/etc/init.d/halt.sh fails on volume groups called "root"$SOMETHING, e.g. rootvg (common name vor VGs e.g. under AIX). Furthermore, /etc/init.d/halt.sh tries to umount /, although it explicitely tries to avoid that. The offending line is: for x in $(awk '!/(^#|proc|devfs|tmpfs|^none|^\/dev\/root|[[:space:]]\/[[:space:]])/ {print $2}' /proc/mounts |sort -r) Perhaps it would be better written as: for x in $(awk '!/(^#|proc|devfs|tmpfs|^none|[[:space:]]\/[[:space:]])/ && ($2 != "/") {print $2}' /proc/mounts |sort -r) Reproducible: Always Steps to Reproduce: 1. Create a VG named "rootvg". 2. Create a logical volume within ths VG. 3. Mount this volume. 4. Reboot. Actual Results: The VG is not shut down properly. Expected Results: Well, umount all LVs within the VG. Portage 2.0.47-r10 (default-x86-1.4, gcc-3.2.2, glibc-2.3.1-r4) ================================================================= System uname: 2.4.20-xfs i686 AMD Athlon(TM) XP2200+ GENTOO_MIRRORS="http://www.fhh.opensource-mirror.de/gentoo.org/ http://linux.rz.ruhr-uni-bochum.de/download/gentoo-mirror/ http://gentoo.oregonstate.edu/ http://www.ibiblio.org/pub/Linux/distributions/gentoo" CONFIG_PROTECT="/etc /var/qmail/control /usr/kde/2/share/config /usr/kde/3/share/config /usr/X11R6/lib/X11/xkb /usr/kde/3.1/share/config /usr/share/config" CONFIG_PROTECT_MASK="/etc/gconf /etc/env.d" PORTDIR="/usr/portage" DISTDIR="/usr/portage/distfiles" PKGDIR="/usr/portage/packages" PORTAGE_TMPDIR="/var/tmp" PORTDIR_OVERLAY="" USE="slang -3dfx 3dnow -aalib acl acpi -afs alsa -apache2 -apm arts -atlas avi berkdb -bonobo -canna cdr -cjk crypt cups -directfb -dga -doc -dvb dvd encode -esd ethereal -evo fbcon flash -freewnn -gb gd gdbm -ggi gif -gnome gphoto2 gpm -gps -gtk -guile -icc imap imlib innodb -ipv6 java jikes jpeg -junit kde -kerberos lcms -ldap -leim -libg++ -libgda libwww lirc -matrix -maildir mbox mikmod mmx motif mozilla mpeg -mule mysql -nas ncurses nls nocardbus -oav -oci8 odbc oggvorbis opengl oss pam -pcmcia pda pdflib perl pic plotutils png -pnp -postgres python qt quicktime readline -ruby samba sasl scanner sdl -slp snmp -socks5 spell sse ssl -static -svga tcltk tcpd tetex tiff truetype -trusted usb -voodoo3 -wavelan wmf X xface xml xml2 xmms xv zlib x86 xft" COMPILER="gcc3" CHOST="i686-pc-linux-gnu" CFLAGS="-march=athlon-xp -O3 -pipe -fomit-frame-pointer" CXXFLAGS="-march=athlon-xp -O3 -pipe -fomit-frame-pointer" ACCEPT_KEYWORDS="x86" MAKEOPTS="-j2" AUTOCLEAN="yes" SYNC="rsync://rsync.de.gentoo.org/gentoo-portage/" FEATURES="sandbox ccache"
Does your changes work ?
Actually, I did fix this already. Could you try baselayout-1.8.6.5, or change that bit to: ----------------------------------------------------------- # Try to unmount all filesystems (no /proc,tmpfs,devfs,etc). # This is needed to make sure we dont have a mounted filesystem # on a LVM volume when shutting LVM down ... ebegin "Unmounting filesystems" for x in `mount | awk '{ if (($5 !~ /(proc|sysfs|devfs|tmpfs)/) && ($1 !~ /^rootfs|^\/dev\/root/) && ($3 != "/")) print $3 }' | sort -r` do umount -f -r ${x} &> /dev/null done eend 0 --------------------------------------------------------------- And let me know if it works ? Thanks.
Ths cannot work either (at least not with a volume group named root$SOMETHING), because ...($1 !~ /^rootfs|^\/dev\/root/)... matches rootvg. As I said, me and two of my colleagues ran into that - we are simply used to call our first volume group "rootvg", it is what AIX does. Furthermore, I think it is a very bad idea to stick to the output of mount - /etc/mtab is not necessarily up to date all of the time, think about chrooted mounts. I did _not_ try the new baselayout yet, because I thought the awk-statement you told me was an excerpt from that. If the new baselayout contains _another_ awk-statement, please let me know. But with that new statement, it doesn't try to umount / too early, so one thing is gone.
It does not match it. If it matches, it will not be in the list to unmount: ---------------------------------------------------------------------------------- workshop root # echo rootvg | awk '{ if ($1 !~ /(^rootfs|^\/dev\/root)/) print }' rootvg workshop root # echo rootfs | awk '{ if ($1 !~ /(^rootfs|^\/dev\/root)/) print }' workshop root # ---------------------------------------------------------------------------------- Also, it check if $3 is "/", and do not add it to the umount list if so. Then, why I do not use /proc/mounts, is because it does not have correct filesystem type for 'bind' mounts, so in a case like this: ----------------------------------------------------------------------- workshop root # mount --bind /dev/ /tmp/ workshop root # mount | grep tmp none on /dev/shm type tmpfs (rw) /dev on /tmp type none (rw,bind) workshop root # cat /proc/mounts | grep tmp none /dev/shm tmpfs rw 0 0 none /tmp devfs rw 0 0 workshop root # ----------------------------------------------------------------------- /tmp do not get unmounted ... Could you at least try the bit in comment #2 ?
Well, ok, I tried that bit - as expected, it did _not_ work, devices on /dev/rootvg didn't get umounted. You chose a bad example to demonstrate, of course it is "/dev/rootvg" which is confused with "/dev/root" and not only "rootvg", see also: foerstes@abyss[~]$ echo /dev/rootvg | awk '{ if ($1 !~ /(^rootfs|^\/dev\/root)/) print }' foerstes@abyss[~]$ Well, why make a loop around these umounts? [tmpfs-stuff, devfs] umount -ttmpfs -a -r [vg-stuff] umount -tnoproc -a -r -v
Ah, ok, missed that, sorry =) Below should fix, please try and let me know. --------------------------------------------------------------------- ebegin "Unmounting filesystems" for x in `mount | awk '{ if (($5 !~ /^(proc|sysfs|devfs|tmpfs)$/) && ($1 !~ /^(rootfs|\/dev\/root)$/) && ($3 != "/")) print $3 }' | sort -r` do umount -f -r ${x} &> /dev/null done ----------------------------------------------------------------------
Yes, this works like it is supposed to do. This means that the new baselayout-package will fix all the bugs of the initial bug report (although I will slightly change these lines for private use). Since using the output of "mount" leaves us with mounts not recorded in /etc/mtab (e.g. mounted in a chroot, / was mountd ro, ...) not being umounted, I wonder if I should open a new bug or if this issue could be included here.
Like I said, using is /proc/mounts is a problem. You cannot detect --bind mounts, so needed things do not get unmounted, as they are seen as /, etc. Is this still lvm releated ? Can't we rather use /proc/lvm to check what volumes is still active, and then check what is mounted in there, and try to unmount those ?
Summary: 1. You can not trust /etc/mtab (chroot, mount -n, / was mounted ro). 2. /proc/mount cannot be trusted because fstype is displayed wrongly when using --bind. 3. Loopback-Mounts can also be --bind'ed. 4. You can not umount /dev when it is on devfs. Would this solved by: a) Send all processes the term and then the kill signal. b) Execute "umount -a -O bind". c) Go through /proc/mounts in reverse order. For every mount which is not /, /dev or /proc: umount /mountpoint, not /device (yes, umount will work with mountpoints not recorded in /etc/mtab). --> We should remain with an active LVM, / and perhaps /dev as devfs mounted. d) Shut lvm down. e) Remount / readonly. [RAID-stuff, UPS] This solves 1.) by using /proc/mounts, 2.) by the fact that we are umounting "--bind"ed mounts regardless of the filesystem-type as long as they are not mounted on / or /dev (and we do not get fooled, because we use the mountpoint, not the device, i.e. we won't try to "umount /" if somebody did a "mount --bind / /foo"), 3.) is solved by the same thing as 2.) and is further simplfied because we call killall5 earlier and also resolves the issue with 4.). Why were you explicitely excluding _all_ mounts of type tmpfs, devfs and procfs? But this is no longer LVM-related, it's more like a complete rewrite of /etc/init.d/halt.sh. If you want me to write this code, just say so. I have no problem with eschaning some scripts on my system, but I won't emerge complete packages which are masked/unstable, sorry.
Umm, well, I meant: [...] 4. You can not umount /dev when it is on devfs, but you can mount every devfs except that from the shutdown script if it is not referenced any longer. [...]
1) "umount -a -O bind" will still fail if it was mounted in chroot, as /proc/mounts do not record that info. 2) halt.sh does try very early (before deactivating swap) to unmount all tmpfs mounts not in use. Thus, because we use tmpfs for /mnt/.init.d, we do not want to force unmount of tmpfs mounts (and the Adelie server stuff need it to be present till the last). What you propose, is still not 100% fail safe. A better approach will be to check /etc/mtab for critical stuff, and then go through /proc/mounts and if its mount point do not match a critical mount, try to unmount: ------------------------------------------------------------------ ebegin "Unmounting filesystems" no_unmount="`mount | awk '{ if (($5 ~ /^(proc|sysfs|devfs|tmpfs)$/) || ($1 ~ /^(rootfs|\/dev\/root|none)$/) || ($3 = "/")) print $3 }' | uniq`" for x in `awk '{ print $2 }' < /proc/mounts | sort -r` do do_unmount="yes" for y in ${no_unmount} do [ "${x}" = "${y}" ] && do_unmount="no" done [ "${do_unmount}" = "yes" ] && umount -f -r ${x} &> /dev/null done eend 0
To your point 1.), "mount -a -O bind" might fail: It doesn't matter if it is successfull, you don't need it at all. However, what you propose is more or less what I said, so this should work reasonably well (although I still think you can ignore any mountpoints except /, /dev and perhapts /mnt/.init.d, I've tested this approach today with a wide variety of mounts and it works fully well for me). It's a "close this bug" as far as I'm concerned.
Ok, thanks for all the input to get this done. BTW: the 'unstable' baselayout is already well tested, there are only a few issues left ... if you want, have a look at bug #14946 .... the change in hostname breaks a few things like apache2, etc. Some input on how it could be fixed will be nice ... also have a look at bug #18158 and bug #18801.
