xref: /openbmc/linux/tools/virtio/vringh_test.c (revision d0e22329)
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 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(to_guest);
165 	pipe(to_host);
166 
167 	CPU_ZERO(&cpu_set);
168 	find_cpus(&first_cpu, &last_cpu);
169 	printf("Using CPUS %u and %u\n", first_cpu, last_cpu);
170 	fflush(stdout);
171 
172 	if (fork() != 0) {
173 		struct vringh vrh;
174 		int status, err, rlen = 0;
175 		char rbuf[5];
176 
177 		/* We are the host: never access guest addresses! */
178 		munmap(guest_map, mapsize);
179 
180 		__user_addr_min = host_map;
181 		__user_addr_max = __user_addr_min + mapsize;
182 		user_addr_offset = host_map - guest_map;
183 		assert(user_addr_offset);
184 
185 		close(to_guest[0]);
186 		close(to_host[1]);
187 
188 		vring_init(&vrh.vring, RINGSIZE, host_map, ALIGN);
189 		vringh_init_user(&vrh, features, RINGSIZE, true,
190 				 vrh.vring.desc, vrh.vring.avail, vrh.vring.used);
191 		CPU_SET(first_cpu, &cpu_set);
192 		if (sched_setaffinity(getpid(), sizeof(cpu_set), &cpu_set))
193 			errx(1, "Could not set affinity to cpu %u", first_cpu);
194 
195 		while (xfers < NUM_XFERS) {
196 			struct iovec host_riov[2], host_wiov[2];
197 			struct vringh_iov riov, wiov;
198 			u16 head, written;
199 
200 			if (fast_vringh) {
201 				for (;;) {
202 					err = vringh_get_head(&vrh, &head);
203 					if (err != 0)
204 						break;
205 					err = vringh_need_notify_user(&vrh);
206 					if (err < 0)
207 						errx(1, "vringh_need_notify_user: %i",
208 						     err);
209 					if (err) {
210 						write(to_guest[1], "", 1);
211 						notifies++;
212 					}
213 				}
214 				if (err != 1)
215 					errx(1, "vringh_get_head");
216 				written = 0;
217 				goto complete;
218 			} else {
219 				vringh_iov_init(&riov,
220 						host_riov,
221 						ARRAY_SIZE(host_riov));
222 				vringh_iov_init(&wiov,
223 						host_wiov,
224 						ARRAY_SIZE(host_wiov));
225 
226 				err = vringh_getdesc_user(&vrh, &riov, &wiov,
227 							  getrange, &head);
228 			}
229 			if (err == 0) {
230 				err = vringh_need_notify_user(&vrh);
231 				if (err < 0)
232 					errx(1, "vringh_need_notify_user: %i",
233 					     err);
234 				if (err) {
235 					write(to_guest[1], "", 1);
236 					notifies++;
237 				}
238 
239 				if (!vringh_notify_enable_user(&vrh))
240 					continue;
241 
242 				/* Swallow all notifies at once. */
243 				if (read(to_host[0], buf, sizeof(buf)) < 1)
244 					break;
245 
246 				vringh_notify_disable_user(&vrh);
247 				receives++;
248 				continue;
249 			}
250 			if (err != 1)
251 				errx(1, "vringh_getdesc_user: %i", err);
252 
253 			/* We simply copy bytes. */
254 			if (riov.used) {
255 				rlen = vringh_iov_pull_user(&riov, rbuf,
256 							    sizeof(rbuf));
257 				if (rlen != 4)
258 					errx(1, "vringh_iov_pull_user: %i",
259 					     rlen);
260 				assert(riov.i == riov.used);
261 				written = 0;
262 			} else {
263 				err = vringh_iov_push_user(&wiov, rbuf, rlen);
264 				if (err != rlen)
265 					errx(1, "vringh_iov_push_user: %i",
266 					     err);
267 				assert(wiov.i == wiov.used);
268 				written = err;
269 			}
270 		complete:
271 			xfers++;
272 
273 			err = vringh_complete_user(&vrh, head, written);
274 			if (err != 0)
275 				errx(1, "vringh_complete_user: %i", err);
276 		}
277 
278 		err = vringh_need_notify_user(&vrh);
279 		if (err < 0)
280 			errx(1, "vringh_need_notify_user: %i", err);
281 		if (err) {
282 			write(to_guest[1], "", 1);
283 			notifies++;
284 		}
285 		wait(&status);
286 		if (!WIFEXITED(status))
287 			errx(1, "Child died with signal %i?", WTERMSIG(status));
288 		if (WEXITSTATUS(status) != 0)
289 			errx(1, "Child exited %i?", WEXITSTATUS(status));
290 		printf("Host: notified %lu, pinged %lu\n", notifies, receives);
291 		return 0;
292 	} else {
293 		struct guest_virtio_device gvdev;
294 		struct virtqueue *vq;
295 		unsigned int *data;
296 		struct vring_desc *indirects;
297 		unsigned int finished = 0;
298 
299 		/* We pass sg[]s pointing into here, but we need RINGSIZE+1 */
300 		data = guest_map + vring_size(RINGSIZE, ALIGN);
301 		indirects = (void *)data + (RINGSIZE + 1) * 2 * sizeof(int);
302 
303 		/* We are the guest. */
304 		munmap(host_map, mapsize);
305 
306 		close(to_guest[1]);
307 		close(to_host[0]);
308 
309 		gvdev.vdev.features = features;
310 		gvdev.to_host_fd = to_host[1];
311 		gvdev.notifies = 0;
312 
313 		CPU_SET(first_cpu, &cpu_set);
314 		if (sched_setaffinity(getpid(), sizeof(cpu_set), &cpu_set))
315 			err(1, "Could not set affinity to cpu %u", first_cpu);
316 
317 		vq = vring_new_virtqueue(0, RINGSIZE, ALIGN, &gvdev.vdev, true,
318 					 false, guest_map,
319 					 fast_vringh ? no_notify_host
320 					 : parallel_notify_host,
321 					 never_callback_guest, "guest vq");
322 
323 		/* Don't kfree indirects. */
324 		__kfree_ignore_start = indirects;
325 		__kfree_ignore_end = indirects + RINGSIZE * 6;
326 
327 		while (xfers < NUM_XFERS) {
328 			struct scatterlist sg[4];
329 			unsigned int num_sg, len;
330 			int *dbuf, err;
331 			bool output = !(xfers % 2);
332 
333 			/* Consume bufs. */
334 			while ((dbuf = virtqueue_get_buf(vq, &len)) != NULL) {
335 				if (len == 4)
336 					assert(*dbuf == finished - 1);
337 				else if (!fast_vringh)
338 					assert(*dbuf == finished);
339 				finished++;
340 			}
341 
342 			/* Produce a buffer. */
343 			dbuf = data + (xfers % (RINGSIZE + 1));
344 
345 			if (output)
346 				*dbuf = xfers;
347 			else
348 				*dbuf = -1;
349 
350 			switch ((xfers / sizeof(*dbuf)) % 4) {
351 			case 0:
352 				/* Nasty three-element sg list. */
353 				sg_init_table(sg, num_sg = 3);
354 				sg_set_buf(&sg[0], (void *)dbuf, 1);
355 				sg_set_buf(&sg[1], (void *)dbuf + 1, 2);
356 				sg_set_buf(&sg[2], (void *)dbuf + 3, 1);
357 				break;
358 			case 1:
359 				sg_init_table(sg, num_sg = 2);
360 				sg_set_buf(&sg[0], (void *)dbuf, 1);
361 				sg_set_buf(&sg[1], (void *)dbuf + 1, 3);
362 				break;
363 			case 2:
364 				sg_init_table(sg, num_sg = 1);
365 				sg_set_buf(&sg[0], (void *)dbuf, 4);
366 				break;
367 			case 3:
368 				sg_init_table(sg, num_sg = 4);
369 				sg_set_buf(&sg[0], (void *)dbuf, 1);
370 				sg_set_buf(&sg[1], (void *)dbuf + 1, 1);
371 				sg_set_buf(&sg[2], (void *)dbuf + 2, 1);
372 				sg_set_buf(&sg[3], (void *)dbuf + 3, 1);
373 				break;
374 			}
375 
376 			/* May allocate an indirect, so force it to allocate
377 			 * user addr */
378 			__kmalloc_fake = indirects + (xfers % RINGSIZE) * 4;
379 			if (output)
380 				err = virtqueue_add_outbuf(vq, sg, num_sg, dbuf,
381 							   GFP_KERNEL);
382 			else
383 				err = virtqueue_add_inbuf(vq, sg, num_sg,
384 							  dbuf, GFP_KERNEL);
385 
386 			if (err == -ENOSPC) {
387 				if (!virtqueue_enable_cb_delayed(vq))
388 					continue;
389 				/* Swallow all notifies at once. */
390 				if (read(to_guest[0], buf, sizeof(buf)) < 1)
391 					break;
392 
393 				receives++;
394 				virtqueue_disable_cb(vq);
395 				continue;
396 			}
397 
398 			if (err)
399 				errx(1, "virtqueue_add_in/outbuf: %i", err);
400 
401 			xfers++;
402 			virtqueue_kick(vq);
403 		}
404 
405 		/* Any extra? */
406 		while (finished != xfers) {
407 			int *dbuf;
408 			unsigned int len;
409 
410 			/* Consume bufs. */
411 			dbuf = virtqueue_get_buf(vq, &len);
412 			if (dbuf) {
413 				if (len == 4)
414 					assert(*dbuf == finished - 1);
415 				else
416 					assert(len == 0);
417 				finished++;
418 				continue;
419 			}
420 
421 			if (!virtqueue_enable_cb_delayed(vq))
422 				continue;
423 			if (read(to_guest[0], buf, sizeof(buf)) < 1)
424 				break;
425 
426 			receives++;
427 			virtqueue_disable_cb(vq);
428 		}
429 
430 		printf("Guest: notified %lu, pinged %lu\n",
431 		       gvdev.notifies, receives);
432 		vring_del_virtqueue(vq);
433 		return 0;
434 	}
435 }
436 
437 int main(int argc, char *argv[])
438 {
439 	struct virtio_device vdev;
440 	struct virtqueue *vq;
441 	struct vringh vrh;
442 	struct scatterlist guest_sg[RINGSIZE], *sgs[2];
443 	struct iovec host_riov[2], host_wiov[2];
444 	struct vringh_iov riov, wiov;
445 	struct vring_used_elem used[RINGSIZE];
446 	char buf[28];
447 	u16 head;
448 	int err;
449 	unsigned i;
450 	void *ret;
451 	bool (*getrange)(struct vringh *vrh, u64 addr, struct vringh_range *r);
452 	bool fast_vringh = false, parallel = false;
453 
454 	getrange = getrange_iov;
455 	vdev.features = 0;
456 
457 	while (argv[1]) {
458 		if (strcmp(argv[1], "--indirect") == 0)
459 			__virtio_set_bit(&vdev, VIRTIO_RING_F_INDIRECT_DESC);
460 		else if (strcmp(argv[1], "--eventidx") == 0)
461 			__virtio_set_bit(&vdev, VIRTIO_RING_F_EVENT_IDX);
462 		else if (strcmp(argv[1], "--virtio-1") == 0)
463 			__virtio_set_bit(&vdev, VIRTIO_F_VERSION_1);
464 		else if (strcmp(argv[1], "--slow-range") == 0)
465 			getrange = getrange_slow;
466 		else if (strcmp(argv[1], "--fast-vringh") == 0)
467 			fast_vringh = true;
468 		else if (strcmp(argv[1], "--parallel") == 0)
469 			parallel = true;
470 		else
471 			errx(1, "Unknown arg %s", argv[1]);
472 		argv++;
473 	}
474 
475 	if (parallel)
476 		return parallel_test(vdev.features, getrange, fast_vringh);
477 
478 	if (posix_memalign(&__user_addr_min, PAGE_SIZE, USER_MEM) != 0)
479 		abort();
480 	__user_addr_max = __user_addr_min + USER_MEM;
481 	memset(__user_addr_min, 0, vring_size(RINGSIZE, ALIGN));
482 
483 	/* Set up guest side. */
484 	vq = vring_new_virtqueue(0, RINGSIZE, ALIGN, &vdev, true, false,
485 				 __user_addr_min,
486 				 never_notify_host, never_callback_guest,
487 				 "guest vq");
488 
489 	/* Set up host side. */
490 	vring_init(&vrh.vring, RINGSIZE, __user_addr_min, ALIGN);
491 	vringh_init_user(&vrh, vdev.features, RINGSIZE, true,
492 			 vrh.vring.desc, vrh.vring.avail, vrh.vring.used);
493 
494 	/* No descriptor to get yet... */
495 	err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
496 	if (err != 0)
497 		errx(1, "vringh_getdesc_user: %i", err);
498 
499 	/* Guest puts in a descriptor. */
500 	memcpy(__user_addr_max - 1, "a", 1);
501 	sg_init_table(guest_sg, 1);
502 	sg_set_buf(&guest_sg[0], __user_addr_max - 1, 1);
503 	sg_init_table(guest_sg+1, 1);
504 	sg_set_buf(&guest_sg[1], __user_addr_max - 3, 2);
505 	sgs[0] = &guest_sg[0];
506 	sgs[1] = &guest_sg[1];
507 
508 	/* May allocate an indirect, so force it to allocate user addr */
509 	__kmalloc_fake = __user_addr_min + vring_size(RINGSIZE, ALIGN);
510 	err = virtqueue_add_sgs(vq, sgs, 1, 1, &err, GFP_KERNEL);
511 	if (err)
512 		errx(1, "virtqueue_add_sgs: %i", err);
513 	__kmalloc_fake = NULL;
514 
515 	/* Host retreives it. */
516 	vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
517 	vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
518 
519 	err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
520 	if (err != 1)
521 		errx(1, "vringh_getdesc_user: %i", err);
522 
523 	assert(riov.used == 1);
524 	assert(riov.iov[0].iov_base == __user_addr_max - 1);
525 	assert(riov.iov[0].iov_len == 1);
526 	if (getrange != getrange_slow) {
527 		assert(wiov.used == 1);
528 		assert(wiov.iov[0].iov_base == __user_addr_max - 3);
529 		assert(wiov.iov[0].iov_len == 2);
530 	} else {
531 		assert(wiov.used == 2);
532 		assert(wiov.iov[0].iov_base == __user_addr_max - 3);
533 		assert(wiov.iov[0].iov_len == 1);
534 		assert(wiov.iov[1].iov_base == __user_addr_max - 2);
535 		assert(wiov.iov[1].iov_len == 1);
536 	}
537 
538 	err = vringh_iov_pull_user(&riov, buf, 5);
539 	if (err != 1)
540 		errx(1, "vringh_iov_pull_user: %i", err);
541 	assert(buf[0] == 'a');
542 	assert(riov.i == 1);
543 	assert(vringh_iov_pull_user(&riov, buf, 5) == 0);
544 
545 	memcpy(buf, "bcdef", 5);
546 	err = vringh_iov_push_user(&wiov, buf, 5);
547 	if (err != 2)
548 		errx(1, "vringh_iov_push_user: %i", err);
549 	assert(memcmp(__user_addr_max - 3, "bc", 2) == 0);
550 	assert(wiov.i == wiov.used);
551 	assert(vringh_iov_push_user(&wiov, buf, 5) == 0);
552 
553 	/* Host is done. */
554 	err = vringh_complete_user(&vrh, head, err);
555 	if (err != 0)
556 		errx(1, "vringh_complete_user: %i", err);
557 
558 	/* Guest should see used token now. */
559 	__kfree_ignore_start = __user_addr_min + vring_size(RINGSIZE, ALIGN);
560 	__kfree_ignore_end = __kfree_ignore_start + 1;
561 	ret = virtqueue_get_buf(vq, &i);
562 	if (ret != &err)
563 		errx(1, "virtqueue_get_buf: %p", ret);
564 	assert(i == 2);
565 
566 	/* Guest puts in a huge descriptor. */
567 	sg_init_table(guest_sg, RINGSIZE);
568 	for (i = 0; i < RINGSIZE; i++) {
569 		sg_set_buf(&guest_sg[i],
570 			   __user_addr_max - USER_MEM/4, USER_MEM/4);
571 	}
572 
573 	/* Fill contents with recognisable garbage. */
574 	for (i = 0; i < USER_MEM/4; i++)
575 		((char *)__user_addr_max - USER_MEM/4)[i] = i;
576 
577 	/* This will allocate an indirect, so force it to allocate user addr */
578 	__kmalloc_fake = __user_addr_min + vring_size(RINGSIZE, ALIGN);
579 	err = virtqueue_add_outbuf(vq, guest_sg, RINGSIZE, &err, GFP_KERNEL);
580 	if (err)
581 		errx(1, "virtqueue_add_outbuf (large): %i", err);
582 	__kmalloc_fake = NULL;
583 
584 	/* Host picks it up (allocates new iov). */
585 	vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
586 	vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
587 
588 	err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
589 	if (err != 1)
590 		errx(1, "vringh_getdesc_user: %i", err);
591 
592 	assert(riov.max_num & VRINGH_IOV_ALLOCATED);
593 	assert(riov.iov != host_riov);
594 	if (getrange != getrange_slow)
595 		assert(riov.used == RINGSIZE);
596 	else
597 		assert(riov.used == RINGSIZE * USER_MEM/4);
598 
599 	assert(!(wiov.max_num & VRINGH_IOV_ALLOCATED));
600 	assert(wiov.used == 0);
601 
602 	/* Pull data back out (in odd chunks), should be as expected. */
603 	for (i = 0; i < RINGSIZE * USER_MEM/4; i += 3) {
604 		err = vringh_iov_pull_user(&riov, buf, 3);
605 		if (err != 3 && i + err != RINGSIZE * USER_MEM/4)
606 			errx(1, "vringh_iov_pull_user large: %i", err);
607 		assert(buf[0] == (char)i);
608 		assert(err < 2 || buf[1] == (char)(i + 1));
609 		assert(err < 3 || buf[2] == (char)(i + 2));
610 	}
611 	assert(riov.i == riov.used);
612 	vringh_iov_cleanup(&riov);
613 	vringh_iov_cleanup(&wiov);
614 
615 	/* Complete using multi interface, just because we can. */
616 	used[0].id = head;
617 	used[0].len = 0;
618 	err = vringh_complete_multi_user(&vrh, used, 1);
619 	if (err)
620 		errx(1, "vringh_complete_multi_user(1): %i", err);
621 
622 	/* Free up those descriptors. */
623 	ret = virtqueue_get_buf(vq, &i);
624 	if (ret != &err)
625 		errx(1, "virtqueue_get_buf: %p", ret);
626 
627 	/* Add lots of descriptors. */
628 	sg_init_table(guest_sg, 1);
629 	sg_set_buf(&guest_sg[0], __user_addr_max - 1, 1);
630 	for (i = 0; i < RINGSIZE; i++) {
631 		err = virtqueue_add_outbuf(vq, guest_sg, 1, &err, GFP_KERNEL);
632 		if (err)
633 			errx(1, "virtqueue_add_outbuf (multiple): %i", err);
634 	}
635 
636 	/* Now get many, and consume them all at once. */
637 	vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
638 	vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
639 
640 	for (i = 0; i < RINGSIZE; i++) {
641 		err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
642 		if (err != 1)
643 			errx(1, "vringh_getdesc_user: %i", err);
644 		used[i].id = head;
645 		used[i].len = 0;
646 	}
647 	/* Make sure it wraps around ring, to test! */
648 	assert(vrh.vring.used->idx % RINGSIZE != 0);
649 	err = vringh_complete_multi_user(&vrh, used, RINGSIZE);
650 	if (err)
651 		errx(1, "vringh_complete_multi_user: %i", err);
652 
653 	/* Free those buffers. */
654 	for (i = 0; i < RINGSIZE; i++) {
655 		unsigned len;
656 		assert(virtqueue_get_buf(vq, &len) != NULL);
657 	}
658 
659 	/* Test weird (but legal!) indirect. */
660 	if (__virtio_test_bit(&vdev, VIRTIO_RING_F_INDIRECT_DESC)) {
661 		char *data = __user_addr_max - USER_MEM/4;
662 		struct vring_desc *d = __user_addr_max - USER_MEM/2;
663 		struct vring vring;
664 
665 		/* Force creation of direct, which we modify. */
666 		__virtio_clear_bit(&vdev, VIRTIO_RING_F_INDIRECT_DESC);
667 		vq = vring_new_virtqueue(0, RINGSIZE, ALIGN, &vdev, true,
668 					 false, __user_addr_min,
669 					 never_notify_host,
670 					 never_callback_guest,
671 					 "guest vq");
672 
673 		sg_init_table(guest_sg, 4);
674 		sg_set_buf(&guest_sg[0], d, sizeof(*d)*2);
675 		sg_set_buf(&guest_sg[1], d + 2, sizeof(*d)*1);
676 		sg_set_buf(&guest_sg[2], data + 6, 4);
677 		sg_set_buf(&guest_sg[3], d + 3, sizeof(*d)*3);
678 
679 		err = virtqueue_add_outbuf(vq, guest_sg, 4, &err, GFP_KERNEL);
680 		if (err)
681 			errx(1, "virtqueue_add_outbuf (indirect): %i", err);
682 
683 		vring_init(&vring, RINGSIZE, __user_addr_min, ALIGN);
684 
685 		/* They're used in order, but double-check... */
686 		assert(vring.desc[0].addr == (unsigned long)d);
687 		assert(vring.desc[1].addr == (unsigned long)(d+2));
688 		assert(vring.desc[2].addr == (unsigned long)data + 6);
689 		assert(vring.desc[3].addr == (unsigned long)(d+3));
690 		vring.desc[0].flags |= VRING_DESC_F_INDIRECT;
691 		vring.desc[1].flags |= VRING_DESC_F_INDIRECT;
692 		vring.desc[3].flags |= VRING_DESC_F_INDIRECT;
693 
694 		/* First indirect */
695 		d[0].addr = (unsigned long)data;
696 		d[0].len = 1;
697 		d[0].flags = VRING_DESC_F_NEXT;
698 		d[0].next = 1;
699 		d[1].addr = (unsigned long)data + 1;
700 		d[1].len = 2;
701 		d[1].flags = 0;
702 
703 		/* Second indirect */
704 		d[2].addr = (unsigned long)data + 3;
705 		d[2].len = 3;
706 		d[2].flags = 0;
707 
708 		/* Third indirect */
709 		d[3].addr = (unsigned long)data + 10;
710 		d[3].len = 5;
711 		d[3].flags = VRING_DESC_F_NEXT;
712 		d[3].next = 1;
713 		d[4].addr = (unsigned long)data + 15;
714 		d[4].len = 6;
715 		d[4].flags = VRING_DESC_F_NEXT;
716 		d[4].next = 2;
717 		d[5].addr = (unsigned long)data + 21;
718 		d[5].len = 7;
719 		d[5].flags = 0;
720 
721 		/* Host picks it up (allocates new iov). */
722 		vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
723 		vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
724 
725 		err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
726 		if (err != 1)
727 			errx(1, "vringh_getdesc_user: %i", err);
728 
729 		if (head != 0)
730 			errx(1, "vringh_getdesc_user: head %i not 0", head);
731 
732 		assert(riov.max_num & VRINGH_IOV_ALLOCATED);
733 		if (getrange != getrange_slow)
734 			assert(riov.used == 7);
735 		else
736 			assert(riov.used == 28);
737 		err = vringh_iov_pull_user(&riov, buf, 29);
738 		assert(err == 28);
739 
740 		/* Data should be linear. */
741 		for (i = 0; i < err; i++)
742 			assert(buf[i] == i);
743 		vringh_iov_cleanup(&riov);
744 	}
745 
746 	/* Don't leak memory... */
747 	vring_del_virtqueue(vq);
748 	free(__user_addr_min);
749 
750 	return 0;
751 }
752