1 #include <linux/linkage.h>
2 #include <linux/lguest.h>
3 #include <asm/lguest_hcall.h>
4 #include <asm/asm-offsets.h>
5 #include <asm/thread_info.h>
6 #include <asm/processor-flags.h>
8 /*G:020 Our story starts with the kernel booting into startup_32 in
9 * arch/x86/kernel/head_32.S. It expects a boot header, which is created by
10 * the bootloader (the Launcher in our case).
12 * The startup_32 function does very little: it clears the uninitialized global
13 * C variables which we expect to be zero (ie. BSS) and then copies the boot
14 * header and kernel command line somewhere safe. Finally it checks the
15 * 'hardware_subarch' field. This was introduced in 2.6.24 for lguest and Xen:
16 * if it's set to '1' (lguest's assigned number), then it calls us here.
18 * WARNING: be very careful here! We're running at addresses equal to physical
19 * addesses (around 0), not above PAGE_OFFSET as most code expectes
20 * (eg. 0xC0000000). Jumps are relative, so they're OK, but we can't touch any
21 * data without remembering to subtract __PAGE_OFFSET!
23 * The .section line puts this code in .init.text so it will be discarded after
25 .section .init.text, "ax", @progbits
27 /* We make the "initialization" hypercall now to tell the Host about
28 * us, and also find out where it put our page tables. */
29 movl $LHCALL_LGUEST_INIT, %eax
30 movl $lguest_data - __PAGE_OFFSET, %ebx
31 .byte 0x0f,0x01,0xc1 /* KVM_HYPERCALL */
33 /* Set up the initial stack so we can run C code. */
34 movl $(init_thread_union+THREAD_SIZE),%esp
36 /* Jumps are relative, and we're running __PAGE_OFFSET too low at the
38 jmp lguest_init+__PAGE_OFFSET
40 /*G:055 We create a macro which puts the assembler code between lgstart_ and
41 * lgend_ markers. These templates are put in the .text section: they can't be
42 * discarded after boot as we may need to patch modules, too. */
44 #define LGUEST_PATCH(name, insns...) \
45 lgstart_##name: insns; lgend_##name:; \
46 .globl lgstart_##name; .globl lgend_##name
48 LGUEST_PATCH(cli, movl $0, lguest_data+LGUEST_DATA_irq_enabled)
49 LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax)
52 /* These demark the EIP range where host should never deliver interrupts. */
53 .global lguest_noirq_start
54 .global lguest_noirq_end
56 /*M:004 When the Host reflects a trap or injects an interrupt into the Guest,
57 * it sets the eflags interrupt bit on the stack based on
58 * lguest_data.irq_enabled, so the Guest iret logic does the right thing when
59 * restoring it. However, when the Host sets the Guest up for direct traps,
60 * such as system calls, the processor is the one to push eflags onto the
61 * stack, and the interrupt bit will be 1 (in reality, interrupts are always
62 * enabled in the Guest).
64 * This turns out to be harmless: the only trap which should happen under Linux
65 * with interrupts disabled is Page Fault (due to our lazy mapping of vmalloc
66 * regions), which has to be reflected through the Host anyway. If another
67 * trap *does* go off when interrupts are disabled, the Guest will panic, and
68 * we'll never get to this iret! :*/
70 /*G:045 There is one final paravirt_op that the Guest implements, and glancing
71 * at it you can see why I left it to last. It's *cool*! It's in *assembler*!
73 * The "iret" instruction is used to return from an interrupt or trap. The
74 * stack looks like this:
76 * old code segment & privilege level
77 * old processor flags ("eflags")
79 * The "iret" instruction pops those values off the stack and restores them all
80 * at once. The only problem is that eflags includes the Interrupt Flag which
81 * the Guest can't change: the CPU will simply ignore it when we do an "iret".
82 * So we have to copy eflags from the stack to lguest_data.irq_enabled before
85 * There are two problems with this: firstly, we need to use a register to do
86 * the copy and secondly, the whole thing needs to be atomic. The first
87 * problem is easy to solve: push %eax on the stack so we can use it, and then
88 * restore it at the end just before the real "iret".
90 * The second is harder: copying eflags to lguest_data.irq_enabled will turn
91 * interrupts on before we're finished, so we could be interrupted before we
92 * return to userspace or wherever. Our solution to this is to surround the
93 * code with lguest_noirq_start: and lguest_noirq_end: labels. We tell the
94 * Host that it is *never* to interrupt us there, even if interrupts seem to be
100 /* Note the %ss: segment prefix here. Normal data accesses use the
101 * "ds" segment, but that will have already been restored for whatever
102 * we're returning to (such as userspace): we can't trust it. The %ss:
103 * prefix makes sure we use the stack segment, which is still valid. */
104 movl %eax,%ss:lguest_data+LGUEST_DATA_irq_enabled