xref: /freebsd/usr.sbin/bhyve/virtio.c (revision 11b249f99e4d4891bd2a21ef796d40073ee15508)
1 /*-
2  * Copyright (c) 2013  Chris Torek <torek @ torek net>
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 THE AUTHOR AND CONTRIBUTORS ``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 THE AUTHOR 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 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include <sys/param.h>
31 #include <sys/uio.h>
32 
33 #include <stdio.h>
34 #include <stdint.h>
35 #include <pthread.h>
36 #include <pthread_np.h>
37 
38 #include "bhyverun.h"
39 #include "pci_emul.h"
40 #include "virtio.h"
41 
42 /*
43  * Functions for dealing with generalized "virtual devices" as
44  * defined by <https://www.google.com/#output=search&q=virtio+spec>
45  */
46 
47 /*
48  * In case we decide to relax the "virtio softc comes at the
49  * front of virtio-based device softc" constraint, let's use
50  * this to convert.
51  */
52 #define DEV_SOFTC(vs) ((void *)(vs))
53 
54 /*
55  * Link a virtio_softc to its constants, the device softc, and
56  * the PCI emulation.
57  */
58 void
59 vi_softc_linkup(struct virtio_softc *vs, struct virtio_consts *vc,
60 		void *dev_softc, struct pci_devinst *pi,
61 		struct vqueue_info *queues)
62 {
63 	int i;
64 
65 	/* vs and dev_softc addresses must match */
66 	assert((void *)vs == dev_softc);
67 	vs->vs_vc = vc;
68 	vs->vs_pi = pi;
69 	pi->pi_arg = vs;
70 
71 	vs->vs_queues = queues;
72 	for (i = 0; i < vc->vc_nvq; i++) {
73 		queues[i].vq_vs = vs;
74 		queues[i].vq_num = i;
75 	}
76 }
77 
78 /*
79  * Reset device (device-wide).  This erases all queues, i.e.,
80  * all the queues become invalid (though we don't wipe out the
81  * internal pointers, we just clear the VQ_ALLOC flag).
82  *
83  * It resets negotiated features to "none".
84  *
85  * If MSI-X is enabled, this also resets all the vectors to NO_VECTOR.
86  */
87 void
88 vi_reset_dev(struct virtio_softc *vs)
89 {
90 	struct vqueue_info *vq;
91 	int i, nvq;
92 
93 	if (vs->vs_mtx)
94 		assert(pthread_mutex_isowned_np(vs->vs_mtx));
95 
96 	nvq = vs->vs_vc->vc_nvq;
97 	for (vq = vs->vs_queues, i = 0; i < nvq; vq++, i++) {
98 		vq->vq_flags = 0;
99 		vq->vq_last_avail = 0;
100 		vq->vq_save_used = 0;
101 		vq->vq_pfn = 0;
102 		vq->vq_msix_idx = VIRTIO_MSI_NO_VECTOR;
103 	}
104 	vs->vs_negotiated_caps = 0;
105 	vs->vs_curq = 0;
106 	/* vs->vs_status = 0; -- redundant */
107 	if (vs->vs_isr)
108 		pci_lintr_deassert(vs->vs_pi);
109 	vs->vs_isr = 0;
110 	vs->vs_msix_cfg_idx = VIRTIO_MSI_NO_VECTOR;
111 }
112 
113 /*
114  * Set I/O BAR (usually 0) to map PCI config registers.
115  */
116 void
117 vi_set_io_bar(struct virtio_softc *vs, int barnum)
118 {
119 	size_t size;
120 
121 	/*
122 	 * ??? should we use CFG0 if MSI-X is disabled?
123 	 * Existing code did not...
124 	 */
125 	size = VTCFG_R_CFG1 + vs->vs_vc->vc_cfgsize;
126 	pci_emul_alloc_bar(vs->vs_pi, barnum, PCIBAR_IO, size);
127 }
128 
129 /*
130  * Initialize MSI-X vector capabilities if we're to use MSI-X,
131  * or MSI capabilities if not.
132  *
133  * We assume we want one MSI-X vector per queue, here, plus one
134  * for the config vec.
135  */
136 int
137 vi_intr_init(struct virtio_softc *vs, int barnum, int use_msix)
138 {
139 	int nvec;
140 
141 	if (use_msix) {
142 		vs->vs_flags |= VIRTIO_USE_MSIX;
143 		VS_LOCK(vs);
144 		vi_reset_dev(vs); /* set all vectors to NO_VECTOR */
145 		VS_UNLOCK(vs);
146 		nvec = vs->vs_vc->vc_nvq + 1;
147 		if (pci_emul_add_msixcap(vs->vs_pi, nvec, barnum))
148 			return (1);
149 	} else
150 		vs->vs_flags &= ~VIRTIO_USE_MSIX;
151 
152 	/* Only 1 MSI vector for bhyve */
153 	pci_emul_add_msicap(vs->vs_pi, 1);
154 
155 	/* Legacy interrupts are mandatory for virtio devices */
156 	pci_lintr_request(vs->vs_pi);
157 
158 	return (0);
159 }
160 
161 /*
162  * Initialize the currently-selected virtio queue (vs->vs_curq).
163  * The guest just gave us a page frame number, from which we can
164  * calculate the addresses of the queue.
165  */
166 void
167 vi_vq_init(struct virtio_softc *vs, uint32_t pfn)
168 {
169 	struct vqueue_info *vq;
170 	uint64_t phys;
171 	size_t size;
172 	char *base;
173 
174 	vq = &vs->vs_queues[vs->vs_curq];
175 	vq->vq_pfn = pfn;
176 	phys = (uint64_t)pfn << VRING_PFN;
177 	size = vring_size(vq->vq_qsize);
178 	base = paddr_guest2host(vs->vs_pi->pi_vmctx, phys, size);
179 
180 	/* First page(s) are descriptors... */
181 	vq->vq_desc = (struct virtio_desc *)base;
182 	base += vq->vq_qsize * sizeof(struct virtio_desc);
183 
184 	/* ... immediately followed by "avail" ring (entirely uint16_t's) */
185 	vq->vq_avail = (struct vring_avail *)base;
186 	base += (2 + vq->vq_qsize + 1) * sizeof(uint16_t);
187 
188 	/* Then it's rounded up to the next page... */
189 	base = (char *)roundup2((uintptr_t)base, VRING_ALIGN);
190 
191 	/* ... and the last page(s) are the used ring. */
192 	vq->vq_used = (struct vring_used *)base;
193 
194 	/* Mark queue as allocated, and start at 0 when we use it. */
195 	vq->vq_flags = VQ_ALLOC;
196 	vq->vq_last_avail = 0;
197 	vq->vq_save_used = 0;
198 }
199 
200 /*
201  * Helper inline for vq_getchain(): record the i'th "real"
202  * descriptor.
203  */
204 static inline void
205 _vq_record(int i, volatile struct virtio_desc *vd, struct vmctx *ctx,
206 	   struct iovec *iov, int n_iov, uint16_t *flags) {
207 
208 	if (i >= n_iov)
209 		return;
210 	iov[i].iov_base = paddr_guest2host(ctx, vd->vd_addr, vd->vd_len);
211 	iov[i].iov_len = vd->vd_len;
212 	if (flags != NULL)
213 		flags[i] = vd->vd_flags;
214 }
215 #define	VQ_MAX_DESCRIPTORS	512	/* see below */
216 
217 /*
218  * Examine the chain of descriptors starting at the "next one" to
219  * make sure that they describe a sensible request.  If so, return
220  * the number of "real" descriptors that would be needed/used in
221  * acting on this request.  This may be smaller than the number of
222  * available descriptors, e.g., if there are two available but
223  * they are two separate requests, this just returns 1.  Or, it
224  * may be larger: if there are indirect descriptors involved,
225  * there may only be one descriptor available but it may be an
226  * indirect pointing to eight more.  We return 8 in this case,
227  * i.e., we do not count the indirect descriptors, only the "real"
228  * ones.
229  *
230  * Basically, this vets the vd_flags and vd_next field of each
231  * descriptor and tells you how many are involved.  Since some may
232  * be indirect, this also needs the vmctx (in the pci_devinst
233  * at vs->vs_pi) so that it can find indirect descriptors.
234  *
235  * As we process each descriptor, we copy and adjust it (guest to
236  * host address wise, also using the vmtctx) into the given iov[]
237  * array (of the given size).  If the array overflows, we stop
238  * placing values into the array but keep processing descriptors,
239  * up to VQ_MAX_DESCRIPTORS, before giving up and returning -1.
240  * So you, the caller, must not assume that iov[] is as big as the
241  * return value (you can process the same thing twice to allocate
242  * a larger iov array if needed, or supply a zero length to find
243  * out how much space is needed).
244  *
245  * If you want to verify the WRITE flag on each descriptor, pass a
246  * non-NULL "flags" pointer to an array of "uint16_t" of the same size
247  * as n_iov and we'll copy each vd_flags field after unwinding any
248  * indirects.
249  *
250  * If some descriptor(s) are invalid, this prints a diagnostic message
251  * and returns -1.  If no descriptors are ready now it simply returns 0.
252  *
253  * You are assumed to have done a vq_ring_ready() if needed (note
254  * that vq_has_descs() does one).
255  */
256 int
257 vq_getchain(struct vqueue_info *vq, uint16_t *pidx,
258 	    struct iovec *iov, int n_iov, uint16_t *flags)
259 {
260 	int i;
261 	u_int ndesc, n_indir;
262 	u_int idx, next;
263 	volatile struct virtio_desc *vdir, *vindir, *vp;
264 	struct vmctx *ctx;
265 	struct virtio_softc *vs;
266 	const char *name;
267 
268 	vs = vq->vq_vs;
269 	name = vs->vs_vc->vc_name;
270 
271 	/*
272 	 * Note: it's the responsibility of the guest not to
273 	 * update vq->vq_avail->va_idx until all of the descriptors
274          * the guest has written are valid (including all their
275          * vd_next fields and vd_flags).
276 	 *
277 	 * Compute (last_avail - va_idx) in integers mod 2**16.  This is
278 	 * the number of descriptors the device has made available
279 	 * since the last time we updated vq->vq_last_avail.
280 	 *
281 	 * We just need to do the subtraction as an unsigned int,
282 	 * then trim off excess bits.
283 	 */
284 	idx = vq->vq_last_avail;
285 	ndesc = (uint16_t)((u_int)vq->vq_avail->va_idx - idx);
286 	if (ndesc == 0)
287 		return (0);
288 	if (ndesc > vq->vq_qsize) {
289 		/* XXX need better way to diagnose issues */
290 		fprintf(stderr,
291 		    "%s: ndesc (%u) out of range, driver confused?\r\n",
292 		    name, (u_int)ndesc);
293 		return (-1);
294 	}
295 
296 	/*
297 	 * Now count/parse "involved" descriptors starting from
298 	 * the head of the chain.
299 	 *
300 	 * To prevent loops, we could be more complicated and
301 	 * check whether we're re-visiting a previously visited
302 	 * index, but we just abort if the count gets excessive.
303 	 */
304 	ctx = vs->vs_pi->pi_vmctx;
305 	*pidx = next = vq->vq_avail->va_ring[idx & (vq->vq_qsize - 1)];
306 	vq->vq_last_avail++;
307 	for (i = 0; i < VQ_MAX_DESCRIPTORS; next = vdir->vd_next) {
308 		if (next >= vq->vq_qsize) {
309 			fprintf(stderr,
310 			    "%s: descriptor index %u out of range, "
311 			    "driver confused?\r\n",
312 			    name, next);
313 			return (-1);
314 		}
315 		vdir = &vq->vq_desc[next];
316 		if ((vdir->vd_flags & VRING_DESC_F_INDIRECT) == 0) {
317 			_vq_record(i, vdir, ctx, iov, n_iov, flags);
318 			i++;
319 		} else if ((vs->vs_vc->vc_hv_caps &
320 		    VIRTIO_RING_F_INDIRECT_DESC) == 0) {
321 			fprintf(stderr,
322 			    "%s: descriptor has forbidden INDIRECT flag, "
323 			    "driver confused?\r\n",
324 			    name);
325 			return (-1);
326 		} else {
327 			n_indir = vdir->vd_len / 16;
328 			if ((vdir->vd_len & 0xf) || n_indir == 0) {
329 				fprintf(stderr,
330 				    "%s: invalid indir len 0x%x, "
331 				    "driver confused?\r\n",
332 				    name, (u_int)vdir->vd_len);
333 				return (-1);
334 			}
335 			vindir = paddr_guest2host(ctx,
336 			    vdir->vd_addr, vdir->vd_len);
337 			/*
338 			 * Indirects start at the 0th, then follow
339 			 * their own embedded "next"s until those run
340 			 * out.  Each one's indirect flag must be off
341 			 * (we don't really have to check, could just
342 			 * ignore errors...).
343 			 */
344 			next = 0;
345 			for (;;) {
346 				vp = &vindir[next];
347 				if (vp->vd_flags & VRING_DESC_F_INDIRECT) {
348 					fprintf(stderr,
349 					    "%s: indirect desc has INDIR flag,"
350 					    " driver confused?\r\n",
351 					    name);
352 					return (-1);
353 				}
354 				_vq_record(i, vp, ctx, iov, n_iov, flags);
355 				if (++i > VQ_MAX_DESCRIPTORS)
356 					goto loopy;
357 				if ((vp->vd_flags & VRING_DESC_F_NEXT) == 0)
358 					break;
359 				next = vp->vd_next;
360 				if (next >= n_indir) {
361 					fprintf(stderr,
362 					    "%s: invalid next %u > %u, "
363 					    "driver confused?\r\n",
364 					    name, (u_int)next, n_indir);
365 					return (-1);
366 				}
367 			}
368 		}
369 		if ((vdir->vd_flags & VRING_DESC_F_NEXT) == 0)
370 			return (i);
371 	}
372 loopy:
373 	fprintf(stderr,
374 	    "%s: descriptor loop? count > %d - driver confused?\r\n",
375 	    name, i);
376 	return (-1);
377 }
378 
379 /*
380  * Return the currently-first request chain back to the available queue.
381  *
382  * (This chain is the one you handled when you called vq_getchain()
383  * and used its positive return value.)
384  */
385 void
386 vq_retchain(struct vqueue_info *vq)
387 {
388 
389 	vq->vq_last_avail--;
390 }
391 
392 /*
393  * Return specified request chain to the guest, setting its I/O length
394  * to the provided value.
395  *
396  * (This chain is the one you handled when you called vq_getchain()
397  * and used its positive return value.)
398  */
399 void
400 vq_relchain(struct vqueue_info *vq, uint16_t idx, uint32_t iolen)
401 {
402 	uint16_t uidx, mask;
403 	volatile struct vring_used *vuh;
404 	volatile struct virtio_used *vue;
405 
406 	/*
407 	 * Notes:
408 	 *  - mask is N-1 where N is a power of 2 so computes x % N
409 	 *  - vuh points to the "used" data shared with guest
410 	 *  - vue points to the "used" ring entry we want to update
411 	 *  - head is the same value we compute in vq_iovecs().
412 	 *
413 	 * (I apologize for the two fields named vu_idx; the
414 	 * virtio spec calls the one that vue points to, "id"...)
415 	 */
416 	mask = vq->vq_qsize - 1;
417 	vuh = vq->vq_used;
418 
419 	uidx = vuh->vu_idx;
420 	vue = &vuh->vu_ring[uidx++ & mask];
421 	vue->vu_idx = idx;
422 	vue->vu_tlen = iolen;
423 	vuh->vu_idx = uidx;
424 }
425 
426 /*
427  * Driver has finished processing "available" chains and calling
428  * vq_relchain on each one.  If driver used all the available
429  * chains, used_all should be set.
430  *
431  * If the "used" index moved we may need to inform the guest, i.e.,
432  * deliver an interrupt.  Even if the used index did NOT move we
433  * may need to deliver an interrupt, if the avail ring is empty and
434  * we are supposed to interrupt on empty.
435  *
436  * Note that used_all_avail is provided by the caller because it's
437  * a snapshot of the ring state when he decided to finish interrupt
438  * processing -- it's possible that descriptors became available after
439  * that point.  (It's also typically a constant 1/True as well.)
440  */
441 void
442 vq_endchains(struct vqueue_info *vq, int used_all_avail)
443 {
444 	struct virtio_softc *vs;
445 	uint16_t event_idx, new_idx, old_idx;
446 	int intr;
447 
448 	/*
449 	 * Interrupt generation: if we're using EVENT_IDX,
450 	 * interrupt if we've crossed the event threshold.
451 	 * Otherwise interrupt is generated if we added "used" entries,
452 	 * but suppressed by VRING_AVAIL_F_NO_INTERRUPT.
453 	 *
454 	 * In any case, though, if NOTIFY_ON_EMPTY is set and the
455 	 * entire avail was processed, we need to interrupt always.
456 	 */
457 	vs = vq->vq_vs;
458 	old_idx = vq->vq_save_used;
459 	vq->vq_save_used = new_idx = vq->vq_used->vu_idx;
460 	if (used_all_avail &&
461 	    (vs->vs_negotiated_caps & VIRTIO_F_NOTIFY_ON_EMPTY))
462 		intr = 1;
463 	else if (vs->vs_negotiated_caps & VIRTIO_RING_F_EVENT_IDX) {
464 		event_idx = VQ_USED_EVENT_IDX(vq);
465 		/*
466 		 * This calculation is per docs and the kernel
467 		 * (see src/sys/dev/virtio/virtio_ring.h).
468 		 */
469 		intr = (uint16_t)(new_idx - event_idx - 1) <
470 			(uint16_t)(new_idx - old_idx);
471 	} else {
472 		intr = new_idx != old_idx &&
473 		    !(vq->vq_avail->va_flags & VRING_AVAIL_F_NO_INTERRUPT);
474 	}
475 	if (intr)
476 		vq_interrupt(vs, vq);
477 }
478 
479 /* Note: these are in sorted order to make for a fast search */
480 static struct config_reg {
481 	uint16_t	cr_offset;	/* register offset */
482 	uint8_t		cr_size;	/* size (bytes) */
483 	uint8_t		cr_ro;		/* true => reg is read only */
484 	const char	*cr_name;	/* name of reg */
485 } config_regs[] = {
486 	{ VTCFG_R_HOSTCAP,	4, 1, "HOSTCAP" },
487 	{ VTCFG_R_GUESTCAP,	4, 0, "GUESTCAP" },
488 	{ VTCFG_R_PFN,		4, 0, "PFN" },
489 	{ VTCFG_R_QNUM,		2, 1, "QNUM" },
490 	{ VTCFG_R_QSEL,		2, 0, "QSEL" },
491 	{ VTCFG_R_QNOTIFY,	2, 0, "QNOTIFY" },
492 	{ VTCFG_R_STATUS,	1, 0, "STATUS" },
493 	{ VTCFG_R_ISR,		1, 0, "ISR" },
494 	{ VTCFG_R_CFGVEC,	2, 0, "CFGVEC" },
495 	{ VTCFG_R_QVEC,		2, 0, "QVEC" },
496 };
497 
498 static inline struct config_reg *
499 vi_find_cr(int offset) {
500 	u_int hi, lo, mid;
501 	struct config_reg *cr;
502 
503 	lo = 0;
504 	hi = sizeof(config_regs) / sizeof(*config_regs) - 1;
505 	while (hi >= lo) {
506 		mid = (hi + lo) >> 1;
507 		cr = &config_regs[mid];
508 		if (cr->cr_offset == offset)
509 			return (cr);
510 		if (cr->cr_offset < offset)
511 			lo = mid + 1;
512 		else
513 			hi = mid - 1;
514 	}
515 	return (NULL);
516 }
517 
518 /*
519  * Handle pci config space reads.
520  * If it's to the MSI-X info, do that.
521  * If it's part of the virtio standard stuff, do that.
522  * Otherwise dispatch to the actual driver.
523  */
524 uint64_t
525 vi_pci_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
526 	    int baridx, uint64_t offset, int size)
527 {
528 	struct virtio_softc *vs = pi->pi_arg;
529 	struct virtio_consts *vc;
530 	struct config_reg *cr;
531 	uint64_t virtio_config_size, max;
532 	const char *name;
533 	uint32_t newoff;
534 	uint32_t value;
535 	int error;
536 
537 	if (vs->vs_flags & VIRTIO_USE_MSIX) {
538 		if (baridx == pci_msix_table_bar(pi) ||
539 		    baridx == pci_msix_pba_bar(pi)) {
540 			return (pci_emul_msix_tread(pi, offset, size));
541 		}
542 	}
543 
544 	/* XXX probably should do something better than just assert() */
545 	assert(baridx == 0);
546 
547 	if (vs->vs_mtx)
548 		pthread_mutex_lock(vs->vs_mtx);
549 
550 	vc = vs->vs_vc;
551 	name = vc->vc_name;
552 	value = size == 1 ? 0xff : size == 2 ? 0xffff : 0xffffffff;
553 
554 	if (size != 1 && size != 2 && size != 4)
555 		goto bad;
556 
557 	if (pci_msix_enabled(pi))
558 		virtio_config_size = VTCFG_R_CFG1;
559 	else
560 		virtio_config_size = VTCFG_R_CFG0;
561 
562 	if (offset >= virtio_config_size) {
563 		/*
564 		 * Subtract off the standard size (including MSI-X
565 		 * registers if enabled) and dispatch to underlying driver.
566 		 * If that fails, fall into general code.
567 		 */
568 		newoff = offset - virtio_config_size;
569 		max = vc->vc_cfgsize ? vc->vc_cfgsize : 0x100000000;
570 		if (newoff + size > max)
571 			goto bad;
572 		error = (*vc->vc_cfgread)(DEV_SOFTC(vs), newoff, size, &value);
573 		if (!error)
574 			goto done;
575 	}
576 
577 bad:
578 	cr = vi_find_cr(offset);
579 	if (cr == NULL || cr->cr_size != size) {
580 		if (cr != NULL) {
581 			/* offset must be OK, so size must be bad */
582 			fprintf(stderr,
583 			    "%s: read from %s: bad size %d\r\n",
584 			    name, cr->cr_name, size);
585 		} else {
586 			fprintf(stderr,
587 			    "%s: read from bad offset/size %jd/%d\r\n",
588 			    name, (uintmax_t)offset, size);
589 		}
590 		goto done;
591 	}
592 
593 	switch (offset) {
594 	case VTCFG_R_HOSTCAP:
595 		value = vc->vc_hv_caps;
596 		break;
597 	case VTCFG_R_GUESTCAP:
598 		value = vs->vs_negotiated_caps;
599 		break;
600 	case VTCFG_R_PFN:
601 		if (vs->vs_curq < vc->vc_nvq)
602 			value = vs->vs_queues[vs->vs_curq].vq_pfn;
603 		break;
604 	case VTCFG_R_QNUM:
605 		value = vs->vs_curq < vc->vc_nvq ?
606 		    vs->vs_queues[vs->vs_curq].vq_qsize : 0;
607 		break;
608 	case VTCFG_R_QSEL:
609 		value = vs->vs_curq;
610 		break;
611 	case VTCFG_R_QNOTIFY:
612 		value = 0;	/* XXX */
613 		break;
614 	case VTCFG_R_STATUS:
615 		value = vs->vs_status;
616 		break;
617 	case VTCFG_R_ISR:
618 		value = vs->vs_isr;
619 		vs->vs_isr = 0;		/* a read clears this flag */
620 		if (value)
621 			pci_lintr_deassert(pi);
622 		break;
623 	case VTCFG_R_CFGVEC:
624 		value = vs->vs_msix_cfg_idx;
625 		break;
626 	case VTCFG_R_QVEC:
627 		value = vs->vs_curq < vc->vc_nvq ?
628 		    vs->vs_queues[vs->vs_curq].vq_msix_idx :
629 		    VIRTIO_MSI_NO_VECTOR;
630 		break;
631 	}
632 done:
633 	if (vs->vs_mtx)
634 		pthread_mutex_unlock(vs->vs_mtx);
635 	return (value);
636 }
637 
638 /*
639  * Handle pci config space writes.
640  * If it's to the MSI-X info, do that.
641  * If it's part of the virtio standard stuff, do that.
642  * Otherwise dispatch to the actual driver.
643  */
644 void
645 vi_pci_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
646 	     int baridx, uint64_t offset, int size, uint64_t value)
647 {
648 	struct virtio_softc *vs = pi->pi_arg;
649 	struct vqueue_info *vq;
650 	struct virtio_consts *vc;
651 	struct config_reg *cr;
652 	uint64_t virtio_config_size, max;
653 	const char *name;
654 	uint32_t newoff;
655 	int error;
656 
657 	if (vs->vs_flags & VIRTIO_USE_MSIX) {
658 		if (baridx == pci_msix_table_bar(pi) ||
659 		    baridx == pci_msix_pba_bar(pi)) {
660 			pci_emul_msix_twrite(pi, offset, size, value);
661 			return;
662 		}
663 	}
664 
665 	/* XXX probably should do something better than just assert() */
666 	assert(baridx == 0);
667 
668 	if (vs->vs_mtx)
669 		pthread_mutex_lock(vs->vs_mtx);
670 
671 	vc = vs->vs_vc;
672 	name = vc->vc_name;
673 
674 	if (size != 1 && size != 2 && size != 4)
675 		goto bad;
676 
677 	if (pci_msix_enabled(pi))
678 		virtio_config_size = VTCFG_R_CFG1;
679 	else
680 		virtio_config_size = VTCFG_R_CFG0;
681 
682 	if (offset >= virtio_config_size) {
683 		/*
684 		 * Subtract off the standard size (including MSI-X
685 		 * registers if enabled) and dispatch to underlying driver.
686 		 */
687 		newoff = offset - virtio_config_size;
688 		max = vc->vc_cfgsize ? vc->vc_cfgsize : 0x100000000;
689 		if (newoff + size > max)
690 			goto bad;
691 		error = (*vc->vc_cfgwrite)(DEV_SOFTC(vs), newoff, size, value);
692 		if (!error)
693 			goto done;
694 	}
695 
696 bad:
697 	cr = vi_find_cr(offset);
698 	if (cr == NULL || cr->cr_size != size || cr->cr_ro) {
699 		if (cr != NULL) {
700 			/* offset must be OK, wrong size and/or reg is R/O */
701 			if (cr->cr_size != size)
702 				fprintf(stderr,
703 				    "%s: write to %s: bad size %d\r\n",
704 				    name, cr->cr_name, size);
705 			if (cr->cr_ro)
706 				fprintf(stderr,
707 				    "%s: write to read-only reg %s\r\n",
708 				    name, cr->cr_name);
709 		} else {
710 			fprintf(stderr,
711 			    "%s: write to bad offset/size %jd/%d\r\n",
712 			    name, (uintmax_t)offset, size);
713 		}
714 		goto done;
715 	}
716 
717 	switch (offset) {
718 	case VTCFG_R_GUESTCAP:
719 		vs->vs_negotiated_caps = value & vc->vc_hv_caps;
720 		if (vc->vc_apply_features)
721 			(*vc->vc_apply_features)(DEV_SOFTC(vs),
722 			    vs->vs_negotiated_caps);
723 		break;
724 	case VTCFG_R_PFN:
725 		if (vs->vs_curq >= vc->vc_nvq)
726 			goto bad_qindex;
727 		vi_vq_init(vs, value);
728 		break;
729 	case VTCFG_R_QSEL:
730 		/*
731 		 * Note that the guest is allowed to select an
732 		 * invalid queue; we just need to return a QNUM
733 		 * of 0 while the bad queue is selected.
734 		 */
735 		vs->vs_curq = value;
736 		break;
737 	case VTCFG_R_QNOTIFY:
738 		if (value >= vc->vc_nvq) {
739 			fprintf(stderr, "%s: queue %d notify out of range\r\n",
740 				name, (int)value);
741 			goto done;
742 		}
743 		vq = &vs->vs_queues[value];
744 		if (vq->vq_notify)
745 			(*vq->vq_notify)(DEV_SOFTC(vs), vq);
746 		else if (vc->vc_qnotify)
747 			(*vc->vc_qnotify)(DEV_SOFTC(vs), vq);
748 		else
749 			fprintf(stderr,
750 			    "%s: qnotify queue %d: missing vq/vc notify\r\n",
751 				name, (int)value);
752 		break;
753 	case VTCFG_R_STATUS:
754 		vs->vs_status = value;
755 		if (value == 0)
756 			(*vc->vc_reset)(DEV_SOFTC(vs));
757 		break;
758 	case VTCFG_R_CFGVEC:
759 		vs->vs_msix_cfg_idx = value;
760 		break;
761 	case VTCFG_R_QVEC:
762 		if (vs->vs_curq >= vc->vc_nvq)
763 			goto bad_qindex;
764 		vq = &vs->vs_queues[vs->vs_curq];
765 		vq->vq_msix_idx = value;
766 		break;
767 	}
768 	goto done;
769 
770 bad_qindex:
771 	fprintf(stderr,
772 	    "%s: write config reg %s: curq %d >= max %d\r\n",
773 	    name, cr->cr_name, vs->vs_curq, vc->vc_nvq);
774 done:
775 	if (vs->vs_mtx)
776 		pthread_mutex_unlock(vs->vs_mtx);
777 }
778