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.
- allocate environment dynamically (still missing some limit?)
- try to recover, if destroy failed (do not destroy lvm here) and free memory
- check strdup() return codes
- report failure to log
- do not print NULL in exclusive lock loop
Activate only the first replicator-dev LV, that activates all other
related LVs from Replicator. In case of error during this activation,
it will not retry again for other heads (less confusing error log).
length(array) is specific to GNU awk and doesn't work in mawk.
Use a return value of "split" function to indicate array size, this is
supported in both gawk and mawk.
This patch fixes the following errors during "make install" when mawk is
installed as a default awk.
mawk: scripts/relpath.awk: line 25: illegal reference to array from
mawk: scripts/relpath.awk: line 25: illegal reference to array to
mawk: scripts/relpath.awk: line 27: illegal reference to array from
mawk: scripts/relpath.awk: line 32: illegal reference to array to
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Adding function _add_partial_replicator_to_dtree() to create
partial tree for Replicator target.
Using dm_tree_node_set_presuspend_node() for Replicator.
As for _process_one_vg() we need similar retry loop for
process_each_lv_in_vg(). This patch retries to process
failed LVs with reopened VGs.
Patch does not add any extra repeated invocations if there is not
found any missing VG during LV processing.
Patch modifes behavior of _process_one_vg().
In the first pass vg_read() collectis for replicator sorted list of
additional VGs during lock_vol().
If any other VG is needed by the replicator and it is not yet opened
then next iteration loop is taken with all collected VGs.
Flag vg->cmd_missing_vgs detects missing VGs.
Introduce struct cmd_vg to store information about needed
volume group name, vgid, flags and the pointer to opened VG.
Keep VGs list in alphabetical order for locking order.
Introduce functions:
cmd_vg_add() add new cmd_vg entry.
cmd_vg_lookup() search cmd_vgs for vg_name.
cmd_vg_read() open VGs in cmd_vgs list.
cmd_vg_release() close VGs in reversed order.
Adding configure.in support for Replicators.
Adding basic lib lvm support for Replicators.
Adding flags REPLICATOR and REPLICATOR_LOG.
Adding segments SEG_REPLICATOR and SEG_REPLICATOR_DEV.
Adding basic methods for handling replicator metadata.
For deactivation of Replicator check in advance that all heads
have open_count == 0. For this presuspend_node is used as all
head nodes are linking this control node.
Introducing dm_tree_node_set_presuspend_node() for presuspending child
node (i.e. replicator control target) before deactivation of parent node
(i.e. replicator-dev target).
This patch presents no functional change to current dtree - only
replicator target currently sets presuspend node for dev nodes.
Patch adds failed_lvnames to the list of parameters for process_each_lv_in_vg().
If the list is not NULL it will be filled with LV names of failing LVs
during function execution.
Application could later reiterate only on failed LVs.
are active mirrors or snapshots.
We don't have the mechanisms in place to change the device-mapper
tables for those targets that have behavioral differences between
cluster and single machine instances. Allowing users to change
the attribute but not changing the target's behavior can lead to
data corruption.
The following bugs are fixed/avoided by this patch:
235123 - vgchange -c [ny] do not change target types when necessary
289331 - RFE: switching from cluster domain to local domain needs to deactivate volume somehow
289541 - when changing from local to cluster, volumes can not appear to be deactivated
This should avoid various races between dmeventd on multiple nodes
in cluster where one node already repairing device and another
run full scan and locks the device.
the device cache file is dumped both in vgscan and clvmd process.
Unfortunately, clvmd calls lvmcache_label_scan,
it properly destroys persistent filter, but during
persistent_filter_dump it merges old cache content back!
This causes that change in filters is not properly propagated
into device cache after vgscan on cluster.
(Only new devices are added.)
https://bugzilla.redhat.com/show_bug.cgi?id=591861
Use Requires.private: instead of Libs.private:
Use UDEV_PC and SELINUX_PC for Require.private:
It looks like usage of Requires.private is prefered from Libs.private.
However pkg-config documentation is really poor here. But here is
short outcome:
There is a difference in Libs.private: and Requires.private: where
we specify libselinux instead of -lselinux -lsepol.
We leave resolving of query like 'pkg-config --libs --static devmapper'
on taking proper selinux and udev libs to their .pc files instead of
hardcoding them into our .pc file which is might give incorrect answer.
- i.e. dependency of libselinux package might change and we may return
wrong list of linked libraries.
http://bugs.freedesktop.org/show_bug.cgi?id=4738http://err.no/personal/blog/tech/2008-03-25-18-07_pkg-config,_sonames_and_Requires.private
A shortcut for --ignorelockingfailure, --ignoremonitoring, --poll n options
and LVM_SUPPRESS_LOCKING_FAILURE_MESSAGES environment variable used all at
once in initialisation scripts (e.g. rc.sysinit or initrd).
being able to remove more images from a mirror than the
number of PVs directly specified for removal.
The effort to fix bug 581611 corrected a bug that was unnoticed
at the time. The loop in _remove_mirror_images that looks over
the specified PVs was allowing devices that were previously
counted and moved to the end of the list to be double-counted.
This resulted in the number of devices needed for removal always
being satisfied - even if the user did not specify enough PVs
for removal to satisfy the request. When 581611 was fixed, this
double-counting no longer took place and the result was to remove
only the minimum of the number of PVs specified or the number
that was asked to be removed.
By simply always setting 'new_area_count' (as used to be done
only in the else statement), we return to the previous behavior.
Indeed, this is exactly what the double-counting was allowing
to happen before the fix of 581611.
Allow lv_remove_with_dependencies() to know the top-level LV that was
requested to be removed (otherwise it recurses and we lose context).
A merging snapshot cannot be removed directly but the associated origin
can be. Disallow removal of a merging snapshot unless the associated
origin is also being removed.
There's no need for foreign udev rules to touch LVM reserved devices
(snapshot, pvmove, _mlog, _mimage, _vorigin) even if they happen to
be visible. The same applies for /dev/disk content - no need to create
any content for these devices (and so no need to run any "blkid" etc.).
This also prevents setting any inotify "watch" from udev rules on such
devices that is a source of race conditions (the rules need to honor
DM_UDEV_DISABLE_OTHER_RULES_FLAG for this to work though).
This version number change reflects the memory handling change
for string-based pv/vg/lv string based attributes.
In addition, when adding support for tags, I forgot to increase
the version number.
Signed-off-by: Dave Wysochanski <dwysocha@redhat.com>
We should write metadata into next position in the ring buffer while calling
vgrename and vgcfgrestore. At this code level (_vg_write_raw), we were not able
to determine if this is a rename or not. If yes, then accompanying VG structure
passed here has a new name set, not the old one.
When looking for a location where to put metadata next, we were given a NULL
value because of failed VG name comparison (in _find_vg_rlocn) between the
name in existing metadata and metadata we're just about to write.
This resets the position in the ring buffer, overwriting any existing metadata
(and also incorrectly updates the cache to "orphan" afterwards).
This patch just adds old_name item in struct volume_group that we can check and use
if necessary and detect renames at lower layers as well.
The same applies for vgcfgrestore, but here we're using a special value of
old_name, an empty string, to disable the check with existing metadata totally.
Internally, we used DM names instead of UUIDs while processing event
handlers. This caused problems while trying to vgrename a VG with active LVs
where the names are being changed and so the devices were not found then.
The patch also contains a little bit of refactoring, moving "build_dlid" code
found in dev_manager.c to "build_dm_uuid", now in lvm-string.c (so we have
build_dm_uuid and build_dm_name at one place).
Patch is inspired by Debian's extra patch.
- removes OWNER & GROUP make vars they are parts of INSTALL command.
- adds INSTALL_PROGRAM for executable, uses $(INSTALL)
- adds INSTALL_DATA for non-executable data, uses ($INSTALL)
- adds INSTALL_WDATA for writable non-executable data, uses ($INSTALL)
- adds configure option --enable-write_install - to support
installatin of writable files used by distribution
- replaces usage of ifeq @LIB_SUFFIX@ with $(LIB_SUFFIX)
- installs .a files from static builds without executable flag
- installs .a files to $(usrlibdir) instead of $(libdir)
- installs all static binaries to $(staticdir)
- create .so links for devel package in $(usrlibdir) instead of
$(libdir)
- makes .so and .so.LIB_VERSION files within builddir
- removes VERSIONED_SHLIB and created versioned LIB_SHARED automagicaly
- install LIB_SHARED via install_lib_shared target
- install plugins via install_lib_shared_plugin target
- prints whole 'install' command during installation instead of less
informative "Installing $(something) $(somewhere)"
- install multiple man pages with one INSTALL command
- use DISTCLEAN_TARGETS instead of creating multiple distclean targets
Usage of VPATH makes troubles when used within $(builddir).
Not only source files are being found through VPATH,
but targets as well. (make --debug=v)
Thus if user builds the code in $(srcdir) and also in some $(builddir)
he gets mangled results as some generated files (i.e. .export.sym)
are 'reused' from $(srcdir) instead of $(builddir).
This patch switches to use vpath were we could explicitly name
suffixes that should be looked via vpath - we must take care,
we do not generate files with these suffixes:
.c, .in, .po, .exported_symbols
When moving parts of striped LVs, pvmove wouldn't care about leaving you with
two stripes on the same disk. Now --alloc anywhere is needed for that.
(Tried and gave up on two alternative approaches before the one committed here.)
to check for presence of this module and avoid using --frames
option for genhtml in this case.
Fix arg list for AC_PATH_PROG for lcov and genhtml.
(detecting empty LCOV and GENHTML string in Makefiles).
Because we have now strong rule for lock ordering:
- VG locks must be taken in alphabetical order
- ORPHAN locks must be the last
vgs_locked() is now not needed.
This fixes problem with orphan locking, e.g.
vgremove VG1 | vgremove VG2
lock(VG1) | lock(VG2)
lock(ORPHAN) | lock(ORPHAN) -> fail, non-blocking
https://bugzilla.redhat.com/show_bug.cgi?id=578413
(More similar places in code.)
Physical segments were still allocated from global
command context mempool.
This leads to very high memory usage when
activating large VG (vgchange).
(Memory usage was about 2G when >3000LVs).
Fix it by properly using vg->vgmem private pool,
so all the memory is released early.
New memory pool parameter is needed here for pv_split_segment
function.
Also fix the same problem in some minor allocations
(vg description, lv segment split).
In addition to previous patch, we really do not need
to search for segment which was just allocated in
split request.
Make pv_split_segment function return newly allocated
(split) segment also.
(So after this patch, there is only one user
of slow find_peg_by_pe).
The function find_peg_by_pe is incredibly inefficient
for Pvs with many segments.
In shiny future there should be binary (or interval) tree
instead of sorted linked list (volunteers?).
Anyway, for now, we can use dirty trick here to optimise this case:
- Allocations are usually applied from the beginning
of PV (we have no alloocation policy which allocates areas
"backwards")
- The only user of find_peg_by_pe is pv_split_segment()
call. In *most* cases it need to split *last* PV segment.
So if we search sorted pv segment list backwards, we
hit the requested segment immediatelly.
This patch applies this tiny change.
(and saves >30% of processing time when >3000LVs segments are on one PV!)
To discourage using this inefficient function from other code,
it is moved to pv_manip.c and used static for now:-)
vg_validate call is an adept to optimisation, it is very
ineeficient and slow.
Anyway, we should call it only before writing data to disk.
The call in lvmcache was just temporary validation,
we realy do not need to revalidate cached metadata
every time.
(Actually, I added that there just to prove that cache works
properly and forgot to remove it.)
Patch removes it from lvmcache completely, this can hit only
internal bug in export function (and this bug must
be detected in any vg_write call anyway before).
The _read_vg uses already hash for PVs to optimise
reading of large VGs and avoiding repeated PV list traversing.
Use the same aproach to speed up parsing VG with many LVs.
If dmeventd runs with -d flag, it doesn't fork into backgroud.
The command kill(getppid(), SIGTERM) attempts to kill the parent dmeventd
process, however, if there is no parent, it kills whatever process spawned
dmeventd. In case of debugging with gdb, the parent is gdb, thus
kill(getppid(), SIGTERM) kills the debugger.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
clvmd does not propagate DMEVENTD_MONITOR_IGNORE.
Update get_activation_monitoring_mode() to check if the VG that the
LV is being activated in is clustered. If so, skip it.
Any get_activation_monitoring_mode() error will cause the associated LV
(or VG) to be skipped during activation. Both vgchange_single() and
lvchange_single(), which call get_activation_monitoring_mode(), are
called by their respective process_each_..() method.
in clvmd, dmevend, man, tests.
Don't include dependency files for clow and cscope.out targets
Improve dependency tracking for dmeventd and liblvm2cmd sources.
This check-in enables the 'mirrored' log type. It can be specified
by using the '--mirrorlog' option as follows:
#> lvcreate -m1 --mirrorlog mirrored -L 5G -n lv vg
I've also included a couple updates to the testsuite. These updates
include tests for the new log type, and some fixes to some of the
*lvconvert* tests.
clvmd's do_lock_lv() already properly controls dmeventd monitoring based
on LCK_DMEVENTD_MONITOR_MODE in lock_flags -- though one small fix was
needed for this to work: _lock_for_cluster() must treat
dmeventd_monitor_mode()'s return as a tri-state value.
Also cleanup do_lock_lv() to:
- explicitly init_dmeventd_monitor() based on LCK_DMEVENTD_MONITOR_MODE
- no longer reset init_dmeventd_monitor() to default at the end of
do_lock_lv() -- it is unnecessary
This is the next preparatory step towards better --alloc anywhere
support and is not intended to break anything that currently works so
please report any problems - segfaults, bogus data in the new debug
messages, or if the code now chooses bizarre allocation layouts.
. Add "monitoring" option to "activation" section of lvm.conf
. Have clvmd consult the lvm.conf "activation/monitoring" too.
. Introduce toollib.c:get_activation_monitoring_mode().
. Error out when both --monitor and --ignoremonitoring are provided.
. Add --monitor and --ignoremonitoring support to lvcreate. Update
lvcreate man page accordingly.
. Clarify that '--monitor' controls the start and stop of monitoring in
the {vg,lv}change man pages.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Prevent lvresize from being able to resize internal LVs: mirror legs
(*_mimage_*), mirror log (*_mlog), snapshot placeholder LVs (snapshot*)
and others. Resizing these would leads to unexpected metadata and
sometimes crashes (in case of growing snapshot*).
This option should be configurable, but for now
do not set it at all.
(lvm2app is used in udisks probers and there
cac cause several nasty races when trying to update
lvmcache during rescan.)
If user try to vgcreate or vgextend non-existent VG,
these messages appears:
# vgcreate xxx /dev/xxx
Internal error: Volume Group xxx was not unlocked
Device /dev/xxx not found (or ignored by filtering).
Unable to add physical volume '/dev/xxx' to volume group 'xxx'.
Internal error: Attempt to unlock unlocked VG xxx.
(the same with existing VG and non-existing PV & vgextend)
# vgextend vg_test /dev/xxx
...
It is caused because code tries to "refresh" cache if
md filter is switched on using cache destroy.
But we can change filters and rescan even without this
machinery now, just use refresh_filters
(and reset md filter afterwards).
(Patch also discovers cache alias bug in vgsplit test,
fix it by using better filter line.)
This patch adds a new implementation of locking function instead
of mlockall() that may lock way too much memory (>100MB).
New function instead uses mlock() system call and selectively locks
memory areas from /proc/self/maps trying to avoid locking areas
unused during lock-ed state.
Patch also adds struct cmd_context to all memlock() calls to have
access to configuration.
For backward compatibility functionality of mlockall()
is preserved with "activation/use_mlockall" flag.
As a simple check, locking and unlocking counts the amount of memory
and compares whether values are matching.
For static builds dependency for SELinux libs is not handled by 'ar'.
Till better solution is found, for static builds STATIC_LIBS is used.
Patch updates SELinux detection to use 3rd & 4th parameter for Success/Fail.
Also removes detection of pthread from this check as we know which
version of libdevmapper we are going to link with lvm after merge.
SELinux header check moved to the SELinux test code.
Create new substituted variable PTHREAD_LIBS and link this library
only with tools/libs which really needs it - i.e. dmeventd.
Check for libpthread only for builds with clvmd or dmeventd.
Remove variable LIB_PTHREAD
The kernel's blk_stack_limits() function may flag a device as
'misaligned'. If it does the alignment_offset will be -1.
Update set_pe_align_offset() to accommodate this corner case.
- increase timeout to 30 secs (on Chrissie request)
- source both cluster and clvmd for options (like all the other cluster
init scripts)
- add clustered_vgs and _lvs commodity fns
- move rh_status* fns at the top, so they can be reused
- heavily cleanup start and stop fns from redundant code and unnecessary
loops
- improve output from different operations
- make the init script lsb compliant
- don´t force kill of the daemon, send only a TERM signal and then wait
for it to exit
- Resolves rhbz#533247
lvm2 devices have always UUID set even if imported from lvm1 metadata.
Patch removes name argument from dev_manager_info call and converts
all activation related calls to use query by UUID.
Also it simplifies mknode call (which is the only user on mknodes parameter).
Add a merging snapshot to the deptree, using the "error" target, rather
than avoid adding it entirely. This allows proper cleanup of the -cow
device without having to rename the -cow to use the origin's name as a
prefix.
Move the preloading of the origin LV, after a merge, from
lv_remove_single() to vg_remove_snapshot(). Having vg_remove_snapshot()
preload the origin allows the -cow device to be released so that it can
be removed via deactivate_lv(). lv_remove_single()'s deactivate_lv()
reliably removes the -cow device because the associated snapshot LV,
that is to be removed when a snapshot-merge completes, is always added
to the deptree (and kernel -- via "error" target).
Now when the snapshot LV is removed both the -cow and -real devices
get removed using uuid rather than device name. This paves the way
for us to switch over to info-by-uuid queries.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
There's a tiny period of time when the _mimage device is visible during
downconversion from mirror to linear. Since it is visible, we need to
create the symlinks, otherwise warning messages will be issued about udev
not creating those symlinks. We have to rely on udev flags completely.
When activating a merging origin it is valid, and expected, to not have
a node in the deptree for both the origin and its merging snapshot. The
_cached_info() caller is only concerned with whether a device is open.
If there isn't a node in the tree the associated device is definitely
not open.
depending on if the mirror has a 'core' or 'disk' log. When there
is a disk log, the new leg is added by stacking a new mirror on
top of the old (one leg is the old mirror and the other leg is the newly
added device). When the log is a 'core' log, the new leg is simply added
to the existing mirror and all the devices are re-synced.
The logic that handles collapsing the stacked 'disk' log mirror was
having the effect of causing 'core' logged mirrors to begin resync'ing
for a second time. I have used the 'CONVERTING' flag to indicate that
a mirror is converting by way of stacking. This is no longer set for
up-converting core logs. The final 'collapse' logic can safely be skipped
for 'core' log mirrors - getting rid of the second resync.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
where we should not expose internal VG names/uuids (the ones with "#" prefix )through the
interface. Otherwise, we could end up with library users opening internal VGs which will
initiate locking mechanism that won't be cleaned up properly.
"#orphans_{lvm1, lvm2, pool}" names are treated in a special way, they are truncated first
to "orphans" and this is used as a part of the lock name then (e.g. while calling lvm_vg_open()).
When library user calls lvm_vg_close(), the original name "orphans_{lvm1, lvm2, pool}"
is used directly and therefore no unlock occurs.
We should exclude internal VG names and uuids in the lists provided by lvmcache:
lvmcache_get_vgids() and lvmcache_get_vgnames().
Allow the number of logical extents to be expressed (for a snapshot) as
a percentage of the total space in the Origin Logical Volume with the
suffix %ORIGIN.
Update the relevant man pages accordingly. Eliminate inconsistencies
between the man pages and tools/commands.h
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
*_safe. This had the effect of segfaulting the log daemon when
converting a mirror from one log type to another.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
When activation of pvmove mirror fails on cluster, some nodes
still possibly succeeded in activation.
- Explicitly deactivate that mirror to be sure
- properly pair suspend/resume calls to not cause memory lock problems in clvmd
Code cannot simply call _finish_pvmove on cluster in this situation, because
changed LVs are suspended twice (causing memory inbalance) and also temporary
mirror is activated when it is not expected (and we know that it failed already).
Patch prepares special function which remove temporary mirror references from
metadata and then resumes changed LVs.
In dev_manager_info 0 means error and 1 info is returned,
not that device exists (that value is part of info struct).
Fix query by uuid only (no name) which returns 0 when device
does not exist.
Support "wait before testing" using '+' in pvmove and lvconvert
interval. Doing so overrides the new default of sleeping after checking
the LV's progress.
Sleeping before checking progress can lead to extraneous polldaemons
being left running. These polldaemons would have otherwise exited had
they checked before sleeping. Checking progress before sleeping helps
workaround the subtly unreliable nature of "finished" state checking
in _percent_run.
Update test/t-mirror-names.sh to use '+' when providing its lvconvert
interval.
more descriptive message if locking fails instead of
"Locking type -1 initialisation failed."
Use read-only locking instead of misleading ignorelocking option
in message.
For mirror repair (and similar tasks) it can happen that full
device rescan is issued from clvmd.
Because code can be in the middle of repair (calling suspend)
clvmd should never try to scan suspended devices
(otherwise it causes deadlock).
Also code must not change ignore_suspended_device flag when
doing refresh_filters (called from lvmcache scan code).
bitmap tracking was switched from the e2fsprogs implementation to
the device-mapper implementation (dm_bitset_t). The latter has a
leading uin32_t field designed to hold the number of bits that are
being tracked. The code was not properly handling this change in
all places. Specifically, when getting the bitmap to/from disk.
Endian adjustments will likely need to be made on the accounting
field as well, since bitmaps are passed between machines on
start-up.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
This spurious 'break' has been here since this code was first committed
in June 2005 and stopped the algorithm behaving as described in the
comment above it and rendered the variable 'already_found_one' useless.
Made .update_metadata optional in 'struct poll_functions' definitions;
eliminated _update_lvconvert_mirror() stub.
Tweak a mirror-specific error message in the generic polldaemon code.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
The logic was that lvconvert repair volumes, marking
PV as MISSING and following vgreduce --removemissing
removes these missing devices.
Previously dmeventd mirror DSO removed all LV and PV
from VG by simply relying on
vgreduce --removemissing --force.
Now, there are two subsequent calls:
lvconvert --repair --use-policies
vgreduce --removemissing
So the VG is locked twice, opening space for all races
between other running lvm processes. If the PV reappears
with old metadata on it (so the winner performs autorepair,
if locking VG for update) the situation is even worse.
Patch simply adds removemissing PV functionality into
lvconcert BUT ONLY if running with --repair and --use-policies
and removing only these empty missing PVs which are
involved in repair.
(This combination is expected to run only from dmeventd.)
Version >= 1.8.0 of the DM snapshot target appends metadata sectors used
to a snapshot's status. This patch allows LVM2 to accurately determine
if the snapshot store is empty. Knowing when a snapshot store is empty
is important in the context of snapshot-merge (means merge is complete).
Also update LVM2 to be aware of the possibility for "Merge failed" in
the snapshot-merge target's status.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
the background polldaemon is allowed to start. It can be used
standalone or in conjunction with --refresh or --available y.
Control over when the background polldaemon starts will be particularly
important for snapshot-merge of a root filesystem.
Dracut will be updated to activate all LVs with: --poll n
The lvm2-monitor initscript will start polling with: --poll y
NOTE: Because we currently have no way of knowing if a background
polldaemon is active for a given LV the following limitations exist and
have been deemed acceptable:
1) it is not possible to stop an active polldaemon; so the lvm2-monitor
initscript doesn't stop running polldaemon(s)
2) redundant polldaemon instances will be started for all specified LVs
if vgchange or lvchange are repeatedly used with '--poll y'
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
This patch tries to correctly track changes in lvmcache related to commit/revert.
For vg_commit: if there is cached precommitted metadata, after successfull commit
these metadata must be tracked as committed.
For vg_revert: remote nodes must drop precommitted metadata and its flag in lvmcache.
(N.B. Patch do not touch LV locks here in any way.)
All this machinery is needed to properly solve remote node cache invalidaton which
cause several problems recently observed.
Lock mode is int masked by LCK_TYPE_MASK, always.
Patch also remove uneccessary masking lock flag on sender side,
if masking is needed, it is don on client side already.
- Add drop_precommitted flag to force drop precommitted metadata
- add lvmcache_commit_metadata() which upgrades precommitted metadata in cache
No functional change in this patch - just preparation for following change.
And decode flags in humar readable form in client.
And clean some trailing whitespaces.
No functional change in this patch (only debugging messages changed).
The use_precommitted flag indicates, that we want to use precommitted metadata
(used in suspend call to preload table with precommitted data).
But if there are no such data, committed metadata are read but the cache
still contains that precommitted flag.
(The problem is that later possible drop_metadata call will not invalidate
device in cache.)
The wrong precommitted state is stored in on remote nodes during normal
suspend/resume cycle _without_ vg_write/commit.
Use the PRECOMMITTED status flag here instead (which is always set if using
precommited metadata here).
If renaming snapshot with virtual origin, the origin is renamed too.
But the code must resume LVs in reverse order to properly
pair memlock (in cluster locking).
(The resume of snapshot resumes origin too and later resume
is ignored otherwise.)
When PV device reappears with old metadata, it is
always updated to new version byt atutomatic metadata
repair.
Remove missing flag if device is empty.
If device contains allocated extents, issue warning that
user must remove volumes and re-add this PV before
manipulating with this volume.
This partially solves bug 547842 when one PV (log) is failed,
dmeventd removes that device and later this device reappears and
is wrongly added into VG marked missing.
The memlock_inc() fix is wrong, memlock count is not
propagated to long living process (clvmd) and just
it underflow there.
Also suspend is needed to pre-load precommited metadata
on other nodes (remapping to error taget in this case).
With explicit suspend we generate lock request and code
can update memlock count.
(Infinitely "locked" memory caused that fs_unlock() was not
called properly and on cluster nodes remains
old links in /dev/mapper for not active devices.)
(N.B. failing of suspend call here is not handled as fatal
error - the LV is going to be removed later anyway.)
The new recovery code first tries to repair LV and then removes failed PV
from VG. It means that during operation there can be VG with PV missing,
and vg_read code handles it like not consistent VG.
We already allows returning "inconsistent" commited metadata,
for mirror repair we need this for precommited too.
(The suspend call prepares precommited metadata to inactive table on
other cluster nodes.)
"Inconsistent" here means - correct metadata, just with some metadata areas
not found (obviously on missing or failed PVs).
The LV locks make sense only for clustered LVs.
Properly check cluster flag and never issue cluster lock here.
There are several places in code, where it is already checked, this
patch add this check to all needed calls.
In previous code the lock behaviour was inconsistent,
for example, the pre/post callback can take lock even for local volume,
but deactivate call do not released this lock and it remains held forever.
The local LV lock request now just let run the underlying activation code
on local node, the same process like in local locking.
(Again, this is important for new mirror repair calls, here for local
mirrors but with cluster locking enabled.)
This is unnoticed regression from commit 31672ff60e
The pre/post callback need to convert lock always, local node
is going to modify metadata in this case, it it fails conversion,
the call is ignored.
Also it fixes bug when the lock is not yet held, we cannot set LKF_CONVERT
in this case, it will fail because this lock do not exist.
Note that the automatic conversion is still disabled in activate
call, so the original fix (reactivation of exlusive LV) should
be still in place.
(Code already not fail if unlocking not locked resource.)
This is needed in pre/post lock_lv call, where we can
request the same lock on local node becuase of suspend call.
- do_command and lock_vg expect flags (no change here)
Bug fixes:
- lock_vg should check for NONBLOCK on lock_cmd, flags have this bit masked-out
- do_pre/post_command expect do not mask flag at all, this causes that
the code inside is never run! (see following patches, these functions
expect plain command without flags)
If there is problem deactivate LV and
_init_mirror_log is called with remove_on_failure = 1,
remove the newly created log LV from metadata.
(This can happen if there is active device with the same name
but different UUID.)
The main reason for this "workaround" patch is to
- do not keep _mlog volume in metadata, so user can repeat the action
- print better error message describing the real problem
# lvcreate -m 2 -n lv1 -l 1 --nosync vg_bar
WARNING: New mirror won't be synchronised. Don't read what you didn't write!
/dev/vg_bar/lv1_mlog: not found: device not cleared
Aborting. Failed to wipe mirror log.
Error locking on node bar-01: Input/output error
Unable to deactivate mirror log LV. Manual intervention required.
Failed to create mirror log.
# lvcreate -m 2 -n lv1 -l 1 --nosync vg_bar
WARNING: New mirror won't be synchronised. Don't read what you didn't write!
Aborting. Unable to deactivate mirror log.
Failed to initialise mirror log.
pvmove suspends all moved LVs + pvmoveX mirrored LV itself.
This suspends even underlying pvmoveX and following explicit
suspend call is just noop.
But in resume the pvmoveX volume is no longer underlying
device for moved LVs, so it performs full resume with memlock
decrease.
Code must call memlock_inc() if suspend is requested, volume
is already suspended and error is not requested.
These are no longer used by anyone. The dm_list defines are all in
libdevmapper.h and libdm/datastruct/list.c contains any function definitions.
There is some code in "old-tests" that still use this but this code is not
being maintained.
Thanks to Zdenek for spotting this.
The physical_volume, volume_group, logical_volume and lv_segment
structures' 'status' member is now uint64_t.
The alignment of these structures was also audited to remove holes. The
movement of some members in 'volume_group' and 'lv_segment' eliminates
holes. The 'physical_volume' structure still has one 4-byte hole after
'pe_size'; the other structures no longer have any holes. Each
structures' size has not changed.
If the vg_read() returned error, no lock was taken,
so always call vg_release().
Otherwise this can happen because of missing FAILED_*:
# vgchange -a y x --ignorelockingfailure
Volume group "x" not found
Internal error: Attempt to unlock unlocked VG x
The sysfs filter initialise hash of available devices using
scan of /sys/block. We need to refresh even this hash
when performing full scan otherwise the newly appeared
device could be rejected, because there is no entry
in sysfs filter.
This easily could happen when attaching new device
to cluster node. (Only force refresh of context
in clvmd -R works here now).
Unfortunately consequences of this are much worse,
missing device part on that node is replaced with missing segment
(even when no partial arg is selected) and this directly
lead to data corruption.
See https://bugzilla.redhat.com/show_bug.cgi?id=538515
Simply fix it by refreshing device filters in lvmcache
before performing the full device scan.
(on one node a storage connection failed):
# vgchange -a y vg_bar ; echo $?
Error locking on node bar-02: Refusing activation of partial LV lv1. Use --partial to override.
1 logical volume(s) in volume group "vg_bar" now active
0
So activation fails on one node, error is correctly printed but
status code is wrong.
This patch fixes the top level (vgchange) to return proper code
(and print # of activated LVs).
(lvchange returns error properly here.)
(This affects only cluster locking because only cluster
locking module set LCK_PRE_MEMLOCK.)
With currect code you get
# vgchange -a n
Internal error: _memlock_count has dropped below 0.
when using cluster locking.
It is caused by _unlock_memory calls here
if ((flags & (LCK_SCOPE_MASK | LCK_TYPE_MASK)) == LCK_LV_RESUME)
memlock_dec();
Unfortunately it is also (wrongly) called in immediate unlock
(when LCK_HOLD is not set) from lock_vol
(LCK_UNLOCK is misinterpreted as LCK_LV_RESUME).
Avoid this by comparing original flags and provide memlock
code type of operation (suspend/resume).
All hidden (not visible) volumes should be activated through
other visible volumes.
(There are already exceptions like snapshot, mirror log and image,
which should be cleaned one day...)
This solves problems for future types of hidden volumes,
which can have special meaning and must not be activated implicitly
(e.g. key store volume).
- fix missing unlocking of VG
lvcreate -l 100%PVS -n lv1 vg_test
Please specify physical volume(s) with %PVS
Internal error: Volume Group vg_test was not unlocked
- if no PVS specified, use all available
Fix segfault if %PVS in lvresize without PVs list.
The DRBD uses underlying device so code should prefer top
device if duplicate is found.
Patch also introduce
dev_subsystem_part_major and dev_subsytem_name
functions to easily handle all these replication susbystems
and not hardcode md_major call.
See https://bugzilla.redhat.com/show_bug.cgi?id=530881
for full problem description.
