| Message ID | 20241025101959.601048-19-eric.auger@redhat.com |
|---|---|
| State | New |
| Headers | show |
| Series | kvm/arm: Introduce a customizable aarch64 KVM host model | expand |
On Fri, Oct 25, 2024 at 12:17:37PM +0200, Eric Auger wrote: > This new cpu model takes by default the host cpu values. > However it exposes uint64 SYSREG properties for writable ID reg > fields exposed by the host kernel. Properties are named > SYSREG_<REG>_<FIELD> with REG and FIELD being those used > in linux arch/arm64/tools/sysreg. This done by matching the > writable fields retrieved from the host kernel against the > generated description of sysregs. > > An example of invocation is: > -cpu custom,SYSREG_ID_AA64ISAR0_EL1_DP=0x0 > which sets DP field of ID_AA64ISAR0_EL1 to 0. "SYSREG_" feels kinda redundant to repeat on every single feature. Also, is this naming convention really the same one that users will see when they look at /proc/cpuinfo to view features ? It feels pretty low level to me ? Naming after the registers & fields, would be like configuring x86 CPU features by asking for "SYSREG_EAX_1_ECX_20" instead of saying "vmx" which is the human friendly name. > Signed-off-by: Eric Auger <eric.auger@redhat.com> > Signed-off-by: Cornelia Huck <cohuck@redhat.com> > > --- > > At the moment, the custom model does not support legacy options > of the host cpu model. We need to understand what we do with those > latter (SVE, ...). This means that related KVM ioctl are > not called yet. It will be pretty painful to have to use different feature terminology for different CPU models. Everything in libvirt assuming feature terminology varies per-arch, not per-CPU model. With regards, Daniel
Hi Daniel, On 10/25/24 15:06, Daniel P. Berrangé wrote: > On Fri, Oct 25, 2024 at 12:17:37PM +0200, Eric Auger wrote: >> This new cpu model takes by default the host cpu values. >> However it exposes uint64 SYSREG properties for writable ID reg >> fields exposed by the host kernel. Properties are named >> SYSREG_<REG>_<FIELD> with REG and FIELD being those used >> in linux arch/arm64/tools/sysreg. This done by matching the >> writable fields retrieved from the host kernel against the >> generated description of sysregs. >> >> An example of invocation is: >> -cpu custom,SYSREG_ID_AA64ISAR0_EL1_DP=0x0 >> which sets DP field of ID_AA64ISAR0_EL1 to 0. > "SYSREG_" feels kinda redundant to repeat on every single > feature. I do agree. To be honest this was mostly driven my implementation need for cpu model expansion. Given the high number of props which are getting exposed, I iterate on all props and having a prefix let me return only those SYSREG props. Most probably we can get rid of the prefix by using some generated code as well. > > Also, is this naming convention really the same one that users > will see when they look at /proc/cpuinfo to view features ? It No it is not. I do agree that the custom cpu model is very low level. It is very well suited to test all series turning ID regs as writable but this would require an extra layer that adapts /proc/cpuinfo feature level to this regid/field abstraction. In /cpu/proc you will see somethink like: Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm lrcpc dcpop asimddp > feels pretty low level to me ? Naming after the registers & > fields, would be like configuring x86 CPU features by asking > for "SYSREG_EAX_1_ECX_20" instead of saying "vmx" which is the > human friendly name. agreed. > > >> Signed-off-by: Eric Auger <eric.auger@redhat.com> >> Signed-off-by: Cornelia Huck <cohuck@redhat.com> >> >> --- >> >> At the moment, the custom model does not support legacy options >> of the host cpu model. We need to understand what we do with those >> latter (SVE, ...). This means that related KVM ioctl are >> not called yet. > It will be pretty painful to have to use different feature > terminology for different CPU models. Everything in libvirt > assuming feature terminology varies per-arch, not per-CPU > model. Actually as far as I understand those regids/fields would fit all kind of aarch64 Cortex-A CPUs. So they wouldn't vary per-CPU (I mean their terminology. Their availability will). Thanks Eric > > > With regards, > Daniel
On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: > Hi Daniel, > > On 10/25/24 15:06, Daniel P. Berrangé wrote: > > On Fri, Oct 25, 2024 at 12:17:37PM +0200, Eric Auger wrote: > >> This new cpu model takes by default the host cpu values. > >> However it exposes uint64 SYSREG properties for writable ID reg > >> fields exposed by the host kernel. Properties are named > >> SYSREG_<REG>_<FIELD> with REG and FIELD being those used > >> in linux arch/arm64/tools/sysreg. This done by matching the > >> writable fields retrieved from the host kernel against the > >> generated description of sysregs. > >> > >> An example of invocation is: > >> -cpu custom,SYSREG_ID_AA64ISAR0_EL1_DP=0x0 > >> which sets DP field of ID_AA64ISAR0_EL1 to 0. > > "SYSREG_" feels kinda redundant to repeat on every single > > feature. > > I do agree. To be honest this was mostly driven my implementation need > for cpu model expansion. Given the high number of props which are > getting exposed, I iterate on all props and having a prefix let me > return only those SYSREG props. Most probably we can get rid of the > prefix by using some generated code as well. > > > > Also, is this naming convention really the same one that users > > will see when they look at /proc/cpuinfo to view features ? It > No it is not. I do agree that the custom cpu model is very low level. It > is very well suited to test all series turning ID regs as writable but > this would require an extra layer that adapts /proc/cpuinfo feature > level to this regid/field abstraction. > > In /cpu/proc you will see somethink like: > Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp > asimdhp cpuid asimdrdm lrcpc dcpop asimddp Right, IMHO, this is the terminology that QEMU must use in user facing APIs. On x86 we have a conversion tables for named features to register bits https://gitlab.com/qemu-project/qemu/-/blob/master/target/i386/cpu.c#L914 and libvirt does similar https://gitlab.com/libvirt/libvirt/-/blob/master/src/cpu_map/x86_features.xml Yep, it is more work, but it is also much better for humans IMHO. > > feels pretty low level to me ? Naming after the registers & > > fields, would be like configuring x86 CPU features by asking > > for "SYSREG_EAX_1_ECX_20" instead of saying "vmx" which is the > > human friendly name. > agreed. > >> At the moment, the custom model does not support legacy options > >> of the host cpu model. We need to understand what we do with those > >> latter (SVE, ...). This means that related KVM ioctl are > >> not called yet. > > It will be pretty painful to have to use different feature > > terminology for different CPU models. Everything in libvirt > > assuming feature terminology varies per-arch, not per-CPU > > model. > Actually as far as I understand those regids/fields would fit all kind > of aarch64 Cortex-A CPUs. So they wouldn't vary per-CPU (I mean their > terminology. Their availability will). What I mean is can we define named models for various different vendor's Cortex-A silicon and just use that without needing to toggle features, except in rare cases. With regards, Daniel
On Fri, Oct 25 2024, Daniel P. Berrangé <berrange@redhat.com> wrote: > On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: >> Hi Daniel, >> >> On 10/25/24 15:06, Daniel P. Berrangé wrote: >> > On Fri, Oct 25, 2024 at 12:17:37PM +0200, Eric Auger wrote: >> >> At the moment, the custom model does not support legacy options >> >> of the host cpu model. We need to understand what we do with those >> >> latter (SVE, ...). This means that related KVM ioctl are >> >> not called yet. >> > It will be pretty painful to have to use different feature >> > terminology for different CPU models. Everything in libvirt >> > assuming feature terminology varies per-arch, not per-CPU >> > model. >> Actually as far as I understand those regids/fields would fit all kind >> of aarch64 Cortex-A CPUs. So they wouldn't vary per-CPU (I mean their >> terminology. Their availability will). > > What I mean is can we define named models for various different > vendor's Cortex-A silicon and just use that without needing to > toggle features, except in rare cases. I'm not sure whether creating named models for various cpus would actually scale; what we probably could do is come up with some Armv8.6, Armv8.7, ... models and have additional models for some well-known cpus. That would also give us some basic sanity checks for feature combinations, if we do not want to support completely custom frankencpus.
On Mon, Oct 28, 2024 at 05:00:46PM +0100, Cornelia Huck wrote: > On Fri, Oct 25 2024, Daniel P. Berrangé <berrange@redhat.com> wrote: > > > On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: > >> Hi Daniel, > >> > >> On 10/25/24 15:06, Daniel P. Berrangé wrote: > >> > On Fri, Oct 25, 2024 at 12:17:37PM +0200, Eric Auger wrote: > >> >> At the moment, the custom model does not support legacy options > >> >> of the host cpu model. We need to understand what we do with those > >> >> latter (SVE, ...). This means that related KVM ioctl are > >> >> not called yet. > >> > It will be pretty painful to have to use different feature > >> > terminology for different CPU models. Everything in libvirt > >> > assuming feature terminology varies per-arch, not per-CPU > >> > model. > >> Actually as far as I understand those regids/fields would fit all kind > >> of aarch64 Cortex-A CPUs. So they wouldn't vary per-CPU (I mean their > >> terminology. Their availability will). > > > > What I mean is can we define named models for various different > > vendor's Cortex-A silicon and just use that without needing to > > toggle features, except in rare cases. > > I'm not sure whether creating named models for various cpus would > actually scale; what we probably could do is come up with some Armv8.6, > Armv8.7, ... models and have additional models for some well-known cpus. > That would also give us some basic sanity checks for feature > combinations, if we do not want to support completely custom > frankencpus. Yep, even on x86 we don't add named models for every conceivable CPU SKU variant, just one so called "typical" variant for a given microarchiteture level [1]. The vendor independent Arm CPU spec versions are a good enough match for typical real world silicon, then it would make sense to focus mostly on such named CPU models, not try to make an exhaustive definition of every vendor's impl. I could see there might be some exceptions where adding a vendor specific named model would be worthwhile, but hopefully that'd be relatively uncommon and -cpu Armv8.6,+feat1,+feat2 would be sufficient. eg toggling 1/2/3 extra features to precise match a vendor's named mode is ok, but adding 20-30 extra features get a match is painful. That's the kind of rule of thumb that could be used when deciding whether its worth adding a vendor specific named model, vs relying on the generic spec based models. With regards, Daniel [1] we do have CPU model versions, as a way to "bugfix" our CPU definitions.
On Fri, 25 Oct 2024 at 14:24, Daniel P. Berrangé <berrange@redhat.com> wrote: > On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: > > On 10/25/24 15:06, Daniel P. Berrangé wrote: > > > Also, is this naming convention really the same one that users > > > will see when they look at /proc/cpuinfo to view features ? It > > No it is not. I do agree that the custom cpu model is very low level. It > > is very well suited to test all series turning ID regs as writable but > > this would require an extra layer that adapts /proc/cpuinfo feature > > level to this regid/field abstraction. > > > > In /cpu/proc you will see somethink like: > > Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp > > asimdhp cpuid asimdrdm lrcpc dcpop asimddp > > Right, IMHO, this is the terminology that QEMU must use in user > facing APIs. /proc/cpuinfo's naming is rather weird for historical reasons (for instance there is only one FEAT_FP16 feature but cpuinfo lists "fphp" and "asimdhp" separately). Currently QEMU only has to care about cpuinfo name tags in linux-user/elfload.c where there's a bunch of data structures for "what hwcaps do we need to advertise given what the CPU has?". I would definitely prefer it if we could use architectural feature names... On other architectures do we do anything to forbid invalid combinations? For Arm there are architectural rules about feature X requiring features Y and Z. Are we going to just let the user create a CPU that the guest OS will barf on if they want to? (The user-experience for that is potentially not very nice, because something like "-cpu cortex-a57,+sve" seems like something you might want to do but isn't actually valid; even listing the direct required dependency of FEAT_SVE like "-cpu cortex-a57,+sve,+fp16" isn't sufficient because SVE is optional-from-v8.2 and so a guest could in theory assume the presence of anything mandatory in v8.1 and v8.2. But even merely diagnosing invalid combinations is a huge pain, and automagically pulling in every mandatory implicit or explicit dependency seems like it might be surprising to users, as well as being annoying to implement. Currently we sidestep all of these problems by (a) only allowing the command line to disable a feature, not to enable it where it didn't previously exist and (b) mostly not providing command line flags for individual features...) thanks -- PMM
On Mon, Oct 28 2024, Peter Maydell <peter.maydell@linaro.org> wrote: > On Fri, 25 Oct 2024 at 14:24, Daniel P. Berrangé <berrange@redhat.com> wrote: >> On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: >> > On 10/25/24 15:06, Daniel P. Berrangé wrote: >> > > Also, is this naming convention really the same one that users >> > > will see when they look at /proc/cpuinfo to view features ? It >> > No it is not. I do agree that the custom cpu model is very low level. It >> > is very well suited to test all series turning ID regs as writable but >> > this would require an extra layer that adapts /proc/cpuinfo feature >> > level to this regid/field abstraction. >> > >> > In /cpu/proc you will see somethink like: >> > Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp >> > asimdhp cpuid asimdrdm lrcpc dcpop asimddp >> >> Right, IMHO, this is the terminology that QEMU must use in user >> facing APIs. > > /proc/cpuinfo's naming is rather weird for historical > reasons (for instance there is only one FEAT_FP16 feature > but cpuinfo lists "fphp" and "asimdhp" separately). > Currently QEMU only has to care about cpuinfo name tags > in linux-user/elfload.c where there's a bunch of data > structures for "what hwcaps do we need to advertise > given what the CPU has?". I would definitely prefer it if > we could use architectural feature names... > > On other architectures do we do anything to forbid > invalid combinations? For Arm there are architectural > rules about feature X requiring features Y and Z. > Are we going to just let the user create a CPU that > the guest OS will barf on if they want to? (The > user-experience for that is potentially not very nice, > because something like "-cpu cortex-a57,+sve" seems like > something you might want to do but isn't actually valid; > even listing the direct required dependency of FEAT_SVE > like "-cpu cortex-a57,+sve,+fp16" isn't sufficient > because SVE is optional-from-v8.2 and so a guest could > in theory assume the presence of anything mandatory in > v8.1 and v8.2. But even merely diagnosing invalid > combinations is a huge pain, and automagically pulling > in every mandatory implicit or explicit dependency > seems like it might be surprising to users, as well as > being annoying to implement. Currently we sidestep > all of these problems by (a) only allowing the command > line to disable a feature, not to enable it where it > didn't previously exist and (b) mostly not providing > command line flags for individual features...) I think s390 does some dependency checking on features, and also rejects features that are "too new".
On Mon, Oct 28, 2024 at 04:16:31PM +0000, Peter Maydell wrote: > On Fri, 25 Oct 2024 at 14:24, Daniel P. Berrangé <berrange@redhat.com> wrote: > > On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: > > > On 10/25/24 15:06, Daniel P. Berrangé wrote: > > > > Also, is this naming convention really the same one that users > > > > will see when they look at /proc/cpuinfo to view features ? It > > > No it is not. I do agree that the custom cpu model is very low level. It > > > is very well suited to test all series turning ID regs as writable but > > > this would require an extra layer that adapts /proc/cpuinfo feature > > > level to this regid/field abstraction. > > > > > > In /cpu/proc you will see somethink like: > > > Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp > > > asimdhp cpuid asimdrdm lrcpc dcpop asimddp > > > > Right, IMHO, this is the terminology that QEMU must use in user > > facing APIs. > > /proc/cpuinfo's naming is rather weird for historical > reasons (for instance there is only one FEAT_FP16 feature > but cpuinfo lists "fphp" and "asimdhp" separately). There's plenty of wierd history in x86 too. In this case I might suggest just picking one of the two common names, and ignoring the other. If we really wanted to, we could alias the 2nd name to the first, but its likely not worth the bother. > Currently QEMU only has to care about cpuinfo name tags > in linux-user/elfload.c where there's a bunch of data > structures for "what hwcaps do we need to advertise > given what the CPU has?". I would definitely prefer it if > we could use architectural feature names... > > On other architectures do we do anything to forbid > invalid combinations? For Arm there are architectural > rules about feature X requiring features Y and Z. > Are we going to just let the user create a CPU that > the guest OS will barf on if they want to? (The > user-experience for that is potentially not very nice, > because something like "-cpu cortex-a57,+sve" seems like > something you might want to do but isn't actually valid; > even listing the direct required dependency of FEAT_SVE > like "-cpu cortex-a57,+sve,+fp16" isn't sufficient > because SVE is optional-from-v8.2 and so a guest could > in theory assume the presence of anything mandatory in > v8.1 and v8.2. But even merely diagnosing invalid > combinations is a huge pain, and automagically pulling > in every mandatory implicit or explicit dependency > seems like it might be surprising to users, as well as > being annoying to implement. Currently we sidestep > all of these problems by (a) only allowing the command > line to disable a feature, not to enable it where it > didn't previously exist and (b) mostly not providing > command line flags for individual features...) AFAICT, in x86 world, there's no generally documented dependancies between features. I imagine there might be deps mentioned in an adhoc basis in docs/specs. There are certainly implicit deps though where apps check for X, and blindly assume it means Y+Z also exist. In the QEMU & libvirt world though, there's no deps. This is generally why we discourage apps/users from trying to enable arbitrarily many features. People should stick to the named CPU models, and just use feature enablement as an exception. Enabling individual features *is* certainly a critically important benefit though, even with the risks from implicit undocumented dependencies. Most notably, it has been used to turn on fixes for hardware CVEs, where microcode introduces a new named feature. Your named CPU model may not be updated, but as long as QEMU knows the feature name, you can turn it on. Overall, while it would be desirable to have a way to express feature dependencies, I wouldn't call that a blocker for enabling the ability to turn features on & off. With regards, Daniel
On Mon, 28 Oct 2024 at 16:35, Daniel P. Berrangé <berrange@redhat.com> wrote: > > On Mon, Oct 28, 2024 at 04:16:31PM +0000, Peter Maydell wrote: > > On Fri, 25 Oct 2024 at 14:24, Daniel P. Berrangé <berrange@redhat.com> wrote: > > > On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: > > > > On 10/25/24 15:06, Daniel P. Berrangé wrote: > > > > > Also, is this naming convention really the same one that users > > > > > will see when they look at /proc/cpuinfo to view features ? It > > > > No it is not. I do agree that the custom cpu model is very low level. It > > > > is very well suited to test all series turning ID regs as writable but > > > > this would require an extra layer that adapts /proc/cpuinfo feature > > > > level to this regid/field abstraction. > > > > > > > > In /cpu/proc you will see somethink like: > > > > Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp > > > > asimdhp cpuid asimdrdm lrcpc dcpop asimddp > > > > > > Right, IMHO, this is the terminology that QEMU must use in user > > > facing APIs. > > > > /proc/cpuinfo's naming is rather weird for historical > > reasons (for instance there is only one FEAT_FP16 feature > > but cpuinfo lists "fphp" and "asimdhp" separately). > > There's plenty of wierd history in x86 too. In this > case I might suggest just picking one of the two > common names, and ignoring the other. > > If we really wanted to, we could alias the 2nd name > to the first, but its likely not worth the bother. Or we could use the standard set of architectural feature names, and not have the problem at all, and not have to document what we mean by our nonstandard names. (cpuinfo names do actually mostly line up with the standard names, just not 100%. Similarly gcc/clang command line options are mostly the architectural feature name.) thanks -- PMM
On Mon, Oct 28, 2024 at 04:48:18PM +0000, Peter Maydell wrote: > On Mon, 28 Oct 2024 at 16:35, Daniel P. Berrangé <berrange@redhat.com> wrote: > > > > On Mon, Oct 28, 2024 at 04:16:31PM +0000, Peter Maydell wrote: > > > On Fri, 25 Oct 2024 at 14:24, Daniel P. Berrangé <berrange@redhat.com> wrote: > > > > On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: > > > > > On 10/25/24 15:06, Daniel P. Berrangé wrote: > > > > > > Also, is this naming convention really the same one that users > > > > > > will see when they look at /proc/cpuinfo to view features ? It > > > > > No it is not. I do agree that the custom cpu model is very low level. It > > > > > is very well suited to test all series turning ID regs as writable but > > > > > this would require an extra layer that adapts /proc/cpuinfo feature > > > > > level to this regid/field abstraction. > > > > > > > > > > In /cpu/proc you will see somethink like: > > > > > Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp > > > > > asimdhp cpuid asimdrdm lrcpc dcpop asimddp > > > > > > > > Right, IMHO, this is the terminology that QEMU must use in user > > > > facing APIs. > > > > > > /proc/cpuinfo's naming is rather weird for historical > > > reasons (for instance there is only one FEAT_FP16 feature > > > but cpuinfo lists "fphp" and "asimdhp" separately). > > > > There's plenty of wierd history in x86 too. In this > > case I might suggest just picking one of the two > > common names, and ignoring the other. > > > > If we really wanted to, we could alias the 2nd name > > to the first, but its likely not worth the bother. > > Or we could use the standard set of architectural > feature names, and not have the problem at all, and not > have to document what we mean by our nonstandard names. +1 There's existing documentation [*] for the standard feature names, which provides: - A short description of what the feature does - Any dependencies a particular feature has (e.g.FEAT_VHE implies FEAT_LSE, FEAT_Debugv8p1, and FEAT_AA64EL2) - The register fields/values that are used to discover the feature. This seems like the most user-friendly option... [*]: https://developer.arm.com/documentation/109697/2024_09
On Mon, Oct 28, 2024 at 04:48:18PM +0000, Peter Maydell wrote: > On Mon, 28 Oct 2024 at 16:35, Daniel P. Berrangé <berrange@redhat.com> wrote: > > > > On Mon, Oct 28, 2024 at 04:16:31PM +0000, Peter Maydell wrote: > > > On Fri, 25 Oct 2024 at 14:24, Daniel P. Berrangé <berrange@redhat.com> wrote: > > > > On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: > > > > > On 10/25/24 15:06, Daniel P. Berrangé wrote: > > > > > > Also, is this naming convention really the same one that users > > > > > > will see when they look at /proc/cpuinfo to view features ? It > > > > > No it is not. I do agree that the custom cpu model is very low level. It > > > > > is very well suited to test all series turning ID regs as writable but > > > > > this would require an extra layer that adapts /proc/cpuinfo feature > > > > > level to this regid/field abstraction. > > > > > > > > > > In /cpu/proc you will see somethink like: > > > > > Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp > > > > > asimdhp cpuid asimdrdm lrcpc dcpop asimddp > > > > > > > > Right, IMHO, this is the terminology that QEMU must use in user > > > > facing APIs. > > > > > > /proc/cpuinfo's naming is rather weird for historical > > > reasons (for instance there is only one FEAT_FP16 feature > > > but cpuinfo lists "fphp" and "asimdhp" separately). > > > > There's plenty of wierd history in x86 too. In this > > case I might suggest just picking one of the two > > common names, and ignoring the other. > > > > If we really wanted to, we could alias the 2nd name > > to the first, but its likely not worth the bother. > > Or we could use the standard set of architectural > feature names, and not have the problem at all, and not > have to document what we mean by our nonstandard names. > (cpuinfo names do actually mostly line up with the > standard names, just not 100%. Similarly gcc/clang command > line options are mostly the architectural feature name.) Ah, right, yes. Sorry I mis-understood you originally to be suggesting the same low level names as this patch. With regards, Daniel
On Mon, Oct 28 2024, Oliver Upton <oliver.upton@linux.dev> wrote: > On Mon, Oct 28, 2024 at 04:48:18PM +0000, Peter Maydell wrote: >> On Mon, 28 Oct 2024 at 16:35, Daniel P. Berrangé <berrange@redhat.com> wrote: >> > >> > On Mon, Oct 28, 2024 at 04:16:31PM +0000, Peter Maydell wrote: >> > > On Fri, 25 Oct 2024 at 14:24, Daniel P. Berrangé <berrange@redhat.com> wrote: >> > > > On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: >> > > > > On 10/25/24 15:06, Daniel P. Berrangé wrote: >> > > > > > Also, is this naming convention really the same one that users >> > > > > > will see when they look at /proc/cpuinfo to view features ? It >> > > > > No it is not. I do agree that the custom cpu model is very low level. It >> > > > > is very well suited to test all series turning ID regs as writable but >> > > > > this would require an extra layer that adapts /proc/cpuinfo feature >> > > > > level to this regid/field abstraction. >> > > > > >> > > > > In /cpu/proc you will see somethink like: >> > > > > Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp >> > > > > asimdhp cpuid asimdrdm lrcpc dcpop asimddp >> > > > >> > > > Right, IMHO, this is the terminology that QEMU must use in user >> > > > facing APIs. >> > > >> > > /proc/cpuinfo's naming is rather weird for historical >> > > reasons (for instance there is only one FEAT_FP16 feature >> > > but cpuinfo lists "fphp" and "asimdhp" separately). >> > >> > There's plenty of wierd history in x86 too. In this >> > case I might suggest just picking one of the two >> > common names, and ignoring the other. >> > >> > If we really wanted to, we could alias the 2nd name >> > to the first, but its likely not worth the bother. >> >> Or we could use the standard set of architectural >> feature names, and not have the problem at all, and not >> have to document what we mean by our nonstandard names. > > +1 > > There's existing documentation [*] for the standard feature names, which > provides: > > - A short description of what the feature does > - Any dependencies a particular feature has (e.g.FEAT_VHE implies > FEAT_LSE, FEAT_Debugv8p1, and FEAT_AA64EL2) > - The register fields/values that are used to discover the feature. > > This seems like the most user-friendly option... > > [*]: https://developer.arm.com/documentation/109697/2024_09 FEAT_XXX sounds good, would be a different approach than this series obviously, since the user resp. upper software layers would operate on a per-feature basis, and QEMU would translate to and from registers. I'm wondering about the amount of translation that would be needed, and what information would be best exposed via QEMU, given that a feature may or may not be toggable not only because of what the Arm revision specifies, but what registers the host kernel allows to be written. I.e. if we have two cpus that differ in whether FEAT_FOO is provided, would it make sense to have an extra QMP command so that you can find out whether FEAT_FOO can be switched off, with QEMU translating from the set of writable id registers to the set of features that can be changed?
On Wed, Oct 30, 2024 at 05:15:05PM +0100, Cornelia Huck wrote: > On Mon, Oct 28 2024, Oliver Upton <oliver.upton@linux.dev> wrote: > > > On Mon, Oct 28, 2024 at 04:48:18PM +0000, Peter Maydell wrote: > >> On Mon, 28 Oct 2024 at 16:35, Daniel P. Berrangé <berrange@redhat.com> wrote: > >> > > >> > On Mon, Oct 28, 2024 at 04:16:31PM +0000, Peter Maydell wrote: > >> > > On Fri, 25 Oct 2024 at 14:24, Daniel P. Berrangé <berrange@redhat.com> wrote: > >> > > > On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: > >> > > > > On 10/25/24 15:06, Daniel P. Berrangé wrote: > >> > > > > > Also, is this naming convention really the same one that users > >> > > > > > will see when they look at /proc/cpuinfo to view features ? It > >> > > > > No it is not. I do agree that the custom cpu model is very low level. It > >> > > > > is very well suited to test all series turning ID regs as writable but > >> > > > > this would require an extra layer that adapts /proc/cpuinfo feature > >> > > > > level to this regid/field abstraction. > >> > > > > > >> > > > > In /cpu/proc you will see somethink like: > >> > > > > Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp > >> > > > > asimdhp cpuid asimdrdm lrcpc dcpop asimddp > >> > > > > >> > > > Right, IMHO, this is the terminology that QEMU must use in user > >> > > > facing APIs. > >> > > > >> > > /proc/cpuinfo's naming is rather weird for historical > >> > > reasons (for instance there is only one FEAT_FP16 feature > >> > > but cpuinfo lists "fphp" and "asimdhp" separately). > >> > > >> > There's plenty of wierd history in x86 too. In this > >> > case I might suggest just picking one of the two > >> > common names, and ignoring the other. > >> > > >> > If we really wanted to, we could alias the 2nd name > >> > to the first, but its likely not worth the bother. > >> > >> Or we could use the standard set of architectural > >> feature names, and not have the problem at all, and not > >> have to document what we mean by our nonstandard names. > > > > +1 > > > > There's existing documentation [*] for the standard feature names, which > > provides: > > > > - A short description of what the feature does > > - Any dependencies a particular feature has (e.g.FEAT_VHE implies > > FEAT_LSE, FEAT_Debugv8p1, and FEAT_AA64EL2) > > - The register fields/values that are used to discover the feature. > > > > This seems like the most user-friendly option... > > > > [*]: https://developer.arm.com/documentation/109697/2024_09 > > FEAT_XXX sounds good, would be a different approach than this series > obviously, since the user resp. upper software layers would operate on a > per-feature basis, and QEMU would translate to and from registers. > > I'm wondering about the amount of translation that would be needed, and > what information would be best exposed via QEMU, given that a feature > may or may not be toggable not only because of what the Arm revision > specifies, but what registers the host kernel allows to be written. > > I.e. if we have two cpus that differ in whether FEAT_FOO is provided, > would it make sense to have an extra QMP command so that you can find > out whether FEAT_FOO can be switched off, with QEMU translating from the > set of writable id registers to the set of features that can be changed? If there are restrictions on what features can be turned off, then yes, we should add that in QMP, as x86 doesn't have such a restriction currently. With regards, Daniel
Hi Daniel, On 10/28/24 18:04, Daniel P. Berrangé wrote: > On Mon, Oct 28, 2024 at 04:48:18PM +0000, Peter Maydell wrote: >> On Mon, 28 Oct 2024 at 16:35, Daniel P. Berrangé <berrange@redhat.com> wrote: >>> On Mon, Oct 28, 2024 at 04:16:31PM +0000, Peter Maydell wrote: >>>> On Fri, 25 Oct 2024 at 14:24, Daniel P. Berrangé <berrange@redhat.com> wrote: >>>>> On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: >>>>>> On 10/25/24 15:06, Daniel P. Berrangé wrote: >>>>>>> Also, is this naming convention really the same one that users >>>>>>> will see when they look at /proc/cpuinfo to view features ? It >>>>>> No it is not. I do agree that the custom cpu model is very low level. It >>>>>> is very well suited to test all series turning ID regs as writable but >>>>>> this would require an extra layer that adapts /proc/cpuinfo feature >>>>>> level to this regid/field abstraction. >>>>>> >>>>>> In /cpu/proc you will see somethink like: >>>>>> Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp >>>>>> asimdhp cpuid asimdrdm lrcpc dcpop asimddp >>>>> Right, IMHO, this is the terminology that QEMU must use in user >>>>> facing APIs. >>>> /proc/cpuinfo's naming is rather weird for historical >>>> reasons (for instance there is only one FEAT_FP16 feature >>>> but cpuinfo lists "fphp" and "asimdhp" separately). >>> There's plenty of wierd history in x86 too. In this >>> case I might suggest just picking one of the two >>> common names, and ignoring the other. >>> >>> If we really wanted to, we could alias the 2nd name >>> to the first, but its likely not worth the bother. >> Or we could use the standard set of architectural >> feature names, and not have the problem at all, and not >> have to document what we mean by our nonstandard names. >> (cpuinfo names do actually mostly line up with the >> standard names, just not 100%. Similarly gcc/clang command >> line options are mostly the architectural feature name.) > Ah, right, yes. Sorry I mis-understood you originally to be suggesting > the same low level names as this patch. If my understanding is correct, Peter suggested to rely on the terminology used in https://developer.arm.com/documentation/109697/2024_09 the doc pointed to by Oliver. So I think the next step is to understand how those "high level" features do map onto low level ID register field values. I think a high level feature can map onto separate fields in separate ID regs. This may not be the most common case though. Thanks Eric > > With regards, > Daniel
Hi Oliver, On 10/28/24 17:56, Oliver Upton wrote: > On Mon, Oct 28, 2024 at 04:48:18PM +0000, Peter Maydell wrote: >> On Mon, 28 Oct 2024 at 16:35, Daniel P. Berrangé <berrange@redhat.com> wrote: >>> On Mon, Oct 28, 2024 at 04:16:31PM +0000, Peter Maydell wrote: >>>> On Fri, 25 Oct 2024 at 14:24, Daniel P. Berrangé <berrange@redhat.com> wrote: >>>>> On Fri, Oct 25, 2024 at 03:18:25PM +0200, Eric Auger wrote: >>>>>> On 10/25/24 15:06, Daniel P. Berrangé wrote: >>>>>>> Also, is this naming convention really the same one that users >>>>>>> will see when they look at /proc/cpuinfo to view features ? It >>>>>> No it is not. I do agree that the custom cpu model is very low level. It >>>>>> is very well suited to test all series turning ID regs as writable but >>>>>> this would require an extra layer that adapts /proc/cpuinfo feature >>>>>> level to this regid/field abstraction. >>>>>> >>>>>> In /cpu/proc you will see somethink like: >>>>>> Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp >>>>>> asimdhp cpuid asimdrdm lrcpc dcpop asimddp >>>>> Right, IMHO, this is the terminology that QEMU must use in user >>>>> facing APIs. >>>> /proc/cpuinfo's naming is rather weird for historical >>>> reasons (for instance there is only one FEAT_FP16 feature >>>> but cpuinfo lists "fphp" and "asimdhp" separately). >>> There's plenty of wierd history in x86 too. In this >>> case I might suggest just picking one of the two >>> common names, and ignoring the other. >>> >>> If we really wanted to, we could alias the 2nd name >>> to the first, but its likely not worth the bother. >> Or we could use the standard set of architectural >> feature names, and not have the problem at all, and not >> have to document what we mean by our nonstandard names. > +1 > > There's existing documentation [*] for the standard feature names, which > provides: > > - A short description of what the feature does > - Any dependencies a particular feature has (e.g.FEAT_VHE implies > FEAT_LSE, FEAT_Debugv8p1, and FEAT_AA64EL2) > - The register fields/values that are used to discover the feature. > > This seems like the most user-friendly option... > > [*]: https://developer.arm.com/documentation/109697/2024_09 Was just wondering if all the writable ID reg fields were associated to an official FEAT_ listed in this feature list. Let's take for instance CTR_EL0.DIC. Is it associated to a given extension or is it just implementation defined? Thanks Eric >
On Mon, 4 Nov 2024 at 17:09, Eric Auger <eric.auger@redhat.com> wrote: > Was just wondering if all the writable ID reg fields were associated to > an official FEAT_ listed in this feature list. > Let's take for instance CTR_EL0.DIC. Is it associated to a given > extension or is it just implementation defined? CTR_EL0.DIC is IMPDEF choice of 0 or 1. There's no associated FEAT_* -- as an implementation, you are effectively just reporting here what your implementation's properties are. (CTR_EL0 is not an ID register in the architectural sense.) thanks -- PMM
Hi Peter, On 11/4/24 18:16, Peter Maydell wrote: > On Mon, 4 Nov 2024 at 17:09, Eric Auger <eric.auger@redhat.com> wrote: >> Was just wondering if all the writable ID reg fields were associated to >> an official FEAT_ listed in this feature list. >> Let's take for instance CTR_EL0.DIC. Is it associated to a given >> extension or is it just implementation defined? > CTR_EL0.DIC is IMPDEF choice of 0 or 1. There's no associated > FEAT_* -- as an implementation, you are effectively just > reporting here what your implementation's properties are. > > (CTR_EL0 is not an ID register in the architectural sense.) thank you for the confirmation. So this means we also need to be able to tune such bit although it does not have any fellow FEAT_*. Eric > > thanks > -- PMM >
diff --git a/target/arm/cpu.c b/target/arm/cpu.c index 454d546feb..e5ac3c3e75 100644 --- a/target/arm/cpu.c +++ b/target/arm/cpu.c @@ -1990,6 +1990,21 @@ static void arm_cpu_realizefn(DeviceState *dev, Error **errp) return; } + /* + * If we failed to retrieve the set of writable ID registers for a "custom" + * CPU model, report it here. + * In case we did get the set of writable ID registers, set the features to + * the configured values here and perform some sanity checks. + */ + if (cpu->writable_id_regs == WRITABLE_ID_REGS_NOT_DISCOVERABLE) { + error_setg(errp, "Host kernel does not support discovering " + "writable id registers"); + return; + } else if (cpu->writable_id_regs == WRITABLE_ID_REGS_FAILED) { + error_setg(errp, "Failed to discover writable id registers"); + return; + } + if (!cpu->gt_cntfrq_hz) { /* * 0 means "the board didn't set a value, use the default". (We also diff --git a/target/arm/cpu64.c b/target/arm/cpu64.c index 86b0797d4b..f10cc4ef8f 100644 --- a/target/arm/cpu64.c +++ b/target/arm/cpu64.c @@ -20,6 +20,7 @@ #include "qemu/osdep.h" #include "qapi/error.h" +#include "qemu/error-report.h" #include "cpu.h" #include "cpregs.h" #include "qemu/module.h" @@ -35,6 +36,8 @@ #include "cpu-features.h" #include "cpregs.h" #include "cpu-custom.h" +#include "cpu-sysregs.h" +#include "trace.h" void arm_cpu_sve_finalize(ARMCPU *cpu, Error **errp) { @@ -742,6 +745,153 @@ static void aarch64_max_initfn(Object *obj) } } +#ifdef CONFIG_KVM + +static ARM64SysRegField *get_field(int i, ARM64SysReg *reg) +{ + GList *l; + + for (l = reg->fields; l; l = l->next) { + ARM64SysRegField *field = (ARM64SysRegField *)l->data; + + if (i >= field->lower && i <= field->upper) { + return field; + } + } + return NULL; +} + +static void set_sysreg_prop(Object *obj, Visitor *v, + const char *name, void *opaque, + Error **errp) +{ + ARM64SysRegField *field = (ARM64SysRegField *)opaque; + ARMCPU *cpu = ARM_CPU(obj); + IdRegMap *idregs = &cpu->isar.idregs; + uint64_t old, value, mask; + int lower = field->lower; + int upper = field->upper; + int length = upper - lower + 1; + int index = field->index; + + if (!visit_type_uint64(v, name, &value, errp)) { + return; + } + + if (length < 64 && value > ((1 << length) - 1)) { + error_setg(errp, + "idreg %s set value (0x%lx) exceeds length of field (%d)!", + name, value, length); + return; + } + + mask = MAKE_64BIT_MASK(lower, length); + value = value << lower; + old = idregs->regs[index]; + idregs->regs[index] = old & ~mask; + idregs->regs[index] |= value; + trace_set_sysreg_prop(name, old, mask, value, idregs->regs[index]); +} + +static void get_sysreg_prop(Object *obj, Visitor *v, + const char *name, void *opaque, + Error **errp) +{ + ARM64SysRegField *field = (ARM64SysRegField *)opaque; + ARMCPU *cpu = ARM_CPU(obj); + IdRegMap *idregs = &cpu->isar.idregs; + int index = field->index; + + error_report("%s %s", __func__, name); + visit_type_uint64(v, name, &idregs->regs[index], errp); + trace_get_sysreg_prop(name, idregs->regs[index]); +} + +/* + * decode_idreg_writemap: Generate props for writable fields + * + * @obj: CPU object + * @index: index of the sysreg + * @map: writable map for the sysreg + * @reg: description of the sysreg + */ +static int +decode_idreg_writemap(Object *obj, int index, uint64_t map, ARM64SysReg *reg) +{ + int i = ctz64(map); + int nb_sysreg_props = 0; + + while (map) { + ARM64SysRegField *field = get_field(i, reg); + int lower, upper; + uint64_t mask; + char *prop_name; + + if (!field) { + /* the field cannot be matched to any know id named field */ + warn_report("%s bit %d of %s is writable but cannot be matched", + __func__, i, reg->name); + warn_report("%s is cpu-sysreg-properties.c up to date?", __func__); + map = map & ~BIT_ULL(i); + i = ctz64(map); + continue; + } + lower = field->lower; + upper = field->upper; + prop_name = g_strdup_printf("SYSREG_%s_%s", reg->name, field->name); + trace_decode_idreg_writemap(field->name, lower, upper, prop_name); + object_property_add(obj, prop_name, "uint64", + get_sysreg_prop, set_sysreg_prop, NULL, field); + nb_sysreg_props++; + + mask = MAKE_64BIT_MASK(lower, upper - lower + 1); + map = map & ~mask; + i = ctz64(map); + } + trace_nb_sysreg_props(reg->name, nb_sysreg_props); + return 0; +} + +/* analyze the writable mask and generate properties for writable fields */ +static int expose_idreg_properties(Object *obj, IdRegMap *map, + ARM64SysReg *regs) +{ + int i; + + for (i = 0; i < NR_ID_REGS; i++) { + uint64_t mask = map->regs[i]; + + if (mask) { + /* reg @i has some writable fields, decode them */ + decode_idreg_writemap(obj, i, mask, ®s[i]); + } + } + return 0; +} + +static void aarch64_customcpu_initfn(Object *obj) +{ + ARMCPU *cpu = ARM_CPU(obj); + int ret; + + cpu->writable_map = g_malloc(sizeof(IdRegMap)); + + /* discover via KVM_ARM_GET_REG_WRITABLE_MASKS */ + ret = kvm_arm_get_writable_id_regs(cpu, cpu->writable_map); + if (ret) { + /* function will have marked an error */ + return; + } + + /* populate from the host (exhaustive) , validate during realize */ + kvm_arm_set_cpu_features_from_host(cpu, true); + + /* generate SYSREG properties according to writable masks */ + expose_idreg_properties(obj, cpu->writable_map, arm64_id_regs); +} + +#endif + static const ARMCPUInfo aarch64_cpus[] = { { .name = "cortex-a57", .initfn = aarch64_a57_initfn }, { .name = "cortex-a53", .initfn = aarch64_a53_initfn }, @@ -749,6 +899,9 @@ static const ARMCPUInfo aarch64_cpus[] = { #if defined(CONFIG_KVM) || defined(CONFIG_HVF) { .name = "host", .initfn = aarch64_host_initfn }, #endif +#ifdef CONFIG_KVM + { .name = "custom", .initfn = aarch64_customcpu_initfn }, +#endif }; static bool aarch64_cpu_get_aarch64(Object *obj, Error **errp) diff --git a/target/arm/trace-events b/target/arm/trace-events index 668acf94ab..1b4bd5ab14 100644 --- a/target/arm/trace-events +++ b/target/arm/trace-events @@ -15,3 +15,9 @@ arm_gt_update_irq(int timer, int irqstate) "gt_update_irq: timer %d irqstate %d" kvm_arm_fixup_msi_route(uint64_t iova, uint64_t gpa) "MSI iova = 0x%"PRIx64" is translated into 0x%"PRIx64 get_host_cpu_idregs(const char *name, uint64_t value) "scratch vcpu gost value for %s is 0x%"PRIx64 kvm_arm_writable_idregs_to_cpreg_list(const char *name, uint64_t previous, uint64_t new) "%s overwrite default 0x%"PRIx64" with 0x%"PRIx64 + +# cpu64.c +decode_idreg_writemap(const char* name, int lower, int upper, char *prop_name) "%s [%d:%d] is writable (prop %s)" +get_sysreg_prop(const char *name, uint64_t value) "%s 0x%"PRIx64 +set_sysreg_prop(const char *name, uint64_t old, uint64_t mask, uint64_t field_value, uint64_t new) "%s old reg value=0x%"PRIx64" mask=0x%"PRIx64" new field value=0x%"PRIx64" new reg value=0x%"PRIx64 +nb_sysreg_props(const char *name, int count) "%s: %d SYSREG properties"