Finish fix for bug 607347: failing both redundant mirror log legs...
A previous check-in added logic to handle the case where both images
of a mirrored log failed. It solved the problem by simply removing
the log entirely - leaving the parent mirror with a 'core' log. This
worked for most cases. However, if there was a small delay between
the failures of the two mirrored log devices, the mirror would hang,
LVM would hang, and no additional LVM commands could be issued.
When the first leg of the log fails, it signals the need for repair.
Before 'lvconvert --repair' is run by dmeventd, the second leg fails.
'lvconvert' would see both devices as failed and try to remove the
log entirely. When it came time to suspend the parent mirror to
update the configuration, the suspend would hang because it couldn't
get any I/O through the mirrored log, which was plugged waiting for
corrective action. The solution is to replace the log with an error
target to clear any pending writes before removing it. This allows
the parent mirror to suspend and make the proper changes.
Pass metadataignore to pv_create, pv_setup, _mda_setup, and add_mda.
Pass metadataignore through PV creation / setup paths.
As a result of this cleanup, we can remove the unnecessary setting
of mda_ignore bits inside pvcreate_single(), after call to pv_create.
For now, just set metadataignore to '0' in some places. This is
equivalent to the prior functionality, although the 0 is given
by the caller not hardcoded in _mda_setup() call.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Cleanups for configure:
Indent updates.
Use AC_HELP_STRING for help string.
Start help string with lower letter.
Add [] around some default values i.e. [TYPE=internal].
Skip some "" around shell assigment when not needed.
Fix typo --with-device-gid=UID string.
Add prompt if using --metadataignore argument with vgmetadatacopies.
When using vgmetadatacopies value other than "umanaged" (0), prompt
the user if the usage of --metadataignore would change the value of
vgmetadatacopies. The main 2 cases are:
1) pvchange --metadataignore
2) vgextend --metadataignore
We leave the prompt check in the tools, and do not change anything
if the user says 'n'.
Examples:
vgextend --metadataignore y vgtest /dev/loop0
Setting metadataignore will override preferred number of copies of VG vgtest metadata.
Are you sure? [y/n]: y
No physical volume label read from /dev/loop0
Physical volume "/dev/loop0" successfully created
Volume group "vgtest" successfully extended
pvchange --metadataignore y /dev/loop3
Setting metadataignore on /dev/loop3 will override preferred number of copies of VG vgtest metadata.
Are you sure? [y/n]: y
WARNING: Changing preferred number of copies of VG vgtest metadata from 3 to 2
Physical volume "/dev/loop3" changed
1 physical volume changed / 0 physical volumes not changed
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Test failed commit of mda on new pv - failed vgextend.
Test the auto-repair capability when we fail committing to an mda
on a new pv adding to a vg. This test should fail until we fix
the auto-repair in this case.
Peter Rajnoha [Wed, 7 Jul 2010 11:22:46 +0000 (11:22 +0000)]
Use "nowatch" udev rule for known inappropriate devices.
For now, this is just a precaution. Normally, all the other (non-dm) rules
should check DM_UDEV_DISABLE_OTHER_RULES_FLAG and therefore avoid setting
any inotify watches as well. But let's make sure.
Support for final assignment of the "nowatch" rule (the use of ":=") will
appear in next udev release, v160. This should also work in previous udev
versions but the setting won't be sealed so any further OPTIONS="watch" will
always prevail there.
We may want to add more specific "nowatch" rules later if needed.
Adjust auto-metadata repair and caching logic to try to cope with empty mdas.
- If a PV contained empty mdas, the auto-recovery code was not kicking in.
- The 'inconsistent' state was getting lost when metadata was cached so
recovery didn't kick in. But leave the behaviour alone when using
precommitted metadata because of a warning in a confusing FIXME.
In my testing, pvs and vgs didn't repair inconsistent metadata like they
used to do. (How many other tools fail similarly now?)
And there should be no need to cache inconsistent metadata because it is
supposed to get repaired under the protection of a write lock immediately it is
discovered.
This code is in need of a redesign based on first principles.
I still see bugs in this code and this commit is risky.
Fix for bug 607347: failing both redundant mirror log legs...
Rather than attempting to remove all the images of a mirrored
log volume via remove_mirror_images, simply remove the log
if all its devices have failed.
Taka was the first to report that there is still an outstanding
issue with handling this case. I've managed to reproduce it
only very rarely, and am still working on identifying the problem.
Failing to handle the problem rarely is better than not handling
the scenario at all, so I'm checking this in.
Milan Broz [Thu, 1 Jul 2010 21:23:47 +0000 (21:23 +0000)]
Remove superfluous suspended device counter from clvmd.
Moreover, in current mirror handling, when it calls activate
on removed but suspended detached log this counter drops below zero
and confuses debug log.
Petr Rockai [Wed, 30 Jun 2010 21:40:27 +0000 (21:40 +0000)]
Maintain memlock balance in clvmd.
When a mirror is being downconverted in a cluster, a series of suspends and
resumes is executed.
With the change to using UUIDs in dev_manager instead of names, the behaviour
has changed with regards to including an _mlog in the deptree of a logical
volume. In the old (pre-UUID-enabled) code, the _mlog would appear in a deptree
of any volume purely based on a name match: a linear volume foo would include
foo_mlog in its dependencies if that happened to exist. This behaviour was
fixed and the mlog is now only included for mirrors.
By a coincidence, this mlog bug had been hiding a different bug in clvmd. When
a mirror is being dismantled (and converted to a linear volume), it is first
suspended as a whole, then later resumed in parts. Nevertheless, the overall
memlock balance is maintained in this operation. The problem kicks in, because
even though the mirror log was suspended as part of the mirror, when the
dismantled mirror is resumed again, it is no longer a mirror and therefore the
mirror log stays suspended. This would not be a problem in itself, since
_delete_lv (from metadata/mirror.c) is called on it subsequently, which does an
activate/deactivate cycle and removes the LV. The activate/deactivate cycle
correctly prompts clvmd to resume the device: however, in doing this, it will
issue an unpaired resume operation (the suspend that caused the mirror log to
be suspended is paired with resuming the dismantled mirror later). We have
concluded that the path in clvmd should never affect memlock_count, since there
should never be an unmatched explicit suspend preceding this resume.
Dave Wysochanski [Wed, 30 Jun 2010 12:17:24 +0000 (12:17 +0000)]
Add --metadataignore to pvcreate.
Allow metadataignore flag to be passed in to pvcreate.
Ideally, more refactoring of the mda allocation / initialization
is warranted, but for now, we just add another parameter to 'add_mda'
to take an existing mda ignored flag. We need to do this or pv_write
loses the state of the mda 'ignored' flag before copying and writing
to disk.
Dave Wysochanski [Tue, 29 Jun 2010 22:37:32 +0000 (22:37 +0000)]
Improve logging for metadata ignore by printing device name.
Print device name when setting or clearing metadata ignore bit.
Example:
label/label.c:160 /dev/loop2: lvm2 label detected
cache/lvmcache.c:1136 lvmcache: /dev/loop2: now in VG #orphans_lvm2 (#orphans_lvm2)
metadata/metadata.c:4142 Setting mda ignored flag for metadata_locn /dev/loop2.
format_text/text_label.c:318 Skipping mda with ignored flag on device /dev/loop2 at offset 4096
Dave Wysochanski [Tue, 29 Jun 2010 22:25:58 +0000 (22:25 +0000)]
Add some log_verbose debug statements related to metadataignore.
Logging isn't ideal, especially for mda_set_ignore. Ideally we'd
like to display the device name and offset in this case but this
requires a bit more work and a per-format 'mda_description' function
pointer definition (we don't have access to mda_context in
metadata.c).
Dave Wysochanski [Tue, 29 Jun 2010 13:29:53 +0000 (13:29 +0000)]
Attempt to fix intermittent failure with non-debug configured vgcfgbackup.
There's an intermittent failure with vgcfgbackup that seems to have been
introduced with the metadataignore / vgmetadatacopies patchset.
Intermittent failures are often the result of uninitialized data,
so this patch calls zalloc in a few places it might matter.
Dave Wysochanski [Mon, 28 Jun 2010 20:40:01 +0000 (20:40 +0000)]
Allow 'all' and 'unmanaged' values for --vgmetadatacopies.
Allowing an 'all' and 'unmanaged' value is more intuitive, and
provides a simple way for users to get back to original LVM behavior
of metadata written to all PVs in the volume group.
If the user requests "--vgmetadatacopies unmanaged", this instructs
LVM not to manage the ignore bits to achieve a specific number of
metadata copies in the volume group. The user is free to use
"pvchange --metadataignore" to control the mdas on a per-PV basis.
If the user requests "--vgmetadatacopies all", this instructs LVM
to do 2 things: 1) clear all ignore bits, and 2) set the "unmanaged"
policy going forward.
Internally, we use the special MAX_UINT32 value to indicate 'all'.
This 'just' works since it's the largest value possible for the
field and so all 'ignore' bits on all mdas in the VG will get
cleared inside _vg_metadata_balance(). However, after we've
called the _vg_metadata_balance function, we check for the special
'all' value, and if set, we write the "unmanaged" value into the
metadata. As such, the 'all' value is never written to disk.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:39:39 +0000 (20:39 +0000)]
Ensure --metadatacopies parameter gets interpreted based on command.
Now that we have both --pvmetadatacopies and --vgmetadatacopies,
we need to make sure --metadatacopies gets interpreted correctly.
For pv commands, --metadatacopies should imply --pvmetadatacopies,
and for vg commands, --vgmetadatacopies.
Note: this will change the behavior of vgcreate with --metadatacopies
to be a synonym for --vgmetadatacopies. Previously, --metadatacopies
would apply to any PVs given with vgcreate that needed an implicit
pvcreate. As a result, one small change is needed to one of the nightly
tests - t-vgcreate-usage.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:39:08 +0000 (20:39 +0000)]
Update _vgmerge_single() to move fid->metadata_areas_ignored.
When vgmerge is called we move the mdas from the source to the
destination. With metadata balancing we now have another mda
list, fid->metadata_areas_ignored, so move the mdas on this list
as well.
This patch should not matter as the code is written today. However
we include it for completeness in the case that _vgmerge_single()
is refactored and/or moved into a library function.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:38:56 +0000 (20:38 +0000)]
Update check in vg_split_mdas to account for ignored mdas list.
The check in vg_split_mdas will trigger an error if the 'from' vg
list is empty. However, this might be ok in some instances now
that we have ignored mdas. Relax this check so an error is triggered
only in the case where there's truly no more mdas in the 'from'
vg.
One example of where this makes a difference is with vgreduce.
If we try to vgreduce a PV with un-ignored mdas, this should trigger
the balancing function to un-ignore mdas on another PV in the VG.
However, we don't get to vg_write() before we fail because this
list size check fails, and we see an error message indicating:
"Cannot remove final metadata area ..."
Another example is with vgsplit into a new VG, where the PVs
being moved contain all ignored mdas. We must move the mdas on
fid->metadata_areas_ignored from 'vg_from' to 'vg_to'.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:38:39 +0000 (20:38 +0000)]
Ensure fid mda lists are populated correctly during vgextend.
The vgextend path calls add_pv_to_vg(). Inside add_pv_to_vg(),
we must ensure we pass the correct mdas list into pv_setup(), as
copies of mdas are placed on the vg->fid list. If we don't place
the mdas on the correct vg->fid list, the various counts may be
incorrect and the metadata balance algorithm will not work when
called from vg_write() path.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:38:06 +0000 (20:38 +0000)]
Set vg_mda_copies when pvchange --metadataignore is given.
When a user explicitly sets a new mda ignore value for a PV, we
should update vg_mda_copies accordingly. When the VG is written
out, the user would not want the new ignore state to get lost as
a result of the vg_mda_copies value and logic in the vg_write
path.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:37:54 +0000 (20:37 +0000)]
Implement _vg_adjust_ignored_mdas and call from vg_write() path.
Compare the value of the newly added vg_mda_copies field
(--vgmetadatacopies parameter) with the current count of
in-use mdas and ignoring or unignoring mdas as necessary to
get to the target count. Also, as a safety check before
returning, ensure we have at least one mda enabled.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:37:37 +0000 (20:37 +0000)]
Update vgchange tool to accept --vgmetadatacopies.
Update logic in vgchange to handle --vgmetadatacopies, allow
--metadatacopies as a synonym to --vgmetadatacopies,
and add these parameters to args.h and commands.h
Forbit both --vgmetadatacopies and --metadatacopies as only
one allowed.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:37:10 +0000 (20:37 +0000)]
Make vg->mda_copies persistent in on disk vg metadata.
This patch adds the ability to read/write the vg->mda_copies values
from/to the vg metadata.
If we read the VG metadata and this field does not exist, we set
mda_copies to the default value of 0. Later in the code, we use
this special '0' value to indicate a disable of metadata balancing.
This should preserve existing LVM behavior and ensure metadata balancing
can be turned off should the need arise.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:36:56 +0000 (20:36 +0000)]
Add vg get/set methods for VG metadata copies.
This patch adds the get and partially implemented set function.
The 'set' function should probably ignore or un-ignore metadata areas
based on new values.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:36:37 +0000 (20:36 +0000)]
Add mda_copies to VG structures and initialization.
Add a field to struct volume_group to later implement metadata
balancing:
- mda_copies: target # of non-ignored mdas in the VG; default 0 (do
not control pv 'ignore mdas' bit.
This patch just adds the parameter to the structures with the default
values but does not modify any commands. Should be no functional change.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:36:18 +0000 (20:36 +0000)]
Define vgmetadatacopies in vgchange man page.
This patch adds a vgmetadatacopies parameter for metadata balancing.
This parameter provides a simple way for users to create a policy for
placing metadata on PVs automatically by LVM. The behavior is implemented
inside LVM by managing the 'ignore' mda bits. We chose the name
'vgmetadatacopies' as this is a natural extension to the existing parameter
'pvmetadatacopies' / 'metadatacopies' in pvcreate.
This is a first step at VG parameter based metadata balancing. Most users
will probably want to state that they want a certain number of PVs to contain
metadata, and they may be less concerned about a specific number of metadata
copies in the volume group. However, for default values (pvmetadatacopies
is 1 by default), the number of metadatacopies in the volume group, and the
number of PVs with metadata are the same. In the future we could add
vgmetadatacopiespvs to define more specifically the number of pvs in the
VG that contain metadata, but for now we start with this parameter.
Another possible future extension would be to define a specific pv tag
to mark the set of PVs that should be used for metadata balancing. This
tag based approach could be used in conjunction with 'vgmetadatacopies'.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:35:49 +0000 (20:35 +0000)]
Before committing each mda, arrange mdas so ignored mdas get committed first.
Arrange mdas so mdas that are to be ignored come first. This is an
optimization that ensures consistency on disk for the longest period of time.
This was noted by agk in review of the v4 patchset of pvchange-based mda
balance.
Note the following example for an explanation of the background:
Assume the initial state on disk is as follows:
PV0 (v1, non-ignored)
PV1 (v1, non-ignored)
PV2 (v1, non-ignored)
PV3 (v1, non-ignored)
If we did not sort the list, we would have a commit sequence something like
this:
PV0 (v2, non-ignored)
PV1 (v2, ignored)
PV2 (v2, ignored)
PV3 (v2, non-ignored)
After the commit of PV0's mdas, we'd have an on-disk state like this:
PV0 (v2, non-ignored)
PV1 (v1, non-ignored)
PV2 (v1, non-ignored)
PV3 (v1, non-ignored)
This is an inconsistent state of the disk. If the machine fails, the next
time it was brought back up, the auto-correct mechanism in vg_read would
update the metadata on PV1-PV3. However, if possible we try to avoid
inconsistent on-disk states. Clearly, because we did not sort, we have
a greater chance of on-disk inconsistency - from the time the commit of
PV0 is complete until the time PV3 is complete.
We could improve the amount of time the on-disk state is consistent by simply
sorting the commit order as follows:
PV1 (v2, ignored)
PV2 (v2, ignored)
PV0 (v2, non-ignored)
PV3 (v2, non-ignored)
Thus, after the first PV is committed (in this case PV1), on-disk we would
have:
PV0 (v1, non-ignored)
PV1 (v2, ignored)
PV2 (v1, non-ignored)
PV3 (v1, non-ignored)
This is clearly a consistent state. PV1 will be read but the mda will be
ignored. All other PVs contain v1 metadata, and no auto-correct will be
required. In fact, if we commit all PVs with ignored mdas first, we'll
only have an inconsistent state when we start writing non-ignored PVs,
and thus the chances we'll get an inconsistent state on disk is much
less with the sorted method.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
Dave Wysochanski [Mon, 28 Jun 2010 20:35:17 +0000 (20:35 +0000)]
Update _vg_read and _text_create_text_instance to use fid_add_mda[s].
When we are constructing the vg, we may need to adjust the list of
metadata_areas if there are ignored mdas. At label read time, we
do not read the metadata of ignored mdas, and as a result, they do
not get placed on vg->fid->metadata_areas inside _text_create_text_instance
since lvmcache does not have these areas attached to vginfo->infos.
However, when we're checking the pvids inside _vg_read, after having
read another metadata area from another PV, we do have the opportunity
to update the metadata_area and metadata_areas_ignored lists based
on the read metadata_area. We need accurate mda lists for the reporting
functions that count the ignored mdas, as well as general correctness
of mda balancing.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>