1 /* Simple test of virtio code, entirely in userpsace. */
5 #include <linux/kernel.h>
7 #include <linux/virtio.h>
8 #include <linux/vringh.h>
9 #include <linux/virtio_ring.h>
10 #include <linux/virtio_config.h>
11 #include <linux/uaccess.h>
12 #include <sys/types.h>
18 #define USER_MEM (1024*1024)
19 void *__user_addr_min, *__user_addr_max;
20 void *__kmalloc_fake, *__kfree_ignore_start, *__kfree_ignore_end;
21 static u64 user_addr_offset;
26 static bool never_notify_host(struct virtqueue *vq)
31 static void never_callback_guest(struct virtqueue *vq)
36 static bool getrange_iov(struct vringh *vrh, u64 addr, struct vringh_range *r)
38 if (addr < (u64)(unsigned long)__user_addr_min - user_addr_offset)
40 if (addr >= (u64)(unsigned long)__user_addr_max - user_addr_offset)
43 r->start = (u64)(unsigned long)__user_addr_min - user_addr_offset;
44 r->end_incl = (u64)(unsigned long)__user_addr_max - 1 - user_addr_offset;
45 r->offset = user_addr_offset;
49 /* We return single byte ranges. */
50 static bool getrange_slow(struct vringh *vrh, u64 addr, struct vringh_range *r)
52 if (addr < (u64)(unsigned long)__user_addr_min - user_addr_offset)
54 if (addr >= (u64)(unsigned long)__user_addr_max - user_addr_offset)
58 r->end_incl = r->start;
59 r->offset = user_addr_offset;
63 struct guest_virtio_device {
64 struct virtio_device vdev;
66 unsigned long notifies;
69 static bool parallel_notify_host(struct virtqueue *vq)
72 struct guest_virtio_device *gvdev;
74 gvdev = container_of(vq->vdev, struct guest_virtio_device, vdev);
75 rc = write(gvdev->to_host_fd, "", 1);
82 static bool no_notify_host(struct virtqueue *vq)
87 #define NUM_XFERS (10000000)
89 /* We aim for two "distant" cpus. */
90 static void find_cpus(unsigned int *first, unsigned int *last)
96 for (i = 0; i < 4096; i++) {
100 if (sched_setaffinity(getpid(), sizeof(set), &set) == 0) {
109 /* Opencoded version for fast mode */
110 static inline int vringh_get_head(struct vringh *vrh, u16 *head)
115 err = get_user(avail_idx, &vrh->vring.avail->idx);
119 if (vrh->last_avail_idx == avail_idx)
122 /* Only get avail ring entries after they have been exposed by guest. */
123 virtio_rmb(vrh->weak_barriers);
125 i = vrh->last_avail_idx & (vrh->vring.num - 1);
127 err = get_user(*head, &vrh->vring.avail->ring[i]);
131 vrh->last_avail_idx++;
135 static int parallel_test(u64 features,
136 bool (*getrange)(struct vringh *vrh,
137 u64 addr, struct vringh_range *r),
140 void *host_map, *guest_map;
141 int fd, mapsize, to_guest[2], to_host[2];
142 unsigned long xfers = 0, notifies = 0, receives = 0;
143 unsigned int first_cpu, last_cpu;
147 /* Create real file to mmap. */
148 fd = open("/tmp/vringh_test-file", O_RDWR|O_CREAT|O_TRUNC, 0600);
150 err(1, "Opening /tmp/vringh_test-file");
152 /* Extra room at the end for some data, and indirects */
153 mapsize = vring_size(RINGSIZE, ALIGN)
154 + RINGSIZE * 2 * sizeof(int)
155 + RINGSIZE * 6 * sizeof(struct vring_desc);
156 mapsize = (mapsize + getpagesize() - 1) & ~(getpagesize() - 1);
157 ftruncate(fd, mapsize);
159 /* Parent and child use separate addresses, to check our mapping logic! */
160 host_map = mmap(NULL, mapsize, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
161 guest_map = mmap(NULL, mapsize, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
167 find_cpus(&first_cpu, &last_cpu);
168 printf("Using CPUS %u and %u\n", first_cpu, last_cpu);
173 int status, err, rlen = 0;
176 /* We are the host: never access guest addresses! */
177 munmap(guest_map, mapsize);
179 __user_addr_min = host_map;
180 __user_addr_max = __user_addr_min + mapsize;
181 user_addr_offset = host_map - guest_map;
182 assert(user_addr_offset);
187 vring_init(&vrh.vring, RINGSIZE, host_map, ALIGN);
188 vringh_init_user(&vrh, features, RINGSIZE, true,
189 vrh.vring.desc, vrh.vring.avail, vrh.vring.used);
190 CPU_SET(first_cpu, &cpu_set);
191 if (sched_setaffinity(getpid(), sizeof(cpu_set), &cpu_set))
192 errx(1, "Could not set affinity to cpu %u", first_cpu);
194 while (xfers < NUM_XFERS) {
195 struct iovec host_riov[2], host_wiov[2];
196 struct vringh_iov riov, wiov;
201 err = vringh_get_head(&vrh, &head);
204 err = vringh_need_notify_user(&vrh);
206 errx(1, "vringh_need_notify_user: %i",
209 write(to_guest[1], "", 1);
214 errx(1, "vringh_get_head");
218 vringh_iov_init(&riov,
220 ARRAY_SIZE(host_riov));
221 vringh_iov_init(&wiov,
223 ARRAY_SIZE(host_wiov));
225 err = vringh_getdesc_user(&vrh, &riov, &wiov,
229 err = vringh_need_notify_user(&vrh);
231 errx(1, "vringh_need_notify_user: %i",
234 write(to_guest[1], "", 1);
238 if (!vringh_notify_enable_user(&vrh))
241 /* Swallow all notifies at once. */
242 if (read(to_host[0], buf, sizeof(buf)) < 1)
245 vringh_notify_disable_user(&vrh);
250 errx(1, "vringh_getdesc_user: %i", err);
252 /* We simply copy bytes. */
254 rlen = vringh_iov_pull_user(&riov, rbuf,
257 errx(1, "vringh_iov_pull_user: %i",
259 assert(riov.i == riov.used);
262 err = vringh_iov_push_user(&wiov, rbuf, rlen);
264 errx(1, "vringh_iov_push_user: %i",
266 assert(wiov.i == wiov.used);
272 err = vringh_complete_user(&vrh, head, written);
274 errx(1, "vringh_complete_user: %i", err);
277 err = vringh_need_notify_user(&vrh);
279 errx(1, "vringh_need_notify_user: %i", err);
281 write(to_guest[1], "", 1);
285 if (!WIFEXITED(status))
286 errx(1, "Child died with signal %i?", WTERMSIG(status));
287 if (WEXITSTATUS(status) != 0)
288 errx(1, "Child exited %i?", WEXITSTATUS(status));
289 printf("Host: notified %lu, pinged %lu\n", notifies, receives);
292 struct guest_virtio_device gvdev;
293 struct virtqueue *vq;
295 struct vring_desc *indirects;
296 unsigned int finished = 0;
298 /* We pass sg[]s pointing into here, but we need RINGSIZE+1 */
299 data = guest_map + vring_size(RINGSIZE, ALIGN);
300 indirects = (void *)data + (RINGSIZE + 1) * 2 * sizeof(int);
302 /* We are the guest. */
303 munmap(host_map, mapsize);
308 gvdev.vdev.features = features;
309 gvdev.to_host_fd = to_host[1];
312 CPU_SET(first_cpu, &cpu_set);
313 if (sched_setaffinity(getpid(), sizeof(cpu_set), &cpu_set))
314 err(1, "Could not set affinity to cpu %u", first_cpu);
316 vq = vring_new_virtqueue(0, RINGSIZE, ALIGN, &gvdev.vdev, true,
318 fast_vringh ? no_notify_host
319 : parallel_notify_host,
320 never_callback_guest, "guest vq");
322 /* Don't kfree indirects. */
323 __kfree_ignore_start = indirects;
324 __kfree_ignore_end = indirects + RINGSIZE * 6;
326 while (xfers < NUM_XFERS) {
327 struct scatterlist sg[4];
328 unsigned int num_sg, len;
330 bool output = !(xfers % 2);
333 while ((dbuf = virtqueue_get_buf(vq, &len)) != NULL) {
335 assert(*dbuf == finished - 1);
336 else if (!fast_vringh)
337 assert(*dbuf == finished);
341 /* Produce a buffer. */
342 dbuf = data + (xfers % (RINGSIZE + 1));
349 switch ((xfers / sizeof(*dbuf)) % 4) {
351 /* Nasty three-element sg list. */
352 sg_init_table(sg, num_sg = 3);
353 sg_set_buf(&sg[0], (void *)dbuf, 1);
354 sg_set_buf(&sg[1], (void *)dbuf + 1, 2);
355 sg_set_buf(&sg[2], (void *)dbuf + 3, 1);
358 sg_init_table(sg, num_sg = 2);
359 sg_set_buf(&sg[0], (void *)dbuf, 1);
360 sg_set_buf(&sg[1], (void *)dbuf + 1, 3);
363 sg_init_table(sg, num_sg = 1);
364 sg_set_buf(&sg[0], (void *)dbuf, 4);
367 sg_init_table(sg, num_sg = 4);
368 sg_set_buf(&sg[0], (void *)dbuf, 1);
369 sg_set_buf(&sg[1], (void *)dbuf + 1, 1);
370 sg_set_buf(&sg[2], (void *)dbuf + 2, 1);
371 sg_set_buf(&sg[3], (void *)dbuf + 3, 1);
375 /* May allocate an indirect, so force it to allocate
377 __kmalloc_fake = indirects + (xfers % RINGSIZE) * 4;
379 err = virtqueue_add_outbuf(vq, sg, num_sg, dbuf,
382 err = virtqueue_add_inbuf(vq, sg, num_sg,
385 if (err == -ENOSPC) {
386 if (!virtqueue_enable_cb_delayed(vq))
388 /* Swallow all notifies at once. */
389 if (read(to_guest[0], buf, sizeof(buf)) < 1)
393 virtqueue_disable_cb(vq);
398 errx(1, "virtqueue_add_in/outbuf: %i", err);
405 while (finished != xfers) {
410 dbuf = virtqueue_get_buf(vq, &len);
413 assert(*dbuf == finished - 1);
420 if (!virtqueue_enable_cb_delayed(vq))
422 if (read(to_guest[0], buf, sizeof(buf)) < 1)
426 virtqueue_disable_cb(vq);
429 printf("Guest: notified %lu, pinged %lu\n",
430 gvdev.notifies, receives);
431 vring_del_virtqueue(vq);
436 int main(int argc, char *argv[])
438 struct virtio_device vdev;
439 struct virtqueue *vq;
441 struct scatterlist guest_sg[RINGSIZE], *sgs[2];
442 struct iovec host_riov[2], host_wiov[2];
443 struct vringh_iov riov, wiov;
444 struct vring_used_elem used[RINGSIZE];
450 bool (*getrange)(struct vringh *vrh, u64 addr, struct vringh_range *r);
451 bool fast_vringh = false, parallel = false;
453 getrange = getrange_iov;
457 if (strcmp(argv[1], "--indirect") == 0)
458 __virtio_set_bit(&vdev, VIRTIO_RING_F_INDIRECT_DESC);
459 else if (strcmp(argv[1], "--eventidx") == 0)
460 __virtio_set_bit(&vdev, VIRTIO_RING_F_EVENT_IDX);
461 else if (strcmp(argv[1], "--virtio-1") == 0)
462 __virtio_set_bit(&vdev, VIRTIO_F_VERSION_1);
463 else if (strcmp(argv[1], "--slow-range") == 0)
464 getrange = getrange_slow;
465 else if (strcmp(argv[1], "--fast-vringh") == 0)
467 else if (strcmp(argv[1], "--parallel") == 0)
470 errx(1, "Unknown arg %s", argv[1]);
475 return parallel_test(vdev.features, getrange, fast_vringh);
477 if (posix_memalign(&__user_addr_min, PAGE_SIZE, USER_MEM) != 0)
479 __user_addr_max = __user_addr_min + USER_MEM;
480 memset(__user_addr_min, 0, vring_size(RINGSIZE, ALIGN));
482 /* Set up guest side. */
483 vq = vring_new_virtqueue(0, RINGSIZE, ALIGN, &vdev, true, false,
485 never_notify_host, never_callback_guest,
488 /* Set up host side. */
489 vring_init(&vrh.vring, RINGSIZE, __user_addr_min, ALIGN);
490 vringh_init_user(&vrh, vdev.features, RINGSIZE, true,
491 vrh.vring.desc, vrh.vring.avail, vrh.vring.used);
493 /* No descriptor to get yet... */
494 err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
496 errx(1, "vringh_getdesc_user: %i", err);
498 /* Guest puts in a descriptor. */
499 memcpy(__user_addr_max - 1, "a", 1);
500 sg_init_table(guest_sg, 1);
501 sg_set_buf(&guest_sg[0], __user_addr_max - 1, 1);
502 sg_init_table(guest_sg+1, 1);
503 sg_set_buf(&guest_sg[1], __user_addr_max - 3, 2);
504 sgs[0] = &guest_sg[0];
505 sgs[1] = &guest_sg[1];
507 /* May allocate an indirect, so force it to allocate user addr */
508 __kmalloc_fake = __user_addr_min + vring_size(RINGSIZE, ALIGN);
509 err = virtqueue_add_sgs(vq, sgs, 1, 1, &err, GFP_KERNEL);
511 errx(1, "virtqueue_add_sgs: %i", err);
512 __kmalloc_fake = NULL;
514 /* Host retreives it. */
515 vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
516 vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
518 err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
520 errx(1, "vringh_getdesc_user: %i", err);
522 assert(riov.used == 1);
523 assert(riov.iov[0].iov_base == __user_addr_max - 1);
524 assert(riov.iov[0].iov_len == 1);
525 if (getrange != getrange_slow) {
526 assert(wiov.used == 1);
527 assert(wiov.iov[0].iov_base == __user_addr_max - 3);
528 assert(wiov.iov[0].iov_len == 2);
530 assert(wiov.used == 2);
531 assert(wiov.iov[0].iov_base == __user_addr_max - 3);
532 assert(wiov.iov[0].iov_len == 1);
533 assert(wiov.iov[1].iov_base == __user_addr_max - 2);
534 assert(wiov.iov[1].iov_len == 1);
537 err = vringh_iov_pull_user(&riov, buf, 5);
539 errx(1, "vringh_iov_pull_user: %i", err);
540 assert(buf[0] == 'a');
542 assert(vringh_iov_pull_user(&riov, buf, 5) == 0);
544 memcpy(buf, "bcdef", 5);
545 err = vringh_iov_push_user(&wiov, buf, 5);
547 errx(1, "vringh_iov_push_user: %i", err);
548 assert(memcmp(__user_addr_max - 3, "bc", 2) == 0);
549 assert(wiov.i == wiov.used);
550 assert(vringh_iov_push_user(&wiov, buf, 5) == 0);
553 err = vringh_complete_user(&vrh, head, err);
555 errx(1, "vringh_complete_user: %i", err);
557 /* Guest should see used token now. */
558 __kfree_ignore_start = __user_addr_min + vring_size(RINGSIZE, ALIGN);
559 __kfree_ignore_end = __kfree_ignore_start + 1;
560 ret = virtqueue_get_buf(vq, &i);
562 errx(1, "virtqueue_get_buf: %p", ret);
565 /* Guest puts in a huge descriptor. */
566 sg_init_table(guest_sg, RINGSIZE);
567 for (i = 0; i < RINGSIZE; i++) {
568 sg_set_buf(&guest_sg[i],
569 __user_addr_max - USER_MEM/4, USER_MEM/4);
572 /* Fill contents with recognisable garbage. */
573 for (i = 0; i < USER_MEM/4; i++)
574 ((char *)__user_addr_max - USER_MEM/4)[i] = i;
576 /* This will allocate an indirect, so force it to allocate user addr */
577 __kmalloc_fake = __user_addr_min + vring_size(RINGSIZE, ALIGN);
578 err = virtqueue_add_outbuf(vq, guest_sg, RINGSIZE, &err, GFP_KERNEL);
580 errx(1, "virtqueue_add_outbuf (large): %i", err);
581 __kmalloc_fake = NULL;
583 /* Host picks it up (allocates new iov). */
584 vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
585 vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
587 err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
589 errx(1, "vringh_getdesc_user: %i", err);
591 assert(riov.max_num & VRINGH_IOV_ALLOCATED);
592 assert(riov.iov != host_riov);
593 if (getrange != getrange_slow)
594 assert(riov.used == RINGSIZE);
596 assert(riov.used == RINGSIZE * USER_MEM/4);
598 assert(!(wiov.max_num & VRINGH_IOV_ALLOCATED));
599 assert(wiov.used == 0);
601 /* Pull data back out (in odd chunks), should be as expected. */
602 for (i = 0; i < RINGSIZE * USER_MEM/4; i += 3) {
603 err = vringh_iov_pull_user(&riov, buf, 3);
604 if (err != 3 && i + err != RINGSIZE * USER_MEM/4)
605 errx(1, "vringh_iov_pull_user large: %i", err);
606 assert(buf[0] == (char)i);
607 assert(err < 2 || buf[1] == (char)(i + 1));
608 assert(err < 3 || buf[2] == (char)(i + 2));
610 assert(riov.i == riov.used);
611 vringh_iov_cleanup(&riov);
612 vringh_iov_cleanup(&wiov);
614 /* Complete using multi interface, just because we can. */
617 err = vringh_complete_multi_user(&vrh, used, 1);
619 errx(1, "vringh_complete_multi_user(1): %i", err);
621 /* Free up those descriptors. */
622 ret = virtqueue_get_buf(vq, &i);
624 errx(1, "virtqueue_get_buf: %p", ret);
626 /* Add lots of descriptors. */
627 sg_init_table(guest_sg, 1);
628 sg_set_buf(&guest_sg[0], __user_addr_max - 1, 1);
629 for (i = 0; i < RINGSIZE; i++) {
630 err = virtqueue_add_outbuf(vq, guest_sg, 1, &err, GFP_KERNEL);
632 errx(1, "virtqueue_add_outbuf (multiple): %i", err);
635 /* Now get many, and consume them all at once. */
636 vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
637 vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
639 for (i = 0; i < RINGSIZE; i++) {
640 err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
642 errx(1, "vringh_getdesc_user: %i", err);
646 /* Make sure it wraps around ring, to test! */
647 assert(vrh.vring.used->idx % RINGSIZE != 0);
648 err = vringh_complete_multi_user(&vrh, used, RINGSIZE);
650 errx(1, "vringh_complete_multi_user: %i", err);
652 /* Free those buffers. */
653 for (i = 0; i < RINGSIZE; i++) {
655 assert(virtqueue_get_buf(vq, &len) != NULL);
658 /* Test weird (but legal!) indirect. */
659 if (__virtio_test_bit(&vdev, VIRTIO_RING_F_INDIRECT_DESC)) {
660 char *data = __user_addr_max - USER_MEM/4;
661 struct vring_desc *d = __user_addr_max - USER_MEM/2;
664 /* Force creation of direct, which we modify. */
665 __virtio_clear_bit(&vdev, VIRTIO_RING_F_INDIRECT_DESC);
666 vq = vring_new_virtqueue(0, RINGSIZE, ALIGN, &vdev, true,
667 false, __user_addr_min,
669 never_callback_guest,
672 sg_init_table(guest_sg, 4);
673 sg_set_buf(&guest_sg[0], d, sizeof(*d)*2);
674 sg_set_buf(&guest_sg[1], d + 2, sizeof(*d)*1);
675 sg_set_buf(&guest_sg[2], data + 6, 4);
676 sg_set_buf(&guest_sg[3], d + 3, sizeof(*d)*3);
678 err = virtqueue_add_outbuf(vq, guest_sg, 4, &err, GFP_KERNEL);
680 errx(1, "virtqueue_add_outbuf (indirect): %i", err);
682 vring_init(&vring, RINGSIZE, __user_addr_min, ALIGN);
684 /* They're used in order, but double-check... */
685 assert(vring.desc[0].addr == (unsigned long)d);
686 assert(vring.desc[1].addr == (unsigned long)(d+2));
687 assert(vring.desc[2].addr == (unsigned long)data + 6);
688 assert(vring.desc[3].addr == (unsigned long)(d+3));
689 vring.desc[0].flags |= VRING_DESC_F_INDIRECT;
690 vring.desc[1].flags |= VRING_DESC_F_INDIRECT;
691 vring.desc[3].flags |= VRING_DESC_F_INDIRECT;
694 d[0].addr = (unsigned long)data;
696 d[0].flags = VRING_DESC_F_NEXT;
698 d[1].addr = (unsigned long)data + 1;
702 /* Second indirect */
703 d[2].addr = (unsigned long)data + 3;
708 d[3].addr = (unsigned long)data + 10;
710 d[3].flags = VRING_DESC_F_NEXT;
712 d[4].addr = (unsigned long)data + 15;
714 d[4].flags = VRING_DESC_F_NEXT;
716 d[5].addr = (unsigned long)data + 21;
720 /* Host picks it up (allocates new iov). */
721 vringh_iov_init(&riov, host_riov, ARRAY_SIZE(host_riov));
722 vringh_iov_init(&wiov, host_wiov, ARRAY_SIZE(host_wiov));
724 err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
726 errx(1, "vringh_getdesc_user: %i", err);
729 errx(1, "vringh_getdesc_user: head %i not 0", head);
731 assert(riov.max_num & VRINGH_IOV_ALLOCATED);
732 if (getrange != getrange_slow)
733 assert(riov.used == 7);
735 assert(riov.used == 28);
736 err = vringh_iov_pull_user(&riov, buf, 29);
739 /* Data should be linear. */
740 for (i = 0; i < err; i++)
742 vringh_iov_cleanup(&riov);
745 /* Don't leak memory... */
746 vring_del_virtqueue(vq);
747 free(__user_addr_min);