lguest: documentation II: Guest
Documentation: The Guest
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff --git a/drivers/lguest/lguest_asm.S b/drivers/lguest/lguest_asm.S
index a3dbf22..3126ae9 100644
--- a/drivers/lguest/lguest_asm.S
+++ b/drivers/lguest/lguest_asm.S
@@ -4,15 +4,15 @@
#include <asm/thread_info.h>
#include <asm/processor-flags.h>
-/*
- * 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
+/*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
* "platform type" field which will guide us here from the normal entry point,
- * but for the moment this suffices. We pass the virtual address of the boot
- * info to lguest_init().
+ * 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).
*
- * We put it in .init.text will be discarded after boot.
- */
+ * The .section line puts this code in .init.text so it will be discarded after
+ * boot. */
.section .init.text, "ax", @progbits
.ascii "GenuineLguest"
/* Set up initial stack. */
@@ -21,7 +21,9 @@
addl $__PAGE_OFFSET, %eax
jmp lguest_init
-/* The templates for inline patching. */
+/*G:055 We create a macro which puts the assembler code between lgstart_ and
+ * lgend_ markers. These templates end up in the .init.text section, so they
+ * are discarded after boot. */
#define LGUEST_PATCH(name, insns...) \
lgstart_##name: insns; lgend_##name:; \
.globl lgstart_##name; .globl lgend_##name
@@ -30,24 +32,47 @@
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)
+/*:*/
.text
/* These demark the EIP range where host should never deliver interrupts. */
.global lguest_noirq_start
.global lguest_noirq_end
-/*
- * We move eflags word to lguest_data.irq_enabled to restore interrupt state.
- * For page faults, gpfs and virtual interrupts, the hypervisor has saved
- * eflags manually, otherwise it was delivered directly and so eflags reflects
- * the real machine IF state, ie. interrupts on. Since the kernel always dies
- * if it takes such a trap with interrupts disabled anyway, turning interrupts
- * back on unconditionally here is OK.
- */
+/*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. */
ENTRY(lguest_iret)
pushl %eax
movl 12(%esp), %eax
lguest_noirq_start:
+ /* 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. */
movl %eax,%ss:lguest_data+LGUEST_DATA_irq_enabled
popl %eax
iret