virtio.c revision 276349
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: stable/10/usr.sbin/bhyve/virtio.c 276349 2014-12-28 21:27:13Z neel $");
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 */
58void
59vi_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 */
87void
88vi_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_pfn = 0;
101		vq->vq_msix_idx = VIRTIO_MSI_NO_VECTOR;
102	}
103	vs->vs_negotiated_caps = 0;
104	vs->vs_curq = 0;
105	/* vs->vs_status = 0; -- redundant */
106	if (vs->vs_isr)
107		pci_lintr_deassert(vs->vs_pi);
108	vs->vs_isr = 0;
109	vs->vs_msix_cfg_idx = VIRTIO_MSI_NO_VECTOR;
110}
111
112/*
113 * Set I/O BAR (usually 0) to map PCI config registers.
114 */
115void
116vi_set_io_bar(struct virtio_softc *vs, int barnum)
117{
118	size_t size;
119
120	/*
121	 * ??? should we use CFG0 if MSI-X is disabled?
122	 * Existing code did not...
123	 */
124	size = VTCFG_R_CFG1 + vs->vs_vc->vc_cfgsize;
125	pci_emul_alloc_bar(vs->vs_pi, barnum, PCIBAR_IO, size);
126}
127
128/*
129 * Initialize MSI-X vector capabilities if we're to use MSI-X,
130 * or MSI capabilities if not.
131 *
132 * We assume we want one MSI-X vector per queue, here, plus one
133 * for the config vec.
134 */
135int
136vi_intr_init(struct virtio_softc *vs, int barnum, int use_msix)
137{
138	int nvec;
139
140	if (use_msix) {
141		vs->vs_flags |= VIRTIO_USE_MSIX;
142		VS_LOCK(vs);
143		vi_reset_dev(vs); /* set all vectors to NO_VECTOR */
144		VS_UNLOCK(vs);
145		nvec = vs->vs_vc->vc_nvq + 1;
146		if (pci_emul_add_msixcap(vs->vs_pi, nvec, barnum))
147			return (1);
148	} else
149		vs->vs_flags &= ~VIRTIO_USE_MSIX;
150	/* Only 1 MSI vector for bhyve */
151	pci_emul_add_msicap(vs->vs_pi, 1);
152	return (0);
153}
154
155/*
156 * Initialize the currently-selected virtio queue (vs->vs_curq).
157 * The guest just gave us a page frame number, from which we can
158 * calculate the addresses of the queue.
159 */
160void
161vi_vq_init(struct virtio_softc *vs, uint32_t pfn)
162{
163	struct vqueue_info *vq;
164	uint64_t phys;
165	size_t size;
166	char *base;
167
168	vq = &vs->vs_queues[vs->vs_curq];
169	vq->vq_pfn = pfn;
170	phys = (uint64_t)pfn << VRING_PFN;
171	size = vring_size(vq->vq_qsize);
172	base = paddr_guest2host(vs->vs_pi->pi_vmctx, phys, size);
173
174	/* First page(s) are descriptors... */
175	vq->vq_desc = (struct virtio_desc *)base;
176	base += vq->vq_qsize * sizeof(struct virtio_desc);
177
178	/* ... immediately followed by "avail" ring (entirely uint16_t's) */
179	vq->vq_avail = (struct vring_avail *)base;
180	base += (2 + vq->vq_qsize + 1) * sizeof(uint16_t);
181
182	/* Then it's rounded up to the next page... */
183	base = (char *)roundup2((uintptr_t)base, VRING_ALIGN);
184
185	/* ... and the last page(s) are the used ring. */
186	vq->vq_used = (struct vring_used *)base;
187
188	/* Mark queue as allocated, and start at 0 when we use it. */
189	vq->vq_flags = VQ_ALLOC;
190	vq->vq_last_avail = 0;
191}
192
193/*
194 * Helper inline for vq_getchain(): record the i'th "real"
195 * descriptor.
196 */
197static inline void
198_vq_record(int i, volatile struct virtio_desc *vd, struct vmctx *ctx,
199	   struct iovec *iov, int n_iov, uint16_t *flags) {
200
201	if (i >= n_iov)
202		return;
203	iov[i].iov_base = paddr_guest2host(ctx, vd->vd_addr, vd->vd_len);
204	iov[i].iov_len = vd->vd_len;
205	if (flags != NULL)
206		flags[i] = vd->vd_flags;
207}
208#define	VQ_MAX_DESCRIPTORS	512	/* see below */
209
210/*
211 * Examine the chain of descriptors starting at the "next one" to
212 * make sure that they describe a sensible request.  If so, return
213 * the number of "real" descriptors that would be needed/used in
214 * acting on this request.  This may be smaller than the number of
215 * available descriptors, e.g., if there are two available but
216 * they are two separate requests, this just returns 1.  Or, it
217 * may be larger: if there are indirect descriptors involved,
218 * there may only be one descriptor available but it may be an
219 * indirect pointing to eight more.  We return 8 in this case,
220 * i.e., we do not count the indirect descriptors, only the "real"
221 * ones.
222 *
223 * Basically, this vets the vd_flags and vd_next field of each
224 * descriptor and tells you how many are involved.  Since some may
225 * be indirect, this also needs the vmctx (in the pci_devinst
226 * at vs->vs_pi) so that it can find indirect descriptors.
227 *
228 * As we process each descriptor, we copy and adjust it (guest to
229 * host address wise, also using the vmtctx) into the given iov[]
230 * array (of the given size).  If the array overflows, we stop
231 * placing values into the array but keep processing descriptors,
232 * up to VQ_MAX_DESCRIPTORS, before giving up and returning -1.
233 * So you, the caller, must not assume that iov[] is as big as the
234 * return value (you can process the same thing twice to allocate
235 * a larger iov array if needed, or supply a zero length to find
236 * out how much space is needed).
237 *
238 * If you want to verify the WRITE flag on each descriptor, pass a
239 * non-NULL "flags" pointer to an array of "uint16_t" of the same size
240 * as n_iov and we'll copy each vd_flags field after unwinding any
241 * indirects.
242 *
243 * If some descriptor(s) are invalid, this prints a diagnostic message
244 * and returns -1.  If no descriptors are ready now it simply returns 0.
245 *
246 * You are assumed to have done a vq_ring_ready() if needed (note
247 * that vq_has_descs() does one).
248 */
249int
250vq_getchain(struct vqueue_info *vq,
251	    struct iovec *iov, int n_iov, uint16_t *flags)
252{
253	int i;
254	u_int ndesc, n_indir;
255	u_int idx, head, next;
256	volatile struct virtio_desc *vdir, *vindir, *vp;
257	struct vmctx *ctx;
258	struct virtio_softc *vs;
259	const char *name;
260
261	vs = vq->vq_vs;
262	name = vs->vs_vc->vc_name;
263
264	/*
265	 * Note: it's the responsibility of the guest not to
266	 * update vq->vq_avail->va_idx until all of the descriptors
267         * the guest has written are valid (including all their
268         * vd_next fields and vd_flags).
269	 *
270	 * Compute (last_avail - va_idx) in integers mod 2**16.  This is
271	 * the number of descriptors the device has made available
272	 * since the last time we updated vq->vq_last_avail.
273	 *
274	 * We just need to do the subtraction as an unsigned int,
275	 * then trim off excess bits.
276	 */
277	idx = vq->vq_last_avail;
278	ndesc = (uint16_t)((u_int)vq->vq_avail->va_idx - idx);
279	if (ndesc == 0)
280		return (0);
281	if (ndesc > vq->vq_qsize) {
282		/* XXX need better way to diagnose issues */
283		fprintf(stderr,
284		    "%s: ndesc (%u) out of range, driver confused?\r\n",
285		    name, (u_int)ndesc);
286		return (-1);
287	}
288
289	/*
290	 * Now count/parse "involved" descriptors starting from
291	 * the head of the chain.
292	 *
293	 * To prevent loops, we could be more complicated and
294	 * check whether we're re-visiting a previously visited
295	 * index, but we just abort if the count gets excessive.
296	 */
297	ctx = vs->vs_pi->pi_vmctx;
298	head = vq->vq_avail->va_ring[idx & (vq->vq_qsize - 1)];
299	next = head;
300	for (i = 0; i < VQ_MAX_DESCRIPTORS; next = vdir->vd_next) {
301		if (next >= vq->vq_qsize) {
302			fprintf(stderr,
303			    "%s: descriptor index %u out of range, "
304			    "driver confused?\r\n",
305			    name, next);
306			return (-1);
307		}
308		vdir = &vq->vq_desc[next];
309		if ((vdir->vd_flags & VRING_DESC_F_INDIRECT) == 0) {
310			_vq_record(i, vdir, ctx, iov, n_iov, flags);
311			i++;
312		} else if ((vs->vs_negotiated_caps &
313		    VIRTIO_RING_F_INDIRECT_DESC) == 0) {
314			fprintf(stderr,
315			    "%s: descriptor has forbidden INDIRECT flag, "
316			    "driver confused?\r\n",
317			    name);
318			return (-1);
319		} else {
320			n_indir = vdir->vd_len / 16;
321			if ((vdir->vd_len & 0xf) || n_indir == 0) {
322				fprintf(stderr,
323				    "%s: invalid indir len 0x%x, "
324				    "driver confused?\r\n",
325				    name, (u_int)vdir->vd_len);
326				return (-1);
327			}
328			vindir = paddr_guest2host(ctx,
329			    vdir->vd_addr, vdir->vd_len);
330			/*
331			 * Indirects start at the 0th, then follow
332			 * their own embedded "next"s until those run
333			 * out.  Each one's indirect flag must be off
334			 * (we don't really have to check, could just
335			 * ignore errors...).
336			 */
337			next = 0;
338			for (;;) {
339				vp = &vindir[next];
340				if (vp->vd_flags & VRING_DESC_F_INDIRECT) {
341					fprintf(stderr,
342					    "%s: indirect desc has INDIR flag,"
343					    " driver confused?\r\n",
344					    name);
345					return (-1);
346				}
347				_vq_record(i, vp, ctx, iov, n_iov, flags);
348				if (++i > VQ_MAX_DESCRIPTORS)
349					goto loopy;
350				if ((vp->vd_flags & VRING_DESC_F_NEXT) == 0)
351					break;
352				next = vp->vd_next;
353				if (next >= n_indir) {
354					fprintf(stderr,
355					    "%s: invalid next %u > %u, "
356					    "driver confused?\r\n",
357					    name, (u_int)next, n_indir);
358					return (-1);
359				}
360			}
361		}
362		if ((vdir->vd_flags & VRING_DESC_F_NEXT) == 0)
363			return (i);
364	}
365loopy:
366	fprintf(stderr,
367	    "%s: descriptor loop? count > %d - driver confused?\r\n",
368	    name, i);
369	return (-1);
370}
371
372/*
373 * Return the currently-first request chain to the guest, setting
374 * its I/O length to the provided value.
375 *
376 * (This chain is the one you handled when you called vq_getchain()
377 * and used its positive return value.)
378 */
379void
380vq_relchain(struct vqueue_info *vq, uint32_t iolen)
381{
382	uint16_t head, uidx, mask;
383	volatile struct vring_used *vuh;
384	volatile struct virtio_used *vue;
385
386	/*
387	 * Notes:
388	 *  - mask is N-1 where N is a power of 2 so computes x % N
389	 *  - vuh points to the "used" data shared with guest
390	 *  - vue points to the "used" ring entry we want to update
391	 *  - head is the same value we compute in vq_iovecs().
392	 *
393	 * (I apologize for the two fields named vu_idx; the
394	 * virtio spec calls the one that vue points to, "id"...)
395	 */
396	mask = vq->vq_qsize - 1;
397	vuh = vq->vq_used;
398	head = vq->vq_avail->va_ring[vq->vq_last_avail++ & mask];
399
400	uidx = vuh->vu_idx;
401	vue = &vuh->vu_ring[uidx++ & mask];
402	vue->vu_idx = head; /* ie, vue->id = head */
403	vue->vu_tlen = iolen;
404	vuh->vu_idx = uidx;
405}
406
407/*
408 * Driver has finished processing "available" chains and calling
409 * vq_relchain on each one.  If driver used all the available
410 * chains, used_all should be set.
411 *
412 * If the "used" index moved we may need to inform the guest, i.e.,
413 * deliver an interrupt.  Even if the used index did NOT move we
414 * may need to deliver an interrupt, if the avail ring is empty and
415 * we are supposed to interrupt on empty.
416 *
417 * Note that used_all_avail is provided by the caller because it's
418 * a snapshot of the ring state when he decided to finish interrupt
419 * processing -- it's possible that descriptors became available after
420 * that point.  (It's also typically a constant 1/True as well.)
421 */
422void
423vq_endchains(struct vqueue_info *vq, int used_all_avail)
424{
425	struct virtio_softc *vs;
426	uint16_t event_idx, new_idx, old_idx;
427	int intr;
428
429	/*
430	 * Interrupt generation: if we're using EVENT_IDX,
431	 * interrupt if we've crossed the event threshold.
432	 * Otherwise interrupt is generated if we added "used" entries,
433	 * but suppressed by VRING_AVAIL_F_NO_INTERRUPT.
434	 *
435	 * In any case, though, if NOTIFY_ON_EMPTY is set and the
436	 * entire avail was processed, we need to interrupt always.
437	 */
438	vs = vq->vq_vs;
439	new_idx = vq->vq_used->vu_idx;
440	old_idx = vq->vq_save_used;
441	if (used_all_avail &&
442	    (vs->vs_negotiated_caps & VIRTIO_F_NOTIFY_ON_EMPTY))
443		intr = 1;
444	else if (vs->vs_negotiated_caps & VIRTIO_RING_F_EVENT_IDX) {
445		event_idx = VQ_USED_EVENT_IDX(vq);
446		/*
447		 * This calculation is per docs and the kernel
448		 * (see src/sys/dev/virtio/virtio_ring.h).
449		 */
450		intr = (uint16_t)(new_idx - event_idx - 1) <
451			(uint16_t)(new_idx - old_idx);
452	} else {
453		intr = new_idx != old_idx &&
454		    !(vq->vq_avail->va_flags & VRING_AVAIL_F_NO_INTERRUPT);
455	}
456	if (intr)
457		vq_interrupt(vs, vq);
458}
459
460/* Note: these are in sorted order to make for a fast search */
461static struct config_reg {
462	uint16_t	cr_offset;	/* register offset */
463	uint8_t		cr_size;	/* size (bytes) */
464	uint8_t		cr_ro;		/* true => reg is read only */
465	const char	*cr_name;	/* name of reg */
466} config_regs[] = {
467	{ VTCFG_R_HOSTCAP,	4, 1, "HOSTCAP" },
468	{ VTCFG_R_GUESTCAP,	4, 0, "GUESTCAP" },
469	{ VTCFG_R_PFN,		4, 0, "PFN" },
470	{ VTCFG_R_QNUM,		2, 1, "QNUM" },
471	{ VTCFG_R_QSEL,		2, 0, "QSEL" },
472	{ VTCFG_R_QNOTIFY,	2, 0, "QNOTIFY" },
473	{ VTCFG_R_STATUS,	1, 0, "STATUS" },
474	{ VTCFG_R_ISR,		1, 0, "ISR" },
475	{ VTCFG_R_CFGVEC,	2, 0, "CFGVEC" },
476	{ VTCFG_R_QVEC,		2, 0, "QVEC" },
477};
478
479static inline struct config_reg *
480vi_find_cr(int offset) {
481	u_int hi, lo, mid;
482	struct config_reg *cr;
483
484	lo = 0;
485	hi = sizeof(config_regs) / sizeof(*config_regs) - 1;
486	while (hi >= lo) {
487		mid = (hi + lo) >> 1;
488		cr = &config_regs[mid];
489		if (cr->cr_offset == offset)
490			return (cr);
491		if (cr->cr_offset < offset)
492			lo = mid + 1;
493		else
494			hi = mid - 1;
495	}
496	return (NULL);
497}
498
499/*
500 * Handle pci config space reads.
501 * If it's to the MSI-X info, do that.
502 * If it's part of the virtio standard stuff, do that.
503 * Otherwise dispatch to the actual driver.
504 */
505uint64_t
506vi_pci_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
507	    int baridx, uint64_t offset, int size)
508{
509	struct virtio_softc *vs = pi->pi_arg;
510	struct virtio_consts *vc;
511	struct config_reg *cr;
512	uint64_t virtio_config_size, max;
513	const char *name;
514	uint32_t newoff;
515	uint32_t value;
516	int error;
517
518	if (vs->vs_flags & VIRTIO_USE_MSIX) {
519		if (baridx == pci_msix_table_bar(pi) ||
520		    baridx == pci_msix_pba_bar(pi)) {
521			return (pci_emul_msix_tread(pi, offset, size));
522		}
523	}
524
525	/* XXX probably should do something better than just assert() */
526	assert(baridx == 0);
527
528	if (vs->vs_mtx)
529		pthread_mutex_lock(vs->vs_mtx);
530
531	vc = vs->vs_vc;
532	name = vc->vc_name;
533	value = size == 1 ? 0xff : size == 2 ? 0xffff : 0xffffffff;
534
535	if (size != 1 && size != 2 && size != 4)
536		goto bad;
537
538	if (pci_msix_enabled(pi))
539		virtio_config_size = VTCFG_R_CFG1;
540	else
541		virtio_config_size = VTCFG_R_CFG0;
542
543	if (offset >= virtio_config_size) {
544		/*
545		 * Subtract off the standard size (including MSI-X
546		 * registers if enabled) and dispatch to underlying driver.
547		 * If that fails, fall into general code.
548		 */
549		newoff = offset - virtio_config_size;
550		max = vc->vc_cfgsize ? vc->vc_cfgsize : 0x100000000;
551		if (newoff + size > max)
552			goto bad;
553		error = (*vc->vc_cfgread)(DEV_SOFTC(vs), newoff, size, &value);
554		if (!error)
555			goto done;
556	}
557
558bad:
559	cr = vi_find_cr(offset);
560	if (cr == NULL || cr->cr_size != size) {
561		if (cr != NULL) {
562			/* offset must be OK, so size must be bad */
563			fprintf(stderr,
564			    "%s: read from %s: bad size %d\r\n",
565			    name, cr->cr_name, size);
566		} else {
567			fprintf(stderr,
568			    "%s: read from bad offset/size %jd/%d\r\n",
569			    name, (uintmax_t)offset, size);
570		}
571		goto done;
572	}
573
574	switch (offset) {
575	case VTCFG_R_HOSTCAP:
576		value = vc->vc_hv_caps;
577		break;
578	case VTCFG_R_GUESTCAP:
579		value = vs->vs_negotiated_caps;
580		break;
581	case VTCFG_R_PFN:
582		if (vs->vs_curq < vc->vc_nvq)
583			value = vs->vs_queues[vs->vs_curq].vq_pfn;
584		break;
585	case VTCFG_R_QNUM:
586		value = vs->vs_curq < vc->vc_nvq ?
587		    vs->vs_queues[vs->vs_curq].vq_qsize : 0;
588		break;
589	case VTCFG_R_QSEL:
590		value = vs->vs_curq;
591		break;
592	case VTCFG_R_QNOTIFY:
593		value = 0;	/* XXX */
594		break;
595	case VTCFG_R_STATUS:
596		value = vs->vs_status;
597		break;
598	case VTCFG_R_ISR:
599		value = vs->vs_isr;
600		vs->vs_isr = 0;		/* a read clears this flag */
601		if (value)
602			pci_lintr_deassert(pi);
603		break;
604	case VTCFG_R_CFGVEC:
605		value = vs->vs_msix_cfg_idx;
606		break;
607	case VTCFG_R_QVEC:
608		value = vs->vs_curq < vc->vc_nvq ?
609		    vs->vs_queues[vs->vs_curq].vq_msix_idx :
610		    VIRTIO_MSI_NO_VECTOR;
611		break;
612	}
613done:
614	if (vs->vs_mtx)
615		pthread_mutex_unlock(vs->vs_mtx);
616	return (value);
617}
618
619/*
620 * Handle pci config space writes.
621 * If it's to the MSI-X info, do that.
622 * If it's part of the virtio standard stuff, do that.
623 * Otherwise dispatch to the actual driver.
624 */
625void
626vi_pci_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
627	     int baridx, uint64_t offset, int size, uint64_t value)
628{
629	struct virtio_softc *vs = pi->pi_arg;
630	struct vqueue_info *vq;
631	struct virtio_consts *vc;
632	struct config_reg *cr;
633	uint64_t virtio_config_size, max;
634	const char *name;
635	uint32_t newoff;
636	int error;
637
638	if (vs->vs_flags & VIRTIO_USE_MSIX) {
639		if (baridx == pci_msix_table_bar(pi) ||
640		    baridx == pci_msix_pba_bar(pi)) {
641			pci_emul_msix_twrite(pi, offset, size, value);
642			return;
643		}
644	}
645
646	/* XXX probably should do something better than just assert() */
647	assert(baridx == 0);
648
649	if (vs->vs_mtx)
650		pthread_mutex_lock(vs->vs_mtx);
651
652	vc = vs->vs_vc;
653	name = vc->vc_name;
654
655	if (size != 1 && size != 2 && size != 4)
656		goto bad;
657
658	if (pci_msix_enabled(pi))
659		virtio_config_size = VTCFG_R_CFG1;
660	else
661		virtio_config_size = VTCFG_R_CFG0;
662
663	if (offset >= virtio_config_size) {
664		/*
665		 * Subtract off the standard size (including MSI-X
666		 * registers if enabled) and dispatch to underlying driver.
667		 */
668		newoff = offset - virtio_config_size;
669		max = vc->vc_cfgsize ? vc->vc_cfgsize : 0x100000000;
670		if (newoff + size > max)
671			goto bad;
672		error = (*vc->vc_cfgwrite)(DEV_SOFTC(vs), newoff, size, value);
673		if (!error)
674			goto done;
675	}
676
677bad:
678	cr = vi_find_cr(offset);
679	if (cr == NULL || cr->cr_size != size || cr->cr_ro) {
680		if (cr != NULL) {
681			/* offset must be OK, wrong size and/or reg is R/O */
682			if (cr->cr_size != size)
683				fprintf(stderr,
684				    "%s: write to %s: bad size %d\r\n",
685				    name, cr->cr_name, size);
686			if (cr->cr_ro)
687				fprintf(stderr,
688				    "%s: write to read-only reg %s\r\n",
689				    name, cr->cr_name);
690		} else {
691			fprintf(stderr,
692			    "%s: write to bad offset/size %jd/%d\r\n",
693			    name, (uintmax_t)offset, size);
694		}
695		goto done;
696	}
697
698	switch (offset) {
699	case VTCFG_R_GUESTCAP:
700		vs->vs_negotiated_caps = value & vc->vc_hv_caps;
701		if (vc->vc_apply_features)
702			(*vc->vc_apply_features)(DEV_SOFTC(vs),
703			    vs->vs_negotiated_caps);
704		break;
705	case VTCFG_R_PFN:
706		if (vs->vs_curq >= vc->vc_nvq)
707			goto bad_qindex;
708		vi_vq_init(vs, value);
709		break;
710	case VTCFG_R_QSEL:
711		/*
712		 * Note that the guest is allowed to select an
713		 * invalid queue; we just need to return a QNUM
714		 * of 0 while the bad queue is selected.
715		 */
716		vs->vs_curq = value;
717		break;
718	case VTCFG_R_QNOTIFY:
719		if (value >= vc->vc_nvq) {
720			fprintf(stderr, "%s: queue %d notify out of range\r\n",
721				name, (int)value);
722			goto done;
723		}
724		vq = &vs->vs_queues[value];
725		if (vq->vq_notify)
726			(*vq->vq_notify)(DEV_SOFTC(vs), vq);
727		else if (vc->vc_qnotify)
728			(*vc->vc_qnotify)(DEV_SOFTC(vs), vq);
729		else
730			fprintf(stderr,
731			    "%s: qnotify queue %d: missing vq/vc notify\r\n",
732				name, (int)value);
733		break;
734	case VTCFG_R_STATUS:
735		vs->vs_status = value;
736		if (value == 0)
737			(*vc->vc_reset)(DEV_SOFTC(vs));
738		break;
739	case VTCFG_R_CFGVEC:
740		vs->vs_msix_cfg_idx = value;
741		break;
742	case VTCFG_R_QVEC:
743		if (vs->vs_curq >= vc->vc_nvq)
744			goto bad_qindex;
745		vq = &vs->vs_queues[vs->vs_curq];
746		vq->vq_msix_idx = value;
747		break;
748	}
749	goto done;
750
751bad_qindex:
752	fprintf(stderr,
753	    "%s: write config reg %s: curq %d >= max %d\r\n",
754	    name, cr->cr_name, vs->vs_curq, vc->vc_nvq);
755done:
756	if (vs->vs_mtx)
757		pthread_mutex_unlock(vs->vs_mtx);
758}
759