Fixed upstream: 

commit a877a1635b8a2cb88170767ca3c28c0d755b21a8
Author: Peter Krempa <pkrempa>
Date:   Tue Jun 14 11:23:23 2016 +0200

    conf: Remove pre-calculation of initial memory size
    
    While we need to know the difference between the total memory stored in
    <memory> and the actual size not included in the possible memory modules
    we can't pre-calculate it reliably. This is due to the fact that
    libvirt's XML is copied via formatting and parsing the XML and the
    initial memory size can be reliably calculated only when certain
    conditions are met due to backwards compatibility.
    
    This patch removes the storage of 'initial_memory' and fixes the helpers
    to recalculate the initial memory size all the time from the total
    memory size. This conversion is possible when we also make sure that
    memory hotplug accounts properly for the update of the total memory size
    and thus the helpers for inserting and removing memory devices need to
    be tweaked too.
    
    This fixes a bug where a cold-plug and cold-remove of a memory device
    would increase the size reported in <memory> in the XML by the size of
    the memory device. This would happen as the persistent definition is
    copied before attaching the device and this would lead to the loss of
    data in 'initial_memory'.

The problem described in the description have been fixed on libvirt-2.0.0-4.el7.x86_64:

1 prepare a running guest:
# virsh dumpxml r7
...
  <maxMemory slots='16' unit='KiB'>15243264</maxMemory>
  <memory unit='KiB'>2097152</memory>
  <currentMemory unit='KiB'>2097152</currentMemory>
...
  <cpu>
    <numa>
      <cell id='0' cpus='0-3' memory='1048576' unit='KiB'/>
    </numa>
  </cpu>
...
    <memory model='dimm'>
      <target>
        <size unit='KiB'>1048576</size>
        <node>0</node>
      </target>
      <alias name='dimm0'/>
      <address type='dimm' slot='0' base='0x100000000'/>
    </memory>
...

2. cold-unplug memory device:

# virsh detach-device r7 memdevice1g.xml --config
Device detached successfully

# cat memdevice1g.xml 
    <memory model='dimm'>
      <target>
        <size unit='KiB'>1048576</size>
        <node>0</node>
      </target>
    </memory>

3. recheck guest inactive xml:

# virsh dumpxml r7 --inactive
...
  <maxMemory slots='16' unit='KiB'>15242882</maxMemory>
  <memory unit='KiB'>1048576</memory>
  <currentMemory unit='KiB'>1048576</currentMemory>
...
  <cpu>
    <numa>
      <cell id='0' cpus='0-3' memory='1048576' unit='KiB'/>
    </numa>
  </cpu>
...

4. test cold-plug for a running guest:

# virsh dumpxml r7
...
  <maxMemory slots='16' unit='KiB'>15243264</maxMemory>
  <memory unit='KiB'>3145728</memory>
  <currentMemory unit='KiB'>3145728</currentMemory>
  <vcpu placement='static' current='6'>10</vcpu>
...
  <cpu>
    <numa>
      <cell id='0' cpus='0-3' memory='1048576' unit='KiB'/>
    </numa>
  </cpu>
...
    <memory model='dimm'>
      <target>
        <size unit='KiB'>1048576</size>
        <node>0</node>
      </target>
      <alias name='dimm0'/>
      <address type='dimm' slot='0' base='0x100000000'/>
    </memory>
    <memory model='dimm'>
      <target>
        <size unit='KiB'>1048576</size>
        <node>0</node>
      </target>
      <alias name='dimm1'/>
      <address type='dimm' slot='1' base='0x140000000'/>
    </memory>
...

# virsh dumpxml r7 --inactive
...
  <maxMemory slots='16' unit='KiB'>15243264</maxMemory>
  <memory unit='KiB'>2097152</memory>
  <currentMemory unit='KiB'>2097152</currentMemory>
  <vcpu placement='static' current='6'>10</vcpu>
...
  <cpu>
    <numa>
      <cell id='0' cpus='0-3' memory='1048576' unit='KiB'/>
    </numa>
  </cpu>
...
    <memory model='dimm'>
      <target>
        <size unit='KiB'>1048576</size>
        <node>0</node>
      </target>
    </memory>

# virsh attach-device r7 memdevice1g.xml --config
Device attached successfully

# virsh attach-device r7 memdevice1g.xml --config
Device attached successfully

# virsh dumpxml r7 --inactive
...
  <maxMemory slots='16' unit='KiB'>15243264</maxMemory>
  <memory unit='KiB'>4194304</memory>
  <currentMemory unit='KiB'>4194304</currentMemory>
...
    <memory model='dimm'>
      <target>
        <size unit='KiB'>1048576</size>
        <node>0</node>
      </target>
    </memory>
    <memory model='dimm'>
      <target>
        <size unit='KiB'>1048576</size>
        <node>0</node>
      </target>
    </memory>
    <memory model='dimm'>
      <target>
        <size unit='KiB'>1048576</size>
        <node>0</node>
      </target>
    </memory>
...

Hi Peter,

I found one problem during test the code introduce in patch:

if there is no numa in guest xml with memory device:

# virsh edit r7
...
  <memory unit='KiB'>1048576</memory>
...
    <memory model='dimm'>
      <target>
        <size unit='KiB'>1048576</size>
        <node>0</node>
      </target>
    </memory>
...

will get error:

error: XML error: Memory size must be specified via <memory> or in the <numa> configuration
Failed. Try again? [y,n,i,f,?]: 

This error looks not correct, since i specified memory in xml

Could you please help to check if this is a problem ?

Thanks,
Luyao

Well, your configuration is wrong. You don't have any startup memory since the memory module size is identical to the size of memory. Unfortunately the error message is not entirely correct in this case but this is a very narrow corner case. If you think it should be fixed please file a new bug.

(In reply to Peter Krempa from comment #6)
> Well, your configuration is wrong. You don't have any startup memory since
> the memory module size is identical to the size of memory. Unfortunately the
> error message is not entirely correct in this case but this is a very narrow
> corner case. If you think it should be fixed please file a new bug.

Thanks for your reply, i think no need to open a new bug for this corner issue.

And verify this bug with comment 4

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory, and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

https://rhn.redhat.com/errata/RHSA-2016-2577.html