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 This is where we begin: head.S notes that the boot header's platform
9 * type field is "1" (lguest), so calls us here. The boot header is in %esi.
11 * WARNING: be very careful here! We're running at addresses equal to physical
12 * addesses (around 0), not above PAGE_OFFSET as most code expectes
13 * (eg. 0xC0000000). Jumps are relative, so they're OK, but we can't touch any
16 * The .section line puts this code in .init.text so it will be discarded after
18 .section .init.text, "ax", @progbits
20 /* Make initial hypercall now, so we can set up the pagetables. */
21 movl $LHCALL_LGUEST_INIT, %eax
22 movl $lguest_data - __PAGE_OFFSET, %edx
23 int $LGUEST_TRAP_ENTRY
25 /* The Host put the toplevel pagetable in lguest_data.pgdir. The movsl
26 * instruction uses %esi implicitly. */
27 movl lguest_data - __PAGE_OFFSET + LGUEST_DATA_pgdir, %esi
29 /* Copy first 32 entries of page directory to __PAGE_OFFSET entries.
30 * This means the first 128M of kernel memory will be mapped at
31 * PAGE_OFFSET where the kernel expects to run. This will get it far
32 * enough through boot to switch to its own pagetables. */
35 addl $((__PAGE_OFFSET >> 22) * 4), %edi
39 /* Set up the initial stack so we can run C code. */
40 movl $(init_thread_union+THREAD_SIZE),%esp
42 /* Jumps are relative, and we're running __PAGE_OFFSET too low at the
44 jmp lguest_init+__PAGE_OFFSET
46 /*G:055 We create a macro which puts the assembler code between lgstart_ and
47 * lgend_ markers. These templates are put in the .text section: they can't be
48 * discarded after boot as we may need to patch modules, too. */
50 #define LGUEST_PATCH(name, insns...) \
51 lgstart_##name: insns; lgend_##name:; \
52 .globl lgstart_##name; .globl lgend_##name
54 LGUEST_PATCH(cli, movl $0, lguest_data+LGUEST_DATA_irq_enabled)
55 LGUEST_PATCH(sti, movl $X86_EFLAGS_IF, lguest_data+LGUEST_DATA_irq_enabled)
56 LGUEST_PATCH(popf, movl %eax, lguest_data+LGUEST_DATA_irq_enabled)
57 LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax)
60 /* These demark the EIP range where host should never deliver interrupts. */
61 .global lguest_noirq_start
62 .global lguest_noirq_end
64 /*M:004 When the Host reflects a trap or injects an interrupt into the Guest,
65 * it sets the eflags interrupt bit on the stack based on
66 * lguest_data.irq_enabled, so the Guest iret logic does the right thing when
67 * restoring it. However, when the Host sets the Guest up for direct traps,
68 * such as system calls, the processor is the one to push eflags onto the
69 * stack, and the interrupt bit will be 1 (in reality, interrupts are always
70 * enabled in the Guest).
72 * This turns out to be harmless: the only trap which should happen under Linux
73 * with interrupts disabled is Page Fault (due to our lazy mapping of vmalloc
74 * regions), which has to be reflected through the Host anyway. If another
75 * trap *does* go off when interrupts are disabled, the Guest will panic, and
76 * we'll never get to this iret! :*/
78 /*G:045 There is one final paravirt_op that the Guest implements, and glancing
79 * at it you can see why I left it to last. It's *cool*! It's in *assembler*!
81 * The "iret" instruction is used to return from an interrupt or trap. The
82 * stack looks like this:
84 * old code segment & privilege level
85 * old processor flags ("eflags")
87 * The "iret" instruction pops those values off the stack and restores them all
88 * at once. The only problem is that eflags includes the Interrupt Flag which
89 * the Guest can't change: the CPU will simply ignore it when we do an "iret".
90 * So we have to copy eflags from the stack to lguest_data.irq_enabled before
93 * There are two problems with this: firstly, we need to use a register to do
94 * the copy and secondly, the whole thing needs to be atomic. The first
95 * problem is easy to solve: push %eax on the stack so we can use it, and then
96 * restore it at the end just before the real "iret".
98 * The second is harder: copying eflags to lguest_data.irq_enabled will turn
99 * interrupts on before we're finished, so we could be interrupted before we
100 * return to userspace or wherever. Our solution to this is to surround the
101 * code with lguest_noirq_start: and lguest_noirq_end: labels. We tell the
102 * Host that it is *never* to interrupt us there, even if interrupts seem to be
108 /* Note the %ss: segment prefix here. Normal data accesses use the
109 * "ds" segment, but that will have already been restored for whatever
110 * we're returning to (such as userspace): we can't trust it. The %ss:
111 * prefix makes sure we use the stack segment, which is still valid. */
112 movl %eax,%ss:lguest_data+LGUEST_DATA_irq_enabled