- 04 Feb, 2021 40 commits
-
-
Ben Gardon authored
Safely rescheduling while holding a spin lock is essential for keeping long running kernel operations running smoothly. Add the facility to cond_resched rwlocks. CC: Ingo Molnar <mingo@redhat.com> CC: Will Deacon <will@kernel.org> Acked-by:
Peter Zijlstra <peterz@infradead.org> Acked-by:
Davidlohr Bueso <dbueso@suse.de> Acked-by:
Waiman Long <longman@redhat.com> Acked-by:
Paolo Bonzini <pbonzini@redhat.com> Signed-off-by:
Ben Gardon <bgardon@google.com> Message-Id: <20210202185734.1680553-9-bgardon@google.com> Signed-off-by:
-
Ben Gardon authored
Contention awareness while holding a spin lock is essential for reducing latency when long running kernel operations can hold that lock. Add the same contention detection interface for read/write spin locks. CC: Ingo Molnar <mingo@redhat.com> CC: Will Deacon <will@kernel.org> Acked-by:
-
Ben Gardon authored
rwlocks do not currently have any facility to detect contention like spinlocks do. In order to allow users of rwlocks to better manage latency, add contention detection for queued rwlocks. CC: Ingo Molnar <mingo@redhat.com> CC: Will Deacon <will@kernel.org> Acked-by:
-
Ben Gardon authored
In order to enable concurrent modifications to the paging structures in the TDP MMU, threads must be able to safely remove pages of page table memory while other threads are traversing the same memory. To ensure threads do not access PT memory after it is freed, protect PT memory with RCU. Protecting concurrent accesses to page table memory from use-after-free bugs could also have been acomplished using walk_shadow_page_lockless_begin/end() and READING_SHADOW_PAGE_TABLES, coupling with the barriers in a TLB flush. The use of RCU for this case has several distinct advantages over that approach. 1. Disabling interrupts for long running operations is not desirable. Future commits will allow operations besides page faults to operate without the exclusive protection of the MMU lock and those operations are too long to disable iterrupts for their duration. 2. The use of RCU here avoids long blocking / spinning operations in perfromance critical paths. By freeing memory with an asynchronous RCU API we avoid the longer wait times TLB flushes experience when overlapping with a thread in walk_shadow_page_lockless_begin/end(). 3. RCU provides a separation of concerns when removing memory from the paging structure. Because the RCU callback to free memory can be scheduled immediately after a TLB flush, there's no need for the thread to manually free a queue of pages later, as commit_zap_pages does. Fixes: 95fb5b02 ("kvm: x86/mmu: Support MMIO in the TDP MMU") Reviewed-by:
Peter Feiner <pfeiner@google.com> Suggested-by:
Sean Christopherson <seanjc@google.com> Signed-off-by:
-
Ben Gardon authored
In clear_dirty_pt_masked, the loop is intended to exit early after processing each of the GFNs with corresponding bits set in mask. This does not work as intended if another thread has already cleared the dirty bit or writable bit on the SPTE. In that case, the loop would proceed to the next iteration early and the bit in mask would not be cleared. As a result the loop could not exit early and would proceed uselessly. Move the unsetting of the mask bit before the check for a no-op SPTE change. Fixes: a6a0b05d ("kvm: x86/mmu: Support dirty logging for the TDP MMU") Suggested-by:
-
Ben Gardon authored
Skip setting SPTEs if no change is expected. Reviewed-by:
-
Ben Gardon authored
Given certain conditions, some TDP MMU functions may not yield reliably / frequently enough. For example, if a paging structure was very large but had few, if any writable entries, wrprot_gfn_range could traverse many entries before finding a writable entry and yielding because the check for yielding only happens after an SPTE is modified. Fix this issue by moving the yield to the beginning of the loop. Fixes: a6a0b05d ("kvm: x86/mmu: Support dirty logging for the TDP MMU") Reviewed-by:
-
Ben Gardon authored
In some functions the TDP iter risks not making forward progress if two threads livelock yielding to one another. This is possible if two threads are trying to execute wrprot_gfn_range. Each could write protect an entry and then yield. This would reset the tdp_iter's walk over the paging structure and the loop would end up repeating the same entry over and over, preventing either thread from making forward progress. Fix this issue by only yielding if the loop has made forward progress since the last yield. Fixes: a6a0b05d ("kvm: x86/mmu: Support dirty logging for the TDP MMU") Reviewed-by:
-
Ben Gardon authored
The goal_gfn field in tdp_iter can be misleading as it implies that it is the iterator's final goal. It is really a target for the lowest gfn mapped by the leaf level SPTE the iterator will traverse towards. Change the field's name to be more precise. Signed-off-by:
-
Ben Gardon authored
The flushing and non-flushing variants of tdp_mmu_iter_cond_resched have almost identical implementations. Merge the two functions and add a flush parameter. Signed-off-by:
-
Ben Gardon authored
No functional change intended. Fixes: 29cf0f50 ("kvm: x86/mmu: NX largepage recovery for TDP MMU") Signed-off-by:
-
Ben Gardon authored
Factor out the code to handle a disconnected subtree of the TDP paging structure from the code to handle the change to an individual SPTE. Future commits will build on this to allow asynchronous page freeing. No functional change intended. Reviewed-by:
-
Ben Gardon authored
The KVM MMU caches already guarantee that shadow page table memory will be zeroed, so there is no reason to re-zero the page in the TDP MMU page fault handler. No functional change intended. Reviewed-by:
-
Ben Gardon authored
Add lockdep to __tdp_mmu_set_spte to ensure that SPTEs are only modified under the MMU lock. No functional change intended. Reviewed-by:
-
Ben Gardon authored
__tdp_mmu_set_spte is a very important function in the TDP MMU which already accepts several arguments and will take more in future commits. To offset this complexity, add a comment to the function describing each of the arguemnts. No functional change intended. Reviewed-by:
-
Ben Gardon authored
Currently the TDP MMU yield / cond_resched functions either return nothing or return true if the TLBs were not flushed. These are confusing semantics, especially when making control flow decisions in calling functions. To clean things up, change both functions to have the same return value semantics as cond_resched: true if the thread yielded, false if it did not. If the function yielded in the _flush_ version, then the TLBs will have been flushed. Reviewed-by:
-
Paolo Bonzini authored
Push the injection of #GP up to the callers, so that they can just use kvm_complete_insn_gp. Signed-off-by: -
Paolo Bonzini authored
kvm_dr6_valid and kvm_dr7_valid check that bits 63:32 are zero. Using them makes it easier to review the code for inconsistencies. Signed-off-by: -
Paolo Bonzini authored
Now that KVM is using static calls, calling vmx_vcpu_run and vmx_sync_pir_to_irr does not incur anymore the cost of a retpoline. Therefore there is no need anymore to handle EXIT_FASTPATH_REENTER_GUEST in vendor code. Signed-off-by: -
Vitaly Kuznetsov authored
Commit 181f4948 ("KVM: x86: fix CPUID entries returned by KVM_GET_CPUID2 ioctl") revealed that we're not testing KVM_GET_CPUID2 ioctl at all. Add a test for it and also check that from inside the guest visible CPUIDs are equal to it's output. Signed-off-by:
Vitaly Kuznetsov <vkuznets@redhat.com> Message-Id: <20210129161821.74635-1-vkuznets@redhat.com> Signed-off-by:
-
Stephen Zhang authored
Given the common pattern: rmap_printk("%s:"..., __func__,...) we could improve this by adding '__func__' in rmap_printk(). Signed-off-by:Stephen Zhang <stephenzhangzsd@gmail.com> Message-Id: <1611713325-3591-1-git-send-email-stephenzhangzsd@gmail.com> Signed-off-by:
-
Krish Sadhukhan authored
Replace the hard-coded value for bit# 1 in EFLAGS, with the available #define. Signed-off-by:
Krish Sadhukhan <krish.sadhukhan@oracle.com> Message-Id: <20210203012842.101447-2-krish.sadhukhan@oracle.com> Signed-off-by:
-
Michael Roth authored
Currently we save host state like user-visible host MSRs, and do some initial guest register setup for MSR_TSC_AUX and MSR_AMD64_TSC_RATIO in svm_vcpu_load(). Defer this until just before we enter the guest by moving the handling to kvm_x86_ops.prepare_guest_switch() similarly to how it is done for the VMX implementation. Additionally, since handling of saving/restoring host user MSRs is the same both with/without SEV-ES enabled, move that handling to common code. Suggested-by:
Michael Roth <michael.roth@amd.com> Message-Id: <20210202190126.2185715-4-michael.roth@amd.com> Signed-off-by:
-
Michael Roth authored
Now that the set of host user MSRs that need to be individually saved/restored are the same with/without SEV-ES, we can drop the .sev_es_restored flag and just iterate through the list unconditionally for both cases. A subsequent patch can then move these loops to a common path. Signed-off-by:
-
Michael Roth authored
Using a guest workload which simply issues 'hlt' in a tight loop to generate VMEXITs, it was observed (on a recent EPYC processor) that a significant amount of the VMEXIT overhead measured on the host was the result of MSR reads/writes in svm_vcpu_load/svm_vcpu_put according to perf: 67.49%--kvm_arch_vcpu_ioctl_run | |--23.13%--vcpu_put | kvm_arch_vcpu_put | | | |--21.31%--native_write_msr | | | --1.27%--svm_set_cr4 | |--16.11%--vcpu_load | | | --15.58%--kvm_arch_vcpu_load | | | |--13.97%--svm_set_cr4 | | | | | |--12.64%--native_read_msr Most of these MSRs relate to 'syscall'/'sysenter' and segment bases, and can be saved/restored using 'vmsave'/'vmload' instructions rather than explicit MSR reads/writes. In doing so there is a significant reduction in the svm_vcpu_load/svm_vcpu_put overhead measured for the above workload: 50.92%--kvm_arch_vcpu_ioctl_run | |--19.28%--disable_nmi_singlestep | |--13.68%--vcpu_load | kvm_arch_vcpu_load | | | |--9.19%--svm_set_cr4 | | | | | --6.44%--native_read_msr | | | --3.55%--native_write_msr | |--6.05%--kvm_inject_nmi |--2.80%--kvm_sev_es_mmio_read |--2.19%--vcpu_put | | | --1.25%--kvm_arch_vcpu_put | native_write_msr Quantifying this further, if we look at the raw cycle counts for a normal iteration of the above workload (according to 'rdtscp'), kvm_arch_vcpu_ioctl_run() takes ~4600 cycles from start to finish with the current behavior. Using 'vmsave'/'vmload', this is reduced to ~2800 cycles, a savings of 39%. While this approach doesn't seem to manifest in any noticeable improvement for more realistic workloads like UnixBench, netperf, and kernel builds, likely due to their exit paths generally involving IO with comparatively high latencies, it does improve overall overhead of KVM_RUN significantly, which may still be noticeable for certain situations. It also simplifies some aspects of the code. With this change, explicit save/restore is no longer needed for the following host MSRs, since they are documented[1] as being part of the VMCB State Save Area: MSR_STAR, MSR_LSTAR, MSR_CSTAR, MSR_SYSCALL_MASK, MSR_KERNEL_GS_BASE, MSR_IA32_SYSENTER_CS, MSR_IA32_SYSENTER_ESP, MSR_IA32_SYSENTER_EIP, MSR_FS_BASE, MSR_GS_BASE and only the following MSR needs individual handling in svm_vcpu_put/svm_vcpu_load: MSR_TSC_AUX We could drop the host_save_user_msrs array/loop and instead handle MSR read/write of MSR_TSC_AUX directly, but we leave that for now as a potential follow-up. Since 'vmsave'/'vmload' also handles the LDTR and FS/GS segment registers (and associated hidden state)[2], some of the code previously used to handle this is no longer needed, so we drop it as well. The first public release of the SVM spec[3] also documents the same handling for the host state in question, so we make these changes unconditionally. Also worth noting is that we 'vmsave' to the same page that is subsequently used by 'vmrun' to record some host additional state. This is okay, since, in accordance with the spec[2], the additional state written to the page by 'vmrun' does not overwrite any fields written by 'vmsave'. This has also been confirmed through testing (for the above CPU, at least). [1] AMD64 Architecture Programmer's Manual, Rev 3.33, Volume 2, Appendix B, Table B-2 [2] AMD64 Architecture Programmer's Manual, Rev 3.31, Volume 3, Chapter 4, VMSAVE/VMLOAD [3] Secure Virtual Machine Architecture Reference Manual, Rev 3.01 Suggested-by:Tom Lendacky <thomas.lendacky@amd.com> Signed-off-by:
-
Sean Christopherson authored
Add svm_asm*() macros, a la the existing vmx_asm*() macros, to handle faults on SVM instructions instead of using the generic __ex(), a.k.a. __kvm_handle_fault_on_reboot(). Using asm goto generates slightly better code as it eliminates the in-line JMP+CALL sequences that are needed by __kvm_handle_fault_on_reboot() to avoid triggering BUG() from fixup (which generates bad stack traces). Using SVM specific macros also drops the last user of __ex() and the the last asm linkage to kvm_spurious_fault(), and adds a helper for VMSAVE, which may gain an addition call site in the future (as part of optimizing the SVM context switching). Signed-off-by:
-
Sean Christopherson authored
Drop kvm_cpu_vmxoff() in favor of the kernel's cpu_vmxoff(). Modify the latter to return -EIO on fault so that KVM can invoke kvm_spurious_fault() when appropriate. In addition to the obvious code reuse, dropping kvm_cpu_vmxoff() also eliminates VMX's last usage of the __ex()/__kvm_handle_fault_on_reboot() macros, thus helping pave the way toward dropping them entirely. Signed-off-by:
-
Sean Christopherson authored
Move the Intel PT tracking outside of the VMXON/VMXOFF helpers so that a future patch can drop KVM's kvm_cpu_vmxoff() in favor of the kernel's cpu_vmxoff() without an associated PT functional change, and without losing symmetry between the VMXON and VMXOFF flows. Barring undocumented behavior, this should have no meaningful effects as Intel PT behavior does not interact with CR4.VMXE. Signed-off-by:
-
Uros Bizjak authored
Replace inline assembly in nested_vmx_check_vmentry_hw with a call to __vmx_vcpu_run. The function is not performance critical, so (double) GPR save/restore in __vmx_vcpu_run can be tolerated, as far as performance effects are concerned. Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Sean Christopherson <seanjc@google.com> Reviewed-and-tested-by:
Uros Bizjak <ubizjak@gmail.com> [sean: dropped versioning info from changelog] Signed-off-by:
-
David P. Reed authored
Explicitly tell the compiler that VMXOFF modifies flags (like all VMX instructions), and mark memory as clobbered since VMXOFF must not be reordered and also may have memory side effects (though the kernel really shouldn't be accessing the root VMCS anyways). Practically speaking, adding the clobbers is most likely a nop; the primary motivation is to properly document VMXOFF's behavior. For the flags clobber, both Clang and GCC automatically mark flags as clobbered; this is noted in commit 4b1e5478 ("KVM/x86: Use assembly instruction mnemonics instead of .byte streams"), which intentionally removed the previous clobber. But, neither Clang nor GCC documents this behavior, and there's no downside to including the clobber. For the memory clobber, the RFLAGS.IF and CR4.VMXE manipulations that immediately follow VMXOFF have compiler barriers of their own, i.e. VMXOFF can't get reordered after clearing CR4.VMXE, which is really what's of interest. Cc: Randy Dunlap <rdunlap@infradead.org> Signed-off-by:
David P. Reed <dpreed@deepplum.com> [sean: rewrote changelog, dropped comment adjustments] Signed-off-by:
-
Sean Christopherson authored
Force all CPUs to do VMXOFF (via NMI shootdown) during an emergency reboot if VMX is _supported_, as VMX being off on the current CPU does not prevent other CPUs from being in VMX root (post-VMXON). This fixes a bug where a crash/panic reboot could leave other CPUs in VMX root and prevent them from being woken via INIT-SIPI-SIPI in the new kernel. Fixes: d176720d ("x86: disable VMX on all CPUs on reboot") Cc: stable@vger.kernel.org Suggested-by:
-
Sean Christopherson authored
Silently ignore all faults on VMXOFF in the reboot flows as such faults are all but guaranteed to be due to the CPU not being in VMX root. Because (a) VMXOFF may be executed in NMI context, e.g. after VMXOFF but before CR4.VMXE is cleared, (b) there's no way to query the CPU's VMX state without faulting, and (c) the whole point is to get out of VMX root, eating faults is the simplest way to achieve the desired behaior. Technically, VMXOFF can fault (or fail) for other reasons, but all other fault and failure scenarios are mode related, i.e. the kernel would have to magically end up in RM, V86, compat mode, at CPL>0, or running with the SMI Transfer Monitor active. The kernel is beyond hosed if any of those scenarios are encountered; trying to do something fancy in the error path to handle them cleanly is pointless. Fixes: 1e993114 ("x86: asm/virtext.h: add cpu_vmxoff() inline function") Reported-by:
-
Jason Baron authored
Convert kvm_x86_ops to use static calls. Note that all kvm_x86_ops are covered here except for 'pmu_ops and 'nested ops'. Here are some numbers running cpuid in a loop of 1 million calls averaged over 5 runs, measured in the vm (lower is better). Intel Xeon 3000MHz: |default |mitigations=off ------------------------------------- vanilla |.671s |.486s static call|.573s(-15%)|.458s(-6%) AMD EPYC 2500MHz: |default |mitigations=off ------------------------------------- vanilla |.710s |.609s static call|.664s(-6%) |.609s(0%) Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Sean Christopherson <seanjc@google.com> Signed-off-by:Jason Baron <jbaron@akamai.com> Message-Id: <e057bf1b8a7ad15652df6eeba3f907ae758d3399.1610680941.git.jbaron@akamai.com> Signed-off-by:
-
Jason Baron authored
Use static calls to improve kvm_x86_ops performance. Introduce the definitions that will be used by a subsequent patch to actualize the savings. Add a new kvm-x86-ops.h header that can be used for the definition of static calls. This header is also intended to be used to simplify the defition of svm_kvm_ops and vmx_x86_ops. Note that all functions in kvm_x86_ops are covered here except for 'pmu_ops' and 'nested ops'. I think they can be covered by static calls in a simlilar manner, but were omitted from this series to reduce scope and because I don't think they have as large of a performance impact. Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Sean Christopherson <seanjc@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by:
-
Jason Baron authored
A subsequent patch introduces macros in preparation for simplifying the definition for vmx_x86_ops and svm_x86_ops. Making the naming more uniform expands the coverage of the macros. Add vmx/svm prefix to the following functions: update_exception_bitmap(), enable_nmi_window(), enable_irq_window(), update_cr8_intercept and enable_smi_window(). Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Sean Christopherson <seanjc@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by:
-
Cun Li authored
The use of 'struct static_key' and 'static_key_false' is deprecated. Use the new API. Signed-off-by:
Cun Li <cun.jia.li@gmail.com> Message-Id: <20210111152435.50275-1-cun.jia.li@gmail.com> [Make it compile. While at it, rename kvm_no_apic_vcpu to kvm_has_noapic_vcpu; the former reads too much like "true if no vCPU has an APIC". - Paolo] Signed-off-by:
-
Wei Huang authored
Under the case of nested on nested (L0, L1, L2 are all hypervisors), we do not support emulation of the vVMLOAD/VMSAVE feature, the L0 hypervisor can inject the proper #VMEXIT to inform L1 of what is happening and L1 can avoid invoking the #GP workaround. For this reason we turns on guest VM's X86_FEATURE_SVME_ADDR_CHK bit for KVM running inside VM to receive the notification and change behavior. Similarly we check if vcpu is under guest mode before emulating the vmware-backdoor instructions. For the case of nested on nested, we let the guest handle it. Co-developed-by:
Bandan Das <bsd@redhat.com> Signed-off-by:
Wei Huang <wei.huang2@amd.com> Tested-by:
Maxim Levitsky <mlevitsk@redhat.com> Reviewed-by:
-
Wei Huang authored
New AMD CPUs have a change that checks #VMEXIT intercept on special SVM instructions before checking their EAX against reserved memory region. This change is indicated by CPUID_0x8000000A_EDX[28]. If it is 1, #VMEXIT is triggered before #GP. KVM doesn't need to intercept and emulate #GP faults as #GP is supposed to be triggered. Co-developed-by:
-
Bandan Das authored
While running SVM related instructions (VMRUN/VMSAVE/VMLOAD), some AMD CPUs check EAX against reserved memory regions (e.g. SMM memory on host) before checking VMCB's instruction intercept. If EAX falls into such memory areas, #GP is triggered before VMEXIT. This causes problem under nested virtualization. To solve this problem, KVM needs to trap #GP and check the instructions triggering #GP. For VM execution instructions, KVM emulates these instructions. Co-developed-by:
-
Wei Huang authored
Move the instruction decode part out of x86_emulate_instruction() for it to be used in other places. Also kvm_clear_exception_queue() is moved inside the if-statement as it doesn't apply when KVM are coming back from userspace. Co-developed-by:
-
