[linux-lvm] Unexptected filesytem unmount with thin provision and autoextend disabled - lvmetad crashed?
zkabelac at redhat.com
Tue May 17 13:48:46 UTC 2016
On 17.5.2016 15:09, Gionatan Danti wrote:
>> Well yeah - ATM we rather take 'early' action and try to stop any user
>> on overfill thin-pool.
> It is a very reasonable standing
>> Basically whenever 'lvresize' failed - dmeventd plugin now tries
>> to unconditionally umount any associated thin-volume with
>> thin-pool above threshold.
>> For now - plugin 'calls' the tool - lvresize --use-policies.
>> If this tool FAILs for ANY reason -> umount will happen.
>> I'll probably put in 'extra' test that 'umount' happens
>> with >=95% values only.
>> dmeventd itself has no idea if there is configure 100 or less - it's
>> the lvresize to see it - so even if you set 100% - and you have enabled
>> monitoring - you will get umount (but no resize)
> Ok, so the "failed to resize" error is also raised when no actual resize
> happens, but the call to the "dummy" lvresize fails. Right?
Yes - in general - you've witnessed general tool failure,
and dmeventd is not 'smart' to recognize the reason of failure.
Normally this 'error' should not happen.
And while I'd even say there could have been a 'shortcut'
without even reading VG 'metadata' - since there is profile support,
it can't be known (100% threshold) without actually reading metadata
(so it's quite tricky case anyway)
>> Well 'lvmetad' shall not crash, ATM this may kill commands - and further
>> stop processing - as we rather 'stop' further usage rather than allowing
>> to cause bigger damage.
>> So if you have unusual system/device setup causing 'lvmetad' crash -
>> open BZ,
>> and meawhile set 'use_lvmetad=0' in your lvm.conf till the bug is fixed.
> My 2 cents are that the last "yum upgrade", which affected the lvm tools,
> needed a system reboot or at least the restart of the lvm-related services
> (dmeventd and lvmetad). The strange thing is that, even if lvmetad crashed, it
> should be restartable via the lvm2-lvmetad.socket systemd unit. Is this a
> wrong expectation?
Assuming you've been bitten by this one:
possibly? targeted by this commit:
More information about the linux-lvm