[linux-lvm] LVM 1.0.4 pvmove is broken

[Poul Petersen]

> It is well known that pvmove of live volumes in LVM1 is broken for a
> couple of reasons, the most obvious being the deadlock under memory
> pressure that you are experiencing.

	If it is so well know, should it be in the known_bugs for 1.0.4? I
saw
the following entries:

- There is a very slight change that pvmove on a live system can
  corrupt your data.  A pvmove on an offline (ie. not mounted) LV
  should be fine.

- There is also a chance under high load that a live pvmove will
  deadlock your system.

	However, what I have been seeing is not a "very slight chance"; it
fails
every time. Also, whether or not the LV is mounted doesn't matter - in all
of the
failures I have seen the LV was *not* mounted. The three critical failure
elements seem 
to be:

	1) use the LVM 1.0.4 Makefile generated patch for 2.4.18
	2) use 1.0.4 pvmove tool.
	3) Use an active VG, that is vgchange -ay

	As long as I do all three things, it will fail every time without
question 
on three different machines that I have tested. If I change any element, say
use the
1.0.3 toolset or use 1.0.4 but simply don't apply the Makefile generated
patch, then
it works fine. Is this still the same problem? Load is also not in question
- the
three machines I have tested were all in a lab environment, so there is no
concern 
about system load, lack of memory, etc. I guess what I am seeing is that if
I follow
the installation directions for 1.0.4 (apply the patch etc) then pvmove is
completely
broken. Is this right? Obviously I wouldn't want to have to deactivate the
entire VG
to move PEs around.

> I would encourage you to try Stephen Tweedies patch and post your
> findings.

	I will try that... Thanks for the pointer, I evidently missed that
post while
searching through the archives before. If pvmove is as broken as his post
suggests, I'll
probably just stay away from it - I only brought up this issue because the
way in which
I was seeing pvmove fail incessantly suggested that there might be a simple
fix.

> In the longer term the solution is to move to LVM2; we are currently
> working on pvmove and will have it in the next beta (beginning of
> June), pvmove is the last major feature from LVM1 that LVM2 doesn't
> yet support.

	Speaking of LVM2, will the transition to LVM2 require recreating the
VGs?


Thanks,

-poul