arch/i386/kernel/paravirt.c

   1 /*  Paravirtualization interfaces
   2     Copyright (C) 2006 Rusty Russell IBM Corporation
   3
   4     This program is free software; you can redistribute it and/or modify
   5     it under the terms of the GNU General Public License as published by
   6     the Free Software Foundation; either version 2 of the License, or
   7     (at your option) any later version.
   8
   9     This program is distributed in the hope that it will be useful,
  10     but WITHOUT ANY WARRANTY; without even the implied warranty of
  11     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12     GNU General Public License for more details.
  13
  14     You should have received a copy of the GNU General Public License
  15     along with this program; if not, write to the Free Software
  16     Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17 */
  18 #include <linux/errno.h>
  19 #include <linux/module.h>
  20 #include <linux/efi.h>
  21 #include <linux/bcd.h>
  22 #include <linux/start_kernel.h>
  23
  24 #include <asm/bug.h>
  25 #include <asm/paravirt.h>
  26 #include <asm/desc.h>
  27 #include <asm/setup.h>
  28 #include <asm/arch_hooks.h>
  29 #include <asm/time.h>
  30 #include <asm/irq.h>
  31 #include <asm/delay.h>
  32 #include <asm/fixmap.h>
  33 #include <asm/apic.h>
  34 #include <asm/tlbflush.h>
  35 #include <asm/timer.h>
  36
  37 /* nop stub */
  38 static void native_nop(void)
  39 {
  40 }
  41
  42 static void __init default_banner(void)
  43 {
  44         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
  45                paravirt_ops.name);
  46 }
  47
  48 char *memory_setup(void)
  49 {
  50         return paravirt_ops.memory_setup();
  51 }
  52
  53 /* Simple instruction patching code. */
  54 #define DEF_NATIVE(name, code)                                  \
  55         extern const char start_##name[], end_##name[];         \
  56         asm("start_" #name ": " code "; end_" #name ":")
  57 DEF_NATIVE(cli, "cli");
  58 DEF_NATIVE(sti, "sti");
  59 DEF_NATIVE(popf, "push %eax; popf");
  60 DEF_NATIVE(pushf, "pushf; pop %eax");
  61 DEF_NATIVE(pushf_cli, "pushf; pop %eax; cli");
  62 DEF_NATIVE(iret, "iret");
  63 DEF_NATIVE(sti_sysexit, "sti; sysexit");
  64
  65 static const struct native_insns
  66 {
  67         const char *start, *end;
  68 } native_insns[] = {
  69         [PARAVIRT_IRQ_DISABLE] = { start_cli, end_cli },
  70         [PARAVIRT_IRQ_ENABLE] = { start_sti, end_sti },
  71         [PARAVIRT_RESTORE_FLAGS] = { start_popf, end_popf },
  72         [PARAVIRT_SAVE_FLAGS] = { start_pushf, end_pushf },
  73         [PARAVIRT_SAVE_FLAGS_IRQ_DISABLE] = { start_pushf_cli, end_pushf_cli },
  74         [PARAVIRT_INTERRUPT_RETURN] = { start_iret, end_iret },
  75         [PARAVIRT_STI_SYSEXIT] = { start_sti_sysexit, end_sti_sysexit },
  76 };
  77
  78 static unsigned native_patch(u8 type, u16 clobbers, void *insns, unsigned len)
  79 {
  80         unsigned int insn_len;
  81
  82         /* Don't touch it if we don't have a replacement */
  83         if (type >= ARRAY_SIZE(native_insns) || !native_insns[type].start)
  84                 return len;
  85
  86         insn_len = native_insns[type].end - native_insns[type].start;
  87
  88         /* Similarly if we can't fit replacement. */
  89         if (len < insn_len)
  90                 return len;
  91
  92         memcpy(insns, native_insns[type].start, insn_len);
  93         return insn_len;
  94 }
  95
  96 void init_IRQ(void)
  97 {
  98         paravirt_ops.init_IRQ();
  99 }
 100
 101 static void native_flush_tlb(void)
 102 {
 103         __native_flush_tlb();
 104 }
 105
 106 /*
 107  * Global pages have to be flushed a bit differently. Not a real
 108  * performance problem because this does not happen often.
 109  */
 110 static void native_flush_tlb_global(void)
 111 {
 112         __native_flush_tlb_global();
 113 }
 114
 115 static void native_flush_tlb_single(u32 addr)
 116 {
 117         __native_flush_tlb_single(addr);
 118 }
 119
 120 #ifndef CONFIG_X86_PAE
 121 static void native_set_pte(pte_t *ptep, pte_t pteval)
 122 {
 123         *ptep = pteval;
 124 }
 125
 126 static void native_set_pte_at(struct mm_struct *mm, u32 addr, pte_t *ptep, pte_t pteval)
 127 {
 128         *ptep = pteval;
 129 }
 130
 131 static void native_set_pmd(pmd_t *pmdp, pmd_t pmdval)
 132 {
 133         *pmdp = pmdval;
 134 }
 135
 136 #else /* CONFIG_X86_PAE */
 137
 138 static void native_set_pte(pte_t *ptep, pte_t pte)
 139 {
 140         ptep->pte_high = pte.pte_high;
 141         smp_wmb();
 142         ptep->pte_low = pte.pte_low;
 143 }
 144
 145 static void native_set_pte_at(struct mm_struct *mm, u32 addr, pte_t *ptep, pte_t pte)
 146 {
 147         ptep->pte_high = pte.pte_high;
 148         smp_wmb();
 149         ptep->pte_low = pte.pte_low;
 150 }
 151
 152 static void native_set_pte_present(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
 153 {
 154         ptep->pte_low = 0;
 155         smp_wmb();
 156         ptep->pte_high = pte.pte_high;
 157         smp_wmb();
 158         ptep->pte_low = pte.pte_low;
 159 }
 160
 161 static void native_set_pte_atomic(pte_t *ptep, pte_t pteval)
 162 {
 163         set_64bit((unsigned long long *)ptep,pte_val(pteval));
 164 }
 165
 166 static void native_set_pmd(pmd_t *pmdp, pmd_t pmdval)
 167 {
 168         set_64bit((unsigned long long *)pmdp,pmd_val(pmdval));
 169 }
 170
 171 static void native_set_pud(pud_t *pudp, pud_t pudval)
 172 {
 173         *pudp = pudval;
 174 }
 175
 176 static void native_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
 177 {
 178         ptep->pte_low = 0;
 179         smp_wmb();
 180         ptep->pte_high = 0;
 181 }
 182
 183 static void native_pmd_clear(pmd_t *pmd)
 184 {
 185         u32 *tmp = (u32 *)pmd;
 186         *tmp = 0;
 187         smp_wmb();
 188         *(tmp + 1) = 0;
 189 }
 190 #endif /* CONFIG_X86_PAE */
 191
 192 /* These are in entry.S */
 193 extern void native_iret(void);
 194 extern void native_irq_enable_sysexit(void);
 195
 196 static int __init print_banner(void)
 197 {
 198         paravirt_ops.banner();
 199         return 0;
 200 }
 201 core_initcall(print_banner);
 202
 203 struct paravirt_ops paravirt_ops = {
 204         .name = "bare hardware",
 205         .paravirt_enabled = 0,
 206         .kernel_rpl = 0,
 207
 208         .patch = native_patch,
 209         .banner = default_banner,
 210         .arch_setup = native_nop,
 211         .memory_setup = machine_specific_memory_setup,
 212         .get_wallclock = native_get_wallclock,
 213         .set_wallclock = native_set_wallclock,
 214         .time_init = hpet_time_init,
 215         .init_IRQ = native_init_IRQ,
 216
 217         .cpuid = native_cpuid,
 218         .get_debugreg = native_get_debugreg,
 219         .set_debugreg = native_set_debugreg,
 220         .clts = native_clts,
 221         .read_cr0 = native_read_cr0,
 222         .write_cr0 = native_write_cr0,
 223         .read_cr2 = native_read_cr2,
 224         .write_cr2 = native_write_cr2,
 225         .read_cr3 = native_read_cr3,
 226         .write_cr3 = native_write_cr3,
 227         .read_cr4 = native_read_cr4,
 228         .read_cr4_safe = native_read_cr4_safe,
 229         .write_cr4 = native_write_cr4,
 230         .save_fl = native_save_fl,
 231         .restore_fl = native_restore_fl,
 232         .irq_disable = native_irq_disable,
 233         .irq_enable = native_irq_enable,
 234         .safe_halt = native_safe_halt,
 235         .halt = native_halt,
 236         .wbinvd = native_wbinvd,
 237         .read_msr = native_read_msr_safe,
 238         .write_msr = native_write_msr_safe,
 239         .read_tsc = native_read_tsc,
 240         .read_pmc = native_read_pmc,
 241         .get_scheduled_cycles = native_read_tsc,
 242         .get_cpu_khz = native_calculate_cpu_khz,
 243         .load_tr_desc = native_load_tr_desc,
 244         .set_ldt = native_set_ldt,
 245         .load_gdt = native_load_gdt,
 246         .load_idt = native_load_idt,
 247         .store_gdt = native_store_gdt,
 248         .store_idt = native_store_idt,
 249         .store_tr = native_store_tr,
 250         .load_tls = native_load_tls,
 251         .write_ldt_entry = write_dt_entry,
 252         .write_gdt_entry = write_dt_entry,
 253         .write_idt_entry = write_dt_entry,
 254         .load_esp0 = native_load_esp0,
 255
 256         .set_iopl_mask = native_set_iopl_mask,
 257         .io_delay = native_io_delay,
 258
 259 #ifdef CONFIG_X86_LOCAL_APIC
 260         .apic_write = native_apic_write,
 261         .apic_write_atomic = native_apic_write_atomic,
 262         .apic_read = native_apic_read,
 263         .setup_boot_clock = setup_boot_APIC_clock,
 264         .setup_secondary_clock = setup_secondary_APIC_clock,
 265 #endif
 266         .set_lazy_mode = (void *)native_nop,
 267
 268         .flush_tlb_user = native_flush_tlb,
 269         .flush_tlb_kernel = native_flush_tlb_global,
 270         .flush_tlb_single = native_flush_tlb_single,
 271
 272         .map_pt_hook = (void *)native_nop,
 273
 274         .alloc_pt = (void *)native_nop,
 275         .alloc_pd = (void *)native_nop,
 276         .alloc_pd_clone = (void *)native_nop,
 277         .release_pt = (void *)native_nop,
 278         .release_pd = (void *)native_nop,
 279
 280         .set_pte = native_set_pte,
 281         .set_pte_at = native_set_pte_at,
 282         .set_pmd = native_set_pmd,
 283         .pte_update = (void *)native_nop,
 284         .pte_update_defer = (void *)native_nop,
 285 #ifdef CONFIG_X86_PAE
 286         .set_pte_atomic = native_set_pte_atomic,
 287         .set_pte_present = native_set_pte_present,
 288         .set_pud = native_set_pud,
 289         .pte_clear = native_pte_clear,
 290         .pmd_clear = native_pmd_clear,
 291 #endif
 292
 293         .irq_enable_sysexit = native_irq_enable_sysexit,
 294         .iret = native_iret,
 295
 296         .startup_ipi_hook = (void *)native_nop,
 297 };
 298
 299 /*
 300  * NOTE: CONFIG_PARAVIRT is experimental and the paravirt_ops
 301  * semantics are subject to change. Hence we only do this
 302  * internal-only export of this, until it gets sorted out and
 303  * all lowlevel CPU ops used by modules are separately exported.
 304  */
 305 EXPORT_SYMBOL_GPL(paravirt_ops);