2007-07-19 15:49:22 +07:00
|
|
|
#include <linux/linkage.h>
|
|
|
|
#include <linux/lguest.h>
|
|
|
|
#include <asm/asm-offsets.h>
|
|
|
|
#include <asm/thread_info.h>
|
2007-07-20 19:12:56 +07:00
|
|
|
#include <asm/processor-flags.h>
|
2007-07-19 15:49:22 +07:00
|
|
|
|
2007-07-27 00:41:02 +07:00
|
|
|
/*G:020 This is where we begin: we have a magic signature which the launcher
|
|
|
|
* looks for. The plan is that the Linux boot protocol will be extended with a
|
2007-07-19 15:49:22 +07:00
|
|
|
* "platform type" field which will guide us here from the normal entry point,
|
2007-07-27 00:41:02 +07:00
|
|
|
* but for the moment this suffices. The normal boot code uses %esi for the
|
|
|
|
* boot header, so we do too. We convert it to a virtual address by adding
|
|
|
|
* PAGE_OFFSET, and hand it to lguest_init() as its argument (ie. %eax).
|
2007-07-19 15:49:22 +07:00
|
|
|
*
|
2007-07-27 00:41:02 +07:00
|
|
|
* The .section line puts this code in .init.text so it will be discarded after
|
|
|
|
* boot. */
|
2007-07-19 15:49:22 +07:00
|
|
|
.section .init.text, "ax", @progbits
|
|
|
|
.ascii "GenuineLguest"
|
|
|
|
/* Set up initial stack. */
|
2007-10-22 08:01:54 +07:00
|
|
|
movl $(init_thread_union+THREAD_SIZE),%esp
|
2007-07-19 15:49:23 +07:00
|
|
|
movl %esi, %eax
|
|
|
|
addl $__PAGE_OFFSET, %eax
|
2007-07-19 15:49:22 +07:00
|
|
|
jmp lguest_init
|
|
|
|
|
2007-07-27 00:41:02 +07:00
|
|
|
/*G:055 We create a macro which puts the assembler code between lgstart_ and
|
2007-09-25 11:24:44 +07:00
|
|
|
* lgend_ markers. These templates are put in the .text section: they can't be
|
|
|
|
* discarded after boot as we may need to patch modules, too. */
|
|
|
|
.text
|
2007-07-19 15:49:22 +07:00
|
|
|
#define LGUEST_PATCH(name, insns...) \
|
|
|
|
lgstart_##name: insns; lgend_##name:; \
|
|
|
|
.globl lgstart_##name; .globl lgend_##name
|
|
|
|
|
|
|
|
LGUEST_PATCH(cli, movl $0, lguest_data+LGUEST_DATA_irq_enabled)
|
|
|
|
LGUEST_PATCH(sti, movl $X86_EFLAGS_IF, lguest_data+LGUEST_DATA_irq_enabled)
|
|
|
|
LGUEST_PATCH(popf, movl %eax, lguest_data+LGUEST_DATA_irq_enabled)
|
|
|
|
LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax)
|
2007-07-27 00:41:02 +07:00
|
|
|
/*:*/
|
2007-07-19 15:49:22 +07:00
|
|
|
|
|
|
|
/* These demark the EIP range where host should never deliver interrupts. */
|
|
|
|
.global lguest_noirq_start
|
|
|
|
.global lguest_noirq_end
|
|
|
|
|
2007-07-27 00:41:05 +07:00
|
|
|
/*M:004 When the Host reflects a trap or injects an interrupt into the Guest,
|
|
|
|
* it sets the eflags interrupt bit on the stack based on
|
|
|
|
* lguest_data.irq_enabled, so the Guest iret logic does the right thing when
|
|
|
|
* restoring it. However, when the Host sets the Guest up for direct traps,
|
|
|
|
* such as system calls, the processor is the one to push eflags onto the
|
|
|
|
* stack, and the interrupt bit will be 1 (in reality, interrupts are always
|
|
|
|
* enabled in the Guest).
|
|
|
|
*
|
|
|
|
* This turns out to be harmless: the only trap which should happen under Linux
|
|
|
|
* with interrupts disabled is Page Fault (due to our lazy mapping of vmalloc
|
|
|
|
* regions), which has to be reflected through the Host anyway. If another
|
|
|
|
* trap *does* go off when interrupts are disabled, the Guest will panic, and
|
|
|
|
* we'll never get to this iret! :*/
|
|
|
|
|
2007-07-27 00:41:02 +07:00
|
|
|
/*G:045 There is one final paravirt_op that the Guest implements, and glancing
|
|
|
|
* at it you can see why I left it to last. It's *cool*! It's in *assembler*!
|
|
|
|
*
|
|
|
|
* The "iret" instruction is used to return from an interrupt or trap. The
|
|
|
|
* stack looks like this:
|
|
|
|
* old address
|
|
|
|
* old code segment & privilege level
|
|
|
|
* old processor flags ("eflags")
|
|
|
|
*
|
|
|
|
* The "iret" instruction pops those values off the stack and restores them all
|
|
|
|
* at once. The only problem is that eflags includes the Interrupt Flag which
|
|
|
|
* the Guest can't change: the CPU will simply ignore it when we do an "iret".
|
|
|
|
* So we have to copy eflags from the stack to lguest_data.irq_enabled before
|
|
|
|
* we do the "iret".
|
|
|
|
*
|
|
|
|
* There are two problems with this: firstly, we need to use a register to do
|
|
|
|
* the copy and secondly, the whole thing needs to be atomic. The first
|
|
|
|
* problem is easy to solve: push %eax on the stack so we can use it, and then
|
|
|
|
* restore it at the end just before the real "iret".
|
|
|
|
*
|
|
|
|
* The second is harder: copying eflags to lguest_data.irq_enabled will turn
|
|
|
|
* interrupts on before we're finished, so we could be interrupted before we
|
|
|
|
* return to userspace or wherever. Our solution to this is to surround the
|
|
|
|
* code with lguest_noirq_start: and lguest_noirq_end: labels. We tell the
|
|
|
|
* Host that it is *never* to interrupt us there, even if interrupts seem to be
|
|
|
|
* enabled. */
|
2007-07-19 15:49:22 +07:00
|
|
|
ENTRY(lguest_iret)
|
|
|
|
pushl %eax
|
|
|
|
movl 12(%esp), %eax
|
|
|
|
lguest_noirq_start:
|
2007-07-27 00:41:02 +07:00
|
|
|
/* Note the %ss: segment prefix here. Normal data accesses use the
|
|
|
|
* "ds" segment, but that will have already been restored for whatever
|
|
|
|
* we're returning to (such as userspace): we can't trust it. The %ss:
|
|
|
|
* prefix makes sure we use the stack segment, which is still valid. */
|
2007-07-19 15:49:22 +07:00
|
|
|
movl %eax,%ss:lguest_data+LGUEST_DATA_irq_enabled
|
|
|
|
popl %eax
|
|
|
|
iret
|
|
|
|
lguest_noirq_end:
|