xref: /freebsd/usr.sbin/bhyve/pci_virtio_net.c (revision 1682a3ab4b725e31e9aab4aa4a8aa2ba400463fe)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2011 NetApp, Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * $FreeBSD$
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include <sys/param.h>
35 #ifndef WITHOUT_CAPSICUM
36 #include <sys/capsicum.h>
37 #endif
38 #include <sys/linker_set.h>
39 #include <sys/select.h>
40 #include <sys/uio.h>
41 #include <sys/ioctl.h>
42 #include <net/ethernet.h>
43 #ifndef NETMAP_WITH_LIBS
44 #define NETMAP_WITH_LIBS
45 #endif
46 #include <net/netmap_user.h>
47 
48 #ifndef WITHOUT_CAPSICUM
49 #include <capsicum_helpers.h>
50 #endif
51 #include <err.h>
52 #include <errno.h>
53 #include <fcntl.h>
54 #include <stdio.h>
55 #include <stdlib.h>
56 #include <stdint.h>
57 #include <string.h>
58 #include <strings.h>
59 #include <unistd.h>
60 #include <assert.h>
61 #include <md5.h>
62 #include <pthread.h>
63 #include <pthread_np.h>
64 #include <sysexits.h>
65 
66 #include "bhyverun.h"
67 #include "pci_emul.h"
68 #include "mevent.h"
69 #include "virtio.h"
70 #include "net_utils.h"
71 
72 #define VTNET_RINGSZ	1024
73 
74 #define VTNET_MAXSEGS	256
75 
76 /*
77  * Host capabilities.  Note that we only offer a few of these.
78  */
79 #define	VIRTIO_NET_F_CSUM	(1 <<  0) /* host handles partial cksum */
80 #define	VIRTIO_NET_F_GUEST_CSUM	(1 <<  1) /* guest handles partial cksum */
81 #define	VIRTIO_NET_F_MAC	(1 <<  5) /* host supplies MAC */
82 #define	VIRTIO_NET_F_GSO_DEPREC	(1 <<  6) /* deprecated: host handles GSO */
83 #define	VIRTIO_NET_F_GUEST_TSO4	(1 <<  7) /* guest can rcv TSOv4 */
84 #define	VIRTIO_NET_F_GUEST_TSO6	(1 <<  8) /* guest can rcv TSOv6 */
85 #define	VIRTIO_NET_F_GUEST_ECN	(1 <<  9) /* guest can rcv TSO with ECN */
86 #define	VIRTIO_NET_F_GUEST_UFO	(1 << 10) /* guest can rcv UFO */
87 #define	VIRTIO_NET_F_HOST_TSO4	(1 << 11) /* host can rcv TSOv4 */
88 #define	VIRTIO_NET_F_HOST_TSO6	(1 << 12) /* host can rcv TSOv6 */
89 #define	VIRTIO_NET_F_HOST_ECN	(1 << 13) /* host can rcv TSO with ECN */
90 #define	VIRTIO_NET_F_HOST_UFO	(1 << 14) /* host can rcv UFO */
91 #define	VIRTIO_NET_F_MRG_RXBUF	(1 << 15) /* host can merge RX buffers */
92 #define	VIRTIO_NET_F_STATUS	(1 << 16) /* config status field available */
93 #define	VIRTIO_NET_F_CTRL_VQ	(1 << 17) /* control channel available */
94 #define	VIRTIO_NET_F_CTRL_RX	(1 << 18) /* control channel RX mode support */
95 #define	VIRTIO_NET_F_CTRL_VLAN	(1 << 19) /* control channel VLAN filtering */
96 #define	VIRTIO_NET_F_GUEST_ANNOUNCE \
97 				(1 << 21) /* guest can send gratuitous pkts */
98 
99 #define VTNET_S_HOSTCAPS      \
100   ( VIRTIO_NET_F_MAC | VIRTIO_NET_F_MRG_RXBUF | VIRTIO_NET_F_STATUS | \
101     VIRTIO_F_NOTIFY_ON_EMPTY | VIRTIO_RING_F_INDIRECT_DESC)
102 
103 /*
104  * PCI config-space "registers"
105  */
106 struct virtio_net_config {
107 	uint8_t  mac[6];
108 	uint16_t status;
109 } __packed;
110 
111 /*
112  * Queue definitions.
113  */
114 #define VTNET_RXQ	0
115 #define VTNET_TXQ	1
116 #define VTNET_CTLQ	2	/* NB: not yet supported */
117 
118 #define VTNET_MAXQ	3
119 
120 /*
121  * Fixed network header size
122  */
123 struct virtio_net_rxhdr {
124 	uint8_t		vrh_flags;
125 	uint8_t		vrh_gso_type;
126 	uint16_t	vrh_hdr_len;
127 	uint16_t	vrh_gso_size;
128 	uint16_t	vrh_csum_start;
129 	uint16_t	vrh_csum_offset;
130 	uint16_t	vrh_bufs;
131 } __packed;
132 
133 /*
134  * Debug printf
135  */
136 static int pci_vtnet_debug;
137 #define DPRINTF(params) if (pci_vtnet_debug) printf params
138 #define WPRINTF(params) printf params
139 
140 /*
141  * Per-device softc
142  */
143 struct pci_vtnet_softc {
144 	struct virtio_softc vsc_vs;
145 	struct vqueue_info vsc_queues[VTNET_MAXQ - 1];
146 	pthread_mutex_t vsc_mtx;
147 	struct mevent	*vsc_mevp;
148 
149 	int		vsc_tapfd;
150 	struct nm_desc	*vsc_nmd;
151 
152 	int		vsc_rx_ready;
153 	int		resetting;	/* protected by tx_mtx */
154 
155 	uint64_t	vsc_features;	/* negotiated features */
156 
157 	struct virtio_net_config vsc_config;
158 
159 	pthread_mutex_t	rx_mtx;
160 	int		rx_vhdrlen;
161 	int		rx_merge;	/* merged rx bufs in use */
162 
163 	pthread_t 	tx_tid;
164 	pthread_mutex_t	tx_mtx;
165 	pthread_cond_t	tx_cond;
166 	int		tx_in_progress;
167 
168 	void (*pci_vtnet_rx)(struct pci_vtnet_softc *sc);
169 	void (*pci_vtnet_tx)(struct pci_vtnet_softc *sc, struct iovec *iov,
170 			     int iovcnt, int len);
171 };
172 
173 static void pci_vtnet_reset(void *);
174 /* static void pci_vtnet_notify(void *, struct vqueue_info *); */
175 static int pci_vtnet_cfgread(void *, int, int, uint32_t *);
176 static int pci_vtnet_cfgwrite(void *, int, int, uint32_t);
177 static void pci_vtnet_neg_features(void *, uint64_t);
178 
179 static struct virtio_consts vtnet_vi_consts = {
180 	"vtnet",		/* our name */
181 	VTNET_MAXQ - 1,		/* we currently support 2 virtqueues */
182 	sizeof(struct virtio_net_config), /* config reg size */
183 	pci_vtnet_reset,	/* reset */
184 	NULL,			/* device-wide qnotify -- not used */
185 	pci_vtnet_cfgread,	/* read PCI config */
186 	pci_vtnet_cfgwrite,	/* write PCI config */
187 	pci_vtnet_neg_features,	/* apply negotiated features */
188 	VTNET_S_HOSTCAPS,	/* our capabilities */
189 };
190 
191 static void
192 pci_vtnet_reset(void *vsc)
193 {
194 	struct pci_vtnet_softc *sc = vsc;
195 
196 	DPRINTF(("vtnet: device reset requested !\n"));
197 
198 	/* Acquire the RX lock to block RX processing. */
199 	pthread_mutex_lock(&sc->rx_mtx);
200 
201 	/* Set sc->resetting and give a chance to the TX thread to stop. */
202 	pthread_mutex_lock(&sc->tx_mtx);
203 	sc->resetting = 1;
204 	while (sc->tx_in_progress) {
205 		pthread_mutex_unlock(&sc->tx_mtx);
206 		usleep(10000);
207 		pthread_mutex_lock(&sc->tx_mtx);
208 	}
209 
210 	sc->vsc_rx_ready = 0;
211 	sc->rx_merge = 1;
212 	sc->rx_vhdrlen = sizeof(struct virtio_net_rxhdr);
213 
214 	/*
215 	 * Now reset rings, MSI-X vectors, and negotiated capabilities.
216 	 * Do that with the TX lock held, since we need to reset
217 	 * sc->resetting.
218 	 */
219 	vi_reset_dev(&sc->vsc_vs);
220 
221 	sc->resetting = 0;
222 	pthread_mutex_unlock(&sc->tx_mtx);
223 	pthread_mutex_unlock(&sc->rx_mtx);
224 }
225 
226 /*
227  * Called to send a buffer chain out to the tap device
228  */
229 static void
230 pci_vtnet_tap_tx(struct pci_vtnet_softc *sc, struct iovec *iov, int iovcnt,
231 		 int len)
232 {
233 	static char pad[60]; /* all zero bytes */
234 
235 	if (sc->vsc_tapfd == -1)
236 		return;
237 
238 	/*
239 	 * If the length is < 60, pad out to that and add the
240 	 * extra zero'd segment to the iov. It is guaranteed that
241 	 * there is always an extra iov available by the caller.
242 	 */
243 	if (len < 60) {
244 		iov[iovcnt].iov_base = pad;
245 		iov[iovcnt].iov_len = 60 - len;
246 		iovcnt++;
247 	}
248 	(void) writev(sc->vsc_tapfd, iov, iovcnt);
249 }
250 
251 /*
252  *  Called when there is read activity on the tap file descriptor.
253  * Each buffer posted by the guest is assumed to be able to contain
254  * an entire ethernet frame + rx header.
255  *  MP note: the dummybuf is only used for discarding frames, so there
256  * is no need for it to be per-vtnet or locked.
257  */
258 static uint8_t dummybuf[2048];
259 
260 static __inline struct iovec *
261 rx_iov_trim(struct iovec *iov, int *niov, int tlen)
262 {
263 	struct iovec *riov;
264 
265 	/* XXX short-cut: assume first segment is >= tlen */
266 	assert(iov[0].iov_len >= tlen);
267 
268 	iov[0].iov_len -= tlen;
269 	if (iov[0].iov_len == 0) {
270 		assert(*niov > 1);
271 		*niov -= 1;
272 		riov = &iov[1];
273 	} else {
274 		iov[0].iov_base = (void *)((uintptr_t)iov[0].iov_base + tlen);
275 		riov = &iov[0];
276 	}
277 
278 	return (riov);
279 }
280 
281 static void
282 pci_vtnet_tap_rx(struct pci_vtnet_softc *sc)
283 {
284 	struct iovec iov[VTNET_MAXSEGS], *riov;
285 	struct vqueue_info *vq;
286 	void *vrx;
287 	int len, n;
288 	uint16_t idx;
289 
290 	/*
291 	 * Should never be called without a valid tap fd
292 	 */
293 	assert(sc->vsc_tapfd != -1);
294 
295 	/*
296 	 * But, will be called when the rx ring hasn't yet
297 	 * been set up.
298 	 */
299 	if (!sc->vsc_rx_ready) {
300 		/*
301 		 * Drop the packet and try later.
302 		 */
303 		(void) read(sc->vsc_tapfd, dummybuf, sizeof(dummybuf));
304 		return;
305 	}
306 
307 	/*
308 	 * Check for available rx buffers
309 	 */
310 	vq = &sc->vsc_queues[VTNET_RXQ];
311 	if (!vq_has_descs(vq)) {
312 		/*
313 		 * Drop the packet and try later.  Interrupt on
314 		 * empty, if that's negotiated.
315 		 */
316 		(void) read(sc->vsc_tapfd, dummybuf, sizeof(dummybuf));
317 		vq_endchains(vq, 1);
318 		return;
319 	}
320 
321 	do {
322 		/*
323 		 * Get descriptor chain.
324 		 */
325 		n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL);
326 		assert(n >= 1 && n <= VTNET_MAXSEGS);
327 
328 		/*
329 		 * Get a pointer to the rx header, and use the
330 		 * data immediately following it for the packet buffer.
331 		 */
332 		vrx = iov[0].iov_base;
333 		riov = rx_iov_trim(iov, &n, sc->rx_vhdrlen);
334 
335 		len = readv(sc->vsc_tapfd, riov, n);
336 
337 		if (len < 0 && errno == EWOULDBLOCK) {
338 			/*
339 			 * No more packets, but still some avail ring
340 			 * entries.  Interrupt if needed/appropriate.
341 			 */
342 			vq_retchain(vq);
343 			vq_endchains(vq, 0);
344 			return;
345 		}
346 
347 		/*
348 		 * The only valid field in the rx packet header is the
349 		 * number of buffers if merged rx bufs were negotiated.
350 		 */
351 		memset(vrx, 0, sc->rx_vhdrlen);
352 
353 		if (sc->rx_merge) {
354 			struct virtio_net_rxhdr *vrxh;
355 
356 			vrxh = vrx;
357 			vrxh->vrh_bufs = 1;
358 		}
359 
360 		/*
361 		 * Release this chain and handle more chains.
362 		 */
363 		vq_relchain(vq, idx, len + sc->rx_vhdrlen);
364 	} while (vq_has_descs(vq));
365 
366 	/* Interrupt if needed, including for NOTIFY_ON_EMPTY. */
367 	vq_endchains(vq, 1);
368 }
369 
370 static __inline int
371 pci_vtnet_netmap_writev(struct nm_desc *nmd, struct iovec *iov, int iovcnt)
372 {
373 	int r, i;
374 	int len = 0;
375 
376 	for (r = nmd->cur_tx_ring; ; ) {
377 		struct netmap_ring *ring = NETMAP_TXRING(nmd->nifp, r);
378 		uint32_t cur, idx;
379 		char *buf;
380 
381 		if (nm_ring_empty(ring)) {
382 			r++;
383 			if (r > nmd->last_tx_ring)
384 				r = nmd->first_tx_ring;
385 			if (r == nmd->cur_tx_ring)
386 				break;
387 			continue;
388 		}
389 		cur = ring->cur;
390 		idx = ring->slot[cur].buf_idx;
391 		buf = NETMAP_BUF(ring, idx);
392 
393 		for (i = 0; i < iovcnt; i++) {
394 			if (len + iov[i].iov_len > 2048)
395 				break;
396 			memcpy(&buf[len], iov[i].iov_base, iov[i].iov_len);
397 			len += iov[i].iov_len;
398 		}
399 		ring->slot[cur].len = len;
400 		ring->head = ring->cur = nm_ring_next(ring, cur);
401 		nmd->cur_tx_ring = r;
402 		ioctl(nmd->fd, NIOCTXSYNC, NULL);
403 		break;
404 	}
405 
406 	return (len);
407 }
408 
409 static __inline int
410 pci_vtnet_netmap_readv(struct nm_desc *nmd, struct iovec *iov, int iovcnt)
411 {
412 	int len = 0;
413 	int i = 0;
414 	int r;
415 
416 	for (r = nmd->cur_rx_ring; ; ) {
417 		struct netmap_ring *ring = NETMAP_RXRING(nmd->nifp, r);
418 		uint32_t cur, idx;
419 		char *buf;
420 		size_t left;
421 
422 		if (nm_ring_empty(ring)) {
423 			r++;
424 			if (r > nmd->last_rx_ring)
425 				r = nmd->first_rx_ring;
426 			if (r == nmd->cur_rx_ring)
427 				break;
428 			continue;
429 		}
430 		cur = ring->cur;
431 		idx = ring->slot[cur].buf_idx;
432 		buf = NETMAP_BUF(ring, idx);
433 		left = ring->slot[cur].len;
434 
435 		for (i = 0; i < iovcnt && left > 0; i++) {
436 			if (iov[i].iov_len > left)
437 				iov[i].iov_len = left;
438 			memcpy(iov[i].iov_base, &buf[len], iov[i].iov_len);
439 			len += iov[i].iov_len;
440 			left -= iov[i].iov_len;
441 		}
442 		ring->head = ring->cur = nm_ring_next(ring, cur);
443 		nmd->cur_rx_ring = r;
444 		ioctl(nmd->fd, NIOCRXSYNC, NULL);
445 		break;
446 	}
447 	for (; i < iovcnt; i++)
448 		iov[i].iov_len = 0;
449 
450 	return (len);
451 }
452 
453 /*
454  * Called to send a buffer chain out to the vale port
455  */
456 static void
457 pci_vtnet_netmap_tx(struct pci_vtnet_softc *sc, struct iovec *iov, int iovcnt,
458 		    int len)
459 {
460 	static char pad[60]; /* all zero bytes */
461 
462 	if (sc->vsc_nmd == NULL)
463 		return;
464 
465 	/*
466 	 * If the length is < 60, pad out to that and add the
467 	 * extra zero'd segment to the iov. It is guaranteed that
468 	 * there is always an extra iov available by the caller.
469 	 */
470 	if (len < 60) {
471 		iov[iovcnt].iov_base = pad;
472 		iov[iovcnt].iov_len = 60 - len;
473 		iovcnt++;
474 	}
475 	(void) pci_vtnet_netmap_writev(sc->vsc_nmd, iov, iovcnt);
476 }
477 
478 static void
479 pci_vtnet_netmap_rx(struct pci_vtnet_softc *sc)
480 {
481 	struct iovec iov[VTNET_MAXSEGS], *riov;
482 	struct vqueue_info *vq;
483 	void *vrx;
484 	int len, n;
485 	uint16_t idx;
486 
487 	/*
488 	 * Should never be called without a valid netmap descriptor
489 	 */
490 	assert(sc->vsc_nmd != NULL);
491 
492 	/*
493 	 * But, will be called when the rx ring hasn't yet
494 	 * been set up.
495 	 */
496 	if (!sc->vsc_rx_ready) {
497 		/*
498 		 * Drop the packet and try later.
499 		 */
500 		(void) nm_nextpkt(sc->vsc_nmd, (void *)dummybuf);
501 		return;
502 	}
503 
504 	/*
505 	 * Check for available rx buffers
506 	 */
507 	vq = &sc->vsc_queues[VTNET_RXQ];
508 	if (!vq_has_descs(vq)) {
509 		/*
510 		 * Drop the packet and try later.  Interrupt on
511 		 * empty, if that's negotiated.
512 		 */
513 		(void) nm_nextpkt(sc->vsc_nmd, (void *)dummybuf);
514 		vq_endchains(vq, 1);
515 		return;
516 	}
517 
518 	do {
519 		/*
520 		 * Get descriptor chain.
521 		 */
522 		n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL);
523 		assert(n >= 1 && n <= VTNET_MAXSEGS);
524 
525 		/*
526 		 * Get a pointer to the rx header, and use the
527 		 * data immediately following it for the packet buffer.
528 		 */
529 		vrx = iov[0].iov_base;
530 		riov = rx_iov_trim(iov, &n, sc->rx_vhdrlen);
531 
532 		len = pci_vtnet_netmap_readv(sc->vsc_nmd, riov, n);
533 
534 		if (len == 0) {
535 			/*
536 			 * No more packets, but still some avail ring
537 			 * entries.  Interrupt if needed/appropriate.
538 			 */
539 			vq_retchain(vq);
540 			vq_endchains(vq, 0);
541 			return;
542 		}
543 
544 		/*
545 		 * The only valid field in the rx packet header is the
546 		 * number of buffers if merged rx bufs were negotiated.
547 		 */
548 		memset(vrx, 0, sc->rx_vhdrlen);
549 
550 		if (sc->rx_merge) {
551 			struct virtio_net_rxhdr *vrxh;
552 
553 			vrxh = vrx;
554 			vrxh->vrh_bufs = 1;
555 		}
556 
557 		/*
558 		 * Release this chain and handle more chains.
559 		 */
560 		vq_relchain(vq, idx, len + sc->rx_vhdrlen);
561 	} while (vq_has_descs(vq));
562 
563 	/* Interrupt if needed, including for NOTIFY_ON_EMPTY. */
564 	vq_endchains(vq, 1);
565 }
566 
567 static void
568 pci_vtnet_rx_callback(int fd, enum ev_type type, void *param)
569 {
570 	struct pci_vtnet_softc *sc = param;
571 
572 	pthread_mutex_lock(&sc->rx_mtx);
573 	sc->pci_vtnet_rx(sc);
574 	pthread_mutex_unlock(&sc->rx_mtx);
575 
576 }
577 
578 static void
579 pci_vtnet_ping_rxq(void *vsc, struct vqueue_info *vq)
580 {
581 	struct pci_vtnet_softc *sc = vsc;
582 
583 	/*
584 	 * A qnotify means that the rx process can now begin
585 	 */
586 	if (sc->vsc_rx_ready == 0) {
587 		sc->vsc_rx_ready = 1;
588 		vq_kick_disable(vq);
589 	}
590 }
591 
592 static void
593 pci_vtnet_proctx(struct pci_vtnet_softc *sc, struct vqueue_info *vq)
594 {
595 	struct iovec iov[VTNET_MAXSEGS + 1];
596 	int i, n;
597 	int plen, tlen;
598 	uint16_t idx;
599 
600 	/*
601 	 * Obtain chain of descriptors.  The first one is
602 	 * really the header descriptor, so we need to sum
603 	 * up two lengths: packet length and transfer length.
604 	 */
605 	n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL);
606 	assert(n >= 1 && n <= VTNET_MAXSEGS);
607 	plen = 0;
608 	tlen = iov[0].iov_len;
609 	for (i = 1; i < n; i++) {
610 		plen += iov[i].iov_len;
611 		tlen += iov[i].iov_len;
612 	}
613 
614 	DPRINTF(("virtio: packet send, %d bytes, %d segs\n\r", plen, n));
615 	sc->pci_vtnet_tx(sc, &iov[1], n - 1, plen);
616 
617 	/* chain is processed, release it and set tlen */
618 	vq_relchain(vq, idx, tlen);
619 }
620 
621 static void
622 pci_vtnet_ping_txq(void *vsc, struct vqueue_info *vq)
623 {
624 	struct pci_vtnet_softc *sc = vsc;
625 
626 	/*
627 	 * Any ring entries to process?
628 	 */
629 	if (!vq_has_descs(vq))
630 		return;
631 
632 	/* Signal the tx thread for processing */
633 	pthread_mutex_lock(&sc->tx_mtx);
634 	vq_kick_disable(vq);
635 	if (sc->tx_in_progress == 0)
636 		pthread_cond_signal(&sc->tx_cond);
637 	pthread_mutex_unlock(&sc->tx_mtx);
638 }
639 
640 /*
641  * Thread which will handle processing of TX desc
642  */
643 static void *
644 pci_vtnet_tx_thread(void *param)
645 {
646 	struct pci_vtnet_softc *sc = param;
647 	struct vqueue_info *vq;
648 	int error;
649 
650 	vq = &sc->vsc_queues[VTNET_TXQ];
651 
652 	/*
653 	 * Let us wait till the tx queue pointers get initialised &
654 	 * first tx signaled
655 	 */
656 	pthread_mutex_lock(&sc->tx_mtx);
657 	error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx);
658 	assert(error == 0);
659 
660 	for (;;) {
661 		/* note - tx mutex is locked here */
662 		while (sc->resetting || !vq_has_descs(vq)) {
663 			vq_kick_enable(vq);
664 			if (!sc->resetting && vq_has_descs(vq))
665 				break;
666 
667 			sc->tx_in_progress = 0;
668 			error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx);
669 			assert(error == 0);
670 		}
671 		vq_kick_disable(vq);
672 		sc->tx_in_progress = 1;
673 		pthread_mutex_unlock(&sc->tx_mtx);
674 
675 		do {
676 			/*
677 			 * Run through entries, placing them into
678 			 * iovecs and sending when an end-of-packet
679 			 * is found
680 			 */
681 			pci_vtnet_proctx(sc, vq);
682 		} while (vq_has_descs(vq));
683 
684 		/*
685 		 * Generate an interrupt if needed.
686 		 */
687 		vq_endchains(vq, 1);
688 
689 		pthread_mutex_lock(&sc->tx_mtx);
690 	}
691 }
692 
693 #ifdef notyet
694 static void
695 pci_vtnet_ping_ctlq(void *vsc, struct vqueue_info *vq)
696 {
697 
698 	DPRINTF(("vtnet: control qnotify!\n\r"));
699 }
700 #endif
701 
702 static void
703 pci_vtnet_tap_setup(struct pci_vtnet_softc *sc, char *devname)
704 {
705 	char tbuf[80];
706 #ifndef WITHOUT_CAPSICUM
707 	cap_rights_t rights;
708 #endif
709 
710 	strcpy(tbuf, "/dev/");
711 	strlcat(tbuf, devname, sizeof(tbuf));
712 
713 	sc->pci_vtnet_rx = pci_vtnet_tap_rx;
714 	sc->pci_vtnet_tx = pci_vtnet_tap_tx;
715 
716 	sc->vsc_tapfd = open(tbuf, O_RDWR);
717 	if (sc->vsc_tapfd == -1) {
718 		WPRINTF(("open of tap device %s failed\n", tbuf));
719 		return;
720 	}
721 
722 	/*
723 	 * Set non-blocking and register for read
724 	 * notifications with the event loop
725 	 */
726 	int opt = 1;
727 	if (ioctl(sc->vsc_tapfd, FIONBIO, &opt) < 0) {
728 		WPRINTF(("tap device O_NONBLOCK failed\n"));
729 		close(sc->vsc_tapfd);
730 		sc->vsc_tapfd = -1;
731 	}
732 
733 #ifndef WITHOUT_CAPSICUM
734 	cap_rights_init(&rights, CAP_EVENT, CAP_READ, CAP_WRITE);
735 	if (caph_rights_limit(sc->vsc_tapfd, &rights) == -1)
736 		errx(EX_OSERR, "Unable to apply rights for sandbox");
737 #endif
738 
739 	sc->vsc_mevp = mevent_add(sc->vsc_tapfd,
740 				  EVF_READ,
741 				  pci_vtnet_rx_callback,
742 				  sc);
743 	if (sc->vsc_mevp == NULL) {
744 		WPRINTF(("Could not register event\n"));
745 		close(sc->vsc_tapfd);
746 		sc->vsc_tapfd = -1;
747 	}
748 }
749 
750 static void
751 pci_vtnet_netmap_setup(struct pci_vtnet_softc *sc, char *ifname)
752 {
753 	sc->pci_vtnet_rx = pci_vtnet_netmap_rx;
754 	sc->pci_vtnet_tx = pci_vtnet_netmap_tx;
755 
756 	sc->vsc_nmd = nm_open(ifname, NULL, 0, 0);
757 	if (sc->vsc_nmd == NULL) {
758 		WPRINTF(("open of netmap device %s failed\n", ifname));
759 		return;
760 	}
761 
762 	sc->vsc_mevp = mevent_add(sc->vsc_nmd->fd,
763 				  EVF_READ,
764 				  pci_vtnet_rx_callback,
765 				  sc);
766 	if (sc->vsc_mevp == NULL) {
767 		WPRINTF(("Could not register event\n"));
768 		nm_close(sc->vsc_nmd);
769 		sc->vsc_nmd = NULL;
770 	}
771 }
772 
773 static int
774 pci_vtnet_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts)
775 {
776 	char tname[MAXCOMLEN + 1];
777 	struct pci_vtnet_softc *sc;
778 	char *devname;
779 	char *vtopts;
780 	int mac_provided;
781 
782 	sc = calloc(1, sizeof(struct pci_vtnet_softc));
783 
784 	pthread_mutex_init(&sc->vsc_mtx, NULL);
785 
786 	vi_softc_linkup(&sc->vsc_vs, &vtnet_vi_consts, sc, pi, sc->vsc_queues);
787 	sc->vsc_vs.vs_mtx = &sc->vsc_mtx;
788 
789 	sc->vsc_queues[VTNET_RXQ].vq_qsize = VTNET_RINGSZ;
790 	sc->vsc_queues[VTNET_RXQ].vq_notify = pci_vtnet_ping_rxq;
791 	sc->vsc_queues[VTNET_TXQ].vq_qsize = VTNET_RINGSZ;
792 	sc->vsc_queues[VTNET_TXQ].vq_notify = pci_vtnet_ping_txq;
793 #ifdef notyet
794 	sc->vsc_queues[VTNET_CTLQ].vq_qsize = VTNET_RINGSZ;
795         sc->vsc_queues[VTNET_CTLQ].vq_notify = pci_vtnet_ping_ctlq;
796 #endif
797 
798 	/*
799 	 * Attempt to open the tap device and read the MAC address
800 	 * if specified
801 	 */
802 	mac_provided = 0;
803 	sc->vsc_tapfd = -1;
804 	sc->vsc_nmd = NULL;
805 	if (opts != NULL) {
806 		int err;
807 
808 		devname = vtopts = strdup(opts);
809 		(void) strsep(&vtopts, ",");
810 
811 		if (vtopts != NULL) {
812 			err = net_parsemac(vtopts, sc->vsc_config.mac);
813 			if (err != 0) {
814 				free(devname);
815 				return (err);
816 			}
817 			mac_provided = 1;
818 		}
819 
820 		if (strncmp(devname, "vale", 4) == 0)
821 			pci_vtnet_netmap_setup(sc, devname);
822 		if (strncmp(devname, "tap", 3) == 0 ||
823 		    strncmp(devname, "vmnet", 5) == 0)
824 			pci_vtnet_tap_setup(sc, devname);
825 
826 		free(devname);
827 	}
828 
829 	if (!mac_provided) {
830 		net_genmac(pi, sc->vsc_config.mac);
831 	}
832 
833 	/* initialize config space */
834 	pci_set_cfgdata16(pi, PCIR_DEVICE, VIRTIO_DEV_NET);
835 	pci_set_cfgdata16(pi, PCIR_VENDOR, VIRTIO_VENDOR);
836 	pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_NETWORK);
837 	pci_set_cfgdata16(pi, PCIR_SUBDEV_0, VIRTIO_TYPE_NET);
838 	pci_set_cfgdata16(pi, PCIR_SUBVEND_0, VIRTIO_VENDOR);
839 
840 	/* Link is up if we managed to open tap device or vale port. */
841 	sc->vsc_config.status = (opts == NULL || sc->vsc_tapfd >= 0 ||
842 	    sc->vsc_nmd != NULL);
843 
844 	/* use BAR 1 to map MSI-X table and PBA, if we're using MSI-X */
845 	if (vi_intr_init(&sc->vsc_vs, 1, fbsdrun_virtio_msix()))
846 		return (1);
847 
848 	/* use BAR 0 to map config regs in IO space */
849 	vi_set_io_bar(&sc->vsc_vs, 0);
850 
851 	sc->resetting = 0;
852 
853 	sc->rx_merge = 1;
854 	sc->rx_vhdrlen = sizeof(struct virtio_net_rxhdr);
855 	pthread_mutex_init(&sc->rx_mtx, NULL);
856 
857 	/*
858 	 * Initialize tx semaphore & spawn TX processing thread.
859 	 * As of now, only one thread for TX desc processing is
860 	 * spawned.
861 	 */
862 	sc->tx_in_progress = 0;
863 	pthread_mutex_init(&sc->tx_mtx, NULL);
864 	pthread_cond_init(&sc->tx_cond, NULL);
865 	pthread_create(&sc->tx_tid, NULL, pci_vtnet_tx_thread, (void *)sc);
866 	snprintf(tname, sizeof(tname), "vtnet-%d:%d tx", pi->pi_slot,
867 	    pi->pi_func);
868 	pthread_set_name_np(sc->tx_tid, tname);
869 
870 	return (0);
871 }
872 
873 static int
874 pci_vtnet_cfgwrite(void *vsc, int offset, int size, uint32_t value)
875 {
876 	struct pci_vtnet_softc *sc = vsc;
877 	void *ptr;
878 
879 	if (offset < 6) {
880 		assert(offset + size <= 6);
881 		/*
882 		 * The driver is allowed to change the MAC address
883 		 */
884 		ptr = &sc->vsc_config.mac[offset];
885 		memcpy(ptr, &value, size);
886 	} else {
887 		/* silently ignore other writes */
888 		DPRINTF(("vtnet: write to readonly reg %d\n\r", offset));
889 	}
890 
891 	return (0);
892 }
893 
894 static int
895 pci_vtnet_cfgread(void *vsc, int offset, int size, uint32_t *retval)
896 {
897 	struct pci_vtnet_softc *sc = vsc;
898 	void *ptr;
899 
900 	ptr = (uint8_t *)&sc->vsc_config + offset;
901 	memcpy(retval, ptr, size);
902 	return (0);
903 }
904 
905 static void
906 pci_vtnet_neg_features(void *vsc, uint64_t negotiated_features)
907 {
908 	struct pci_vtnet_softc *sc = vsc;
909 
910 	sc->vsc_features = negotiated_features;
911 
912 	if (!(sc->vsc_features & VIRTIO_NET_F_MRG_RXBUF)) {
913 		sc->rx_merge = 0;
914 		/* non-merge rx header is 2 bytes shorter */
915 		sc->rx_vhdrlen -= 2;
916 	}
917 }
918 
919 struct pci_devemu pci_de_vnet = {
920 	.pe_emu = 	"virtio-net",
921 	.pe_init =	pci_vtnet_init,
922 	.pe_barwrite =	vi_pci_write,
923 	.pe_barread =	vi_pci_read
924 };
925 PCI_EMUL_SET(pci_de_vnet);
926