We were using ->component_size while it hadn't been set.
This effectively meant that 'blocks' wasn't multiplied by
16 and reshape was even slower than it should have been.
Signed-off-by: NeilBrown <neilb@suse.de>
If an array goes degraded during reshape, we need to
adjust the devices we read from so as not to back up
stale data.
Signed-off-by: NeilBrown <neilb@suse.de>
The 'arraystart' is in sectors while restore_stripes requires
bytes, so we need a conversion.
Without this, backups get restored to the wrong offset.
Reported-by: "KueiHuan Chen" <kueihuan.chen@gmail.com>
Signed-off-by: NeilBrown <neilb@suse.de>
As mdadm is normally a short-lived program it isn't always necessary
to free memory that was allocated, as the 'exit()' call will
automatically free everything. But it is more obviously correct if
the 'free' is there.
So this patch add a few calls to 'free'
Signed-off-by: NeilBrown <neilb@suse.de>
Originally the backup-metadata was only written once at the
start of a raid5 reshape that made the array bigger. So we only
set the mtime once.
Now that we can be writing metadata continually during an in-place
reshape, we need to update the mtime more often.
Also, allow the metadata mtime to be slightly in advance of the
array mtime. Normally the difference will be less than a second,
so 10 minutes should be plenty. This guards against an old backup
file being used to restart an array. but starting two reshapes in the
10 minutes is sufficiently unlikely, and the possibility of an
accident is already sufficiently small, that 10 minutes is probably
fine.
Thanks to Guy Martin <gmsoft@tuxicoman.be> for discovering and
reporting that .mtime wasn't being updated properly.
Signed-off-by: NeilBrown <neilb@suse.de>
I think we sometime get way ahead of udev and devices disappear
and appear almost at random. So add some settling.
Signed-off-by: NeilBrown <neilb@suse.de>
ie. the percent increments after which RebuildNN event is generated
This is particulary useful when using --program option, rather than
(only) syslog for alerts.
Signed-off-by: Zdenek Behan <rain@matfyz.cz>
Signed-off-by: NeilBrown <neilb@suse.de>
mlockall(MCL_FUTURE) only locks mappings that have not yet
been created. To lock all memory used by the process, we need
MCL_CURRENT | MCL_FUTURE
Signed-off-by: NeilBrown <neilb@suse.de>
FIXME this is wrong . what direction does reshape_position move?
If the device count in an array is shrinking, the critical
region is different so the tests need to be different when
restarting.
Signed-off-by: NeilBrown <neilb@suse.de>
On small (test) arrays, multiplying by 16 can make the 'chunk' size
larger than half the array, which is a problem.
Signed-off-by: NeilBrown <neilb@suse.de>
Connect to the monitor in the old namespace and use that connection for
WaitClean requests when stopping the victim mdmon instance. This allows
ping_monitor() to work post chroot().
Cc: Hans de Goede <hdegoede@redhat.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Try to execute mdmon from the target namespace. When used for initramfs
handovers we need to drop all references to the initramfs filesystem for
that memory to be freed.
Cc: Hans de Goede <hdegoede@redhat.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
When killing a previous monitor be careful not to cause writes to the
filesystem until the reads necessary to get the monitor operational have
completed.
The code is already prepared for errors creating the pid and socket
files, so simply defer creation of these files until after the first
call to manage().
Cc: Hans de Goede <hdegoede@redhat.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
The load_super() from an mdadm --detail call may race against an mdmon
update. When this happens the load_super sees an inconsistent metadata
block and returns an error. The fallback path to use the map file
contents lacks uuid reporting, so provide __fname_from_uuid for
generically printing a uuid.
Reported-by: Hans de Goede <hdegoede@redhat.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Provide a test to sanity check assembly and reassembly in the presence
of conflicting family number information.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
When disks have conflicting container memberships (same container ids
but incompatible member arrays) --update=uuid can be used to move
offenders to a new container id by changing 'orig_family_num'.
Note that this only supports random updates of the uuid as the actual
uuid is synthesized. We also need to communicate the new
'orig_family_num' value to all disks involved in the update. A new
field 'update_private' is added to struct mdinfo to allow this
information to be transmitted.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
The full fix would be to support updating ddf metadata, but this minimal
fix just prevents the superblock from being zeroed when someone
inadvertently passes an unsupported --update option during assembly.
Reported-by: Hans de Goede <hdegoede@redhat.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Fix init_super_imsm() to return an empty mpb when info == NULL, and
teach store_super_imsm() to simply write out the passed in mpb.
Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=523320
Reported-by: Hans de Goede <hdegoede@redhat.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
imsm_activate_spare() in the manager thread may race against
write_super_imsm_spares() in the monitor thread. Give
write_super_imsm_spares() its own private mpb buffer to prevent
confusing the manager.
This change uncovered cases where spares were not being assembled due to
a failed metadata version number check. Spares can freely associate
across metadata version number, so reduce the scope of the version check
in the spare assembly case.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
The last time wait_backup is called, it might see reshape
finish and so return an error indicator.
But this is not an error, and we must go ahead and prepare
the array for full access.
Signed-off-by: NeilBrown <neilb@suse.de>