Loading...
1/* Copyright (C) 2009 Red Hat, Inc.
2 * Author: Michael S. Tsirkin <mst@redhat.com>
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2.
5 *
6 * virtio-net server in host kernel.
7 */
8
9#include <linux/compat.h>
10#include <linux/eventfd.h>
11#include <linux/vhost.h>
12#include <linux/virtio_net.h>
13#include <linux/miscdevice.h>
14#include <linux/module.h>
15#include <linux/moduleparam.h>
16#include <linux/mutex.h>
17#include <linux/workqueue.h>
18#include <linux/file.h>
19#include <linux/slab.h>
20#include <linux/vmalloc.h>
21
22#include <linux/net.h>
23#include <linux/if_packet.h>
24#include <linux/if_arp.h>
25#include <linux/if_tun.h>
26#include <linux/if_macvlan.h>
27#include <linux/if_vlan.h>
28
29#include <net/sock.h>
30
31#include "vhost.h"
32
33static int experimental_zcopytx = 1;
34module_param(experimental_zcopytx, int, 0444);
35MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
36 " 1 -Enable; 0 - Disable");
37
38/* Max number of bytes transferred before requeueing the job.
39 * Using this limit prevents one virtqueue from starving others. */
40#define VHOST_NET_WEIGHT 0x80000
41
42/* MAX number of TX used buffers for outstanding zerocopy */
43#define VHOST_MAX_PEND 128
44#define VHOST_GOODCOPY_LEN 256
45
46/*
47 * For transmit, used buffer len is unused; we override it to track buffer
48 * status internally; used for zerocopy tx only.
49 */
50/* Lower device DMA failed */
51#define VHOST_DMA_FAILED_LEN ((__force __virtio32)3)
52/* Lower device DMA done */
53#define VHOST_DMA_DONE_LEN ((__force __virtio32)2)
54/* Lower device DMA in progress */
55#define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1)
56/* Buffer unused */
57#define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0)
58
59#define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
60
61enum {
62 VHOST_NET_FEATURES = VHOST_FEATURES |
63 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
64 (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
65 (1ULL << VIRTIO_F_IOMMU_PLATFORM)
66};
67
68enum {
69 VHOST_NET_VQ_RX = 0,
70 VHOST_NET_VQ_TX = 1,
71 VHOST_NET_VQ_MAX = 2,
72};
73
74struct vhost_net_ubuf_ref {
75 /* refcount follows semantics similar to kref:
76 * 0: object is released
77 * 1: no outstanding ubufs
78 * >1: outstanding ubufs
79 */
80 atomic_t refcount;
81 wait_queue_head_t wait;
82 struct vhost_virtqueue *vq;
83};
84
85struct vhost_net_virtqueue {
86 struct vhost_virtqueue vq;
87 size_t vhost_hlen;
88 size_t sock_hlen;
89 /* vhost zerocopy support fields below: */
90 /* last used idx for outstanding DMA zerocopy buffers */
91 int upend_idx;
92 /* first used idx for DMA done zerocopy buffers */
93 int done_idx;
94 /* an array of userspace buffers info */
95 struct ubuf_info *ubuf_info;
96 /* Reference counting for outstanding ubufs.
97 * Protected by vq mutex. Writers must also take device mutex. */
98 struct vhost_net_ubuf_ref *ubufs;
99};
100
101struct vhost_net {
102 struct vhost_dev dev;
103 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
104 struct vhost_poll poll[VHOST_NET_VQ_MAX];
105 /* Number of TX recently submitted.
106 * Protected by tx vq lock. */
107 unsigned tx_packets;
108 /* Number of times zerocopy TX recently failed.
109 * Protected by tx vq lock. */
110 unsigned tx_zcopy_err;
111 /* Flush in progress. Protected by tx vq lock. */
112 bool tx_flush;
113};
114
115static unsigned vhost_net_zcopy_mask __read_mostly;
116
117static void vhost_net_enable_zcopy(int vq)
118{
119 vhost_net_zcopy_mask |= 0x1 << vq;
120}
121
122static struct vhost_net_ubuf_ref *
123vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
124{
125 struct vhost_net_ubuf_ref *ubufs;
126 /* No zero copy backend? Nothing to count. */
127 if (!zcopy)
128 return NULL;
129 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
130 if (!ubufs)
131 return ERR_PTR(-ENOMEM);
132 atomic_set(&ubufs->refcount, 1);
133 init_waitqueue_head(&ubufs->wait);
134 ubufs->vq = vq;
135 return ubufs;
136}
137
138static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
139{
140 int r = atomic_sub_return(1, &ubufs->refcount);
141 if (unlikely(!r))
142 wake_up(&ubufs->wait);
143 return r;
144}
145
146static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
147{
148 vhost_net_ubuf_put(ubufs);
149 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
150}
151
152static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
153{
154 vhost_net_ubuf_put_and_wait(ubufs);
155 kfree(ubufs);
156}
157
158static void vhost_net_clear_ubuf_info(struct vhost_net *n)
159{
160 int i;
161
162 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
163 kfree(n->vqs[i].ubuf_info);
164 n->vqs[i].ubuf_info = NULL;
165 }
166}
167
168static int vhost_net_set_ubuf_info(struct vhost_net *n)
169{
170 bool zcopy;
171 int i;
172
173 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
174 zcopy = vhost_net_zcopy_mask & (0x1 << i);
175 if (!zcopy)
176 continue;
177 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
178 UIO_MAXIOV, GFP_KERNEL);
179 if (!n->vqs[i].ubuf_info)
180 goto err;
181 }
182 return 0;
183
184err:
185 vhost_net_clear_ubuf_info(n);
186 return -ENOMEM;
187}
188
189static void vhost_net_vq_reset(struct vhost_net *n)
190{
191 int i;
192
193 vhost_net_clear_ubuf_info(n);
194
195 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
196 n->vqs[i].done_idx = 0;
197 n->vqs[i].upend_idx = 0;
198 n->vqs[i].ubufs = NULL;
199 n->vqs[i].vhost_hlen = 0;
200 n->vqs[i].sock_hlen = 0;
201 }
202
203}
204
205static void vhost_net_tx_packet(struct vhost_net *net)
206{
207 ++net->tx_packets;
208 if (net->tx_packets < 1024)
209 return;
210 net->tx_packets = 0;
211 net->tx_zcopy_err = 0;
212}
213
214static void vhost_net_tx_err(struct vhost_net *net)
215{
216 ++net->tx_zcopy_err;
217}
218
219static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
220{
221 /* TX flush waits for outstanding DMAs to be done.
222 * Don't start new DMAs.
223 */
224 return !net->tx_flush &&
225 net->tx_packets / 64 >= net->tx_zcopy_err;
226}
227
228static bool vhost_sock_zcopy(struct socket *sock)
229{
230 return unlikely(experimental_zcopytx) &&
231 sock_flag(sock->sk, SOCK_ZEROCOPY);
232}
233
234/* In case of DMA done not in order in lower device driver for some reason.
235 * upend_idx is used to track end of used idx, done_idx is used to track head
236 * of used idx. Once lower device DMA done contiguously, we will signal KVM
237 * guest used idx.
238 */
239static void vhost_zerocopy_signal_used(struct vhost_net *net,
240 struct vhost_virtqueue *vq)
241{
242 struct vhost_net_virtqueue *nvq =
243 container_of(vq, struct vhost_net_virtqueue, vq);
244 int i, add;
245 int j = 0;
246
247 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
248 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
249 vhost_net_tx_err(net);
250 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
251 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
252 ++j;
253 } else
254 break;
255 }
256 while (j) {
257 add = min(UIO_MAXIOV - nvq->done_idx, j);
258 vhost_add_used_and_signal_n(vq->dev, vq,
259 &vq->heads[nvq->done_idx], add);
260 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
261 j -= add;
262 }
263}
264
265static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
266{
267 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
268 struct vhost_virtqueue *vq = ubufs->vq;
269 int cnt;
270
271 rcu_read_lock_bh();
272
273 /* set len to mark this desc buffers done DMA */
274 vq->heads[ubuf->desc].len = success ?
275 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
276 cnt = vhost_net_ubuf_put(ubufs);
277
278 /*
279 * Trigger polling thread if guest stopped submitting new buffers:
280 * in this case, the refcount after decrement will eventually reach 1.
281 * We also trigger polling periodically after each 16 packets
282 * (the value 16 here is more or less arbitrary, it's tuned to trigger
283 * less than 10% of times).
284 */
285 if (cnt <= 1 || !(cnt % 16))
286 vhost_poll_queue(&vq->poll);
287
288 rcu_read_unlock_bh();
289}
290
291static inline unsigned long busy_clock(void)
292{
293 return local_clock() >> 10;
294}
295
296static bool vhost_can_busy_poll(struct vhost_dev *dev,
297 unsigned long endtime)
298{
299 return likely(!need_resched()) &&
300 likely(!time_after(busy_clock(), endtime)) &&
301 likely(!signal_pending(current)) &&
302 !vhost_has_work(dev);
303}
304
305static void vhost_net_disable_vq(struct vhost_net *n,
306 struct vhost_virtqueue *vq)
307{
308 struct vhost_net_virtqueue *nvq =
309 container_of(vq, struct vhost_net_virtqueue, vq);
310 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
311 if (!vq->private_data)
312 return;
313 vhost_poll_stop(poll);
314}
315
316static int vhost_net_enable_vq(struct vhost_net *n,
317 struct vhost_virtqueue *vq)
318{
319 struct vhost_net_virtqueue *nvq =
320 container_of(vq, struct vhost_net_virtqueue, vq);
321 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
322 struct socket *sock;
323
324 sock = vq->private_data;
325 if (!sock)
326 return 0;
327
328 return vhost_poll_start(poll, sock->file);
329}
330
331static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
332 struct vhost_virtqueue *vq,
333 struct iovec iov[], unsigned int iov_size,
334 unsigned int *out_num, unsigned int *in_num)
335{
336 unsigned long uninitialized_var(endtime);
337 int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
338 out_num, in_num, NULL, NULL);
339
340 if (r == vq->num && vq->busyloop_timeout) {
341 preempt_disable();
342 endtime = busy_clock() + vq->busyloop_timeout;
343 while (vhost_can_busy_poll(vq->dev, endtime) &&
344 vhost_vq_avail_empty(vq->dev, vq))
345 cpu_relax();
346 preempt_enable();
347 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
348 out_num, in_num, NULL, NULL);
349 }
350
351 return r;
352}
353
354/* Expects to be always run from workqueue - which acts as
355 * read-size critical section for our kind of RCU. */
356static void handle_tx(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 unsigned out, in;
361 int head;
362 struct msghdr msg = {
363 .msg_name = NULL,
364 .msg_namelen = 0,
365 .msg_control = NULL,
366 .msg_controllen = 0,
367 .msg_flags = MSG_DONTWAIT,
368 };
369 size_t len, total_len = 0;
370 int err;
371 size_t hdr_size;
372 struct socket *sock;
373 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
374 bool zcopy, zcopy_used;
375
376 mutex_lock(&vq->mutex);
377 sock = vq->private_data;
378 if (!sock)
379 goto out;
380
381 if (!vq_iotlb_prefetch(vq))
382 goto out;
383
384 vhost_disable_notify(&net->dev, vq);
385
386 hdr_size = nvq->vhost_hlen;
387 zcopy = nvq->ubufs;
388
389 for (;;) {
390 /* Release DMAs done buffers first */
391 if (zcopy)
392 vhost_zerocopy_signal_used(net, vq);
393
394 /* If more outstanding DMAs, queue the work.
395 * Handle upend_idx wrap around
396 */
397 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
398 % UIO_MAXIOV == nvq->done_idx))
399 break;
400
401 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov,
402 ARRAY_SIZE(vq->iov),
403 &out, &in);
404 /* On error, stop handling until the next kick. */
405 if (unlikely(head < 0))
406 break;
407 /* Nothing new? Wait for eventfd to tell us they refilled. */
408 if (head == vq->num) {
409 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
410 vhost_disable_notify(&net->dev, vq);
411 continue;
412 }
413 break;
414 }
415 if (in) {
416 vq_err(vq, "Unexpected descriptor format for TX: "
417 "out %d, int %d\n", out, in);
418 break;
419 }
420 /* Skip header. TODO: support TSO. */
421 len = iov_length(vq->iov, out);
422 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
423 iov_iter_advance(&msg.msg_iter, hdr_size);
424 /* Sanity check */
425 if (!msg_data_left(&msg)) {
426 vq_err(vq, "Unexpected header len for TX: "
427 "%zd expected %zd\n",
428 len, hdr_size);
429 break;
430 }
431 len = msg_data_left(&msg);
432
433 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
434 && (nvq->upend_idx + 1) % UIO_MAXIOV !=
435 nvq->done_idx
436 && vhost_net_tx_select_zcopy(net);
437
438 /* use msg_control to pass vhost zerocopy ubuf info to skb */
439 if (zcopy_used) {
440 struct ubuf_info *ubuf;
441 ubuf = nvq->ubuf_info + nvq->upend_idx;
442
443 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
444 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
445 ubuf->callback = vhost_zerocopy_callback;
446 ubuf->ctx = nvq->ubufs;
447 ubuf->desc = nvq->upend_idx;
448 msg.msg_control = ubuf;
449 msg.msg_controllen = sizeof(ubuf);
450 ubufs = nvq->ubufs;
451 atomic_inc(&ubufs->refcount);
452 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
453 } else {
454 msg.msg_control = NULL;
455 ubufs = NULL;
456 }
457 /* TODO: Check specific error and bomb out unless ENOBUFS? */
458 err = sock->ops->sendmsg(sock, &msg, len);
459 if (unlikely(err < 0)) {
460 if (zcopy_used) {
461 vhost_net_ubuf_put(ubufs);
462 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
463 % UIO_MAXIOV;
464 }
465 vhost_discard_vq_desc(vq, 1);
466 break;
467 }
468 if (err != len)
469 pr_debug("Truncated TX packet: "
470 " len %d != %zd\n", err, len);
471 if (!zcopy_used)
472 vhost_add_used_and_signal(&net->dev, vq, head, 0);
473 else
474 vhost_zerocopy_signal_used(net, vq);
475 total_len += len;
476 vhost_net_tx_packet(net);
477 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
478 vhost_poll_queue(&vq->poll);
479 break;
480 }
481 }
482out:
483 mutex_unlock(&vq->mutex);
484}
485
486static int peek_head_len(struct sock *sk)
487{
488 struct socket *sock = sk->sk_socket;
489 struct sk_buff *head;
490 int len = 0;
491 unsigned long flags;
492
493 if (sock->ops->peek_len)
494 return sock->ops->peek_len(sock);
495
496 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
497 head = skb_peek(&sk->sk_receive_queue);
498 if (likely(head)) {
499 len = head->len;
500 if (skb_vlan_tag_present(head))
501 len += VLAN_HLEN;
502 }
503
504 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
505 return len;
506}
507
508static int sk_has_rx_data(struct sock *sk)
509{
510 struct socket *sock = sk->sk_socket;
511
512 if (sock->ops->peek_len)
513 return sock->ops->peek_len(sock);
514
515 return skb_queue_empty(&sk->sk_receive_queue);
516}
517
518static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
519{
520 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
521 struct vhost_virtqueue *vq = &nvq->vq;
522 unsigned long uninitialized_var(endtime);
523 int len = peek_head_len(sk);
524
525 if (!len && vq->busyloop_timeout) {
526 /* Both tx vq and rx socket were polled here */
527 mutex_lock(&vq->mutex);
528 vhost_disable_notify(&net->dev, vq);
529
530 preempt_disable();
531 endtime = busy_clock() + vq->busyloop_timeout;
532
533 while (vhost_can_busy_poll(&net->dev, endtime) &&
534 !sk_has_rx_data(sk) &&
535 vhost_vq_avail_empty(&net->dev, vq))
536 cpu_relax();
537
538 preempt_enable();
539
540 if (vhost_enable_notify(&net->dev, vq))
541 vhost_poll_queue(&vq->poll);
542 mutex_unlock(&vq->mutex);
543
544 len = peek_head_len(sk);
545 }
546
547 return len;
548}
549
550/* This is a multi-buffer version of vhost_get_desc, that works if
551 * vq has read descriptors only.
552 * @vq - the relevant virtqueue
553 * @datalen - data length we'll be reading
554 * @iovcount - returned count of io vectors we fill
555 * @log - vhost log
556 * @log_num - log offset
557 * @quota - headcount quota, 1 for big buffer
558 * returns number of buffer heads allocated, negative on error
559 */
560static int get_rx_bufs(struct vhost_virtqueue *vq,
561 struct vring_used_elem *heads,
562 int datalen,
563 unsigned *iovcount,
564 struct vhost_log *log,
565 unsigned *log_num,
566 unsigned int quota)
567{
568 unsigned int out, in;
569 int seg = 0;
570 int headcount = 0;
571 unsigned d;
572 int r, nlogs = 0;
573 /* len is always initialized before use since we are always called with
574 * datalen > 0.
575 */
576 u32 uninitialized_var(len);
577
578 while (datalen > 0 && headcount < quota) {
579 if (unlikely(seg >= UIO_MAXIOV)) {
580 r = -ENOBUFS;
581 goto err;
582 }
583 r = vhost_get_vq_desc(vq, vq->iov + seg,
584 ARRAY_SIZE(vq->iov) - seg, &out,
585 &in, log, log_num);
586 if (unlikely(r < 0))
587 goto err;
588
589 d = r;
590 if (d == vq->num) {
591 r = 0;
592 goto err;
593 }
594 if (unlikely(out || in <= 0)) {
595 vq_err(vq, "unexpected descriptor format for RX: "
596 "out %d, in %d\n", out, in);
597 r = -EINVAL;
598 goto err;
599 }
600 if (unlikely(log)) {
601 nlogs += *log_num;
602 log += *log_num;
603 }
604 heads[headcount].id = cpu_to_vhost32(vq, d);
605 len = iov_length(vq->iov + seg, in);
606 heads[headcount].len = cpu_to_vhost32(vq, len);
607 datalen -= len;
608 ++headcount;
609 seg += in;
610 }
611 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
612 *iovcount = seg;
613 if (unlikely(log))
614 *log_num = nlogs;
615
616 /* Detect overrun */
617 if (unlikely(datalen > 0)) {
618 r = UIO_MAXIOV + 1;
619 goto err;
620 }
621 return headcount;
622err:
623 vhost_discard_vq_desc(vq, headcount);
624 return r;
625}
626
627/* Expects to be always run from workqueue - which acts as
628 * read-size critical section for our kind of RCU. */
629static void handle_rx(struct vhost_net *net)
630{
631 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
632 struct vhost_virtqueue *vq = &nvq->vq;
633 unsigned uninitialized_var(in), log;
634 struct vhost_log *vq_log;
635 struct msghdr msg = {
636 .msg_name = NULL,
637 .msg_namelen = 0,
638 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
639 .msg_controllen = 0,
640 .msg_flags = MSG_DONTWAIT,
641 };
642 struct virtio_net_hdr hdr = {
643 .flags = 0,
644 .gso_type = VIRTIO_NET_HDR_GSO_NONE
645 };
646 size_t total_len = 0;
647 int err, mergeable;
648 s16 headcount;
649 size_t vhost_hlen, sock_hlen;
650 size_t vhost_len, sock_len;
651 struct socket *sock;
652 struct iov_iter fixup;
653 __virtio16 num_buffers;
654
655 mutex_lock(&vq->mutex);
656 sock = vq->private_data;
657 if (!sock)
658 goto out;
659
660 if (!vq_iotlb_prefetch(vq))
661 goto out;
662
663 vhost_disable_notify(&net->dev, vq);
664 vhost_net_disable_vq(net, vq);
665
666 vhost_hlen = nvq->vhost_hlen;
667 sock_hlen = nvq->sock_hlen;
668
669 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
670 vq->log : NULL;
671 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
672
673 while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) {
674 sock_len += sock_hlen;
675 vhost_len = sock_len + vhost_hlen;
676 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
677 &in, vq_log, &log,
678 likely(mergeable) ? UIO_MAXIOV : 1);
679 /* On error, stop handling until the next kick. */
680 if (unlikely(headcount < 0))
681 goto out;
682 /* On overrun, truncate and discard */
683 if (unlikely(headcount > UIO_MAXIOV)) {
684 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
685 err = sock->ops->recvmsg(sock, &msg,
686 1, MSG_DONTWAIT | MSG_TRUNC);
687 pr_debug("Discarded rx packet: len %zd\n", sock_len);
688 continue;
689 }
690 /* OK, now we need to know about added descriptors. */
691 if (!headcount) {
692 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
693 /* They have slipped one in as we were
694 * doing that: check again. */
695 vhost_disable_notify(&net->dev, vq);
696 continue;
697 }
698 /* Nothing new? Wait for eventfd to tell us
699 * they refilled. */
700 goto out;
701 }
702 /* We don't need to be notified again. */
703 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
704 fixup = msg.msg_iter;
705 if (unlikely((vhost_hlen))) {
706 /* We will supply the header ourselves
707 * TODO: support TSO.
708 */
709 iov_iter_advance(&msg.msg_iter, vhost_hlen);
710 }
711 err = sock->ops->recvmsg(sock, &msg,
712 sock_len, MSG_DONTWAIT | MSG_TRUNC);
713 /* Userspace might have consumed the packet meanwhile:
714 * it's not supposed to do this usually, but might be hard
715 * to prevent. Discard data we got (if any) and keep going. */
716 if (unlikely(err != sock_len)) {
717 pr_debug("Discarded rx packet: "
718 " len %d, expected %zd\n", err, sock_len);
719 vhost_discard_vq_desc(vq, headcount);
720 continue;
721 }
722 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
723 if (unlikely(vhost_hlen)) {
724 if (copy_to_iter(&hdr, sizeof(hdr),
725 &fixup) != sizeof(hdr)) {
726 vq_err(vq, "Unable to write vnet_hdr "
727 "at addr %p\n", vq->iov->iov_base);
728 goto out;
729 }
730 } else {
731 /* Header came from socket; we'll need to patch
732 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
733 */
734 iov_iter_advance(&fixup, sizeof(hdr));
735 }
736 /* TODO: Should check and handle checksum. */
737
738 num_buffers = cpu_to_vhost16(vq, headcount);
739 if (likely(mergeable) &&
740 copy_to_iter(&num_buffers, sizeof num_buffers,
741 &fixup) != sizeof num_buffers) {
742 vq_err(vq, "Failed num_buffers write");
743 vhost_discard_vq_desc(vq, headcount);
744 goto out;
745 }
746 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
747 headcount);
748 if (unlikely(vq_log))
749 vhost_log_write(vq, vq_log, log, vhost_len);
750 total_len += vhost_len;
751 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
752 vhost_poll_queue(&vq->poll);
753 goto out;
754 }
755 }
756 vhost_net_enable_vq(net, vq);
757out:
758 mutex_unlock(&vq->mutex);
759}
760
761static void handle_tx_kick(struct vhost_work *work)
762{
763 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
764 poll.work);
765 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
766
767 handle_tx(net);
768}
769
770static void handle_rx_kick(struct vhost_work *work)
771{
772 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
773 poll.work);
774 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
775
776 handle_rx(net);
777}
778
779static void handle_tx_net(struct vhost_work *work)
780{
781 struct vhost_net *net = container_of(work, struct vhost_net,
782 poll[VHOST_NET_VQ_TX].work);
783 handle_tx(net);
784}
785
786static void handle_rx_net(struct vhost_work *work)
787{
788 struct vhost_net *net = container_of(work, struct vhost_net,
789 poll[VHOST_NET_VQ_RX].work);
790 handle_rx(net);
791}
792
793static int vhost_net_open(struct inode *inode, struct file *f)
794{
795 struct vhost_net *n;
796 struct vhost_dev *dev;
797 struct vhost_virtqueue **vqs;
798 int i;
799
800 n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
801 if (!n) {
802 n = vmalloc(sizeof *n);
803 if (!n)
804 return -ENOMEM;
805 }
806 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
807 if (!vqs) {
808 kvfree(n);
809 return -ENOMEM;
810 }
811
812 dev = &n->dev;
813 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
814 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
815 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
816 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
817 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
818 n->vqs[i].ubufs = NULL;
819 n->vqs[i].ubuf_info = NULL;
820 n->vqs[i].upend_idx = 0;
821 n->vqs[i].done_idx = 0;
822 n->vqs[i].vhost_hlen = 0;
823 n->vqs[i].sock_hlen = 0;
824 }
825 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
826
827 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
828 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
829
830 f->private_data = n;
831
832 return 0;
833}
834
835static struct socket *vhost_net_stop_vq(struct vhost_net *n,
836 struct vhost_virtqueue *vq)
837{
838 struct socket *sock;
839
840 mutex_lock(&vq->mutex);
841 sock = vq->private_data;
842 vhost_net_disable_vq(n, vq);
843 vq->private_data = NULL;
844 mutex_unlock(&vq->mutex);
845 return sock;
846}
847
848static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
849 struct socket **rx_sock)
850{
851 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
852 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
853}
854
855static void vhost_net_flush_vq(struct vhost_net *n, int index)
856{
857 vhost_poll_flush(n->poll + index);
858 vhost_poll_flush(&n->vqs[index].vq.poll);
859}
860
861static void vhost_net_flush(struct vhost_net *n)
862{
863 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
864 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
865 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
866 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
867 n->tx_flush = true;
868 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
869 /* Wait for all lower device DMAs done. */
870 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
871 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
872 n->tx_flush = false;
873 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
874 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
875 }
876}
877
878static int vhost_net_release(struct inode *inode, struct file *f)
879{
880 struct vhost_net *n = f->private_data;
881 struct socket *tx_sock;
882 struct socket *rx_sock;
883
884 vhost_net_stop(n, &tx_sock, &rx_sock);
885 vhost_net_flush(n);
886 vhost_dev_stop(&n->dev);
887 vhost_dev_cleanup(&n->dev, false);
888 vhost_net_vq_reset(n);
889 if (tx_sock)
890 sockfd_put(tx_sock);
891 if (rx_sock)
892 sockfd_put(rx_sock);
893 /* Make sure no callbacks are outstanding */
894 synchronize_rcu_bh();
895 /* We do an extra flush before freeing memory,
896 * since jobs can re-queue themselves. */
897 vhost_net_flush(n);
898 kfree(n->dev.vqs);
899 kvfree(n);
900 return 0;
901}
902
903static struct socket *get_raw_socket(int fd)
904{
905 struct {
906 struct sockaddr_ll sa;
907 char buf[MAX_ADDR_LEN];
908 } uaddr;
909 int uaddr_len = sizeof uaddr, r;
910 struct socket *sock = sockfd_lookup(fd, &r);
911
912 if (!sock)
913 return ERR_PTR(-ENOTSOCK);
914
915 /* Parameter checking */
916 if (sock->sk->sk_type != SOCK_RAW) {
917 r = -ESOCKTNOSUPPORT;
918 goto err;
919 }
920
921 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
922 &uaddr_len, 0);
923 if (r)
924 goto err;
925
926 if (uaddr.sa.sll_family != AF_PACKET) {
927 r = -EPFNOSUPPORT;
928 goto err;
929 }
930 return sock;
931err:
932 sockfd_put(sock);
933 return ERR_PTR(r);
934}
935
936static struct socket *get_tap_socket(int fd)
937{
938 struct file *file = fget(fd);
939 struct socket *sock;
940
941 if (!file)
942 return ERR_PTR(-EBADF);
943 sock = tun_get_socket(file);
944 if (!IS_ERR(sock))
945 return sock;
946 sock = macvtap_get_socket(file);
947 if (IS_ERR(sock))
948 fput(file);
949 return sock;
950}
951
952static struct socket *get_socket(int fd)
953{
954 struct socket *sock;
955
956 /* special case to disable backend */
957 if (fd == -1)
958 return NULL;
959 sock = get_raw_socket(fd);
960 if (!IS_ERR(sock))
961 return sock;
962 sock = get_tap_socket(fd);
963 if (!IS_ERR(sock))
964 return sock;
965 return ERR_PTR(-ENOTSOCK);
966}
967
968static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
969{
970 struct socket *sock, *oldsock;
971 struct vhost_virtqueue *vq;
972 struct vhost_net_virtqueue *nvq;
973 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
974 int r;
975
976 mutex_lock(&n->dev.mutex);
977 r = vhost_dev_check_owner(&n->dev);
978 if (r)
979 goto err;
980
981 if (index >= VHOST_NET_VQ_MAX) {
982 r = -ENOBUFS;
983 goto err;
984 }
985 vq = &n->vqs[index].vq;
986 nvq = &n->vqs[index];
987 mutex_lock(&vq->mutex);
988
989 /* Verify that ring has been setup correctly. */
990 if (!vhost_vq_access_ok(vq)) {
991 r = -EFAULT;
992 goto err_vq;
993 }
994 sock = get_socket(fd);
995 if (IS_ERR(sock)) {
996 r = PTR_ERR(sock);
997 goto err_vq;
998 }
999
1000 /* start polling new socket */
1001 oldsock = vq->private_data;
1002 if (sock != oldsock) {
1003 ubufs = vhost_net_ubuf_alloc(vq,
1004 sock && vhost_sock_zcopy(sock));
1005 if (IS_ERR(ubufs)) {
1006 r = PTR_ERR(ubufs);
1007 goto err_ubufs;
1008 }
1009
1010 vhost_net_disable_vq(n, vq);
1011 vq->private_data = sock;
1012 r = vhost_vq_init_access(vq);
1013 if (r)
1014 goto err_used;
1015 r = vhost_net_enable_vq(n, vq);
1016 if (r)
1017 goto err_used;
1018
1019 oldubufs = nvq->ubufs;
1020 nvq->ubufs = ubufs;
1021
1022 n->tx_packets = 0;
1023 n->tx_zcopy_err = 0;
1024 n->tx_flush = false;
1025 }
1026
1027 mutex_unlock(&vq->mutex);
1028
1029 if (oldubufs) {
1030 vhost_net_ubuf_put_wait_and_free(oldubufs);
1031 mutex_lock(&vq->mutex);
1032 vhost_zerocopy_signal_used(n, vq);
1033 mutex_unlock(&vq->mutex);
1034 }
1035
1036 if (oldsock) {
1037 vhost_net_flush_vq(n, index);
1038 sockfd_put(oldsock);
1039 }
1040
1041 mutex_unlock(&n->dev.mutex);
1042 return 0;
1043
1044err_used:
1045 vq->private_data = oldsock;
1046 vhost_net_enable_vq(n, vq);
1047 if (ubufs)
1048 vhost_net_ubuf_put_wait_and_free(ubufs);
1049err_ubufs:
1050 sockfd_put(sock);
1051err_vq:
1052 mutex_unlock(&vq->mutex);
1053err:
1054 mutex_unlock(&n->dev.mutex);
1055 return r;
1056}
1057
1058static long vhost_net_reset_owner(struct vhost_net *n)
1059{
1060 struct socket *tx_sock = NULL;
1061 struct socket *rx_sock = NULL;
1062 long err;
1063 struct vhost_umem *umem;
1064
1065 mutex_lock(&n->dev.mutex);
1066 err = vhost_dev_check_owner(&n->dev);
1067 if (err)
1068 goto done;
1069 umem = vhost_dev_reset_owner_prepare();
1070 if (!umem) {
1071 err = -ENOMEM;
1072 goto done;
1073 }
1074 vhost_net_stop(n, &tx_sock, &rx_sock);
1075 vhost_net_flush(n);
1076 vhost_dev_reset_owner(&n->dev, umem);
1077 vhost_net_vq_reset(n);
1078done:
1079 mutex_unlock(&n->dev.mutex);
1080 if (tx_sock)
1081 sockfd_put(tx_sock);
1082 if (rx_sock)
1083 sockfd_put(rx_sock);
1084 return err;
1085}
1086
1087static int vhost_net_set_features(struct vhost_net *n, u64 features)
1088{
1089 size_t vhost_hlen, sock_hlen, hdr_len;
1090 int i;
1091
1092 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1093 (1ULL << VIRTIO_F_VERSION_1))) ?
1094 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1095 sizeof(struct virtio_net_hdr);
1096 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1097 /* vhost provides vnet_hdr */
1098 vhost_hlen = hdr_len;
1099 sock_hlen = 0;
1100 } else {
1101 /* socket provides vnet_hdr */
1102 vhost_hlen = 0;
1103 sock_hlen = hdr_len;
1104 }
1105 mutex_lock(&n->dev.mutex);
1106 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1107 !vhost_log_access_ok(&n->dev))
1108 goto out_unlock;
1109
1110 if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1111 if (vhost_init_device_iotlb(&n->dev, true))
1112 goto out_unlock;
1113 }
1114
1115 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1116 mutex_lock(&n->vqs[i].vq.mutex);
1117 n->vqs[i].vq.acked_features = features;
1118 n->vqs[i].vhost_hlen = vhost_hlen;
1119 n->vqs[i].sock_hlen = sock_hlen;
1120 mutex_unlock(&n->vqs[i].vq.mutex);
1121 }
1122 mutex_unlock(&n->dev.mutex);
1123 return 0;
1124
1125out_unlock:
1126 mutex_unlock(&n->dev.mutex);
1127 return -EFAULT;
1128}
1129
1130static long vhost_net_set_owner(struct vhost_net *n)
1131{
1132 int r;
1133
1134 mutex_lock(&n->dev.mutex);
1135 if (vhost_dev_has_owner(&n->dev)) {
1136 r = -EBUSY;
1137 goto out;
1138 }
1139 r = vhost_net_set_ubuf_info(n);
1140 if (r)
1141 goto out;
1142 r = vhost_dev_set_owner(&n->dev);
1143 if (r)
1144 vhost_net_clear_ubuf_info(n);
1145 vhost_net_flush(n);
1146out:
1147 mutex_unlock(&n->dev.mutex);
1148 return r;
1149}
1150
1151static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1152 unsigned long arg)
1153{
1154 struct vhost_net *n = f->private_data;
1155 void __user *argp = (void __user *)arg;
1156 u64 __user *featurep = argp;
1157 struct vhost_vring_file backend;
1158 u64 features;
1159 int r;
1160
1161 switch (ioctl) {
1162 case VHOST_NET_SET_BACKEND:
1163 if (copy_from_user(&backend, argp, sizeof backend))
1164 return -EFAULT;
1165 return vhost_net_set_backend(n, backend.index, backend.fd);
1166 case VHOST_GET_FEATURES:
1167 features = VHOST_NET_FEATURES;
1168 if (copy_to_user(featurep, &features, sizeof features))
1169 return -EFAULT;
1170 return 0;
1171 case VHOST_SET_FEATURES:
1172 if (copy_from_user(&features, featurep, sizeof features))
1173 return -EFAULT;
1174 if (features & ~VHOST_NET_FEATURES)
1175 return -EOPNOTSUPP;
1176 return vhost_net_set_features(n, features);
1177 case VHOST_RESET_OWNER:
1178 return vhost_net_reset_owner(n);
1179 case VHOST_SET_OWNER:
1180 return vhost_net_set_owner(n);
1181 default:
1182 mutex_lock(&n->dev.mutex);
1183 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1184 if (r == -ENOIOCTLCMD)
1185 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1186 else
1187 vhost_net_flush(n);
1188 mutex_unlock(&n->dev.mutex);
1189 return r;
1190 }
1191}
1192
1193#ifdef CONFIG_COMPAT
1194static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1195 unsigned long arg)
1196{
1197 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1198}
1199#endif
1200
1201static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1202{
1203 struct file *file = iocb->ki_filp;
1204 struct vhost_net *n = file->private_data;
1205 struct vhost_dev *dev = &n->dev;
1206 int noblock = file->f_flags & O_NONBLOCK;
1207
1208 return vhost_chr_read_iter(dev, to, noblock);
1209}
1210
1211static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1212 struct iov_iter *from)
1213{
1214 struct file *file = iocb->ki_filp;
1215 struct vhost_net *n = file->private_data;
1216 struct vhost_dev *dev = &n->dev;
1217
1218 return vhost_chr_write_iter(dev, from);
1219}
1220
1221static unsigned int vhost_net_chr_poll(struct file *file, poll_table *wait)
1222{
1223 struct vhost_net *n = file->private_data;
1224 struct vhost_dev *dev = &n->dev;
1225
1226 return vhost_chr_poll(file, dev, wait);
1227}
1228
1229static const struct file_operations vhost_net_fops = {
1230 .owner = THIS_MODULE,
1231 .release = vhost_net_release,
1232 .read_iter = vhost_net_chr_read_iter,
1233 .write_iter = vhost_net_chr_write_iter,
1234 .poll = vhost_net_chr_poll,
1235 .unlocked_ioctl = vhost_net_ioctl,
1236#ifdef CONFIG_COMPAT
1237 .compat_ioctl = vhost_net_compat_ioctl,
1238#endif
1239 .open = vhost_net_open,
1240 .llseek = noop_llseek,
1241};
1242
1243static struct miscdevice vhost_net_misc = {
1244 .minor = VHOST_NET_MINOR,
1245 .name = "vhost-net",
1246 .fops = &vhost_net_fops,
1247};
1248
1249static int vhost_net_init(void)
1250{
1251 if (experimental_zcopytx)
1252 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1253 return misc_register(&vhost_net_misc);
1254}
1255module_init(vhost_net_init);
1256
1257static void vhost_net_exit(void)
1258{
1259 misc_deregister(&vhost_net_misc);
1260}
1261module_exit(vhost_net_exit);
1262
1263MODULE_VERSION("0.0.1");
1264MODULE_LICENSE("GPL v2");
1265MODULE_AUTHOR("Michael S. Tsirkin");
1266MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1267MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1268MODULE_ALIAS("devname:vhost-net");
1/* Copyright (C) 2009 Red Hat, Inc.
2 * Author: Michael S. Tsirkin <mst@redhat.com>
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2.
5 *
6 * virtio-net server in host kernel.
7 */
8
9#include <linux/compat.h>
10#include <linux/eventfd.h>
11#include <linux/vhost.h>
12#include <linux/virtio_net.h>
13#include <linux/miscdevice.h>
14#include <linux/module.h>
15#include <linux/moduleparam.h>
16#include <linux/mutex.h>
17#include <linux/workqueue.h>
18#include <linux/file.h>
19#include <linux/slab.h>
20#include <linux/vmalloc.h>
21
22#include <linux/net.h>
23#include <linux/if_packet.h>
24#include <linux/if_arp.h>
25#include <linux/if_tun.h>
26#include <linux/if_macvlan.h>
27#include <linux/if_vlan.h>
28
29#include <net/sock.h>
30
31#include "vhost.h"
32
33static int experimental_zcopytx = 1;
34module_param(experimental_zcopytx, int, 0444);
35MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
36 " 1 -Enable; 0 - Disable");
37
38/* Max number of bytes transferred before requeueing the job.
39 * Using this limit prevents one virtqueue from starving others. */
40#define VHOST_NET_WEIGHT 0x80000
41
42/* MAX number of TX used buffers for outstanding zerocopy */
43#define VHOST_MAX_PEND 128
44#define VHOST_GOODCOPY_LEN 256
45
46/*
47 * For transmit, used buffer len is unused; we override it to track buffer
48 * status internally; used for zerocopy tx only.
49 */
50/* Lower device DMA failed */
51#define VHOST_DMA_FAILED_LEN ((__force __virtio32)3)
52/* Lower device DMA done */
53#define VHOST_DMA_DONE_LEN ((__force __virtio32)2)
54/* Lower device DMA in progress */
55#define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1)
56/* Buffer unused */
57#define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0)
58
59#define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
60
61enum {
62 VHOST_NET_FEATURES = VHOST_FEATURES |
63 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
64 (1ULL << VIRTIO_NET_F_MRG_RXBUF)
65};
66
67enum {
68 VHOST_NET_VQ_RX = 0,
69 VHOST_NET_VQ_TX = 1,
70 VHOST_NET_VQ_MAX = 2,
71};
72
73struct vhost_net_ubuf_ref {
74 /* refcount follows semantics similar to kref:
75 * 0: object is released
76 * 1: no outstanding ubufs
77 * >1: outstanding ubufs
78 */
79 atomic_t refcount;
80 wait_queue_head_t wait;
81 struct vhost_virtqueue *vq;
82};
83
84struct vhost_net_virtqueue {
85 struct vhost_virtqueue vq;
86 size_t vhost_hlen;
87 size_t sock_hlen;
88 /* vhost zerocopy support fields below: */
89 /* last used idx for outstanding DMA zerocopy buffers */
90 int upend_idx;
91 /* first used idx for DMA done zerocopy buffers */
92 int done_idx;
93 /* an array of userspace buffers info */
94 struct ubuf_info *ubuf_info;
95 /* Reference counting for outstanding ubufs.
96 * Protected by vq mutex. Writers must also take device mutex. */
97 struct vhost_net_ubuf_ref *ubufs;
98};
99
100struct vhost_net {
101 struct vhost_dev dev;
102 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
103 struct vhost_poll poll[VHOST_NET_VQ_MAX];
104 /* Number of TX recently submitted.
105 * Protected by tx vq lock. */
106 unsigned tx_packets;
107 /* Number of times zerocopy TX recently failed.
108 * Protected by tx vq lock. */
109 unsigned tx_zcopy_err;
110 /* Flush in progress. Protected by tx vq lock. */
111 bool tx_flush;
112};
113
114static unsigned vhost_net_zcopy_mask __read_mostly;
115
116static void vhost_net_enable_zcopy(int vq)
117{
118 vhost_net_zcopy_mask |= 0x1 << vq;
119}
120
121static struct vhost_net_ubuf_ref *
122vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
123{
124 struct vhost_net_ubuf_ref *ubufs;
125 /* No zero copy backend? Nothing to count. */
126 if (!zcopy)
127 return NULL;
128 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
129 if (!ubufs)
130 return ERR_PTR(-ENOMEM);
131 atomic_set(&ubufs->refcount, 1);
132 init_waitqueue_head(&ubufs->wait);
133 ubufs->vq = vq;
134 return ubufs;
135}
136
137static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
138{
139 int r = atomic_sub_return(1, &ubufs->refcount);
140 if (unlikely(!r))
141 wake_up(&ubufs->wait);
142 return r;
143}
144
145static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
146{
147 vhost_net_ubuf_put(ubufs);
148 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
149}
150
151static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
152{
153 vhost_net_ubuf_put_and_wait(ubufs);
154 kfree(ubufs);
155}
156
157static void vhost_net_clear_ubuf_info(struct vhost_net *n)
158{
159 int i;
160
161 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
162 kfree(n->vqs[i].ubuf_info);
163 n->vqs[i].ubuf_info = NULL;
164 }
165}
166
167static int vhost_net_set_ubuf_info(struct vhost_net *n)
168{
169 bool zcopy;
170 int i;
171
172 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
173 zcopy = vhost_net_zcopy_mask & (0x1 << i);
174 if (!zcopy)
175 continue;
176 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
177 UIO_MAXIOV, GFP_KERNEL);
178 if (!n->vqs[i].ubuf_info)
179 goto err;
180 }
181 return 0;
182
183err:
184 vhost_net_clear_ubuf_info(n);
185 return -ENOMEM;
186}
187
188static void vhost_net_vq_reset(struct vhost_net *n)
189{
190 int i;
191
192 vhost_net_clear_ubuf_info(n);
193
194 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
195 n->vqs[i].done_idx = 0;
196 n->vqs[i].upend_idx = 0;
197 n->vqs[i].ubufs = NULL;
198 n->vqs[i].vhost_hlen = 0;
199 n->vqs[i].sock_hlen = 0;
200 }
201
202}
203
204static void vhost_net_tx_packet(struct vhost_net *net)
205{
206 ++net->tx_packets;
207 if (net->tx_packets < 1024)
208 return;
209 net->tx_packets = 0;
210 net->tx_zcopy_err = 0;
211}
212
213static void vhost_net_tx_err(struct vhost_net *net)
214{
215 ++net->tx_zcopy_err;
216}
217
218static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
219{
220 /* TX flush waits for outstanding DMAs to be done.
221 * Don't start new DMAs.
222 */
223 return !net->tx_flush &&
224 net->tx_packets / 64 >= net->tx_zcopy_err;
225}
226
227static bool vhost_sock_zcopy(struct socket *sock)
228{
229 return unlikely(experimental_zcopytx) &&
230 sock_flag(sock->sk, SOCK_ZEROCOPY);
231}
232
233/* In case of DMA done not in order in lower device driver for some reason.
234 * upend_idx is used to track end of used idx, done_idx is used to track head
235 * of used idx. Once lower device DMA done contiguously, we will signal KVM
236 * guest used idx.
237 */
238static void vhost_zerocopy_signal_used(struct vhost_net *net,
239 struct vhost_virtqueue *vq)
240{
241 struct vhost_net_virtqueue *nvq =
242 container_of(vq, struct vhost_net_virtqueue, vq);
243 int i, add;
244 int j = 0;
245
246 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
247 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
248 vhost_net_tx_err(net);
249 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
250 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
251 ++j;
252 } else
253 break;
254 }
255 while (j) {
256 add = min(UIO_MAXIOV - nvq->done_idx, j);
257 vhost_add_used_and_signal_n(vq->dev, vq,
258 &vq->heads[nvq->done_idx], add);
259 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
260 j -= add;
261 }
262}
263
264static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
265{
266 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
267 struct vhost_virtqueue *vq = ubufs->vq;
268 int cnt;
269
270 rcu_read_lock_bh();
271
272 /* set len to mark this desc buffers done DMA */
273 vq->heads[ubuf->desc].len = success ?
274 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
275 cnt = vhost_net_ubuf_put(ubufs);
276
277 /*
278 * Trigger polling thread if guest stopped submitting new buffers:
279 * in this case, the refcount after decrement will eventually reach 1.
280 * We also trigger polling periodically after each 16 packets
281 * (the value 16 here is more or less arbitrary, it's tuned to trigger
282 * less than 10% of times).
283 */
284 if (cnt <= 1 || !(cnt % 16))
285 vhost_poll_queue(&vq->poll);
286
287 rcu_read_unlock_bh();
288}
289
290static inline unsigned long busy_clock(void)
291{
292 return local_clock() >> 10;
293}
294
295static bool vhost_can_busy_poll(struct vhost_dev *dev,
296 unsigned long endtime)
297{
298 return likely(!need_resched()) &&
299 likely(!time_after(busy_clock(), endtime)) &&
300 likely(!signal_pending(current)) &&
301 !vhost_has_work(dev);
302}
303
304static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
305 struct vhost_virtqueue *vq,
306 struct iovec iov[], unsigned int iov_size,
307 unsigned int *out_num, unsigned int *in_num)
308{
309 unsigned long uninitialized_var(endtime);
310 int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
311 out_num, in_num, NULL, NULL);
312
313 if (r == vq->num && vq->busyloop_timeout) {
314 preempt_disable();
315 endtime = busy_clock() + vq->busyloop_timeout;
316 while (vhost_can_busy_poll(vq->dev, endtime) &&
317 vhost_vq_avail_empty(vq->dev, vq))
318 cpu_relax_lowlatency();
319 preempt_enable();
320 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
321 out_num, in_num, NULL, NULL);
322 }
323
324 return r;
325}
326
327/* Expects to be always run from workqueue - which acts as
328 * read-size critical section for our kind of RCU. */
329static void handle_tx(struct vhost_net *net)
330{
331 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
332 struct vhost_virtqueue *vq = &nvq->vq;
333 unsigned out, in;
334 int head;
335 struct msghdr msg = {
336 .msg_name = NULL,
337 .msg_namelen = 0,
338 .msg_control = NULL,
339 .msg_controllen = 0,
340 .msg_flags = MSG_DONTWAIT,
341 };
342 size_t len, total_len = 0;
343 int err;
344 size_t hdr_size;
345 struct socket *sock;
346 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
347 bool zcopy, zcopy_used;
348
349 mutex_lock(&vq->mutex);
350 sock = vq->private_data;
351 if (!sock)
352 goto out;
353
354 vhost_disable_notify(&net->dev, vq);
355
356 hdr_size = nvq->vhost_hlen;
357 zcopy = nvq->ubufs;
358
359 for (;;) {
360 /* Release DMAs done buffers first */
361 if (zcopy)
362 vhost_zerocopy_signal_used(net, vq);
363
364 /* If more outstanding DMAs, queue the work.
365 * Handle upend_idx wrap around
366 */
367 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
368 % UIO_MAXIOV == nvq->done_idx))
369 break;
370
371 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov,
372 ARRAY_SIZE(vq->iov),
373 &out, &in);
374 /* On error, stop handling until the next kick. */
375 if (unlikely(head < 0))
376 break;
377 /* Nothing new? Wait for eventfd to tell us they refilled. */
378 if (head == vq->num) {
379 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
380 vhost_disable_notify(&net->dev, vq);
381 continue;
382 }
383 break;
384 }
385 if (in) {
386 vq_err(vq, "Unexpected descriptor format for TX: "
387 "out %d, int %d\n", out, in);
388 break;
389 }
390 /* Skip header. TODO: support TSO. */
391 len = iov_length(vq->iov, out);
392 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
393 iov_iter_advance(&msg.msg_iter, hdr_size);
394 /* Sanity check */
395 if (!msg_data_left(&msg)) {
396 vq_err(vq, "Unexpected header len for TX: "
397 "%zd expected %zd\n",
398 len, hdr_size);
399 break;
400 }
401 len = msg_data_left(&msg);
402
403 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
404 && (nvq->upend_idx + 1) % UIO_MAXIOV !=
405 nvq->done_idx
406 && vhost_net_tx_select_zcopy(net);
407
408 /* use msg_control to pass vhost zerocopy ubuf info to skb */
409 if (zcopy_used) {
410 struct ubuf_info *ubuf;
411 ubuf = nvq->ubuf_info + nvq->upend_idx;
412
413 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
414 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
415 ubuf->callback = vhost_zerocopy_callback;
416 ubuf->ctx = nvq->ubufs;
417 ubuf->desc = nvq->upend_idx;
418 msg.msg_control = ubuf;
419 msg.msg_controllen = sizeof(ubuf);
420 ubufs = nvq->ubufs;
421 atomic_inc(&ubufs->refcount);
422 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
423 } else {
424 msg.msg_control = NULL;
425 ubufs = NULL;
426 }
427 /* TODO: Check specific error and bomb out unless ENOBUFS? */
428 err = sock->ops->sendmsg(sock, &msg, len);
429 if (unlikely(err < 0)) {
430 if (zcopy_used) {
431 vhost_net_ubuf_put(ubufs);
432 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
433 % UIO_MAXIOV;
434 }
435 vhost_discard_vq_desc(vq, 1);
436 break;
437 }
438 if (err != len)
439 pr_debug("Truncated TX packet: "
440 " len %d != %zd\n", err, len);
441 if (!zcopy_used)
442 vhost_add_used_and_signal(&net->dev, vq, head, 0);
443 else
444 vhost_zerocopy_signal_used(net, vq);
445 total_len += len;
446 vhost_net_tx_packet(net);
447 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
448 vhost_poll_queue(&vq->poll);
449 break;
450 }
451 }
452out:
453 mutex_unlock(&vq->mutex);
454}
455
456static int peek_head_len(struct sock *sk)
457{
458 struct sk_buff *head;
459 int len = 0;
460 unsigned long flags;
461
462 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
463 head = skb_peek(&sk->sk_receive_queue);
464 if (likely(head)) {
465 len = head->len;
466 if (skb_vlan_tag_present(head))
467 len += VLAN_HLEN;
468 }
469
470 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
471 return len;
472}
473
474static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
475{
476 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
477 struct vhost_virtqueue *vq = &nvq->vq;
478 unsigned long uninitialized_var(endtime);
479 int len = peek_head_len(sk);
480
481 if (!len && vq->busyloop_timeout) {
482 /* Both tx vq and rx socket were polled here */
483 mutex_lock(&vq->mutex);
484 vhost_disable_notify(&net->dev, vq);
485
486 preempt_disable();
487 endtime = busy_clock() + vq->busyloop_timeout;
488
489 while (vhost_can_busy_poll(&net->dev, endtime) &&
490 skb_queue_empty(&sk->sk_receive_queue) &&
491 vhost_vq_avail_empty(&net->dev, vq))
492 cpu_relax_lowlatency();
493
494 preempt_enable();
495
496 if (vhost_enable_notify(&net->dev, vq))
497 vhost_poll_queue(&vq->poll);
498 mutex_unlock(&vq->mutex);
499
500 len = peek_head_len(sk);
501 }
502
503 return len;
504}
505
506/* This is a multi-buffer version of vhost_get_desc, that works if
507 * vq has read descriptors only.
508 * @vq - the relevant virtqueue
509 * @datalen - data length we'll be reading
510 * @iovcount - returned count of io vectors we fill
511 * @log - vhost log
512 * @log_num - log offset
513 * @quota - headcount quota, 1 for big buffer
514 * returns number of buffer heads allocated, negative on error
515 */
516static int get_rx_bufs(struct vhost_virtqueue *vq,
517 struct vring_used_elem *heads,
518 int datalen,
519 unsigned *iovcount,
520 struct vhost_log *log,
521 unsigned *log_num,
522 unsigned int quota)
523{
524 unsigned int out, in;
525 int seg = 0;
526 int headcount = 0;
527 unsigned d;
528 int r, nlogs = 0;
529 /* len is always initialized before use since we are always called with
530 * datalen > 0.
531 */
532 u32 uninitialized_var(len);
533
534 while (datalen > 0 && headcount < quota) {
535 if (unlikely(seg >= UIO_MAXIOV)) {
536 r = -ENOBUFS;
537 goto err;
538 }
539 r = vhost_get_vq_desc(vq, vq->iov + seg,
540 ARRAY_SIZE(vq->iov) - seg, &out,
541 &in, log, log_num);
542 if (unlikely(r < 0))
543 goto err;
544
545 d = r;
546 if (d == vq->num) {
547 r = 0;
548 goto err;
549 }
550 if (unlikely(out || in <= 0)) {
551 vq_err(vq, "unexpected descriptor format for RX: "
552 "out %d, in %d\n", out, in);
553 r = -EINVAL;
554 goto err;
555 }
556 if (unlikely(log)) {
557 nlogs += *log_num;
558 log += *log_num;
559 }
560 heads[headcount].id = cpu_to_vhost32(vq, d);
561 len = iov_length(vq->iov + seg, in);
562 heads[headcount].len = cpu_to_vhost32(vq, len);
563 datalen -= len;
564 ++headcount;
565 seg += in;
566 }
567 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
568 *iovcount = seg;
569 if (unlikely(log))
570 *log_num = nlogs;
571
572 /* Detect overrun */
573 if (unlikely(datalen > 0)) {
574 r = UIO_MAXIOV + 1;
575 goto err;
576 }
577 return headcount;
578err:
579 vhost_discard_vq_desc(vq, headcount);
580 return r;
581}
582
583/* Expects to be always run from workqueue - which acts as
584 * read-size critical section for our kind of RCU. */
585static void handle_rx(struct vhost_net *net)
586{
587 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
588 struct vhost_virtqueue *vq = &nvq->vq;
589 unsigned uninitialized_var(in), log;
590 struct vhost_log *vq_log;
591 struct msghdr msg = {
592 .msg_name = NULL,
593 .msg_namelen = 0,
594 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
595 .msg_controllen = 0,
596 .msg_flags = MSG_DONTWAIT,
597 };
598 struct virtio_net_hdr hdr = {
599 .flags = 0,
600 .gso_type = VIRTIO_NET_HDR_GSO_NONE
601 };
602 size_t total_len = 0;
603 int err, mergeable;
604 s16 headcount;
605 size_t vhost_hlen, sock_hlen;
606 size_t vhost_len, sock_len;
607 struct socket *sock;
608 struct iov_iter fixup;
609 __virtio16 num_buffers;
610
611 mutex_lock(&vq->mutex);
612 sock = vq->private_data;
613 if (!sock)
614 goto out;
615 vhost_disable_notify(&net->dev, vq);
616
617 vhost_hlen = nvq->vhost_hlen;
618 sock_hlen = nvq->sock_hlen;
619
620 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
621 vq->log : NULL;
622 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
623
624 while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) {
625 sock_len += sock_hlen;
626 vhost_len = sock_len + vhost_hlen;
627 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
628 &in, vq_log, &log,
629 likely(mergeable) ? UIO_MAXIOV : 1);
630 /* On error, stop handling until the next kick. */
631 if (unlikely(headcount < 0))
632 break;
633 /* On overrun, truncate and discard */
634 if (unlikely(headcount > UIO_MAXIOV)) {
635 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
636 err = sock->ops->recvmsg(sock, &msg,
637 1, MSG_DONTWAIT | MSG_TRUNC);
638 pr_debug("Discarded rx packet: len %zd\n", sock_len);
639 continue;
640 }
641 /* OK, now we need to know about added descriptors. */
642 if (!headcount) {
643 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
644 /* They have slipped one in as we were
645 * doing that: check again. */
646 vhost_disable_notify(&net->dev, vq);
647 continue;
648 }
649 /* Nothing new? Wait for eventfd to tell us
650 * they refilled. */
651 break;
652 }
653 /* We don't need to be notified again. */
654 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
655 fixup = msg.msg_iter;
656 if (unlikely((vhost_hlen))) {
657 /* We will supply the header ourselves
658 * TODO: support TSO.
659 */
660 iov_iter_advance(&msg.msg_iter, vhost_hlen);
661 }
662 err = sock->ops->recvmsg(sock, &msg,
663 sock_len, MSG_DONTWAIT | MSG_TRUNC);
664 /* Userspace might have consumed the packet meanwhile:
665 * it's not supposed to do this usually, but might be hard
666 * to prevent. Discard data we got (if any) and keep going. */
667 if (unlikely(err != sock_len)) {
668 pr_debug("Discarded rx packet: "
669 " len %d, expected %zd\n", err, sock_len);
670 vhost_discard_vq_desc(vq, headcount);
671 continue;
672 }
673 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
674 if (unlikely(vhost_hlen)) {
675 if (copy_to_iter(&hdr, sizeof(hdr),
676 &fixup) != sizeof(hdr)) {
677 vq_err(vq, "Unable to write vnet_hdr "
678 "at addr %p\n", vq->iov->iov_base);
679 break;
680 }
681 } else {
682 /* Header came from socket; we'll need to patch
683 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
684 */
685 iov_iter_advance(&fixup, sizeof(hdr));
686 }
687 /* TODO: Should check and handle checksum. */
688
689 num_buffers = cpu_to_vhost16(vq, headcount);
690 if (likely(mergeable) &&
691 copy_to_iter(&num_buffers, sizeof num_buffers,
692 &fixup) != sizeof num_buffers) {
693 vq_err(vq, "Failed num_buffers write");
694 vhost_discard_vq_desc(vq, headcount);
695 break;
696 }
697 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
698 headcount);
699 if (unlikely(vq_log))
700 vhost_log_write(vq, vq_log, log, vhost_len);
701 total_len += vhost_len;
702 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
703 vhost_poll_queue(&vq->poll);
704 break;
705 }
706 }
707out:
708 mutex_unlock(&vq->mutex);
709}
710
711static void handle_tx_kick(struct vhost_work *work)
712{
713 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
714 poll.work);
715 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
716
717 handle_tx(net);
718}
719
720static void handle_rx_kick(struct vhost_work *work)
721{
722 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
723 poll.work);
724 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
725
726 handle_rx(net);
727}
728
729static void handle_tx_net(struct vhost_work *work)
730{
731 struct vhost_net *net = container_of(work, struct vhost_net,
732 poll[VHOST_NET_VQ_TX].work);
733 handle_tx(net);
734}
735
736static void handle_rx_net(struct vhost_work *work)
737{
738 struct vhost_net *net = container_of(work, struct vhost_net,
739 poll[VHOST_NET_VQ_RX].work);
740 handle_rx(net);
741}
742
743static int vhost_net_open(struct inode *inode, struct file *f)
744{
745 struct vhost_net *n;
746 struct vhost_dev *dev;
747 struct vhost_virtqueue **vqs;
748 int i;
749
750 n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
751 if (!n) {
752 n = vmalloc(sizeof *n);
753 if (!n)
754 return -ENOMEM;
755 }
756 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
757 if (!vqs) {
758 kvfree(n);
759 return -ENOMEM;
760 }
761
762 dev = &n->dev;
763 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
764 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
765 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
766 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
767 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
768 n->vqs[i].ubufs = NULL;
769 n->vqs[i].ubuf_info = NULL;
770 n->vqs[i].upend_idx = 0;
771 n->vqs[i].done_idx = 0;
772 n->vqs[i].vhost_hlen = 0;
773 n->vqs[i].sock_hlen = 0;
774 }
775 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
776
777 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
778 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
779
780 f->private_data = n;
781
782 return 0;
783}
784
785static void vhost_net_disable_vq(struct vhost_net *n,
786 struct vhost_virtqueue *vq)
787{
788 struct vhost_net_virtqueue *nvq =
789 container_of(vq, struct vhost_net_virtqueue, vq);
790 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
791 if (!vq->private_data)
792 return;
793 vhost_poll_stop(poll);
794}
795
796static int vhost_net_enable_vq(struct vhost_net *n,
797 struct vhost_virtqueue *vq)
798{
799 struct vhost_net_virtqueue *nvq =
800 container_of(vq, struct vhost_net_virtqueue, vq);
801 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
802 struct socket *sock;
803
804 sock = vq->private_data;
805 if (!sock)
806 return 0;
807
808 return vhost_poll_start(poll, sock->file);
809}
810
811static struct socket *vhost_net_stop_vq(struct vhost_net *n,
812 struct vhost_virtqueue *vq)
813{
814 struct socket *sock;
815
816 mutex_lock(&vq->mutex);
817 sock = vq->private_data;
818 vhost_net_disable_vq(n, vq);
819 vq->private_data = NULL;
820 mutex_unlock(&vq->mutex);
821 return sock;
822}
823
824static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
825 struct socket **rx_sock)
826{
827 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
828 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
829}
830
831static void vhost_net_flush_vq(struct vhost_net *n, int index)
832{
833 vhost_poll_flush(n->poll + index);
834 vhost_poll_flush(&n->vqs[index].vq.poll);
835}
836
837static void vhost_net_flush(struct vhost_net *n)
838{
839 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
840 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
841 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
842 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
843 n->tx_flush = true;
844 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
845 /* Wait for all lower device DMAs done. */
846 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
847 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
848 n->tx_flush = false;
849 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
850 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
851 }
852}
853
854static int vhost_net_release(struct inode *inode, struct file *f)
855{
856 struct vhost_net *n = f->private_data;
857 struct socket *tx_sock;
858 struct socket *rx_sock;
859
860 vhost_net_stop(n, &tx_sock, &rx_sock);
861 vhost_net_flush(n);
862 vhost_dev_stop(&n->dev);
863 vhost_dev_cleanup(&n->dev, false);
864 vhost_net_vq_reset(n);
865 if (tx_sock)
866 sockfd_put(tx_sock);
867 if (rx_sock)
868 sockfd_put(rx_sock);
869 /* Make sure no callbacks are outstanding */
870 synchronize_rcu_bh();
871 /* We do an extra flush before freeing memory,
872 * since jobs can re-queue themselves. */
873 vhost_net_flush(n);
874 kfree(n->dev.vqs);
875 kvfree(n);
876 return 0;
877}
878
879static struct socket *get_raw_socket(int fd)
880{
881 struct {
882 struct sockaddr_ll sa;
883 char buf[MAX_ADDR_LEN];
884 } uaddr;
885 int uaddr_len = sizeof uaddr, r;
886 struct socket *sock = sockfd_lookup(fd, &r);
887
888 if (!sock)
889 return ERR_PTR(-ENOTSOCK);
890
891 /* Parameter checking */
892 if (sock->sk->sk_type != SOCK_RAW) {
893 r = -ESOCKTNOSUPPORT;
894 goto err;
895 }
896
897 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
898 &uaddr_len, 0);
899 if (r)
900 goto err;
901
902 if (uaddr.sa.sll_family != AF_PACKET) {
903 r = -EPFNOSUPPORT;
904 goto err;
905 }
906 return sock;
907err:
908 sockfd_put(sock);
909 return ERR_PTR(r);
910}
911
912static struct socket *get_tap_socket(int fd)
913{
914 struct file *file = fget(fd);
915 struct socket *sock;
916
917 if (!file)
918 return ERR_PTR(-EBADF);
919 sock = tun_get_socket(file);
920 if (!IS_ERR(sock))
921 return sock;
922 sock = macvtap_get_socket(file);
923 if (IS_ERR(sock))
924 fput(file);
925 return sock;
926}
927
928static struct socket *get_socket(int fd)
929{
930 struct socket *sock;
931
932 /* special case to disable backend */
933 if (fd == -1)
934 return NULL;
935 sock = get_raw_socket(fd);
936 if (!IS_ERR(sock))
937 return sock;
938 sock = get_tap_socket(fd);
939 if (!IS_ERR(sock))
940 return sock;
941 return ERR_PTR(-ENOTSOCK);
942}
943
944static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
945{
946 struct socket *sock, *oldsock;
947 struct vhost_virtqueue *vq;
948 struct vhost_net_virtqueue *nvq;
949 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
950 int r;
951
952 mutex_lock(&n->dev.mutex);
953 r = vhost_dev_check_owner(&n->dev);
954 if (r)
955 goto err;
956
957 if (index >= VHOST_NET_VQ_MAX) {
958 r = -ENOBUFS;
959 goto err;
960 }
961 vq = &n->vqs[index].vq;
962 nvq = &n->vqs[index];
963 mutex_lock(&vq->mutex);
964
965 /* Verify that ring has been setup correctly. */
966 if (!vhost_vq_access_ok(vq)) {
967 r = -EFAULT;
968 goto err_vq;
969 }
970 sock = get_socket(fd);
971 if (IS_ERR(sock)) {
972 r = PTR_ERR(sock);
973 goto err_vq;
974 }
975
976 /* start polling new socket */
977 oldsock = vq->private_data;
978 if (sock != oldsock) {
979 ubufs = vhost_net_ubuf_alloc(vq,
980 sock && vhost_sock_zcopy(sock));
981 if (IS_ERR(ubufs)) {
982 r = PTR_ERR(ubufs);
983 goto err_ubufs;
984 }
985
986 vhost_net_disable_vq(n, vq);
987 vq->private_data = sock;
988 r = vhost_vq_init_access(vq);
989 if (r)
990 goto err_used;
991 r = vhost_net_enable_vq(n, vq);
992 if (r)
993 goto err_used;
994
995 oldubufs = nvq->ubufs;
996 nvq->ubufs = ubufs;
997
998 n->tx_packets = 0;
999 n->tx_zcopy_err = 0;
1000 n->tx_flush = false;
1001 }
1002
1003 mutex_unlock(&vq->mutex);
1004
1005 if (oldubufs) {
1006 vhost_net_ubuf_put_wait_and_free(oldubufs);
1007 mutex_lock(&vq->mutex);
1008 vhost_zerocopy_signal_used(n, vq);
1009 mutex_unlock(&vq->mutex);
1010 }
1011
1012 if (oldsock) {
1013 vhost_net_flush_vq(n, index);
1014 sockfd_put(oldsock);
1015 }
1016
1017 mutex_unlock(&n->dev.mutex);
1018 return 0;
1019
1020err_used:
1021 vq->private_data = oldsock;
1022 vhost_net_enable_vq(n, vq);
1023 if (ubufs)
1024 vhost_net_ubuf_put_wait_and_free(ubufs);
1025err_ubufs:
1026 sockfd_put(sock);
1027err_vq:
1028 mutex_unlock(&vq->mutex);
1029err:
1030 mutex_unlock(&n->dev.mutex);
1031 return r;
1032}
1033
1034static long vhost_net_reset_owner(struct vhost_net *n)
1035{
1036 struct socket *tx_sock = NULL;
1037 struct socket *rx_sock = NULL;
1038 long err;
1039 struct vhost_memory *memory;
1040
1041 mutex_lock(&n->dev.mutex);
1042 err = vhost_dev_check_owner(&n->dev);
1043 if (err)
1044 goto done;
1045 memory = vhost_dev_reset_owner_prepare();
1046 if (!memory) {
1047 err = -ENOMEM;
1048 goto done;
1049 }
1050 vhost_net_stop(n, &tx_sock, &rx_sock);
1051 vhost_net_flush(n);
1052 vhost_dev_reset_owner(&n->dev, memory);
1053 vhost_net_vq_reset(n);
1054done:
1055 mutex_unlock(&n->dev.mutex);
1056 if (tx_sock)
1057 sockfd_put(tx_sock);
1058 if (rx_sock)
1059 sockfd_put(rx_sock);
1060 return err;
1061}
1062
1063static int vhost_net_set_features(struct vhost_net *n, u64 features)
1064{
1065 size_t vhost_hlen, sock_hlen, hdr_len;
1066 int i;
1067
1068 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1069 (1ULL << VIRTIO_F_VERSION_1))) ?
1070 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1071 sizeof(struct virtio_net_hdr);
1072 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1073 /* vhost provides vnet_hdr */
1074 vhost_hlen = hdr_len;
1075 sock_hlen = 0;
1076 } else {
1077 /* socket provides vnet_hdr */
1078 vhost_hlen = 0;
1079 sock_hlen = hdr_len;
1080 }
1081 mutex_lock(&n->dev.mutex);
1082 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1083 !vhost_log_access_ok(&n->dev)) {
1084 mutex_unlock(&n->dev.mutex);
1085 return -EFAULT;
1086 }
1087 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1088 mutex_lock(&n->vqs[i].vq.mutex);
1089 n->vqs[i].vq.acked_features = features;
1090 n->vqs[i].vhost_hlen = vhost_hlen;
1091 n->vqs[i].sock_hlen = sock_hlen;
1092 mutex_unlock(&n->vqs[i].vq.mutex);
1093 }
1094 mutex_unlock(&n->dev.mutex);
1095 return 0;
1096}
1097
1098static long vhost_net_set_owner(struct vhost_net *n)
1099{
1100 int r;
1101
1102 mutex_lock(&n->dev.mutex);
1103 if (vhost_dev_has_owner(&n->dev)) {
1104 r = -EBUSY;
1105 goto out;
1106 }
1107 r = vhost_net_set_ubuf_info(n);
1108 if (r)
1109 goto out;
1110 r = vhost_dev_set_owner(&n->dev);
1111 if (r)
1112 vhost_net_clear_ubuf_info(n);
1113 vhost_net_flush(n);
1114out:
1115 mutex_unlock(&n->dev.mutex);
1116 return r;
1117}
1118
1119static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1120 unsigned long arg)
1121{
1122 struct vhost_net *n = f->private_data;
1123 void __user *argp = (void __user *)arg;
1124 u64 __user *featurep = argp;
1125 struct vhost_vring_file backend;
1126 u64 features;
1127 int r;
1128
1129 switch (ioctl) {
1130 case VHOST_NET_SET_BACKEND:
1131 if (copy_from_user(&backend, argp, sizeof backend))
1132 return -EFAULT;
1133 return vhost_net_set_backend(n, backend.index, backend.fd);
1134 case VHOST_GET_FEATURES:
1135 features = VHOST_NET_FEATURES;
1136 if (copy_to_user(featurep, &features, sizeof features))
1137 return -EFAULT;
1138 return 0;
1139 case VHOST_SET_FEATURES:
1140 if (copy_from_user(&features, featurep, sizeof features))
1141 return -EFAULT;
1142 if (features & ~VHOST_NET_FEATURES)
1143 return -EOPNOTSUPP;
1144 return vhost_net_set_features(n, features);
1145 case VHOST_RESET_OWNER:
1146 return vhost_net_reset_owner(n);
1147 case VHOST_SET_OWNER:
1148 return vhost_net_set_owner(n);
1149 default:
1150 mutex_lock(&n->dev.mutex);
1151 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1152 if (r == -ENOIOCTLCMD)
1153 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1154 else
1155 vhost_net_flush(n);
1156 mutex_unlock(&n->dev.mutex);
1157 return r;
1158 }
1159}
1160
1161#ifdef CONFIG_COMPAT
1162static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1163 unsigned long arg)
1164{
1165 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1166}
1167#endif
1168
1169static const struct file_operations vhost_net_fops = {
1170 .owner = THIS_MODULE,
1171 .release = vhost_net_release,
1172 .unlocked_ioctl = vhost_net_ioctl,
1173#ifdef CONFIG_COMPAT
1174 .compat_ioctl = vhost_net_compat_ioctl,
1175#endif
1176 .open = vhost_net_open,
1177 .llseek = noop_llseek,
1178};
1179
1180static struct miscdevice vhost_net_misc = {
1181 .minor = VHOST_NET_MINOR,
1182 .name = "vhost-net",
1183 .fops = &vhost_net_fops,
1184};
1185
1186static int vhost_net_init(void)
1187{
1188 if (experimental_zcopytx)
1189 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1190 return misc_register(&vhost_net_misc);
1191}
1192module_init(vhost_net_init);
1193
1194static void vhost_net_exit(void)
1195{
1196 misc_deregister(&vhost_net_misc);
1197}
1198module_exit(vhost_net_exit);
1199
1200MODULE_VERSION("0.0.1");
1201MODULE_LICENSE("GPL v2");
1202MODULE_AUTHOR("Michael S. Tsirkin");
1203MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1204MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1205MODULE_ALIAS("devname:vhost-net");