Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation. |
| 3 | * Copyright (C) 2007, Jes Sorensen <jes@sgi.com> SGI. |
| 4 | * |
| 5 | * This program is free software; you can redistribute it and/or modify |
| 6 | * it under the terms of the GNU General Public License as published by |
| 7 | * the Free Software Foundation; either version 2 of the License, or |
| 8 | * (at your option) any later version. |
| 9 | * |
| 10 | * This program is distributed in the hope that it will be useful, but |
| 11 | * WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 | * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or |
| 13 | * NON INFRINGEMENT. See the GNU General Public License for more |
| 14 | * details. |
| 15 | * |
| 16 | * You should have received a copy of the GNU General Public License |
| 17 | * along with this program; if not, write to the Free Software |
| 18 | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
| 19 | */ |
| 20 | #include <linux/kernel.h> |
| 21 | #include <linux/start_kernel.h> |
| 22 | #include <linux/string.h> |
| 23 | #include <linux/console.h> |
| 24 | #include <linux/screen_info.h> |
| 25 | #include <linux/irq.h> |
| 26 | #include <linux/interrupt.h> |
| 27 | #include <linux/clocksource.h> |
| 28 | #include <linux/clockchips.h> |
| 29 | #include <linux/cpu.h> |
| 30 | #include <linux/lguest.h> |
| 31 | #include <linux/lguest_launcher.h> |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 32 | #include <asm/paravirt.h> |
| 33 | #include <asm/param.h> |
| 34 | #include <asm/page.h> |
| 35 | #include <asm/pgtable.h> |
| 36 | #include <asm/desc.h> |
| 37 | #include <asm/setup.h> |
| 38 | #include <asm/lguest.h> |
| 39 | #include <asm/uaccess.h> |
| 40 | #include <asm/i387.h> |
| 41 | #include "../lg.h" |
| 42 | |
| 43 | static int cpu_had_pge; |
| 44 | |
| 45 | static struct { |
| 46 | unsigned long offset; |
| 47 | unsigned short segment; |
| 48 | } lguest_entry; |
| 49 | |
| 50 | /* Offset from where switcher.S was compiled to where we've copied it */ |
| 51 | static unsigned long switcher_offset(void) |
| 52 | { |
| 53 | return SWITCHER_ADDR - (unsigned long)start_switcher_text; |
| 54 | } |
| 55 | |
| 56 | /* This cpu's struct lguest_pages. */ |
| 57 | static struct lguest_pages *lguest_pages(unsigned int cpu) |
| 58 | { |
| 59 | return &(((struct lguest_pages *) |
| 60 | (SWITCHER_ADDR + SHARED_SWITCHER_PAGES*PAGE_SIZE))[cpu]); |
| 61 | } |
| 62 | |
Glauber de Oliveira Costa | c40a9f4 | 2008-01-17 19:11:20 -0200 | [diff] [blame] | 63 | static DEFINE_PER_CPU(struct lg_cpu *, last_cpu); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 64 | |
| 65 | /*S:010 |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 66 | * We approach the Switcher. |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 67 | * |
| 68 | * Remember that each CPU has two pages which are visible to the Guest when it |
| 69 | * runs on that CPU. This has to contain the state for that Guest: we copy the |
| 70 | * state in just before we run the Guest. |
| 71 | * |
| 72 | * Each Guest has "changed" flags which indicate what has changed in the Guest |
| 73 | * since it last ran. We saw this set in interrupts_and_traps.c and |
| 74 | * segments.c. |
| 75 | */ |
Glauber de Oliveira Costa | d0953d4 | 2008-01-07 11:05:25 -0200 | [diff] [blame] | 76 | static void copy_in_guest_info(struct lg_cpu *cpu, struct lguest_pages *pages) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 77 | { |
| 78 | /* Copying all this data can be quite expensive. We usually run the |
| 79 | * same Guest we ran last time (and that Guest hasn't run anywhere else |
| 80 | * meanwhile). If that's not the case, we pretend everything in the |
| 81 | * Guest has changed. */ |
Glauber de Oliveira Costa | f34f8c5 | 2008-01-17 19:13:26 -0200 | [diff] [blame] | 82 | if (__get_cpu_var(last_cpu) != cpu || cpu->last_pages != pages) { |
Glauber de Oliveira Costa | c40a9f4 | 2008-01-17 19:11:20 -0200 | [diff] [blame] | 83 | __get_cpu_var(last_cpu) = cpu; |
Glauber de Oliveira Costa | f34f8c5 | 2008-01-17 19:13:26 -0200 | [diff] [blame] | 84 | cpu->last_pages = pages; |
Glauber de Oliveira Costa | ae3749d | 2008-01-17 19:14:46 -0200 | [diff] [blame] | 85 | cpu->changed = CHANGED_ALL; |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 86 | } |
| 87 | |
| 88 | /* These copies are pretty cheap, so we do them unconditionally: */ |
| 89 | /* Save the current Host top-level page directory. */ |
| 90 | pages->state.host_cr3 = __pa(current->mm->pgd); |
| 91 | /* Set up the Guest's page tables to see this CPU's pages (and no |
| 92 | * other CPU's pages). */ |
Glauber de Oliveira Costa | 0c78441 | 2008-01-07 11:05:30 -0200 | [diff] [blame] | 93 | map_switcher_in_guest(cpu, pages); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 94 | /* Set up the two "TSS" members which tell the CPU what stack to use |
| 95 | * for traps which do directly into the Guest (ie. traps at privilege |
| 96 | * level 1). */ |
Glauber de Oliveira Costa | 4665ac8 | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 97 | pages->state.guest_tss.esp1 = cpu->esp1; |
| 98 | pages->state.guest_tss.ss1 = cpu->ss1; |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 99 | |
| 100 | /* Copy direct-to-Guest trap entries. */ |
Glauber de Oliveira Costa | ae3749d | 2008-01-17 19:14:46 -0200 | [diff] [blame] | 101 | if (cpu->changed & CHANGED_IDT) |
Glauber de Oliveira Costa | fc708b3 | 2008-01-07 11:05:33 -0200 | [diff] [blame] | 102 | copy_traps(cpu, pages->state.guest_idt, default_idt_entries); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 103 | |
| 104 | /* Copy all GDT entries which the Guest can change. */ |
Glauber de Oliveira Costa | ae3749d | 2008-01-17 19:14:46 -0200 | [diff] [blame] | 105 | if (cpu->changed & CHANGED_GDT) |
Glauber de Oliveira Costa | fc708b3 | 2008-01-07 11:05:33 -0200 | [diff] [blame] | 106 | copy_gdt(cpu, pages->state.guest_gdt); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 107 | /* If only the TLS entries have changed, copy them. */ |
Glauber de Oliveira Costa | ae3749d | 2008-01-17 19:14:46 -0200 | [diff] [blame] | 108 | else if (cpu->changed & CHANGED_GDT_TLS) |
Glauber de Oliveira Costa | fc708b3 | 2008-01-07 11:05:33 -0200 | [diff] [blame] | 109 | copy_gdt_tls(cpu, pages->state.guest_gdt); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 110 | |
| 111 | /* Mark the Guest as unchanged for next time. */ |
Glauber de Oliveira Costa | ae3749d | 2008-01-17 19:14:46 -0200 | [diff] [blame] | 112 | cpu->changed = 0; |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 113 | } |
| 114 | |
| 115 | /* Finally: the code to actually call into the Switcher to run the Guest. */ |
Glauber de Oliveira Costa | d0953d4 | 2008-01-07 11:05:25 -0200 | [diff] [blame] | 116 | static void run_guest_once(struct lg_cpu *cpu, struct lguest_pages *pages) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 117 | { |
| 118 | /* This is a dummy value we need for GCC's sake. */ |
| 119 | unsigned int clobber; |
| 120 | |
| 121 | /* Copy the guest-specific information into this CPU's "struct |
| 122 | * lguest_pages". */ |
Glauber de Oliveira Costa | d0953d4 | 2008-01-07 11:05:25 -0200 | [diff] [blame] | 123 | copy_in_guest_info(cpu, pages); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 124 | |
| 125 | /* Set the trap number to 256 (impossible value). If we fault while |
| 126 | * switching to the Guest (bad segment registers or bug), this will |
| 127 | * cause us to abort the Guest. */ |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 128 | cpu->regs->trapnum = 256; |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 129 | |
| 130 | /* Now: we push the "eflags" register on the stack, then do an "lcall". |
| 131 | * This is how we change from using the kernel code segment to using |
| 132 | * the dedicated lguest code segment, as well as jumping into the |
| 133 | * Switcher. |
| 134 | * |
| 135 | * The lcall also pushes the old code segment (KERNEL_CS) onto the |
| 136 | * stack, then the address of this call. This stack layout happens to |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 137 | * exactly match the stack layout created by an interrupt... */ |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 138 | asm volatile("pushf; lcall *lguest_entry" |
| 139 | /* This is how we tell GCC that %eax ("a") and %ebx ("b") |
| 140 | * are changed by this routine. The "=" means output. */ |
| 141 | : "=a"(clobber), "=b"(clobber) |
| 142 | /* %eax contains the pages pointer. ("0" refers to the |
| 143 | * 0-th argument above, ie "a"). %ebx contains the |
| 144 | * physical address of the Guest's top-level page |
| 145 | * directory. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame^] | 146 | : "0"(pages), "1"(__pa(cpu->lg->pgdirs[cpu->cpu_pgd].pgdir)) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 147 | /* We tell gcc that all these registers could change, |
| 148 | * which means we don't have to save and restore them in |
| 149 | * the Switcher. */ |
| 150 | : "memory", "%edx", "%ecx", "%edi", "%esi"); |
| 151 | } |
| 152 | /*:*/ |
| 153 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 154 | /*M:002 There are hooks in the scheduler which we can register to tell when we |
| 155 | * get kicked off the CPU (preempt_notifier_register()). This would allow us |
| 156 | * to lazily disable SYSENTER which would regain some performance, and should |
| 157 | * also simplify copy_in_guest_info(). Note that we'd still need to restore |
| 158 | * things when we exit to Launcher userspace, but that's fairly easy. |
| 159 | * |
| 160 | * The hooks were designed for KVM, but we can also put them to good use. :*/ |
| 161 | |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 162 | /*H:040 This is the i386-specific code to setup and run the Guest. Interrupts |
| 163 | * are disabled: we own the CPU. */ |
Glauber de Oliveira Costa | d0953d4 | 2008-01-07 11:05:25 -0200 | [diff] [blame] | 164 | void lguest_arch_run_guest(struct lg_cpu *cpu) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 165 | { |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 166 | /* Remember the awfully-named TS bit? If the Guest has asked to set it |
| 167 | * we set it now, so we can trap and pass that trap to the Guest if it |
| 168 | * uses the FPU. */ |
Glauber de Oliveira Costa | 4665ac8 | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 169 | if (cpu->ts) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 170 | lguest_set_ts(); |
| 171 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 172 | /* SYSENTER is an optimized way of doing system calls. We can't allow |
| 173 | * it because it always jumps to privilege level 0. A normal Guest |
| 174 | * won't try it because we don't advertise it in CPUID, but a malicious |
| 175 | * Guest (or malicious Guest userspace program) could, so we tell the |
| 176 | * CPU to disable it before running the Guest. */ |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 177 | if (boot_cpu_has(X86_FEATURE_SEP)) |
| 178 | wrmsr(MSR_IA32_SYSENTER_CS, 0, 0); |
| 179 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 180 | /* Now we actually run the Guest. It will return when something |
| 181 | * interesting happens, and we can examine its registers to see what it |
| 182 | * was doing. */ |
Glauber de Oliveira Costa | d0953d4 | 2008-01-07 11:05:25 -0200 | [diff] [blame] | 183 | run_guest_once(cpu, lguest_pages(raw_smp_processor_id())); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 184 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 185 | /* Note that the "regs" pointer contains two extra entries which are |
| 186 | * not really registers: a trap number which says what interrupt or |
| 187 | * trap made the switcher code come back, and an error code which some |
| 188 | * traps set. */ |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 189 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 190 | /* If the Guest page faulted, then the cr2 register will tell us the |
| 191 | * bad virtual address. We have to grab this now, because once we |
| 192 | * re-enable interrupts an interrupt could fault and thus overwrite |
| 193 | * cr2, or we could even move off to a different CPU. */ |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 194 | if (cpu->regs->trapnum == 14) |
Glauber de Oliveira Costa | fc708b3 | 2008-01-07 11:05:33 -0200 | [diff] [blame] | 195 | cpu->arch.last_pagefault = read_cr2(); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 196 | /* Similarly, if we took a trap because the Guest used the FPU, |
| 197 | * we have to restore the FPU it expects to see. */ |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 198 | else if (cpu->regs->trapnum == 7) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 199 | math_state_restore(); |
| 200 | |
| 201 | /* Restore SYSENTER if it's supposed to be on. */ |
| 202 | if (boot_cpu_has(X86_FEATURE_SEP)) |
| 203 | wrmsr(MSR_IA32_SYSENTER_CS, __KERNEL_CS, 0); |
| 204 | } |
| 205 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 206 | /*H:130 Now we've examined the hypercall code; our Guest can make requests. |
| 207 | * Our Guest is usually so well behaved; it never tries to do things it isn't |
| 208 | * allowed to, and uses hypercalls instead. Unfortunately, Linux's paravirtual |
| 209 | * infrastructure isn't quite complete, because it doesn't contain replacements |
| 210 | * for the Intel I/O instructions. As a result, the Guest sometimes fumbles |
| 211 | * across one during the boot process as it probes for various things which are |
| 212 | * usually attached to a PC. |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 213 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 214 | * When the Guest uses one of these instructions, we get a trap (General |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 215 | * Protection Fault) and come here. We see if it's one of those troublesome |
| 216 | * instructions and skip over it. We return true if we did. */ |
Glauber de Oliveira Costa | a3863f6 | 2008-01-07 11:05:31 -0200 | [diff] [blame] | 217 | static int emulate_insn(struct lg_cpu *cpu) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 218 | { |
| 219 | u8 insn; |
| 220 | unsigned int insnlen = 0, in = 0, shift = 0; |
| 221 | /* The eip contains the *virtual* address of the Guest's instruction: |
| 222 | * guest_pa just subtracts the Guest's page_offset. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame] | 223 | unsigned long physaddr = guest_pa(cpu, cpu->regs->eip); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 224 | |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 225 | /* This must be the Guest kernel trying to do something, not userspace! |
| 226 | * The bottom two bits of the CS segment register are the privilege |
| 227 | * level. */ |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 228 | if ((cpu->regs->cs & 3) != GUEST_PL) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 229 | return 0; |
| 230 | |
| 231 | /* Decoding x86 instructions is icky. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame^] | 232 | insn = lgread(cpu, physaddr, u8); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 233 | |
| 234 | /* 0x66 is an "operand prefix". It means it's using the upper 16 bits |
| 235 | of the eax register. */ |
| 236 | if (insn == 0x66) { |
| 237 | shift = 16; |
| 238 | /* The instruction is 1 byte so far, read the next byte. */ |
| 239 | insnlen = 1; |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame^] | 240 | insn = lgread(cpu, physaddr + insnlen, u8); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 241 | } |
| 242 | |
| 243 | /* We can ignore the lower bit for the moment and decode the 4 opcodes |
| 244 | * we need to emulate. */ |
| 245 | switch (insn & 0xFE) { |
| 246 | case 0xE4: /* in <next byte>,%al */ |
| 247 | insnlen += 2; |
| 248 | in = 1; |
| 249 | break; |
| 250 | case 0xEC: /* in (%dx),%al */ |
| 251 | insnlen += 1; |
| 252 | in = 1; |
| 253 | break; |
| 254 | case 0xE6: /* out %al,<next byte> */ |
| 255 | insnlen += 2; |
| 256 | break; |
| 257 | case 0xEE: /* out %al,(%dx) */ |
| 258 | insnlen += 1; |
| 259 | break; |
| 260 | default: |
| 261 | /* OK, we don't know what this is, can't emulate. */ |
| 262 | return 0; |
| 263 | } |
| 264 | |
| 265 | /* If it was an "IN" instruction, they expect the result to be read |
| 266 | * into %eax, so we change %eax. We always return all-ones, which |
| 267 | * traditionally means "there's nothing there". */ |
| 268 | if (in) { |
| 269 | /* Lower bit tells is whether it's a 16 or 32 bit access */ |
| 270 | if (insn & 0x1) |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 271 | cpu->regs->eax = 0xFFFFFFFF; |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 272 | else |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 273 | cpu->regs->eax |= (0xFFFF << shift); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 274 | } |
| 275 | /* Finally, we've "done" the instruction, so move past it. */ |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 276 | cpu->regs->eip += insnlen; |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 277 | /* Success! */ |
| 278 | return 1; |
| 279 | } |
| 280 | |
| 281 | /*H:050 Once we've re-enabled interrupts, we look at why the Guest exited. */ |
Glauber de Oliveira Costa | 73044f0 | 2008-01-07 11:05:27 -0200 | [diff] [blame] | 282 | void lguest_arch_handle_trap(struct lg_cpu *cpu) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 283 | { |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 284 | switch (cpu->regs->trapnum) { |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 285 | case 13: /* We've intercepted a General Protection Fault. */ |
| 286 | /* Check if this was one of those annoying IN or OUT |
| 287 | * instructions which we need to emulate. If so, we just go |
| 288 | * back into the Guest after we've done it. */ |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 289 | if (cpu->regs->errcode == 0) { |
Glauber de Oliveira Costa | a3863f6 | 2008-01-07 11:05:31 -0200 | [diff] [blame] | 290 | if (emulate_insn(cpu)) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 291 | return; |
| 292 | } |
| 293 | break; |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 294 | case 14: /* We've intercepted a Page Fault. */ |
| 295 | /* The Guest accessed a virtual address that wasn't mapped. |
| 296 | * This happens a lot: we don't actually set up most of the |
| 297 | * page tables for the Guest at all when we start: as it runs |
| 298 | * it asks for more and more, and we set them up as |
| 299 | * required. In this case, we don't even tell the Guest that |
| 300 | * the fault happened. |
| 301 | * |
| 302 | * The errcode tells whether this was a read or a write, and |
| 303 | * whether kernel or userspace code. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame] | 304 | if (demand_page(cpu, cpu->arch.last_pagefault, |
| 305 | cpu->regs->errcode)) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 306 | return; |
| 307 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 308 | /* OK, it's really not there (or not OK): the Guest needs to |
| 309 | * know. We write out the cr2 value so it knows where the |
| 310 | * fault occurred. |
| 311 | * |
| 312 | * Note that if the Guest were really messed up, this could |
| 313 | * happen before it's done the LHCALL_LGUEST_INIT hypercall, so |
| 314 | * lg->lguest_data could be NULL */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame^] | 315 | if (cpu->lg->lguest_data && |
| 316 | put_user(cpu->arch.last_pagefault, |
| 317 | &cpu->lg->lguest_data->cr2)) |
| 318 | kill_guest(cpu, "Writing cr2"); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 319 | break; |
| 320 | case 7: /* We've intercepted a Device Not Available fault. */ |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 321 | /* If the Guest doesn't want to know, we already restored the |
| 322 | * Floating Point Unit, so we just continue without telling |
| 323 | * it. */ |
Glauber de Oliveira Costa | 4665ac8 | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 324 | if (!cpu->ts) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 325 | return; |
| 326 | break; |
| 327 | case 32 ... 255: |
Rusty Russell | cc6d4fb | 2007-10-22 11:03:30 +1000 | [diff] [blame] | 328 | /* These values mean a real interrupt occurred, in which case |
| 329 | * the Host handler has already been run. We just do a |
| 330 | * friendly check if another process should now be run, then |
| 331 | * return to run the Guest again */ |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 332 | cond_resched(); |
Rusty Russell | cc6d4fb | 2007-10-22 11:03:30 +1000 | [diff] [blame] | 333 | return; |
| 334 | case LGUEST_TRAP_ENTRY: |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 335 | /* Our 'struct hcall_args' maps directly over our regs: we set |
| 336 | * up the pointer now to indicate a hypercall is pending. */ |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 337 | cpu->hcall = (struct hcall_args *)cpu->regs; |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 338 | return; |
| 339 | } |
| 340 | |
| 341 | /* We didn't handle the trap, so it needs to go to the Guest. */ |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 342 | if (!deliver_trap(cpu, cpu->regs->trapnum)) |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 343 | /* If the Guest doesn't have a handler (either it hasn't |
| 344 | * registered any yet, or it's one of the faults we don't let |
| 345 | * it handle), it dies with a cryptic error message. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame^] | 346 | kill_guest(cpu, "unhandled trap %li at %#lx (%#lx)", |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 347 | cpu->regs->trapnum, cpu->regs->eip, |
Glauber de Oliveira Costa | fc708b3 | 2008-01-07 11:05:33 -0200 | [diff] [blame] | 348 | cpu->regs->trapnum == 14 ? cpu->arch.last_pagefault |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 349 | : cpu->regs->errcode); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 350 | } |
| 351 | |
| 352 | /* Now we can look at each of the routines this calls, in increasing order of |
| 353 | * complexity: do_hypercalls(), emulate_insn(), maybe_do_interrupt(), |
| 354 | * deliver_trap() and demand_page(). After all those, we'll be ready to |
| 355 | * examine the Switcher, and our philosophical understanding of the Host/Guest |
| 356 | * duality will be complete. :*/ |
| 357 | static void adjust_pge(void *on) |
| 358 | { |
| 359 | if (on) |
| 360 | write_cr4(read_cr4() | X86_CR4_PGE); |
| 361 | else |
| 362 | write_cr4(read_cr4() & ~X86_CR4_PGE); |
| 363 | } |
| 364 | |
| 365 | /*H:020 Now the Switcher is mapped and every thing else is ready, we need to do |
| 366 | * some more i386-specific initialization. */ |
| 367 | void __init lguest_arch_host_init(void) |
| 368 | { |
| 369 | int i; |
| 370 | |
| 371 | /* Most of the i386/switcher.S doesn't care that it's been moved; on |
| 372 | * Intel, jumps are relative, and it doesn't access any references to |
| 373 | * external code or data. |
| 374 | * |
| 375 | * The only exception is the interrupt handlers in switcher.S: their |
| 376 | * addresses are placed in a table (default_idt_entries), so we need to |
| 377 | * update the table with the new addresses. switcher_offset() is a |
| 378 | * convenience function which returns the distance between the builtin |
| 379 | * switcher code and the high-mapped copy we just made. */ |
| 380 | for (i = 0; i < IDT_ENTRIES; i++) |
| 381 | default_idt_entries[i] += switcher_offset(); |
| 382 | |
| 383 | /* |
| 384 | * Set up the Switcher's per-cpu areas. |
| 385 | * |
| 386 | * Each CPU gets two pages of its own within the high-mapped region |
| 387 | * (aka. "struct lguest_pages"). Much of this can be initialized now, |
| 388 | * but some depends on what Guest we are running (which is set up in |
| 389 | * copy_in_guest_info()). |
| 390 | */ |
| 391 | for_each_possible_cpu(i) { |
| 392 | /* lguest_pages() returns this CPU's two pages. */ |
| 393 | struct lguest_pages *pages = lguest_pages(i); |
| 394 | /* This is a convenience pointer to make the code fit one |
| 395 | * statement to a line. */ |
| 396 | struct lguest_ro_state *state = &pages->state; |
| 397 | |
| 398 | /* The Global Descriptor Table: the Host has a different one |
| 399 | * for each CPU. We keep a descriptor for the GDT which says |
| 400 | * where it is and how big it is (the size is actually the last |
| 401 | * byte, not the size, hence the "-1"). */ |
| 402 | state->host_gdt_desc.size = GDT_SIZE-1; |
| 403 | state->host_gdt_desc.address = (long)get_cpu_gdt_table(i); |
| 404 | |
| 405 | /* All CPUs on the Host use the same Interrupt Descriptor |
| 406 | * Table, so we just use store_idt(), which gets this CPU's IDT |
| 407 | * descriptor. */ |
| 408 | store_idt(&state->host_idt_desc); |
| 409 | |
| 410 | /* The descriptors for the Guest's GDT and IDT can be filled |
| 411 | * out now, too. We copy the GDT & IDT into ->guest_gdt and |
| 412 | * ->guest_idt before actually running the Guest. */ |
| 413 | state->guest_idt_desc.size = sizeof(state->guest_idt)-1; |
| 414 | state->guest_idt_desc.address = (long)&state->guest_idt; |
| 415 | state->guest_gdt_desc.size = sizeof(state->guest_gdt)-1; |
| 416 | state->guest_gdt_desc.address = (long)&state->guest_gdt; |
| 417 | |
| 418 | /* We know where we want the stack to be when the Guest enters |
| 419 | * the switcher: in pages->regs. The stack grows upwards, so |
| 420 | * we start it at the end of that structure. */ |
| 421 | state->guest_tss.esp0 = (long)(&pages->regs + 1); |
| 422 | /* And this is the GDT entry to use for the stack: we keep a |
| 423 | * couple of special LGUEST entries. */ |
| 424 | state->guest_tss.ss0 = LGUEST_DS; |
| 425 | |
| 426 | /* x86 can have a finegrained bitmap which indicates what I/O |
| 427 | * ports the process can use. We set it to the end of our |
| 428 | * structure, meaning "none". */ |
| 429 | state->guest_tss.io_bitmap_base = sizeof(state->guest_tss); |
| 430 | |
| 431 | /* Some GDT entries are the same across all Guests, so we can |
| 432 | * set them up now. */ |
| 433 | setup_default_gdt_entries(state); |
| 434 | /* Most IDT entries are the same for all Guests, too.*/ |
| 435 | setup_default_idt_entries(state, default_idt_entries); |
| 436 | |
| 437 | /* The Host needs to be able to use the LGUEST segments on this |
| 438 | * CPU, too, so put them in the Host GDT. */ |
| 439 | get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT; |
| 440 | get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT; |
| 441 | } |
| 442 | |
| 443 | /* In the Switcher, we want the %cs segment register to use the |
| 444 | * LGUEST_CS GDT entry: we've put that in the Host and Guest GDTs, so |
| 445 | * it will be undisturbed when we switch. To change %cs and jump we |
| 446 | * need this structure to feed to Intel's "lcall" instruction. */ |
| 447 | lguest_entry.offset = (long)switch_to_guest + switcher_offset(); |
| 448 | lguest_entry.segment = LGUEST_CS; |
| 449 | |
| 450 | /* Finally, we need to turn off "Page Global Enable". PGE is an |
| 451 | * optimization where page table entries are specially marked to show |
| 452 | * they never change. The Host kernel marks all the kernel pages this |
| 453 | * way because it's always present, even when userspace is running. |
| 454 | * |
| 455 | * Lguest breaks this: unbeknownst to the rest of the Host kernel, we |
| 456 | * switch to the Guest kernel. If you don't disable this on all CPUs, |
| 457 | * you'll get really weird bugs that you'll chase for two days. |
| 458 | * |
| 459 | * I used to turn PGE off every time we switched to the Guest and back |
| 460 | * on when we return, but that slowed the Switcher down noticibly. */ |
| 461 | |
| 462 | /* We don't need the complexity of CPUs coming and going while we're |
| 463 | * doing this. */ |
Gautham R Shenoy | 86ef5c9 | 2008-01-25 21:08:02 +0100 | [diff] [blame] | 464 | get_online_cpus(); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 465 | if (cpu_has_pge) { /* We have a broader idea of "global". */ |
| 466 | /* Remember that this was originally set (for cleanup). */ |
| 467 | cpu_had_pge = 1; |
| 468 | /* adjust_pge is a helper function which sets or unsets the PGE |
| 469 | * bit on its CPU, depending on the argument (0 == unset). */ |
| 470 | on_each_cpu(adjust_pge, (void *)0, 0, 1); |
| 471 | /* Turn off the feature in the global feature set. */ |
| 472 | clear_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability); |
| 473 | } |
Gautham R Shenoy | 86ef5c9 | 2008-01-25 21:08:02 +0100 | [diff] [blame] | 474 | put_online_cpus(); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 475 | }; |
| 476 | /*:*/ |
| 477 | |
| 478 | void __exit lguest_arch_host_fini(void) |
| 479 | { |
| 480 | /* If we had PGE before we started, turn it back on now. */ |
Gautham R Shenoy | 86ef5c9 | 2008-01-25 21:08:02 +0100 | [diff] [blame] | 481 | get_online_cpus(); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 482 | if (cpu_had_pge) { |
| 483 | set_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability); |
| 484 | /* adjust_pge's argument "1" means set PGE. */ |
| 485 | on_each_cpu(adjust_pge, (void *)1, 0, 1); |
| 486 | } |
Gautham R Shenoy | 86ef5c9 | 2008-01-25 21:08:02 +0100 | [diff] [blame] | 487 | put_online_cpus(); |
Jes Sorensen | 625efab | 2007-10-22 11:03:28 +1000 | [diff] [blame] | 488 | } |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 489 | |
| 490 | |
| 491 | /*H:122 The i386-specific hypercalls simply farm out to the right functions. */ |
Glauber de Oliveira Costa | 73044f0 | 2008-01-07 11:05:27 -0200 | [diff] [blame] | 492 | int lguest_arch_do_hcall(struct lg_cpu *cpu, struct hcall_args *args) |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 493 | { |
| 494 | switch (args->arg0) { |
| 495 | case LHCALL_LOAD_GDT: |
Glauber de Oliveira Costa | fc708b3 | 2008-01-07 11:05:33 -0200 | [diff] [blame] | 496 | load_guest_gdt(cpu, args->arg1, args->arg2); |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 497 | break; |
| 498 | case LHCALL_LOAD_IDT_ENTRY: |
Glauber de Oliveira Costa | fc708b3 | 2008-01-07 11:05:33 -0200 | [diff] [blame] | 499 | load_guest_idt_entry(cpu, args->arg1, args->arg2, args->arg3); |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 500 | break; |
| 501 | case LHCALL_LOAD_TLS: |
Glauber de Oliveira Costa | fc708b3 | 2008-01-07 11:05:33 -0200 | [diff] [blame] | 502 | guest_load_tls(cpu, args->arg1); |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 503 | break; |
| 504 | default: |
| 505 | /* Bad Guest. Bad! */ |
| 506 | return -EIO; |
| 507 | } |
| 508 | return 0; |
| 509 | } |
| 510 | |
| 511 | /*H:126 i386-specific hypercall initialization: */ |
Glauber de Oliveira Costa | 73044f0 | 2008-01-07 11:05:27 -0200 | [diff] [blame] | 512 | int lguest_arch_init_hypercalls(struct lg_cpu *cpu) |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 513 | { |
| 514 | u32 tsc_speed; |
| 515 | |
| 516 | /* The pointer to the Guest's "struct lguest_data" is the only |
| 517 | * argument. We check that address now. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame^] | 518 | if (!lguest_address_ok(cpu->lg, cpu->hcall->arg1, |
| 519 | sizeof(*cpu->lg->lguest_data))) |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 520 | return -EFAULT; |
| 521 | |
| 522 | /* Having checked it, we simply set lg->lguest_data to point straight |
| 523 | * into the Launcher's memory at the right place and then use |
| 524 | * copy_to_user/from_user from now on, instead of lgread/write. I put |
| 525 | * this in to show that I'm not immune to writing stupid |
| 526 | * optimizations. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame^] | 527 | cpu->lg->lguest_data = cpu->lg->mem_base + cpu->hcall->arg1; |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 528 | |
| 529 | /* We insist that the Time Stamp Counter exist and doesn't change with |
| 530 | * cpu frequency. Some devious chip manufacturers decided that TSC |
| 531 | * changes could be handled in software. I decided that time going |
| 532 | * backwards might be good for benchmarks, but it's bad for users. |
| 533 | * |
| 534 | * We also insist that the TSC be stable: the kernel detects unreliable |
| 535 | * TSCs for its own purposes, and we use that here. */ |
| 536 | if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) && !check_tsc_unstable()) |
| 537 | tsc_speed = tsc_khz; |
| 538 | else |
| 539 | tsc_speed = 0; |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame^] | 540 | if (put_user(tsc_speed, &cpu->lg->lguest_data->tsc_khz)) |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 541 | return -EFAULT; |
| 542 | |
Rusty Russell | c18acd7 | 2007-10-22 11:03:35 +1000 | [diff] [blame] | 543 | /* The interrupt code might not like the system call vector. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame^] | 544 | if (!check_syscall_vector(cpu->lg)) |
| 545 | kill_guest(cpu, "bad syscall vector"); |
Rusty Russell | c18acd7 | 2007-10-22 11:03:35 +1000 | [diff] [blame] | 546 | |
Jes Sorensen | b410e7b | 2007-10-22 11:03:31 +1000 | [diff] [blame] | 547 | return 0; |
| 548 | } |
Jes Sorensen | d612cde | 2007-10-22 11:03:32 +1000 | [diff] [blame] | 549 | |
| 550 | /*L:030 lguest_arch_setup_regs() |
| 551 | * |
| 552 | * Most of the Guest's registers are left alone: we used get_zeroed_page() to |
| 553 | * allocate the structure, so they will be 0. */ |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 554 | void lguest_arch_setup_regs(struct lg_cpu *cpu, unsigned long start) |
Jes Sorensen | d612cde | 2007-10-22 11:03:32 +1000 | [diff] [blame] | 555 | { |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 556 | struct lguest_regs *regs = cpu->regs; |
Jes Sorensen | d612cde | 2007-10-22 11:03:32 +1000 | [diff] [blame] | 557 | |
| 558 | /* There are four "segment" registers which the Guest needs to boot: |
| 559 | * The "code segment" register (cs) refers to the kernel code segment |
| 560 | * __KERNEL_CS, and the "data", "extra" and "stack" segment registers |
| 561 | * refer to the kernel data segment __KERNEL_DS. |
| 562 | * |
| 563 | * The privilege level is packed into the lower bits. The Guest runs |
| 564 | * at privilege level 1 (GUEST_PL).*/ |
| 565 | regs->ds = regs->es = regs->ss = __KERNEL_DS|GUEST_PL; |
| 566 | regs->cs = __KERNEL_CS|GUEST_PL; |
| 567 | |
| 568 | /* The "eflags" register contains miscellaneous flags. Bit 1 (0x002) |
| 569 | * is supposed to always be "1". Bit 9 (0x200) controls whether |
| 570 | * interrupts are enabled. We always leave interrupts enabled while |
| 571 | * running the Guest. */ |
Rusty Russell | 25c47bb | 2007-10-25 14:09:53 +1000 | [diff] [blame] | 572 | regs->eflags = X86_EFLAGS_IF | 0x2; |
Jes Sorensen | d612cde | 2007-10-22 11:03:32 +1000 | [diff] [blame] | 573 | |
| 574 | /* The "Extended Instruction Pointer" register says where the Guest is |
| 575 | * running. */ |
| 576 | regs->eip = start; |
| 577 | |
| 578 | /* %esi points to our boot information, at physical address 0, so don't |
| 579 | * touch it. */ |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 580 | |
Jes Sorensen | d612cde | 2007-10-22 11:03:32 +1000 | [diff] [blame] | 581 | /* There are a couple of GDT entries the Guest expects when first |
| 582 | * booting. */ |
Glauber de Oliveira Costa | fc708b3 | 2008-01-07 11:05:33 -0200 | [diff] [blame] | 583 | setup_guest_gdt(cpu); |
Jes Sorensen | d612cde | 2007-10-22 11:03:32 +1000 | [diff] [blame] | 584 | } |