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 fact that vg repair is implemented as a part of vg read
has led to a messy and complicated implementation of vg_read,
and limited and uncontrolled repair capability. This splits
read and repair apart.
Summary
-------
- take all kinds of various repairs out of vg_read
- vg_read no longer writes anything
- vg_read now simply reads and returns vg metadata
- vg_read ignores bad or old copies of metadata
- vg_read proceeds with a single good copy of metadata
- improve error checks and handling when reading
- keep track of bad (corrupt) copies of metadata in lvmcache
- keep track of old (seqno) copies of metadata in lvmcache
- keep track of outdated PVs in lvmcache
- vg_write will do basic repairs
- new command vgck --updatemetdata will do all repairs
Details
-------
- In scan, do not delete dev from lvmcache if reading/processing fails;
the dev is still present, and removing it makes it look like the dev
is not there. Records are now kept about the problems with each PV
so they be fixed/repaired in the appropriate places.
- In scan, record a bad mda on failure, and delete the mda from
mda in use list so it will not be used by vg_read or vg_write,
only by repair.
- In scan, succeed if any good mda on a device is found, instead of
failing if any is bad. The bad/old copies of metadata should not
interfere with normal usage while good copies can be used.
- In scan, add a record of old mdas in lvmcache for later, do not repair
them while reading, and do not let them prevent us from finding and
using a good copy of metadata from elsewhere. One result is that
"inconsistent metadata" is no longer a read error, but instead a
record in lvmcache that can be addressed separate from the read.
- Treat a dev with no good mdas like a dev with no mdas, which is an
existing case we already handle.
- Don't use a fake vg "handle" for returning an error from vg_read,
or the vg_read_error function for getting that error number;
just return null if the vg cannot be read or used, and an error_flags
arg with flags set for the specific kind of error (which can be used
later for determining the kind of repair.)
- Saving an original copy of the vg metadata, for purposes of reverting
a write, is now done explicitly in vg_read instead of being hidden in
the vg_make_handle function.
- When a vg is not accessible due to "access restrictions" but is
otherwise fine, return the vg through the new error_vg arg so that
process_each_pv can skip the PVs in the VG while processing.
(This is a temporary accomodation for the way process_each_pv
tracks which devs have been looked at, and can be dropped later
when process_each_pv implementation dev tracking is changed.)
- vg_read does not try to fix or recover a vg, but now just reads the
metadata, checks access restrictions and returns it.
(Checking access restrictions might be better done outside of vg_read,
but this is a later improvement.)
- _vg_read now simply makes one attempt to read metadata from
each mda, and uses the most recent copy to return to the caller
in the form of a 'vg' struct.
(bad mdas were excluded during the scan and are not retried)
(old mdas were not excluded during scan and are retried here)
- vg_read uses _vg_read to get the latest copy of metadata from mdas,
and then makes various checks against it to produce warnings,
and to check if VG access is allowed (access restrictions include:
writable, foreign, shared, clustered, missing pvs).
- Things that were previously silently/automatically written by vg_read
that are now done by vg_write, based on the records made in lvmcache
during the scan and read:
. clearing the missing flag
. updating old copies of metadata
. clearing outdated pvs
. updating pv header flags
- Bad/corrupt metadata are now repaired; they were not before.
Test changes
------------
- A read command no longer writes the VG to repair it, so add a write
command to do a repair.
(inconsistent-metadata, unlost-pv)
- When a missing PV is removed from a VG, and then the device is
enabled again, vgck --updatemetadata is needed to clear the
outdated PV before it can be used again, where it wasn't before.
(lvconvert-repair-policy, lvconvert-repair-raid, lvconvert-repair,
mirror-vgreduce-removemissing, pv-ext-flags, unlost-pv)
Reading bad/old metadata
------------------------
- "bad metadata": the mda_header or metadata text has invalid fields
or can't be parsed by lvm. This is a form of corruption that would
not be caused by known failure scenarios. A checksum error is
typically included among the errors reported.
- "old metadata": a valid copy of the metadata that has a smaller seqno
than other copies of the metadata. This can happen if the device
failed, or io failed, or lvm failed while commiting new metadata
to all the metadata areas. Old metadata on a PV that has been
removed from the VG is the "outdated" case below.
When a VG has some PVs with bad/old metadata, lvm can simply ignore
the bad/old copies, and use a good copy. This is why there are
multiple copies of the metadata -- so it's available even when some
of the copies cannot be used. The bad/old copies do not have to be
repaired before the VG can be used (the repair can happen later.)
A PV with no good copies of the metadata simply falls back to being
treated like a PV with no mdas; a common and harmless configuration.
When bad/old metadata exists, lvm warns the user about it, and
suggests repairing it using a new metadata repair command.
Bad metadata in particular is something that users will want to
investigate and repair themselves, since it should not happen and
may indicate some other problem that needs to be fixed.
PVs with bad/old metadata are not the same as missing devices.
Missing devices will block various kinds of VG modification or
activation, but bad/old metadata will not.
Previously, lvm would attempt to repair bad/old metadata whenever
it was read. This was unnecessary since lvm does not require every
copy of the metadata to be used. It would also hide potential
problems that should be investigated by the user. It was also
dangerous in cases where the VG was on shared storage. The user
is now allowed to investigate potential problems and decide how
and when to repair them.
Repairing bad/old metadata
--------------------------
When label scan sees bad metadata in an mda, that mda is removed
from the lvmcache info->mdas list. This means that vg_read will
skip it, and not attempt to read/process it again. If it was
the only in-use mda on a PV, that PV is treated like a PV with
no mdas. It also means that vg_write will also skip the bad mda,
and not attempt to write new metadata to it. The only way to
repair bad metadata is with the metadata repair command.
When label scan sees old metadata in an mda, that mda is kept
in the lvmcache info->mdas list. This means that vg_read will
read/process it again, and likely see the same mismatch with
the other copies of the metadata. Like the label_scan, the
vg_read will simply ignore the old copy of the metadata and
use the latest copy. If the command is modifying the vg
(e.g. lvcreate), then vg_write, which writes new metadata to
every mda on info->mdas, will write the new metadata to the
mda that had the old version. If successful, this will resolve
the old metadata problem (without needing to run a metadata
repair command.)
Outdated PVs
------------
An outdated PV is a PV that has an old copy of VG metadata
that shows it is a member of the VG, but the latest copy of
the VG metadata does not include this PV. This happens if
the PV is disconnected, vgreduce --removemissing is run to
remove the PV from the VG, then the PV is reconnected.
In this case, the outdated PV needs have its outdated metadata
removed and the PV used flag needs to be cleared. This repair
will be done by the subsequent repair command. It is also done
if vgremove is run on the VG.
MISSING PVs
-----------
When a device is missing, most commands will refuse to modify
the VG. This is the simple case. More complicated is when
a command is allowed to modify the VG while it is missing a
device.
When a VG is written while a device is missing for one of it's PVs,
the VG metadata is written to disk with the MISSING flag on the PV
with the missing device. When the VG is next used, it is treated
as if the PV with the MISSING flag still has a missing device, even
if that device has reappeared.
If all LVs that were using a PV with the MISSING flag are removed
or repaired so that the MISSING PV is no longer used, then the
next time the VG metadata is written, the MISSING flag will be
dropped.
Alternative methods of clearing the MISSING flag are:
vgreduce --removemissing will remove PVs with missing devices,
or PVs with the MISSING flag where the device has reappeared.
vgextend --restoremissing will clear the MISSING flag on PVs
where the device has reappeared, allowing the VG to be used
normally. This must be done with caution since the reappeared
device may have old data that is inconsistent with data on other PVs.
Bad mda repair
--------------
The new command:
vgck --updatemetadata VG
first uses vg_write to repair old metadata, and other basic
issues mentioned above (old metadata, outdated PVs, pv_header
flags, MISSING_PV flags). It will also go further and repair
bad metadata:
. text metadata that has a bad checksum
. text metadata that is not parsable
. corrupt mda_header checksum and version fields
(To keep a clean diff, #if 0 is added around functions that
are replaced by new code. These commented functions are
removed by the following commit.)
There have been two file locks used to protect lvm
"global state": "ORPHANS" and "GLOBAL".
Commands that used the ORPHAN flock in exclusive mode:
pvcreate, pvremove, vgcreate, vgextend, vgremove,
vgcfgrestore
Commands that used the ORPHAN flock in shared mode:
vgimportclone, pvs, pvscan, pvresize, pvmove,
pvdisplay, pvchange, fullreport
Commands that used the GLOBAL flock in exclusive mode:
pvchange, pvscan, vgimportclone, vgscan
Commands that used the GLOBAL flock in shared mode:
pvscan --cache, pvs
The ORPHAN lock covers the important cases of serializing
the use of orphan PVs. It also partially covers the
reporting of orphan PVs (although not correctly as
explained below.)
The GLOBAL lock doesn't seem to have a clear purpose
(it may have eroded over time.)
Neither lock correctly protects the VG namespace, or
orphan PV properties.
To simplify and correct these issues, the two separate
flocks are combined into the one GLOBAL flock, and this flock
is used from the locking sites that are in place for the
lvmlockd global lock.
The logic behind the lvmlockd (distributed) global lock is
that any command that changes "global state" needs to take
the global lock in ex mode. Global state in lvm is: the list
of VG names, the set of orphan PVs, and any properties of
orphan PVs. Reading this global state can use the global lock
in sh mode to ensure it doesn't change while being reported.
The locking of global state now looks like:
lockd_global()
previously named lockd_gl(), acquires the distributed
global lock through lvmlockd. This is unchanged.
It serializes distributed lvm commands that are changing
global state. This is a no-op when lvmlockd is not in use.
lockf_global()
acquires an flock on a local file. It serializes local lvm
commands that are changing global state.
lock_global()
first calls lockf_global() to acquire the local flock for
global state, and if this succeeds, it calls lockd_global()
to acquire the distributed lock for global state.
Replace instances of lockd_gl() with lock_global(), so that the
existing sites for lvmlockd global state locking are now also
used for local file locking of global state. Remove the previous
file locking calls lock_vol(GLOBAL) and lock_vol(ORPHAN).
The following commands which change global state are now
serialized with the exclusive global flock:
pvchange (of orphan), pvresize (of orphan), pvcreate, pvremove,
vgcreate, vgextend, vgremove, vgreduce, vgrename,
vgcfgrestore, vgimportclone, vgmerge, vgsplit
Commands that use a shared flock to read global state (and will
be serialized against the prior list) are those that use
process_each functions that are based on processing a list of
all VG names, or all PVs. The list of all VGs or all PVs is
global state and the shared lock prevents those lists from
changing while the command is processing them.
The ORPHAN lock previously attempted to produce an accurate
listing of orphan PVs, but it was only acquired at the end of
the command during the fake vg_read of the fake orphan vg.
This is not when orphan PVs were determined; they were
determined by elimination beforehand by processing all real
VGs, and subtracting the PVs in the real VGs from the list
of all PVs that had been identified during the initial scan.
This is fixed by holding the single global lock in shared mode
while processing all VGs to determine the list of orphan PVs.
Native disk scanning is now both reduced and
async/parallel, which makes it comparable in
performance (and often faster) when compared
to lvm using lvmetad.
Autoactivation now uses local temp files to record
online PVs, and no longer requires lvmetad.
There should be no apparent command-level change
in behavior.
Different flavors of activate_lv() and lv_is_active()
which are meaningful in a clustered VG can be eliminated
and replaced with whatever that flavor already falls back
to in a local VG.
e.g. lv_is_active_exclusive_locally() is distinct from
lv_is_active() in a clustered VG, but in a local VG they
are equivalent. So, all instances of the variant are
replaced with the basic local equivalent.
For local VGs, the same behavior remains as before.
For shared VGs, lvmlockd was written with the explicit
requirement of local behavior from these functions
(lvmlockd requires locking_type 1), so the behavior
in shared VGs also remains the same.
As we start refactoring the code to break dependencies (see doc/refactoring.txt),
I want us to use full paths in the includes (eg, #include "base/data-struct/list.h").
This makes it more obvious when we're breaking abstraction boundaries, eg, including a file in
metadata/ from base/
ATM it's a bit ugly to enforce flushing of 'stdio' here, but works as quick
hot-fix.
log_print*() is using buffered I/O.
But for pooling with typical 1s interval this may take a while before
buffer about continues progress gets flushed.
So ATM fflush().
TODO: either add log_print*_with_flush() or maybe directly use just
line buffering with log_print() and only log_debug() keep using buffered
I/O mode.
When pvmove was run in background mode and forks
instead of using lvmpolld, the child pvmove process
was not clearing the bcache from the parent, so all
the aio ops in the child were failing.
The lvm fullreport works per VG and as such, the vg, lv, pv, seg and
pvseg subreport is done for each VG. However, if the PV is not part of
any VG yet, we still want to display pv and pvseg subreports for these
"orphan" PVs - so enable this for lvm fullreport's process_each_vg call.
If there's parent processing handle, we don't need to create completely
new report group and status report - we'll just reuse the one already
initialized for the parent.
Currently, the situation where this matter is when doing internal report
to do the selection for processing commands where we have parent processing
handle for the command itself and processing handle for the selection
part (that is selection for non-reporting tools).
Pass the single vgname as a new process_each_vg arg
instead of setting a cmd flag to tell process_each_vg
to take only the first vgname arg from argv.
Other commands with different argv formats will be
able to use it this way.
CONVERTING status flag is a tricky one. It's not set when converting
a non-mirror LV type to the mirror type, i.e.: linear -> two leg mirror.
Also the conversion itself is instant and doesn't require to be polled.
When mirror reaches sync state there's no final update on VG metadata
for lvmpolld to be made thereby report_progress in fact doesn't report
percentage of mirror being converted but percentage of mirror
being in sync. Perhaps we should reword the lvconvert output here.
On the other hand CONVERTING is set while we upconvert the mirror
from i.e. two leg mirror to four leg mirror. In such case the operation
is required to be polled so that lvmpolld can cleanup temporary
conversion log when the conversion is over.
Ignore CONVERTING lv_type for the moment and match LVs only by uuids
during 'mirror conversion'/'waiting for a sync to finish'.
The unlock call will fail in expected and normal cases,
and should not cause the command to fail. (An actual
unlock in the lock manager should never fail.)
tools/polldaemon.c:465: uninit_use_in_call: Using uninitialized value "id.vg_name" when calling "print_log".
tools/polldaemon.c:465: uninit_use_in_call: Using uninitialized value "id.lv_name" when calling "print_log".
. the poll check will eventually call finish which will
write the VG, so an ex VG lock is needed from lvmlockd.
. fix missing unlock on poll error path
. remove the lockd locking while monitoring the progress
of the command, as suggested by the earlier FIXME comment,
as it's not needed.
tools/polldaemon.c:457: array_null: Comparing an array to null is not useful: "lv->lvid.s"
The lv->lvid.s is never NULL. The check was supposed to be *lv->lvid.s
to check if the string is not empty.
... Using uninitialized value "lockd_state" when calling "lockd_vg"
(even though lockd_vg assigns 0 to the lockd_state, but it looks at
previous state of lockd_state just before that so we need to have
that properly initialized!)
libdm/libdm-report.c:2934: uninit_use_in_call: Using uninitialized value "tm". Field "tm.tm_gmtoff" is uninitialized when calling "_get_final_time".
daemons/lvmlockd/lvmlockctl.c:273: uninit_use_in_call: Using uninitialized element of array "r_name" when calling "format_info_r_action". (just added FIXME as this looks unfinished?)
There's a race when asking lvmpolld about progress_status and
actually reading the progress info from kernel:
Even with lvmpolld being used we read status info from
LVM2 command issued by a user (client side from lvmpolld perspective).
The whole cycle may look like following:
1) set up an operation that requires polling (i.e. pvmove /dev/sda)
2) notify lvmpolld about such operation (lvmpolld_poll_init())
3) in case 1) was not called with --background it would continue with:
4) Ask lvmpolld about progress status. it may respond with one of:
a) in_progress
b) not_found
c) finished
d) any low level error
5) provided the answer was 4a) try to read progress info from polling LV
(i.e. vg00/pvmove1). Repeat steps 4) and 5) until the answer is != 4a).
And now we got into racy configuration: lvmpolld answered with in_progress
but it may be the that in_between 4) and 5) the operation has already
finished and polling LV is already gone or there's nothing to ask for.
Up to now, 5) would report warning and it could print such warning many
times if --interval was set to 0.
We don't want to scary users by warnings in such situation so let's just
print these messages in verbose mode. Error messages due to error while
reading kernel status info (on existing, active and locked LV) remained
the same.
currently in wait_for_single_lv() fn trying to poll missing pvmove LV
is considered success. It may have been already finished by another
instance of polldaemon. either by another forked off polldaemon
or by lvmpolld.
Let's try to handle the mirror conversion and snapshot merge the same
way.
These wrappers have been replaced by direct calls
to vg_read() and find_lv() in previous commits.
This commit should have no functional impact since
all bits were already unreachable.
let's call dev_close_all() only before we're about to 'sleep'
for at least one second during the polling.
(it's questionable whether to call dev_close_all() at all in
polldaemon code. Natural extension would be to drop it completely)
querying future lvmpolld with zero wait time is highly undesirable
and can cause serious performance drop of the future daemon. The new
wrapper function may avoid immediate return from syscal by
introducing minimal wait time on demand.
Routines responsible for polling of in-progress pvmove, snapshot merge
or mirror conversion each used custom lookup functions to find vg and
lv involved in polling.
Especially pvmove used pvname to lookup pvmove in-progress. The future
lvmpolld will poll each operation by vg/lv name (internally by lvid).
Also there're plans to make pvmove able to move non-overlaping ranges
of extents instead of single PVs as of now. This would also require
to identify the opertion in different manner.
The poll_operation_id structure together with daemon_parms structure they
identify unambiguously the polling task.
Waiting even after _check_lv_status returned success and
'finished' flag was set to true doesn't make much sense.
Note that while we skip the wait() we also skip the
init_full_scan_done(0) inside the routine. This should
have no impact as long as the code after _wait_for_single_lv
doesn't presume anything about the state of the cache.
as a part of bigger effort to unify polling intefaces
poll_get_copy_lv should be able to look up LVs based
on theirs lv->status field.
Effective after pvmove starts using poll_get_copy_lv
fn as well.
If the device name is not found in our metadata,
we cannot call strdup few lines later with NULL name.
More intersting story goes behind how it happens -
pvmove removal is unfortunatelly 'multi-state' process
and at some point (for now) we have in lvm2 metadata
LV pvmove0 as stripe and mirror image as error.
If such metadata are left - we fail with any further removal.
we do not allow 0 interval for pvmove command issued
without parameters with classical polldaemon. It would
query the kernel too often with possibly many pvmoves
in-progress.
