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.
lvm uses a bcache block size of 128K. A bcache block
at the end of the metadata area will overlap the PEs
from which LVs are allocated. How much depends on
alignments. When lvm reads and writes one of these
bcache blocks to update VG metadata, it can also be
reading and writing PEs that belong to an LV.
If these overlapping PEs are being written to by the
LV user (e.g. filesystem) at the same time that lvm
is modifying VG metadata in the overlapping bcache
block, then the user's updates to the PEs can be lost.
This patch is a quick hack to prevent lvm from writing
past the end of the metadata area.
Previously the size was limited by checking if the
old and new copies of the metadata overlapped.
This generally limited the size to about half of
the total space, but it could be larger given the
size differences between old and new. Now add a
direct check to limit the size to half the space.
Remove another instance of an invalid check for metadata
overflow during read. The previous instance was removed
in commit 5fb15b193.
This was checking for metadata that that overflowed the
circular disk metadata buffer during read, but such metadata
cannot be written, so it shouldn't be possible to find see.
Also, the check was incorrect and could trigger when there
was no overflow.
The vg_write/vg_commit code was imprecise, uncommented, and
hard to understand. Rewrite it with clearer, cleaner code,
extensive comments, descriptions of how it works, and add
more info in debugging output.
The minor changes in behavior are to things that were
either incorrect or probably unintended:
- vg_write/vg_commit no longer check that the current vgname at
the start of the text metadata matches the vgname being written.
This has already been done at least twice by the time they are
called, and repeating it again against the same cached data has
no use.
- A fragment of old removed code had been left behind that checked
if the old unused alignment policy would wrap. It was still
being checked to decide if the metadata area was full, which
could possibly cause an incorrect full metadata failure.
- vg_remove now clears both the raw_locns in the mda_header that
point to committed metadata (raw_locn slot 0) and precommitted
metadata (raw_locn slot 1). Previously it fully cleared the
committed slot, and would only clear the offset field in the
precommitted slot if it saw a problem with the metadata in the
vg being removed.
- read_metadata_location_summary was wrongly comparing the number
of wrapped bytes with an offset to report an error about the
metadata being too large. This wrong check is removed, it
could have resulted in erroneous errors.
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.
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/
There are likely more bits of code that can be removed,
e.g. lvm1/pool-specific bits of code that were identified
using FMT flags.
The vgconvert command can likely be reduced further.
The lvm1-specific config settings should probably have
some other fields set for proper deprecation.
For reporting commands (pvs,vgs,lvs,pvdisplay,vgdisplay,lvdisplay)
we do not need to repeat the label scan of devices in vg_read if
they all had matching metadata in the initial label scan. The
data read by label scan can just be reused for the vg_read.
This cuts the amount of device i/o in half, from two reads of
each device to one. We have to be careful to avoid repairing
the VG if we've skipped rescanning. (The VG repair code is very
poor, and will be redone soon.)
The improved detection of bad metadata when scanning
(where errors were ignored before) means we now have to
override some errors when forcibly erasing damaged metadata.
This is a temporary hacky workaround to the problem of
reads going through bcache and writes not using bcache.
The write path wants to read parts of data that it is
incrementally writing to disk, but the reads (using
bcache) don't work because the writes are not in the
bcache. For now, add a dev to bcache before each attempt
to read it in case it's being used on the write path.
Create a new dev->bcache_fd that the scanning code owns
and is in charge of opening/closing. This prevents other
parts of lvm code (which do various open/close) from
interfering with the bcache fd. A number of dev_open
and dev_close are removed from the reading path since
the read path now uses the bcache.
With that in place, open(O_EXCL) for pvcreate/pvremove
can then be fixed. That wouldn't work previously because
of other open fds.
When process_each_pv() calls vg_read() on the orphan VG, the
internal implementation was doing an unnecessary
lvmcache_label_scan() and two unnecessary label_read() calls
on each orphan. Some of those unnecessary label scans/reads
would sometimes be skipped due to caching, but the code was
always doing at least one unnecessary read on each orphan.
The common format_text case was also unecessarily calling into
the format-specific pv_read() function which actually did nothing.
By analyzing each case in which vg_read() was being called on
the orphan VG, we can say that all of the label scans/reads
in vg_read_orphans are unnecessary:
1. reporting commands: the information saved in lvmcache by
the original label scan can be reported. There is no advantage
to repeating the label scan on the orphans a second time before
reporting it.
2. pvcreate/vgcreate/vgextend: these all share a common
implementation in pvcreate_each_device(). That function
already rescans labels after acquiring the orphan VG lock,
which ensures that the command is using valid lvmcache
information.
This fixes the use of lvmcache_label_rescan_vg() in the previous
commit for the special case of independent metadata areas.
label scan is about discovering VG name to device associations
using information from disks, but devices in VGs with
independent metadata areas have no information on disk, so
the label scan does nothing for these VGs/devices.
With independent metadata areas, only the VG metadata found
in files is used. This metadata is found and read in
vg_read in the processing phase.
lvmcache_label_rescan_vg() drops lvmcache info for the VG devices
before repeating the label scan on them. In the case of
independent metadata areas, there is no metadata on devices, so the
label scan of the devices will find nothing, so will not recreate
the necessary vginfo/info data in lvmcache for the VG. Fix this
by setting a flag in the lvmcache vginfo struct indicating that
the VG uses independent metadata areas, and label rescanning should
be skipped.
In the case of independent metadata areas, it is the metadata
processing in the vg_read phase that sets up the lvmcache
vginfo/info information, and label scan has no role.
Move the location of scans to make it clearer and avoid
unnecessary repeated scanning. There should be one scan
at the start of a command which is then used through the
rest of command processing.
Previously, the initial label scan was called as a side effect
from various utility functions. This would lead to it being called
unnecessarily. It is an expensive operation, and should only be
called when necessary. Also, this is a primary step in the
function of the command, and as such it should be called prominently
at the top level of command processing, not as a hidden side effect
of a utility function. lvm knows exactly where and when the
label scan needs to be done. Because of this, move the label scan
calls from the internal functions to the top level of processing.
Other specific instances of lvmcache_label_scan() are still called
unnecessarily or unclearly by specific commands that do not use
the common process_each functions. These will be improved in
future commits.
During the processing phase, rescanning labels for devices in a VG
needs to be done after the VG lock is acquired in case things have
changed since the initial label scan. This was being done by way
of rescanning devices that had the INVALID flag set in lvmcache.
This usually approximated the right set of devices, but it was not
exact, and obfuscated the real requirement. Correct this by using
a new function that rescans the devices in the VG:
lvmcache_label_rescan_vg().
Apart from being inexact, the rescanning was extremely well hidden.
_vg_read() would call ->create_instance(), _text_create_text_instance(),
_create_vg_text_instance() which would call lvmcache_label_scan()
which would call _scan_invalid() which repeats the label scan on
devices flagged INVALID. lvmcache_label_rescan_vg() is now called
prominently by _vg_read() directly.
To do label scanning, lvm code calls lvmcache_label_scan().
Change lvmcache_label_scan() to use the new label_scan()
based on bcache.
Also add lvmcache_label_rescan_vg() which calls the new
label_scan_devs() which does label scanning on only the
specified devices. This is for a subsequent commit and
is not yet used.
New label_scan function populates bcache for each device
on the system.
The two read paths are updated to get data from bcache.
The bcache is not yet used for writing. bcache blocks
for a device are invalidated when the device is written.
Callers that read larger amounts of data now get a pointer to read-only
data directly without copying it through an intermediate buffer. This
data is owned by the device layer so the callers no longer free it.
Rename dev_read() to dev_read_buf() - the function that reads data
into a supplied buffer.
Introduce a new dev_read() that allocates the buffer it returns and
switch the important users over to this. No caller may change the
returned data. (For now, callers are responsible for freeing it after
use, but later the device layer will take full ownership.)
dev_read_buf() should only be used for tiny buffers or unimportant code
(such as the old disk formats).
The creation of wrapped around metadata - where the start of metadata is
written up to the end of the buffer and the remainder follows back at
the start of the buffer - is now restricted to cases where writing the
metadata in one piece wouldn't fit. This shouldn't happen in 'normal'
usage so let's begin treating the code for this as a special case that
can be ignored when optimising 'normal' cases.
If there is sufficient space in the metadata area, align the next
metadata to a disk offset that is a multiple of 4096 bytes and
don't write it circularly. If it doesn't all fit at the end
of the metadata area, go back to the start and write it all there
contiguously.
If there is insufficient space to use the new stricter rules, revert to
the original behaviour, aligning on 512-byte boundaries wrapping around
the circular buffer as required.
Use new ALIGN_ABSOLUTE macro when calculating the start location
of new metadata and adjust the end of buffer detection so that
there is no longer an imposed gap between old and new metadata.
Currently both start and offset should always be divisible by alignment,
so this should have no effect, but a later patch will increase alignment
so these variables can no longer be optimised out.
Expand out the metadata wrapping calculations to prepare
to support a larger alignment.
The current alignment is 512 bytes so
(mdac_area_start + rlocn->offset) % alignment is zero.
Mark the first metadata area on each text format PV as MDA_PRIMARY.
Pass this information down to the device layer so that when
there are two metadata areas on a block device, we can easily
distinguish two independent streams of I/O.
Introduce enum dev_io_reason to categorise block device I/O
in debug messages so it's obvious what it is for.
DEV_IO_SIGNATURES /* Scanning device signatures */
DEV_IO_LABEL /* LVM PV disk label */
DEV_IO_MDA_HEADER /* Text format metadata area header */
DEV_IO_MDA_CONTENT /* Text format metadata area content */
DEV_IO_FMT1 /* Original LVM1 metadata format */
DEV_IO_POOL /* Pool metadata format */
DEV_IO_LV /* Content written to an LV */
DEV_IO_LOG /* Logging messages */
When an ignored metadata area gets flagged for use again, make sure the
code doesn't try to parse its old metadata. Firstly by trying to detect
this situation and skipping the read (while still remembering the
position reached in the circular buffer), and secondly by clearing the
invalid live metadata location on disk as a precaution when subsequently
writing out the precommitted metadata.
Problems showed up when a metadata area in one VG got moved to
another VG in ignored state (still holding metadata for the original
VG) and then later got brought into use in the new VG - only the header
should be read in this case, not any of the metadata content.
vgmerge suffers from a similar problem to the one fixed in commit
8146548d25 ("vgsplit: Fix intermediate
metadata corruption.")
When merging, splitting or renaming VGs, use a new PV status flag
PV_MOVED_VG to mark the PVs that hold metadata with the old VG name and
use this to provide PV-level granularity instead of incorrectly assuming
all PVs in the VG are the same.
Changing the VG of a PV uses the same on-disk mechanism as vgrename.
This relies on recognising both the old and new VG names. Prior to this
patch the vgsplit code incorrectly provided the new VG name twice
instead of the old and new ones. This lead the low-level mechanism not
to recognise the device as already belonging to a VG and so paying no
attention to the location of its existing metadata, sometimes partly
overwriting it and then later trying to read the corrupt metadata and
issuing a checksum error.
Fix code checking that the 2nd mda which is at the end of disk really
fits the available free space and avoid any DA and MDA interleaving when
we already have DA preallocated. This mainly applies when we're restoring
a PV from VG backup using pvcreate --restorefile where we may already have
some DA preallocated - this means the PV was in a VG before with already
allocated space from it (the LVs were created). Hence we need to avoid
stepping into DA - the MDA can never ever be inside in such case!
The code responsible for this calculation was already in
_text_pv_add_metadata_area fn, but it had a bug in the calculation where
we subtracted one more sector by mistake and then the code could still
incorrectly allocate the MDA inside existing DA. The patch also renames
the variable in the code so it doesn't confuse us in future.
Also, if the 2nd mda doesn't fit, don't silently continue with just 1
MDA (at the start of the disk). If 2nd mda was requested and we can't
create that due to unavailable space, error out correctly (the patch
also adds a test to shell/pvcreate-operation.sh for this case).
A number of places are working on a specific dev when they
call lvmcache_info_from_pvid() to look up an info struct
based on a pvid. In those cases, pass the dev being used
to lvmcache_info_from_pvid(). When a dev is specified,
lvmcache_info_from_pvid() will verify that the cached
info it's using matches the dev being processed before
returning the info. Calling code will not mistakenly
get info for the wrong dev when duplicate devs exist.
This confusion was happening when scanning labels when
duplicate devs existed. label_read for the first dev
would add an info struct to lvmcache for that dev/pvid.
label_read for the second dev would see the pvid in
lvmcache from first dev, and mistakenly conclude that
the label_read from the second dev can be skipped
because it's already been done. By verifying that the
dev for the cached pvid matches the dev being read,
this mismatch is avoided and the label is actually read
from the second duplicate.
The lvmetad connection is created within the
init_connections() path during command startup,
rather than via the old lvmetad_active() check.
The old lvmetad_active() checks are replaced
with lvmetad_used() which is a simple check that
tests if the command is using/connected to lvmetad.
The old lvmetad_set_active(cmd, 0) calls, which
stopped the command from using lvmetad (to revert to
disk scanning), are replaced with lvmetad_make_unused(cmd).
Use process_each_vg() to lock and read the old VG,
and then call the main vgrename code.
When real VG names are used (not a UUID in place of the
old name), the command still pre-locks the new name
(when strcmp wants it locked first), before calling
process_each_vg on the old name.
In the case where the old name is replaced with a UUID,
process_each_vg now translates that UUID into the real
VG name, which it locks and reads. In this case, we
cannot do pre-locking to maintain lock ordering because
the old name is unknown. So, in this case the strcmp
based lock ordering is suppressed and the old name is
always locked first. This opens a remote chance for
lock ordering conflict between racing vgrenames between
two names where one or both commands use the UUID.
When checking minimum mda size, make sure the mda_size after alignment
and calculation is more than 0 - if there's no place for an MDA at the
end of the disk, the _text_pv_add_metadata_area does not try to add it
there and it returns (because we already have the MDA at the start of
the disk at least).
Actually, we don't need extra condition as introduced in commit
00348c0a63. We should fix the last
condition:
(mdac->rlocn.size >= mdah->size)
...which should be:
(MDA_HEADER_SIZE + (rlocn ? rlocn->size : 0) + mdac->rlocn.size >= mdah->size))
Where the "mdac" is new metadata, the "rlocn" is old metadata.
So the main problem with the previous condition was that it
didn't count in MDA_HEADER_SIZE properly (and possible existing
metadata - the "rlocn"). This could have caused the error state
where metadata in ring buffer overlap to not be hit.
Replace the new condition introduced in 00348c0a63
with the improved one for the condition that existed there
already but it was just incomplete.
We're already checking whether old and new meta do not overlap in
ring buffer (as we need to keep both old and new meta during vg_write
up until vg_commit).
We also need to check whether the new metadata do not overlap
themselves in case we don't have old metadata yet (...because
we're in vgcreate). This could happen if we're creating a VG so
that the very first metadata written are long enough that it wraps
themselves in metadata ring buffer.
Although we limited the minimum metadata area size better with the
previous commit ccb8da404d which
makes the initial VG metadata overlap in ring buffer to be less
probable, the risk of hitting this overlap condition is still there
if we still manage to generate big enough metadata somehow.
For example, users can provide many and/or long VG tags during vgcreate
so that the VG metadata is long enough to start to wrap in the ring
buffer again...
Also, leave out the note about "circular buffer" which is
an internal imeplementation detail anyway and not quite
informational for users:
Before this patch:
$ vgcreate vg1 /dev/sda
VG vg1 metadata too large for circular buffer
Failed to write VG vg1.
With this patch applied:
$ vgcreate vg1 /dev/sda
VG vg1 metadata too large: size of metadata to write is 691 bytes while PV metadata area size on /dev/sda is 512 bytes.
Failed to write VG vg1.
When using lvm shell, some structures which are cached in memory may be
reused. This happens for the struct label (a part of lvmcache_info
structure) when lvmetad is used in which case the PV scan is not
done that would normally overwrite these label structures in memory
and making them up-to-date.
This is all consequence of the fact that struct lvmcache_info and
struct label are not always assigned in the same part of the code.
For example, if lvmetad *is not* used, parts of the struct label are
reassigned in label_read fn while struct lvmcache_info is created
elsewhere. No part of the code reused struct label (and its "dev"
field) before calling label_read fn. That's why the real bug is
hidden when using lvm shell without lvmetad.
However, with lvmetad and lvm shell, the situation is a bit different.
The label_read fn is not called if lvmetad *is* used, hence the
struct label may have ended up not initialized properly.
There was missing assignment for the dev field in struct label
in _text_pv_write fn which caused this problem to appear in
lvm shell with lvmetad, for example:
Before this patch:
lvm> pvcreate /dev/sda
Physical volume "/dev/sda" successfully created
lvm> pvs /dev/sda
PV VG Fmt Attr PSize PFree
unknown device lvm2 --- 128.00m 128.00m
With this patch applied:
lvm> pvcreate /dev/sda
Physical volume "/dev/sda" successfully created
lvm> pvs /dev/sda
PV VG Fmt Attr PSize PFree
/dev/sda lvm2 --- 128.00m 128.00m
Also, this problem had not appeared before changes introduced
by commits e1a63905d1 through
3a6f91d713 which, among other
things, added proper label field type reporting. Before, label
reporting was the same as using struct physical_volume which
has its own dev field assigned and so this problem was not exposed.
Use 64bit arithmentic for PV size calculation (Coverity).
Also remove sector shift for compared PV size, since all
values are already held in sectors.
This fixes validatio of PV size when restoring PV
from vg metadata backup file.
When performing initial allocation (so there is nothing yet to
cling to), use the list of tags in allocation/cling_tag_list to
partition the PVs. We implement this by maintaining a list of
tags that have been "used up" as we proceed and ignoring further
devices that have a tag on the list.
https://bugzilla.redhat.com/983600
pv_write is called both to write orphans and to rewrite PV headers
of PVs in VGs. It needs to select the correct VG id so that the
internal cache state gets updated correctly.
It only affected commands that involved further steps after
the pv_write and was often masked because the metadata would
be re-read off disk and correct itself.
"Incorrect metadata area header checksum" warnings appeared.
Example:
Create vg1 containing dev1, dev2 and dev3.
Hide dev1 and dev2 from the system.
Fix up vg1 with vgreduce --removemissing.
Bring back dev1 and dev2.
In a single operation reinstate dev1 and dev2 into vg1 (vgextend).
Done as separate operations (automatically fix-up dev1 and dev2 as orphans,
then vgextend) it worked, but done all in one go the internal cache got
corrupted and warnings about checksum errors appeared.
The code never mixes reads of committed and precommitted metadata,
so there's no need to attempt to set PRECOMMITTED when
*use_previous_vg is being set.
Refactor the recent metadata-reading optimisation patches.
Remove the recently-added cache fields from struct labeller
and struct format_instance.
Instead, introduce struct lvmcache_vgsummary to wrap the VG information
that lvmcache holds and add the metadata size and checksum to it.
Allow this VG summary information to be looked up by metadata size +
checksum. Adjust the debug log messages to make it clear when this
shortcut has been successful.
(This changes the optimisation slightly, and might be extendable
further.)
Add struct cached_vg_fmtdata to format-specific vg_read calls to
preserve state alongside the VG across separate calls and indicate
if the details supplied match, avoiding the need to read and
process the VG metadata again.
Use similar logic as with text_vg_import_fd() and avoid repeated
parsing of same mda and its config tree for vgname_from_mda().
Remember last parsed vgname, vgid and creation_host in labeller
structure and if the metadata have the same size and checksum,
return this stored info.
TODO: The reuse of labeller struct is not ideal, some lvmcache API for
this functionality would be nicer.
