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/time.h>
  33 #include <asm/irq.h>
  34 #include <asm/delay.h>
  35 #include <asm/fixmap.h>
  36 #include <asm/apic.h>
  37 #include <asm/tlbflush.h>
  38 #include <asm/timer.h>
  39
  40 /* nop stub */
  41 void _paravirt_nop(void)
  42 {
  43 }
  44
  45 static void __init default_banner(void)
  46 {
  47         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
  48                pv_info.name);
  49 }
  50
  51 char *memory_setup(void)
  52 {
  53         return pv_init_ops.memory_setup();
  54 }
  55
  56 /* Simple instruction patching code. */
  57 #define DEF_NATIVE(ops, name, code)                                     \
  58         extern const char start_##ops##_##name[], end_##ops##_##name[]; \
  59         asm("start_" #ops "_" #name ": " code "; end_" #ops "_" #name ":")
  60
  61 DEF_NATIVE(pv_irq_ops, irq_disable, "cli");
  62 DEF_NATIVE(pv_irq_ops, irq_enable, "sti");
  63 DEF_NATIVE(pv_irq_ops, restore_fl, "push %eax; popf");
  64 DEF_NATIVE(pv_irq_ops, save_fl, "pushf; pop %eax");
  65 DEF_NATIVE(pv_cpu_ops, iret, "iret");
  66 DEF_NATIVE(pv_cpu_ops, irq_enable_syscall_ret, "sti; sysexit");
  67 DEF_NATIVE(pv_mmu_ops, read_cr2, "mov %cr2, %eax");
  68 DEF_NATIVE(pv_mmu_ops, write_cr3, "mov %eax, %cr3");
  69 DEF_NATIVE(pv_mmu_ops, read_cr3, "mov %cr3, %eax");
  70 DEF_NATIVE(pv_cpu_ops, clts, "clts");
  71 DEF_NATIVE(pv_cpu_ops, read_tsc, "rdtsc");
  72
  73 /* Undefined instruction for dealing with missing ops pointers. */
  74 static const unsigned char ud2a[] = { 0x0f, 0x0b };
  75
  76 static unsigned native_patch(u8 type, u16 clobbers, void *ibuf,
  77                              unsigned long addr, unsigned len)
  78 {
  79         const unsigned char *start, *end;
  80         unsigned ret;
  81
  82         switch(type) {
  83 #define SITE(ops, x)                                            \
  84         case PARAVIRT_PATCH(ops.x):                             \
  85                 start = start_##ops##_##x;                      \
  86                 end = end_##ops##_##x;                          \
  87                 goto patch_site
  88
  89         SITE(pv_irq_ops, irq_disable);
  90         SITE(pv_irq_ops, irq_enable);
  91         SITE(pv_irq_ops, restore_fl);
  92         SITE(pv_irq_ops, save_fl);
  93         SITE(pv_cpu_ops, iret);
  94         SITE(pv_cpu_ops, irq_enable_syscall_ret);
  95         SITE(pv_mmu_ops, read_cr2);
  96         SITE(pv_mmu_ops, read_cr3);
  97         SITE(pv_mmu_ops, write_cr3);
  98         SITE(pv_cpu_ops, clts);
  99         SITE(pv_cpu_ops, read_tsc);
 100 #undef SITE
 101
 102         patch_site:
 103                 ret = paravirt_patch_insns(ibuf, len, start, end);
 104                 break;
 105
 106         default:
 107                 ret = paravirt_patch_default(type, clobbers, ibuf, addr, len);
 108                 break;
 109         }
 110
 111         return ret;
 112 }
 113
 114 unsigned paravirt_patch_nop(void)
 115 {
 116         return 0;
 117 }
 118
 119 unsigned paravirt_patch_ignore(unsigned len)
 120 {
 121         return len;
 122 }
 123
 124 struct branch {
 125         unsigned char opcode;
 126         u32 delta;
 127 } __attribute__((packed));
 128
 129 unsigned paravirt_patch_call(void *insnbuf,
 130                              const void *target, u16 tgt_clobbers,
 131                              unsigned long addr, u16 site_clobbers,
 132                              unsigned len)
 133 {
 134         struct branch *b = insnbuf;
 135         unsigned long delta = (unsigned long)target - (addr+5);
 136
 137         if (tgt_clobbers & ~site_clobbers)
 138                 return len;     /* target would clobber too much for this site */
 139         if (len < 5)
 140                 return len;     /* call too long for patch site */
 141
 142         b->opcode = 0xe8; /* call */
 143         b->delta = delta;
 144         BUILD_BUG_ON(sizeof(*b) != 5);
 145
 146         return 5;
 147 }
 148
 149 unsigned paravirt_patch_jmp(void *insnbuf, const void *target,
 150                             unsigned long addr, unsigned len)
 151 {
 152         struct branch *b = insnbuf;
 153         unsigned long delta = (unsigned long)target - (addr+5);
 154
 155         if (len < 5)
 156                 return len;     /* call too long for patch site */
 157
 158         b->opcode = 0xe9;       /* jmp */
 159         b->delta = delta;
 160
 161         return 5;
 162 }
 163
 164 /* Neat trick to map patch type back to the call within the
 165  * corresponding structure. */
 166 static void *get_call_destination(u8 type)
 167 {
 168         struct paravirt_patch_template tmpl = {
 169                 .pv_init_ops = pv_init_ops,
 170                 .pv_time_ops = pv_time_ops,
 171                 .pv_cpu_ops = pv_cpu_ops,
 172                 .pv_irq_ops = pv_irq_ops,
 173                 .pv_apic_ops = pv_apic_ops,
 174                 .pv_mmu_ops = pv_mmu_ops,
 175         };
 176         return *((void **)&tmpl + type);
 177 }
 178
 179 unsigned paravirt_patch_default(u8 type, u16 clobbers, void *insnbuf,
 180                                 unsigned long addr, unsigned len)
 181 {
 182         void *opfunc = get_call_destination(type);
 183         unsigned ret;
 184
 185         if (opfunc == NULL)
 186                 /* If there's no function, patch it with a ud2a (BUG) */
 187                 ret = paravirt_patch_insns(insnbuf, len, ud2a, ud2a+sizeof(ud2a));
 188         else if (opfunc == paravirt_nop)
 189                 /* If the operation is a nop, then nop the callsite */
 190                 ret = paravirt_patch_nop();
 191         else if (type == PARAVIRT_PATCH(pv_cpu_ops.iret) ||
 192                  type == PARAVIRT_PATCH(pv_cpu_ops.irq_enable_syscall_ret))
 193                 /* If operation requires a jmp, then jmp */
 194                 ret = paravirt_patch_jmp(insnbuf, opfunc, addr, len);
 195         else
 196                 /* Otherwise call the function; assume target could
 197                    clobber any caller-save reg */
 198                 ret = paravirt_patch_call(insnbuf, opfunc, CLBR_ANY,
 199                                           addr, clobbers, len);
 200
 201         return ret;
 202 }
 203
 204 unsigned paravirt_patch_insns(void *insnbuf, unsigned len,
 205                               const char *start, const char *end)
 206 {
 207         unsigned insn_len = end - start;
 208
 209         if (insn_len > len || start == NULL)
 210                 insn_len = len;
 211         else
 212                 memcpy(insnbuf, start, insn_len);
 213
 214         return insn_len;
 215 }
 216
 217 void init_IRQ(void)
 218 {
 219         pv_irq_ops.init_IRQ();
 220 }
 221
 222 static void native_flush_tlb(void)
 223 {
 224         __native_flush_tlb();
 225 }
 226
 227 /*
 228  * Global pages have to be flushed a bit differently. Not a real
 229  * performance problem because this does not happen often.
 230  */
 231 static void native_flush_tlb_global(void)
 232 {
 233         __native_flush_tlb_global();
 234 }
 235
 236 static void native_flush_tlb_single(unsigned long addr)
 237 {
 238         __native_flush_tlb_single(addr);
 239 }
 240
 241 /* These are in entry.S */
 242 extern void native_iret(void);
 243 extern void native_irq_enable_syscall_ret(void);
 244
 245 static int __init print_banner(void)
 246 {
 247         pv_init_ops.banner();
 248         return 0;
 249 }
 250 core_initcall(print_banner);
 251
 252 static struct resource reserve_ioports = {
 253         .start = 0,
 254         .end = IO_SPACE_LIMIT,
 255         .name = "paravirt-ioport",
 256         .flags = IORESOURCE_IO | IORESOURCE_BUSY,
 257 };
 258
 259 static struct resource reserve_iomem = {
 260         .start = 0,
 261         .end = -1,
 262         .name = "paravirt-iomem",
 263         .flags = IORESOURCE_MEM | IORESOURCE_BUSY,
 264 };
 265
 266 /*
 267  * Reserve the whole legacy IO space to prevent any legacy drivers
 268  * from wasting time probing for their hardware.  This is a fairly
 269  * brute-force approach to disabling all non-virtual drivers.
 270  *
 271  * Note that this must be called very early to have any effect.
 272  */
 273 int paravirt_disable_iospace(void)
 274 {
 275         int ret;
 276
 277         ret = request_resource(&ioport_resource, &reserve_ioports);
 278         if (ret == 0) {
 279                 ret = request_resource(&iomem_resource, &reserve_iomem);
 280                 if (ret)
 281                         release_resource(&reserve_ioports);
 282         }
 283
 284         return ret;
 285 }
 286
 287 static DEFINE_PER_CPU(enum paravirt_lazy_mode, paravirt_lazy_mode) = PARAVIRT_LAZY_NONE;
 288
 289 static inline void enter_lazy(enum paravirt_lazy_mode mode)
 290 {
 291         BUG_ON(x86_read_percpu(paravirt_lazy_mode) != PARAVIRT_LAZY_NONE);
 292         BUG_ON(preemptible());
 293
 294         x86_write_percpu(paravirt_lazy_mode, mode);
 295 }
 296
 297 void paravirt_leave_lazy(enum paravirt_lazy_mode mode)
 298 {
 299         BUG_ON(x86_read_percpu(paravirt_lazy_mode) != mode);
 300         BUG_ON(preemptible());
 301
 302         x86_write_percpu(paravirt_lazy_mode, PARAVIRT_LAZY_NONE);
 303 }
 304
 305 void paravirt_enter_lazy_mmu(void)
 306 {
 307         enter_lazy(PARAVIRT_LAZY_MMU);
 308 }
 309
 310 void paravirt_leave_lazy_mmu(void)
 311 {
 312         paravirt_leave_lazy(PARAVIRT_LAZY_MMU);
 313 }
 314
 315 void paravirt_enter_lazy_cpu(void)
 316 {
 317         enter_lazy(PARAVIRT_LAZY_CPU);
 318 }
 319
 320 void paravirt_leave_lazy_cpu(void)
 321 {
 322         paravirt_leave_lazy(PARAVIRT_LAZY_CPU);
 323 }
 324
 325 enum paravirt_lazy_mode paravirt_get_lazy_mode(void)
 326 {
 327         return x86_read_percpu(paravirt_lazy_mode);
 328 }
 329
 330 struct pv_info pv_info = {
 331         .name = "bare hardware",
 332         .paravirt_enabled = 0,
 333         .kernel_rpl = 0,
 334         .shared_kernel_pmd = 1, /* Only used when CONFIG_X86_PAE is set */
 335 };
 336
 337 struct pv_init_ops pv_init_ops = {
 338         .patch = native_patch,
 339         .banner = default_banner,
 340         .arch_setup = paravirt_nop,
 341         .memory_setup = machine_specific_memory_setup,
 342 };
 343
 344 struct pv_time_ops pv_time_ops = {
 345         .time_init = hpet_time_init,
 346         .get_wallclock = native_get_wallclock,
 347         .set_wallclock = native_set_wallclock,
 348         .sched_clock = native_sched_clock,
 349         .get_cpu_khz = native_calculate_cpu_khz,
 350 };
 351
 352 struct pv_irq_ops pv_irq_ops = {
 353         .init_IRQ = native_init_IRQ,
 354         .save_fl = native_save_fl,
 355         .restore_fl = native_restore_fl,
 356         .irq_disable = native_irq_disable,
 357         .irq_enable = native_irq_enable,
 358         .safe_halt = native_safe_halt,
 359         .halt = native_halt,
 360 };
 361
 362 struct pv_cpu_ops pv_cpu_ops = {
 363         .cpuid = native_cpuid,
 364         .get_debugreg = native_get_debugreg,
 365         .set_debugreg = native_set_debugreg,
 366         .clts = native_clts,
 367         .read_cr0 = native_read_cr0,
 368         .write_cr0 = native_write_cr0,
 369         .read_cr4 = native_read_cr4,
 370         .read_cr4_safe = native_read_cr4_safe,
 371         .write_cr4 = native_write_cr4,
 372         .wbinvd = native_wbinvd,
 373         .read_msr = native_read_msr_safe,
 374         .write_msr = native_write_msr_safe,
 375         .read_tsc = native_read_tsc,
 376         .read_pmc = native_read_pmc,
 377         .read_tscp = native_read_tscp,
 378         .load_tr_desc = native_load_tr_desc,
 379         .set_ldt = native_set_ldt,
 380         .load_gdt = native_load_gdt,
 381         .load_idt = native_load_idt,
 382         .store_gdt = native_store_gdt,
 383         .store_idt = native_store_idt,
 384         .store_tr = native_store_tr,
 385         .load_tls = native_load_tls,
 386         .write_ldt_entry = native_write_ldt_entry,
 387         .write_gdt_entry = native_write_gdt_entry,
 388         .write_idt_entry = native_write_idt_entry,
 389         .load_sp0 = native_load_sp0,
 390
 391         .irq_enable_syscall_ret = native_irq_enable_syscall_ret,
 392         .iret = native_iret,
 393         .swapgs = native_swapgs,
 394
 395         .set_iopl_mask = native_set_iopl_mask,
 396         .io_delay = native_io_delay,
 397
 398         .lazy_mode = {
 399                 .enter = paravirt_nop,
 400                 .leave = paravirt_nop,
 401         },
 402 };
 403
 404 struct pv_apic_ops pv_apic_ops = {
 405 #ifdef CONFIG_X86_LOCAL_APIC
 406         .apic_write = native_apic_write,
 407         .apic_write_atomic = native_apic_write_atomic,
 408         .apic_read = native_apic_read,
 409         .setup_boot_clock = setup_boot_APIC_clock,
 410         .setup_secondary_clock = setup_secondary_APIC_clock,
 411         .startup_ipi_hook = paravirt_nop,
 412 #endif
 413 };
 414
 415 struct pv_mmu_ops pv_mmu_ops = {
 416         .pagetable_setup_start = native_pagetable_setup_start,
 417         .pagetable_setup_done = native_pagetable_setup_done,
 418
 419         .read_cr2 = native_read_cr2,
 420         .write_cr2 = native_write_cr2,
 421         .read_cr3 = native_read_cr3,
 422         .write_cr3 = native_write_cr3,
 423
 424         .flush_tlb_user = native_flush_tlb,
 425         .flush_tlb_kernel = native_flush_tlb_global,
 426         .flush_tlb_single = native_flush_tlb_single,
 427         .flush_tlb_others = native_flush_tlb_others,
 428
 429         .alloc_pt = paravirt_nop,
 430         .alloc_pd = paravirt_nop,
 431         .alloc_pd_clone = paravirt_nop,
 432         .release_pt = paravirt_nop,
 433         .release_pd = paravirt_nop,
 434
 435         .set_pte = native_set_pte,
 436         .set_pte_at = native_set_pte_at,
 437         .set_pmd = native_set_pmd,
 438         .pte_update = paravirt_nop,
 439         .pte_update_defer = paravirt_nop,
 440
 441 #ifdef CONFIG_HIGHPTE
 442         .kmap_atomic_pte = kmap_atomic,
 443 #endif
 444
 445 #ifdef CONFIG_X86_PAE
 446         .set_pte_atomic = native_set_pte_atomic,
 447         .set_pte_present = native_set_pte_present,
 448         .set_pud = native_set_pud,
 449         .pte_clear = native_pte_clear,
 450         .pmd_clear = native_pmd_clear,
 451
 452         .pmd_val = native_pmd_val,
 453         .make_pmd = native_make_pmd,
 454 #endif
 455
 456         .pte_val = native_pte_val,
 457         .pgd_val = native_pgd_val,
 458
 459         .make_pte = native_make_pte,
 460         .make_pgd = native_make_pgd,
 461
 462         .dup_mmap = paravirt_nop,
 463         .exit_mmap = paravirt_nop,
 464         .activate_mm = paravirt_nop,
 465
 466         .lazy_mode = {
 467                 .enter = paravirt_nop,
 468                 .leave = paravirt_nop,
 469         },
 470 };
 471
 472 EXPORT_SYMBOL_GPL(pv_time_ops);
 473 EXPORT_SYMBOL    (pv_cpu_ops);
 474 EXPORT_SYMBOL    (pv_mmu_ops);
 475 EXPORT_SYMBOL_GPL(pv_apic_ops);
 476 EXPORT_SYMBOL_GPL(pv_info);
 477 EXPORT_SYMBOL    (pv_irq_ops);