- 26 Sep, 2019 1 commit
-
-
Wanpeng Li authored
5000 guest cycles delta is easy to encounter on desktop, per-vCPU lapic_timer_advance_ns always keeps at 1000ns initial value, let's loosen the filter a bit to let adaptive tuning make progress. Signed-off-by:
Wanpeng Li <wanpengli@tencent.com> Signed-off-by:
Paolo Bonzini <pbonzini@redhat.com>
-
- 25 Sep, 2019 9 commits
-
-
Paolo Bonzini authored
KVM was incorrectly checking vmcs12->host_ia32_efer even if the "load IA32_EFER" exit control was reset. Also, some checks were not using the new CC macro for tracing. Cleanup everything so that the vCPU's 64-bit mode is determined directly from EFER_LMA and the VMCS checks are based on that, which matches section 26.2.4 of the SDM. Cc: Sean Christopherson <sean.j.christopherson@intel.com> Cc: Krish Sadhukhan <krish.sadhukhan@oracle.com> Fixes: 5845038cReviewed-by:
Jim Mattson <jmattson@google.com> Signed-off-by:
-
Vitaly Kuznetsov authored
The following was reported on i386: arch/x86/kvm/vmx/vmx.c: In function 'hv_enable_direct_tlbflush': arch/x86/kvm/vmx/vmx.c:503:10: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast] pr_debugs() in this function are more or less useless, let's just remove them. evmcs->hv_vm_id can use 'unsigned long' instead of 'u64'. Also, simplify the code a little bit. Reported-by:
kbuild test robot <lkp@intel.com> Signed-off-by:
Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by:
-
Sean Christopherson authored
Remove the kvm_rebooting check from VMX/SVM instruction exception fixup now that kvm_spurious_fault() conditions its BUG() on !kvm_rebooting. Because the 'cleanup_insn' functionally is also gone, deferring to kvm_spurious_fault() means __kvm_handle_fault_on_reboot() can eliminate its .fixup code entirely and have its exception table entry branch directly to the call to kvm_spurious_fault(). Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Signed-off-by:
Sean Christopherson <sean.j.christopherson@intel.com> Signed-off-by:
-
Sean Christopherson authored
Remove the variation of __kvm_handle_fault_on_reboot() that accepts a post-fault cleanup instruction now that its sole user (VMREAD) uses a different method for handling faults. Acked-by:
-
Sean Christopherson authored
Now that VMREAD flows require a taken branch, courtesy of commit 3901336e ("x86/kvm: Don't call kvm_spurious_fault() from .fixup") bite the bullet and add full error handling to VMREAD, i.e. replace the JMP added by __ex()/____kvm_handle_fault_on_reboot() with a hinted Jcc. To minimize the code footprint, add a helper function, vmread_error(), to handle both faults and failures so that the inline flow has a single CALL. Acked-by:
-
Sean Christopherson authored
Rework the VMX instruction helpers using asm-goto to branch directly to error/fault "handlers" in lieu of using __ex(), i.e. the generic ____kvm_handle_fault_on_reboot(). Branching directly to fault handling code during fixup avoids the extra JMP that is inserted after every VMX instruction when using the generic "fault on reboot" (see commit 3901336e, "x86/kvm: Don't call kvm_spurious_fault() from .fixup"). Opportunistically clean up the helpers so that they all have consistent error handling and messages. Leave the usage of ____kvm_handle_fault_on_reboot() (via __ex()) in kvm_cpu_vmxoff() and nested_vmx_check_vmentry_hw() as is. The VMXOFF case is not a fast path, i.e. the cleanliness of __ex() is worth the JMP, and the extra JMP in nested_vmx_check_vmentry_hw() is unavoidable. Note, VMREAD cannot get the asm-goto treatment as output operands aren't compatible with GCC's asm-goto due to internal compiler restrictions. Acked-by:
-
Sean Christopherson authored
Explicitly check kvm_rebooting in kvm_spurious_fault() prior to invoking BUG(), as opposed to assuming the caller has already done so. Letting kvm_spurious_fault() be called "directly" will allow VMX to better optimize its low level assembly flows. As a happy side effect, kvm_spurious_fault() no longer needs to be marked as a dead end since it doesn't unconditionally BUG(). Acked-by:
-
Vitaly Kuznetsov authored
After commit e8bb4755eea2("KVM: selftests: Split ucall.c into architecture specific files") selftests which use ucall on x86 started segfaulting and apparently it's gcc to blame: it "optimizes" ucall() function throwing away va_start/va_end part because it thinks the structure is not being used. Previously, it couldn't do that because the there was also MMIO version and the decision which particular implementation to use was done at runtime. With older gccs it's possible to solve the problem by adding 'volatile' to 'struct ucall' but at least with gcc-8.3 this trick doesn't work. 'memory' clobber seems to do the job. Signed-off-by: -
Wanpeng Li authored
This patch reverts commit 75437bb3 (locking/pvqspinlock: Don't wait if vCPU is preempted). A large performance regression was caused by this commit. on over-subscription scenarios. The test was run on a Xeon Skylake box, 2 sockets, 40 cores, 80 threads, with three VMs of 80 vCPUs each. The score of ebizzy -M is reduced from 13000-14000 records/s to 1700-1800 records/s: Host Guest score vanilla w/o kvm optimizations upstream 1700-1800 records/s vanilla w/o kvm optimizations revert 13000-14000 records/s vanilla w/ kvm optimizations upstream 4500-5000 records/s vanilla w/ kvm optimizations revert 14000-15500 records/s Exit from aggressive wait-early mechanism can result in premature yield and extra scheduling latency. Actually, only 6% of wait_early events are caused by vcpu_is_preempted() being true. However, when one vCPU voluntarily releases its vCPU, all the subsequently waiters in the queue will do the same and the cascading effect leads to bad performance. kvm optimizations: [1] commit d73eb57b (KVM: Boost vCPUs that are delivering interrupts) [2] commit 266e85a5 (KVM: X86: Boost queue head vCPU to mitigate lock waiter preemption) Tested-by: loobinliu@tencent.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@kernel.org> Cc: Waiman Long <longman@redhat.com> Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Radim Krčmář <rkrcmar@redhat.com> Cc: loobinliu@tencent.com Cc: stable@vger.kernel.org Fixes: 75437bb3 (locking/pvqspinlock: Don't wait if vCPU is preempted) Signed-off-by:
-
- 24 Sep, 2019 30 commits
-
-
Marc Orr authored
Allowing an unlimited number of MSRs to be specified via the VMX load/store MSR lists (e.g., vm-entry MSR load list) is bad for two reasons. First, a guest can specify an unreasonable number of MSRs, forcing KVM to process all of them in software. Second, the SDM bounds the number of MSRs allowed to be packed into the atomic switch MSR lists. Quoting the "Miscellaneous Data" section in the "VMX Capability Reporting Facility" appendix: "Bits 27:25 is used to compute the recommended maximum number of MSRs that should appear in the VM-exit MSR-store list, the VM-exit MSR-load list, or the VM-entry MSR-load list. Specifically, if the value bits 27:25 of IA32_VMX_MISC is N, then 512 * (N + 1) is the recommended maximum number of MSRs to be included in each list. If the limit is exceeded, undefined processor behavior may result (including a machine check during the VMX transition)." Because KVM needs to protect itself and can't model "undefined processor behavior", arbitrarily force a VM-entry to fail due to MSR loading when the MSR load list is too large. Similarly, trigger an abort during a VM exit that encounters an MSR load list or MSR store list that is too large. The MSR list size is intentionally not pre-checked so as to maintain compatibility with hardware inasmuch as possible. Test these new checks with the kvm-unit-test "x86: nvmx: test max atomic switch MSRs". Suggested-by:
Peter Shier <pshier@google.com> Signed-off-by:
Marc Orr <marcorr@google.com> Signed-off-by:
-
Jim Mattson authored
The RDPRU instruction gives the guest read access to the IA32_APERF MSR and the IA32_MPERF MSR. According to volume 3 of the APM, "When virtualization is enabled, this instruction can be intercepted by the Hypervisor. The intercept bit is at VMCB byte offset 10h, bit 14." Since we don't enumerate the instruction in KVM_SUPPORTED_CPUID, intercept it and synthesize #UD. Signed-off-by:
Drew Schmitt <dasch@google.com> Reviewed-by:
Jacob Xu <jacobhxu@google.com> Reviewed-by:
Krish Sadhukhan <krish.sadhukhan@oracle.com> Signed-off-by:
-
Jim Mattson authored
According to the Intel SDM, volume 2, "CPUID," the index is significant (or partially significant) for CPUID leaves 0FH, 10H, 12H, 17H, 18H, and 1FH. Add the corresponding flag to these CPUID leaves in do_host_cpuid(). Signed-off-by:
Steve Rutherford <srutherford@google.com> Fixes: a87f2d3a ("KVM: x86: Add Intel CPUID.1F cpuid emulation support") Reviewed-by:
-
Sean Christopherson authored
Do not skip invalid shadow pages when zapping obsolete pages if the pages' root_count has reached zero, in which case the page can be immediately zapped and freed. Update the comment accordingly. Signed-off-by:
-
Sean Christopherson authored
Toggle mmu_valid_gen between '0' and '1' instead of blindly incrementing the generation. Because slots_lock is held for the entire duration of zapping obsolete pages, it's impossible for there to be multiple invalid generations associated with shadow pages at any given time. Toggling between the two generations (valid vs. invalid) allows changing mmu_valid_gen from an unsigned long to a u8, which reduces the size of struct kvm_mmu_page from 160 to 152 bytes on 64-bit KVM, i.e. reduces KVM's memory footprint by 8 bytes per shadow page. Set sp->mmu_valid_gen before it is added to active_mmu_pages. Functionally this has no effect as kvm_mmu_alloc_page() has a single caller that sets sp->mmu_valid_gen soon thereafter, but visually it is jarring to see a shadow page being added to the list without its mmu_valid_gen first being set. Signed-off-by:
-
Sean Christopherson authored
Now that the fast invalidate mechanism has been reintroduced, restore the performance tweaks for fast invalidation that existed prior to its removal. Paraphrasing the original changelog (commit 5ff05683 was itself a partial revert): Don't force reloading the remote mmu when zapping an obsolete page, as a MMU_RELOAD request has already been issued by kvm_mmu_zap_all_fast() immediately after incrementing mmu_valid_gen, i.e. after marking pages obsolete. This reverts commit 5ff05683. Signed-off-by:
-
Sean Christopherson authored
Now that the fast invalidate mechanism has been reintroduced, restore the performance tweaks for fast invalidation that existed prior to its removal. Paraphrashing the original changelog: Introduce a per-VM list to track obsolete shadow pages, i.e. pages which have been deleted from the mmu cache but haven't yet been freed. When page reclaiming is needed, zap/free the deleted pages first. This reverts commit 52d5dedc. Signed-off-by:
-
Sean Christopherson authored
Now that the fast invalidate mechanism has been reintroduced, restore the performance tweaks for fast invalidation that existed prior to its removal. Paraphrashing the original changelog: Reload the mmu on all vCPUs after updating the generation number so that obsolete pages are not used by any vCPUs. This allows collapsing all TLB flushes during obsolete page zapping into a single flush, as there is no need to flush when dropping mmu_lock (to reschedule). Note: a remote TLB flush is still needed before freeing the pages as other vCPUs may be doing a lockless shadow page walk. Opportunstically improve the comments restored by the revert (the code itself is a true revert). This reverts commit f34d251d. Signed-off-by:
-
Sean Christopherson authored
Now that the fast invalidate mechanism has been reintroduced, restore the performance tweaks for fast invalidation that existed prior to its removal. Paraphrashing the original changelog: Zap at least 10 shadow pages before releasing mmu_lock to reduce the overhead associated with re-acquiring the lock. Note: "10" is an arbitrary number, speculated to be high enough so that a vCPU isn't stuck zapping obsolete pages for an extended period, but small enough so that other vCPUs aren't starved waiting for mmu_lock. This reverts commit 43d2b14b. Signed-off-by:
-
Sean Christopherson authored
Now that the fast invalidate mechanism has been reintroduced, restore the tracepoint associated with said mechanism. Note, the name of the tracepoint deviates from the original tracepoint so as to match KVM's current nomenclature. This reverts commit 42560fb1. Signed-off-by:
-
Sean Christopherson authored
Now that the fast invalidate mechanism has been reintroduced, restore tracing of the generation number in shadow page tracepoints. This reverts commit b59c4830. Signed-off-by:
-
Sean Christopherson authored
Use the fast invalidate mechasim to zap MMIO sptes on a MMIO generation wrap. The fast invalidate flow was reintroduced to fix a livelock bug in kvm_mmu_zap_all() that can occur if kvm_mmu_zap_all() is invoked when the guest has live vCPUs. I.e. using kvm_mmu_zap_all() to handle the MMIO generation wrap is theoretically susceptible to the livelock bug. This effectively reverts commit 4771450c ("Revert "KVM: MMU: drop kvm_mmu_zap_mmio_sptes""), i.e. restores the behavior of commit a8eca9dc ("KVM: MMU: drop kvm_mmu_zap_mmio_sptes"). Note, this actually fixes commit 571c5af0 ("KVM: x86/mmu: Voluntarily reschedule as needed when zapping MMIO sptes"), but there is no need to incrementally revert back to using fast invalidate, e.g. doing so doesn't provide any bisection or stability benefits. Fixes: 571c5af0 ("KVM: x86/mmu: Voluntarily reschedule as needed when zapping MMIO sptes") Cc: stable@vger.kernel.org Signed-off-by:
-
Sean Christopherson authored
Treat invalid shadow pages as obsolete to fix a bug where an obsolete and invalid page with a non-zero root count could become non-obsolete due to mmu_valid_gen wrapping. The bug is largely theoretical with the current code base, as an unsigned long will effectively never wrap on 64-bit KVM, and userspace would have to deliberately stall a vCPU in order to keep an obsolete invalid page on the active list while simultaneously modifying memslots billions of times to trigger a wrap. The obvious alternative is to use a 64-bit value for mmu_valid_gen, but it's actually desirable to go in the opposite direction, i.e. using a smaller 8-bit value to reduce KVM's memory footprint by 8 bytes per shadow page, and relying on proper treatment of invalid pages instead of preventing the generation from wrapping. Note, "Fixes" points at a commit that was at one point reverted, but has since been restored. Fixes: 5304b8d3 ("KVM: MMU: fast invalidate all pages") Signed-off-by:
-
Wanpeng Li authored
Filter out drastic fluctuation and random fluctuation, remove timer_advance_adjust_done altogether, the adjustment would be continuous. Signed-off-by:
-
Tao Xu authored
As the latest Intel 64 and IA-32 Architectures Software Developer's Manual, UMWAIT and TPAUSE instructions cause a VM exit if the RDTSC exiting and enable user wait and pause VM-execution controls are both 1. Because KVM never enable RDTSC exiting, the vm-exit for UMWAIT and TPAUSE should never happen. Considering EXIT_REASON_XSAVES and EXIT_REASON_XRSTORS is also unexpected VM-exit for KVM. Introduce a common exit helper handle_unexpected_vmexit() to handle these unexpected VM-exit. Suggested-by:
Jingqi Liu <jingqi.liu@intel.com> Signed-off-by:
Tao Xu <tao3.xu@intel.com> Signed-off-by:
-
Tao Xu authored
UMWAIT and TPAUSE instructions use 32bit IA32_UMWAIT_CONTROL at MSR index E1H to determines the maximum time in TSC-quanta that the processor can reside in either C0.1 or C0.2. This patch emulates MSR IA32_UMWAIT_CONTROL in guest and differentiate IA32_UMWAIT_CONTROL between host and guest. The variable mwait_control_cached in arch/x86/kernel/cpu/umwait.c caches the MSR value, so this patch uses it to avoid frequently rdmsr of IA32_UMWAIT_CONTROL. Co-developed-by:
-
Tao Xu authored
UMONITOR, UMWAIT and TPAUSE are a set of user wait instructions. This patch adds support for user wait instructions in KVM. Availability of the user wait instructions is indicated by the presence of the CPUID feature flag WAITPKG CPUID.0x07.0x0:ECX[5]. User wait instructions may be executed at any privilege level, and use 32bit IA32_UMWAIT_CONTROL MSR to set the maximum time. The behavior of user wait instructions in VMX non-root operation is determined first by the setting of the "enable user wait and pause" secondary processor-based VM-execution control bit 26. If the VM-execution control is 0, UMONITOR/UMWAIT/TPAUSE cause an invalid-opcode exception (#UD). If the VM-execution control is 1, treatment is based on the setting of the “RDTSC exiting†VM-execution control. Because KVM never enables RDTSC exiting, if the instruction causes a delay, the amount of time delayed is called here the physical delay. The physical delay is first computed by determining the virtual delay. If IA32_UMWAIT_CONTROL[31:2] is zero, the virtual delay is the value in EDX:EAX minus the value that RDTSC would return; if IA32_UMWAIT_CONTROL[31:2] is not zero, the virtual delay is the minimum of that difference and AND(IA32_UMWAIT_CONTROL,FFFFFFFCH). Because umwait and tpause can put a (psysical) CPU into a power saving state, by default we dont't expose it to kvm and enable it only when guest CPUID has it. Detailed information about user wait instructions can be found in the latest Intel 64 and IA-32 Architectures Software Developer's Manual. Co-developed-by:
-
Sean Christopherson authored
Document the intended usage of each emulation type as each exists to handle an edge case of one kind or another and can be easily misinterpreted at first glance. Cc: Liran Alon <liran.alon@oracle.com> Signed-off-by:
-
Sean Christopherson authored
VMX's EPT misconfig flow to handle fast-MMIO path falls back to decoding the instruction to determine the instruction length when running as a guest (Hyper-V doesn't fill VMCS.VM_EXIT_INSTRUCTION_LEN because it's technically not defined for EPT misconfigs). Rather than implement the slow skip in VMX's generic skip_emulated_instruction(), handle_ept_misconfig() directly calls kvm_emulate_instruction() with EMULTYPE_SKIP, which intentionally doesn't do single-step detection, and so handle_ept_misconfig() misses a single-step #DB. Rework the EPT misconfig fallback case to route it through kvm_skip_emulated_instruction() so that single-step #DBs and interrupt shadow updates are handled automatically. I.e. make VMX's slow skip logic match SVM's and have the SVM flow not intentionally avoid the shadow update. Alternatively, the handle_ept_misconfig() could manually handle single- step detection, but that results in EMULTYPE_SKIP having split logic for the interrupt shadow vs. single-step #DBs, and split emulator logic is largely what led to this mess in the first place. Modifying SVM to mirror VMX flow isn't really an option as SVM's case isn't limited to a specific exit reason, i.e. handling the slow skip in skip_emulated_instruction() is mandatory for all intents and purposes. Drop VMX's skip_emulated_instruction() wrapper since it can now fail, and instead WARN if it fails unexpectedly, e.g. if exit_reason somehow becomes corrupted. Cc: Vitaly Kuznetsov <vkuznets@redhat.com> Fixes: d391f120 ("x86/kvm/vmx: do not use vm-exit instruction length for fast MMIO when running nested") Signed-off-by:
-
Sean Christopherson authored
Deferring emulation failure handling (in some cases) to the caller of x86_emulate_instruction() has proven fragile, e.g. multiple instances of KVM not setting run->exit_reason on EMULATE_FAIL, largely due to it being difficult to discern what emulation types can return what result, and which combination of types and results are handled where. Now that x86_emulate_instruction() always handles emulation failure, i.e. EMULATION_FAIL is only referenced in callers, remove the emulation_result enums entirely. Per KVM's existing exit handling conventions, return '0' and '1' for "exit to userspace" and "resume guest" respectively. Doing so cleans up many callers, e.g. they can return kvm_emulate_instruction() directly instead of having to interpret its result. Signed-off-by:
-
Sean Christopherson authored
Now that EMULATE_FAIL is completely unused, remove the last remaning usage where KVM does something functional in response to EMULATE_FAIL. Leave the check in place as a WARN_ON_ONCE to provide a better paper trail when EMULATE_{DONE,FAIL,USER_EXIT} are completely removed. Opportunistically remove the gotos in handle_invalid_guest_state(). With the EMULATE_FAIL handling gone there is no need to have a common handler for emulation failure and the gotos only complicate things, e.g. the signal_pending() check always returns '1', but this is far from obvious when glancing through the code. Signed-off-by: -
Sean Christopherson authored
Request triple fault in kvm_inject_realmode_interrupt() instead of returning EMULATE_FAIL and deferring to the caller. All existing callers request triple fault and it's highly unlikely Real Mode is going to acquire new features. While this consolidates a small amount of code, the real goal is to remove the last reference to EMULATE_FAIL. No functional change intended. Signed-off-by:
-
Sean Christopherson authored
Consolidate the reporting of emulation failure into kvm_task_switch() so that it can return EMULATE_USER_EXIT. This helps pave the way for removing EMULATE_FAIL altogether. This also fixes a theoretical bug where task switch interception could suppress an EMULATE_USER_EXIT return. Signed-off-by:
-
Sean Christopherson authored
Kill a few birds with one stone by forcing an exit to userspace on skip emulation failure. This removes a reference to EMULATE_FAIL, fixes a bug in handle_ept_misconfig() where it would exit to userspace without setting run->exit_reason, and fixes a theoretical bug in SVM's task_switch_interception() where it would overwrite run->exit_reason on a return of EMULATE_USER_EXIT. Note, this technically doesn't fully fix task_switch_interception() as it now incorrectly handles EMULATE_FAIL, but in practice there is no bug as EMULATE_FAIL will never be returned for EMULTYPE_SKIP. Signed-off-by:
-
Sean Christopherson authored
Immediately inject a #UD and return EMULATE done if emulation fails when handling an intercepted #UD. This helps pave the way for removing EMULATE_FAIL altogether. Signed-off-by:
-
Sean Christopherson authored
Add an explicit emulation type for forced #UD emulation and use it to detect that KVM should unconditionally inject a #UD instead of falling into its standard emulation failure handling. Signed-off-by:
-
Sean Christopherson authored
Immediately inject a #GP when VMware emulation fails and return EMULATE_DONE instead of propagating EMULATE_FAIL up the stack. This helps pave the way for removing EMULATE_FAIL altogether. Rename EMULTYPE_VMWARE to EMULTYPE_VMWARE_GP to document that the x86 emulator is called to handle VMware #GP interception, e.g. why a #GP is injected on emulation failure for EMULTYPE_VMWARE_GP. Drop EMULTYPE_NO_UD_ON_FAIL as a standalone type. The "no #UD on fail" is used only in the VMWare case and is obsoleted by having the emulator itself reinject #GP. Signed-off-by:
Liran Alon <liran.alon@oracle.com> Signed-off-by:
-
Sean Christopherson authored
The VMware backdoor hooks #GP faults on IN{S}, OUT{S}, and RDPMC, none of which generate a non-zero error code for their #GP. Re-injecting #GP instead of attempting emulation on a non-zero error code will allow a future patch to move #GP injection (for emulation failure) into kvm_emulate_instruction() without having to plumb in the error code. Reviewed-and-tested-by: -
Sean Christopherson authored
Return the single-step emulation result directly instead of via an out param. Presumably at some point in the past kvm_vcpu_do_singlestep() could be called with *r==EMULATE_USER_EXIT, but that is no longer the case, i.e. all callers are happy to overwrite their own return variable. Reviewed-by:
-
Sean Christopherson authored
When handling emulation failure, return the emulation result directly instead of capturing it in a local variable. Future patches will move additional cases into handle_emulation_failure(), clean up the cruft before so there isn't an ugly mix of setting a local variable and returning directly. Reviewed-by:
-
