git.oblomov.eu Git - linux-2.6/blob - arch/i386/kernel/vmi.c

   1 /*
   2  * VMI specific paravirt-ops implementation
   3  *
   4  * Copyright (C) 2005, VMware, Inc.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License as published by
   8  * the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful, but
  12  * WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  14  * NON INFRINGEMENT.  See the GNU General Public License for more
  15  * details.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * along with this program; if not, write to the Free Software
  19  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  * Send feedback to zach@vmware.com
  22  *
  23  */
  24
  25 #include <linux/module.h>
  26 #include <linux/license.h>
  27 #include <linux/cpu.h>
  28 #include <linux/bootmem.h>
  29 #include <linux/mm.h>
  30 #include <asm/vmi.h>
  31 #include <asm/io.h>
  32 #include <asm/fixmap.h>
  33 #include <asm/apicdef.h>
  34 #include <asm/apic.h>
  35 #include <asm/processor.h>
  36 #include <asm/timer.h>
  37 #include <asm/vmi_time.h>
  38
  39 /* Convenient for calling VMI functions indirectly in the ROM */
  40 typedef u32 __attribute__((regparm(1))) (VROMFUNC)(void);
  41 typedef u64 __attribute__((regparm(2))) (VROMLONGFUNC)(int);
  42
  43 #define call_vrom_func(rom,func) \
  44    (((VROMFUNC *)(rom->func))())
  45
  46 #define call_vrom_long_func(rom,func,arg) \
  47    (((VROMLONGFUNC *)(rom->func)) (arg))
  48
  49 static struct vrom_header *vmi_rom;
  50 static int license_gplok;
  51 static int disable_nodelay;
  52 static int disable_pge;
  53 static int disable_pse;
  54 static int disable_sep;
  55 static int disable_tsc;
  56 static int disable_mtrr;
  57 static int disable_noidle;
  58
  59 /* Cached VMI operations */
  60 struct {
  61         void (*cpuid)(void /* non-c */);
  62         void (*_set_ldt)(u32 selector);
  63         void (*set_tr)(u32 selector);
  64         void (*set_kernel_stack)(u32 selector, u32 esp0);
  65         void (*allocate_page)(u32, u32, u32, u32, u32);
  66         void (*release_page)(u32, u32);
  67         void (*set_pte)(pte_t, pte_t *, unsigned);
  68         void (*update_pte)(pte_t *, unsigned);
  69         void (*set_linear_mapping)(int, u32, u32, u32);
  70         void (*flush_tlb)(int);
  71         void (*set_initial_ap_state)(int, int);
  72         void (*halt)(void);
  73 } vmi_ops;
  74
  75 /* XXX move this to alternative.h */
  76 extern struct paravirt_patch __start_parainstructions[],
  77         __stop_parainstructions[];
  78
  79 /*
  80  * VMI patching routines.
  81  */
  82 #define MNEM_CALL 0xe8
  83 #define MNEM_JMP  0xe9
  84 #define MNEM_RET  0xc3
  85
  86 static char irq_save_disable_callout[] = {
  87         MNEM_CALL, 0, 0, 0, 0,
  88         MNEM_CALL, 0, 0, 0, 0,
  89         MNEM_RET
  90 };
  91 #define IRQ_PATCH_INT_MASK 0
  92 #define IRQ_PATCH_DISABLE  5
  93
  94 static inline void patch_offset(unsigned char *eip, unsigned char *dest)
  95 {
  96         *(unsigned long *)(eip+1) = dest-eip-5;
  97 }
  98
  99 static unsigned patch_internal(int call, unsigned len, void *insns)
 100 {
 101         u64 reloc;
 102         struct vmi_relocation_info *const rel = (struct vmi_relocation_info *)&reloc;
 103         reloc = call_vrom_long_func(vmi_rom, get_reloc, call);
 104         switch(rel->type) {
 105                 case VMI_RELOCATION_CALL_REL:
 106                         BUG_ON(len < 5);
 107                         *(char *)insns = MNEM_CALL;
 108                         patch_offset(insns, rel->eip);
 109                         return 5;
 110
 111                 case VMI_RELOCATION_JUMP_REL:
 112                         BUG_ON(len < 5);
 113                         *(char *)insns = MNEM_JMP;
 114                         patch_offset(insns, rel->eip);
 115                         return 5;
 116
 117                 case VMI_RELOCATION_NOP:
 118                         /* obliterate the whole thing */
 119                         return 0;
 120
 121                 case VMI_RELOCATION_NONE:
 122                         /* leave native code in place */
 123                         break;
 124
 125                 default:
 126                         BUG();
 127         }
 128         return len;
 129 }
 130
 131 /*
 132  * Apply patch if appropriate, return length of new instruction
 133  * sequence.  The callee does nop padding for us.
 134  */
 135 static unsigned vmi_patch(u8 type, u16 clobbers, void *insns, unsigned len)
 136 {
 137         switch (type) {
 138                 case PARAVIRT_IRQ_DISABLE:
 139                         return patch_internal(VMI_CALL_DisableInterrupts, len, insns);
 140                 case PARAVIRT_IRQ_ENABLE:
 141                         return patch_internal(VMI_CALL_EnableInterrupts, len, insns);
 142                 case PARAVIRT_RESTORE_FLAGS:
 143                         return patch_internal(VMI_CALL_SetInterruptMask, len, insns);
 144                 case PARAVIRT_SAVE_FLAGS:
 145                         return patch_internal(VMI_CALL_GetInterruptMask, len, insns);
 146                 case PARAVIRT_SAVE_FLAGS_IRQ_DISABLE:
 147                         if (len >= 10) {
 148                                 patch_internal(VMI_CALL_GetInterruptMask, len, insns);
 149                                 patch_internal(VMI_CALL_DisableInterrupts, len-5, insns+5);
 150                                 return 10;
 151                         } else {
 152                                 /*
 153                                  * You bastards didn't leave enough room to
 154                                  * patch save_flags_irq_disable inline.  Patch
 155                                  * to a helper
 156                                  */
 157                                 BUG_ON(len < 5);
 158                                 *(char *)insns = MNEM_CALL;
 159                                 patch_offset(insns, irq_save_disable_callout);
 160                                 return 5;
 161                         }
 162                 case PARAVIRT_INTERRUPT_RETURN:
 163                         return patch_internal(VMI_CALL_IRET, len, insns);
 164                 case PARAVIRT_STI_SYSEXIT:
 165                         return patch_internal(VMI_CALL_SYSEXIT, len, insns);
 166                 default:
 167                         break;
 168         }
 169         return len;
 170 }
 171
 172 /* CPUID has non-C semantics, and paravirt-ops API doesn't match hardware ISA */
 173 static void vmi_cpuid(unsigned int *eax, unsigned int *ebx,
 174                                unsigned int *ecx, unsigned int *edx)
 175 {
 176         int override = 0;
 177         if (*eax == 1)
 178                 override = 1;
 179         asm volatile ("call *%6"
 180                       : "=a" (*eax),
 181                         "=b" (*ebx),
 182                         "=c" (*ecx),
 183                         "=d" (*edx)
 184                       : "0" (*eax), "2" (*ecx), "r" (vmi_ops.cpuid));
 185         if (override) {
 186                 if (disable_pse)
 187                         *edx &= ~X86_FEATURE_PSE;
 188                 if (disable_pge)
 189                         *edx &= ~X86_FEATURE_PGE;
 190                 if (disable_sep)
 191                         *edx &= ~X86_FEATURE_SEP;
 192                 if (disable_tsc)
 193                         *edx &= ~X86_FEATURE_TSC;
 194                 if (disable_mtrr)
 195                         *edx &= ~X86_FEATURE_MTRR;
 196         }
 197 }
 198
 199 static inline void vmi_maybe_load_tls(struct desc_struct *gdt, int nr, struct desc_struct *new)
 200 {
 201         if (gdt[nr].a != new->a || gdt[nr].b != new->b)
 202                 write_gdt_entry(gdt, nr, new->a, new->b);
 203 }
 204
 205 static void vmi_load_tls(struct thread_struct *t, unsigned int cpu)
 206 {
 207         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
 208         vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 0, &t->tls_array[0]);
 209         vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 1, &t->tls_array[1]);
 210         vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 2, &t->tls_array[2]);
 211 }
 212
 213 static void vmi_set_ldt(const void *addr, unsigned entries)
 214 {
 215         unsigned cpu = smp_processor_id();
 216         u32 low, high;
 217
 218         pack_descriptor(&low, &high, (unsigned long)addr,
 219                         entries * sizeof(struct desc_struct) - 1,
 220                         DESCTYPE_LDT, 0);
 221         write_gdt_entry(get_cpu_gdt_table(cpu), GDT_ENTRY_LDT, low, high);
 222         vmi_ops._set_ldt(entries ? GDT_ENTRY_LDT*sizeof(struct desc_struct) : 0);
 223 }
 224
 225 static void vmi_set_tr(void)
 226 {
 227         vmi_ops.set_tr(GDT_ENTRY_TSS*sizeof(struct desc_struct));
 228 }
 229
 230 static void vmi_load_esp0(struct tss_struct *tss,
 231                                    struct thread_struct *thread)
 232 {
 233         tss->esp0 = thread->esp0;
 234
 235         /* This can only happen when SEP is enabled, no need to test "SEP"arately */
 236         if (unlikely(tss->ss1 != thread->sysenter_cs)) {
 237                 tss->ss1 = thread->sysenter_cs;
 238                 wrmsr(MSR_IA32_SYSENTER_CS, thread->sysenter_cs, 0);
 239         }
 240         vmi_ops.set_kernel_stack(__KERNEL_DS, tss->esp0);
 241 }
 242
 243 static void vmi_flush_tlb_user(void)
 244 {
 245         vmi_ops.flush_tlb(VMI_FLUSH_TLB);
 246 }
 247
 248 static void vmi_flush_tlb_kernel(void)
 249 {
 250         vmi_ops.flush_tlb(VMI_FLUSH_TLB | VMI_FLUSH_GLOBAL);
 251 }
 252
 253 /* Stub to do nothing at all; used for delays and unimplemented calls */
 254 static void vmi_nop(void)
 255 {
 256 }
 257
 258 /* For NO_IDLE_HZ, we stop the clock when halting the kernel */
 259 static fastcall void vmi_safe_halt(void)
 260 {
 261         int idle = vmi_stop_hz_timer();
 262         vmi_ops.halt();
 263         if (idle) {
 264                 local_irq_disable();
 265                 vmi_account_time_restart_hz_timer();
 266                 local_irq_enable();
 267         }
 268 }
 269
 270 #ifdef CONFIG_DEBUG_PAGE_TYPE
 271
 272 #ifdef CONFIG_X86_PAE
 273 #define MAX_BOOT_PTS (2048+4+1)
 274 #else
 275 #define MAX_BOOT_PTS (1024+1)
 276 #endif
 277
 278 /*
 279  * During boot, mem_map is not yet available in paging_init, so stash
 280  * all the boot page allocations here.
 281  */
 282 static struct {
 283         u32 pfn;
 284         int type;
 285 } boot_page_allocations[MAX_BOOT_PTS];
 286 static int num_boot_page_allocations;
 287 static int boot_allocations_applied;
 288
 289 void vmi_apply_boot_page_allocations(void)
 290 {
 291         int i;
 292         BUG_ON(!mem_map);
 293         for (i = 0; i < num_boot_page_allocations; i++) {
 294                 struct page *page = pfn_to_page(boot_page_allocations[i].pfn);
 295                 page->type = boot_page_allocations[i].type;
 296                 page->type = boot_page_allocations[i].type &
 297                                 ~(VMI_PAGE_ZEROED | VMI_PAGE_CLONE);
 298         }
 299         boot_allocations_applied = 1;
 300 }
 301
 302 static void record_page_type(u32 pfn, int type)
 303 {
 304         BUG_ON(num_boot_page_allocations >= MAX_BOOT_PTS);
 305         boot_page_allocations[num_boot_page_allocations].pfn = pfn;
 306         boot_page_allocations[num_boot_page_allocations].type = type;
 307         num_boot_page_allocations++;
 308 }
 309
 310 static void check_zeroed_page(u32 pfn, int type, struct page *page)
 311 {
 312         u32 *ptr;
 313         int i;
 314         int limit = PAGE_SIZE / sizeof(int);
 315
 316         if (page_address(page))
 317                 ptr = (u32 *)page_address(page);
 318         else
 319                 ptr = (u32 *)__va(pfn << PAGE_SHIFT);
 320         /*
 321          * When cloning the root in non-PAE mode, only the userspace
 322          * pdes need to be zeroed.
 323          */
 324         if (type & VMI_PAGE_CLONE)
 325                 limit = USER_PTRS_PER_PGD;
 326         for (i = 0; i < limit; i++)
 327                 BUG_ON(ptr[i]);
 328 }
 329
 330 /*
 331  * We stash the page type into struct page so we can verify the page
 332  * types are used properly.
 333  */
 334 static void vmi_set_page_type(u32 pfn, int type)
 335 {
 336         /* PAE can have multiple roots per page - don't track */
 337         if (PTRS_PER_PMD > 1 && (type & VMI_PAGE_PDP))
 338                 return;
 339
 340         if (boot_allocations_applied) {
 341                 struct page *page = pfn_to_page(pfn);
 342                 if (type != VMI_PAGE_NORMAL)
 343                         BUG_ON(page->type);
 344                 else
 345                         BUG_ON(page->type == VMI_PAGE_NORMAL);
 346                 page->type = type & ~(VMI_PAGE_ZEROED | VMI_PAGE_CLONE);
 347                 if (type & VMI_PAGE_ZEROED)
 348                         check_zeroed_page(pfn, type, page);
 349         } else {
 350                 record_page_type(pfn, type);
 351         }
 352 }
 353
 354 static void vmi_check_page_type(u32 pfn, int type)
 355 {
 356         /* PAE can have multiple roots per page - skip checks */
 357         if (PTRS_PER_PMD > 1 && (type & VMI_PAGE_PDP))
 358                 return;
 359
 360         type &= ~(VMI_PAGE_ZEROED | VMI_PAGE_CLONE);
 361         if (boot_allocations_applied) {
 362                 struct page *page = pfn_to_page(pfn);
 363                 BUG_ON((page->type ^ type) & VMI_PAGE_PAE);
 364                 BUG_ON(type == VMI_PAGE_NORMAL && page->type);
 365                 BUG_ON((type & page->type) == 0);
 366         }
 367 }
 368 #else
 369 #define vmi_set_page_type(p,t) do { } while (0)
 370 #define vmi_check_page_type(p,t) do { } while (0)
 371 #endif
 372
 373 static void vmi_allocate_pt(u32 pfn)
 374 {
 375         vmi_set_page_type(pfn, VMI_PAGE_L1);
 376         vmi_ops.allocate_page(pfn, VMI_PAGE_L1, 0, 0, 0);
 377 }
 378
 379 static void vmi_allocate_pd(u32 pfn)
 380 {
 381         /*
 382          * This call comes in very early, before mem_map is setup.
 383          * It is called only for swapper_pg_dir, which already has
 384          * data on it.
 385          */
 386         vmi_set_page_type(pfn, VMI_PAGE_L2);
 387         vmi_ops.allocate_page(pfn, VMI_PAGE_L2, 0, 0, 0);
 388 }
 389
 390 static void vmi_allocate_pd_clone(u32 pfn, u32 clonepfn, u32 start, u32 count)
 391 {
 392         vmi_set_page_type(pfn, VMI_PAGE_L2 | VMI_PAGE_CLONE);
 393         vmi_check_page_type(clonepfn, VMI_PAGE_L2);
 394         vmi_ops.allocate_page(pfn, VMI_PAGE_L2 | VMI_PAGE_CLONE, clonepfn, start, count);
 395 }
 396
 397 static void vmi_release_pt(u32 pfn)
 398 {
 399         vmi_ops.release_page(pfn, VMI_PAGE_L1);
 400         vmi_set_page_type(pfn, VMI_PAGE_NORMAL);
 401 }
 402
 403 static void vmi_release_pd(u32 pfn)
 404 {
 405         vmi_ops.release_page(pfn, VMI_PAGE_L2);
 406         vmi_set_page_type(pfn, VMI_PAGE_NORMAL);
 407 }
 408
 409 /*
 410  * Helper macros for MMU update flags.  We can defer updates until a flush
 411  * or page invalidation only if the update is to the current address space
 412  * (otherwise, there is no flush).  We must check against init_mm, since
 413  * this could be a kernel update, which usually passes init_mm, although
 414  * sometimes this check can be skipped if we know the particular function
 415  * is only called on user mode PTEs.  We could change the kernel to pass
 416  * current->active_mm here, but in particular, I was unsure if changing
 417  * mm/highmem.c to do this would still be correct on other architectures.
 418  */
 419 #define is_current_as(mm, mustbeuser) ((mm) == current->active_mm ||    \
 420                                        (!mustbeuser && (mm) == &init_mm))
 421 #define vmi_flags_addr(mm, addr, level, user)                           \
 422         ((level) | (is_current_as(mm, user) ?                           \
 423                 (VMI_PAGE_CURRENT_AS | ((addr) & VMI_PAGE_VA_MASK)) : 0))
 424 #define vmi_flags_addr_defer(mm, addr, level, user)                     \
 425         ((level) | (is_current_as(mm, user) ?                           \
 426                 (VMI_PAGE_DEFER | VMI_PAGE_CURRENT_AS | ((addr) & VMI_PAGE_VA_MASK)) : 0))
 427
 428 static void vmi_update_pte(struct mm_struct *mm, u32 addr, pte_t *ptep)
 429 {
 430         vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
 431         vmi_ops.update_pte(ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
 432 }
 433
 434 static void vmi_update_pte_defer(struct mm_struct *mm, u32 addr, pte_t *ptep)
 435 {
 436         vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
 437         vmi_ops.update_pte(ptep, vmi_flags_addr_defer(mm, addr, VMI_PAGE_PT, 0));
 438 }
 439
 440 static void vmi_set_pte(pte_t *ptep, pte_t pte)
 441 {
 442         /* XXX because of set_pmd_pte, this can be called on PT or PD layers */
 443         vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE | VMI_PAGE_PD);
 444         vmi_ops.set_pte(pte, ptep, VMI_PAGE_PT);
 445 }
 446
 447 static void vmi_set_pte_at(struct mm_struct *mm, u32 addr, pte_t *ptep, pte_t pte)
 448 {
 449         vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
 450         vmi_ops.set_pte(pte, ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
 451 }
 452
 453 static void vmi_set_pmd(pmd_t *pmdp, pmd_t pmdval)
 454 {
 455 #ifdef CONFIG_X86_PAE
 456         const pte_t pte = { pmdval.pmd, pmdval.pmd >> 32 };
 457         vmi_check_page_type(__pa(pmdp) >> PAGE_SHIFT, VMI_PAGE_PMD);
 458 #else
 459         const pte_t pte = { pmdval.pud.pgd.pgd };
 460         vmi_check_page_type(__pa(pmdp) >> PAGE_SHIFT, VMI_PAGE_PGD);
 461 #endif
 462         vmi_ops.set_pte(pte, (pte_t *)pmdp, VMI_PAGE_PD);
 463 }
 464
 465 #ifdef CONFIG_X86_PAE
 466
 467 static void vmi_set_pte_atomic(pte_t *ptep, pte_t pteval)
 468 {
 469         /*
 470          * XXX This is called from set_pmd_pte, but at both PT
 471          * and PD layers so the VMI_PAGE_PT flag is wrong.  But
 472          * it is only called for large page mapping changes,
 473          * the Xen backend, doesn't support large pages, and the
 474          * ESX backend doesn't depend on the flag.
 475          */
 476         set_64bit((unsigned long long *)ptep,pte_val(pteval));
 477         vmi_ops.update_pte(ptep, VMI_PAGE_PT);
 478 }
 479
 480 static void vmi_set_pte_present(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
 481 {
 482         vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
 483         vmi_ops.set_pte(pte, ptep, vmi_flags_addr_defer(mm, addr, VMI_PAGE_PT, 1));
 484 }
 485
 486 static void vmi_set_pud(pud_t *pudp, pud_t pudval)
 487 {
 488         /* Um, eww */
 489         const pte_t pte = { pudval.pgd.pgd, pudval.pgd.pgd >> 32 };
 490         vmi_check_page_type(__pa(pudp) >> PAGE_SHIFT, VMI_PAGE_PGD);
 491         vmi_ops.set_pte(pte, (pte_t *)pudp, VMI_PAGE_PDP);
 492 }
 493
 494 static void vmi_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
 495 {
 496         const pte_t pte = { 0 };
 497         vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
 498         vmi_ops.set_pte(pte, ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
 499 }
 500
 501 void vmi_pmd_clear(pmd_t *pmd)
 502 {
 503         const pte_t pte = { 0 };
 504         vmi_check_page_type(__pa(pmd) >> PAGE_SHIFT, VMI_PAGE_PMD);
 505         vmi_ops.set_pte(pte, (pte_t *)pmd, VMI_PAGE_PD);
 506 }
 507 #endif
 508
 509 #ifdef CONFIG_SMP
 510 struct vmi_ap_state ap;
 511 extern void setup_pda(void);
 512
 513 static void __init /* XXX cpu hotplug */
 514 vmi_startup_ipi_hook(int phys_apicid, unsigned long start_eip,
 515                      unsigned long start_esp)
 516 {
 517         /* Default everything to zero.  This is fine for most GPRs. */
 518         memset(&ap, 0, sizeof(struct vmi_ap_state));
 519
 520         ap.gdtr_limit = GDT_SIZE - 1;
 521         ap.gdtr_base = (unsigned long) get_cpu_gdt_table(phys_apicid);
 522
 523         ap.idtr_limit = IDT_ENTRIES * 8 - 1;
 524         ap.idtr_base = (unsigned long) idt_table;
 525
 526         ap.ldtr = 0;
 527
 528         ap.cs = __KERNEL_CS;
 529         ap.eip = (unsigned long) start_eip;
 530         ap.ss = __KERNEL_DS;
 531         ap.esp = (unsigned long) start_esp;
 532
 533         ap.ds = __USER_DS;
 534         ap.es = __USER_DS;
 535         ap.fs = __KERNEL_PDA;
 536         ap.gs = 0;
 537
 538         ap.eflags = 0;
 539
 540         setup_pda();
 541
 542 #ifdef CONFIG_X86_PAE
 543         /* efer should match BSP efer. */
 544         if (cpu_has_nx) {
 545                 unsigned l, h;
 546                 rdmsr(MSR_EFER, l, h);
 547                 ap.efer = (unsigned long long) h << 32 | l;
 548         }
 549 #endif
 550
 551         ap.cr3 = __pa(swapper_pg_dir);
 552         /* Protected mode, paging, AM, WP, NE, MP. */
 553         ap.cr0 = 0x80050023;
 554         ap.cr4 = mmu_cr4_features;
 555         vmi_ops.set_initial_ap_state(__pa(&ap), phys_apicid);
 556 }
 557 #endif
 558
 559 static inline int __init check_vmi_rom(struct vrom_header *rom)
 560 {
 561         struct pci_header *pci;
 562         struct pnp_header *pnp;
 563         const char *manufacturer = "UNKNOWN";
 564         const char *product = "UNKNOWN";
 565         const char *license = "unspecified";
 566
 567         if (rom->rom_signature != 0xaa55)
 568                 return 0;
 569         if (rom->vrom_signature != VMI_SIGNATURE)
 570                 return 0;
 571         if (rom->api_version_maj != VMI_API_REV_MAJOR ||
 572             rom->api_version_min+1 < VMI_API_REV_MINOR+1) {
 573                 printk(KERN_WARNING "VMI: Found mismatched rom version %d.%d\n",
 574                                 rom->api_version_maj,
 575                                 rom->api_version_min);
 576                 return 0;
 577         }
 578
 579         /*
 580          * Relying on the VMI_SIGNATURE field is not 100% safe, so check
 581          * the PCI header and device type to make sure this is really a
 582          * VMI device.
 583          */
 584         if (!rom->pci_header_offs) {
 585                 printk(KERN_WARNING "VMI: ROM does not contain PCI header.\n");
 586                 return 0;
 587         }
 588
 589         pci = (struct pci_header *)((char *)rom+rom->pci_header_offs);
 590         if (pci->vendorID != PCI_VENDOR_ID_VMWARE ||
 591             pci->deviceID != PCI_DEVICE_ID_VMWARE_VMI) {
 592                 /* Allow it to run... anyways, but warn */
 593                 printk(KERN_WARNING "VMI: ROM from unknown manufacturer\n");
 594         }
 595
 596         if (rom->pnp_header_offs) {
 597                 pnp = (struct pnp_header *)((char *)rom+rom->pnp_header_offs);
 598                 if (pnp->manufacturer_offset)
 599                         manufacturer = (const char *)rom+pnp->manufacturer_offset;
 600                 if (pnp->product_offset)
 601                         product = (const char *)rom+pnp->product_offset;
 602         }
 603
 604         if (rom->license_offs)
 605                 license = (char *)rom+rom->license_offs;
 606
 607         printk(KERN_INFO "VMI: Found %s %s, API version %d.%d, ROM version %d.%d\n",
 608                 manufacturer, product,
 609                 rom->api_version_maj, rom->api_version_min,
 610                 pci->rom_version_maj, pci->rom_version_min);
 611
 612         license_gplok = license_is_gpl_compatible(license);
 613         if (!license_gplok) {
 614                 printk(KERN_WARNING "VMI: ROM license '%s' taints kernel... "
 615                        "inlining disabled\n",
 616                        license);
 617                 add_taint(TAINT_PROPRIETARY_MODULE);
 618         }
 619         return 1;
 620 }
 621
 622 /*
 623  * Probe for the VMI option ROM
 624  */
 625 static inline int __init probe_vmi_rom(void)
 626 {
 627         unsigned long base;
 628
 629         /* VMI ROM is in option ROM area, check signature */
 630         for (base = 0xC0000; base < 0xE0000; base += 2048) {
 631                 struct vrom_header *romstart;
 632                 romstart = (struct vrom_header *)isa_bus_to_virt(base);
 633                 if (check_vmi_rom(romstart)) {
 634                         vmi_rom = romstart;
 635                         return 1;
 636                 }
 637         }
 638         return 0;
 639 }
 640
 641 /*
 642  * VMI setup common to all processors
 643  */
 644 void vmi_bringup(void)
 645 {
 646         /* We must establish the lowmem mapping for MMU ops to work */
 647         if (vmi_rom)
 648                 vmi_ops.set_linear_mapping(0, __PAGE_OFFSET, max_low_pfn, 0);
 649 }
 650
 651 /*
 652  * Return a pointer to the VMI function or a NOP stub
 653  */
 654 static void *vmi_get_function(int vmicall)
 655 {
 656         u64 reloc;
 657         const struct vmi_relocation_info *rel = (struct vmi_relocation_info *)&reloc;
 658         reloc = call_vrom_long_func(vmi_rom, get_reloc, vmicall);
 659         BUG_ON(rel->type == VMI_RELOCATION_JUMP_REL);
 660         if (rel->type == VMI_RELOCATION_CALL_REL)
 661                 return (void *)rel->eip;
 662         else
 663                 return (void *)vmi_nop;
 664 }
 665
 666 /*
 667  * Helper macro for making the VMI paravirt-ops fill code readable.
 668  * For unimplemented operations, fall back to default.
 669  */
 670 #define para_fill(opname, vmicall)                              \
 671 do {                                                            \
 672         reloc = call_vrom_long_func(vmi_rom, get_reloc,         \
 673                                     VMI_CALL_##vmicall);        \
 674         if (rel->type != VMI_RELOCATION_NONE) {                 \
 675                 BUG_ON(rel->type != VMI_RELOCATION_CALL_REL);   \
 676                 paravirt_ops.opname = (void *)rel->eip;         \
 677         }                                                       \
 678 } while (0)
 679
 680 /*
 681  * Activate the VMI interface and switch into paravirtualized mode
 682  */
 683 static inline int __init activate_vmi(void)
 684 {
 685         short kernel_cs;
 686         u64 reloc;
 687         const struct vmi_relocation_info *rel = (struct vmi_relocation_info *)&reloc;
 688
 689         if (call_vrom_func(vmi_rom, vmi_init) != 0) {
 690                 printk(KERN_ERR "VMI ROM failed to initialize!");
 691                 return 0;
 692         }
 693         savesegment(cs, kernel_cs);
 694
 695         paravirt_ops.paravirt_enabled = 1;
 696         paravirt_ops.kernel_rpl = kernel_cs & SEGMENT_RPL_MASK;
 697
 698         paravirt_ops.patch = vmi_patch;
 699         paravirt_ops.name = "vmi";
 700
 701         /*
 702          * Many of these operations are ABI compatible with VMI.
 703          * This means we can fill in the paravirt-ops with direct
 704          * pointers into the VMI ROM.  If the calling convention for
 705          * these operations changes, this code needs to be updated.
 706          *
 707          * Exceptions
 708          *  CPUID paravirt-op uses pointers, not the native ISA
 709          *  halt has no VMI equivalent; all VMI halts are "safe"
 710          *  no MSR support yet - just trap and emulate.  VMI uses the
 711          *    same ABI as the native ISA, but Linux wants exceptions
 712          *    from bogus MSR read / write handled
 713          *  rdpmc is not yet used in Linux
 714          */
 715
 716         /* CPUID is special, so very special */
 717         reloc = call_vrom_long_func(vmi_rom, get_reloc, VMI_CALL_CPUID);
 718         if (rel->type != VMI_RELOCATION_NONE) {
 719                 BUG_ON(rel->type != VMI_RELOCATION_CALL_REL);
 720                 vmi_ops.cpuid = (void *)rel->eip;
 721                 paravirt_ops.cpuid = vmi_cpuid;
 722         }
 723
 724         para_fill(clts, CLTS);
 725         para_fill(get_debugreg, GetDR);
 726         para_fill(set_debugreg, SetDR);
 727         para_fill(read_cr0, GetCR0);
 728         para_fill(read_cr2, GetCR2);
 729         para_fill(read_cr3, GetCR3);
 730         para_fill(read_cr4, GetCR4);
 731         para_fill(write_cr0, SetCR0);
 732         para_fill(write_cr2, SetCR2);
 733         para_fill(write_cr3, SetCR3);
 734         para_fill(write_cr4, SetCR4);
 735         para_fill(save_fl, GetInterruptMask);
 736         para_fill(restore_fl, SetInterruptMask);
 737         para_fill(irq_disable, DisableInterrupts);
 738         para_fill(irq_enable, EnableInterrupts);
 739         /* irq_save_disable !!! sheer pain */
 740         patch_offset(&irq_save_disable_callout[IRQ_PATCH_INT_MASK],
 741                      (char *)paravirt_ops.save_fl);
 742         patch_offset(&irq_save_disable_callout[IRQ_PATCH_DISABLE],
 743                      (char *)paravirt_ops.irq_disable);
 744
 745         para_fill(wbinvd, WBINVD);
 746         /* paravirt_ops.read_msr = vmi_rdmsr */
 747         /* paravirt_ops.write_msr = vmi_wrmsr */
 748         para_fill(read_tsc, RDTSC);
 749         /* paravirt_ops.rdpmc = vmi_rdpmc */
 750
 751         /* TR interface doesn't pass TR value */
 752         reloc = call_vrom_long_func(vmi_rom, get_reloc, VMI_CALL_SetTR);
 753         if (rel->type != VMI_RELOCATION_NONE) {
 754                 BUG_ON(rel->type != VMI_RELOCATION_CALL_REL);
 755                 vmi_ops.set_tr = (void *)rel->eip;
 756                 paravirt_ops.load_tr_desc = vmi_set_tr;
 757         }
 758
 759         /* LDT is special, too */
 760         reloc = call_vrom_long_func(vmi_rom, get_reloc, VMI_CALL_SetLDT);
 761         if (rel->type != VMI_RELOCATION_NONE) {
 762                 BUG_ON(rel->type != VMI_RELOCATION_CALL_REL);
 763                 vmi_ops._set_ldt = (void *)rel->eip;
 764                 paravirt_ops.set_ldt = vmi_set_ldt;
 765         }
 766
 767         para_fill(load_gdt, SetGDT);
 768         para_fill(load_idt, SetIDT);
 769         para_fill(store_gdt, GetGDT);
 770         para_fill(store_idt, GetIDT);
 771         para_fill(store_tr, GetTR);
 772         paravirt_ops.load_tls = vmi_load_tls;
 773         para_fill(write_ldt_entry, WriteLDTEntry);
 774         para_fill(write_gdt_entry, WriteGDTEntry);
 775         para_fill(write_idt_entry, WriteIDTEntry);
 776         reloc = call_vrom_long_func(vmi_rom, get_reloc,
 777                                     VMI_CALL_UpdateKernelStack);
 778         if (rel->type != VMI_RELOCATION_NONE) {
 779                 BUG_ON(rel->type != VMI_RELOCATION_CALL_REL);
 780                 vmi_ops.set_kernel_stack = (void *)rel->eip;
 781                 paravirt_ops.load_esp0 = vmi_load_esp0;
 782         }
 783
 784         para_fill(set_iopl_mask, SetIOPLMask);
 785         paravirt_ops.io_delay = (void *)vmi_nop;
 786         if (!disable_nodelay) {
 787                 paravirt_ops.const_udelay = (void *)vmi_nop;
 788         }
 789
 790         para_fill(set_lazy_mode, SetLazyMode);
 791
 792         reloc = call_vrom_long_func(vmi_rom, get_reloc, VMI_CALL_FlushTLB);
 793         if (rel->type != VMI_RELOCATION_NONE) {
 794                 vmi_ops.flush_tlb = (void *)rel->eip;
 795                 paravirt_ops.flush_tlb_user = vmi_flush_tlb_user;
 796                 paravirt_ops.flush_tlb_kernel = vmi_flush_tlb_kernel;
 797         }
 798         para_fill(flush_tlb_single, InvalPage);
 799
 800         /*
 801          * Until a standard flag format can be agreed on, we need to
 802          * implement these as wrappers in Linux.  Get the VMI ROM
 803          * function pointers for the two backend calls.
 804          */
 805 #ifdef CONFIG_X86_PAE
 806         vmi_ops.set_pte = vmi_get_function(VMI_CALL_SetPxELong);
 807         vmi_ops.update_pte = vmi_get_function(VMI_CALL_UpdatePxELong);
 808 #else
 809         vmi_ops.set_pte = vmi_get_function(VMI_CALL_SetPxE);
 810         vmi_ops.update_pte = vmi_get_function(VMI_CALL_UpdatePxE);
 811 #endif
 812         vmi_ops.set_linear_mapping = vmi_get_function(VMI_CALL_SetLinearMapping);
 813         vmi_ops.allocate_page = vmi_get_function(VMI_CALL_AllocatePage);
 814         vmi_ops.release_page = vmi_get_function(VMI_CALL_ReleasePage);
 815
 816         paravirt_ops.alloc_pt = vmi_allocate_pt;
 817         paravirt_ops.alloc_pd = vmi_allocate_pd;
 818         paravirt_ops.alloc_pd_clone = vmi_allocate_pd_clone;
 819         paravirt_ops.release_pt = vmi_release_pt;
 820         paravirt_ops.release_pd = vmi_release_pd;
 821         paravirt_ops.set_pte = vmi_set_pte;
 822         paravirt_ops.set_pte_at = vmi_set_pte_at;
 823         paravirt_ops.set_pmd = vmi_set_pmd;
 824         paravirt_ops.pte_update = vmi_update_pte;
 825         paravirt_ops.pte_update_defer = vmi_update_pte_defer;
 826 #ifdef CONFIG_X86_PAE
 827         paravirt_ops.set_pte_atomic = vmi_set_pte_atomic;
 828         paravirt_ops.set_pte_present = vmi_set_pte_present;
 829         paravirt_ops.set_pud = vmi_set_pud;
 830         paravirt_ops.pte_clear = vmi_pte_clear;
 831         paravirt_ops.pmd_clear = vmi_pmd_clear;
 832 #endif
 833         /*
 834          * These MUST always be patched.  Don't support indirect jumps
 835          * through these operations, as the VMI interface may use either
 836          * a jump or a call to get to these operations, depending on
 837          * the backend.  They are performance critical anyway, so requiring
 838          * a patch is not a big problem.
 839          */
 840         paravirt_ops.irq_enable_sysexit = (void *)0xfeedbab0;
 841         paravirt_ops.iret = (void *)0xbadbab0;
 842
 843 #ifdef CONFIG_SMP
 844         paravirt_ops.startup_ipi_hook = vmi_startup_ipi_hook;
 845         vmi_ops.set_initial_ap_state = vmi_get_function(VMI_CALL_SetInitialAPState);
 846 #endif
 847
 848 #ifdef CONFIG_X86_LOCAL_APIC
 849         paravirt_ops.apic_read = vmi_get_function(VMI_CALL_APICRead);
 850         paravirt_ops.apic_write = vmi_get_function(VMI_CALL_APICWrite);
 851         paravirt_ops.apic_write_atomic = vmi_get_function(VMI_CALL_APICWrite);
 852 #endif
 853
 854         /*
 855          * Check for VMI timer functionality by probing for a cycle frequency method
 856          */
 857         reloc = call_vrom_long_func(vmi_rom, get_reloc, VMI_CALL_GetCycleFrequency);
 858         if (rel->type != VMI_RELOCATION_NONE) {
 859                 vmi_timer_ops.get_cycle_frequency = (void *)rel->eip;
 860                 vmi_timer_ops.get_cycle_counter =
 861                         vmi_get_function(VMI_CALL_GetCycleCounter);
 862                 vmi_timer_ops.get_wallclock =
 863                         vmi_get_function(VMI_CALL_GetWallclockTime);
 864                 vmi_timer_ops.wallclock_updated =
 865                         vmi_get_function(VMI_CALL_WallclockUpdated);
 866                 vmi_timer_ops.set_alarm = vmi_get_function(VMI_CALL_SetAlarm);
 867                 vmi_timer_ops.cancel_alarm =
 868                          vmi_get_function(VMI_CALL_CancelAlarm);
 869                 paravirt_ops.time_init = vmi_time_init;
 870                 paravirt_ops.get_wallclock = vmi_get_wallclock;
 871                 paravirt_ops.set_wallclock = vmi_set_wallclock;
 872 #ifdef CONFIG_X86_LOCAL_APIC
 873                 paravirt_ops.setup_boot_clock = vmi_timer_setup_boot_alarm;
 874                 paravirt_ops.setup_secondary_clock = vmi_timer_setup_secondary_alarm;
 875 #endif
 876                 paravirt_ops.get_scheduled_cycles = vmi_get_sched_cycles;
 877         }
 878         if (!disable_noidle)
 879                 para_fill(safe_halt, Halt);
 880         else {
 881                 vmi_ops.halt = vmi_get_function(VMI_CALL_Halt);
 882                 paravirt_ops.safe_halt = vmi_safe_halt;
 883         }
 884
 885         /*
 886          * Alternative instruction rewriting doesn't happen soon enough
 887          * to convert VMI_IRET to a call instead of a jump; so we have
 888          * to do this before IRQs get reenabled.  Fortunately, it is
 889          * idempotent.
 890          */
 891         apply_paravirt(__start_parainstructions, __stop_parainstructions);
 892
 893         vmi_bringup();
 894
 895         return 1;
 896 }
 897
 898 #undef para_fill
 899
 900 void __init vmi_init(void)
 901 {
 902         unsigned long flags;
 903
 904         if (!vmi_rom)
 905                 probe_vmi_rom();
 906         else
 907                 check_vmi_rom(vmi_rom);
 908
 909         /* In case probing for or validating the ROM failed, basil */
 910         if (!vmi_rom)
 911                 return;
 912
 913         reserve_top_address(-vmi_rom->virtual_top);
 914
 915         local_irq_save(flags);
 916         activate_vmi();
 917
 918 #ifdef CONFIG_X86_IO_APIC
 919         no_timer_check = 1;
 920 #endif
 921
 922         local_irq_restore(flags & X86_EFLAGS_IF);
 923 }
 924
 925 static int __init parse_vmi(char *arg)
 926 {
 927         if (!arg)
 928                 return -EINVAL;
 929
 930         if (!strcmp(arg, "disable_nodelay"))
 931                 disable_nodelay = 1;
 932         else if (!strcmp(arg, "disable_pge")) {
 933                 clear_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability);
 934                 disable_pge = 1;
 935         } else if (!strcmp(arg, "disable_pse")) {
 936                 clear_bit(X86_FEATURE_PSE, boot_cpu_data.x86_capability);
 937                 disable_pse = 1;
 938         } else if (!strcmp(arg, "disable_sep")) {
 939                 clear_bit(X86_FEATURE_SEP, boot_cpu_data.x86_capability);
 940                 disable_sep = 1;
 941         } else if (!strcmp(arg, "disable_tsc")) {
 942                 clear_bit(X86_FEATURE_TSC, boot_cpu_data.x86_capability);
 943                 disable_tsc = 1;
 944         } else if (!strcmp(arg, "disable_mtrr")) {
 945                 clear_bit(X86_FEATURE_MTRR, boot_cpu_data.x86_capability);
 946                 disable_mtrr = 1;
 947         } else if (!strcmp(arg, "disable_noidle"))
 948                 disable_noidle = 1;
 949         return 0;
 950 }
 951
 952 early_param("vmi", parse_vmi);