err.no Git - linux-2.6/blob - arch/x86/kernel/paravirt_32.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/highmem.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 void _paravirt_nop(void)
  39 {
  40 }
  41
  42 static void __init default_banner(void)
  43 {
  44         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
  45                pv_info.name);
  46 }
  47
  48 char *memory_setup(void)
  49 {
  50         return pv_init_ops.memory_setup();
  51 }
  52
  53 /* Simple instruction patching code. */
  54 #define DEF_NATIVE(ops, name, code)                                     \
  55         extern const char start_##ops##_##name[], end_##ops##_##name[]; \
  56         asm("start_" #ops "_" #name ": " code "; end_" #ops "_" #name ":")
  57
  58 DEF_NATIVE(pv_irq_ops, irq_disable, "cli");
  59 DEF_NATIVE(pv_irq_ops, irq_enable, "sti");
  60 DEF_NATIVE(pv_irq_ops, restore_fl, "push %eax; popf");
  61 DEF_NATIVE(pv_irq_ops, save_fl, "pushf; pop %eax");
  62 DEF_NATIVE(pv_cpu_ops, iret, "iret");
  63 DEF_NATIVE(pv_cpu_ops, irq_enable_sysexit, "sti; sysexit");
  64 DEF_NATIVE(pv_mmu_ops, read_cr2, "mov %cr2, %eax");
  65 DEF_NATIVE(pv_mmu_ops, write_cr3, "mov %eax, %cr3");
  66 DEF_NATIVE(pv_mmu_ops, read_cr3, "mov %cr3, %eax");
  67 DEF_NATIVE(pv_cpu_ops, clts, "clts");
  68 DEF_NATIVE(pv_cpu_ops, read_tsc, "rdtsc");
  69
  70 /* Undefined instruction for dealing with missing ops pointers. */
  71 static const unsigned char ud2a[] = { 0x0f, 0x0b };
  72
  73 static unsigned native_patch(u8 type, u16 clobbers, void *ibuf,
  74                              unsigned long addr, unsigned len)
  75 {
  76         const unsigned char *start, *end;
  77         unsigned ret;
  78
  79         switch(type) {
  80 #define SITE(ops, x)                                            \
  81         case PARAVIRT_PATCH(ops.x):                             \
  82                 start = start_##ops##_##x;                      \
  83                 end = end_##ops##_##x;                          \
  84                 goto patch_site
  85
  86         SITE(pv_irq_ops, irq_disable);
  87         SITE(pv_irq_ops, irq_enable);
  88         SITE(pv_irq_ops, restore_fl);
  89         SITE(pv_irq_ops, save_fl);
  90         SITE(pv_cpu_ops, iret);
  91         SITE(pv_cpu_ops, irq_enable_sysexit);
  92         SITE(pv_mmu_ops, read_cr2);
  93         SITE(pv_mmu_ops, read_cr3);
  94         SITE(pv_mmu_ops, write_cr3);
  95         SITE(pv_cpu_ops, clts);
  96         SITE(pv_cpu_ops, read_tsc);
  97 #undef SITE
  98
  99         patch_site:
 100                 ret = paravirt_patch_insns(ibuf, len, start, end);
 101                 break;
 102
 103         default:
 104                 ret = paravirt_patch_default(type, clobbers, ibuf, addr, len);
 105                 break;
 106         }
 107
 108         return ret;
 109 }
 110
 111 unsigned paravirt_patch_nop(void)
 112 {
 113         return 0;
 114 }
 115
 116 unsigned paravirt_patch_ignore(unsigned len)
 117 {
 118         return len;
 119 }
 120
 121 struct branch {
 122         unsigned char opcode;
 123         u32 delta;
 124 } __attribute__((packed));
 125
 126 unsigned paravirt_patch_call(void *insnbuf,
 127                              const void *target, u16 tgt_clobbers,
 128                              unsigned long addr, u16 site_clobbers,
 129                              unsigned len)
 130 {
 131         struct branch *b = insnbuf;
 132         unsigned long delta = (unsigned long)target - (addr+5);
 133
 134         if (tgt_clobbers & ~site_clobbers)
 135                 return len;     /* target would clobber too much for this site */
 136         if (len < 5)
 137                 return len;     /* call too long for patch site */
 138
 139         b->opcode = 0xe8; /* call */
 140         b->delta = delta;
 141         BUILD_BUG_ON(sizeof(*b) != 5);
 142
 143         return 5;
 144 }
 145
 146 unsigned paravirt_patch_jmp(void *insnbuf, const void *target,
 147                             unsigned long addr, unsigned len)
 148 {
 149         struct branch *b = insnbuf;
 150         unsigned long delta = (unsigned long)target - (addr+5);
 151
 152         if (len < 5)
 153                 return len;     /* call too long for patch site */
 154
 155         b->opcode = 0xe9;       /* jmp */
 156         b->delta = delta;
 157
 158         return 5;
 159 }
 160
 161 /* Neat trick to map patch type back to the call within the
 162  * corresponding structure. */
 163 static void *get_call_destination(u8 type)
 164 {
 165         struct paravirt_patch_template tmpl = {
 166                 .pv_init_ops = pv_init_ops,
 167                 .pv_misc_ops = pv_misc_ops,
 168                 .pv_time_ops = pv_time_ops,
 169                 .pv_cpu_ops = pv_cpu_ops,
 170                 .pv_irq_ops = pv_irq_ops,
 171                 .pv_apic_ops = pv_apic_ops,
 172                 .pv_mmu_ops = pv_mmu_ops,
 173         };
 174         return *((void **)&tmpl + type);
 175 }
 176
 177 unsigned paravirt_patch_default(u8 type, u16 clobbers, void *insnbuf,
 178                                 unsigned long addr, unsigned len)
 179 {
 180         void *opfunc = get_call_destination(type);
 181         unsigned ret;
 182
 183         if (opfunc == NULL)
 184                 /* If there's no function, patch it with a ud2a (BUG) */
 185                 ret = paravirt_patch_insns(insnbuf, len, ud2a, ud2a+sizeof(ud2a));
 186         else if (opfunc == paravirt_nop)
 187                 /* If the operation is a nop, then nop the callsite */
 188                 ret = paravirt_patch_nop();
 189         else if (type == PARAVIRT_PATCH(pv_cpu_ops.iret) ||
 190                  type == PARAVIRT_PATCH(pv_cpu_ops.irq_enable_sysexit))
 191                 /* If operation requires a jmp, then jmp */
 192                 ret = paravirt_patch_jmp(insnbuf, opfunc, addr, len);
 193         else
 194                 /* Otherwise call the function; assume target could
 195                    clobber any caller-save reg */
 196                 ret = paravirt_patch_call(insnbuf, opfunc, CLBR_ANY,
 197                                           addr, clobbers, len);
 198
 199         return ret;
 200 }
 201
 202 unsigned paravirt_patch_insns(void *insnbuf, unsigned len,
 203                               const char *start, const char *end)
 204 {
 205         unsigned insn_len = end - start;
 206
 207         if (insn_len > len || start == NULL)
 208                 insn_len = len;
 209         else
 210                 memcpy(insnbuf, start, insn_len);
 211
 212         return insn_len;
 213 }
 214
 215 void init_IRQ(void)
 216 {
 217         pv_irq_ops.init_IRQ();
 218 }
 219
 220 static void native_flush_tlb(void)
 221 {
 222         __native_flush_tlb();
 223 }
 224
 225 /*
 226  * Global pages have to be flushed a bit differently. Not a real
 227  * performance problem because this does not happen often.
 228  */
 229 static void native_flush_tlb_global(void)
 230 {
 231         __native_flush_tlb_global();
 232 }
 233
 234 static void native_flush_tlb_single(unsigned long addr)
 235 {
 236         __native_flush_tlb_single(addr);
 237 }
 238
 239 /* These are in entry.S */
 240 extern void native_iret(void);
 241 extern void native_irq_enable_sysexit(void);
 242
 243 static int __init print_banner(void)
 244 {
 245         pv_init_ops.banner();
 246         return 0;
 247 }
 248 core_initcall(print_banner);
 249
 250 static struct resource reserve_ioports = {
 251         .start = 0,
 252         .end = IO_SPACE_LIMIT,
 253         .name = "paravirt-ioport",
 254         .flags = IORESOURCE_IO | IORESOURCE_BUSY,
 255 };
 256
 257 static struct resource reserve_iomem = {
 258         .start = 0,
 259         .end = -1,
 260         .name = "paravirt-iomem",
 261         .flags = IORESOURCE_MEM | IORESOURCE_BUSY,
 262 };
 263
 264 /*
 265  * Reserve the whole legacy IO space to prevent any legacy drivers
 266  * from wasting time probing for their hardware.  This is a fairly
 267  * brute-force approach to disabling all non-virtual drivers.
 268  *
 269  * Note that this must be called very early to have any effect.
 270  */
 271 int paravirt_disable_iospace(void)
 272 {
 273         int ret;
 274
 275         ret = request_resource(&ioport_resource, &reserve_ioports);
 276         if (ret == 0) {
 277                 ret = request_resource(&iomem_resource, &reserve_iomem);
 278                 if (ret)
 279                         release_resource(&reserve_ioports);
 280         }
 281
 282         return ret;
 283 }
 284
 285 struct pv_info pv_info = {
 286         .name = "bare hardware",
 287         .paravirt_enabled = 0,
 288         .kernel_rpl = 0,
 289         .shared_kernel_pmd = 1, /* Only used when CONFIG_X86_PAE is set */
 290 };
 291
 292 struct pv_init_ops pv_init_ops = {
 293         .patch = native_patch,
 294         .banner = default_banner,
 295         .arch_setup = paravirt_nop,
 296         .memory_setup = machine_specific_memory_setup,
 297 };
 298
 299 struct pv_time_ops pv_time_ops = {
 300         .time_init = hpet_time_init,
 301         .get_wallclock = native_get_wallclock,
 302         .set_wallclock = native_set_wallclock,
 303         .sched_clock = native_sched_clock,
 304         .get_cpu_khz = native_calculate_cpu_khz,
 305 };
 306
 307 struct pv_irq_ops pv_irq_ops = {
 308         .init_IRQ = native_init_IRQ,
 309         .save_fl = native_save_fl,
 310         .restore_fl = native_restore_fl,
 311         .irq_disable = native_irq_disable,
 312         .irq_enable = native_irq_enable,
 313         .safe_halt = native_safe_halt,
 314         .halt = native_halt,
 315 };
 316
 317 struct pv_cpu_ops pv_cpu_ops = {
 318         .cpuid = native_cpuid,
 319         .get_debugreg = native_get_debugreg,
 320         .set_debugreg = native_set_debugreg,
 321         .clts = native_clts,
 322         .read_cr0 = native_read_cr0,
 323         .write_cr0 = native_write_cr0,
 324         .read_cr4 = native_read_cr4,
 325         .read_cr4_safe = native_read_cr4_safe,
 326         .write_cr4 = native_write_cr4,
 327         .wbinvd = native_wbinvd,
 328         .read_msr = native_read_msr_safe,
 329         .write_msr = native_write_msr_safe,
 330         .read_tsc = native_read_tsc,
 331         .read_pmc = native_read_pmc,
 332         .load_tr_desc = native_load_tr_desc,
 333         .set_ldt = native_set_ldt,
 334         .load_gdt = native_load_gdt,
 335         .load_idt = native_load_idt,
 336         .store_gdt = native_store_gdt,
 337         .store_idt = native_store_idt,
 338         .store_tr = native_store_tr,
 339         .load_tls = native_load_tls,
 340         .write_ldt_entry = write_dt_entry,
 341         .write_gdt_entry = write_dt_entry,
 342         .write_idt_entry = write_dt_entry,
 343         .load_esp0 = native_load_esp0,
 344
 345         .irq_enable_sysexit = native_irq_enable_sysexit,
 346         .iret = native_iret,
 347
 348         .set_iopl_mask = native_set_iopl_mask,
 349         .io_delay = native_io_delay,
 350 };
 351
 352 struct pv_apic_ops pv_apic_ops = {
 353 #ifdef CONFIG_X86_LOCAL_APIC
 354         .apic_write = native_apic_write,
 355         .apic_write_atomic = native_apic_write_atomic,
 356         .apic_read = native_apic_read,
 357         .setup_boot_clock = setup_boot_APIC_clock,
 358         .setup_secondary_clock = setup_secondary_APIC_clock,
 359         .startup_ipi_hook = paravirt_nop,
 360 #endif
 361 };
 362
 363 struct pv_misc_ops pv_misc_ops = {
 364         .set_lazy_mode = paravirt_nop,
 365 };
 366
 367 struct pv_mmu_ops pv_mmu_ops = {
 368         .pagetable_setup_start = native_pagetable_setup_start,
 369         .pagetable_setup_done = native_pagetable_setup_done,
 370
 371         .read_cr2 = native_read_cr2,
 372         .write_cr2 = native_write_cr2,
 373         .read_cr3 = native_read_cr3,
 374         .write_cr3 = native_write_cr3,
 375
 376         .flush_tlb_user = native_flush_tlb,
 377         .flush_tlb_kernel = native_flush_tlb_global,
 378         .flush_tlb_single = native_flush_tlb_single,
 379         .flush_tlb_others = native_flush_tlb_others,
 380
 381         .alloc_pt = paravirt_nop,
 382         .alloc_pd = paravirt_nop,
 383         .alloc_pd_clone = paravirt_nop,
 384         .release_pt = paravirt_nop,
 385         .release_pd = paravirt_nop,
 386
 387         .set_pte = native_set_pte,
 388         .set_pte_at = native_set_pte_at,
 389         .set_pmd = native_set_pmd,
 390         .pte_update = paravirt_nop,
 391         .pte_update_defer = paravirt_nop,
 392
 393 #ifdef CONFIG_HIGHPTE
 394         .kmap_atomic_pte = kmap_atomic,
 395 #endif
 396
 397 #ifdef CONFIG_X86_PAE
 398         .set_pte_atomic = native_set_pte_atomic,
 399         .set_pte_present = native_set_pte_present,
 400         .set_pud = native_set_pud,
 401         .pte_clear = native_pte_clear,
 402         .pmd_clear = native_pmd_clear,
 403
 404         .pmd_val = native_pmd_val,
 405         .make_pmd = native_make_pmd,
 406 #endif
 407
 408         .pte_val = native_pte_val,
 409         .pgd_val = native_pgd_val,
 410
 411         .make_pte = native_make_pte,
 412         .make_pgd = native_make_pgd,
 413
 414         .dup_mmap = paravirt_nop,
 415         .exit_mmap = paravirt_nop,
 416         .activate_mm = paravirt_nop,
 417 };
 418
 419 EXPORT_SYMBOL_GPL(pv_time_ops);
 420 EXPORT_SYMBOL_GPL(pv_cpu_ops);
 421 EXPORT_SYMBOL_GPL(pv_mmu_ops);
 422 EXPORT_SYMBOL_GPL(pv_apic_ops);
 423 EXPORT_SYMBOL_GPL(pv_info);
 424 EXPORT_SYMBOL    (pv_irq_ops);