[libvirt-users] virDomainMemoryPeek: bad behavior under workload

NoxDaFox noxdafox at gmail.com
Sat Sep 1 08:20:30 UTC 2012


On 31/08/12 18:09, Daniel P. Berrange wrote:
> On Fri, Aug 31, 2012 at 03:23:18PM +0300, NoxDaFox wrote:
>> Greetings,
>>
>> I am working on a platform for analysis automation.
>> I need to run several Virtual Environments concurrently and record
>> information about their behavior.
>>
>> I wrote some months ago about the capability of reading the Memory
>> during the Environment's execution (in paused state).
>> What do I need is the complete linear memory image, byte per byte,
>> nothing special; I will give this output to tools and parsers like
>> Volatility to get the value from it.
> If you want the complete memory image, perhaps you can just run the
> virDomainCoreDump() command, with the VIR_DUMP_MEMORY_ONLY flag
> (though this flag only works on very recent QEMU)
I'm working on Debian Wheezy, the flag you're talking about is in 
libvirt 0.9.13, it's in experimental now, do you know if it's going to 
be included into the next stable release?
I'm using libvirt 0.9.13 and QEMU 1.1.1 (from experimental as well) and 
the result is an error which tells the feature is not supported by QEMU, 
do I need an even more recent version?
>
>> I looked around and the only way to get the memory in such a way is
>> using the QEMU monitor command `pmemsave`.
>> I am using libvirt through its Python bindings and the
>> virDomainQemuMonitorCommand seems not to be exposed by the API so, as
>> suggested in some mails I read into the mailig list, I switched to
>> virDomainMemoryPeek.
>>
>> Using this function keeps up to 14-16 seconds to read 512Mb of memory
>> with the 64Kb limitation and 2-3 seconds with the 1Mb one; but the
>> most annoying thing is that I can't run several environment
>> concurrently as the function keeps failing.
> FYI, the virDomainMemoryPeek command was not really designed with
> scalability in mind, in particular not really intended for dumping
> the entire of guest memory. Its use case was tools like the virt-dmesg
> command, where you just want to peek at a handful of small memory
> regions.
What I realized is a simple Python iterator that allows to iterate 
through Disk and Memory with chunk of 64K or 1M depending on the libvirt 
version.
The users may have the need to read the entire memory before finding 
what they're looking for.

Do you think would be feasible to rework on this command allowing more 
scalability at least?
>
>
>> Here's the typical output:
>>
>>    File "/home/nox/workspace/NOX/src/NOX/hooks.py", line 134, in trigger
>>      hook.trigger(event)
>>    File "/home/nox/workspace/NOX/src/NOX/hooks.py", line 33, in trigger
>>      self.handlers[event]()
>>    File "/home/nox/workspace/NOX/hooks/volatility.py", line 81, in memory_dump
>>      for block in Memory(self.ctx):
>>    File "/home/see/workspace/NOX/src/NOX/lib/libtools.py", line 179, in next
>>      libvirt.VIR_MEMORY_PHYSICAL)
>>    File "/usr/lib/python2.7/dist-packages/libvirt.py", line 1759, in memoryPeek
>>      ret = libvirtmod.virDomainMemoryPeek(self._o, start, size, flags)
>> SystemError: error return without exception set
> Hmm, that's a peculiar message to see - I can't find anywhere in the
> libvirt code that uses that particular messages, so I'm not sure what
> has gone wrong here.
Is a Python message, it's triggered once a C method returns NULL.
>> I can't run more than 3 environments concurrently on a Xeon Quad with
>> 8Gb of memory.
>>
>> I guess the RPC reply goes in timeout because the system is under
>> heavy load but I'm not sure as the error output is quite obscure.
>> Is there any solution to this issue? Is it possible to raise the RPC
>> reply timeout value so that, even if slowly, I eventually get the
>> memory dump?
> For the memory peek API, we invoked a QEMU monitor command - we should
> not timeout on this at all, unless you are trying to invoke other
> monitor commands against the same QEMU process concurrently
So there's some other problem; I run several tests and seems to happen 
randomly, is not depending which block of memory is reading.
It is enough to read memory from 3 different running guests at the same 
time to see this issue.
>> If through virsh I use the QEMU `pmemsave` command, I get the memory
>> dump in less than one second; is there any way to obtain the same
>> performance?
> If virsh works properly, then this suggests the problem is somewhere
> in the python code, either libvirt's python binding, or your apps
> usage.
>
The virsh monitor command uses an unexposed API method: 
virDomainQemuMonitorCommand. This method is not available in the Python 
APIs neither is documented in libvirt documentation. I assume is for 
internal use only.

I can use this strategy as a workaround, but spawning a virsh process 
just for run this command while I can access to such resources as the 
connection to the Hypervisor and the Domain doesn't seem a clever 
solution to me.

NoxDaFox




More information about the libvirt-users mailing list