[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [libvirt] [PATCHv4 12/18] blockjob: implement block copy for qemu



On 04/13/2012 08:46 AM, Jiri Denemark wrote:
> On Mon, Apr 09, 2012 at 21:52:21 -0600, Eric Blake wrote:
>> Minimal patch to wire up all the pieces in the previous patches
>> to actually enable a block copy job.  By minimal, I mean that
>> qemu creates the file (that is, no REUSE_EXT flag support yet),
>> SELinux must be disabled, a lock manager is not informed, and
>> the audit logs aren't updated.  But those will be added as
>> improvements in future patches.
>>
>> * src/qemu/qemu_driver.c (qemuDomainBlockCopy): New function.
>> (qemuDomainBlockRebase): Call it when appropriate.
>> ---
>>  src/qemu/qemu_driver.c |  114 +++++++++++++++++++++++++++++++++++++++++++++++-
>>  1 files changed, 113 insertions(+), 1 deletions(-)

>> +
>> +    priv = vm->privateData;
>> +    if (!(qemuCapsGet(priv->qemuCaps, QEMU_CAPS_DRIVE_MIRROR) &&
>> +          qemuCapsGet(priv->qemuCaps, QEMU_CAPS_DRIVE_REOPEN))) {
>> +        qemuReportError(VIR_ERR_OPERATION_INVALID, "%s",
> 
> We usually use VIR_ERR_CONFIG_UNSUPPORTED in such cases.

Sure, easy to fix.  Actually, looking at 'git grep -B3 -i "qemu binary"
src/qemu', I counted an INTERNAL_ERROR, an OPERATION_FAILED, and a
couple of OPERATION_INVALID that should all be cleaned up, to match the
fact that the majority of uses did indeed favor CONFIG_UNSUPPORTED.
I'll split the cleanup into a separate patch.

> 
>> +                        _("block copy is not supported with this QEMU binary"));
>> +        goto cleanup;
>> +    }
>> +    if (vm->persistent) {
>> +        qemuReportError(VIR_ERR_OPERATION_INVALID, "%s",
>> +                        _("domain is not transient"));
>> +        goto cleanup;
>> +    }
> 
> I guess I wasn't paying enough attention somewhere but why do we forbid block
> copy for persistent domains? I understand why we want to forbid certain
> operations when block copy is active but I don't see a reason for penalizing
> persistent domains.

It was in patch 8/18 where I added the restrictions, and hopefully
documented in that commit message why limiting things to transient is a
good first step:

1. the first client of live migration is oVirt, which uses transient domains

2. qemu does not (yet) provide a way to resume a mirror when restarting
a domain, so anything that would require restoring a domain from saved
state is broken: incoming migration, virsh restore, and virsh start
after a managed save.  But users would be upset if they saved a domain,
only to find out that they cannot then restore it, so I squelched things
one step earlier in the process, by preventing any save of a domain so
that we never have a broken save image in the first place.

My worry now comes from the fact that managedsave is on the list of
forbidden operations.  If a domain is transient, managedsave is already
forbidden (it is assumed that you either don't care about the domain if
the host dies, or that you are running a higher-level app like oVirt
that knows how to rebuild the guest on a different host).  But if a
guest is persistent, and you use the libvirt-guests init script, then
you have a right to assume that rebooting your host will resume your
guests in the same state that they were prior to the host going down -
because libvirt-guests uses managedsave.  If we allow a mirror job on a
persistent domain, we violate this assumption (libvirt-guests will fail
to save the guest).  Therefore, I forbid to start a mirror job on a
persistent domain, just as I forbid to 'virsh define' a transient domain
into a persistent domain if a mirror job is active.

If, at a later date, qemu comes up with a way to resume mirroring when
restarting a domain, we can relax these restrictions.

>> +
>> +    /* Actually start the mirroring */
>> +    qemuDomainObjEnterMonitorWithDriver(driver, vm);
>> +    ret = qemuMonitorDriveMirror(priv->mon, NULL, device, dest, format, flags);
>> +    if (ret == 0 && bandwidth != 0)
>> +        ret = qemuMonitorBlockJob(priv->mon, device, NULL, bandwidth, NULL,
>> +                                  BLOCK_JOB_SPEED_INTERNAL);
>> +    qemuDomainObjExitMonitorWithDriver(driver, vm);
> 
> Can BLOCK_JOB_SPEED_INTERNAL fail while block copy remains running? If so, we
> should try to abort the job in case we fail to set the speed, otherwise we
> will report an error and forget about the job while qemu will keep copying
> data.

Good catch, and virDomainBlockPull() suffers from the same
ramifications.  I think both code paths need the same fix.

>> @@ -11889,6 +12000,7 @@ static int
>>  qemuDomainBlockPull(virDomainPtr dom, const char *path, unsigned long bandwidth,
>>                      unsigned int flags)
>>  {
>> +    virCheckFlags(0, -1);
>>      return qemuDomainBlockRebase(dom, path, NULL, bandwidth, flags);
>>  }
> 
> Hmm, personally, I would make qemuDomainBlockPull call qemuDomainBlockJobImpl
> directly instead of going through qemuDomainBlockRebase while adding the flags
> check...

Easy enough to change.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

Attachment: signature.asc
Description: OpenPGP digital signature


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]