xref: /freebsd/sys/dev/virtio/virtqueue.c (revision f677a9e2672665f4eb3dd4111c07ee8f1f954262)
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/virtqueue.h>
52 #include <dev/virtio/virtio_ring.h>
53 
54 #include "virtio_bus_if.h"
55 
56 struct virtqueue {
57 	device_t		 vq_dev;
58 	char			 vq_name[VIRTQUEUE_MAX_NAME_SZ];
59 	uint16_t		 vq_queue_index;
60 	uint16_t		 vq_nentries;
61 	uint32_t		 vq_flags;
62 #define	VIRTQUEUE_FLAG_INDIRECT	 0x0001
63 #define	VIRTQUEUE_FLAG_EVENT_IDX 0x0002
64 
65 	int			 vq_alignment;
66 	int			 vq_ring_size;
67 	void			*vq_ring_mem;
68 	int			 vq_max_indirect_size;
69 	int			 vq_indirect_mem_size;
70 	virtqueue_intr_t	*vq_intrhand;
71 	void			*vq_intrhand_arg;
72 
73 	struct vring		 vq_ring;
74 	uint16_t		 vq_free_cnt;
75 	uint16_t		 vq_queued_cnt;
76 	/*
77 	 * Head of the free chain in the descriptor table. If
78 	 * there are no free descriptors, this will be set to
79 	 * VQ_RING_DESC_CHAIN_END.
80 	 */
81 	uint16_t		 vq_desc_head_idx;
82 	/*
83 	 * Last consumed descriptor in the used table,
84 	 * trails vq_ring.used->idx.
85 	 */
86 	uint16_t		 vq_used_cons_idx;
87 
88 	struct vq_desc_extra {
89 		void		  *cookie;
90 		struct vring_desc *indirect;
91 		vm_paddr_t	   indirect_paddr;
92 		uint16_t	   ndescs;
93 	} vq_descx[0];
94 };
95 
96 /*
97  * The maximum virtqueue size is 2^15. Use that value as the end of
98  * descriptor chain terminator since it will never be a valid index
99  * in the descriptor table. This is used to verify we are correctly
100  * handling vq_free_cnt.
101  */
102 #define VQ_RING_DESC_CHAIN_END 32768
103 
104 #define VQASSERT(_vq, _exp, _msg, ...)				\
105     KASSERT((_exp),("%s: %s - "_msg, __func__, (_vq)->vq_name,	\
106 	##__VA_ARGS__))
107 
108 #define VQ_RING_ASSERT_VALID_IDX(_vq, _idx)			\
109     VQASSERT((_vq), (_idx) < (_vq)->vq_nentries,		\
110 	"invalid ring index: %d, max: %d", (_idx),		\
111 	(_vq)->vq_nentries)
112 
113 #define VQ_RING_ASSERT_CHAIN_TERM(_vq)				\
114     VQASSERT((_vq), (_vq)->vq_desc_head_idx ==			\
115 	VQ_RING_DESC_CHAIN_END,	"full ring terminated "		\
116 	"incorrectly: head idx: %d", (_vq)->vq_desc_head_idx)
117 
118 static int	virtqueue_init_indirect(struct virtqueue *vq, int);
119 static void	virtqueue_free_indirect(struct virtqueue *vq);
120 static void	virtqueue_init_indirect_list(struct virtqueue *,
121 		    struct vring_desc *);
122 
123 static void	vq_ring_init(struct virtqueue *);
124 static void	vq_ring_update_avail(struct virtqueue *, uint16_t);
125 static uint16_t	vq_ring_enqueue_segments(struct virtqueue *,
126 		    struct vring_desc *, uint16_t, struct sglist *, int, int);
127 static int	vq_ring_use_indirect(struct virtqueue *, int);
128 static void	vq_ring_enqueue_indirect(struct virtqueue *, void *,
129 		    struct sglist *, int, int);
130 static int	vq_ring_enable_interrupt(struct virtqueue *, uint16_t);
131 static int	vq_ring_must_notify_host(struct virtqueue *);
132 static void	vq_ring_notify_host(struct virtqueue *);
133 static void	vq_ring_free_chain(struct virtqueue *, uint16_t);
134 
135 uint64_t
136 virtqueue_filter_features(uint64_t features)
137 {
138 	uint64_t mask;
139 
140 	mask = (1 << VIRTIO_TRANSPORT_F_START) - 1;
141 	mask |= VIRTIO_RING_F_INDIRECT_DESC;
142 	mask |= VIRTIO_RING_F_EVENT_IDX;
143 
144 	return (features & mask);
145 }
146 
147 int
148 virtqueue_alloc(device_t dev, uint16_t queue, uint16_t size, int align,
149     vm_paddr_t highaddr, struct vq_alloc_info *info, struct virtqueue **vqp)
150 {
151 	struct virtqueue *vq;
152 	int error;
153 
154 	*vqp = NULL;
155 	error = 0;
156 
157 	if (size == 0) {
158 		device_printf(dev,
159 		    "virtqueue %d (%s) does not exist (size is zero)\n",
160 		    queue, info->vqai_name);
161 		return (ENODEV);
162 	} else if (!powerof2(size)) {
163 		device_printf(dev,
164 		    "virtqueue %d (%s) size is not a power of 2: %d\n",
165 		    queue, info->vqai_name, size);
166 		return (ENXIO);
167 	} else if (info->vqai_maxindirsz > VIRTIO_MAX_INDIRECT) {
168 		device_printf(dev, "virtqueue %d (%s) requested too many "
169 		    "indirect descriptors: %d, max %d\n",
170 		    queue, info->vqai_name, info->vqai_maxindirsz,
171 		    VIRTIO_MAX_INDIRECT);
172 		return (EINVAL);
173 	}
174 
175 	vq = malloc(sizeof(struct virtqueue) +
176 	    size * sizeof(struct vq_desc_extra), M_DEVBUF, M_NOWAIT | M_ZERO);
177 	if (vq == NULL) {
178 		device_printf(dev, "cannot allocate virtqueue\n");
179 		return (ENOMEM);
180 	}
181 
182 	vq->vq_dev = dev;
183 	strlcpy(vq->vq_name, info->vqai_name, sizeof(vq->vq_name));
184 	vq->vq_queue_index = queue;
185 	vq->vq_alignment = align;
186 	vq->vq_nentries = size;
187 	vq->vq_free_cnt = size;
188 	vq->vq_intrhand = info->vqai_intr;
189 	vq->vq_intrhand_arg = info->vqai_intr_arg;
190 
191 	if (VIRTIO_BUS_WITH_FEATURE(dev, VIRTIO_RING_F_EVENT_IDX) != 0)
192 		vq->vq_flags |= VIRTQUEUE_FLAG_EVENT_IDX;
193 
194 	if (info->vqai_maxindirsz > 1) {
195 		error = virtqueue_init_indirect(vq, info->vqai_maxindirsz);
196 		if (error)
197 			goto fail;
198 	}
199 
200 	vq->vq_ring_size = round_page(vring_size(size, align));
201 	vq->vq_ring_mem = contigmalloc(vq->vq_ring_size, M_DEVBUF,
202 	    M_NOWAIT | M_ZERO, 0, highaddr, PAGE_SIZE, 0);
203 	if (vq->vq_ring_mem == NULL) {
204 		device_printf(dev,
205 		    "cannot allocate memory for virtqueue ring\n");
206 		error = ENOMEM;
207 		goto fail;
208 	}
209 
210 	vq_ring_init(vq);
211 	virtqueue_disable_intr(vq);
212 
213 	*vqp = vq;
214 
215 fail:
216 	if (error)
217 		virtqueue_free(vq);
218 
219 	return (error);
220 }
221 
222 static int
223 virtqueue_init_indirect(struct virtqueue *vq, int indirect_size)
224 {
225 	device_t dev;
226 	struct vq_desc_extra *dxp;
227 	int i, size;
228 
229 	dev = vq->vq_dev;
230 
231 	if (VIRTIO_BUS_WITH_FEATURE(dev, VIRTIO_RING_F_INDIRECT_DESC) == 0) {
232 		/*
233 		 * Indirect descriptors requested by the driver but not
234 		 * negotiated. Return zero to keep the initialization
235 		 * going: we'll run fine without.
236 		 */
237 		if (bootverbose)
238 			device_printf(dev, "virtqueue %d (%s) requested "
239 			    "indirect descriptors but not negotiated\n",
240 			    vq->vq_queue_index, vq->vq_name);
241 		return (0);
242 	}
243 
244 	size = indirect_size * sizeof(struct vring_desc);
245 	vq->vq_max_indirect_size = indirect_size;
246 	vq->vq_indirect_mem_size = size;
247 	vq->vq_flags |= VIRTQUEUE_FLAG_INDIRECT;
248 
249 	for (i = 0; i < vq->vq_nentries; i++) {
250 		dxp = &vq->vq_descx[i];
251 
252 		dxp->indirect = malloc(size, M_DEVBUF, M_NOWAIT);
253 		if (dxp->indirect == NULL) {
254 			device_printf(dev, "cannot allocate indirect list\n");
255 			return (ENOMEM);
256 		}
257 
258 		dxp->indirect_paddr = vtophys(dxp->indirect);
259 		virtqueue_init_indirect_list(vq, dxp->indirect);
260 	}
261 
262 	return (0);
263 }
264 
265 static void
266 virtqueue_free_indirect(struct virtqueue *vq)
267 {
268 	struct vq_desc_extra *dxp;
269 	int i;
270 
271 	for (i = 0; i < vq->vq_nentries; i++) {
272 		dxp = &vq->vq_descx[i];
273 
274 		if (dxp->indirect == NULL)
275 			break;
276 
277 		free(dxp->indirect, M_DEVBUF);
278 		dxp->indirect = NULL;
279 		dxp->indirect_paddr = 0;
280 	}
281 
282 	vq->vq_flags &= ~VIRTQUEUE_FLAG_INDIRECT;
283 	vq->vq_indirect_mem_size = 0;
284 }
285 
286 static void
287 virtqueue_init_indirect_list(struct virtqueue *vq,
288     struct vring_desc *indirect)
289 {
290 	int i;
291 
292 	bzero(indirect, vq->vq_indirect_mem_size);
293 
294 	for (i = 0; i < vq->vq_max_indirect_size - 1; i++)
295 		indirect[i].next = i + 1;
296 	indirect[i].next = VQ_RING_DESC_CHAIN_END;
297 }
298 
299 int
300 virtqueue_reinit(struct virtqueue *vq, uint16_t size)
301 {
302 	struct vq_desc_extra *dxp;
303 	int i;
304 
305 	if (vq->vq_nentries != size) {
306 		device_printf(vq->vq_dev,
307 		    "%s: '%s' changed size; old=%hu, new=%hu\n",
308 		    __func__, vq->vq_name, vq->vq_nentries, size);
309 		return (EINVAL);
310 	}
311 
312 	/* Warn if the virtqueue was not properly cleaned up. */
313 	if (vq->vq_free_cnt != vq->vq_nentries) {
314 		device_printf(vq->vq_dev,
315 		    "%s: warning '%s' virtqueue not empty, "
316 		    "leaking %d entries\n", __func__, vq->vq_name,
317 		    vq->vq_nentries - vq->vq_free_cnt);
318 	}
319 
320 	vq->vq_desc_head_idx = 0;
321 	vq->vq_used_cons_idx = 0;
322 	vq->vq_queued_cnt = 0;
323 	vq->vq_free_cnt = vq->vq_nentries;
324 
325 	/* To be safe, reset all our allocated memory. */
326 	bzero(vq->vq_ring_mem, vq->vq_ring_size);
327 	for (i = 0; i < vq->vq_nentries; i++) {
328 		dxp = &vq->vq_descx[i];
329 		dxp->cookie = NULL;
330 		dxp->ndescs = 0;
331 		if (vq->vq_flags & VIRTQUEUE_FLAG_INDIRECT)
332 			virtqueue_init_indirect_list(vq, dxp->indirect);
333 	}
334 
335 	vq_ring_init(vq);
336 	virtqueue_disable_intr(vq);
337 
338 	return (0);
339 }
340 
341 void
342 virtqueue_free(struct virtqueue *vq)
343 {
344 
345 	if (vq->vq_free_cnt != vq->vq_nentries) {
346 		device_printf(vq->vq_dev, "%s: freeing non-empty virtqueue, "
347 		    "leaking %d entries\n", vq->vq_name,
348 		    vq->vq_nentries - vq->vq_free_cnt);
349 	}
350 
351 	if (vq->vq_flags & VIRTQUEUE_FLAG_INDIRECT)
352 		virtqueue_free_indirect(vq);
353 
354 	if (vq->vq_ring_mem != NULL) {
355 		contigfree(vq->vq_ring_mem, vq->vq_ring_size, M_DEVBUF);
356 		vq->vq_ring_size = 0;
357 		vq->vq_ring_mem = NULL;
358 	}
359 
360 	free(vq, M_DEVBUF);
361 }
362 
363 vm_paddr_t
364 virtqueue_paddr(struct virtqueue *vq)
365 {
366 
367 	return (vtophys(vq->vq_ring_mem));
368 }
369 
370 int
371 virtqueue_size(struct virtqueue *vq)
372 {
373 
374 	return (vq->vq_nentries);
375 }
376 
377 int
378 virtqueue_empty(struct virtqueue *vq)
379 {
380 
381 	return (vq->vq_nentries == vq->vq_free_cnt);
382 }
383 
384 int
385 virtqueue_full(struct virtqueue *vq)
386 {
387 
388 	return (vq->vq_free_cnt == 0);
389 }
390 
391 void
392 virtqueue_notify(struct virtqueue *vq)
393 {
394 
395 	/* Ensure updated avail->idx is visible to host. */
396 	mb();
397 
398 	if (vq_ring_must_notify_host(vq))
399 		vq_ring_notify_host(vq);
400 	vq->vq_queued_cnt = 0;
401 }
402 
403 int
404 virtqueue_nused(struct virtqueue *vq)
405 {
406 	uint16_t used_idx, nused;
407 
408 	used_idx = vq->vq_ring.used->idx;
409 
410 	nused = (uint16_t)(used_idx - vq->vq_used_cons_idx);
411 	VQASSERT(vq, nused <= vq->vq_nentries, "used more than available");
412 
413 	return (nused);
414 }
415 
416 int
417 virtqueue_intr_filter(struct virtqueue *vq)
418 {
419 
420 	if (vq->vq_used_cons_idx == vq->vq_ring.used->idx)
421 		return (0);
422 
423 	virtqueue_disable_intr(vq);
424 
425 	return (1);
426 }
427 
428 void
429 virtqueue_intr(struct virtqueue *vq)
430 {
431 
432 	vq->vq_intrhand(vq->vq_intrhand_arg);
433 }
434 
435 int
436 virtqueue_enable_intr(struct virtqueue *vq)
437 {
438 
439 	return (vq_ring_enable_interrupt(vq, 0));
440 }
441 
442 int
443 virtqueue_postpone_intr(struct virtqueue *vq, vq_postpone_t hint)
444 {
445 	uint16_t ndesc, avail_idx;
446 
447 	avail_idx = vq->vq_ring.avail->idx;
448 	ndesc = (uint16_t)(avail_idx - vq->vq_used_cons_idx);
449 
450 	switch (hint) {
451 	case VQ_POSTPONE_SHORT:
452 		ndesc = ndesc / 4;
453 		break;
454 	case VQ_POSTPONE_LONG:
455 		ndesc = (ndesc * 3) / 4;
456 		break;
457 	case VQ_POSTPONE_EMPTIED:
458 		break;
459 	}
460 
461 	return (vq_ring_enable_interrupt(vq, ndesc));
462 }
463 
464 /*
465  * Note this is only considered a hint to the host.
466  */
467 void
468 virtqueue_disable_intr(struct virtqueue *vq)
469 {
470 
471 	if (vq->vq_flags & VIRTQUEUE_FLAG_EVENT_IDX) {
472 		vring_used_event(&vq->vq_ring) = vq->vq_used_cons_idx -
473 		    vq->vq_nentries - 1;
474 	} else
475 		vq->vq_ring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
476 }
477 
478 int
479 virtqueue_enqueue(struct virtqueue *vq, void *cookie, struct sglist *sg,
480     int readable, int writable)
481 {
482 	struct vq_desc_extra *dxp;
483 	int needed;
484 	uint16_t head_idx, idx;
485 
486 	needed = readable + writable;
487 
488 	VQASSERT(vq, cookie != NULL, "enqueuing with no cookie");
489 	VQASSERT(vq, needed == sg->sg_nseg,
490 	    "segment count mismatch, %d, %d", needed, sg->sg_nseg);
491 	VQASSERT(vq,
492 	    needed <= vq->vq_nentries || needed <= vq->vq_max_indirect_size,
493 	    "too many segments to enqueue: %d, %d/%d", needed,
494 	    vq->vq_nentries, vq->vq_max_indirect_size);
495 
496 	if (needed < 1)
497 		return (EINVAL);
498 	if (vq->vq_free_cnt == 0)
499 		return (ENOSPC);
500 
501 	if (vq_ring_use_indirect(vq, needed)) {
502 		vq_ring_enqueue_indirect(vq, cookie, sg, readable, writable);
503 		return (0);
504 	} else if (vq->vq_free_cnt < needed)
505 		return (EMSGSIZE);
506 
507 	head_idx = vq->vq_desc_head_idx;
508 	VQ_RING_ASSERT_VALID_IDX(vq, head_idx);
509 	dxp = &vq->vq_descx[head_idx];
510 
511 	VQASSERT(vq, dxp->cookie == NULL,
512 	    "cookie already exists for index %d", head_idx);
513 	dxp->cookie = cookie;
514 	dxp->ndescs = needed;
515 
516 	idx = vq_ring_enqueue_segments(vq, vq->vq_ring.desc, head_idx,
517 	    sg, readable, writable);
518 
519 	vq->vq_desc_head_idx = idx;
520 	vq->vq_free_cnt -= needed;
521 	if (vq->vq_free_cnt == 0)
522 		VQ_RING_ASSERT_CHAIN_TERM(vq);
523 	else
524 		VQ_RING_ASSERT_VALID_IDX(vq, idx);
525 
526 	vq_ring_update_avail(vq, head_idx);
527 
528 	return (0);
529 }
530 
531 void *
532 virtqueue_dequeue(struct virtqueue *vq, uint32_t *len)
533 {
534 	struct vring_used_elem *uep;
535 	void *cookie;
536 	uint16_t used_idx, desc_idx;
537 
538 	if (vq->vq_used_cons_idx == vq->vq_ring.used->idx)
539 		return (NULL);
540 
541 	used_idx = vq->vq_used_cons_idx++ & (vq->vq_nentries - 1);
542 	uep = &vq->vq_ring.used->ring[used_idx];
543 
544 	rmb();
545 	desc_idx = (uint16_t) uep->id;
546 	if (len != NULL)
547 		*len = uep->len;
548 
549 	vq_ring_free_chain(vq, desc_idx);
550 
551 	cookie = vq->vq_descx[desc_idx].cookie;
552 	VQASSERT(vq, cookie != NULL, "no cookie for index %d", desc_idx);
553 	vq->vq_descx[desc_idx].cookie = NULL;
554 
555 	return (cookie);
556 }
557 
558 void *
559 virtqueue_poll(struct virtqueue *vq, uint32_t *len)
560 {
561 	void *cookie;
562 
563 	while ((cookie = virtqueue_dequeue(vq, len)) == NULL)
564 		cpu_spinwait();
565 
566 	return (cookie);
567 }
568 
569 void *
570 virtqueue_drain(struct virtqueue *vq, int *last)
571 {
572 	void *cookie;
573 	int idx;
574 
575 	cookie = NULL;
576 	idx = *last;
577 
578 	while (idx < vq->vq_nentries && cookie == NULL) {
579 		if ((cookie = vq->vq_descx[idx].cookie) != NULL) {
580 			vq->vq_descx[idx].cookie = NULL;
581 			/* Free chain to keep free count consistent. */
582 			vq_ring_free_chain(vq, idx);
583 		}
584 		idx++;
585 	}
586 
587 	*last = idx;
588 
589 	return (cookie);
590 }
591 
592 void
593 virtqueue_dump(struct virtqueue *vq)
594 {
595 
596 	if (vq == NULL)
597 		return;
598 
599 	printf("VQ: %s - size=%d; free=%d; used=%d; queued=%d; "
600 	    "desc_head_idx=%d; avail.idx=%d; used_cons_idx=%d; "
601 	    "used.idx=%d; avail.flags=0x%x; used.flags=0x%x\n",
602 	    vq->vq_name, vq->vq_nentries, vq->vq_free_cnt,
603 	    virtqueue_nused(vq), vq->vq_queued_cnt, vq->vq_desc_head_idx,
604 	    vq->vq_ring.avail->idx, vq->vq_used_cons_idx,
605 	    vq->vq_ring.used->idx, vq->vq_ring.avail->flags,
606 	    vq->vq_ring.used->flags);
607 }
608 
609 static void
610 vq_ring_init(struct virtqueue *vq)
611 {
612 	struct vring *vr;
613 	char *ring_mem;
614 	int i, size;
615 
616 	ring_mem = vq->vq_ring_mem;
617 	size = vq->vq_nentries;
618 	vr = &vq->vq_ring;
619 
620 	vring_init(vr, size, ring_mem, vq->vq_alignment);
621 
622 	for (i = 0; i < size - 1; i++)
623 		vr->desc[i].next = i + 1;
624 	vr->desc[i].next = VQ_RING_DESC_CHAIN_END;
625 }
626 
627 static void
628 vq_ring_update_avail(struct virtqueue *vq, uint16_t desc_idx)
629 {
630 	uint16_t avail_idx;
631 
632 	/*
633 	 * Place the head of the descriptor chain into the next slot and make
634 	 * it usable to the host. The chain is made available now rather than
635 	 * deferring to virtqueue_notify() in the hopes that if the host is
636 	 * currently running on another CPU, we can keep it processing the new
637 	 * descriptor.
638 	 */
639 	avail_idx = vq->vq_ring.avail->idx & (vq->vq_nentries - 1);
640 	vq->vq_ring.avail->ring[avail_idx] = desc_idx;
641 
642 	wmb();
643 	vq->vq_ring.avail->idx++;
644 
645 	/* Keep pending count until virtqueue_notify(). */
646 	vq->vq_queued_cnt++;
647 }
648 
649 static uint16_t
650 vq_ring_enqueue_segments(struct virtqueue *vq, struct vring_desc *desc,
651     uint16_t head_idx, struct sglist *sg, int readable, int writable)
652 {
653 	struct sglist_seg *seg;
654 	struct vring_desc *dp;
655 	int i, needed;
656 	uint16_t idx;
657 
658 	needed = readable + writable;
659 
660 	for (i = 0, idx = head_idx, seg = sg->sg_segs;
661 	     i < needed;
662 	     i++, idx = dp->next, seg++) {
663 		VQASSERT(vq, idx != VQ_RING_DESC_CHAIN_END,
664 		    "premature end of free desc chain");
665 
666 		dp = &desc[idx];
667 		dp->addr = seg->ss_paddr;
668 		dp->len = seg->ss_len;
669 		dp->flags = 0;
670 
671 		if (i < needed - 1)
672 			dp->flags |= VRING_DESC_F_NEXT;
673 		if (i >= readable)
674 			dp->flags |= VRING_DESC_F_WRITE;
675 	}
676 
677 	return (idx);
678 }
679 
680 static int
681 vq_ring_use_indirect(struct virtqueue *vq, int needed)
682 {
683 
684 	if ((vq->vq_flags & VIRTQUEUE_FLAG_INDIRECT) == 0)
685 		return (0);
686 
687 	if (vq->vq_max_indirect_size < needed)
688 		return (0);
689 
690 	if (needed < 2)
691 		return (0);
692 
693 	return (1);
694 }
695 
696 static void
697 vq_ring_enqueue_indirect(struct virtqueue *vq, void *cookie,
698     struct sglist *sg, int readable, int writable)
699 {
700 	struct vring_desc *dp;
701 	struct vq_desc_extra *dxp;
702 	int needed;
703 	uint16_t head_idx;
704 
705 	needed = readable + writable;
706 	VQASSERT(vq, needed <= vq->vq_max_indirect_size,
707 	    "enqueuing too many indirect descriptors");
708 
709 	head_idx = vq->vq_desc_head_idx;
710 	VQ_RING_ASSERT_VALID_IDX(vq, head_idx);
711 	dp = &vq->vq_ring.desc[head_idx];
712 	dxp = &vq->vq_descx[head_idx];
713 
714 	VQASSERT(vq, dxp->cookie == NULL,
715 	    "cookie already exists for index %d", head_idx);
716 	dxp->cookie = cookie;
717 	dxp->ndescs = 1;
718 
719 	dp->addr = dxp->indirect_paddr;
720 	dp->len = needed * sizeof(struct vring_desc);
721 	dp->flags = VRING_DESC_F_INDIRECT;
722 
723 	vq_ring_enqueue_segments(vq, dxp->indirect, 0,
724 	    sg, readable, writable);
725 
726 	vq->vq_desc_head_idx = dp->next;
727 	vq->vq_free_cnt--;
728 	if (vq->vq_free_cnt == 0)
729 		VQ_RING_ASSERT_CHAIN_TERM(vq);
730 	else
731 		VQ_RING_ASSERT_VALID_IDX(vq, vq->vq_desc_head_idx);
732 
733 	vq_ring_update_avail(vq, head_idx);
734 }
735 
736 static int
737 vq_ring_enable_interrupt(struct virtqueue *vq, uint16_t ndesc)
738 {
739 
740 	/*
741 	 * Enable interrupts, making sure we get the latest index of
742 	 * what's already been consumed.
743 	 */
744 	if (vq->vq_flags & VIRTQUEUE_FLAG_EVENT_IDX)
745 		vring_used_event(&vq->vq_ring) = vq->vq_used_cons_idx + ndesc;
746 	else
747 		vq->vq_ring.avail->flags &= ~VRING_AVAIL_F_NO_INTERRUPT;
748 
749 	mb();
750 
751 	/*
752 	 * Enough items may have already been consumed to meet our threshold
753 	 * since we last checked. Let our caller know so it processes the new
754 	 * entries.
755 	 */
756 	if (virtqueue_nused(vq) > ndesc)
757 		return (1);
758 
759 	return (0);
760 }
761 
762 static int
763 vq_ring_must_notify_host(struct virtqueue *vq)
764 {
765 	uint16_t new_idx, prev_idx, event_idx;
766 
767 	if (vq->vq_flags & VIRTQUEUE_FLAG_EVENT_IDX) {
768 		new_idx = vq->vq_ring.avail->idx;
769 		prev_idx = new_idx - vq->vq_queued_cnt;
770 		event_idx = vring_avail_event(&vq->vq_ring);
771 
772 		return (vring_need_event(event_idx, new_idx, prev_idx) != 0);
773 	}
774 
775 	return ((vq->vq_ring.used->flags & VRING_USED_F_NO_NOTIFY) == 0);
776 }
777 
778 static void
779 vq_ring_notify_host(struct virtqueue *vq)
780 {
781 
782 	VIRTIO_BUS_NOTIFY_VQ(vq->vq_dev, vq->vq_queue_index);
783 }
784 
785 static void
786 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
787 {
788 	struct vring_desc *dp;
789 	struct vq_desc_extra *dxp;
790 
791 	VQ_RING_ASSERT_VALID_IDX(vq, desc_idx);
792 	dp = &vq->vq_ring.desc[desc_idx];
793 	dxp = &vq->vq_descx[desc_idx];
794 
795 	if (vq->vq_free_cnt == 0)
796 		VQ_RING_ASSERT_CHAIN_TERM(vq);
797 
798 	vq->vq_free_cnt += dxp->ndescs;
799 	dxp->ndescs--;
800 
801 	if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
802 		while (dp->flags & VRING_DESC_F_NEXT) {
803 			VQ_RING_ASSERT_VALID_IDX(vq, dp->next);
804 			dp = &vq->vq_ring.desc[dp->next];
805 			dxp->ndescs--;
806 		}
807 	}
808 
809 	VQASSERT(vq, dxp->ndescs == 0,
810 	    "failed to free entire desc chain, remaining: %d", dxp->ndescs);
811 
812 	/*
813 	 * We must append the existing free chain, if any, to the end of
814 	 * newly freed chain. If the virtqueue was completely used, then
815 	 * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
816 	 */
817 	dp->next = vq->vq_desc_head_idx;
818 	vq->vq_desc_head_idx = desc_idx;
819 }
820