Andi Kleen wrote: > >> + jnz restore_all_enable_events # != 0 => reenable event delivery >> +#endif >> RESTORE_REGS >> addl $4, %esp >> 1: iret >> .section .fixup,"ax" >> iret_exc: >> - sti >> +#ifndef CONFIG_XEN >> + ENABLE_INTERRUPTS >> +#endif >> pushl $0 # no error code >> pushl $do_iret_error >> jmp error_code >> @@ -269,6 +317,7 @@ iret_exc: >> .long 1b,iret_exc >> .previous >> >> +#ifndef CONFIG_XEN >> ldt_ss: >> > > So are you sure that problem this ugly piece of code tries to work around > isn't in Xen kernels too? Or do you just not care? If yes add a comment. > This code would otherwise be broken. ENABLE_INTERRUPTS in Xen requires access to the data segment, and the data segment is not available at this point. Plus, it corrupts the %esi register. Hint - use %ebp. The LDT SS code is broken as well because the iret onto a 16-bit stack is a pretty crippling blow to transparency in this code. Then, you don't have data or even stack segments that are reliable for calling out to hypervisor assist code. We never really fixed this code either in our implementation, although we did consider several approaches. Leaving it out does break userspace applications. Zach
