i386: prepare shared kernel/vmi.c
authorThomas Gleixner <tglx@linutronix.de>
Thu, 11 Oct 2007 09:11:57 +0000 (11:11 +0200)
committerThomas Gleixner <tglx@linutronix.de>
Thu, 11 Oct 2007 09:11:57 +0000 (11:11 +0200)
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/i386/kernel/Makefile
arch/i386/kernel/vmi.c [deleted file]
arch/i386/kernel/vmi_32.c [new file with mode: 0644]

index 6f3a7f2118118c0cbc9f26f796580f46ac7551b1..559f1f77185be95cf7062fa1d32521c70ccac3d4 100644 (file)
@@ -41,7 +41,7 @@ obj-$(CONFIG_HPET_TIMER)      += hpet.o
 obj-$(CONFIG_K8_NB)            += k8.o
 obj-$(CONFIG_MGEODE_LX)                += geode.o
 
-obj-$(CONFIG_VMI)              += vmi.o vmiclock_32.o
+obj-$(CONFIG_VMI)              += vmi_32.o vmiclock_32.o
 obj-$(CONFIG_PARAVIRT)         += paravirt.o
 obj-y                          += pcspeaker.o
 
diff --git a/arch/i386/kernel/vmi.c b/arch/i386/kernel/vmi.c
deleted file mode 100644 (file)
index 18673e0..0000000
+++ /dev/null
@@ -1,981 +0,0 @@
-/*
- * VMI specific paravirt-ops implementation
- *
- * Copyright (C) 2005, VMware, Inc.
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License as published by
- * the Free Software Foundation; either version 2 of the License, or
- * (at your option) any later version.
- *
- * This program is distributed in the hope that it will be useful, but
- * WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
- * NON INFRINGEMENT.  See the GNU General Public License for more
- * details.
- *
- * You should have received a copy of the GNU General Public License
- * along with this program; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
- *
- * Send feedback to zach@vmware.com
- *
- */
-
-#include <linux/module.h>
-#include <linux/cpu.h>
-#include <linux/bootmem.h>
-#include <linux/mm.h>
-#include <linux/highmem.h>
-#include <linux/sched.h>
-#include <asm/vmi.h>
-#include <asm/io.h>
-#include <asm/fixmap.h>
-#include <asm/apicdef.h>
-#include <asm/apic.h>
-#include <asm/processor.h>
-#include <asm/timer.h>
-#include <asm/vmi_time.h>
-#include <asm/kmap_types.h>
-
-/* Convenient for calling VMI functions indirectly in the ROM */
-typedef u32 __attribute__((regparm(1))) (VROMFUNC)(void);
-typedef u64 __attribute__((regparm(2))) (VROMLONGFUNC)(int);
-
-#define call_vrom_func(rom,func) \
-   (((VROMFUNC *)(rom->func))())
-
-#define call_vrom_long_func(rom,func,arg) \
-   (((VROMLONGFUNC *)(rom->func)) (arg))
-
-static struct vrom_header *vmi_rom;
-static int disable_pge;
-static int disable_pse;
-static int disable_sep;
-static int disable_tsc;
-static int disable_mtrr;
-static int disable_noidle;
-static int disable_vmi_timer;
-
-/* Cached VMI operations */
-static struct {
-       void (*cpuid)(void /* non-c */);
-       void (*_set_ldt)(u32 selector);
-       void (*set_tr)(u32 selector);
-       void (*set_kernel_stack)(u32 selector, u32 esp0);
-       void (*allocate_page)(u32, u32, u32, u32, u32);
-       void (*release_page)(u32, u32);
-       void (*set_pte)(pte_t, pte_t *, unsigned);
-       void (*update_pte)(pte_t *, unsigned);
-       void (*set_linear_mapping)(int, void *, u32, u32);
-       void (*_flush_tlb)(int);
-       void (*set_initial_ap_state)(int, int);
-       void (*halt)(void);
-       void (*set_lazy_mode)(int mode);
-} vmi_ops;
-
-/* Cached VMI operations */
-struct vmi_timer_ops vmi_timer_ops;
-
-/*
- * VMI patching routines.
- */
-#define MNEM_CALL 0xe8
-#define MNEM_JMP  0xe9
-#define MNEM_RET  0xc3
-
-#define IRQ_PATCH_INT_MASK 0
-#define IRQ_PATCH_DISABLE  5
-
-static inline void patch_offset(void *insnbuf,
-                               unsigned long eip, unsigned long dest)
-{
-        *(unsigned long *)(insnbuf+1) = dest-eip-5;
-}
-
-static unsigned patch_internal(int call, unsigned len, void *insnbuf,
-                              unsigned long eip)
-{
-       u64 reloc;
-       struct vmi_relocation_info *const rel = (struct vmi_relocation_info *)&reloc;
-       reloc = call_vrom_long_func(vmi_rom, get_reloc, call);
-       switch(rel->type) {
-               case VMI_RELOCATION_CALL_REL:
-                       BUG_ON(len < 5);
-                       *(char *)insnbuf = MNEM_CALL;
-                       patch_offset(insnbuf, eip, (unsigned long)rel->eip);
-                       return 5;
-
-               case VMI_RELOCATION_JUMP_REL:
-                       BUG_ON(len < 5);
-                       *(char *)insnbuf = MNEM_JMP;
-                       patch_offset(insnbuf, eip, (unsigned long)rel->eip);
-                       return 5;
-
-               case VMI_RELOCATION_NOP:
-                       /* obliterate the whole thing */
-                       return 0;
-
-               case VMI_RELOCATION_NONE:
-                       /* leave native code in place */
-                       break;
-
-               default:
-                       BUG();
-       }
-       return len;
-}
-
-/*
- * Apply patch if appropriate, return length of new instruction
- * sequence.  The callee does nop padding for us.
- */
-static unsigned vmi_patch(u8 type, u16 clobbers, void *insns,
-                         unsigned long eip, unsigned len)
-{
-       switch (type) {
-               case PARAVIRT_PATCH(irq_disable):
-                       return patch_internal(VMI_CALL_DisableInterrupts, len,
-                                             insns, eip);
-               case PARAVIRT_PATCH(irq_enable):
-                       return patch_internal(VMI_CALL_EnableInterrupts, len,
-                                             insns, eip);
-               case PARAVIRT_PATCH(restore_fl):
-                       return patch_internal(VMI_CALL_SetInterruptMask, len,
-                                             insns, eip);
-               case PARAVIRT_PATCH(save_fl):
-                       return patch_internal(VMI_CALL_GetInterruptMask, len,
-                                             insns, eip);
-               case PARAVIRT_PATCH(iret):
-                       return patch_internal(VMI_CALL_IRET, len, insns, eip);
-               case PARAVIRT_PATCH(irq_enable_sysexit):
-                       return patch_internal(VMI_CALL_SYSEXIT, len, insns, eip);
-               default:
-                       break;
-       }
-       return len;
-}
-
-/* CPUID has non-C semantics, and paravirt-ops API doesn't match hardware ISA */
-static void vmi_cpuid(unsigned int *eax, unsigned int *ebx,
-                               unsigned int *ecx, unsigned int *edx)
-{
-       int override = 0;
-       if (*eax == 1)
-               override = 1;
-        asm volatile ("call *%6"
-                      : "=a" (*eax),
-                        "=b" (*ebx),
-                        "=c" (*ecx),
-                        "=d" (*edx)
-                      : "0" (*eax), "2" (*ecx), "r" (vmi_ops.cpuid));
-       if (override) {
-               if (disable_pse)
-                       *edx &= ~X86_FEATURE_PSE;
-               if (disable_pge)
-                       *edx &= ~X86_FEATURE_PGE;
-               if (disable_sep)
-                       *edx &= ~X86_FEATURE_SEP;
-               if (disable_tsc)
-                       *edx &= ~X86_FEATURE_TSC;
-               if (disable_mtrr)
-                       *edx &= ~X86_FEATURE_MTRR;
-       }
-}
-
-static inline void vmi_maybe_load_tls(struct desc_struct *gdt, int nr, struct desc_struct *new)
-{
-       if (gdt[nr].a != new->a || gdt[nr].b != new->b)
-               write_gdt_entry(gdt, nr, new->a, new->b);
-}
-
-static void vmi_load_tls(struct thread_struct *t, unsigned int cpu)
-{
-       struct desc_struct *gdt = get_cpu_gdt_table(cpu);
-       vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 0, &t->tls_array[0]);
-       vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 1, &t->tls_array[1]);
-       vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 2, &t->tls_array[2]);
-}
-
-static void vmi_set_ldt(const void *addr, unsigned entries)
-{
-       unsigned cpu = smp_processor_id();
-       u32 low, high;
-
-       pack_descriptor(&low, &high, (unsigned long)addr,
-                       entries * sizeof(struct desc_struct) - 1,
-                       DESCTYPE_LDT, 0);
-       write_gdt_entry(get_cpu_gdt_table(cpu), GDT_ENTRY_LDT, low, high);
-       vmi_ops._set_ldt(entries ? GDT_ENTRY_LDT*sizeof(struct desc_struct) : 0);
-}
-
-static void vmi_set_tr(void)
-{
-       vmi_ops.set_tr(GDT_ENTRY_TSS*sizeof(struct desc_struct));
-}
-
-static void vmi_load_esp0(struct tss_struct *tss,
-                                  struct thread_struct *thread)
-{
-       tss->x86_tss.esp0 = thread->esp0;
-
-       /* This can only happen when SEP is enabled, no need to test "SEP"arately */
-       if (unlikely(tss->x86_tss.ss1 != thread->sysenter_cs)) {
-               tss->x86_tss.ss1 = thread->sysenter_cs;
-               wrmsr(MSR_IA32_SYSENTER_CS, thread->sysenter_cs, 0);
-       }
-       vmi_ops.set_kernel_stack(__KERNEL_DS, tss->x86_tss.esp0);
-}
-
-static void vmi_flush_tlb_user(void)
-{
-       vmi_ops._flush_tlb(VMI_FLUSH_TLB);
-}
-
-static void vmi_flush_tlb_kernel(void)
-{
-       vmi_ops._flush_tlb(VMI_FLUSH_TLB | VMI_FLUSH_GLOBAL);
-}
-
-/* Stub to do nothing at all; used for delays and unimplemented calls */
-static void vmi_nop(void)
-{
-}
-
-#ifdef CONFIG_DEBUG_PAGE_TYPE
-
-#ifdef CONFIG_X86_PAE
-#define MAX_BOOT_PTS (2048+4+1)
-#else
-#define MAX_BOOT_PTS (1024+1)
-#endif
-
-/*
- * During boot, mem_map is not yet available in paging_init, so stash
- * all the boot page allocations here.
- */
-static struct {
-       u32 pfn;
-       int type;
-} boot_page_allocations[MAX_BOOT_PTS];
-static int num_boot_page_allocations;
-static int boot_allocations_applied;
-
-void vmi_apply_boot_page_allocations(void)
-{
-       int i;
-       BUG_ON(!mem_map);
-       for (i = 0; i < num_boot_page_allocations; i++) {
-               struct page *page = pfn_to_page(boot_page_allocations[i].pfn);
-               page->type = boot_page_allocations[i].type;
-               page->type = boot_page_allocations[i].type &
-                               ~(VMI_PAGE_ZEROED | VMI_PAGE_CLONE);
-       }
-       boot_allocations_applied = 1;
-}
-
-static void record_page_type(u32 pfn, int type)
-{
-       BUG_ON(num_boot_page_allocations >= MAX_BOOT_PTS);
-       boot_page_allocations[num_boot_page_allocations].pfn = pfn;
-       boot_page_allocations[num_boot_page_allocations].type = type;
-       num_boot_page_allocations++;
-}
-
-static void check_zeroed_page(u32 pfn, int type, struct page *page)
-{
-       u32 *ptr;
-       int i;
-       int limit = PAGE_SIZE / sizeof(int);
-
-       if (page_address(page))
-               ptr = (u32 *)page_address(page);
-       else
-               ptr = (u32 *)__va(pfn << PAGE_SHIFT);
-       /*
-        * When cloning the root in non-PAE mode, only the userspace
-        * pdes need to be zeroed.
-        */
-       if (type & VMI_PAGE_CLONE)
-               limit = USER_PTRS_PER_PGD;
-       for (i = 0; i < limit; i++)
-               BUG_ON(ptr[i]);
-}
-
-/*
- * We stash the page type into struct page so we can verify the page
- * types are used properly.
- */
-static void vmi_set_page_type(u32 pfn, int type)
-{
-       /* PAE can have multiple roots per page - don't track */
-       if (PTRS_PER_PMD > 1 && (type & VMI_PAGE_PDP))
-               return;
-
-       if (boot_allocations_applied) {
-               struct page *page = pfn_to_page(pfn);
-               if (type != VMI_PAGE_NORMAL)
-                       BUG_ON(page->type);
-               else
-                       BUG_ON(page->type == VMI_PAGE_NORMAL);
-               page->type = type & ~(VMI_PAGE_ZEROED | VMI_PAGE_CLONE);
-               if (type & VMI_PAGE_ZEROED)
-                       check_zeroed_page(pfn, type, page);
-       } else {
-               record_page_type(pfn, type);
-       }
-}
-
-static void vmi_check_page_type(u32 pfn, int type)
-{
-       /* PAE can have multiple roots per page - skip checks */
-       if (PTRS_PER_PMD > 1 && (type & VMI_PAGE_PDP))
-               return;
-
-       type &= ~(VMI_PAGE_ZEROED | VMI_PAGE_CLONE);
-       if (boot_allocations_applied) {
-               struct page *page = pfn_to_page(pfn);
-               BUG_ON((page->type ^ type) & VMI_PAGE_PAE);
-               BUG_ON(type == VMI_PAGE_NORMAL && page->type);
-               BUG_ON((type & page->type) == 0);
-       }
-}
-#else
-#define vmi_set_page_type(p,t) do { } while (0)
-#define vmi_check_page_type(p,t) do { } while (0)
-#endif
-
-#ifdef CONFIG_HIGHPTE
-static void *vmi_kmap_atomic_pte(struct page *page, enum km_type type)
-{
-       void *va = kmap_atomic(page, type);
-
-       /*
-        * Internally, the VMI ROM must map virtual addresses to physical
-        * addresses for processing MMU updates.  By the time MMU updates
-        * are issued, this information is typically already lost.
-        * Fortunately, the VMI provides a cache of mapping slots for active
-        * page tables.
-        *
-        * We use slot zero for the linear mapping of physical memory, and
-        * in HIGHPTE kernels, slot 1 and 2 for KM_PTE0 and KM_PTE1.
-        *
-        *  args:                 SLOT                 VA    COUNT PFN
-        */
-       BUG_ON(type != KM_PTE0 && type != KM_PTE1);
-       vmi_ops.set_linear_mapping((type - KM_PTE0)+1, va, 1, page_to_pfn(page));
-
-       return va;
-}
-#endif
-
-static void vmi_allocate_pt(struct mm_struct *mm, u32 pfn)
-{
-       vmi_set_page_type(pfn, VMI_PAGE_L1);
-       vmi_ops.allocate_page(pfn, VMI_PAGE_L1, 0, 0, 0);
-}
-
-static void vmi_allocate_pd(u32 pfn)
-{
-       /*
-        * This call comes in very early, before mem_map is setup.
-        * It is called only for swapper_pg_dir, which already has
-        * data on it.
-        */
-       vmi_set_page_type(pfn, VMI_PAGE_L2);
-       vmi_ops.allocate_page(pfn, VMI_PAGE_L2, 0, 0, 0);
-}
-
-static void vmi_allocate_pd_clone(u32 pfn, u32 clonepfn, u32 start, u32 count)
-{
-       vmi_set_page_type(pfn, VMI_PAGE_L2 | VMI_PAGE_CLONE);
-       vmi_check_page_type(clonepfn, VMI_PAGE_L2);
-       vmi_ops.allocate_page(pfn, VMI_PAGE_L2 | VMI_PAGE_CLONE, clonepfn, start, count);
-}
-
-static void vmi_release_pt(u32 pfn)
-{
-       vmi_ops.release_page(pfn, VMI_PAGE_L1);
-       vmi_set_page_type(pfn, VMI_PAGE_NORMAL);
-}
-
-static void vmi_release_pd(u32 pfn)
-{
-       vmi_ops.release_page(pfn, VMI_PAGE_L2);
-       vmi_set_page_type(pfn, VMI_PAGE_NORMAL);
-}
-
-/*
- * Helper macros for MMU update flags.  We can defer updates until a flush
- * or page invalidation only if the update is to the current address space
- * (otherwise, there is no flush).  We must check against init_mm, since
- * this could be a kernel update, which usually passes init_mm, although
- * sometimes this check can be skipped if we know the particular function
- * is only called on user mode PTEs.  We could change the kernel to pass
- * current->active_mm here, but in particular, I was unsure if changing
- * mm/highmem.c to do this would still be correct on other architectures.
- */
-#define is_current_as(mm, mustbeuser) ((mm) == current->active_mm ||    \
-                                       (!mustbeuser && (mm) == &init_mm))
-#define vmi_flags_addr(mm, addr, level, user)                           \
-        ((level) | (is_current_as(mm, user) ?                           \
-                (VMI_PAGE_CURRENT_AS | ((addr) & VMI_PAGE_VA_MASK)) : 0))
-#define vmi_flags_addr_defer(mm, addr, level, user)                     \
-        ((level) | (is_current_as(mm, user) ?                           \
-                (VMI_PAGE_DEFER | VMI_PAGE_CURRENT_AS | ((addr) & VMI_PAGE_VA_MASK)) : 0))
-
-static void vmi_update_pte(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
-{
-       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
-       vmi_ops.update_pte(ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
-}
-
-static void vmi_update_pte_defer(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
-{
-       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
-       vmi_ops.update_pte(ptep, vmi_flags_addr_defer(mm, addr, VMI_PAGE_PT, 0));
-}
-
-static void vmi_set_pte(pte_t *ptep, pte_t pte)
-{
-       /* XXX because of set_pmd_pte, this can be called on PT or PD layers */
-       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE | VMI_PAGE_PD);
-       vmi_ops.set_pte(pte, ptep, VMI_PAGE_PT);
-}
-
-static void vmi_set_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
-{
-       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
-       vmi_ops.set_pte(pte, ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
-}
-
-static void vmi_set_pmd(pmd_t *pmdp, pmd_t pmdval)
-{
-#ifdef CONFIG_X86_PAE
-       const pte_t pte = { pmdval.pmd, pmdval.pmd >> 32 };
-       vmi_check_page_type(__pa(pmdp) >> PAGE_SHIFT, VMI_PAGE_PMD);
-#else
-       const pte_t pte = { pmdval.pud.pgd.pgd };
-       vmi_check_page_type(__pa(pmdp) >> PAGE_SHIFT, VMI_PAGE_PGD);
-#endif
-       vmi_ops.set_pte(pte, (pte_t *)pmdp, VMI_PAGE_PD);
-}
-
-#ifdef CONFIG_X86_PAE
-
-static void vmi_set_pte_atomic(pte_t *ptep, pte_t pteval)
-{
-       /*
-        * XXX This is called from set_pmd_pte, but at both PT
-        * and PD layers so the VMI_PAGE_PT flag is wrong.  But
-        * it is only called for large page mapping changes,
-        * the Xen backend, doesn't support large pages, and the
-        * ESX backend doesn't depend on the flag.
-        */
-       set_64bit((unsigned long long *)ptep,pte_val(pteval));
-       vmi_ops.update_pte(ptep, VMI_PAGE_PT);
-}
-
-static void vmi_set_pte_present(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
-{
-       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
-       vmi_ops.set_pte(pte, ptep, vmi_flags_addr_defer(mm, addr, VMI_PAGE_PT, 1));
-}
-
-static void vmi_set_pud(pud_t *pudp, pud_t pudval)
-{
-       /* Um, eww */
-       const pte_t pte = { pudval.pgd.pgd, pudval.pgd.pgd >> 32 };
-       vmi_check_page_type(__pa(pudp) >> PAGE_SHIFT, VMI_PAGE_PGD);
-       vmi_ops.set_pte(pte, (pte_t *)pudp, VMI_PAGE_PDP);
-}
-
-static void vmi_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
-{
-       const pte_t pte = { 0 };
-       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
-       vmi_ops.set_pte(pte, ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
-}
-
-static void vmi_pmd_clear(pmd_t *pmd)
-{
-       const pte_t pte = { 0 };
-       vmi_check_page_type(__pa(pmd) >> PAGE_SHIFT, VMI_PAGE_PMD);
-       vmi_ops.set_pte(pte, (pte_t *)pmd, VMI_PAGE_PD);
-}
-#endif
-
-#ifdef CONFIG_SMP
-static void __devinit
-vmi_startup_ipi_hook(int phys_apicid, unsigned long start_eip,
-                    unsigned long start_esp)
-{
-       struct vmi_ap_state ap;
-
-       /* Default everything to zero.  This is fine for most GPRs. */
-       memset(&ap, 0, sizeof(struct vmi_ap_state));
-
-       ap.gdtr_limit = GDT_SIZE - 1;
-       ap.gdtr_base = (unsigned long) get_cpu_gdt_table(phys_apicid);
-
-       ap.idtr_limit = IDT_ENTRIES * 8 - 1;
-       ap.idtr_base = (unsigned long) idt_table;
-
-       ap.ldtr = 0;
-
-       ap.cs = __KERNEL_CS;
-       ap.eip = (unsigned long) start_eip;
-       ap.ss = __KERNEL_DS;
-       ap.esp = (unsigned long) start_esp;
-
-       ap.ds = __USER_DS;
-       ap.es = __USER_DS;
-       ap.fs = __KERNEL_PERCPU;
-       ap.gs = 0;
-
-       ap.eflags = 0;
-
-#ifdef CONFIG_X86_PAE
-       /* efer should match BSP efer. */
-       if (cpu_has_nx) {
-               unsigned l, h;
-               rdmsr(MSR_EFER, l, h);
-               ap.efer = (unsigned long long) h << 32 | l;
-       }
-#endif
-
-       ap.cr3 = __pa(swapper_pg_dir);
-       /* Protected mode, paging, AM, WP, NE, MP. */
-       ap.cr0 = 0x80050023;
-       ap.cr4 = mmu_cr4_features;
-       vmi_ops.set_initial_ap_state((u32)&ap, phys_apicid);
-}
-#endif
-
-static void vmi_set_lazy_mode(enum paravirt_lazy_mode mode)
-{
-       static DEFINE_PER_CPU(enum paravirt_lazy_mode, lazy_mode);
-
-       if (!vmi_ops.set_lazy_mode)
-               return;
-
-       /* Modes should never nest or overlap */
-       BUG_ON(__get_cpu_var(lazy_mode) && !(mode == PARAVIRT_LAZY_NONE ||
-                                            mode == PARAVIRT_LAZY_FLUSH));
-
-       if (mode == PARAVIRT_LAZY_FLUSH) {
-               vmi_ops.set_lazy_mode(0);
-               vmi_ops.set_lazy_mode(__get_cpu_var(lazy_mode));
-       } else {
-               vmi_ops.set_lazy_mode(mode);
-               __get_cpu_var(lazy_mode) = mode;
-       }
-}
-
-static inline int __init check_vmi_rom(struct vrom_header *rom)
-{
-       struct pci_header *pci;
-       struct pnp_header *pnp;
-       const char *manufacturer = "UNKNOWN";
-       const char *product = "UNKNOWN";
-       const char *license = "unspecified";
-
-       if (rom->rom_signature != 0xaa55)
-               return 0;
-       if (rom->vrom_signature != VMI_SIGNATURE)
-               return 0;
-       if (rom->api_version_maj != VMI_API_REV_MAJOR ||
-           rom->api_version_min+1 < VMI_API_REV_MINOR+1) {
-               printk(KERN_WARNING "VMI: Found mismatched rom version %d.%d\n",
-                               rom->api_version_maj,
-                               rom->api_version_min);
-               return 0;
-       }
-
-       /*
-        * Relying on the VMI_SIGNATURE field is not 100% safe, so check
-        * the PCI header and device type to make sure this is really a
-        * VMI device.
-        */
-       if (!rom->pci_header_offs) {
-               printk(KERN_WARNING "VMI: ROM does not contain PCI header.\n");
-               return 0;
-       }
-
-       pci = (struct pci_header *)((char *)rom+rom->pci_header_offs);
-       if (pci->vendorID != PCI_VENDOR_ID_VMWARE ||
-           pci->deviceID != PCI_DEVICE_ID_VMWARE_VMI) {
-               /* Allow it to run... anyways, but warn */
-               printk(KERN_WARNING "VMI: ROM from unknown manufacturer\n");
-       }
-
-       if (rom->pnp_header_offs) {
-               pnp = (struct pnp_header *)((char *)rom+rom->pnp_header_offs);
-               if (pnp->manufacturer_offset)
-                       manufacturer = (const char *)rom+pnp->manufacturer_offset;
-               if (pnp->product_offset)
-                       product = (const char *)rom+pnp->product_offset;
-       }
-
-       if (rom->license_offs)
-               license = (char *)rom+rom->license_offs;
-
-       printk(KERN_INFO "VMI: Found %s %s, API version %d.%d, ROM version %d.%d\n",
-               manufacturer, product,
-               rom->api_version_maj, rom->api_version_min,
-               pci->rom_version_maj, pci->rom_version_min);
-
-       /* Don't allow BSD/MIT here for now because we don't want to end up
-          with any binary only shim layers */
-       if (strcmp(license, "GPL") && strcmp(license, "GPL v2")) {
-               printk(KERN_WARNING "VMI: Non GPL license `%s' found for ROM. Not used.\n",
-                       license);
-               return 0;
-       }
-
-       return 1;
-}
-
-/*
- * Probe for the VMI option ROM
- */
-static inline int __init probe_vmi_rom(void)
-{
-       unsigned long base;
-
-       /* VMI ROM is in option ROM area, check signature */
-       for (base = 0xC0000; base < 0xE0000; base += 2048) {
-               struct vrom_header *romstart;
-               romstart = (struct vrom_header *)isa_bus_to_virt(base);
-               if (check_vmi_rom(romstart)) {
-                       vmi_rom = romstart;
-                       return 1;
-               }
-       }
-       return 0;
-}
-
-/*
- * VMI setup common to all processors
- */
-void vmi_bringup(void)
-{
-       /* We must establish the lowmem mapping for MMU ops to work */
-       if (vmi_ops.set_linear_mapping)
-               vmi_ops.set_linear_mapping(0, (void *)__PAGE_OFFSET, max_low_pfn, 0);
-}
-
-/*
- * Return a pointer to a VMI function or NULL if unimplemented
- */
-static void *vmi_get_function(int vmicall)
-{
-       u64 reloc;
-       const struct vmi_relocation_info *rel = (struct vmi_relocation_info *)&reloc;
-       reloc = call_vrom_long_func(vmi_rom, get_reloc, vmicall);
-       BUG_ON(rel->type == VMI_RELOCATION_JUMP_REL);
-       if (rel->type == VMI_RELOCATION_CALL_REL)
-               return (void *)rel->eip;
-       else
-               return NULL;
-}
-
-/*
- * Helper macro for making the VMI paravirt-ops fill code readable.
- * For unimplemented operations, fall back to default, unless nop
- * is returned by the ROM.
- */
-#define para_fill(opname, vmicall)                             \
-do {                                                           \
-       reloc = call_vrom_long_func(vmi_rom, get_reloc,         \
-                                   VMI_CALL_##vmicall);        \
-       if (rel->type == VMI_RELOCATION_CALL_REL)               \
-               paravirt_ops.opname = (void *)rel->eip;         \
-       else if (rel->type == VMI_RELOCATION_NOP)               \
-               paravirt_ops.opname = (void *)vmi_nop;          \
-       else if (rel->type != VMI_RELOCATION_NONE)              \
-               printk(KERN_WARNING "VMI: Unknown relocation "  \
-                                   "type %d for " #vmicall"\n",\
-                                       rel->type);             \
-} while (0)
-
-/*
- * Helper macro for making the VMI paravirt-ops fill code readable.
- * For cached operations which do not match the VMI ROM ABI and must
- * go through a tranlation stub.  Ignore NOPs, since it is not clear
- * a NOP * VMI function corresponds to a NOP paravirt-op when the
- * functions are not in 1-1 correspondence.
- */
-#define para_wrap(opname, wrapper, cache, vmicall)             \
-do {                                                           \
-       reloc = call_vrom_long_func(vmi_rom, get_reloc,         \
-                                   VMI_CALL_##vmicall);        \
-       BUG_ON(rel->type == VMI_RELOCATION_JUMP_REL);           \
-       if (rel->type == VMI_RELOCATION_CALL_REL) {             \
-               paravirt_ops.opname = wrapper;                  \
-               vmi_ops.cache = (void *)rel->eip;               \
-       }                                                       \
-} while (0)
-
-/*
- * Activate the VMI interface and switch into paravirtualized mode
- */
-static inline int __init activate_vmi(void)
-{
-       short kernel_cs;
-       u64 reloc;
-       const struct vmi_relocation_info *rel = (struct vmi_relocation_info *)&reloc;
-
-       if (call_vrom_func(vmi_rom, vmi_init) != 0) {
-               printk(KERN_ERR "VMI ROM failed to initialize!");
-               return 0;
-       }
-       savesegment(cs, kernel_cs);
-
-       paravirt_ops.paravirt_enabled = 1;
-       paravirt_ops.kernel_rpl = kernel_cs & SEGMENT_RPL_MASK;
-
-       paravirt_ops.patch = vmi_patch;
-       paravirt_ops.name = "vmi";
-
-       /*
-        * Many of these operations are ABI compatible with VMI.
-        * This means we can fill in the paravirt-ops with direct
-        * pointers into the VMI ROM.  If the calling convention for
-        * these operations changes, this code needs to be updated.
-        *
-        * Exceptions
-        *  CPUID paravirt-op uses pointers, not the native ISA
-        *  halt has no VMI equivalent; all VMI halts are "safe"
-        *  no MSR support yet - just trap and emulate.  VMI uses the
-        *    same ABI as the native ISA, but Linux wants exceptions
-        *    from bogus MSR read / write handled
-        *  rdpmc is not yet used in Linux
-        */
-
-       /* CPUID is special, so very special it gets wrapped like a present */
-       para_wrap(cpuid, vmi_cpuid, cpuid, CPUID);
-
-       para_fill(clts, CLTS);
-       para_fill(get_debugreg, GetDR);
-       para_fill(set_debugreg, SetDR);
-       para_fill(read_cr0, GetCR0);
-       para_fill(read_cr2, GetCR2);
-       para_fill(read_cr3, GetCR3);
-       para_fill(read_cr4, GetCR4);
-       para_fill(write_cr0, SetCR0);
-       para_fill(write_cr2, SetCR2);
-       para_fill(write_cr3, SetCR3);
-       para_fill(write_cr4, SetCR4);
-       para_fill(save_fl, GetInterruptMask);
-       para_fill(restore_fl, SetInterruptMask);
-       para_fill(irq_disable, DisableInterrupts);
-       para_fill(irq_enable, EnableInterrupts);
-
-       para_fill(wbinvd, WBINVD);
-       para_fill(read_tsc, RDTSC);
-
-       /* The following we emulate with trap and emulate for now */
-       /* paravirt_ops.read_msr = vmi_rdmsr */
-       /* paravirt_ops.write_msr = vmi_wrmsr */
-       /* paravirt_ops.rdpmc = vmi_rdpmc */
-
-       /* TR interface doesn't pass TR value, wrap */
-       para_wrap(load_tr_desc, vmi_set_tr, set_tr, SetTR);
-
-       /* LDT is special, too */
-       para_wrap(set_ldt, vmi_set_ldt, _set_ldt, SetLDT);
-
-       para_fill(load_gdt, SetGDT);
-       para_fill(load_idt, SetIDT);
-       para_fill(store_gdt, GetGDT);
-       para_fill(store_idt, GetIDT);
-       para_fill(store_tr, GetTR);
-       paravirt_ops.load_tls = vmi_load_tls;
-       para_fill(write_ldt_entry, WriteLDTEntry);
-       para_fill(write_gdt_entry, WriteGDTEntry);
-       para_fill(write_idt_entry, WriteIDTEntry);
-       para_wrap(load_esp0, vmi_load_esp0, set_kernel_stack, UpdateKernelStack);
-       para_fill(set_iopl_mask, SetIOPLMask);
-       para_fill(io_delay, IODelay);
-       para_wrap(set_lazy_mode, vmi_set_lazy_mode, set_lazy_mode, SetLazyMode);
-
-       /* user and kernel flush are just handled with different flags to FlushTLB */
-       para_wrap(flush_tlb_user, vmi_flush_tlb_user, _flush_tlb, FlushTLB);
-       para_wrap(flush_tlb_kernel, vmi_flush_tlb_kernel, _flush_tlb, FlushTLB);
-       para_fill(flush_tlb_single, InvalPage);
-
-       /*
-        * Until a standard flag format can be agreed on, we need to
-        * implement these as wrappers in Linux.  Get the VMI ROM
-        * function pointers for the two backend calls.
-        */
-#ifdef CONFIG_X86_PAE
-       vmi_ops.set_pte = vmi_get_function(VMI_CALL_SetPxELong);
-       vmi_ops.update_pte = vmi_get_function(VMI_CALL_UpdatePxELong);
-#else
-       vmi_ops.set_pte = vmi_get_function(VMI_CALL_SetPxE);
-       vmi_ops.update_pte = vmi_get_function(VMI_CALL_UpdatePxE);
-#endif
-
-       if (vmi_ops.set_pte) {
-               paravirt_ops.set_pte = vmi_set_pte;
-               paravirt_ops.set_pte_at = vmi_set_pte_at;
-               paravirt_ops.set_pmd = vmi_set_pmd;
-#ifdef CONFIG_X86_PAE
-               paravirt_ops.set_pte_atomic = vmi_set_pte_atomic;
-               paravirt_ops.set_pte_present = vmi_set_pte_present;
-               paravirt_ops.set_pud = vmi_set_pud;
-               paravirt_ops.pte_clear = vmi_pte_clear;
-               paravirt_ops.pmd_clear = vmi_pmd_clear;
-#endif
-       }
-
-       if (vmi_ops.update_pte) {
-               paravirt_ops.pte_update = vmi_update_pte;
-               paravirt_ops.pte_update_defer = vmi_update_pte_defer;
-       }
-
-       vmi_ops.allocate_page = vmi_get_function(VMI_CALL_AllocatePage);
-       if (vmi_ops.allocate_page) {
-               paravirt_ops.alloc_pt = vmi_allocate_pt;
-               paravirt_ops.alloc_pd = vmi_allocate_pd;
-               paravirt_ops.alloc_pd_clone = vmi_allocate_pd_clone;
-       }
-
-       vmi_ops.release_page = vmi_get_function(VMI_CALL_ReleasePage);
-       if (vmi_ops.release_page) {
-               paravirt_ops.release_pt = vmi_release_pt;
-               paravirt_ops.release_pd = vmi_release_pd;
-       }
-
-       /* Set linear is needed in all cases */
-       vmi_ops.set_linear_mapping = vmi_get_function(VMI_CALL_SetLinearMapping);
-#ifdef CONFIG_HIGHPTE
-       if (vmi_ops.set_linear_mapping)
-               paravirt_ops.kmap_atomic_pte = vmi_kmap_atomic_pte;
-#endif
-
-       /*
-        * These MUST always be patched.  Don't support indirect jumps
-        * through these operations, as the VMI interface may use either
-        * a jump or a call to get to these operations, depending on
-        * the backend.  They are performance critical anyway, so requiring
-        * a patch is not a big problem.
-        */
-       paravirt_ops.irq_enable_sysexit = (void *)0xfeedbab0;
-       paravirt_ops.iret = (void *)0xbadbab0;
-
-#ifdef CONFIG_SMP
-       para_wrap(startup_ipi_hook, vmi_startup_ipi_hook, set_initial_ap_state, SetInitialAPState);
-#endif
-
-#ifdef CONFIG_X86_LOCAL_APIC
-       para_fill(apic_read, APICRead);
-       para_fill(apic_write, APICWrite);
-       para_fill(apic_write_atomic, APICWrite);
-#endif
-
-       /*
-        * Check for VMI timer functionality by probing for a cycle frequency method
-        */
-       reloc = call_vrom_long_func(vmi_rom, get_reloc, VMI_CALL_GetCycleFrequency);
-       if (!disable_vmi_timer && rel->type != VMI_RELOCATION_NONE) {
-               vmi_timer_ops.get_cycle_frequency = (void *)rel->eip;
-               vmi_timer_ops.get_cycle_counter =
-                       vmi_get_function(VMI_CALL_GetCycleCounter);
-               vmi_timer_ops.get_wallclock =
-                       vmi_get_function(VMI_CALL_GetWallclockTime);
-               vmi_timer_ops.wallclock_updated =
-                       vmi_get_function(VMI_CALL_WallclockUpdated);
-               vmi_timer_ops.set_alarm = vmi_get_function(VMI_CALL_SetAlarm);
-               vmi_timer_ops.cancel_alarm =
-                        vmi_get_function(VMI_CALL_CancelAlarm);
-               paravirt_ops.time_init = vmi_time_init;
-               paravirt_ops.get_wallclock = vmi_get_wallclock;
-               paravirt_ops.set_wallclock = vmi_set_wallclock;
-#ifdef CONFIG_X86_LOCAL_APIC
-               paravirt_ops.setup_boot_clock = vmi_time_bsp_init;
-               paravirt_ops.setup_secondary_clock = vmi_time_ap_init;
-#endif
-               paravirt_ops.sched_clock = vmi_sched_clock;
-               paravirt_ops.get_cpu_khz = vmi_cpu_khz;
-
-               /* We have true wallclock functions; disable CMOS clock sync */
-               no_sync_cmos_clock = 1;
-       } else {
-               disable_noidle = 1;
-               disable_vmi_timer = 1;
-       }
-
-       para_fill(safe_halt, Halt);
-
-       /*
-        * Alternative instruction rewriting doesn't happen soon enough
-        * to convert VMI_IRET to a call instead of a jump; so we have
-        * to do this before IRQs get reenabled.  Fortunately, it is
-        * idempotent.
-        */
-       apply_paravirt(__parainstructions, __parainstructions_end);
-
-       vmi_bringup();
-
-       return 1;
-}
-
-#undef para_fill
-
-void __init vmi_init(void)
-{
-       unsigned long flags;
-
-       if (!vmi_rom)
-               probe_vmi_rom();
-       else
-               check_vmi_rom(vmi_rom);
-
-       /* In case probing for or validating the ROM failed, basil */
-       if (!vmi_rom)
-               return;
-
-       reserve_top_address(-vmi_rom->virtual_top);
-
-       local_irq_save(flags);
-       activate_vmi();
-
-#ifdef CONFIG_X86_IO_APIC
-       /* This is virtual hardware; timer routing is wired correctly */
-       no_timer_check = 1;
-#endif
-       local_irq_restore(flags & X86_EFLAGS_IF);
-}
-
-static int __init parse_vmi(char *arg)
-{
-       if (!arg)
-               return -EINVAL;
-
-       if (!strcmp(arg, "disable_pge")) {
-               clear_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability);
-               disable_pge = 1;
-       } else if (!strcmp(arg, "disable_pse")) {
-               clear_bit(X86_FEATURE_PSE, boot_cpu_data.x86_capability);
-               disable_pse = 1;
-       } else if (!strcmp(arg, "disable_sep")) {
-               clear_bit(X86_FEATURE_SEP, boot_cpu_data.x86_capability);
-               disable_sep = 1;
-       } else if (!strcmp(arg, "disable_tsc")) {
-               clear_bit(X86_FEATURE_TSC, boot_cpu_data.x86_capability);
-               disable_tsc = 1;
-       } else if (!strcmp(arg, "disable_mtrr")) {
-               clear_bit(X86_FEATURE_MTRR, boot_cpu_data.x86_capability);
-               disable_mtrr = 1;
-       } else if (!strcmp(arg, "disable_timer")) {
-               disable_vmi_timer = 1;
-               disable_noidle = 1;
-       } else if (!strcmp(arg, "disable_noidle"))
-               disable_noidle = 1;
-       return 0;
-}
-
-early_param("vmi", parse_vmi);
diff --git a/arch/i386/kernel/vmi_32.c b/arch/i386/kernel/vmi_32.c
new file mode 100644 (file)
index 0000000..18673e0
--- /dev/null
@@ -0,0 +1,981 @@
+/*
+ * VMI specific paravirt-ops implementation
+ *
+ * Copyright (C) 2005, VMware, Inc.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ * NON INFRINGEMENT.  See the GNU General Public License for more
+ * details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ *
+ * Send feedback to zach@vmware.com
+ *
+ */
+
+#include <linux/module.h>
+#include <linux/cpu.h>
+#include <linux/bootmem.h>
+#include <linux/mm.h>
+#include <linux/highmem.h>
+#include <linux/sched.h>
+#include <asm/vmi.h>
+#include <asm/io.h>
+#include <asm/fixmap.h>
+#include <asm/apicdef.h>
+#include <asm/apic.h>
+#include <asm/processor.h>
+#include <asm/timer.h>
+#include <asm/vmi_time.h>
+#include <asm/kmap_types.h>
+
+/* Convenient for calling VMI functions indirectly in the ROM */
+typedef u32 __attribute__((regparm(1))) (VROMFUNC)(void);
+typedef u64 __attribute__((regparm(2))) (VROMLONGFUNC)(int);
+
+#define call_vrom_func(rom,func) \
+   (((VROMFUNC *)(rom->func))())
+
+#define call_vrom_long_func(rom,func,arg) \
+   (((VROMLONGFUNC *)(rom->func)) (arg))
+
+static struct vrom_header *vmi_rom;
+static int disable_pge;
+static int disable_pse;
+static int disable_sep;
+static int disable_tsc;
+static int disable_mtrr;
+static int disable_noidle;
+static int disable_vmi_timer;
+
+/* Cached VMI operations */
+static struct {
+       void (*cpuid)(void /* non-c */);
+       void (*_set_ldt)(u32 selector);
+       void (*set_tr)(u32 selector);
+       void (*set_kernel_stack)(u32 selector, u32 esp0);
+       void (*allocate_page)(u32, u32, u32, u32, u32);
+       void (*release_page)(u32, u32);
+       void (*set_pte)(pte_t, pte_t *, unsigned);
+       void (*update_pte)(pte_t *, unsigned);
+       void (*set_linear_mapping)(int, void *, u32, u32);
+       void (*_flush_tlb)(int);
+       void (*set_initial_ap_state)(int, int);
+       void (*halt)(void);
+       void (*set_lazy_mode)(int mode);
+} vmi_ops;
+
+/* Cached VMI operations */
+struct vmi_timer_ops vmi_timer_ops;
+
+/*
+ * VMI patching routines.
+ */
+#define MNEM_CALL 0xe8
+#define MNEM_JMP  0xe9
+#define MNEM_RET  0xc3
+
+#define IRQ_PATCH_INT_MASK 0
+#define IRQ_PATCH_DISABLE  5
+
+static inline void patch_offset(void *insnbuf,
+                               unsigned long eip, unsigned long dest)
+{
+        *(unsigned long *)(insnbuf+1) = dest-eip-5;
+}
+
+static unsigned patch_internal(int call, unsigned len, void *insnbuf,
+                              unsigned long eip)
+{
+       u64 reloc;
+       struct vmi_relocation_info *const rel = (struct vmi_relocation_info *)&reloc;
+       reloc = call_vrom_long_func(vmi_rom, get_reloc, call);
+       switch(rel->type) {
+               case VMI_RELOCATION_CALL_REL:
+                       BUG_ON(len < 5);
+                       *(char *)insnbuf = MNEM_CALL;
+                       patch_offset(insnbuf, eip, (unsigned long)rel->eip);
+                       return 5;
+
+               case VMI_RELOCATION_JUMP_REL:
+                       BUG_ON(len < 5);
+                       *(char *)insnbuf = MNEM_JMP;
+                       patch_offset(insnbuf, eip, (unsigned long)rel->eip);
+                       return 5;
+
+               case VMI_RELOCATION_NOP:
+                       /* obliterate the whole thing */
+                       return 0;
+
+               case VMI_RELOCATION_NONE:
+                       /* leave native code in place */
+                       break;
+
+               default:
+                       BUG();
+       }
+       return len;
+}
+
+/*
+ * Apply patch if appropriate, return length of new instruction
+ * sequence.  The callee does nop padding for us.
+ */
+static unsigned vmi_patch(u8 type, u16 clobbers, void *insns,
+                         unsigned long eip, unsigned len)
+{
+       switch (type) {
+               case PARAVIRT_PATCH(irq_disable):
+                       return patch_internal(VMI_CALL_DisableInterrupts, len,
+                                             insns, eip);
+               case PARAVIRT_PATCH(irq_enable):
+                       return patch_internal(VMI_CALL_EnableInterrupts, len,
+                                             insns, eip);
+               case PARAVIRT_PATCH(restore_fl):
+                       return patch_internal(VMI_CALL_SetInterruptMask, len,
+                                             insns, eip);
+               case PARAVIRT_PATCH(save_fl):
+                       return patch_internal(VMI_CALL_GetInterruptMask, len,
+                                             insns, eip);
+               case PARAVIRT_PATCH(iret):
+                       return patch_internal(VMI_CALL_IRET, len, insns, eip);
+               case PARAVIRT_PATCH(irq_enable_sysexit):
+                       return patch_internal(VMI_CALL_SYSEXIT, len, insns, eip);
+               default:
+                       break;
+       }
+       return len;
+}
+
+/* CPUID has non-C semantics, and paravirt-ops API doesn't match hardware ISA */
+static void vmi_cpuid(unsigned int *eax, unsigned int *ebx,
+                               unsigned int *ecx, unsigned int *edx)
+{
+       int override = 0;
+       if (*eax == 1)
+               override = 1;
+        asm volatile ("call *%6"
+                      : "=a" (*eax),
+                        "=b" (*ebx),
+                        "=c" (*ecx),
+                        "=d" (*edx)
+                      : "0" (*eax), "2" (*ecx), "r" (vmi_ops.cpuid));
+       if (override) {
+               if (disable_pse)
+                       *edx &= ~X86_FEATURE_PSE;
+               if (disable_pge)
+                       *edx &= ~X86_FEATURE_PGE;
+               if (disable_sep)
+                       *edx &= ~X86_FEATURE_SEP;
+               if (disable_tsc)
+                       *edx &= ~X86_FEATURE_TSC;
+               if (disable_mtrr)
+                       *edx &= ~X86_FEATURE_MTRR;
+       }
+}
+
+static inline void vmi_maybe_load_tls(struct desc_struct *gdt, int nr, struct desc_struct *new)
+{
+       if (gdt[nr].a != new->a || gdt[nr].b != new->b)
+               write_gdt_entry(gdt, nr, new->a, new->b);
+}
+
+static void vmi_load_tls(struct thread_struct *t, unsigned int cpu)
+{
+       struct desc_struct *gdt = get_cpu_gdt_table(cpu);
+       vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 0, &t->tls_array[0]);
+       vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 1, &t->tls_array[1]);
+       vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 2, &t->tls_array[2]);
+}
+
+static void vmi_set_ldt(const void *addr, unsigned entries)
+{
+       unsigned cpu = smp_processor_id();
+       u32 low, high;
+
+       pack_descriptor(&low, &high, (unsigned long)addr,
+                       entries * sizeof(struct desc_struct) - 1,
+                       DESCTYPE_LDT, 0);
+       write_gdt_entry(get_cpu_gdt_table(cpu), GDT_ENTRY_LDT, low, high);
+       vmi_ops._set_ldt(entries ? GDT_ENTRY_LDT*sizeof(struct desc_struct) : 0);
+}
+
+static void vmi_set_tr(void)
+{
+       vmi_ops.set_tr(GDT_ENTRY_TSS*sizeof(struct desc_struct));
+}
+
+static void vmi_load_esp0(struct tss_struct *tss,
+                                  struct thread_struct *thread)
+{
+       tss->x86_tss.esp0 = thread->esp0;
+
+       /* This can only happen when SEP is enabled, no need to test "SEP"arately */
+       if (unlikely(tss->x86_tss.ss1 != thread->sysenter_cs)) {
+               tss->x86_tss.ss1 = thread->sysenter_cs;
+               wrmsr(MSR_IA32_SYSENTER_CS, thread->sysenter_cs, 0);
+       }
+       vmi_ops.set_kernel_stack(__KERNEL_DS, tss->x86_tss.esp0);
+}
+
+static void vmi_flush_tlb_user(void)
+{
+       vmi_ops._flush_tlb(VMI_FLUSH_TLB);
+}
+
+static void vmi_flush_tlb_kernel(void)
+{
+       vmi_ops._flush_tlb(VMI_FLUSH_TLB | VMI_FLUSH_GLOBAL);
+}
+
+/* Stub to do nothing at all; used for delays and unimplemented calls */
+static void vmi_nop(void)
+{
+}
+
+#ifdef CONFIG_DEBUG_PAGE_TYPE
+
+#ifdef CONFIG_X86_PAE
+#define MAX_BOOT_PTS (2048+4+1)
+#else
+#define MAX_BOOT_PTS (1024+1)
+#endif
+
+/*
+ * During boot, mem_map is not yet available in paging_init, so stash
+ * all the boot page allocations here.
+ */
+static struct {
+       u32 pfn;
+       int type;
+} boot_page_allocations[MAX_BOOT_PTS];
+static int num_boot_page_allocations;
+static int boot_allocations_applied;
+
+void vmi_apply_boot_page_allocations(void)
+{
+       int i;
+       BUG_ON(!mem_map);
+       for (i = 0; i < num_boot_page_allocations; i++) {
+               struct page *page = pfn_to_page(boot_page_allocations[i].pfn);
+               page->type = boot_page_allocations[i].type;
+               page->type = boot_page_allocations[i].type &
+                               ~(VMI_PAGE_ZEROED | VMI_PAGE_CLONE);
+       }
+       boot_allocations_applied = 1;
+}
+
+static void record_page_type(u32 pfn, int type)
+{
+       BUG_ON(num_boot_page_allocations >= MAX_BOOT_PTS);
+       boot_page_allocations[num_boot_page_allocations].pfn = pfn;
+       boot_page_allocations[num_boot_page_allocations].type = type;
+       num_boot_page_allocations++;
+}
+
+static void check_zeroed_page(u32 pfn, int type, struct page *page)
+{
+       u32 *ptr;
+       int i;
+       int limit = PAGE_SIZE / sizeof(int);
+
+       if (page_address(page))
+               ptr = (u32 *)page_address(page);
+       else
+               ptr = (u32 *)__va(pfn << PAGE_SHIFT);
+       /*
+        * When cloning the root in non-PAE mode, only the userspace
+        * pdes need to be zeroed.
+        */
+       if (type & VMI_PAGE_CLONE)
+               limit = USER_PTRS_PER_PGD;
+       for (i = 0; i < limit; i++)
+               BUG_ON(ptr[i]);
+}
+
+/*
+ * We stash the page type into struct page so we can verify the page
+ * types are used properly.
+ */
+static void vmi_set_page_type(u32 pfn, int type)
+{
+       /* PAE can have multiple roots per page - don't track */
+       if (PTRS_PER_PMD > 1 && (type & VMI_PAGE_PDP))
+               return;
+
+       if (boot_allocations_applied) {
+               struct page *page = pfn_to_page(pfn);
+               if (type != VMI_PAGE_NORMAL)
+                       BUG_ON(page->type);
+               else
+                       BUG_ON(page->type == VMI_PAGE_NORMAL);
+               page->type = type & ~(VMI_PAGE_ZEROED | VMI_PAGE_CLONE);
+               if (type & VMI_PAGE_ZEROED)
+                       check_zeroed_page(pfn, type, page);
+       } else {
+               record_page_type(pfn, type);
+       }
+}
+
+static void vmi_check_page_type(u32 pfn, int type)
+{
+       /* PAE can have multiple roots per page - skip checks */
+       if (PTRS_PER_PMD > 1 && (type & VMI_PAGE_PDP))
+               return;
+
+       type &= ~(VMI_PAGE_ZEROED | VMI_PAGE_CLONE);
+       if (boot_allocations_applied) {
+               struct page *page = pfn_to_page(pfn);
+               BUG_ON((page->type ^ type) & VMI_PAGE_PAE);
+               BUG_ON(type == VMI_PAGE_NORMAL && page->type);
+               BUG_ON((type & page->type) == 0);
+       }
+}
+#else
+#define vmi_set_page_type(p,t) do { } while (0)
+#define vmi_check_page_type(p,t) do { } while (0)
+#endif
+
+#ifdef CONFIG_HIGHPTE
+static void *vmi_kmap_atomic_pte(struct page *page, enum km_type type)
+{
+       void *va = kmap_atomic(page, type);
+
+       /*
+        * Internally, the VMI ROM must map virtual addresses to physical
+        * addresses for processing MMU updates.  By the time MMU updates
+        * are issued, this information is typically already lost.
+        * Fortunately, the VMI provides a cache of mapping slots for active
+        * page tables.
+        *
+        * We use slot zero for the linear mapping of physical memory, and
+        * in HIGHPTE kernels, slot 1 and 2 for KM_PTE0 and KM_PTE1.
+        *
+        *  args:                 SLOT                 VA    COUNT PFN
+        */
+       BUG_ON(type != KM_PTE0 && type != KM_PTE1);
+       vmi_ops.set_linear_mapping((type - KM_PTE0)+1, va, 1, page_to_pfn(page));
+
+       return va;
+}
+#endif
+
+static void vmi_allocate_pt(struct mm_struct *mm, u32 pfn)
+{
+       vmi_set_page_type(pfn, VMI_PAGE_L1);
+       vmi_ops.allocate_page(pfn, VMI_PAGE_L1, 0, 0, 0);
+}
+
+static void vmi_allocate_pd(u32 pfn)
+{
+       /*
+        * This call comes in very early, before mem_map is setup.
+        * It is called only for swapper_pg_dir, which already has
+        * data on it.
+        */
+       vmi_set_page_type(pfn, VMI_PAGE_L2);
+       vmi_ops.allocate_page(pfn, VMI_PAGE_L2, 0, 0, 0);
+}
+
+static void vmi_allocate_pd_clone(u32 pfn, u32 clonepfn, u32 start, u32 count)
+{
+       vmi_set_page_type(pfn, VMI_PAGE_L2 | VMI_PAGE_CLONE);
+       vmi_check_page_type(clonepfn, VMI_PAGE_L2);
+       vmi_ops.allocate_page(pfn, VMI_PAGE_L2 | VMI_PAGE_CLONE, clonepfn, start, count);
+}
+
+static void vmi_release_pt(u32 pfn)
+{
+       vmi_ops.release_page(pfn, VMI_PAGE_L1);
+       vmi_set_page_type(pfn, VMI_PAGE_NORMAL);
+}
+
+static void vmi_release_pd(u32 pfn)
+{
+       vmi_ops.release_page(pfn, VMI_PAGE_L2);
+       vmi_set_page_type(pfn, VMI_PAGE_NORMAL);
+}
+
+/*
+ * Helper macros for MMU update flags.  We can defer updates until a flush
+ * or page invalidation only if the update is to the current address space
+ * (otherwise, there is no flush).  We must check against init_mm, since
+ * this could be a kernel update, which usually passes init_mm, although
+ * sometimes this check can be skipped if we know the particular function
+ * is only called on user mode PTEs.  We could change the kernel to pass
+ * current->active_mm here, but in particular, I was unsure if changing
+ * mm/highmem.c to do this would still be correct on other architectures.
+ */
+#define is_current_as(mm, mustbeuser) ((mm) == current->active_mm ||    \
+                                       (!mustbeuser && (mm) == &init_mm))
+#define vmi_flags_addr(mm, addr, level, user)                           \
+        ((level) | (is_current_as(mm, user) ?                           \
+                (VMI_PAGE_CURRENT_AS | ((addr) & VMI_PAGE_VA_MASK)) : 0))
+#define vmi_flags_addr_defer(mm, addr, level, user)                     \
+        ((level) | (is_current_as(mm, user) ?                           \
+                (VMI_PAGE_DEFER | VMI_PAGE_CURRENT_AS | ((addr) & VMI_PAGE_VA_MASK)) : 0))
+
+static void vmi_update_pte(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
+{
+       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
+       vmi_ops.update_pte(ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
+}
+
+static void vmi_update_pte_defer(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
+{
+       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
+       vmi_ops.update_pte(ptep, vmi_flags_addr_defer(mm, addr, VMI_PAGE_PT, 0));
+}
+
+static void vmi_set_pte(pte_t *ptep, pte_t pte)
+{
+       /* XXX because of set_pmd_pte, this can be called on PT or PD layers */
+       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE | VMI_PAGE_PD);
+       vmi_ops.set_pte(pte, ptep, VMI_PAGE_PT);
+}
+
+static void vmi_set_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
+{
+       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
+       vmi_ops.set_pte(pte, ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
+}
+
+static void vmi_set_pmd(pmd_t *pmdp, pmd_t pmdval)
+{
+#ifdef CONFIG_X86_PAE
+       const pte_t pte = { pmdval.pmd, pmdval.pmd >> 32 };
+       vmi_check_page_type(__pa(pmdp) >> PAGE_SHIFT, VMI_PAGE_PMD);
+#else
+       const pte_t pte = { pmdval.pud.pgd.pgd };
+       vmi_check_page_type(__pa(pmdp) >> PAGE_SHIFT, VMI_PAGE_PGD);
+#endif
+       vmi_ops.set_pte(pte, (pte_t *)pmdp, VMI_PAGE_PD);
+}
+
+#ifdef CONFIG_X86_PAE
+
+static void vmi_set_pte_atomic(pte_t *ptep, pte_t pteval)
+{
+       /*
+        * XXX This is called from set_pmd_pte, but at both PT
+        * and PD layers so the VMI_PAGE_PT flag is wrong.  But
+        * it is only called for large page mapping changes,
+        * the Xen backend, doesn't support large pages, and the
+        * ESX backend doesn't depend on the flag.
+        */
+       set_64bit((unsigned long long *)ptep,pte_val(pteval));
+       vmi_ops.update_pte(ptep, VMI_PAGE_PT);
+}
+
+static void vmi_set_pte_present(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
+{
+       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
+       vmi_ops.set_pte(pte, ptep, vmi_flags_addr_defer(mm, addr, VMI_PAGE_PT, 1));
+}
+
+static void vmi_set_pud(pud_t *pudp, pud_t pudval)
+{
+       /* Um, eww */
+       const pte_t pte = { pudval.pgd.pgd, pudval.pgd.pgd >> 32 };
+       vmi_check_page_type(__pa(pudp) >> PAGE_SHIFT, VMI_PAGE_PGD);
+       vmi_ops.set_pte(pte, (pte_t *)pudp, VMI_PAGE_PDP);
+}
+
+static void vmi_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
+{
+       const pte_t pte = { 0 };
+       vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
+       vmi_ops.set_pte(pte, ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
+}
+
+static void vmi_pmd_clear(pmd_t *pmd)
+{
+       const pte_t pte = { 0 };
+       vmi_check_page_type(__pa(pmd) >> PAGE_SHIFT, VMI_PAGE_PMD);
+       vmi_ops.set_pte(pte, (pte_t *)pmd, VMI_PAGE_PD);
+}
+#endif
+
+#ifdef CONFIG_SMP
+static void __devinit
+vmi_startup_ipi_hook(int phys_apicid, unsigned long start_eip,
+                    unsigned long start_esp)
+{
+       struct vmi_ap_state ap;
+
+       /* Default everything to zero.  This is fine for most GPRs. */
+       memset(&ap, 0, sizeof(struct vmi_ap_state));
+
+       ap.gdtr_limit = GDT_SIZE - 1;
+       ap.gdtr_base = (unsigned long) get_cpu_gdt_table(phys_apicid);
+
+       ap.idtr_limit = IDT_ENTRIES * 8 - 1;
+       ap.idtr_base = (unsigned long) idt_table;
+
+       ap.ldtr = 0;
+
+       ap.cs = __KERNEL_CS;
+       ap.eip = (unsigned long) start_eip;
+       ap.ss = __KERNEL_DS;
+       ap.esp = (unsigned long) start_esp;
+
+       ap.ds = __USER_DS;
+       ap.es = __USER_DS;
+       ap.fs = __KERNEL_PERCPU;
+       ap.gs = 0;
+
+       ap.eflags = 0;
+
+#ifdef CONFIG_X86_PAE
+       /* efer should match BSP efer. */
+       if (cpu_has_nx) {
+               unsigned l, h;
+               rdmsr(MSR_EFER, l, h);
+               ap.efer = (unsigned long long) h << 32 | l;
+       }
+#endif
+
+       ap.cr3 = __pa(swapper_pg_dir);
+       /* Protected mode, paging, AM, WP, NE, MP. */
+       ap.cr0 = 0x80050023;
+       ap.cr4 = mmu_cr4_features;
+       vmi_ops.set_initial_ap_state((u32)&ap, phys_apicid);
+}
+#endif
+
+static void vmi_set_lazy_mode(enum paravirt_lazy_mode mode)
+{
+       static DEFINE_PER_CPU(enum paravirt_lazy_mode, lazy_mode);
+
+       if (!vmi_ops.set_lazy_mode)
+               return;
+
+       /* Modes should never nest or overlap */
+       BUG_ON(__get_cpu_var(lazy_mode) && !(mode == PARAVIRT_LAZY_NONE ||
+                                            mode == PARAVIRT_LAZY_FLUSH));
+
+       if (mode == PARAVIRT_LAZY_FLUSH) {
+               vmi_ops.set_lazy_mode(0);
+               vmi_ops.set_lazy_mode(__get_cpu_var(lazy_mode));
+       } else {
+               vmi_ops.set_lazy_mode(mode);
+               __get_cpu_var(lazy_mode) = mode;
+       }
+}
+
+static inline int __init check_vmi_rom(struct vrom_header *rom)
+{
+       struct pci_header *pci;
+       struct pnp_header *pnp;
+       const char *manufacturer = "UNKNOWN";
+       const char *product = "UNKNOWN";
+       const char *license = "unspecified";
+
+       if (rom->rom_signature != 0xaa55)
+               return 0;
+       if (rom->vrom_signature != VMI_SIGNATURE)
+               return 0;
+       if (rom->api_version_maj != VMI_API_REV_MAJOR ||
+           rom->api_version_min+1 < VMI_API_REV_MINOR+1) {
+               printk(KERN_WARNING "VMI: Found mismatched rom version %d.%d\n",
+                               rom->api_version_maj,
+                               rom->api_version_min);
+               return 0;
+       }
+
+       /*
+        * Relying on the VMI_SIGNATURE field is not 100% safe, so check
+        * the PCI header and device type to make sure this is really a
+        * VMI device.
+        */
+       if (!rom->pci_header_offs) {
+               printk(KERN_WARNING "VMI: ROM does not contain PCI header.\n");
+               return 0;
+       }
+
+       pci = (struct pci_header *)((char *)rom+rom->pci_header_offs);
+       if (pci->vendorID != PCI_VENDOR_ID_VMWARE ||
+           pci->deviceID != PCI_DEVICE_ID_VMWARE_VMI) {
+               /* Allow it to run... anyways, but warn */
+               printk(KERN_WARNING "VMI: ROM from unknown manufacturer\n");
+       }
+
+       if (rom->pnp_header_offs) {
+               pnp = (struct pnp_header *)((char *)rom+rom->pnp_header_offs);
+               if (pnp->manufacturer_offset)
+                       manufacturer = (const char *)rom+pnp->manufacturer_offset;
+               if (pnp->product_offset)
+                       product = (const char *)rom+pnp->product_offset;
+       }
+
+       if (rom->license_offs)
+               license = (char *)rom+rom->license_offs;
+
+       printk(KERN_INFO "VMI: Found %s %s, API version %d.%d, ROM version %d.%d\n",
+               manufacturer, product,
+               rom->api_version_maj, rom->api_version_min,
+               pci->rom_version_maj, pci->rom_version_min);
+
+       /* Don't allow BSD/MIT here for now because we don't want to end up
+          with any binary only shim layers */
+       if (strcmp(license, "GPL") && strcmp(license, "GPL v2")) {
+               printk(KERN_WARNING "VMI: Non GPL license `%s' found for ROM. Not used.\n",
+                       license);
+               return 0;
+       }
+
+       return 1;
+}
+
+/*
+ * Probe for the VMI option ROM
+ */
+static inline int __init probe_vmi_rom(void)
+{
+       unsigned long base;
+
+       /* VMI ROM is in option ROM area, check signature */
+       for (base = 0xC0000; base < 0xE0000; base += 2048) {
+               struct vrom_header *romstart;
+               romstart = (struct vrom_header *)isa_bus_to_virt(base);
+               if (check_vmi_rom(romstart)) {
+                       vmi_rom = romstart;
+                       return 1;
+               }
+       }
+       return 0;
+}
+
+/*
+ * VMI setup common to all processors
+ */
+void vmi_bringup(void)
+{
+       /* We must establish the lowmem mapping for MMU ops to work */
+       if (vmi_ops.set_linear_mapping)
+               vmi_ops.set_linear_mapping(0, (void *)__PAGE_OFFSET, max_low_pfn, 0);
+}
+
+/*
+ * Return a pointer to a VMI function or NULL if unimplemented
+ */
+static void *vmi_get_function(int vmicall)
+{
+       u64 reloc;
+       const struct vmi_relocation_info *rel = (struct vmi_relocation_info *)&reloc;
+       reloc = call_vrom_long_func(vmi_rom, get_reloc, vmicall);
+       BUG_ON(rel->type == VMI_RELOCATION_JUMP_REL);
+       if (rel->type == VMI_RELOCATION_CALL_REL)
+               return (void *)rel->eip;
+       else
+               return NULL;
+}
+
+/*
+ * Helper macro for making the VMI paravirt-ops fill code readable.
+ * For unimplemented operations, fall back to default, unless nop
+ * is returned by the ROM.
+ */
+#define para_fill(opname, vmicall)                             \
+do {                                                           \
+       reloc = call_vrom_long_func(vmi_rom, get_reloc,         \
+                                   VMI_CALL_##vmicall);        \
+       if (rel->type == VMI_RELOCATION_CALL_REL)               \
+               paravirt_ops.opname = (void *)rel->eip;         \
+       else if (rel->type == VMI_RELOCATION_NOP)               \
+               paravirt_ops.opname = (void *)vmi_nop;          \
+       else if (rel->type != VMI_RELOCATION_NONE)              \
+               printk(KERN_WARNING "VMI: Unknown relocation "  \
+                                   "type %d for " #vmicall"\n",\
+                                       rel->type);             \
+} while (0)
+
+/*
+ * Helper macro for making the VMI paravirt-ops fill code readable.
+ * For cached operations which do not match the VMI ROM ABI and must
+ * go through a tranlation stub.  Ignore NOPs, since it is not clear
+ * a NOP * VMI function corresponds to a NOP paravirt-op when the
+ * functions are not in 1-1 correspondence.
+ */
+#define para_wrap(opname, wrapper, cache, vmicall)             \
+do {                                                           \
+       reloc = call_vrom_long_func(vmi_rom, get_reloc,         \
+                                   VMI_CALL_##vmicall);        \
+       BUG_ON(rel->type == VMI_RELOCATION_JUMP_REL);           \
+       if (rel->type == VMI_RELOCATION_CALL_REL) {             \
+               paravirt_ops.opname = wrapper;                  \
+               vmi_ops.cache = (void *)rel->eip;               \
+       }                                                       \
+} while (0)
+
+/*
+ * Activate the VMI interface and switch into paravirtualized mode
+ */
+static inline int __init activate_vmi(void)
+{
+       short kernel_cs;
+       u64 reloc;
+       const struct vmi_relocation_info *rel = (struct vmi_relocation_info *)&reloc;
+
+       if (call_vrom_func(vmi_rom, vmi_init) != 0) {
+               printk(KERN_ERR "VMI ROM failed to initialize!");
+               return 0;
+       }
+       savesegment(cs, kernel_cs);
+
+       paravirt_ops.paravirt_enabled = 1;
+       paravirt_ops.kernel_rpl = kernel_cs & SEGMENT_RPL_MASK;
+
+       paravirt_ops.patch = vmi_patch;
+       paravirt_ops.name = "vmi";
+
+       /*
+        * Many of these operations are ABI compatible with VMI.
+        * This means we can fill in the paravirt-ops with direct
+        * pointers into the VMI ROM.  If the calling convention for
+        * these operations changes, this code needs to be updated.
+        *
+        * Exceptions
+        *  CPUID paravirt-op uses pointers, not the native ISA
+        *  halt has no VMI equivalent; all VMI halts are "safe"
+        *  no MSR support yet - just trap and emulate.  VMI uses the
+        *    same ABI as the native ISA, but Linux wants exceptions
+        *    from bogus MSR read / write handled
+        *  rdpmc is not yet used in Linux
+        */
+
+       /* CPUID is special, so very special it gets wrapped like a present */
+       para_wrap(cpuid, vmi_cpuid, cpuid, CPUID);
+
+       para_fill(clts, CLTS);
+       para_fill(get_debugreg, GetDR);
+       para_fill(set_debugreg, SetDR);
+       para_fill(read_cr0, GetCR0);
+       para_fill(read_cr2, GetCR2);
+       para_fill(read_cr3, GetCR3);
+       para_fill(read_cr4, GetCR4);
+       para_fill(write_cr0, SetCR0);
+       para_fill(write_cr2, SetCR2);
+       para_fill(write_cr3, SetCR3);
+       para_fill(write_cr4, SetCR4);
+       para_fill(save_fl, GetInterruptMask);
+       para_fill(restore_fl, SetInterruptMask);
+       para_fill(irq_disable, DisableInterrupts);
+       para_fill(irq_enable, EnableInterrupts);
+
+       para_fill(wbinvd, WBINVD);
+       para_fill(read_tsc, RDTSC);
+
+       /* The following we emulate with trap and emulate for now */
+       /* paravirt_ops.read_msr = vmi_rdmsr */
+       /* paravirt_ops.write_msr = vmi_wrmsr */
+       /* paravirt_ops.rdpmc = vmi_rdpmc */
+
+       /* TR interface doesn't pass TR value, wrap */
+       para_wrap(load_tr_desc, vmi_set_tr, set_tr, SetTR);
+
+       /* LDT is special, too */
+       para_wrap(set_ldt, vmi_set_ldt, _set_ldt, SetLDT);
+
+       para_fill(load_gdt, SetGDT);
+       para_fill(load_idt, SetIDT);
+       para_fill(store_gdt, GetGDT);
+       para_fill(store_idt, GetIDT);
+       para_fill(store_tr, GetTR);
+       paravirt_ops.load_tls = vmi_load_tls;
+       para_fill(write_ldt_entry, WriteLDTEntry);
+       para_fill(write_gdt_entry, WriteGDTEntry);
+       para_fill(write_idt_entry, WriteIDTEntry);
+       para_wrap(load_esp0, vmi_load_esp0, set_kernel_stack, UpdateKernelStack);
+       para_fill(set_iopl_mask, SetIOPLMask);
+       para_fill(io_delay, IODelay);
+       para_wrap(set_lazy_mode, vmi_set_lazy_mode, set_lazy_mode, SetLazyMode);
+
+       /* user and kernel flush are just handled with different flags to FlushTLB */
+       para_wrap(flush_tlb_user, vmi_flush_tlb_user, _flush_tlb, FlushTLB);
+       para_wrap(flush_tlb_kernel, vmi_flush_tlb_kernel, _flush_tlb, FlushTLB);
+       para_fill(flush_tlb_single, InvalPage);
+
+       /*
+        * Until a standard flag format can be agreed on, we need to
+        * implement these as wrappers in Linux.  Get the VMI ROM
+        * function pointers for the two backend calls.
+        */
+#ifdef CONFIG_X86_PAE
+       vmi_ops.set_pte = vmi_get_function(VMI_CALL_SetPxELong);
+       vmi_ops.update_pte = vmi_get_function(VMI_CALL_UpdatePxELong);
+#else
+       vmi_ops.set_pte = vmi_get_function(VMI_CALL_SetPxE);
+       vmi_ops.update_pte = vmi_get_function(VMI_CALL_UpdatePxE);
+#endif
+
+       if (vmi_ops.set_pte) {
+               paravirt_ops.set_pte = vmi_set_pte;
+               paravirt_ops.set_pte_at = vmi_set_pte_at;
+               paravirt_ops.set_pmd = vmi_set_pmd;
+#ifdef CONFIG_X86_PAE
+               paravirt_ops.set_pte_atomic = vmi_set_pte_atomic;
+               paravirt_ops.set_pte_present = vmi_set_pte_present;
+               paravirt_ops.set_pud = vmi_set_pud;
+               paravirt_ops.pte_clear = vmi_pte_clear;
+               paravirt_ops.pmd_clear = vmi_pmd_clear;
+#endif
+       }
+
+       if (vmi_ops.update_pte) {
+               paravirt_ops.pte_update = vmi_update_pte;
+               paravirt_ops.pte_update_defer = vmi_update_pte_defer;
+       }
+
+       vmi_ops.allocate_page = vmi_get_function(VMI_CALL_AllocatePage);
+       if (vmi_ops.allocate_page) {
+               paravirt_ops.alloc_pt = vmi_allocate_pt;
+               paravirt_ops.alloc_pd = vmi_allocate_pd;
+               paravirt_ops.alloc_pd_clone = vmi_allocate_pd_clone;
+       }
+
+       vmi_ops.release_page = vmi_get_function(VMI_CALL_ReleasePage);
+       if (vmi_ops.release_page) {
+               paravirt_ops.release_pt = vmi_release_pt;
+               paravirt_ops.release_pd = vmi_release_pd;
+       }
+
+       /* Set linear is needed in all cases */
+       vmi_ops.set_linear_mapping = vmi_get_function(VMI_CALL_SetLinearMapping);
+#ifdef CONFIG_HIGHPTE
+       if (vmi_ops.set_linear_mapping)
+               paravirt_ops.kmap_atomic_pte = vmi_kmap_atomic_pte;
+#endif
+
+       /*
+        * These MUST always be patched.  Don't support indirect jumps
+        * through these operations, as the VMI interface may use either
+        * a jump or a call to get to these operations, depending on
+        * the backend.  They are performance critical anyway, so requiring
+        * a patch is not a big problem.
+        */
+       paravirt_ops.irq_enable_sysexit = (void *)0xfeedbab0;
+       paravirt_ops.iret = (void *)0xbadbab0;
+
+#ifdef CONFIG_SMP
+       para_wrap(startup_ipi_hook, vmi_startup_ipi_hook, set_initial_ap_state, SetInitialAPState);
+#endif
+
+#ifdef CONFIG_X86_LOCAL_APIC
+       para_fill(apic_read, APICRead);
+       para_fill(apic_write, APICWrite);
+       para_fill(apic_write_atomic, APICWrite);
+#endif
+
+       /*
+        * Check for VMI timer functionality by probing for a cycle frequency method
+        */
+       reloc = call_vrom_long_func(vmi_rom, get_reloc, VMI_CALL_GetCycleFrequency);
+       if (!disable_vmi_timer && rel->type != VMI_RELOCATION_NONE) {
+               vmi_timer_ops.get_cycle_frequency = (void *)rel->eip;
+               vmi_timer_ops.get_cycle_counter =
+                       vmi_get_function(VMI_CALL_GetCycleCounter);
+               vmi_timer_ops.get_wallclock =
+                       vmi_get_function(VMI_CALL_GetWallclockTime);
+               vmi_timer_ops.wallclock_updated =
+                       vmi_get_function(VMI_CALL_WallclockUpdated);
+               vmi_timer_ops.set_alarm = vmi_get_function(VMI_CALL_SetAlarm);
+               vmi_timer_ops.cancel_alarm =
+                        vmi_get_function(VMI_CALL_CancelAlarm);
+               paravirt_ops.time_init = vmi_time_init;
+               paravirt_ops.get_wallclock = vmi_get_wallclock;
+               paravirt_ops.set_wallclock = vmi_set_wallclock;
+#ifdef CONFIG_X86_LOCAL_APIC
+               paravirt_ops.setup_boot_clock = vmi_time_bsp_init;
+               paravirt_ops.setup_secondary_clock = vmi_time_ap_init;
+#endif
+               paravirt_ops.sched_clock = vmi_sched_clock;
+               paravirt_ops.get_cpu_khz = vmi_cpu_khz;
+
+               /* We have true wallclock functions; disable CMOS clock sync */
+               no_sync_cmos_clock = 1;
+       } else {
+               disable_noidle = 1;
+               disable_vmi_timer = 1;
+       }
+
+       para_fill(safe_halt, Halt);
+
+       /*
+        * Alternative instruction rewriting doesn't happen soon enough
+        * to convert VMI_IRET to a call instead of a jump; so we have
+        * to do this before IRQs get reenabled.  Fortunately, it is
+        * idempotent.
+        */
+       apply_paravirt(__parainstructions, __parainstructions_end);
+
+       vmi_bringup();
+
+       return 1;
+}
+
+#undef para_fill
+
+void __init vmi_init(void)
+{
+       unsigned long flags;
+
+       if (!vmi_rom)
+               probe_vmi_rom();
+       else
+               check_vmi_rom(vmi_rom);
+
+       /* In case probing for or validating the ROM failed, basil */
+       if (!vmi_rom)
+               return;
+
+       reserve_top_address(-vmi_rom->virtual_top);
+
+       local_irq_save(flags);
+       activate_vmi();
+
+#ifdef CONFIG_X86_IO_APIC
+       /* This is virtual hardware; timer routing is wired correctly */
+       no_timer_check = 1;
+#endif
+       local_irq_restore(flags & X86_EFLAGS_IF);
+}
+
+static int __init parse_vmi(char *arg)
+{
+       if (!arg)
+               return -EINVAL;
+
+       if (!strcmp(arg, "disable_pge")) {
+               clear_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability);
+               disable_pge = 1;
+       } else if (!strcmp(arg, "disable_pse")) {
+               clear_bit(X86_FEATURE_PSE, boot_cpu_data.x86_capability);
+               disable_pse = 1;
+       } else if (!strcmp(arg, "disable_sep")) {
+               clear_bit(X86_FEATURE_SEP, boot_cpu_data.x86_capability);
+               disable_sep = 1;
+       } else if (!strcmp(arg, "disable_tsc")) {
+               clear_bit(X86_FEATURE_TSC, boot_cpu_data.x86_capability);
+               disable_tsc = 1;
+       } else if (!strcmp(arg, "disable_mtrr")) {
+               clear_bit(X86_FEATURE_MTRR, boot_cpu_data.x86_capability);
+               disable_mtrr = 1;
+       } else if (!strcmp(arg, "disable_timer")) {
+               disable_vmi_timer = 1;
+               disable_noidle = 1;
+       } else if (!strcmp(arg, "disable_noidle"))
+               disable_noidle = 1;
+       return 0;
+}
+
+early_param("vmi", parse_vmi);