virtio_transport.c

  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * virtio transport for vsock
  4 *
  5 * Copyright (C) 2013-2015 Red Hat, Inc.
  6 * Author: Asias He <asias@redhat.com>
  7 *         Stefan Hajnoczi <stefanha@redhat.com>
  8 *
  9 * Some of the code is take from Gerd Hoffmann <kraxel@redhat.com>'s
 10 * early virtio-vsock proof-of-concept bits.
 11 */
 12#include <linux/spinlock.h>
 13#include <linux/module.h>
 14#include <linux/list.h>
 15#include <linux/atomic.h>
 16#include <linux/virtio.h>
 17#include <linux/virtio_ids.h>
 18#include <linux/virtio_config.h>
 19#include <linux/virtio_vsock.h>
 20#include <net/sock.h>
 21#include <linux/mutex.h>
 22#include <net/af_vsock.h>
 23
 24static struct workqueue_struct *virtio_vsock_workqueue;
 25static struct virtio_vsock __rcu *the_virtio_vsock;
 26static DEFINE_MUTEX(the_virtio_vsock_mutex); /* protects the_virtio_vsock */
 27static struct virtio_transport virtio_transport; /* forward declaration */
 28
 29struct virtio_vsock {
 30	struct virtio_device *vdev;
 31	struct virtqueue *vqs[VSOCK_VQ_MAX];
 32
 33	/* Virtqueue processing is deferred to a workqueue */
 34	struct work_struct tx_work;
 35	struct work_struct rx_work;
 36	struct work_struct event_work;
 37
 38	/* The following fields are protected by tx_lock.  vqs[VSOCK_VQ_TX]
 39	 * must be accessed with tx_lock held.
 40	 */
 41	struct mutex tx_lock;
 42	bool tx_run;
 43
 44	struct work_struct send_pkt_work;
 45	struct sk_buff_head send_pkt_queue;
 46
 47	atomic_t queued_replies;
 48
 49	/* The following fields are protected by rx_lock.  vqs[VSOCK_VQ_RX]
 50	 * must be accessed with rx_lock held.
 51	 */
 52	struct mutex rx_lock;
 53	bool rx_run;
 54	int rx_buf_nr;
 55	int rx_buf_max_nr;
 56
 57	/* The following fields are protected by event_lock.
 58	 * vqs[VSOCK_VQ_EVENT] must be accessed with event_lock held.
 59	 */
 60	struct mutex event_lock;
 61	bool event_run;
 62	struct virtio_vsock_event event_list[8];
 63
 64	u32 guest_cid;
 65	bool seqpacket_allow;
 66
 67	/* These fields are used only in tx path in function
 68	 * 'virtio_transport_send_pkt_work()', so to save
 69	 * stack space in it, place both of them here. Each
 70	 * pointer from 'out_sgs' points to the corresponding
 71	 * element in 'out_bufs' - this is initialized in
 72	 * 'virtio_vsock_probe()'. Both fields are protected
 73	 * by 'tx_lock'. +1 is needed for packet header.
 74	 */
 75	struct scatterlist *out_sgs[MAX_SKB_FRAGS + 1];
 76	struct scatterlist out_bufs[MAX_SKB_FRAGS + 1];
 77};
 78
 79static u32 virtio_transport_get_local_cid(void)
 80{
 81	struct virtio_vsock *vsock;
 82	u32 ret;
 83
 84	rcu_read_lock();
 85	vsock = rcu_dereference(the_virtio_vsock);
 86	if (!vsock) {
 87		ret = VMADDR_CID_ANY;
 88		goto out_rcu;
 89	}
 90
 91	ret = vsock->guest_cid;
 92out_rcu:
 93	rcu_read_unlock();
 94	return ret;
 95}
 96
 97/* Caller need to hold vsock->tx_lock on vq */
 98static int virtio_transport_send_skb(struct sk_buff *skb, struct virtqueue *vq,
 99				     struct virtio_vsock *vsock, gfp_t gfp)
100{
101	int ret, in_sg = 0, out_sg = 0;
102	struct scatterlist **sgs;
103
104	sgs = vsock->out_sgs;
105	sg_init_one(sgs[out_sg], virtio_vsock_hdr(skb),
106		    sizeof(*virtio_vsock_hdr(skb)));
107	out_sg++;
108
109	if (!skb_is_nonlinear(skb)) {
110		if (skb->len > 0) {
111			sg_init_one(sgs[out_sg], skb->data, skb->len);
112			out_sg++;
113		}
114	} else {
115		struct skb_shared_info *si;
116		int i;
117
118		/* If skb is nonlinear, then its buffer must contain
119		 * only header and nothing more. Data is stored in
120		 * the fragged part.
121		 */
122		WARN_ON_ONCE(skb_headroom(skb) != sizeof(*virtio_vsock_hdr(skb)));
123
124		si = skb_shinfo(skb);
125
126		for (i = 0; i < si->nr_frags; i++) {
127			skb_frag_t *skb_frag = &si->frags[i];
128			void *va;
129
130			/* We will use 'page_to_virt()' for the userspace page
131			 * here, because virtio or dma-mapping layers will call
132			 * 'virt_to_phys()' later to fill the buffer descriptor.
133			 * We don't touch memory at "virtual" address of this page.
134			 */
135			va = page_to_virt(skb_frag_page(skb_frag));
136			sg_init_one(sgs[out_sg],
137				    va + skb_frag_off(skb_frag),
138				    skb_frag_size(skb_frag));
139			out_sg++;
140		}
141	}
142
143	ret = virtqueue_add_sgs(vq, sgs, out_sg, in_sg, skb, gfp);
144	/* Usually this means that there is no more space available in
145	 * the vq
146	 */
147	if (ret < 0)
148		return ret;
149
150	virtio_transport_deliver_tap_pkt(skb);
151	return 0;
152}
153
154static void
155virtio_transport_send_pkt_work(struct work_struct *work)
156{
157	struct virtio_vsock *vsock =
158		container_of(work, struct virtio_vsock, send_pkt_work);
159	struct virtqueue *vq;
160	bool added = false;
161	bool restart_rx = false;
162
163	mutex_lock(&vsock->tx_lock);
164
165	if (!vsock->tx_run)
166		goto out;
167
168	vq = vsock->vqs[VSOCK_VQ_TX];
169
170	for (;;) {
171		struct sk_buff *skb;
172		bool reply;
173		int ret;
174
175		skb = virtio_vsock_skb_dequeue(&vsock->send_pkt_queue);
176		if (!skb)
177			break;
178
179		reply = virtio_vsock_skb_reply(skb);
180
181		ret = virtio_transport_send_skb(skb, vq, vsock, GFP_KERNEL);
182		if (ret < 0) {
183			virtio_vsock_skb_queue_head(&vsock->send_pkt_queue, skb);
184			break;
185		}
186
187		if (reply) {
188			struct virtqueue *rx_vq = vsock->vqs[VSOCK_VQ_RX];
189			int val;
190
191			val = atomic_dec_return(&vsock->queued_replies);
192
193			/* Do we now have resources to resume rx processing? */
194			if (val + 1 == virtqueue_get_vring_size(rx_vq))
195				restart_rx = true;
196		}
197
198		added = true;
199	}
200
201	if (added)
202		virtqueue_kick(vq);
203
204out:
205	mutex_unlock(&vsock->tx_lock);
206
207	if (restart_rx)
208		queue_work(virtio_vsock_workqueue, &vsock->rx_work);
209}
210
211/* Caller need to hold RCU for vsock.
212 * Returns 0 if the packet is successfully put on the vq.
213 */
214static int virtio_transport_send_skb_fast_path(struct virtio_vsock *vsock, struct sk_buff *skb)
215{
216	struct virtqueue *vq = vsock->vqs[VSOCK_VQ_TX];
217	int ret;
218
219	/* Inside RCU, can't sleep! */
220	ret = mutex_trylock(&vsock->tx_lock);
221	if (unlikely(ret == 0))
222		return -EBUSY;
223
224	ret = virtio_transport_send_skb(skb, vq, vsock, GFP_ATOMIC);
225	if (ret == 0)
226		virtqueue_kick(vq);
227
228	mutex_unlock(&vsock->tx_lock);
229
230	return ret;
231}
232
233static int
234virtio_transport_send_pkt(struct sk_buff *skb)
235{
236	struct virtio_vsock_hdr *hdr;
237	struct virtio_vsock *vsock;
238	int len = skb->len;
239
240	hdr = virtio_vsock_hdr(skb);
241
242	rcu_read_lock();
243	vsock = rcu_dereference(the_virtio_vsock);
244	if (!vsock) {
245		kfree_skb(skb);
246		len = -ENODEV;
247		goto out_rcu;
248	}
249
250	if (le64_to_cpu(hdr->dst_cid) == vsock->guest_cid) {
251		kfree_skb(skb);
252		len = -ENODEV;
253		goto out_rcu;
254	}
255
256	/* If send_pkt_queue is empty, we can safely bypass this queue
257	 * because packet order is maintained and (try) to put the packet
258	 * on the virtqueue using virtio_transport_send_skb_fast_path.
259	 * If this fails we simply put the packet on the intermediate
260	 * queue and schedule the worker.
261	 */
262	if (!skb_queue_empty_lockless(&vsock->send_pkt_queue) ||
263	    virtio_transport_send_skb_fast_path(vsock, skb)) {
264		if (virtio_vsock_skb_reply(skb))
265			atomic_inc(&vsock->queued_replies);
266
267		virtio_vsock_skb_queue_tail(&vsock->send_pkt_queue, skb);
268		queue_work(virtio_vsock_workqueue, &vsock->send_pkt_work);
269	}
270
271out_rcu:
272	rcu_read_unlock();
273	return len;
274}
275
276static int
277virtio_transport_cancel_pkt(struct vsock_sock *vsk)
278{
279	struct virtio_vsock *vsock;
280	int cnt = 0, ret;
281
282	rcu_read_lock();
283	vsock = rcu_dereference(the_virtio_vsock);
284	if (!vsock) {
285		ret = -ENODEV;
286		goto out_rcu;
287	}
288
289	cnt = virtio_transport_purge_skbs(vsk, &vsock->send_pkt_queue);
290
291	if (cnt) {
292		struct virtqueue *rx_vq = vsock->vqs[VSOCK_VQ_RX];
293		int new_cnt;
294
295		new_cnt = atomic_sub_return(cnt, &vsock->queued_replies);
296		if (new_cnt + cnt >= virtqueue_get_vring_size(rx_vq) &&
297		    new_cnt < virtqueue_get_vring_size(rx_vq))
298			queue_work(virtio_vsock_workqueue, &vsock->rx_work);
299	}
300
301	ret = 0;
302
303out_rcu:
304	rcu_read_unlock();
305	return ret;
306}
307
308static void virtio_vsock_rx_fill(struct virtio_vsock *vsock)
309{
310	int total_len = VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE;
311	struct scatterlist pkt, *p;
312	struct virtqueue *vq;
313	struct sk_buff *skb;
314	int ret;
315
316	vq = vsock->vqs[VSOCK_VQ_RX];
317
318	do {
319		skb = virtio_vsock_alloc_linear_skb(total_len, GFP_KERNEL);
320		if (!skb)
321			break;
322
323		memset(skb->head, 0, VIRTIO_VSOCK_SKB_HEADROOM);
324		sg_init_one(&pkt, virtio_vsock_hdr(skb), total_len);
325		p = &pkt;
326		ret = virtqueue_add_sgs(vq, &p, 0, 1, skb, GFP_KERNEL);
327		if (ret < 0) {
328			kfree_skb(skb);
329			break;
330		}
331
332		vsock->rx_buf_nr++;
333	} while (vq->num_free);
334	if (vsock->rx_buf_nr > vsock->rx_buf_max_nr)
335		vsock->rx_buf_max_nr = vsock->rx_buf_nr;
336	virtqueue_kick(vq);
337}
338
339static void virtio_transport_tx_work(struct work_struct *work)
340{
341	struct virtio_vsock *vsock =
342		container_of(work, struct virtio_vsock, tx_work);
343	struct virtqueue *vq;
344	bool added = false;
345
346	vq = vsock->vqs[VSOCK_VQ_TX];
347	mutex_lock(&vsock->tx_lock);
348
349	if (!vsock->tx_run)
350		goto out;
351
352	do {
353		struct sk_buff *skb;
354		unsigned int len;
355
356		virtqueue_disable_cb(vq);
357		while ((skb = virtqueue_get_buf(vq, &len)) != NULL) {
358			virtio_transport_consume_skb_sent(skb, true);
359			added = true;
360		}
361	} while (!virtqueue_enable_cb(vq));
362
363out:
364	mutex_unlock(&vsock->tx_lock);
365
366	if (added)
367		queue_work(virtio_vsock_workqueue, &vsock->send_pkt_work);
368}
369
370/* Is there space left for replies to rx packets? */
371static bool virtio_transport_more_replies(struct virtio_vsock *vsock)
372{
373	struct virtqueue *vq = vsock->vqs[VSOCK_VQ_RX];
374	int val;
375
376	smp_rmb(); /* paired with atomic_inc() and atomic_dec_return() */
377	val = atomic_read(&vsock->queued_replies);
378
379	return val < virtqueue_get_vring_size(vq);
380}
381
382/* event_lock must be held */
383static int virtio_vsock_event_fill_one(struct virtio_vsock *vsock,
384				       struct virtio_vsock_event *event)
385{
386	struct scatterlist sg;
387	struct virtqueue *vq;
388
389	vq = vsock->vqs[VSOCK_VQ_EVENT];
390
391	sg_init_one(&sg, event, sizeof(*event));
392
393	return virtqueue_add_inbuf(vq, &sg, 1, event, GFP_KERNEL);
394}
395
396/* event_lock must be held */
397static void virtio_vsock_event_fill(struct virtio_vsock *vsock)
398{
399	size_t i;
400
401	for (i = 0; i < ARRAY_SIZE(vsock->event_list); i++) {
402		struct virtio_vsock_event *event = &vsock->event_list[i];
403
404		virtio_vsock_event_fill_one(vsock, event);
405	}
406
407	virtqueue_kick(vsock->vqs[VSOCK_VQ_EVENT]);
408}
409
410static void virtio_vsock_reset_sock(struct sock *sk)
411{
412	/* vmci_transport.c doesn't take sk_lock here either.  At least we're
413	 * under vsock_table_lock so the sock cannot disappear while we're
414	 * executing.
415	 */
416
417	sk->sk_state = TCP_CLOSE;
418	sk->sk_err = ECONNRESET;
419	sk_error_report(sk);
420}
421
422static void virtio_vsock_update_guest_cid(struct virtio_vsock *vsock)
423{
424	struct virtio_device *vdev = vsock->vdev;
425	__le64 guest_cid;
426
427	vdev->config->get(vdev, offsetof(struct virtio_vsock_config, guest_cid),
428			  &guest_cid, sizeof(guest_cid));
429	vsock->guest_cid = le64_to_cpu(guest_cid);
430}
431
432/* event_lock must be held */
433static void virtio_vsock_event_handle(struct virtio_vsock *vsock,
434				      struct virtio_vsock_event *event)
435{
436	switch (le32_to_cpu(event->id)) {
437	case VIRTIO_VSOCK_EVENT_TRANSPORT_RESET:
438		virtio_vsock_update_guest_cid(vsock);
439		vsock_for_each_connected_socket(&virtio_transport.transport,
440						virtio_vsock_reset_sock);
441		break;
442	}
443}
444
445static void virtio_transport_event_work(struct work_struct *work)
446{
447	struct virtio_vsock *vsock =
448		container_of(work, struct virtio_vsock, event_work);
449	struct virtqueue *vq;
450
451	vq = vsock->vqs[VSOCK_VQ_EVENT];
452
453	mutex_lock(&vsock->event_lock);
454
455	if (!vsock->event_run)
456		goto out;
457
458	do {
459		struct virtio_vsock_event *event;
460		unsigned int len;
461
462		virtqueue_disable_cb(vq);
463		while ((event = virtqueue_get_buf(vq, &len)) != NULL) {
464			if (len == sizeof(*event))
465				virtio_vsock_event_handle(vsock, event);
466
467			virtio_vsock_event_fill_one(vsock, event);
468		}
469	} while (!virtqueue_enable_cb(vq));
470
471	virtqueue_kick(vsock->vqs[VSOCK_VQ_EVENT]);
472out:
473	mutex_unlock(&vsock->event_lock);
474}
475
476static void virtio_vsock_event_done(struct virtqueue *vq)
477{
478	struct virtio_vsock *vsock = vq->vdev->priv;
479
480	if (!vsock)
481		return;
482	queue_work(virtio_vsock_workqueue, &vsock->event_work);
483}
484
485static void virtio_vsock_tx_done(struct virtqueue *vq)
486{
487	struct virtio_vsock *vsock = vq->vdev->priv;
488
489	if (!vsock)
490		return;
491	queue_work(virtio_vsock_workqueue, &vsock->tx_work);
492}
493
494static void virtio_vsock_rx_done(struct virtqueue *vq)
495{
496	struct virtio_vsock *vsock = vq->vdev->priv;
497
498	if (!vsock)
499		return;
500	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
501}
502
503static bool virtio_transport_can_msgzerocopy(int bufs_num)
504{
505	struct virtio_vsock *vsock;
506	bool res = false;
507
508	rcu_read_lock();
509
510	vsock = rcu_dereference(the_virtio_vsock);
511	if (vsock) {
512		struct virtqueue *vq = vsock->vqs[VSOCK_VQ_TX];
513
514		/* Check that tx queue is large enough to keep whole
515		 * data to send. This is needed, because when there is
516		 * not enough free space in the queue, current skb to
517		 * send will be reinserted to the head of tx list of
518		 * the socket to retry transmission later, so if skb
519		 * is bigger than whole queue, it will be reinserted
520		 * again and again, thus blocking other skbs to be sent.
521		 * Each page of the user provided buffer will be added
522		 * as a single buffer to the tx virtqueue, so compare
523		 * number of pages against maximum capacity of the queue.
524		 */
525		if (bufs_num <= vq->num_max)
526			res = true;
527	}
528
529	rcu_read_unlock();
530
531	return res;
532}
533
534static bool virtio_transport_msgzerocopy_allow(void)
535{
536	return true;
537}
538
539static bool virtio_transport_seqpacket_allow(u32 remote_cid);
540
541static struct virtio_transport virtio_transport = {
542	.transport = {
543		.module                   = THIS_MODULE,
544
545		.get_local_cid            = virtio_transport_get_local_cid,
546
547		.init                     = virtio_transport_do_socket_init,
548		.destruct                 = virtio_transport_destruct,
549		.release                  = virtio_transport_release,
550		.connect                  = virtio_transport_connect,
551		.shutdown                 = virtio_transport_shutdown,
552		.cancel_pkt               = virtio_transport_cancel_pkt,
553
554		.dgram_bind               = virtio_transport_dgram_bind,
555		.dgram_dequeue            = virtio_transport_dgram_dequeue,
556		.dgram_enqueue            = virtio_transport_dgram_enqueue,
557		.dgram_allow              = virtio_transport_dgram_allow,
558
559		.stream_dequeue           = virtio_transport_stream_dequeue,
560		.stream_enqueue           = virtio_transport_stream_enqueue,
561		.stream_has_data          = virtio_transport_stream_has_data,
562		.stream_has_space         = virtio_transport_stream_has_space,
563		.stream_rcvhiwat          = virtio_transport_stream_rcvhiwat,
564		.stream_is_active         = virtio_transport_stream_is_active,
565		.stream_allow             = virtio_transport_stream_allow,
566
567		.seqpacket_dequeue        = virtio_transport_seqpacket_dequeue,
568		.seqpacket_enqueue        = virtio_transport_seqpacket_enqueue,
569		.seqpacket_allow          = virtio_transport_seqpacket_allow,
570		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
571
572		.msgzerocopy_allow        = virtio_transport_msgzerocopy_allow,
573
574		.notify_poll_in           = virtio_transport_notify_poll_in,
575		.notify_poll_out          = virtio_transport_notify_poll_out,
576		.notify_recv_init         = virtio_transport_notify_recv_init,
577		.notify_recv_pre_block    = virtio_transport_notify_recv_pre_block,
578		.notify_recv_pre_dequeue  = virtio_transport_notify_recv_pre_dequeue,
579		.notify_recv_post_dequeue = virtio_transport_notify_recv_post_dequeue,
580		.notify_send_init         = virtio_transport_notify_send_init,
581		.notify_send_pre_block    = virtio_transport_notify_send_pre_block,
582		.notify_send_pre_enqueue  = virtio_transport_notify_send_pre_enqueue,
583		.notify_send_post_enqueue = virtio_transport_notify_send_post_enqueue,
584		.notify_buffer_size       = virtio_transport_notify_buffer_size,
585		.notify_set_rcvlowat      = virtio_transport_notify_set_rcvlowat,
586
587		.unsent_bytes             = virtio_transport_unsent_bytes,
588
589		.read_skb = virtio_transport_read_skb,
590	},
591
592	.send_pkt = virtio_transport_send_pkt,
593	.can_msgzerocopy = virtio_transport_can_msgzerocopy,
594};
595
596static bool virtio_transport_seqpacket_allow(u32 remote_cid)
597{
598	struct virtio_vsock *vsock;
599	bool seqpacket_allow;
600
601	seqpacket_allow = false;
602	rcu_read_lock();
603	vsock = rcu_dereference(the_virtio_vsock);
604	if (vsock)
605		seqpacket_allow = vsock->seqpacket_allow;
606	rcu_read_unlock();
607
608	return seqpacket_allow;
609}
610
611static void virtio_transport_rx_work(struct work_struct *work)
612{
613	struct virtio_vsock *vsock =
614		container_of(work, struct virtio_vsock, rx_work);
615	struct virtqueue *vq;
616
617	vq = vsock->vqs[VSOCK_VQ_RX];
618
619	mutex_lock(&vsock->rx_lock);
620
621	if (!vsock->rx_run)
622		goto out;
623
624	do {
625		virtqueue_disable_cb(vq);
626		for (;;) {
627			unsigned int len, payload_len;
628			struct virtio_vsock_hdr *hdr;
629			struct sk_buff *skb;
630
631			if (!virtio_transport_more_replies(vsock)) {
632				/* Stop rx until the device processes already
633				 * pending replies.  Leave rx virtqueue
634				 * callbacks disabled.
635				 */
636				goto out;
637			}
638
639			skb = virtqueue_get_buf(vq, &len);
640			if (!skb)
641				break;
642
643			vsock->rx_buf_nr--;
644
645			/* Drop short/long packets */
646			if (unlikely(len < sizeof(*hdr) ||
647				     len > virtio_vsock_skb_len(skb))) {
648				kfree_skb(skb);
649				continue;
650			}
651
652			hdr = virtio_vsock_hdr(skb);
653			payload_len = le32_to_cpu(hdr->len);
654			if (unlikely(payload_len > len - sizeof(*hdr))) {
655				kfree_skb(skb);
656				continue;
657			}
658
659			if (payload_len)
660				virtio_vsock_skb_put(skb, payload_len);
661
662			virtio_transport_deliver_tap_pkt(skb);
663			virtio_transport_recv_pkt(&virtio_transport, skb);
664		}
665	} while (!virtqueue_enable_cb(vq));
666
667out:
668	if (vsock->rx_buf_nr < vsock->rx_buf_max_nr / 2)
669		virtio_vsock_rx_fill(vsock);
670	mutex_unlock(&vsock->rx_lock);
671}
672
673static int virtio_vsock_vqs_init(struct virtio_vsock *vsock)
674{
675	struct virtio_device *vdev = vsock->vdev;
676	struct virtqueue_info vqs_info[] = {
677		{ "rx", virtio_vsock_rx_done },
678		{ "tx", virtio_vsock_tx_done },
679		{ "event", virtio_vsock_event_done },
680	};
681	int ret;
682
683	mutex_lock(&vsock->rx_lock);
684	vsock->rx_buf_nr = 0;
685	vsock->rx_buf_max_nr = 0;
686	mutex_unlock(&vsock->rx_lock);
687
688	atomic_set(&vsock->queued_replies, 0);
689
690	ret = virtio_find_vqs(vdev, VSOCK_VQ_MAX, vsock->vqs, vqs_info, NULL);
691	if (ret < 0)
692		return ret;
693
694	virtio_vsock_update_guest_cid(vsock);
695
696	virtio_device_ready(vdev);
697
698	return 0;
699}
700
701static void virtio_vsock_vqs_start(struct virtio_vsock *vsock)
702{
703	mutex_lock(&vsock->tx_lock);
704	vsock->tx_run = true;
705	mutex_unlock(&vsock->tx_lock);
706
707	mutex_lock(&vsock->rx_lock);
708	virtio_vsock_rx_fill(vsock);
709	vsock->rx_run = true;
710	mutex_unlock(&vsock->rx_lock);
711
712	mutex_lock(&vsock->event_lock);
713	virtio_vsock_event_fill(vsock);
714	vsock->event_run = true;
715	mutex_unlock(&vsock->event_lock);
716
717	/* virtio_transport_send_pkt() can queue packets once
718	 * the_virtio_vsock is set, but they won't be processed until
719	 * vsock->tx_run is set to true. We queue vsock->send_pkt_work
720	 * when initialization finishes to send those packets queued
721	 * earlier.
722	 * We don't need to queue the other workers (rx, event) because
723	 * as long as we don't fill the queues with empty buffers, the
724	 * host can't send us any notification.
725	 */
726	queue_work(virtio_vsock_workqueue, &vsock->send_pkt_work);
727}
728
729static void virtio_vsock_vqs_del(struct virtio_vsock *vsock)
730{
731	struct virtio_device *vdev = vsock->vdev;
732	struct sk_buff *skb;
733
734	/* Reset all connected sockets when the VQs disappear */
735	vsock_for_each_connected_socket(&virtio_transport.transport,
736					virtio_vsock_reset_sock);
737
738	/* Stop all work handlers to make sure no one is accessing the device,
739	 * so we can safely call virtio_reset_device().
740	 */
741	mutex_lock(&vsock->rx_lock);
742	vsock->rx_run = false;
743	mutex_unlock(&vsock->rx_lock);
744
745	mutex_lock(&vsock->tx_lock);
746	vsock->tx_run = false;
747	mutex_unlock(&vsock->tx_lock);
748
749	mutex_lock(&vsock->event_lock);
750	vsock->event_run = false;
751	mutex_unlock(&vsock->event_lock);
752
753	/* Flush all device writes and interrupts, device will not use any
754	 * more buffers.
755	 */
756	virtio_reset_device(vdev);
757
758	mutex_lock(&vsock->rx_lock);
759	while ((skb = virtqueue_detach_unused_buf(vsock->vqs[VSOCK_VQ_RX])))
760		kfree_skb(skb);
761	mutex_unlock(&vsock->rx_lock);
762
763	mutex_lock(&vsock->tx_lock);
764	while ((skb = virtqueue_detach_unused_buf(vsock->vqs[VSOCK_VQ_TX])))
765		kfree_skb(skb);
766	mutex_unlock(&vsock->tx_lock);
767
768	virtio_vsock_skb_queue_purge(&vsock->send_pkt_queue);
769
770	/* Delete virtqueues and flush outstanding callbacks if any */
771	vdev->config->del_vqs(vdev);
772}
773
774static int virtio_vsock_probe(struct virtio_device *vdev)
775{
776	struct virtio_vsock *vsock = NULL;
777	int ret;
778	int i;
779
780	ret = mutex_lock_interruptible(&the_virtio_vsock_mutex);
781	if (ret)
782		return ret;
783
784	/* Only one virtio-vsock device per guest is supported */
785	if (rcu_dereference_protected(the_virtio_vsock,
786				lockdep_is_held(&the_virtio_vsock_mutex))) {
787		ret = -EBUSY;
788		goto out;
789	}
790
791	vsock = kzalloc(sizeof(*vsock), GFP_KERNEL);
792	if (!vsock) {
793		ret = -ENOMEM;
794		goto out;
795	}
796
797	vsock->vdev = vdev;
798
799
800	mutex_init(&vsock->tx_lock);
801	mutex_init(&vsock->rx_lock);
802	mutex_init(&vsock->event_lock);
803	skb_queue_head_init(&vsock->send_pkt_queue);
804	INIT_WORK(&vsock->rx_work, virtio_transport_rx_work);
805	INIT_WORK(&vsock->tx_work, virtio_transport_tx_work);
806	INIT_WORK(&vsock->event_work, virtio_transport_event_work);
807	INIT_WORK(&vsock->send_pkt_work, virtio_transport_send_pkt_work);
808
809	if (virtio_has_feature(vdev, VIRTIO_VSOCK_F_SEQPACKET))
810		vsock->seqpacket_allow = true;
811
812	vdev->priv = vsock;
813
814	ret = virtio_vsock_vqs_init(vsock);
815	if (ret < 0)
816		goto out;
817
818	for (i = 0; i < ARRAY_SIZE(vsock->out_sgs); i++)
819		vsock->out_sgs[i] = &vsock->out_bufs[i];
820
821	rcu_assign_pointer(the_virtio_vsock, vsock);
822	virtio_vsock_vqs_start(vsock);
823
824	mutex_unlock(&the_virtio_vsock_mutex);
825
826	return 0;
827
828out:
829	kfree(vsock);
830	mutex_unlock(&the_virtio_vsock_mutex);
831	return ret;
832}
833
834static void virtio_vsock_remove(struct virtio_device *vdev)
835{
836	struct virtio_vsock *vsock = vdev->priv;
837
838	mutex_lock(&the_virtio_vsock_mutex);
839
840	vdev->priv = NULL;
841	rcu_assign_pointer(the_virtio_vsock, NULL);
842	synchronize_rcu();
843
844	virtio_vsock_vqs_del(vsock);
845
846	/* Other works can be queued before 'config->del_vqs()', so we flush
847	 * all works before to free the vsock object to avoid use after free.
848	 */
849	flush_work(&vsock->rx_work);
850	flush_work(&vsock->tx_work);
851	flush_work(&vsock->event_work);
852	flush_work(&vsock->send_pkt_work);
853
854	mutex_unlock(&the_virtio_vsock_mutex);
855
856	kfree(vsock);
857}
858
859#ifdef CONFIG_PM_SLEEP
860static int virtio_vsock_freeze(struct virtio_device *vdev)
861{
862	struct virtio_vsock *vsock = vdev->priv;
863
864	mutex_lock(&the_virtio_vsock_mutex);
865
866	rcu_assign_pointer(the_virtio_vsock, NULL);
867	synchronize_rcu();
868
869	virtio_vsock_vqs_del(vsock);
870
871	mutex_unlock(&the_virtio_vsock_mutex);
872
873	return 0;
874}
875
876static int virtio_vsock_restore(struct virtio_device *vdev)
877{
878	struct virtio_vsock *vsock = vdev->priv;
879	int ret;
880
881	mutex_lock(&the_virtio_vsock_mutex);
882
883	/* Only one virtio-vsock device per guest is supported */
884	if (rcu_dereference_protected(the_virtio_vsock,
885				lockdep_is_held(&the_virtio_vsock_mutex))) {
886		ret = -EBUSY;
887		goto out;
888	}
889
890	ret = virtio_vsock_vqs_init(vsock);
891	if (ret < 0)
892		goto out;
893
894	rcu_assign_pointer(the_virtio_vsock, vsock);
895	virtio_vsock_vqs_start(vsock);
896
897out:
898	mutex_unlock(&the_virtio_vsock_mutex);
899	return ret;
900}
901#endif /* CONFIG_PM_SLEEP */
902
903static struct virtio_device_id id_table[] = {
904	{ VIRTIO_ID_VSOCK, VIRTIO_DEV_ANY_ID },
905	{ 0 },
906};
907
908static unsigned int features[] = {
909	VIRTIO_VSOCK_F_SEQPACKET
910};
911
912static struct virtio_driver virtio_vsock_driver = {
913	.feature_table = features,
914	.feature_table_size = ARRAY_SIZE(features),
915	.driver.name = KBUILD_MODNAME,
916	.id_table = id_table,
917	.probe = virtio_vsock_probe,
918	.remove = virtio_vsock_remove,
919#ifdef CONFIG_PM_SLEEP
920	.freeze = virtio_vsock_freeze,
921	.restore = virtio_vsock_restore,
922#endif
923};
924
925static int __init virtio_vsock_init(void)
926{
927	int ret;
928
929	virtio_vsock_workqueue = alloc_workqueue("virtio_vsock", WQ_PERCPU, 0);
930	if (!virtio_vsock_workqueue)
931		return -ENOMEM;
932
933	ret = vsock_core_register(&virtio_transport.transport,
934				  VSOCK_TRANSPORT_F_G2H);
935	if (ret)
936		goto out_wq;
937
938	ret = register_virtio_driver(&virtio_vsock_driver);
939	if (ret)
940		goto out_vci;
941
942	return 0;
943
944out_vci:
945	vsock_core_unregister(&virtio_transport.transport);
946out_wq:
947	destroy_workqueue(virtio_vsock_workqueue);
948	return ret;
949}
950
951static void __exit virtio_vsock_exit(void)
952{
953	unregister_virtio_driver(&virtio_vsock_driver);
954	vsock_core_unregister(&virtio_transport.transport);
955	destroy_workqueue(virtio_vsock_workqueue);
956}
957
958module_init(virtio_vsock_init);
959module_exit(virtio_vsock_exit);
960MODULE_LICENSE("GPL v2");
961MODULE_AUTHOR("Asias He");
962MODULE_DESCRIPTION("virtio transport for vsock");
963MODULE_DEVICE_TABLE(virtio, id_table);