On Fri, Aug 25, 2023 at 12:40 PM Jason Gunthorpe <jgg@xxxxxxxxxx> wrote: > > On Fri, Aug 25, 2023 at 11:49:53AM -0700, Nick Desaulniers wrote: > > I think this example demonstrates a little clearer what's going on: > > https://godbolt.org/z/9d6scv1hE > > > > So based on the warning, it seems like symbol_get() can only be used > > before the parameter it's passed is defined, otherwise the "weak" and > > "visibility(hidden)" attributes are ignored. These attributes are > > merge-able upon redeclaration up until the first definition. > > No, that doesn't make sense. The macro does: > > #define symbol_get(x) ({ extern typeof(x) x __attribute__((weak,visibility("hidden"))); &(x); }) > > x must be defined for typeof(x) to work. x must be //declared// for typeof(x) to work, but x doesn't need to be //defined//. https://godbolt.org/z/qv6zPWf9K But I think it's kind of moot, given the kconfig discussion below. > > I think the issue is that: > > # CONFIG_VFIO is not set > CONFIG_KVM_VFIO=y > > And > > kvm-$(CONFIG_KVM_VFIO) += $(KVM)/vfio.o > > Which is a combination that doesn't make any sense. > > Looks like CONFIG_KVM_VFIO should probably be called CONFIG_KVM_ARCH_VFIO > > And then > > config KVM_VFIO > bool > depends on KVM_ARCH_VFIO > depends on VFIO > > (or similar) > > So we don't even attempt to compile kvm/vfio.c if we don't have VFIO > support turned on. Do we need to split the Kconfig? ``` diff --git a/virt/kvm/Kconfig b/virt/kvm/Kconfig index 484d0873061c..e5ebf89de855 100644 --- a/virt/kvm/Kconfig +++ b/virt/kvm/Kconfig @@ -61,6 +61,7 @@ config HAVE_KVM_CPU_RELAX_INTERCEPT config KVM_VFIO bool + depends on VFIO config HAVE_KVM_INVALID_WAKEUPS bool ``` $ ./scripts/config -e VFIO $ make LLVM=1 ARCH=powerpc -j128 olddefconfig arch/powerpc/kvm/../../../virt/kvm/vfio.o # # configuration written to .config # SYNC include/config/auto.conf.cmd CALL scripts/checksyscalls.sh CC arch/powerpc/kvm/../../../virt/kvm/vfio.o $ -- Thanks, ~Nick Desaulniers
