is an unlikely error path, 482 * keep it simple and avoid more complex bulk update for the 483 * used pages 484 */ 485 for (i = 0; i < nvq->batched_xdp; ++i) 486 put_page(virt_to_head_page(nvq->xdp[i].data)); 487 nvq->batched_xdp = 0; 488 nvq->done_idx = 0; 489 return; 490 } 491 492 signal_used: 493 vhost_net_signal_used(nvq); 494 nvq->batched_xdp = 0; 495 } 496 497 static int sock_has_rx_data(struct socket *sock) 498 { 499 if (unlikely(!sock)) 500 return 0; 501 502 if (sock->ops->peek_len) 503 return sock->ops->peek_len(sock); 504 505 return skb_queue_empty(&sock->sk->sk_receive_queue); 506 } 507 508 static void vhost_net_busy_poll_try_queue(struct vhost_net *net, 509 struct vhost_virtqueue *vq) 510 { 511 if (!vhost_vq_avail_empty(&net->dev, vq)) { 512 vhost_poll_queue(&vq->poll); 513 } else if (unlikely(vhost_enable_notify(&net->dev, vq))) { 514 vhost_disable_notify(&net->dev, vq); 515 vhost_poll_queue(&vq->poll); 516 } 517 } 518 519 static void vhost_net_busy_poll(struct vhost_net *net, 520 struct vhost_virtqueue *rvq, 521 struct vhost_virtqueue *tvq, 522 bool *busyloop_intr, 523 bool poll_rx) 524 { 525 unsigned long busyloop_timeout; 526 unsigned long endtime; 527 struct socket *sock; 528 struct vhost_virtqueue *vq = poll_rx ? tvq : rvq; 529 530 /* Try to hold the vq mutex of the paired virtqueue. We can't 531 * use mutex_lock() here since we could not guarantee a 532 * consistenet lock ordering. 533 */ 534 if (!mutex_trylock(&vq->mutex)) 535 return; 536 537 vhost_disable_notify(&net->dev, vq); 538 sock = vhost_vq_get_backend(rvq); 539 540 busyloop_timeout = poll_rx ? rvq->busyloop_timeout: 541 tvq->busyloop_timeout; 542 543 preempt_disable(); 544 endtime = busy_clock() + busyloop_timeout; 545 546 while (vhost_can_busy_poll(endtime)) { 547 if (vhost_has_work(&net->dev)) { 548 *busyloop_intr = true; 549 break; 550 } 551 552 if ((sock_has_rx_data(sock) && 553 !vhost_vq_avail_empty(&net->dev, rvq)) || 554 !vhost_vq_avail_empty(&net->dev, tvq)) 555 break; 556 557 cpu_relax(); 558 } 559 560 preempt_enable(); 561 562 if (poll_rx || sock_has_rx_data(sock)) 563 vhost_net_busy_poll_try_queue(net, vq); 564 else if (!poll_rx) /* On tx here, sock has no rx data. */ 565 vhost_enable_notify(&net->dev, rvq); 566 567 mutex_unlock(&vq->mutex); 568 } 569 570 static int vhost_net_tx_get_vq_desc(struct vhost_net *net, 571 struct vhost_net_virtqueue *tnvq, 572 unsigned int *out_num, unsigned int *in_num, 573 struct msghdr *msghdr, bool *busyloop_intr) 574 { 575 struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX]; 576 struct vhost_virtqueue *rvq = &rnvq->vq; 577 struct vhost_virtqueue *tvq = &tnvq->vq; 578 579 int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov), 580 out_num, in_num, NULL, NULL); 581 582 if (r == tvq->num && tvq->busyloop_timeout) { 583 /* Flush batched packets first */ 584 if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq))) 585 vhost_tx_batch(net, tnvq, 586 vhost_vq_get_backend(tvq), 587 msghdr); 588 589 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false); 590 591 r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov), 592 out_num, in_num, NULL, NULL); 593 } 594 595 return r; 596 } 597 598 static bool vhost_exceeds_maxpend(struct vhost_net *net) 599 { 600 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 601 struct vhost_virtqueue *vq = &nvq->vq; 602 603 return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV > 604 min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2); 605 } 606 607 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter, 608 size_t hdr_size, int out) 609 { 610 /* Skip header. TODO: support TSO. */ 611 size_t len = iov_length(vq->iov, out); 612 613 iov_iter_init(iter, WRITE, vq->iov, out, len); 614 iov_iter_advance(iter, hdr_size); 615 616 return iov_iter_count(iter); 617 } 618 619 static int get_tx_bufs(struct vhost_net *net, 620 struct vhost_net_virtqueue *nvq, 621 struct msghdr *msg, 622 unsigned int *out, unsigned int *in, 623 size_t *len, bool *busyloop_intr) 624 { 625 struct vhost_virtqueue *vq = &nvq->vq; 626 int ret; 627 628 ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr); 629 630 if (ret < 0 || ret == vq->num) 631 return ret; 632 633 if (*in) { 634 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n", 635 *out, *in); 636 return -EFAULT; 637 } 638 639 /* Sanity check */ 640 *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out); 641 if (*len == 0) { 642 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n", 643 *len, nvq->vhost_hlen); 644 return -EFAULT; 645 } 646 647 return ret; 648 } 649 650 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len) 651 { 652 return total_len < VHOST_NET_WEIGHT && 653 !vhost_vq_avail_empty(vq->dev, vq); 654 } 655 656 static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz, 657 struct page_frag *pfrag, gfp_t gfp) 658 { 659 if (pfrag->page) { 660 if (pfrag->offset + sz <= pfrag->size) 661 return true; 662 __page_frag_cache_drain(pfrag->page, net->refcnt_bias); 663 } 664 665 pfrag->offset = 0; 666 net->refcnt_bias = 0; 667 if (SKB_FRAG_PAGE_ORDER) { 668 /* Avoid direct reclaim but allow kswapd to wake */ 669 pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) | 670 __GFP_COMP | __GFP_NOWARN | 671 __GFP_NORETRY, 672 SKB_FRAG_PAGE_ORDER); 673 if (likely(pfrag->page)) { 674 pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER; 675 goto done; 676 } 677 } 678 pfrag->page = alloc_page(gfp); 679 if (likely(pfrag->page)) { 680 pfrag->size = PAGE_SIZE; 681 goto done; 682 } 683 return false; 684 685 done: 686 net->refcnt_bias = USHRT_MAX; 687 page_ref_add(pfrag->page, USHRT_MAX - 1); 688 return true; 689 } 690 691 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD) 692 693 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq, 694 struct iov_iter *from) 695 { 696 struct vhost_virtqueue *vq = &nvq->vq; 697 struct vhost_net *net = container_of(vq->dev, struct vhost_net, 698 dev); 699 struct socket *sock = vhost_vq_get_backend(vq); 700 struct page_frag *alloc_frag = &net->page_frag; 701 struct virtio_net_hdr *gso; 702 struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp]; 703 struct tun_xdp_hdr *hdr; 704 size_t len = iov_iter_count(from); 705 int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0; 706 int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 707 int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen); 708 int sock_hlen = nvq->sock_hlen; 709 void *buf; 710 int copied; 711 712 if (unlikely(len < nvq->sock_hlen)) 713 return -EFAULT; 714 715 if (SKB_DATA_ALIGN(len + pad) + 716 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE) 717 return -ENOSPC; 718 719 buflen += SKB_DATA_ALIGN(len + pad); 720 alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES); 721 if (unlikely(!vhost_net_page_frag_refill(net, buflen, 722 alloc_frag, GFP_KERNEL))) 723 return -ENOMEM; 724 725 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 726 copied = copy_page_from_iter(alloc_frag->page, 727 alloc_frag->offset + 728 offsetof(struct tun_xdp_hdr, gso), 729 sock_hlen, from); 730 if (copied != sock_hlen) 731 return -EFAULT; 732 733 hdr = buf; 734 gso = &hdr->gso; 735 736 if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) && 737 vhost16_to_cpu(vq, gso->csum_start) + 738 vhost16_to_cpu(vq, gso->csum_offset) + 2 > 739 vhost16_to_cpu(vq, gso->hdr_len)) { 740 gso->hdr_len = cpu_to_vhost16(vq, 741 vhost16_to_cpu(vq, gso->csum_start) + 742 vhost16_to_cpu(vq, gso->csum_offset) + 2); 743 744 if (vhost16_to_cpu(vq, gso->hdr_len) > len) 745 return -EINVAL; 746 } 747 748 len -= sock_hlen; 749 copied = copy_page_from_iter(alloc_frag->page, 750 alloc_frag->offset + pad, 751 len, from); 752 if (copied != len) 753 return -EFAULT; 754 755 xdp_init_buff(xdp, buflen, NULL); 756 xdp_prepare_buff(xdp, buf, pad, len, true); 757 hdr->buflen = buflen; 758 759 --net->refcnt_bias; 760 alloc_frag->offset += buflen; 761 762 ++nvq->batched_xdp; 763 764 return 0; 765 } 766 767 static void handle_tx_copy(struct vhost_net *net, struct socket *sock) 768 { 769 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 770 struct vhost_virtqueue *vq = &nvq->vq; 771 unsigned out, in; 772 int head; 773 struct msghdr msg = { 774 .msg_name = NULL, 775 .msg_namelen = 0, 776 .msg_control = NULL, 777 .msg_controllen = 0, 778 .msg_flags = MSG_DONTWAIT, 779 }; 780 size_t len, total_len = 0; 781 int err; 782 int sent_pkts = 0; 783 bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX); 784 785 do { 786 bool busyloop_intr = false; 787 788 if (nvq->done_idx == VHOST_NET_BATCH) 789 vhost_tx_batch(net, nvq, sock, &msg); 790 791 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len, 792 &busyloop_intr); 793 /* On error, stop handling until the next kick. */ 794 if (unlikely(head < 0)) 795 break; 796 /* Nothing new? Wait for eventfd to tell us they refilled. */ 797 if (head == vq->num) { 798 if (unlikely(busyloop_intr)) { 799 vhost_poll_queue(&vq->poll); 800 } else if (unlikely(vhost_enable_notify(&net->dev, 801 vq))) { 802 vhost_disable_notify(&net->dev, vq); 803 continue; 804 } 805 break; 806 } 807 808 total_len += len; 809 810 /* For simplicity, TX batching is only enabled if 811 * sndbuf is unlimited. 812 */ 813 if (sock_can_batch) { 814 err = vhost_net_build_xdp(nvq, &msg.msg_iter); 815 if (!err) { 816 goto done; 817 } else if (unlikely(err != -ENOSPC)) { 818 vhost_tx_batch(net, nvq, sock, &msg); 819 vhost_discard_vq_desc(vq, 1); 820 vhost_net_enable_vq(net, vq); 821 break; 822 } 823 824 /* We can't build XDP buff, go for single 825 * packet path but let's flush batched 826 * packets. 827 */ 828 vhost_tx_batch(net, nvq, sock, &msg); 829 msg.msg_control = NULL; 830 } else { 831 if (tx_can_batch(vq, total_len)) 832 msg.msg_flags |= MSG_MORE; 833 else 834 msg.msg_flags &= ~MSG_MORE; 835 } 836 837 err = sock->ops->sendmsg(sock, &msg, len); 838 if (unlikely(err < 0)) { 839 if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) { 840 vhost_discard_vq_desc(vq, 1); 841 vhost_net_enable_vq(net, vq); 842 break; 843 } 844 pr_debug("Fail to send packet: err %d", err); 845 } else if (unlikely(err != len)) 846 pr_debug("Truncated TX packet: len %d != %zd\n", 847 err, len); 848 done: 849 vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head); 850 vq->heads[nvq->done_idx].len = 0; 851 ++nvq->done_idx; 852 } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len))); 853 854 vhost_tx_batch(net, nvq, sock, &msg); 855 } 856 857 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock) 858 { 859 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 860 struct vhost_virtqueue *vq = &nvq->vq; 861 unsigned out, in; 862 int head; 863 struct msghdr msg = { 864 .msg_name = NULL, 865 .msg_namelen = 0, 866 .msg_control = NULL, 867 .msg_controllen = 0, 868 .msg_flags = MSG_DONTWAIT, 869 }; 870 struct tun_msg_ctl ctl; 871 size_t len, total_len = 0; 872 int err; 873 struct vhost_net_ubuf_ref *ubufs; 874 struct ubuf_info *ubuf; 875 bool zcopy_used; 876 int sent_pkts = 0; 877 878 do { 879 bool busyloop_intr; 880 881 /* Release DMAs done buffers first */ 882 vhost_zerocopy_signal_used(net, vq); 883 884 busyloop_intr = false; 885 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len, 886 &busyloop_intr); 887 /* On error, stop handling until the next kick. */ 888 if (unlikely(head < 0)) 889 break; 890 /* Nothing new? Wait for eventfd to tell us they refilled. */ 891 if (head == vq->num) { 892 if (unlikely(busyloop_intr)) { 893 vhost_poll_queue(&vq->poll); 894 } else if (unlikely(vhost_enable_notify(&net->dev, vq))) { 895 vhost_disable_notify(&net->dev, vq); 896 continue; 897 } 898 break; 899 } 900 901 zcopy_used = len >= VHOST_GOODCOPY_LEN 902 && !vhost_exceeds_maxpend(net) 903 && vhost_net_tx_select_zcopy(net); 904 905 /* use msg_control to pass vhost zerocopy ubuf info to skb */ 906 if (zcopy_used) { 907 ubuf = nvq->ubuf_info + nvq->upend_idx; 908 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head); 909 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS; 910 ubuf->callback = vhost_zerocopy_callback; 911 ubuf->ctx = nvq->ubufs; 912 ubuf->desc = nvq->upend_idx; 913 ubuf->flags = SKBFL_ZEROCOPY_FRAG; 914 refcount_set(&ubuf->refcnt, 1); 915 msg.msg_control = &ctl; 916 ctl.type = TUN_MSG_UBUF; 917 ctl.ptr = ubuf; 918 msg.msg_controllen = sizeof(ctl); 919 ubufs = nvq->ubufs; 920 atomic_inc(&ubufs->refcount); 921 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV; 922 } else { 923 msg.msg_control = NULL; 924 ubufs = NULL; 925 } 926 total_len += len; 927 if (tx_can_batch(vq, total_len) && 928 likely(!vhost_exceeds_maxpend(net))) { 929 msg.msg_flags |= MSG_MORE; 930 } else { 931 msg.msg_flags &= ~MSG_MORE; 932 } 933 934 err = sock->ops->sendmsg(sock, &msg, len); 935 if (unlikely(err < 0)) { 936 if (zcopy_used) { 937 if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS) 938 vhost_net_ubuf_put(ubufs); 939 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1) 940 % UIO_MAXIOV; 941 } 942 if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) { 943 vhost_discard_vq_desc(vq, 1); 944 vhost_net_enable_vq(net, vq); 945 break; 946 } 947 pr_debug("Fail to send packet: err %d", err); 948 } else if (unlikely(err != len)) 949 pr_debug("Truncated TX packet: " 950 " len %d != %zd\n", err, len); 951 if (!zcopy_used) 952 vhost_add_used_and_signal(&net->dev, vq, head, 0); 953 else 954 vhost_zerocopy_signal_used(net, vq); 955 vhost_net_tx_packet(net); 956 } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len))); 957 } 958 959 /* Expects to be always run from workqueue - which acts as 960 * read-size critical section for our kind of RCU. */ 961 static void handle_tx(struct vhost_net *net) 962 { 963 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX]; 964 struct vhost_virtqueue *vq = &nvq->vq; 965 struct socket *sock; 966 967 mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX); 968 sock = vhost_vq_get_backend(vq); 969 if (!sock) 970 goto out; 971 972 if (!vq_meta_prefetch(vq)) 973 goto out; 974 975 vhost_disable_notify(&net->dev, vq); 976 vhost_net_disable_vq(net, vq); 977 978 if (vhost_sock_zcopy(sock)) 979 handle_tx_zerocopy(net, sock); 980 else 981 handle_tx_copy(net, sock); 982 983 out: 984 mutex_unlock(&vq->mutex); 985 } 986 987 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk) 988 { 989 struct sk_buff *head; 990 int len = 0; 991 unsigned long flags; 992 993 if (rvq->rx_ring) 994 return vhost_net_buf_peek(rvq); 995 996 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags); 997 head = skb_peek(&sk->sk_receive_queue); 998 if (likely(head)) { 999 len = head->len; 1000 if (skb_vlan_tag_present(head)) 1001 len += VLAN_HLEN; 1002 } 1003 1004 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags); 1005 return len; 1006 } 1007 1008 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk, 1009 bool *busyloop_intr) 1010 { 1011 struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX]; 1012 struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX]; 1013 struct vhost_virtqueue *rvq = &rnvq->vq; 1014 struct vhost_virtqueue *tvq = &tnvq->vq; 1015 int len = peek_head_len(rnvq, sk); 1016 1017 if (!len && rvq->busyloop_timeout) { 1018 /* Flush batched heads first */ 1019 vhost_net_signal_used(rnvq); 1020 /* Both tx vq and rx socket were polled here */ 1021 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true); 1022 1023 len = peek_head_len(rnvq, sk); 1024 } 1025 1026 return len; 1027 } 1028 1029 /* This is a multi-buffer version of vhost_get_desc, that works if 1030 * vq has read descriptors only. 1031 * @vq - the relevant virtqueue 1032 * @datalen - data length we'll be reading 1033 * @iovcount - returned count of io vectors we fill 1034 * @log - vhost log 1035 * @log_num - log offset 1036 * @quota - headcount quota, 1 for big buffer 1037 * returns number of buffer heads allocated, negative on error 1038 */ 1039 static int get_rx_bufs(struct vhost_virtqueue *vq, 1040 struct vring_used_elem *heads, 1041 int datalen, 1042 unsigned *iovcount, 1043 struct vhost_log *log, 1044 unsigned *log_num, 1045 unsigned int quota) 1046 { 1047 unsigned int out, in; 1048 int seg = 0; 1049 int headcount = 0; 1050 unsigned d; 1051 int r, nlogs = 0; 1052 /* len is always initialized before use since we are always called with 1053 * datalen > 0. 1054 */ 1055 u32 len; 1056 1057 while (datalen > 0 && headcount < quota) { 1058 if (unlikely(seg >= UIO_MAXIOV)) { 1059 r = -ENOBUFS; 1060 goto err; 1061 } 1062 r = vhost_get_vq_desc(vq, vq->iov + seg, 1063 ARRAY_SIZE(vq->iov) - seg, &out, 1064 &in, log, log_num); 1065 if (unlikely(r < 0)) 1066 goto err; 1067 1068 d = r; 1069 if (d == vq->num) { 1070 r = 0; 1071 goto err; 1072 } 1073 if (unlikely(out || in <= 0)) { 1074 vq_err(vq, "unexpected descriptor format for RX: " 1075 "out %d, in %d\n", out, in); 1076 r = -EINVAL; 1077 goto err; 1078 } 1079 if (unlikely(log)) { 1080 nlogs += *log_num; 1081 log += *log_num; 1082 } 1083 heads[headcount].id = cpu_to_vhost32(vq, d); 1084 len = iov_length(vq->iov + seg, in); 1085 heads[headcount].len = cpu_to_vhost32(vq, len); 1086 datalen -= len; 1087 ++headcount; 1088 seg += in; 1089 } 1090 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen); 1091 *iovcount = seg; 1092 if (unlikely(log)) 1093 *log_num = nlogs; 1094 1095 /* Detect overrun */ 1096 if (unlikely(datalen > 0)) { 1097 r = UIO_MAXIOV + 1; 1098 goto err; 1099 } 1100 return headcount; 1101 err: 1102 vhost_discard_vq_desc(vq, headcount); 1103 return r; 1104 } 1105 1106 /* Expects to be always run from workqueue - which acts as 1107 * read-size critical section for our kind of RCU. */ 1108 static void handle_rx(struct vhost_net *net) 1109 { 1110 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX]; 1111 struct vhost_virtqueue *vq = &nvq->vq; 1112 unsigned in, log; 1113 struct vhost_log *vq_log; 1114 struct msghdr msg = { 1115 .msg_name = NULL, 1116 .msg_namelen = 0, 1117 .msg_control = NULL, /* FIXME: get and handle RX aux data. */ 1118 .msg_controllen = 0, 1119 .msg_flags = MSG_DONTWAIT, 1120 }; 1121 struct virtio_net_hdr hdr = { 1122 .flags = 0, 1123 .gso_type = VIRTIO_NET_HDR_GSO_NONE 1124 }; 1125 size_t total_len = 0; 1126 int err, mergeable; 1127 s16 headcount; 1128 size_t vhost_hlen, sock_hlen; 1129 size_t vhost_len, sock_len; 1130 bool busyloop_intr = false; 1131 struct socket *sock; 1132 struct iov_iter fixup; 1133 __virtio16 num_buffers; 1134 int recv_pkts = 0; 1135 1136 mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX); 1137 sock = vhost_vq_get_backend(vq); 1138 if (!sock) 1139 goto out; 1140 1141 if (!vq_meta_prefetch(vq)) 1142 goto out; 1143 1144 vhost_disable_notify(&net->dev, vq); 1145 vhost_net_disable_vq(net, vq); 1146 1147 vhost_hlen = nvq->vhost_hlen; 1148 sock_hlen = nvq->sock_hlen; 1149 1150 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ? 1151 vq->log : NULL; 1152 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF); 1153 1154 do { 1155 sock_len = vhost_net_rx_peek_head_len(net, sock->sk, 1156 &busyloop_intr); 1157 if (!sock_len) 1158 break; 1159 sock_len += sock_hlen; 1160 vhost_len = sock_len + vhost_hlen; 1161 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx, 1162 vhost_len, &in, vq_log, &log, 1163 likely(mergeable) ? UIO_MAXIOV : 1); 1164 /* On error, stop handling until the next kick. */ 1165 if (unlikely(headcount < 0)) 1166 goto out; 1167 /* OK, now we need to know about added descriptors. */ 1168 if (!headcount) { 1169 if (unlikely(busyloop_intr)) { 1170 vhost_poll_queue(&vq->poll); 1171 } else if (unlikely(vhost_enable_notify(&net->dev, vq))) { 1172 /* They have slipped one in as we were 1173 * doing that: check again. */ 1174 vhost_disable_notify(&net->dev, vq); 1175 continue; 1176 } 1177 /* Nothing new? Wait for eventfd to tell us 1178 * they refilled. */ 1179 goto out; 1180 } 1181 busyloop_intr = false; 1182 if (nvq->rx_ring) 1183 msg.msg_control = vhost_net_buf_consume(&nvq->rxq); 1184 /* On overrun, truncate and discard */ 1185 if (unlikely(headcount > UIO_MAXIOV)) { 1186 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1); 1187 err = sock->ops->recvmsg(sock, &msg, 1188 1, MSG_DONTWAIT | MSG_TRUNC); 1189 pr_debug("Discarded rx packet: len %zd\n", sock_len); 1190 continue; 1191 } 1192 /* We don't need to be notified again. */ 1193 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len); 1194 fixup = msg.msg_iter; 1195 if (unlikely((vhost_hlen))) { 1196 /* We will supply the header ourselves 1197 * TODO: support TSO. 1198 */ 1199 iov_iter_advance(&msg.msg_iter, vhost_hlen); 1200 } 1201 err = sock->ops->recvmsg(sock, &msg, 1202 sock_len, MSG_DONTWAIT | MSG_TRUNC); 1203 /* Userspace might have consumed the packet meanwhile: 1204 * it's not supposed to do this usually, but might be hard 1205 * to prevent. Discard data we got (if any) and keep going. */ 1206 if (unlikely(err != sock_len)) { 1207 pr_debug("Discarded rx packet: " 1208 " len %d, expected %zd\n", err, sock_len); 1209 vhost_discard_vq_desc(vq, headcount); 1210 continue; 1211 } 1212 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */ 1213 if (unlikely(vhost_hlen)) { 1214 if (copy_to_iter(&hdr, sizeof(hdr), 1215 &fixup) != sizeof(hdr)) { 1216 vq_err(vq, "Unable to write vnet_hdr " 1217 "at addr %p\n", vq->iov->iov_base); 1218 goto out; 1219 } 1220 } else { 1221 /* Header came from socket; we'll need to patch 1222 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF 1223 */ 1224 iov_iter_advance(&fixup, sizeof(hdr)); 1225 } 1226 /* TODO: Should check and handle checksum. */ 1227 1228 num_buffers = cpu_to_vhost16(vq, headcount); 1229 if (likely(mergeable) && 1230 copy_to_iter(&num_buffers, sizeof num_buffers, 1231 &fixup) != sizeof num_buffers) { 1232 vq_err(vq, "Failed num_buffers write"); 1233 vhost_discard_vq_desc(vq, headcount); 1234 goto out; 1235 } 1236 nvq->done_idx += headcount; 1237 if (nvq->done_idx > VHOST_NET_BATCH) 1238 vhost_net_signal_used(nvq); 1239 if (unlikely(vq_log)) 1240 vhost_log_write(vq, vq_log, log, vhost_len, 1241 vq->iov, in); 1242 total_len += vhost_len; 1243 } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len))); 1244 1245 if (unlikely(busyloop_intr)) 1246 vhost_poll_queue(&vq->poll); 1247 else if (!sock_len) 1248 vhost_net_enable_vq(net, vq); 1249 out: 1250 vhost_net_signal_used(nvq); 1251 mutex_unlock(&vq->mutex); 1252 } 1253 1254 static void handle_tx_kick(struct vhost_work *work) 1255 { 1256 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, 1257 poll.work); 1258 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); 1259 1260 handle_tx(net); 1261 } 1262 1263 static void handle_rx_kick(struct vhost_work *work) 1264 { 1265 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, 1266 poll.work); 1267 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); 1268 1269 handle_rx(net); 1270 } 1271 1272 static void handle_tx_net(struct vhost_work *work) 1273 { 1274 struct vhost_net *net = container_of(work, struct vhost_net, 1275 poll[VHOST_NET_VQ_TX].work); 1276 handle_tx(net); 1277 } 1278 1279 static void handle_rx_net(struct vhost_work *work) 1280 { 1281 struct vhost_net *net = container_of(work, struct vhost_net, 1282 poll[VHOST_NET_VQ_RX].work); 1283 handle_rx(net); 1284 } 1285 1286 static int vhost_net_open(struct inode *inode, struct file *f) 1287 { 1288 struct vhost_net *n; 1289 struct vhost_dev *dev; 1290 struct vhost_virtqueue **vqs; 1291 void **queue; 1292 struct xdp_buff *xdp; 1293 int i; 1294 1295 n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL); 1296 if (!n) 1297 return -ENOMEM; 1298 vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL); 1299 if (!vqs) { 1300 kvfree(n); 1301 return -ENOMEM; 1302 } 1303 1304 queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *), 1305 GFP_KERNEL); 1306 if (!queue) { 1307 kfree(vqs); 1308 kvfree(n); 1309 return -ENOMEM; 1310 } 1311 n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue; 1312 1313 xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL); 1314 if (!xdp) { 1315 kfree(vqs); 1316 kvfree(n); 1317 kfree(queue); 1318 return -ENOMEM; 1319 } 1320 n->vqs[VHOST_NET_VQ_TX].xdp = xdp; 1321 1322 dev = &n->dev; 1323 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq; 1324 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq; 1325 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick; 1326 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick; 1327 for (i = 0; i < VHOST_NET_VQ_MAX; i++) { 1328 n->vqs[i].ubufs = NULL; 1329 n->vqs[i].ubuf_info = NULL; 1330 n->vqs[i].upend_idx = 0; 1331 n->vqs[i].done_idx = 0; 1332 n->vqs[i].batched_xdp = 0; 1333 n->vqs[i].vhost_hlen = 0; 1334 n->vqs[i].sock_hlen = 0; 1335 n->vqs[i].rx_ring = NULL; 1336 vhost_net_buf_init(&n->vqs[i].rxq); 1337 } 1338 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX, 1339 UIO_MAXIOV + VHOST_NET_BATCH, 1340 VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true, 1341 NULL); 1342 1343 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev); 1344 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev); 1345 1346 f->private_data = n; 1347 n->page_frag.page = NULL; 1348 n->refcnt_bias = 0; 1349 1350 return 0; 1351 } 1352 1353 static struct socket *vhost_net_stop_vq(struct vhost_net *n, 1354 struct vhost_virtqueue *vq) 1355 { 1356 struct socket *sock; 1357 struct vhost_net_virtqueue *nvq = 1358 container_of(vq, struct vhost_net_virtqueue, vq); 1359 1360 mutex_lock(&vq->mutex); 1361 sock = vhost_vq_get_backend(vq); 1362 vhost_net_disable_vq(n, vq); 1363 vhost_vq_set_backend(vq, NULL); 1364 vhost_net_buf_unproduce(nvq); 1365 nvq->rx_ring = NULL; 1366 mutex_unlock(&vq->mutex); 1367 return sock; 1368 } 1369 1370 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock, 1371 struct socket **rx_sock) 1372 { 1373 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq); 1374 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq); 1375 } 1376 1377 static void vhost_net_flush_vq(struct vhost_net *n, int index) 1378 { 1379 vhost_poll_flush(n->poll + index); 1380 vhost_poll_flush(&n->vqs[index].vq.poll); 1381 } 1382 1383 static void vhost_net_flush(struct vhost_net *n) 1384 { 1385 vhost_net_flush_vq(n, VHOST_NET_VQ_TX); 1386 vhost_net_flush_vq(n, VHOST_NET_VQ_RX); 1387 if (n->vqs[VHOST_NET_VQ_TX].ubufs) { 1388 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 1389 n->tx_flush = true; 1390 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 1391 /* Wait for all lower device DMAs done. */ 1392 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs); 1393 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 1394 n->tx_flush = false; 1395 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1); 1396 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex); 1397 } 1398 } 1399 1400 static int vhost_net_release(struct inode *inode, struct file *f) 1401 { 1402 struct vhost_net *n = f->private_data; 1403 struct socket *tx_sock; 1404 struct socket *rx_sock; 1405 1406 vhost_net_stop(n, &tx_sock, &rx_sock); 1407 vhost_net_flush(n); 1408 vhost_dev_stop(&n->dev); 1409 vhost_dev_cleanup(&n->dev); 1410 vhost_net_vq_reset(n); 1411 if (tx_sock) 1412 sockfd_put(tx_sock); 1413 if (rx_sock) 1414 sockfd_put(rx_sock); 1415 /* Make sure no callbacks are outstanding */ 1416 synchronize_rcu(); 1417 /* We do an extra flush before freeing memory, 1418 * since jobs can re-queue themselves. */ 1419 vhost_net_flush(n); 1420 kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue); 1421 kfree(n->vqs[VHOST_NET_VQ_TX].xdp); 1422 kfree(n->dev.vqs); 1423 if (n->page_frag.page) 1424 __page_frag_cache_drain(n->page_frag.page, n->refcnt_bias); 1425 kvfree(n); 1426 return 0; 1427 } 1428 1429 static struct socket *get_raw_socket(int fd) 1430 { 1431 int r; 1432 struct socket *sock = sockfd_lookup(fd, &r); 1433 1434 if (!sock) 1435 return ERR_PTR(-ENOTSOCK); 1436 1437 /* Parameter checking */ 1438 if (sock->sk->sk_type != SOCK_RAW) { 1439 r = -ESOCKTNOSUPPORT; 1440 goto err; 1441 } 1442 1443 if (sock->sk->sk_family != AF_PACKET) { 1444 r = -EPFNOSUPPORT; 1445 goto err; 1446 } 1447 return sock; 1448 err: 1449 sockfd_put(sock); 1450 return ERR_PTR(r); 1451 } 1452 1453 static struct ptr_ring *get_tap_ptr_ring(int fd) 1454 { 1455 struct ptr_ring *ring; 1456 struct file *file = fget(fd); 1457 1458 if (!file) 1459 return NULL; 1460 ring = tun_get_tx_ring(file); 1461 if (!IS_ERR(ring)) 1462 goto out; 1463 ring = tap_get_ptr_ring(file); 1464 if (!IS_ERR(ring)) 1465 goto out; 1466 ring = NULL; 1467 out: 1468 fput(file); 1469 return ring; 1470 } 1471 1472 static struct socket *get_tap_socket(int fd) 1473 { 1474 struct file *file = fget(fd); 1475 struct socket *sock; 1476 1477 if (!file) 1478 return ERR_PTR(-EBADF); 1479 sock = tun_get_socket(file); 1480 if (!IS_ERR(sock)) 1481 return sock; 1482 sock = tap_get_socket(file); 1483 if (IS_ERR(sock)) 1484 fput(file); 1485 return sock; 1486 } 1487 1488 static struct socket *get_socket(int fd) 1489 { 1490 struct socket *sock; 1491 1492 /* special case to disable backend */ 1493 if (fd == -1) 1494 return NULL; 1495 sock = get_raw_socket(fd); 1496 if (!IS_ERR(sock)) 1497 return sock; 1498 sock = get_tap_socket(fd); 1499 if (!IS_ERR(sock)) 1500 return sock; 1501 return ERR_PTR(-ENOTSOCK); 1502 } 1503 1504 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd) 1505 { 1506 struct socket *sock, *oldsock; 1507 struct vhost_virtqueue *vq; 1508 struct vhost_net_virtqueue *nvq; 1509 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL; 1510 int r; 1511 1512 mutex_lock(&n->dev.mutex); 1513 r = vhost_dev_check_owner(&n->dev); 1514 if (r) 1515 goto err; 1516 1517 if (index >= VHOST_NET_VQ_MAX) { 1518 r = -ENOBUFS; 1519 goto err; 1520 } 1521 vq = &n->vqs[index].vq; 1522 nvq = &n->vqs[index]; 1523 mutex_lock(&vq->mutex); 1524 1525 /* Verify that ring has been setup correctly. */ 1526 if (!vhost_vq_access_ok(vq)) { 1527 r = -EFAULT; 1528 goto err_vq; 1529 } 1530 sock = get_socket(fd); 1531 if (IS_ERR(sock)) { 1532 r = PTR_ERR(sock); 1533 goto err_vq; 1534 } 1535 1536 /* start polling new socket */ 1537 oldsock = vhost_vq_get_backend(vq); 1538 if (sock != oldsock) { 1539 ubufs = vhost_net_ubuf_alloc(vq, 1540 sock && vhost_sock_zcopy(sock)); 1541 if (IS_ERR(ubufs)) { 1542 r = PTR_ERR(ubufs); 1543 goto err_ubufs; 1544 } 1545 1546 vhost_net_disable_vq(n, vq); 1547 vhost_vq_set_backend(vq, sock); 1548 vhost_net_buf_unproduce(nvq); 1549 r = vhost_vq_init_access(vq); 1550 if (r) 1551 goto err_used; 1552 r = vhost_net_enable_vq(n, vq); 1553 if (r) 1554 goto err_used; 1555 if (index == VHOST_NET_VQ_RX) 1556 nvq->rx_ring = get_tap_ptr_ring(fd); 1557 1558 oldubufs = nvq->ubufs; 1559 nvq->ubufs = ubufs; 1560 1561 n->tx_packets = 0; 1562 n->tx_zcopy_err = 0; 1563 n->tx_flush = false; 1564 } 1565 1566 mutex_unlock(&vq->mutex); 1567 1568 if (oldubufs) { 1569 vhost_net_ubuf_put_wait_and_free(oldubufs); 1570 mutex_lock(&vq->mutex); 1571 vhost_zerocopy_signal_used(n, vq); 1572 mutex_unlock(&vq->mutex); 1573 } 1574 1575 if (oldsock) { 1576 vhost_net_flush_vq(n, index); 1577 sockfd_put(oldsock); 1578 } 1579 1580 mutex_unlock(&n->dev.mutex); 1581 return 0; 1582 1583 err_used: 1584 vhost_vq_set_backend(vq, oldsock); 1585 vhost_net_enable_vq(n, vq); 1586 if (ubufs) 1587 vhost_net_ubuf_put_wait_and_free(ubufs); 1588 err_ubufs: 1589 if (sock) 1590 sockfd_put(sock); 1591 err_vq: 1592 mutex_unlock(&vq->mutex); 1593 err: 1594 mutex_unlock(&n->dev.mutex); 1595 return r; 1596 } 1597 1598 static long vhost_net_reset_owner(struct vhost_net *n) 1599 { 1600 struct socket *tx_sock = NULL; 1601 struct socket *rx_sock = NULL; 1602 long err; 1603 struct vhost_iotlb *umem; 1604 1605 mutex_lock(&n->dev.mutex); 1606 err = vhost_dev_check_owner(&n->dev); 1607 if (err) 1608 goto done; 1609 umem = vhost_dev_reset_owner_prepare(); 1610 if (!umem) { 1611 err = -ENOMEM; 1612 goto done; 1613 } 1614 vhost_net_stop(n, &tx_sock, &rx_sock); 1615 vhost_net_flush(n); 1616 vhost_dev_stop(&n->dev); 1617 vhost_dev_reset_owner(&n->dev, umem); 1618 vhost_net_vq_reset(n); 1619 done: 1620 mutex_unlock(&n->dev.mutex); 1621 if (tx_sock) 1622 sockfd_put(tx_sock); 1623 if (rx_sock) 1624 sockfd_put(rx_sock); 1625 return err; 1626 } 1627 1628 static int vhost_net_set_features(struct vhost_net *n, u64 features) 1629 { 1630 size_t vhost_hlen, sock_hlen, hdr_len; 1631 int i; 1632 1633 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) | 1634 (1ULL << VIRTIO_F_VERSION_1))) ? 1635 sizeof(struct virtio_net_hdr_mrg_rxbuf) : 1636 sizeof(struct virtio_net_hdr); 1637 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) { 1638 /* vhost provides vnet_hdr */ 1639 vhost_hlen = hdr_len; 1640 sock_hlen = 0; 1641 } else { 1642 /* socket provides vnet_hdr */ 1643 vhost_hlen = 0; 1644 sock_hlen = hdr_len; 1645 } 1646 mutex_lock(&n->dev.mutex); 1647 if ((features & (1 << VHOST_F_LOG_ALL)) && 1648 !vhost_log_access_ok(&n->dev)) 1649 goto out_unlock; 1650 1651 if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) { 1652 if (vhost_init_device_iotlb(&n->dev, true)) 1653 goto out_unlock; 1654 } 1655 1656 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { 1657 mutex_lock(&n->vqs[i].vq.mutex); 1658 n->vqs[i].vq.acked_features = features; 1659 n->vqs[i].vhost_hlen = vhost_hlen; 1660 n->vqs[i].sock_hlen = sock_hlen; 1661 mutex_unlock(&n->vqs[i].vq.mutex); 1662 } 1663 mutex_unlock(&n->dev.mutex); 1664 return 0; 1665 1666 out_unlock: 1667 mutex_unlock(&n->dev.mutex); 1668 return -EFAULT; 1669 } 1670 1671 static long vhost_net_set_owner(struct vhost_net *n) 1672 { 1673 int r; 1674 1675 mutex_lock(&n->dev.mutex); 1676 if (vhost_dev_has_owner(&n->dev)) { 1677 r = -EBUSY; 1678 goto out; 1679 } 1680 r = vhost_net_set_ubuf_info(n); 1681 if (r) 1682 goto out; 1683 r = vhost_dev_set_owner(&n->dev); 1684 if (r) 1685 vhost_net_clear_ubuf_info(n); 1686 vhost_net_flush(n); 1687 out: 1688 mutex_unlock(&n->dev.mutex); 1689 return r; 1690 } 1691 1692 static long vhost_net_ioctl(struct file *f, unsigned int ioctl, 1693 unsigned long arg) 1694 { 1695 struct vhost_net *n = f->private_data; 1696 void __user *argp = (void __user *)arg; 1697 u64 __user *featurep = argp; 1698 struct vhost_vring_file backend; 1699 u64 features; 1700 int r; 1701 1702 switch (ioctl) { 1703 case VHOST_NET_SET_BACKEND: 1704 if (copy_from_user(&backend, argp, sizeof backend)) 1705 return -EFAULT; 1706 return vhost_net_set_backend(n, backend.index, backend.fd); 1707 case VHOST_GET_FEATURES: 1708 features = VHOST_NET_FEATURES; 1709 if (copy_to_user(featurep, &features, sizeof features)) 1710 return -EFAULT; 1711 return 0; 1712 case VHOST_SET_FEATURES: 1713 if (copy_from_user(&features, featurep, sizeof features)) 1714 return -EFAULT; 1715 if (features & ~VHOST_NET_FEATURES) 1716 return -EOPNOTSUPP; 1717 return vhost_net_set_features(n, features); 1718 case VHOST_GET_BACKEND_FEATURES: 1719 features = VHOST_NET_BACKEND_FEATURES; 1720 if (copy_to_user(featurep, &features, sizeof(features))) 1721 return -EFAULT; 1722 return 0; 1723 case VHOST_SET_BACKEND_FEATURES: 1724 if (copy_from_user(&features, featurep, sizeof(features))) 1725 return -EFAULT; 1726 if (features & ~VHOST_NET_BACKEND_FEATURES) 1727 return -EOPNOTSUPP; 1728 vhost_set_backend_features(&n->dev, features); 1729 return 0; 1730 case VHOST_RESET_OWNER: 1731 return vhost_net_reset_owner(n); 1732 case VHOST_SET_OWNER: 1733 return vhost_net_set_owner(n); 1734 default: 1735 mutex_lock(&n->dev.mutex); 1736 r = vhost_dev_ioctl(&n->dev, ioctl, argp); 1737 if (r == -ENOIOCTLCMD) 1738 r = vhost_vring_ioctl(&n->dev, ioctl, argp); 1739 else 1740 vhost_net_flush(n); 1741 mutex_unlock(&n->dev.mutex); 1742 return r; 1743 } 1744 } 1745 1746 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to) 1747 { 1748 struct file *file = iocb->ki_filp; 1749 struct vhost_net *n = file->private_data; 1750 struct vhost_dev *dev = &n->dev; 1751 int noblock = file->f_flags & O_NONBLOCK; 1752 1753 return vhost_chr_read_iter(dev, to, noblock); 1754 } 1755 1756 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb, 1757 struct iov_iter *from) 1758 { 1759 struct file *file = iocb->ki_filp; 1760 struct vhost_net *n = file->private_data; 1761 struct vhost_dev *dev = &n->dev; 1762 1763 return vhost_chr_write_iter(dev, from); 1764 } 1765 1766 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait) 1767 { 1768 struct vhost_net *n = file->private_data; 1769 struct vhost_dev *dev = &n->dev; 1770 1771 return vhost_chr_poll(file, dev, wait); 1772 } 1773 1774 static const struct file_operations vhost_net_fops = { 1775 .owner = THIS_MODULE, 1776 .release = vhost_net_release, 1777 .read_iter = vhost_net_chr_read_iter, 1778 .write_iter = vhost_net_chr_write_iter, 1779 .poll = vhost_net_chr_poll, 1780 .unlocked_ioctl = vhost_net_ioctl, 1781 .compat_ioctl = compat_ptr_ioctl, 1782 .open = vhost_net_open, 1783 .llseek = noop_llseek, 1784 }; 1785 1786 static struct miscdevice vhost_net_misc = { 1787 .minor = VHOST_NET_MINOR, 1788 .name = "vhost-net", 1789 .fops = &vhost_net_fops, 1790 }; 1791 1792 static int vhost_net_init(void) 1793 { 1794 if (experimental_zcopytx) 1795 vhost_net_enable_zcopy(VHOST_NET_VQ_TX); 1796 return misc_register(&vhost_net_misc); 1797 } 1798 module_init(vhost_net_init); 1799 1800 static void vhost_net_exit(void) 1801 { 1802 misc_deregister(&vhost_net_misc); 1803 } 1804 module_exit(vhost_net_exit); 1805 1806 MODULE_VERSION("0.0.1"); 1807 MODULE_LICENSE("GPL v2"); 1808 MODULE_AUTHOR("Michael S. Tsirkin"); 1809 MODULE_DESCRIPTION("Host kernel accelerator for virtio net"); 1810 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR); 1811 MODULE_ALIAS("devname:vhost-net"); 1812