- 16 Mar, 2020 40 commits
-
-
Wanpeng Li authored
PMU is not exposed to guest by most of products from cloud providers since the bad performance of PMU emulation and security concern. However, it calls perf_guest_switch_get_msrs() and clear_atomic_switch_msr() unconditionally even if PMU is not exposed to the guest before each vmentry. ~2% vmexit time reduced can be observed by kvm-unit-tests/vmexit.flat on my SKX server. Before patch: vmcall 1559 After patch: vmcall 1529 Signed-off-by:
Wanpeng Li <wanpengli@tencent.com> Signed-off-by:
Paolo Bonzini <pbonzini@redhat.com>
-
Andrew Jones authored
s390 requires 1M aligned guest sizes. Embedding the rounding in vm_adjust_num_guest_pages() allows us to remove it from a few other places. Signed-off-by:
Andrew Jones <drjones@redhat.com> Signed-off-by:
-
Suravee Suthikulpanit authored
GA Log tracepoint is useful when debugging AVIC performance issue as it can be used with perf to count the number of times IOMMU AVIC injects interrupts through the slow-path instead of directly inject interrupts to the target vcpu. Signed-off-by:
Suravee Suthikulpanit <suravee.suthikulpanit@amd.com> Signed-off-by:
-
Peter Xu authored
Clarify locking.rst to mention early that we're not enabling fast page fault for indirect sps. The previous wording is confusing, in that it seems the proposed solution has been already implemented but it has not. Signed-off-by:
Peter Xu <peterx@redhat.com> Signed-off-by:
-
Paolo Bonzini authored
This patch reproduces for nSVM the change that was made for nVMX in commit b5861e5c ("KVM: nVMX: Fix loss of pending IRQ/NMI before entering L2"). While I do not have a test that breaks without it, I cannot see why it would not be necessary since all events are unblocked by VMRUN's setting of GIF back to 1. Signed-off-by:
-
Paolo Bonzini authored
The current implementation of physical interrupt delivery to a nested guest is quite broken. It relies on svm_interrupt_allowed returning false if VINTR=1 so that the interrupt can be injected from enable_irq_window, but this does not work for guests that do not intercept HLT or that rely on clearing the host IF to block physical interrupts while L2 runs. This patch can be split in two logical parts, but including only one breaks tests so I am combining both changes together. The first and easiest is simply to return true for svm_interrupt_allowed if HF_VINTR_MASK is set and HIF is set. This way the semantics of svm_interrupt_allowed are respected: svm_interrupt_allowed being false does not mean "call enable_irq_window", it means "interrupts cannot be injected now". After doing this, however, we need another place to inject the interrupt, and fortunately we already have one, check_nested_events, which nested SVM does not implement but which is meant exactly for this purpose. It is called before interrupts are injected, and it can therefore do the L2->L1 switch while leaving inject_pending_event none the wiser. This patch was developed together with Cathy Avery, who wrote the test and did a lot of the initial debugging. Signed-off-by: -
Paolo Bonzini authored
If a nested VM is started while an IRQ was pending and with V_INTR_MASKING=1, the behavior of the guest depends on host IF. If it is 1, the VM should exit immediately, before executing the first instruction of the guest, because VMRUN sets GIF back to 1. If it is 0 and the host has VGIF, however, at the time of the VMRUN instruction L0 is running the guest with a pending interrupt window request. This interrupt window request is completely irrelevant to L2, since IF only controls virtual interrupts, so this patch drops INTERCEPT_VINTR from the VMCB while running L2 under these circumstances. To simplify the code, both steps of enabling the interrupt window (setting the VINTR intercept and requesting a fake virtual interrupt in svm_inject_irq) are grouped in the svm_set_vintr function, and likewise for dismissing the interrupt window request in svm_clear_vintr. Signed-off-by: -
Paolo Bonzini authored
Instead of touching the host intercepts so that the bitwise OR in recalc_intercepts just works, mask away uninteresting intercepts directly in recalc_intercepts. This is cleaner and keeps the logic in one place even for intercepts that can change even while L2 is running. Signed-off-by: -
Paolo Bonzini authored
The set_cr3 callback is not setting the guest CR3, it is setting the root of the guest page tables, either shadow or two-dimensional. To make this clearer as well as to indicate that the MMU calls it via kvm_mmu_load_cr3, rename it to load_mmu_pgd. Signed-off-by: -
Paolo Bonzini authored
Similar to what kvm-intel.ko is doing, provide a single callback that merges svm_set_cr3, set_tdp_cr3 and nested_svm_set_tdp_cr3. This lets us unify the set_cr3 and set_tdp_cr3 entries in kvm_x86_ops. I'm doing that in this same patch because splitting it adds quite a bit of churn due to the need for forward declarations. For the same reason the assignment to vcpu->arch.mmu->set_cr3 is moved to kvm_init_shadow_mmu from init_kvm_softmmu and nested_svm_init_mmu_context. Signed-off-by: -
Sean Christopherson authored
Invert and rename the kvm_cpuid() param that controls out-of-range logic to better reflect the semantics of the affected callers, i.e. callers that bypass the out-of-range logic do so because they are looking up an exact guest CPUID entry, e.g. to query the maxphyaddr. Similarly, rename kvm_cpuid()'s internal "found" to "exact" to clarify that it tracks whether or not the exact requested leaf was found, as opposed to any usable leaf being found. No functional change intended. Signed-off-by:
Sean Christopherson <sean.j.christopherson@intel.com> Signed-off-by:
-
Sean Christopherson authored
Move all of the out-of-range logic into a single helper, get_out_of_range_cpuid_entry(), to avoid an extra lookup of CPUID.0.0 and to provide a single location for documenting the out-of-range behavior. No functional change intended. Cc: Jim Mattson <jmattson@google.com> Signed-off-by:
-
Sean Christopherson authored
Rework the masking in the out-of-range CPUID logic to handle the Hypervisor sub-classes, as well as the Centaur class if the guest virtual CPU vendor is Centaur. Masking against 0x80000000 only handles basic and extended leafs, which results in Hypervisor range checks being performed against the basic CPUID class, and Centuar range checks being performed against the Extended class. E.g. if CPUID.0x40000000.EAX returns 0x4000000A and there is no entry for CPUID.0x40000006, then function 0x40000006 would be incorrectly reported as out of bounds. While there is no official definition of what constitutes a class, the convention established for Hypervisor classes effectively uses bits 31:8 as the mask by virtue of checking for different bases in increments of 0x100, e.g. KVM advertises its CPUID functions starting at 0x40000100 when HyperV features are advertised at the default base of 0x40000000. The bad range check doesn't cause functional problems for any known VMM because out-of-range semantics only come into play if the exact entry isn't found, and VMMs either support a very limited Hypervisor range, e.g. the official KVM range is 0x40000000-0x40000001 (effectively no room for undefined leafs) or explicitly defines gaps to be zero, e.g. Qemu explicitly creates zeroed entries up to the Centaur and Hypervisor limits (the latter comes into play when providing HyperV features). The bad behavior can be visually confirmed by dumping CPUID output in the guest when running Qemu with a stable TSC, as Qemu extends the limit of range 0x40000000 to 0x40000010 to advertise VMware's cpuid_freq, without defining zeroed entries for 0x40000002 - 0x4000000f. Note, documentation of Centaur/VIA CPUs is hard to come by. Designating 0xc0000000 - 0xcfffffff as the Centaur class is a best guess as to the behavior of a real Centaur/VIA CPU. Fixes: 43561123 ("kvm: x86: Improve emulation of CPUID leaves 0BH and 1FH") Cc: Jim Mattson <jmattson@google.com> Signed-off-by:
-
Sean Christopherson authored
Extend guest_cpuid_is_amd() to cover Hygon virtual CPUs and rename it accordingly. Hygon CPUs use an AMD-based core and so have the same basic behavior as AMD CPUs. Fixes: b8f4abb6 ("x86/kvm: Add Hygon Dhyana support to KVM") Cc: Pu Wen <puwen@hygon.cn> Signed-off-by:
-
Sean Christopherson authored
Add helpers to provide CPUID-based guest vendor checks, i.e. to do the ugly register comparisons. Use the new helpers to check for an AMD guest vendor in guest_cpuid_is_amd() as well as in the existing emulator flows. Using the new helpers fixes a _very_ theoretical bug where guest_cpuid_is_amd() would get a false positive on a non-AMD virtual CPU with a vendor string beginning with "Auth" due to the previous logic only checking EBX. It also fixes a marginally less theoretically bug where guest_cpuid_is_amd() would incorrectly return false for a guest CPU with "AMDisbetter!" as its vendor string. Fixes: a0c0feb5 ("KVM: x86: reserve bit 8 of non-leaf PDPEs and PML4Es in 64-bit mode on AMD") Signed-off-by:
-
Jan Kiszka authored
Trace the requested CPUID function instead of the effective function, e.g. if the requested function is out-of-range and KVM is emulating an Intel CPU, as the intent of the tracepoint is to show if the output came from the actual leaf as opposed to the max basic leaf via redirection. Similarly, leave "found" as is, i.e. report that an entry was found if and only if the requested entry was found. Fixes: 43561123 ("kvm: x86: Improve emulation of CPUID leaves 0BH and 1FH") Signed-off-by:
Jan Kiszka <jan.kiszka@siemens.com> [Sean: Drop "found" semantic change, reword changelong accordingly ] Signed-off-by:
-
Paolo Bonzini authored
Current CPUID 0xd enumeration code does not support supervisor states, because KVM only supports setting IA32_XSS to zero. Change it instead to use a new variable supported_xss, to be set from the hardware_setup callback which is in charge of CPU capabilities. Signed-off-by: -
Sean Christopherson authored
Handle CPUID 0x8000000A in the main switch in __do_cpuid_func() and drop ->set_supported_cpuid() now that both VMX and SVM implementations are empty. Like leaf 0x14 (Intel PT) and leaf 0x8000001F (SEV), leaf 0x8000000A is is (obviously) vendor specific but can be queried in common code while respecting SVM's wishes by querying kvm_cpu_cap_has(). Suggested-by:
Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by:
-
Sean Christopherson authored
Set NRIPS in KVM capabilities if and only if nrips=true, which naturally incorporates the boot_cpu_has() check, and set nrips_enabled only if the KVM capability is enabled. Note, previously KVM would set nrips_enabled based purely on userspace input, but at worst that would cause KVM to propagate garbage into L1, i.e. userspace would simply be hosing its VM. Signed-off-by:
-
Sean Christopherson authored
Set SVM feature bits in KVM capabilities if and only if nested=true, KVM shouldn't advertise features that realistically can't be used. Use kvm_cpu_cap_has(X86_FEATURE_SVM) to indirectly query "nested" in svm_set_supported_cpuid() in anticipation of moving CPUID 0x8000000A adjustments into common x86 code. Suggested-by:
-
Sean Christopherson authored
Move host_efer to common x86 code and use it for CPUID's is_efer_nx() to avoid constantly re-reading the MSR. No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Drop largepages_enabled, kvm_largepages_enabled() and kvm_disable_largepages() now that all users are gone. Note, largepages_enabled was an x86-only flag that got left in common KVM code when KVM gained support for multiple architectures. No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Stop propagating MMU large page support into a memslot's disallow_lpage now that the MMU's max_page_level handles the scenario where VMX's EPT is enabled and EPT doesn't support 2M pages. No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Configure the max page level during hardware setup to avoid a retpoline in the page fault handler. Drop ->get_lpage_level() as the page fault handler was the last user. No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Combine kvm_enable_tdp() and kvm_disable_tdp() into a single function, kvm_configure_mmu(), in preparation for doing additional configuration during hardware setup. And because having separate helpers is silly. No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Tweak SVM's logging of NPT enabled/disabled to handle the logging in a single pr_info() in preparation for merging kvm_enable_tdp() and kvm_disable_tdp() into a single function. No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Use vmx_pt_mode_is_host_guest() in intel_pmu_refresh() instead of bouncing through kvm_x86_ops->pt_supported, and remove ->pt_supported() as the PMU code was the last remaining user. Opportunistically clean up the wording of a comment that referenced kvm_x86_ops->pt_supported(). No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Use kvm_cpu_cap_has() to check for Intel PT when processing the list of virtualized MSRs to pave the way toward removing ->pt_supported(). No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Use cpu_has_vmx_rdtscp() directly when computing secondary exec controls and drop the now defunct vmx_rdtscp_supported(). No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Check for MSR_TSC_AUX virtualization via kvm_cpu_cap_has() and drop ->rdtscp_supported(). Note, vmx_rdtscp_supported() needs to hang around a tiny bit longer due other usage in VMX code. No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Perform the capacity checks on the userspace provided kvm_cpuid_array in the lower __do_cpuid_func() and __do_cpuid_func_emulated(). Pre-checking the array in do_cpuid_func() no longer adds value now that __do_cpuid_func() has been trimmed down to size, i.e. doesn't invoke a big pile of retpolined functions before doing anything useful. Note, __do_cpuid_func() already checks the array capacity via do_host_cpuid(), "moving" the check to __do_cpuid_func() simply means removing a WARN_ON(). Suggested-by:
-
Sean Christopherson authored
Check for Intel PT using kvm_cpu_cap_has() to pave the way toward eliminating ->pt_supported(). No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Set emulated and transmuted (set based on other features) feature bits via kvm_cpu_caps now that the CPUID output for KVM_GET_SUPPORTED_CPUID is direcly overidden with kvm_cpu_caps. Note, VMX emulation of UMIP already sets kvm_cpu_caps. No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Override CPUID entries with kvm_cpu_caps during KVM_GET_SUPPORTED_CPUID instead of masking the host CPUID result, which is redundant now that the host CPUID is incorporated into kvm_cpu_caps at runtime. No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Mask kvm_cpu_caps based on host CPUID in preparation for overriding the CPUID results during KVM_GET_SUPPORTED_CPUID instead of doing the masking at runtime. Note, masking may or may not be necessary, e.g. the kernel rarely, if ever, sets real CPUID bits that are not supported by hardware. But, the code is cheap and only runs once at load, so an abundance of caution is warranted. No functional change intended. Reviewed-by:
-
Sean Christopherson authored
Remove the code for handling stateful CPUID 0x2 and mark the associated flags as deprecated. WARN if host CPUID 0x2.0.AL > 1, i.e. if by some miracle a host with stateful CPUID 0x2 is encountered. No known CPU exists that supports hardware accelerated virtualization _and_ a stateful CPUID 0x2. Barring an extremely contrived nested virtualization scenario, stateful CPUID support is dead code. Suggested-by:
-
Sean Christopherson authored
Rework CPUID 0x2.0 to be a normal CPUID leaf if it returns "01" in AL, i.e. EAX & 0xff, as a step towards removing KVM's stateful CPUID code altogether. Long ago, Intel documented CPUID 0x2.0 as being a stateful leaf, e.g. a version of the SDM circa 1995 states: The least-significant byte in register EAX (register AL) indicates the number of times the CPUID instruction must be executed with an input value of 2 to get a complete description of the processors's caches and TLBs. The Pentium Pro family of processors will return a 1. A 2000 version of the SDM only updated the paragraph to reference Intel's new processory family: The first member of the family of Pentium 4 processors will return a 1. Fast forward to the present, and Intel's SDM now states: The least-significant byte in register EAX (register AL) will always return 01H. Software should ignore this value and not interpret it as an information descriptor. AMD's APM simply states that CPUID 0x2 is reserved. Given that CPUID itself was introduced in the Pentium, odds are good that the only Intel CPU family that *maybe* implemented a stateful CPUID was the P5. Which obviously did not support VMX, or KVM. In other words, KVM's emulation of a stateful CPUID 0x2.0 has likely been dead code from the day it was introduced. This is backed up by commit 0fdf8e59 ("KVM: Fix cpuid iteration on multiple leaves per eac"), which shows that the stateful iteration code was completely broken when it was introduced by commit 07716717 ("KVM: Enhance guest cpuid management"), i.e. not actually tested. Annotate all stateful code paths as "unlikely", but defer its removal to a future patch to simplify reinstating the code if by some miracle there is someone running KVM on a CPU with a stateful CPUID 0x2. Reviewed-by:
-
Sean Christopherson authored
Explicitly handle CPUID 0x7 sub-leaf 1. The kernel is currently aware of exactly one feature in CPUID 0x7.1, which means there is room for another 127 features before CPUID 0x7.2 will see the light of day, i.e. the looping is likely to be dead code for years to come. Reviewed-by:
-
Sean Christopherson authored
Move the CPUID 0x7 masking back into __do_cpuid_func() now that the size of the code has been trimmed down significantly. Tweak the WARN case, which is impossible to hit unless the CPU is completely broken, to break the loop before creating the bogus entry. Opportunustically reorder the cpuid_entry_set() calls and shorten the comment about emulation to further reduce the footprint of CPUID 0x7. Reviewed-by:
-
Sean Christopherson authored
Set UMIP in kvm_cpu_caps when it is emulated by VMX, even though the bit will effectively be dropped by do_host_cpuid(). This allows checking for UMIP emulation via kvm_cpu_caps instead of a dedicated kvm_x86_ops callback. No functional change intended. Reviewed-by:
-
