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