xref: /linux/drivers/vhost/net.c (revision f6e0a4984c2e7244689ea87b62b433bed9d07e94)
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 			 (1ULL << VIRTIO_F_RING_RESET)
78 };
79 
80 enum {
81 	VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
82 };
83 
84 enum {
85 	VHOST_NET_VQ_RX = 0,
86 	VHOST_NET_VQ_TX = 1,
87 	VHOST_NET_VQ_MAX = 2,
88 };
89 
90 struct vhost_net_ubuf_ref {
91 	/* refcount follows semantics similar to kref:
92 	 *  0: object is released
93 	 *  1: no outstanding ubufs
94 	 * >1: outstanding ubufs
95 	 */
96 	atomic_t refcount;
97 	wait_queue_head_t wait;
98 	struct vhost_virtqueue *vq;
99 };
100 
101 #define VHOST_NET_BATCH 64
102 struct vhost_net_buf {
103 	void **queue;
104 	int tail;
105 	int head;
106 };
107 
108 struct vhost_net_virtqueue {
109 	struct vhost_virtqueue vq;
110 	size_t vhost_hlen;
111 	size_t sock_hlen;
112 	/* vhost zerocopy support fields below: */
113 	/* last used idx for outstanding DMA zerocopy buffers */
114 	int upend_idx;
115 	/* For TX, first used idx for DMA done zerocopy buffers
116 	 * For RX, number of batched heads
117 	 */
118 	int done_idx;
119 	/* Number of XDP frames batched */
120 	int batched_xdp;
121 	/* an array of userspace buffers info */
122 	struct ubuf_info_msgzc *ubuf_info;
123 	/* Reference counting for outstanding ubufs.
124 	 * Protected by vq mutex. Writers must also take device mutex. */
125 	struct vhost_net_ubuf_ref *ubufs;
126 	struct ptr_ring *rx_ring;
127 	struct vhost_net_buf rxq;
128 	/* Batched XDP buffs */
129 	struct xdp_buff *xdp;
130 };
131 
132 struct vhost_net {
133 	struct vhost_dev dev;
134 	struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
135 	struct vhost_poll poll[VHOST_NET_VQ_MAX];
136 	/* Number of TX recently submitted.
137 	 * Protected by tx vq lock. */
138 	unsigned tx_packets;
139 	/* Number of times zerocopy TX recently failed.
140 	 * Protected by tx vq lock. */
141 	unsigned tx_zcopy_err;
142 	/* Flush in progress. Protected by tx vq lock. */
143 	bool tx_flush;
144 	/* Private page frag cache */
145 	struct page_frag_cache pf_cache;
146 };
147 
148 static unsigned vhost_net_zcopy_mask __read_mostly;
149 
150 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
151 {
152 	if (rxq->tail != rxq->head)
153 		return rxq->queue[rxq->head];
154 	else
155 		return NULL;
156 }
157 
158 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
159 {
160 	return rxq->tail - rxq->head;
161 }
162 
163 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
164 {
165 	return rxq->tail == rxq->head;
166 }
167 
168 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
169 {
170 	void *ret = vhost_net_buf_get_ptr(rxq);
171 	++rxq->head;
172 	return ret;
173 }
174 
175 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
176 {
177 	struct vhost_net_buf *rxq = &nvq->rxq;
178 
179 	rxq->head = 0;
180 	rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
181 					      VHOST_NET_BATCH);
182 	return rxq->tail;
183 }
184 
185 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
186 {
187 	struct vhost_net_buf *rxq = &nvq->rxq;
188 
189 	if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
190 		ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
191 				   vhost_net_buf_get_size(rxq),
192 				   tun_ptr_free);
193 		rxq->head = rxq->tail = 0;
194 	}
195 }
196 
197 static int vhost_net_buf_peek_len(void *ptr)
198 {
199 	if (tun_is_xdp_frame(ptr)) {
200 		struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
201 
202 		return xdpf->len;
203 	}
204 
205 	return __skb_array_len_with_tag(ptr);
206 }
207 
208 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
209 {
210 	struct vhost_net_buf *rxq = &nvq->rxq;
211 
212 	if (!vhost_net_buf_is_empty(rxq))
213 		goto out;
214 
215 	if (!vhost_net_buf_produce(nvq))
216 		return 0;
217 
218 out:
219 	return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
220 }
221 
222 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
223 {
224 	rxq->head = rxq->tail = 0;
225 }
226 
227 static void vhost_net_enable_zcopy(int vq)
228 {
229 	vhost_net_zcopy_mask |= 0x1 << vq;
230 }
231 
232 static struct vhost_net_ubuf_ref *
233 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
234 {
235 	struct vhost_net_ubuf_ref *ubufs;
236 	/* No zero copy backend? Nothing to count. */
237 	if (!zcopy)
238 		return NULL;
239 	ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
240 	if (!ubufs)
241 		return ERR_PTR(-ENOMEM);
242 	atomic_set(&ubufs->refcount, 1);
243 	init_waitqueue_head(&ubufs->wait);
244 	ubufs->vq = vq;
245 	return ubufs;
246 }
247 
248 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
249 {
250 	int r = atomic_sub_return(1, &ubufs->refcount);
251 	if (unlikely(!r))
252 		wake_up(&ubufs->wait);
253 	return r;
254 }
255 
256 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
257 {
258 	vhost_net_ubuf_put(ubufs);
259 	wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
260 }
261 
262 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
263 {
264 	vhost_net_ubuf_put_and_wait(ubufs);
265 	kfree(ubufs);
266 }
267 
268 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
269 {
270 	int i;
271 
272 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
273 		kfree(n->vqs[i].ubuf_info);
274 		n->vqs[i].ubuf_info = NULL;
275 	}
276 }
277 
278 static int vhost_net_set_ubuf_info(struct vhost_net *n)
279 {
280 	bool zcopy;
281 	int i;
282 
283 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
284 		zcopy = vhost_net_zcopy_mask & (0x1 << i);
285 		if (!zcopy)
286 			continue;
287 		n->vqs[i].ubuf_info =
288 			kmalloc_array(UIO_MAXIOV,
289 				      sizeof(*n->vqs[i].ubuf_info),
290 				      GFP_KERNEL);
291 		if  (!n->vqs[i].ubuf_info)
292 			goto err;
293 	}
294 	return 0;
295 
296 err:
297 	vhost_net_clear_ubuf_info(n);
298 	return -ENOMEM;
299 }
300 
301 static void vhost_net_vq_reset(struct vhost_net *n)
302 {
303 	int i;
304 
305 	vhost_net_clear_ubuf_info(n);
306 
307 	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
308 		n->vqs[i].done_idx = 0;
309 		n->vqs[i].upend_idx = 0;
310 		n->vqs[i].ubufs = NULL;
311 		n->vqs[i].vhost_hlen = 0;
312 		n->vqs[i].sock_hlen = 0;
313 		vhost_net_buf_init(&n->vqs[i].rxq);
314 	}
315 
316 }
317 
318 static void vhost_net_tx_packet(struct vhost_net *net)
319 {
320 	++net->tx_packets;
321 	if (net->tx_packets < 1024)
322 		return;
323 	net->tx_packets = 0;
324 	net->tx_zcopy_err = 0;
325 }
326 
327 static void vhost_net_tx_err(struct vhost_net *net)
328 {
329 	++net->tx_zcopy_err;
330 }
331 
332 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
333 {
334 	/* TX flush waits for outstanding DMAs to be done.
335 	 * Don't start new DMAs.
336 	 */
337 	return !net->tx_flush &&
338 		net->tx_packets / 64 >= net->tx_zcopy_err;
339 }
340 
341 static bool vhost_sock_zcopy(struct socket *sock)
342 {
343 	return unlikely(experimental_zcopytx) &&
344 		sock_flag(sock->sk, SOCK_ZEROCOPY);
345 }
346 
347 static bool vhost_sock_xdp(struct socket *sock)
348 {
349 	return sock_flag(sock->sk, SOCK_XDP);
350 }
351 
352 /* In case of DMA done not in order in lower device driver for some reason.
353  * upend_idx is used to track end of used idx, done_idx is used to track head
354  * of used idx. Once lower device DMA done contiguously, we will signal KVM
355  * guest used idx.
356  */
357 static void vhost_zerocopy_signal_used(struct vhost_net *net,
358 				       struct vhost_virtqueue *vq)
359 {
360 	struct vhost_net_virtqueue *nvq =
361 		container_of(vq, struct vhost_net_virtqueue, vq);
362 	int i, add;
363 	int j = 0;
364 
365 	for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
366 		if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
367 			vhost_net_tx_err(net);
368 		if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
369 			vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
370 			++j;
371 		} else
372 			break;
373 	}
374 	while (j) {
375 		add = min(UIO_MAXIOV - nvq->done_idx, j);
376 		vhost_add_used_and_signal_n(vq->dev, vq,
377 					    &vq->heads[nvq->done_idx], add);
378 		nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
379 		j -= add;
380 	}
381 }
382 
383 static void vhost_zerocopy_callback(struct sk_buff *skb,
384 				    struct ubuf_info *ubuf_base, bool success)
385 {
386 	struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
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_vq_has_work(vq)) {
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, ITER_SOURCE, 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 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
657 
658 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
659 			       struct iov_iter *from)
660 {
661 	struct vhost_virtqueue *vq = &nvq->vq;
662 	struct vhost_net *net = container_of(vq->dev, struct vhost_net,
663 					     dev);
664 	struct socket *sock = vhost_vq_get_backend(vq);
665 	struct virtio_net_hdr *gso;
666 	struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
667 	struct tun_xdp_hdr *hdr;
668 	size_t len = iov_iter_count(from);
669 	int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
670 	int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
671 	int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
672 	int sock_hlen = nvq->sock_hlen;
673 	void *buf;
674 	int copied;
675 	int ret;
676 
677 	if (unlikely(len < nvq->sock_hlen))
678 		return -EFAULT;
679 
680 	if (SKB_DATA_ALIGN(len + pad) +
681 	    SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
682 		return -ENOSPC;
683 
684 	buflen += SKB_DATA_ALIGN(len + pad);
685 	buf = page_frag_alloc_align(&net->pf_cache, buflen, GFP_KERNEL,
686 				    SMP_CACHE_BYTES);
687 	if (unlikely(!buf))
688 		return -ENOMEM;
689 
690 	copied = copy_from_iter(buf + offsetof(struct tun_xdp_hdr, gso),
691 				sock_hlen, from);
692 	if (copied != sock_hlen) {
693 		ret = -EFAULT;
694 		goto err;
695 	}
696 
697 	hdr = buf;
698 	gso = &hdr->gso;
699 
700 	if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
701 	    vhost16_to_cpu(vq, gso->csum_start) +
702 	    vhost16_to_cpu(vq, gso->csum_offset) + 2 >
703 	    vhost16_to_cpu(vq, gso->hdr_len)) {
704 		gso->hdr_len = cpu_to_vhost16(vq,
705 			       vhost16_to_cpu(vq, gso->csum_start) +
706 			       vhost16_to_cpu(vq, gso->csum_offset) + 2);
707 
708 		if (vhost16_to_cpu(vq, gso->hdr_len) > len) {
709 			ret = -EINVAL;
710 			goto err;
711 		}
712 	}
713 
714 	len -= sock_hlen;
715 	copied = copy_from_iter(buf + pad, len, from);
716 	if (copied != len) {
717 		ret = -EFAULT;
718 		goto err;
719 	}
720 
721 	xdp_init_buff(xdp, buflen, NULL);
722 	xdp_prepare_buff(xdp, buf, pad, len, true);
723 	hdr->buflen = buflen;
724 
725 	++nvq->batched_xdp;
726 
727 	return 0;
728 
729 err:
730 	page_frag_free(buf);
731 	return ret;
732 }
733 
734 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
735 {
736 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
737 	struct vhost_virtqueue *vq = &nvq->vq;
738 	unsigned out, in;
739 	int head;
740 	struct msghdr msg = {
741 		.msg_name = NULL,
742 		.msg_namelen = 0,
743 		.msg_control = NULL,
744 		.msg_controllen = 0,
745 		.msg_flags = MSG_DONTWAIT,
746 	};
747 	size_t len, total_len = 0;
748 	int err;
749 	int sent_pkts = 0;
750 	bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
751 
752 	do {
753 		bool busyloop_intr = false;
754 
755 		if (nvq->done_idx == VHOST_NET_BATCH)
756 			vhost_tx_batch(net, nvq, sock, &msg);
757 
758 		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
759 				   &busyloop_intr);
760 		/* On error, stop handling until the next kick. */
761 		if (unlikely(head < 0))
762 			break;
763 		/* Nothing new?  Wait for eventfd to tell us they refilled. */
764 		if (head == vq->num) {
765 			if (unlikely(busyloop_intr)) {
766 				vhost_poll_queue(&vq->poll);
767 			} else if (unlikely(vhost_enable_notify(&net->dev,
768 								vq))) {
769 				vhost_disable_notify(&net->dev, vq);
770 				continue;
771 			}
772 			break;
773 		}
774 
775 		total_len += len;
776 
777 		/* For simplicity, TX batching is only enabled if
778 		 * sndbuf is unlimited.
779 		 */
780 		if (sock_can_batch) {
781 			err = vhost_net_build_xdp(nvq, &msg.msg_iter);
782 			if (!err) {
783 				goto done;
784 			} else if (unlikely(err != -ENOSPC)) {
785 				vhost_tx_batch(net, nvq, sock, &msg);
786 				vhost_discard_vq_desc(vq, 1);
787 				vhost_net_enable_vq(net, vq);
788 				break;
789 			}
790 
791 			/* We can't build XDP buff, go for single
792 			 * packet path but let's flush batched
793 			 * packets.
794 			 */
795 			vhost_tx_batch(net, nvq, sock, &msg);
796 			msg.msg_control = NULL;
797 		} else {
798 			if (tx_can_batch(vq, total_len))
799 				msg.msg_flags |= MSG_MORE;
800 			else
801 				msg.msg_flags &= ~MSG_MORE;
802 		}
803 
804 		err = sock->ops->sendmsg(sock, &msg, len);
805 		if (unlikely(err < 0)) {
806 			if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) {
807 				vhost_discard_vq_desc(vq, 1);
808 				vhost_net_enable_vq(net, vq);
809 				break;
810 			}
811 			pr_debug("Fail to send packet: err %d", err);
812 		} else if (unlikely(err != len))
813 			pr_debug("Truncated TX packet: len %d != %zd\n",
814 				 err, len);
815 done:
816 		vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
817 		vq->heads[nvq->done_idx].len = 0;
818 		++nvq->done_idx;
819 	} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
820 
821 	vhost_tx_batch(net, nvq, sock, &msg);
822 }
823 
824 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
825 {
826 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
827 	struct vhost_virtqueue *vq = &nvq->vq;
828 	unsigned out, in;
829 	int head;
830 	struct msghdr msg = {
831 		.msg_name = NULL,
832 		.msg_namelen = 0,
833 		.msg_control = NULL,
834 		.msg_controllen = 0,
835 		.msg_flags = MSG_DONTWAIT,
836 	};
837 	struct tun_msg_ctl ctl;
838 	size_t len, total_len = 0;
839 	int err;
840 	struct vhost_net_ubuf_ref *ubufs;
841 	struct ubuf_info_msgzc *ubuf;
842 	bool zcopy_used;
843 	int sent_pkts = 0;
844 
845 	do {
846 		bool busyloop_intr;
847 
848 		/* Release DMAs done buffers first */
849 		vhost_zerocopy_signal_used(net, vq);
850 
851 		busyloop_intr = false;
852 		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
853 				   &busyloop_intr);
854 		/* On error, stop handling until the next kick. */
855 		if (unlikely(head < 0))
856 			break;
857 		/* Nothing new?  Wait for eventfd to tell us they refilled. */
858 		if (head == vq->num) {
859 			if (unlikely(busyloop_intr)) {
860 				vhost_poll_queue(&vq->poll);
861 			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
862 				vhost_disable_notify(&net->dev, vq);
863 				continue;
864 			}
865 			break;
866 		}
867 
868 		zcopy_used = len >= VHOST_GOODCOPY_LEN
869 			     && !vhost_exceeds_maxpend(net)
870 			     && vhost_net_tx_select_zcopy(net);
871 
872 		/* use msg_control to pass vhost zerocopy ubuf info to skb */
873 		if (zcopy_used) {
874 			ubuf = nvq->ubuf_info + nvq->upend_idx;
875 			vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
876 			vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
877 			ubuf->ctx = nvq->ubufs;
878 			ubuf->desc = nvq->upend_idx;
879 			ubuf->ubuf.callback = vhost_zerocopy_callback;
880 			ubuf->ubuf.flags = SKBFL_ZEROCOPY_FRAG;
881 			refcount_set(&ubuf->ubuf.refcnt, 1);
882 			msg.msg_control = &ctl;
883 			ctl.type = TUN_MSG_UBUF;
884 			ctl.ptr = &ubuf->ubuf;
885 			msg.msg_controllen = sizeof(ctl);
886 			ubufs = nvq->ubufs;
887 			atomic_inc(&ubufs->refcount);
888 			nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
889 		} else {
890 			msg.msg_control = NULL;
891 			ubufs = NULL;
892 		}
893 		total_len += len;
894 		if (tx_can_batch(vq, total_len) &&
895 		    likely(!vhost_exceeds_maxpend(net))) {
896 			msg.msg_flags |= MSG_MORE;
897 		} else {
898 			msg.msg_flags &= ~MSG_MORE;
899 		}
900 
901 		err = sock->ops->sendmsg(sock, &msg, len);
902 		if (unlikely(err < 0)) {
903 			bool retry = err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS;
904 
905 			if (zcopy_used) {
906 				if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
907 					vhost_net_ubuf_put(ubufs);
908 				if (retry)
909 					nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
910 						% UIO_MAXIOV;
911 				else
912 					vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
913 			}
914 			if (retry) {
915 				vhost_discard_vq_desc(vq, 1);
916 				vhost_net_enable_vq(net, vq);
917 				break;
918 			}
919 			pr_debug("Fail to send packet: err %d", err);
920 		} else if (unlikely(err != len))
921 			pr_debug("Truncated TX packet: "
922 				 " len %d != %zd\n", err, len);
923 		if (!zcopy_used)
924 			vhost_add_used_and_signal(&net->dev, vq, head, 0);
925 		else
926 			vhost_zerocopy_signal_used(net, vq);
927 		vhost_net_tx_packet(net);
928 	} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
929 }
930 
931 /* Expects to be always run from workqueue - which acts as
932  * read-size critical section for our kind of RCU. */
933 static void handle_tx(struct vhost_net *net)
934 {
935 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
936 	struct vhost_virtqueue *vq = &nvq->vq;
937 	struct socket *sock;
938 
939 	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
940 	sock = vhost_vq_get_backend(vq);
941 	if (!sock)
942 		goto out;
943 
944 	if (!vq_meta_prefetch(vq))
945 		goto out;
946 
947 	vhost_disable_notify(&net->dev, vq);
948 	vhost_net_disable_vq(net, vq);
949 
950 	if (vhost_sock_zcopy(sock))
951 		handle_tx_zerocopy(net, sock);
952 	else
953 		handle_tx_copy(net, sock);
954 
955 out:
956 	mutex_unlock(&vq->mutex);
957 }
958 
959 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
960 {
961 	struct sk_buff *head;
962 	int len = 0;
963 	unsigned long flags;
964 
965 	if (rvq->rx_ring)
966 		return vhost_net_buf_peek(rvq);
967 
968 	spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
969 	head = skb_peek(&sk->sk_receive_queue);
970 	if (likely(head)) {
971 		len = head->len;
972 		if (skb_vlan_tag_present(head))
973 			len += VLAN_HLEN;
974 	}
975 
976 	spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
977 	return len;
978 }
979 
980 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
981 				      bool *busyloop_intr)
982 {
983 	struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
984 	struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
985 	struct vhost_virtqueue *rvq = &rnvq->vq;
986 	struct vhost_virtqueue *tvq = &tnvq->vq;
987 	int len = peek_head_len(rnvq, sk);
988 
989 	if (!len && rvq->busyloop_timeout) {
990 		/* Flush batched heads first */
991 		vhost_net_signal_used(rnvq);
992 		/* Both tx vq and rx socket were polled here */
993 		vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
994 
995 		len = peek_head_len(rnvq, sk);
996 	}
997 
998 	return len;
999 }
1000 
1001 /* This is a multi-buffer version of vhost_get_desc, that works if
1002  *	vq has read descriptors only.
1003  * @vq		- the relevant virtqueue
1004  * @datalen	- data length we'll be reading
1005  * @iovcount	- returned count of io vectors we fill
1006  * @log		- vhost log
1007  * @log_num	- log offset
1008  * @quota       - headcount quota, 1 for big buffer
1009  *	returns number of buffer heads allocated, negative on error
1010  */
1011 static int get_rx_bufs(struct vhost_virtqueue *vq,
1012 		       struct vring_used_elem *heads,
1013 		       int datalen,
1014 		       unsigned *iovcount,
1015 		       struct vhost_log *log,
1016 		       unsigned *log_num,
1017 		       unsigned int quota)
1018 {
1019 	unsigned int out, in;
1020 	int seg = 0;
1021 	int headcount = 0;
1022 	unsigned d;
1023 	int r, nlogs = 0;
1024 	/* len is always initialized before use since we are always called with
1025 	 * datalen > 0.
1026 	 */
1027 	u32 len;
1028 
1029 	while (datalen > 0 && headcount < quota) {
1030 		if (unlikely(seg >= UIO_MAXIOV)) {
1031 			r = -ENOBUFS;
1032 			goto err;
1033 		}
1034 		r = vhost_get_vq_desc(vq, vq->iov + seg,
1035 				      ARRAY_SIZE(vq->iov) - seg, &out,
1036 				      &in, log, log_num);
1037 		if (unlikely(r < 0))
1038 			goto err;
1039 
1040 		d = r;
1041 		if (d == vq->num) {
1042 			r = 0;
1043 			goto err;
1044 		}
1045 		if (unlikely(out || in <= 0)) {
1046 			vq_err(vq, "unexpected descriptor format for RX: "
1047 				"out %d, in %d\n", out, in);
1048 			r = -EINVAL;
1049 			goto err;
1050 		}
1051 		if (unlikely(log)) {
1052 			nlogs += *log_num;
1053 			log += *log_num;
1054 		}
1055 		heads[headcount].id = cpu_to_vhost32(vq, d);
1056 		len = iov_length(vq->iov + seg, in);
1057 		heads[headcount].len = cpu_to_vhost32(vq, len);
1058 		datalen -= len;
1059 		++headcount;
1060 		seg += in;
1061 	}
1062 	heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1063 	*iovcount = seg;
1064 	if (unlikely(log))
1065 		*log_num = nlogs;
1066 
1067 	/* Detect overrun */
1068 	if (unlikely(datalen > 0)) {
1069 		r = UIO_MAXIOV + 1;
1070 		goto err;
1071 	}
1072 	return headcount;
1073 err:
1074 	vhost_discard_vq_desc(vq, headcount);
1075 	return r;
1076 }
1077 
1078 /* Expects to be always run from workqueue - which acts as
1079  * read-size critical section for our kind of RCU. */
1080 static void handle_rx(struct vhost_net *net)
1081 {
1082 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1083 	struct vhost_virtqueue *vq = &nvq->vq;
1084 	unsigned in, log;
1085 	struct vhost_log *vq_log;
1086 	struct msghdr msg = {
1087 		.msg_name = NULL,
1088 		.msg_namelen = 0,
1089 		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
1090 		.msg_controllen = 0,
1091 		.msg_flags = MSG_DONTWAIT,
1092 	};
1093 	struct virtio_net_hdr hdr = {
1094 		.flags = 0,
1095 		.gso_type = VIRTIO_NET_HDR_GSO_NONE
1096 	};
1097 	size_t total_len = 0;
1098 	int err, mergeable;
1099 	s16 headcount;
1100 	size_t vhost_hlen, sock_hlen;
1101 	size_t vhost_len, sock_len;
1102 	bool busyloop_intr = false;
1103 	struct socket *sock;
1104 	struct iov_iter fixup;
1105 	__virtio16 num_buffers;
1106 	int recv_pkts = 0;
1107 
1108 	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1109 	sock = vhost_vq_get_backend(vq);
1110 	if (!sock)
1111 		goto out;
1112 
1113 	if (!vq_meta_prefetch(vq))
1114 		goto out;
1115 
1116 	vhost_disable_notify(&net->dev, vq);
1117 	vhost_net_disable_vq(net, vq);
1118 
1119 	vhost_hlen = nvq->vhost_hlen;
1120 	sock_hlen = nvq->sock_hlen;
1121 
1122 	vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1123 		vq->log : NULL;
1124 	mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1125 
1126 	do {
1127 		sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1128 						      &busyloop_intr);
1129 		if (!sock_len)
1130 			break;
1131 		sock_len += sock_hlen;
1132 		vhost_len = sock_len + vhost_hlen;
1133 		headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1134 					vhost_len, &in, vq_log, &log,
1135 					likely(mergeable) ? UIO_MAXIOV : 1);
1136 		/* On error, stop handling until the next kick. */
1137 		if (unlikely(headcount < 0))
1138 			goto out;
1139 		/* OK, now we need to know about added descriptors. */
1140 		if (!headcount) {
1141 			if (unlikely(busyloop_intr)) {
1142 				vhost_poll_queue(&vq->poll);
1143 			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1144 				/* They have slipped one in as we were
1145 				 * doing that: check again. */
1146 				vhost_disable_notify(&net->dev, vq);
1147 				continue;
1148 			}
1149 			/* Nothing new?  Wait for eventfd to tell us
1150 			 * they refilled. */
1151 			goto out;
1152 		}
1153 		busyloop_intr = false;
1154 		if (nvq->rx_ring)
1155 			msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1156 		/* On overrun, truncate and discard */
1157 		if (unlikely(headcount > UIO_MAXIOV)) {
1158 			iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, 1, 1);
1159 			err = sock->ops->recvmsg(sock, &msg,
1160 						 1, MSG_DONTWAIT | MSG_TRUNC);
1161 			pr_debug("Discarded rx packet: len %zd\n", sock_len);
1162 			continue;
1163 		}
1164 		/* We don't need to be notified again. */
1165 		iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, in, vhost_len);
1166 		fixup = msg.msg_iter;
1167 		if (unlikely((vhost_hlen))) {
1168 			/* We will supply the header ourselves
1169 			 * TODO: support TSO.
1170 			 */
1171 			iov_iter_advance(&msg.msg_iter, vhost_hlen);
1172 		}
1173 		err = sock->ops->recvmsg(sock, &msg,
1174 					 sock_len, MSG_DONTWAIT | MSG_TRUNC);
1175 		/* Userspace might have consumed the packet meanwhile:
1176 		 * it's not supposed to do this usually, but might be hard
1177 		 * to prevent. Discard data we got (if any) and keep going. */
1178 		if (unlikely(err != sock_len)) {
1179 			pr_debug("Discarded rx packet: "
1180 				 " len %d, expected %zd\n", err, sock_len);
1181 			vhost_discard_vq_desc(vq, headcount);
1182 			continue;
1183 		}
1184 		/* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1185 		if (unlikely(vhost_hlen)) {
1186 			if (copy_to_iter(&hdr, sizeof(hdr),
1187 					 &fixup) != sizeof(hdr)) {
1188 				vq_err(vq, "Unable to write vnet_hdr "
1189 				       "at addr %p\n", vq->iov->iov_base);
1190 				goto out;
1191 			}
1192 		} else {
1193 			/* Header came from socket; we'll need to patch
1194 			 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1195 			 */
1196 			iov_iter_advance(&fixup, sizeof(hdr));
1197 		}
1198 		/* TODO: Should check and handle checksum. */
1199 
1200 		num_buffers = cpu_to_vhost16(vq, headcount);
1201 		if (likely(mergeable) &&
1202 		    copy_to_iter(&num_buffers, sizeof num_buffers,
1203 				 &fixup) != sizeof num_buffers) {
1204 			vq_err(vq, "Failed num_buffers write");
1205 			vhost_discard_vq_desc(vq, headcount);
1206 			goto out;
1207 		}
1208 		nvq->done_idx += headcount;
1209 		if (nvq->done_idx > VHOST_NET_BATCH)
1210 			vhost_net_signal_used(nvq);
1211 		if (unlikely(vq_log))
1212 			vhost_log_write(vq, vq_log, log, vhost_len,
1213 					vq->iov, in);
1214 		total_len += vhost_len;
1215 	} while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1216 
1217 	if (unlikely(busyloop_intr))
1218 		vhost_poll_queue(&vq->poll);
1219 	else if (!sock_len)
1220 		vhost_net_enable_vq(net, vq);
1221 out:
1222 	vhost_net_signal_used(nvq);
1223 	mutex_unlock(&vq->mutex);
1224 }
1225 
1226 static void handle_tx_kick(struct vhost_work *work)
1227 {
1228 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1229 						  poll.work);
1230 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1231 
1232 	handle_tx(net);
1233 }
1234 
1235 static void handle_rx_kick(struct vhost_work *work)
1236 {
1237 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1238 						  poll.work);
1239 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1240 
1241 	handle_rx(net);
1242 }
1243 
1244 static void handle_tx_net(struct vhost_work *work)
1245 {
1246 	struct vhost_net *net = container_of(work, struct vhost_net,
1247 					     poll[VHOST_NET_VQ_TX].work);
1248 	handle_tx(net);
1249 }
1250 
1251 static void handle_rx_net(struct vhost_work *work)
1252 {
1253 	struct vhost_net *net = container_of(work, struct vhost_net,
1254 					     poll[VHOST_NET_VQ_RX].work);
1255 	handle_rx(net);
1256 }
1257 
1258 static int vhost_net_open(struct inode *inode, struct file *f)
1259 {
1260 	struct vhost_net *n;
1261 	struct vhost_dev *dev;
1262 	struct vhost_virtqueue **vqs;
1263 	void **queue;
1264 	struct xdp_buff *xdp;
1265 	int i;
1266 
1267 	n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1268 	if (!n)
1269 		return -ENOMEM;
1270 	vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1271 	if (!vqs) {
1272 		kvfree(n);
1273 		return -ENOMEM;
1274 	}
1275 
1276 	queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1277 			      GFP_KERNEL);
1278 	if (!queue) {
1279 		kfree(vqs);
1280 		kvfree(n);
1281 		return -ENOMEM;
1282 	}
1283 	n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1284 
1285 	xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1286 	if (!xdp) {
1287 		kfree(vqs);
1288 		kvfree(n);
1289 		kfree(queue);
1290 		return -ENOMEM;
1291 	}
1292 	n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1293 
1294 	dev = &n->dev;
1295 	vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1296 	vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1297 	n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1298 	n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1299 	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1300 		n->vqs[i].ubufs = NULL;
1301 		n->vqs[i].ubuf_info = NULL;
1302 		n->vqs[i].upend_idx = 0;
1303 		n->vqs[i].done_idx = 0;
1304 		n->vqs[i].batched_xdp = 0;
1305 		n->vqs[i].vhost_hlen = 0;
1306 		n->vqs[i].sock_hlen = 0;
1307 		n->vqs[i].rx_ring = NULL;
1308 		vhost_net_buf_init(&n->vqs[i].rxq);
1309 	}
1310 	vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1311 		       UIO_MAXIOV + VHOST_NET_BATCH,
1312 		       VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true,
1313 		       NULL);
1314 
1315 	vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev,
1316 			vqs[VHOST_NET_VQ_TX]);
1317 	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev,
1318 			vqs[VHOST_NET_VQ_RX]);
1319 
1320 	f->private_data = n;
1321 	n->pf_cache.va = NULL;
1322 
1323 	return 0;
1324 }
1325 
1326 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1327 					struct vhost_virtqueue *vq)
1328 {
1329 	struct socket *sock;
1330 	struct vhost_net_virtqueue *nvq =
1331 		container_of(vq, struct vhost_net_virtqueue, vq);
1332 
1333 	mutex_lock(&vq->mutex);
1334 	sock = vhost_vq_get_backend(vq);
1335 	vhost_net_disable_vq(n, vq);
1336 	vhost_vq_set_backend(vq, NULL);
1337 	vhost_net_buf_unproduce(nvq);
1338 	nvq->rx_ring = NULL;
1339 	mutex_unlock(&vq->mutex);
1340 	return sock;
1341 }
1342 
1343 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1344 			   struct socket **rx_sock)
1345 {
1346 	*tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1347 	*rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1348 }
1349 
1350 static void vhost_net_flush(struct vhost_net *n)
1351 {
1352 	vhost_dev_flush(&n->dev);
1353 	if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1354 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1355 		n->tx_flush = true;
1356 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1357 		/* Wait for all lower device DMAs done. */
1358 		vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1359 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1360 		n->tx_flush = false;
1361 		atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1362 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1363 	}
1364 }
1365 
1366 static int vhost_net_release(struct inode *inode, struct file *f)
1367 {
1368 	struct vhost_net *n = f->private_data;
1369 	struct socket *tx_sock;
1370 	struct socket *rx_sock;
1371 
1372 	vhost_net_stop(n, &tx_sock, &rx_sock);
1373 	vhost_net_flush(n);
1374 	vhost_dev_stop(&n->dev);
1375 	vhost_dev_cleanup(&n->dev);
1376 	vhost_net_vq_reset(n);
1377 	if (tx_sock)
1378 		sockfd_put(tx_sock);
1379 	if (rx_sock)
1380 		sockfd_put(rx_sock);
1381 	/* Make sure no callbacks are outstanding */
1382 	synchronize_rcu();
1383 	/* We do an extra flush before freeing memory,
1384 	 * since jobs can re-queue themselves. */
1385 	vhost_net_flush(n);
1386 	kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1387 	kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1388 	kfree(n->dev.vqs);
1389 	page_frag_cache_drain(&n->pf_cache);
1390 	kvfree(n);
1391 	return 0;
1392 }
1393 
1394 static struct socket *get_raw_socket(int fd)
1395 {
1396 	int r;
1397 	struct socket *sock = sockfd_lookup(fd, &r);
1398 
1399 	if (!sock)
1400 		return ERR_PTR(-ENOTSOCK);
1401 
1402 	/* Parameter checking */
1403 	if (sock->sk->sk_type != SOCK_RAW) {
1404 		r = -ESOCKTNOSUPPORT;
1405 		goto err;
1406 	}
1407 
1408 	if (sock->sk->sk_family != AF_PACKET) {
1409 		r = -EPFNOSUPPORT;
1410 		goto err;
1411 	}
1412 	return sock;
1413 err:
1414 	sockfd_put(sock);
1415 	return ERR_PTR(r);
1416 }
1417 
1418 static struct ptr_ring *get_tap_ptr_ring(struct file *file)
1419 {
1420 	struct ptr_ring *ring;
1421 	ring = tun_get_tx_ring(file);
1422 	if (!IS_ERR(ring))
1423 		goto out;
1424 	ring = tap_get_ptr_ring(file);
1425 	if (!IS_ERR(ring))
1426 		goto out;
1427 	ring = NULL;
1428 out:
1429 	return ring;
1430 }
1431 
1432 static struct socket *get_tap_socket(int fd)
1433 {
1434 	struct file *file = fget(fd);
1435 	struct socket *sock;
1436 
1437 	if (!file)
1438 		return ERR_PTR(-EBADF);
1439 	sock = tun_get_socket(file);
1440 	if (!IS_ERR(sock))
1441 		return sock;
1442 	sock = tap_get_socket(file);
1443 	if (IS_ERR(sock))
1444 		fput(file);
1445 	return sock;
1446 }
1447 
1448 static struct socket *get_socket(int fd)
1449 {
1450 	struct socket *sock;
1451 
1452 	/* special case to disable backend */
1453 	if (fd == -1)
1454 		return NULL;
1455 	sock = get_raw_socket(fd);
1456 	if (!IS_ERR(sock))
1457 		return sock;
1458 	sock = get_tap_socket(fd);
1459 	if (!IS_ERR(sock))
1460 		return sock;
1461 	return ERR_PTR(-ENOTSOCK);
1462 }
1463 
1464 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1465 {
1466 	struct socket *sock, *oldsock;
1467 	struct vhost_virtqueue *vq;
1468 	struct vhost_net_virtqueue *nvq;
1469 	struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1470 	int r;
1471 
1472 	mutex_lock(&n->dev.mutex);
1473 	r = vhost_dev_check_owner(&n->dev);
1474 	if (r)
1475 		goto err;
1476 
1477 	if (index >= VHOST_NET_VQ_MAX) {
1478 		r = -ENOBUFS;
1479 		goto err;
1480 	}
1481 	vq = &n->vqs[index].vq;
1482 	nvq = &n->vqs[index];
1483 	mutex_lock(&vq->mutex);
1484 
1485 	if (fd == -1)
1486 		vhost_clear_msg(&n->dev);
1487 
1488 	/* Verify that ring has been setup correctly. */
1489 	if (!vhost_vq_access_ok(vq)) {
1490 		r = -EFAULT;
1491 		goto err_vq;
1492 	}
1493 	sock = get_socket(fd);
1494 	if (IS_ERR(sock)) {
1495 		r = PTR_ERR(sock);
1496 		goto err_vq;
1497 	}
1498 
1499 	/* start polling new socket */
1500 	oldsock = vhost_vq_get_backend(vq);
1501 	if (sock != oldsock) {
1502 		ubufs = vhost_net_ubuf_alloc(vq,
1503 					     sock && vhost_sock_zcopy(sock));
1504 		if (IS_ERR(ubufs)) {
1505 			r = PTR_ERR(ubufs);
1506 			goto err_ubufs;
1507 		}
1508 
1509 		vhost_net_disable_vq(n, vq);
1510 		vhost_vq_set_backend(vq, sock);
1511 		vhost_net_buf_unproduce(nvq);
1512 		r = vhost_vq_init_access(vq);
1513 		if (r)
1514 			goto err_used;
1515 		r = vhost_net_enable_vq(n, vq);
1516 		if (r)
1517 			goto err_used;
1518 		if (index == VHOST_NET_VQ_RX) {
1519 			if (sock)
1520 				nvq->rx_ring = get_tap_ptr_ring(sock->file);
1521 			else
1522 				nvq->rx_ring = NULL;
1523 		}
1524 
1525 		oldubufs = nvq->ubufs;
1526 		nvq->ubufs = ubufs;
1527 
1528 		n->tx_packets = 0;
1529 		n->tx_zcopy_err = 0;
1530 		n->tx_flush = false;
1531 	}
1532 
1533 	mutex_unlock(&vq->mutex);
1534 
1535 	if (oldubufs) {
1536 		vhost_net_ubuf_put_wait_and_free(oldubufs);
1537 		mutex_lock(&vq->mutex);
1538 		vhost_zerocopy_signal_used(n, vq);
1539 		mutex_unlock(&vq->mutex);
1540 	}
1541 
1542 	if (oldsock) {
1543 		vhost_dev_flush(&n->dev);
1544 		sockfd_put(oldsock);
1545 	}
1546 
1547 	mutex_unlock(&n->dev.mutex);
1548 	return 0;
1549 
1550 err_used:
1551 	vhost_vq_set_backend(vq, oldsock);
1552 	vhost_net_enable_vq(n, vq);
1553 	if (ubufs)
1554 		vhost_net_ubuf_put_wait_and_free(ubufs);
1555 err_ubufs:
1556 	if (sock)
1557 		sockfd_put(sock);
1558 err_vq:
1559 	mutex_unlock(&vq->mutex);
1560 err:
1561 	mutex_unlock(&n->dev.mutex);
1562 	return r;
1563 }
1564 
1565 static long vhost_net_reset_owner(struct vhost_net *n)
1566 {
1567 	struct socket *tx_sock = NULL;
1568 	struct socket *rx_sock = NULL;
1569 	long err;
1570 	struct vhost_iotlb *umem;
1571 
1572 	mutex_lock(&n->dev.mutex);
1573 	err = vhost_dev_check_owner(&n->dev);
1574 	if (err)
1575 		goto done;
1576 	umem = vhost_dev_reset_owner_prepare();
1577 	if (!umem) {
1578 		err = -ENOMEM;
1579 		goto done;
1580 	}
1581 	vhost_net_stop(n, &tx_sock, &rx_sock);
1582 	vhost_net_flush(n);
1583 	vhost_dev_stop(&n->dev);
1584 	vhost_dev_reset_owner(&n->dev, umem);
1585 	vhost_net_vq_reset(n);
1586 done:
1587 	mutex_unlock(&n->dev.mutex);
1588 	if (tx_sock)
1589 		sockfd_put(tx_sock);
1590 	if (rx_sock)
1591 		sockfd_put(rx_sock);
1592 	return err;
1593 }
1594 
1595 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1596 {
1597 	size_t vhost_hlen, sock_hlen, hdr_len;
1598 	int i;
1599 
1600 	hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1601 			       (1ULL << VIRTIO_F_VERSION_1))) ?
1602 			sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1603 			sizeof(struct virtio_net_hdr);
1604 	if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1605 		/* vhost provides vnet_hdr */
1606 		vhost_hlen = hdr_len;
1607 		sock_hlen = 0;
1608 	} else {
1609 		/* socket provides vnet_hdr */
1610 		vhost_hlen = 0;
1611 		sock_hlen = hdr_len;
1612 	}
1613 	mutex_lock(&n->dev.mutex);
1614 	if ((features & (1 << VHOST_F_LOG_ALL)) &&
1615 	    !vhost_log_access_ok(&n->dev))
1616 		goto out_unlock;
1617 
1618 	if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) {
1619 		if (vhost_init_device_iotlb(&n->dev))
1620 			goto out_unlock;
1621 	}
1622 
1623 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1624 		mutex_lock(&n->vqs[i].vq.mutex);
1625 		n->vqs[i].vq.acked_features = features;
1626 		n->vqs[i].vhost_hlen = vhost_hlen;
1627 		n->vqs[i].sock_hlen = sock_hlen;
1628 		mutex_unlock(&n->vqs[i].vq.mutex);
1629 	}
1630 	mutex_unlock(&n->dev.mutex);
1631 	return 0;
1632 
1633 out_unlock:
1634 	mutex_unlock(&n->dev.mutex);
1635 	return -EFAULT;
1636 }
1637 
1638 static long vhost_net_set_owner(struct vhost_net *n)
1639 {
1640 	int r;
1641 
1642 	mutex_lock(&n->dev.mutex);
1643 	if (vhost_dev_has_owner(&n->dev)) {
1644 		r = -EBUSY;
1645 		goto out;
1646 	}
1647 	r = vhost_net_set_ubuf_info(n);
1648 	if (r)
1649 		goto out;
1650 	r = vhost_dev_set_owner(&n->dev);
1651 	if (r)
1652 		vhost_net_clear_ubuf_info(n);
1653 	vhost_net_flush(n);
1654 out:
1655 	mutex_unlock(&n->dev.mutex);
1656 	return r;
1657 }
1658 
1659 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1660 			    unsigned long arg)
1661 {
1662 	struct vhost_net *n = f->private_data;
1663 	void __user *argp = (void __user *)arg;
1664 	u64 __user *featurep = argp;
1665 	struct vhost_vring_file backend;
1666 	u64 features;
1667 	int r;
1668 
1669 	switch (ioctl) {
1670 	case VHOST_NET_SET_BACKEND:
1671 		if (copy_from_user(&backend, argp, sizeof backend))
1672 			return -EFAULT;
1673 		return vhost_net_set_backend(n, backend.index, backend.fd);
1674 	case VHOST_GET_FEATURES:
1675 		features = VHOST_NET_FEATURES;
1676 		if (copy_to_user(featurep, &features, sizeof features))
1677 			return -EFAULT;
1678 		return 0;
1679 	case VHOST_SET_FEATURES:
1680 		if (copy_from_user(&features, featurep, sizeof features))
1681 			return -EFAULT;
1682 		if (features & ~VHOST_NET_FEATURES)
1683 			return -EOPNOTSUPP;
1684 		return vhost_net_set_features(n, features);
1685 	case VHOST_GET_BACKEND_FEATURES:
1686 		features = VHOST_NET_BACKEND_FEATURES;
1687 		if (copy_to_user(featurep, &features, sizeof(features)))
1688 			return -EFAULT;
1689 		return 0;
1690 	case VHOST_SET_BACKEND_FEATURES:
1691 		if (copy_from_user(&features, featurep, sizeof(features)))
1692 			return -EFAULT;
1693 		if (features & ~VHOST_NET_BACKEND_FEATURES)
1694 			return -EOPNOTSUPP;
1695 		vhost_set_backend_features(&n->dev, features);
1696 		return 0;
1697 	case VHOST_RESET_OWNER:
1698 		return vhost_net_reset_owner(n);
1699 	case VHOST_SET_OWNER:
1700 		return vhost_net_set_owner(n);
1701 	default:
1702 		mutex_lock(&n->dev.mutex);
1703 		r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1704 		if (r == -ENOIOCTLCMD)
1705 			r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1706 		else
1707 			vhost_net_flush(n);
1708 		mutex_unlock(&n->dev.mutex);
1709 		return r;
1710 	}
1711 }
1712 
1713 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1714 {
1715 	struct file *file = iocb->ki_filp;
1716 	struct vhost_net *n = file->private_data;
1717 	struct vhost_dev *dev = &n->dev;
1718 	int noblock = file->f_flags & O_NONBLOCK;
1719 
1720 	return vhost_chr_read_iter(dev, to, noblock);
1721 }
1722 
1723 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1724 					struct iov_iter *from)
1725 {
1726 	struct file *file = iocb->ki_filp;
1727 	struct vhost_net *n = file->private_data;
1728 	struct vhost_dev *dev = &n->dev;
1729 
1730 	return vhost_chr_write_iter(dev, from);
1731 }
1732 
1733 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1734 {
1735 	struct vhost_net *n = file->private_data;
1736 	struct vhost_dev *dev = &n->dev;
1737 
1738 	return vhost_chr_poll(file, dev, wait);
1739 }
1740 
1741 static const struct file_operations vhost_net_fops = {
1742 	.owner          = THIS_MODULE,
1743 	.release        = vhost_net_release,
1744 	.read_iter      = vhost_net_chr_read_iter,
1745 	.write_iter     = vhost_net_chr_write_iter,
1746 	.poll           = vhost_net_chr_poll,
1747 	.unlocked_ioctl = vhost_net_ioctl,
1748 	.compat_ioctl   = compat_ptr_ioctl,
1749 	.open           = vhost_net_open,
1750 	.llseek		= noop_llseek,
1751 };
1752 
1753 static struct miscdevice vhost_net_misc = {
1754 	.minor = VHOST_NET_MINOR,
1755 	.name = "vhost-net",
1756 	.fops = &vhost_net_fops,
1757 };
1758 
1759 static int __init vhost_net_init(void)
1760 {
1761 	if (experimental_zcopytx)
1762 		vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1763 	return misc_register(&vhost_net_misc);
1764 }
1765 module_init(vhost_net_init);
1766 
1767 static void __exit vhost_net_exit(void)
1768 {
1769 	misc_deregister(&vhost_net_misc);
1770 }
1771 module_exit(vhost_net_exit);
1772 
1773 MODULE_VERSION("0.0.1");
1774 MODULE_LICENSE("GPL v2");
1775 MODULE_AUTHOR("Michael S. Tsirkin");
1776 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1777 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1778 MODULE_ALIAS("devname:vhost-net");
1779