Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 1 | #include <linux/linkage.h> |
| 2 | #include <linux/lguest.h> |
| 3 | #include <asm/asm-offsets.h> |
| 4 | #include <asm/thread_info.h> |
Rusty Russell | 876be9d | 2007-07-20 22:12:56 +1000 | [diff] [blame] | 5 | #include <asm/processor-flags.h> |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 6 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 7 | /*G:020 This is where we begin: we have a magic signature which the launcher |
| 8 | * looks for. The plan is that the Linux boot protocol will be extended with a |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 9 | * "platform type" field which will guide us here from the normal entry point, |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 10 | * but for the moment this suffices. The normal boot code uses %esi for the |
| 11 | * boot header, so we do too. We convert it to a virtual address by adding |
| 12 | * PAGE_OFFSET, and hand it to lguest_init() as its argument (ie. %eax). |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 13 | * |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 14 | * The .section line puts this code in .init.text so it will be discarded after |
| 15 | * boot. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 16 | .section .init.text, "ax", @progbits |
| 17 | .ascii "GenuineLguest" |
| 18 | /* Set up initial stack. */ |
| 19 | movl $(init_thread_union+THREAD_SIZE),%esp |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 20 | movl %esi, %eax |
| 21 | addl $__PAGE_OFFSET, %eax |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 22 | jmp lguest_init |
| 23 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 24 | /*G:055 We create a macro which puts the assembler code between lgstart_ and |
Rusty Russell | bbbd2bf | 2007-09-24 21:24:44 -0700 | [diff] [blame] | 25 | * lgend_ markers. These templates are put in the .text section: they can't be |
| 26 | * discarded after boot as we may need to patch modules, too. */ |
| 27 | .text |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 28 | #define LGUEST_PATCH(name, insns...) \ |
| 29 | lgstart_##name: insns; lgend_##name:; \ |
| 30 | .globl lgstart_##name; .globl lgend_##name |
| 31 | |
| 32 | LGUEST_PATCH(cli, movl $0, lguest_data+LGUEST_DATA_irq_enabled) |
| 33 | LGUEST_PATCH(sti, movl $X86_EFLAGS_IF, lguest_data+LGUEST_DATA_irq_enabled) |
| 34 | LGUEST_PATCH(popf, movl %eax, lguest_data+LGUEST_DATA_irq_enabled) |
| 35 | LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax) |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 36 | /*:*/ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 37 | |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 38 | /* These demark the EIP range where host should never deliver interrupts. */ |
| 39 | .global lguest_noirq_start |
| 40 | .global lguest_noirq_end |
| 41 | |
Rusty Russell | f56a384 | 2007-07-26 10:41:05 -0700 | [diff] [blame] | 42 | /*M:004 When the Host reflects a trap or injects an interrupt into the Guest, |
| 43 | * it sets the eflags interrupt bit on the stack based on |
| 44 | * lguest_data.irq_enabled, so the Guest iret logic does the right thing when |
| 45 | * restoring it. However, when the Host sets the Guest up for direct traps, |
| 46 | * such as system calls, the processor is the one to push eflags onto the |
| 47 | * stack, and the interrupt bit will be 1 (in reality, interrupts are always |
| 48 | * enabled in the Guest). |
| 49 | * |
| 50 | * This turns out to be harmless: the only trap which should happen under Linux |
| 51 | * with interrupts disabled is Page Fault (due to our lazy mapping of vmalloc |
| 52 | * regions), which has to be reflected through the Host anyway. If another |
| 53 | * trap *does* go off when interrupts are disabled, the Guest will panic, and |
| 54 | * we'll never get to this iret! :*/ |
| 55 | |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 56 | /*G:045 There is one final paravirt_op that the Guest implements, and glancing |
| 57 | * at it you can see why I left it to last. It's *cool*! It's in *assembler*! |
| 58 | * |
| 59 | * The "iret" instruction is used to return from an interrupt or trap. The |
| 60 | * stack looks like this: |
| 61 | * old address |
| 62 | * old code segment & privilege level |
| 63 | * old processor flags ("eflags") |
| 64 | * |
| 65 | * The "iret" instruction pops those values off the stack and restores them all |
| 66 | * at once. The only problem is that eflags includes the Interrupt Flag which |
| 67 | * the Guest can't change: the CPU will simply ignore it when we do an "iret". |
| 68 | * So we have to copy eflags from the stack to lguest_data.irq_enabled before |
| 69 | * we do the "iret". |
| 70 | * |
| 71 | * There are two problems with this: firstly, we need to use a register to do |
| 72 | * the copy and secondly, the whole thing needs to be atomic. The first |
| 73 | * problem is easy to solve: push %eax on the stack so we can use it, and then |
| 74 | * restore it at the end just before the real "iret". |
| 75 | * |
| 76 | * The second is harder: copying eflags to lguest_data.irq_enabled will turn |
| 77 | * interrupts on before we're finished, so we could be interrupted before we |
| 78 | * return to userspace or wherever. Our solution to this is to surround the |
| 79 | * code with lguest_noirq_start: and lguest_noirq_end: labels. We tell the |
| 80 | * Host that it is *never* to interrupt us there, even if interrupts seem to be |
| 81 | * enabled. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 82 | ENTRY(lguest_iret) |
| 83 | pushl %eax |
| 84 | movl 12(%esp), %eax |
| 85 | lguest_noirq_start: |
Rusty Russell | b2b47c2 | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 86 | /* Note the %ss: segment prefix here. Normal data accesses use the |
| 87 | * "ds" segment, but that will have already been restored for whatever |
| 88 | * we're returning to (such as userspace): we can't trust it. The %ss: |
| 89 | * prefix makes sure we use the stack segment, which is still valid. */ |
Rusty Russell | 07ad157 | 2007-07-19 01:49:22 -0700 | [diff] [blame] | 90 | movl %eax,%ss:lguest_data+LGUEST_DATA_irq_enabled |
| 91 | popl %eax |
| 92 | iret |
| 93 | lguest_noirq_end: |