tools/virtio/vringh_test.c

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