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_complete(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 const struct ubuf_info_ops vhost_ubuf_ops = {
412 	.complete = vhost_zerocopy_complete,
413 };
414 
415 static inline unsigned long busy_clock(void)
416 {
417 	return local_clock() >> 10;
418 }
419 
420 static bool vhost_can_busy_poll(unsigned long endtime)
421 {
422 	return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
423 		      !signal_pending(current));
424 }
425 
426 static void vhost_net_disable_vq(struct vhost_net *n,
427 				 struct vhost_virtqueue *vq)
428 {
429 	struct vhost_net_virtqueue *nvq =
430 		container_of(vq, struct vhost_net_virtqueue, vq);
431 	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
432 	if (!vhost_vq_get_backend(vq))
433 		return;
434 	vhost_poll_stop(poll);
435 }
436 
437 static int vhost_net_enable_vq(struct vhost_net *n,
438 				struct vhost_virtqueue *vq)
439 {
440 	struct vhost_net_virtqueue *nvq =
441 		container_of(vq, struct vhost_net_virtqueue, vq);
442 	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
443 	struct socket *sock;
444 
445 	sock = vhost_vq_get_backend(vq);
446 	if (!sock)
447 		return 0;
448 
449 	return vhost_poll_start(poll, sock->file);
450 }
451 
452 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
453 {
454 	struct vhost_virtqueue *vq = &nvq->vq;
455 	struct vhost_dev *dev = vq->dev;
456 
457 	if (!nvq->done_idx)
458 		return;
459 
460 	vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
461 	nvq->done_idx = 0;
462 }
463 
464 static void vhost_tx_batch(struct vhost_net *net,
465 			   struct vhost_net_virtqueue *nvq,
466 			   struct socket *sock,
467 			   struct msghdr *msghdr)
468 {
469 	struct tun_msg_ctl ctl = {
470 		.type = TUN_MSG_PTR,
471 		.num = nvq->batched_xdp,
472 		.ptr = nvq->xdp,
473 	};
474 	int i, err;
475 
476 	if (nvq->batched_xdp == 0)
477 		goto signal_used;
478 
479 	msghdr->msg_control = &ctl;
480 	msghdr->msg_controllen = sizeof(ctl);
481 	err = sock->ops->sendmsg(sock, msghdr, 0);
482 	if (unlikely(err < 0)) {
483 		vq_err(&nvq->vq, "Fail to batch sending packets\n");
484 
485 		/* free pages owned by XDP; since this is an unlikely error path,
486 		 * keep it simple and avoid more complex bulk update for the
487 		 * used pages
488 		 */
489 		for (i = 0; i < nvq->batched_xdp; ++i)
490 			put_page(virt_to_head_page(nvq->xdp[i].data));
491 		nvq->batched_xdp = 0;
492 		nvq->done_idx = 0;
493 		return;
494 	}
495 
496 signal_used:
497 	vhost_net_signal_used(nvq);
498 	nvq->batched_xdp = 0;
499 }
500 
501 static int sock_has_rx_data(struct socket *sock)
502 {
503 	if (unlikely(!sock))
504 		return 0;
505 
506 	if (sock->ops->peek_len)
507 		return sock->ops->peek_len(sock);
508 
509 	return skb_queue_empty(&sock->sk->sk_receive_queue);
510 }
511 
512 static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
513 					  struct vhost_virtqueue *vq)
514 {
515 	if (!vhost_vq_avail_empty(&net->dev, vq)) {
516 		vhost_poll_queue(&vq->poll);
517 	} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
518 		vhost_disable_notify(&net->dev, vq);
519 		vhost_poll_queue(&vq->poll);
520 	}
521 }
522 
523 static void vhost_net_busy_poll(struct vhost_net *net,
524 				struct vhost_virtqueue *rvq,
525 				struct vhost_virtqueue *tvq,
526 				bool *busyloop_intr,
527 				bool poll_rx)
528 {
529 	unsigned long busyloop_timeout;
530 	unsigned long endtime;
531 	struct socket *sock;
532 	struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
533 
534 	/* Try to hold the vq mutex of the paired virtqueue. We can't
535 	 * use mutex_lock() here since we could not guarantee a
536 	 * consistenet lock ordering.
537 	 */
538 	if (!mutex_trylock(&vq->mutex))
539 		return;
540 
541 	vhost_disable_notify(&net->dev, vq);
542 	sock = vhost_vq_get_backend(rvq);
543 
544 	busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
545 				     tvq->busyloop_timeout;
546 
547 	preempt_disable();
548 	endtime = busy_clock() + busyloop_timeout;
549 
550 	while (vhost_can_busy_poll(endtime)) {
551 		if (vhost_vq_has_work(vq)) {
552 			*busyloop_intr = true;
553 			break;
554 		}
555 
556 		if ((sock_has_rx_data(sock) &&
557 		     !vhost_vq_avail_empty(&net->dev, rvq)) ||
558 		    !vhost_vq_avail_empty(&net->dev, tvq))
559 			break;
560 
561 		cpu_relax();
562 	}
563 
564 	preempt_enable();
565 
566 	if (poll_rx || sock_has_rx_data(sock))
567 		vhost_net_busy_poll_try_queue(net, vq);
568 	else if (!poll_rx) /* On tx here, sock has no rx data. */
569 		vhost_enable_notify(&net->dev, rvq);
570 
571 	mutex_unlock(&vq->mutex);
572 }
573 
574 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
575 				    struct vhost_net_virtqueue *tnvq,
576 				    unsigned int *out_num, unsigned int *in_num,
577 				    struct msghdr *msghdr, bool *busyloop_intr)
578 {
579 	struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
580 	struct vhost_virtqueue *rvq = &rnvq->vq;
581 	struct vhost_virtqueue *tvq = &tnvq->vq;
582 
583 	int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
584 				  out_num, in_num, NULL, NULL);
585 
586 	if (r == tvq->num && tvq->busyloop_timeout) {
587 		/* Flush batched packets first */
588 		if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq)))
589 			vhost_tx_batch(net, tnvq,
590 				       vhost_vq_get_backend(tvq),
591 				       msghdr);
592 
593 		vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
594 
595 		r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
596 				      out_num, in_num, NULL, NULL);
597 	}
598 
599 	return r;
600 }
601 
602 static bool vhost_exceeds_maxpend(struct vhost_net *net)
603 {
604 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
605 	struct vhost_virtqueue *vq = &nvq->vq;
606 
607 	return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
608 	       min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
609 }
610 
611 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
612 			    size_t hdr_size, int out)
613 {
614 	/* Skip header. TODO: support TSO. */
615 	size_t len = iov_length(vq->iov, out);
616 
617 	iov_iter_init(iter, ITER_SOURCE, vq->iov, out, len);
618 	iov_iter_advance(iter, hdr_size);
619 
620 	return iov_iter_count(iter);
621 }
622 
623 static int get_tx_bufs(struct vhost_net *net,
624 		       struct vhost_net_virtqueue *nvq,
625 		       struct msghdr *msg,
626 		       unsigned int *out, unsigned int *in,
627 		       size_t *len, bool *busyloop_intr)
628 {
629 	struct vhost_virtqueue *vq = &nvq->vq;
630 	int ret;
631 
632 	ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
633 
634 	if (ret < 0 || ret == vq->num)
635 		return ret;
636 
637 	if (*in) {
638 		vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
639 			*out, *in);
640 		return -EFAULT;
641 	}
642 
643 	/* Sanity check */
644 	*len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
645 	if (*len == 0) {
646 		vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
647 			*len, nvq->vhost_hlen);
648 		return -EFAULT;
649 	}
650 
651 	return ret;
652 }
653 
654 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
655 {
656 	return total_len < VHOST_NET_WEIGHT &&
657 	       !vhost_vq_avail_empty(vq->dev, vq);
658 }
659 
660 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
661 
662 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
663 			       struct iov_iter *from)
664 {
665 	struct vhost_virtqueue *vq = &nvq->vq;
666 	struct vhost_net *net = container_of(vq->dev, struct vhost_net,
667 					     dev);
668 	struct socket *sock = vhost_vq_get_backend(vq);
669 	struct virtio_net_hdr *gso;
670 	struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
671 	struct tun_xdp_hdr *hdr;
672 	size_t len = iov_iter_count(from);
673 	int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
674 	int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
675 	int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
676 	int sock_hlen = nvq->sock_hlen;
677 	void *buf;
678 	int copied;
679 	int ret;
680 
681 	if (unlikely(len < nvq->sock_hlen))
682 		return -EFAULT;
683 
684 	if (SKB_DATA_ALIGN(len + pad) +
685 	    SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
686 		return -ENOSPC;
687 
688 	buflen += SKB_DATA_ALIGN(len + pad);
689 	buf = page_frag_alloc_align(&net->pf_cache, buflen, GFP_KERNEL,
690 				    SMP_CACHE_BYTES);
691 	if (unlikely(!buf))
692 		return -ENOMEM;
693 
694 	copied = copy_from_iter(buf + offsetof(struct tun_xdp_hdr, gso),
695 				sock_hlen, from);
696 	if (copied != sock_hlen) {
697 		ret = -EFAULT;
698 		goto err;
699 	}
700 
701 	hdr = buf;
702 	gso = &hdr->gso;
703 
704 	if (!sock_hlen)
705 		memset(buf, 0, pad);
706 
707 	if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
708 	    vhost16_to_cpu(vq, gso->csum_start) +
709 	    vhost16_to_cpu(vq, gso->csum_offset) + 2 >
710 	    vhost16_to_cpu(vq, gso->hdr_len)) {
711 		gso->hdr_len = cpu_to_vhost16(vq,
712 			       vhost16_to_cpu(vq, gso->csum_start) +
713 			       vhost16_to_cpu(vq, gso->csum_offset) + 2);
714 
715 		if (vhost16_to_cpu(vq, gso->hdr_len) > len) {
716 			ret = -EINVAL;
717 			goto err;
718 		}
719 	}
720 
721 	len -= sock_hlen;
722 	copied = copy_from_iter(buf + pad, len, from);
723 	if (copied != len) {
724 		ret = -EFAULT;
725 		goto err;
726 	}
727 
728 	xdp_init_buff(xdp, buflen, NULL);
729 	xdp_prepare_buff(xdp, buf, pad, len, true);
730 	hdr->buflen = buflen;
731 
732 	++nvq->batched_xdp;
733 
734 	return 0;
735 
736 err:
737 	page_frag_free(buf);
738 	return ret;
739 }
740 
741 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
742 {
743 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
744 	struct vhost_virtqueue *vq = &nvq->vq;
745 	unsigned out, in;
746 	int head;
747 	struct msghdr msg = {
748 		.msg_name = NULL,
749 		.msg_namelen = 0,
750 		.msg_control = NULL,
751 		.msg_controllen = 0,
752 		.msg_flags = MSG_DONTWAIT,
753 	};
754 	size_t len, total_len = 0;
755 	int err;
756 	int sent_pkts = 0;
757 	bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
758 	bool busyloop_intr;
759 
760 	do {
761 		busyloop_intr = false;
762 		if (nvq->done_idx == VHOST_NET_BATCH)
763 			vhost_tx_batch(net, nvq, sock, &msg);
764 
765 		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
766 				   &busyloop_intr);
767 		/* On error, stop handling until the next kick. */
768 		if (unlikely(head < 0))
769 			break;
770 		/* Nothing new?  Wait for eventfd to tell us they refilled. */
771 		if (head == vq->num) {
772 			/* Kicks are disabled at this point, break loop and
773 			 * process any remaining batched packets. Queue will
774 			 * be re-enabled afterwards.
775 			 */
776 			break;
777 		}
778 
779 		total_len += len;
780 
781 		/* For simplicity, TX batching is only enabled if
782 		 * sndbuf is unlimited.
783 		 */
784 		if (sock_can_batch) {
785 			err = vhost_net_build_xdp(nvq, &msg.msg_iter);
786 			if (!err) {
787 				goto done;
788 			} else if (unlikely(err != -ENOSPC)) {
789 				vhost_tx_batch(net, nvq, sock, &msg);
790 				vhost_discard_vq_desc(vq, 1);
791 				vhost_net_enable_vq(net, vq);
792 				break;
793 			}
794 
795 			/* We can't build XDP buff, go for single
796 			 * packet path but let's flush batched
797 			 * packets.
798 			 */
799 			vhost_tx_batch(net, nvq, sock, &msg);
800 			msg.msg_control = NULL;
801 		} else {
802 			if (tx_can_batch(vq, total_len))
803 				msg.msg_flags |= MSG_MORE;
804 			else
805 				msg.msg_flags &= ~MSG_MORE;
806 		}
807 
808 		err = sock->ops->sendmsg(sock, &msg, len);
809 		if (unlikely(err < 0)) {
810 			if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) {
811 				vhost_discard_vq_desc(vq, 1);
812 				vhost_net_enable_vq(net, vq);
813 				break;
814 			}
815 			pr_debug("Fail to send packet: err %d", err);
816 		} else if (unlikely(err != len))
817 			pr_debug("Truncated TX packet: len %d != %zd\n",
818 				 err, len);
819 done:
820 		vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
821 		vq->heads[nvq->done_idx].len = 0;
822 		++nvq->done_idx;
823 	} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
824 
825 	/* Kicks are still disabled, dispatch any remaining batched msgs. */
826 	vhost_tx_batch(net, nvq, sock, &msg);
827 
828 	if (unlikely(busyloop_intr))
829 		/* If interrupted while doing busy polling, requeue the
830 		 * handler to be fair handle_rx as well as other tasks
831 		 * waiting on cpu.
832 		 */
833 		vhost_poll_queue(&vq->poll);
834 	else
835 		/* All of our work has been completed; however, before
836 		 * leaving the TX handler, do one last check for work,
837 		 * and requeue handler if necessary. If there is no work,
838 		 * queue will be reenabled.
839 		 */
840 		vhost_net_busy_poll_try_queue(net, vq);
841 }
842 
843 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
844 {
845 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
846 	struct vhost_virtqueue *vq = &nvq->vq;
847 	unsigned out, in;
848 	int head;
849 	struct msghdr msg = {
850 		.msg_name = NULL,
851 		.msg_namelen = 0,
852 		.msg_control = NULL,
853 		.msg_controllen = 0,
854 		.msg_flags = MSG_DONTWAIT,
855 	};
856 	struct tun_msg_ctl ctl;
857 	size_t len, total_len = 0;
858 	int err;
859 	struct vhost_net_ubuf_ref *ubufs;
860 	struct ubuf_info_msgzc *ubuf;
861 	bool zcopy_used;
862 	int sent_pkts = 0;
863 
864 	do {
865 		bool busyloop_intr;
866 
867 		/* Release DMAs done buffers first */
868 		vhost_zerocopy_signal_used(net, vq);
869 
870 		busyloop_intr = false;
871 		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
872 				   &busyloop_intr);
873 		/* On error, stop handling until the next kick. */
874 		if (unlikely(head < 0))
875 			break;
876 		/* Nothing new?  Wait for eventfd to tell us they refilled. */
877 		if (head == vq->num) {
878 			if (unlikely(busyloop_intr)) {
879 				vhost_poll_queue(&vq->poll);
880 			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
881 				vhost_disable_notify(&net->dev, vq);
882 				continue;
883 			}
884 			break;
885 		}
886 
887 		zcopy_used = len >= VHOST_GOODCOPY_LEN
888 			     && !vhost_exceeds_maxpend(net)
889 			     && vhost_net_tx_select_zcopy(net);
890 
891 		/* use msg_control to pass vhost zerocopy ubuf info to skb */
892 		if (zcopy_used) {
893 			ubuf = nvq->ubuf_info + nvq->upend_idx;
894 			vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
895 			vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
896 			ubuf->ctx = nvq->ubufs;
897 			ubuf->desc = nvq->upend_idx;
898 			ubuf->ubuf.ops = &vhost_ubuf_ops;
899 			ubuf->ubuf.flags = SKBFL_ZEROCOPY_FRAG;
900 			refcount_set(&ubuf->ubuf.refcnt, 1);
901 			msg.msg_control = &ctl;
902 			ctl.type = TUN_MSG_UBUF;
903 			ctl.ptr = &ubuf->ubuf;
904 			msg.msg_controllen = sizeof(ctl);
905 			ubufs = nvq->ubufs;
906 			atomic_inc(&ubufs->refcount);
907 			nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
908 		} else {
909 			msg.msg_control = NULL;
910 			ubufs = NULL;
911 		}
912 		total_len += len;
913 		if (tx_can_batch(vq, total_len) &&
914 		    likely(!vhost_exceeds_maxpend(net))) {
915 			msg.msg_flags |= MSG_MORE;
916 		} else {
917 			msg.msg_flags &= ~MSG_MORE;
918 		}
919 
920 		err = sock->ops->sendmsg(sock, &msg, len);
921 		if (unlikely(err < 0)) {
922 			bool retry = err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS;
923 
924 			if (zcopy_used) {
925 				if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
926 					vhost_net_ubuf_put(ubufs);
927 				if (retry)
928 					nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
929 						% UIO_MAXIOV;
930 				else
931 					vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
932 			}
933 			if (retry) {
934 				vhost_discard_vq_desc(vq, 1);
935 				vhost_net_enable_vq(net, vq);
936 				break;
937 			}
938 			pr_debug("Fail to send packet: err %d", err);
939 		} else if (unlikely(err != len))
940 			pr_debug("Truncated TX packet: "
941 				 " len %d != %zd\n", err, len);
942 		if (!zcopy_used)
943 			vhost_add_used_and_signal(&net->dev, vq, head, 0);
944 		else
945 			vhost_zerocopy_signal_used(net, vq);
946 		vhost_net_tx_packet(net);
947 	} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
948 }
949 
950 /* Expects to be always run from workqueue - which acts as
951  * read-size critical section for our kind of RCU. */
952 static void handle_tx(struct vhost_net *net)
953 {
954 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
955 	struct vhost_virtqueue *vq = &nvq->vq;
956 	struct socket *sock;
957 
958 	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
959 	sock = vhost_vq_get_backend(vq);
960 	if (!sock)
961 		goto out;
962 
963 	if (!vq_meta_prefetch(vq))
964 		goto out;
965 
966 	vhost_disable_notify(&net->dev, vq);
967 	vhost_net_disable_vq(net, vq);
968 
969 	if (vhost_sock_zcopy(sock))
970 		handle_tx_zerocopy(net, sock);
971 	else
972 		handle_tx_copy(net, sock);
973 
974 out:
975 	mutex_unlock(&vq->mutex);
976 }
977 
978 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
979 {
980 	struct sk_buff *head;
981 	int len = 0;
982 	unsigned long flags;
983 
984 	if (rvq->rx_ring)
985 		return vhost_net_buf_peek(rvq);
986 
987 	spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
988 	head = skb_peek(&sk->sk_receive_queue);
989 	if (likely(head)) {
990 		len = head->len;
991 		if (skb_vlan_tag_present(head))
992 			len += VLAN_HLEN;
993 	}
994 
995 	spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
996 	return len;
997 }
998 
999 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
1000 				      bool *busyloop_intr)
1001 {
1002 	struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1003 	struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1004 	struct vhost_virtqueue *rvq = &rnvq->vq;
1005 	struct vhost_virtqueue *tvq = &tnvq->vq;
1006 	int len = peek_head_len(rnvq, sk);
1007 
1008 	if (!len && rvq->busyloop_timeout) {
1009 		/* Flush batched heads first */
1010 		vhost_net_signal_used(rnvq);
1011 		/* Both tx vq and rx socket were polled here */
1012 		vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1013 
1014 		len = peek_head_len(rnvq, sk);
1015 	}
1016 
1017 	return len;
1018 }
1019 
1020 /* This is a multi-buffer version of vhost_get_desc, that works if
1021  *	vq has read descriptors only.
1022  * @vq		- the relevant virtqueue
1023  * @datalen	- data length we'll be reading
1024  * @iovcount	- returned count of io vectors we fill
1025  * @log		- vhost log
1026  * @log_num	- log offset
1027  * @quota       - headcount quota, 1 for big buffer
1028  *	returns number of buffer heads allocated, negative on error
1029  */
1030 static int get_rx_bufs(struct vhost_virtqueue *vq,
1031 		       struct vring_used_elem *heads,
1032 		       int datalen,
1033 		       unsigned *iovcount,
1034 		       struct vhost_log *log,
1035 		       unsigned *log_num,
1036 		       unsigned int quota)
1037 {
1038 	unsigned int out, in;
1039 	int seg = 0;
1040 	int headcount = 0;
1041 	unsigned d;
1042 	int r, nlogs = 0;
1043 	/* len is always initialized before use since we are always called with
1044 	 * datalen > 0.
1045 	 */
1046 	u32 len;
1047 
1048 	while (datalen > 0 && headcount < quota) {
1049 		if (unlikely(seg >= UIO_MAXIOV)) {
1050 			r = -ENOBUFS;
1051 			goto err;
1052 		}
1053 		r = vhost_get_vq_desc(vq, vq->iov + seg,
1054 				      ARRAY_SIZE(vq->iov) - seg, &out,
1055 				      &in, log, log_num);
1056 		if (unlikely(r < 0))
1057 			goto err;
1058 
1059 		d = r;
1060 		if (d == vq->num) {
1061 			r = 0;
1062 			goto err;
1063 		}
1064 		if (unlikely(out || in <= 0)) {
1065 			vq_err(vq, "unexpected descriptor format for RX: "
1066 				"out %d, in %d\n", out, in);
1067 			r = -EINVAL;
1068 			goto err;
1069 		}
1070 		if (unlikely(log)) {
1071 			nlogs += *log_num;
1072 			log += *log_num;
1073 		}
1074 		heads[headcount].id = cpu_to_vhost32(vq, d);
1075 		len = iov_length(vq->iov + seg, in);
1076 		heads[headcount].len = cpu_to_vhost32(vq, len);
1077 		datalen -= len;
1078 		++headcount;
1079 		seg += in;
1080 	}
1081 	heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1082 	*iovcount = seg;
1083 	if (unlikely(log))
1084 		*log_num = nlogs;
1085 
1086 	/* Detect overrun */
1087 	if (unlikely(datalen > 0)) {
1088 		r = UIO_MAXIOV + 1;
1089 		goto err;
1090 	}
1091 	return headcount;
1092 err:
1093 	vhost_discard_vq_desc(vq, headcount);
1094 	return r;
1095 }
1096 
1097 /* Expects to be always run from workqueue - which acts as
1098  * read-size critical section for our kind of RCU. */
1099 static void handle_rx(struct vhost_net *net)
1100 {
1101 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1102 	struct vhost_virtqueue *vq = &nvq->vq;
1103 	unsigned in, log;
1104 	struct vhost_log *vq_log;
1105 	struct msghdr msg = {
1106 		.msg_name = NULL,
1107 		.msg_namelen = 0,
1108 		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
1109 		.msg_controllen = 0,
1110 		.msg_flags = MSG_DONTWAIT,
1111 	};
1112 	struct virtio_net_hdr hdr = {
1113 		.flags = 0,
1114 		.gso_type = VIRTIO_NET_HDR_GSO_NONE
1115 	};
1116 	size_t total_len = 0;
1117 	int err, mergeable;
1118 	s16 headcount;
1119 	size_t vhost_hlen, sock_hlen;
1120 	size_t vhost_len, sock_len;
1121 	bool busyloop_intr = false;
1122 	bool set_num_buffers;
1123 	struct socket *sock;
1124 	struct iov_iter fixup;
1125 	__virtio16 num_buffers;
1126 	int recv_pkts = 0;
1127 
1128 	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1129 	sock = vhost_vq_get_backend(vq);
1130 	if (!sock)
1131 		goto out;
1132 
1133 	if (!vq_meta_prefetch(vq))
1134 		goto out;
1135 
1136 	vhost_disable_notify(&net->dev, vq);
1137 	vhost_net_disable_vq(net, vq);
1138 
1139 	vhost_hlen = nvq->vhost_hlen;
1140 	sock_hlen = nvq->sock_hlen;
1141 
1142 	vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1143 		vq->log : NULL;
1144 	mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1145 	set_num_buffers = mergeable ||
1146 			  vhost_has_feature(vq, VIRTIO_F_VERSION_1);
1147 
1148 	do {
1149 		sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1150 						      &busyloop_intr);
1151 		if (!sock_len)
1152 			break;
1153 		sock_len += sock_hlen;
1154 		vhost_len = sock_len + vhost_hlen;
1155 		headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1156 					vhost_len, &in, vq_log, &log,
1157 					likely(mergeable) ? UIO_MAXIOV : 1);
1158 		/* On error, stop handling until the next kick. */
1159 		if (unlikely(headcount < 0))
1160 			goto out;
1161 		/* OK, now we need to know about added descriptors. */
1162 		if (!headcount) {
1163 			if (unlikely(busyloop_intr)) {
1164 				vhost_poll_queue(&vq->poll);
1165 			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1166 				/* They have slipped one in as we were
1167 				 * doing that: check again. */
1168 				vhost_disable_notify(&net->dev, vq);
1169 				continue;
1170 			}
1171 			/* Nothing new?  Wait for eventfd to tell us
1172 			 * they refilled. */
1173 			goto out;
1174 		}
1175 		busyloop_intr = false;
1176 		if (nvq->rx_ring)
1177 			msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1178 		/* On overrun, truncate and discard */
1179 		if (unlikely(headcount > UIO_MAXIOV)) {
1180 			iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, 1, 1);
1181 			err = sock->ops->recvmsg(sock, &msg,
1182 						 1, MSG_DONTWAIT | MSG_TRUNC);
1183 			pr_debug("Discarded rx packet: len %zd\n", sock_len);
1184 			continue;
1185 		}
1186 		/* We don't need to be notified again. */
1187 		iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, in, vhost_len);
1188 		fixup = msg.msg_iter;
1189 		if (unlikely((vhost_hlen))) {
1190 			/* We will supply the header ourselves
1191 			 * TODO: support TSO.
1192 			 */
1193 			iov_iter_advance(&msg.msg_iter, vhost_hlen);
1194 		}
1195 		err = sock->ops->recvmsg(sock, &msg,
1196 					 sock_len, MSG_DONTWAIT | MSG_TRUNC);
1197 		/* Userspace might have consumed the packet meanwhile:
1198 		 * it's not supposed to do this usually, but might be hard
1199 		 * to prevent. Discard data we got (if any) and keep going. */
1200 		if (unlikely(err != sock_len)) {
1201 			pr_debug("Discarded rx packet: "
1202 				 " len %d, expected %zd\n", err, sock_len);
1203 			vhost_discard_vq_desc(vq, headcount);
1204 			continue;
1205 		}
1206 		/* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1207 		if (unlikely(vhost_hlen)) {
1208 			if (copy_to_iter(&hdr, sizeof(hdr),
1209 					 &fixup) != sizeof(hdr)) {
1210 				vq_err(vq, "Unable to write vnet_hdr "
1211 				       "at addr %p\n", vq->iov->iov_base);
1212 				goto out;
1213 			}
1214 		} else {
1215 			/* Header came from socket; we'll need to patch
1216 			 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1217 			 */
1218 			iov_iter_advance(&fixup, sizeof(hdr));
1219 		}
1220 		/* TODO: Should check and handle checksum. */
1221 
1222 		num_buffers = cpu_to_vhost16(vq, headcount);
1223 		if (likely(set_num_buffers) &&
1224 		    copy_to_iter(&num_buffers, sizeof num_buffers,
1225 				 &fixup) != sizeof num_buffers) {
1226 			vq_err(vq, "Failed num_buffers write");
1227 			vhost_discard_vq_desc(vq, headcount);
1228 			goto out;
1229 		}
1230 		nvq->done_idx += headcount;
1231 		if (nvq->done_idx > VHOST_NET_BATCH)
1232 			vhost_net_signal_used(nvq);
1233 		if (unlikely(vq_log))
1234 			vhost_log_write(vq, vq_log, log, vhost_len,
1235 					vq->iov, in);
1236 		total_len += vhost_len;
1237 	} while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1238 
1239 	if (unlikely(busyloop_intr))
1240 		vhost_poll_queue(&vq->poll);
1241 	else if (!sock_len)
1242 		vhost_net_enable_vq(net, vq);
1243 out:
1244 	vhost_net_signal_used(nvq);
1245 	mutex_unlock(&vq->mutex);
1246 }
1247 
1248 static void handle_tx_kick(struct vhost_work *work)
1249 {
1250 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1251 						  poll.work);
1252 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1253 
1254 	handle_tx(net);
1255 }
1256 
1257 static void handle_rx_kick(struct vhost_work *work)
1258 {
1259 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1260 						  poll.work);
1261 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1262 
1263 	handle_rx(net);
1264 }
1265 
1266 static void handle_tx_net(struct vhost_work *work)
1267 {
1268 	struct vhost_net *net = container_of(work, struct vhost_net,
1269 					     poll[VHOST_NET_VQ_TX].work);
1270 	handle_tx(net);
1271 }
1272 
1273 static void handle_rx_net(struct vhost_work *work)
1274 {
1275 	struct vhost_net *net = container_of(work, struct vhost_net,
1276 					     poll[VHOST_NET_VQ_RX].work);
1277 	handle_rx(net);
1278 }
1279 
1280 static int vhost_net_open(struct inode *inode, struct file *f)
1281 {
1282 	struct vhost_net *n;
1283 	struct vhost_dev *dev;
1284 	struct vhost_virtqueue **vqs;
1285 	void **queue;
1286 	struct xdp_buff *xdp;
1287 	int i;
1288 
1289 	n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1290 	if (!n)
1291 		return -ENOMEM;
1292 	vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1293 	if (!vqs) {
1294 		kvfree(n);
1295 		return -ENOMEM;
1296 	}
1297 
1298 	queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1299 			      GFP_KERNEL);
1300 	if (!queue) {
1301 		kfree(vqs);
1302 		kvfree(n);
1303 		return -ENOMEM;
1304 	}
1305 	n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1306 
1307 	xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1308 	if (!xdp) {
1309 		kfree(vqs);
1310 		kvfree(n);
1311 		kfree(queue);
1312 		return -ENOMEM;
1313 	}
1314 	n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1315 
1316 	dev = &n->dev;
1317 	vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1318 	vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1319 	n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1320 	n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1321 	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1322 		n->vqs[i].ubufs = NULL;
1323 		n->vqs[i].ubuf_info = NULL;
1324 		n->vqs[i].upend_idx = 0;
1325 		n->vqs[i].done_idx = 0;
1326 		n->vqs[i].batched_xdp = 0;
1327 		n->vqs[i].vhost_hlen = 0;
1328 		n->vqs[i].sock_hlen = 0;
1329 		n->vqs[i].rx_ring = NULL;
1330 		vhost_net_buf_init(&n->vqs[i].rxq);
1331 	}
1332 	vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1333 		       UIO_MAXIOV + VHOST_NET_BATCH,
1334 		       VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true,
1335 		       NULL);
1336 
1337 	vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev,
1338 			vqs[VHOST_NET_VQ_TX]);
1339 	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev,
1340 			vqs[VHOST_NET_VQ_RX]);
1341 
1342 	f->private_data = n;
1343 	page_frag_cache_init(&n->pf_cache);
1344 
1345 	return 0;
1346 }
1347 
1348 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1349 					struct vhost_virtqueue *vq)
1350 {
1351 	struct socket *sock;
1352 	struct vhost_net_virtqueue *nvq =
1353 		container_of(vq, struct vhost_net_virtqueue, vq);
1354 
1355 	mutex_lock(&vq->mutex);
1356 	sock = vhost_vq_get_backend(vq);
1357 	vhost_net_disable_vq(n, vq);
1358 	vhost_vq_set_backend(vq, NULL);
1359 	vhost_net_buf_unproduce(nvq);
1360 	nvq->rx_ring = NULL;
1361 	mutex_unlock(&vq->mutex);
1362 	return sock;
1363 }
1364 
1365 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1366 			   struct socket **rx_sock)
1367 {
1368 	*tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1369 	*rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1370 }
1371 
1372 static void vhost_net_flush(struct vhost_net *n)
1373 {
1374 	vhost_dev_flush(&n->dev);
1375 	if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1376 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1377 		n->tx_flush = true;
1378 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1379 		/* Wait for all lower device DMAs done. */
1380 		vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1381 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1382 		n->tx_flush = false;
1383 		atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1384 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1385 	}
1386 }
1387 
1388 static int vhost_net_release(struct inode *inode, struct file *f)
1389 {
1390 	struct vhost_net *n = f->private_data;
1391 	struct socket *tx_sock;
1392 	struct socket *rx_sock;
1393 
1394 	vhost_net_stop(n, &tx_sock, &rx_sock);
1395 	vhost_net_flush(n);
1396 	vhost_dev_stop(&n->dev);
1397 	vhost_dev_cleanup(&n->dev);
1398 	vhost_net_vq_reset(n);
1399 	if (tx_sock)
1400 		sockfd_put(tx_sock);
1401 	if (rx_sock)
1402 		sockfd_put(rx_sock);
1403 	/* Make sure no callbacks are outstanding */
1404 	synchronize_rcu();
1405 	/* We do an extra flush before freeing memory,
1406 	 * since jobs can re-queue themselves. */
1407 	vhost_net_flush(n);
1408 	kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1409 	kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1410 	kfree(n->dev.vqs);
1411 	page_frag_cache_drain(&n->pf_cache);
1412 	kvfree(n);
1413 	return 0;
1414 }
1415 
1416 static struct socket *get_raw_socket(int fd)
1417 {
1418 	int r;
1419 	struct socket *sock = sockfd_lookup(fd, &r);
1420 
1421 	if (!sock)
1422 		return ERR_PTR(-ENOTSOCK);
1423 
1424 	/* Parameter checking */
1425 	if (sock->sk->sk_type != SOCK_RAW) {
1426 		r = -ESOCKTNOSUPPORT;
1427 		goto err;
1428 	}
1429 
1430 	if (sock->sk->sk_family != AF_PACKET) {
1431 		r = -EPFNOSUPPORT;
1432 		goto err;
1433 	}
1434 	return sock;
1435 err:
1436 	sockfd_put(sock);
1437 	return ERR_PTR(r);
1438 }
1439 
1440 static struct ptr_ring *get_tap_ptr_ring(struct file *file)
1441 {
1442 	struct ptr_ring *ring;
1443 	ring = tun_get_tx_ring(file);
1444 	if (!IS_ERR(ring))
1445 		goto out;
1446 	ring = tap_get_ptr_ring(file);
1447 	if (!IS_ERR(ring))
1448 		goto out;
1449 	ring = NULL;
1450 out:
1451 	return ring;
1452 }
1453 
1454 static struct socket *get_tap_socket(int fd)
1455 {
1456 	struct file *file = fget(fd);
1457 	struct socket *sock;
1458 
1459 	if (!file)
1460 		return ERR_PTR(-EBADF);
1461 	sock = tun_get_socket(file);
1462 	if (!IS_ERR(sock))
1463 		return sock;
1464 	sock = tap_get_socket(file);
1465 	if (IS_ERR(sock))
1466 		fput(file);
1467 	return sock;
1468 }
1469 
1470 static struct socket *get_socket(int fd)
1471 {
1472 	struct socket *sock;
1473 
1474 	/* special case to disable backend */
1475 	if (fd == -1)
1476 		return NULL;
1477 	sock = get_raw_socket(fd);
1478 	if (!IS_ERR(sock))
1479 		return sock;
1480 	sock = get_tap_socket(fd);
1481 	if (!IS_ERR(sock))
1482 		return sock;
1483 	return ERR_PTR(-ENOTSOCK);
1484 }
1485 
1486 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1487 {
1488 	struct socket *sock, *oldsock;
1489 	struct vhost_virtqueue *vq;
1490 	struct vhost_net_virtqueue *nvq;
1491 	struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1492 	int r;
1493 
1494 	mutex_lock(&n->dev.mutex);
1495 	r = vhost_dev_check_owner(&n->dev);
1496 	if (r)
1497 		goto err;
1498 
1499 	if (index >= VHOST_NET_VQ_MAX) {
1500 		r = -ENOBUFS;
1501 		goto err;
1502 	}
1503 	vq = &n->vqs[index].vq;
1504 	nvq = &n->vqs[index];
1505 	mutex_lock(&vq->mutex);
1506 
1507 	if (fd == -1)
1508 		vhost_clear_msg(&n->dev);
1509 
1510 	/* Verify that ring has been setup correctly. */
1511 	if (!vhost_vq_access_ok(vq)) {
1512 		r = -EFAULT;
1513 		goto err_vq;
1514 	}
1515 	sock = get_socket(fd);
1516 	if (IS_ERR(sock)) {
1517 		r = PTR_ERR(sock);
1518 		goto err_vq;
1519 	}
1520 
1521 	/* start polling new socket */
1522 	oldsock = vhost_vq_get_backend(vq);
1523 	if (sock != oldsock) {
1524 		ubufs = vhost_net_ubuf_alloc(vq,
1525 					     sock && vhost_sock_zcopy(sock));
1526 		if (IS_ERR(ubufs)) {
1527 			r = PTR_ERR(ubufs);
1528 			goto err_ubufs;
1529 		}
1530 
1531 		vhost_net_disable_vq(n, vq);
1532 		vhost_vq_set_backend(vq, sock);
1533 		vhost_net_buf_unproduce(nvq);
1534 		r = vhost_vq_init_access(vq);
1535 		if (r)
1536 			goto err_used;
1537 		r = vhost_net_enable_vq(n, vq);
1538 		if (r)
1539 			goto err_used;
1540 		if (index == VHOST_NET_VQ_RX) {
1541 			if (sock)
1542 				nvq->rx_ring = get_tap_ptr_ring(sock->file);
1543 			else
1544 				nvq->rx_ring = NULL;
1545 		}
1546 
1547 		oldubufs = nvq->ubufs;
1548 		nvq->ubufs = ubufs;
1549 
1550 		n->tx_packets = 0;
1551 		n->tx_zcopy_err = 0;
1552 		n->tx_flush = false;
1553 	}
1554 
1555 	mutex_unlock(&vq->mutex);
1556 
1557 	if (oldubufs) {
1558 		vhost_net_ubuf_put_wait_and_free(oldubufs);
1559 		mutex_lock(&vq->mutex);
1560 		vhost_zerocopy_signal_used(n, vq);
1561 		mutex_unlock(&vq->mutex);
1562 	}
1563 
1564 	if (oldsock) {
1565 		vhost_dev_flush(&n->dev);
1566 		sockfd_put(oldsock);
1567 	}
1568 
1569 	mutex_unlock(&n->dev.mutex);
1570 	return 0;
1571 
1572 err_used:
1573 	vhost_vq_set_backend(vq, oldsock);
1574 	vhost_net_enable_vq(n, vq);
1575 	if (ubufs)
1576 		vhost_net_ubuf_put_wait_and_free(ubufs);
1577 err_ubufs:
1578 	if (sock)
1579 		sockfd_put(sock);
1580 err_vq:
1581 	mutex_unlock(&vq->mutex);
1582 err:
1583 	mutex_unlock(&n->dev.mutex);
1584 	return r;
1585 }
1586 
1587 static long vhost_net_reset_owner(struct vhost_net *n)
1588 {
1589 	struct socket *tx_sock = NULL;
1590 	struct socket *rx_sock = NULL;
1591 	long err;
1592 	struct vhost_iotlb *umem;
1593 
1594 	mutex_lock(&n->dev.mutex);
1595 	err = vhost_dev_check_owner(&n->dev);
1596 	if (err)
1597 		goto done;
1598 	umem = vhost_dev_reset_owner_prepare();
1599 	if (!umem) {
1600 		err = -ENOMEM;
1601 		goto done;
1602 	}
1603 	vhost_net_stop(n, &tx_sock, &rx_sock);
1604 	vhost_net_flush(n);
1605 	vhost_dev_stop(&n->dev);
1606 	vhost_dev_reset_owner(&n->dev, umem);
1607 	vhost_net_vq_reset(n);
1608 done:
1609 	mutex_unlock(&n->dev.mutex);
1610 	if (tx_sock)
1611 		sockfd_put(tx_sock);
1612 	if (rx_sock)
1613 		sockfd_put(rx_sock);
1614 	return err;
1615 }
1616 
1617 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1618 {
1619 	size_t vhost_hlen, sock_hlen, hdr_len;
1620 	int i;
1621 
1622 	hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1623 			       (1ULL << VIRTIO_F_VERSION_1))) ?
1624 			sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1625 			sizeof(struct virtio_net_hdr);
1626 	if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1627 		/* vhost provides vnet_hdr */
1628 		vhost_hlen = hdr_len;
1629 		sock_hlen = 0;
1630 	} else {
1631 		/* socket provides vnet_hdr */
1632 		vhost_hlen = 0;
1633 		sock_hlen = hdr_len;
1634 	}
1635 	mutex_lock(&n->dev.mutex);
1636 	if ((features & (1 << VHOST_F_LOG_ALL)) &&
1637 	    !vhost_log_access_ok(&n->dev))
1638 		goto out_unlock;
1639 
1640 	if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) {
1641 		if (vhost_init_device_iotlb(&n->dev))
1642 			goto out_unlock;
1643 	}
1644 
1645 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1646 		mutex_lock(&n->vqs[i].vq.mutex);
1647 		n->vqs[i].vq.acked_features = features;
1648 		n->vqs[i].vhost_hlen = vhost_hlen;
1649 		n->vqs[i].sock_hlen = sock_hlen;
1650 		mutex_unlock(&n->vqs[i].vq.mutex);
1651 	}
1652 	mutex_unlock(&n->dev.mutex);
1653 	return 0;
1654 
1655 out_unlock:
1656 	mutex_unlock(&n->dev.mutex);
1657 	return -EFAULT;
1658 }
1659 
1660 static long vhost_net_set_owner(struct vhost_net *n)
1661 {
1662 	int r;
1663 
1664 	mutex_lock(&n->dev.mutex);
1665 	if (vhost_dev_has_owner(&n->dev)) {
1666 		r = -EBUSY;
1667 		goto out;
1668 	}
1669 	r = vhost_net_set_ubuf_info(n);
1670 	if (r)
1671 		goto out;
1672 	r = vhost_dev_set_owner(&n->dev);
1673 	if (r)
1674 		vhost_net_clear_ubuf_info(n);
1675 	vhost_net_flush(n);
1676 out:
1677 	mutex_unlock(&n->dev.mutex);
1678 	return r;
1679 }
1680 
1681 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1682 			    unsigned long arg)
1683 {
1684 	struct vhost_net *n = f->private_data;
1685 	void __user *argp = (void __user *)arg;
1686 	u64 __user *featurep = argp;
1687 	struct vhost_vring_file backend;
1688 	u64 features;
1689 	int r;
1690 
1691 	switch (ioctl) {
1692 	case VHOST_NET_SET_BACKEND:
1693 		if (copy_from_user(&backend, argp, sizeof backend))
1694 			return -EFAULT;
1695 		return vhost_net_set_backend(n, backend.index, backend.fd);
1696 	case VHOST_GET_FEATURES:
1697 		features = VHOST_NET_FEATURES;
1698 		if (copy_to_user(featurep, &features, sizeof features))
1699 			return -EFAULT;
1700 		return 0;
1701 	case VHOST_SET_FEATURES:
1702 		if (copy_from_user(&features, featurep, sizeof features))
1703 			return -EFAULT;
1704 		if (features & ~VHOST_NET_FEATURES)
1705 			return -EOPNOTSUPP;
1706 		return vhost_net_set_features(n, features);
1707 	case VHOST_GET_BACKEND_FEATURES:
1708 		features = VHOST_NET_BACKEND_FEATURES;
1709 		if (copy_to_user(featurep, &features, sizeof(features)))
1710 			return -EFAULT;
1711 		return 0;
1712 	case VHOST_SET_BACKEND_FEATURES:
1713 		if (copy_from_user(&features, featurep, sizeof(features)))
1714 			return -EFAULT;
1715 		if (features & ~VHOST_NET_BACKEND_FEATURES)
1716 			return -EOPNOTSUPP;
1717 		vhost_set_backend_features(&n->dev, features);
1718 		return 0;
1719 	case VHOST_RESET_OWNER:
1720 		return vhost_net_reset_owner(n);
1721 	case VHOST_SET_OWNER:
1722 		return vhost_net_set_owner(n);
1723 	default:
1724 		mutex_lock(&n->dev.mutex);
1725 		r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1726 		if (r == -ENOIOCTLCMD)
1727 			r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1728 		else
1729 			vhost_net_flush(n);
1730 		mutex_unlock(&n->dev.mutex);
1731 		return r;
1732 	}
1733 }
1734 
1735 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1736 {
1737 	struct file *file = iocb->ki_filp;
1738 	struct vhost_net *n = file->private_data;
1739 	struct vhost_dev *dev = &n->dev;
1740 	int noblock = file->f_flags & O_NONBLOCK;
1741 
1742 	return vhost_chr_read_iter(dev, to, noblock);
1743 }
1744 
1745 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1746 					struct iov_iter *from)
1747 {
1748 	struct file *file = iocb->ki_filp;
1749 	struct vhost_net *n = file->private_data;
1750 	struct vhost_dev *dev = &n->dev;
1751 
1752 	return vhost_chr_write_iter(dev, from);
1753 }
1754 
1755 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1756 {
1757 	struct vhost_net *n = file->private_data;
1758 	struct vhost_dev *dev = &n->dev;
1759 
1760 	return vhost_chr_poll(file, dev, wait);
1761 }
1762 
1763 static const struct file_operations vhost_net_fops = {
1764 	.owner          = THIS_MODULE,
1765 	.release        = vhost_net_release,
1766 	.read_iter      = vhost_net_chr_read_iter,
1767 	.write_iter     = vhost_net_chr_write_iter,
1768 	.poll           = vhost_net_chr_poll,
1769 	.unlocked_ioctl = vhost_net_ioctl,
1770 	.compat_ioctl   = compat_ptr_ioctl,
1771 	.open           = vhost_net_open,
1772 	.llseek		= noop_llseek,
1773 };
1774 
1775 static struct miscdevice vhost_net_misc = {
1776 	.minor = VHOST_NET_MINOR,
1777 	.name = "vhost-net",
1778 	.fops = &vhost_net_fops,
1779 };
1780 
1781 static int __init vhost_net_init(void)
1782 {
1783 	if (experimental_zcopytx)
1784 		vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1785 	return misc_register(&vhost_net_misc);
1786 }
1787 module_init(vhost_net_init);
1788 
1789 static void __exit vhost_net_exit(void)
1790 {
1791 	misc_deregister(&vhost_net_misc);
1792 }
1793 module_exit(vhost_net_exit);
1794 
1795 MODULE_VERSION("0.0.1");
1796 MODULE_LICENSE("GPL v2");
1797 MODULE_AUTHOR("Michael S. Tsirkin");
1798 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1799 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1800 MODULE_ALIAS("devname:vhost-net");
1801