Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Each per-group watcher process writes its own instance status file. Once
that's done it tries to acquire an exclusive lock on the global file and
will proceed to read all status file, merging them based on each file's
mtime. If an instance is moved to another group, the newer status will
supersede that of an older file which hasn't yet been updated.

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

This will be used by the watcher to store the file's fstat(2). It must
be done from the filehandle.

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: René Nussbaumer <rn@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: René Nussbaumer <rn@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: Agata Murawska <agatamurawska@google.com>
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

We just need an object that has a list_owned method.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

Remove 15 second sleep when wait_for_sync is not set. LUInstanceCreate already
calls _WaitForSync with oneshot=True, which already performs an internal
wait-loop for disks to start syncing.

Signed-off-by: Apollon Oikonomopoulos <apollon@noc.grnet.gr>
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

lu.glm.list_owned becomes lu.owned_locks, which is clearer for the
reader.

Also rename three variables (which were before named owned_locks) to
make clearer what they track.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

The module is called “objects”, not “object”.

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

This is quite similar to evacuating a group, but the locking
is different.

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: René Nussbaumer <rn@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Watcher state files can stay around if node groups are removed. With
this patch they're removed after 21 days.

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

This patch brings a huge change to ganeti-watcher to make it aware of
node groups. Each node group is processed in its own subprocess,
reducing the impact of long-running operations.

The global watcher state file, $datadir/ganeti/watcher.data, is replaced
with a state file per node group ($datadir/ganeti/watcher.${uuid}.data).

Previously a lock on the state file was used to ensure only one instance
of watcher was running at the same time. Some operations, e.g.
“gnt-cluster renew-crypto”, blocked the watcher by acquiring an
exclusive lock on the state file. Since the watcher processes now use
different files, this method is no longer usable. Locking multiple files
isn't atomic. Instead a dedicated lock file is used and every watcher
process acquires a shared lock on it. If a Ganeti command wants to block
the watcher it acquires the lock in exclusive mode.

Each per-nodegroup watcher process also acquires an exclusive lock on
its state file. This prevents multiple watchers from running for the
same nodegroup.

The code is reorganized heavily to clear up dependencies between
functions and to get rid of the global “client” variable. The utility
class “Watcher” is removed in favour of stand-alone utility functions.

Since the parent watcher process won't wait for its children by
default, a new option (--wait-children) was added. It is used, for
example, by QA.

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

All potential target nodes should be locked while calculating
a group evacuation.

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

- Use OpPrereqError in CheckPrereq
- Clarify command synopsis

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

The same logic will be used for changing an instance's group.

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Including the designs which were actually implemented.

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

When wait_for_sync is set to False in LUInstanceCreate, Ganeti lets DRBD sync
in the background while performing the rest of the installation steps,
including OS installation.

However, OS installation is a very disk-intensive task that intereferes badly
with the background I/O caused by DRBD's initial sync. To this end, we pause
the background sync before OS installation and unpause it afterwards, which
yields a significant speed boost for OS installation. The following should be
noted:

a) The user has requested not to wait for sync, i.e. the instance will be
   non-redundant for an unspecified interval anyway and delaying this by a
   couple of minutes is not a big compromise.

b) This approach is also followed during disk wiping.

Signed-off-by: Apollon Oikonomopoulos <apollon@noc.grnet.gr>
[iustin@google.com: simplify an if check]
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Explain that we only start the instance on the new node if it was
originally running.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Just explain a bit the relation between node evacuate and instance
commands.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

“--dry-run” starts with two dashes.

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

This uses now the component for the transfer (if available), otherwise
(e.g. in installs/renames) nothing.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This modifies the RPC layer to export the component name too to the
backend, so that it can be used in log files and messages.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Currently, transfer data is done mainly with just the instance name,
but when we have instances with multiple disks this is not enough to
distinguish between the different transfers being done for the
instance.

Some parts of the code do have knowledge of the part being transferred
(i.e. DiskTransfer.name), but if I understood correctly not all, so I
decided to add a new argument to the respective disk import/disk
export classes.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Similar to the previous patch, this adds a helper function to
eliminate repeated calls info ConfigWriter.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This adds a new ConfigWriter.GetMultiNodeInfo function and replaces
multiple/looping calls to GetNodeInfo with it.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

It turns out that the only use of the operator module was for
itemgetter, so patch eb62069e should have removed that import too.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Signed-off-by: René Nussbaumer <rn@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

operator.itemgetter(0) → fst
operator.itemgetter(1) → snd

snd is not used yet, but it makes sense to add both.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

Signed-off-by: Pedro Macedo <pmacedo@google.com>
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Iallocator mode reloc, parameter reloc_from takes a list; half of the
code already forced this parameter to list, we add the other two cases
where it is needed.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Currently we use absolute symlinks, but this doesn't work when we
install remotely (due to install first to local temp dir, then rsync
to remote machines). To fix, we change to manually-computed relative
paths, which is not best, but it works.

One possible alternative would be to use hard-links…

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

By sleeping for 100ms after receiving a notification for a changed job
file the job is given some additional time to change again. This
significantly reduces the number of LUXI calls for WaitForJobChanges
(depending on the job, in my tests with “gnt-cluster verify
--debug-simulate-errors” by about 80%), and improves performance (the
same job went from around 7 seconds to around 3.5 seconds).

This method is not perfect. The algorithm could be made more complex,
e.g. by increasing the delay on each change, etc., but for now this
simple change provides a good improvement.

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>