1 // SPDX-License-Identifier: GPL-2.0
2 /* Simple test of virtio code, entirely in userpsace. */
6 #include <linux/kernel.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>
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;
27 static bool never_notify_host(struct virtqueue *vq)
32 static void never_callback_guest(struct virtqueue *vq)
37 static bool getrange_iov(struct vringh *vrh, u64 addr, struct vringh_range *r)
39 if (addr < (u64)(unsigned long)__user_addr_min - user_addr_offset)
41 if (addr >= (u64)(unsigned long)__user_addr_max - user_addr_offset)
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;
50 /* We return single byte ranges. */
51 static bool getrange_slow(struct vringh *vrh, u64 addr, struct vringh_range *r)
53 if (addr < (u64)(unsigned long)__user_addr_min - user_addr_offset)
55 if (addr >= (u64)(unsigned long)__user_addr_max - user_addr_offset)
59 r->end_incl = r->start;
60 r->offset = user_addr_offset;
64 struct guest_virtio_device {
65 struct virtio_device vdev;
67 unsigned long notifies;
70 static bool parallel_notify_host(struct virtqueue *vq)
73 struct guest_virtio_device *gvdev;
75 gvdev = container_of(vq->vdev, struct guest_virtio_device, vdev);
76 rc = write(gvdev->to_host_fd, "", 1);
83 static bool no_notify_host(struct virtqueue *vq)
88 #define NUM_XFERS (10000000)
90 /* We aim for two "distant" cpus. */
91 static void find_cpus(unsigned int *first, unsigned int *last)
97 for (i = 0; i < 4096; i++) {
101 if (sched_setaffinity(getpid(), sizeof(set), &set) == 0) {
110 /* Opencoded version for fast mode */
111 static inline int vringh_get_head(struct vringh *vrh, u16 *head)
116 err = get_user(avail_idx, &vrh->vring.avail->idx);
120 if (vrh->last_avail_idx == avail_idx)
123 /* Only get avail ring entries after they have been exposed by guest. */
124 virtio_rmb(vrh->weak_barriers);
126 i = vrh->last_avail_idx & (vrh->vring.num - 1);
128 err = get_user(*head, &vrh->vring.avail->ring[i]);
132 vrh->last_avail_idx++;
136 static int parallel_test(u64 features,
137 bool (*getrange)(struct vringh *vrh,
138 u64 addr, struct vringh_range *r),
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;
148 /* Create real file to mmap. */
149 fd = open("/tmp/vringh_test-file", O_RDWR|O_CREAT|O_TRUNC, 0600);
151 err(1, "Opening /tmp/vringh_test-file");
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);
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);
168 find_cpus(&first_cpu, &last_cpu);
169 printf("Using CPUS %u and %u\n", first_cpu, last_cpu);
174 int status, err, rlen = 0;
177 /* We are the host: never access guest addresses! */
178 munmap(guest_map, mapsize);
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);
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);
195 while (xfers < NUM_XFERS) {
196 struct iovec host_riov[2], host_wiov[2];
197 struct vringh_iov riov, wiov;
202 err = vringh_get_head(&vrh, &head);
205 err = vringh_need_notify_user(&vrh);
207 errx(1, "vringh_need_notify_user: %i",
210 write(to_guest[1], "", 1);
215 errx(1, "vringh_get_head");
219 vringh_iov_init(&riov,
221 ARRAY_SIZE(host_riov));
222 vringh_iov_init(&wiov,
224 ARRAY_SIZE(host_wiov));
226 err = vringh_getdesc_user(&vrh, &riov, &wiov,
230 err = vringh_need_notify_user(&vrh);
232 errx(1, "vringh_need_notify_user: %i",
235 write(to_guest[1], "", 1);
239 if (!vringh_notify_enable_user(&vrh))
242 /* Swallow all notifies at once. */
243 if (read(to_host[0], buf, sizeof(buf)) < 1)
246 vringh_notify_disable_user(&vrh);
251 errx(1, "vringh_getdesc_user: %i", err);
253 /* We simply copy bytes. */
255 rlen = vringh_iov_pull_user(&riov, rbuf,
258 errx(1, "vringh_iov_pull_user: %i",
260 assert(riov.i == riov.used);
263 err = vringh_iov_push_user(&wiov, rbuf, rlen);
265 errx(1, "vringh_iov_push_user: %i",
267 assert(wiov.i == wiov.used);
273 err = vringh_complete_user(&vrh, head, written);
275 errx(1, "vringh_complete_user: %i", err);
278 err = vringh_need_notify_user(&vrh);
280 errx(1, "vringh_need_notify_user: %i", err);
282 write(to_guest[1], "", 1);
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);
293 struct guest_virtio_device gvdev;
294 struct virtqueue *vq;
296 struct vring_desc *indirects;
297 unsigned int finished = 0;
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);
303 /* We are the guest. */
304 munmap(host_map, mapsize);
309 gvdev.vdev.features = features;
310 gvdev.to_host_fd = to_host[1];
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);
317 vq = vring_new_virtqueue(0, RINGSIZE, ALIGN, &gvdev.vdev, true,
319 fast_vringh ? no_notify_host
320 : parallel_notify_host,
321 never_callback_guest, "guest vq");
323 /* Don't kfree indirects. */
324 __kfree_ignore_start = indirects;
325 __kfree_ignore_end = indirects + RINGSIZE * 6;
327 while (xfers < NUM_XFERS) {
328 struct scatterlist sg[4];
329 unsigned int num_sg, len;
331 bool output = !(xfers % 2);
334 while ((dbuf = virtqueue_get_buf(vq, &len)) != NULL) {
336 assert(*dbuf == finished - 1);
337 else if (!fast_vringh)
338 assert(*dbuf == finished);
342 /* Produce a buffer. */
343 dbuf = data + (xfers % (RINGSIZE + 1));
350 switch ((xfers / sizeof(*dbuf)) % 4) {
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);
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);
364 sg_init_table(sg, num_sg = 1);
365 sg_set_buf(&sg[0], (void *)dbuf, 4);
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);
376 /* May allocate an indirect, so force it to allocate
378 __kmalloc_fake = indirects + (xfers % RINGSIZE) * 4;
380 err = virtqueue_add_outbuf(vq, sg, num_sg, dbuf,
383 err = virtqueue_add_inbuf(vq, sg, num_sg,
386 if (err == -ENOSPC) {
387 if (!virtqueue_enable_cb_delayed(vq))
389 /* Swallow all notifies at once. */
390 if (read(to_guest[0], buf, sizeof(buf)) < 1)
394 virtqueue_disable_cb(vq);
399 errx(1, "virtqueue_add_in/outbuf: %i", err);
406 while (finished != xfers) {
411 dbuf = virtqueue_get_buf(vq, &len);
414 assert(*dbuf == finished - 1);
421 if (!virtqueue_enable_cb_delayed(vq))
423 if (read(to_guest[0], buf, sizeof(buf)) < 1)
427 virtqueue_disable_cb(vq);
430 printf("Guest: notified %lu, pinged %lu\n",
431 gvdev.notifies, receives);
432 vring_del_virtqueue(vq);
437 int main(int argc, char *argv[])
439 struct virtio_device vdev;
440 struct virtqueue *vq;
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];
451 bool (*getrange)(struct vringh *vrh, u64 addr, struct vringh_range *r);
452 bool fast_vringh = false, parallel = false;
454 getrange = getrange_iov;
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)
468 else if (strcmp(argv[1], "--parallel") == 0)
471 errx(1, "Unknown arg %s", argv[1]);
476 return parallel_test(vdev.features, getrange, fast_vringh);
478 if (posix_memalign(&__user_addr_min, PAGE_SIZE, USER_MEM) != 0)
480 __user_addr_max = __user_addr_min + USER_MEM;
481 memset(__user_addr_min, 0, vring_size(RINGSIZE, ALIGN));
483 /* Set up guest side. */
484 vq = vring_new_virtqueue(0, RINGSIZE, ALIGN, &vdev, true, false,
486 never_notify_host, never_callback_guest,
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);
494 /* No descriptor to get yet... */
495 err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
497 errx(1, "vringh_getdesc_user: %i", err);
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];
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);
512 errx(1, "virtqueue_add_sgs: %i", err);
513 __kmalloc_fake = NULL;
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));
519 err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
521 errx(1, "vringh_getdesc_user: %i", err);
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);
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);
538 err = vringh_iov_pull_user(&riov, buf, 5);
540 errx(1, "vringh_iov_pull_user: %i", err);
541 assert(buf[0] == 'a');
543 assert(vringh_iov_pull_user(&riov, buf, 5) == 0);
545 memcpy(buf, "bcdef", 5);
546 err = vringh_iov_push_user(&wiov, buf, 5);
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);
554 err = vringh_complete_user(&vrh, head, err);
556 errx(1, "vringh_complete_user: %i", err);
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);
563 errx(1, "virtqueue_get_buf: %p", ret);
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);
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;
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);
581 errx(1, "virtqueue_add_outbuf (large): %i", err);
582 __kmalloc_fake = NULL;
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));
588 err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
590 errx(1, "vringh_getdesc_user: %i", err);
592 assert(riov.max_num & VRINGH_IOV_ALLOCATED);
593 assert(riov.iov != host_riov);
594 if (getrange != getrange_slow)
595 assert(riov.used == RINGSIZE);
597 assert(riov.used == RINGSIZE * USER_MEM/4);
599 assert(!(wiov.max_num & VRINGH_IOV_ALLOCATED));
600 assert(wiov.used == 0);
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));
611 assert(riov.i == riov.used);
612 vringh_iov_cleanup(&riov);
613 vringh_iov_cleanup(&wiov);
615 /* Complete using multi interface, just because we can. */
618 err = vringh_complete_multi_user(&vrh, used, 1);
620 errx(1, "vringh_complete_multi_user(1): %i", err);
622 /* Free up those descriptors. */
623 ret = virtqueue_get_buf(vq, &i);
625 errx(1, "virtqueue_get_buf: %p", ret);
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);
633 errx(1, "virtqueue_add_outbuf (multiple): %i", err);
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));
640 for (i = 0; i < RINGSIZE; i++) {
641 err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
643 errx(1, "vringh_getdesc_user: %i", err);
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);
651 errx(1, "vringh_complete_multi_user: %i", err);
653 /* Free those buffers. */
654 for (i = 0; i < RINGSIZE; i++) {
656 assert(virtqueue_get_buf(vq, &len) != NULL);
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;
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,
670 never_callback_guest,
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);
679 err = virtqueue_add_outbuf(vq, guest_sg, 4, &err, GFP_KERNEL);
681 errx(1, "virtqueue_add_outbuf (indirect): %i", err);
683 vring_init(&vring, RINGSIZE, __user_addr_min, ALIGN);
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;
695 d[0].addr = (unsigned long)data;
697 d[0].flags = VRING_DESC_F_NEXT;
699 d[1].addr = (unsigned long)data + 1;
703 /* Second indirect */
704 d[2].addr = (unsigned long)data + 3;
709 d[3].addr = (unsigned long)data + 10;
711 d[3].flags = VRING_DESC_F_NEXT;
713 d[4].addr = (unsigned long)data + 15;
715 d[4].flags = VRING_DESC_F_NEXT;
717 d[5].addr = (unsigned long)data + 21;
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));
725 err = vringh_getdesc_user(&vrh, &riov, &wiov, getrange, &head);
727 errx(1, "vringh_getdesc_user: %i", err);
730 errx(1, "vringh_getdesc_user: head %i not 0", head);
732 assert(riov.max_num & VRINGH_IOV_ALLOCATED);
733 if (getrange != getrange_slow)
734 assert(riov.used == 7);
736 assert(riov.used == 28);
737 err = vringh_iov_pull_user(&riov, buf, 29);
740 /* Data should be linear. */
741 for (i = 0; i < err; i++)
743 vringh_iov_cleanup(&riov);
746 /* Don't leak memory... */
747 vring_del_virtqueue(vq);
748 free(__user_addr_min);