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> 7 8/*G:020 9 * Our story starts with the kernel booting into startup_32 in 10 * arch/x86/kernel/head_32.S. It expects a boot header, which is created by 11 * the bootloader (the Launcher in our case). 12 * 13 * The startup_32 function does very little: it clears the uninitialized global 14 * C variables which we expect to be zero (ie. BSS) and then copies the boot 15 * header and kernel command line somewhere safe. Finally it checks the 16 * 'hardware_subarch' field. This was introduced in 2.6.24 for lguest and Xen: 17 * if it's set to '1' (lguest's assigned number), then it calls us here. 18 * 19 * WARNING: be very careful here! We're running at addresses equal to physical 20 * addesses (around 0), not above PAGE_OFFSET as most code expectes 21 * (eg. 0xC0000000). Jumps are relative, so they're OK, but we can't touch any 22 * data without remembering to subtract __PAGE_OFFSET! 23 * 24 * The .section line puts this code in .init.text so it will be discarded after 25 * boot. 26 */ 27.section .init.text, "ax", @progbits 28ENTRY(lguest_entry) 29 /* 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 */ 33 movl $LHCALL_LGUEST_INIT, %eax 34 movl $lguest_data - __PAGE_OFFSET, %ebx 35 int $LGUEST_TRAP_ENTRY 36 37 /* Set up the initial stack so we can run C code. */ 38 movl $(init_thread_union+THREAD_SIZE),%esp 39 40 /* Jumps are relative: we're running __PAGE_OFFSET too low. */ 41 jmp lguest_init+__PAGE_OFFSET 42 43/*G:055 44 * We create a macro which puts the assembler code between lgstart_ and lgend_ 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 */ 48.text 49#define LGUEST_PATCH(name, insns...) \ 50 lgstart_##name: insns; lgend_##name:; \ 51 .globl lgstart_##name; .globl lgend_##name 52 53LGUEST_PATCH(cli, movl $0, lguest_data+LGUEST_DATA_irq_enabled) 54LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax) 55 56/*G:033 57 * But using those wrappers is inefficient (we'll see why that doesn't matter 58 * for save_fl and irq_disable later). If we write our routines carefully in 59 * assembler, we can avoid clobbering any registers and avoid jumping through 60 * the wrapper functions. 61 * 62 * I skipped over our first piece of assembler, but this one is worth studying 63 * in a bit more detail so I'll describe in easy stages. First, the routine to 64 * enable interrupts: 65 */ 66ENTRY(lg_irq_enable) 67 /* 68 * The reverse of irq_disable, this sets lguest_data.irq_enabled to 69 * X86_EFLAGS_IF (ie. "Interrupts enabled"). 70 */ 71 movl $X86_EFLAGS_IF, lguest_data+LGUEST_DATA_irq_enabled 72 /* 73 * But now we need to check if the Host wants to know: there might have 74 * been interrupts waiting to be delivered, in which case it will have 75 * set lguest_data.irq_pending to X86_EFLAGS_IF. If it's not zero, we 76 * jump to send_interrupts, otherwise we're done. 77 */ 78 testl $0, lguest_data+LGUEST_DATA_irq_pending 79 jnz send_interrupts 80 /* 81 * One cool thing about x86 is that you can do many things without using 82 * a register. In this case, the normal path hasn't needed to save or 83 * restore any registers at all! 84 */ 85 ret 86send_interrupts: 87 /* 88 * OK, now we need a register: eax is used for the hypercall number, 89 * which is LHCALL_SEND_INTERRUPTS. 90 * 91 * We used not to bother with this pending detection at all, which was 92 * much simpler. Sooner or later the Host would realize it had to 93 * send us an interrupt. But that turns out to make performance 7 94 * times worse on a simple tcp benchmark. So now we do this the hard 95 * way. 96 */ 97 pushl %eax 98 movl $LHCALL_SEND_INTERRUPTS, %eax 99 /* 100 * This is a vmcall instruction (same thing that KVM uses). Older 101 * assembler versions might not know the "vmcall" instruction, so we 102 * create one manually here. 103 */ 104 .byte 0x0f,0x01,0xc1 /* KVM_HYPERCALL */ 105 /* Put eax back the way we found it. */ 106 popl %eax 107 ret 108 109/* 110 * Finally, the "popf" or "restore flags" routine. The %eax register holds the 111 * flags (in practice, either X86_EFLAGS_IF or 0): if it's X86_EFLAGS_IF we're 112 * enabling interrupts again, if it's 0 we're leaving them off. 113 */ 114ENTRY(lg_restore_fl) 115 /* This is just "lguest_data.irq_enabled = flags;" */ 116 movl %eax, lguest_data+LGUEST_DATA_irq_enabled 117 /* 118 * Now, if the %eax value has enabled interrupts and 119 * lguest_data.irq_pending is set, we want to tell the Host so it can 120 * deliver any outstanding interrupts. Fortunately, both values will 121 * be X86_EFLAGS_IF (ie. 512) in that case, and the "testl" 122 * instruction will AND them together for us. If both are set, we 123 * jump to send_interrupts. 124 */ 125 testl lguest_data+LGUEST_DATA_irq_pending, %eax 126 jnz send_interrupts 127 /* Again, the normal path has used no extra registers. Clever, huh? */ 128 ret 129/*:*/ 130 131/* These demark the EIP range where host should never deliver interrupts. */ 132.global lguest_noirq_start 133.global lguest_noirq_end 134 135/*M:004 136 * When the Host reflects a trap or injects an interrupt into the Guest, it 137 * sets the eflags interrupt bit on the stack based on lguest_data.irq_enabled, 138 * so the Guest iret logic does the right thing when restoring it. However, 139 * when the Host sets the Guest up for direct traps, such as system calls, the 140 * processor is the one to push eflags onto the stack, and the interrupt bit 141 * will be 1 (in reality, interrupts are always enabled in the Guest). 142 * 143 * This turns out to be harmless: the only trap which should happen under Linux 144 * with interrupts disabled is Page Fault (due to our lazy mapping of vmalloc 145 * regions), which has to be reflected through the Host anyway. If another 146 * trap *does* go off when interrupts are disabled, the Guest will panic, and 147 * we'll never get to this iret! 148:*/ 149 150/*G:045 151 * 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*! 153 * 154 * The "iret" instruction is used to return from an interrupt or trap. The 155 * stack looks like this: 156 * old address 157 * old code segment & privilege level 158 * old processor flags ("eflags") 159 * 160 * The "iret" instruction pops those values off the stack and restores them all 161 * at once. The only problem is that eflags includes the Interrupt Flag which 162 * the Guest can't change: the CPU will simply ignore it when we do an "iret". 163 * So we have to copy eflags from the stack to lguest_data.irq_enabled before 164 * we do the "iret". 165 * 166 * There are two problems with this: firstly, we need to use a register to do 167 * the copy and secondly, the whole thing needs to be atomic. The first 168 * problem is easy to solve: push %eax on the stack so we can use it, and then 169 * restore it at the end just before the real "iret". 170 * 171 * The second is harder: copying eflags to lguest_data.irq_enabled will turn 172 * interrupts on before we're finished, so we could be interrupted before we 173 * return to userspace or wherever. Our solution to this is to surround the 174 * code with lguest_noirq_start: and lguest_noirq_end: labels. We tell the 175 * Host that it is *never* to interrupt us there, even if interrupts seem to be 176 * enabled. 177 */ 178ENTRY(lguest_iret) 179 pushl %eax 180 movl 12(%esp), %eax 181lguest_noirq_start: 182 /* 183 * Note the %ss: segment prefix here. Normal data accesses use the 184 * "ds" segment, but that will have already been restored for whatever 185 * we're returning to (such as userspace): we can't trust it. The %ss: 186 * prefix makes sure we use the stack segment, which is still valid. 187 */ 188 movl %eax,%ss:lguest_data+LGUEST_DATA_irq_enabled 189 popl %eax 190 iret 191lguest_noirq_end: 192