[Cooker] Re: [PATCH REVIEW] mkinitrd, lvm, UUID

[Andrey Borzenkov]

Andreas Hasenack wrote:

> This is about https://bugzilla.mandriva.com/show_bug.cgi?id=36457
> 
> Summary:
> a) lvm tools don't work on UUID=... style devices
>   -> "lvdisplay UUID=c52cc68a-fcff-4a11-b034-31023bf5025c" won't work
> 
> b) lvm tools don't work with devices outside /dev/mapper/* or
>   /dev/<volumegroupname>/*
>   -> "lvdisplay /dev/dm-5" won't work, even if /dev/dm-5 is a literal
>       copy of /dev/mapper/vg0-root for example
> 
> c) nash's mount command, used in the initrd, doesn't do mount by UUIDs
>   -> "mount -t ext3 -o ro UUID=c52cc68a-fcff-4a11-b034-31023bf5025c
>   /sysroot" won't work
> 
> d) mkinitrd was doing stat(1) on some UUID=... "devices" instead of the
>    real device. stat(1) doesn't support UUID.
> 
> 
> [b] is working as intended, according to the #lvm IRC channel:
> <agk_>   dmsetup uses the device-mapper namespace to refer to devices
> <agk_>   it does not use /dev
> <ahasenack>      ok, that's why "dmsetup info vg0-root" works
> <ahasenack>      but not /dev/vg0/root, at most only /dev/mapper/vg0-root
> <agk_>   if someone sent me a patch to enhance dmsetup to accept /dev
> names, I'd accept it
> <agk_>   but noboduy has yet
> <agk_>   it would involve looking up the major/minor of the supplied
> /dev/dm-5 inode
> <ahasenack>      and searching for a match in the device-mapper namespace
> <agk_>   then changing the command into 'dmsetup -j <major> -m <minor>
> ...'
> <ahasenack>      is this namespace the same issue why lvdisplay doesn't
> work on /dev/dm-5?
> <agk_>   lvdisplay uses a different namespace agin
> <agk_>     vg/lv
> <agk_>   "/dev" is an optional prefix that lvm2 simply deletes
> <agk_>   so lvdisplay /dev/vg/lv  becomes lvdisplay vg/lv  which is
> processed
> <agk_>   - the arg is not treated as a file
> <ahasenack>      and since dm-5 is not a logical volume name, it fails
> <agk_>   we added special processing so /dev/mapper/* would work
> 
> 
> Bottom line: if you use LVM on / and mount by UUID (new with 2008.1),
> your system won't boot.
> 
> 
> Attached are patches to try to workaround this:
> 
> - nash-mount-by-uuid.patch: adds mount-by-uuid to nash's builtin mount
>   command (fixing [c])

The getpathbyspec() part looks fine; I missed this one because at the time
mkswap did not support UUID so resume=UUID= did not work at all.

I do not quite like mount changes. mkrootdev already supports UUID. Just let
rootdev=/dev/root if fstabrootdev=(LABEL|UUID) for all cases; nash will
handle it automatically already. Actually this will allow you to remove the
LABEL block in mountCommand.

> - mkinitrd-4.2.17-uuid_lvm.patch: deals with [a], [b] and [d]
> 

looks fine with one comment - you can actually set different prefix for LVM
namespace like

    # Where do you want your volume groups to appear ?
    dir = "/dev/lvm"

so this will be not valid:

+        rootdev=/dev/mapper/$rootvg-`echo $lv | cut -d : -f 2`

because mapper goes inside of dir:

{pts/0}% sudo lvdisplay
  --- Logical volume ---
  LV Name                /dev/lvm/test/v1
..
{pts/0}% LC_ALL=C ll /dev/mapper /dev/lvm/mapper
/dev/lvm/mapper:
total 0
crw------- 1 root root  10, 62 Dec 29 17:20 control
brw------- 1 root root 253,  0 Jan  4 22:12 test-v1
brw------- 1 root root 253,  1 Jan  4 22:36 test-v2

/dev/mapper:
total 0
crw-rw---- 1 root root 10, 62 Dec 29 17:19 control

personally I consider this a bug - dir should affect only volume names
not /dev/mapper - but well ... and IMHO placing VG name directly under /dev
is wrong as well (what about having group zero or like?)

> The good: my system boots again, and a plain installation (without lvm)
> works too.
> The bad: didn't test RAID on / yet, nor LVM1 (and I won't test LVM1).
> The ugly: feels sort of hackish, we should find a way to deal with all
> these "dm"s in one good way. Who knows what other surprises are there.