Fix splitmirror in cluster having different DM/LVM views of storage.
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.
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.
This patch fixes issues with improper udev flags on sub-LVs.
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.
#vgsplit -v "TestA" "TestB" "/dev/sdq" "/dev/sdo" "/dev/sdi" "/dev/sds"
"/dev/sde"
Checking for volume group "TestA"
Checking for new volume group "TestB"
Archiving volume group "TestA" metadata (seqno 213).
Can't split mirror MyMirror between two Volume Groups
AFTER FIX:
[root@bp-01 ~]# lvs -a -o name,vg_name,devices vg new
Volume group "new" not found
Skipping volume group new
LV VG Devices
lv vg lv_mimage_0(0),lv_mimage_1(0)
[lv_mimage_0] vg /dev/sdb1(0)
[lv_mimage_1] vg /dev/sdc1(0)
[lv_mlog] vg lv_mlog_mimage_0(0),lv_mlog_mimage_1(0)
[lv_mlog_mimage_0] vg /dev/sdh1(0)
[lv_mlog_mimage_1] vg /dev/sdi1(0)
[root@bp-01 ~]# vgsplit vg new /dev/sd[bchi]1
New volume group "new" successfully split from "vg"
[root@bp-01 ~]# lvs -a -o name,vg_name,devices vg new
LV VG Devices
lv new lv_mimage_0(0),lv_mimage_1(0)
[lv_mimage_0] new /dev/sdb1(0)
[lv_mimage_1] new /dev/sdc1(0)
[lv_mlog] new lv_mlog_mimage_0(0),lv_mlog_mimage_1(0)
[lv_mlog_mimage_0] new /dev/sdh1(0)
[lv_mlog_mimage_1] new /dev/sdi1(0)
Abort if _finish_pvmove suspend_lvs fails instead of cleaning up incompletely.
Change suspend_lvs to call vg_revert internally.
Change vg_revert to void and remove superfluous calls after failed vg_commit.
Replace test for NULL of root->child with test for NULL l
It's 100% equivalent test - since it always happen for the first iteration.
But the check for 'l' is understandable with analyzers - since analyzer
is not smart enough to deduce connection between root->child == NULL.
Avoid sending garbage to terminal in verbose mode.
When read in drain returned <0 value, terminal content has been trashed.
Remove unneeded memset() and use whole buffer.
Free readbuf before exit (valgrind).
Test whether nodes could be used on given filesystem where TMP
dir is being used and skip teardown quicker in fail case.
(makes the problem quickly obvious if you try to such fs).
Skip teardown_dev if we have not created any devs yet.
and do not mkdir /dev/mapper dir when LVM_TEST_DEVDIR is set.
Bugfix:
Add (most probably unfinished) support for -E arg with list of exclusive
locks. (During clvmd restart all exclusive locks would have been lost and
in fact, if there would have been an exclusive lock, usage text would be
printed and clvmd exits.)
Instead of parsing list options multiple times every time some lock UUID is
checked - put them straight into the hash table - make the code easier to
understand as well.
Remove was_ex_lock() function (replaced with dm_hash_lookup()).
Swap return value for get_initial_state() (1 means success).
Update the RAID design doc to reflect some of the new options introduce (e.g.
--merge and --trackchanges) and document the coding steps of up/down-conversion,
splitting RAID1 images, and merging RAID1 images.
Peter Rajnoha [Thu, 22 Sep 2011 17:33:50 +0000 (17:33 +0000)]
Replace open_count check with holders/mounted_fs check on lvremove path.
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.
Peter Rajnoha [Thu, 22 Sep 2011 17:23:35 +0000 (17:23 +0000)]
Add dm_device_has_holders fn to to check use of the device by another device.
Add dm_device_has_mounted_fs fn to check mounted filesystem on a device.
This requires sysfs directory to be correctly set via dm_set_sysfs_dir
(/sys by default). If sysfs dir is not used or it's set incorrectly,
dm_device_has_{holders,mounted_fs} will return 0!
Disallow the creation of mirrors (mirror or raid1 segtype) with only one leg.
If you specify the segment type (e.g. --type mirror) and the mirrors argument
as zero, it would result in a mirrored LV with only one image. While the device
may be valid in theory, it should not be allowed in practice. It also makes it
difficult on the conversion tools, since they react badly to single-image
mirrors.
When up-converting a RAID1 array, we need to allocate new larger arrays for
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.
Patch fixes Clang warnings about possible access via lv_name NULL pointer.
Replaces allocation of memory (strdup) with just pointer assignment
(since execve is being called anyway).
Checks for !*lv_name only when lv_name is defined.
(and as I'm not quite sure what state this really is - putting a FIXME
around - as this rather looks suspicios ??).
Check for failing 'stat' and skip this loop iteration
(since data in statbuf are invalid).
Check whether sysconf managed to find _SC_PAGESIZE.
Report at least debug warning about failing unlink
(logging scheme here seems to be a different then in lvm).
Duplicate terminal FDs and use similar code as is made in clvmd
and cleanup warns about missing open/close tests.
FIXME: Looks like we already have 3 instancies of the same code in lvm repo.
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.
When fsadm is test - it needs to execute lvm and fsadm from non-standard path
setting. So adding a support in fsadm script when user set LVM_BINARY, then
the lvm command invoced from fsadm will have the same PATH setting as before
entering fsadm command.
Fix Bug 738832 - core to disk log conversion fails with internal error
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: 2d8a2f35c77fdeef1dbe0ef791db8530d07826eb
The solution is to recognise in _alloc_init that if there are no mirrors
or stripes specified, then 'new_extents' should be zero.
After suspend/resume following a splitmirror op, call sync_local_dev_names
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.