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