diff options
Diffstat (limited to 'results/classifier/accel-gemma3:12b/kvm/2710')
| -rw-r--r-- | results/classifier/accel-gemma3:12b/kvm/2710 | 127 |
1 files changed, 127 insertions, 0 deletions
diff --git a/results/classifier/accel-gemma3:12b/kvm/2710 b/results/classifier/accel-gemma3:12b/kvm/2710 new file mode 100644 index 000000000..3bfbf7c4c --- /dev/null +++ b/results/classifier/accel-gemma3:12b/kvm/2710 @@ -0,0 +1,127 @@ + +QEMU can't detect guest debug support on older (pre v5.7) x86 host kernels due to missing KVM_CAP_SET_GUEST_DEBUG +Description of problem: +``` +qemu-system-x86_64: -s: gdbstub: current accelerator doesn't support guest debugging +``` +Additional information: +I initially located the QEMU source code to determine whether KVM supports gdbstub by checking for `KVM_CAP_SET_GUEST_DEBUG`. The corresponding code can be found at: +```c +// qemu/accel/kvm/kvm-all.c:2695 +#ifdef TARGET_KVM_HAVE_GUEST_DEBUG + kvm_has_guest_debug = + (kvm_check_extension(s, KVM_CAP_SET_GUEST_DEBUG) > 0); +#endif +``` +It can be observed that if the return value is <= 0 (in practice, this function only returns 0 on failure), the debug_flag is set to false. + +Upon further investigation of the Linux 4.15 kernel code, I discovered that in earlier versions, support for checking VM debugging capabilities via `KVM_CAP_SET_GUEST_DEBUG` was almost non-existent (it was only supported on arm64). However, for x86_64, VM debugging is supported on the 4.15 kernel. + +```c +// linu4.15/arch/x86/kvm/x86.c:2672 +int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext) +{ + int r; + + switch (ext) { + case KVM_CAP_IRQCHIP: + case KVM_CAP_HLT: + case KVM_CAP_MMU_SHADOW_CACHE_CONTROL: + case KVM_CAP_SET_TSS_ADDR: + case KVM_CAP_EXT_CPUID: + case KVM_CAP_EXT_EMUL_CPUID: + case KVM_CAP_CLOCKSOURCE: + case KVM_CAP_PIT: + case KVM_CAP_NOP_IO_DELAY: + case KVM_CAP_MP_STATE: + case KVM_CAP_SYNC_MMU: + case KVM_CAP_USER_NMI: + case KVM_CAP_REINJECT_CONTROL: + case KVM_CAP_IRQ_INJECT_STATUS: + case KVM_CAP_IOEVENTFD: + case KVM_CAP_IOEVENTFD_NO_LENGTH: + case KVM_CAP_PIT2: + case KVM_CAP_PIT_STATE2: + case KVM_CAP_SET_IDENTITY_MAP_ADDR: + case KVM_CAP_XEN_HVM: + case KVM_CAP_VCPU_EVENTS: + case KVM_CAP_HYPERV: + case KVM_CAP_HYPERV_VAPIC: + case KVM_CAP_HYPERV_SPIN: + case KVM_CAP_HYPERV_SYNIC: + case KVM_CAP_HYPERV_SYNIC2: + case KVM_CAP_HYPERV_VP_INDEX: + case KVM_CAP_PCI_SEGMENT: + case KVM_CAP_DEBUGREGS: + case KVM_CAP_X86_ROBUST_SINGLESTEP: + case KVM_CAP_XSAVE: + case KVM_CAP_ASYNC_PF: + case KVM_CAP_GET_TSC_KHZ: + case KVM_CAP_KVMCLOCK_CTRL: + case KVM_CAP_READONLY_MEM: + case KVM_CAP_HYPERV_TIME: + case KVM_CAP_IOAPIC_POLARITY_IGNORED: + case KVM_CAP_TSC_DEADLINE_TIMER: + case KVM_CAP_ENABLE_CAP_VM: + case KVM_CAP_DISABLE_QUIRKS: + case KVM_CAP_SET_BOOT_CPU_ID: + case KVM_CAP_SPLIT_IRQCHIP: + case KVM_CAP_IMMEDIATE_EXIT: + r = 1; + break; + case KVM_CAP_ADJUST_CLOCK: + r = KVM_CLOCK_TSC_STABLE; + break; + case KVM_CAP_X86_GUEST_MWAIT: + r = kvm_mwait_in_guest(); + break; + case KVM_CAP_X86_SMM: + /* SMBASE is usually relocated above 1M on modern chipsets, + * and SMM handlers might indeed rely on 4G segment limits, + * so do not report SMM to be available if real mode is + * emulated via vm86 mode. Still, do not go to great lengths + * to avoid userspace's usage of the feature, because it is a + * fringe case that is not enabled except via specific settings + * of the module parameters. + */ + r = kvm_x86_ops->cpu_has_high_real_mode_segbase(); + break; + case KVM_CAP_VAPIC: + r = !kvm_x86_ops->cpu_has_accelerated_tpr(); + break; + case KVM_CAP_NR_VCPUS: + r = KVM_SOFT_MAX_VCPUS; + break; + case KVM_CAP_MAX_VCPUS: + r = KVM_MAX_VCPUS; + break; + case KVM_CAP_NR_MEMSLOTS: + r = KVM_USER_MEM_SLOTS; + break; + case KVM_CAP_PV_MMU: /* obsolete */ + r = 0; + break; + case KVM_CAP_MCE: + r = KVM_MAX_MCE_BANKS; + break; + case KVM_CAP_XCRS: + r = boot_cpu_has(X86_FEATURE_XSAVE); + break; + case KVM_CAP_TSC_CONTROL: + r = kvm_has_tsc_control; + break; + case KVM_CAP_X2APIC_API: + r = KVM_X2APIC_API_VALID_FLAGS; + break; + default: + r = 0; + break; + } + return r; + +} +``` + +I attempted to bypass this check in QEMU and verified that the QEMU gdbstub works normally on the 4.15 kernel. + +For modifications related to this part in QEMU, you can refer to the email: https://lore.kernel.org/all/20211111110604.207376-5-pbonzini@redhat.com/. |