[libvirt] Qemu, libvirt, and CPU models
Hi, Sorry for the long message, but I didn't find a way to summarize the questions and issues and make it shorter. For people who don't know me: I have started to work recently on the Qemu CPU model code. I have been looking at how things work on libvirt+Qemu today w.r.t. CPU models, and I have some points I would like to understand better and see if they can be improved. I have two main points I would like to understand/discuss: 1) The relationship between libvirt's cpu_map.xml and the Qemu CPU model definitions. 2) How we could properly allow CPU models to be changed without breaking existing virtual machines? Note that for all the questions below, I don't expect that we design the whole solution and discuss every single detail in this thread. I just want to collectn suggestions, information about libvirt requirements and assumptions, and warnings about expected pitfalls before I start working on a solution on Qemu. 1) Qemu and cpu_map.xml I would like to understand how cpu_map.xml is supposed to be used, and how it is supposed to interact with the CPU model definitions provided by Qemu. More precisely: 1.1) Do we want to eliminate the duplication between the Qemu CPU definitions and cpu_map.xml? 1.1.1) If we want to eliminate the duplication, how can we accomplish that? What interfaces you miss, that Qemu could provide? 1.1.2) If the duplication has a purpose and you want to keep cpu_map.xml, then: - First, I would like to understand why libvirt needs cpu_map.xml? Is it part of the "public" interface of libvirt, or is it just an internal file where libvirt stores non-user-visible data? - How can we make sure there is no confusion between libvirt and Qemu about the CPU models? For example, what if cpu_map.xml says model 'Moo' has the flag 'foo' enabled, but Qemu disagrees? How do we guarantee that libvirt gets exactly what it expects from Qemu when it asks for a CPU model? We have "-cpu ?dump" today, but it's not the better interface we could have. Do you miss something in special in the Qemu<->libvirt interface, to help on that? 1.2) About the probing of available features on the host system: Qemu has code specialized to query KVM about the available features, and to check what can be enabled and what can't be enabled in a VM. On many cases, the available features match exactly what is returned by the CPUID instruction on the host system, but there are some exceptions: - Some features can be enabled even when the host CPU doesn't support it (because they are completely emulated by KVM, e.g. x2apic). - On many other cases, the feature may be available but we have to check if Qemu+KVM are really able to expose it to the guest (many features work this way, as many depend on specific support by the KVM kernel module and/or Qemu). I suppose libvirt does want to check which flags can be enabled in a VM, as it already have checks for host CPU features (e.g. src/cpu/cpu_x86.c:x86Compute()). But I also suppose that libvirt doesn't want to duplicate the KVM feature probing code present on Qemu, and in this case we could have an interface where libvirt could query for the actually-available CPU features. Would it be useful for libvirt? What's the best way to expose this interface? 1.3) Some features are not plain CPU feature bits: e.g. level=X can be set in "-cpu" argument, and other features are enabled/disabled by exposing specific CPUID leafs and not just a feature bit (e.g. PMU CPUID leaf support). I suppose libvirt wants to be able to probe for those features too, and be able to enable/disable them, right? 2) How to change an existing model and keep existing VMs working? Sometimes we have to update a CPU model definition because of some bug. Eamples: - The CPU models Conroe, Penrym and Nehalem, have level=2 set. This works most times, but it breaks CPU core/thread topology enumeration. We have to change those CPU models to use level=4 to fix the bug. - This can happen with plain CPU feature bits, too, not just "level": sometimes real-world CPU models have a feature that is not supported by Qemu+KVM yet, but when the kernel and Qemu finally starts to support it, we may want to enable it on existing CPU models. Sometimes a model simply has the wrong set of feature bits, and we have to fix it to have the right set of features. But if we simply change the existing model definition, this will break existing machines: - Today, it would break on live migration, but that's slightly easy to fix: we have to migrate the CPUID information too, to make sure we won't change the CPU under the guest OS feet. - Even if we fix live migration, simple "cold" migration will make the guest OS see a different CPU after a reboot, and that's undesirable too. Even if the Qemu developers disagree with me and decide that this is not a problem, libvirt may want to expose a more stable CPU to the guest, and some cooperation from Qemu would be ncessary. So, my questions are: About the libvirt<->Qemu interface: 2.1) What's the best mechanism to have different versions of a CPU model? An alias system like the one used by machine-types? How to implement this without confusing the existing libvirt probing code? 2.2) We have to make the CPU model version-choosing mechanism depend on the machine-type. e.g. if the user has a pc-1.0 machine using the Nehalem CPU model, we have to keep using the level=2 version of that CPU. But if the user chose a newer machine-type version, we can safely get the latest-and-greates version of the Nehalem CPU model. How to make this work without confusing libvirt? About the user<->libvirt interface: 2.3) How all this will interact with cpu_map.xml? Right now there's the assumption that the CPU model definitions are immutable, right? 2.4) How do you think libvirt would expose this "CPU model version" to the user? Should it just expose the unversioned CPU models to the user, and let Qemu or libvirt choose the right version based on machine-type? Should it expose only the versioned CPU models (because they are immutable) and never expose the unversioned aliases? Should it expose the unversioned alias, but change the Domain XML definition automatically to the versioned immutable one (like it happens with machine-type)? I don't plan to interfere on the libvirt interface design, but I suppose that libvirt design assumptions will be impacted by the solution we choose on Qemu. For example: right now libvirt seems to assume that CPU models are immutable. Are you going to keep this assumption in the libvirt interfaces? Because I am already willing to break this assumption on Qemu, although I would like to cooperate with libvirt and not break any requirements/assumptions without warning. -- Eduardo
On Tue, Mar 06, 2012 at 03:27:53PM -0300, Eduardo Habkost wrote:
Hi,
Sorry for the long message, but I didn't find a way to summarize the questions and issues and make it shorter.
For people who don't know me: I have started to work recently on the Qemu CPU model code. I have been looking at how things work on libvirt+Qemu today w.r.t. CPU models, and I have some points I would like to understand better and see if they can be improved.
I have two main points I would like to understand/discuss:
1) The relationship between libvirt's cpu_map.xml and the Qemu CPU model definitions.
We have several areas of code in which we use CPU definitions - Reporting the host CPU definition (virsh capabilities) - Calculating host CPU compatibility / baseline definitions - Checking guest / host CPU compatibility - Configuring the guest CPU definition libvirt targets multiple platforms, and our CPU handling code is designed to be common & sharable across all the libvirt drivers, VMWare, Xen, KVM, LXC, etc. Obviously for container based virt, only the host side of things is relevant. The libvirt CPU XML definition consists of - Model name - Vendor name - zero or more feature flags added/removed. A model name is basically just an alias for a bunch of feature flags, so that the CPU XML definitions are a) reasonably short b) have some sensible default baselines. The cpu_map.xml is the database of the CPU models that libvirt supports. We use this database to transform the CPU definition from the guest XML, into the hypervisor's own format. As luck would have it, the cpu_map.xml file contents match what QEMU has. This need not be the case though. If there is a model in the libvirt cpu_map.xml that QEMU doesn't know, we'll just pick the nearest matching QEMU cpu model & specify the fature flags to compensate. We could go one step further and just write out a cpu.conf file that we load in QEMU with -loadconfig. On Xen we would use the cpu_map.xml to generate the CPUID masks that Xen expects. Similarly for VMWare.
2) How we could properly allow CPU models to be changed without breaking existing virtual machines?
What is the scope of changes expected to CPU models ?
1) Qemu and cpu_map.xml
I would like to understand how cpu_map.xml is supposed to be used, and how it is supposed to interact with the CPU model definitions provided by Qemu. More precisely:
1.1) Do we want to eliminate the duplication between the Qemu CPU definitions and cpu_map.xml?
It isn't possible for us to the libvirt cpu_map.xml, since we need that across all our hypervisor targets.
1.1.1) If we want to eliminate the duplication, how can we accomplish that? What interfaces you miss, that Qemu could provide?
1.1.2) If the duplication has a purpose and you want to keep cpu_map.xml, then: - First, I would like to understand why libvirt needs cpu_map.xml? Is it part of the "public" interface of libvirt, or is it just an internal file where libvirt stores non-user-visible data? - How can we make sure there is no confusion between libvirt and Qemu about the CPU models? For example, what if cpu_map.xml says model 'Moo' has the flag 'foo' enabled, but Qemu disagrees? How do we guarantee that libvirt gets exactly what it expects from Qemu when it asks for a CPU model? We have "-cpu ?dump" today, but it's not the better interface we could have. Do you miss something in special in the Qemu<->libvirt interface, to help on that?
1.2) About the probing of available features on the host system: Qemu has code specialized to query KVM about the available features, and to check what can be enabled and what can't be enabled in a VM. On many cases, the available features match exactly what is returned by the CPUID instruction on the host system, but there are some exceptions: - Some features can be enabled even when the host CPU doesn't support it (because they are completely emulated by KVM, e.g. x2apic). - On many other cases, the feature may be available but we have to check if Qemu+KVM are really able to expose it to the guest (many features work this way, as many depend on specific support by the KVM kernel module and/or Qemu).
I suppose libvirt does want to check which flags can be enabled in a VM, as it already have checks for host CPU features (e.g. src/cpu/cpu_x86.c:x86Compute()). But I also suppose that libvirt doesn't want to duplicate the KVM feature probing code present on Qemu, and in this case we could have an interface where libvirt could query for the actually-available CPU features. Would it be useful for libvirt? What's the best way to expose this interface?
1.3) Some features are not plain CPU feature bits: e.g. level=X can be set in "-cpu" argument, and other features are enabled/disabled by exposing specific CPUID leafs and not just a feature bit (e.g. PMU CPUID leaf support). I suppose libvirt wants to be able to probe for those features too, and be able to enable/disable them, right?
The libvirt CPU definition does not currently store info about the level, family, model, stepping, xlevel or model_id items. We really ought to fix this, so that libvirt does have that info. Then we'd be able to write out a QEMU config that fully specified the exact model.
2) How to change an existing model and keep existing VMs working?
Sometimes we have to update a CPU model definition because of some bug. Eamples:
- The CPU models Conroe, Penrym and Nehalem, have level=2 set. This works most times, but it breaks CPU core/thread topology enumeration. We have to change those CPU models to use level=4 to fix the bug.
This is an example of why libvirt needs to represent the level/family etc in its CPU definition. That way, when a guest is first created, the XML will save the CPU model, feature flags, level, family, etc it is created with. Should the level be changed later, existing guests would then not be affected, only new guests would get the level=4
- This can happen with plain CPU feature bits, too, not just "level": sometimes real-world CPU models have a feature that is not supported by Qemu+KVM yet, but when the kernel and Qemu finally starts to support it, we may want to enable it on existing CPU models. Sometimes a model simply has the wrong set of feature bits, and we have to fix it to have the right set of features.
2.3) How all this will interact with cpu_map.xml? Right now there's the assumption that the CPU model definitions are immutable, right?
2.4) How do you think libvirt would expose this "CPU model version" to the user? Should it just expose the unversioned CPU models to the user, and let Qemu or libvirt choose the right version based on machine-type? Should it expose only the versioned CPU models (because they are immutable) and never expose the unversioned aliases? Should it expose the unversioned alias, but change the Domain XML definition automatically to the versioned immutable one (like it happens with machine-type)?
We should only expose unversioned CPU models, but then record the precise details of the current version in the guest XML.
I don't plan to interfere on the libvirt interface design, but I suppose that libvirt design assumptions will be impacted by the solution we choose on Qemu. For example: right now libvirt seems to assume that CPU models are immutable. Are you going to keep this assumption in the libvirt interfaces? Because I am already willing to break this assumption on Qemu, although I would like to cooperate with libvirt and not break any requirements/assumptions without warning.
Daniel -- |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| |: http://libvirt.org -o- http://virt-manager.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: http://entangle-photo.org -o- http://live.gnome.org/gtk-vnc :|
Thanks a lot for the explanations, Daniel. Comments about specific items inline. On Wed, Mar 07, 2012 at 02:18:28PM +0000, Daniel P. Berrange wrote:
I have two main points I would like to understand/discuss:
1) The relationship between libvirt's cpu_map.xml and the Qemu CPU model definitions.
We have several areas of code in which we use CPU definitions
- Reporting the host CPU definition (virsh capabilities) - Calculating host CPU compatibility / baseline definitions - Checking guest / host CPU compatibility - Configuring the guest CPU definition
libvirt targets multiple platforms, and our CPU handling code is designed to be common & sharable across all the libvirt drivers, VMWare, Xen, KVM, LXC, etc. Obviously for container based virt, only the host side of things is relevant.
The libvirt CPU XML definition consists of
- Model name - Vendor name - zero or more feature flags added/removed.
A model name is basically just an alias for a bunch of feature flags, so that the CPU XML definitions are a) reasonably short b) have some sensible default baselines.
The cpu_map.xml is the database of the CPU models that libvirt supports. We use this database to transform the CPU definition from the guest XML, into the hypervisor's own format.
Understood. Makes sense.
As luck would have it, the cpu_map.xml file contents match what QEMU has. This need not be the case though. If there is a model in the libvirt cpu_map.xml that QEMU doesn't know, we'll just pick the nearest matching QEMU cpu model & specify the fature flags to compensate.
Awesome. So, if Qemu and libvirt disagrees, libvirt will know that and add the necessary flags? That was my main worry. If disagreement between Qemu and libvirt is not a problem, it would make things much easier. ...but: Is that really implemented? I simply don't see libvirt doing that. I see code that calls "-cpu ?" to list the available CPU models, but no code calling "-cpu ?dump", or parsing the Qemu CPU definition config file. I even removed some random flags from the Nehalem model on my machine (running Fedora 16), and no additional flags were added.
We could go one step further and just write out a cpu.conf file that we load in QEMU with -loadconfig.
Sounds good. Anyway, I want to make everything configurable on the cpudef config file configurable on the command-line too, so both options (command-line or config file) would work.
On Xen we would use the cpu_map.xml to generate the CPUID masks that Xen expects. Similarly for VMWare.
2) How we could properly allow CPU models to be changed without breaking existing virtual machines?
What is the scope of changes expected to CPU models ?
We already have at least four cases, affecting different fields of the CPU definitions: A) Adding/removing flags. Exampes: - When the current set of flags is simply incorrect. See commit df07ec56 on qemu.git, where lots of flags that weren't supposed to be set were removed from some models. - When a new feature is now supported by Qemu+KVM and it's present on real-world CPUs, but our CPU definitions don't have the feature yet. e.g. x2apic, that is present on real-world Westmere CPUs but disabled on Qemu Westmere CPU definition. B) Changing "level" for some reason. One example: Conroe, Penrym and Nehalem have level=2, but need to have level>=4 to make CPU topology work, so they have to be changed. C) Enabling/disabling or overriding specific CPUID leafs. This isn't even configurable on the config files today, but I plan to allow it to be configured, otherwise users won't be able to enable/disable some features that are probed by the guest by simply looking at a CPUID leaf (e.g. the 0xA CPUID leaf that contains PMU information). The PMU leaf is an example where a CPU looks different by simply using a different Qemu or kernel version, and libvirt can't control the visibility of that feature to the guest: - If you start a Virtual Machine using Qemu-1.0 today, with the "pc-1.0" machine-type, the PMU CPUID leaf won't be visible to the guest (as Qemu-1.0 doesn't support the PMU leaf). - If you start a Virtual Machine using Qemu-1.1 in the future, using the "pc-1.1" machine-type, with a recent kernel, the PMU CPUID leaf _will_ be visible to the guest (as the qemu.git master branch supports it). Up to now, it is OK because the machine-type in theory help us control the feature, but we have a problem on this case: - If you start a Virtual Machine using Qemu-1.1 in the future, using the "pc-1.1" machine-type, using exactly the same command-line as above, but using an old kernel, the PMU CPUID leaf will _not_ be visible to the guest.
1) Qemu and cpu_map.xml
I would like to understand how cpu_map.xml is supposed to be used, and how it is supposed to interact with the CPU model definitions provided by Qemu. More precisely:
1.1) Do we want to eliminate the duplication between the Qemu CPU definitions and cpu_map.xml?
It isn't possible for us to the libvirt cpu_map.xml, since we need that across all our hypervisor targets.
OK, as you already explained. It's not a problem to me as long as things work as expected when Qemu and libvirt disagree about a CPU model definition. So, about the specific questions:
1.1.1) If we want to eliminate the duplication, how can we accomplish that? What interfaces you miss, that Qemu could provide?
1.1.2) If the duplication has a purpose and you want to keep cpu_map.xml, then: - First, I would like to understand why libvirt needs cpu_map.xml? Is it part of the "public" interface of libvirt, or is it just an internal file where libvirt stores non-user-visible data?
You answered that above.
- How can we make sure there is no confusion between libvirt and Qemu about the CPU models? For example, what if cpu_map.xml says model 'Moo' has the flag 'foo' enabled, but Qemu disagrees? How do we guarantee that libvirt gets exactly what it expects from Qemu when it asks for a CPU model? We have "-cpu ?dump" today, but it's not the better interface we could have. Do you miss something in special in the Qemu<->libvirt interface, to help on that?
So, it looks like either I am missing something on my tests or libvirt is _not_ probing the Qemu CPU model definitions to make sure libvirt gets all the features it expects. Also, I would like to ask if you have suggestions to implement the equivalent of "-cpu ?dump" in a more friendly and extensible way. Would a QMP command be a good alternative? Would a command-line option with json output be good enough? (Do we have any case of capability-querying being made using QMP before starting any actual VM, today?)
1.2) About the probing of available features on the host system: Qemu has code specialized to query KVM about the available features, and to check what can be enabled and what can't be enabled in a VM. On many cases, the available features match exactly what is returned by the CPUID instruction on the host system, but there are some exceptions: - Some features can be enabled even when the host CPU doesn't support it (because they are completely emulated by KVM, e.g. x2apic). - On many other cases, the feature may be available but we have to check if Qemu+KVM are really able to expose it to the guest (many features work this way, as many depend on specific support by the KVM kernel module and/or Qemu).
I suppose libvirt does want to check which flags can be enabled in a VM, as it already have checks for host CPU features (e.g. src/cpu/cpu_x86.c:x86Compute()). But I also suppose that libvirt doesn't want to duplicate the KVM feature probing code present on Qemu, and in this case we could have an interface where libvirt could query for the actually-available CPU features. Would it be useful for libvirt? What's the best way to expose this interface?
So, about the above: the cases where libvirt thinks a feature is available but Qemu knows it is not available are sort-of OK today, because Qemu would simply refuse to start and an error message would be returned to the user. But what about the features that are not available on the host CPU, libvirt will think it can't be enabled, but that _can_ be enabled? x2apic seems to be the only case today, but we may have others in the future.
1.3) Some features are not plain CPU feature bits: e.g. level=X can be set in "-cpu" argument, and other features are enabled/disabled by exposing specific CPUID leafs and not just a feature bit (e.g. PMU CPUID leaf support). I suppose libvirt wants to be able to probe for those features too, and be able to enable/disable them, right?
The libvirt CPU definition does not currently store info about the level, family, model, stepping, xlevel or model_id items. We really ought to fix this, so that libvirt does have that info. Then we'd be able to write out a QEMU config that fully specified the exact model.
OK, good to know that this is being planned.
2) How to change an existing model and keep existing VMs working?
Sometimes we have to update a CPU model definition because of some bug. Eamples:
- The CPU models Conroe, Penrym and Nehalem, have level=2 set. This works most times, but it breaks CPU core/thread topology enumeration. We have to change those CPU models to use level=4 to fix the bug.
This is an example of why libvirt needs to represent the level/family etc in its CPU definition. That way, when a guest is first created, the XML will save the CPU model, feature flags, level, family, etc it is created with. Should the level be changed later, existing guests would then not be affected, only new guests would get the level=4
Correct.
- This can happen with plain CPU feature bits, too, not just "level": sometimes real-world CPU models have a feature that is not supported by Qemu+KVM yet, but when the kernel and Qemu finally starts to support it, we may want to enable it on existing CPU models. Sometimes a model simply has the wrong set of feature bits, and we have to fix it to have the right set of features.
2.3) How all this will interact with cpu_map.xml? Right now there's the assumption that the CPU model definitions are immutable, right?
2.4) How do you think libvirt would expose this "CPU model version" to the user? Should it just expose the unversioned CPU models to the user, and let Qemu or libvirt choose the right version based on machine-type? Should it expose only the versioned CPU models (because they are immutable) and never expose the unversioned aliases? Should it expose the unversioned alias, but change the Domain XML definition automatically to the versioned immutable one (like it happens with machine-type)?
We should only expose unversioned CPU models, but then record the precise details of the current version in the guest XML.
Sounds good to me. That answers most of my questions about how libvirt would handle changes on CPU models. Now we need good mechanisms that allow libvirt to do that. If you have specific requirements or suggestions in mind, please let me know. -- Eduardo
On 03/07/2012 03:26 PM, Eduardo Habkost wrote:
Thanks a lot for the explanations, Daniel.
Comments about specific items inline.
- How can we make sure there is no confusion between libvirt and Qemu about the CPU models? For example, what if cpu_map.xml says model 'Moo' has the flag 'foo' enabled, but Qemu disagrees? How do we guarantee that libvirt gets exactly what it expects from Qemu when it asks for a CPU model? We have "-cpu ?dump" today, but it's not the better interface we could have. Do you miss something in special in the Qemu<->libvirt interface, to help on that?
So, it looks like either I am missing something on my tests or libvirt is _not_ probing the Qemu CPU model definitions to make sure libvirt gets all the features it expects.
Also, I would like to ask if you have suggestions to implement the equivalent of "-cpu ?dump" in a more friendly and extensible way. Would a QMP command be a good alternative? Would a command-line option with json output be good enough?
I'm not sure where we are are using "-cpu ?dump", but it sounds like we should be.
(Do we have any case of capability-querying being made using QMP before starting any actual VM, today?)
Right now, we have two levels of queries - the 'qemu -help' and 'qemu -device ?' output is gathered up front (we really need to patch things to cache that, rather than repeating it for every VM start). Then we start qemu with -S, query QMP, all before starting the guest (qemu -S is in fact necessary for setting some options that cannot be set in the current CLI but can be set via the monitor) - but right now that is the only point where we query QMP capabilities. If QMP can alter the CPU model prior to the initial start of the guest, then that would be a sufficient interface. But I'm worried that once we start qemu, even with qemu -S, that it's too late to alter the CPU model in use by that guest, and that libvirt should instead start querying these things in advance. We definitely want a machine-parseable construct, so querying over QMP rather than '-cpu ?dump' sounds like it might be nicer, but it would also be more work to set up libvirt to do a dry-run query of QMP capabilities without also starting a real guest.
But what about the features that are not available on the host CPU, libvirt will think it can't be enabled, but that _can_ be enabled? x2apic seems to be the only case today, but we may have others in the future.
That's where having an interface to probe qemu to see what capabilities are possible for any given cpu model would be worthwhile, so that libvirt can correlate the feature sets properly.
That answers most of my questions about how libvirt would handle changes on CPU models. Now we need good mechanisms that allow libvirt to do that. If you have specific requirements or suggestions in mind, please let me know.
I'll let others chime in with more responses, but I do appreciate you taking the time to coordinate this. -- Eric Blake eblake@redhat.com +1-919-301-3266 Libvirt virtualization library http://libvirt.org
On Wed, 2012-03-07 at 16:07 -0700, Eric Blake wrote:
On 03/07/2012 03:26 PM, Eduardo Habkost wrote:
Thanks a lot for the explanations, Daniel.
Comments about specific items inline.
- How can we make sure there is no confusion between libvirt and Qemu about the CPU models? For example, what if cpu_map.xml says model 'Moo' has the flag 'foo' enabled, but Qemu disagrees? How do we guarantee that libvirt gets exactly what it expects from Qemu when it asks for a CPU model? We have "-cpu ?dump" today, but it's not the better interface we could have. Do you miss something in special in the Qemu<->libvirt interface, to help on that?
So, it looks like either I am missing something on my tests or libvirt is _not_ probing the Qemu CPU model definitions to make sure libvirt gets all the features it expects.
Also, I would like to ask if you have suggestions to implement the equivalent of "-cpu ?dump" in a more friendly and extensible way. Would a QMP command be a good alternative? Would a command-line option with json output be good enough?
I'm not sure where we are are using "-cpu ?dump", but it sounds like we should be.
(Do we have any case of capability-querying being made using QMP before starting any actual VM, today?)
Right now, we have two levels of queries - the 'qemu -help' and 'qemu -device ?' output is gathered up front (we really need to patch things to cache that, rather than repeating it for every VM start).
Eric: In addition to VM start, it appears that the libvirt qemu driver also runs both the 32-bit and 64-bit qemu binaries 3 times each when fetching capabilities that appears to occur when fetching VM state. Noticed this on an openstack/nova compute node that queries vm state periodically. Seemed to be taking a long time. stracing libvirtd during these queries showed this sequence for each query: 6461 17:15:25.269464 execve("/usr/bin/qemu", ["/usr/bin/qemu", "-cpu", "?"], [/* 2 vars */]) = 0 6462 17:15:25.335300 execve("/usr/bin/qemu", ["/usr/bin/qemu", "-help"], [/* 2 vars */]) = 0 6463 17:15:25.393786 execve("/usr/bin/qemu", ["/usr/bin/qemu", "-device", "?", "-device", "pci-assign,?", "-device", "virtio-blk-pci,?"], [/* 2 vars */]) = 0 6466 17:15:25.841086 execve("/usr/bin/qemu-system-x86_64", ["/usr/bin/qemu-system-x86_64", "-cpu", "?"], [/* 2 vars */]) = 0 6468 17:15:25.906746 execve("/usr/bin/qemu-system-x86_64", ["/usr/bin/qemu-system-x86_64", "-help"], [/* 2 vars */]) = 0 6469 17:15:25.980520 execve("/usr/bin/qemu-system-x86_64", ["/usr/bin/qemu-system-x86_64", "-device", "?", "-device", "pci-assign,?", "-device", "virtio-blk-pci,?"], [/* 2 vars */]) = 0 Seems to add about a second per VM running on the host. The periodic scan thus takes a couple of minutes on a heavily loaded host -- several 10s of VMs. Not a killer, but we'd like to eliminate it. I see that libvirt does some level of caching of capabilities, checking the st_mtime of the binaries to detect changes. I haven't figured out when that caching comes into effect, but it doesn't prevent the execs above. So, I created a patch series that caches the results of parsing the output of these calls that I will post shortly for RFC. It eliminates most of such execs. I think it might obviate the existing capabilities caching, but I'm not sure. Haven't had time to look into it. Later, Lee Schermerhorn HPCS
Then we start qemu with -S, query QMP, all before starting the guest (qemu -S is in fact necessary for setting some options that cannot be set in the current CLI but can be set via the monitor) - but right now that is the only point where we query QMP capabilities. <snip>
On Wed, Mar 07, 2012 at 04:07:06PM -0700, Eric Blake wrote:
(Do we have any case of capability-querying being made using QMP before starting any actual VM, today?)
Right now, we have two levels of queries - the 'qemu -help' and 'qemu -device ?' output is gathered up front (we really need to patch things to cache that, rather than repeating it for every VM start).
That's what I feared. I was wondering if we had a better machine-friendly interface to make some of these queries, today.
Then we start qemu with -S, query QMP, all before starting the guest (qemu -S is in fact necessary for setting some options that cannot be set in the current CLI but can be set via the monitor) - but right now that is the only point where we query QMP capabilities.
If QMP can alter the CPU model prior to the initial start of the guest, then that would be a sufficient interface. But I'm worried that once we start qemu, even with qemu -S, that it's too late to alter the CPU model in use by that guest, and that libvirt should instead start querying these things in advance.
This is probably true, and I don't see this being changed in the near future. Even if we fix that for CPU initialization, there are many other initialization steps involved that would have to be reworked to allow all capability querying to be made to the same Qemu process that would run the VM later.
We definitely want a machine-parseable construct, so querying over QMP rather than '-cpu ?dump' sounds like it might be nicer, but it would also be more work to set up libvirt to do a dry-run query of QMP capabilities without also starting a real guest.
On the other hand, with QMP we would have a better interface that could be used for all other queries libvirt has to run. Instead of running Qemu multiple times for capability querying, just start a single Qemu process and make the capability queries using QMP. I don't know if this was discussed or considered before.
But what about the features that are not available on the host CPU, libvirt will think it can't be enabled, but that _can_ be enabled? x2apic seems to be the only case today, but we may have others in the future.
That's where having an interface to probe qemu to see what capabilities are possible for any given cpu model would be worthwhile, so that libvirt can correlate the feature sets properly.
Yes. The issue currently is that many things don't depend just on static CPU model or machine-type definitions, libvirt has to know what capabilities the kernel provides and Qemu will really be able to use. It will be a long way to fix this. Some features are simply not configurable yet, even on the command-line. They are just automatically used by Qemu when provided by the kernel. -- Eduardo
On Wed, Mar 07, 2012 at 19:26:25 -0300, Eduardo Habkost wrote:
Awesome. So, if Qemu and libvirt disagrees, libvirt will know that and add the necessary flags? That was my main worry. If disagreement between Qemu and libvirt is not a problem, it would make things much easier.
...but:
Is that really implemented? I simply don't see libvirt doing that. I see code that calls "-cpu ?" to list the available CPU models, but no code calling "-cpu ?dump", or parsing the Qemu CPU definition config file. I even removed some random flags from the Nehalem model on my machine (running Fedora 16), and no additional flags were added.
Right, currently we only detect if Qemu knows requested CPU model and use another one if not. We should really start using something like -cpu ?dump. However, since qemu may decide to change parts of the model according to, e.g., machine type, we would need something more dynamic. Something like, "hey Qemu, this is the machine type and CPU model we want to use, these are the features we want in this model, and we also want few additional features, please, tell us what the resulting CPU configuration is (if it is even possible to deliver such CPU on current host)". And the result would be complete CPU model, which may of course be different from what the qemu's configuration file says. We could then use the result to update domain XML (in a way similar to how we handle machine types) so that we can guarantee the guest will always see the same CPU. Once CPU is updated, we could just check with Qemu if it can provide such CPU and start (or refuse to start) the domain. Does it seem reasonable?
We could go one step further and just write out a cpu.conf file that we load in QEMU with -loadconfig.
Sounds good. Anyway, I want to make everything configurable on the cpudef config file configurable on the command-line too, so both options (command-line or config file) would work.
I'd be afraid of hitting the command line length limit if we specified all CPU details in it :-)
So, it looks like either I am missing something on my tests or libvirt is _not_ probing the Qemu CPU model definitions to make sure libvirt gets all the features it expects.
Also, I would like to ask if you have suggestions to implement the equivalent of "-cpu ?dump" in a more friendly and extensible way. Would a QMP command be a good alternative? Would a command-line option with json output be good enough?
I quite like the possible solution Anthony (or perhaps someone else) suggested some time ago (it may however be biased by my memory): qemu could provide a command line option that would take QMP command(s) and the result would be QMP response on stdout. We could use this interface for all kinds of probes with easily parsed output.
(Do we have any case of capability-querying being made using QMP before starting any actual VM, today?)
Not really. We only query QMP while for available QMP commands that we can used further on when the domain is running.
So, about the above: the cases where libvirt thinks a feature is available but Qemu knows it is not available are sort-of OK today, because Qemu would simply refuse to start and an error message would be returned to the user.
Really? In my experience qemu just ignored the feature it didn't know about without any error message and started the domain happily. It might be because libvirt doesn't use anything like -cpu ...,check or whatever is needed to make it fail. However, I think we should fix it.
But what about the features that are not available on the host CPU, libvirt will think it can't be enabled, but that _can_ be enabled? x2apic seems to be the only case today, but we may have others in the future.
I think qemu could tell us about those features during the probe phase (my first paragraph) and we would either use them with policy='force' or something similar. Jirka
On Thu, Mar 08, 2012 at 02:41:54PM +0100, Jiri Denemark wrote:
On Wed, Mar 07, 2012 at 19:26:25 -0300, Eduardo Habkost wrote:
Awesome. So, if Qemu and libvirt disagrees, libvirt will know that and add the necessary flags? That was my main worry. If disagreement between Qemu and libvirt is not a problem, it would make things much easier.
...but:
Is that really implemented? I simply don't see libvirt doing that. I see code that calls "-cpu ?" to list the available CPU models, but no code calling "-cpu ?dump", or parsing the Qemu CPU definition config file. I even removed some random flags from the Nehalem model on my machine (running Fedora 16), and no additional flags were added.
Right, currently we only detect if Qemu knows requested CPU model and use another one if not. We should really start using something like -cpu ?dump. However, since qemu may decide to change parts of the model according to, e.g., machine type, we would need something more dynamic. Something like, "hey Qemu, this is the machine type and CPU model we want to use, these are the features we want in this model, and we also want few additional features, please, tell us what the resulting CPU configuration is (if it is even possible to deliver such CPU on current host)". And the result would be complete CPU model, which may of course be different from what the qemu's configuration file says. We could then use the result to update domain XML (in a way similar to how we handle machine types) so that we can guarantee the guest will always see the same CPU. Once CPU is updated, we could just check with Qemu if it can provide such CPU and start (or refuse to start) the domain. Does it seem reasonable?
Absolutely. I would even advise libvirt to refrain from using "-cpu ?dump", as its semantics are likely to change.
We could go one step further and just write out a cpu.conf file that we load in QEMU with -loadconfig.
Sounds good. Anyway, I want to make everything configurable on the cpudef config file configurable on the command-line too, so both options (command-line or config file) would work.
I'd be afraid of hitting the command line length limit if we specified all CPU details in it :-)
True. I am already afraid of hitting the command-line length limit with Qemu as-is right now. ;-)
So, it looks like either I am missing something on my tests or libvirt is _not_ probing the Qemu CPU model definitions to make sure libvirt gets all the features it expects.
Also, I would like to ask if you have suggestions to implement the equivalent of "-cpu ?dump" in a more friendly and extensible way. Would a QMP command be a good alternative? Would a command-line option with json output be good enough?
I quite like the possible solution Anthony (or perhaps someone else) suggested some time ago (it may however be biased by my memory): qemu could provide a command line option that would take QMP command(s) and the result would be QMP response on stdout. We could use this interface for all kinds of probes with easily parsed output.
This is another case where command-line limits could be hit, isn't it? Reading QMP commands from a normal chardev (a socket, or even stdio) is already available, we just need to make sure the "query QMP without ever initializing a machine" use-case is working and really supported by Qemu.
So, about the above: the cases where libvirt thinks a feature is available but Qemu knows it is not available are sort-of OK today, because Qemu would simply refuse to start and an error message would be returned to the user.
Really? In my experience qemu just ignored the feature it didn't know about without any error message and started the domain happily. It might be because libvirt doesn't use anything like -cpu ...,check or whatever is needed to make it fail. However, I think we should fix it.
Correct, I was assuming that 'enforce' was being used. I forgot that libvirt doesn't use it today. I really think libvirt should be using 'enforce', the only problem is that there may be cases where an existing VM was working (but with a result unpredictable by by libvirt), and with 'enforce' it would stop working. This is very likely to happen when using the defualt "qemu64" CPU model, that has some AMD-only CPUID:8000_0000h bits set, but everybody probably expects it to work on Intel CPU hosts too.
But what about the features that are not available on the host CPU, libvirt will think it can't be enabled, but that _can_ be enabled? x2apic seems to be the only case today, but we may have others in the future.
I think qemu could tell us about those features during the probe phase (my first paragraph) and we would either use them with policy='force' or something similar.
Yes, that's the conclusion I was trying to reach: we really need better CPU feature probing. -- Eduardo
participants (5)
-
Daniel P. Berrange -
Eduardo Habkost -
Eric Blake -
Jiri Denemark -
Lee Schermerhorn