arch/x86/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     2007 - x86_64 support added by Glauber de Oliveira Costa, Red Hat Inc
  19 */
  20
  21 #include <linux/errno.h>
  22 #include <linux/module.h>
  23 #include <linux/efi.h>
  24 #include <linux/bcd.h>
  25 #include <linux/highmem.h>
  26
  27 #include <asm/bug.h>
  28 #include <asm/paravirt.h>
  29 #include <asm/desc.h>
  30 #include <asm/setup.h>
  31 #include <asm/arch_hooks.h>
  32 #include <asm/pgtable.h>
  33 #include <asm/time.h>
  34 #include <asm/pgalloc.h>
  35 #include <asm/irq.h>
  36 #include <asm/delay.h>
  37 #include <asm/fixmap.h>
  38 #include <asm/apic.h>
  39 #include <asm/tlbflush.h>
  40 #include <asm/timer.h>
  41
  42 /* nop stub */
  43 void _paravirt_nop(void)
  44 {
  45 }
  46
  47 /* identity function, which can be inlined */
  48 u32 _paravirt_ident_32(u32 x)
  49 {
  50         return x;
  51 }
  52
  53 u64 _paravirt_ident_64(u64 x)
  54 {
  55         return x;
  56 }
  57
  58 static void __init default_banner(void)
  59 {
  60         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
  61                pv_info.name);
  62 }
  63
  64 char *memory_setup(void)
  65 {
  66         return pv_init_ops.memory_setup();
  67 }
  68
  69 /* Simple instruction patching code. */
  70 #define DEF_NATIVE(ops, name, code)                                     \
  71         extern const char start_##ops##_##name[], end_##ops##_##name[]; \
  72         asm("start_" #ops "_" #name ": " code "; end_" #ops "_" #name ":")
  73
  74 /* Undefined instruction for dealing with missing ops pointers. */
  75 static const unsigned char ud2a[] = { 0x0f, 0x0b };
  76
  77 unsigned paravirt_patch_nop(void)
  78 {
  79         return 0;
  80 }
  81
  82 unsigned paravirt_patch_ignore(unsigned len)
  83 {
  84         return len;
  85 }
  86
  87 struct branch {
  88         unsigned char opcode;
  89         u32 delta;
  90 } __attribute__((packed));
  91
  92 unsigned paravirt_patch_call(void *insnbuf,
  93                              const void *target, u16 tgt_clobbers,
  94                              unsigned long addr, u16 site_clobbers,
  95                              unsigned len)
  96 {
  97         struct branch *b = insnbuf;
  98         unsigned long delta = (unsigned long)target - (addr+5);
  99
 100         if (tgt_clobbers & ~site_clobbers)
 101                 return len;     /* target would clobber too much for this site */
 102         if (len < 5)
 103                 return len;     /* call too long for patch site */
 104
 105         b->opcode = 0xe8; /* call */
 106         b->delta = delta;
 107         BUILD_BUG_ON(sizeof(*b) != 5);
 108
 109         return 5;
 110 }
 111
 112 unsigned paravirt_patch_jmp(void *insnbuf, const void *target,
 113                             unsigned long addr, unsigned len)
 114 {
 115         struct branch *b = insnbuf;
 116         unsigned long delta = (unsigned long)target - (addr+5);
 117
 118         if (len < 5)
 119                 return len;     /* call too long for patch site */
 120
 121         b->opcode = 0xe9;       /* jmp */
 122         b->delta = delta;
 123
 124         return 5;
 125 }
 126
 127 /* Neat trick to map patch type back to the call within the
 128  * corresponding structure. */
 129 static void *get_call_destination(u8 type)
 130 {
 131         struct paravirt_patch_template tmpl = {
 132                 .pv_init_ops = pv_init_ops,
 133                 .pv_time_ops = pv_time_ops,
 134                 .pv_cpu_ops = pv_cpu_ops,
 135                 .pv_irq_ops = pv_irq_ops,
 136                 .pv_apic_ops = pv_apic_ops,
 137                 .pv_mmu_ops = pv_mmu_ops,
 138                 .pv_lock_ops = pv_lock_ops,
 139         };
 140         return *((void **)&tmpl + type);
 141 }
 142
 143 unsigned paravirt_patch_default(u8 type, u16 clobbers, void *insnbuf,
 144                                 unsigned long addr, unsigned len)
 145 {
 146         void *opfunc = get_call_destination(type);
 147         unsigned ret;
 148
 149         if (opfunc == NULL)
 150                 /* If there's no function, patch it with a ud2a (BUG) */
 151                 ret = paravirt_patch_insns(insnbuf, len, ud2a, ud2a+sizeof(ud2a));
 152         else if (opfunc == _paravirt_nop)
 153                 /* If the operation is a nop, then nop the callsite */
 154                 ret = paravirt_patch_nop();
 155
 156         /* identity functions just return their single argument */
 157         else if (opfunc == _paravirt_ident_32)
 158                 ret = paravirt_patch_ident_32(insnbuf, len);
 159         else if (opfunc == _paravirt_ident_64)
 160                 ret = paravirt_patch_ident_64(insnbuf, len);
 161
 162         else if (type == PARAVIRT_PATCH(pv_cpu_ops.iret) ||
 163                  type == PARAVIRT_PATCH(pv_cpu_ops.irq_enable_sysexit) ||
 164                  type == PARAVIRT_PATCH(pv_cpu_ops.usergs_sysret32) ||
 165                  type == PARAVIRT_PATCH(pv_cpu_ops.usergs_sysret64))
 166                 /* If operation requires a jmp, then jmp */
 167                 ret = paravirt_patch_jmp(insnbuf, opfunc, addr, len);
 168         else
 169                 /* Otherwise call the function; assume target could
 170                    clobber any caller-save reg */
 171                 ret = paravirt_patch_call(insnbuf, opfunc, CLBR_ANY,
 172                                           addr, clobbers, len);
 173
 174         return ret;
 175 }
 176
 177 unsigned paravirt_patch_insns(void *insnbuf, unsigned len,
 178                               const char *start, const char *end)
 179 {
 180         unsigned insn_len = end - start;
 181
 182         if (insn_len > len || start == NULL)
 183                 insn_len = len;
 184         else
 185                 memcpy(insnbuf, start, insn_len);
 186
 187         return insn_len;
 188 }
 189
 190 void init_IRQ(void)
 191 {
 192         pv_irq_ops.init_IRQ();
 193 }
 194
 195 static void native_flush_tlb(void)
 196 {
 197         __native_flush_tlb();
 198 }
 199
 200 /*
 201  * Global pages have to be flushed a bit differently. Not a real
 202  * performance problem because this does not happen often.
 203  */
 204 static void native_flush_tlb_global(void)
 205 {
 206         __native_flush_tlb_global();
 207 }
 208
 209 static void native_flush_tlb_single(unsigned long addr)
 210 {
 211         __native_flush_tlb_single(addr);
 212 }
 213
 214 /* These are in entry.S */
 215 extern void native_iret(void);
 216 extern void native_irq_enable_sysexit(void);
 217 extern void native_usergs_sysret32(void);
 218 extern void native_usergs_sysret64(void);
 219
 220 static int __init print_banner(void)
 221 {
 222         pv_init_ops.banner();
 223         return 0;
 224 }
 225 core_initcall(print_banner);
 226
 227 static struct resource reserve_ioports = {
 228         .start = 0,
 229         .end = IO_SPACE_LIMIT,
 230         .name = "paravirt-ioport",
 231         .flags = IORESOURCE_IO | IORESOURCE_BUSY,
 232 };
 233
 234 /*
 235  * Reserve the whole legacy IO space to prevent any legacy drivers
 236  * from wasting time probing for their hardware.  This is a fairly
 237  * brute-force approach to disabling all non-virtual drivers.
 238  *
 239  * Note that this must be called very early to have any effect.
 240  */
 241 int paravirt_disable_iospace(void)
 242 {
 243         return request_resource(&ioport_resource, &reserve_ioports);
 244 }
 245
 246 static DEFINE_PER_CPU(enum paravirt_lazy_mode, paravirt_lazy_mode) = PARAVIRT_LAZY_NONE;
 247
 248 static inline void enter_lazy(enum paravirt_lazy_mode mode)
 249 {
 250         BUG_ON(__get_cpu_var(paravirt_lazy_mode) != PARAVIRT_LAZY_NONE);
 251         BUG_ON(preemptible());
 252
 253         __get_cpu_var(paravirt_lazy_mode) = mode;
 254 }
 255
 256 void paravirt_leave_lazy(enum paravirt_lazy_mode mode)
 257 {
 258         BUG_ON(__get_cpu_var(paravirt_lazy_mode) != mode);
 259         BUG_ON(preemptible());
 260
 261         __get_cpu_var(paravirt_lazy_mode) = PARAVIRT_LAZY_NONE;
 262 }
 263
 264 void paravirt_enter_lazy_mmu(void)
 265 {
 266         enter_lazy(PARAVIRT_LAZY_MMU);
 267 }
 268
 269 void paravirt_leave_lazy_mmu(void)
 270 {
 271         paravirt_leave_lazy(PARAVIRT_LAZY_MMU);
 272 }
 273
 274 void paravirt_enter_lazy_cpu(void)
 275 {
 276         enter_lazy(PARAVIRT_LAZY_CPU);
 277 }
 278
 279 void paravirt_leave_lazy_cpu(void)
 280 {
 281         paravirt_leave_lazy(PARAVIRT_LAZY_CPU);
 282 }
 283
 284 enum paravirt_lazy_mode paravirt_get_lazy_mode(void)
 285 {
 286         return __get_cpu_var(paravirt_lazy_mode);
 287 }
 288
 289 void arch_flush_lazy_mmu_mode(void)
 290 {
 291         preempt_disable();
 292
 293         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU) {
 294                 WARN_ON(preempt_count() == 1);
 295                 arch_leave_lazy_mmu_mode();
 296                 arch_enter_lazy_mmu_mode();
 297         }
 298
 299         preempt_enable();
 300 }
 301
 302 void arch_flush_lazy_cpu_mode(void)
 303 {
 304         preempt_disable();
 305
 306         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
 307                 WARN_ON(preempt_count() == 1);
 308                 arch_leave_lazy_cpu_mode();
 309                 arch_enter_lazy_cpu_mode();
 310         }
 311
 312         preempt_enable();
 313 }
 314
 315 struct pv_info pv_info = {
 316         .name = "bare hardware",
 317         .paravirt_enabled = 0,
 318         .kernel_rpl = 0,
 319         .shared_kernel_pmd = 1, /* Only used when CONFIG_X86_PAE is set */
 320 };
 321
 322 struct pv_init_ops pv_init_ops = {
 323         .patch = native_patch,
 324         .banner = default_banner,
 325         .arch_setup = paravirt_nop,
 326         .memory_setup = machine_specific_memory_setup,
 327 };
 328
 329 struct pv_time_ops pv_time_ops = {
 330         .time_init = hpet_time_init,
 331         .get_wallclock = native_get_wallclock,
 332         .set_wallclock = native_set_wallclock,
 333         .sched_clock = native_sched_clock,
 334         .get_tsc_khz = native_calibrate_tsc,
 335 };
 336
 337 struct pv_irq_ops pv_irq_ops = {
 338         .init_IRQ = native_init_IRQ,
 339         .save_fl = __PV_IS_CALLEE_SAVE(native_save_fl),
 340         .restore_fl = __PV_IS_CALLEE_SAVE(native_restore_fl),
 341         .irq_disable = __PV_IS_CALLEE_SAVE(native_irq_disable),
 342         .irq_enable = __PV_IS_CALLEE_SAVE(native_irq_enable),
 343         .safe_halt = native_safe_halt,
 344         .halt = native_halt,
 345 #ifdef CONFIG_X86_64
 346         .adjust_exception_frame = paravirt_nop,
 347 #endif
 348 };
 349
 350 struct pv_cpu_ops pv_cpu_ops = {
 351         .cpuid = native_cpuid,
 352         .get_debugreg = native_get_debugreg,
 353         .set_debugreg = native_set_debugreg,
 354         .clts = native_clts,
 355         .read_cr0 = native_read_cr0,
 356         .write_cr0 = native_write_cr0,
 357         .read_cr4 = native_read_cr4,
 358         .read_cr4_safe = native_read_cr4_safe,
 359         .write_cr4 = native_write_cr4,
 360 #ifdef CONFIG_X86_64
 361         .read_cr8 = native_read_cr8,
 362         .write_cr8 = native_write_cr8,
 363 #endif
 364         .wbinvd = native_wbinvd,
 365         .read_msr = native_read_msr_safe,
 366         .read_msr_amd = native_read_msr_amd_safe,
 367         .write_msr = native_write_msr_safe,
 368         .read_tsc = native_read_tsc,
 369         .read_pmc = native_read_pmc,
 370         .read_tscp = native_read_tscp,
 371         .load_tr_desc = native_load_tr_desc,
 372         .set_ldt = native_set_ldt,
 373         .load_gdt = native_load_gdt,
 374         .load_idt = native_load_idt,
 375         .store_gdt = native_store_gdt,
 376         .store_idt = native_store_idt,
 377         .store_tr = native_store_tr,
 378         .load_tls = native_load_tls,
 379 #ifdef CONFIG_X86_64
 380         .load_gs_index = native_load_gs_index,
 381 #endif
 382         .write_ldt_entry = native_write_ldt_entry,
 383         .write_gdt_entry = native_write_gdt_entry,
 384         .write_idt_entry = native_write_idt_entry,
 385
 386         .alloc_ldt = paravirt_nop,
 387         .free_ldt = paravirt_nop,
 388
 389         .load_sp0 = native_load_sp0,
 390
 391 #if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION)
 392         .irq_enable_sysexit = native_irq_enable_sysexit,
 393 #endif
 394 #ifdef CONFIG_X86_64
 395 #ifdef CONFIG_IA32_EMULATION
 396         .usergs_sysret32 = native_usergs_sysret32,
 397 #endif
 398         .usergs_sysret64 = native_usergs_sysret64,
 399 #endif
 400         .iret = native_iret,
 401         .swapgs = native_swapgs,
 402
 403         .set_iopl_mask = native_set_iopl_mask,
 404         .io_delay = native_io_delay,
 405
 406         .lazy_mode = {
 407                 .enter = paravirt_nop,
 408                 .leave = paravirt_nop,
 409         },
 410 };
 411
 412 struct pv_apic_ops pv_apic_ops = {
 413 #ifdef CONFIG_X86_LOCAL_APIC
 414         .setup_boot_clock = setup_boot_APIC_clock,
 415         .setup_secondary_clock = setup_secondary_APIC_clock,
 416         .startup_ipi_hook = paravirt_nop,
 417 #endif
 418 };
 419
 420 #if defined(CONFIG_X86_32) && !defined(CONFIG_X86_PAE)
 421 /* 32-bit pagetable entries */
 422 #define PTE_IDENT       __PV_IS_CALLEE_SAVE(_paravirt_ident_32)
 423 #else
 424 /* 64-bit pagetable entries */
 425 #define PTE_IDENT       __PV_IS_CALLEE_SAVE(_paravirt_ident_64)
 426 #endif
 427
 428 struct pv_mmu_ops pv_mmu_ops = {
 429 #ifndef CONFIG_X86_64
 430         .pagetable_setup_start = native_pagetable_setup_start,
 431         .pagetable_setup_done = native_pagetable_setup_done,
 432 #else
 433         .pagetable_setup_start = paravirt_nop,
 434         .pagetable_setup_done = paravirt_nop,
 435 #endif
 436
 437         .read_cr2 = native_read_cr2,
 438         .write_cr2 = native_write_cr2,
 439         .read_cr3 = native_read_cr3,
 440         .write_cr3 = native_write_cr3,
 441
 442         .flush_tlb_user = native_flush_tlb,
 443         .flush_tlb_kernel = native_flush_tlb_global,
 444         .flush_tlb_single = native_flush_tlb_single,
 445         .flush_tlb_others = native_flush_tlb_others,
 446
 447         .pgd_alloc = __paravirt_pgd_alloc,
 448         .pgd_free = paravirt_nop,
 449
 450         .alloc_pte = paravirt_nop,
 451         .alloc_pmd = paravirt_nop,
 452         .alloc_pmd_clone = paravirt_nop,
 453         .alloc_pud = paravirt_nop,
 454         .release_pte = paravirt_nop,
 455         .release_pmd = paravirt_nop,
 456         .release_pud = paravirt_nop,
 457
 458         .set_pte = native_set_pte,
 459         .set_pte_at = native_set_pte_at,
 460         .set_pmd = native_set_pmd,
 461         .pte_update = paravirt_nop,
 462         .pte_update_defer = paravirt_nop,
 463
 464         .ptep_modify_prot_start = __ptep_modify_prot_start,
 465         .ptep_modify_prot_commit = __ptep_modify_prot_commit,
 466
 467 #ifdef CONFIG_HIGHPTE
 468         .kmap_atomic_pte = kmap_atomic,
 469 #endif
 470
 471 #if PAGETABLE_LEVELS >= 3
 472 #ifdef CONFIG_X86_PAE
 473         .set_pte_atomic = native_set_pte_atomic,
 474         .set_pte_present = native_set_pte_present,
 475         .pte_clear = native_pte_clear,
 476         .pmd_clear = native_pmd_clear,
 477 #endif
 478         .set_pud = native_set_pud,
 479
 480         .pmd_val = PTE_IDENT,
 481         .make_pmd = PTE_IDENT,
 482
 483 #if PAGETABLE_LEVELS == 4
 484         .pud_val = PTE_IDENT,
 485         .make_pud = PTE_IDENT,
 486
 487         .set_pgd = native_set_pgd,
 488 #endif
 489 #endif /* PAGETABLE_LEVELS >= 3 */
 490
 491         .pte_val = PTE_IDENT,
 492         .pgd_val = PTE_IDENT,
 493
 494         .make_pte = PTE_IDENT,
 495         .make_pgd = PTE_IDENT,
 496
 497         .dup_mmap = paravirt_nop,
 498         .exit_mmap = paravirt_nop,
 499         .activate_mm = paravirt_nop,
 500
 501         .lazy_mode = {
 502                 .enter = paravirt_nop,
 503                 .leave = paravirt_nop,
 504         },
 505
 506         .set_fixmap = native_set_fixmap,
 507 };
 508
 509 EXPORT_SYMBOL_GPL(pv_time_ops);
 510 EXPORT_SYMBOL    (pv_cpu_ops);
 511 EXPORT_SYMBOL    (pv_mmu_ops);
 512 EXPORT_SYMBOL_GPL(pv_apic_ops);
 513 EXPORT_SYMBOL_GPL(pv_info);
 514 EXPORT_SYMBOL    (pv_irq_ops);