Change in vdsm[master]: Live Merge: Proactively extend merge target volume

Wednesday, 26 November 2014

Adam Litke has uploaded a new change for review.

Change subject: Live Merge: Proactively extend merge target volume
......................................................................

Live Merge: Proactively extend merge target volume

Since libvirt is not yet providing the write watermark information for
all volumes, we need to employ a workaround to support live merge on
block-based storage.  When starting a merge, we can request the base
volume (merge target) to be extended to the size of the top volume
(merge source) plus one chunk to accomodate active writing.  This may
cause some over-extension of the merge target volume but the upside is
that the merge should complete without active monitoring.  Future
versions of oVirt will take advantage of an upcoming libvirt enhancement
to report write watermarks for all volumes which will allow on-demand
volume extension.

Change-Id: I3d483f9dfe6e19a9f9fec84950bf2ea0c7ed7946
Bug-Url: https://bugzilla.redhat.com/show_bug.cgi?id=1155583
Signed-off-by: Adam Litke <alitke(a)redhat.com&gt;
---
M vdsm/virt/vm.py
1 file changed, 29 insertions(+), 73 deletions(-)


  git pull ssh://gerrit.ovirt.org:29418/vdsm refs/changes/14/35614/1

diff --git a/vdsm/virt/vm.py b/vdsm/virt/vm.py
index 1ae33b1..ab71892 100644
--- a/vdsm/virt/vm.py
+++ b/vdsm/virt/vm.py
@@ -870,8 +870,9 @@
         to be extended.  For the leaf volume curSize == self.apparentsize.
         For internal volumes it is discovered by calling irs.getVolumeSize().
         """
-        return (self.volExtensionChunk +
-                ((curSize + constants.MEGAB - 1) / constants.MEGAB))
+        nextSize = (self.volExtensionChunk +
+                    ((curSize + constants.MEGAB - 1) / constants.MEGAB))
+        return min(nextSize, self.truesize)
 
     @property
     def networkDev(self):
@@ -1832,69 +1833,12 @@
         with self._confLock:
             self.conf['timeOffset'] = newTimeOffset
 
-    def _getMergeWriteWatermarks(self):
-        drives = [drive for drive in self.getDiskDevices()
-                  if isVdsmImage(drive) and drive.blockDev]
-        allChains = self._driveGetActualVolumeChain(drives).values()
-        return dict((entry.uuid, entry.allocation)
-                    for chain in allChains
-                    for entry in chain)
-
-    def _getLiveMergeExtendCandidates(self):
-        ret = {}
-
-        # The common case is that there are no active jobs.
-        if not self.conf['_blockJobs'].values():
-            return ret
-
-        try:
-            watermarks = self._getMergeWriteWatermarks()
-        except LookupError:
-            self.log.warning("Failed to look up watermark information")
-            return ret
-
-        for job in self.conf['_blockJobs'].values():
-            try:
-                drive = self._findDriveByUUIDs(job['disk'])
-            except LookupError:
-                # After an active layer merge completes the vdsm metadata will
-                # be out of sync for a brief period.  If we cannot find the old
-                # disk then it's safe to skip it.
-                continue
-            if not drive.blockDev or drive.format != 'cow':
-                continue
-
-            if job['strategy'] == 'commit':
-                volumeID = job['baseVolume']
-            else:
-                self.log.debug("Unrecognized merge strategy '%s'",
-                               job['strategy'])
-                continue
-            volSize = self.cif.irs.getVolumeSize(drive.domainID, drive.poolID,
-                                                 drive.imageID, volumeID)
-            if volSize['status']['code'] != 0:
-                self.log.error("Unable to get the size of volume %s (domain: "
-                               "%s image: %s)", volumeID, drive.domainID,
-                               drive.imageID)
-                continue
-
-            if volumeID in watermarks:
-                self.log.debug("Adding live merge extension candidate: "
-                               "volume=%s", volumeID)
-                ret[drive.imageID] = {
-                    'alloc': watermarks[volumeID],
-                    'physical': int(volSize['truesize']),
-                    'capacity': drive.apparentsize,
-                    'volumeID': volumeID}
-            else:
-                self.log.warning("No watermark info available for %s",
-                                 volumeID)
-        return ret
-
     def _getExtendCandidates(self):
         ret = []
 
-        mergeCandidates = self._getLiveMergeExtendCandidates()
+        # FIXME: mergeCandidates should be a dictionary of candidate volumes
+        # once libvirt starts reporting watermark information for all volumes.
+        mergeCandidates = {}
         for drive in self._devices[DISK_DEVICES]:
             if not drive.blockDev or drive.format != 'cow':
                 continue
@@ -5172,7 +5116,6 @@
         if res['info']['voltype'] == 'SHARED':
             self.log.error("merge: Refusing to merge into a shared volume")
             return errCode['mergeErr']
-        baseSize = int(res['info']['apparentsize'])
 
         # Indicate that we expect libvirt to maintain the relative paths of
         # backing files.  This is necessary to ensure that a volume chain is
@@ -5186,6 +5129,19 @@
             # transitioned to the second phase.  We must then tell libvirt to
             # pivot to the new active layer (baseVolUUID).
             flags |= libvirt.VIR_DOMAIN_BLOCK_COMMIT_ACTIVE
+
+            # If top is the active layer, it's allocated size is stored in
+            # drive.apparantsize.
+            topSize = drive.apparentsize
+        else:
+            # If top is an internal volume, we must call getVolumeInfo
+            res = self.cif.irs.getVolumeInfo(drive.domainID, drive.poolID,
+                                             drive.imageID, topVolUUID)
+            if res['status']['code'] != 0:
+                self.log.error("Unable to get volume info for '%s'",
+                               topVolUUID)
+                return errCode['mergeErr']
+            topSize = int(res['info']['apparentsize'])
 
         # Take the jobs lock here to protect the new job we are tracking from
         # being cleaned up by queryBlockJobs() since it won't exist right away
@@ -5210,9 +5166,13 @@
 
         # blockCommit will cause data to be written into the base volume.
         # Perform an initial extension to ensure there is enough space to
-        # start copying.  The normal monitoring code will take care of any
-        # future internal volume extensions that may be necessary
-        self.extendDriveVolume(drive, baseVolUUID, baseSize)
+        # copy all the required data.  Normally we'd use monitoring to extend
+        # the volume on-demand but internal watermark information is not being
+        # reported by libvirt so we must do the full extension up front.  In
+        # the worst case, we'll need to extend 'base' to the same size as
'top'
+        # plus a bit more to accomodate additional writes to 'top' during the
+        # live merge operation.
+        self.extendDriveVolume(drive, baseVolUUID, topSize)
 
         # Trigger the collection of stats before returning so that callers
         # of getVmStats after this returns will see the new job
@@ -5240,14 +5200,10 @@
                 break
             sourceAttr = ('file', 'dev')[drive.blockDev]
             path = sourceXML.getAttribute(sourceAttr)
+
+            # TODO: Allocation information is not available in the XML.  Switch
+            # to the new interface once it becomes available in libvirt.
             alloc = None
-            if drive.blockDev:
-                allocXML = find_element_by_name(diskXML, 'allocation')
-                if not allocXML:
-                    self.log.debug("<allocation/> missing from backing "
-                                   "chain for block device %s", drive.name)
-                    break
-                alloc = int(allocXML.firstChild.nodeValue)
             bsXML = find_element_by_name(diskXML, 'backingStore')
             if not bsXML:
                 self.log.warning("<backingStore/> missing from backing "


-- 
To view, visit http://gerrit.ovirt.org/35614
To unsubscribe, visit http://gerrit.ovirt.org/settings

Gerrit-MessageType: newchange
Gerrit-Change-Id: I3d483f9dfe6e19a9f9fec84950bf2ea0c7ed7946
Gerrit-PatchSet: 1
Gerrit-Project: vdsm
Gerrit-Branch: master
Gerrit-Owner: Adam Litke <alitke(a)redhat.com&gt;

    

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011