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.
The code fail to account for the case where we just need a single device
in a RAID 4/5/6 array. There is no good way to tell the allocation functions
that we don't need parity devices when we are allocating just a single device.
So, I've used a bit of a hack. If we are allocating an area_count that is <=
the parity count, then we can assume we are simply allocating a replacement
device (i.e. no need to include parity devices in the calculations). This
should make sense in most cases. If we need to allocate replacement devices
due to failure (or moving), we will never allocate more than the parity count;
or we would cause the array to become unusable. If we are creating a new device,
we should always create more stripes than parity devices.
Read lvm.conf setting for monitoring for each command. So we should not
activate monitoring if the default compilation is set to monitor during
lvconvert commnads.
Patch also removes check for clustered VG and allows to disable monitoring
for clustered VG with the assumption, the problem with monitoring and dmeventd
flag passing for INGNORE is already fixed.
It was not possible to pass down the DM_[FORCE|NO]SYNC flags to
'dm_tree_node_add_raid_target'. This meant that converting to 'raid1' from
'mirror' would cause a full resync. (It also meant that '--nosync' was
ineffective when creating a 'raid1' LV.)
I've taken the 'reserved' parameter in 'dm_tree_node_add_raid_target' and
used it for the "flags" parameter. Now it is possible to pass the sync
flags and any other flags that may come up.
Commit 02f6f4902f introduced a bug that caused
RAID devices to fail to activate if the device for a single sub-LV failed.
The special case of LVM mirror was handled, but not LVM RAID.
EXAMPLE:
[root@bp-01 ~]# devices vg
LV Copy% Devices
lv 100.00 lv_rimage_0(0),lv_rimage_1(0)
[lv_rimage_0] /dev/sde1(1)
[lv_rimage_1] /dev/sdh1(1)
[lv_rmeta_0] /dev/sde1(0)
[lv_rmeta_1] /dev/sdh1(0)
[root@bp-01 ~]# vgchange -an vg
0 logical volume(s) in volume group "vg" now active
[root@bp-01 ~]# off.sh sdh
Turning off sdh
[root@bp-01 ~]# vgchange -ay vg --partial
Partial mode. Incomplete logical volumes will be processed.
Couldn't find device with uuid fbI0YO-GX7x-firU-Vy5o-vzwx-vAKZ-feRxfF.
Cannot activate vg/lv_rimage_1: all segments missing.
0 logical volume(s) in volume group "vg" now active
AFTER this patch:
[root@bp-01 ~]# vgchange -ay vg --partial
Partial mode. Incomplete logical volumes will be processed.
Couldn't find device with uuid fbI0YO-GX7x-firU-Vy5o-vzwx-vAKZ-feRxfF.
1 logical volume(s) in volume group "vg" now active
[root@bp-01 ~]# devices vg
Couldn't find device with uuid fbI0YO-GX7x-firU-Vy5o-vzwx-vAKZ-feRxfF.
LV Copy% Devices
lv 100.00 lv_rimage_0(0),lv_rimage_1(0)
[lv_rimage_0] /dev/sde1(1)
[lv_rimage_1] unknown device(1)
[lv_rmeta_0] /dev/sde1(0)
[lv_rmeta_1] unknown device(0)
[root@bp-01 ~]# dmsetup table vg-lv; dmsetup status vg-lv
0 1024000 raid raid1 3 0 region_size 1024 2 253:2 253:3 - -
0 1024000 raid raid1 2 AD 1024000/1024000
No WHATSNEW update necessary because this is an intrarelease fix.
brassow
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.
label_exit() is called destroy_toolcontext() and we are now
using standard dm_list function for destroy, we have to make sure
dm_list gets initialized properly.
As atoi may return negative value - test for both limits.
Test log_args for limits before calling alloca().
Code from dmeventd mirror plugin should probably share same code as
we have in mirrored.c.
Since the function dev_close() has code path, which really could close
file (for unlocked vg) and destroy dev handler, stay on safe side and move
the close few lines later, even our current use case shouldn't trigger
such scenario.
Count number of error and existing areas and if there is no existing area
for the LV avoid its activation.
Always disable partial activatio for thin volumes.
For mirrors currently put in hack to let it pass with a special name
since current mirror code needs to activate such LV during some operations.
For reading % of mapped size of thin volume use as origin for
old style snapshot '-real' device needs to be queried.
Fix log_error report given for lvs -a in this case.
Before removing thin pool LV always make sure, stacked message
for previous run are cleared - but allow to remove any
device that should have been created
(i.e. creation of snapshot failed - so the message for snapshot creation
may be replaced with delete message within unfinished transaction).
Also commit messages after lv remove - so free space is released in pool.
Extend the usage of origin_only flag to allow resume of thin pool LV
(when it's active) to pass only the messages.
origin_only flag will skip detection of already resumed tree for thin_pool,
so we do not need to suspend the tree and we just send messages.
Pass in the origin_only flag also for thin volumes - but curently the flag
is not used to its best.
FIXME: achieve the state where only thin volume snapshot origin is
suspended without its childrens - let's explore whether this may
happen automatically inside libdm (might be generic for other targets).
So the code would not need to annotate the node for this.
Extend lv_activate_opts with bool flag to know for which purpose
dtree is created - and add message only for activation tree
(since that's the only place that may send them).
Extend validation check for thin snapshot creation and test whether
active snapshot origin is suspended before its snapshot is created
(useful in recover scenarios) - in this case also detect, whether
transaction has been already completed and avoid such suspend check
failure in that case.
Add pool_has_message and use it in attach_pool_message.
Also update header to make more obvious which segment type is
expected as parameter.
Rename 'read_only' to 'no_update' (no auto update transaction_id)
to better fit how it's used.
Fix problem when there was only one stacked message replaced with delete
message that caused unwanted transaction_id increase.
Failure to do so results in "Performing unsafe table load while X device(s) are
known to be suspended" errors. While fixing the problem in this way works and
is consistent with the way the mirror segment type does it, it would be nice
to find a solution that uses the generic suspend/resume calls.
Also included in this check-in are additions to the test suite that perform
conversions on RAID LVs under a snapshot. These tests are disabled for the
time being due to a kernel bug that is yet to be tracked down.
Similar to the "mirror" segment type's log device, _add_dev_to_dtree should
be called and not _add_lv_to_dtree when adding metadata sub-LVs to the deptree.
Since _add_lv_to_dtree was being called, 'origin_only' could be set if a
snapshot sits on top of the RAID device. This would cause the actual device
that needed to be added to be skipped in favor of the non-existant device,
"<foo>-real".
Reformat name and path how the LV is represented with lvm1 compatible option,
to switch to the old way - which had number of problem - i.e. many links
do not exist - since for private devices we are not creating them.
Add more info about thin pools and volumes.
Since striped name function knows when to report 'linear' instead of
'stripe' type name - drop it from this place.
This fixes problem when reporting segtype e.g. for thin-pool which
is also using area_count=1 to store thin data device reference.
It also returns properly strduped memory instead of badly casted const char*.
This patch to the suspend code - like the similar change for resume -
queries the lock mode of a cluster volume and records whether it is active
exclusively. This is necessary for suspend due to the possibility of
preloading targets. Failure to check to exclusivity causes the cluster target
of an exclusively activated mirror to be used when converting - rather than
the single machine target.
Since snapshot needs to suspend origin - it might lead to pool userspace
deadlock (as the pool will wait for new space in case it would be overfilled,
but dmeventd would not be able to resize it, as the lvcreate operation would
have kept the VG lock.)
To minimize the risk of such scenario - we prevent to create new snapshot
in case we are over the threshold - but beware, there is still small timewindow,
so keep threshold at some reasonable level!
New field Data% is able to display info about
thin_pool, thin, snapshot and has generic meaning here.
Simple Time/Host field are here to display host and time creation.
Basic support to keep info when the LV was created.
Host and time is stored into LV mda section.
FIXME: Current version doesn't support configurable string via lvm.conf
and used fixed version strftime "%Y-%m-%d %T %z".
This value returns percentage of 'mapped' size compared with total LV size.
(Without passed seg pointer it return highest mapped size - but it's
not used yet.)
LVM- prefix.
Try harder not to leave stray empty devices around (locally or remotely) when
reverting changes after failures while there are inactive tables.
Since the !(dev->flags & DEV_REGULAR) code path just called
dev_name_confirmed() which has just called 'stat()' inside,
remove duplicate second stat() call here.
When both path have identical prefix i.e. /dev/disk/by-id
skip 2 x lstat() for /dev /dev/disk /dev/disk/by-id
and directly lstat() only different part of the path.
Reduces amount of lstat calls on system with lots of devices.
The RAID plug-in for dmeventd now calls 'lvconvert --repair' to address failures
of devices in a RAID logical volume. The action taken can be either to "warn"
or "allocate" a new device from any spares that may be available in the
volume group. The action is designated by setting 'raid_fault_policy' in
lvm.conf - the default being "warn".
Also, don't allow a splitmirror operation on a RAID LV that is already tracking
a split, unless the operation is to stop the tracking and complete the split.
Example:
~> lvconvert --splitmirrors 1 --trackchanges vg/lv /dev/sdc1
# Now tracking changes - image can be merged back or split-off for good
~> lvconvert --splitmirrors 1 -n new_name vg/lv /dev/sdc1
# ^ Completes split ^
If a split is performed on a RAID that is tracking an already split image and
PVs are provided, we must ensure that
1) the already split LV is represented in the PVs
2) we are careful to split only the tracked image
RAID is not like traditional LVM mirroring. LVM mirroring required failed
devices to be removed or the logical volume would simply hang. RAID arrays can
keep on running with failed devices. In fact, for RAID types other than RAID1,
removing a device would mean substituting an error target or converting to a
lower level RAID (e.g. RAID6 -> RAID5, or RAID4/5 to RAID0). Therefore, rather
than removing a failed device unconditionally and potentially allocating a
replacement, RAID allows the user to "replace" a device with a new one. This
approach is a 1-step solution vs the current 2-step solution.
example> lvconvert --replace <dev_to_remove> vg/lv [possible_replacement_PVs]
'--replace' can be specified more than once.
example> lvconvert --replace /dev/sdb1 --replace /dev/sdc1 vg/lv
LVM metadata knows only of striped segments - not linear ones.
The activation code detects segments with a single stripe and switches
them to use the linear target.
If the new lvm.conf setting is set to 0 (e.g. in a test script), this
'optimisation' is turned off.
Remove FIXMES - there should not be any pool free call since
the memory pool is from device manager, and pool is detroyed
after the operation, so doing extra free here would not help here.
However lv_has_target_type() is using cmd mempool so here the extra
call for dm_pool_free makes sence.
"result_independent_of_operands: ((dev->dev & 0xfff00UL) >> 8) ==
18446744073709551615UL /* -1 */ is always false regardless of the values
of its operands (logical operand of if)."
'dev->dev' is set in dev-cache.c _insert() and it's not expectable
st_rdev would have '-1'
This code has been introduced with drbd support commit and code never
worked - so eliminated.
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.
Removal of an inactive origin removes also all related snapshots.
When we now support 'old' external snapshots with thin volumes,
removal of pool will not only drop all thin volumes, but as
a consequence also all snapshots - which might be seen a bit
unexpected for the user - so add a query to confirm such action.
lvremove -f will skip the prompt.
Update region_size only for mirror and raid targets.
This fixes warning messages when vg is using small
extent size like 1KiB and no mirror/raid is created,
but the user still got the message:
$> vgcreate -s 1K vg <pvs>
$> lvcreate -L10K vg
Using reduced mirror region size of 4 sectors
Since we support snapshots of thin volumes, we could have more layers,
so we have to check whether tpool layer is going to be inserted.
As the _add_segment_to_dtree() is the only place that adds tpool
segment, we may just check pointer (no strcmp for layer).
Switch to use seg_is_ function instead of lv_is_.
When a PV label write is deferred to a vg_write call (as introduced by a patch
in 2.02.86), the PV is flagged with the internal UNLABELLED_PV flag. However,
when calling vg_archive before vg_write, we still have the PV labelled with the
UNLABELLED_PV flag which was not recognised as a proper flag while exporting
VG metadata:
# vgcreate vg /dev/sda
No physical volume label read from /dev/sda
Metadata inconsistency: Not all flags successfully exported.
Metadata inconsistency: Not all flags successfully exported.
Writing physical volume data to disk "/dev/sda"
Physical volume "/dev/sda" successfully created
Volume group "vg" successfully created
Remove DM_THIN_ERROR_DEVICE_ID from API.
Remove API warning.
Drop code that was using DM_THIN_ERROR_DEVICE_ID (already commented)
Remove debug message which slipped in through some previous commit.
Add filter which tries to check if scanned device is part
of active multipath.
Firstly, only SCSI major number devices are handled in filter.
Then it checks if device has exactly one holder (in sysfs) and
if it is device-mapper device and DM-UUID is prefixed by "MPATH-".
If so, this device is filtered out.
The whole filter can be switched off by setting
mpath_component_detection in lvm.conf.
https://bugzilla.redhat.com/show_bug.cgi?id=597010
Signed-off-by: Milan Broz <mbroz@redhat.com>
If the extent_size is smaller then the chunk_size we may try
to find better aligment (wasting less space).
i.e. using 4KB extent_size and 64KB chunk size will
lead to creation of 64KB aligned thin volume.
Support allocation of metadata from the same PV, if the VG
is build only from one PV.
As thinp is not mirror - we do not require 2 PVs
for basic thin usage as user is losing only perfomance.
Replace detach_pool_messages with update_pool_lv.
Move creation code from to 'if' condition into 1.
Ensure creation has finished all previous message operations.
Let's put the overlay device over real thin pool device.
So we can get the proper locking on cluster.
Overwise the pool LV would be activate once implicitely
and in other case explicitely, confusing locking mechanism.
This patch make the activation of pool LV independent on
activation of thin LV since they will both implicitely use
real -thin pool device.
All thins are created with the next activation and VG is updated
without messages. Only some basic commands works.
(i.e. lvcreate -an -V10 -T mvg/pool)
There can be some combination to confuse this system.
This functionality for snapshots is going to be interesting.
If something fails during creation of thin LV remove such LV
and deactivate in case it's been already tried to activate
(i.e. thin kernel driver fails for some reason.)
To ensure we properly handle LV cluster locking - explicitely do
not allow to change the availability of the thin pool that is in use
for some thin LV.
As soon as the thin volume is created the only way to activate pool
is via implicit dependency.
Ignore thinpool open count for lv/vgchange operations.
Before adding a new virtual segment to LV, check first whether
the last segment isn't already of the same type. In this case
extend last segment instead of creating the new one.
Thin volumes should have always only 1 virtual segment, but it
helps also to virtual snapshot or error segtype..
Git commit ID 0864378250 was meant to disallow
'mirrored' logs for cluster mirrors. However, when add_mirror_log is used
to create the log (as is now the case when using 'lvcreate' or converting only
the log) the check is bypassed.
This patch adds the check to add_mirror_log.
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>
Since thin is not able to use _lv_insert_empty_sublvs,
remove its appearence from this function.
Start to use extend_pool() function for desired functionality
and modify lv_extend() for this.
Code in _lv_insert_empty_sublvs was not able to provide proper
initialization order for thin pool LV.
New function extend_pool() first adds metadata segment to pool LV which
is still visible. Such LV is activate and cleared.
Then new meta LV is created and metadata segments are moved there.
Now the preallocated pool data segment is attached to the pool LV
and layer _tpool is created. Finaly segment is marked as thin_pool.
This function could be useful for other _manip source files.
Use dm_list manipulation function for provided functionality,
which make the code more readable and avoid touching list
internal details here.
This was the intended behaviour, as described in the lvchange man page, so you
have complete control through volume_list in lvm.conf, but the code seems to
have been treating -ae as local-only for a very long time.
Couple FIXMEs put into the code for parts of the code which may be
improved later, since we might be able to add 'lazy' device creation later.
For now require exclusive activation.
lvm part of messaging.
Each message is now stored it's own thin pool section:
message1 {
create = lv
}
Messages are queued to thin pool dm target when this target
is going to be resumed or used through some dependency.
Currently 'delete' message are purely queued and processed
with next thin pool resume operation (i.e. create_thin).
WARNING - thin provisioning support is developmental code.
When verify_udev_operations was disable, code for stacking fs operation for
lvm links was completely disable - but this code was also used for collecting
information, that a new node is being created.
Add a new flag which is set when a creation of lv symlinks is requested which
should restore old behaviour of lv_info function, that has called fs_sync()
before quere for open count on device.
Go with just 64KiB for stack.
Closer inspection should be made, whether we actually need to play with
settings at all.
Since default stack size is 8MB and gets mapped via page locking thus,
it seems there is no big help with preallocation of stack to some value.
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.
Using log_warn to report missing symlinks as warning, since the command
itself returns as successful, we should not produce log_error().
log_warn is better fit here.
Cosmetic, since r is already 0 for the error path, no need to assign it there,
and r is assigned to 1 after switch command.
Also makes the code more readable.
Example:
~> lvconvert --type raid1 vg/mirror_lv
Steps to convert "mirror" to "raid1"
1) Allocate a RAID metadata LV for each mirror image from the same PVs
on which they are located.
2) Clear the metadata LVs. This involves writing LVM metadata, so we don't
change any aspects of the mirror LV before this so that the user can easily
remove LVs from the failed convert attempt while retaining the original
mirror.
3) Remove the mirror log, if it exists.
4) Add metadata LVs to mirror LV
5) Rename mirror sub-lvs (s/mimage/rimage/)
6) Change flags and segtype from mirror to raid1
Example:
~> lvconvert --type raid1 -m 1 vg/lv
The following steps are performed to convert linear to RAID1:
1) Allocate a metadata device from the same PV as the linear device
to provide the metadata/data LV pair required for all RAID components.
2) Allocate the required number of metadata/data LV pairs for the
remaining additional images.
3) Clear the metadata LVs. This performs a LVM metadata update.
4) Create the top-level RAID LV and add the component devices.
We want to make any failure easy to unwind. This is why we don't create the
top-level LV and add the components until the last step. Should anything
happen before that, the user could simply remove the unnecessary images. Also,
we want to ensure that the metadata LVs are cleared before forming the array to
prevent stale information from polluting the new array.
A new macro 'seg_is_linear' was added to allow us to distinguish linear LVs
from striped LVs.
This patch allows a mirror to be extended without an initial resync of the
extended portion. It compliments the existing '--nosync' option to lvcreate.
This action can be done implicitly if the mirror was created with the '--nosync'
option, or explicitly if the '--nosync' option is used when extending the device.
Here are the operational criteria:
1) A mirror created with '--nosync' should extend with 'nosync' implicitly
[EXAMPLE]# lvs vg; lvextend -L +5G vg/lv ; lvs vg
LV VG Attr LSize Pool Origin Snap% Move Log Copy% Convert
lv vg Mwi-a-m- 5.00g lv_mlog 100.00
Extending 2 mirror images.
Extending logical volume lv to 10.00 GiB
Logical volume lv successfully resized
LV VG Attr LSize Pool Origin Snap% Move Log Copy% Convert
lv vg Mwi-a-m- 10.00g lv_mlog 100.00
2) The 'M' attribute ('M' signifies a mirror created with '--nosync', while 'm'
signifies a mirror created w/o '--nosync') must be preserved when extending a
mirror created with '--nosync'. See #1 for example of 'M' attribute.
3) A mirror created without '--nosync' should extend with 'nosync' only when
'--nosync' is explicitly used when extending.
[EXAMPLE]# lvs vg; lvextend -L +5G vg/lv; lvs vg
LV VG Attr LSize Pool Origin Snap% Move Log Copy% Convert
lv vg mwi-a-m- 20.00m lv_mlog 100.00
Extending 2 mirror images.
Extending logical volume lv to 5.02 GiB
Logical volume lv successfully resized
LV VG Attr LSize Pool Origin Snap% Move Log Copy% Convert
lv vg mwi-a-m- 5.02g lv_mlog 0.39
vs.
[EXAMPLE]# lvs vg; lvextend -L +5G vg/lv --nosync; lvs vg
LV VG Attr LSize Pool Origin Snap% Move Log Copy% Convert
lv vg mwi-a-m- 20.00m lv_mlog 100.00
Extending 2 mirror images.
Extending logical volume lv to 5.02 GiB
Logical volume lv successfully resized
LV VG Attr LSize Pool Origin Snap% Move Log Copy% Convert
lv vg Mwi-a-m- 5.02g lv_mlog 100.00
4) The 'm' attribute must change to 'M' when extending a mirror created without
'--nosync' is extended with the '--nosync' option. (See #3 examples above.)
5) An inactive mirror's sync percent cannot be determined definitively, so it
must not be allowed to skip resync. Instead, the extend should ask the user if
they want to extend while performing a resync.
[EXAMPLE]# lvchange -an vg/lv
[EXAMPLE]# lvextend -L +5G vg/lv
Extending 2 mirror images.
Extending logical volume lv to 10.00 GiB
vg/lv is not active. Unable to get sync percent.
Do full resync of extended portion of vg/lv? [y/n]: y
Logical volume lv successfully resized
6) A mirror that is performing recovery (as opposed to an initial sync) - like
after a failure - is not allowed to extend with either an implicit or
explicit nosync option. [You can simulate this with a 'corelog' mirror because
when it is reactivated, it must be recovered every time.]
[EXAMPLE]# lvcreate -m1 -L 5G -n lv vg --nosync --corelog
WARNING: New mirror won't be synchronised. Don't read what you didn't write!
Logical volume "lv" created
[EXAMPLE]# lvs vg
LV VG Attr LSize Pool Origin Snap% Move Log Copy% Convert
lv vg Mwi-a-m- 5.00g 100.00
[EXAMPLE]# lvchange -an vg/lv; lvchange -ay vg/lv; lvs vg
LV VG Attr LSize Pool Origin Snap% Move Log Copy% Convert
lv vg Mwi-a-m- 5.00g 0.08
[EXAMPLE]# lvextend -L +5G vg/lv
Extending 2 mirror images.
Extending logical volume lv to 10.00 GiB
vg/lv cannot be extended while it is recovering.
7) If 'no' is selected in #5 or if the condition in #6 is hit, it should not
result in the mirror being resized or the 'm/M' attribute being changed.
NOTE: A mirror created with '--nosync' behaves differently than one created
without it when performing an extension. The former cannot be extended when
the mirror is recovering (unless in-active), while the latter can. This is
a reasonable thing to do since recovery of a mirror doesn't take long (at
least in the case of an on-disk log) and it would cause far more time in
degraded mode if the extension w/o '--nosync' was allowed. It might be
reasonable to add the ability to force the operation in the future. This
should /not/ force a nosync extension, but rather force a sync'ed extension.
IOW, the user would be saying, "Yes, yes... I know recovery won't take long
and that I'll be adding significantly to the time spent in degraded mode, but
I need the extra space right now!".
This patch also does some clean-up of the splitmirrors code.
I've attempted to clean-up the splitmirrors code to make it easier to
understand with fewer operations. I've tried to reduce the number of
metadata operations without compromising the intermediate stages which
are necessary for easy clean-up in the even of failure.
These changes now correctly handle cluster situations - including exclusive
cluster mirrors. Whereas before, a splitmirror operation would result in
remote nodes having LVM commands report the newly split LV with a proper
name while DM commands would report the old (pre-split) names of the device.
IOW, there was a kernel/userspace mismatch.
The original commit comments can be located via this git commit ID:
7d8e615c0b
There were three possible solutions to the original problem proposed in the
initial check-in. The one chosen was as follows:
2) Do like _remove_mirror_images does and suspend the original, then suspend
the sub-lv (the error target), then resume the sub-lv, and finally resume the
original LV. This seems like extra pointless operations to me, but it doesn't
produce the error message (although, I'm not sure why) and it allows us to
leave the visible flag in place.
Turns out, the cluster also views the extra suspend/resume operations as
pointless too and ignores them. So, this solution doesn't work in a cluster.
Further, I've noticed that in addition to the remote cluster nodes still getting
I/O errors from scanning the error target, they also have a different LVM and
DM views of the same LV. IOW, while the LVM level (gotten from the LVM metadata)
sees the correct name for the newly split LV, device-mapper still maintains the
old names.
Because the original fix failed to completely fix the problem (or work-around it)
and because a better solution must be found to address the additional cluster
issue of device renaming, I am reverting the above mentioned commit.
The current code does not always assign proper udev flags to sub-LVs (e.g.
mirror images and log LVs). This shows up especially during a splitmirror
operation in which an image is split off from a mirror to form a new LV.
A mirror with a disk log is actually composed of 4 different LVs: the 2
mirror images, the log, and the top-level LV that "glues" them all together.
When a 2-way mirror is split into two linear LVs, two of those LVs must be
removed. The segments of the image which is not split off to form the new
LV are transferred to the top-level LV. This is done so that the original
LV can maintain its major/minor, UUID, and name. The sub-lv from which the
segments were transferred gets an error segment as a transitory process
before it is eventually removed. (Note that if the error target was not put
in place, a resume_lv would result in two LVs pointing to the same segment!
If the machine crashes before the eventual removal of the sub-LV, the result
would be a residual LV with the same mapping as the original (now linear) LV.)
So, the two LVs that need to be removed are now the log device and the sub-LV
with the error segment. If udev_flags are not properly set, a resume will
cause the error LV to come up and be scanned by udev. This causes I/O errors.
Additionally, when udev scans sub-LVs (or former sub-LVs), it can cause races
when we are trying to remove those LVs. This is especially bad during failure
conditions.
When the mirror is suspended, the top-level along with its sub-LVs are
suspended. The changes (now 2 linear devices and the yet-to-be-removed log
and error LV) are committed. When the resume takes place on the original
LV, there are no longer links to the other sub-lvs through the LVM metadata.
The links are implicitly handled by querying the kernel for a list of
dependencies. This is done in the '_add_dev' function (which is recursively
called for each dependency found) - called through the following chain:
_add_dev
dm_tree_add_dev_with_udev_flags
<*** DM / LVM divide ***>
_add_dev_to_dtree
_add_lv_to_dtree
_create_partial_dtree
_tree_action
dev_manager_activate
_lv_activate_lv
_lv_resume
lv_resume_if_active
When udev flags are calculated by '_get_udev_flags', it is done by referencing
the 'logical_volume' structure. Those flags are then passed down into
'dm_tree_add_dev_with_udev_flags', which in turn passes them to '_add_dev'.
Unfortunately, when '_add_dev' is finding the dependencies, it has no way to
calculate their proper udev_flags. This is because it is below the DM/LVM
divide - it doesn't have access to the logical_volume structure. In fact,
'_add_dev' simply reuses the udev_flags given for the initial device! This
virtually guarentees the udev_flags are wrong for all the dependencies unless
they are reset by some other mechanism. The current code provides no such
mechanism. Even if '_add_new_lv_to_dtree' were called on the sub-devices -
which it isn't - entries already in the tree are simply passed over, failing
to reset any udev_flags. The solution must retain its implicit nature of
discovering dependencies and be able to go back over the dependencies found
to properly set the udev_flags.
My solution simply calls a new function before leaving '_add_new_lv_to_dtree'
that iterates over the dtree nodes to properly reset the udev_flags of any
children. It is important that this function occur after the '_add_dev' has
done its job of querying the kernel for a list of dependencies. It is this
list of children that we use to look up their respective LVs and properly
calculate the udev_flags.
This solution has worked for single machine, cluster, and cluster w/ exclusive
activation.
Make limits for thin data_block_size and device_id part of public API.
FIXME: read them possible from some kernel header file in the future ?
But we may need to support different values for different versions ?
Before, we used to display "Can't remove open logical volume" which was
generic. There 3 possibilities of how a device could be opened:
- used by another device
- having a filesystem on that device which is mounted
- opened directly by an application
With the help of sysfs info, we can distinguish the first two situations.
The third one will be subject to "remove retry" logic - if it's opened
quickly (e.g. a parallel scan from within a udev rule run), this will
finish quickly and we can remove it once it has finished. If it's a
legitimate application that keeps the device opened, we'll do our best
to remove the device, but we will fail finally after a few retries.
seg->areas and seg->meta_areas. We also need to copy the memory from the
old arrays to the newly allocated arrays. The amount of memory to copy was
determined by seg->area_count. However, seg->area_count was being set to the
higher value after copying the 'seg->areas' information, but before copying
the 'seg->meta_areas' information. This means we were copying more memory
than necessary for 'seg->meta_areas' - something that could lead to a segfault.
Compiler says variable may be used uninitialized. It can't be, but we
initialize the variable to NULL anyway. Also, remove the double initialization
of another variable.
This bug showed up when trying to add a log to a mirror whose images are on
multiple devices. This is an intra-release regression and no WHATS_NEW
entry will be added. The error was introduce in the following commit:
2d8a2f35c7
The solution is to recognise in _alloc_init that if there are no mirrors
or stripes specified, then 'new_extents' should be zero.
to settle udev before calling deactivate_lv.
This is an intra-release regression (no WHATS_NEW entry required). It is
part of the fix for the current WHATS_NEW entry:
Work around resume_lv causing error LV scanning during splitmirror operation.
When user wants to remove thin pool - check if there are no thin volumes using it.
If so - query before removal (or -ff for no question) and remove them first.
When LV is unlinked, we want to catch problem in vg_validate,
that LV has changed.
i.e. catch LV has been removed and is no long thin_pool while still
being referenced by some thin volume.
Revert John patch, which fixed only 1 place where ~LVM_WRITE was in use and
convert ommited LVM_READ/WRITE flags to 64bit constants as well.
(Since both 'status' flags for LV and VG are 64bit.)
Changing lv_mirror_count to only count the AREA_LVs made the function
stop working for PVMOVE mirrors. A conditional has been added to fix
that problem. Additionally, when counting the images in a mirror stack,
we don't need to subtract 1 from the count we get back from the
lv_mirror_count call on the temporary mirror layer. (This is because we
are no falsely counting the top layer of the temporary mirror.)
lv_mirror_count was not able to handle mirrors of stripes properly. When a
failed device is removed, the MIRRORED status flag is removed from the LV
conditionally based on the results of lv_mirror_count. However, lv_mirror_count
trusted the MIRRORED flag - thinking any such LV must be mirrored. It would
happily assign first_seg(lv)->area_count as the number of mirrors, but when
a mirrored striped LV was reduced to a simple striped LV area_count would be
the number of /stripes/ not the number of /mirrors/. A result higher than 1
would be returned from lv_mirror_count, the MIRRORED flag would not be cleared,
and the LV would fail to be up-converted properly in lvconvert_mirrors_aux
because of it.
The operation of deactivating the residual error target LV after removing a
mirror layer can cause a "device in-use" conflict with udev. Giving udev a
poke before calling deactivate_lv eliminates the conflict. The stick used
to poke udev is 'sync_local_dev_names'.
Kernel requires a mirror to be at least 1 region large. So,
if our mirror log is itself a mirror, it must be at least
1 region large. This restriction may not be necessary for
non-mirrored logs, but we apply the rule anyway.
(The other option is to make the region size of the log
mirror smaller than the mirror it is acting as a log for,
but that really complicates things. It's much easier to
keep the region_size the same for both.)
WHATS_NEW entry:
Fix log size calculation when only a log is being added to a mirror.
The original fix pass the mirror LV to allocate_extents (rather than
passing NULL) so that _alloc_init could correctly determine the necessary
size of the mirror log. In the previous check-in, I noted:
In order to get a decent value computed, we need to pass in the 'lv' argument
to allocate_extents. This would normally imply a desire for cling/contiguous
allocation to the given LV, but since we are not allocating any parallel
extents and only log extents, it works fine.
However, passing in the LV did have unintended consequences on the placement of
the log. The better solution is to pass in the number of extext that are in
the mirror LV instead of the LV itself. This will not cause the allocator to
reserve that number of extents, because 'stripes' and 'mirrors' are specified
as 0. Thus, 'extents' is used to calculate the size of the log, but won't
affect how much is allocated.
LVM_WRITE is a 32-bit flag. Now that RAID[_IMAGE|_META] are 64-bit,
and'ing a RAID LV's status against LVM_WRITE can reset the higher order
flags.
A similar thing will affect thinp flags if not careful.
_alloc_init calculates the number of necessary log extents via
'mirror_log_extents'. 'mirror_log_extents' takes 3 arguments: region_size,
pe_size, and size of the mirror LV. Unfortunately, _alloc_init is guessing at
the mirror size by using 'ah->new_extents / ah->area_multiple' - the number of
extents that the mirror images have. However, this is /always/ wrong when
allocating the log separately. Further, the log is always allocated separately
unless we are up-converting the mirror at the same time. It was by luck alone
that a default value of '1' reflects what we want in most cases.
In order to get a decent value computed, we need to pass in the 'lv' argument
to allocate_extents. This would normally imply a desire for cling/contiguous
allocation to the given LV, but since we are not allocating any parallel
extents and only log extents, it works fine.
When an image is split from a 2-way mirror, the original mirror is converted to
a linear device. To do this, the top "layer" must be removed. The segments
are transferred from the sub-lv to the top-level LV and the link is severed.
The former sub-lv - having its segments transferred - now contains a temporary
error target.
When the original LV is resumed, the old sub-lv that now contains an error
segment is activated and scanned. This is what causes the I/O error messages.
There are three ways to fix this problem:
1) Do not set the sub-lv which contains the error target as "visible" before
suspending the original LV. This way, when the original is resumed, the sub-lv
device node is not created and it is not scanned - avoiding the error messages.
The problem with this approach is that if the machine crashes after the
resume, it leaves the *hidden* LV in place and the user has a more difficult
time noticing that it needs to be cleaned up. Thus, this type of processing is
frowned upon.
2) Do like _remove_mirror_images does and suspend the original, then suspend
the sub-lv (the error target), then resume the sub-lv, and finally resume the
original LV. This seems like extra pointless operations to me, but it does not
produce the error message (although, I'm not sure why) and it allows us to
leave the visible flag in place.
3) Flag the sub-lv (error target) with a "do not scan" flag. This seems like
the cleanest approach, but I have been unable to find the method for doing
this. LVs get tagged in such a way by _get_udev_flags, but in this case the
resume of the original LV also resumes the error target LV without running it
through _get_udev_flags (likely because they are no longer linked). Could
there be something wrong in resume_lv?
Option #2 was chosen to fix this bug, but it seems like more of a workaround
for now.
A gentle reminder that anyone relying on the output of reporting commands
like lvs in scripts must use -o to guarantee they get the fields they expect.
The default sequence of fields can change from release to release.
Equally, the 'attr' fields can have new values introduced and/or characters
appended to them.
Makes dumpconfig whole-section output wrong in a different way from before,
but we should be able to merge cft_cmdline properly into cmd->cft now and
remove cascade.
There was a bad sequence:
*) Make changes to LV layout to split images (e.g. 4-way -> 2-way/2-way)
1) vg_write, suspend_lv(original_mirror), vg_commit
2) activate_lv(newly_split_lv)
3) resume_lv(original_mirror)
Step #2 is not allowed. However, without it, the resume of the original
mirror will also resume its former sub-LVs - making it impossible to
activate the newly split LV due to the changes in layering, pointers, and
names that had already been made. Additionally, the resume or the original
brings the sub-lv's online with names that differ from the metadata on disk -
also a no-no. Thus, the split must be done in stages such that the active LVs
always reflect what is in the committed LVM metadata.
First, alter the original mirror by releasing the images. The images are made
visible and independent as an intermediate stage. (This way, we can have
consistency between LVM metadata and active LVs.) The second stage collects
the recently split LVs, deactivates them, forms them into a mirror if necessary,
and then activates them. It is a bit of a circuitous method, but it is the only
way to split a mirror from a mirror and obey these general rules:
1) Never [de]activate sub-lvs when the top-level LV is suspended
2) Avoid having active LVs that differ from the description in the LVM metadata
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
functionality. A number of bugs (copied and pasted all over the code) should
disappear:
- most string lookup based on dm_config_find_node would segfault when
encountering a non-zero integer (the intention there was to print an
error message instead)
- check for required sections in metadata would have been satisfied by
values as well (i.e. not sections)
- encountering a section in place of expected flag value would have
segfaulted (due to assumed but unchecked cn->v != NULL)
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.
There's a very high memory usage when calling _pv_analyse_mda_raw (e.g. while
executing pvck) that can end up with "out of memory".
_pv_analyse_mda_raw scans for metadata in the MDA, iteratively increasing the
size to scan with SECTOR_SIZE until we find a probable config section or we're
at the edge of the metadata area. However, when using a memory pool, we're also
iteratively chasing for bigger and bigger mempool chunk which can't be found
and so we're always allocating a new one, consuming more and more memory...
This patch just changes the mempool to direct memory allocation in this
problematic part of the code.
