Dan Kenigsberg has posted comments on this change.
Change subject: clientIF: automatically unpause vms in EIO when SD becomes
active
......................................................................
Patch Set 7: I would prefer that you didn't submit this
(8 inline comments)
....................................................
File vdsm/clientIF.py
Line 74: self.log = log
Line 75: self._recovery = True
Line 76: self.channelListener = Listener(self.log)
Line 77: self._generationID = str(uuid.uuid4())
Line 78: self.domainVmIds = {}
better define as _private.
it would be nicer if this dictionary contained the vmObjs directly, instead of
their IDs, saving one or two lookups.
But a more fundamental question: why should we bother to maintain the
domain->vms mapping, while a vm->domains already exists?
Line 79: self.domainVmIdsLock = threading.Lock()
Line 80: self._initIRS()
Line 81: self.irs.registerDomainStateChangeCallback(self.contEIOVms)
Line 82: self.mom = None
Line 77: self._generationID = str(uuid.uuid4())
Line 78: self.domainVmIds = {}
Line 79: self.domainVmIdsLock = threading.Lock()
Line 80: self._initIRS()
Line 81: self.irs.registerDomainStateChangeCallback(self.contEIOVms)
this has to be tucked inside _initIRS(), since we have an evil try-except block
there, making it possible for
self.irs == None
here.
Line 82: self.mom = None
Line 83: if _glusterEnabled:
Line 84: self.gluster = gapi.GlusterApi(self, log)
Line 85: else:
Line 141: if not isDomainStateValid:
Line 142: return
Line 143:
Line 144: libvirtCon = libvirtconnection.get()
Line 145: libvirtVms = libvirtCon.listAllDomains(
why should we re-collect this information from libvirt? Vdsm should already
have full information about all Vms, their state, and the reason for their
being paused.
Line 146: libvirt.VIR_CONNECT_LIST_DOMAINS_PAUSED)
Line 147:
Line 148: if sdUUID in self.domainVmIds:
Line 149: with self.vmContainerLock:
Line 406:
Line 407: return {'status': 0, 'alignment': aligning}
Line 408:
Line 409: def createVm(self, vmParams):
Line 410: with self.vmContainerLock:
valid cleanup, but unrelated to patch.
Line 411: self.log.info("vmContainerLock acquired by vm %s",
Line 412: vmParams['vmId'])
Line 413: try:
Line 414: if 'recover' not in vmParams:
....................................................
File vdsm/storage/hsm.py
Line 399: storageRefreshThread.daemon = True
Line 400: storageRefreshThread.start()
Line 401:
Line 402: @public
Line 403: def registerDomainStateChangeCallback(self, callbackFunc):
Doesn't this expose Vdsm to attack from its client, which could call the new
verb directly, passing specially-concocted callbackFunc pointers?
We trust our clients already, but making this @public over xmlrpc (albeit
undocumented) is not pretty. A better client/server API would be to block until
a domain state change (or timeout) occurs.
P.s. a real "register" verb should have an accompanying "unregister" one.
Line 404: """
Line 405: Currently this method assumes that all registrations
Line 406: are done *prior* to connecting to pool.
Line 407: """
Line 1052: res = pool.connect(hostID, scsiKey, msdUUID,
masterVersion)
Line 1053: if res:
Line 1054: self.pools[spUUID] = pool
Line 1055: for callback in self.domainStateChangeCallbacks:
Line 1056: self.pools[spUUID].domainMonitor.\
nit: better use "pool" directly, instead of looking it up time and again in
self.pools.
Line 1057: onDomainStateChange.register(callback)
Line 1058: return res
Line 1059:
Line 1060: @public
Line 1053: if res:
Line 1054: self.pools[spUUID] = pool
Line 1055: for callback in self.domainStateChangeCallbacks:
Line 1056: self.pools[spUUID].domainMonitor.\
Line 1057: onDomainStateChange.register(callback)
this may block for quite some time on libvirt's operations (listAllDomains,
cont). Can we do it after releasing the pool resource?
With this, a single hung-up qemu process can jam the storage system as a whole.
Line 1058: return res
Line 1059:
Line 1060: @public
Line 1061: def disconnectStoragePool(self, spUUID, hostID, scsiKey,
remove=False,
....................................................
File vdsm/vm.py
Line 2072: def isDisksStatsCollectionEnabled(self):
Line 2073: return self._volumesPrepared
Line 2074:
Line 2075: def preparePaths(self, drives):
Line 2076: domains = set()
in vm.py, "domain" usually means libvirt.virDomain. I think that being explicit
and calling this "storagedoms" would keep confusion checked.
Line 2077: for drive in drives:
Line 2078: with self._volPrepareLock:
Line 2079: if self.destroyed:
Line 2080: # A destroy request has been issued, exit early
--
To view, visit http://gerrit.ovirt.org/16244
To unsubscribe, visit http://gerrit.ovirt.org/settings
Gerrit-MessageType: comment
Gerrit-Change-Id: I0f88ebcab07dea20a0288f6b75c5888a74d9e440
Gerrit-PatchSet: 7
Gerrit-Project: vdsm
Gerrit-Branch: master
Gerrit-Owner: Yeela Kaplan <ykap...@redhat.com>
Gerrit-Reviewer: Arik Hadas <aha...@redhat.com>
Gerrit-Reviewer: Ayal Baron <aba...@redhat.com>
Gerrit-Reviewer: Dan Kenigsberg <dan...@redhat.com>
Gerrit-Reviewer: Eduardo <ewars...@redhat.com>
Gerrit-Reviewer: Federico Simoncelli <fsimo...@redhat.com>
Gerrit-Reviewer: Yeela Kaplan <ykap...@redhat.com>
Gerrit-Reviewer: oVirt Jenkins CI Server
_______________________________________________
vdsm-patches mailing list
vdsm-patches@lists.fedorahosted.org
https://lists.fedorahosted.org/mailman/listinfo/vdsm-patches
