Rusty Russell wrote: > I'd expect (1) higher-level additions to paravirt_ops, and (2) binary > patching similar to the VMI patches for interrupt operations. > > Name: Function Call Abstraction for Paravirtualization > Status: Booted on 2.6.16-rc2-git7 > Depends: Paravirt/abstract-sensitive.patch.gz > Signed-off-by: Rusty Russell <rusty@xxxxxxxxxxxxxxx> > > This patch does the dumbest possible replacement of paravirtualized > instructions: calls through a "paravirt_ops" structure. Currently > these are function implementations of native hardware: hypervisors > will override the ops structure with their own variants. > Well, performance will not be the best, and this native stub seems awkward (especially the debug register bit) - but as a first step, I think this is a fine direction. > > +#define IRET jmp *paravirt_ops+PARAVIRT_iret > +#define CLI pushl %eax; pushl %ecx; pushl %edx; call *paravirt_ops+PARAVIRT_irq_disable; popl %edx; popl %ecx; popl %eax > +#define STI pushl %eax; pushl %ecx; pushl %edx; call *paravirt_ops+PARAVIRT_irq_enable; popl %edx; popl %ecx; popl %eax > +#define STI_SYSEXIT jmp *paravirt_ops+PARAVIRT_irq_enable_sysexit > +#define GET_CR0 pushl %ecx; pushl %edx; call *paravirt_ops+PARAVIRT_read_cr0; popl %edx; popl %ecx > +#define RDMSR pushl %ecx; pushl %edx; pushl %ecx; call *paravirt_ops+PARAVIRT_read_msr; addl $4, %esp; popl %edx; popl %ecx > +#define WRMSR pushl %eax; pushl %ecx; pushl %edx; pushl %eax; pushl %ecx; call *paravirt_ops+PARAVIRT_write_msr; addl $8, %esp; popl %edx; popl %ecx > +#define CPUID pushl %edx; pushl %ecx; pushl %ebx; pushl %eax; leal $16(%esp), %eax; pushl %eax; leal $16(%esp), %eax; pushl %eax; leal $16(%esp), %eax; pushl %eax; leal $16(%esp), %eax; pushl %eax; call *paravirt_ops+PARAVIRT_cpuid; addl $16; popl %eax; popl %ebx; popl %ecx; popl %edx > + I'm pretty uncomfortable with this bit - the extra push / pops here are really redundant if you implement the CLI/STI in assembly. And it is pretty clear that they have to be for performance. That said, I'm ok with this as a transitory step to building a working interface first, and getting performance optimal second. Also, RDMSR / WRMSR need not be defined for assembly code - this was legacy from my original patch. I defined them there because of the non-C semantics of the instructions register input and output until I could get an inline assembler version. Zach
