After updating the kernel from 5.1.17 to 5.2.1, when I use the g++ compiler in qemu with kvm enabled, often the compiler launched in the guest for compile my sources ends with this error: "exception in floating point". Moreever it is not possible to update the bsd system because the process (that use g++ compiler) exits with the same error. Same thing for some system utilities, like "pkg update". But if I boot the system with the previous 5.1.17 kernel and launch qemu, as indicated above, everything works fine and I don't detect errors. Command line is: /usr/bin/qemu-system-x86_64 -k it -machine accel=kvm -m 4096 -no-fd-bootchk -show-cursor -drive file="vmdragon.img",if=ide,media=disk -boot once=c,menu=off -net none -rtc base=localtime -name "vmdragon" -smp 8 -vga std -device qemu-xhci,id=xhci I can't test without kvm because the process is very slow, but I think the problem could be in some changes made to the kvm module. Thanks, Antonio
I can confirm this issue. It occurs since 5.2, when the FPU state changes were introduced. I send an e-mail about this yesterday, seems I should have included the kvm mailing list: https://lkml.org/lkml/2019/7/17/758 - Thomas
I tested the patch indicated on https://lkml.org/lkml/2019/7/19/644 with simple row position adjustments for the kernel 5.2.1 and it seems to work. --- a/arch/x86/kvm/x86.c 2019-07-19 20:17:35.358848175 +0200 +++ b/arch/x86/kvm/x86.c 2019-07-19 20:17:17.956692942 +0200 @@ -3264,6 +3264,12 @@ kvm_x86_ops->vcpu_load(vcpu, cpu); + + // fix floating point error kvm guest + if (test_thread_flag(TIF_NEED_FPU_LOAD)) + switch_fpu_return(); + + /* Apply any externally detected TSC adjustments (due to suspend) */ if (unlikely(vcpu->arch.tsc_offset_adjustment)) { adjust_tsc_offset_host(vcpu, vcpu->arch.tsc_offset_adjustment); @@ -7955,9 +7961,11 @@ wait_lapic_expire(vcpu); guest_enter_irqoff(); - fpregs_assert_state_consistent(); - if (test_thread_flag(TIF_NEED_FPU_LOAD)) - switch_fpu_return(); +// fix floating point error kvm guest +// +// fpregs_assert_state_consistent(); +// if (test_thread_flag(TIF_NEED_FPU_LOAD)) +// switch_fpu_return(); if (unlikely(vcpu->arch.switch_db_regs)) { set_debugreg(0, 7);
Today I saw the following commit: >Revert "kvm: x86: Use task structs fpu field for user" >commit ec269475cba7bcdd1eb8fdf8e87f4c6c81a376fe upstream. > >This reverts commit 240c35a3783ab9b3a0afaba0dde7291295680a6b >("kvm: x86: Use task structs fpu field for user", 2018-11-06). >The commit is broken and causes QEMU's FPU state to be destroyed >when KVM_RUN is preempted. > >Fixes: 240c35a3783a ("kvm: x86: Use task structs fpu field for user") applied to the 5.2.4 kernel and I thought it could relate to the reported bug, but recompiling the kernel without previous indicated patch, the guest report "fpu exception" error again: it was necessary to re-include the patch and recompile the kernel for it to work.
Tested with Kernel 5.2.5: problem solved. Thanks, Antonio