userland: avl is broken (was: libzfs export fails to umount all filesystems)

[lundman]

During zpool export, the userland libzfs iterates the mounted filesystems to issue zfs_umount. However the mnttab logic is empty.

libzfs_dataset.c:

libzfs_mnttab_add(libzfs_handle_t *hdl, const char *special,
    const char *mountp, const char *mntopts)
{   
    mnttab_node_t *mtn;

    printf("mnttab ADDING '%s'\n", mountp);

# ./zpool.sh create BOOM pool-image.bin
mnttab ADDING 'BOOM'

on export;

libzfs_mount.c:

    printf("disable datasets\n");

    namelen = strlen(zhp->zpool_name);

    used = alloc = 0;
    for (mntn = libzfs_mnttab_first(hdl); mntn != NULL;
         mntn = libzfs_mnttab_next(hdl, mntn)) {
        struct mnttab *mt = &mntn->mtn_mt;

        printf("disable_datasets: '%s'\n",mt->mnt_mountp);

#./zpool.sh export BOOM
disable datasets

Alas, even though we add BOOM to mnttab, the call libzfs_mnttab_first() returns NULL.

[lundman]

After adding a new mount, if I can avl_first, it works as expected.
When it comes in from zpool export, avl_first returns NULL.

I am unsure where AVL are stored, on disk? part of the dataset?

There is a temporary replacement using getfsinfo() in place until it can be fixed.

[lundman]

12db36a

Incidentally, the libzfs_mnttab code essentially call mnttab (or rather getmntinfo() on OSX) and builds an avl_list of the mounted file systems, so the temporary work around was not all that bad. But now the greater libzfs_mnttab functions all work as expected and all temporary work-around have been rolled-back.