xref: /freebsd/sys/dev/virtio/virtqueue.c (revision 068d8643adb13434b57fdd5febf43a11194d508f)
1 /*-
2  * Copyright (c) 2011, Bryan Venteicher <bryanv@FreeBSD.org>
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 unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 /*
28  * Implements the virtqueue interface as basically described
29  * in the original VirtIO paper.
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/malloc.h>
39 #include <sys/sglist.h>
40 #include <vm/vm.h>
41 #include <vm/pmap.h>
42 
43 #include <machine/cpu.h>
44 #include <machine/bus.h>
45 #include <machine/atomic.h>
46 #include <machine/resource.h>
47 #include <sys/bus.h>
48 #include <sys/rman.h>
49 
50 #include <dev/virtio/virtio.h>
51 #include <dev/virtio/virtio_config.h>
52 #include <dev/virtio/virtqueue.h>
53 #include <dev/virtio/virtio_ring.h>
54 
55 #include "virtio_bus_if.h"
56 
57 struct virtqueue {
58 	device_t		 vq_dev;
59 	char			 vq_name[VIRTQUEUE_MAX_NAME_SZ];
60 	uint16_t		 vq_queue_index;
61 	uint16_t		 vq_nentries;
62 	uint32_t		 vq_flags;
63 #define	VIRTQUEUE_FLAG_INDIRECT	 0x0001
64 #define	VIRTQUEUE_FLAG_EVENT_IDX 0x0002
65 
66 	int			 vq_alignment;
67 	int			 vq_ring_size;
68 	void			*vq_ring_mem;
69 	int			 vq_max_indirect_size;
70 	int			 vq_indirect_mem_size;
71 	virtqueue_intr_t	*vq_intrhand;
72 	void			*vq_intrhand_arg;
73 
74 	struct vring		 vq_ring;
75 	uint16_t		 vq_free_cnt;
76 	uint16_t		 vq_queued_cnt;
77 	/*
78 	 * Head of the free chain in the descriptor table. If
79 	 * there are no free descriptors, this will be set to
80 	 * VQ_RING_DESC_CHAIN_END.
81 	 */
82 	uint16_t		 vq_desc_head_idx;
83 	/*
84 	 * Last consumed descriptor in the used table,
85 	 * trails vq_ring.used->idx.
86 	 */
87 	uint16_t		 vq_used_cons_idx;
88 
89 	struct vq_desc_extra {
90 		void		  *cookie;
91 		struct vring_desc *indirect;
92 		vm_paddr_t	   indirect_paddr;
93 		uint16_t	   ndescs;
94 	} vq_descx[0];
95 };
96 
97 /*
98  * The maximum virtqueue size is 2^15. Use that value as the end of
99  * descriptor chain terminator since it will never be a valid index
100  * in the descriptor table. This is used to verify we are correctly
101  * handling vq_free_cnt.
102  */
103 #define VQ_RING_DESC_CHAIN_END 32768
104 
105 #define VQASSERT(_vq, _exp, _msg, ...)				\
106     KASSERT((_exp),("%s: %s - "_msg, __func__, (_vq)->vq_name,	\
107 	##__VA_ARGS__))
108 
109 #define VQ_RING_ASSERT_VALID_IDX(_vq, _idx)			\
110     VQASSERT((_vq), (_idx) < (_vq)->vq_nentries,		\
111 	"invalid ring index: %d, max: %d", (_idx),		\
112 	(_vq)->vq_nentries)
113 
114 #define VQ_RING_ASSERT_CHAIN_TERM(_vq)				\
115     VQASSERT((_vq), (_vq)->vq_desc_head_idx ==			\
116 	VQ_RING_DESC_CHAIN_END,	"full ring terminated "		\
117 	"incorrectly: head idx: %d", (_vq)->vq_desc_head_idx)
118 
119 static int	virtqueue_init_indirect(struct virtqueue *vq, int);
120 static void	virtqueue_free_indirect(struct virtqueue *vq);
121 static void	virtqueue_init_indirect_list(struct virtqueue *,
122 		    struct vring_desc *);
123 
124 static void	vq_ring_init(struct virtqueue *);
125 static void	vq_ring_update_avail(struct virtqueue *, uint16_t);
126 static uint16_t	vq_ring_enqueue_segments(struct virtqueue *,
127 		    struct vring_desc *, uint16_t, struct sglist *, int, int);
128 static int	vq_ring_use_indirect(struct virtqueue *, int);
129 static void	vq_ring_enqueue_indirect(struct virtqueue *, void *,
130 		    struct sglist *, int, int);
131 static int	vq_ring_enable_interrupt(struct virtqueue *, uint16_t);
132 static int	vq_ring_must_notify_host(struct virtqueue *);
133 static void	vq_ring_notify_host(struct virtqueue *);
134 static void	vq_ring_free_chain(struct virtqueue *, uint16_t);
135 
136 uint64_t
137 virtqueue_filter_features(uint64_t features)
138 {
139 	uint64_t mask;
140 
141 	mask = (1 << VIRTIO_TRANSPORT_F_START) - 1;
142 	mask |= VIRTIO_RING_F_INDIRECT_DESC;
143 	mask |= VIRTIO_RING_F_EVENT_IDX;
144 
145 	return (features & mask);
146 }
147 
148 int
149 virtqueue_alloc(device_t dev, uint16_t queue, uint16_t size, int align,
150     vm_paddr_t highaddr, struct vq_alloc_info *info, struct virtqueue **vqp)
151 {
152 	struct virtqueue *vq;
153 	int error;
154 
155 	*vqp = NULL;
156 	error = 0;
157 
158 	if (size == 0) {
159 		device_printf(dev,
160 		    "virtqueue %d (%s) does not exist (size is zero)\n",
161 		    queue, info->vqai_name);
162 		return (ENODEV);
163 	} else if (!powerof2(size)) {
164 		device_printf(dev,
165 		    "virtqueue %d (%s) size is not a power of 2: %d\n",
166 		    queue, info->vqai_name, size);
167 		return (ENXIO);
168 	} else if (info->vqai_maxindirsz > VIRTIO_MAX_INDIRECT) {
169 		device_printf(dev, "virtqueue %d (%s) requested too many "
170 		    "indirect descriptors: %d, max %d\n",
171 		    queue, info->vqai_name, info->vqai_maxindirsz,
172 		    VIRTIO_MAX_INDIRECT);
173 		return (EINVAL);
174 	}
175 
176 	vq = malloc(sizeof(struct virtqueue) +
177 	    size * sizeof(struct vq_desc_extra), M_DEVBUF, M_NOWAIT | M_ZERO);
178 	if (vq == NULL) {
179 		device_printf(dev, "cannot allocate virtqueue\n");
180 		return (ENOMEM);
181 	}
182 
183 	vq->vq_dev = dev;
184 	strlcpy(vq->vq_name, info->vqai_name, sizeof(vq->vq_name));
185 	vq->vq_queue_index = queue;
186 	vq->vq_alignment = align;
187 	vq->vq_nentries = size;
188 	vq->vq_free_cnt = size;
189 	vq->vq_intrhand = info->vqai_intr;
190 	vq->vq_intrhand_arg = info->vqai_intr_arg;
191 
192 	if (VIRTIO_BUS_WITH_FEATURE(dev, VIRTIO_RING_F_EVENT_IDX) != 0)
193 		vq->vq_flags |= VIRTQUEUE_FLAG_EVENT_IDX;
194 
195 	if (info->vqai_maxindirsz > 1) {
196 		error = virtqueue_init_indirect(vq, info->vqai_maxindirsz);
197 		if (error)
198 			goto fail;
199 	}
200 
201 	vq->vq_ring_size = round_page(vring_size(size, align));
202 	vq->vq_ring_mem = contigmalloc(vq->vq_ring_size, M_DEVBUF,
203 	    M_NOWAIT | M_ZERO, 0, highaddr, PAGE_SIZE, 0);
204 	if (vq->vq_ring_mem == NULL) {
205 		device_printf(dev,
206 		    "cannot allocate memory for virtqueue ring\n");
207 		error = ENOMEM;
208 		goto fail;
209 	}
210 
211 	vq_ring_init(vq);
212 	virtqueue_disable_intr(vq);
213 
214 	*vqp = vq;
215 
216 fail:
217 	if (error)
218 		virtqueue_free(vq);
219 
220 	return (error);
221 }
222 
223 static int
224 virtqueue_init_indirect(struct virtqueue *vq, int indirect_size)
225 {
226 	device_t dev;
227 	struct vq_desc_extra *dxp;
228 	int i, size;
229 
230 	dev = vq->vq_dev;
231 
232 	if (VIRTIO_BUS_WITH_FEATURE(dev, VIRTIO_RING_F_INDIRECT_DESC) == 0) {
233 		/*
234 		 * Indirect descriptors requested by the driver but not
235 		 * negotiated. Return zero to keep the initialization
236 		 * going: we'll run fine without.
237 		 */
238 		if (bootverbose)
239 			device_printf(dev, "virtqueue %d (%s) requested "
240 			    "indirect descriptors but not negotiated\n",
241 			    vq->vq_queue_index, vq->vq_name);
242 		return (0);
243 	}
244 
245 	size = indirect_size * sizeof(struct vring_desc);
246 	vq->vq_max_indirect_size = indirect_size;
247 	vq->vq_indirect_mem_size = size;
248 	vq->vq_flags |= VIRTQUEUE_FLAG_INDIRECT;
249 
250 	for (i = 0; i < vq->vq_nentries; i++) {
251 		dxp = &vq->vq_descx[i];
252 
253 		dxp->indirect = malloc(size, M_DEVBUF, M_NOWAIT);
254 		if (dxp->indirect == NULL) {
255 			device_printf(dev, "cannot allocate indirect list\n");
256 			return (ENOMEM);
257 		}
258 
259 		dxp->indirect_paddr = vtophys(dxp->indirect);
260 		virtqueue_init_indirect_list(vq, dxp->indirect);
261 	}
262 
263 	return (0);
264 }
265 
266 static void
267 virtqueue_free_indirect(struct virtqueue *vq)
268 {
269 	struct vq_desc_extra *dxp;
270 	int i;
271 
272 	for (i = 0; i < vq->vq_nentries; i++) {
273 		dxp = &vq->vq_descx[i];
274 
275 		if (dxp->indirect == NULL)
276 			break;
277 
278 		free(dxp->indirect, M_DEVBUF);
279 		dxp->indirect = NULL;
280 		dxp->indirect_paddr = 0;
281 	}
282 
283 	vq->vq_flags &= ~VIRTQUEUE_FLAG_INDIRECT;
284 	vq->vq_indirect_mem_size = 0;
285 }
286 
287 static void
288 virtqueue_init_indirect_list(struct virtqueue *vq,
289     struct vring_desc *indirect)
290 {
291 	int i;
292 
293 	bzero(indirect, vq->vq_indirect_mem_size);
294 
295 	for (i = 0; i < vq->vq_max_indirect_size - 1; i++)
296 		indirect[i].next = i + 1;
297 	indirect[i].next = VQ_RING_DESC_CHAIN_END;
298 }
299 
300 int
301 virtqueue_reinit(struct virtqueue *vq, uint16_t size)
302 {
303 	struct vq_desc_extra *dxp;
304 	int i;
305 
306 	if (vq->vq_nentries != size) {
307 		device_printf(vq->vq_dev,
308 		    "%s: '%s' changed size; old=%hu, new=%hu\n",
309 		    __func__, vq->vq_name, vq->vq_nentries, size);
310 		return (EINVAL);
311 	}
312 
313 	/* Warn if the virtqueue was not properly cleaned up. */
314 	if (vq->vq_free_cnt != vq->vq_nentries) {
315 		device_printf(vq->vq_dev,
316 		    "%s: warning '%s' virtqueue not empty, "
317 		    "leaking %d entries\n", __func__, vq->vq_name,
318 		    vq->vq_nentries - vq->vq_free_cnt);
319 	}
320 
321 	vq->vq_desc_head_idx = 0;
322 	vq->vq_used_cons_idx = 0;
323 	vq->vq_queued_cnt = 0;
324 	vq->vq_free_cnt = vq->vq_nentries;
325 
326 	/* To be safe, reset all our allocated memory. */
327 	bzero(vq->vq_ring_mem, vq->vq_ring_size);
328 	for (i = 0; i < vq->vq_nentries; i++) {
329 		dxp = &vq->vq_descx[i];
330 		dxp->cookie = NULL;
331 		dxp->ndescs = 0;
332 		if (vq->vq_flags & VIRTQUEUE_FLAG_INDIRECT)
333 			virtqueue_init_indirect_list(vq, dxp->indirect);
334 	}
335 
336 	vq_ring_init(vq);
337 	virtqueue_disable_intr(vq);
338 
339 	return (0);
340 }
341 
342 void
343 virtqueue_free(struct virtqueue *vq)
344 {
345 
346 	if (vq->vq_free_cnt != vq->vq_nentries) {
347 		device_printf(vq->vq_dev, "%s: freeing non-empty virtqueue, "
348 		    "leaking %d entries\n", vq->vq_name,
349 		    vq->vq_nentries - vq->vq_free_cnt);
350 	}
351 
352 	if (vq->vq_flags & VIRTQUEUE_FLAG_INDIRECT)
353 		virtqueue_free_indirect(vq);
354 
355 	if (vq->vq_ring_mem != NULL) {
356 		contigfree(vq->vq_ring_mem, vq->vq_ring_size, M_DEVBUF);
357 		vq->vq_ring_size = 0;
358 		vq->vq_ring_mem = NULL;
359 	}
360 
361 	free(vq, M_DEVBUF);
362 }
363 
364 vm_paddr_t
365 virtqueue_paddr(struct virtqueue *vq)
366 {
367 
368 	return (vtophys(vq->vq_ring_mem));
369 }
370 
371 int
372 virtqueue_size(struct virtqueue *vq)
373 {
374 
375 	return (vq->vq_nentries);
376 }
377 
378 int
379 virtqueue_empty(struct virtqueue *vq)
380 {
381 
382 	return (vq->vq_nentries == vq->vq_free_cnt);
383 }
384 
385 int
386 virtqueue_full(struct virtqueue *vq)
387 {
388 
389 	return (vq->vq_free_cnt == 0);
390 }
391 
392 void
393 virtqueue_notify(struct virtqueue *vq)
394 {
395 
396 	/* Ensure updated avail->idx is visible to host. */
397 	mb();
398 
399 	if (vq_ring_must_notify_host(vq))
400 		vq_ring_notify_host(vq);
401 	vq->vq_queued_cnt = 0;
402 }
403 
404 int
405 virtqueue_nused(struct virtqueue *vq)
406 {
407 	uint16_t used_idx, nused;
408 
409 	used_idx = vq->vq_ring.used->idx;
410 
411 	nused = (uint16_t)(used_idx - vq->vq_used_cons_idx);
412 	VQASSERT(vq, nused <= vq->vq_nentries, "used more than available");
413 
414 	return (nused);
415 }
416 
417 int
418 virtqueue_intr_filter(struct virtqueue *vq)
419 {
420 
421 	if (vq->vq_used_cons_idx == vq->vq_ring.used->idx)
422 		return (0);
423 
424 	virtqueue_disable_intr(vq);
425 
426 	return (1);
427 }
428 
429 void
430 virtqueue_intr(struct virtqueue *vq)
431 {
432 
433 	vq->vq_intrhand(vq->vq_intrhand_arg);
434 }
435 
436 int
437 virtqueue_enable_intr(struct virtqueue *vq)
438 {
439 
440 	return (vq_ring_enable_interrupt(vq, 0));
441 }
442 
443 int
444 virtqueue_postpone_intr(struct virtqueue *vq, vq_postpone_t hint)
445 {
446 	uint16_t ndesc, avail_idx;
447 
448 	avail_idx = vq->vq_ring.avail->idx;
449 	ndesc = (uint16_t)(avail_idx - vq->vq_used_cons_idx);
450 
451 	switch (hint) {
452 	case VQ_POSTPONE_SHORT:
453 		ndesc = ndesc / 4;
454 		break;
455 	case VQ_POSTPONE_LONG:
456 		ndesc = (ndesc * 3) / 4;
457 		break;
458 	case VQ_POSTPONE_EMPTIED:
459 		break;
460 	}
461 
462 	return (vq_ring_enable_interrupt(vq, ndesc));
463 }
464 
465 /*
466  * Note this is only considered a hint to the host.
467  */
468 void
469 virtqueue_disable_intr(struct virtqueue *vq)
470 {
471 
472 	if (vq->vq_flags & VIRTQUEUE_FLAG_EVENT_IDX) {
473 		vring_used_event(&vq->vq_ring) = vq->vq_used_cons_idx -
474 		    vq->vq_nentries - 1;
475 	} else
476 		vq->vq_ring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
477 }
478 
479 int
480 virtqueue_enqueue(struct virtqueue *vq, void *cookie, struct sglist *sg,
481     int readable, int writable)
482 {
483 	struct vq_desc_extra *dxp;
484 	int needed;
485 	uint16_t head_idx, idx;
486 
487 	needed = readable + writable;
488 
489 	VQASSERT(vq, cookie != NULL, "enqueuing with no cookie");
490 	VQASSERT(vq, needed == sg->sg_nseg,
491 	    "segment count mismatch, %d, %d", needed, sg->sg_nseg);
492 	VQASSERT(vq,
493 	    needed <= vq->vq_nentries || needed <= vq->vq_max_indirect_size,
494 	    "too many segments to enqueue: %d, %d/%d", needed,
495 	    vq->vq_nentries, vq->vq_max_indirect_size);
496 
497 	if (needed < 1)
498 		return (EINVAL);
499 	if (vq->vq_free_cnt == 0)
500 		return (ENOSPC);
501 
502 	if (vq_ring_use_indirect(vq, needed)) {
503 		vq_ring_enqueue_indirect(vq, cookie, sg, readable, writable);
504 		return (0);
505 	} else if (vq->vq_free_cnt < needed)
506 		return (EMSGSIZE);
507 
508 	head_idx = vq->vq_desc_head_idx;
509 	VQ_RING_ASSERT_VALID_IDX(vq, head_idx);
510 	dxp = &vq->vq_descx[head_idx];
511 
512 	VQASSERT(vq, dxp->cookie == NULL,
513 	    "cookie already exists for index %d", head_idx);
514 	dxp->cookie = cookie;
515 	dxp->ndescs = needed;
516 
517 	idx = vq_ring_enqueue_segments(vq, vq->vq_ring.desc, head_idx,
518 	    sg, readable, writable);
519 
520 	vq->vq_desc_head_idx = idx;
521 	vq->vq_free_cnt -= needed;
522 	if (vq->vq_free_cnt == 0)
523 		VQ_RING_ASSERT_CHAIN_TERM(vq);
524 	else
525 		VQ_RING_ASSERT_VALID_IDX(vq, idx);
526 
527 	vq_ring_update_avail(vq, head_idx);
528 
529 	return (0);
530 }
531 
532 void *
533 virtqueue_dequeue(struct virtqueue *vq, uint32_t *len)
534 {
535 	struct vring_used_elem *uep;
536 	void *cookie;
537 	uint16_t used_idx, desc_idx;
538 
539 	if (vq->vq_used_cons_idx == vq->vq_ring.used->idx)
540 		return (NULL);
541 
542 	used_idx = vq->vq_used_cons_idx++ & (vq->vq_nentries - 1);
543 	uep = &vq->vq_ring.used->ring[used_idx];
544 
545 	rmb();
546 	desc_idx = (uint16_t) uep->id;
547 	if (len != NULL)
548 		*len = uep->len;
549 
550 	vq_ring_free_chain(vq, desc_idx);
551 
552 	cookie = vq->vq_descx[desc_idx].cookie;
553 	VQASSERT(vq, cookie != NULL, "no cookie for index %d", desc_idx);
554 	vq->vq_descx[desc_idx].cookie = NULL;
555 
556 	return (cookie);
557 }
558 
559 void *
560 virtqueue_poll(struct virtqueue *vq, uint32_t *len)
561 {
562 	void *cookie;
563 
564 	while ((cookie = virtqueue_dequeue(vq, len)) == NULL)
565 		cpu_spinwait();
566 
567 	return (cookie);
568 }
569 
570 void *
571 virtqueue_drain(struct virtqueue *vq, int *last)
572 {
573 	void *cookie;
574 	int idx;
575 
576 	cookie = NULL;
577 	idx = *last;
578 
579 	while (idx < vq->vq_nentries && cookie == NULL) {
580 		if ((cookie = vq->vq_descx[idx].cookie) != NULL) {
581 			vq->vq_descx[idx].cookie = NULL;
582 			/* Free chain to keep free count consistent. */
583 			vq_ring_free_chain(vq, idx);
584 		}
585 		idx++;
586 	}
587 
588 	*last = idx;
589 
590 	return (cookie);
591 }
592 
593 void
594 virtqueue_dump(struct virtqueue *vq)
595 {
596 
597 	if (vq == NULL)
598 		return;
599 
600 	printf("VQ: %s - size=%d; free=%d; used=%d; queued=%d; "
601 	    "desc_head_idx=%d; avail.idx=%d; used_cons_idx=%d; "
602 	    "used.idx=%d; avail.flags=0x%x; used.flags=0x%x\n",
603 	    vq->vq_name, vq->vq_nentries, vq->vq_free_cnt,
604 	    virtqueue_nused(vq), vq->vq_queued_cnt, vq->vq_desc_head_idx,
605 	    vq->vq_ring.avail->idx, vq->vq_used_cons_idx,
606 	    vq->vq_ring.used->idx, vq->vq_ring.avail->flags,
607 	    vq->vq_ring.used->flags);
608 }
609 
610 static void
611 vq_ring_init(struct virtqueue *vq)
612 {
613 	struct vring *vr;
614 	char *ring_mem;
615 	int i, size;
616 
617 	ring_mem = vq->vq_ring_mem;
618 	size = vq->vq_nentries;
619 	vr = &vq->vq_ring;
620 
621 	vring_init(vr, size, ring_mem, vq->vq_alignment);
622 
623 	for (i = 0; i < size - 1; i++)
624 		vr->desc[i].next = i + 1;
625 	vr->desc[i].next = VQ_RING_DESC_CHAIN_END;
626 }
627 
628 static void
629 vq_ring_update_avail(struct virtqueue *vq, uint16_t desc_idx)
630 {
631 	uint16_t avail_idx;
632 
633 	/*
634 	 * Place the head of the descriptor chain into the next slot and make
635 	 * it usable to the host. The chain is made available now rather than
636 	 * deferring to virtqueue_notify() in the hopes that if the host is
637 	 * currently running on another CPU, we can keep it processing the new
638 	 * descriptor.
639 	 */
640 	avail_idx = vq->vq_ring.avail->idx & (vq->vq_nentries - 1);
641 	vq->vq_ring.avail->ring[avail_idx] = desc_idx;
642 
643 	wmb();
644 	vq->vq_ring.avail->idx++;
645 
646 	/* Keep pending count until virtqueue_notify(). */
647 	vq->vq_queued_cnt++;
648 }
649 
650 static uint16_t
651 vq_ring_enqueue_segments(struct virtqueue *vq, struct vring_desc *desc,
652     uint16_t head_idx, struct sglist *sg, int readable, int writable)
653 {
654 	struct sglist_seg *seg;
655 	struct vring_desc *dp;
656 	int i, needed;
657 	uint16_t idx;
658 
659 	needed = readable + writable;
660 
661 	for (i = 0, idx = head_idx, seg = sg->sg_segs;
662 	     i < needed;
663 	     i++, idx = dp->next, seg++) {
664 		VQASSERT(vq, idx != VQ_RING_DESC_CHAIN_END,
665 		    "premature end of free desc chain");
666 
667 		dp = &desc[idx];
668 		dp->addr = seg->ss_paddr;
669 		dp->len = seg->ss_len;
670 		dp->flags = 0;
671 
672 		if (i < needed - 1)
673 			dp->flags |= VRING_DESC_F_NEXT;
674 		if (i >= readable)
675 			dp->flags |= VRING_DESC_F_WRITE;
676 	}
677 
678 	return (idx);
679 }
680 
681 static int
682 vq_ring_use_indirect(struct virtqueue *vq, int needed)
683 {
684 
685 	if ((vq->vq_flags & VIRTQUEUE_FLAG_INDIRECT) == 0)
686 		return (0);
687 
688 	if (vq->vq_max_indirect_size < needed)
689 		return (0);
690 
691 	if (needed < 2)
692 		return (0);
693 
694 	return (1);
695 }
696 
697 static void
698 vq_ring_enqueue_indirect(struct virtqueue *vq, void *cookie,
699     struct sglist *sg, int readable, int writable)
700 {
701 	struct vring_desc *dp;
702 	struct vq_desc_extra *dxp;
703 	int needed;
704 	uint16_t head_idx;
705 
706 	needed = readable + writable;
707 	VQASSERT(vq, needed <= vq->vq_max_indirect_size,
708 	    "enqueuing too many indirect descriptors");
709 
710 	head_idx = vq->vq_desc_head_idx;
711 	VQ_RING_ASSERT_VALID_IDX(vq, head_idx);
712 	dp = &vq->vq_ring.desc[head_idx];
713 	dxp = &vq->vq_descx[head_idx];
714 
715 	VQASSERT(vq, dxp->cookie == NULL,
716 	    "cookie already exists for index %d", head_idx);
717 	dxp->cookie = cookie;
718 	dxp->ndescs = 1;
719 
720 	dp->addr = dxp->indirect_paddr;
721 	dp->len = needed * sizeof(struct vring_desc);
722 	dp->flags = VRING_DESC_F_INDIRECT;
723 
724 	vq_ring_enqueue_segments(vq, dxp->indirect, 0,
725 	    sg, readable, writable);
726 
727 	vq->vq_desc_head_idx = dp->next;
728 	vq->vq_free_cnt--;
729 	if (vq->vq_free_cnt == 0)
730 		VQ_RING_ASSERT_CHAIN_TERM(vq);
731 	else
732 		VQ_RING_ASSERT_VALID_IDX(vq, vq->vq_desc_head_idx);
733 
734 	vq_ring_update_avail(vq, head_idx);
735 }
736 
737 static int
738 vq_ring_enable_interrupt(struct virtqueue *vq, uint16_t ndesc)
739 {
740 
741 	/*
742 	 * Enable interrupts, making sure we get the latest index of
743 	 * what's already been consumed.
744 	 */
745 	if (vq->vq_flags & VIRTQUEUE_FLAG_EVENT_IDX)
746 		vring_used_event(&vq->vq_ring) = vq->vq_used_cons_idx + ndesc;
747 	else
748 		vq->vq_ring.avail->flags &= ~VRING_AVAIL_F_NO_INTERRUPT;
749 
750 	mb();
751 
752 	/*
753 	 * Enough items may have already been consumed to meet our threshold
754 	 * since we last checked. Let our caller know so it processes the new
755 	 * entries.
756 	 */
757 	if (virtqueue_nused(vq) > ndesc)
758 		return (1);
759 
760 	return (0);
761 }
762 
763 static int
764 vq_ring_must_notify_host(struct virtqueue *vq)
765 {
766 	uint16_t new_idx, prev_idx, event_idx;
767 
768 	if (vq->vq_flags & VIRTQUEUE_FLAG_EVENT_IDX) {
769 		new_idx = vq->vq_ring.avail->idx;
770 		prev_idx = new_idx - vq->vq_queued_cnt;
771 		event_idx = vring_avail_event(&vq->vq_ring);
772 
773 		return (vring_need_event(event_idx, new_idx, prev_idx) != 0);
774 	}
775 
776 	return ((vq->vq_ring.used->flags & VRING_USED_F_NO_NOTIFY) == 0);
777 }
778 
779 static void
780 vq_ring_notify_host(struct virtqueue *vq)
781 {
782 
783 	VIRTIO_BUS_NOTIFY_VQ(vq->vq_dev, vq->vq_queue_index);
784 }
785 
786 static void
787 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
788 {
789 	struct vring_desc *dp;
790 	struct vq_desc_extra *dxp;
791 
792 	VQ_RING_ASSERT_VALID_IDX(vq, desc_idx);
793 	dp = &vq->vq_ring.desc[desc_idx];
794 	dxp = &vq->vq_descx[desc_idx];
795 
796 	if (vq->vq_free_cnt == 0)
797 		VQ_RING_ASSERT_CHAIN_TERM(vq);
798 
799 	vq->vq_free_cnt += dxp->ndescs;
800 	dxp->ndescs--;
801 
802 	if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
803 		while (dp->flags & VRING_DESC_F_NEXT) {
804 			VQ_RING_ASSERT_VALID_IDX(vq, dp->next);
805 			dp = &vq->vq_ring.desc[dp->next];
806 			dxp->ndescs--;
807 		}
808 	}
809 
810 	VQASSERT(vq, dxp->ndescs == 0,
811 	    "failed to free entire desc chain, remaining: %d", dxp->ndescs);
812 
813 	/*
814 	 * We must append the existing free chain, if any, to the end of
815 	 * newly freed chain. If the virtqueue was completely used, then
816 	 * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
817 	 */
818 	dp->next = vq->vq_desc_head_idx;
819 	vq->vq_desc_head_idx = desc_idx;
820 }
821