* Glauber de Oliveira Costa (gcosta@xxxxxxxxxx) wrote: > Only caveat, is that it has to be done before smp gets in the game, and > with interrupts disabled. (which makes the function in vsmp.c not eligible). > > My current option is to force VSMP to use PARAVIRT, as said before, and > then fill paravirt_arch_setup, which is currently unused, with code to > replace the needed paravirt_ops.fn. > > I don't know if there is any method to dynamically determine (at this > point) that we are in a vsmp arch, and if there are not, it will have to > get ifdefs anyway. But at least, they are far more local. between __cacheline_aligned_in_smp and other compile time bits based on VSMP specific INTERNODE_CACHE, etc. I think compile time the way to go. > I am okay with both, but after all the explanation, I don't think that > adding a new pvops is a bad idea. It would make things less cumbersome > in this case. Also, hacks like this save_fl may require changes to the > hypervisor, right? I don't even know where such hypervisor is, and how > easy it is to replace it (it may be deeply hidden in firmware) No hypervisor change needed. Just the pv backend needs to return 0 or X86_EFLAGS_IF for save_flags (and similar translation on restore_flags). Xen uses a simple shared memory flag and does something which you could roughly translate into this: xen_save_flags() if (xen_vcpu_interrupts_enabled) return X86_EFLAGS_IF; else return 0; This doesn't require any hypervisor changes. Similarly, VSMP could do something along the lines of: vsmp_save_flags() flags = native_save_flags(); if (flags & X86_EFLAGS_IF) || (flags & X86_EFLAGS_AC) return X86_EFLAGS_IF; else return 0; > A question raises here: Would vsmp turn paravirt_enabled to 1 ? Probably not. It's mostly native and I'm not sure it would want the bits disabled from if (paravirt_enabled()) tests. _______________________________________________ Virtualization mailing list Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/virtualization
