On 12/15/2011 05:30 PM, Jiri Denemark wrote:

On Thu, Dec 15, 2011 at 08:58:55 -0600, Anthony Liguori wrote:

...

Pass '-readconfig /etc/qemu/target-x86_64.conf' to pick up those models and if you are absolutely insistent on not giving the user any ability to change things on their own, cp the file from qemu.git into libvirt.git and install it in a safe place.

Ah, this looks like a good idea (and we could even generate that file dynamically if we add support for family/stepping/... and other things that we do not model now). However, separating these definitions from qemu may result in incompatibilities with older qemu versions. I guess mainly because our configuration file would mention a CPU feature that an installed qemu version doesn't understand. Currently, qemu seems to just ignore such feature (although it prints an error) and continues happily without it. Is there any way for us to ask qemu what CPU features it knows about so that we could avoid using a CPU models which include features qemu doesn't understand?

Libvirt should use the 'enforce' option and there is also a 'check' one too:

qemu-kvm -cpu Conroe,+avx,enforce warning: host cpuid 0000_0001 lacks requested flag 'avx' [0x10000000] Unable to find x86 CPU definition

Qemu supports backward compatibility of the command line and if libvirt stores additional cpuid related definitions it should still work on other qemu releases. btw: I do prompt the usage of the 'formal' cpu models as much as it can. If a user likes to toggle various options it should do it on top of our models. Regards, Dor

Jirka

On Thu, Dec 15, 2011 at 03:42:44PM +0000, Daniel P. Berrange wrote:

On Thu, Dec 15, 2011 at 03:54:15PM +0100, Jiri Denemark wrote:

...

Hi,

Recently I realized that all modern CPU models defined in /etc/qemu/target-x86_64.conf are useless when qemu is used through libvirt. That's because we start qemu with -nodefconfig which results in qemu ignoring that file with CPU model definitions. We have a very good reason for using -nodefconfig because we need to control the ABI presented to a guest OS and we don't want any configuration file that can contain lots of things including device definitions to be read by qemu. However, we would really like the new CPU models to be understood by qemu even if used through libvirt. What would be the best way to solve this?

Ideally libvirt would just write out a config file with the CPU that was configured in libvirt and pass -readconfig /var/lib/libvirt/qemu/guestcpu.conf That way libvirt can configure CPU models without regard for what the particular QEMU version might have in its config.

And how libvirt would know that particular QEMU/kvm version can create this CPU. Managing cpuid shouldn't be libvirt busyness. -- Gleb.

On Fri, Mar 09, 2012 at 05:56:52PM -0300, Eduardo Habkost wrote:

Resurrecting an old thread:

I didn't see any clear conclusion in this thread (this is why I am resurrecting it), except that many were arguing that libvirt should simply copy and/or generate the CPU model definitions from Qemu. I really don't think it's reasonable to expect that.

On Thu, Dec 15, 2011 at 03:54:15PM +0100, Jiri Denemark wrote:

...

Hi,

Recently I realized that all modern CPU models defined in /etc/qemu/target-x86_64.conf are useless when qemu is used through libvirt. That's because we start qemu with -nodefconfig which results in qemu ignoring that file with CPU model definitions. We have a very good reason for using -nodefconfig because we need to control the ABI presented to a guest OS and we don't want any configuration file that can contain lots of things including device definitions to be read by qemu. However, we would really like the new CPU models to be understood by qemu even if used through libvirt. What would be the best way to solve this?

I suspect this could have been already discussed in the past but obviously a workable solution was either not found or just not implemented.

So, our problem today is basically:

A) libvirt uses -nodefconfig; B) -nodefconfig makes Qemu not load the config file containing the CPU model definitions; and C) libvirt expects the full CPU model list from Qemu to be available.

I could have sworn we had this discussion a year ago or so, and had decided that the default CPU models would be in something like /usr/share/qemu/cpu-x86_64.conf and loaded regardless of the -nodefconfig setting. /etc/qemu/target-x86_64.conf would be solely for end user configuration changes, not for QEMU builtin defaults. But looking at the code in QEMU, it doesn't seem we ever implemented this ? 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 :|

On Fri, Mar 09, 2012 at 03:15:26PM -0600, Anthony Liguori wrote:

On 03/09/2012 03:04 PM, Daniel P. Berrange wrote:

...

On Fri, Mar 09, 2012 at 05:56:52PM -0300, Eduardo Habkost wrote:

...

Resurrecting an old thread:

I didn't see any clear conclusion in this thread (this is why I am resurrecting it), except that many were arguing that libvirt should simply copy and/or generate the CPU model definitions from Qemu. I really don't think it's reasonable to expect that.

On Thu, Dec 15, 2011 at 03:54:15PM +0100, Jiri Denemark wrote:

...

Hi,

Recently I realized that all modern CPU models defined in /etc/qemu/target-x86_64.conf are useless when qemu is used through libvirt. That's because we start qemu with -nodefconfig which results in qemu ignoring that file with CPU model definitions. We have a very good reason for using -nodefconfig because we need to control the ABI presented to a guest OS and we don't want any configuration file that can contain lots of things including device definitions to be read by qemu. However, we would really like the new CPU models to be understood by qemu even if used through libvirt. What would be the best way to solve this?

I suspect this could have been already discussed in the past but obviously a workable solution was either not found or just not implemented.

So, our problem today is basically:

A) libvirt uses -nodefconfig; B) -nodefconfig makes Qemu not load the config file containing the CPU model definitions; and C) libvirt expects the full CPU model list from Qemu to be available.

I could have sworn we had this discussion a year ago or so, and had decided that the default CPU models would be in something like /usr/share/qemu/cpu-x86_64.conf and loaded regardless of the -nodefconfig setting. /etc/qemu/target-x86_64.conf would be solely for end user configuration changes, not for QEMU builtin defaults.

But looking at the code in QEMU, it doesn't seem we ever implemented this ?

I don't remember that discussion and really don't think I agree with the conclusion.

If libvirt wants to define CPU models on their own, they can. If libvirt wants to use the user's definitions, don't use -nodefconfig.

CPU models aren't a QEMU concept. The reason it's in the configuration file is to allow a user to add their own as they see fit. There is no right model names. It's strictly a policy.

Honestly, I don't think every single detail of CPU definitions should be considered user-provided configuration, the cpudefs are a low-level definitions of known-to-work combinations of CPUID bits, not a proper way for a user (or even a upper layer such as libvirt) to configure a CPU from scratch. At least a set of sane defaults to be used as base should be part of what defines a "machine type". If we don't do that, we will be unable to fix bugs on those CPU definitions without having to tell users to edit their config files, and libvirt users won't be able to use any new CPU feature until support for them is implemented on libvirt. And we would confuse two sets of users: 1) the ones that are using old cpudefs but don't understand why new features aren't available even on newer machine-types (and will have to read the Intel or AMD developer manuals to understand how to enable the feature); 2) the ones that get their config files upgraded automatically and don't understand why they are getting a different virtual machine after the upgrade, even when using an older machine-type. -- Eduardo

On Fri, Mar 09, 2012 at 09:04:03PM +0000, Daniel P. Berrange wrote:

On Fri, Mar 09, 2012 at 05:56:52PM -0300, Eduardo Habkost wrote:

...

Resurrecting an old thread:

I didn't see any clear conclusion in this thread (this is why I am resurrecting it), except that many were arguing that libvirt should simply copy and/or generate the CPU model definitions from Qemu. I really don't think it's reasonable to expect that.

On Thu, Dec 15, 2011 at 03:54:15PM +0100, Jiri Denemark wrote:

...

Hi,

Recently I realized that all modern CPU models defined in /etc/qemu/target-x86_64.conf are useless when qemu is used through libvirt. That's because we start qemu with -nodefconfig which results in qemu ignoring that file with CPU model definitions. We have a very good reason for using -nodefconfig because we need to control the ABI presented to a guest OS and we don't want any configuration file that can contain lots of things including device definitions to be read by qemu. However, we would really like the new CPU models to be understood by qemu even if used through libvirt. What would be the best way to solve this?

I suspect this could have been already discussed in the past but obviously a workable solution was either not found or just not implemented.

So, our problem today is basically:

A) libvirt uses -nodefconfig; B) -nodefconfig makes Qemu not load the config file containing the CPU model definitions; and C) libvirt expects the full CPU model list from Qemu to be available.

I could have sworn we had this discussion a year ago or so, and had decided that the default CPU models would be in something like /usr/share/qemu/cpu-x86_64.conf and loaded regardless of the -nodefconfig setting. /etc/qemu/target-x86_64.conf would be solely for end user configuration changes, not for QEMU builtin defaults.

But looking at the code in QEMU, it doesn't seem we ever implemented this ?

Arrrgggh. It seems this was implemented as a patch in RHEL-6 qemu RPMs but, contrary to our normal RHEL development practice, it was not based on a cherry-pick of an upstream patch :-( For sake of reference, I'm attaching the two patches from the RHEL6 source RPM that do what I'm describing NB, I'm not neccessarily advocating these patches for upstream. I still maintain that libvirt should write out a config file containing the exact CPU model description it desires and specify that with -readconfig. The end result would be identical from QEMU's POV and it would avoid playing games with QEMU's config loading code. Regards, 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 :|

On 03/10/2012 09:58 AM, Eduardo Habkost wrote:

On Sat, Mar 10, 2012 at 12:42:46PM +0000, Daniel P. Berrange wrote:

...

...

I could have sworn we had this discussion a year ago or so, and had decided that the default CPU models would be in something like /usr/share/qemu/cpu-x86_64.conf and loaded regardless of the -nodefconfig setting. /etc/qemu/target-x86_64.conf would be solely for end user configuration changes, not for QEMU builtin defaults.

But looking at the code in QEMU, it doesn't seem we ever implemented this ?

Arrrgggh. It seems this was implemented as a patch in RHEL-6 qemu RPMs but, contrary to our normal RHEL development practice, it was not based on a cherry-pick of an upstream patch :-(

For sake of reference, I'm attaching the two patches from the RHEL6 source RPM that do what I'm describing

NB, I'm not neccessarily advocating these patches for upstream. I still maintain that libvirt should write out a config file containing the exact CPU model description it desires and specify that with -readconfig. The end result would be identical from QEMU's POV and it would avoid playing games with QEMU's config loading code.

I agree that libvirt should just write the config somewhere. The problem here is to define: 1) what information should be mandatory on that config data; 2) who should be responsible to test and maintain sane defaults (and where should they be maintained).

The current cpudef definitions are simply too low-level to require it to be written from scratch. Lots of testing have to be done to make sure we have working combinations of CPUID bits defined, so they can be used as defaults or templates. Not facilitating reuse of those tested defauls/templates by libvirt is duplication of efforts.

Really, if we expect libvirt to define all the CPU bits from scratch on a config file, we could as well just expect libvirt to open /dev/kvm itself and call the all CPUID setup ioctl()s itself. That's how low-level some of the cpudef bits are.

Let's step back here. Why are you writing these patches? It's probably not because you have a desire to say -cpu Westmere when you run QEMU on your laptop. I'd wager to say that no human has ever done that or that if they had, they did so by accident because they read documentation and thought they had to. Humans probably do one of two things: 1) no cpu option or 2) -cpu host. So then why are you introducing -cpu Westmere? Because ovirt-engine has a concept of datacenters and the entire datacenter has to use a compatible CPU model to allow migration compatibility. Today, the interface that ovirt-engine exposes is based on CPU codenames. Presumably ovirt-engine wants to add a Westmere CPU group and as such have levied a requirement down the stack to QEMU. But there's no intrinsic reason why it uses CPU model names. VMware doesn't do this. It has a concept of compatibility groups[1]. oVirt could just as well define compatibility groups like GroupA, GroupB, GroupC, etc. and then the -cpu option we would be discussing would be -cpu GroupA. This is why it's a configuration option and not builtin to QEMU. It's a user interface as as such, should be defined at a higher level. Perhaps it really should be VDSM that is providing the model info to libvirt? Then they can add whatever groups then want whenever they want as long as we have the appropriate feature bits. P.S. I spent 30 minutes the other day helping a user who was attempting to figure out whether his processor was a Conroe, Penryn, etc. Making this determination is fairly difficult and it makes me wonder whether having CPU code names is even the best interface for oVirt.. [1] http://kb.vmware.com/selfservice/microsites/search.do?language=en_US&cmd=displayKC&externalId=1991 Regards, Anthony Liguori

(Also, there are additional low-level bits that really have to be maintained somewhere, just to have sane defaults. Currently many CPUID leafs are exposed to the guest without letting the user control them, and worse: without keeping stability of guest-visible bits when upgrading Qemu or the host kernel. And that's what machine-types are for: to have sane defaults to be used as base.)

Let me give you a practical example: I had a bug report about improper CPU topology information[1]. After investigating it, I have found out that the "level" cpudef field is too low; CPU core topology information is provided on CPUID leaf 4, and most of the Intel CPU models on Qemu have level=2 today (I don't know why). So, Qemu is responsible for exposing CPU topology information set using '-smp' to the guest OS, but libvirt would have to be responsible for choosing a proper "level" value that makes that information visible to the guest. We can _allow_ libvirt to fiddle with these low-level bits, of course, but requiring every management layer to build this low-level information from scratch is just a recipe to waste developer time.

(And I really hope that there's no plan to require all those low-level bits to appear as-is on the libvirt XML definitions. Because that would require users to read the Intel 64 and IA-32 Architectures Software Developer's Manual, or the AMD64 Architecture Programmer's Manual and BIOS and Kernel Developer's Guides, just to understand why something is not working on his Virtual Machine.)

[1] https://bugzilla.redhat.com/show_bug.cgi?id=689665

On Sat, Mar 10, 2012 at 12:24 PM, Anthony Liguori <anthony@codemonkey.ws> wrote:

On 03/10/2012 09:58 AM, Eduardo Habkost wrote:

...

On Sat, Mar 10, 2012 at 12:42:46PM +0000, Daniel P. Berrange wrote:

...

...

I could have sworn we had this discussion a year ago or so, and had decided that the default CPU models would be in something like /usr/share/qemu/cpu-x86_64.conf and loaded regardless of the -nodefconfig setting. /etc/qemu/target-x86_64.conf would be solely for end user configuration changes, not for QEMU builtin defaults.

But looking at the code in QEMU, it doesn't seem we ever implemented this ?

Arrrgggh. It seems this was implemented as a patch in RHEL-6 qemu RPMs but, contrary to our normal RHEL development practice, it was not based on a cherry-pick of an upstream patch :-(

For sake of reference, I'm attaching the two patches from the RHEL6 source RPM that do what I'm describing

NB, I'm not neccessarily advocating these patches for upstream. I still maintain that libvirt should write out a config file containing the exact CPU model description it desires and specify that with -readconfig. The end result would be identical from QEMU's POV and it would avoid playing games with QEMU's config loading code.

I agree that libvirt should just write the config somewhere. The problem here is to define: 1) what information should be mandatory on that config data; 2) who should be responsible to test and maintain sane defaults (and where should they be maintained).

The current cpudef definitions are simply too low-level to require it to be written from scratch. Lots of testing have to be done to make sure we have working combinations of CPUID bits defined, so they can be used as defaults or templates. Not facilitating reuse of those tested defauls/templates by libvirt is duplication of efforts.

Really, if we expect libvirt to define all the CPU bits from scratch on a config file, we could as well just expect libvirt to open /dev/kvm itself and call the all CPUID setup ioctl()s itself. That's how low-level some of the cpudef bits are.

Let's step back here.

Why are you writing these patches?  It's probably not because you have a desire to say -cpu Westmere when you run QEMU on your laptop.  I'd wager to say that no human has ever done that or that if they had, they did so by accident because they read documentation and thought they had to.

Humans probably do one of two things: 1) no cpu option or 2) -cpu host.

So then why are you introducing -cpu Westmere?  Because ovirt-engine has a concept of datacenters and the entire datacenter has to use a compatible CPU model to allow migration compatibility.  Today, the interface that ovirt-engine exposes is based on CPU codenames.  Presumably ovirt-engine wants to add a Westmere CPU group and as such have levied a requirement down the stack to QEMU.

But there's no intrinsic reason why it uses CPU model names.  VMware doesn't do this.  It has a concept of compatibility groups[1].

oVirt could just as well define compatibility groups like GroupA, GroupB, GroupC, etc. and then the -cpu option we would be discussing would be -cpu GroupA.

This is why it's a configuration option and not builtin to QEMU.  It's a user interface as as such, should be defined at a higher level.

Perhaps it really should be VDSM that is providing the model info to libvirt? Then they can add whatever groups then want whenever they want as long as we have the appropriate feature bits.

P.S. I spent 30 minutes the other day helping a user who was attempting to figure out whether his processor was a Conroe, Penryn, etc.  Making this determination is fairly difficult and it makes me wonder whether having CPU code names is even the best interface for oVirt..

[1] http://kb.vmware.com/selfservice/microsites/search.do?language=en_US&cmd=displayKC&externalId=1991

Regards,

Anthony Liguori

FWIW, as a user this would be a good improvement. As it stands right now when a cluster of machines is established as being redundant migratable machines for each other I must do the following for each machine: virsh -c qemu://machine/system capabilities | xpath /capabilities/host/cpu > machine-cpu.xml Once I have that data I combine them together and use virsh cpu-baseline, which is a handy addition from the past of doing it manually, but still not optimal. This gives me a model which is mostly meaningless and uninteresting to me, but I know all the guests must use Penryn for example. If ovirt and by extension libvirt let me know that guest X is running on CPU-A, I know I could migrate it to any other machine supporting CPU-A or CPU-B (assuming B is a super set of A). -- Doug Goldstein

On Sat, Mar 10, 2012 at 12:24:47PM -0600, Anthony Liguori wrote:

Let's step back here.

Why are you writing these patches? It's probably not because you have a desire to say -cpu Westmere when you run QEMU on your laptop. I'd wager to say that no human has ever done that or that if they had, they did so by accident because they read documentation and thought they had to.

I'd be glad if QEMU will chose -cpu Westmere for me if it detects Westmere host CPU as a default.

Humans probably do one of two things: 1) no cpu option or 2) -cpu host.

And both are not optimal. Actually both are bad. First one because default cpu is very conservative and the second because there is no guaranty that guest will continue to work after qemu or kernel upgrade. Let me elaborate about the later. Suppose host CPU has kill_guest feature and at the time a guest was installed it was not implemented by kvm. Since it was not implemented by kvm it was not present in vcpu during installation and the guest didn't install "workaround kill_guest" module. Now unsuspecting user upgrades the kernel and tries to restart the guest and fails. He writes angry letter to qemu-devel and is asked to reinstall his guest and move along.

So then why are you introducing -cpu Westmere? Because ovirt-engine has a concept of datacenters and the entire datacenter has to use a compatible CPU model to allow migration compatibility. Today, the interface that ovirt-engine exposes is based on CPU codenames. Presumably ovirt-engine wants to add a Westmere CPU group and as such have levied a requirement down the stack to QEMU.

First of all this is not about live migration only. Guest visible vcpu should not change after guest reboot (or hibernate/resume) too. And second this concept exists with only your laptop and single guest on it too. There are three inputs into a "CPU model module": 1) host cpu, 2) qemu capabilities, 3) kvm capabilities. With datacenters scenario all three can change, with your laptop only last two can change (first one can change too when you'll get new laptop) , but the net result is that guest visible cpuid can change and it shouldn't. This is the goal of introducing -cpu Westmere, to prevent it from happening.

But there's no intrinsic reason why it uses CPU model names. VMware doesn't do this. It has a concept of compatibility groups[1].

As Andrew noted, not any more. There is no intrinsic reason, but people are more familiar with Intel terminology than random hypervisor terminology.

oVirt could just as well define compatibility groups like GroupA, GroupB, GroupC, etc. and then the -cpu option we would be discussing would be -cpu GroupA. It could, but I can't see why this is less confusing.

This is why it's a configuration option and not builtin to QEMU. It's a user interface as as such, should be defined at a higher level.

This is not the only configuration that is builtin in QEMU. As it stands now QEMU does not even allow configuring cpuid enough to define those compatibility groups outside of QEMU. And after the work is done to allow enough configurability there is no much left to provide compatibility groups in QEMU itself.

Perhaps it really should be VDSM that is providing the model info to libvirt? Then they can add whatever groups then want whenever they want as long as we have the appropriate feature bits.

P.S. I spent 30 minutes the other day helping a user who was attempting to figure out whether his processor was a Conroe, Penryn, etc. Making this determination is fairly difficult and it makes me wonder whether having CPU code names is even the best interface for oVirt..

[1] http://kb.vmware.com/selfservice/microsites/search.do?language=en_US&cmd=displayKC&externalId=1991

Regards,

Anthony Liguori

...

(Also, there are additional low-level bits that really have to be maintained somewhere, just to have sane defaults. Currently many CPUID leafs are exposed to the guest without letting the user control them, and worse: without keeping stability of guest-visible bits when upgrading Qemu or the host kernel. And that's what machine-types are for: to have sane defaults to be used as base.)

Let me give you a practical example: I had a bug report about improper CPU topology information[1]. After investigating it, I have found out that the "level" cpudef field is too low; CPU core topology information is provided on CPUID leaf 4, and most of the Intel CPU models on Qemu have level=2 today (I don't know why). So, Qemu is responsible for exposing CPU topology information set using '-smp' to the guest OS, but libvirt would have to be responsible for choosing a proper "level" value that makes that information visible to the guest. We can _allow_ libvirt to fiddle with these low-level bits, of course, but requiring every management layer to build this low-level information from scratch is just a recipe to waste developer time.

(And I really hope that there's no plan to require all those low-level bits to appear as-is on the libvirt XML definitions. Because that would require users to read the Intel 64 and IA-32 Architectures Software Developer's Manual, or the AMD64 Architecture Programmer's Manual and BIOS and Kernel Developer's Guides, just to understand why something is not working on his Virtual Machine.)

[1] https://bugzilla.redhat.com/show_bug.cgi?id=689665

-- Gleb.

On Sun, Mar 11, 2012 at 09:12:58AM -0500, Anthony Liguori wrote:

On 03/11/2012 08:27 AM, Gleb Natapov wrote:

...

On Sat, Mar 10, 2012 at 12:24:47PM -0600, Anthony Liguori wrote:

...

Let's step back here.

Why are you writing these patches? It's probably not because you have a desire to say -cpu Westmere when you run QEMU on your laptop. I'd wager to say that no human has ever done that or that if they had, they did so by accident because they read documentation and thought they had to.

I'd be glad if QEMU will chose -cpu Westmere for me if it detects Westmere host CPU as a default.

This is -cpu best that Alex proposed FWIW.

...

...

Humans probably do one of two things: 1) no cpu option or 2) -cpu host.

And both are not optimal. Actually both are bad. First one because default cpu is very conservative and the second because there is no guaranty that guest will continue to work after qemu or kernel upgrade.

Let me elaborate about the later. Suppose host CPU has kill_guest feature and at the time a guest was installed it was not implemented by kvm. Since it was not implemented by kvm it was not present in vcpu during installation and the guest didn't install "workaround kill_guest" module. Now unsuspecting user upgrades the kernel and tries to restart the guest and fails. He writes angry letter to qemu-devel and is asked to reinstall his guest and move along.

-cpu best wouldn't solve this. You need a read/write configuration file where QEMU probes the available CPU and records it to be used for the lifetime of the VM. That what I thought too, but this shouldn't be the case (Avi's idea). We need two things: 1) CPU model config should be per machine type.

I didn't look at exact implementation but I doubt it does exactly what we need because currently we do not have infrastructure for that. If qemu is upgraded with support for new cpuid bits and -cpu best will pass them to a guest on next boot then this is not the same. -cpu Westmere can mean different thing for different machine types with proper infrastructure in place. 2) QEMU should refuse to start if it cannot create cpu exactly as specified by model config. With two conditions above if user creates VM with qemu 1.0 and cpu model Westmere which has no kill_guest feature he will still be able to run it in QEMU 1.1 (where kill_guest is added to Westmere model) and new kvm that support kill_guest by providing -M pc-1.0 flag (old definition of Westmere will be used). If user will try to create VM with QEMU 1.1 on a kernel that does not support kill_guest QEMU will refuse to start.

...

...

So then why are you introducing -cpu Westmere? Because ovirt-engine has a concept of datacenters and the entire datacenter has to use a compatible CPU model to allow migration compatibility. Today, the interface that ovirt-engine exposes is based on CPU codenames. Presumably ovirt-engine wants to add a Westmere CPU group and as such have levied a requirement down the stack to QEMU.

First of all this is not about live migration only. Guest visible vcpu should not change after guest reboot (or hibernate/resume) too. And second this concept exists with only your laptop and single guest on it too. There are three inputs into a "CPU model module": 1) host cpu, 2) qemu capabilities, 3) kvm capabilities. With datacenters scenario all three can change, with your laptop only last two can change (first one can change too when you'll get new laptop) , but the net result is that guest visible cpuid can change and it shouldn't. This is the goal of introducing -cpu Westmere, to prevent it from happening.

This discussion isn't about whether QEMU should have a Westmere processor definition. In fact, I think I already applied that patch.

It's a discussion about how we handle this up and down the stack.

The question is who should define and manage CPU compatibility. Right now QEMU does to a certain degree, libvirt discards this and does it's own thing, and VDSM/ovirt-engine assume that we're providing something and has built a UI around it.

If we want QEMU to be usable without management layer then QEMU should provide stable CPU models. Stable in a sense that qemu, kernel or CPU upgrade does not change what guest sees. If libvirt wants to override QEMU we should have a way to allow that, but than compatibility becomes libvirt problem. Figuring out what minimal CPU model that can be used across a cluster of different machines should be ovirt task.

What I'm proposing we consider: have VDSM manage CPU definitions in order to provide a specific user experience in ovirt-engine.

We would continue to have Westmere/etc in QEMU exposed as part of the user configuration. But I don't think it makes a lot of sense to have to modify QEMU any time a new CPU comes out.

If new cpu does not provide any new instruction set or capability that can be passed to a guest then there is no point creating CPU model for it in QEMU. If it does it is just a matter of updating config file. New CPUs are not something that pops up twice a month. -- Gleb.

On Sun, Mar 11, 2012 at 10:33:15AM -0500, Anthony Liguori wrote:

On 03/11/2012 09:56 AM, Gleb Natapov wrote:

...

...

-cpu best wouldn't solve this. You need a read/write configuration file where QEMU probes the available CPU and records it to be used for the lifetime of the VM. That what I thought too, but this shouldn't be the case (Avi's idea). We need two things: 1) CPU model config should be per machine type.

On Sun, Mar 11, 2012 at 09:12:58AM -0500, Anthony Liguori wrote: 2) QEMU should refuse to start if it cannot create cpu exactly as specified by model config.

This would either mean:

A. pc-1.1 uses -cpu best with a fixed mask for 1.1

B. pc-1.1 hardcodes Westmere or some other family

This would mean neither A nor B. May be it wasn't clear but I didn't talk about -cpu best above. I am talking about any CPU model with fixed meaning (not host or best which are host cpu dependant). Lets take Nehalem for example (just to move from Westmere :)). Currently it has level=2. Eduardo wants to fix it to be 11, but old guests, installed with -cpu Nehalem, should see the same CPU exactly. How do you do it? Have different Nehalem definition for pc-1.0 (which level=2) and pc-1.1 (with level=11). Lets get back to Westmere. It actually has level=11, but that's only expose another problem. Kernel 3.3 and qemu-1.1 combo will support architectural PMU which is exposed in cpuid leaf 10. We do not want guests installed with -cpu Westmere and qemu-1.0 to see architectural PMU after upgrade. How do you do it? Have different Westmere definitions for pc-1.0 (does not report PMU) and pc-1.1 (reports PMU). What happens if you'll try to run qemu-1.1 -cpu Westmere on Kernel < 3.3 (without PMU support)? Qemu will fail to start.

(A) would imply a different CPU if you moved the machine from one system to another. I would think this would be very problematic from a user's perspective.

(B) would imply that we had to choose the least common denominator which is essentially what we do today with qemu64. If you want to just switch qemu64 to Conroe, I don't think that's a huge difference from what we have today.

...

...

It's a discussion about how we handle this up and down the stack.

The question is who should define and manage CPU compatibility. Right now QEMU does to a certain degree, libvirt discards this and does it's own thing, and VDSM/ovirt-engine assume that we're providing something and has built a UI around it. If we want QEMU to be usable without management layer then QEMU should provide stable CPU models. Stable in a sense that qemu, kernel or CPU upgrade does not change what guest sees.

We do this today by exposing -cpu qemu64 by default. If all you're advocating is doing -cpu Conroe by default, that's fine. I am not advocating that. I am saying we should be able to amend qemu64 definition without breaking older guests that use it.

But I fail to see where this fits into the larger discussion here. The problem to solve is: I want to use the largest possible subset of CPU features available uniformly throughout my datacenter.

QEMU and libvirt have single node views so they cannot solve this problem on their own. Whether that subset is a generic Westmere-like processor that never existed IRL or a specific Westmere processor seems like a decision that should be made by the datacenter level manager with the node level view.

If I have a homogeneous environments of Xeon 7540, I would probably like to see a Xeon 7540 in my guest. Doesn't it make sense to enable the management tool to make this decision?

Of course neither QEMU nor libvirt can't made a cluster wide decision. If QEMU provides sane CPU model definitions (usable even with -nodefconfig) it would be always possible to find the model that fits best. If the oldest CPU in data center is Nehalem then probably -cpu Nehalem will do. But our CPU model definitions have a lot of shortcomings and we were talking with Edurado how to fix them when he brought this thread back to life, so may be I stirred the discussion a little bit in the wrong direction, but I do think those things are connected. If QEMU CPU model definitions are not stable across upgrades how can we say to management that it is safe to use them? Instead they insist in reimplementing the same logic in mngmt layer and do it badly (because the lack of info). -- Gleb.

On Mon, Mar 12, 2012 at 09:48:11AM -0500, Anthony Liguori wrote:

On 03/11/2012 11:16 AM, Gleb Natapov wrote:

...

On Sun, Mar 11, 2012 at 10:33:15AM -0500, Anthony Liguori wrote:

...

On 03/11/2012 09:56 AM, Gleb Natapov wrote:

...

...

-cpu best wouldn't solve this. You need a read/write configuration file where QEMU probes the available CPU and records it to be used for the lifetime of the VM. That what I thought too, but this shouldn't be the case (Avi's idea). We need two things: 1) CPU model config should be per machine type.

On Sun, Mar 11, 2012 at 09:12:58AM -0500, Anthony Liguori wrote: 2) QEMU should refuse to start if it cannot create cpu exactly as specified by model config.

This would either mean:

A. pc-1.1 uses -cpu best with a fixed mask for 1.1

B. pc-1.1 hardcodes Westmere or some other family

This would mean neither A nor B. May be it wasn't clear but I didn't talk about -cpu best above. I am talking about any CPU model with fixed meaning (not host or best which are host cpu dependant). Lets take Nehalem for example (just to move from Westmere :)). Currently it has level=2. Eduardo wants to fix it to be 11, but old guests, installed with -cpu Nehalem, should see the same CPU exactly. How do you do it? Have different Nehalem definition for pc-1.0 (which level=2) and pc-1.1 (with level=11). Lets get back to Westmere. It actually has level=11, but that's only expose another problem. Kernel 3.3 and qemu-1.1 combo will support architectural PMU which is exposed in cpuid leaf 10. We do not want guests installed with -cpu Westmere and qemu-1.0 to see architectural PMU after upgrade. How do you do it? Have different Westmere definitions for pc-1.0 (does not report PMU) and pc-1.1 (reports PMU). What happens if you'll try to run qemu-1.1 -cpu Westmere on Kernel< 3.3 (without PMU support)? Qemu will fail to start.

So, you're essentially proposing that -cpu Westmere becomes a machine option and that we let the machines interpret it as they see fit?

So --machine pc-1.0,cpu=Westmere would result in something different than --machine pc-1.1,cpu=Westmere?

Exactly.

That's something pretty different than what we're doing today. I think that we would have a single CPUX86 object and that part of the pc initialization process was to create an appropriately configured CPUx86 object.

Yes, that's different from what we're doing today, and it has to be fixed. (And, BTW, I'm really worried about your proposal that machine-types would suddenly disappear when using -nodefconfig in case we decide to move machine-type data to an external file one day. Design decisions aside, this would break an interface that management tools already have today.) -- Eduardo

On Mon, Mar 12, 2012 at 04:49:47PM +0100, Andreas Färber wrote:

Am 11.03.2012 17:16, schrieb Gleb Natapov:

...

On Sun, Mar 11, 2012 at 10:33:15AM -0500, Anthony Liguori wrote:

...

On 03/11/2012 09:56 AM, Gleb Natapov wrote:

...

...

-cpu best wouldn't solve this. You need a read/write configuration file where QEMU probes the available CPU and records it to be used for the lifetime of the VM. That what I thought too, but this shouldn't be the case (Avi's idea). We need two things: 1) CPU model config should be per machine type.

On Sun, Mar 11, 2012 at 09:12:58AM -0500, Anthony Liguori wrote: 2) QEMU should refuse to start if it cannot create cpu exactly as specified by model config.

This would either mean:

A. pc-1.1 uses -cpu best with a fixed mask for 1.1

B. pc-1.1 hardcodes Westmere or some other family

This would mean neither A nor B. May be it wasn't clear but I didn't talk about -cpu best above. I am talking about any CPU model with fixed meaning (not host or best which are host cpu dependant). Lets take Nehalem for example (just to move from Westmere :)). Currently it has level=2. Eduardo wants to fix it to be 11, but old guests, installed with -cpu Nehalem, should see the same CPU exactly. How do you do it? Have different Nehalem definition for pc-1.0 (which level=2) and pc-1.1 (with level=11). Lets get back to Westmere. It actually has level=11, but that's only expose another problem. Kernel 3.3 and qemu-1.1 combo will support architectural PMU which is exposed in cpuid leaf 10. We do not want guests installed with -cpu Westmere and qemu-1.0 to see architectural PMU after upgrade. How do you do it? Have different Westmere definitions for pc-1.0 (does not report PMU) and pc-1.1 (reports PMU). What happens if you'll try to run qemu-1.1 -cpu Westmere on Kernel < 3.3 (without PMU support)? Qemu will fail to start.

This sounds pretty much like what Liu Jinsong and Jan are discussing in the TSC thread on qemu-devel. (cc'ing)

I'll look for that thread. Thanks!

IMO interpreting an explicit -cpu parameter depending on -M would be wrong. Changing the default CPU based on -M is fine with me. For an explicit argument we would need Westmere-1.0 analog to pc-1.0. Then the user gets what the user asks for, without unexpected magic.

It is not unexpected magic. It would be a documented mechanism: "-cpu Nehalem-1.0" and "-cpu Nehalem-1.1" would have the same meaning every time, with any machine-type, but "-cpu Nehalem" would be an alias, whose meaning depends on the machine-type. Otherwise we would be stuck with a broken "Nehalem" model forever, and we don't want that.

Note that on my qom-cpu-wip branch [1] (that I hope to have cleaned up and sent out by tomorrow), all built-in CPUs become statically registered QOM types. The external definitions that get passed in via -cpudef become dynamically registered QOM types; I took care to allow overriding existing classes with the specified -cpudef fields (but untested). Setting family, level, etc. for -cpu is done on the X86CPU object instance. [2] What I don't have yet are QOM properties to set the fields from, e.g., machine code, but those should be fairly easy to add.

Sounds interesting. I will have to take a look at the code to understand how it affects what's being discussed in this thread.

Andreas

[1] http://repo.or.cz/w/qemu/afaerber.git/shortlog/refs/heads/qom-cpu-wip

[2] http://repo.or.cz/w/qemu/afaerber.git/commit/8a6ede101a2722b790489989f21cad3...

-- SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg

-- Eduardo

On 12 March 2012 17:41, Andreas Färber <afaerber@suse.de> wrote:

Also keep in mind linux-user. There's no concept of a machine there, but there's a cpu_copy() function used for forking that tries to re-create the CPU based on its model.

Incidentally, do you know why the linux-user code calls cpu_reset on the newly copied CPU state but only for TARGET_I386/SPARC/PPC ? That looks very odd to me... -- PMM

On Mon, Mar 12, 2012 at 06:53:27PM +0100, Andreas Färber wrote:

Am 12.03.2012 18:47, schrieb Peter Maydell:

...

On 12 March 2012 17:41, Andreas Färber <afaerber@suse.de> wrote:

...

Also keep in mind linux-user. There's no concept of a machine there, but there's a cpu_copy() function used for forking that tries to re-create the CPU based on its model.

Incidentally, do you know why the linux-user code calls cpu_reset on the newly copied CPU state but only for TARGET_I386/SPARC/PPC ? That looks very odd to me...

Incidentally for i386 I do: cpu_reset() is intentionally not part of cpu_init() there because afterwards the machine or something sets whether this CPU is a "bsp" (Board Support Package? ;)) and only then Boot Strap Processor I guess :)

resets it.

For ppc and sparc I don't know but I'd be surprised if it's necessary for ppc... Alex?

Andreas

-- SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg

-- Gleb.

On Mon, Mar 12, 2012 at 06:41:06PM +0100, Andreas Färber wrote:

Am 12.03.2012 17:50, schrieb Eduardo Habkost:

...

On Mon, Mar 12, 2012 at 04:49:47PM +0100, Andreas Färber wrote:

...

Am 11.03.2012 17:16, schrieb Gleb Natapov:

...

On Sun, Mar 11, 2012 at 10:33:15AM -0500, Anthony Liguori wrote:

...

On 03/11/2012 09:56 AM, Gleb Natapov wrote:

...

On Sun, Mar 11, 2012 at 09:12:58AM -0500, Anthony Liguori wrote: > -cpu best wouldn't solve this. You need a read/write configuration > file where QEMU probes the available CPU and records it to be used > for the lifetime of the VM. That what I thought too, but this shouldn't be the case (Avi's idea). We need two things: 1) CPU model config should be per machine type. 2) QEMU should refuse to start if it cannot create cpu exactly as specified by model config.

This would either mean:

A. pc-1.1 uses -cpu best with a fixed mask for 1.1

B. pc-1.1 hardcodes Westmere or some other family

This would mean neither A nor B. May be it wasn't clear but I didn't talk about -cpu best above. I am talking about any CPU model with fixed meaning (not host or best which are host cpu dependant). Lets take Nehalem for example (just to move from Westmere :)). Currently it has level=2. Eduardo wants to fix it to be 11, but old guests, installed with -cpu Nehalem, should see the same CPU exactly. How do you do it? Have different Nehalem definition for pc-1.0 (which level=2) and pc-1.1 (with level=11). Lets get back to Westmere. It actually has level=11, but that's only expose another problem. Kernel 3.3 and qemu-1.1 combo will support architectural PMU which is exposed in cpuid leaf 10. We do not want guests installed with -cpu Westmere and qemu-1.0 to see architectural PMU after upgrade. How do you do it? Have different Westmere definitions for pc-1.0 (does not report PMU) and pc-1.1 (reports PMU). What happens if you'll try to run qemu-1.1 -cpu Westmere on Kernel < 3.3 (without PMU support)? Qemu will fail to start. [...] IMO interpreting an explicit -cpu parameter depending on -M would be wrong. Changing the default CPU based on -M is fine with me. For an explicit argument we would need Westmere-1.0 analog to pc-1.0. Then the user gets what the user asks for, without unexpected magic.

It is not unexpected magic. It would be a documented mechanism: "-cpu Nehalem-1.0" and "-cpu Nehalem-1.1" would have the same meaning every time, with any machine-type, but "-cpu Nehalem" would be an alias, whose meaning depends on the machine-type.

Otherwise we would be stuck with a broken "Nehalem" model forever, and we don't want that.

Not quite what I meant: In light of QOM we should be able to instantiate a CPU based on its name and optional parameters IMO. No dependency on the machine, please. An alias sure, but if the user explicitly says -cpu Nehalem then on 1.1 it should always be an alias to Nehalem-1.1 whether the machine is -M pc-0.15 or pc. If no -cpu was specified by the user, then choosing a default of Nehalem-1.0 for pc-1.0 is fine. Just trying to keep separate things separate here.

Those things are not separate. If user will get Nehalem-1.1 with -M pc-0.15 on qemu-1.1 it will get broken VM. If user uses -M pc-0.15 it should get exactly same machine it gets by running qemu-0.15. Guest should not be able to tell the difference. This is the reason -M exists, anything else is a bug. -- Gleb.

On 03/12/2012 01:30 PM, Eduardo Habkost wrote:

On Mon, Mar 12, 2012 at 06:41:06PM +0100, Andreas Färber wrote:

...

Am 12.03.2012 17:50, schrieb Eduardo Habkost:

...

On Mon, Mar 12, 2012 at 04:49:47PM +0100, Andreas Färber wrote: [...]

...

IMO interpreting an explicit -cpu parameter depending on -M would be wrong. Changing the default CPU based on -M is fine with me. For an explicit argument we would need Westmere-1.0 analog to pc-1.0. Then the user gets what the user asks for, without unexpected magic.

It is not unexpected magic. It would be a documented mechanism: "-cpu Nehalem-1.0" and "-cpu Nehalem-1.1" would have the same meaning every time, with any machine-type, but "-cpu Nehalem" would be an alias, whose meaning depends on the machine-type.

Otherwise we would be stuck with a broken "Nehalem" model forever, and we don't want that.

Not quite what I meant: In light of QOM we should be able to instantiate a CPU based on its name and optional parameters IMO. No dependency on the machine, please. An alias sure, but if the user explicitly says -cpu Nehalem then on 1.1 it should always be an alias to Nehalem-1.1 whether the machine is -M pc-0.15 or pc. If no -cpu was specified by the user, then choosing a default of Nehalem-1.0 for pc-1.0 is fine. Just trying to keep separate things separate here.

As Gleb explained, things aren't really separated: "qemu-1.1 -M pc-1.0 -cpu Nehalem" should result in the same machine as "qemu-1.0 -cpu Nehalem", no difference should be visible to the guest. simply make incompatible changes.

So this is easy. CPU's need to be qdev/QOM and the various cpuid settings need to be done through qdev properties. Then you can just add globals to the machine definition. No different than what we do with virtio-blk. Regards, Anthony Liguori

...

Also keep in mind linux-user. There's no concept of a machine there, but there's a cpu_copy() function used for forking that tries to re-create the CPU based on its model. So currently cpu_*_init(env->cpu_model_str) needs to be able to recreate an identical CPU through the central code path, without access to a QEMUMachine.

So just translate the CPU alias given to "-cpu" to the true CPU model name as soon as possible, at the command-line-handling code, so the rest of the code always see the true CPU model name.

After all, the need to make the aliases is a command-line interface compatibility problem, so it makes sense to handle this at the command-line-handling code.

On 03/12/2012 01:53 PM, Itamar Heim wrote:

On 03/11/2012 05:33 PM, Anthony Liguori wrote:

...

On 03/11/2012 09:56 AM, Gleb Natapov wrote:

...

...

-cpu best wouldn't solve this. You need a read/write configuration file where QEMU probes the available CPU and records it to be used for the lifetime of the VM. That what I thought too, but this shouldn't be the case (Avi's idea). We need two things: 1) CPU model config should be per machine type.

On Sun, Mar 11, 2012 at 09:12:58AM -0500, Anthony Liguori wrote: 2) QEMU should refuse to start if it cannot create cpu exactly as specified by model config.

This would either mean:

A. pc-1.1 uses -cpu best with a fixed mask for 1.1

B. pc-1.1 hardcodes Westmere or some other family

(A) would imply a different CPU if you moved the machine from one system to another. I would think this would be very problematic from a user's perspective.

(B) would imply that we had to choose the least common denominator which is essentially what we do today with qemu64. If you want to just switch qemu64 to Conroe, I don't think that's a huge difference from what we have today.

...

...

It's a discussion about how we handle this up and down the stack.

The question is who should define and manage CPU compatibility. Right now QEMU does to a certain degree, libvirt discards this and does it's own thing, and VDSM/ovirt-engine assume that we're providing something and has built a UI around it. If we want QEMU to be usable without management layer then QEMU should provide stable CPU models. Stable in a sense that qemu, kernel or CPU upgrade does not change what guest sees.

We do this today by exposing -cpu qemu64 by default. If all you're advocating is doing -cpu Conroe by default, that's fine.

But I fail to see where this fits into the larger discussion here. The problem to solve is: I want to use the largest possible subset of CPU features available uniformly throughout my datacenter.

QEMU and libvirt have single node views so they cannot solve this problem on their own. Whether that subset is a generic Westmere-like processor that never existed IRL or a specific Westmere processor seems like a decision that should be made by the datacenter level manager with the node level view.

If I have a homogeneous environments of Xeon 7540, I would probably like to see a Xeon 7540 in my guest. Doesn't it make sense to enable the management tool to make this decision?

literally, or in capabilities? literally means you want to allow passing the cpu name to be exposed to the guest?

Yes, literally. Xen exposes the host CPUID to the guest for PV. Both PHYP (IBM System P) and z/VM (IBM System Z) do the same. What does VMware expose to guests by default?

if in capabilities, how would it differ from choosing the correct "cpu family"? it wouldn't really be identical (say, number of cores/sockets and no VT for time being)

It's a trade off. From a RAS perspective, it's helpful to have information about the host available in the guest. If you're already exposing a compatible family, exposing the actual processor seems to be worth the extra effort.

ovirt allows to set "cpu family" per cluster. assume tomorrow it could do it an even more granular way. it could also do it automatically based on subset of flags on all hosts - but would it really make sense to expose a set of capabilities which doesn't exist in the real world (which iiuc, is pretty much aligned with the cpu families?), that users understand?

No, I think the lesson we've learned in QEMU (the hard way) is that exposing a CPU that never existed will cause something to break. Often times, that something is glibc or GCC which tends to be rather epic in terms of failure. Regards, Anthony Liguori

On 03/12/2012 02:12 PM, Itamar Heim wrote:

On 03/12/2012 09:01 PM, Anthony Liguori wrote:

...

It's a trade off. From a RAS perspective, it's helpful to have information about the host available in the guest.

If you're already exposing a compatible family, exposing the actual processor seems to be worth the extra effort.

only if the entire cluster is (and will be?) identical cpu.

At least in my experience, this isn't unusual.

or if you don't care about live migration i guess, which could be hte case for clouds, then again, not sure a cloud provider would want to expose the physical cpu to the tenant.

Depends on the type of cloud you're building, I guess.

...

...

ovirt allows to set "cpu family" per cluster. assume tomorrow it could do it an even more granular way. it could also do it automatically based on subset of flags on all hosts - but would it really make sense to expose a set of capabilities which doesn't exist in the real world (which iiuc, is pretty much aligned with the cpu families?), that users understand?

No, I think the lesson we've learned in QEMU (the hard way) is that exposing a CPU that never existed will cause something to break. Often times, that something is glibc or GCC which tends to be rather epic in terms of failure.

good to hear - I think this is the important part. so from that perspective, cpu families sounds the right abstraction for general use case to me. for ovirt, could improve on smaller/dynamic subsets of migration domains rather than current clusters and sounds like you would want to see "expose host cpu for non migratable guests, or for identical clusters".

Would it be possible to have a "best available" option in oVirt-engine that would assume that all processors are of the same class and fail an attempt to add something that's an older class? I think that most people probably would start with "best available" and then after adding a node fails, revisit the decision and start lowering the minimum CPU family (I'm assuming that it's possible to modify the CPU family over time). From a QEMU perspective, I think that means having per-family CPU options and then Alex's '-cpu best'. But presumably it's also necessary to be able to figure out in virsh capabilities what '-cpu best' would be. Regards, Anthony Liguori

_______________________________________________ Arch mailing list Arch@ovirt.org http://lists.ovirt.org/mailman/listinfo/arch

----- Original Message -----

On 03/12/2012 02:12 PM, Itamar Heim wrote:

...

On 03/12/2012 09:01 PM, Anthony Liguori wrote:

...

It's a trade off. From a RAS perspective, it's helpful to have information about the host available in the guest.

If you're already exposing a compatible family, exposing the actual processor seems to be worth the extra effort.

only if the entire cluster is (and will be?) identical cpu.

At least in my experience, this isn't unusual.

I can definitely see places choosing homogeneous hardware and upgrading every few years. Giving them max capabilities for their cluster sounds logical to me. Esp. cloud providers.

...

or if you don't care about live migration i guess, which could be hte case for clouds, then again, not sure a cloud provider would want to expose the physical cpu to the tenant.

Depends on the type of cloud you're building, I guess.

Wouldn't this affect a simple startup of a VM with a different CPU (if motherboard changed as well cause reactivation issues in windows and fun things like that)? Even if the cloud doesn't support live migration, they don't pin VMs to a host. User could shut it down and start it up again and it might run on a different node. Your ephemeral storage would be lost, but persistent image storage could still contain os info pertinent to cpu type. Btw, I don't see why internally they would not support live migration even for when they need to put a host in maintenance etc. live storage migration could take care of the ephemeral storage if that's the issue (albeit take a million years to finish).

...

...

...

ovirt allows to set "cpu family" per cluster. assume tomorrow it could do it an even more granular way. it could also do it automatically based on subset of flags on all hosts - but would it really make sense to expose a set of capabilities which doesn't exist in the real world (which iiuc, is pretty much aligned with the cpu families?), that users understand?

No, I think the lesson we've learned in QEMU (the hard way) is that exposing a CPU that never existed will cause something to break. Often times, that something is glibc or GCC which tends to be rather epic in terms of failure.

good to hear - I think this is the important part. so from that perspective, cpu families sounds the right abstraction for general use case to me. for ovirt, could improve on smaller/dynamic subsets of migration domains rather than current clusters and sounds like you would want to see "expose host cpu for non migratable guests, or for identical clusters".

Would it be possible to have a "best available" option in oVirt-engine that would assume that all processors are of the same class and fail an attempt to add something that's an older class?

I think that most people probably would start with "best available" and then after adding a node fails, revisit the decision and start lowering the minimum CPU family (I'm assuming that it's possible to modify the CPU family over time).

But then they'd already have VMs that were started with the better CPU and now it'd change under their feet? or would we start them up with the best and fail to start these VMs on the newly added hosts which have the lower cpu family/type?

From a QEMU perspective, I think that means having per-family CPU options and then Alex's '-cpu best'. But presumably it's also necessary to be able to figure out in virsh capabilities what '-cpu best' would be.

Regards,

Anthony Liguori

...

_______________________________________________ Arch mailing list Arch@ovirt.org http://lists.ovirt.org/mailman/listinfo/arch

-- libvir-list mailing list libvir-list@redhat.com https://www.redhat.com/mailman/listinfo/libvir-list

----- Original Message -----

On 03/12/2012 10:19 PM, Ayal Baron wrote:

...

----- Original Message -----

...

On 03/12/2012 02:12 PM, Itamar Heim wrote:

...

On 03/12/2012 09:01 PM, Anthony Liguori wrote:

...

It's a trade off. From a RAS perspective, it's helpful to have information about the host available in the guest.

If you're already exposing a compatible family, exposing the actual processor seems to be worth the extra effort.

only if the entire cluster is (and will be?) identical cpu.

At least in my experience, this isn't unusual.

I can definitely see places choosing homogeneous hardware and upgrading every few years. Giving them max capabilities for their cluster sounds logical to me. Esp. cloud providers.

they would get same performance as from the matching "cpu family". only difference would be if the guest known the name of the host cpu.

...

...

...

or if you don't care about live migration i guess, which could be hte case for clouds, then again, not sure a cloud provider would want to expose the physical cpu to the tenant.

Depends on the type of cloud you're building, I guess.

Wouldn't this affect a simple startup of a VM with a different CPU (if motherboard changed as well cause reactivation issues in windows and fun things like that)?

that's an interesting question, I have to assume this works though, since we didn't see issues with changing the cpu family for guests so far.

assumption... :) I'd try changing twice in a row (run VM, stop, change family, restart VM, stop, change family restart VM).

On Mon, Mar 12, 2012 at 09:52:27AM -0300, Eduardo Habkost wrote:

On Sun, Mar 11, 2012 at 09:12:58AM -0500, Anthony Liguori wrote:

...

On 03/11/2012 08:27 AM, Gleb Natapov wrote:

...

On Sat, Mar 10, 2012 at 12:24:47PM -0600, Anthony Liguori wrote:

...

Let's step back here.

Why are you writing these patches? It's probably not because you have a desire to say -cpu Westmere when you run QEMU on your laptop. I'd wager to say that no human has ever done that or that if they had, they did so by accident because they read documentation and thought they had to.

No, it's because libvirt doesn't handle all the tiny small details involved in specifying a CPU. All libvirty knows about are a set of CPU flag bits, but it knows nothing about 'level', 'family', and 'xlevel', but we would like to allow it to expose a Westmere-like CPU to the guest.

This is easily fixable in libvirt - so for the point of going discussion, IMHO, we can assume libvirt will support level, family, xlevel, etc. 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 :|

On Mon, Mar 12, 2012 at 03:15:32PM +0200, Gleb Natapov wrote:

On Mon, Mar 12, 2012 at 01:04:19PM +0000, Daniel P. Berrange wrote:

...

On Mon, Mar 12, 2012 at 09:52:27AM -0300, Eduardo Habkost wrote:

...

On Sun, Mar 11, 2012 at 09:12:58AM -0500, Anthony Liguori wrote:

...

On 03/11/2012 08:27 AM, Gleb Natapov wrote:

...

On Sat, Mar 10, 2012 at 12:24:47PM -0600, Anthony Liguori wrote:

...

Let's step back here.

Why are you writing these patches? It's probably not because you have a desire to say -cpu Westmere when you run QEMU on your laptop. I'd wager to say that no human has ever done that or that if they had, they did so by accident because they read documentation and thought they had to.

No, it's because libvirt doesn't handle all the tiny small details involved in specifying a CPU. All libvirty knows about are a set of CPU flag bits, but it knows nothing about 'level', 'family', and 'xlevel', but we would like to allow it to expose a Westmere-like CPU to the guest.

This is easily fixable in libvirt - so for the point of going discussion, IMHO, we can assume libvirt will support level, family, xlevel, etc.

And fill in all cpuid leafs by querying /dev/kvm when needed or, if TCG is used, replicating QEMU logic? And since QEMU should be usable without libvirt the same logic should be implemented in QEMU anyway.

To implement this properly, libvirt will need a proper probing interface to know what exactly is available and can be enabled. I plan to implement that. I am have no problem in giving to libvirt the power to shoot itself in the foot. I believe libvirt developers can handle that. I have a problem with requiring every user (human or machine) to handle a weapon that can shoot their foot (that means, requiring the user to write the CPU model definition from scratch, or requiring the user to blindly copy&paste the default config file). -- Eduardo

On Mon, Mar 12, 2012 at 03:15:32PM +0200, Gleb Natapov wrote:

On Mon, Mar 12, 2012 at 01:04:19PM +0000, Daniel P. Berrange wrote:

...

On Mon, Mar 12, 2012 at 09:52:27AM -0300, Eduardo Habkost wrote:

...

On Sun, Mar 11, 2012 at 09:12:58AM -0500, Anthony Liguori wrote:

...

On 03/11/2012 08:27 AM, Gleb Natapov wrote:

...

On Sat, Mar 10, 2012 at 12:24:47PM -0600, Anthony Liguori wrote:

...

Let's step back here.

Why are you writing these patches? It's probably not because you have a desire to say -cpu Westmere when you run QEMU on your laptop. I'd wager to say that no human has ever done that or that if they had, they did so by accident because they read documentation and thought they had to.

No, it's because libvirt doesn't handle all the tiny small details involved in specifying a CPU. All libvirty knows about are a set of CPU flag bits, but it knows nothing about 'level', 'family', and 'xlevel', but we would like to allow it to expose a Westmere-like CPU to the guest.

This is easily fixable in libvirt - so for the point of going discussion, IMHO, we can assume libvirt will support level, family, xlevel, etc.

And fill in all cpuid leafs by querying /dev/kvm when needed or, if TCG is used, replicating QEMU logic? And since QEMU should be usable without libvirt the same logic should be implemented in QEMU anyway.

I'm not refering to that. I'm saying that any data QEMU has in its config file (/etc/qemu/target-x86_64.conf) should be represented in the libvirt CPU XML. family, model, stepping, xlevel and model_id are currently in QEMU CPU configs, but not in libvirt XML, which is something we will fix. The other issues you mention are completely independant of that. Regards, 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 :|

On Mon, Mar 12, 2012 at 03:53:38PM +0200, Gleb Natapov wrote:

On Mon, Mar 12, 2012 at 01:50:18PM +0000, Daniel P. Berrange wrote:

...

On Mon, Mar 12, 2012 at 03:15:32PM +0200, Gleb Natapov wrote:

...

On Mon, Mar 12, 2012 at 01:04:19PM +0000, Daniel P. Berrange wrote:

...

On Mon, Mar 12, 2012 at 09:52:27AM -0300, Eduardo Habkost wrote:

...

On Sun, Mar 11, 2012 at 09:12:58AM -0500, Anthony Liguori wrote:

...

On 03/11/2012 08:27 AM, Gleb Natapov wrote: >On Sat, Mar 10, 2012 at 12:24:47PM -0600, Anthony Liguori wrote: >>Let's step back here. >> >>Why are you writing these patches? It's probably not because you >>have a desire to say -cpu Westmere when you run QEMU on your laptop. >>I'd wager to say that no human has ever done that or that if they >>had, they did so by accident because they read documentation and >>thought they had to.

No, it's because libvirt doesn't handle all the tiny small details involved in specifying a CPU. All libvirty knows about are a set of CPU flag bits, but it knows nothing about 'level', 'family', and 'xlevel', but we would like to allow it to expose a Westmere-like CPU to the guest.

This is easily fixable in libvirt - so for the point of going discussion, IMHO, we can assume libvirt will support level, family, xlevel, etc.

And fill in all cpuid leafs by querying /dev/kvm when needed or, if TCG is used, replicating QEMU logic? And since QEMU should be usable without libvirt the same logic should be implemented in QEMU anyway.

I'm not refering to that. I'm saying that any data QEMU has in its config file (/etc/qemu/target-x86_64.conf) should be represented in the libvirt CPU XML. family, model, stepping, xlevel and model_id are currently in QEMU CPU configs, but not in libvirt XML, which is something we will fix. The other issues you mention are completely independant of that.

Eduardo is going to extend what can be configured in /etc/qemu/target-x86_64.conf and make CPU models name per machine type. What QEMU has now is not good enough. I doubt libvirt goal is to be as bad as QEMU :)

Of course not - libvirt will obviously be extended to cope with this too 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 :|

On 03/11/2012 04:12 PM, Anthony Liguori wrote:

...

Let me elaborate about the later. Suppose host CPU has kill_guest feature and at the time a guest was installed it was not implemented by kvm. Since it was not implemented by kvm it was not present in vcpu during installation and the guest didn't install "workaround kill_guest" module. Now unsuspecting user upgrades the kernel and tries to restart the guest and fails. He writes angry letter to qemu-devel and is asked to reinstall his guest and move along.

-cpu best wouldn't solve this. You need a read/write configuration file where QEMU probes the available CPU and records it to be used for the lifetime of the VM.

This doesn't work with live migration, and makes templating harder. The only persistent storage we can count on are disk images. The current approach is simple. The management tool determines the configuration, qemu applies it. Unidirectional information flow. This also lends itself to the management tool scanning a cluster and determining a GCD.

This discussion isn't about whether QEMU should have a Westmere processor definition. In fact, I think I already applied that patch.

It's a discussion about how we handle this up and down the stack.

The question is who should define and manage CPU compatibility. Right now QEMU does to a certain degree, libvirt discards this and does it's own thing, and VDSM/ovirt-engine assume that we're providing something and has built a UI around it.

What I'm proposing we consider: have VDSM manage CPU definitions in order to provide a specific user experience in ovirt-engine.

We would continue to have Westmere/etc in QEMU exposed as part of the user configuration. But I don't think it makes a lot of sense to have to modify QEMU any time a new CPU comes out.

We have to. New features often come with new MSRs which need to be live migrated, and of course the cpu flags as well. We may push all these to qemu data files, but this is still qemu. We can't let a management tool decide that cpu feature X is safe to use on qemu version Y. -- error compiling committee.c: too many arguments to function

On 03/25/2012 03:26 PM, Anthony Liguori wrote:

...

...

We would continue to have Westmere/etc in QEMU exposed as part of the user configuration. But I don't think it makes a lot of sense to have to modify QEMU any time a new CPU comes out.

We have to. New features often come with new MSRs which need to be live migrated, and of course the cpu flags as well. We may push all these to qemu data files, but this is still qemu. We can't let a management tool decide that cpu feature X is safe to use on qemu version Y.

I think QEMU should own CPU definitions.

Agree.

I think a management tool should have the choice of whether they are used though because they are a policy IMHO.

It's okay for QEMU to implement some degree of policy as long as a management tool can override it with a different policy.

Sure. We can have something like # default machine's westmere qemu -cpu westmere # pc-1.0's westmere qemu -M pc-1.0 -cpu westmere # pc-1.0's westmere, without nx-less qemu -M pc-1.0 -cpu westmere,-nx # specify everything in painful detail qemu -cpu vendor=Foo,family=17,model=19,stepping=3,maxleaf=12,+fpu,+vme,leaf10eax=0x1234567,+etc -- error compiling committee.c: too many arguments to function

On 03/11/2012 07:41 AM, Gleb Natapov wrote:

On Sat, Mar 10, 2012 at 12:58:43PM -0300, Eduardo Habkost wrote:

...

On Sat, Mar 10, 2012 at 12:42:46PM +0000, Daniel P. Berrange wrote:

...

...

I could have sworn we had this discussion a year ago or so, and had decided that the default CPU models would be in something like /usr/share/qemu/cpu-x86_64.conf and loaded regardless of the -nodefconfig setting. /etc/qemu/target-x86_64.conf would be solely for end user configuration changes, not for QEMU builtin defaults.

But looking at the code in QEMU, it doesn't seem we ever implemented this ?

Arrrgggh. It seems this was implemented as a patch in RHEL-6 qemu RPMs but, contrary to our normal RHEL development practice, it was not based on a cherry-pick of an upstream patch :-(

For sake of reference, I'm attaching the two patches from the RHEL6 source RPM that do what I'm describing

NB, I'm not neccessarily advocating these patches for upstream. I still maintain that libvirt should write out a config file containing the exact CPU model description it desires and specify that with -readconfig. The end result would be identical from QEMU's POV and it would avoid playing games with QEMU's config loading code.

I agree that libvirt should just write the config somewhere. The problem here is to define: 1) what information should be mandatory on that config data; 2) who should be responsible to test and maintain sane defaults (and where should they be maintained).

The current cpudef definitions are simply too low-level to require it to be written from scratch. Lots of testing have to be done to make sure we have working combinations of CPUID bits defined, so they can be used as defaults or templates. Not facilitating reuse of those tested defauls/templates by libvirt is duplication of efforts.

Really, if we expect libvirt to define all the CPU bits from scratch on a config file, we could as well just expect libvirt to open /dev/kvm itself and call the all CPUID setup ioctl()s itself. That's how low-level some of the cpudef bits are.

s/some/all

If libvirt assumes anything about what kvm actually supports it is working only by sheer luck.

Well the simple answer for libvirt is don't use -nodefconfig and then it can reuse the CPU definitions (including any that the user adds). Really, what's the point of having a layer of management if we're saying that doing policy management is too complicated for that layer? What does that layer exist to provide then?

...

(Also, there are additional low-level bits that really have to be maintained somewhere, just to have sane defaults. Currently many CPUID leafs are exposed to the guest without letting the user control them, and worse: without keeping stability of guest-visible bits when upgrading Qemu or the host kernel. And that's what machine-types are for: to have sane defaults to be used as base.)

Let me give you a practical example: I had a bug report about improper CPU topology information[1]. After investigating it, I have found out that the "level" cpudef field is too low; CPU core topology information is provided on CPUID leaf 4, and most of the Intel CPU models on Qemu have level=2 today (I don't know why). So, Qemu is responsible for exposing CPU topology information set using '-smp' to the guest OS, but libvirt would have to be responsible for choosing a proper "level" value that makes that information visible to the guest. We can _allow_ libvirt to fiddle with these low-level bits, of course, but requiring every management layer to build this low-level information from scratch is just a recipe to waste developer time. And QEMU become even less usable from a command line. One more point to kvm-tool I guess.

I'm not sure what your point is. We're talking about an option that humans don't use. How is this a discussion about QEMU usability? Regards, Anthony Liguori

On 03/11/2012 10:12 AM, Gleb Natapov wrote:

...

...

If libvirt assumes anything about what kvm actually supports it is working only by sheer luck.

Well the simple answer for libvirt is don't use -nodefconfig and then it can reuse the CPU definitions (including any that the user adds). CPU models should be usable even with -nodefconfig. CPU model is more

On Sun, Mar 11, 2012 at 09:16:49AM -0500, Anthony Liguori wrote: like device. By -cpu Nehalem I am saying I want Nehalem device in my machine.

Let's say we moved CPU definitions to /usr/share/qemu/cpu-models.xml. Obviously, we'd want a command line option to be able to change that location so we'd introduce -cpu-models PATH. But we want all of our command line options to be settable by the global configuration file so we would have a cpu-model=PATH to the configuration file. But why hard code a path when we can just set the default path in the configuration file so let's avoid hard coding and just put cpu-models=/usr/share/qemu/cpu-models.xml in the default configuration file. But now when libvirt uses -nodefconfig, those models go away. -nodefconfig means start QEMU in the most minimal state possible. You get what you pay for if you use it. We'll have the same problem with machine configuration files. At some point in time, -nodefconfig will make machine models disappear. Regards, Anthony Liguori

On Sun, Mar 11, 2012 at 10:41:32AM -0500, Anthony Liguori wrote:

On 03/11/2012 10:12 AM, Gleb Natapov wrote:

...

...

...

If libvirt assumes anything about what kvm actually supports it is working only by sheer luck.

Well the simple answer for libvirt is don't use -nodefconfig and then it can reuse the CPU definitions (including any that the user adds). CPU models should be usable even with -nodefconfig. CPU model is more

On Sun, Mar 11, 2012 at 09:16:49AM -0500, Anthony Liguori wrote: like device. By -cpu Nehalem I am saying I want Nehalem device in my machine.

Let's say we moved CPU definitions to /usr/share/qemu/cpu-models.xml.

Obviously, we'd want a command line option to be able to change that location so we'd introduce -cpu-models PATH.

But we want all of our command line options to be settable by the global configuration file so we would have a cpu-model=PATH to the configuration file.

But why hard code a path when we can just set the default path in the configuration file so let's avoid hard coding and just put cpu-models=/usr/share/qemu/cpu-models.xml in the default configuration file.

We wouldn't do the above. -nodefconfig should disable the loading of files on /etc, but it shouldn't disable loading internal non-configurable data that we just happened to choose to store outside the qemu binary because it makes development easier. Really, the requirement of a "default configuration file" is a problem by itself. Qemu should not require a default configuration file to work, and it shouldn't require users to copy the default configuration file to change options from the default. Doing this would make it impossible to deploy fixes to users if we evern find out that the default configuration file had a serious bug. What if a bug in our default configuration file has a serious security implication?

But now when libvirt uses -nodefconfig, those models go away. -nodefconfig means start QEMU in the most minimal state possible. You get what you pay for if you use it.

We'll have the same problem with machine configuration files. At some point in time, -nodefconfig will make machine models disappear.

It shouldn't. Machine-types are defaults to be used as base, they are not user-provided configuration. And the fact that we decided to store some data outside of the Qemu binary is orthogonal the design decisions in the Qemu command-line and configuration interface. As I said previously, requiring generation of opaque config files (and "copy the default config file and change it" is included on my definition of "generation of opaque config files") is poor design, IMO. I bet this even has an entry in some design anti-pattern catalog somewhere. -- Eduardo

On Tue, Mar 13, 2012 at 11:53:19AM -0300, Eduardo Habkost wrote:

So, trying to summarize what was discussed in the call:

On Mon, Mar 12, 2012 at 10:08:10AM -0300, Eduardo Habkost wrote:

...

...

Let's say we moved CPU definitions to /usr/share/qemu/cpu-models.xml.

Obviously, we'd want a command line option to be able to change that location so we'd introduce -cpu-models PATH.

But we want all of our command line options to be settable by the global configuration file so we would have a cpu-model=PATH to the configuration file.

But why hard code a path when we can just set the default path in the configuration file so let's avoid hard coding and just put cpu-models=/usr/share/qemu/cpu-models.xml in the default configuration file.

We wouldn't do the above.

-nodefconfig should disable the loading of files on /etc, but it shouldn't disable loading internal non-configurable data that we just happened to choose to store outside the qemu binary because it makes development easier.

The statement above is the one not fulfilled by the compromise solution: -nodefconfig would really disable the loading of files on /usr/share.

What does this mean? Will -nodefconfig disable loading of bios.bin, option roms, keymaps?

...

Really, the requirement of a "default configuration file" is a problem by itself. Qemu should not require a default configuration file to work, and it shouldn't require users to copy the default configuration file to change options from the default.

The statement above is only partly true. The default configuration file would be still needed, but if defaults are stored on /usr/share, I will be happy with it.

My main problem was with the need to _copy_ or edit a non-trivial default config file. If the not-often-edited defaults/templates are easily found on /usr/share to be used with -readconfig, I will be happy with this solution, even if -nodefconfig disable the files on /usr/share.

...

Doing this would make it impossible to deploy fixes to users if we evern find out that the default configuration file had a serious bug. What if a bug in our default configuration file has a serious security implication?

The answer to this is: if the broken templates/defaults are on /usr/share, it would be easy to deploy the fix.

So, the compromise solution is:

- We can move some configuration data (especially defaults/templates) to /usr/share (machine-types and CPU models could go there). This way we can easily deploy fixes to the defaults, if necessary. - To reuse Qemu models, or machine-types, and not define everything from scratch, libvirt will have to use something like: "-nodefconfig -readconfig /usr/share/qemu/cpu-models-x86.conf"

cpu-models-x86.conf is not a configuration file. It is hardware description file. QEMU should not lose capability just because you run it with -nodefconfig. -nodefconfig means that QEMU does not create machine for you, but all parts needed to create a machine that would have been created without -nodefconfig are still present. Not been able to create Nehalem CPU after specifying -nodefconfig is the same as not been able to create virtio-net i.e the bug.

(the item below is not something discussed on the call, just something I want to add)

To make this work better, we can allow users (humans or machines) to "extend" CPU models on the config file, instead of having to define everything from scratch. So, on /etc (or on a libvirt-generated config) we could have something like:

============= [cpu] base_cpudef = Nehalem add_features = "vmx" =============

Then, as long as /usr/share/cpu-models-x86.conf is loaded, the user will be able to reuse the Nehalem CPU model provided by Qemu.

And if it will not be loaded?

...

...

But now when libvirt uses -nodefconfig, those models go away. -nodefconfig means start QEMU in the most minimal state possible. You get what you pay for if you use it.

We'll have the same problem with machine configuration files. At some point in time, -nodefconfig will make machine models disappear.

It shouldn't. Machine-types are defaults to be used as base, they are not user-provided configuration. And the fact that we decided to store some data outside of the Qemu binary is orthogonal the design decisions in the Qemu command-line and configuration interface.

So, this problem is solved if the defaults are easily found on /usr/share.

What problem is solved and why are we mixing machine configuration files and cpu configuration files? They are different and should be treated differently. -nodefconfig exists only because there is not machine configuration files currently. With machine configuration files libvirt does not need -nodefconfig because it can create its own machine file and make QEMU use it. So specifying machine file on QEMU's command line implies -nodefconfig. The option itself loses its meaning and can be dropped.

We still have the backwards compatibility problem for pc-1.0, pc-1.1, and so on. But that can be discussed later, when we actually move machine-types to somewhere outside .c files.

...

As I said previously, requiring generation of opaque config files (and "copy the default config file and change it" is included on my definition of "generation of opaque config files") is poor design, IMO. I bet this even has an entry in some design anti-pattern catalog somewhere.

This problem is also solved if the defaults are deployed on /usr/share and just reused/included by the config files on /etc.

-- Gleb.

On Thu, Mar 22, 2012 at 10:31:21AM -0300, Eduardo Habkost wrote:

On Thu, Mar 22, 2012 at 11:32:44AM +0200, Gleb Natapov wrote:

...

On Tue, Mar 13, 2012 at 11:53:19AM -0300, Eduardo Habkost wrote:

...

So, trying to summarize what was discussed in the call:

On Mon, Mar 12, 2012 at 10:08:10AM -0300, Eduardo Habkost wrote:

...

...

Let's say we moved CPU definitions to /usr/share/qemu/cpu-models.xml.

Obviously, we'd want a command line option to be able to change that location so we'd introduce -cpu-models PATH.

But we want all of our command line options to be settable by the global configuration file so we would have a cpu-model=PATH to the configuration file.

But why hard code a path when we can just set the default path in the configuration file so let's avoid hard coding and just put cpu-models=/usr/share/qemu/cpu-models.xml in the default configuration file.

We wouldn't do the above.

-nodefconfig should disable the loading of files on /etc, but it shouldn't disable loading internal non-configurable data that we just happened to choose to store outside the qemu binary because it makes development easier.

The statement above is the one not fulfilled by the compromise solution: -nodefconfig would really disable the loading of files on /usr/share.

What does this mean? Will -nodefconfig disable loading of bios.bin, option roms, keymaps?

Correcting myself: loading of _config_ files on /usr/share. ROM images are opaque data to be presented to the guest somehow, just like a disk image or kernel binary. But maybe keymaps will become "configuration" someday, I really don't know.

Where do you draw the line between "opaque data" and configuration. CPU models are also something that is present to a guest somehow. Are you consider ROMs to be "opaque data" because they are binary and CPU models to be config just because it is ascii file? What if we pre-process CPU models into binary for QEMU to read will it magically stop being configuration?

...

...

...

Doing this would make it impossible to deploy fixes to users if we evern find out that the default configuration file had a serious bug. What if a bug in our default configuration file has a serious security implication?

The answer to this is: if the broken templates/defaults are on /usr/share, it would be easy to deploy the fix.

So, the compromise solution is:

- We can move some configuration data (especially defaults/templates) to /usr/share (machine-types and CPU models could go there). This way we can easily deploy fixes to the defaults, if necessary. - To reuse Qemu models, or machine-types, and not define everything from scratch, libvirt will have to use something like: "-nodefconfig -readconfig /usr/share/qemu/cpu-models-x86.conf"

cpu-models-x86.conf is not a configuration file. It is hardware description file. QEMU should not lose capability just because you run it with -nodefconfig. -nodefconfig means that QEMU does not create machine for you, but all parts needed to create a machine that would have been created without -nodefconfig are still present. Not been able to create Nehalem CPU after specifying -nodefconfig is the same as not been able to create virtio-net i.e the bug.

The current design direction Qemu seems to be following is different from that: hardware description is also considered "configuration" just like actual machine configuration. Anthony, please correct me if I am wrong.

That's a bug. Why trying to rationalize it now instead of fixing it. It was fixed in RHEL by the same person who introduced it in upstream in the first place. He just forgot to send the fix upstream. Does bug that is present for a long time is promoted to a feature?

What you propose is to have two levels of "configuration" (or descriptions, or whatever we call it):

1) Hardware descriptions (or templates, or models, whatever we call it), that are not editable by the user (and not disabled by -nodefconfig). This may include CPU models, hardware emulation implemented in another language, machine-types, and other stuff that is part of "what Qemu always provides". 2) Actual machine configuration file, that is configurable and editable by the user, and normally loaded from /etc on from the command-line.

The only problem is: the Qemu design simply doesn't have this distinction today (well, it _has_, the only difference is that today item (1) is almost completely coded inside tables in .c files). So if we want to go in that direction we have to agree this will be part of the Qemu design.

I am not strongly inclined either way. Both approaches look good to me, we just have to decide where we are going, because we're are in this weird position todady because we never decided it explicitly, libvirt expected one thing, and we implemented something else.

On the one hand I think the "two-layer" design gives us more freedom to move stuff outside .c files and change implementation details, and fits how we have been doing until today with machine types and built-in CPU models, keymaps, etc.

On the other hand, I think not having this distinction between "machine configuration" and "hardware description" may be a good thing.

For example: today there are two different ways of enabling a feature on a CPU: defining a new model, and adding a flag to "-cpu". And I think this asymmetry shouldn't be there: you just need a good system to define a CPU, a good set of defaults/templates, and a good system to base your configuration on those defaults/teampltes, no need to have two separate "CPU definition languages". Also, we wouldn't have to code things twice if we want to load an internal hardware description externally _and_ make some hardware details configurable one day: we just do it once: decide how it will be configured, and store it on the configuration file.

Maybe the same thing can be applied to machine types: machine types can become "configuration" too, but just a configuration template to be used as base, and can be augmented/extended on the user-provided configuration file.

...

...

(the item below is not something discussed on the call, just something I want to add)

To make this work better, we can allow users (humans or machines) to "extend" CPU models on the config file, instead of having to define everything from scratch. So, on /etc (or on a libvirt-generated config) we could have something like:

============= [cpu] base_cpudef = Nehalem add_features = "vmx" =============

Then, as long as /usr/share/cpu-models-x86.conf is loaded, the user will be able to reuse the Nehalem CPU model provided by Qemu.

And if it will not be loaded?

If it is not loaded, it is a configuration mistake. If you are reusing something defined somewhere, it would be your responsibility to make sure the file where the model is defined is present. On most cases you wouldn't use -nodefconfig and it would be shipped with Qemu and you shouldn't worry. If you used -nodefconfig, you load the CPU models file explicitly using -readconfig.

Regular user has no idea that we describe some cpu models in .c code and others in config file and those that are described in config files are disappear if -nodefconfig is used and that if they disappear they should be loaded back and how exactly they should be loaded back. The only user of -nodefconfig is libvirt and currently it does unexpected things for its only user.

...

...

...

...

But now when libvirt uses -nodefconfig, those models go away. -nodefconfig means start QEMU in the most minimal state possible. You get what you pay for if you use it.

We'll have the same problem with machine configuration files. At some point in time, -nodefconfig will make machine models disappear.

It shouldn't. Machine-types are defaults to be used as base, they are not user-provided configuration. And the fact that we decided to store some data outside of the Qemu binary is orthogonal the design decisions in the Qemu command-line and configuration interface.

So, this problem is solved if the defaults are easily found on /usr/share.

What problem is solved and why are we mixing machine configuration files and cpu configuration files? They are different and should be treated differently.

This is the root of the disagreement, it seems: they are not considered different today. Today cpudefs are on a config file inside /etc. One may argue that this was a mistake in the first place, but that's the design we have today.

Machine configuration files do not exist today! Machine configuration is done in .c code and -nodefconfig skips the code that does that, but it does not skip the code that describes cpu models and resides in .c. So as far as .c code is concerned machine creation and cpu description are two different things, but for some reason the file with rest of cpu model descriptions is not loaded. This is madness, not design.

...

-nodefconfig exists only because there is not machine configuration files currently. With machine configuration files libvirt does not need -nodefconfig because it can create its own machine file and make QEMU use it. So specifying machine file on QEMU's command line implies -nodefconfig. The option itself loses its meaning and can be dropped.

I think the approach today is:

There is not such thing as todays approach since there is not machine config file today.

- Qemu loads defaults from default config files; - Machine description files would be given using -readconfig and they would _augment_ the defaults from the default config files.

Jut make it possible to include one machine config file from another and you do not need -nodefconfig again. But this is completely orthogonal to the discussion if cpu models are configuration or not.

With this interface, -nodefconfig is necessary and useful. But if we consider that "configuration" = "machine description file", and a machine description file would never include the CPU models themselves, that would be a different approach.

-- Eduardo

-- Gleb.

On Thu, Mar 22, 2012 at 12:50:18PM -0300, Eduardo Habkost wrote:

On Thu, Mar 22, 2012 at 04:30:55PM +0200, Gleb Natapov wrote:

...

On Thu, Mar 22, 2012 at 10:31:21AM -0300, Eduardo Habkost wrote:

...

On Thu, Mar 22, 2012 at 11:32:44AM +0200, Gleb Natapov wrote:

...

What does this mean? Will -nodefconfig disable loading of bios.bin, option roms, keymaps?

Correcting myself: loading of _config_ files on /usr/share. ROM images are opaque data to be presented to the guest somehow, just like a disk image or kernel binary. But maybe keymaps will become "configuration" someday, I really don't know.

Where do you draw the line between "opaque data" and configuration. CPU models are also something that is present to a guest somehow.

Just the fact that it's in a structured key=value format that Qemu itself will interpret before exposing something to the guest. Yes, it's a bit arbitrary. If we could make everything "configuration data", we would (or that's what I think Anthony is pushing for--I hope he will reply and clarify that).

It is not a "bit arbitrary" it is completely arbitrary.

...

Are you consider ROMs to be "opaque data" because they are binary and CPU models to be config just because it is ascii file?

Not just "ascii file", but structured (and relatively small) [section]key=value data. If BIOSes were not opaque binary code and could be converted to a small set of config sections and key=values just like cpudefs, one could argue that BIOSes could become configuration data, too.

There is no argument I can make about it since there is no logic to refute to begin with :) Well may be except that when cpu model file will support configuring all couid leafs for each cpu model it will not be so small :)

...

What if we pre-process CPU models into binary for QEMU to read will it magically stop being configuration?

Doing the reverse (transforming simple [section]key=value data to a binary format) would just be silly and wouldn't gain us anything. The point here is that we (Qemu) seem to be moving towards a design where most things are structured "configuration data" that fits on a [section]key=value model, and Qemu just interprets that structured data to build a virtual machine.

Nothing silly about it. You can move data parsing outside of QEMU and just mmap cpu definitions in QEMU.

(That's why I said that perhaps keymaps could become configuration someday. Because maybe they can be converted to a key=value model relatively easily)

Such whole sale approach is harmful since it starts to affect design decisions. So now if it seams logical to move something outside the code one can decide against it just because it will become "configuration" due to that design.

...

...

...

...

...

Doing this would make it impossible to deploy fixes to users if we evern find out that the default configuration file had a serious bug. What if a bug in our default configuration file has a serious security implication?

The answer to this is: if the broken templates/defaults are on /usr/share, it would be easy to deploy the fix.

So, the compromise solution is:

- We can move some configuration data (especially defaults/templates) to /usr/share (machine-types and CPU models could go there). This way we can easily deploy fixes to the defaults, if necessary. - To reuse Qemu models, or machine-types, and not define everything from scratch, libvirt will have to use something like: "-nodefconfig -readconfig /usr/share/qemu/cpu-models-x86.conf"

cpu-models-x86.conf is not a configuration file. It is hardware description file. QEMU should not lose capability just because you run it with -nodefconfig. -nodefconfig means that QEMU does not create machine for you, but all parts needed to create a machine that would have been created without -nodefconfig are still present. Not been able to create Nehalem CPU after specifying -nodefconfig is the same as not been able to create virtio-net i.e the bug.

The current design direction Qemu seems to be following is different from that: hardware description is also considered "configuration" just like actual machine configuration. Anthony, please correct me if I am wrong. That's a bug. Why trying to rationalize it now instead of fixing it.

It's just a bug for you because you disagree with the current design. You can call it rationalization, yes; I am just accepting Anthony's proposal because it's equally good (to me) as what you are proposing.

...

It was fixed in RHEL by the same person who introduced it in upstream in the first place. He just forgot to send the fix upstream. Does bug that is present for a long time is promoted to a feature?

John didn't forget it, he knew that upstream could go in a different direction. The RHEL6 patch description has this:

"Note this is intended as an interim work-around for rhel6.0. While the new location of the config file should remain the same, the mechanism to direct qemu to it will likely differ going forward."

That's even worse. It is very sad state of affairs for upstream if he knew that the fix will not be accepted. As a result QEMU upstream does not work properly with libvirt for actually users.

...

...

...

...

(the item below is not something discussed on the call, just something I want to add)

To make this work better, we can allow users (humans or machines) to "extend" CPU models on the config file, instead of having to define everything from scratch. So, on /etc (or on a libvirt-generated config) we could have something like:

============= [cpu] base_cpudef = Nehalem add_features = "vmx" =============

Then, as long as /usr/share/cpu-models-x86.conf is loaded, the user will be able to reuse the Nehalem CPU model provided by Qemu.

And if it will not be loaded?

If it is not loaded, it is a configuration mistake. If you are reusing something defined somewhere, it would be your responsibility to make sure the file where the model is defined is present. On most cases you wouldn't use -nodefconfig and it would be shipped with Qemu and you shouldn't worry. If you used -nodefconfig, you load the CPU models file explicitly using -readconfig.

Regular user has no idea that we describe some cpu models in .c code and others in config file and those that are described in config files are disappear if -nodefconfig is used and that if they disappear they should be loaded back and how exactly they should be loaded back. The only user of -nodefconfig is libvirt and currently it does unexpected things for its only user.

And one could argue that this only user has wrong expectations and Qemu does it this way by design, and we would be running in circles forever. QEMU is new GNOME3?

The only problem for us is that if we keep running in circles, Qemu will keep the current design (or bug, if you want to call it), until Qemu maintainers agree to change the current behavior.

I mean: you're right into arguing for it, but somehow I feel I am the wrong person to reply to your arguments, because IMO both approaches are valid. I really hope Anthony will reply to your points, too, or that we discuss that on the next Qemu call. Because if you convince only me, we would gain nothing if the patches implementing the design we agreed with get rejected. It's really a pity that your arguments weren't exposed last week during the Qemu call, when I and Anthony discussed this and agreed on the "compromise solution" I described.

If I will convince you we can both convince Anthony afterwords :) You are keeping rationalizing the current behaviour, and since you are the only one who is doing it I have no other points to reply to.

...

...

...

...

...

> > But now when libvirt uses -nodefconfig, those models go away. > -nodefconfig means start QEMU in the most minimal state possible. > You get what you pay for if you use it. > > We'll have the same problem with machine configuration files. At > some point in time, -nodefconfig will make machine models disappear.

It shouldn't. Machine-types are defaults to be used as base, they are not user-provided configuration. And the fact that we decided to store some data outside of the Qemu binary is orthogonal the design decisions in the Qemu command-line and configuration interface.

So, this problem is solved if the defaults are easily found on /usr/share.

What problem is solved and why are we mixing machine configuration files and cpu configuration files? They are different and should be treated differently.

This is the root of the disagreement, it seems: they are not considered different today. Today cpudefs are on a config file inside /etc. One may argue that this was a mistake in the first place, but that's the design we have today.

Machine configuration files do not exist today! Machine configuration is done in .c code and -nodefconfig skips the code that does that, but it does not skip the code that describes cpu models and resides in .c. So as far as .c code is concerned machine creation and cpu description are two different things, but for some reason the file with rest of cpu model descriptions is not loaded. This is madness, not design.

Let's try to agree on terminology here: what's "machine configuration"? I understand it as the configuration of a specific virtual machine for a specific Qemu instance (i.e. what we put on the command-line today), I am not talking about machine-types (yet ;).

So, considering this definition, machine configuration exists today, there is stuff you can define on a configuration file, and you can give this configuration to -readconfig today. The config file is

My /etc/qemu/ contains no machine configuration file after fresh QEMU install. The only thing there is the thing that shouldn't be there: cpu modes definition. So everything I wrote above is correct. i.e -nodefconfig skips the _code_ that creates machine (not reading a config file) and it does not skip cpu model definitions that are done in .c file. I would gladly hear rationalization for this behaviour and you just ignored it in your reply.

unfortunately not as powerful as the command-line (some things can be set only on the command-line today), but it exists. Anthony even submitted a RFC series to make the config files as powerful as the command-line, by adding a [system] section. I don't agree completely with that specific implementation, but I agree with the direction it is taking.

...

...

...

-nodefconfig exists only because there is not machine configuration files currently. With machine configuration files libvirt does not need -nodefconfig because it can create its own machine file and make QEMU use it. So specifying machine file on QEMU's command line implies -nodefconfig. The option itself loses its meaning and can be dropped.

I think the approach today is:

There is not such thing as todays approach since there is not machine config file today.

There are config files today, and they work (but they are not as powerful as the command-line). But the interface is the one I describe below. Again, this is not how default machine is created, so hardly relevant to -nodefconfig behaviour, no?

...

...

- Qemu loads defaults from default config files; - Machine description files would be given using -readconfig and they would _augment_ the defaults from the default config files.

Jut make it possible to include one machine config file from another and you do not need -nodefconfig again.

That would be a valid approach too, but this is just not how the config files work today. We may change it, yes. But today the config file system is based on getting config data from multiple places (/etc/qemu/qemu.conf, extend it with /etc/qemu/target-x86_64.conf, and extend the result with the command-line arguments), instead of using includes.

Nothing wrong with that! But cpu models should not be part of /etc/qemu/target-x86_64.conf :)

...

But this is completely orthogonal to the discussion if cpu models are configuration or not.

You are right. How the "machine configuration" is going to be loaded is orthogonal to the main question. The main question here is wheter cpudefs should be included on what we call "machine configuration" or not. (agreed?)

-- Eduardo

-- Gleb.

On 03/22/2012 04:32 AM, Gleb Natapov wrote:

On Tue, Mar 13, 2012 at 11:53:19AM -0300, Eduardo Habkost wrote:

...

So, trying to summarize what was discussed in the call:

On Mon, Mar 12, 2012 at 10:08:10AM -0300, Eduardo Habkost wrote:

...

...

Let's say we moved CPU definitions to /usr/share/qemu/cpu-models.xml.

Obviously, we'd want a command line option to be able to change that location so we'd introduce -cpu-models PATH.

But we want all of our command line options to be settable by the global configuration file so we would have a cpu-model=PATH to the configuration file.

But why hard code a path when we can just set the default path in the configuration file so let's avoid hard coding and just put cpu-models=/usr/share/qemu/cpu-models.xml in the default configuration file.

We wouldn't do the above.

-nodefconfig should disable the loading of files on /etc, but it shouldn't disable loading internal non-configurable data that we just happened to choose to store outside the qemu binary because it makes development easier.

The statement above is the one not fulfilled by the compromise solution: -nodefconfig would really disable the loading of files on /usr/share.

What does this mean? Will -nodefconfig disable loading of bios.bin, option roms, keymaps?

...

...

Really, the requirement of a "default configuration file" is a problem by itself. Qemu should not require a default configuration file to work, and it shouldn't require users to copy the default configuration file to change options from the default.

The statement above is only partly true. The default configuration file would be still needed, but if defaults are stored on /usr/share, I will be happy with it.

My main problem was with the need to _copy_ or edit a non-trivial default config file. If the not-often-edited defaults/templates are easily found on /usr/share to be used with -readconfig, I will be happy with this solution, even if -nodefconfig disable the files on /usr/share.

...

Doing this would make it impossible to deploy fixes to users if we evern find out that the default configuration file had a serious bug. What if a bug in our default configuration file has a serious security implication?

The answer to this is: if the broken templates/defaults are on /usr/share, it would be easy to deploy the fix.

So, the compromise solution is:

- We can move some configuration data (especially defaults/templates) to /usr/share (machine-types and CPU models could go there). This way we can easily deploy fixes to the defaults, if necessary. - To reuse Qemu models, or machine-types, and not define everything from scratch, libvirt will have to use something like: "-nodefconfig -readconfig /usr/share/qemu/cpu-models-x86.conf"

cpu-models-x86.conf is not a configuration file. It is hardware description file. QEMU should not lose capability just because you run it with -nodefconfig. -nodefconfig means that QEMU does not create machine for you, but all parts needed to create a machine that would have been created without -nodefconfig are still present. Not been able to create Nehalem CPU after specifying -nodefconfig is the same as not been able to create virtio-net i.e the bug.

...

(the item below is not something discussed on the call, just something I want to add)

To make this work better, we can allow users (humans or machines) to "extend" CPU models on the config file, instead of having to define everything from scratch. So, on /etc (or on a libvirt-generated config) we could have something like:

============= [cpu] base_cpudef = Nehalem add_features = "vmx" =============

Then, as long as /usr/share/cpu-models-x86.conf is loaded, the user will be able to reuse the Nehalem CPU model provided by Qemu.

And if it will not be loaded?

...

...

...

But now when libvirt uses -nodefconfig, those models go away. -nodefconfig means start QEMU in the most minimal state possible. You get what you pay for if you use it.

We'll have the same problem with machine configuration files. At some point in time, -nodefconfig will make machine models disappear.

It shouldn't. Machine-types are defaults to be used as base, they are not user-provided configuration. And the fact that we decided to store some data outside of the Qemu binary is orthogonal the design decisions in the Qemu command-line and configuration interface.

So, this problem is solved if the defaults are easily found on /usr/share.

What problem is solved and why are we mixing machine configuration files and cpu configuration files? They are different and should be treated differently. -nodefconfig exists only because there is not machine configuration files currently. With machine configuration files libvirt does not need -nodefconfig because it can create its own machine file and make QEMU use it. So specifying machine file on QEMU's command line implies -nodefconfig. The option itself loses its meaning and can be dropped.

No, -nodefconfig means "no default config". As with many projects, we can have *some* configuration required. The default configure should have a: [system] readconfig=@SYSCONFDIR@/cpu-models-x86_64.cfg Stanza by default. If libvirt wants to reuse this, they can use -readconfig if they use -nodefconfig. Regards, Anthony Liguori

...

We still have the backwards compatibility problem for pc-1.0, pc-1.1, and so on. But that can be discussed later, when we actually move machine-types to somewhere outside .c files.

...

As I said previously, requiring generation of opaque config files (and "copy the default config file and change it" is included on my definition of "generation of opaque config files") is poor design, IMO. I bet this even has an entry in some design anti-pattern catalog somewhere.

This problem is also solved if the defaults are deployed on /usr/share and just reused/included by the config files on /etc.

-- Gleb.

On 03/22/2012 12:14 PM, Eduardo Habkost wrote:

On Thu, Mar 22, 2012 at 11:37:39AM -0500, Anthony Liguori wrote:

...

On 03/22/2012 04:32 AM, Gleb Natapov wrote:

...

On Tue, Mar 13, 2012 at 11:53:19AM -0300, Eduardo Habkost wrote:

...

So, this problem is solved if the defaults are easily found on /usr/share.

What problem is solved and why are we mixing machine configuration files and cpu configuration files? They are different and should be treated differently. -nodefconfig exists only because there is not machine configuration files currently. With machine configuration files libvirt does not need -nodefconfig because it can create its own machine file and make QEMU use it. So specifying machine file on QEMU's command line implies -nodefconfig. The option itself loses its meaning and can be dropped.

No, -nodefconfig means "no default config".

As with many projects, we can have *some* configuration required.

The default configure should have a:

[system] readconfig=@SYSCONFDIR@/cpu-models-x86_64.cfg

Not @SYSCONFDIR@, but @DATADIR@. CPU models belong to /usr/share because they aren't meant to be changed by the user (I think I already explained why: because we have to be able to deploy fixes to them).

...

Stanza by default. If libvirt wants to reuse this, they can use -readconfig if they use -nodefconfig.

You are just repeating how you believe it should work based on the premise that "cpudefs are configuration". We're discussing/questioning this exact premise, here, and I would really appreciate to hear why the model Gleb is proposing is not valid.

More precisely, this part:

...

...

cpu-models-x86.conf is not a configuration file. It is hardware description file. QEMU should not lose capability just because you run it with -nodefconfig. -nodefconfig means that QEMU does not create machine for you, but all parts needed to create a machine that would have been created without -nodefconfig are still present. Not been able to create Nehalem CPU after specifying -nodefconfig is the same as not been able to create virtio-net i.e the bug.

And the related points Gleb mentioned further in this thread.

Because the next patch series that would follow would be a -cpu-defs-path that would be a one-off hack with a global variable and a -no-cpu-defs. So let's avoid that and start by having a positive configuration mechanism that the user can use to change the path and exclude it. My suggestion eliminate the need for two future command line options. Regards, Anthony Liguori

On 03/25/2012 04:49 AM, Gleb Natapov wrote:

On Thu, Mar 22, 2012 at 03:01:17PM -0500, Anthony Liguori wrote:

...

So let's avoid that and start by having a positive configuration mechanism that the user can use to change the path and exclude it. My suggestion eliminate the need for two future command line options.

If cpu models are not part of configuration they should not be affected by configuration mechanism. You are just avoiding addressing the real question that if asked above.

I think you're just refusing to listen. The stated direction of QEMU, for literally years now, is that we want to arrive at the following: QEMU is composed of a series of objects who's relationships can be fully described by an external configuration file. Much of the current baked in concepts (like machines) would then become configuration files. qemu -M pc Would effectively be short hand for -readconfig /usr/share/qemu/machines/pc.cfg There are some valid points that were raised in this thread, namely that the user needs to have a file that acts as strictly config (stored in /etc). I'm totally happy moving the CPU configuration to /usr/share in order to address this. I think the thread has reduced to: should /usr/share configuration files be read by default or just treated as additional configuration files. It seems pretty obvious to me that they should be treated as normal configuration files. This gives you the user the ability to have fine grain control over which files are used including the ability to change the location for each file. Maybe RHEL only wants to expose supported CPUs and supported machines, wouldn't it be better to not have to patch QEMU to do that? Wouldn't it be even better if you could drop in a separate CPU configuration file with the supported CPU types and then change the default /etc config to use that instead? Regards, Anthony Liguori

On 03/25/2012 08:08 AM, Avi Kivity wrote:

On 03/25/2012 02:55 PM, Anthony Liguori wrote:

...

...

If cpu models are not part of configuration they should not be affected by configuration mechanism. You are just avoiding addressing the real question that if asked above.

I think you're just refusing to listen.

The stated direction of QEMU, for literally years now, is that we want to arrive at the following:

QEMU is composed of a series of objects who's relationships can be fully described by an external configuration file. Much of the current baked in concepts (like machines) would then become configuration files.

qemu -M pc

Would effectively be short hand for -readconfig /usr/share/qemu/machines/pc.cfg

In that case

qemu -cpu westmere

is shorthand for -readconfig /usr/share/qemu/cpus/westmere.cfg.

This is not a bad suggestion, although it would make -cpu ? a bit awkward. Do you see an advantage to this over having /usr/share/qemu/target-x86_64-cpus.cfg that's read early on?

...

I think the thread has reduced to: should /usr/share configuration files be read by default or just treated as additional configuration files.

If they're read as soon as they're referenced, what's the difference?

I suspect libvirt would not be happy with reading configuration files on demand.. Regards, Anthony Liguori

On 03/25/2012 08:14 AM, Avi Kivity wrote:

On 03/25/2012 03:12 PM, Anthony Liguori wrote:

...

...

...

qemu -M pc

Would effectively be short hand for -readconfig /usr/share/qemu/machines/pc.cfg

In that case

qemu -cpu westmere

is shorthand for -readconfig /usr/share/qemu/cpus/westmere.cfg.

This is not a bad suggestion, although it would make -cpu ? a bit awkward. Do you see an advantage to this over having /usr/share/qemu/target-x86_64-cpus.cfg that's read early on?

Nope. As long as qemu -nodefconfig -cpu westmere works, I'm happy.

Why? What's wrong with: qemu -nodefconfig -readconfig /usr/share/qemu/cpus/target-x86_64-cpus.cfg \ -cpu westmere And if that's not okay, would: qemu -nodefconfig -nocpudefconfig -cpu Westmere Not working be a problem?

The reasoning is, loading target-x86_64-cpus.cfg does not alter the current instance's configuration, so reading it doesn't violate -nodefconfig.

I think we have a different view of what -nodefconfig does. We have a couple options today: -nodefconfig Don't read the default configuration files. By default, we read /etc/qemu/qemu.cfg and /etc/qemu/target-$(ARCH).cfg -nodefaults Don't create default devices. -vga none Don't create the default VGA device (not covered by -nodefaults). With these two options, the semantics you get an absolutely minimalistic instance of QEMU. Tools like libguestfs really want to create the simplest guest and do the least amount of processing so the guest runs as fast as possible. It does suck a lot that this isn't a single option. I would much prefer -nodefaults to be implied by -nodefconfig. Likewise, I would prefer that -nodefaults implied -vga none.

...

...

...

files be read by default or just treated as additional configuration files.

If they're read as soon as they're referenced, what's the difference? I think the thread has reduced to: should /usr/share configuration

I suspect libvirt would not be happy with reading configuration files on demand..

Why not?

It implies a bunch of SELinux labeling to make sVirt work. libvirt tries very hard to avoid having QEMU read *any* files at all when it starts up. Regards, Anthony Liguori

On 03/25/2012 08:34 AM, Avi Kivity wrote:

On 03/25/2012 03:22 PM, Anthony Liguori wrote:

...

...

...

...

In that case

qemu -cpu westmere

is shorthand for -readconfig /usr/share/qemu/cpus/westmere.cfg.

This is not a bad suggestion, although it would make -cpu ? a bit awkward. Do you see an advantage to this over having /usr/share/qemu/target-x86_64-cpus.cfg that's read early on?

Nope. As long as qemu -nodefconfig -cpu westmere works, I'm happy.

Why? What's wrong with:

qemu -nodefconfig -readconfig /usr/share/qemu/cpus/target-x86_64-cpus.cfg \ -cpu westmere

And if that's not okay, would:

qemu -nodefconfig -nocpudefconfig -cpu Westmere

Not working be a problem?

Apart from the command line length, it confuses configuration with definition.

There is no distinction with what we have today. Our configuration file basically corresponds to command line options and as there is no distinction in command line options, there's no distinction in the configuration format.

target-x86_64-cpus.cfg does not configure qemu for anything, it's merely the equivalent of

#define westmere (x86_def_t) { ... } #define nehalem (x86_def_t) { ... } #define bulldozer (x86_def_t) { ... } // for PC

so it should be read at each invocation. On the other hand, pc.cfg and westmere.cfg (as used previously) are shorthand for

machine = (QEMUMachine) { ... }; cpu = (x86_def_t) { ... };

so they should only be read if requested explicitly (or indirectly).

This doesn't make a lot of sense to me. Here's what I'm proposing: 1) QEMU would have a target-x86_64-cpu.cfg.in that is installed by default in /etc/qemu. It would contain: [system] # Load default CPU definitions readconfig = @DATADIR@/target-x86_64-cpus.cfg 2) target-x86_64-cpus.cfg would be installed to @DATADIR@ and would contain: [cpudef] name = "Westmere" ... This has the following properties: A) QEMU has no builtin notion of CPU definitions. It just has a "cpu factory". -cpudef will create a new class called Westmere that can then be enumerated through qom-type-list and created via qom-create. B) A management tool has complete control over cpu definitions without modifying the underlying filesystem. -nodefconfig will prevent it from loading and the management tool can explicitly load the QEMU definition (via -readconfig, potentially using a /dev/fd/N path) or it can define it's own cpu definitions. C) This model maps to any other type of class factory. Machines will eventually be expressed as a class factory. When we implement this, we would change the default target-x86_64-cpu.cfg to: [system] # Load default CPU definitions readconfig = @DATADIR@/target-x86_64-cpus.cfg # Load default machines readconfig = @DATADIR@/target-x86_64-machines.cfg A machine definition would look like: [machinedef] name = pc-0.15 virtio-blk.class_code = 32 ... Loading a file based on -cpu doesn't generalize well unless we try to load a definition for any possible QOM type to find the class factory for it. I don't think this is a good idea.

...

...

The reasoning is, loading target-x86_64-cpus.cfg does not alter the current instance's configuration, so reading it doesn't violate -nodefconfig.

I think we have a different view of what -nodefconfig does.

We have a couple options today:

-nodefconfig

Don't read the default configuration files. By default, we read /etc/qemu/qemu.cfg and /etc/qemu/target-$(ARCH).cfg

The latter seems meaningless to avoid reading. It's just a set of #defines, what do you get by not reading it?

In my target-$(ARCH).cfg, I have: [machine] enable-kvm = "on" Which means I don't have to use -enable-kvm anymore. But if you look at a tool like libguestfs, start up time is the most important thing so avoiding unnecessary I/O and processing is critical.

...

-nodefaults

Don't create default devices.

-vga none

Don't create the default VGA device (not covered by -nodefaults).

With these two options, the semantics you get an absolutely minimalistic instance of QEMU. Tools like libguestfs really want to create the simplest guest and do the least amount of processing so the guest runs as fast as possible.

It does suck a lot that this isn't a single option. I would much prefer -nodefaults to be implied by -nodefconfig. Likewise, I would prefer that -nodefaults implied -vga none.

I don't have a qemu.cfg so can't comment on it, but in what way does reading target-x86_64.cfg affect the current instance (that is, why is -nodefconfig needed over -nodefaults -vga look-at-the-previous-option?)

It depends on what the user configures it to do. Regards, Anthony Liguori

On 03/25/2012 09:46 AM, Avi Kivity wrote:

On 03/25/2012 04:36 PM, Anthony Liguori wrote:

...

...

Apart from the command line length, it confuses configuration with definition.

There is no distinction with what we have today. Our configuration file basically corresponds to command line options and as there is no distinction in command line options, there's no distinction in the configuration format.

We don't have command line options for defining, only configuring.

That's an oversight. There should be a -cpudef option. It's a QemuOptsList.

Again, defining = #define

I think -global fits your definition of #define...

Configuring = modifying current instance

...

...

target-x86_64-cpus.cfg does not configure qemu for anything, it's merely the equivalent of

#define westmere (x86_def_t) { ... } #define nehalem (x86_def_t) { ... } #define bulldozer (x86_def_t) { ... } // for PC

so it should be read at each invocation. On the other hand, pc.cfg and westmere.cfg (as used previously) are shorthand for

machine = (QEMUMachine) { ... }; cpu = (x86_def_t) { ... };

so they should only be read if requested explicitly (or indirectly).

This doesn't make a lot of sense to me. Here's what I'm proposing:

1) QEMU would have a target-x86_64-cpu.cfg.in that is installed by default in /etc/qemu. It would contain:

[system] # Load default CPU definitions readconfig = @DATADIR@/target-x86_64-cpus.cfg

2) target-x86_64-cpus.cfg would be installed to @DATADIR@ and would contain:

[cpudef] name = "Westmere" ...

This has the following properties:

A) QEMU has no builtin notion of CPU definitions. It just has a "cpu factory". -cpudef will create a new class called Westmere that can then be enumerated through qom-type-list and created via qom-create.

B) A management tool has complete control over cpu definitions without modifying the underlying filesystem. -nodefconfig will prevent it from loading and the management tool can explicitly load the QEMU definition (via -readconfig, potentially using a /dev/fd/N path) or it can define it's own cpu definitions.

Why does -nodefconfig affect anything?

Because -nodefconfig means "don't load *any* default configuration files".

The file defines westmere as an alias for a grab bag of options. Whether it's loaded or not is immaterial, unless someone uses one of the names within.

But you would agree, a management tool should be able to control whether class factories get loaded, right? So what's the mechanism to do this?

...

C) This model maps to any other type of class factory. Machines will eventually be expressed as a class factory. When we implement this, we would change the default target-x86_64-cpu.cfg to:

[system] # Load default CPU definitions readconfig = @DATADIR@/target-x86_64-cpus.cfg # Load default machines readconfig = @DATADIR@/target-x86_64-machines.cfg

A machine definition would look like:

[machinedef] name = pc-0.15 virtio-blk.class_code = 32 ...

Loading a file based on -cpu doesn't generalize well unless we try to load a definition for any possible QOM type to find the class factory for it. I don't think this is a good idea.

Why not load all class factories? Just don't instantiate any objects.

Unless we have two different config syntaxes, I think it will lead to a lot of confusion. Having some parts of a config file be parsed and others not is fairly strange.

Otherwise, the meaning of -nodefconfig changes as more stuff is moved out of .c and into .cfg.

What's the problem with this?

...

...

...

...

The reasoning is, loading target-x86_64-cpus.cfg does not alter the current instance's configuration, so reading it doesn't violate -nodefconfig.

I think we have a different view of what -nodefconfig does.

We have a couple options today:

-nodefconfig

Don't read the default configuration files. By default, we read /etc/qemu/qemu.cfg and /etc/qemu/target-$(ARCH).cfg

The latter seems meaningless to avoid reading. It's just a set of #defines, what do you get by not reading it?

In my target-$(ARCH).cfg, I have:

[machine] enable-kvm = "on"

Which means I don't have to use -enable-kvm anymore. But if you look at a tool like libguestfs, start up time is the most important thing so avoiding unnecessary I/O and processing is critical.

So this is definitely configuration (applies to the current instance) as opposed to target-x86_64.cfg, which doesn't.

I'm not sure which part you're responding to..

...

...

...

-nodefaults

Don't create default devices.

-vga none

Don't create the default VGA device (not covered by -nodefaults).

With these two options, the semantics you get an absolutely minimalistic instance of QEMU. Tools like libguestfs really want to create the simplest guest and do the least amount of processing so the guest runs as fast as possible.

It does suck a lot that this isn't a single option. I would much prefer -nodefaults to be implied by -nodefconfig. Likewise, I would prefer that -nodefaults implied -vga none.

I don't have a qemu.cfg so can't comment on it, but in what way does reading target-x86_64.cfg affect the current instance (that is, why is -nodefconfig needed over -nodefaults -vga look-at-the-previous-option?)

It depends on what the user configures it to do.

How?

As far as I can tell, the only difference is that -nodefconfig -cpu westmere will error out instead of working. But if you don't supply -cpu westmere, the configuration is identical.

What configuration? Let me ask, what do you think the semantics of -nodefconfig should be? I'm not sure I understand what you're advocating for. Regards, Anthony Liguori

On 03/25/2012 10:16 AM, Avi Kivity wrote:

On 03/25/2012 04:59 PM, Anthony Liguori wrote:

...

On 03/25/2012 09:46 AM, Avi Kivity wrote:

...

On 03/25/2012 04:36 PM, Anthony Liguori wrote:

...

...

Apart from the command line length, it confuses configuration with definition.

There is no distinction with what we have today. Our configuration file basically corresponds to command line options and as there is no distinction in command line options, there's no distinction in the configuration format.

We don't have command line options for defining, only configuring.

That's an oversight. There should be a -cpudef option. It's a QemuOptsList.

...

Again, defining = #define

I think -global fits your definition of #define...

Yes (apart from the corner case of modifying a default-instantiated device).

...

...

...

B) A management tool has complete control over cpu definitions without modifying the underlying filesystem. -nodefconfig will prevent it from loading and the management tool can explicitly load the QEMU definition (via -readconfig, potentially using a /dev/fd/N path) or it can define it's own cpu definitions.

Why does -nodefconfig affect anything?

Because -nodefconfig means "don't load *any* default configuration files".

Put the emphasis around *configuration*.

So how about: 1) Load ['@SYSCONFDIR@/qemu/qemu.cfg', '@SYSCONFDIR@/qemu/target-@ARCH@.cfg', '@DATADIR@/system.cfg', '@DATADIR@/system-@ARCH@.cfg'] 2) system-@ARCH@.cfg will contain: [system] readconfig=@DATADIR@/target-@ARCH@-cpus.cfg readconfig=@DATADIR@/target-@ARCH@-machine.cfg 3) -nodefconfig will not load any configuration files from DATADIR or SYSCONFDIR. -no-user-config will not load any configuration files from SYSCONFDIR.

"#define westmere blah" is not configuration, otherwise the meaning of configuration will drift over time.

-cpu blah is, of course.

It's the same mechanism, but the above would create two classes of default configuration files and then it becomes a question of how they're used.

...

...

The file defines westmere as an alias for a grab bag of options. Whether it's loaded or not is immaterial, unless someone uses one of the names within.

But you would agree, a management tool should be able to control whether class factories get loaded, right?

No, why? But perhaps I don't entirely get what you mean by "class factories".

Aren't they just implementations of

virtual Device *new_instance(...) = 0?

if so, why not load them?

No, a class factory creates a new type of class. -cpudef will ultimately call type_register() to create a new QOM visible type. From a management tools perspective, the type is no different than a built-in type.

...

...

Otherwise, the meaning of -nodefconfig changes as more stuff is moved out of .c and into .cfg.

What's the problem with this?

The command line becomes unstable if you use -nodefconfig.

-no-user-config solves this but I fully expect libvirt would continue to use -nodefconfig.

...

...

...

In my target-$(ARCH).cfg, I have:

[machine] enable-kvm = "on"

Which means I don't have to use -enable-kvm anymore. But if you look at a tool like libguestfs, start up time is the most important thing so avoiding unnecessary I/O and processing is critical.

So this is definitely configuration (applies to the current instance) as opposed to target-x86_64.cfg, which doesn't.

I'm not sure which part you're responding to..

I was saying that target-x86_64.cfg appears to be definitions, not configuration, and was asking about qemu.cfg (which is configuration).

...

...

As far as I can tell, the only difference is that -nodefconfig -cpu westmere will error out instead of working. But if you don't supply -cpu westmere, the configuration is identical.

What configuration?

Let me ask, what do you think the semantics of -nodefconfig should be? I'm not sure I understand what you're advocating for.

-nodefconfig = create an empty machine, don't assume anything (=don't read qemu.cfg) let me build it out of all those lego bricks. Those can be defined in code or in definition files in /usr/share, I don't care.

Maybe that's -nodevices -vga none. But in this case I don't see the point in -nodefconfig. Not loading target_x86-64.cfg doesn't buy the user anything, since it wouldn't affect the guest in any way.

-nodefconfig doesn't mean what you think it means. -nodefconfig doesn't say anything about the user visible machine. -nodefconfig tells QEMU not to read any configuration files at start up. This has an undefined affect on the user visible machine that depends on the specific version of QEMU. Regards, Anthony Liguori

On 03/25/2012 10:40 AM, Avi Kivity wrote:

On 03/25/2012 05:26 PM, Anthony Liguori wrote:

...

...

Put the emphasis around *configuration*.

So how about:

1) Load ['@SYSCONFDIR@/qemu/qemu.cfg', '@SYSCONFDIR@/qemu/target-@ARCH@.cfg', '@DATADIR@/system.cfg', '@DATADIR@/system-@ARCH@.cfg']

2) system-@ARCH@.cfg will contain:

[system] readconfig=@DATADIR@/target-@ARCH@-cpus.cfg readconfig=@DATADIR@/target-@ARCH@-machine.cfg

3) -nodefconfig will not load any configuration files from DATADIR or SYSCONFDIR. -no-user-config will not load any configuration files from SYSCONFDIR.

What, more options?

Okay, we can just drop -no-user-config and then if a management tool wants to do the equivalent, they can do -nodefconfig + '-readconfig @DATADIR@/system-@ARCH@.cfg'. I'm equally happy with that :-)

I don't think -nodefconfig (as defined) is usable, since there is no way for the user to tell what it means short of reading those files.

*if the user doesn't know specifics about this QEMU version. You make the assumption that all users are going to throw arbitrary options at arbitrary QEMU versions. That's certainly an important use-case but it's not the only one.

-no-user-config is usable, I think it needs also to mean that qemu without -M/-cpu/-m options will error out?

You're confusing -nodefaults (or something stronger than -nodefaults) with -no-user-config. Yes, the distinctions are confusing. It's not all fixable tomorrow. If we take my config refactoring series, we can get 90% of the way there soon but Paolo has a more thorough refactoring..

...

...

"#define westmere blah" is not configuration, otherwise the meaning of configuration will drift over time.

-cpu blah is, of course.

It's the same mechanism, but the above would create two classes of default configuration files and then it becomes a question of how they're used.

Confused.

We don't have a formal concept of -read-definition-config and -read-configuration-config There's no easy or obvious way to create such a concept either nor do I think the distinction is meaningful to users.

...

...

...

...

The file defines westmere as an alias for a grab bag of options. Whether it's loaded or not is immaterial, unless someone uses one of the names within.

But you would agree, a management tool should be able to control whether class factories get loaded, right?

No, why? But perhaps I don't entirely get what you mean by "class factories".

Aren't they just implementations of

virtual Device *new_instance(...) = 0?

if so, why not load them?

No, a class factory creates a new type of class. -cpudef will ultimately call type_register() to create a new QOM visible type. From a management tools perspective, the type is no different than a built-in type.

Exactly. The types are no different, so there's no reason to discriminate against types that happen to live in qemu-provided data files vs. qemu code. They aren't instantiated, so we lose nothing by creating the factories (just so long as the factories aren't mass-producing objects).

At some point, I'd like to have type modules that are shared objects. I'd like QEMU to start with almost no builtin types and allow the user to configure which modules get loaded. In the long term, I'd like QEMU to be a small, robust core with the vast majority of code relegated to modules with the user ultimately in control of module loading. Yes, I'd want some module autoloading system but there should always be a way to launch QEMU without loading any modules and then load a very specific set of modules (as defined by the user). You can imagine this being useful for something like Common Criteria certifications.

...

...

...

...

Otherwise, the meaning of -nodefconfig changes as more stuff is moved out of .c and into .cfg.

What's the problem with this?

The command line becomes unstable if you use -nodefconfig.

-no-user-config solves this but I fully expect libvirt would continue to use -nodefconfig.

I don't see how libvirt can use -nodefconfig with the fluid meaning you attach to it, or what it gains from it.

...

...

-nodefconfig = create an empty machine, don't assume anything (=don't read qemu.cfg) let me build it out of all those lego bricks. Those can be defined in code or in definition files in /usr/share, I don't care.

Maybe that's -nodevices -vga none. But in this case I don't see the point in -nodefconfig. Not loading target_x86-64.cfg doesn't buy the user anything, since it wouldn't affect the guest in any way.

-nodefconfig doesn't mean what you think it means. -nodefconfig doesn't say anything about the user visible machine.

-nodefconfig tells QEMU not to read any configuration files at start up. This has an undefined affect on the user visible machine that depends on the specific version of QEMU.

Then it's broken. How can anyone use something that has an undefined effect?

It's obviously defined for a given release, just not defined long term.

If I see something like -nodefconfig, I assume it will create a bare bones guest that will not depend on any qemu defaults and will be stable across releases.

That's not even close to what -nodefconfig is. That's pretty much what -nodefaults is but -nodefaults has also had a fluid definition historically. Regards, Anthony Liguori

I don't think anyone will understand -nodefconfig to be something version dependent without reading the qemu management tool author's guide.

On Mon, Mar 26, 2012 at 11:08:16AM +0200, Avi Kivity wrote:

...

...

Exactly. The types are no different, so there's no reason to discriminate against types that happen to live in qemu-provided data files vs. qemu code. They aren't instantiated, so we lose nothing by creating the factories (just so long as the factories aren't mass-producing objects).

At some point, I'd like to have type modules that are shared objects. I'd like QEMU to start with almost no builtin types and allow the user to configure which modules get loaded.

In the long term, I'd like QEMU to be a small, robust core with the vast majority of code relegated to modules with the user ultimately in control of module loading.

Yes, I'd want some module autoloading system but there should always be a way to launch QEMU without loading any modules and then load a very specific set of modules (as defined by the user).

You can imagine this being useful for something like Common Criteria certifications.

Okay.

Modularised minimal QEMU may be a good thing, but how -nodefconfig helps here? Won't you have the same effect if QEMU will load modules on demand, only when they are actually needed (regardless of -nodefconfig). i.e virtio-blk is loaded only if there is -device virtio-blk somewhere in configuration.

...

It's obviously defined for a given release, just not defined long term.

...

If I see something like -nodefconfig, I assume it will create a bare bones guest that will not depend on any qemu defaults and will be stable across releases.

That's not even close to what -nodefconfig is. That's pretty much what -nodefaults is but -nodefaults has also had a fluid definition historically.

Okay. Let's just make sure to document -nodefconfig as version specific and -nodefaults as the stable way to create a bare bones guest (and define exactly what that means).

What is the reason libvirt uses -nodefconfig instead of -nodefaults now? What the former does for them that the later doesn't? -- Gleb.

On 03/26/2012 09:03 PM, Anthony Liguori wrote:

I think what we want to move toward is a -no-machine option which allows a user to explicitly build a machine from scratch. That is:

qemu -no-machine -device i440fx,id=host -device isa-serial,chr=chr0 ...

I'd call it -M bare-1.1, so that it can be used to override driver properties in 1.2+. So we'd have # default machine for this version qemu / qemu -M pc # an older version's pc qemu -M pc-1.1 # just a chassis, bring your own screwdriver qemu -M bare # previous generation chassis, beige qemu -M bare-1.1 That is because -M not only specifies the components that go into the machine, it also alters other devices you add to it. This also helps preserve the planet's dwindling supply of command line options. -- error compiling committee.c: too many arguments to function

On Sun, Mar 25, 2012 at 10:26:57 -0500, Anthony Liguori wrote:

On 03/25/2012 10:16 AM, Avi Kivity wrote:

...

On 03/25/2012 04:59 PM, Anthony Liguori wrote: So how about:

1) Load ['@SYSCONFDIR@/qemu/qemu.cfg', '@SYSCONFDIR@/qemu/target-@ARCH@.cfg', '@DATADIR@/system.cfg', '@DATADIR@/system-@ARCH@.cfg']

2) system-@ARCH@.cfg will contain:

[system] readconfig=@DATADIR@/target-@ARCH@-cpus.cfg readconfig=@DATADIR@/target-@ARCH@-machine.cfg

3) -nodefconfig will not load any configuration files from DATADIR or SYSCONFDIR. -no-user-config will not load any configuration files from SYSCONFDIR.

...

...

The command line becomes unstable if you use -nodefconfig.

-no-user-config solves this but I fully expect libvirt would continue to use -nodefconfig.

Libvirt uses -nodefaults -nodefconfig because it wants to fully control how the virtual machine will look like (mainly in terms of devices). In other words, we don't want any devices to just magically appear without libvirt knowing about them. -nodefaults gets rid of default devices that are built directly in qemu. Since users can set any devices or command line options (such as enable-kvm) into qemu configuration files in @SYSCONFDIR@, we need to avoid reading those files as well. Hence we use -nodefconfig. However, we would still like qemu to read CPU definitions, machine types, etc. once they become externally loaded configuration (or however we decide to call it). That said, when CPU definitions are moved into @DATADIR@, and -no-user-config is introduced, I don't see any reason for libvirt to keep using -nodefconfig. I actually like -no-user-config more than -nodefconfig -readconfig @DATADIR@/... since it would avoid additional magic to detect what files libvirt should explicitly pass to -readconfig but basically any approach that would allow us to do read files only from @DATADIR@ is much better than what we have with -nodefconfig now. Jirka

On Mon, Mar 26, 2012 at 01:59:05PM +0200, Avi Kivity wrote:

On 03/26/2012 01:24 PM, Jiri Denemark wrote:

...

...

...

...

The command line becomes unstable if you use -nodefconfig.

-no-user-config solves this but I fully expect libvirt would continue to use -nodefconfig.

Libvirt uses -nodefaults -nodefconfig because it wants to fully control how the virtual machine will look like (mainly in terms of devices). In other words, we don't want any devices to just magically appear without libvirt knowing about them. -nodefaults gets rid of default devices that are built directly in qemu. Since users can set any devices or command line options (such as enable-kvm) into qemu configuration files in @SYSCONFDIR@, we need to avoid reading those files as well. Hence we use -nodefconfig. However, we would still like qemu to read CPU definitions, machine types, etc. once they become externally loaded configuration (or however we decide to call it). That said, when CPU definitions are moved into @DATADIR@, and -no-user-config is introduced, I don't see any reason for libvirt to keep using -nodefconfig.

I actually like -no-user-config more than -nodefconfig -readconfig @DATADIR@/... since it would avoid additional magic to detect what files libvirt should explicitly pass to -readconfig but basically any approach that would allow us to do read files only from @DATADIR@ is much better than what we have with -nodefconfig now.

That's how I see it as well.

+1 except that instead of -no-user-config we can do what most other programs do. If config file is specified during invocation default one is not used. After implementing -no-user-config (or similar) we can drop -nodefconfig entirely since its only user will be gone it its semantics is not clear. -- Gleb.

On 03/26/2012 11:14 AM, Eduardo Habkost wrote:

On Mon, Mar 26, 2012 at 02:03:21PM +0200, Gleb Natapov wrote:

...

On Mon, Mar 26, 2012 at 01:59:05PM +0200, Avi Kivity wrote:

...

On 03/26/2012 01:24 PM, Jiri Denemark wrote:

...

...

...

...

The command line becomes unstable if you use -nodefconfig.

-no-user-config solves this but I fully expect libvirt would continue to use -nodefconfig.

Libvirt uses -nodefaults -nodefconfig because it wants to fully control how the virtual machine will look like (mainly in terms of devices). In other words, we don't want any devices to just magically appear without libvirt knowing about them. -nodefaults gets rid of default devices that are built directly in qemu. Since users can set any devices or command line options (such as enable-kvm) into qemu configuration files in @SYSCONFDIR@, we need to avoid reading those files as well. Hence we use -nodefconfig. However, we would still like qemu to read CPU definitions, machine types, etc. once they become externally loaded configuration (or however we decide to call it). That said, when CPU definitions are moved into @DATADIR@, and -no-user-config is introduced, I don't see any reason for libvirt to keep using -nodefconfig.

ACK.

...

...

...

I actually like -no-user-config more than -nodefconfig -readconfig @DATADIR@/... since it would avoid additional magic to detect what files libvirt should explicitly pass to -readconfig but basically any approach that would allow us to do read files only from @DATADIR@ is much better than what we have with -nodefconfig now.

That's how I see it as well.

+1

except that instead of -no-user-config we can do what most other programs do. If config file is specified during invocation default one is not used. After implementing -no-user-config (or similar) we can drop -nodefconfig entirely since its only user will be gone it its semantics is not clear.

Awesome. It looks like we have a solution now? Anthony, do you agree with that? Daniel, it looks good for you?

We cannot and should not drop -nodefconfig. But yes, I agree that we should introduce -no-user-config and use the semantics I specified earlier in the thread. Regards, Anthony Liguori

On 03/25/2012 09:58 AM, Gleb Natapov wrote:

On Sun, Mar 25, 2012 at 03:14:56PM +0200, Avi Kivity wrote:

...

On 03/25/2012 03:12 PM, Anthony Liguori wrote:

...

...

...

qemu -M pc

Would effectively be short hand for -readconfig /usr/share/qemu/machines/pc.cfg

In that case

qemu -cpu westmere

is shorthand for -readconfig /usr/share/qemu/cpus/westmere.cfg.

This is not a bad suggestion, although it would make -cpu ? a bit awkward. Do you see an advantage to this over having /usr/share/qemu/target-x86_64-cpus.cfg that's read early on?

Nope. As long as qemu -nodefconfig -cpu westmere works, I'm happy.

As log as qemu -nodefconfig -cpu westmere -M pc1.1

-nodefconfig is going to eventually mean that -cpu westmere and -M pc-1.1 will not work. This is where QEMU is going. There is no reason that a normal user should ever use -nodefconfig. Regards, Anthony Liguori

On 03/25/2012 10:18 AM, Avi Kivity wrote:

On 03/25/2012 05:07 PM, Anthony Liguori wrote:

...

...

As log as qemu -nodefconfig -cpu westmere -M pc1.1

-nodefconfig is going to eventually mean that -cpu westmere and -M pc-1.1 will not work.

This is where QEMU is going. There is no reason that a normal user should ever use -nodefconfig.

I don't think anyone or anything can use it, since it's meaning is not well defined. "not read any configuration files" where parts of qemu are continually moved out to configuration files means its a moving target.

I think you assume that all QEMU users care about forward and backwards compatibility on the command line about all else. That's really not true. The libvirt folks have stated repeatedly that command line backwards compatibility is not critical to them. They are happy to require that a new version of QEMU requires a new version of libvirt. I'm not saying that backwards compat isn't important--it is. But there are users who are happy to live on the bleeding edge.

Suppose we define the southbridge via a configuration file. Does that mean we don't load it any more?

Yes. If I want the leanest and meanest version of QEMU that will start in the smallest number of milliseconds, then being able to tell QEMU not to load configuration files and create a very specific machine is a Good Thing. Why exclude users from being able to do this? Regards, Anthony Liguori

On 03/25/2012 10:45 AM, Avi Kivity wrote:

On 03/25/2012 05:30 PM, Anthony Liguori wrote:

...

On 03/25/2012 10:18 AM, Avi Kivity wrote:

...

On 03/25/2012 05:07 PM, Anthony Liguori wrote:

...

...

As log as qemu -nodefconfig -cpu westmere -M pc1.1

-nodefconfig is going to eventually mean that -cpu westmere and -M pc-1.1 will not work.

This is where QEMU is going. There is no reason that a normal user should ever use -nodefconfig.

I don't think anyone or anything can use it, since it's meaning is not well defined. "not read any configuration files" where parts of qemu are continually moved out to configuration files means its a moving target.

I think you assume that all QEMU users care about forward and backwards compatibility on the command line about all else.

That's really not true. The libvirt folks have stated repeatedly that command line backwards compatibility is not critical to them. They are happy to require that a new version of QEMU requires a new version of libvirt.

I don't think this came out of happiness, but despair. Seriously, keeping compatibility is one of the things we work hardest to achieve, and we can't manage it for our command line?

I hate to burst your bubble, but we struggle and rarely maintain the level of compatibility you're seeking to have. I agree with you that we need to do a better job maintaining compatibility which is why I'm trying to clearly separate the things that we will never break from the things that will change over time. -nodefconfig is a moving target. If you want stability, don't use it. If you just want to prevent the user's /etc/qemu stuff from being loaded, use -no-user-config.

...

I'm not saying that backwards compat isn't important--it is. But there are users who are happy to live on the bleeding edge.

That's fine, but I don't see how -nodefconfig helps them. All it does is take away the building blocks (definitions) that they can use when setting up their configuration.

Yes, this is a feature.

...

...

Suppose we define the southbridge via a configuration file. Does that mean we don't load it any more?

Yes. If I want the leanest and meanest version of QEMU that will start in the smallest number of milliseconds, then being able to tell QEMU not to load configuration files and create a very specific machine is a Good Thing. Why exclude users from being able to do this?

So is this the point? Reducing startup time?

Yes, that's one reason. But maybe a user wants to have a whole different set of machine types and doesn't care to have the ones we provide. Why prevent a user from doing this? Maybe they have a management tool that attempts to totally hide QEMU from the end user and exposes a different set of machine types. It's certainly more convenient for something like the Android emulator to only have to deal with QEMU knowing about the 4 types of machines that it specifically supports. Regards, Anthony Liguori

On 03/25/2012 01:09 PM, Avi Kivity wrote:

...

...

...

...

Suppose we define the southbridge via a configuration file. Does that mean we don't load it any more?

Yes. If I want the leanest and meanest version of QEMU that will start in the smallest number of milliseconds, then being able to tell QEMU not to load configuration files and create a very specific machine is a Good Thing. Why exclude users from being able to do this?

So is this the point? Reducing startup time?

Yes, that's one reason. But maybe a user wants to have a whole different set of machine types and doesn't care to have the ones we provide. Why prevent a user from doing this?

How are we preventing a user from doing it? In what way is -nodefconfig helping it?

Let me explain it in a different way, perhaps. We launch smbd in QEMU in order to do file sharing over slirp. One of the historic problems we've had is that we don't assume root privileges, yet want to be able to run smbd without using any of the system configuration files. You can do this by specify -s with the config file, and then in the config file you can overload the various default paths (like private dir, lock dir, etc.). In some cases, earlier versions of smbd didn't allow you to change private dir. You should be able to tell a well behaved tool not to read any configuration/data files and explicitly tell it where/how to read them. We cannot exhaustively anticipate every future use case of QEMU. But beyond the justification for -nodefconfig, the fact is that it exists today, and has a specific semantic. If we want to have a different semantic, we should introduce a new option (-no-user-config). Regards, Anthony Liguori

...

Maybe they have a management tool that attempts to totally hide QEMU from the end user and exposes a different set of machine types. It's certainly more convenient for something like the Android emulator to only have to deal with QEMU knowing about the 4 types of machines that it specifically supports.

If it supports four types, it should always pass one of them to qemu. The only thing -nodefconfig adds is breakage when qemu moves something that one of those four machines relies on to a config file.

I added a summary of the changes I am planning, at: http://wiki.qemu.org/Features/CPUModels#Current_Issues_and_proposed_changes I'm sure I forgot lots of details, so feel free to send me questions and suggestions, or even edit the wiki page directly. On Fri, Mar 09, 2012 at 05:56:52PM -0300, Eduardo Habkost wrote:

Resurrecting an old thread:

[...] -- Eduardo