linux/arch/x86/lguest/boot.c
<<
>>
Prefs
   1/*P:010
   2 * A hypervisor allows multiple Operating Systems to run on a single machine.
   3 * To quote David Wheeler: "Any problem in computer science can be solved with
   4 * another layer of indirection."
   5 *
   6 * We keep things simple in two ways.  First, we start with a normal Linux
   7 * kernel and insert a module (lg.ko) which allows us to run other Linux
   8 * kernels the same way we'd run processes.  We call the first kernel the Host,
   9 * and the others the Guests.  The program which sets up and configures Guests
  10 * (such as the example in Documentation/lguest/lguest.c) is called the
  11 * Launcher.
  12 *
  13 * Secondly, we only run specially modified Guests, not normal kernels: setting
  14 * CONFIG_LGUEST_GUEST to "y" compiles this file into the kernel so it knows
  15 * how to be a Guest at boot time.  This means that you can use the same kernel
  16 * you boot normally (ie. as a Host) as a Guest.
  17 *
  18 * These Guests know that they cannot do privileged operations, such as disable
  19 * interrupts, and that they have to ask the Host to do such things explicitly.
  20 * This file consists of all the replacements for such low-level native
  21 * hardware operations: these special Guest versions call the Host.
  22 *
  23 * So how does the kernel know it's a Guest?  We'll see that later, but let's
  24 * just say that we end up here where we replace the native functions various
  25 * "paravirt" structures with our Guest versions, then boot like normal. :*/
  26
  27/*
  28 * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation.
  29 *
  30 * This program is free software; you can redistribute it and/or modify
  31 * it under the terms of the GNU General Public License as published by
  32 * the Free Software Foundation; either version 2 of the License, or
  33 * (at your option) any later version.
  34 *
  35 * This program is distributed in the hope that it will be useful, but
  36 * WITHOUT ANY WARRANTY; without even the implied warranty of
  37 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  38 * NON INFRINGEMENT.  See the GNU General Public License for more
  39 * details.
  40 *
  41 * You should have received a copy of the GNU General Public License
  42 * along with this program; if not, write to the Free Software
  43 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  44 */
  45#include <linux/kernel.h>
  46#include <linux/start_kernel.h>
  47#include <linux/string.h>
  48#include <linux/console.h>
  49#include <linux/screen_info.h>
  50#include <linux/irq.h>
  51#include <linux/interrupt.h>
  52#include <linux/clocksource.h>
  53#include <linux/clockchips.h>
  54#include <linux/lguest.h>
  55#include <linux/lguest_launcher.h>
  56#include <linux/virtio_console.h>
  57#include <linux/pm.h>
  58#include <asm/apic.h>
  59#include <asm/lguest.h>
  60#include <asm/paravirt.h>
  61#include <asm/param.h>
  62#include <asm/page.h>
  63#include <asm/pgtable.h>
  64#include <asm/desc.h>
  65#include <asm/setup.h>
  66#include <asm/e820.h>
  67#include <asm/mce.h>
  68#include <asm/io.h>
  69#include <asm/i387.h>
  70#include <asm/reboot.h>         /* for struct machine_ops */
  71
  72/*G:010 Welcome to the Guest!
  73 *
  74 * The Guest in our tale is a simple creature: identical to the Host but
  75 * behaving in simplified but equivalent ways.  In particular, the Guest is the
  76 * same kernel as the Host (or at least, built from the same source code). :*/
  77
  78struct lguest_data lguest_data = {
  79        .hcall_status = { [0 ... LHCALL_RING_SIZE-1] = 0xFF },
  80        .noirq_start = (u32)lguest_noirq_start,
  81        .noirq_end = (u32)lguest_noirq_end,
  82        .kernel_address = PAGE_OFFSET,
  83        .blocked_interrupts = { 1 }, /* Block timer interrupts */
  84        .syscall_vec = SYSCALL_VECTOR,
  85};
  86
  87/*G:037 async_hcall() is pretty simple: I'm quite proud of it really.  We have a
  88 * ring buffer of stored hypercalls which the Host will run though next time we
  89 * do a normal hypercall.  Each entry in the ring has 4 slots for the hypercall
  90 * arguments, and a "hcall_status" word which is 0 if the call is ready to go,
  91 * and 255 once the Host has finished with it.
  92 *
  93 * If we come around to a slot which hasn't been finished, then the table is
  94 * full and we just make the hypercall directly.  This has the nice side
  95 * effect of causing the Host to run all the stored calls in the ring buffer
  96 * which empties it for next time! */
  97static void async_hcall(unsigned long call, unsigned long arg1,
  98                        unsigned long arg2, unsigned long arg3)
  99{
 100        /* Note: This code assumes we're uniprocessor. */
 101        static unsigned int next_call;
 102        unsigned long flags;
 103
 104        /* Disable interrupts if not already disabled: we don't want an
 105         * interrupt handler making a hypercall while we're already doing
 106         * one! */
 107        local_irq_save(flags);
 108        if (lguest_data.hcall_status[next_call] != 0xFF) {
 109                /* Table full, so do normal hcall which will flush table. */
 110                hcall(call, arg1, arg2, arg3);
 111        } else {
 112                lguest_data.hcalls[next_call].arg0 = call;
 113                lguest_data.hcalls[next_call].arg1 = arg1;
 114                lguest_data.hcalls[next_call].arg2 = arg2;
 115                lguest_data.hcalls[next_call].arg3 = arg3;
 116                /* Arguments must all be written before we mark it to go */
 117                wmb();
 118                lguest_data.hcall_status[next_call] = 0;
 119                if (++next_call == LHCALL_RING_SIZE)
 120                        next_call = 0;
 121        }
 122        local_irq_restore(flags);
 123}
 124
 125/*G:035 Notice the lazy_hcall() above, rather than hcall().  This is our first
 126 * real optimization trick!
 127 *
 128 * When lazy_mode is set, it means we're allowed to defer all hypercalls and do
 129 * them as a batch when lazy_mode is eventually turned off.  Because hypercalls
 130 * are reasonably expensive, batching them up makes sense.  For example, a
 131 * large munmap might update dozens of page table entries: that code calls
 132 * paravirt_enter_lazy_mmu(), does the dozen updates, then calls
 133 * lguest_leave_lazy_mode().
 134 *
 135 * So, when we're in lazy mode, we call async_hcall() to store the call for
 136 * future processing: */
 137static void lazy_hcall(unsigned long call,
 138                       unsigned long arg1,
 139                       unsigned long arg2,
 140                       unsigned long arg3)
 141{
 142        if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
 143                hcall(call, arg1, arg2, arg3);
 144        else
 145                async_hcall(call, arg1, arg2, arg3);
 146}
 147
 148/* When lazy mode is turned off reset the per-cpu lazy mode variable and then
 149 * issue the do-nothing hypercall to flush any stored calls. */
 150static void lguest_leave_lazy_mode(void)
 151{
 152        paravirt_leave_lazy(paravirt_get_lazy_mode());
 153        hcall(LHCALL_FLUSH_ASYNC, 0, 0, 0);
 154}
 155
 156/*G:033
 157 * After that diversion we return to our first native-instruction
 158 * replacements: four functions for interrupt control.
 159 *
 160 * The simplest way of implementing these would be to have "turn interrupts
 161 * off" and "turn interrupts on" hypercalls.  Unfortunately, this is too slow:
 162 * these are by far the most commonly called functions of those we override.
 163 *
 164 * So instead we keep an "irq_enabled" field inside our "struct lguest_data",
 165 * which the Guest can update with a single instruction.  The Host knows to
 166 * check there before it tries to deliver an interrupt.
 167 */
 168
 169/* save_flags() is expected to return the processor state (ie. "flags").  The
 170 * flags word contains all kind of stuff, but in practice Linux only cares
 171 * about the interrupt flag.  Our "save_flags()" just returns that. */
 172static unsigned long save_fl(void)
 173{
 174        return lguest_data.irq_enabled;
 175}
 176
 177/* restore_flags() just sets the flags back to the value given. */
 178static void restore_fl(unsigned long flags)
 179{
 180        lguest_data.irq_enabled = flags;
 181}
 182
 183/* Interrupts go off... */
 184static void irq_disable(void)
 185{
 186        lguest_data.irq_enabled = 0;
 187}
 188
 189/* Interrupts go on... */
 190static void irq_enable(void)
 191{
 192        lguest_data.irq_enabled = X86_EFLAGS_IF;
 193}
 194/*:*/
 195/*M:003 Note that we don't check for outstanding interrupts when we re-enable
 196 * them (or when we unmask an interrupt).  This seems to work for the moment,
 197 * since interrupts are rare and we'll just get the interrupt on the next timer
 198 * tick, but now we can run with CONFIG_NO_HZ, we should revisit this.  One way
 199 * would be to put the "irq_enabled" field in a page by itself, and have the
 200 * Host write-protect it when an interrupt comes in when irqs are disabled.
 201 * There will then be a page fault as soon as interrupts are re-enabled.
 202 *
 203 * A better method is to implement soft interrupt disable generally for x86:
 204 * instead of disabling interrupts, we set a flag.  If an interrupt does come
 205 * in, we then disable them for real.  This is uncommon, so we could simply use
 206 * a hypercall for interrupt control and not worry about efficiency. :*/
 207
 208/*G:034
 209 * The Interrupt Descriptor Table (IDT).
 210 *
 211 * The IDT tells the processor what to do when an interrupt comes in.  Each
 212 * entry in the table is a 64-bit descriptor: this holds the privilege level,
 213 * address of the handler, and... well, who cares?  The Guest just asks the
 214 * Host to make the change anyway, because the Host controls the real IDT.
 215 */
 216static void lguest_write_idt_entry(gate_desc *dt,
 217                                   int entrynum, const gate_desc *g)
 218{
 219        /* The gate_desc structure is 8 bytes long: we hand it to the Host in
 220         * two 32-bit chunks.  The whole 32-bit kernel used to hand descriptors
 221         * around like this; typesafety wasn't a big concern in Linux's early
 222         * years. */
 223        u32 *desc = (u32 *)g;
 224        /* Keep the local copy up to date. */
 225        native_write_idt_entry(dt, entrynum, g);
 226        /* Tell Host about this new entry. */
 227        hcall(LHCALL_LOAD_IDT_ENTRY, entrynum, desc[0], desc[1]);
 228}
 229
 230/* Changing to a different IDT is very rare: we keep the IDT up-to-date every
 231 * time it is written, so we can simply loop through all entries and tell the
 232 * Host about them. */
 233static void lguest_load_idt(const struct desc_ptr *desc)
 234{
 235        unsigned int i;
 236        struct desc_struct *idt = (void *)desc->address;
 237
 238        for (i = 0; i < (desc->size+1)/8; i++)
 239                hcall(LHCALL_LOAD_IDT_ENTRY, i, idt[i].a, idt[i].b);
 240}
 241
 242/*
 243 * The Global Descriptor Table.
 244 *
 245 * The Intel architecture defines another table, called the Global Descriptor
 246 * Table (GDT).  You tell the CPU where it is (and its size) using the "lgdt"
 247 * instruction, and then several other instructions refer to entries in the
 248 * table.  There are three entries which the Switcher needs, so the Host simply
 249 * controls the entire thing and the Guest asks it to make changes using the
 250 * LOAD_GDT hypercall.
 251 *
 252 * This is the opposite of the IDT code where we have a LOAD_IDT_ENTRY
 253 * hypercall and use that repeatedly to load a new IDT.  I don't think it
 254 * really matters, but wouldn't it be nice if they were the same?  Wouldn't
 255 * it be even better if you were the one to send the patch to fix it?
 256 */
 257static void lguest_load_gdt(const struct desc_ptr *desc)
 258{
 259        BUG_ON((desc->size+1)/8 != GDT_ENTRIES);
 260        hcall(LHCALL_LOAD_GDT, __pa(desc->address), GDT_ENTRIES, 0);
 261}
 262
 263/* For a single GDT entry which changes, we do the lazy thing: alter our GDT,
 264 * then tell the Host to reload the entire thing.  This operation is so rare
 265 * that this naive implementation is reasonable. */
 266static void lguest_write_gdt_entry(struct desc_struct *dt, int entrynum,
 267                                   const void *desc, int type)
 268{
 269        native_write_gdt_entry(dt, entrynum, desc, type);
 270        hcall(LHCALL_LOAD_GDT, __pa(dt), GDT_ENTRIES, 0);
 271}
 272
 273/* OK, I lied.  There are three "thread local storage" GDT entries which change
 274 * on every context switch (these three entries are how glibc implements
 275 * __thread variables).  So we have a hypercall specifically for this case. */
 276static void lguest_load_tls(struct thread_struct *t, unsigned int cpu)
 277{
 278        /* There's one problem which normal hardware doesn't have: the Host
 279         * can't handle us removing entries we're currently using.  So we clear
 280         * the GS register here: if it's needed it'll be reloaded anyway. */
 281        loadsegment(gs, 0);
 282        lazy_hcall(LHCALL_LOAD_TLS, __pa(&t->tls_array), cpu, 0);
 283}
 284
 285/*G:038 That's enough excitement for now, back to ploughing through each of
 286 * the different pv_ops structures (we're about 1/3 of the way through).
 287 *
 288 * This is the Local Descriptor Table, another weird Intel thingy.  Linux only
 289 * uses this for some strange applications like Wine.  We don't do anything
 290 * here, so they'll get an informative and friendly Segmentation Fault. */
 291static void lguest_set_ldt(const void *addr, unsigned entries)
 292{
 293}
 294
 295/* This loads a GDT entry into the "Task Register": that entry points to a
 296 * structure called the Task State Segment.  Some comments scattered though the
 297 * kernel code indicate that this used for task switching in ages past, along
 298 * with blood sacrifice and astrology.
 299 *
 300 * Now there's nothing interesting in here that we don't get told elsewhere.
 301 * But the native version uses the "ltr" instruction, which makes the Host
 302 * complain to the Guest about a Segmentation Fault and it'll oops.  So we
 303 * override the native version with a do-nothing version. */
 304static void lguest_load_tr_desc(void)
 305{
 306}
 307
 308/* The "cpuid" instruction is a way of querying both the CPU identity
 309 * (manufacturer, model, etc) and its features.  It was introduced before the
 310 * Pentium in 1993 and keeps getting extended by both Intel, AMD and others.
 311 * As you might imagine, after a decade and a half this treatment, it is now a
 312 * giant ball of hair.  Its entry in the current Intel manual runs to 28 pages.
 313 *
 314 * This instruction even it has its own Wikipedia entry.  The Wikipedia entry
 315 * has been translated into 4 languages.  I am not making this up!
 316 *
 317 * We could get funky here and identify ourselves as "GenuineLguest", but
 318 * instead we just use the real "cpuid" instruction.  Then I pretty much turned
 319 * off feature bits until the Guest booted.  (Don't say that: you'll damage
 320 * lguest sales!)  Shut up, inner voice!  (Hey, just pointing out that this is
 321 * hardly future proof.)  Noone's listening!  They don't like you anyway,
 322 * parenthetic weirdo!
 323 *
 324 * Replacing the cpuid so we can turn features off is great for the kernel, but
 325 * anyone (including userspace) can just use the raw "cpuid" instruction and
 326 * the Host won't even notice since it isn't privileged.  So we try not to get
 327 * too worked up about it. */
 328static void lguest_cpuid(unsigned int *ax, unsigned int *bx,
 329                         unsigned int *cx, unsigned int *dx)
 330{
 331        int function = *ax;
 332
 333        native_cpuid(ax, bx, cx, dx);
 334        switch (function) {
 335        case 1: /* Basic feature request. */
 336                /* We only allow kernel to see SSE3, CMPXCHG16B and SSSE3 */
 337                *cx &= 0x00002201;
 338                /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, TSC, FPU. */
 339                *dx &= 0x07808111;
 340                /* The Host can do a nice optimization if it knows that the
 341                 * kernel mappings (addresses above 0xC0000000 or whatever
 342                 * PAGE_OFFSET is set to) haven't changed.  But Linux calls
 343                 * flush_tlb_user() for both user and kernel mappings unless
 344                 * the Page Global Enable (PGE) feature bit is set. */
 345                *dx |= 0x00002000;
 346                break;
 347        case 0x80000000:
 348                /* Futureproof this a little: if they ask how much extended
 349                 * processor information there is, limit it to known fields. */
 350                if (*ax > 0x80000008)
 351                        *ax = 0x80000008;
 352                break;
 353        }
 354}
 355
 356/* Intel has four control registers, imaginatively named cr0, cr2, cr3 and cr4.
 357 * I assume there's a cr1, but it hasn't bothered us yet, so we'll not bother
 358 * it.  The Host needs to know when the Guest wants to change them, so we have
 359 * a whole series of functions like read_cr0() and write_cr0().
 360 *
 361 * We start with cr0.  cr0 allows you to turn on and off all kinds of basic
 362 * features, but Linux only really cares about one: the horrifically-named Task
 363 * Switched (TS) bit at bit 3 (ie. 8)
 364 *
 365 * What does the TS bit do?  Well, it causes the CPU to trap (interrupt 7) if
 366 * the floating point unit is used.  Which allows us to restore FPU state
 367 * lazily after a task switch, and Linux uses that gratefully, but wouldn't a
 368 * name like "FPUTRAP bit" be a little less cryptic?
 369 *
 370 * We store cr0 locally because the Host never changes it.  The Guest sometimes
 371 * wants to read it and we'd prefer not to bother the Host unnecessarily. */
 372static unsigned long current_cr0;
 373static void lguest_write_cr0(unsigned long val)
 374{
 375        lazy_hcall(LHCALL_TS, val & X86_CR0_TS, 0, 0);
 376        current_cr0 = val;
 377}
 378
 379static unsigned long lguest_read_cr0(void)
 380{
 381        return current_cr0;
 382}
 383
 384/* Intel provided a special instruction to clear the TS bit for people too cool
 385 * to use write_cr0() to do it.  This "clts" instruction is faster, because all
 386 * the vowels have been optimized out. */
 387static void lguest_clts(void)
 388{
 389        lazy_hcall(LHCALL_TS, 0, 0, 0);
 390        current_cr0 &= ~X86_CR0_TS;
 391}
 392
 393/* cr2 is the virtual address of the last page fault, which the Guest only ever
 394 * reads.  The Host kindly writes this into our "struct lguest_data", so we
 395 * just read it out of there. */
 396static unsigned long lguest_read_cr2(void)
 397{
 398        return lguest_data.cr2;
 399}
 400
 401/* See lguest_set_pte() below. */
 402static bool cr3_changed = false;
 403
 404/* cr3 is the current toplevel pagetable page: the principle is the same as
 405 * cr0.  Keep a local copy, and tell the Host when it changes.  The only
 406 * difference is that our local copy is in lguest_data because the Host needs
 407 * to set it upon our initial hypercall. */
 408static void lguest_write_cr3(unsigned long cr3)
 409{
 410        lguest_data.pgdir = cr3;
 411        lazy_hcall(LHCALL_NEW_PGTABLE, cr3, 0, 0);
 412        cr3_changed = true;
 413}
 414
 415static unsigned long lguest_read_cr3(void)
 416{
 417        return lguest_data.pgdir;
 418}
 419
 420/* cr4 is used to enable and disable PGE, but we don't care. */
 421static unsigned long lguest_read_cr4(void)
 422{
 423        return 0;
 424}
 425
 426static void lguest_write_cr4(unsigned long val)
 427{
 428}
 429
 430/*
 431 * Page Table Handling.
 432 *
 433 * Now would be a good time to take a rest and grab a coffee or similarly
 434 * relaxing stimulant.  The easy parts are behind us, and the trek gradually
 435 * winds uphill from here.
 436 *
 437 * Quick refresher: memory is divided into "pages" of 4096 bytes each.  The CPU
 438 * maps virtual addresses to physical addresses using "page tables".  We could
 439 * use one huge index of 1 million entries: each address is 4 bytes, so that's
 440 * 1024 pages just to hold the page tables.   But since most virtual addresses
 441 * are unused, we use a two level index which saves space.  The cr3 register
 442 * contains the physical address of the top level "page directory" page, which
 443 * contains physical addresses of up to 1024 second-level pages.  Each of these
 444 * second level pages contains up to 1024 physical addresses of actual pages,
 445 * or Page Table Entries (PTEs).
 446 *
 447 * Here's a diagram, where arrows indicate physical addresses:
 448 *
 449 * cr3 ---> +---------+
 450 *          |      --------->+---------+
 451 *          |         |      | PADDR1  |
 452 *        Top-level   |      | PADDR2  |
 453 *        (PMD) page  |      |         |
 454 *          |         |    Lower-level |
 455 *          |         |    (PTE) page  |
 456 *          |         |      |         |
 457 *            ....               ....
 458 *
 459 * So to convert a virtual address to a physical address, we look up the top
 460 * level, which points us to the second level, which gives us the physical
 461 * address of that page.  If the top level entry was not present, or the second
 462 * level entry was not present, then the virtual address is invalid (we
 463 * say "the page was not mapped").
 464 *
 465 * Put another way, a 32-bit virtual address is divided up like so:
 466 *
 467 *  1 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 468 * |<---- 10 bits ---->|<---- 10 bits ---->|<------ 12 bits ------>|
 469 *    Index into top     Index into second      Offset within page
 470 *  page directory page    pagetable page
 471 *
 472 * The kernel spends a lot of time changing both the top-level page directory
 473 * and lower-level pagetable pages.  The Guest doesn't know physical addresses,
 474 * so while it maintains these page tables exactly like normal, it also needs
 475 * to keep the Host informed whenever it makes a change: the Host will create
 476 * the real page tables based on the Guests'.
 477 */
 478
 479/* The Guest calls this to set a second-level entry (pte), ie. to map a page
 480 * into a process' address space.  We set the entry then tell the Host the
 481 * toplevel and address this corresponds to.  The Guest uses one pagetable per
 482 * process, so we need to tell the Host which one we're changing (mm->pgd). */
 483static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr,
 484                              pte_t *ptep, pte_t pteval)
 485{
 486        *ptep = pteval;
 487        lazy_hcall(LHCALL_SET_PTE, __pa(mm->pgd), addr, pteval.pte_low);
 488}
 489
 490/* The Guest calls this to set a top-level entry.  Again, we set the entry then
 491 * tell the Host which top-level page we changed, and the index of the entry we
 492 * changed. */
 493static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
 494{
 495        *pmdp = pmdval;
 496        lazy_hcall(LHCALL_SET_PMD, __pa(pmdp)&PAGE_MASK,
 497                   (__pa(pmdp)&(PAGE_SIZE-1))/4, 0);
 498}
 499
 500/* There are a couple of legacy places where the kernel sets a PTE, but we
 501 * don't know the top level any more.  This is useless for us, since we don't
 502 * know which pagetable is changing or what address, so we just tell the Host
 503 * to forget all of them.  Fortunately, this is very rare.
 504 *
 505 * ... except in early boot when the kernel sets up the initial pagetables,
 506 * which makes booting astonishingly slow: 1.83 seconds!  So we don't even tell
 507 * the Host anything changed until we've done the first page table switch,
 508 * which brings boot back to 0.25 seconds. */
 509static void lguest_set_pte(pte_t *ptep, pte_t pteval)
 510{
 511        *ptep = pteval;
 512        if (cr3_changed)
 513                lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0);
 514}
 515
 516/* Unfortunately for Lguest, the pv_mmu_ops for page tables were based on
 517 * native page table operations.  On native hardware you can set a new page
 518 * table entry whenever you want, but if you want to remove one you have to do
 519 * a TLB flush (a TLB is a little cache of page table entries kept by the CPU).
 520 *
 521 * So the lguest_set_pte_at() and lguest_set_pmd() functions above are only
 522 * called when a valid entry is written, not when it's removed (ie. marked not
 523 * present).  Instead, this is where we come when the Guest wants to remove a
 524 * page table entry: we tell the Host to set that entry to 0 (ie. the present
 525 * bit is zero). */
 526static void lguest_flush_tlb_single(unsigned long addr)
 527{
 528        /* Simply set it to zero: if it was not, it will fault back in. */
 529        lazy_hcall(LHCALL_SET_PTE, lguest_data.pgdir, addr, 0);
 530}
 531
 532/* This is what happens after the Guest has removed a large number of entries.
 533 * This tells the Host that any of the page table entries for userspace might
 534 * have changed, ie. virtual addresses below PAGE_OFFSET. */
 535static void lguest_flush_tlb_user(void)
 536{
 537        lazy_hcall(LHCALL_FLUSH_TLB, 0, 0, 0);
 538}
 539
 540/* This is called when the kernel page tables have changed.  That's not very
 541 * common (unless the Guest is using highmem, which makes the Guest extremely
 542 * slow), so it's worth separating this from the user flushing above. */
 543static void lguest_flush_tlb_kernel(void)
 544{
 545        lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0);
 546}
 547
 548/*
 549 * The Unadvanced Programmable Interrupt Controller.
 550 *
 551 * This is an attempt to implement the simplest possible interrupt controller.
 552 * I spent some time looking though routines like set_irq_chip_and_handler,
 553 * set_irq_chip_and_handler_name, set_irq_chip_data and set_phasers_to_stun and
 554 * I *think* this is as simple as it gets.
 555 *
 556 * We can tell the Host what interrupts we want blocked ready for using the
 557 * lguest_data.interrupts bitmap, so disabling (aka "masking") them is as
 558 * simple as setting a bit.  We don't actually "ack" interrupts as such, we
 559 * just mask and unmask them.  I wonder if we should be cleverer?
 560 */
 561static void disable_lguest_irq(unsigned int irq)
 562{
 563        set_bit(irq, lguest_data.blocked_interrupts);
 564}
 565
 566static void enable_lguest_irq(unsigned int irq)
 567{
 568        clear_bit(irq, lguest_data.blocked_interrupts);
 569}
 570
 571/* This structure describes the lguest IRQ controller. */
 572static struct irq_chip lguest_irq_controller = {
 573        .name           = "lguest",
 574        .mask           = disable_lguest_irq,
 575        .mask_ack       = disable_lguest_irq,
 576        .unmask         = enable_lguest_irq,
 577};
 578
 579/* This sets up the Interrupt Descriptor Table (IDT) entry for each hardware
 580 * interrupt (except 128, which is used for system calls), and then tells the
 581 * Linux infrastructure that each interrupt is controlled by our level-based
 582 * lguest interrupt controller. */
 583static void __init lguest_init_IRQ(void)
 584{
 585        unsigned int i;
 586
 587        for (i = 0; i < LGUEST_IRQS; i++) {
 588                int vector = FIRST_EXTERNAL_VECTOR + i;
 589                /* Some systems map "vectors" to interrupts weirdly.  Lguest has
 590                 * a straightforward 1 to 1 mapping, so force that here. */
 591                __get_cpu_var(vector_irq)[vector] = i;
 592                if (vector != SYSCALL_VECTOR) {
 593                        set_intr_gate(vector, interrupt[vector]);
 594                        set_irq_chip_and_handler_name(i, &lguest_irq_controller,
 595                                                      handle_level_irq,
 596                                                      "level");
 597                }
 598        }
 599        /* This call is required to set up for 4k stacks, where we have
 600         * separate stacks for hard and soft interrupts. */
 601        irq_ctx_init(smp_processor_id());
 602}
 603
 604/*
 605 * Time.
 606 *
 607 * It would be far better for everyone if the Guest had its own clock, but
 608 * until then the Host gives us the time on every interrupt.
 609 */
 610static unsigned long lguest_get_wallclock(void)
 611{
 612        return lguest_data.time.tv_sec;
 613}
 614
 615/* The TSC is an Intel thing called the Time Stamp Counter.  The Host tells us
 616 * what speed it runs at, or 0 if it's unusable as a reliable clock source.
 617 * This matches what we want here: if we return 0 from this function, the x86
 618 * TSC clock will give up and not register itself. */
 619static unsigned long lguest_tsc_khz(void)
 620{
 621        return lguest_data.tsc_khz;
 622}
 623
 624/* If we can't use the TSC, the kernel falls back to our lower-priority
 625 * "lguest_clock", where we read the time value given to us by the Host. */
 626static cycle_t lguest_clock_read(void)
 627{
 628        unsigned long sec, nsec;
 629
 630        /* Since the time is in two parts (seconds and nanoseconds), we risk
 631         * reading it just as it's changing from 99 & 0.999999999 to 100 and 0,
 632         * and getting 99 and 0.  As Linux tends to come apart under the stress
 633         * of time travel, we must be careful: */
 634        do {
 635                /* First we read the seconds part. */
 636                sec = lguest_data.time.tv_sec;
 637                /* This read memory barrier tells the compiler and the CPU that
 638                 * this can't be reordered: we have to complete the above
 639                 * before going on. */
 640                rmb();
 641                /* Now we read the nanoseconds part. */
 642                nsec = lguest_data.time.tv_nsec;
 643                /* Make sure we've done that. */
 644                rmb();
 645                /* Now if the seconds part has changed, try again. */
 646        } while (unlikely(lguest_data.time.tv_sec != sec));
 647
 648        /* Our lguest clock is in real nanoseconds. */
 649        return sec*1000000000ULL + nsec;
 650}
 651
 652/* This is the fallback clocksource: lower priority than the TSC clocksource. */
 653static struct clocksource lguest_clock = {
 654        .name           = "lguest",
 655        .rating         = 200,
 656        .read           = lguest_clock_read,
 657        .mask           = CLOCKSOURCE_MASK(64),
 658        .mult           = 1 << 22,
 659        .shift          = 22,
 660        .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
 661};
 662
 663/* We also need a "struct clock_event_device": Linux asks us to set it to go
 664 * off some time in the future.  Actually, James Morris figured all this out, I
 665 * just applied the patch. */
 666static int lguest_clockevent_set_next_event(unsigned long delta,
 667                                           struct clock_event_device *evt)
 668{
 669        /* FIXME: I don't think this can ever happen, but James tells me he had
 670         * to put this code in.  Maybe we should remove it now.  Anyone? */
 671        if (delta < LG_CLOCK_MIN_DELTA) {
 672                if (printk_ratelimit())
 673                        printk(KERN_DEBUG "%s: small delta %lu ns\n",
 674                               __func__, delta);
 675                return -ETIME;
 676        }
 677
 678        /* Please wake us this far in the future. */
 679        hcall(LHCALL_SET_CLOCKEVENT, delta, 0, 0);
 680        return 0;
 681}
 682
 683static void lguest_clockevent_set_mode(enum clock_event_mode mode,
 684                                      struct clock_event_device *evt)
 685{
 686        switch (mode) {
 687        case CLOCK_EVT_MODE_UNUSED:
 688        case CLOCK_EVT_MODE_SHUTDOWN:
 689                /* A 0 argument shuts the clock down. */
 690                hcall(LHCALL_SET_CLOCKEVENT, 0, 0, 0);
 691                break;
 692        case CLOCK_EVT_MODE_ONESHOT:
 693                /* This is what we expect. */
 694                break;
 695        case CLOCK_EVT_MODE_PERIODIC:
 696                BUG();
 697        case CLOCK_EVT_MODE_RESUME:
 698                break;
 699        }
 700}
 701
 702/* This describes our primitive timer chip. */
 703static struct clock_event_device lguest_clockevent = {
 704        .name                   = "lguest",
 705        .features               = CLOCK_EVT_FEAT_ONESHOT,
 706        .set_next_event         = lguest_clockevent_set_next_event,
 707        .set_mode               = lguest_clockevent_set_mode,
 708        .rating                 = INT_MAX,
 709        .mult                   = 1,
 710        .shift                  = 0,
 711        .min_delta_ns           = LG_CLOCK_MIN_DELTA,
 712        .max_delta_ns           = LG_CLOCK_MAX_DELTA,
 713};
 714
 715/* This is the Guest timer interrupt handler (hardware interrupt 0).  We just
 716 * call the clockevent infrastructure and it does whatever needs doing. */
 717static void lguest_time_irq(unsigned int irq, struct irq_desc *desc)
 718{
 719        unsigned long flags;
 720
 721        /* Don't interrupt us while this is running. */
 722        local_irq_save(flags);
 723        lguest_clockevent.event_handler(&lguest_clockevent);
 724        local_irq_restore(flags);
 725}
 726
 727/* At some point in the boot process, we get asked to set up our timing
 728 * infrastructure.  The kernel doesn't expect timer interrupts before this, but
 729 * we cleverly initialized the "blocked_interrupts" field of "struct
 730 * lguest_data" so that timer interrupts were blocked until now. */
 731static void lguest_time_init(void)
 732{
 733        /* Set up the timer interrupt (0) to go to our simple timer routine */
 734        set_irq_handler(0, lguest_time_irq);
 735
 736        clocksource_register(&lguest_clock);
 737
 738        /* We can't set cpumask in the initializer: damn C limitations!  Set it
 739         * here and register our timer device. */
 740        lguest_clockevent.cpumask = cpumask_of_cpu(0);
 741        clockevents_register_device(&lguest_clockevent);
 742
 743        /* Finally, we unblock the timer interrupt. */
 744        enable_lguest_irq(0);
 745}
 746
 747/*
 748 * Miscellaneous bits and pieces.
 749 *
 750 * Here is an oddball collection of functions which the Guest needs for things
 751 * to work.  They're pretty simple.
 752 */
 753
 754/* The Guest needs to tell the Host what stack it expects traps to use.  For
 755 * native hardware, this is part of the Task State Segment mentioned above in
 756 * lguest_load_tr_desc(), but to help hypervisors there's this special call.
 757 *
 758 * We tell the Host the segment we want to use (__KERNEL_DS is the kernel data
 759 * segment), the privilege level (we're privilege level 1, the Host is 0 and
 760 * will not tolerate us trying to use that), the stack pointer, and the number
 761 * of pages in the stack. */
 762static void lguest_load_sp0(struct tss_struct *tss,
 763                            struct thread_struct *thread)
 764{
 765        lazy_hcall(LHCALL_SET_STACK, __KERNEL_DS|0x1, thread->sp0,
 766                   THREAD_SIZE/PAGE_SIZE);
 767}
 768
 769/* Let's just say, I wouldn't do debugging under a Guest. */
 770static void lguest_set_debugreg(int regno, unsigned long value)
 771{
 772        /* FIXME: Implement */
 773}
 774
 775/* There are times when the kernel wants to make sure that no memory writes are
 776 * caught in the cache (that they've all reached real hardware devices).  This
 777 * doesn't matter for the Guest which has virtual hardware.
 778 *
 779 * On the Pentium 4 and above, cpuid() indicates that the Cache Line Flush
 780 * (clflush) instruction is available and the kernel uses that.  Otherwise, it
 781 * uses the older "Write Back and Invalidate Cache" (wbinvd) instruction.
 782 * Unlike clflush, wbinvd can only be run at privilege level 0.  So we can
 783 * ignore clflush, but replace wbinvd.
 784 */
 785static void lguest_wbinvd(void)
 786{
 787}
 788
 789/* If the Guest expects to have an Advanced Programmable Interrupt Controller,
 790 * we play dumb by ignoring writes and returning 0 for reads.  So it's no
 791 * longer Programmable nor Controlling anything, and I don't think 8 lines of
 792 * code qualifies for Advanced.  It will also never interrupt anything.  It
 793 * does, however, allow us to get through the Linux boot code. */
 794#ifdef CONFIG_X86_LOCAL_APIC
 795static void lguest_apic_write(u32 reg, u32 v)
 796{
 797}
 798
 799static u32 lguest_apic_read(u32 reg)
 800{
 801        return 0;
 802}
 803
 804static u64 lguest_apic_icr_read(void)
 805{
 806        return 0;
 807}
 808
 809static void lguest_apic_icr_write(u32 low, u32 id)
 810{
 811        /* Warn to see if there's any stray references */
 812        WARN_ON(1);
 813}
 814
 815static void lguest_apic_wait_icr_idle(void)
 816{
 817        return;
 818}
 819
 820static u32 lguest_apic_safe_wait_icr_idle(void)
 821{
 822        return 0;
 823}
 824
 825static struct apic_ops lguest_basic_apic_ops = {
 826        .read = lguest_apic_read,
 827        .write = lguest_apic_write,
 828        .icr_read = lguest_apic_icr_read,
 829        .icr_write = lguest_apic_icr_write,
 830        .wait_icr_idle = lguest_apic_wait_icr_idle,
 831        .safe_wait_icr_idle = lguest_apic_safe_wait_icr_idle,
 832};
 833#endif
 834
 835/* STOP!  Until an interrupt comes in. */
 836static void lguest_safe_halt(void)
 837{
 838        hcall(LHCALL_HALT, 0, 0, 0);
 839}
 840
 841/* The SHUTDOWN hypercall takes a string to describe what's happening, and
 842 * an argument which says whether this to restart (reboot) the Guest or not.
 843 *
 844 * Note that the Host always prefers that the Guest speak in physical addresses
 845 * rather than virtual addresses, so we use __pa() here. */
 846static void lguest_power_off(void)
 847{
 848        hcall(LHCALL_SHUTDOWN, __pa("Power down"), LGUEST_SHUTDOWN_POWEROFF, 0);
 849}
 850
 851/*
 852 * Panicing.
 853 *
 854 * Don't.  But if you did, this is what happens.
 855 */
 856static int lguest_panic(struct notifier_block *nb, unsigned long l, void *p)
 857{
 858        hcall(LHCALL_SHUTDOWN, __pa(p), LGUEST_SHUTDOWN_POWEROFF, 0);
 859        /* The hcall won't return, but to keep gcc happy, we're "done". */
 860        return NOTIFY_DONE;
 861}
 862
 863static struct notifier_block paniced = {
 864        .notifier_call = lguest_panic
 865};
 866
 867/* Setting up memory is fairly easy. */
 868static __init char *lguest_memory_setup(void)
 869{
 870        /* We do this here and not earlier because lockcheck used to barf if we
 871         * did it before start_kernel().  I think we fixed that, so it'd be
 872         * nice to move it back to lguest_init.  Patch welcome... */
 873        atomic_notifier_chain_register(&panic_notifier_list, &paniced);
 874
 875        /* The Linux bootloader header contains an "e820" memory map: the
 876         * Launcher populated the first entry with our memory limit. */
 877        e820_add_region(boot_params.e820_map[0].addr,
 878                          boot_params.e820_map[0].size,
 879                          boot_params.e820_map[0].type);
 880
 881        /* This string is for the boot messages. */
 882        return "LGUEST";
 883}
 884
 885/* We will eventually use the virtio console device to produce console output,
 886 * but before that is set up we use LHCALL_NOTIFY on normal memory to produce
 887 * console output. */
 888static __init int early_put_chars(u32 vtermno, const char *buf, int count)
 889{
 890        char scratch[17];
 891        unsigned int len = count;
 892
 893        /* We use a nul-terminated string, so we have to make a copy.  Icky,
 894         * huh? */
 895        if (len > sizeof(scratch) - 1)
 896                len = sizeof(scratch) - 1;
 897        scratch[len] = '\0';
 898        memcpy(scratch, buf, len);
 899        hcall(LHCALL_NOTIFY, __pa(scratch), 0, 0);
 900
 901        /* This routine returns the number of bytes actually written. */
 902        return len;
 903}
 904
 905/* Rebooting also tells the Host we're finished, but the RESTART flag tells the
 906 * Launcher to reboot us. */
 907static void lguest_restart(char *reason)
 908{
 909        hcall(LHCALL_SHUTDOWN, __pa(reason), LGUEST_SHUTDOWN_RESTART, 0);
 910}
 911
 912/*G:050
 913 * Patching (Powerfully Placating Performance Pedants)
 914 *
 915 * We have already seen that pv_ops structures let us replace simple native
 916 * instructions with calls to the appropriate back end all throughout the
 917 * kernel.  This allows the same kernel to run as a Guest and as a native
 918 * kernel, but it's slow because of all the indirect branches.
 919 *
 920 * Remember that David Wheeler quote about "Any problem in computer science can
 921 * be solved with another layer of indirection"?  The rest of that quote is
 922 * "... But that usually will create another problem."  This is the first of
 923 * those problems.
 924 *
 925 * Our current solution is to allow the paravirt back end to optionally patch
 926 * over the indirect calls to replace them with something more efficient.  We
 927 * patch the four most commonly called functions: disable interrupts, enable
 928 * interrupts, restore interrupts and save interrupts.  We usually have 6 or 10
 929 * bytes to patch into: the Guest versions of these operations are small enough
 930 * that we can fit comfortably.
 931 *
 932 * First we need assembly templates of each of the patchable Guest operations,
 933 * and these are in lguest_asm.S. */
 934
 935/*G:060 We construct a table from the assembler templates: */
 936static const struct lguest_insns
 937{
 938        const char *start, *end;
 939} lguest_insns[] = {
 940        [PARAVIRT_PATCH(pv_irq_ops.irq_disable)] = { lgstart_cli, lgend_cli },
 941        [PARAVIRT_PATCH(pv_irq_ops.irq_enable)] = { lgstart_sti, lgend_sti },
 942        [PARAVIRT_PATCH(pv_irq_ops.restore_fl)] = { lgstart_popf, lgend_popf },
 943        [PARAVIRT_PATCH(pv_irq_ops.save_fl)] = { lgstart_pushf, lgend_pushf },
 944};
 945
 946/* Now our patch routine is fairly simple (based on the native one in
 947 * paravirt.c).  If we have a replacement, we copy it in and return how much of
 948 * the available space we used. */
 949static unsigned lguest_patch(u8 type, u16 clobber, void *ibuf,
 950                             unsigned long addr, unsigned len)
 951{
 952        unsigned int insn_len;
 953
 954        /* Don't do anything special if we don't have a replacement */
 955        if (type >= ARRAY_SIZE(lguest_insns) || !lguest_insns[type].start)
 956                return paravirt_patch_default(type, clobber, ibuf, addr, len);
 957
 958        insn_len = lguest_insns[type].end - lguest_insns[type].start;
 959
 960        /* Similarly if we can't fit replacement (shouldn't happen, but let's
 961         * be thorough). */
 962        if (len < insn_len)
 963                return paravirt_patch_default(type, clobber, ibuf, addr, len);
 964
 965        /* Copy in our instructions. */
 966        memcpy(ibuf, lguest_insns[type].start, insn_len);
 967        return insn_len;
 968}
 969
 970/*G:030 Once we get to lguest_init(), we know we're a Guest.  The various
 971 * pv_ops structures in the kernel provide points for (almost) every routine we
 972 * have to override to avoid privileged instructions. */
 973__init void lguest_init(void)
 974{
 975        /* We're under lguest, paravirt is enabled, and we're running at
 976         * privilege level 1, not 0 as normal. */
 977        pv_info.name = "lguest";
 978        pv_info.paravirt_enabled = 1;
 979        pv_info.kernel_rpl = 1;
 980
 981        /* We set up all the lguest overrides for sensitive operations.  These
 982         * are detailed with the operations themselves. */
 983
 984        /* interrupt-related operations */
 985        pv_irq_ops.init_IRQ = lguest_init_IRQ;
 986        pv_irq_ops.save_fl = save_fl;
 987        pv_irq_ops.restore_fl = restore_fl;
 988        pv_irq_ops.irq_disable = irq_disable;
 989        pv_irq_ops.irq_enable = irq_enable;
 990        pv_irq_ops.safe_halt = lguest_safe_halt;
 991
 992        /* init-time operations */
 993        pv_init_ops.memory_setup = lguest_memory_setup;
 994        pv_init_ops.patch = lguest_patch;
 995
 996        /* Intercepts of various cpu instructions */
 997        pv_cpu_ops.load_gdt = lguest_load_gdt;
 998        pv_cpu_ops.cpuid = lguest_cpuid;
 999        pv_cpu_ops.load_idt = lguest_load_idt;
1000        pv_cpu_ops.iret = lguest_iret;
1001        pv_cpu_ops.load_sp0 = lguest_load_sp0;
1002        pv_cpu_ops.load_tr_desc = lguest_load_tr_desc;
1003        pv_cpu_ops.set_ldt = lguest_set_ldt;
1004        pv_cpu_ops.load_tls = lguest_load_tls;
1005        pv_cpu_ops.set_debugreg = lguest_set_debugreg;
1006        pv_cpu_ops.clts = lguest_clts;
1007        pv_cpu_ops.read_cr0 = lguest_read_cr0;
1008        pv_cpu_ops.write_cr0 = lguest_write_cr0;
1009        pv_cpu_ops.read_cr4 = lguest_read_cr4;
1010        pv_cpu_ops.write_cr4 = lguest_write_cr4;
1011        pv_cpu_ops.write_gdt_entry = lguest_write_gdt_entry;
1012        pv_cpu_ops.write_idt_entry = lguest_write_idt_entry;
1013        pv_cpu_ops.wbinvd = lguest_wbinvd;
1014        pv_cpu_ops.lazy_mode.enter = paravirt_enter_lazy_cpu;
1015        pv_cpu_ops.lazy_mode.leave = lguest_leave_lazy_mode;
1016
1017        /* pagetable management */
1018        pv_mmu_ops.write_cr3 = lguest_write_cr3;
1019        pv_mmu_ops.flush_tlb_user = lguest_flush_tlb_user;
1020        pv_mmu_ops.flush_tlb_single = lguest_flush_tlb_single;
1021        pv_mmu_ops.flush_tlb_kernel = lguest_flush_tlb_kernel;
1022        pv_mmu_ops.set_pte = lguest_set_pte;
1023        pv_mmu_ops.set_pte_at = lguest_set_pte_at;
1024        pv_mmu_ops.set_pmd = lguest_set_pmd;
1025        pv_mmu_ops.read_cr2 = lguest_read_cr2;
1026        pv_mmu_ops.read_cr3 = lguest_read_cr3;
1027        pv_mmu_ops.lazy_mode.enter = paravirt_enter_lazy_mmu;
1028        pv_mmu_ops.lazy_mode.leave = lguest_leave_lazy_mode;
1029
1030#ifdef CONFIG_X86_LOCAL_APIC
1031        /* apic read/write intercepts */
1032        apic_ops = &lguest_basic_apic_ops;
1033#endif
1034
1035        /* time operations */
1036        pv_time_ops.get_wallclock = lguest_get_wallclock;
1037        pv_time_ops.time_init = lguest_time_init;
1038        pv_time_ops.get_tsc_khz = lguest_tsc_khz;
1039
1040        /* Now is a good time to look at the implementations of these functions
1041         * before returning to the rest of lguest_init(). */
1042
1043        /*G:070 Now we've seen all the paravirt_ops, we return to
1044         * lguest_init() where the rest of the fairly chaotic boot setup
1045         * occurs. */
1046
1047        /* The native boot code sets up initial page tables immediately after
1048         * the kernel itself, and sets init_pg_tables_end so they're not
1049         * clobbered.  The Launcher places our initial pagetables somewhere at
1050         * the top of our physical memory, so we don't need extra space: set
1051         * init_pg_tables_end to the end of the kernel. */
1052        init_pg_tables_start = __pa(pg0);
1053        init_pg_tables_end = __pa(pg0);
1054
1055        /* As described in head_32.S, we map the first 128M of memory. */
1056        max_pfn_mapped = (128*1024*1024) >> PAGE_SHIFT;
1057
1058        /* Load the %fs segment register (the per-cpu segment register) with
1059         * the normal data segment to get through booting. */
1060        asm volatile ("mov %0, %%fs" : : "r" (__KERNEL_DS) : "memory");
1061
1062        /* The Host<->Guest Switcher lives at the top of our address space, and
1063         * the Host told us how big it is when we made LGUEST_INIT hypercall:
1064         * it put the answer in lguest_data.reserve_mem  */
1065        reserve_top_address(lguest_data.reserve_mem);
1066
1067        /* If we don't initialize the lock dependency checker now, it crashes
1068         * paravirt_disable_iospace. */
1069        lockdep_init();
1070
1071        /* The IDE code spends about 3 seconds probing for disks: if we reserve
1072         * all the I/O ports up front it can't get them and so doesn't probe.
1073         * Other device drivers are similar (but less severe).  This cuts the
1074         * kernel boot time on my machine from 4.1 seconds to 0.45 seconds. */
1075        paravirt_disable_iospace();
1076
1077        /* This is messy CPU setup stuff which the native boot code does before
1078         * start_kernel, so we have to do, too: */
1079        cpu_detect(&new_cpu_data);
1080        /* head.S usually sets up the first capability word, so do it here. */
1081        new_cpu_data.x86_capability[0] = cpuid_edx(1);
1082
1083        /* Math is always hard! */
1084        new_cpu_data.hard_math = 1;
1085
1086        /* We don't have features.  We have puppies!  Puppies! */
1087#ifdef CONFIG_X86_MCE
1088        mce_disabled = 1;
1089#endif
1090#ifdef CONFIG_ACPI
1091        acpi_disabled = 1;
1092        acpi_ht = 0;
1093#endif
1094
1095        /* We set the perferred console to "hvc".  This is the "hypervisor
1096         * virtual console" driver written by the PowerPC people, which we also
1097         * adapted for lguest's use. */
1098        add_preferred_console("hvc", 0, NULL);
1099
1100        /* Register our very early console. */
1101        virtio_cons_early_init(early_put_chars);
1102
1103        /* Last of all, we set the power management poweroff hook to point to
1104         * the Guest routine to power off, and the reboot hook to our restart
1105         * routine. */
1106        pm_power_off = lguest_power_off;
1107        machine_ops.restart = lguest_restart;
1108
1109        /* Now we're set up, call i386_start_kernel() in head32.c and we proceed
1110         * to boot as normal.  It never returns. */
1111        i386_start_kernel();
1112}
1113/*
1114 * This marks the end of stage II of our journey, The Guest.
1115 *
1116 * It is now time for us to explore the layer of virtual drivers and complete
1117 * our understanding of the Guest in "make Drivers".
1118 */
1119