xref: /linux/tools/virtio/vringh_test.c (revision 5ac79730324c6f37106ce397586020ffe6e8e234)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Simple test of virtio code, entirely in userpsace. */
3 #define _GNU_SOURCE
4 #include <sched.h>
5 #include <err.h>
6 #include <linux/kernel.h>
7 #include <linux/err.h>
8 #include <linux/virtio.h>
9 #include <linux/vringh.h>
10 #include <linux/virtio_ring.h>
11 #include <linux/virtio_config.h>
12 #include <linux/uaccess.h>
13 #include <sys/types.h>
14 #include <sys/stat.h>
15 #include <sys/mman.h>
16 #include <sys/wait.h>
17 #include <fcntl.h>
18 
19 #define USER_MEM (1024*1024)
20 void *__user_addr_min, *__user_addr_max;
21 void *__kmalloc_fake, *__kfree_ignore_start, *__kfree_ignore_end;
22 static u64 user_addr_offset;
23 
24 #define RINGSIZE 256
25 #define ALIGN 4096
26 
27 static bool never_notify_host(struct virtqueue *vq)
28 {
29 	abort();
30 }
31 
32 static void never_callback_guest(struct virtqueue *vq)
33 {
34 	abort();
35 }
36 
37 static bool getrange_iov(struct vringh *vrh, u64 addr, struct vringh_range *r)
38 {
39 	if (addr < (u64)(unsigned long)__user_addr_min - user_addr_offset)
40 		return false;
41 	if (addr >= (u64)(unsigned long)__user_addr_max - user_addr_offset)
42 		return false;
43 
44 	r->start = (u64)(unsigned long)__user_addr_min - user_addr_offset;
45 	r->end_incl = (u64)(unsigned long)__user_addr_max - 1 - user_addr_offset;
46 	r->offset = user_addr_offset;
47 	return true;
48 }
49 
50 /* We return single byte ranges. */
51 static bool getrange_slow(struct vringh *vrh, u64 addr, struct vringh_range *r)
52 {
53 	if (addr < (u64)(unsigned long)__user_addr_min - user_addr_offset)
54 		return false;
55 	if (addr >= (u64)(unsigned long)__user_addr_max - user_addr_offset)
56 		return false;
57 
58 	r->start = addr;
59 	r->end_incl = r->start;
60 	r->offset = user_addr_offset;
61 	return true;
62 }
63 
64 struct guest_virtio_device {
65 	struct virtio_device vdev;
66 	int to_host_fd;
67 	unsigned long notifies;
68 };
69 
70 static bool parallel_notify_host(struct virtqueue *vq)
71 {
72 	int rc;
73 	struct guest_virtio_device *gvdev;
74 
75 	gvdev = container_of(vq->vdev, struct guest_virtio_device, vdev);
76 	rc = write(gvdev->to_host_fd, "", 1);
77 	if (rc < 0)
78 		return false;
79 	gvdev->notifies++;
80 	return true;
81 }
82 
83 static bool no_notify_host(struct virtqueue *vq)
84 {
85 	return true;
86 }
87 
88 #define NUM_XFERS (10000000)
89 
90 /* We aim for two "distant" cpus. */
91 static void find_cpus(unsigned int *first, unsigned int *last)
92 {
93 	unsigned int i;
94 
95 	*first = -1U;
96 	*last = 0;
97 	for (i = 0; i < 4096; i++) {
98 		cpu_set_t set;
99 		CPU_ZERO(&set);
100 		CPU_SET(i, &set);
101 		if (sched_setaffinity(getpid(), sizeof(set), &set) == 0) {
102 			if (i < *first)
103 				*first = i;
104 			if (i > *last)
105 				*last = i;
106 		}
107 	}
108 }
109 
110 /* Opencoded version for fast mode */
111 static inline int vringh_get_head(struct vringh *vrh, u16 *head)
112 {
113 	u16 avail_idx, i;
114 	int err;
115 
116 	err = get_user(avail_idx, &vrh->vring.avail->idx);
117 	if (err)
118 		return err;
119 
120 	if (vrh->last_avail_idx == avail_idx)
121 		return 0;
122 
123 	/* Only get avail ring entries after they have been exposed by guest. */
124 	virtio_rmb(vrh->weak_barriers);
125 
126 	i = vrh->last_avail_idx & (vrh->vring.num - 1);
127 
128 	err = get_user(*head, &vrh->vring.avail->ring[i]);
129 	if (err)
130 		return err;
131 
132 	vrh->last_avail_idx++;
133 	return 1;
134 }
135 
136 static int parallel_test(u64 features,
137 			 bool (*getrange)(struct vringh *vrh,
138 					  u64 addr, struct vringh_range *r),
139 			 bool fast_vringh)
140 {
141 	void *host_map, *guest_map;
142 	int pipe_ret, fd, mapsize, to_guest[2], to_host[2];
143 	unsigned long xfers = 0, notifies = 0, receives = 0;
144 	unsigned int first_cpu, last_cpu;
145 	cpu_set_t cpu_set;
146 	char buf[128];
147 
148 	/* Create real file to mmap. */
149 	fd = open("/tmp/vringh_test-file", O_RDWR|O_CREAT|O_TRUNC, 0600);
150 	if (fd < 0)
151 		err(1, "Opening /tmp/vringh_test-file");
152 
153 	/* Extra room at the end for some data, and indirects */
154 	mapsize = vring_size(RINGSIZE, ALIGN)
155 		+ RINGSIZE * 2 * sizeof(int)
156 		+ RINGSIZE * 6 * sizeof(struct vring_desc);
157 	mapsize = (mapsize + getpagesize() - 1) & ~(getpagesize() - 1);
158 	ftruncate(fd, mapsize);
159 
160 	/* Parent and child use separate addresses, to check our mapping logic! */
161 	host_map = mmap(NULL, mapsize, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
162 	guest_map = mmap(NULL, mapsize, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
163 
164 	pipe_ret = pipe(to_guest);
165 	assert(!pipe_ret);
166 
167 	pipe_ret = pipe(to_host);
168 	assert(!pipe_ret);
169 
170 	CPU_ZERO(&cpu_set);
171 	find_cpus(&first_cpu, &last_cpu);
172 	printf("Using CPUS %u and %u\n", first_cpu, last_cpu);
173 	fflush(stdout);
174 
175 	if (fork() != 0) {
176 		struct vringh vrh;
177 		int status, err, rlen = 0;
178 		char rbuf[5];
179 
180 		/* We are the host: never access guest addresses! */
181 		munmap(guest_map, mapsize);
182 
183 		__user_addr_min = host_map;
184 		__user_addr_max = __user_addr_min + mapsize;
185 		user_addr_offset = host_map - guest_map;
186 		assert(user_addr_offset);
187 
188 		close(to_guest[0]);
189 		close(to_host[1]);
190 
191 		vring_init(&vrh.vring, RINGSIZE, host_map, ALIGN);
192 		vringh_init_user(&vrh, features, RINGSIZE, true,
193 				 vrh.vring.desc, vrh.vring.avail, vrh.vring.used);
194 		CPU_SET(first_cpu, &cpu_set);
195 		if (sched_setaffinity(getpid(), sizeof(cpu_set), &cpu_set))
196 			errx(1, "Could not set affinity to cpu %u", first_cpu);
197 
198 		while (xfers < NUM_XFERS) {
199 			struct iovec host_riov[2], host_wiov[2];
200 			struct vringh_iov riov, wiov;
201 			u16 head, written;
202 
203 			if (fast_vringh) {
204 				for (;;) {
205 					err = vringh_get_head(&vrh, &head);
206 					if (err != 0)
207 						break;
208 					err = vringh_need_notify_user(&vrh);
209 					if (err < 0)
210 						errx(1, "vringh_need_notify_user: %i",
211 						     err);
212 					if (err) {
213 						write(to_guest[1], "", 1);
214 						notifies++;
215 					}
216 				}
217 				if (err != 1)
218 					errx(1, "vringh_get_head");
219 				written = 0;
220 				goto complete;
221 			} else {
222 				vringh_iov_init(&riov,
223 						host_riov,
224 						ARRAY_SIZE(host_riov));
225 				vringh_iov_init(&wiov,
226 						host_wiov,
227 						ARRAY_SIZE(host_wiov));
228 
229 				err = vringh_getdesc_user(&vrh, &riov, &wiov,
230 							  getrange, &head);
231 			}
232 			if (err == 0) {
233 				err = vringh_need_notify_user(&vrh);
234 				if (err < 0)
235 					errx(1, "vringh_need_notify_user: %i",
236 					     err);
237 				if (err) {
238 					write(to_guest[1], "", 1);
239 					notifies++;
240 				}
241 
242 				if (!vringh_notify_enable_user(&vrh))
243 					continue;
244 
245 				/* Swallow all notifies at once. */
246 				if (read(to_host[0], buf, sizeof(buf)) < 1)
247 					break;
248 
249 				vringh_notify_disable_user(&vrh);
250 				receives++;
251 				continue;
252 			}
253 			if (err != 1)
254 				errx(1, "vringh_getdesc_user: %i", err);
255 
256 			/* We simply copy bytes. */
257 			if (riov.used) {
258 				rlen = vringh_iov_pull_user(&riov, rbuf,
259 							    sizeof(rbuf));
260 				if (rlen != 4)
261 					errx(1, "vringh_iov_pull_user: %i",
262 					     rlen);
263 				assert(riov.i == riov.used);
264 				written = 0;
265 			} else {
266 				err = vringh_iov_push_user(&wiov, rbuf, rlen);
267 				if (err != rlen)
268 					errx(1, "vringh_iov_push_user: %i",
269 					     err);
270 				assert(wiov.i == wiov.used);
271 				written = err;
272 			}
273 		complete:
274 			xfers++;
275 
276 			err = vringh_complete_user(&vrh, head, written);
277 			if (err != 0)
278 				errx(1, "vringh_complete_user: %i", err);
279 		}
280 
281 		err = vringh_need_notify_user(&vrh);
282 		if (err < 0)
283 			errx(1, "vringh_need_notify_user: %i", err);
284 		if (err) {
285 			write(to_guest[1], "", 1);
286 			notifies++;
287 		}
288 		wait(&status);
289 		if (!WIFEXITED(status))
290 			errx(1, "Child died with signal %i?", WTERMSIG(status));
291 		if (WEXITSTATUS(status) != 0)
292 			errx(1, "Child exited %i?", WEXITSTATUS(status));
293 		printf("Host: notified %lu, pinged %lu\n", notifies, receives);
294 		return 0;
295 	} else {
296 		struct guest_virtio_device gvdev;
297 		struct virtqueue *vq;
298 		unsigned int *data;
299 		struct vring_desc *indirects;
300 		unsigned int finished = 0;
301 
302 		/* We pass sg[]s pointing into here, but we need RINGSIZE+1 */
303 		data = guest_map + vring_size(RINGSIZE, ALIGN);
304 		indirects = (void *)data + (RINGSIZE + 1) * 2 * sizeof(int);
305 
306 		/* We are the guest. */
307 		munmap(host_map, mapsize);
308 
309 		close(to_guest[1]);
310 		close(to_host[0]);
311 
312 		gvdev.vdev.features = features;
313 		INIT_LIST_HEAD(&gvdev.vdev.vqs);
314 		spin_lock_init(&gvdev.vdev.vqs_list_lock);
315 		gvdev.to_host_fd = to_host[1];
316 		gvdev.notifies = 0;
317 
318 		CPU_SET(first_cpu, &cpu_set);
319 		if (sched_setaffinity(getpid(), sizeof(cpu_set), &cpu_set))
320 			err(1, "Could not set affinity to cpu %u", first_cpu);
321 
322 		vq = vring_new_virtqueue(0, RINGSIZE, ALIGN, &gvdev.vdev, true,
323 					 false, guest_map,
324 					 fast_vringh ? no_notify_host
325 					 : parallel_notify_host,
326 					 never_callback_guest, "guest vq");
327 
328 		/* Don't kfree indirects. */
329 		__kfree_ignore_start = indirects;
330 		__kfree_ignore_end = indirects + RINGSIZE * 6;
331 
332 		while (xfers < NUM_XFERS) {
333 			struct scatterlist sg[4];
334 			unsigned int num_sg, len;
335 			int *dbuf, err;
336 			bool output = !(xfers % 2);
337 
338 			/* Consume bufs. */
339 			while ((dbuf = virtqueue_get_buf(vq, &len)) != NULL) {
340 				if (len == 4)
341 					assert(*dbuf == finished - 1);
342 				else if (!fast_vringh)
343 					assert(*dbuf == finished);
344 				finished++;
345 			}
346 
347 			/* Produce a buffer. */
348 			dbuf = data + (xfers % (RINGSIZE + 1));
349 
350 			if (output)
351 				*dbuf = xfers;
352 			else
353 				*dbuf = -1;
354 
355 			switch ((xfers / sizeof(*dbuf)) % 4) {
356 			case 0:
357 				/* Nasty three-element sg list. */
358 				sg_init_table(sg, num_sg = 3);
359 				sg_set_buf(&sg[0], (void *)dbuf, 1);
360 				sg_set_buf(&sg[1], (void *)dbuf + 1, 2);
361 				sg_set_buf(&sg[2], (void *)dbuf + 3, 1);
362 				break;
363 			case 1:
364 				sg_init_table(sg, num_sg = 2);
365 				sg_set_buf(&sg[0], (void *)dbuf, 1);
366 				sg_set_buf(&sg[1], (void *)dbuf + 1, 3);
367 				break;
368 			case 2:
369 				sg_init_table(sg, num_sg = 1);
370 				sg_set_buf(&sg[0], (void *)dbuf, 4);
371 				break;
372 			case 3:
373 				sg_init_table(sg, num_sg = 4);
374 				sg_set_buf(&sg[0], (void *)dbuf, 1);
375 				sg_set_buf(&sg[1], (void *)dbuf + 1, 1);
376 				sg_set_buf(&sg[2], (void *)dbuf + 2, 1);
377 				sg_set_buf(&sg[3], (void *)dbuf + 3, 1);
378 				break;
379 			}
380 
381 			/* May allocate an indirect, so force it to allocate
382 			 * user addr */
383 			__kmalloc_fake = indirects + (xfers % RINGSIZE) * 4;
384 			if (output)
385 				err = virtqueue_add_outbuf(vq, sg, num_sg, dbuf,
386 							   GFP_KERNEL);
387 			else
388 				err = virtqueue_add_inbuf(vq, sg, num_sg,
389 							  dbuf, GFP_KERNEL);
390 
391 			if (err == -ENOSPC) {
392 				if (!virtqueue_enable_cb_delayed(vq))
393 					continue;
394 				/* Swallow all notifies at once. */
395 				if (read(to_guest[0], buf, sizeof(buf)) < 1)
396 					break;
397 
398 				receives++;
399 				virtqueue_disable_cb(vq);
400 				continue;
401 			}
402 
403 			if (err)
404 				errx(1, "virtqueue_add_in/outbuf: %i", err);
405 
406 			xfers++;
407 			virtqueue_kick(vq);
408 		}
409 
410 		/* Any extra? */
411 		while (finished != xfers) {
412 			int *dbuf;
413 			unsigned int len;
414 
415 			/* Consume bufs. */
416 			dbuf = virtqueue_get_buf(vq, &len);
417 			if (dbuf) {
418 				if (len == 4)
419 					assert(*dbuf == finished - 1);
420 				else
421 					assert(len == 0);
422 				finished++;
423 				continue;
424 			}
425 
426 			if (!virtqueue_enable_cb_delayed(vq))
427 				continue;
428 			if (read(to_guest[0], buf, sizeof(buf)) < 1)
429 				break;
430 
431 			receives++;
432 			virtqueue_disable_cb(vq);
433 		}
434 
435 		printf("Guest: notified %lu, pinged %lu\n",
436 		       gvdev.notifies, receives);
437 		vring_del_virtqueue(vq);
438 		return 0;
439 	}
440 }
441 
442 int main(int argc, char *argv[])
443 {
444 	struct virtio_device vdev;
445 	struct virtqueue *vq;
446 	struct vringh vrh;
447 	struct scatterlist guest_sg[RINGSIZE], *sgs[2];
448 	struct iovec host_riov[2], host_wiov[2];
449 	struct vringh_iov riov, wiov;
450 	struct vring_used_elem used[RINGSIZE];
451 	char buf[28];
452 	u16 head;
453 	int err;
454 	unsigned i;
455 	void *ret;
456 	bool (*getrange)(struct vringh *vrh, u64 addr, struct vringh_range *r);
457 	bool fast_vringh = false, parallel = false;
458 
459 	getrange = getrange_iov;
460 	vdev.features = 0;
461 	INIT_LIST_HEAD(&vdev.vqs);
462 	spin_lock_init(&vdev.vqs_list_lock);
463 
464 	while (argv[1]) {
465 		if (strcmp(argv[1], "--indirect") == 0)
466 			__virtio_set_bit(&vdev, VIRTIO_RING_F_INDIRECT_DESC);
467 		else if (strcmp(argv[1], "--eventidx") == 0)
468 			__virtio_set_bit(&vdev, VIRTIO_RING_F_EVENT_IDX);
469 		else if (strcmp(argv[1], "--virtio-1") == 0)
470 			__virtio_set_bit(&vdev, VIRTIO_F_VERSION_1);
471 		else if (strcmp(argv[1], "--slow-range") == 0)
472 			getrange = getrange_slow;
473 		else if (strcmp(argv[1], "--fast-vringh") == 0)
474 			fast_vringh = true;
475 		else if (strcmp(argv[1], "--parallel") == 0)
476 			parallel = true;
477 		else
478 			errx(1, "Unknown arg %s", argv[1]);
479 		argv++;
480 	}
481 
482 	if (parallel)
483 		return parallel_test(vdev.features, getrange, fast_vringh);
484 
485 	if (posix_memalign(&__user_addr_min, PAGE_SIZE, USER_MEM) != 0)
486 		abort();
487 	__user_addr_max = __user_addr_min + USER_MEM;
488 	memset(__user_addr_min, 0, vring_size(RINGSIZE, ALIGN));
489 
490 	/* Set up guest side. */
491 	vq = vring_new_virtqueue(0, RINGSIZE, ALIGN, &vdev, true, false,
492 				 __user_addr_min,
493 				 never_notify_host, never_callback_guest,
494 				 "guest vq");
495 
496 	/* Set up host side. */
497 	vring_init(&vrh.vring, RINGSIZE, __user_addr_min, ALIGN);
498 	vringh_init_user(&vrh, vdev.features, RINGSIZE, true,
499 			 vrh.vring.desc, vrh.vring.avail, vrh.vring.used);
500 
501 	/* No descriptor to get yet... */
502 	err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
503 	if (err != 0)
504 		errx(1, "vringh_getdesc_user: %i", err);
505 
506 	/* Guest puts in a descriptor. */
507 	memcpy(__user_addr_max - 1, "a", 1);
508 	sg_init_table(guest_sg, 1);
509 	sg_set_buf(&guest_sg[0], __user_addr_max - 1, 1);
510 	sg_init_table(guest_sg+1, 1);
511 	sg_set_buf(&guest_sg[1], __user_addr_max - 3, 2);
512 	sgs[0] = &guest_sg[0];
513 	sgs[1] = &guest_sg[1];
514 
515 	/* May allocate an indirect, so force it to allocate user addr */
516 	__kmalloc_fake = __user_addr_min + vring_size(RINGSIZE, ALIGN);
517 	err = virtqueue_add_sgs(vq, sgs, 1, 1, &err, GFP_KERNEL);
518 	if (err)
519 		errx(1, "virtqueue_add_sgs: %i", err);
520 	__kmalloc_fake = NULL;
521 
522 	/* Host retreives it. */
523 	vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
524 	vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
525 
526 	err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
527 	if (err != 1)
528 		errx(1, "vringh_getdesc_user: %i", err);
529 
530 	assert(riov.used == 1);
531 	assert(riov.iov[0].iov_base == __user_addr_max - 1);
532 	assert(riov.iov[0].iov_len == 1);
533 	if (getrange != getrange_slow) {
534 		assert(wiov.used == 1);
535 		assert(wiov.iov[0].iov_base == __user_addr_max - 3);
536 		assert(wiov.iov[0].iov_len == 2);
537 	} else {
538 		assert(wiov.used == 2);
539 		assert(wiov.iov[0].iov_base == __user_addr_max - 3);
540 		assert(wiov.iov[0].iov_len == 1);
541 		assert(wiov.iov[1].iov_base == __user_addr_max - 2);
542 		assert(wiov.iov[1].iov_len == 1);
543 	}
544 
545 	err = vringh_iov_pull_user(&riov, buf, 5);
546 	if (err != 1)
547 		errx(1, "vringh_iov_pull_user: %i", err);
548 	assert(buf[0] == 'a');
549 	assert(riov.i == 1);
550 	assert(vringh_iov_pull_user(&riov, buf, 5) == 0);
551 
552 	memcpy(buf, "bcdef", 5);
553 	err = vringh_iov_push_user(&wiov, buf, 5);
554 	if (err != 2)
555 		errx(1, "vringh_iov_push_user: %i", err);
556 	assert(memcmp(__user_addr_max - 3, "bc", 2) == 0);
557 	assert(wiov.i == wiov.used);
558 	assert(vringh_iov_push_user(&wiov, buf, 5) == 0);
559 
560 	/* Host is done. */
561 	err = vringh_complete_user(&vrh, head, err);
562 	if (err != 0)
563 		errx(1, "vringh_complete_user: %i", err);
564 
565 	/* Guest should see used token now. */
566 	__kfree_ignore_start = __user_addr_min + vring_size(RINGSIZE, ALIGN);
567 	__kfree_ignore_end = __kfree_ignore_start + 1;
568 	ret = virtqueue_get_buf(vq, &i);
569 	if (ret != &err)
570 		errx(1, "virtqueue_get_buf: %p", ret);
571 	assert(i == 2);
572 
573 	/* Guest puts in a huge descriptor. */
574 	sg_init_table(guest_sg, RINGSIZE);
575 	for (i = 0; i < RINGSIZE; i++) {
576 		sg_set_buf(&guest_sg[i],
577 			   __user_addr_max - USER_MEM/4, USER_MEM/4);
578 	}
579 
580 	/* Fill contents with recognisable garbage. */
581 	for (i = 0; i < USER_MEM/4; i++)
582 		((char *)__user_addr_max - USER_MEM/4)[i] = i;
583 
584 	/* This will allocate an indirect, so force it to allocate user addr */
585 	__kmalloc_fake = __user_addr_min + vring_size(RINGSIZE, ALIGN);
586 	err = virtqueue_add_outbuf(vq, guest_sg, RINGSIZE, &err, GFP_KERNEL);
587 	if (err)
588 		errx(1, "virtqueue_add_outbuf (large): %i", err);
589 	__kmalloc_fake = NULL;
590 
591 	/* Host picks it up (allocates new iov). */
592 	vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
593 	vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
594 
595 	err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
596 	if (err != 1)
597 		errx(1, "vringh_getdesc_user: %i", err);
598 
599 	assert(riov.max_num & VRINGH_IOV_ALLOCATED);
600 	assert(riov.iov != host_riov);
601 	if (getrange != getrange_slow)
602 		assert(riov.used == RINGSIZE);
603 	else
604 		assert(riov.used == RINGSIZE * USER_MEM/4);
605 
606 	assert(!(wiov.max_num & VRINGH_IOV_ALLOCATED));
607 	assert(wiov.used == 0);
608 
609 	/* Pull data back out (in odd chunks), should be as expected. */
610 	for (i = 0; i < RINGSIZE * USER_MEM/4; i += 3) {
611 		err = vringh_iov_pull_user(&riov, buf, 3);
612 		if (err != 3 && i + err != RINGSIZE * USER_MEM/4)
613 			errx(1, "vringh_iov_pull_user large: %i", err);
614 		assert(buf[0] == (char)i);
615 		assert(err < 2 || buf[1] == (char)(i + 1));
616 		assert(err < 3 || buf[2] == (char)(i + 2));
617 	}
618 	assert(riov.i == riov.used);
619 	vringh_iov_cleanup(&riov);
620 	vringh_iov_cleanup(&wiov);
621 
622 	/* Complete using multi interface, just because we can. */
623 	used[0].id = head;
624 	used[0].len = 0;
625 	err = vringh_complete_multi_user(&vrh, used, 1);
626 	if (err)
627 		errx(1, "vringh_complete_multi_user(1): %i", err);
628 
629 	/* Free up those descriptors. */
630 	ret = virtqueue_get_buf(vq, &i);
631 	if (ret != &err)
632 		errx(1, "virtqueue_get_buf: %p", ret);
633 
634 	/* Add lots of descriptors. */
635 	sg_init_table(guest_sg, 1);
636 	sg_set_buf(&guest_sg[0], __user_addr_max - 1, 1);
637 	for (i = 0; i < RINGSIZE; i++) {
638 		err = virtqueue_add_outbuf(vq, guest_sg, 1, &err, GFP_KERNEL);
639 		if (err)
640 			errx(1, "virtqueue_add_outbuf (multiple): %i", err);
641 	}
642 
643 	/* Now get many, and consume them all at once. */
644 	vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
645 	vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
646 
647 	for (i = 0; i < RINGSIZE; i++) {
648 		err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
649 		if (err != 1)
650 			errx(1, "vringh_getdesc_user: %i", err);
651 		used[i].id = head;
652 		used[i].len = 0;
653 	}
654 	/* Make sure it wraps around ring, to test! */
655 	assert(vrh.vring.used->idx % RINGSIZE != 0);
656 	err = vringh_complete_multi_user(&vrh, used, RINGSIZE);
657 	if (err)
658 		errx(1, "vringh_complete_multi_user: %i", err);
659 
660 	/* Free those buffers. */
661 	for (i = 0; i < RINGSIZE; i++) {
662 		unsigned len;
663 		assert(virtqueue_get_buf(vq, &len) != NULL);
664 	}
665 
666 	/* Test weird (but legal!) indirect. */
667 	if (__virtio_test_bit(&vdev, VIRTIO_RING_F_INDIRECT_DESC)) {
668 		char *data = __user_addr_max - USER_MEM/4;
669 		struct vring_desc *d = __user_addr_max - USER_MEM/2;
670 		struct vring vring;
671 
672 		/* Force creation of direct, which we modify. */
673 		__virtio_clear_bit(&vdev, VIRTIO_RING_F_INDIRECT_DESC);
674 		vq = vring_new_virtqueue(0, RINGSIZE, ALIGN, &vdev, true,
675 					 false, __user_addr_min,
676 					 never_notify_host,
677 					 never_callback_guest,
678 					 "guest vq");
679 
680 		sg_init_table(guest_sg, 4);
681 		sg_set_buf(&guest_sg[0], d, sizeof(*d)*2);
682 		sg_set_buf(&guest_sg[1], d + 2, sizeof(*d)*1);
683 		sg_set_buf(&guest_sg[2], data + 6, 4);
684 		sg_set_buf(&guest_sg[3], d + 3, sizeof(*d)*3);
685 
686 		err = virtqueue_add_outbuf(vq, guest_sg, 4, &err, GFP_KERNEL);
687 		if (err)
688 			errx(1, "virtqueue_add_outbuf (indirect): %i", err);
689 
690 		vring_init(&vring, RINGSIZE, __user_addr_min, ALIGN);
691 
692 		/* They're used in order, but double-check... */
693 		assert(vring.desc[0].addr == (unsigned long)d);
694 		assert(vring.desc[1].addr == (unsigned long)(d+2));
695 		assert(vring.desc[2].addr == (unsigned long)data + 6);
696 		assert(vring.desc[3].addr == (unsigned long)(d+3));
697 		vring.desc[0].flags |= VRING_DESC_F_INDIRECT;
698 		vring.desc[1].flags |= VRING_DESC_F_INDIRECT;
699 		vring.desc[3].flags |= VRING_DESC_F_INDIRECT;
700 
701 		/* First indirect */
702 		d[0].addr = (unsigned long)data;
703 		d[0].len = 1;
704 		d[0].flags = VRING_DESC_F_NEXT;
705 		d[0].next = 1;
706 		d[1].addr = (unsigned long)data + 1;
707 		d[1].len = 2;
708 		d[1].flags = 0;
709 
710 		/* Second indirect */
711 		d[2].addr = (unsigned long)data + 3;
712 		d[2].len = 3;
713 		d[2].flags = 0;
714 
715 		/* Third indirect */
716 		d[3].addr = (unsigned long)data + 10;
717 		d[3].len = 5;
718 		d[3].flags = VRING_DESC_F_NEXT;
719 		d[3].next = 1;
720 		d[4].addr = (unsigned long)data + 15;
721 		d[4].len = 6;
722 		d[4].flags = VRING_DESC_F_NEXT;
723 		d[4].next = 2;
724 		d[5].addr = (unsigned long)data + 21;
725 		d[5].len = 7;
726 		d[5].flags = 0;
727 
728 		/* Host picks it up (allocates new iov). */
729 		vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
730 		vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
731 
732 		err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
733 		if (err != 1)
734 			errx(1, "vringh_getdesc_user: %i", err);
735 
736 		if (head != 0)
737 			errx(1, "vringh_getdesc_user: head %i not 0", head);
738 
739 		assert(riov.max_num & VRINGH_IOV_ALLOCATED);
740 		if (getrange != getrange_slow)
741 			assert(riov.used == 7);
742 		else
743 			assert(riov.used == 28);
744 		err = vringh_iov_pull_user(&riov, buf, 29);
745 		assert(err == 28);
746 
747 		/* Data should be linear. */
748 		for (i = 0; i < err; i++)
749 			assert(buf[i] == i);
750 		vringh_iov_cleanup(&riov);
751 	}
752 
753 	/* Don't leak memory... */
754 	vring_del_virtqueue(vq);
755 	free(__user_addr_min);
756 
757 	return 0;
758 }
759