xref: /freebsd/usr.sbin/bhyve/pci_virtio_net.c (revision 3fc36ee018bb836bd1796067cf4ef8683f166ebc)
1 /*-
2  * Copyright (c) 2011 NetApp, Inc.
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  *
26  * $FreeBSD$
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/linker_set.h>
34 #include <sys/select.h>
35 #include <sys/uio.h>
36 #include <sys/ioctl.h>
37 #include <machine/atomic.h>
38 #include <net/ethernet.h>
39 #ifndef NETMAP_WITH_LIBS
40 #define NETMAP_WITH_LIBS
41 #endif
42 #include <net/netmap_user.h>
43 
44 #include <errno.h>
45 #include <fcntl.h>
46 #include <stdio.h>
47 #include <stdlib.h>
48 #include <stdint.h>
49 #include <string.h>
50 #include <strings.h>
51 #include <unistd.h>
52 #include <assert.h>
53 #include <md5.h>
54 #include <pthread.h>
55 #include <pthread_np.h>
56 
57 #include "bhyverun.h"
58 #include "pci_emul.h"
59 #include "mevent.h"
60 #include "virtio.h"
61 
62 #define VTNET_RINGSZ	1024
63 
64 #define VTNET_MAXSEGS	256
65 
66 /*
67  * Host capabilities.  Note that we only offer a few of these.
68  */
69 #define	VIRTIO_NET_F_CSUM	(1 <<  0) /* host handles partial cksum */
70 #define	VIRTIO_NET_F_GUEST_CSUM	(1 <<  1) /* guest handles partial cksum */
71 #define	VIRTIO_NET_F_MAC	(1 <<  5) /* host supplies MAC */
72 #define	VIRTIO_NET_F_GSO_DEPREC	(1 <<  6) /* deprecated: host handles GSO */
73 #define	VIRTIO_NET_F_GUEST_TSO4	(1 <<  7) /* guest can rcv TSOv4 */
74 #define	VIRTIO_NET_F_GUEST_TSO6	(1 <<  8) /* guest can rcv TSOv6 */
75 #define	VIRTIO_NET_F_GUEST_ECN	(1 <<  9) /* guest can rcv TSO with ECN */
76 #define	VIRTIO_NET_F_GUEST_UFO	(1 << 10) /* guest can rcv UFO */
77 #define	VIRTIO_NET_F_HOST_TSO4	(1 << 11) /* host can rcv TSOv4 */
78 #define	VIRTIO_NET_F_HOST_TSO6	(1 << 12) /* host can rcv TSOv6 */
79 #define	VIRTIO_NET_F_HOST_ECN	(1 << 13) /* host can rcv TSO with ECN */
80 #define	VIRTIO_NET_F_HOST_UFO	(1 << 14) /* host can rcv UFO */
81 #define	VIRTIO_NET_F_MRG_RXBUF	(1 << 15) /* host can merge RX buffers */
82 #define	VIRTIO_NET_F_STATUS	(1 << 16) /* config status field available */
83 #define	VIRTIO_NET_F_CTRL_VQ	(1 << 17) /* control channel available */
84 #define	VIRTIO_NET_F_CTRL_RX	(1 << 18) /* control channel RX mode support */
85 #define	VIRTIO_NET_F_CTRL_VLAN	(1 << 19) /* control channel VLAN filtering */
86 #define	VIRTIO_NET_F_GUEST_ANNOUNCE \
87 				(1 << 21) /* guest can send gratuitous pkts */
88 
89 #define VTNET_S_HOSTCAPS      \
90   ( VIRTIO_NET_F_MAC | VIRTIO_NET_F_MRG_RXBUF | VIRTIO_NET_F_STATUS | \
91     VIRTIO_F_NOTIFY_ON_EMPTY | VIRTIO_RING_F_INDIRECT_DESC)
92 
93 /*
94  * PCI config-space "registers"
95  */
96 struct virtio_net_config {
97 	uint8_t  mac[6];
98 	uint16_t status;
99 } __packed;
100 
101 /*
102  * Queue definitions.
103  */
104 #define VTNET_RXQ	0
105 #define VTNET_TXQ	1
106 #define VTNET_CTLQ	2	/* NB: not yet supported */
107 
108 #define VTNET_MAXQ	3
109 
110 /*
111  * Fixed network header size
112  */
113 struct virtio_net_rxhdr {
114 	uint8_t		vrh_flags;
115 	uint8_t		vrh_gso_type;
116 	uint16_t	vrh_hdr_len;
117 	uint16_t	vrh_gso_size;
118 	uint16_t	vrh_csum_start;
119 	uint16_t	vrh_csum_offset;
120 	uint16_t	vrh_bufs;
121 } __packed;
122 
123 /*
124  * Debug printf
125  */
126 static int pci_vtnet_debug;
127 #define DPRINTF(params) if (pci_vtnet_debug) printf params
128 #define WPRINTF(params) printf params
129 
130 /*
131  * Per-device softc
132  */
133 struct pci_vtnet_softc {
134 	struct virtio_softc vsc_vs;
135 	struct vqueue_info vsc_queues[VTNET_MAXQ - 1];
136 	pthread_mutex_t vsc_mtx;
137 	struct mevent	*vsc_mevp;
138 
139 	int		vsc_tapfd;
140 	struct nm_desc	*vsc_nmd;
141 
142 	int		vsc_rx_ready;
143 	volatile int	resetting;	/* set and checked outside lock */
144 
145 	uint64_t	vsc_features;	/* negotiated features */
146 
147 	struct virtio_net_config vsc_config;
148 
149 	pthread_mutex_t	rx_mtx;
150 	int		rx_in_progress;
151 	int		rx_vhdrlen;
152 	int		rx_merge;	/* merged rx bufs in use */
153 
154 	pthread_t 	tx_tid;
155 	pthread_mutex_t	tx_mtx;
156 	pthread_cond_t	tx_cond;
157 	int		tx_in_progress;
158 
159 	void (*pci_vtnet_rx)(struct pci_vtnet_softc *sc);
160 	void (*pci_vtnet_tx)(struct pci_vtnet_softc *sc, struct iovec *iov,
161 			     int iovcnt, int len);
162 };
163 
164 static void pci_vtnet_reset(void *);
165 /* static void pci_vtnet_notify(void *, struct vqueue_info *); */
166 static int pci_vtnet_cfgread(void *, int, int, uint32_t *);
167 static int pci_vtnet_cfgwrite(void *, int, int, uint32_t);
168 static void pci_vtnet_neg_features(void *, uint64_t);
169 
170 static struct virtio_consts vtnet_vi_consts = {
171 	"vtnet",		/* our name */
172 	VTNET_MAXQ - 1,		/* we currently support 2 virtqueues */
173 	sizeof(struct virtio_net_config), /* config reg size */
174 	pci_vtnet_reset,	/* reset */
175 	NULL,			/* device-wide qnotify -- not used */
176 	pci_vtnet_cfgread,	/* read PCI config */
177 	pci_vtnet_cfgwrite,	/* write PCI config */
178 	pci_vtnet_neg_features,	/* apply negotiated features */
179 	VTNET_S_HOSTCAPS,	/* our capabilities */
180 };
181 
182 /*
183  * If the transmit thread is active then stall until it is done.
184  */
185 static void
186 pci_vtnet_txwait(struct pci_vtnet_softc *sc)
187 {
188 
189 	pthread_mutex_lock(&sc->tx_mtx);
190 	while (sc->tx_in_progress) {
191 		pthread_mutex_unlock(&sc->tx_mtx);
192 		usleep(10000);
193 		pthread_mutex_lock(&sc->tx_mtx);
194 	}
195 	pthread_mutex_unlock(&sc->tx_mtx);
196 }
197 
198 /*
199  * If the receive thread is active then stall until it is done.
200  */
201 static void
202 pci_vtnet_rxwait(struct pci_vtnet_softc *sc)
203 {
204 
205 	pthread_mutex_lock(&sc->rx_mtx);
206 	while (sc->rx_in_progress) {
207 		pthread_mutex_unlock(&sc->rx_mtx);
208 		usleep(10000);
209 		pthread_mutex_lock(&sc->rx_mtx);
210 	}
211 	pthread_mutex_unlock(&sc->rx_mtx);
212 }
213 
214 static void
215 pci_vtnet_reset(void *vsc)
216 {
217 	struct pci_vtnet_softc *sc = vsc;
218 
219 	DPRINTF(("vtnet: device reset requested !\n"));
220 
221 	sc->resetting = 1;
222 
223 	/*
224 	 * Wait for the transmit and receive threads to finish their
225 	 * processing.
226 	 */
227 	pci_vtnet_txwait(sc);
228 	pci_vtnet_rxwait(sc);
229 
230 	sc->vsc_rx_ready = 0;
231 	sc->rx_merge = 1;
232 	sc->rx_vhdrlen = sizeof(struct virtio_net_rxhdr);
233 
234 	/* now reset rings, MSI-X vectors, and negotiated capabilities */
235 	vi_reset_dev(&sc->vsc_vs);
236 
237 	sc->resetting = 0;
238 }
239 
240 /*
241  * Called to send a buffer chain out to the tap device
242  */
243 static void
244 pci_vtnet_tap_tx(struct pci_vtnet_softc *sc, struct iovec *iov, int iovcnt,
245 		 int len)
246 {
247 	static char pad[60]; /* all zero bytes */
248 
249 	if (sc->vsc_tapfd == -1)
250 		return;
251 
252 	/*
253 	 * If the length is < 60, pad out to that and add the
254 	 * extra zero'd segment to the iov. It is guaranteed that
255 	 * there is always an extra iov available by the caller.
256 	 */
257 	if (len < 60) {
258 		iov[iovcnt].iov_base = pad;
259 		iov[iovcnt].iov_len = 60 - len;
260 		iovcnt++;
261 	}
262 	(void) writev(sc->vsc_tapfd, iov, iovcnt);
263 }
264 
265 /*
266  *  Called when there is read activity on the tap file descriptor.
267  * Each buffer posted by the guest is assumed to be able to contain
268  * an entire ethernet frame + rx header.
269  *  MP note: the dummybuf is only used for discarding frames, so there
270  * is no need for it to be per-vtnet or locked.
271  */
272 static uint8_t dummybuf[2048];
273 
274 static __inline struct iovec *
275 rx_iov_trim(struct iovec *iov, int *niov, int tlen)
276 {
277 	struct iovec *riov;
278 
279 	/* XXX short-cut: assume first segment is >= tlen */
280 	assert(iov[0].iov_len >= tlen);
281 
282 	iov[0].iov_len -= tlen;
283 	if (iov[0].iov_len == 0) {
284 		assert(*niov > 1);
285 		*niov -= 1;
286 		riov = &iov[1];
287 	} else {
288 		iov[0].iov_base = (void *)((uintptr_t)iov[0].iov_base + tlen);
289 		riov = &iov[0];
290 	}
291 
292 	return (riov);
293 }
294 
295 static void
296 pci_vtnet_tap_rx(struct pci_vtnet_softc *sc)
297 {
298 	struct iovec iov[VTNET_MAXSEGS], *riov;
299 	struct vqueue_info *vq;
300 	void *vrx;
301 	int len, n;
302 	uint16_t idx;
303 
304 	/*
305 	 * Should never be called without a valid tap fd
306 	 */
307 	assert(sc->vsc_tapfd != -1);
308 
309 	/*
310 	 * But, will be called when the rx ring hasn't yet
311 	 * been set up or the guest is resetting the device.
312 	 */
313 	if (!sc->vsc_rx_ready || sc->resetting) {
314 		/*
315 		 * Drop the packet and try later.
316 		 */
317 		(void) read(sc->vsc_tapfd, dummybuf, sizeof(dummybuf));
318 		return;
319 	}
320 
321 	/*
322 	 * Check for available rx buffers
323 	 */
324 	vq = &sc->vsc_queues[VTNET_RXQ];
325 	if (!vq_has_descs(vq)) {
326 		/*
327 		 * Drop the packet and try later.  Interrupt on
328 		 * empty, if that's negotiated.
329 		 */
330 		(void) read(sc->vsc_tapfd, dummybuf, sizeof(dummybuf));
331 		vq_endchains(vq, 1);
332 		return;
333 	}
334 
335 	do {
336 		/*
337 		 * Get descriptor chain.
338 		 */
339 		n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL);
340 		assert(n >= 1 && n <= VTNET_MAXSEGS);
341 
342 		/*
343 		 * Get a pointer to the rx header, and use the
344 		 * data immediately following it for the packet buffer.
345 		 */
346 		vrx = iov[0].iov_base;
347 		riov = rx_iov_trim(iov, &n, sc->rx_vhdrlen);
348 
349 		len = readv(sc->vsc_tapfd, riov, n);
350 
351 		if (len < 0 && errno == EWOULDBLOCK) {
352 			/*
353 			 * No more packets, but still some avail ring
354 			 * entries.  Interrupt if needed/appropriate.
355 			 */
356 			vq_retchain(vq);
357 			vq_endchains(vq, 0);
358 			return;
359 		}
360 
361 		/*
362 		 * The only valid field in the rx packet header is the
363 		 * number of buffers if merged rx bufs were negotiated.
364 		 */
365 		memset(vrx, 0, sc->rx_vhdrlen);
366 
367 		if (sc->rx_merge) {
368 			struct virtio_net_rxhdr *vrxh;
369 
370 			vrxh = vrx;
371 			vrxh->vrh_bufs = 1;
372 		}
373 
374 		/*
375 		 * Release this chain and handle more chains.
376 		 */
377 		vq_relchain(vq, idx, len + sc->rx_vhdrlen);
378 	} while (vq_has_descs(vq));
379 
380 	/* Interrupt if needed, including for NOTIFY_ON_EMPTY. */
381 	vq_endchains(vq, 1);
382 }
383 
384 static __inline int
385 pci_vtnet_netmap_writev(struct nm_desc *nmd, struct iovec *iov, int iovcnt)
386 {
387 	int r, i;
388 	int len = 0;
389 
390 	for (r = nmd->cur_tx_ring; ; ) {
391 		struct netmap_ring *ring = NETMAP_TXRING(nmd->nifp, r);
392 		uint32_t cur, idx;
393 		char *buf;
394 
395 		if (nm_ring_empty(ring)) {
396 			r++;
397 			if (r > nmd->last_tx_ring)
398 				r = nmd->first_tx_ring;
399 			if (r == nmd->cur_tx_ring)
400 				break;
401 			continue;
402 		}
403 		cur = ring->cur;
404 		idx = ring->slot[cur].buf_idx;
405 		buf = NETMAP_BUF(ring, idx);
406 
407 		for (i = 0; i < iovcnt; i++) {
408 			if (len + iov[i].iov_len > 2048)
409 				break;
410 			memcpy(&buf[len], iov[i].iov_base, iov[i].iov_len);
411 			len += iov[i].iov_len;
412 		}
413 		ring->slot[cur].len = len;
414 		ring->head = ring->cur = nm_ring_next(ring, cur);
415 		nmd->cur_tx_ring = r;
416 		ioctl(nmd->fd, NIOCTXSYNC, NULL);
417 		break;
418 	}
419 
420 	return (len);
421 }
422 
423 static __inline int
424 pci_vtnet_netmap_readv(struct nm_desc *nmd, struct iovec *iov, int iovcnt)
425 {
426 	int len = 0;
427 	int i = 0;
428 	int r;
429 
430 	for (r = nmd->cur_rx_ring; ; ) {
431 		struct netmap_ring *ring = NETMAP_RXRING(nmd->nifp, r);
432 		uint32_t cur, idx;
433 		char *buf;
434 		size_t left;
435 
436 		if (nm_ring_empty(ring)) {
437 			r++;
438 			if (r > nmd->last_rx_ring)
439 				r = nmd->first_rx_ring;
440 			if (r == nmd->cur_rx_ring)
441 				break;
442 			continue;
443 		}
444 		cur = ring->cur;
445 		idx = ring->slot[cur].buf_idx;
446 		buf = NETMAP_BUF(ring, idx);
447 		left = ring->slot[cur].len;
448 
449 		for (i = 0; i < iovcnt && left > 0; i++) {
450 			if (iov[i].iov_len > left)
451 				iov[i].iov_len = left;
452 			memcpy(iov[i].iov_base, &buf[len], iov[i].iov_len);
453 			len += iov[i].iov_len;
454 			left -= iov[i].iov_len;
455 		}
456 		ring->head = ring->cur = nm_ring_next(ring, cur);
457 		nmd->cur_rx_ring = r;
458 		ioctl(nmd->fd, NIOCRXSYNC, NULL);
459 		break;
460 	}
461 	for (; i < iovcnt; i++)
462 		iov[i].iov_len = 0;
463 
464 	return (len);
465 }
466 
467 /*
468  * Called to send a buffer chain out to the vale port
469  */
470 static void
471 pci_vtnet_netmap_tx(struct pci_vtnet_softc *sc, struct iovec *iov, int iovcnt,
472 		    int len)
473 {
474 	static char pad[60]; /* all zero bytes */
475 
476 	if (sc->vsc_nmd == NULL)
477 		return;
478 
479 	/*
480 	 * If the length is < 60, pad out to that and add the
481 	 * extra zero'd segment to the iov. It is guaranteed that
482 	 * there is always an extra iov available by the caller.
483 	 */
484 	if (len < 60) {
485 		iov[iovcnt].iov_base = pad;
486 		iov[iovcnt].iov_len = 60 - len;
487 		iovcnt++;
488 	}
489 	(void) pci_vtnet_netmap_writev(sc->vsc_nmd, iov, iovcnt);
490 }
491 
492 static void
493 pci_vtnet_netmap_rx(struct pci_vtnet_softc *sc)
494 {
495 	struct iovec iov[VTNET_MAXSEGS], *riov;
496 	struct vqueue_info *vq;
497 	void *vrx;
498 	int len, n;
499 	uint16_t idx;
500 
501 	/*
502 	 * Should never be called without a valid netmap descriptor
503 	 */
504 	assert(sc->vsc_nmd != NULL);
505 
506 	/*
507 	 * But, will be called when the rx ring hasn't yet
508 	 * been set up or the guest is resetting the device.
509 	 */
510 	if (!sc->vsc_rx_ready || sc->resetting) {
511 		/*
512 		 * Drop the packet and try later.
513 		 */
514 		(void) nm_nextpkt(sc->vsc_nmd, (void *)dummybuf);
515 		return;
516 	}
517 
518 	/*
519 	 * Check for available rx buffers
520 	 */
521 	vq = &sc->vsc_queues[VTNET_RXQ];
522 	if (!vq_has_descs(vq)) {
523 		/*
524 		 * Drop the packet and try later.  Interrupt on
525 		 * empty, if that's negotiated.
526 		 */
527 		(void) nm_nextpkt(sc->vsc_nmd, (void *)dummybuf);
528 		vq_endchains(vq, 1);
529 		return;
530 	}
531 
532 	do {
533 		/*
534 		 * Get descriptor chain.
535 		 */
536 		n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL);
537 		assert(n >= 1 && n <= VTNET_MAXSEGS);
538 
539 		/*
540 		 * Get a pointer to the rx header, and use the
541 		 * data immediately following it for the packet buffer.
542 		 */
543 		vrx = iov[0].iov_base;
544 		riov = rx_iov_trim(iov, &n, sc->rx_vhdrlen);
545 
546 		len = pci_vtnet_netmap_readv(sc->vsc_nmd, riov, n);
547 
548 		if (len == 0) {
549 			/*
550 			 * No more packets, but still some avail ring
551 			 * entries.  Interrupt if needed/appropriate.
552 			 */
553 			vq_retchain(vq);
554 			vq_endchains(vq, 0);
555 			return;
556 		}
557 
558 		/*
559 		 * The only valid field in the rx packet header is the
560 		 * number of buffers if merged rx bufs were negotiated.
561 		 */
562 		memset(vrx, 0, sc->rx_vhdrlen);
563 
564 		if (sc->rx_merge) {
565 			struct virtio_net_rxhdr *vrxh;
566 
567 			vrxh = vrx;
568 			vrxh->vrh_bufs = 1;
569 		}
570 
571 		/*
572 		 * Release this chain and handle more chains.
573 		 */
574 		vq_relchain(vq, idx, len + sc->rx_vhdrlen);
575 	} while (vq_has_descs(vq));
576 
577 	/* Interrupt if needed, including for NOTIFY_ON_EMPTY. */
578 	vq_endchains(vq, 1);
579 }
580 
581 static void
582 pci_vtnet_rx_callback(int fd, enum ev_type type, void *param)
583 {
584 	struct pci_vtnet_softc *sc = param;
585 
586 	pthread_mutex_lock(&sc->rx_mtx);
587 	sc->rx_in_progress = 1;
588 	sc->pci_vtnet_rx(sc);
589 	sc->rx_in_progress = 0;
590 	pthread_mutex_unlock(&sc->rx_mtx);
591 
592 }
593 
594 static void
595 pci_vtnet_ping_rxq(void *vsc, struct vqueue_info *vq)
596 {
597 	struct pci_vtnet_softc *sc = vsc;
598 
599 	/*
600 	 * A qnotify means that the rx process can now begin
601 	 */
602 	if (sc->vsc_rx_ready == 0) {
603 		sc->vsc_rx_ready = 1;
604 		vq->vq_used->vu_flags |= VRING_USED_F_NO_NOTIFY;
605 	}
606 }
607 
608 static void
609 pci_vtnet_proctx(struct pci_vtnet_softc *sc, struct vqueue_info *vq)
610 {
611 	struct iovec iov[VTNET_MAXSEGS + 1];
612 	int i, n;
613 	int plen, tlen;
614 	uint16_t idx;
615 
616 	/*
617 	 * Obtain chain of descriptors.  The first one is
618 	 * really the header descriptor, so we need to sum
619 	 * up two lengths: packet length and transfer length.
620 	 */
621 	n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL);
622 	assert(n >= 1 && n <= VTNET_MAXSEGS);
623 	plen = 0;
624 	tlen = iov[0].iov_len;
625 	for (i = 1; i < n; i++) {
626 		plen += iov[i].iov_len;
627 		tlen += iov[i].iov_len;
628 	}
629 
630 	DPRINTF(("virtio: packet send, %d bytes, %d segs\n\r", plen, n));
631 	sc->pci_vtnet_tx(sc, &iov[1], n - 1, plen);
632 
633 	/* chain is processed, release it and set tlen */
634 	vq_relchain(vq, idx, tlen);
635 }
636 
637 static void
638 pci_vtnet_ping_txq(void *vsc, struct vqueue_info *vq)
639 {
640 	struct pci_vtnet_softc *sc = vsc;
641 
642 	/*
643 	 * Any ring entries to process?
644 	 */
645 	if (!vq_has_descs(vq))
646 		return;
647 
648 	/* Signal the tx thread for processing */
649 	pthread_mutex_lock(&sc->tx_mtx);
650 	vq->vq_used->vu_flags |= VRING_USED_F_NO_NOTIFY;
651 	if (sc->tx_in_progress == 0)
652 		pthread_cond_signal(&sc->tx_cond);
653 	pthread_mutex_unlock(&sc->tx_mtx);
654 }
655 
656 /*
657  * Thread which will handle processing of TX desc
658  */
659 static void *
660 pci_vtnet_tx_thread(void *param)
661 {
662 	struct pci_vtnet_softc *sc = param;
663 	struct vqueue_info *vq;
664 	int error;
665 
666 	vq = &sc->vsc_queues[VTNET_TXQ];
667 
668 	/*
669 	 * Let us wait till the tx queue pointers get initialised &
670 	 * first tx signaled
671 	 */
672 	pthread_mutex_lock(&sc->tx_mtx);
673 	error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx);
674 	assert(error == 0);
675 
676 	for (;;) {
677 		/* note - tx mutex is locked here */
678 		while (sc->resetting || !vq_has_descs(vq)) {
679 			vq->vq_used->vu_flags &= ~VRING_USED_F_NO_NOTIFY;
680 			mb();
681 			if (!sc->resetting && vq_has_descs(vq))
682 				break;
683 
684 			sc->tx_in_progress = 0;
685 			error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx);
686 			assert(error == 0);
687 		}
688 		vq->vq_used->vu_flags |= VRING_USED_F_NO_NOTIFY;
689 		sc->tx_in_progress = 1;
690 		pthread_mutex_unlock(&sc->tx_mtx);
691 
692 		do {
693 			/*
694 			 * Run through entries, placing them into
695 			 * iovecs and sending when an end-of-packet
696 			 * is found
697 			 */
698 			pci_vtnet_proctx(sc, vq);
699 		} while (vq_has_descs(vq));
700 
701 		/*
702 		 * Generate an interrupt if needed.
703 		 */
704 		vq_endchains(vq, 1);
705 
706 		pthread_mutex_lock(&sc->tx_mtx);
707 	}
708 }
709 
710 #ifdef notyet
711 static void
712 pci_vtnet_ping_ctlq(void *vsc, struct vqueue_info *vq)
713 {
714 
715 	DPRINTF(("vtnet: control qnotify!\n\r"));
716 }
717 #endif
718 
719 static int
720 pci_vtnet_parsemac(char *mac_str, uint8_t *mac_addr)
721 {
722         struct ether_addr *ea;
723         char *tmpstr;
724         char zero_addr[ETHER_ADDR_LEN] = { 0, 0, 0, 0, 0, 0 };
725 
726         tmpstr = strsep(&mac_str,"=");
727 
728         if ((mac_str != NULL) && (!strcmp(tmpstr,"mac"))) {
729                 ea = ether_aton(mac_str);
730 
731                 if (ea == NULL || ETHER_IS_MULTICAST(ea->octet) ||
732                     memcmp(ea->octet, zero_addr, ETHER_ADDR_LEN) == 0) {
733 			fprintf(stderr, "Invalid MAC %s\n", mac_str);
734                         return (EINVAL);
735                 } else
736                         memcpy(mac_addr, ea->octet, ETHER_ADDR_LEN);
737         }
738 
739         return (0);
740 }
741 
742 static void
743 pci_vtnet_tap_setup(struct pci_vtnet_softc *sc, char *devname)
744 {
745 	char tbuf[80];
746 
747 	strcpy(tbuf, "/dev/");
748 	strlcat(tbuf, devname, sizeof(tbuf));
749 
750 	sc->pci_vtnet_rx = pci_vtnet_tap_rx;
751 	sc->pci_vtnet_tx = pci_vtnet_tap_tx;
752 
753 	sc->vsc_tapfd = open(tbuf, O_RDWR);
754 	if (sc->vsc_tapfd == -1) {
755 		WPRINTF(("open of tap device %s failed\n", tbuf));
756 		return;
757 	}
758 
759 	/*
760 	 * Set non-blocking and register for read
761 	 * notifications with the event loop
762 	 */
763 	int opt = 1;
764 	if (ioctl(sc->vsc_tapfd, FIONBIO, &opt) < 0) {
765 		WPRINTF(("tap device O_NONBLOCK failed\n"));
766 		close(sc->vsc_tapfd);
767 		sc->vsc_tapfd = -1;
768 	}
769 
770 	sc->vsc_mevp = mevent_add(sc->vsc_tapfd,
771 				  EVF_READ,
772 				  pci_vtnet_rx_callback,
773 				  sc);
774 	if (sc->vsc_mevp == NULL) {
775 		WPRINTF(("Could not register event\n"));
776 		close(sc->vsc_tapfd);
777 		sc->vsc_tapfd = -1;
778 	}
779 }
780 
781 static void
782 pci_vtnet_netmap_setup(struct pci_vtnet_softc *sc, char *ifname)
783 {
784 	sc->pci_vtnet_rx = pci_vtnet_netmap_rx;
785 	sc->pci_vtnet_tx = pci_vtnet_netmap_tx;
786 
787 	sc->vsc_nmd = nm_open(ifname, NULL, 0, 0);
788 	if (sc->vsc_nmd == NULL) {
789 		WPRINTF(("open of netmap device %s failed\n", ifname));
790 		return;
791 	}
792 
793 	sc->vsc_mevp = mevent_add(sc->vsc_nmd->fd,
794 				  EVF_READ,
795 				  pci_vtnet_rx_callback,
796 				  sc);
797 	if (sc->vsc_mevp == NULL) {
798 		WPRINTF(("Could not register event\n"));
799 		nm_close(sc->vsc_nmd);
800 		sc->vsc_nmd = NULL;
801 	}
802 }
803 
804 static int
805 pci_vtnet_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts)
806 {
807 	MD5_CTX mdctx;
808 	unsigned char digest[16];
809 	char nstr[80];
810 	char tname[MAXCOMLEN + 1];
811 	struct pci_vtnet_softc *sc;
812 	char *devname;
813 	char *vtopts;
814 	int mac_provided;
815 
816 	sc = calloc(1, sizeof(struct pci_vtnet_softc));
817 
818 	pthread_mutex_init(&sc->vsc_mtx, NULL);
819 
820 	vi_softc_linkup(&sc->vsc_vs, &vtnet_vi_consts, sc, pi, sc->vsc_queues);
821 	sc->vsc_vs.vs_mtx = &sc->vsc_mtx;
822 
823 	sc->vsc_queues[VTNET_RXQ].vq_qsize = VTNET_RINGSZ;
824 	sc->vsc_queues[VTNET_RXQ].vq_notify = pci_vtnet_ping_rxq;
825 	sc->vsc_queues[VTNET_TXQ].vq_qsize = VTNET_RINGSZ;
826 	sc->vsc_queues[VTNET_TXQ].vq_notify = pci_vtnet_ping_txq;
827 #ifdef notyet
828 	sc->vsc_queues[VTNET_CTLQ].vq_qsize = VTNET_RINGSZ;
829         sc->vsc_queues[VTNET_CTLQ].vq_notify = pci_vtnet_ping_ctlq;
830 #endif
831 
832 	/*
833 	 * Attempt to open the tap device and read the MAC address
834 	 * if specified
835 	 */
836 	mac_provided = 0;
837 	sc->vsc_tapfd = -1;
838 	sc->vsc_nmd = NULL;
839 	if (opts != NULL) {
840 		int err;
841 
842 		devname = vtopts = strdup(opts);
843 		(void) strsep(&vtopts, ",");
844 
845 		if (vtopts != NULL) {
846 			err = pci_vtnet_parsemac(vtopts, sc->vsc_config.mac);
847 			if (err != 0) {
848 				free(devname);
849 				return (err);
850 			}
851 			mac_provided = 1;
852 		}
853 
854 		if (strncmp(devname, "vale", 4) == 0)
855 			pci_vtnet_netmap_setup(sc, devname);
856 		if (strncmp(devname, "tap", 3) == 0 ||
857 		    strncmp(devname, "vmnet", 5) == 0)
858 			pci_vtnet_tap_setup(sc, devname);
859 
860 		free(devname);
861 	}
862 
863 	/*
864 	 * The default MAC address is the standard NetApp OUI of 00-a0-98,
865 	 * followed by an MD5 of the PCI slot/func number and dev name
866 	 */
867 	if (!mac_provided) {
868 		snprintf(nstr, sizeof(nstr), "%d-%d-%s", pi->pi_slot,
869 		    pi->pi_func, vmname);
870 
871 		MD5Init(&mdctx);
872 		MD5Update(&mdctx, nstr, strlen(nstr));
873 		MD5Final(digest, &mdctx);
874 
875 		sc->vsc_config.mac[0] = 0x00;
876 		sc->vsc_config.mac[1] = 0xa0;
877 		sc->vsc_config.mac[2] = 0x98;
878 		sc->vsc_config.mac[3] = digest[0];
879 		sc->vsc_config.mac[4] = digest[1];
880 		sc->vsc_config.mac[5] = digest[2];
881 	}
882 
883 	/* initialize config space */
884 	pci_set_cfgdata16(pi, PCIR_DEVICE, VIRTIO_DEV_NET);
885 	pci_set_cfgdata16(pi, PCIR_VENDOR, VIRTIO_VENDOR);
886 	pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_NETWORK);
887 	pci_set_cfgdata16(pi, PCIR_SUBDEV_0, VIRTIO_TYPE_NET);
888 	pci_set_cfgdata16(pi, PCIR_SUBVEND_0, VIRTIO_VENDOR);
889 
890 	/* Link is up if we managed to open tap device or vale port. */
891 	sc->vsc_config.status = (opts == NULL || sc->vsc_tapfd >= 0 ||
892 	    sc->vsc_nmd != NULL);
893 
894 	/* use BAR 1 to map MSI-X table and PBA, if we're using MSI-X */
895 	if (vi_intr_init(&sc->vsc_vs, 1, fbsdrun_virtio_msix()))
896 		return (1);
897 
898 	/* use BAR 0 to map config regs in IO space */
899 	vi_set_io_bar(&sc->vsc_vs, 0);
900 
901 	sc->resetting = 0;
902 
903 	sc->rx_merge = 1;
904 	sc->rx_vhdrlen = sizeof(struct virtio_net_rxhdr);
905 	sc->rx_in_progress = 0;
906 	pthread_mutex_init(&sc->rx_mtx, NULL);
907 
908 	/*
909 	 * Initialize tx semaphore & spawn TX processing thread.
910 	 * As of now, only one thread for TX desc processing is
911 	 * spawned.
912 	 */
913 	sc->tx_in_progress = 0;
914 	pthread_mutex_init(&sc->tx_mtx, NULL);
915 	pthread_cond_init(&sc->tx_cond, NULL);
916 	pthread_create(&sc->tx_tid, NULL, pci_vtnet_tx_thread, (void *)sc);
917 	snprintf(tname, sizeof(tname), "vtnet-%d:%d tx", pi->pi_slot,
918 	    pi->pi_func);
919         pthread_set_name_np(sc->tx_tid, tname);
920 
921 	return (0);
922 }
923 
924 static int
925 pci_vtnet_cfgwrite(void *vsc, int offset, int size, uint32_t value)
926 {
927 	struct pci_vtnet_softc *sc = vsc;
928 	void *ptr;
929 
930 	if (offset < 6) {
931 		assert(offset + size <= 6);
932 		/*
933 		 * The driver is allowed to change the MAC address
934 		 */
935 		ptr = &sc->vsc_config.mac[offset];
936 		memcpy(ptr, &value, size);
937 	} else {
938 		/* silently ignore other writes */
939 		DPRINTF(("vtnet: write to readonly reg %d\n\r", offset));
940 	}
941 
942 	return (0);
943 }
944 
945 static int
946 pci_vtnet_cfgread(void *vsc, int offset, int size, uint32_t *retval)
947 {
948 	struct pci_vtnet_softc *sc = vsc;
949 	void *ptr;
950 
951 	ptr = (uint8_t *)&sc->vsc_config + offset;
952 	memcpy(retval, ptr, size);
953 	return (0);
954 }
955 
956 static void
957 pci_vtnet_neg_features(void *vsc, uint64_t negotiated_features)
958 {
959 	struct pci_vtnet_softc *sc = vsc;
960 
961 	sc->vsc_features = negotiated_features;
962 
963 	if (!(sc->vsc_features & VIRTIO_NET_F_MRG_RXBUF)) {
964 		sc->rx_merge = 0;
965 		/* non-merge rx header is 2 bytes shorter */
966 		sc->rx_vhdrlen -= 2;
967 	}
968 }
969 
970 struct pci_devemu pci_de_vnet = {
971 	.pe_emu = 	"virtio-net",
972 	.pe_init =	pci_vtnet_init,
973 	.pe_barwrite =	vi_pci_write,
974 	.pe_barread =	vi_pci_read
975 };
976 PCI_EMUL_SET(pci_de_vnet);
977