- 12 Dec, 2020 4 commits
-
-
Kyung Min Park authored
Enumerate AVX512 Half-precision floating point (FP16) CPUID feature flag. Compared with using FP32, using FP16 cut the number of bits required for storage in half, reducing the exponent from 8 bits to 5, and the mantissa from 23 bits to 10. Using FP16 also enables developers to train and run inference on deep learning models fast when all precision or magnitude (FP32) is not needed. A processor supports AVX512 FP16 if CPUID.(EAX=7,ECX=0):EDX[bit 23] is present. The AVX512 FP16 requires AVX512BW feature be implemented since the instructions for manipulating 32bit masks are associated with AVX512BW. The only in-kernel usage of this is kvm passthrough. The CPU feature flag is shown as "avx512_fp16" in /proc/cpuinfo. Signed-off-by:
Kyung Min Park <kyung.min.park@intel.com> Acked-by:
Dave Hansen <dave.hansen@intel.com> Reviewed-by:
Tony Luck <tony.luck@intel.com> Message-Id: <20201208033441.28207-2-kyung.min.park@intel.com> Acked-by:
Borislav Petkov <bp@suse.de> Signed-off-by:
Paolo Bonzini <pbonzini@redhat.com>
-
Aaron Lewis authored
Both user_msr_test and userspace_msr_exit_test tests the functionality of kvm_msr_filter. Instead of testing this feature in two tests, merge them together, so there is only one test for this feature. Signed-off-by:
Aaron Lewis <aaronlewis@google.com> Message-Id: <20201204172530.2958493-1-aaronlewis@google.com> Signed-off-by:
-
Aaron Lewis authored
Add a selftest to test that when the ioctl KVM_X86_SET_MSR_FILTER is called with an MSR list, those MSRs exit to userspace. This test uses 3 MSRs to test this: 1. MSR_IA32_XSS, an MSR the kernel knows about. 2. MSR_IA32_FLUSH_CMD, an MSR the kernel does not know about. 3. MSR_NON_EXISTENT, an MSR invented in this test for the purposes of passing a fake MSR from the guest to userspace. KVM just acts as a pass through. Userspace is also able to inject a #GP. This is demonstrated when MSR_IA32_XSS and MSR_IA32_FLUSH_CMD are misused in the test. When this happens a #GP is initiated in userspace to be thrown in the guest which is handled gracefully by the exception handling framework introduced earlier in this series. Tests for the generic instruction emulator were also added. For this to work the module parameter kvm.force_emulation_prefix=1 has to be enabled. If it isn't enabled the tests will be skipped. A test was also added to ensure the MSR permission bitmap is being set correctly by executing reads and writes of MSR_FS_BASE and MSR_GS_BASE in the guest while alternating which MSR userspace should intercept. If the permission bitmap is being set correctly only one of the MSRs should be coming through at a time, and the guest should be able to read and write the other one directly. Signed-off-by:Alexander Graf <graf@amazon.com> Message-Id: <20201012194716.3950330-5-aaronlewis@google.com> Signed-off-by:
-
Uros Bizjak authored
Saves one byte in __vmx_vcpu_run for the same functionality. Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Sean Christopherson <sean.j.christopherson@intel.com> Signed-off-by:
Uros Bizjak <ubizjak@gmail.com> Message-Id: <20201029140457.126965-1-ubizjak@gmail.com> Signed-off-by:
-
- 09 Dec, 2020 1 commit
-
-
Maxim Levitsky authored
In the commit 1c96dcce ("KVM: x86: fix apic_accept_events vs check_nested_events"), we accidently started latching SIPIs that are received while the cpu is not waiting for them. This causes vCPUs to never enter a halted state. Fixes: 1c96dcce ("KVM: x86: fix apic_accept_events vs check_nested_events") Signed-off-by:
Maxim Levitsky <mlevitsk@redhat.com> Message-Id: <20201203143319.159394-2-mlevitsk@redhat.com> Signed-off-by:
-
- 03 Dec, 2020 1 commit
-
-
Paolo Bonzini authored
Since the ASID is now stored in svm->asid, pre_sev_run should also place it there and not directly in the VMCB control area. Reported-by:
Ashish Kalra <Ashish.Kalra@amd.com> Signed-off-by:
-
- 27 Nov, 2020 1 commit
-
-
Paolo Bonzini authored
SVM generally ignores fixed-1 bits. Set them manually so that we do not end up by mistake without those bits set in struct kvm_vcpu; it is part of userspace API that KVM always returns value with the bits set. Signed-off-by:
-
- 19 Nov, 2020 2 commits
-
-
Ben Gardon authored
Add an extremely verbose trace point to the TDP MMU to log all SPTE changes, regardless of callstack / motivation. This is useful when a complete picture of the paging structure is needed or a change cannot be explained with the other, existing trace points. Tested: ran the demand paging selftest on an Intel Skylake machine with all the trace points used by the TDP MMU enabled and observed them firing with expected values. This patch can be viewed in Gerrit at: https://linux-review.googlesource.com/c/virt/kvm/kvm/+/3813Signed-off-by:
Ben Gardon <bgardon@google.com> Message-Id: <20201027175944.1183301-2-bgardon@google.com> Signed-off-by:
-
Ben Gardon authored
The TDP MMU was initially implemented without some of the usual tracepoints found in mmu.c. Correct this discrepancy by adding the missing trace points to the TDP MMU. Tested: ran the demand paging selftest on an Intel Skylake machine with all the trace points used by the TDP MMU enabled and observed them firing with expected values. This patch can be viewed in Gerrit at: https://linux-review.googlesource.com/c/virt/kvm/kvm/+/3812Signed-off-by:
-
- 16 Nov, 2020 5 commits
-
-
Paolo Bonzini authored
Similarly to what vmx/vmx.c does, use vcpu->arch.cr4 to check if CR4 bits PGE, PKE and OSXSAVE have changed. When switching between VMCB01 and VMCB02, CPUID has to be adjusted every time if CR4.PKE or CR4.OSXSAVE change; without this patch, instead, CR4 would be checked against the previous value for L2 on vmentry, and against the previous value for L1 on vmexit, and CPUID would not be updated. Signed-off-by: -
Cathy Avery authored
KVM does not have separate ASIDs for L1 and L2; either the nested hypervisor and nested guests share a single ASID, or on older processor the ASID is used only to implement TLB flushing. Either way, ASIDs are handled at the VM level. In preparation for having different VMCBs passed to VMLOAD/VMRUN/VMSAVE for L1 and L2, store the current ASID to struct vcpu_svm and only move it to the VMCB in svm_vcpu_run. This way, TLB flushes can be applied no matter which VMCB will be active during the next svm_vcpu_run. Signed-off-by:
Cathy Avery <cavery@redhat.com> Message-Id: <20201011184818.3609-2-cavery@redhat.com> Signed-off-by:
-
Alex Shi authored
This macro is useless, and could cause gcc warning: arch/x86/kernel/kvmclock.c:47:0: warning: macro "HV_CLOCK_SIZE" is not used [-Wunused-macros] Let's remove it. Signed-off-by:
Alex Shi <alex.shi@linux.alibaba.com> Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Sean Christopherson <sean.j.christopherson@intel.com> Cc: Vitaly Kuznetsov <vkuznets@redhat.com> Cc: Wanpeng Li <wanpengli@tencent.com> Cc: Jim Mattson <jmattson@google.com> Cc: Joerg Roedel <joro@8bytes.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Borislav Petkov <bp@alien8.de> Cc: x86@kernel.org Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: kvm@vger.kernel.org Cc: linux-kernel@vger.kernel.org Message-Id: <1604651963-10067-1-git-send-email-alex.shi@linux.alibaba.com> Signed-off-by:
-
Andrew Jones authored
Almost all tests do this anyway and the ones that don't don't appear to care. Only vmx_set_nested_state_test assumes that a feature (VMX) is disabled until later setting the supported CPUIDs. It's better to disable that explicitly anyway. Signed-off-by:
Andrew Jones <drjones@redhat.com> Message-Id: <20201111122636.73346-11-drjones@redhat.com> [Restore CPUID_VMX, or vmx_set_nested_state breaks. - Paolo] Signed-off-by:
-
Andrew Jones authored
Signed-off-by:
-
- 15 Nov, 2020 26 commits
-
-
Andrew Jones authored
Signed-off-by:
-
Andrew Jones authored
Introduce new vm_create variants that also takes a number of vcpus, an amount of per-vcpu pages, and optionally a list of vcpuids. These variants will create default VMs with enough additional pages to cover the vcpu stacks, per-vcpu pages, and pagetable pages for all. The new 'default' variant uses VM_MODE_DEFAULT, whereas the other new variant accepts the mode as a parameter. Reviewed-by:
Peter Xu <peterx@redhat.com> Reviewed-by:
-
Andrew Jones authored
The code is almost 100% the same anyway. Just move it to common and add a few arch-specific macros. Reviewed-by:
-
Paolo Bonzini authored
Nothing sets USE_CLEAR_DIRTY_LOG anymore, so anything it surrounds is dead code. However, it is the recommended way to use the dirty page bitmap for new enough kernel, so use it whenever KVM has the KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 capability. Signed-off-by: -
Jim Mattson authored
On emulated VM-entry and VM-exit, update the CPUID bits that reflect CR4.OSXSAVE and CR4.PKE. This fixes a bug where the CPUID bits could continue to reflect L2 CR4 values after emulated VM-exit to L1. It also fixes a related bug where the CPUID bits could continue to reflect L1 CR4 values after emulated VM-entry to L2. The latter bug is mainly relevant to SVM, wherein CPUID is not a required intercept. However, it could also be relevant to VMX, because the code to conditionally update these CPUID bits assumes that the guest CPUID and the guest CR4 are always in sync. Fixes: 8eb3f87d ("KVM: nVMX: fix guest CR4 loading when emulating L2 to L1 exit") Fixes: 2acf923e ("KVM: VMX: Enable XSAVE/XRSTOR for guest") Fixes: b9baba86 ("KVM, pkeys: expose CPUID/CR4 to guest") Reported-by:
Abhiroop Dabral <adabral@paloaltonetworks.com> Signed-off-by:
Jim Mattson <jmattson@google.com> Reviewed-by:
Ricardo Koller <ricarkol@google.com> Reviewed-by:
Peter Shier <pshier@google.com> Cc: Haozhong Zhang <haozhong.zhang@intel.com> Cc: Dexuan Cui <dexuan.cui@intel.com> Cc: Huaitong Han <huaitong.han@intel.com> Message-Id: <20201029170648.483210-1-jmattson@google.com>
-
Paolo Bonzini authored
Add tsc_msrs_test, remove clear_dirty_log_test and alphabetize everything. Signed-off-by: -
Peter Xu authored
It's only used to override the existing dirty ring size/count. If with a bigger ring count, we test async of dirty ring. If with a smaller ring count, we test ring full code path. Async is default. It has no use for non-dirty-ring tests. Reviewed-by:
-
Peter Xu authored
Previously the dirty ring test was working in synchronous way, because only with a vmexit (with that it was the ring full event) we'll know the hardware dirty bits will be flushed to the dirty ring. With this patch we first introduce a vcpu kick mechanism using SIGUSR1, which guarantees a vmexit and also therefore the flushing of hardware dirty bits. Once this is in place, we can keep the vcpu dirty work asynchronous of the whole collection procedure now. Still, we need to be very careful that when reaching the ring buffer soft limit (KVM_EXIT_DIRTY_RING_FULL) we must collect the dirty bits before continuing the vcpu. Further increase the dirty ring size to current maximum to make sure we torture more on the no-ring-full case, which should be the major scenario when the hypervisors like QEMU would like to use this feature. Reviewed-by:
-
Peter Xu authored
Add the initial dirty ring buffer test. The current test implements the userspace dirty ring collection, by only reaping the dirty ring when the ring is full. So it's still running synchronously like this: vcpu main thread 1. vcpu dirties pages 2. vcpu gets dirty ring full (userspace exit) 3. main thread waits until full (so hardware buffers flushed) 4. main thread collects 5. main thread continues vcpu 6. vcpu continues, goes back to 1 We can't directly collects dirty bits during vcpu execution because otherwise we can't guarantee the hardware dirty bits were flushed when we collect and we're very strict on the dirty bits so otherwise we can fail the future verify procedure. A follow up patch will make this test to support async just like the existing dirty log test, by adding a vcpu kick mechanism. Signed-off-by: -
Peter Xu authored
Provide a hook for the checks after vcpu_run() completes. Preparation for the dirty ring test because we'll need to take care of another exit reason. Reviewed-by:
-
Peter Xu authored
Because kvm dirty rings and kvm dirty log is used in an exclusive way, Let's avoid creating the dirty_bitmap when kvm dirty ring is enabled. At the meantime, since the dirty_bitmap will be conditionally created now, we can't use it as a sign of "whether this memory slot enabled dirty tracking". Change users like that to check against the kvm memory slot flags. Note that there still can be chances where the kvm memory slot got its dirty_bitmap allocated, _if_ the memory slots are created before enabling of the dirty rings and at the same time with the dirty tracking capability enabled, they'll still with the dirty_bitmap. However it should not hurt much (e.g., the bitmaps will always be freed if they are there), and the real users normally won't trigger this because dirty bit tracking flag should in most cases only be applied to kvm slots only before migration starts, that should be far latter than kvm initializes (VM starts). Signed-off-by:
-
Peter Xu authored
There's no good reason to use both the dirty bitmap logging and the new dirty ring buffer to track dirty bits. We should be able to even support both of them at the same time, but it could complicate things which could actually help little. Let's simply make it the rule before we enable dirty ring on any arch, that we don't allow these two interfaces to be used together. The big world switch would be KVM_CAP_DIRTY_LOG_RING capability enablement. That's where we'll switch from the default dirty logging way to the dirty ring way. As long as kvm->dirty_ring_size is setup correctly, we'll once and for all switch to the dirty ring buffer mode for the current virtual machine. Signed-off-by:
-
Peter Xu authored
This patch is heavily based on previous work from Lei Cao <lei.cao@stratus.com> and Paolo Bonzini <pbonzini@redhat.com>. [1] KVM currently uses large bitmaps to track dirty memory. These bitmaps are copied to userspace when userspace queries KVM for its dirty page information. The use of bitmaps is mostly sufficient for live migration, as large parts of memory are be dirtied from one log-dirty pass to another. However, in a checkpointing system, the number of dirty pages is small and in fact it is often bounded---the VM is paused when it has dirtied a pre-defined number of pages. Traversing a large, sparsely populated bitmap to find set bits is time-consuming, as is copying the bitmap to user-space. A similar issue will be there for live migration when the guest memory is huge while the page dirty procedure is trivial. In that case for each dirty sync we need to pull the whole dirty bitmap to userspace and analyse every bit even if it's mostly zeros. The preferred data structure for above scenarios is a dense list of guest frame numbers (GFN). This patch series stores the dirty list in kernel memory that can be memory mapped into userspace to allow speedy harvesting. This patch enables dirty ring for X86 only. However it should be easily extended to other archs as well. [1] https://patchwork.kernel.org/patch/10471409/Signed-off-by:
Lei Cao <lei.cao@stratus.com> Signed-off-by:
-
Peter Xu authored
The context will be needed to implement the kvm dirty ring. Signed-off-by:
-
Paolo Bonzini authored
kvm_clear_guest_page is not used anymore after "KVM: X86: Don't track dirty for KVM_SET_[TSS_ADDR|IDENTITY_MAP_ADDR]", except from kvm_clear_guest. We can just inline it in its sole user. Signed-off-by: -
Peter Xu authored
Originally, we have three code paths that can dirty a page without vcpu context for X86: - init_rmode_identity_map - init_rmode_tss - kvmgt_rw_gpa init_rmode_identity_map and init_rmode_tss will be setup on destination VM no matter what (and the guest cannot even see them), so it does not make sense to track them at all. To do this, allow __x86_set_memory_region() to return the userspace address that just allocated to the caller. Then in both of the functions we directly write to the userspace address instead of calling kvm_write_*() APIs. Another trivial change is that we don't need to explicitly clear the identity page table root in init_rmode_identity_map() because no matter what we'll write to the whole page with 4M huge page entries. Suggested-by:
Sean Christopherson <sean.j.christopherson@intel.com> Signed-off-by:
-
Vitaly Kuznetsov authored
KVM_GET_SUPPORTED_HV_CPUID is now supported as both vCPU and VM ioctl, test that. Signed-off-by:
Vitaly Kuznetsov <vkuznets@redhat.com> Message-Id: <20200929150944.1235688-3-vkuznets@redhat.com> Signed-off-by:
-
Vitaly Kuznetsov authored
KVM_GET_SUPPORTED_HV_CPUID is a vCPU ioctl but its output is now independent from vCPU and in some cases VMMs may want to use it as a system ioctl instead. In particular, QEMU doesn CPU feature expansion before any vCPU gets created so KVM_GET_SUPPORTED_HV_CPUID can't be used. Convert KVM_GET_SUPPORTED_HV_CPUID to 'dual' system/vCPU ioctl with the same meaning. Signed-off-by:
-
David Woodhouse authored
Don't allow the events to accumulate in the eventfd counter, drain them as they are handled. Signed-off-by:
David Woodhouse <dwmw@amazon.co.uk> Message-Id: <20201027135523.646811-4-dwmw2@infradead.org> Signed-off-by:
-
David Woodhouse authored
Don't allow the events to accumulate in the eventfd counter, drain them as they are handled. Signed-off-by:
Alex Williamson <alex.williamson@redhat.com>
-
David Woodhouse authored
Where events are consumed in the kernel, for example by KVM's irqfd_wakeup() and VFIO's virqfd_wakeup(), they currently lack a mechanism to drain the eventfd's counter. Since the wait queue is already locked while the wakeup functions are invoked, all they really need to do is call eventfd_ctx_do_read(). Add a check for the lock, and export it for them. Signed-off-by:
-
David Woodhouse authored
As far as I can tell, when we use posted interrupts we silently cut off the events from userspace, if it's listening on the same eventfd that feeds the irqfd. I like that behaviour. Let's do it all the time, even without posted interrupts. It makes it much easier to handle IRQ remapping invalidation without having to constantly add/remove the fd from the userspace poll set. We can just leave userspace polling on it, and the bypass will... well... bypass it. Signed-off-by:
-
David Woodhouse authored
This allows an exclusive wait_queue_entry to be added at the head of the queue, instead of the tail as normal. Thus, it gets to consume events first without allowing non-exclusive waiters to be woken at all. The (first) intended use is for KVM IRQFD, which currently has inconsistent behaviour depending on whether posted interrupts are available or not. If they are, KVM will bypass the eventfd completely and deliver interrupts directly to the appropriate vCPU. If not, events are delivered through the eventfd and userspace will receive them when polling on the eventfd. By using add_wait_queue_priority(), KVM will be able to consistently consume events within the kernel without accidentally exposing them to userspace when they're supposed to be bypassed. This, in turn, means that userspace doesn't have to jump through hoops to avoid listening on the erroneously noisy eventfd and injecting duplicate interrupts. Signed-off-by:
-
Yadong Qi authored
Background: We have a lightweight HV, it needs INIT-VMExit and SIPI-VMExit to wake-up APs for guests since it do not monitor the Local APIC. But currently virtual wait-for-SIPI(WFS) state is not supported in nVMX, so when running on top of KVM, the L1 HV cannot receive the INIT-VMExit and SIPI-VMExit which cause the L2 guest cannot wake up the APs. According to Intel SDM Chapter 25.2 Other Causes of VM Exits, SIPIs cause VM exits when a logical processor is in wait-for-SIPI state. In this patch: 1. introduce SIPI exit reason, 2. introduce wait-for-SIPI state for nVMX, 3. advertise wait-for-SIPI support to guest. When L1 hypervisor is not monitoring Local APIC, L0 need to emulate INIT-VMExit and SIPI-VMExit to L1 to emulate INIT-SIPI-SIPI for L2. L2 LAPIC write would be traped by L0 Hypervisor(KVM), L0 should emulate the INIT/SIPI vmexit to L1 hypervisor to set proper state for L2's vcpu state. Handle procdure: Source vCPU: L2 write LAPIC.ICR(INIT). L0 trap LAPIC.ICR write(INIT): inject a latched INIT event to target vCPU. Target vCPU: L0 emulate an INIT VMExit to L1 if is guest mode. L1 set guest VMCS, guest_activity_state=WAIT_SIPI, vmresume. L0 set vcpu.mp_state to INIT_RECEIVED if (vmcs12.guest_activity_state == WAIT_SIPI). Source vCPU: L2 write LAPIC.ICR(SIPI). L0 trap LAPIC.ICR write(INIT): inject a latched SIPI event to traget vCPU. Target vCPU: L0 emulate an SIPI VMExit to L1 if (vcpu.mp_state == INIT_RECEIVED). L1 set CS:IP, guest_activity_state=ACTIVE, vmresume. L0 resume to L2. L2 start-up. Signed-off-by:Yadong Qi <yadong.qi@intel.com> Message-Id: <20200922052343.84388-1-yadong.qi@intel.com> Signed-off-by:
-
Paolo Bonzini authored
vmx_apic_init_signal_blocked is buggy in that it returns true even in VMX non-root mode. In non-root mode, however, INITs are not latched, they just cause a vmexit. Previously, KVM was waiting for them to be processed when kvm_apic_accept_events and in the meanwhile it ate the SIPIs that the processor received. However, in order to implement the wait-for-SIPI activity state, KVM will have to process KVM_APIC_SIPI in vmx_check_nested_events, and it will not be possible anymore to disregard SIPIs in non-root mode as the code is currently doing. By calling kvm_x86_ops.nested_ops->check_events, we can force a vmexit (with the side-effect of latching INITs) before incorrectly injecting an INIT or SIPI in a guest, and therefore vmx_apic_init_signal_blocked can do the right thing. Signed-off-by: -
Sean Christopherson authored
Extend the KVM_SET_SREGS test to verify that all supported CR4 bits, as enumerated by KVM, can be set before KVM_SET_CPUID2, i.e. without first defining the vCPU model. KVM is supposed to skip guest CPUID checks when host userspace is stuffing guest state. Check the inverse as well, i.e. that KVM rejects KVM_SET_REGS if CR4 has one or more unsupported bits set. Signed-off-by:
-
