IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
When reading a foreign VG we cannot write it, since
it belongs to another host. When reading a shared VG
we cannot write it because we may not have an ex lock.
(Or we may be reading the shared VG while not using
lvmlockd in which case it's like reading a foreign VG.)
Add the same checks for wiping outdated PVs. We may
read a foreign or shared VG, or see the PVs, while
another host is part way through writing a new version
of the VG to the PVs. This might cause us to think
some of the PVs are outdated. We do not want to
write another host's PVs, especially when we may
wrongly conclude they are outdated.
Running "vgremove -f VG & pvs" results in the pvs
command reporting that the VG is not found or is
inconsistent. If the VG is gone or being removed,
the pvs command should just skip it and not print
errors about it.
"Not found" is because the pvs command created the
list of VGs to process, including VG, then vgremove
removed the VG, then the pvs command came to to read
the VG to process it and did not find it.
An "inconsistent" error could be reported if vgremove
had only partially completed removing VG when pvs did
vg_read on the VG to process it, causing pvs to find
the VG in a partially-removed state.
This fix adds a flag that pvs uses to ignore a VG
that can't be read or is inconsistent.
When a command does a sequence of
vg_write + vg_commit + vg_write + vg_commit,
initialization of non-PV devices happens during the
first vg_write, and does not need to be repeated by
the second vg_write.
When creating a lockd VG, this sequence occurs because
the VG is first created, then the lockd data is created,
then the lockd data is then written to the VG metadata.
The vgchange/lvchange activation commands read the VG, and
don't write it, so they acquire a shared VG lock from lvmlockd.
When other commands fail to acquire a shared VG lock from
lvmlockd, a warning is printed and they continue without it.
(Without it, the VG metadata they display from lvmetad may
not be up to date.)
vgchange/lvchange -a shouldn't continue without the shared
lock for a couple reasons:
. Usually they will just continue on and fail to acquire the
LV locks for activation, so continuing is pointless.
. More importantly, without the sh VG lock, the VG metadata
used by the command may be stale, and the LV locks shown
in the VG metadata may no longer be current. In the
case of sanlock, this would result in odd, unpredictable
errors when lvmlockd doesn't find the expected lock on
disk. In the case of dlm, the invalid LV lock could be
granted for the non-existing LV.
The solution is to not continue after the shared lock fails,
in the same way that a command fails if an exclusive lock fails.
There are two different failure conditions detected in
access_vg_lock_type() that should have different error
messages. This adds another failure flag so the two
cases can be distinguished to avoid printing a misleading
error message.
There are at least a couple instances where
the lock_args check does not work correctly,
(listed in the comment), so disable the
NULL check for lock_args until those are
resolved.
In process_each_{vg,lv,pv} when no vgname args are given,
the first step is to get a list of all vgid/vgname on the
system. This is exactly what lvmetad returns from a
vg_list request. The current code is doing a vg_lookup
on each VG after the vg_list and populating lvmcache with
the info for each VG. These preliminary vg_lookup's are
unnecessary, because they will be done again when the
processing functions call vg_read. This patch eliminates
the initial round of vg_lookup's, which can roughly cut in
half the number of lvmetad requests and save a lot of extra work.
Do not keep dangling LVs if they're removed from the vg->lvs list and
move them to vg->removed_lvs instead (this is actually similar to already
existing vg->removed_pvs list, just it's for LVs now).
Once we have this vg->removed_lvs list indexed so it's possible to
do lookups for LVs quickly, we can remove the LV_REMOVED flag as
that one won't be needed anymore - instead of checking the flag,
we can directly check the vg->removed_lvs list if the LV is present
there or not and to say if the LV is removed or not then. For now,
we don't have this index, but it may be implemented in the future.
Refactor the recent metadata-reading optimisation patches.
Remove the recently-added cache fields from struct labeller
and struct format_instance.
Instead, introduce struct lvmcache_vgsummary to wrap the VG information
that lvmcache holds and add the metadata size and checksum to it.
Allow this VG summary information to be looked up by metadata size +
checksum. Adjust the debug log messages to make it clear when this
shortcut has been successful.
(This changes the optimisation slightly, and might be extendable
further.)
Add struct cached_vg_fmtdata to format-specific vg_read calls to
preserve state alongside the VG across separate calls and indicate
if the details supplied match, avoiding the need to read and
process the VG metadata again.
Since we take a lock inside vg_lock_newname() and we do a full
detection of presence of vgname inside all scanned labels,
there is no point to do this for second time to be sure
there is no such vg.
The only side-effect of such call would be a full validation of
some already exising VG metadata - but that's not the task for
vgcreate when create a new VG.
This call noticable reduces number of scans during 'vgcreate'.
When reading VG mda from multiple PVs - do all the validation only
when mda is seen for the first time and when mda checksum and length
is same just return already existing VG pointer.
(i.e. using 300PVs for a VG would lead to create and destroy 300 config trees....)
Previous versions of lvm will not obey the restrictions
imposed by the new system_id, and would allow such a VG
to be written. So, a VG with a new system_id is further
changed to force previous lvm versions to treat it as
read-only. This is done by removing the WRITE flag from
the metadata status line of these VGs, and putting a new
WRITE_LOCKED flag in the flags line of the metadata.
Versions of lvm that recognize WRITE_LOCKED, also obey the
new system_id. For these lvm versions, WRITE_LOCKED is
identical to WRITE, and the rules associated with matching
system_id's are imposed.
A new VG lock_type field is also added that causes the same
WRITE/WRITE_LOCKED transformation when set. A previous
version of lvm will also see a VG with lock_type as read-only.
Versions of lvm that recognize WRITE_LOCKED, must also obey
the lock_type setting. Until the lock_type feature is added,
lvm will fail to read any VG with lock_type set and report an
error about an unsupported lock_type. Once the lock_type
feature is added, lvm will allow VGs with lock_type to be
used according to the rules imposed by the lock_type.
When both system_id and lock_type settings are removed, a VG
is written with the old WRITE status flag, and without the
new WRITE_LOCKED flag. This allows old versions of lvm to
use the VG as before.
In log messages refer to it as system ID (not System ID).
Do not put quotes around the system_id string when printing.
On the command line use systemid.
In code, metadata, and config files use system_id.
In lvmsystemid refer to the concept/entity as system_id.
The only realistic way for a host to have active LVs in a
foreign VG is if the host's system_id (or system_id_source)
is changed while LVs are active.
In this case, the active LVs produce an warning, and access
to the VG is implicitly allowed (without requiring --foreign.)
This allows the active LVs to be deactivated.
In this case, rescanning PVs for the VG offers no benefit.
It is not possible that rescanning would reveal an LV that
is active but wasn't previously in the VG metadata.
A foreign VG should be silently ignored by a reporting/display
command like 'vgs'. If the reporting/display command specifies
a foreign VG by name on the command line, it should produce an
error message.
Scanning commands pvscan/vgscan/lvscan are always allowed to
read and update caches from all PVs, including those that belong
to foreign VGs.
Other non-report/display/scan commands always ignore a foreign
VG, or report an error if they attempt to use a foreign VG.
vgimport should always invalidate the lvmetad cache because
lvmetad likely holds a pre-vgexported copy of the VG.
(This is unrelated to using foreign VGs; the pre-vgexported
VG may have had no system_id at all.)
When checking whether the system ID permits access to a VG, check for
each permitted situation first, and only then issue the appropriate
error message. Always issue a message for now. (We'll try to
suppress some of those later when the VG concerned wasn't explicitly
requested.)
Add more messages to try to ensure every return code is checked and
every error path (and only an error path) contains a log_error().
Add self-correction to vgchange -c to deal with situations where
the cluster state and system ID state are out-of-sync (e.g. if
old tools were used).
Move the lvm1 sys ID into vg->lvm1_system_id and reenable the #if 0
LVM1 code. Still display the new-style system ID in the same
reporting field, though, as only one can be set.
Add a format feature flag FMT_SYSTEM_ON_PVS for LVM1 and disallow
access to LVM1 VGs if a new-style system ID has been set.
Treat the new vg->system_id as const.
The dev ext source must be reset for the dev_cache_get call
(which evaluates filters), not lvmcache_label_scan - so fix
original commit 727c7ff85d.
Also, add comments in _pvcreate_check fn explaining why
refresh filter and rescan is needed and exactly in which
situations.
Before, we refreshed filters and we did full rescan of devices if
we passed through wiping (wipe_known_signatures fn call). However,
this fn returns success even if no signatures were found and so
nothing was wiped. In this case, it's not necessary to do the
filter refresh/rescan of devices as nothing changed clearly.
This patch exports number of wiped signatures from all the
wiping functions below. The caller (_pvcreate_check) then checks
whether any wiping was done at all and if not, no refresh/rescan
is done, saving some time and resources.
pvcreate code path executes signature wiping if there are any signatures
found on device to prepare the device for PV. When the signature is wiped,
the WATCH udev rule triggers the event which then updates udev database
with fresh info, clearing the old record about previous signature.
However, when we're using udev db as dev-ext source, we'd need to wait
for this WATCH-triggered event. But we can't synchronize against such
events (at least not at this moment). Without this sync, if the code
continues, the device could still be marked as containing the old
signature if reading udev db. This may end up even with the device
to be still filtered, though the signature is already wiped.
This problem is then exposed as (an example with md components):
$ mdadm --create /dev/md0 --level=1 --raid-devices=2 /dev/sda /dev/sdb --run
$ mdadm -S /dev/md0
$ pvcreate -y /dev/sda
Wiping linux_raid_member signature on /dev/sda.
/dev/sda: Couldn't find device. Check your filters?
$ echo $?
5
So we need to temporarily switch off "udev" dev-ext source here
in this part of pvcreate code until we find a way how to sync
with WATCH events.
(This problem does not occur with signature wiping which we do
on newly created LVs since we already handle this properly with
our udev flags - the LV_NOSCAN/LV_TEMPORARY flag. But we can't use
this technique for non-dm devices to keep WATCH rule under control.)
Better than previous patch which changed log_warn to log_error -
we can have multiple MDAs and if one of them fails to be written,
we can still continue with other MDAs if we're in a mode where
we can handle missing PVs - so keep the log_warn for single
failed MDA write as it was before.
However, add log_error with "Failed to write VG <vg_name>." in
case we're not handling missing PVs or no MDA was written at all
during VG write process. This also prevents an internal error in
which the vg_write fails and we're not issuing any other log_error
in vg_write caller or above, so we end up with:
"Internal error: Failed command did not use log_error".
Use log_warn when we are effectively not creating an error -
we 'allowed' inconsistent read for a reason - so it's just warning
level we process inconsistent VG - it's upto caller later to decide
error level of command return value and in case of error it needs
to use log_error then.
Failed recovery provides different (NULL) VG then FAILED_INCONSISTENT.
Mark it with different failure bit - since FAILED_INCONSISTENT is
supposed to contain something 'usable' (thought inconsistent).
Add API call to calculate extents from percentage value.
Size is based in DM_PERCENT_1 units.
(Supporting decimal point number).
This commit is preparing functionality for more global
usage of % with i.e. --size option.
find_pv_in_vg fn iterates over the list of PVs covered by the VG and
each PV's pvl->pv->dev is compared with device acquired from device
cache. However, in case pvl->pv->dev is NULL as well as device cache
returns NULL (e.g. when device is filtered), we'll get incorrect match
and the code calling find_pv_in_vg uses incorrect PV (as it thinks
it's the exact PV with the pv_name). The INTERNAL_ERROR covers this
situation and errors out immediately.
The warnings arg was used to enable logging of warnings
when reading a PV. This arg is turned into a set of flags
with the WARN_PV_READ flag matching the existing behavior.
A new flag WARN_INCONSISTENT is added that will cause
vg_read_internal() to log the "VG is not consistent"
warning so the various callers do not need to log
this warning themselves.
A new vg_read flag READ_WARN_INCONSISTENT is used from
reporting to enable the WARN_INCONSISTENT flag in
vg_read_internal.
[Committed by agk with cosmetic changes and tweaks.]
There are actually three filter chains if lvmetad is used:
- cmd->lvmetad_filter used when when scanning devices for lvmetad
- cmd->filter used when processing lvmetad responses
- cmd->full_fiilter (which is just cmd->lvmetad_filter + cmd->filter chained together) used
for remaining situations
This patch adds the third one - "cmd->full_filter" - currently this is
used if device processing does not fall into any of the groups before,
for example, devices which does not have the PV label yet and we're just
creating a new one or we're processing the devices where the list of the
devices (PVs) is not returned by lvmetad initially.
Currently, the cmd->full_filter is used exactly in these functions:
- lvmcache_label_scan
- _pvcreate_check
- pvcreate_vol
- lvmdiskscan
- pvscan
- _process_each_label
If lvmetad is used, then simply cmd->full_filter == cmd->filter because
cmd->lvmetad_filter is NULL in this case.
When vg_ondisk is NULL we do not need to search
through the whole VG to find out the same LV.
NOTE: as of now - VG locking is not enabled as some code parts
are breaking memory locking rules (lvm2app).
Once we enforce VG locking for read-only commands the effect
will be much better for larger VGs.
Use lv_is_* macros throughout the code base, introducing
lv_is_pvmove, lv_is_locked, lv_is_converting and lv_is_merging.
lv_is_mirror_type no longer includes pvmove.
_pvcreate_check() has two missing requirements:
After refreshing filters there must be a rescan.
(Otherwise the persistent filter may remain empty.)
After wiping a signature, the filters must be refreshed.
(A device that was previously excluded by the filter due to
its signature might now need to be included.)
If several devices are added at once, the repeated scanning isn't
strictly needed, but we can address that later as part of the command
processing restructuring (by grouping the devices).
Replace the new pvcreate code added by commit
54685c20fc "filters: fix regression caused
by commit e80884cd080cad7e10be4588e3493b9000649426"
with this change to _pvcreate_check().
The filter refresh problem dates back to commit
acb4b5e4de "Fix pvcreate device check."
Commit e80884cd08 tried to dump filters
for them to be reevaluated when creating a PV to avoid overwriting
any existing signature that may have been created after last
scan/filtering.
However, we need to call refresh_filters instead of
persistent_filter->dump since dump requires proper rescannig to fill
up the persistent filter again. However, this is true only for pvcreate
but not for vgcreate with PV creation where the scanning happens before
this PV creation and hence the next rescan (if not full scan), does not
fill the persistent filter.
Also, move refresh_filters so that it's called sooner and only for
pvcreate, vgcreate already calls lvmcache_label_scan(cmd, 2) which
then calls refresh_filters itself, so no need to reevaluate this again.
This caused the persistent filter (/etc/lvm/cache/.cache file) to be
wrong and contain only the PV just being processed with
vgcreate <vg_name> <pv_name_to_create>.
This regression caused other block devices to be filtered out in case
the vgcreate with PV creation was used and then the persistent filter
is used by any other LVM command afterwards.
This is addendum to commit 2e82a070f3
which fixed these spurious messages that appeared after commit
651d5093ed ("avoid pv_read in
find_pv_by_name").
There was one more "not found" message issued in case the device
could not be found in device cache (commit 2e82a07 fixed this only
for PV lookup itself). But if we "allow_unformatted" for
find_pv_by_name, we should not issue this message even in case
the device can't be found in dev cache as we just need to know
whether there's a PV or not for the code to decide on next steps
and we don't want to issue any messages if either device itself
is not found or PV is not found.
For example, when we were creating a new PV (and so allow_unformatted = 1)
and the device had a signature on it which caused it to be filtered
by device filter (e.g. MD signature if md filtering is enabled),
or it was part of some other subsystem (e.g. multipath), this message
was issued on find_pv_by_name call which was misleading.
Also, remove misleading "stack" call in case find_pv_by_name
returns NULL in pvcreate_check - any error state is reported
later by pvcreate_check code so no need to "stack" here.
There's one more and proper check to issue "not found" message if
the device can't be found in device cache within pvcreate_check fn
so this situation is still covered properly later in the code.
Before this patch (/dev/sda contains MD signature and is therefore filtered):
$ pvcreate /dev/sda
Physical volume /dev/sda not found
WARNING: linux_raid_member signature detected on /dev/sda at offset 4096. Wipe it? [y/n]:
With this patch applied:
$ pvcreate /dev/sda
WARNING: linux_raid_member signature detected on /dev/sda at offset 4096. Wipe it? [y/n]:
Non-existent devices are still caught properly:
$ pvcreate /dev/sdx
Device /dev/sdx not found (or ignored by filtering).
This is addendum for commit 6dc7b783c8.
LVM1 format stores the ALLOCATABLE flag directly in PV header, not
in VG metadata. So the code needs to be fixed further to work
properly for lvm1 format so that the correct PV header is written
(the flag is set only if the PV is in some VG, unset otherwise).
...to avoid using cached value (persistent filter) and therefore
not noticing any change made after last scan/filtering - the state
of the device may have changed, for example new signatures added.
$ lvm dumpconfig --type diff
allocation {
use_blkid_wiping=0
}
devices {
obtain_device_list_from_udev=0
}
$ cat /etc/lvm/cache/.cache | grep sda
$ vgscan
Reading all physical volumes. This may take a while...
Found volume group "fedora" using metadata type lvm2
$ cat /etc/lvm/cache/.cache | grep sda
"/dev/sda",
$ parted /dev/sda mklabel gpt
Information: You may need to update /etc/fstab.
$ parted /dev/sda print
Model: QEMU QEMU HARDDISK (scsi)
Disk /dev/sda: 134MB
Sector size (logical/physical): 512B/512B
Partition Table: gpt
Disk Flags:
Number Start End Size File system Name Flags
$ cat /etc/lvm/cache/.cache | grep sda
"/dev/sda",
====
Before this patch:
$ pvcreate /dev/sda
Physical volume "/dev/sda" successfully created
With this patch applied:
$ pvcreate /dev/sda
Physical volume /dev/sda not found
Device /dev/sda not found (or ignored by filtering).
The list of strings is used quite frequently and we'd like to reuse
this simple structure for report selection support too. Make it part
of libdevmapper for general reuse throughout the code.
This also simplifies the LVM code a bit since we don't need to
include and manage lvm-types.h anymore (the string list was the
only structure defined there).
Given a named mirror LV, vgsplit will look for the PVs that compose it
and move them to a new VG. It does this by first looking at the log
and then the legs. If the log is on the same device as one of the mirror
images, a problem occurs. This is because the PV is moved to the new VG
as the log is processed and thus cannot be found in the current VG when
the image is processed. The solution is to check and see if the PV we are
looking for has already been moved to the new VG. If so, it is not an
error.
Parsing vg structure during supend/commit/resume may require a lot of
memory - so move this into vg_write.
FIXME: there are now multiple cache layers which our doing some thing
multiple times at different levels. Moreover there is now different
caching path with and without lvmetad - this should be unified
and both path should use same mechanism.
The libblkid can detect DM_snapshot_cow signature and when creating
new LVs with blkid wiping used (allocation/use_blkid_wiping=1 lvm.conf
setting and --wipe y used at the same time - which it is by default).
Do not issue any prompts about this signature when new LV is created
and just wipe it right away without asking questions. Still keep the
log in verbose mode though.
When LV is scanned for its dependencies - scan also origin's snapshots,
and thin external origins.
So if any PV from snapshot or external origin device is missing - lvm2 will
avoid trying to activate such device.
Replacement of pv_read by find_pv_by_name in commit
651d5093ed caused spurious
error messages when running pvcreate or vgextend against an
unformatted device.
Physical volume /dev/loop4 not found
Physical volume "/dev/loop4" successfully created
Physical volume /dev/loop4 not found
Physical volume /dev/loop4 not found
Physical volume "/dev/loop4" successfully created
Volume group "vg1" successfully extended
If we're calling pvcreate on a device that already has a PV label,
the blkid detects the existing PV and then we consider it for wiping
before we continue creating the new PV label and we issue a warning
with a prompt whether such old PV label should be removed. We don't
do this with native signature detection code. Let's make it consistent
with old behaviour.
But still keep this "PV" (identified as "LVM1_member" or "LVM2_member"
by blkid) detection when creating new LVs to avoid unexpected PV label
appeareance inside LV.
This is actually the wipefs functionailty as a matter of fact
(wipefs uses the same libblkid calls).
libblkid is more rich when it comes to detecting various
signatures, including filesystems and users can better
decide what to erase and what should be kept.
The code is shared for both pvcreate (where wiping is necessary
to complete the pvcreate operation) and lvcreate where it's up
to the user to decide.
The verbose output contains a bit more information about the
signature like LABEL and UUID.
For example:
raw/~ # lvcreate -L16m vg
WARNING: linux_raid_member signature detected on /dev/vg/lvol0 at offset 4096. Wipe it? [y/n]
or more verbose one:
raw/~ # lvcreate -L16m vg -v
...
Found existing signature on /dev/vg/lvol0 at offset 4096: LABEL="raw.virt:0" UUID="da6af139-8403-5d06-b8c4-13f6f24b73b1" TYPE="linux_raid_member" USAGE="raid"
WARNING: linux_raid_member signature detected on /dev/vg/lvol0 at offset 4096. Wipe it? [y/n]
The verbose output is the same output as found in blkid.
The wipe_known_signatures fn now wraps the _wipe_signature fn that is called
for each known signature (currently md, swap and luks). This patch makes the
code more readable, not repeating the same sequence when used anywhere in the
code. We're going to reuse this code later...
Add a PV create which takes a paramters object that
has get/set method to configure PV creation.
Current get/set operations include:
- size
- pvmetadatacopies
- pvmetadatasize
- data_alignment
- data_alignment_offset
- zero
Reference: https://bugzilla.redhat.com/show_bug.cgi?id=880395
Signed-off-by: Tony Asleson <tasleson@redhat.com>
Replace the code with the refactored vgreduce_single instead
of calling its own implementation.
Corrects bug: https://bugzilla.redhat.com/show_bug.cgi?id=989174
Signed-off-by: Tony Asleson <tasleson@redhat.com>
Accept --ignoreskippedcluster with pvs, vgs, lvs, pvdisplay, vgdisplay,
lvdisplay, vgchange and lvchange to avoid the 'Skipping clustered
VG' errors when requesting information about a clustered VG
without using clustered locking and still exit with success.
The messages can still be seen with -v.
The pv resize code required that a lvm_vg_write be done
to commit the change. When the method to add the ability
to list all PVs, including ones that are not assocated with
a VG we had no way for the user to make the change persistent.
Thus additional resize code was move and now liblvm calls into
a resize function that does indeed write the changes out, thus
not requiring the user to explicitly write out he changes.
Signed-off-by: Tony Asleson <tasleson@redhat.com>
As locks are held, you need to call the included function
to release the memory and locks when done transversing the
list of physical volumes.
V2: Rebase fix
V3: Prevent VGs from getting cached and then write protected.
Signed-off-by: Tony Asleson <tasleson@redhat.com>
Changes:
- move device type registration out of "type filter" (filter.c)
to a separate and new dev-type.[ch] for common use throughout the code
- the structure for keeping the major numbers detected for available
device types and available partitioning available is stored in
"dev_types" structure now
- move common partitioning detection code to dev-type.[ch] as well
together with other device-related functions bound to dev_types
(see dev-type.h for the interface)
The dev-type interface contains all common functions used to detect
subsystems/device types, signature/superblock recognition code,
type-specific device properties and other common device properties
(bound to dev_types), including partitioning support.
- add dev_types instance to cmd context as cmd->dev_types for common use
- use cmd->dev_types throughout as a central point for providing
information about device types
Previously, we have relied on UUIDs alone, and on lvmcache to make getting a
"new copy" of VG metadata fast. If the code which triggers the activation has
the correct VG metadata at hand (the version which is currently on disk), it can
now hand it to the activation code directly.
This allows us to get the current on-disk version of the metadata whenever we
have the current in-flight version, without a recourse to scanning or lvmcache.
Avoid hitting memory corruption (double free) in code path,
where PV FID has been already destroyed and the released pointer
was left in PV structure and could have been tried to be released
from there 2nd. time with final context destruction.
This fixes a long standing regression since LVM2 2.02.74 (commit 4efb1d9c,
"Update heuristic used for default and detected data alignment.")
The default PE alignment could be used (via MAX()) even if it was
determined that the device's MD stripe width, or minimal_io_size or
optimal_io_size were not factors of the default PE alignment (either 64K
or the newer default of 1MB, etc). This bug would manifest if the
default PE alignment was larger than the overriding hint that the
device provided (e.g. default of 1MB vs optimal_io_size of 768K).
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
The pv_by_path might be also dangerous to use as it does not
count with any other metadata areas but the ones found on the PV
itself. If metadata was not found on the PV referenced by the path,
it returned no PV though it might have been referenced by metadata
elsewhere (on other PVs...).
If extending a VG and including a PV with 0 MDAs that was already
a part of a VG, the vgextend allowed that PV to be added and we
ended up *with one PV in two VGs*!
The vgextend code used the 'pv_by_path' fn that returned a PV for
a given path. However, when the PV did not have any metadata areas,
the fn just returned a PV without any reference to existing VG.
Consequently, any checks for the existing VG failed.
[0] raw/~ # pvcreate --metadatacopies 0 /dev/sda
Physical volume "/dev/sda" successfully created
[0] raw/~ # pvcreate --metadatacopies 1 /dev/sdb
Physical volume "/dev/sdb" successfully created
[0] raw/~ # vgcreate vg1 /dev/sda /dev/sdb
Volume group "vg1" successfully created
[0] raw/~ # pvcreate --metadatacopies 1 /dev/sdc
Physical volume "/dev/sdc" successfully created
[0] raw/~ # vgcreate vg2 /dev/sdc
Volume group "vg2" successfully created
Before this patch (incorrect):
[0] raw/~ # vgextend vg2 /dev/sda
Volume group "vg2" successfully extended
With this patch (correct):
[0] raw/~ # vgextend vg2 /dev/sda
Physical volume '/dev/sda' is already in volume group 'vg1'
Unable to add physical volume '/dev/sda' to volume group 'vg2'.
Before, the find_pv_by_name call always failed if the PV found was orphan.
However, we might use this function even for a PV that is not part of any VG.
This patch adds 'allow_orphan' arg to find_pv_by_name fn that allows that.
_find_pv_by_name -> find_pv_by_name
_find_pv_in_vg -> find_pv_in_vg
_find_pv_in_vg_by_uuid -> find_pv_in_vg_by_uuid
The only callers of the underscored variants were their wrappers
without the underscore. No other part of the code referenced the
underscored variants.
For example, the old call and reference:
find_config_tree_str(cmd, "devices/dir", DEFAULT_DEV_DIR)
...now becomes:
find_config_tree_str(cmd, devices_dir_CFG)
So we're referring to the named configuration ID instead
of passing the configuration path and the default value
is taken from central config definition in config_settings.h
automatically.
The PV header extension information (PV header extension version, flags
and list of Embedding Area locations) is stored just beyond the PV header base.
When calculating the Embedding Area start value (ea_start), the same logic is
used as when calculating the pe_start value for Data Area - the value must
follow exactly the same alignment restrictions for its start value
(the alignment detected automatically or provided via command line using
the --dataalignment and --dataalignmentoffset arguments).
The Embedding Area is placed at the very start of the PV, starting at
ea_start. The Data Area starting at pe_start is placed next. The pe_start is
still properly aligned. Due to the pe_start alignment, it's possible that the
resulting Embedding Area size (ea_size) ends up bigger in size than requested
(but never less than requested).
Extract restorable PV creation parameters from struct pvcreate_params into
a separate struct pvcreate_restorable_params for clarity and also for better
maintainability when adding any new items later.
If zero metadata copies are used, there's no further recalculation of
PV alignment that happens when adding metadata areas to the PV and
which actually calculates the alignment correctly as a matter of fact.
So fix this for "PV without MDA" case as well.
Before this patch:
[1] raw/~ # pvcreate --dataalignment 8m --dataalignmentoffset 4m
--metadatacopies 1 /dev/sda
Physical volume "/dev/sda" successfully created
[1] raw/~ # pvs -o pv_name,pe_start
PV 1st PE
/dev/sda 12.00m
[1] raw/~ # pvcreate --dataalignment 8m --dataalignmentoffset 4m
--metadatacopies 0 /dev/sda
Physical volume "/dev/sda" successfully created
[1] raw/~ # pvs -o pv_name,pe_start
PV 1st PE
/dev/sda 8.00m
After this patch:
[1] raw/~ # pvcreate --dataalignment 8m --dataalignmentoffset 4m
--metadatacopies 1 /dev/sda
Physical volume "/dev/sda" successfully created
[1] raw/~ # pvs -o pv_name,pe_start
PV 1st PE
/dev/sda 12.00m
[1] raw/~ # pvcreate --dataalignment 8m --dataalignmentoffset 4m
--metadatacopies 0 /dev/sda
Physical volume "/dev/sda" successfully created
[1] raw/~ # pvs -o pv_name,pe_start
PV 1st PE
/dev/sda 12.00m
Also, remove a superfluous condition "pv->pe_start < pv->pe_align" in:
if (pe_start == PV_PE_START_CALC && pv->pe_start < pv->pe_align)
pv->pe_start = pv->pe_align ...
This part of the condition is not reachable as with the PV_PE_START_CALC,
we always have pv->pe_start set to 0 from the PV struct initialisation
(...the pv->pe_start value is just being calculated).
fmt1 doesn't have a separate commit function: updates take effect
immediately vg_write is called, so we must update lvmetad at this
point if we're going to go on and ask lvmetad for the VG metadata
again before calling the commit function (though that's probably an
unsupported and pointless thing to do anyway as the client must
already have that data and it cannot have changed because it's locked
and with devs suspended we shouldn't be communicating with lvmetad;
so when that's fixed properly, this fix here can be reverted).
This problem showed up as an internal error when lvremoving an LVM1
snapshot.
> Internal error: LV snap1 (00000000000000000000000000000001) missing from preload metadata
https://bugzilla.redhat.com/891855
If the lvmcache_info_from_pvid() fails to find valid
info, invoke the lookup by dev, and only in this case
call lvmcache_info_from_pvid() again.
Also check for the result of info and return
error directly, so the NULL is not passed
to lvmcache_get_label().
Accept -q as the short form of --quiet.
Suppress non-essential standard output if -q is given twice.
Treat log/silent in lvm.conf as equivalent to -qq.
Review all log_print messages and change some to
log_print_unless_silent.
When silent, the following commands still produce output:
dumpconfig, lvdisplay, lvmdiskscan, lvs, pvck, pvdisplay,
pvs, version, vgcfgrestore -l, vgdisplay, vgs.
[Needs checking.]
Non-essential messages are shifted from log level 4 to log level 5
for syslog and lvm2_log_fn purposes.
Adding couple INTERNAL_ERROR reports for unwanted parameters:
Ensure the 'top' metadata node cannot be NULL for lvmetad.
Make obvious vginfo2 cannot be NULL.
Report internal error if handler and vg is undefined.
Check for handle in poll_vg().
Ensure seg is not NULL in dev_manager_transient().
Report missing read_ahead for _lv_read_ahead_single().
Check for report handler in dm_report_object().
Check missing VG in _vgreduce_single().
Define an 'activation_handler' that gets called automatically on
PV appearance/disappearance while processing the lvmetad_pv_found
and lvmetad_pv_gone functions that are supposed to update the
lvmetad state based on PV availability state. For now, the actual
support is for PV appearance only, leaving room for PV disappearance
support as well (which is a more complex problem to solve as this
needs to count with possible device stack).
Add a new activation change mode - CHANGE_AAY exposed as
'--activate ay/-aay' argument ('activate automatically').
Factor out the vgchange activation functionality for use in other
tools (like pvscan...).
Add 3rd daemon return state "unknown" for lookups that are carried out
successfully but don't find the item requested.
Avoid issuing error messages when it's expected that a device that's
being looked up in lvmetad might not be there.
Adding at least stack traces with some FIXMEs for cases,
where we might want to do something cleaver - maybe fail command
or give user hints something is not going well ?
For remote_backup is stack probably 'good' enough for now.
Move commod code to destroy orphan VG into free_orphan_vg() function.
Use orphan vgmem for creation of PV lists.
Remove some free_pv_fid() calls (FIXME: check all of them)
FIXME: Check whether we could merge release_vg back again for all VGs.
Use static buffer instead of stack allocated buffer.
This reduces stack size usage of lvm tool and the
change is very simple.
Since the whole library is not thread safe - it should not
add any new problems - and if there will be some conversion
it's easy to convert this to use some preallocated buffer.
For write we do not need to hold memory locked.
This relaxes many conditions and avoid problems when allocating
a lot of memory for writting metadata buffers.
(In case of huge MDA size this would lead to mismatch between
locked and unlocked memory region size).
Add also internal check we are not writing in critical section.
pvck prints 'extra' character from the label since there is no '\0'
after the struct label entry and just uint64_t follows directly.
So avoid it by limiting 8 chars to be printed.
https://www.redhat.com/archives/lvm-devel/2011-January/msg00109.html
Signed-off-by: Paul Bolle <pebolle tiscali nl>
Properly detect if the filters were refreshed properly.
(May needs few more fixes ??)
Filter refresh may fail because it may be out of free file descriptors
when clvmd gets overloaded.
leaving behind the LVM-specific parts of the code (convenience wrappers that
handle `struct device` and `struct cmd_context`, basically). A number of
functions have been renamed (in addition to getting a dm_ prefix) -- namely,
all of the config interface now has a dm_config_ prefix.
Use debug pool locking functionality. So the command could check,
whether the memory in the pool has not been modified.
For lv_postoder() instead of unlocking and locking for every changed
struct status member do it once when entering and leaving function.
(mprotect would trap each such memory access).
Currently lv_postoder() does not modify other part of vg structure
then status flags of each LV with flags that are reverted back to
its original state after function exit.
Extend vginfo cache with cached VG structure. So if the same metadata
are use, skip mda decoding in the case, the same data are in use.
This helps for operations like activation of all LVs in one VG,
where same data were decoded giving the same output result.
Patch adds 1-to-1 connection between volume_group and lvmcache_vginfo.
Move the free_vg() to vg.c and replace free_vg with release_vg
and make the _free_vg internal.
Patch is needed for sharing VG in vginfo cache so the release_vg function name
is a better fit here.
As this flag could not have been set by the current code - removing it.
Note: because of the wrong code logic this call:
lvmcache_update_vg(correct_vg, correct_vg->status & PRECOMMITTED &
(inconsistent ? INCONSISTENT_VG : 0));
had always passed '0' - now after flag removal it's passing
PRECOMMITTED flag in - this present functinal change in this patch.
To match the original functionality - 0 had to be always passed.
More testing is needed here.
Last usage was removed in Petr's commit related to VG mda repair fix
where relaxed check starts to ignore inconsistencies coming from
PVs that are marked MISSING - thus removing unused variable.
It's useful to keep the partial flag cached - so just move the call
for vg_mark_partil_lvs() into import_vg_from_config_tree() so it gets
evaluated before it goes through the lvmcache.
This patch should not present any functional change.
Note: It is rather temporal solution - proper place is probably inside the
'read' call back - but needs some more discussion.
For now using this minor hack.
transient error), stemming from the following sequence of events:
1) devices fail IO, triggering repair
2) dmeventd starts fixing up the mirror
3) during the downconversion, a new metadata version is written
--> the devices come back online here
4) the mirror device suspend/resume is called to update DM tables
5) during the suspend/resume cycle, *pre*-commit metadata is read;
however, since the failed devices are now back online, we get back
inconsistent set of precommit metadata and the whole operation fails
The patch relaxes the check that fails in step 5 above, namely by ignoring
inconsistencies coming from PVs that are marked MISSING.
Avoid using of already released memory when duplicated MDA is found.
As get_pv_from_vg_by_id() may call lvmcache_label_scan() use the local copy
of the vgname and vgid on the stack as vginfo may dissapear and code was
then accessing garbage in memory.
i.e. pvs /dev/loop0
(when /dev/loop0 and /dev/loop1 has same MDA content)
Invalid read of size 1
at 0x523C986: dm_hash_lookup (hash.c:325)
by 0x440C8C: vginfo_from_vgname (lvmcache.c:399)
by 0x4605C0: _create_vg_text_instance (format-text.c:1882)
by 0x46140D: _text_create_text_instance (format-text.c:2243)
by 0x47EB49: _vg_read (metadata.c:2887)
by 0x47FBD8: vg_read_internal (metadata.c:3231)
by 0x477594: get_pv_from_vg_by_id (metadata.c:344)
by 0x45F07A: _get_pv_if_in_vg (format-text.c:1400)
by 0x45F0B9: _populate_pv_fields (format-text.c:1414)
by 0x45F40F: _text_pv_read (format-text.c:1493)
by 0x480431: _pv_read (metadata.c:3500)
by 0x4802B2: pv_read (metadata.c:3462)
Address 0x652ab80 is 0 bytes inside a block of size 4 free'd
at 0x4C2756E: free (vg_replace_malloc.c:366)
by 0x442277: _free_vginfo (lvmcache.c:963)
by 0x44235E: _drop_vginfo (lvmcache.c:992)
by 0x442B23: _lvmcache_update_vgname (lvmcache.c:1165)
by 0x443449: lvmcache_update_vgname_and_id (lvmcache.c:1358)
by 0x443C07: lvmcache_add (lvmcache.c:1492)
by 0x46588C: _text_read (text_label.c:271)
by 0x466A65: label_read (label.c:289)
by 0x4413FC: lvmcache_label_scan (lvmcache.c:635)
by 0x4605AD: _create_vg_text_instance (format-text.c:1881)
by 0x46140D: _text_create_text_instance (format-text.c:2243)
by 0x47EB49: _vg_read (metadata.c:2887)
Add testing script
Actually, we can call vg_set_fid(vg, NULL) instead of calling
destroy_instance for all PV structs and a VG struct - it's the same
code we already have in the vg_set_fid.
Also, allow exactly the same fid to be set again for the same PV/VG
Before, this could end up with the fid destroyed because we destroyed
existing fid first and then we used the new one and we didn't care
whether existing one == new one by chance.
Instead of searching linear list of all LVs, PVs - use created hash tables
also for quick mapping between LV.
(Note - for small number of PVs or LVs the overhead of the hash is bigger).
TODO: Use hash tables in volume_group structure directly.
Attach \0 for proper char* display - otherwise somewhat random message could
be displayed in debug more and read of unpredictable read of uninitilized
memory values could happen.
As code uses strncpy(system_id, NAME_LEN) and doesn't set '\0'
Fix it by always allocating NAME_LEN + 1 buffer size and with zalloc
we always get '\0' as the last byte.
This bug may trigger some unexpected behavior of the string operation
code - depends on the pool allocator.
FIXME: refactor this code to alloc_vg.
Missing free_vg on error_path in lvmcache_get_vg fn. Call destroy_instance
only if the fid is not part of the vg in backup_read_vg fn (otherwise it's
part of the VG we're returning and we definitely don't want to destroy it!).
This is necessary for proper format instance ref_count support. We iterate
over vg->pvs and vg->removed_pvs list and the ref_count is decremented and
then it is destroyed if not referenced anymore.
Since format instances will use own memory pool, it's necessary to properly
deallocate it. For now, only fid is deallocated. The PV structure itself
still uses cmd mempool mostly, but anytime we'd like to add a mempool
in the struct physical_volume, we can just rename this fn to free_pv and
add the code (like we have free_vg fn for VGs).
This is essential for proper format instance ref_count support. We must
use these functions to set the fid everywhere from now on, even the NULL
value!
Format instances can be created anytime on demand and it contains
metadata area information mostly (at least for now, but in the future,
we may store more things here to update/edit in a PV/VG). In case we
have lots of metadata areas, memory consumption will rise. Using cmd
context mempool is not quite optimal here because it is destroyed too
late. So let's use a separate mempool for format instances.
Reference counting is used because fids could be shared, e.g. each PV
has either a PV-based fid or VG-based fid. If it's VG-based, each PV has
a shared fid with the VG - a reference to VG's fid.
Add _lv_postorder_vg() - for calling _lv_postorder() for every LV from VG.
We use this in 2 places - vg_mark_partial_lvs() and vg_validate()
so make it as a one function.
Benefit here is - to use only one cleanup code and avoid
potentially duplicate scans of same LVs.
Accelerate validation loop by using lvname, lvid, pvid hash tables.
Also merge pvl loop into one cycle now - no need to scan the list twice.
List scan is stopped when dm_hash_insert fails.
The error message with loop_counter1 is no longer provided - however
the message has been misleading anyway.
Create new function alloc_vg() to allocate VG structure.
It takes pool_name (for easier debugging).
and also take vg_name to futher simplify code.
Move remainder of _build_vg_from_pds to _pool_vg_read
and use vg memory pool for import functions.
(it's been using smem -> fid mempool -> cmd mempool)
(FIXME: remove mempool parameter for import functions and use vg).
Move remainder of the _build_vg to _format1_vg_read
We allow writing non-orphan PVs only for resize now. The "orphan PV" assert
in pv_write fn uses the "allow_non_orphan" parameter to control this assert.
However, we should find a more elaborate solution so we can remove this
restriction altogether (pv_write together with vg_write is not atomic, we
need to find a safe mechanism so there's an easy revert possible in case of
an error).
If the PV is already part of the VG (so the pv->fid == vg->fid), it makes no
sense to attach the mdas information from PV to a VG. Instead, we read new
PV metadata information from cache and attach it to the VG fid.
This function also sets a reference to a new VG format instance for all PVs
that are part of the VG so the PV-VG interconnection is consistent after the
change.
Add supporting functions to work with the format instance and metadata area
structures stored within the format instance. Add support for simple indexing
of metadata areas using PV id and mda order (for on-disk PV only for now, we
can extend the indexing even for other mdas if needed - we only need to define
a proper key for the index).
Fixing some const warnings - with API change in:
int vg_extend(struct volume_group *vg, int pv_count, const char *const *pv_names,
Change is needed - as lvm2api expects const behaviour here.
So vg_extend() is doing local strdup for unescaping.
skip_dev_dir return const char* from const char* vg_name.
Rest of the patch is cleanup of related warnings.
Also using dm_report_filed_string() API change to simplify
casting in _string_disp and _lvname_disp.
New strategy for memory locking to decrease the number of call to
to un/lock memory when processing critical lvm functions.
Introducing functions for critical section.
Inside the critical section - memory is always locked.
When leaving the critical section, the memory stays locked
until memlock_unlock() is called - this happens with
sync_local_dev_names() and sync_dev_names() function call.
memlock_reset() is needed to reset locking numbers after fork
(polldaemon).
The patch itself is mostly rename:
memlock_inc -> critical_section_inc
memlock_dec -> critical_section_dec
memlock -> critical_section
Daemons (clmvd, dmevent) are using memlock_daemon_inc&dec
(mlockall()) thus they will never release or relock memory they've
already locked memory.
Macros sync_local_dev_names() and sync_dev_names() are functions.
It's better for debugging - and also we do not need to add memlock.h
to locking.h header (for memlock_unlock() prototyp).
Add configurable option to define minimal size of
of block device usable as a PV.
pv_min_size() is added to lvm-globals and it's being
initialized through _process_config.
Macro PV_MIN_SIZE is unused and removed.
New define DEFAULT_PV_MIN_SIZE_KB is added to lvm-global
and unlike PV_MIN_SIZE it uses KB units.
Should help users with various slow devices attached to the system,
which cannot be easily filtered out (like FDD on /dev/sdX):
https://bugzilla.redhat.com/show_bug.cgi?id=644578
Set cmd->independent_metadata_areas if metadata/dirs or disk_areas in use.
- Identify and record this state.
Don't skip full scan when independent mdas are present even if memlock is set.
- Clusters and OOM aren't supported, so no problem doing the proper scans.
Avoid revalidating the label cache immediately after scanning.
- A simple optimisation.
Support scanning for a single VG in independent mdas.
- Not used by the fix but I left it in anyway as later patches might use it.
This reset of vgmem pointer causes access of already released memory.
(_vg_make_handle allocates vg from vgmem pool itself - which is a bit tricky)
Interestingly this memory fault was missed by our test suite.
Set vg to NULL after releasing it as the following memlock() test may
lead to goto for the second call of vg_release() with the already
released vg pointer.
Use _even_rand() function instead of floor() in _bitset_with_random_bits().
floor() function is missing in dietlibc (on architectures other than x86).
Moreover using floor() to clip rand results does not assure even result
distribution. _even_rand() uses integer arithmetic only and is designed to
return evenly distributed results.
> Looks OK to me. It took a while to decipher what is the exact meaning of
> the loop in _even_rand (to a non-pseudorandomness-expert) but I am
> fairly comfortable with it now. If I understand this correctly, it
> rejects numbers that come from an "incomplete" slice of the RAND_MAX
> space (considering the number space [0, RAND_MAX] is divided into some
> "max"-sized slices and at most a single smaller slice, between [n*max,
> RAND_MAX] for suitable n -- numbers from this last slice are discarded
> because they could distort the distribution in favour of smaller
> numbers).
Signed-off-by: Przemyslaw Iskra <sparky <at> pld-linux.org>
Reviewed-by: Petr Rockai <prockai <at> redhat.com>
In other LVM memory structures such as volume_group, the field
used to store flags is called "status", and on-disk fields are called
'flags', so rename the one inside metadata_area to be consistent.
Not only is it more consistent with existing code but is cleaner
to say "the status of this mda is ignored".
Background for this patch - prajnoha pinged me on IRC this morning
about a fix he was working on related to metadataignore when
metadata/dirs was set. I was reviewing my patches from this year
and realized the 'flags' field was probably not the best choice
when I originally did the metadataignore patches.
Move the creating of the 'attr' strings into a common function so
they can be called from the 'disp' functions as well as the new
'get' property functions.
Add "_dup" suffix to indicate memory is allocated.
Refactor pvstatus_disp to take pv argument and call pv_attr_dup().
This patch is similar to the other patches for pv and vg
functionality, and separates lv functionality into separate
files, concentrating on reporting fields and simple functions.
The metadata.[ch] files are very large. This patch makes a first
attempt at separating out pv functions and data, particularly
related to the reporting fields calculations.
More code could be moved here but for now I'm stopping at reporting
functions 'get' / 'set' functions.
The metadata.[ch] files are very large. This patch makes a first
attempt at separating out vg functions and data, particularly
related to the reporting fields calculations.
Add "devices/default_data_alignment" to lvm.conf to control the internal
default that LVM2 uses: 0==64k, 1==1MB, 2==2MB, etc.
If --dataalignment (or lvm.conf's "devices/data_alignment") is specified
then it is always used to align the start of the data area. This means
the md_chunk_alignment and data_alignment_detection are disabled if set.
(Same now applies to pvcreate --dataalignmentoffset, the specified value
will be used instead of the result from data_alignment_offset_detection)
set_pe_align() still looks to use the determined default alignment
(based on lvm.conf's default_data_alignment) if the default is a
multiple of the MD or topology detected values.
In all top vg read functions only LCK_VG_READ/WRITE can be used.
All other vg lock definitions are low-level backend machinery.
Moreover, LCK_WRITE cannot be tested through bitmask.
This patch fixes these mistakes.
For _recover_vg() we do not need lock_flags, it can be only
two of above and we always upgrading to LCK_VG_WRITE lock there.
(N.B. that code is racy)
There is no functional change in code (despite wrong masking
it produces correct bits:-)
One shiny day we should use libblkid here. But now using LUKS is
very common together with LVM and pvcreate destroys LUKS completely.
So for user's convenience, try to detect LUKS signature and allow abort.
pvcreate detects MD and swap signature.
The logic hidden there is not only documented but it is also
user unfriendly. Who invented this logic should run pvcreate
on its own critical MD device to see why;-)
This patch
- creates one function instead of duplication code
- asks if user want to overwrite signature
- allows aborting (!)
(Please note that writing LVM signatute without wiping old
is wrong, it confuses blkid, MD will not work anyway and
swap and LUKS is broken too.)
The new standard in the storage industry is to default alignment of data
areas to 1MB. fdisk, parted, and mdadm have all been updated to this
default.
Update LVM to align the PV's data area start (pe_start) to 1MB. This
provides a more useful default than the previous default of 64K (which
generally ended up being a 192K pe_start once the first metadata area
was created).
Before this patch:
# pvs -o name,vg_mda_size,pe_start
PV VMdaSize 1st PE
/dev/sdd 188.00k 192.00k
After this patch:
# pvs -o name,vg_mda_size,pe_start
PV VMdaSize 1st PE
/dev/sdd 1020.00k 1.00m
The heuristic for setting the default alignment for LVM data areas is:
- If the default value (1MB) is a multiple of the detected alignment
then just use the default.
- Otherwise, use the detected value.
In practice this means we'll almost always use 1MB -- that is unless:
- the alignment was explicitly specified with --dataalignment
- or MD's full stripe width, or the {minimum,optimal}_io_size exceeds
1MB
- or the specified/detected value is not a power-of-2
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>
- 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.
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.
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
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).
In preparation to call this from both pvcreate as well as pvchange,
move the guts of metadataignore into a library function.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
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>
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>
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>
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>
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>
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>
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>
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>
With the addition of ignored mdas, we replace all checks for an empty
mda list with a new function to look for either an empty mda list or
ignored mdas.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
