virtio.c revision 271685
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 271685 2014-09-16 19:08:54Z grehan $");
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
37#include "bhyverun.h"
38#include "pci_emul.h"
39#include "virtio.h"
40
41/*
42 * Functions for dealing with generalized "virtual devices" as
43 * defined by <https://www.google.com/#output=search&q=virtio+spec>
44 */
45
46/*
47 * In case we decide to relax the "virtio softc comes at the
48 * front of virtio-based device softc" constraint, let's use
49 * this to convert.
50 */
51#define DEV_SOFTC(vs) ((void *)(vs))
52
53/*
54 * Link a virtio_softc to its constants, the device softc, and
55 * the PCI emulation.
56 */
57void
58vi_softc_linkup(struct virtio_softc *vs, struct virtio_consts *vc,
59		void *dev_softc, struct pci_devinst *pi,
60		struct vqueue_info *queues)
61{
62	int i;
63
64	/* vs and dev_softc addresses must match */
65	assert((void *)vs == dev_softc);
66	vs->vs_vc = vc;
67	vs->vs_pi = pi;
68	pi->pi_arg = vs;
69
70	vs->vs_queues = queues;
71	for (i = 0; i < vc->vc_nvq; i++) {
72		queues[i].vq_vs = vs;
73		queues[i].vq_num = i;
74	}
75}
76
77/*
78 * Reset device (device-wide).  This erases all queues, i.e.,
79 * all the queues become invalid (though we don't wipe out the
80 * internal pointers, we just clear the VQ_ALLOC flag).
81 *
82 * It resets negotiated features to "none".
83 *
84 * If MSI-X is enabled, this also resets all the vectors to NO_VECTOR.
85 */
86void
87vi_reset_dev(struct virtio_softc *vs)
88{
89	struct vqueue_info *vq;
90	int i, nvq;
91
92	nvq = vs->vs_vc->vc_nvq;
93	for (vq = vs->vs_queues, i = 0; i < nvq; vq++, i++) {
94		vq->vq_flags = 0;
95		vq->vq_last_avail = 0;
96		vq->vq_pfn = 0;
97		vq->vq_msix_idx = VIRTIO_MSI_NO_VECTOR;
98	}
99	vs->vs_negotiated_caps = 0;
100	vs->vs_curq = 0;
101	/* vs->vs_status = 0; -- redundant */
102	VS_LOCK(vs);
103	if (vs->vs_isr)
104		pci_lintr_deassert(vs->vs_pi);
105	vs->vs_isr = 0;
106	VS_UNLOCK(vs);
107	vs->vs_msix_cfg_idx = VIRTIO_MSI_NO_VECTOR;
108}
109
110/*
111 * Set I/O BAR (usually 0) to map PCI config registers.
112 */
113void
114vi_set_io_bar(struct virtio_softc *vs, int barnum)
115{
116	size_t size;
117
118	/*
119	 * ??? should we use CFG0 if MSI-X is disabled?
120	 * Existing code did not...
121	 */
122	size = VTCFG_R_CFG1 + vs->vs_vc->vc_cfgsize;
123	pci_emul_alloc_bar(vs->vs_pi, barnum, PCIBAR_IO, size);
124}
125
126/*
127 * Initialize MSI-X vector capabilities if we're to use MSI-X,
128 * or MSI capabilities if not.
129 *
130 * We assume we want one MSI-X vector per queue, here, plus one
131 * for the config vec.
132 */
133int
134vi_intr_init(struct virtio_softc *vs, int barnum, int use_msix)
135{
136	int nvec;
137
138	if (use_msix) {
139		vs->vs_flags |= VIRTIO_USE_MSIX;
140		vi_reset_dev(vs); /* set all vectors to NO_VECTOR */
141		nvec = vs->vs_vc->vc_nvq + 1;
142		if (pci_emul_add_msixcap(vs->vs_pi, nvec, barnum))
143			return (1);
144	} else
145		vs->vs_flags &= ~VIRTIO_USE_MSIX;
146	/* Only 1 MSI vector for bhyve */
147	pci_emul_add_msicap(vs->vs_pi, 1);
148	return (0);
149}
150
151/*
152 * Initialize the currently-selected virtio queue (vs->vs_curq).
153 * The guest just gave us a page frame number, from which we can
154 * calculate the addresses of the queue.
155 */
156void
157vi_vq_init(struct virtio_softc *vs, uint32_t pfn)
158{
159	struct vqueue_info *vq;
160	uint64_t phys;
161	size_t size;
162	char *base;
163
164	vq = &vs->vs_queues[vs->vs_curq];
165	vq->vq_pfn = pfn;
166	phys = (uint64_t)pfn << VRING_PFN;
167	size = vring_size(vq->vq_qsize);
168	base = paddr_guest2host(vs->vs_pi->pi_vmctx, phys, size);
169
170	/* First page(s) are descriptors... */
171	vq->vq_desc = (struct virtio_desc *)base;
172	base += vq->vq_qsize * sizeof(struct virtio_desc);
173
174	/* ... immediately followed by "avail" ring (entirely uint16_t's) */
175	vq->vq_avail = (struct vring_avail *)base;
176	base += (2 + vq->vq_qsize + 1) * sizeof(uint16_t);
177
178	/* Then it's rounded up to the next page... */
179	base = (char *)roundup2((uintptr_t)base, VRING_ALIGN);
180
181	/* ... and the last page(s) are the used ring. */
182	vq->vq_used = (struct vring_used *)base;
183
184	/* Mark queue as allocated, and start at 0 when we use it. */
185	vq->vq_flags = VQ_ALLOC;
186	vq->vq_last_avail = 0;
187}
188
189/*
190 * Helper inline for vq_getchain(): record the i'th "real"
191 * descriptor.
192 */
193static inline void
194_vq_record(int i, volatile struct virtio_desc *vd, struct vmctx *ctx,
195	   struct iovec *iov, int n_iov, uint16_t *flags) {
196
197	if (i >= n_iov)
198		return;
199	iov[i].iov_base = paddr_guest2host(ctx, vd->vd_addr, vd->vd_len);
200	iov[i].iov_len = vd->vd_len;
201	if (flags != NULL)
202		flags[i] = vd->vd_flags;
203}
204#define	VQ_MAX_DESCRIPTORS	512	/* see below */
205
206/*
207 * Examine the chain of descriptors starting at the "next one" to
208 * make sure that they describe a sensible request.  If so, return
209 * the number of "real" descriptors that would be needed/used in
210 * acting on this request.  This may be smaller than the number of
211 * available descriptors, e.g., if there are two available but
212 * they are two separate requests, this just returns 1.  Or, it
213 * may be larger: if there are indirect descriptors involved,
214 * there may only be one descriptor available but it may be an
215 * indirect pointing to eight more.  We return 8 in this case,
216 * i.e., we do not count the indirect descriptors, only the "real"
217 * ones.
218 *
219 * Basically, this vets the vd_flags and vd_next field of each
220 * descriptor and tells you how many are involved.  Since some may
221 * be indirect, this also needs the vmctx (in the pci_devinst
222 * at vs->vs_pi) so that it can find indirect descriptors.
223 *
224 * As we process each descriptor, we copy and adjust it (guest to
225 * host address wise, also using the vmtctx) into the given iov[]
226 * array (of the given size).  If the array overflows, we stop
227 * placing values into the array but keep processing descriptors,
228 * up to VQ_MAX_DESCRIPTORS, before giving up and returning -1.
229 * So you, the caller, must not assume that iov[] is as big as the
230 * return value (you can process the same thing twice to allocate
231 * a larger iov array if needed, or supply a zero length to find
232 * out how much space is needed).
233 *
234 * If you want to verify the WRITE flag on each descriptor, pass a
235 * non-NULL "flags" pointer to an array of "uint16_t" of the same size
236 * as n_iov and we'll copy each vd_flags field after unwinding any
237 * indirects.
238 *
239 * If some descriptor(s) are invalid, this prints a diagnostic message
240 * and returns -1.  If no descriptors are ready now it simply returns 0.
241 *
242 * You are assumed to have done a vq_ring_ready() if needed (note
243 * that vq_has_descs() does one).
244 */
245int
246vq_getchain(struct vqueue_info *vq,
247	    struct iovec *iov, int n_iov, uint16_t *flags)
248{
249	int i;
250	u_int ndesc, n_indir;
251	u_int idx, head, next;
252	volatile struct virtio_desc *vdir, *vindir, *vp;
253	struct vmctx *ctx;
254	struct virtio_softc *vs;
255	const char *name;
256
257	vs = vq->vq_vs;
258	name = vs->vs_vc->vc_name;
259
260	/*
261	 * Note: it's the responsibility of the guest not to
262	 * update vq->vq_avail->va_idx until all of the descriptors
263         * the guest has written are valid (including all their
264         * vd_next fields and vd_flags).
265	 *
266	 * Compute (last_avail - va_idx) in integers mod 2**16.  This is
267	 * the number of descriptors the device has made available
268	 * since the last time we updated vq->vq_last_avail.
269	 *
270	 * We just need to do the subtraction as an unsigned int,
271	 * then trim off excess bits.
272	 */
273	idx = vq->vq_last_avail;
274	ndesc = (uint16_t)((u_int)vq->vq_avail->va_idx - idx);
275	if (ndesc == 0)
276		return (0);
277	if (ndesc > vq->vq_qsize) {
278		/* XXX need better way to diagnose issues */
279		fprintf(stderr,
280		    "%s: ndesc (%u) out of range, driver confused?\r\n",
281		    name, (u_int)ndesc);
282		return (-1);
283	}
284
285	/*
286	 * Now count/parse "involved" descriptors starting from
287	 * the head of the chain.
288	 *
289	 * To prevent loops, we could be more complicated and
290	 * check whether we're re-visiting a previously visited
291	 * index, but we just abort if the count gets excessive.
292	 */
293	ctx = vs->vs_pi->pi_vmctx;
294	head = vq->vq_avail->va_ring[idx & (vq->vq_qsize - 1)];
295	next = head;
296	for (i = 0; i < VQ_MAX_DESCRIPTORS; next = vdir->vd_next) {
297		if (next >= vq->vq_qsize) {
298			fprintf(stderr,
299			    "%s: descriptor index %u out of range, "
300			    "driver confused?\r\n",
301			    name, next);
302			return (-1);
303		}
304		vdir = &vq->vq_desc[next];
305		if ((vdir->vd_flags & VRING_DESC_F_INDIRECT) == 0) {
306			_vq_record(i, vdir, ctx, iov, n_iov, flags);
307			i++;
308		} else if ((vs->vs_negotiated_caps &
309		    VIRTIO_RING_F_INDIRECT_DESC) == 0) {
310			fprintf(stderr,
311			    "%s: descriptor has forbidden INDIRECT flag, "
312			    "driver confused?\r\n",
313			    name);
314			return (-1);
315		} else {
316			n_indir = vdir->vd_len / 16;
317			if ((vdir->vd_len & 0xf) || n_indir == 0) {
318				fprintf(stderr,
319				    "%s: invalid indir len 0x%x, "
320				    "driver confused?\r\n",
321				    name, (u_int)vdir->vd_len);
322				return (-1);
323			}
324			vindir = paddr_guest2host(ctx,
325			    vdir->vd_addr, vdir->vd_len);
326			/*
327			 * Indirects start at the 0th, then follow
328			 * their own embedded "next"s until those run
329			 * out.  Each one's indirect flag must be off
330			 * (we don't really have to check, could just
331			 * ignore errors...).
332			 */
333			next = 0;
334			for (;;) {
335				vp = &vindir[next];
336				if (vp->vd_flags & VRING_DESC_F_INDIRECT) {
337					fprintf(stderr,
338					    "%s: indirect desc has INDIR flag,"
339					    " driver confused?\r\n",
340					    name);
341					return (-1);
342				}
343				_vq_record(i, vp, ctx, iov, n_iov, flags);
344				if (++i > VQ_MAX_DESCRIPTORS)
345					goto loopy;
346				if ((vp->vd_flags & VRING_DESC_F_NEXT) == 0)
347					break;
348				next = vp->vd_next;
349				if (next >= n_indir) {
350					fprintf(stderr,
351					    "%s: invalid next %u > %u, "
352					    "driver confused?\r\n",
353					    name, (u_int)next, n_indir);
354					return (-1);
355				}
356			}
357		}
358		if ((vdir->vd_flags & VRING_DESC_F_NEXT) == 0)
359			return (i);
360	}
361loopy:
362	fprintf(stderr,
363	    "%s: descriptor loop? count > %d - driver confused?\r\n",
364	    name, i);
365	return (-1);
366}
367
368/*
369 * Return the currently-first request chain to the guest, setting
370 * its I/O length to the provided value.
371 *
372 * (This chain is the one you handled when you called vq_getchain()
373 * and used its positive return value.)
374 */
375void
376vq_relchain(struct vqueue_info *vq, uint32_t iolen)
377{
378	uint16_t head, uidx, mask;
379	volatile struct vring_used *vuh;
380	volatile struct virtio_used *vue;
381
382	/*
383	 * Notes:
384	 *  - mask is N-1 where N is a power of 2 so computes x % N
385	 *  - vuh points to the "used" data shared with guest
386	 *  - vue points to the "used" ring entry we want to update
387	 *  - head is the same value we compute in vq_iovecs().
388	 *
389	 * (I apologize for the two fields named vu_idx; the
390	 * virtio spec calls the one that vue points to, "id"...)
391	 */
392	mask = vq->vq_qsize - 1;
393	vuh = vq->vq_used;
394	head = vq->vq_avail->va_ring[vq->vq_last_avail++ & mask];
395
396	uidx = vuh->vu_idx;
397	vue = &vuh->vu_ring[uidx++ & mask];
398	vue->vu_idx = head; /* ie, vue->id = head */
399	vue->vu_tlen = iolen;
400	vuh->vu_idx = uidx;
401}
402
403/*
404 * Driver has finished processing "available" chains and calling
405 * vq_relchain on each one.  If driver used all the available
406 * chains, used_all should be set.
407 *
408 * If the "used" index moved we may need to inform the guest, i.e.,
409 * deliver an interrupt.  Even if the used index did NOT move we
410 * may need to deliver an interrupt, if the avail ring is empty and
411 * we are supposed to interrupt on empty.
412 *
413 * Note that used_all_avail is provided by the caller because it's
414 * a snapshot of the ring state when he decided to finish interrupt
415 * processing -- it's possible that descriptors became available after
416 * that point.  (It's also typically a constant 1/True as well.)
417 */
418void
419vq_endchains(struct vqueue_info *vq, int used_all_avail)
420{
421	struct virtio_softc *vs;
422	uint16_t event_idx, new_idx, old_idx;
423	int intr;
424
425	/*
426	 * Interrupt generation: if we're using EVENT_IDX,
427	 * interrupt if we've crossed the event threshold.
428	 * Otherwise interrupt is generated if we added "used" entries,
429	 * but suppressed by VRING_AVAIL_F_NO_INTERRUPT.
430	 *
431	 * In any case, though, if NOTIFY_ON_EMPTY is set and the
432	 * entire avail was processed, we need to interrupt always.
433	 */
434	vs = vq->vq_vs;
435	new_idx = vq->vq_used->vu_idx;
436	old_idx = vq->vq_save_used;
437	if (used_all_avail &&
438	    (vs->vs_negotiated_caps & VIRTIO_F_NOTIFY_ON_EMPTY))
439		intr = 1;
440	else if (vs->vs_negotiated_caps & VIRTIO_RING_F_EVENT_IDX) {
441		event_idx = VQ_USED_EVENT_IDX(vq);
442		/*
443		 * This calculation is per docs and the kernel
444		 * (see src/sys/dev/virtio/virtio_ring.h).
445		 */
446		intr = (uint16_t)(new_idx - event_idx - 1) <
447			(uint16_t)(new_idx - old_idx);
448	} else {
449		intr = new_idx != old_idx &&
450		    !(vq->vq_avail->va_flags & VRING_AVAIL_F_NO_INTERRUPT);
451	}
452	if (intr)
453		vq_interrupt(vs, vq);
454}
455
456/* Note: these are in sorted order to make for a fast search */
457static struct config_reg {
458	uint16_t	cr_offset;	/* register offset */
459	uint8_t		cr_size;	/* size (bytes) */
460	uint8_t		cr_ro;		/* true => reg is read only */
461	const char	*cr_name;	/* name of reg */
462} config_regs[] = {
463	{ VTCFG_R_HOSTCAP,	4, 1, "HOSTCAP" },
464	{ VTCFG_R_GUESTCAP,	4, 0, "GUESTCAP" },
465	{ VTCFG_R_PFN,		4, 0, "PFN" },
466	{ VTCFG_R_QNUM,		2, 1, "QNUM" },
467	{ VTCFG_R_QSEL,		2, 0, "QSEL" },
468	{ VTCFG_R_QNOTIFY,	2, 0, "QNOTIFY" },
469	{ VTCFG_R_STATUS,	1, 0, "STATUS" },
470	{ VTCFG_R_ISR,		1, 0, "ISR" },
471	{ VTCFG_R_CFGVEC,	2, 0, "CFGVEC" },
472	{ VTCFG_R_QVEC,		2, 0, "QVEC" },
473};
474
475static inline struct config_reg *
476vi_find_cr(int offset) {
477	u_int hi, lo, mid;
478	struct config_reg *cr;
479
480	lo = 0;
481	hi = sizeof(config_regs) / sizeof(*config_regs) - 1;
482	while (hi >= lo) {
483		mid = (hi + lo) >> 1;
484		cr = &config_regs[mid];
485		if (cr->cr_offset == offset)
486			return (cr);
487		if (cr->cr_offset < offset)
488			lo = mid + 1;
489		else
490			hi = mid - 1;
491	}
492	return (NULL);
493}
494
495/*
496 * Handle pci config space reads.
497 * If it's to the MSI-X info, do that.
498 * If it's part of the virtio standard stuff, do that.
499 * Otherwise dispatch to the actual driver.
500 */
501uint64_t
502vi_pci_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
503	    int baridx, uint64_t offset, int size)
504{
505	struct virtio_softc *vs = pi->pi_arg;
506	struct virtio_consts *vc;
507	struct config_reg *cr;
508	uint64_t virtio_config_size, max;
509	const char *name;
510	uint32_t newoff;
511	uint32_t value;
512	int error;
513
514	if (vs->vs_flags & VIRTIO_USE_MSIX) {
515		if (baridx == pci_msix_table_bar(pi) ||
516		    baridx == pci_msix_pba_bar(pi)) {
517			return (pci_emul_msix_tread(pi, offset, size));
518		}
519	}
520
521	/* XXX probably should do something better than just assert() */
522	assert(baridx == 0);
523
524	if (vs->vs_mtx)
525		pthread_mutex_lock(vs->vs_mtx);
526
527	vc = vs->vs_vc;
528	name = vc->vc_name;
529	value = size == 1 ? 0xff : size == 2 ? 0xffff : 0xffffffff;
530
531	if (size != 1 && size != 2 && size != 4)
532		goto bad;
533
534	if (pci_msix_enabled(pi))
535		virtio_config_size = VTCFG_R_CFG1;
536	else
537		virtio_config_size = VTCFG_R_CFG0;
538
539	if (offset >= virtio_config_size) {
540		/*
541		 * Subtract off the standard size (including MSI-X
542		 * registers if enabled) and dispatch to underlying driver.
543		 * If that fails, fall into general code.
544		 */
545		newoff = offset - virtio_config_size;
546		max = vc->vc_cfgsize ? vc->vc_cfgsize : 0x100000000;
547		if (newoff + size > max)
548			goto bad;
549		error = (*vc->vc_cfgread)(DEV_SOFTC(vs), newoff, size, &value);
550		if (!error)
551			goto done;
552	}
553
554bad:
555	cr = vi_find_cr(offset);
556	if (cr == NULL || cr->cr_size != size) {
557		if (cr != NULL) {
558			/* offset must be OK, so size must be bad */
559			fprintf(stderr,
560			    "%s: read from %s: bad size %d\r\n",
561			    name, cr->cr_name, size);
562		} else {
563			fprintf(stderr,
564			    "%s: read from bad offset/size %jd/%d\r\n",
565			    name, (uintmax_t)offset, size);
566		}
567		goto done;
568	}
569
570	switch (offset) {
571	case VTCFG_R_HOSTCAP:
572		value = vc->vc_hv_caps;
573		break;
574	case VTCFG_R_GUESTCAP:
575		value = vs->vs_negotiated_caps;
576		break;
577	case VTCFG_R_PFN:
578		if (vs->vs_curq < vc->vc_nvq)
579			value = vs->vs_queues[vs->vs_curq].vq_pfn;
580		break;
581	case VTCFG_R_QNUM:
582		value = vs->vs_curq < vc->vc_nvq ?
583		    vs->vs_queues[vs->vs_curq].vq_qsize : 0;
584		break;
585	case VTCFG_R_QSEL:
586		value = vs->vs_curq;
587		break;
588	case VTCFG_R_QNOTIFY:
589		value = 0;	/* XXX */
590		break;
591	case VTCFG_R_STATUS:
592		value = vs->vs_status;
593		break;
594	case VTCFG_R_ISR:
595		value = vs->vs_isr;
596		vs->vs_isr = 0;		/* a read clears this flag */
597		if (value)
598			pci_lintr_deassert(pi);
599		break;
600	case VTCFG_R_CFGVEC:
601		value = vs->vs_msix_cfg_idx;
602		break;
603	case VTCFG_R_QVEC:
604		value = vs->vs_curq < vc->vc_nvq ?
605		    vs->vs_queues[vs->vs_curq].vq_msix_idx :
606		    VIRTIO_MSI_NO_VECTOR;
607		break;
608	}
609done:
610	if (vs->vs_mtx)
611		pthread_mutex_unlock(vs->vs_mtx);
612	return (value);
613}
614
615/*
616 * Handle pci config space writes.
617 * If it's to the MSI-X info, do that.
618 * If it's part of the virtio standard stuff, do that.
619 * Otherwise dispatch to the actual driver.
620 */
621void
622vi_pci_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
623	     int baridx, uint64_t offset, int size, uint64_t value)
624{
625	struct virtio_softc *vs = pi->pi_arg;
626	struct vqueue_info *vq;
627	struct virtio_consts *vc;
628	struct config_reg *cr;
629	uint64_t virtio_config_size, max;
630	const char *name;
631	uint32_t newoff;
632	int error;
633
634	if (vs->vs_flags & VIRTIO_USE_MSIX) {
635		if (baridx == pci_msix_table_bar(pi) ||
636		    baridx == pci_msix_pba_bar(pi)) {
637			pci_emul_msix_twrite(pi, offset, size, value);
638			return;
639		}
640	}
641
642	/* XXX probably should do something better than just assert() */
643	assert(baridx == 0);
644
645	if (vs->vs_mtx)
646		pthread_mutex_lock(vs->vs_mtx);
647
648	vc = vs->vs_vc;
649	name = vc->vc_name;
650
651	if (size != 1 && size != 2 && size != 4)
652		goto bad;
653
654	if (pci_msix_enabled(pi))
655		virtio_config_size = VTCFG_R_CFG1;
656	else
657		virtio_config_size = VTCFG_R_CFG0;
658
659	if (offset >= virtio_config_size) {
660		/*
661		 * Subtract off the standard size (including MSI-X
662		 * registers if enabled) and dispatch to underlying driver.
663		 */
664		newoff = offset - virtio_config_size;
665		max = vc->vc_cfgsize ? vc->vc_cfgsize : 0x100000000;
666		if (newoff + size > max)
667			goto bad;
668		error = (*vc->vc_cfgwrite)(DEV_SOFTC(vs), newoff, size, value);
669		if (!error)
670			goto done;
671	}
672
673bad:
674	cr = vi_find_cr(offset);
675	if (cr == NULL || cr->cr_size != size || cr->cr_ro) {
676		if (cr != NULL) {
677			/* offset must be OK, wrong size and/or reg is R/O */
678			if (cr->cr_size != size)
679				fprintf(stderr,
680				    "%s: write to %s: bad size %d\r\n",
681				    name, cr->cr_name, size);
682			if (cr->cr_ro)
683				fprintf(stderr,
684				    "%s: write to read-only reg %s\r\n",
685				    name, cr->cr_name);
686		} else {
687			fprintf(stderr,
688			    "%s: write to bad offset/size %jd/%d\r\n",
689			    name, (uintmax_t)offset, size);
690		}
691		goto done;
692	}
693
694	switch (offset) {
695	case VTCFG_R_GUESTCAP:
696		vs->vs_negotiated_caps = value & vc->vc_hv_caps;
697		if (vc->vc_apply_features)
698			(*vc->vc_apply_features)(DEV_SOFTC(vs),
699			    vs->vs_negotiated_caps);
700		break;
701	case VTCFG_R_PFN:
702		if (vs->vs_curq >= vc->vc_nvq)
703			goto bad_qindex;
704		vi_vq_init(vs, value);
705		break;
706	case VTCFG_R_QSEL:
707		/*
708		 * Note that the guest is allowed to select an
709		 * invalid queue; we just need to return a QNUM
710		 * of 0 while the bad queue is selected.
711		 */
712		vs->vs_curq = value;
713		break;
714	case VTCFG_R_QNOTIFY:
715		if (value >= vc->vc_nvq) {
716			fprintf(stderr, "%s: queue %d notify out of range\r\n",
717				name, (int)value);
718			goto done;
719		}
720		vq = &vs->vs_queues[value];
721		if (vq->vq_notify)
722			(*vq->vq_notify)(DEV_SOFTC(vs), vq);
723		else if (vc->vc_qnotify)
724			(*vc->vc_qnotify)(DEV_SOFTC(vs), vq);
725		else
726			fprintf(stderr,
727			    "%s: qnotify queue %d: missing vq/vc notify\r\n",
728				name, (int)value);
729		break;
730	case VTCFG_R_STATUS:
731		vs->vs_status = value;
732		if (value == 0)
733			(*vc->vc_reset)(DEV_SOFTC(vs));
734		break;
735	case VTCFG_R_CFGVEC:
736		vs->vs_msix_cfg_idx = value;
737		break;
738	case VTCFG_R_QVEC:
739		if (vs->vs_curq >= vc->vc_nvq)
740			goto bad_qindex;
741		vq = &vs->vs_queues[vs->vs_curq];
742		vq->vq_msix_idx = value;
743		break;
744	}
745	goto done;
746
747bad_qindex:
748	fprintf(stderr,
749	    "%s: write config reg %s: curq %d >= max %d\r\n",
750	    name, cr->cr_name, vs->vs_curq, vc->vc_nvq);
751done:
752	if (vs->vs_mtx)
753		pthread_mutex_unlock(vs->vs_mtx);
754}
755