diff options
Diffstat (limited to 'arch/x86/lguest/i386_head.S')
| -rw-r--r-- | arch/x86/lguest/i386_head.S | 112 |
1 files changed, 70 insertions, 42 deletions
diff --git a/arch/x86/lguest/i386_head.S b/arch/x86/lguest/i386_head.S index a9c8cfe61cd4..27eac0faee48 100644 --- a/arch/x86/lguest/i386_head.S +++ b/arch/x86/lguest/i386_head.S | |||
| @@ -5,7 +5,8 @@ | |||
| 5 | #include <asm/thread_info.h> | 5 | #include <asm/thread_info.h> |
| 6 | #include <asm/processor-flags.h> | 6 | #include <asm/processor-flags.h> |
| 7 | 7 | ||
| 8 | /*G:020 Our story starts with the kernel booting into startup_32 in | 8 | /*G:020 |
| 9 | * 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 | * arch/x86/kernel/head_32.S. It expects a boot header, which is created by |
| 10 | * the bootloader (the Launcher in our case). | 11 | * the bootloader (the Launcher in our case). |
| 11 | * | 12 | * |
| @@ -21,11 +22,14 @@ | |||
| 21 | * data without remembering to subtract __PAGE_OFFSET! | 22 | * data without remembering to subtract __PAGE_OFFSET! |
| 22 | * | 23 | * |
| 23 | * The .section line puts this code in .init.text so it will be discarded after | 24 | * The .section line puts this code in .init.text so it will be discarded after |
| 24 | * boot. */ | 25 | * boot. |
| 26 | */ | ||
| 25 | .section .init.text, "ax", @progbits | 27 | .section .init.text, "ax", @progbits |
| 26 | ENTRY(lguest_entry) | 28 | ENTRY(lguest_entry) |
| 27 | /* We make the "initialization" hypercall now to tell the Host about | 29 | /* |
| 28 | * us, and also find out where it put our page tables. */ | 30 | * We make the "initialization" hypercall now to tell the Host about |
| 31 | * us, and also find out where it put our page tables. | ||
| 32 | */ | ||
| 29 | movl $LHCALL_LGUEST_INIT, %eax | 33 | movl $LHCALL_LGUEST_INIT, %eax |
| 30 | movl $lguest_data - __PAGE_OFFSET, %ebx | 34 | movl $lguest_data - __PAGE_OFFSET, %ebx |
| 31 | .byte 0x0f,0x01,0xc1 /* KVM_HYPERCALL */ | 35 | .byte 0x0f,0x01,0xc1 /* KVM_HYPERCALL */ |
| @@ -33,13 +37,14 @@ ENTRY(lguest_entry) | |||
| 33 | /* Set up the initial stack so we can run C code. */ | 37 | /* Set up the initial stack so we can run C code. */ |
| 34 | movl $(init_thread_union+THREAD_SIZE),%esp | 38 | movl $(init_thread_union+THREAD_SIZE),%esp |
| 35 | 39 | ||
| 36 | /* Jumps are relative, and we're running __PAGE_OFFSET too low at the | 40 | /* Jumps are relative: we're running __PAGE_OFFSET too low. */ |
| 37 | * moment. */ | ||
| 38 | jmp lguest_init+__PAGE_OFFSET | 41 | jmp lguest_init+__PAGE_OFFSET |
| 39 | 42 | ||
| 40 | /*G:055 We create a macro which puts the assembler code between lgstart_ and | 43 | /*G:055 |
| 41 | * lgend_ markers. These templates are put in the .text section: they can't be | 44 | * We create a macro which puts the assembler code between lgstart_ and lgend_ |
| 42 | * discarded after boot as we may need to patch modules, too. */ | 45 | * markers. These templates are put in the .text section: they can't be |
| 46 | * discarded after boot as we may need to patch modules, too. | ||
| 47 | */ | ||
| 43 | .text | 48 | .text |
| 44 | #define LGUEST_PATCH(name, insns...) \ | 49 | #define LGUEST_PATCH(name, insns...) \ |
| 45 | lgstart_##name: insns; lgend_##name:; \ | 50 | lgstart_##name: insns; lgend_##name:; \ |
| @@ -48,83 +53,103 @@ ENTRY(lguest_entry) | |||
| 48 | LGUEST_PATCH(cli, movl $0, lguest_data+LGUEST_DATA_irq_enabled) | 53 | LGUEST_PATCH(cli, movl $0, lguest_data+LGUEST_DATA_irq_enabled) |
| 49 | LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax) | 54 | LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax) |
| 50 | 55 | ||
| 51 | /*G:033 But using those wrappers is inefficient (we'll see why that doesn't | 56 | /*G:033 |
| 52 | * matter for save_fl and irq_disable later). If we write our routines | 57 | * But using those wrappers is inefficient (we'll see why that doesn't matter |
| 53 | * carefully in assembler, we can avoid clobbering any registers and avoid | 58 | * for save_fl and irq_disable later). If we write our routines carefully in |
| 54 | * jumping through the wrapper functions. | 59 | * assembler, we can avoid clobbering any registers and avoid jumping through |
| 60 | * the wrapper functions. | ||
| 55 | * | 61 | * |
| 56 | * I skipped over our first piece of assembler, but this one is worth studying | 62 | * I skipped over our first piece of assembler, but this one is worth studying |
| 57 | * in a bit more detail so I'll describe in easy stages. First, the routine | 63 | * in a bit more detail so I'll describe in easy stages. First, the routine to |
| 58 | * to enable interrupts: */ | 64 | * enable interrupts: |
| 65 | */ | ||
| 59 | ENTRY(lg_irq_enable) | 66 | ENTRY(lg_irq_enable) |
| 60 | /* The reverse of irq_disable, this sets lguest_data.irq_enabled to | 67 | /* |
| 61 | * X86_EFLAGS_IF (ie. "Interrupts enabled"). */ | 68 | * The reverse of irq_disable, this sets lguest_data.irq_enabled to |
| 69 | * X86_EFLAGS_IF (ie. "Interrupts enabled"). | ||
| 70 | */ | ||
| 62 | movl $X86_EFLAGS_IF, lguest_data+LGUEST_DATA_irq_enabled | 71 | movl $X86_EFLAGS_IF, lguest_data+LGUEST_DATA_irq_enabled |
| 63 | /* But now we need to check if the Host wants to know: there might have | 72 | /* |
| 73 | * But now we need to check if the Host wants to know: there might have | ||
| 64 | * been interrupts waiting to be delivered, in which case it will have | 74 | * been interrupts waiting to be delivered, in which case it will have |
| 65 | * set lguest_data.irq_pending to X86_EFLAGS_IF. If it's not zero, we | 75 | * set lguest_data.irq_pending to X86_EFLAGS_IF. If it's not zero, we |
| 66 | * jump to send_interrupts, otherwise we're done. */ | 76 | * jump to send_interrupts, otherwise we're done. |
| 77 | */ | ||
| 67 | testl $0, lguest_data+LGUEST_DATA_irq_pending | 78 | testl $0, lguest_data+LGUEST_DATA_irq_pending |
| 68 | jnz send_interrupts | 79 | jnz send_interrupts |
| 69 | /* One cool thing about x86 is that you can do many things without using | 80 | /* |
| 81 | * One cool thing about x86 is that you can do many things without using | ||
| 70 | * a register. In this case, the normal path hasn't needed to save or | 82 | * a register. In this case, the normal path hasn't needed to save or |
| 71 | * restore any registers at all! */ | 83 | * restore any registers at all! |
| 84 | */ | ||
| 72 | ret | 85 | ret |
| 73 | send_interrupts: | 86 | send_interrupts: |
| 74 | /* OK, now we need a register: eax is used for the hypercall number, | 87 | /* |
| 88 | * OK, now we need a register: eax is used for the hypercall number, | ||
| 75 | * which is LHCALL_SEND_INTERRUPTS. | 89 | * which is LHCALL_SEND_INTERRUPTS. |
| 76 | * | 90 | * |
| 77 | * We used not to bother with this pending detection at all, which was | 91 | * We used not to bother with this pending detection at all, which was |
| 78 | * much simpler. Sooner or later the Host would realize it had to | 92 | * much simpler. Sooner or later the Host would realize it had to |
| 79 | * send us an interrupt. But that turns out to make performance 7 | 93 | * send us an interrupt. But that turns out to make performance 7 |
| 80 | * times worse on a simple tcp benchmark. So now we do this the hard | 94 | * times worse on a simple tcp benchmark. So now we do this the hard |
| 81 | * way. */ | 95 | * way. |
| 96 | */ | ||
| 82 | pushl %eax | 97 | pushl %eax |
| 83 | movl $LHCALL_SEND_INTERRUPTS, %eax | 98 | movl $LHCALL_SEND_INTERRUPTS, %eax |
| 84 | /* This is a vmcall instruction (same thing that KVM uses). Older | 99 | /* |
| 100 | * This is a vmcall instruction (same thing that KVM uses). Older | ||
| 85 | * assembler versions might not know the "vmcall" instruction, so we | 101 | * assembler versions might not know the "vmcall" instruction, so we |
| 86 | * create one manually here. */ | 102 | * create one manually here. |
| 103 | */ | ||
| 87 | .byte 0x0f,0x01,0xc1 /* KVM_HYPERCALL */ | 104 | .byte 0x0f,0x01,0xc1 /* KVM_HYPERCALL */ |
| 105 | /* Put eax back the way we found it. */ | ||
| 88 | popl %eax | 106 | popl %eax |
| 89 | ret | 107 | ret |
| 90 | 108 | ||
| 91 | /* Finally, the "popf" or "restore flags" routine. The %eax register holds the | 109 | /* |
| 110 | * Finally, the "popf" or "restore flags" routine. The %eax register holds the | ||
| 92 | * flags (in practice, either X86_EFLAGS_IF or 0): if it's X86_EFLAGS_IF we're | 111 | * flags (in practice, either X86_EFLAGS_IF or 0): if it's X86_EFLAGS_IF we're |
| 93 | * enabling interrupts again, if it's 0 we're leaving them off. */ | 112 | * enabling interrupts again, if it's 0 we're leaving them off. |
| 113 | */ | ||
| 94 | ENTRY(lg_restore_fl) | 114 | ENTRY(lg_restore_fl) |
| 95 | /* This is just "lguest_data.irq_enabled = flags;" */ | 115 | /* This is just "lguest_data.irq_enabled = flags;" */ |
| 96 | movl %eax, lguest_data+LGUEST_DATA_irq_enabled | 116 | movl %eax, lguest_data+LGUEST_DATA_irq_enabled |
| 97 | /* Now, if the %eax value has enabled interrupts and | 117 | /* |
| 118 | * Now, if the %eax value has enabled interrupts and | ||
| 98 | * lguest_data.irq_pending is set, we want to tell the Host so it can | 119 | * lguest_data.irq_pending is set, we want to tell the Host so it can |
| 99 | * deliver any outstanding interrupts. Fortunately, both values will | 120 | * deliver any outstanding interrupts. Fortunately, both values will |
| 100 | * be X86_EFLAGS_IF (ie. 512) in that case, and the "testl" | 121 | * be X86_EFLAGS_IF (ie. 512) in that case, and the "testl" |
| 101 | * instruction will AND them together for us. If both are set, we | 122 | * instruction will AND them together for us. If both are set, we |
| 102 | * jump to send_interrupts. */ | 123 | * jump to send_interrupts. |
| 124 | */ | ||
| 103 | testl lguest_data+LGUEST_DATA_irq_pending, %eax | 125 | testl lguest_data+LGUEST_DATA_irq_pending, %eax |
| 104 | jnz send_interrupts | 126 | jnz send_interrupts |
| 105 | /* Again, the normal path has used no extra registers. Clever, huh? */ | 127 | /* Again, the normal path has used no extra registers. Clever, huh? */ |
| 106 | ret | 128 | ret |
| 129 | /*:*/ | ||
| 107 | 130 | ||
| 108 | /* These demark the EIP range where host should never deliver interrupts. */ | 131 | /* These demark the EIP range where host should never deliver interrupts. */ |
| 109 | .global lguest_noirq_start | 132 | .global lguest_noirq_start |
| 110 | .global lguest_noirq_end | 133 | .global lguest_noirq_end |
| 111 | 134 | ||
| 112 | /*M:004 When the Host reflects a trap or injects an interrupt into the Guest, | 135 | /*M:004 |
| 113 | * it sets the eflags interrupt bit on the stack based on | 136 | * When the Host reflects a trap or injects an interrupt into the Guest, it |
| 114 | * lguest_data.irq_enabled, so the Guest iret logic does the right thing when | 137 | * sets the eflags interrupt bit on the stack based on lguest_data.irq_enabled, |
| 115 | * restoring it. However, when the Host sets the Guest up for direct traps, | 138 | * so the Guest iret logic does the right thing when restoring it. However, |
| 116 | * such as system calls, the processor is the one to push eflags onto the | 139 | * when the Host sets the Guest up for direct traps, such as system calls, the |
| 117 | * stack, and the interrupt bit will be 1 (in reality, interrupts are always | 140 | * processor is the one to push eflags onto the stack, and the interrupt bit |
| 118 | * enabled in the Guest). | 141 | * will be 1 (in reality, interrupts are always enabled in the Guest). |
| 119 | * | 142 | * |
| 120 | * This turns out to be harmless: the only trap which should happen under Linux | 143 | * This turns out to be harmless: the only trap which should happen under Linux |
| 121 | * with interrupts disabled is Page Fault (due to our lazy mapping of vmalloc | 144 | * with interrupts disabled is Page Fault (due to our lazy mapping of vmalloc |
| 122 | * regions), which has to be reflected through the Host anyway. If another | 145 | * regions), which has to be reflected through the Host anyway. If another |
| 123 | * trap *does* go off when interrupts are disabled, the Guest will panic, and | 146 | * trap *does* go off when interrupts are disabled, the Guest will panic, and |
| 124 | * we'll never get to this iret! :*/ | 147 | * we'll never get to this iret! |
| 148 | :*/ | ||
| 125 | 149 | ||
| 126 | /*G:045 There is one final paravirt_op that the Guest implements, and glancing | 150 | /*G:045 |
| 127 | * at it you can see why I left it to last. It's *cool*! It's in *assembler*! | 151 | * There is one final paravirt_op that the Guest implements, and glancing at it |
| 152 | * you can see why I left it to last. It's *cool*! It's in *assembler*! | ||
| 128 | * | 153 | * |
| 129 | * The "iret" instruction is used to return from an interrupt or trap. The | 154 | * The "iret" instruction is used to return from an interrupt or trap. The |
| 130 | * stack looks like this: | 155 | * stack looks like this: |
| @@ -148,15 +173,18 @@ ENTRY(lg_restore_fl) | |||
| 148 | * return to userspace or wherever. Our solution to this is to surround the | 173 | * return to userspace or wherever. Our solution to this is to surround the |
| 149 | * code with lguest_noirq_start: and lguest_noirq_end: labels. We tell the | 174 | * code with lguest_noirq_start: and lguest_noirq_end: labels. We tell the |
| 150 | * Host that it is *never* to interrupt us there, even if interrupts seem to be | 175 | * Host that it is *never* to interrupt us there, even if interrupts seem to be |
| 151 | * enabled. */ | 176 | * enabled. |
| 177 | */ | ||
| 152 | ENTRY(lguest_iret) | 178 | ENTRY(lguest_iret) |
| 153 | pushl %eax | 179 | pushl %eax |
| 154 | movl 12(%esp), %eax | 180 | movl 12(%esp), %eax |
| 155 | lguest_noirq_start: | 181 | lguest_noirq_start: |
| 156 | /* Note the %ss: segment prefix here. Normal data accesses use the | 182 | /* |
| 183 | * Note the %ss: segment prefix here. Normal data accesses use the | ||
| 157 | * "ds" segment, but that will have already been restored for whatever | 184 | * "ds" segment, but that will have already been restored for whatever |
| 158 | * we're returning to (such as userspace): we can't trust it. The %ss: | 185 | * we're returning to (such as userspace): we can't trust it. The %ss: |
| 159 | * prefix makes sure we use the stack segment, which is still valid. */ | 186 | * prefix makes sure we use the stack segment, which is still valid. |
| 187 | */ | ||
| 160 | movl %eax,%ss:lguest_data+LGUEST_DATA_irq_enabled | 188 | movl %eax,%ss:lguest_data+LGUEST_DATA_irq_enabled |
| 161 | popl %eax | 189 | popl %eax |
| 162 | iret | 190 | iret |
