[Bug 237397] 'zfs mount -a' mounts filesystems in incorrect order

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=237397

            Bug ID: 237397
           Summary: 'zfs mount -a' mounts filesystems in incorrect order
           Product: Base System
           Version: 12.0-STABLE
          Hardware: amd64
                OS: Any
            Status: New
          Severity: Affects Only Me
          Priority: ---
         Component: kern
          Assignee: bugs at FreeBSD.org
          Reporter: freebsd-bugzilla at umpquanet.com

I have a system that boots from ZFS.  I recently upgraded it from
FreeBSD 11.2-STABLE to 12.0-STABLE.

FreeBSD electron 12.0-STABLE FreeBSD 12.0-STABLE r346293 GENERIC  amd64

I upgraded source, and rebooted into the new 12-STABLE kernel with no
problems.  After installworld and mergemaster, I discovered that certain
ZFS filesystems were not auto-mounting in the correct sequence,
"correct" meaning that /a mounts before /a/b mounts before /a/b/c.

Here are the mountable filesystems on that machine (canmount != off):

electron : 15:36:37 /root# zfs get -Hrt filesystem canmount | grep -vw off |
while read FS b; do echo $FS; done | xargs zfs list -o
name,canmount,mounted,mountpoint
NAME                           CANMOUNT  MOUNTED  MOUNTPOINT
electron                             on       no  none
electron/data                        on       no  /data
electron/data/backup                 on       no  /data/backup
electron/data/backup/jimsdesk        on       no  /data/backup/jimsdesk
electron/system                      on       no  none
electron/system/ROOT                 on       no  none
electron/system/ROOT/default         on      yes  /
electron/system/home                 on      yes  /home
electron/system/tmp                  on      yes  /tmp
electron/system/usr                  on      yes  /usr
electron/system/usr/local            on      yes  /usr/local
electron/system/var                  on      yes  /var

The easiest tree to demonstrate with is electron/data and children, although
this problem is also happening with electron/system/usr and children.

I have not found any way to query the ZFS stack as to what it thinks the
correct mount order should be, and thus demonstrate whether the
automatic mounting order is correct.  However, I have fashioned this
test script to highlight the observable differences between a correct,
manual mounting, and the incorrect mount order used by zfs mount -a:


--snip--
#!/usr/bin/env bash

# correct mount ordering:
MANUAL="
        zfs mount electron/data
        zfs mount electron/data/backup
        zfs mount electron/data/backup/jimsdesk
"

# automatic mount ordering:
AUTO="
        zfs mount -va
"


case $1
in
   manual)
        MOUNT="$MANUAL"
        ;;

   auto)
        MOUNT="$AUTO"
        ;;
   *)
        exit
esac


cat << EOF | sh -v
        zfs mount | grep '/data'
        mkdir /data
        find  /data
$MOUNT
        zfs mount | grep '/data'
        find /data | head -3000000 | wc -l
        zfs umount /data/backup/jimsdesk
        zfs umount /data/backup
        zfs umount /data
        rmdir /data
        find  /data
EOF
--snip--

Here's a run of that script using the manual mounting order.  The
mountpoint /data is created fresh at the start.  When the filesystems
are correctly mounted, there are over 3 million files under /data.
After unmounting, /data is empty, and can be rmdir'ed.

--snip--
electron : 12:26:00 /root# ./zfs-mount-test.sh manual
        zfs mount | grep '/data'
        mkdir /data
        find  /data
/data

        zfs mount electron/data
        zfs mount electron/data/backup
        zfs mount electron/data/backup/jimsdesk

        zfs mount | grep '/data'
electron/data                   /data
electron/data/backup            /data/backup
electron/data/backup/jimsdesk   /data/backup/jimsdesk
        find /data | head -3000000 | wc -l
 3000000
        zfs umount /data/backup/jimsdesk
        zfs umount /data/backup
        zfs umount /data
        rmdir /data
        find  /data
find: /data: No such file or directory
--snip--

Here's a run of that script using the automatic mounting order.  When
the filesystems are automatically mounted, fewer than 1.5 million files
are present under /data.  Unmounting everything generates an error
message, /data is not clean, and rmdir returns an error.

--snip--
electron : 12:26:17 /root# ./zfs-mount-test.sh auto
        zfs mount | grep '/data'
        mkdir /data
        find  /data
/data

        zfs mount -va

        zfs mount | grep '/data'
electron/data                   /data
electron/data/backup            /data/backup
electron/data/backup/jimsdesk   /data/backup/jimsdesk
        find /data | head -3000000 | wc -l
 1481445
        zfs umount /data/backup/jimsdesk
        zfs umount /data/backup
cannot unmount '/data/backup': not a mountpoint
        zfs umount /data
        rmdir /data
rmdir: /data: Directory not empty
        find  /data
/data
/data/backup
--snip--


Please let me know how I can provide additional information or testing.

Thank you!

Jim

-- 
You are receiving this mail because:
You are the assignee for the bug.