xref: /illumos-gate/usr/src/cmd/bhyve/pci_virtio_net.c (revision 0554d5ecd11d9644cbb915be31b5a0b7abb40122)
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 #include <sys/linker_set.h>
36 #include <sys/select.h>
37 #include <sys/uio.h>
38 #include <sys/ioctl.h>
39 #include <net/ethernet.h>
40 #include <net/if.h> /* IFNAMSIZ */
41 
42 #include <err.h>
43 #include <errno.h>
44 #include <fcntl.h>
45 #include <stdio.h>
46 #include <stdlib.h>
47 #include <stdint.h>
48 #include <string.h>
49 #include <strings.h>
50 #include <unistd.h>
51 #include <assert.h>
52 #include <pthread.h>
53 #include <pthread_np.h>
54 
55 #include "bhyverun.h"
56 #include "config.h"
57 #include "debug.h"
58 #include "pci_emul.h"
59 #include "mevent.h"
60 #include "virtio.h"
61 #include "net_utils.h"
62 #include "net_backends.h"
63 #include "iov.h"
64 
65 #define VTNET_RINGSZ	1024
66 
67 #define VTNET_MAXSEGS	256
68 
69 #define VTNET_MAX_PKT_LEN	(65536 + 64)
70 
71 #define VTNET_MIN_MTU	ETHERMIN
72 #define VTNET_MAX_MTU	65535
73 
74 #define VTNET_S_HOSTCAPS      \
75   ( VIRTIO_NET_F_MAC | VIRTIO_NET_F_STATUS | \
76     VIRTIO_F_NOTIFY_ON_EMPTY | VIRTIO_RING_F_INDIRECT_DESC)
77 
78 /*
79  * PCI config-space "registers"
80  */
81 struct virtio_net_config {
82 	uint8_t  mac[6];
83 	uint16_t status;
84 	uint16_t max_virtqueue_pairs;
85 	uint16_t mtu;
86 } __packed;
87 
88 /*
89  * Queue definitions.
90  */
91 #define VTNET_RXQ	0
92 #define VTNET_TXQ	1
93 #define VTNET_CTLQ	2	/* NB: not yet supported */
94 
95 #define VTNET_MAXQ	3
96 
97 /*
98  * Debug printf
99  */
100 static int pci_vtnet_debug;
101 #define DPRINTF(params) if (pci_vtnet_debug) PRINTLN params
102 #define WPRINTF(params) PRINTLN params
103 
104 /*
105  * Per-device softc
106  */
107 struct pci_vtnet_softc {
108 	struct virtio_softc vsc_vs;
109 	struct vqueue_info vsc_queues[VTNET_MAXQ - 1];
110 	pthread_mutex_t vsc_mtx;
111 
112 	net_backend_t	*vsc_be;
113 
114 	bool    features_negotiated;	/* protected by rx_mtx */
115 
116 	int		resetting;	/* protected by tx_mtx */
117 
118 	uint64_t	vsc_features;	/* negotiated features */
119 
120 	pthread_mutex_t	rx_mtx;
121 	int		rx_merge;	/* merged rx bufs in use */
122 
123 	pthread_t 	tx_tid;
124 	pthread_mutex_t	tx_mtx;
125 	pthread_cond_t	tx_cond;
126 	int		tx_in_progress;
127 
128 	size_t		vhdrlen;
129 	size_t		be_vhdrlen;
130 
131 	struct virtio_net_config vsc_config;
132 	struct virtio_consts vsc_consts;
133 };
134 
135 static void pci_vtnet_reset(void *);
136 /* static void pci_vtnet_notify(void *, struct vqueue_info *); */
137 static int pci_vtnet_cfgread(void *, int, int, uint32_t *);
138 static int pci_vtnet_cfgwrite(void *, int, int, uint32_t);
139 static void pci_vtnet_neg_features(void *, uint64_t);
140 
141 static struct virtio_consts vtnet_vi_consts = {
142 	"vtnet",		/* our name */
143 	VTNET_MAXQ - 1,		/* we currently support 2 virtqueues */
144 	sizeof(struct virtio_net_config), /* config reg size */
145 	pci_vtnet_reset,	/* reset */
146 	NULL,			/* device-wide qnotify -- not used */
147 	pci_vtnet_cfgread,	/* read PCI config */
148 	pci_vtnet_cfgwrite,	/* write PCI config */
149 	pci_vtnet_neg_features,	/* apply negotiated features */
150 	VTNET_S_HOSTCAPS,	/* our capabilities */
151 };
152 
153 static void
154 pci_vtnet_reset(void *vsc)
155 {
156 	struct pci_vtnet_softc *sc = vsc;
157 
158 	DPRINTF(("vtnet: device reset requested !"));
159 
160 	/* Acquire the RX lock to block RX processing. */
161 	pthread_mutex_lock(&sc->rx_mtx);
162 
163 	/*
164 	 * Make sure receive operation is disabled at least until we
165 	 * re-negotiate the features, since receive operation depends
166 	 * on the value of sc->rx_merge and the header length, which
167 	 * are both set in pci_vtnet_neg_features().
168 	 * Receive operation will be enabled again once the guest adds
169 	 * the first receive buffers and kicks us.
170 	 */
171 	sc->features_negotiated = false;
172 	netbe_rx_disable(sc->vsc_be);
173 
174 	/* Set sc->resetting and give a chance to the TX thread to stop. */
175 	pthread_mutex_lock(&sc->tx_mtx);
176 	sc->resetting = 1;
177 	while (sc->tx_in_progress) {
178 		pthread_mutex_unlock(&sc->tx_mtx);
179 		usleep(10000);
180 		pthread_mutex_lock(&sc->tx_mtx);
181 	}
182 
183 	/*
184 	 * Now reset rings, MSI-X vectors, and negotiated capabilities.
185 	 * Do that with the TX lock held, since we need to reset
186 	 * sc->resetting.
187 	 */
188 	vi_reset_dev(&sc->vsc_vs);
189 
190 	sc->resetting = 0;
191 	pthread_mutex_unlock(&sc->tx_mtx);
192 	pthread_mutex_unlock(&sc->rx_mtx);
193 }
194 
195 static __inline struct iovec *
196 iov_trim_hdr(struct iovec *iov, int *iovcnt, unsigned int hlen)
197 {
198 	struct iovec *riov;
199 
200 	if (iov[0].iov_len < hlen) {
201 		/*
202 		 * Not enough header space in the first fragment.
203 		 * That's not ok for us.
204 		 */
205 		return NULL;
206 	}
207 
208 	iov[0].iov_len -= hlen;
209 	if (iov[0].iov_len == 0) {
210 		*iovcnt -= 1;
211 		if (*iovcnt == 0) {
212 			/*
213 			 * Only space for the header. That's not
214 			 * enough for us.
215 			 */
216 			return NULL;
217 		}
218 		riov = &iov[1];
219 	} else {
220 		iov[0].iov_base = (void *)((uintptr_t)iov[0].iov_base + hlen);
221 		riov = &iov[0];
222 	}
223 
224 	return (riov);
225 }
226 
227 struct virtio_mrg_rxbuf_info {
228 	uint16_t idx;
229 	uint16_t pad;
230 	uint32_t len;
231 };
232 
233 static void
234 pci_vtnet_rx(struct pci_vtnet_softc *sc)
235 {
236 	int prepend_hdr_len = sc->vhdrlen - sc->be_vhdrlen;
237 	struct virtio_mrg_rxbuf_info info[VTNET_MAXSEGS];
238 	struct iovec iov[VTNET_MAXSEGS + 1];
239 	struct vqueue_info *vq;
240 
241 	vq = &sc->vsc_queues[VTNET_RXQ];
242 
243 	/* Features must be negotiated */
244 	if (!sc->features_negotiated) {
245 		return;
246 	}
247 
248 	for (;;) {
249 		struct virtio_net_rxhdr *hdr;
250 		uint32_t riov_bytes;
251 		struct iovec *riov;
252 		uint32_t ulen;
253 		int riov_len;
254 		int n_chains;
255 		ssize_t rlen;
256 		ssize_t plen;
257 
258 		plen = netbe_peek_recvlen(sc->vsc_be);
259 		if (plen <= 0) {
260 			/*
261 			 * No more packets (plen == 0), or backend errored
262 			 * (plen < 0). Interrupt if needed and stop.
263 			 */
264 			vq_endchains(vq, /*used_all_avail=*/0);
265 			return;
266 		}
267 		plen += prepend_hdr_len;
268 
269 		/*
270 		 * Get a descriptor chain to store the next ingress
271 		 * packet. In case of mergeable rx buffers, get as
272 		 * many chains as necessary in order to make room
273 		 * for plen bytes.
274 		 */
275 		riov_bytes = 0;
276 		riov_len = 0;
277 		riov = iov;
278 		n_chains = 0;
279 		do {
280 			int n = vq_getchain(vq, &info[n_chains].idx, riov,
281 			    VTNET_MAXSEGS - riov_len, NULL);
282 
283 			if (n == 0) {
284 				/*
285 				 * No rx buffers. Enable RX kicks and double
286 				 * check.
287 				 */
288 				vq_kick_enable(vq);
289 				if (!vq_has_descs(vq)) {
290 					/*
291 					 * Still no buffers. Return the unused
292 					 * chains (if any), interrupt if needed
293 					 * (including for NOTIFY_ON_EMPTY), and
294 					 * disable the backend until the next
295 					 * kick.
296 					 */
297 					vq_retchains(vq, n_chains);
298 					vq_endchains(vq, /*used_all_avail=*/1);
299 					netbe_rx_disable(sc->vsc_be);
300 					return;
301 				}
302 
303 				/* More rx buffers found, so keep going. */
304 				vq_kick_disable(vq);
305 				continue;
306 			}
307 #ifndef __FreeBSD__
308 			if (n == -1) {
309 				/*
310 				 * An error from vq_getchain() means that
311 				 * an invalid descriptor was found.
312 				 */
313 				vq_retchains(vq, n_chains);
314 				vq_endchains(vq, /*used_all_avail=*/0);
315 				return;
316 			}
317 #endif
318 			assert(n >= 1 && riov_len + n <= VTNET_MAXSEGS);
319 			riov_len += n;
320 			if (!sc->rx_merge) {
321 				n_chains = 1;
322 				break;
323 			}
324 #ifndef __FreeBSD__
325 			size_t c = count_iov(riov, n);
326 			if (c > UINT32_MAX) {
327 				vq_retchains(vq, n_chains);
328 				vq_endchains(vq, /*used_all_avail=*/0);
329 				return;
330 			}
331 			info[n_chains].len = (uint32_t)c;
332 #else
333 			info[n_chains].len = (uint32_t)count_iov(riov, n);
334 #endif
335 			riov_bytes += info[n_chains].len;
336 			riov += n;
337 			n_chains++;
338 		} while (riov_bytes < plen && riov_len < VTNET_MAXSEGS);
339 
340 		riov = iov;
341 #ifdef __FreeBSD__
342 		hdr = riov[0].iov_base;
343 #else
344 		hdr = (struct virtio_net_rxhdr *)riov[0].iov_base;
345 #endif
346 		if (prepend_hdr_len > 0) {
347 			/*
348 			 * The frontend uses a virtio-net header, but the
349 			 * backend does not. We need to prepend a zeroed
350 			 * header.
351 			 */
352 			riov = iov_trim_hdr(riov, &riov_len, prepend_hdr_len);
353 			if (riov == NULL) {
354 				/*
355 				 * The first collected chain is nonsensical,
356 				 * as it is not even enough to store the
357 				 * virtio-net header. Just drop it.
358 				 */
359 				vq_relchain(vq, info[0].idx, 0);
360 				vq_retchains(vq, n_chains - 1);
361 				continue;
362 			}
363 			memset(hdr, 0, prepend_hdr_len);
364 		}
365 
366 		rlen = netbe_recv(sc->vsc_be, riov, riov_len);
367 		if (rlen != plen - prepend_hdr_len) {
368 			/*
369 			 * If this happens it means there is something
370 			 * wrong with the backend (e.g., some other
371 			 * process is stealing our packets).
372 			 */
373 			WPRINTF(("netbe_recv: expected %zd bytes, "
374 				"got %zd", plen - prepend_hdr_len, rlen));
375 			vq_retchains(vq, n_chains);
376 			continue;
377 		}
378 
379 		ulen = (uint32_t)plen;
380 
381 		/*
382 		 * Publish the used buffers to the guest, reporting the
383 		 * number of bytes that we wrote.
384 		 */
385 		if (!sc->rx_merge) {
386 			vq_relchain(vq, info[0].idx, ulen);
387 		} else {
388 			uint32_t iolen;
389 			int i = 0;
390 
391 			do {
392 				iolen = info[i].len;
393 				if (iolen > ulen) {
394 					iolen = ulen;
395 				}
396 				vq_relchain_prepare(vq, info[i].idx, iolen);
397 				ulen -= iolen;
398 				i++;
399 			} while (ulen > 0);
400 
401 			hdr->vrh_bufs = i;
402 			vq_relchain_publish(vq);
403 			assert(i == n_chains);
404 		}
405 	}
406 
407 }
408 
409 /*
410  * Called when there is read activity on the backend file descriptor.
411  * Each buffer posted by the guest is assumed to be able to contain
412  * an entire ethernet frame + rx header.
413  */
414 static void
415 pci_vtnet_rx_callback(int fd, enum ev_type type, void *param)
416 {
417 	struct pci_vtnet_softc *sc = param;
418 
419 	pthread_mutex_lock(&sc->rx_mtx);
420 	pci_vtnet_rx(sc);
421 	pthread_mutex_unlock(&sc->rx_mtx);
422 
423 }
424 
425 /* Called on RX kick. */
426 static void
427 pci_vtnet_ping_rxq(void *vsc, struct vqueue_info *vq)
428 {
429 	struct pci_vtnet_softc *sc = vsc;
430 
431 	/*
432 	 * A qnotify means that the rx process can now begin.
433 	 * Enable RX only if features are negotiated.
434 	 */
435 	pthread_mutex_lock(&sc->rx_mtx);
436 	if (!sc->features_negotiated) {
437 		pthread_mutex_unlock(&sc->rx_mtx);
438 		return;
439 	}
440 
441 	vq_kick_disable(vq);
442 	netbe_rx_enable(sc->vsc_be);
443 	pthread_mutex_unlock(&sc->rx_mtx);
444 }
445 
446 /* TX virtqueue processing, called by the TX thread. */
447 static void
448 pci_vtnet_proctx(struct pci_vtnet_softc *sc, struct vqueue_info *vq)
449 {
450 	struct iovec iov[VTNET_MAXSEGS + 1];
451 	struct iovec *siov = iov;
452 	uint16_t idx;
453 	ssize_t len;
454 	int n;
455 
456 	/*
457 	 * Obtain chain of descriptors. The first descriptor also
458 	 * contains the virtio-net header.
459 	 */
460 	n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL);
461 	assert(n >= 1 && n <= VTNET_MAXSEGS);
462 
463 	if (sc->vhdrlen != sc->be_vhdrlen) {
464 		/*
465 		 * The frontend uses a virtio-net header, but the backend
466 		 * does not. We simply strip the header and ignore it, as
467 		 * it should be zero-filled.
468 		 */
469 		siov = iov_trim_hdr(siov, &n, sc->vhdrlen);
470 	}
471 
472 	if (siov == NULL) {
473 		/* The chain is nonsensical. Just drop it. */
474 		len = 0;
475 	} else {
476 		len = netbe_send(sc->vsc_be, siov, n);
477 		if (len < 0) {
478 			/*
479 			 * If send failed, report that 0 bytes
480 			 * were read.
481 			 */
482 			len = 0;
483 		}
484 	}
485 
486 	/*
487 	 * Return the processed chain to the guest, reporting
488 	 * the number of bytes that we read.
489 	 */
490 	vq_relchain(vq, idx, len);
491 }
492 
493 /* Called on TX kick. */
494 static void
495 pci_vtnet_ping_txq(void *vsc, struct vqueue_info *vq)
496 {
497 	struct pci_vtnet_softc *sc = vsc;
498 
499 	/*
500 	 * Any ring entries to process?
501 	 */
502 	if (!vq_has_descs(vq))
503 		return;
504 
505 	/* Signal the tx thread for processing */
506 	pthread_mutex_lock(&sc->tx_mtx);
507 	vq_kick_disable(vq);
508 	if (sc->tx_in_progress == 0)
509 		pthread_cond_signal(&sc->tx_cond);
510 	pthread_mutex_unlock(&sc->tx_mtx);
511 }
512 
513 /*
514  * Thread which will handle processing of TX desc
515  */
516 static void *
517 pci_vtnet_tx_thread(void *param)
518 {
519 	struct pci_vtnet_softc *sc = param;
520 	struct vqueue_info *vq;
521 	int error;
522 
523 	vq = &sc->vsc_queues[VTNET_TXQ];
524 
525 	/*
526 	 * Let us wait till the tx queue pointers get initialised &
527 	 * first tx signaled
528 	 */
529 	pthread_mutex_lock(&sc->tx_mtx);
530 	error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx);
531 	assert(error == 0);
532 
533 	for (;;) {
534 		/* note - tx mutex is locked here */
535 		while (sc->resetting || !vq_has_descs(vq)) {
536 			vq_kick_enable(vq);
537 			if (!sc->resetting && vq_has_descs(vq))
538 				break;
539 
540 			sc->tx_in_progress = 0;
541 			error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx);
542 			assert(error == 0);
543 		}
544 		vq_kick_disable(vq);
545 		sc->tx_in_progress = 1;
546 		pthread_mutex_unlock(&sc->tx_mtx);
547 
548 		do {
549 			/*
550 			 * Run through entries, placing them into
551 			 * iovecs and sending when an end-of-packet
552 			 * is found
553 			 */
554 			pci_vtnet_proctx(sc, vq);
555 		} while (vq_has_descs(vq));
556 
557 		/*
558 		 * Generate an interrupt if needed.
559 		 */
560 		vq_endchains(vq, /*used_all_avail=*/1);
561 
562 		pthread_mutex_lock(&sc->tx_mtx);
563 	}
564 #ifndef __FreeBSD__
565 	return (NULL);
566 #endif
567 }
568 
569 #ifdef notyet
570 static void
571 pci_vtnet_ping_ctlq(void *vsc, struct vqueue_info *vq)
572 {
573 
574 	DPRINTF(("vtnet: control qnotify!"));
575 }
576 #endif
577 
578 static int
579 pci_vtnet_init(struct vmctx *ctx, struct pci_devinst *pi, nvlist_t *nvl)
580 {
581 	struct pci_vtnet_softc *sc;
582 	const char *value;
583 	char tname[MAXCOMLEN + 1];
584 	unsigned long mtu = ETHERMTU;
585 	int err;
586 
587 	/*
588 	 * Allocate data structures for further virtio initializations.
589 	 * sc also contains a copy of vtnet_vi_consts, since capabilities
590 	 * change depending on the backend.
591 	 */
592 	sc = calloc(1, sizeof(struct pci_vtnet_softc));
593 
594 	sc->vsc_consts = vtnet_vi_consts;
595 	pthread_mutex_init(&sc->vsc_mtx, NULL);
596 
597 	sc->vsc_queues[VTNET_RXQ].vq_qsize = VTNET_RINGSZ;
598 	sc->vsc_queues[VTNET_RXQ].vq_notify = pci_vtnet_ping_rxq;
599 	sc->vsc_queues[VTNET_TXQ].vq_qsize = VTNET_RINGSZ;
600 	sc->vsc_queues[VTNET_TXQ].vq_notify = pci_vtnet_ping_txq;
601 #ifdef notyet
602 	sc->vsc_queues[VTNET_CTLQ].vq_qsize = VTNET_RINGSZ;
603         sc->vsc_queues[VTNET_CTLQ].vq_notify = pci_vtnet_ping_ctlq;
604 #endif
605 
606 	value = get_config_value_node(nvl, "mac");
607 	if (value != NULL) {
608 		err = net_parsemac(value, sc->vsc_config.mac);
609 		if (err) {
610 			free(sc);
611 			return (err);
612 		}
613 	} else
614 		net_genmac(pi, sc->vsc_config.mac);
615 
616 	value = get_config_value_node(nvl, "mtu");
617 	if (value != NULL) {
618 		err = net_parsemtu(value, &mtu);
619 		if (err) {
620 			free(sc);
621 			return (err);
622 		}
623 
624 		if (mtu < VTNET_MIN_MTU || mtu > VTNET_MAX_MTU) {
625 			err = EINVAL;
626 			errno = EINVAL;
627 			free(sc);
628 			return (err);
629 		}
630 		sc->vsc_consts.vc_hv_caps |= VIRTIO_NET_F_MTU;
631 	}
632 	sc->vsc_config.mtu = mtu;
633 
634 	/* Permit interfaces without a configured backend. */
635 	if (get_config_value_node(nvl, "backend") != NULL) {
636 		err = netbe_init(&sc->vsc_be, nvl, pci_vtnet_rx_callback, sc);
637 		if (err) {
638 			free(sc);
639 			return (err);
640 		}
641 #ifndef __FreeBSD__
642 		size_t buflen = sizeof (sc->vsc_config.mac);
643 
644 		err = netbe_get_mac(sc->vsc_be, sc->vsc_config.mac, &buflen);
645 		if (err != 0) {
646 		       free(sc);
647 		       return (err);
648 		}
649 #endif
650 	}
651 
652 	sc->vsc_consts.vc_hv_caps |= VIRTIO_NET_F_MRG_RXBUF |
653 	    netbe_get_cap(sc->vsc_be);
654 
655 	/*
656 	 * Since we do not actually support multiqueue,
657 	 * set the maximum virtqueue pairs to 1.
658 	 */
659 	sc->vsc_config.max_virtqueue_pairs = 1;
660 
661 	/* initialize config space */
662 	pci_set_cfgdata16(pi, PCIR_DEVICE, VIRTIO_DEV_NET);
663 	pci_set_cfgdata16(pi, PCIR_VENDOR, VIRTIO_VENDOR);
664 	pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_NETWORK);
665 	pci_set_cfgdata16(pi, PCIR_SUBDEV_0, VIRTIO_ID_NETWORK);
666 	pci_set_cfgdata16(pi, PCIR_SUBVEND_0, VIRTIO_VENDOR);
667 
668 	/* Link is always up. */
669 	sc->vsc_config.status = 1;
670 
671 	vi_softc_linkup(&sc->vsc_vs, &sc->vsc_consts, sc, pi, sc->vsc_queues);
672 	sc->vsc_vs.vs_mtx = &sc->vsc_mtx;
673 
674 	/* use BAR 1 to map MSI-X table and PBA, if we're using MSI-X */
675 	if (vi_intr_init(&sc->vsc_vs, 1, fbsdrun_virtio_msix())) {
676 		free(sc);
677 		return (1);
678 	}
679 
680 	/* use BAR 0 to map config regs in IO space */
681 	vi_set_io_bar(&sc->vsc_vs, 0);
682 
683 	sc->resetting = 0;
684 
685 	sc->rx_merge = 0;
686 	sc->vhdrlen = sizeof(struct virtio_net_rxhdr) - 2;
687 	pthread_mutex_init(&sc->rx_mtx, NULL);
688 
689 	/*
690 	 * Initialize tx semaphore & spawn TX processing thread.
691 	 * As of now, only one thread for TX desc processing is
692 	 * spawned.
693 	 */
694 	sc->tx_in_progress = 0;
695 	pthread_mutex_init(&sc->tx_mtx, NULL);
696 	pthread_cond_init(&sc->tx_cond, NULL);
697 	pthread_create(&sc->tx_tid, NULL, pci_vtnet_tx_thread, (void *)sc);
698 	snprintf(tname, sizeof(tname), "vtnet-%d:%d tx", pi->pi_slot,
699 	    pi->pi_func);
700 	pthread_set_name_np(sc->tx_tid, tname);
701 
702 	return (0);
703 }
704 
705 static int
706 pci_vtnet_cfgwrite(void *vsc, int offset, int size, uint32_t value)
707 {
708 	struct pci_vtnet_softc *sc = vsc;
709 	void *ptr;
710 
711 	if (offset < (int)sizeof(sc->vsc_config.mac)) {
712 		assert(offset + size <= (int)sizeof(sc->vsc_config.mac));
713 		/*
714 		 * The driver is allowed to change the MAC address
715 		 */
716 		ptr = &sc->vsc_config.mac[offset];
717 		memcpy(ptr, &value, size);
718 	} else {
719 		/* silently ignore other writes */
720 		DPRINTF(("vtnet: write to readonly reg %d", offset));
721 	}
722 
723 	return (0);
724 }
725 
726 static int
727 pci_vtnet_cfgread(void *vsc, int offset, int size, uint32_t *retval)
728 {
729 	struct pci_vtnet_softc *sc = vsc;
730 	void *ptr;
731 
732 	ptr = (uint8_t *)&sc->vsc_config + offset;
733 	memcpy(retval, ptr, size);
734 	return (0);
735 }
736 
737 static void
738 pci_vtnet_neg_features(void *vsc, uint64_t negotiated_features)
739 {
740 	struct pci_vtnet_softc *sc = vsc;
741 
742 	sc->vsc_features = negotiated_features;
743 
744 	if (negotiated_features & VIRTIO_NET_F_MRG_RXBUF) {
745 		sc->vhdrlen = sizeof(struct virtio_net_rxhdr);
746 		sc->rx_merge = 1;
747 	} else {
748 		/*
749 		 * Without mergeable rx buffers, virtio-net header is 2
750 		 * bytes shorter than sizeof(struct virtio_net_rxhdr).
751 		 */
752 		sc->vhdrlen = sizeof(struct virtio_net_rxhdr) - 2;
753 		sc->rx_merge = 0;
754 	}
755 
756 	/* Tell the backend to enable some capabilities it has advertised. */
757 	netbe_set_cap(sc->vsc_be, negotiated_features, sc->vhdrlen);
758 	sc->be_vhdrlen = netbe_get_vnet_hdr_len(sc->vsc_be);
759 	assert(sc->be_vhdrlen == 0 || sc->be_vhdrlen == sc->vhdrlen);
760 
761 	pthread_mutex_lock(&sc->rx_mtx);
762 	sc->features_negotiated = true;
763 	pthread_mutex_unlock(&sc->rx_mtx);
764 }
765 
766 static struct pci_devemu pci_de_vnet = {
767 	.pe_emu = 	"virtio-net",
768 	.pe_init =	pci_vtnet_init,
769 	.pe_legacy_config = netbe_legacy_config,
770 	.pe_barwrite =	vi_pci_write,
771 	.pe_barread =	vi_pci_read,
772 };
773 PCI_EMUL_SET(pci_de_vnet);
774