2 * VMI specific paravirt-ops implementation
4 * Copyright (C) 2005, VMware, Inc.
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.
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
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.
21 * Send feedback to zach@vmware.com
25 #include <linux/module.h>
26 #include <linux/license.h>
27 #include <linux/cpu.h>
28 #include <linux/bootmem.h>
32 #include <asm/fixmap.h>
33 #include <asm/apicdef.h>
35 #include <asm/processor.h>
36 #include <asm/timer.h>
37 #include <asm/vmi_time.h>
38 #include <asm/kmap_types.h>
40 /* Convenient for calling VMI functions indirectly in the ROM */
41 typedef u32 __attribute__((regparm(1))) (VROMFUNC)(void);
42 typedef u64 __attribute__((regparm(2))) (VROMLONGFUNC)(int);
44 #define call_vrom_func(rom,func) \
45 (((VROMFUNC *)(rom->func))())
47 #define call_vrom_long_func(rom,func,arg) \
48 (((VROMLONGFUNC *)(rom->func)) (arg))
50 static struct vrom_header *vmi_rom;
51 static int license_gplok;
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 static int disable_vmi_timer;
60 /* Cached VMI operations */
62 void (*cpuid)(void /* non-c */);
63 void (*_set_ldt)(u32 selector);
64 void (*set_tr)(u32 selector);
65 void (*set_kernel_stack)(u32 selector, u32 esp0);
66 void (*allocate_page)(u32, u32, u32, u32, u32);
67 void (*release_page)(u32, u32);
68 void (*set_pte)(pte_t, pte_t *, unsigned);
69 void (*update_pte)(pte_t *, unsigned);
70 void (*set_linear_mapping)(int, u32, u32, u32);
71 void (*flush_tlb)(int);
72 void (*set_initial_ap_state)(int, int);
76 /* XXX move this to alternative.h */
77 extern struct paravirt_patch __start_parainstructions[],
78 __stop_parainstructions[];
81 * VMI patching routines.
83 #define MNEM_CALL 0xe8
87 static char irq_save_disable_callout[] = {
88 MNEM_CALL, 0, 0, 0, 0,
89 MNEM_CALL, 0, 0, 0, 0,
92 #define IRQ_PATCH_INT_MASK 0
93 #define IRQ_PATCH_DISABLE 5
95 static inline void patch_offset(unsigned char *eip, unsigned char *dest)
97 *(unsigned long *)(eip+1) = dest-eip-5;
100 static unsigned patch_internal(int call, unsigned len, void *insns)
103 struct vmi_relocation_info *const rel = (struct vmi_relocation_info *)&reloc;
104 reloc = call_vrom_long_func(vmi_rom, get_reloc, call);
106 case VMI_RELOCATION_CALL_REL:
108 *(char *)insns = MNEM_CALL;
109 patch_offset(insns, rel->eip);
112 case VMI_RELOCATION_JUMP_REL:
114 *(char *)insns = MNEM_JMP;
115 patch_offset(insns, rel->eip);
118 case VMI_RELOCATION_NOP:
119 /* obliterate the whole thing */
122 case VMI_RELOCATION_NONE:
123 /* leave native code in place */
133 * Apply patch if appropriate, return length of new instruction
134 * sequence. The callee does nop padding for us.
136 static unsigned vmi_patch(u8 type, u16 clobbers, void *insns, unsigned len)
139 case PARAVIRT_IRQ_DISABLE:
140 return patch_internal(VMI_CALL_DisableInterrupts, len, insns);
141 case PARAVIRT_IRQ_ENABLE:
142 return patch_internal(VMI_CALL_EnableInterrupts, len, insns);
143 case PARAVIRT_RESTORE_FLAGS:
144 return patch_internal(VMI_CALL_SetInterruptMask, len, insns);
145 case PARAVIRT_SAVE_FLAGS:
146 return patch_internal(VMI_CALL_GetInterruptMask, len, insns);
147 case PARAVIRT_SAVE_FLAGS_IRQ_DISABLE:
149 patch_internal(VMI_CALL_GetInterruptMask, len, insns);
150 patch_internal(VMI_CALL_DisableInterrupts, len-5, insns+5);
154 * You bastards didn't leave enough room to
155 * patch save_flags_irq_disable inline. Patch
159 *(char *)insns = MNEM_CALL;
160 patch_offset(insns, irq_save_disable_callout);
163 case PARAVIRT_INTERRUPT_RETURN:
164 return patch_internal(VMI_CALL_IRET, len, insns);
165 case PARAVIRT_STI_SYSEXIT:
166 return patch_internal(VMI_CALL_SYSEXIT, len, insns);
173 /* CPUID has non-C semantics, and paravirt-ops API doesn't match hardware ISA */
174 static void vmi_cpuid(unsigned int *eax, unsigned int *ebx,
175 unsigned int *ecx, unsigned int *edx)
180 asm volatile ("call *%6"
185 : "0" (*eax), "2" (*ecx), "r" (vmi_ops.cpuid));
188 *edx &= ~X86_FEATURE_PSE;
190 *edx &= ~X86_FEATURE_PGE;
192 *edx &= ~X86_FEATURE_SEP;
194 *edx &= ~X86_FEATURE_TSC;
196 *edx &= ~X86_FEATURE_MTRR;
200 static inline void vmi_maybe_load_tls(struct desc_struct *gdt, int nr, struct desc_struct *new)
202 if (gdt[nr].a != new->a || gdt[nr].b != new->b)
203 write_gdt_entry(gdt, nr, new->a, new->b);
206 static void vmi_load_tls(struct thread_struct *t, unsigned int cpu)
208 struct desc_struct *gdt = get_cpu_gdt_table(cpu);
209 vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 0, &t->tls_array[0]);
210 vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 1, &t->tls_array[1]);
211 vmi_maybe_load_tls(gdt, GDT_ENTRY_TLS_MIN + 2, &t->tls_array[2]);
214 static void vmi_set_ldt(const void *addr, unsigned entries)
216 unsigned cpu = smp_processor_id();
219 pack_descriptor(&low, &high, (unsigned long)addr,
220 entries * sizeof(struct desc_struct) - 1,
222 write_gdt_entry(get_cpu_gdt_table(cpu), GDT_ENTRY_LDT, low, high);
223 vmi_ops._set_ldt(entries ? GDT_ENTRY_LDT*sizeof(struct desc_struct) : 0);
226 static void vmi_set_tr(void)
228 vmi_ops.set_tr(GDT_ENTRY_TSS*sizeof(struct desc_struct));
231 static void vmi_load_esp0(struct tss_struct *tss,
232 struct thread_struct *thread)
234 tss->esp0 = thread->esp0;
236 /* This can only happen when SEP is enabled, no need to test "SEP"arately */
237 if (unlikely(tss->ss1 != thread->sysenter_cs)) {
238 tss->ss1 = thread->sysenter_cs;
239 wrmsr(MSR_IA32_SYSENTER_CS, thread->sysenter_cs, 0);
241 vmi_ops.set_kernel_stack(__KERNEL_DS, tss->esp0);
244 static void vmi_flush_tlb_user(void)
246 vmi_ops.flush_tlb(VMI_FLUSH_TLB);
249 static void vmi_flush_tlb_kernel(void)
251 vmi_ops.flush_tlb(VMI_FLUSH_TLB | VMI_FLUSH_GLOBAL);
254 /* Stub to do nothing at all; used for delays and unimplemented calls */
255 static void vmi_nop(void)
259 /* For NO_IDLE_HZ, we stop the clock when halting the kernel */
260 static fastcall void vmi_safe_halt(void)
262 int idle = vmi_stop_hz_timer();
266 vmi_account_time_restart_hz_timer();
271 #ifdef CONFIG_DEBUG_PAGE_TYPE
273 #ifdef CONFIG_X86_PAE
274 #define MAX_BOOT_PTS (2048+4+1)
276 #define MAX_BOOT_PTS (1024+1)
280 * During boot, mem_map is not yet available in paging_init, so stash
281 * all the boot page allocations here.
286 } boot_page_allocations[MAX_BOOT_PTS];
287 static int num_boot_page_allocations;
288 static int boot_allocations_applied;
290 void vmi_apply_boot_page_allocations(void)
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);
300 boot_allocations_applied = 1;
303 static void record_page_type(u32 pfn, int type)
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++;
311 static void check_zeroed_page(u32 pfn, int type, struct page *page)
315 int limit = PAGE_SIZE / sizeof(int);
317 if (page_address(page))
318 ptr = (u32 *)page_address(page);
320 ptr = (u32 *)__va(pfn << PAGE_SHIFT);
322 * When cloning the root in non-PAE mode, only the userspace
323 * pdes need to be zeroed.
325 if (type & VMI_PAGE_CLONE)
326 limit = USER_PTRS_PER_PGD;
327 for (i = 0; i < limit; i++)
332 * We stash the page type into struct page so we can verify the page
333 * types are used properly.
335 static void vmi_set_page_type(u32 pfn, int type)
337 /* PAE can have multiple roots per page - don't track */
338 if (PTRS_PER_PMD > 1 && (type & VMI_PAGE_PDP))
341 if (boot_allocations_applied) {
342 struct page *page = pfn_to_page(pfn);
343 if (type != VMI_PAGE_NORMAL)
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);
351 record_page_type(pfn, type);
355 static void vmi_check_page_type(u32 pfn, int type)
357 /* PAE can have multiple roots per page - skip checks */
358 if (PTRS_PER_PMD > 1 && (type & VMI_PAGE_PDP))
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);
370 #define vmi_set_page_type(p,t) do { } while (0)
371 #define vmi_check_page_type(p,t) do { } while (0)
374 static void vmi_map_pt_hook(int type, pte_t *va, u32 pfn)
377 * Internally, the VMI ROM must map virtual addresses to physical
378 * addresses for processing MMU updates. By the time MMU updates
379 * are issued, this information is typically already lost.
380 * Fortunately, the VMI provides a cache of mapping slots for active
383 * We use slot zero for the linear mapping of physical memory, and
384 * in HIGHPTE kernels, slot 1 and 2 for KM_PTE0 and KM_PTE1.
386 * args: SLOT VA COUNT PFN
388 BUG_ON(type != KM_PTE0 && type != KM_PTE1);
389 vmi_ops.set_linear_mapping((type - KM_PTE0)+1, (u32)va, 1, pfn);
392 static void vmi_allocate_pt(u32 pfn)
394 vmi_set_page_type(pfn, VMI_PAGE_L1);
395 vmi_ops.allocate_page(pfn, VMI_PAGE_L1, 0, 0, 0);
398 static void vmi_allocate_pd(u32 pfn)
401 * This call comes in very early, before mem_map is setup.
402 * It is called only for swapper_pg_dir, which already has
405 vmi_set_page_type(pfn, VMI_PAGE_L2);
406 vmi_ops.allocate_page(pfn, VMI_PAGE_L2, 0, 0, 0);
409 static void vmi_allocate_pd_clone(u32 pfn, u32 clonepfn, u32 start, u32 count)
411 vmi_set_page_type(pfn, VMI_PAGE_L2 | VMI_PAGE_CLONE);
412 vmi_check_page_type(clonepfn, VMI_PAGE_L2);
413 vmi_ops.allocate_page(pfn, VMI_PAGE_L2 | VMI_PAGE_CLONE, clonepfn, start, count);
416 static void vmi_release_pt(u32 pfn)
418 vmi_ops.release_page(pfn, VMI_PAGE_L1);
419 vmi_set_page_type(pfn, VMI_PAGE_NORMAL);
422 static void vmi_release_pd(u32 pfn)
424 vmi_ops.release_page(pfn, VMI_PAGE_L2);
425 vmi_set_page_type(pfn, VMI_PAGE_NORMAL);
429 * Helper macros for MMU update flags. We can defer updates until a flush
430 * or page invalidation only if the update is to the current address space
431 * (otherwise, there is no flush). We must check against init_mm, since
432 * this could be a kernel update, which usually passes init_mm, although
433 * sometimes this check can be skipped if we know the particular function
434 * is only called on user mode PTEs. We could change the kernel to pass
435 * current->active_mm here, but in particular, I was unsure if changing
436 * mm/highmem.c to do this would still be correct on other architectures.
438 #define is_current_as(mm, mustbeuser) ((mm) == current->active_mm || \
439 (!mustbeuser && (mm) == &init_mm))
440 #define vmi_flags_addr(mm, addr, level, user) \
441 ((level) | (is_current_as(mm, user) ? \
442 (VMI_PAGE_CURRENT_AS | ((addr) & VMI_PAGE_VA_MASK)) : 0))
443 #define vmi_flags_addr_defer(mm, addr, level, user) \
444 ((level) | (is_current_as(mm, user) ? \
445 (VMI_PAGE_DEFER | VMI_PAGE_CURRENT_AS | ((addr) & VMI_PAGE_VA_MASK)) : 0))
447 static void vmi_update_pte(struct mm_struct *mm, u32 addr, pte_t *ptep)
449 vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
450 vmi_ops.update_pte(ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
453 static void vmi_update_pte_defer(struct mm_struct *mm, u32 addr, pte_t *ptep)
455 vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
456 vmi_ops.update_pte(ptep, vmi_flags_addr_defer(mm, addr, VMI_PAGE_PT, 0));
459 static void vmi_set_pte(pte_t *ptep, pte_t pte)
461 /* XXX because of set_pmd_pte, this can be called on PT or PD layers */
462 vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE | VMI_PAGE_PD);
463 vmi_ops.set_pte(pte, ptep, VMI_PAGE_PT);
466 static void vmi_set_pte_at(struct mm_struct *mm, u32 addr, pte_t *ptep, pte_t pte)
468 vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
469 vmi_ops.set_pte(pte, ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
472 static void vmi_set_pmd(pmd_t *pmdp, pmd_t pmdval)
474 #ifdef CONFIG_X86_PAE
475 const pte_t pte = { pmdval.pmd, pmdval.pmd >> 32 };
476 vmi_check_page_type(__pa(pmdp) >> PAGE_SHIFT, VMI_PAGE_PMD);
478 const pte_t pte = { pmdval.pud.pgd.pgd };
479 vmi_check_page_type(__pa(pmdp) >> PAGE_SHIFT, VMI_PAGE_PGD);
481 vmi_ops.set_pte(pte, (pte_t *)pmdp, VMI_PAGE_PD);
484 #ifdef CONFIG_X86_PAE
486 static void vmi_set_pte_atomic(pte_t *ptep, pte_t pteval)
489 * XXX This is called from set_pmd_pte, but at both PT
490 * and PD layers so the VMI_PAGE_PT flag is wrong. But
491 * it is only called for large page mapping changes,
492 * the Xen backend, doesn't support large pages, and the
493 * ESX backend doesn't depend on the flag.
495 set_64bit((unsigned long long *)ptep,pte_val(pteval));
496 vmi_ops.update_pte(ptep, VMI_PAGE_PT);
499 static void vmi_set_pte_present(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
501 vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
502 vmi_ops.set_pte(pte, ptep, vmi_flags_addr_defer(mm, addr, VMI_PAGE_PT, 1));
505 static void vmi_set_pud(pud_t *pudp, pud_t pudval)
508 const pte_t pte = { pudval.pgd.pgd, pudval.pgd.pgd >> 32 };
509 vmi_check_page_type(__pa(pudp) >> PAGE_SHIFT, VMI_PAGE_PGD);
510 vmi_ops.set_pte(pte, (pte_t *)pudp, VMI_PAGE_PDP);
513 static void vmi_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
515 const pte_t pte = { 0 };
516 vmi_check_page_type(__pa(ptep) >> PAGE_SHIFT, VMI_PAGE_PTE);
517 vmi_ops.set_pte(pte, ptep, vmi_flags_addr(mm, addr, VMI_PAGE_PT, 0));
520 void vmi_pmd_clear(pmd_t *pmd)
522 const pte_t pte = { 0 };
523 vmi_check_page_type(__pa(pmd) >> PAGE_SHIFT, VMI_PAGE_PMD);
524 vmi_ops.set_pte(pte, (pte_t *)pmd, VMI_PAGE_PD);
529 extern void setup_pda(void);
531 static void __devinit
532 vmi_startup_ipi_hook(int phys_apicid, unsigned long start_eip,
533 unsigned long start_esp)
535 struct vmi_ap_state ap;
537 /* Default everything to zero. This is fine for most GPRs. */
538 memset(&ap, 0, sizeof(struct vmi_ap_state));
540 ap.gdtr_limit = GDT_SIZE - 1;
541 ap.gdtr_base = (unsigned long) get_cpu_gdt_table(phys_apicid);
543 ap.idtr_limit = IDT_ENTRIES * 8 - 1;
544 ap.idtr_base = (unsigned long) idt_table;
549 ap.eip = (unsigned long) start_eip;
551 ap.esp = (unsigned long) start_esp;
555 ap.fs = __KERNEL_PDA;
562 #ifdef CONFIG_X86_PAE
563 /* efer should match BSP efer. */
566 rdmsr(MSR_EFER, l, h);
567 ap.efer = (unsigned long long) h << 32 | l;
571 ap.cr3 = __pa(swapper_pg_dir);
572 /* Protected mode, paging, AM, WP, NE, MP. */
574 ap.cr4 = mmu_cr4_features;
575 vmi_ops.set_initial_ap_state((u32)&ap, phys_apicid);
579 static inline int __init check_vmi_rom(struct vrom_header *rom)
581 struct pci_header *pci;
582 struct pnp_header *pnp;
583 const char *manufacturer = "UNKNOWN";
584 const char *product = "UNKNOWN";
585 const char *license = "unspecified";
587 if (rom->rom_signature != 0xaa55)
589 if (rom->vrom_signature != VMI_SIGNATURE)
591 if (rom->api_version_maj != VMI_API_REV_MAJOR ||
592 rom->api_version_min+1 < VMI_API_REV_MINOR+1) {
593 printk(KERN_WARNING "VMI: Found mismatched rom version %d.%d\n",
594 rom->api_version_maj,
595 rom->api_version_min);
600 * Relying on the VMI_SIGNATURE field is not 100% safe, so check
601 * the PCI header and device type to make sure this is really a
604 if (!rom->pci_header_offs) {
605 printk(KERN_WARNING "VMI: ROM does not contain PCI header.\n");
609 pci = (struct pci_header *)((char *)rom+rom->pci_header_offs);
610 if (pci->vendorID != PCI_VENDOR_ID_VMWARE ||
611 pci->deviceID != PCI_DEVICE_ID_VMWARE_VMI) {
612 /* Allow it to run... anyways, but warn */
613 printk(KERN_WARNING "VMI: ROM from unknown manufacturer\n");
616 if (rom->pnp_header_offs) {
617 pnp = (struct pnp_header *)((char *)rom+rom->pnp_header_offs);
618 if (pnp->manufacturer_offset)
619 manufacturer = (const char *)rom+pnp->manufacturer_offset;
620 if (pnp->product_offset)
621 product = (const char *)rom+pnp->product_offset;
624 if (rom->license_offs)
625 license = (char *)rom+rom->license_offs;
627 printk(KERN_INFO "VMI: Found %s %s, API version %d.%d, ROM version %d.%d\n",
628 manufacturer, product,
629 rom->api_version_maj, rom->api_version_min,
630 pci->rom_version_maj, pci->rom_version_min);
632 license_gplok = license_is_gpl_compatible(license);
633 if (!license_gplok) {
634 printk(KERN_WARNING "VMI: ROM license '%s' taints kernel... "
635 "inlining disabled\n",
637 add_taint(TAINT_PROPRIETARY_MODULE);
643 * Probe for the VMI option ROM
645 static inline int __init probe_vmi_rom(void)
649 /* VMI ROM is in option ROM area, check signature */
650 for (base = 0xC0000; base < 0xE0000; base += 2048) {
651 struct vrom_header *romstart;
652 romstart = (struct vrom_header *)isa_bus_to_virt(base);
653 if (check_vmi_rom(romstart)) {
662 * VMI setup common to all processors
664 void vmi_bringup(void)
666 /* We must establish the lowmem mapping for MMU ops to work */
667 if (vmi_ops.set_linear_mapping)
668 vmi_ops.set_linear_mapping(0, __PAGE_OFFSET, max_low_pfn, 0);
672 * Return a pointer to a VMI function or NULL if unimplemented
674 static void *vmi_get_function(int vmicall)
677 const struct vmi_relocation_info *rel = (struct vmi_relocation_info *)&reloc;
678 reloc = call_vrom_long_func(vmi_rom, get_reloc, vmicall);
679 BUG_ON(rel->type == VMI_RELOCATION_JUMP_REL);
680 if (rel->type == VMI_RELOCATION_CALL_REL)
681 return (void *)rel->eip;
687 * Helper macro for making the VMI paravirt-ops fill code readable.
688 * For unimplemented operations, fall back to default, unless nop
689 * is returned by the ROM.
691 #define para_fill(opname, vmicall) \
693 reloc = call_vrom_long_func(vmi_rom, get_reloc, \
694 VMI_CALL_##vmicall); \
695 if (rel->type != VMI_RELOCATION_NONE) { \
696 BUG_ON(rel->type != VMI_RELOCATION_CALL_REL); \
697 paravirt_ops.opname = (void *)rel->eip; \
698 } else if (rel->type == VMI_RELOCATION_NOP) \
699 paravirt_ops.opname = (void *)vmi_nop; \
703 * Helper macro for making the VMI paravirt-ops fill code readable.
704 * For cached operations which do not match the VMI ROM ABI and must
705 * go through a tranlation stub. Ignore NOPs, since it is not clear
706 * a NOP * VMI function corresponds to a NOP paravirt-op when the
707 * functions are not in 1-1 correspondence.
709 #define para_wrap(opname, wrapper, cache, vmicall) \
711 reloc = call_vrom_long_func(vmi_rom, get_reloc, \
712 VMI_CALL_##vmicall); \
713 BUG_ON(rel->type == VMI_RELOCATION_JUMP_REL); \
714 if (rel->type == VMI_RELOCATION_CALL_REL) { \
715 paravirt_ops.opname = wrapper; \
716 vmi_ops.cache = (void *)rel->eip; \
722 * Activate the VMI interface and switch into paravirtualized mode
724 static inline int __init activate_vmi(void)
728 const struct vmi_relocation_info *rel = (struct vmi_relocation_info *)&reloc;
730 if (call_vrom_func(vmi_rom, vmi_init) != 0) {
731 printk(KERN_ERR "VMI ROM failed to initialize!");
734 savesegment(cs, kernel_cs);
736 paravirt_ops.paravirt_enabled = 1;
737 paravirt_ops.kernel_rpl = kernel_cs & SEGMENT_RPL_MASK;
739 paravirt_ops.patch = vmi_patch;
740 paravirt_ops.name = "vmi";
743 * Many of these operations are ABI compatible with VMI.
744 * This means we can fill in the paravirt-ops with direct
745 * pointers into the VMI ROM. If the calling convention for
746 * these operations changes, this code needs to be updated.
749 * CPUID paravirt-op uses pointers, not the native ISA
750 * halt has no VMI equivalent; all VMI halts are "safe"
751 * no MSR support yet - just trap and emulate. VMI uses the
752 * same ABI as the native ISA, but Linux wants exceptions
753 * from bogus MSR read / write handled
754 * rdpmc is not yet used in Linux
757 /* CPUID is special, so very special it gets wrapped like a present */
758 para_wrap(cpuid, vmi_cpuid, cpuid, CPUID);
760 para_fill(clts, CLTS);
761 para_fill(get_debugreg, GetDR);
762 para_fill(set_debugreg, SetDR);
763 para_fill(read_cr0, GetCR0);
764 para_fill(read_cr2, GetCR2);
765 para_fill(read_cr3, GetCR3);
766 para_fill(read_cr4, GetCR4);
767 para_fill(write_cr0, SetCR0);
768 para_fill(write_cr2, SetCR2);
769 para_fill(write_cr3, SetCR3);
770 para_fill(write_cr4, SetCR4);
771 para_fill(save_fl, GetInterruptMask);
772 para_fill(restore_fl, SetInterruptMask);
773 para_fill(irq_disable, DisableInterrupts);
774 para_fill(irq_enable, EnableInterrupts);
776 /* irq_save_disable !!! sheer pain */
777 patch_offset(&irq_save_disable_callout[IRQ_PATCH_INT_MASK],
778 (char *)paravirt_ops.save_fl);
779 patch_offset(&irq_save_disable_callout[IRQ_PATCH_DISABLE],
780 (char *)paravirt_ops.irq_disable);
782 para_fill(wbinvd, WBINVD);
783 para_fill(read_tsc, RDTSC);
785 /* The following we emulate with trap and emulate for now */
786 /* paravirt_ops.read_msr = vmi_rdmsr */
787 /* paravirt_ops.write_msr = vmi_wrmsr */
788 /* paravirt_ops.rdpmc = vmi_rdpmc */
790 /* TR interface doesn't pass TR value, wrap */
791 para_wrap(load_tr_desc, vmi_set_tr, set_tr, SetTR);
793 /* LDT is special, too */
794 para_wrap(set_ldt, vmi_set_ldt, _set_ldt, SetLDT);
796 para_fill(load_gdt, SetGDT);
797 para_fill(load_idt, SetIDT);
798 para_fill(store_gdt, GetGDT);
799 para_fill(store_idt, GetIDT);
800 para_fill(store_tr, GetTR);
801 paravirt_ops.load_tls = vmi_load_tls;
802 para_fill(write_ldt_entry, WriteLDTEntry);
803 para_fill(write_gdt_entry, WriteGDTEntry);
804 para_fill(write_idt_entry, WriteIDTEntry);
805 para_wrap(load_esp0, vmi_load_esp0, set_kernel_stack, UpdateKernelStack);
806 para_fill(set_iopl_mask, SetIOPLMask);
807 para_fill(io_delay, IODelay);
808 para_fill(set_lazy_mode, SetLazyMode);
810 /* user and kernel flush are just handled with different flags to FlushTLB */
811 para_wrap(flush_tlb_user, vmi_flush_tlb_user, flush_tlb, FlushTLB);
812 para_wrap(flush_tlb_kernel, vmi_flush_tlb_kernel, flush_tlb, FlushTLB);
813 para_fill(flush_tlb_single, InvalPage);
816 * Until a standard flag format can be agreed on, we need to
817 * implement these as wrappers in Linux. Get the VMI ROM
818 * function pointers for the two backend calls.
820 #ifdef CONFIG_X86_PAE
821 vmi_ops.set_pte = vmi_get_function(VMI_CALL_SetPxELong);
822 vmi_ops.update_pte = vmi_get_function(VMI_CALL_UpdatePxELong);
824 vmi_ops.set_pte = vmi_get_function(VMI_CALL_SetPxE);
825 vmi_ops.update_pte = vmi_get_function(VMI_CALL_UpdatePxE);
828 if (vmi_ops.set_pte) {
829 paravirt_ops.set_pte = vmi_set_pte;
830 paravirt_ops.set_pte_at = vmi_set_pte_at;
831 paravirt_ops.set_pmd = vmi_set_pmd;
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;
841 if (vmi_ops.update_pte) {
842 paravirt_ops.pte_update = vmi_update_pte;
843 paravirt_ops.pte_update_defer = vmi_update_pte_defer;
846 vmi_ops.allocate_page = vmi_get_function(VMI_CALL_AllocatePage);
847 if (vmi_ops.allocate_page) {
848 paravirt_ops.alloc_pt = vmi_allocate_pt;
849 paravirt_ops.alloc_pd = vmi_allocate_pd;
850 paravirt_ops.alloc_pd_clone = vmi_allocate_pd_clone;
853 vmi_ops.release_page = vmi_get_function(VMI_CALL_ReleasePage);
854 if (vmi_ops.release_page) {
855 paravirt_ops.release_pt = vmi_release_pt;
856 paravirt_ops.release_pd = vmi_release_pd;
858 para_wrap(map_pt_hook, vmi_map_pt_hook, set_linear_mapping,
862 * These MUST always be patched. Don't support indirect jumps
863 * through these operations, as the VMI interface may use either
864 * a jump or a call to get to these operations, depending on
865 * the backend. They are performance critical anyway, so requiring
866 * a patch is not a big problem.
868 paravirt_ops.irq_enable_sysexit = (void *)0xfeedbab0;
869 paravirt_ops.iret = (void *)0xbadbab0;
872 para_wrap(startup_ipi_hook, vmi_startup_ipi_hook, set_initial_ap_state, SetInitialAPState);
875 #ifdef CONFIG_X86_LOCAL_APIC
876 para_fill(apic_read, APICRead);
877 para_fill(apic_write, APICWrite);
878 para_fill(apic_write_atomic, APICWrite);
882 * Check for VMI timer functionality by probing for a cycle frequency method
884 reloc = call_vrom_long_func(vmi_rom, get_reloc, VMI_CALL_GetCycleFrequency);
885 if (!disable_vmi_timer && rel->type != VMI_RELOCATION_NONE) {
886 vmi_timer_ops.get_cycle_frequency = (void *)rel->eip;
887 vmi_timer_ops.get_cycle_counter =
888 vmi_get_function(VMI_CALL_GetCycleCounter);
889 vmi_timer_ops.get_wallclock =
890 vmi_get_function(VMI_CALL_GetWallclockTime);
891 vmi_timer_ops.wallclock_updated =
892 vmi_get_function(VMI_CALL_WallclockUpdated);
893 vmi_timer_ops.set_alarm = vmi_get_function(VMI_CALL_SetAlarm);
894 vmi_timer_ops.cancel_alarm =
895 vmi_get_function(VMI_CALL_CancelAlarm);
896 paravirt_ops.time_init = vmi_time_init;
897 paravirt_ops.get_wallclock = vmi_get_wallclock;
898 paravirt_ops.set_wallclock = vmi_set_wallclock;
899 #ifdef CONFIG_X86_LOCAL_APIC
900 paravirt_ops.setup_boot_clock = vmi_timer_setup_boot_alarm;
901 paravirt_ops.setup_secondary_clock = vmi_timer_setup_secondary_alarm;
903 paravirt_ops.get_scheduled_cycles = vmi_get_sched_cycles;
904 paravirt_ops.get_cpu_khz = vmi_cpu_khz;
906 /* We have true wallclock functions; disable CMOS clock sync */
907 no_sync_cmos_clock = 1;
910 disable_vmi_timer = 1;
913 /* No idle HZ mode only works if VMI timer and no idle is enabled */
914 if (disable_noidle || disable_vmi_timer)
915 para_fill(safe_halt, Halt);
917 para_wrap(safe_halt, vmi_safe_halt, halt, Halt);
920 * Alternative instruction rewriting doesn't happen soon enough
921 * to convert VMI_IRET to a call instead of a jump; so we have
922 * to do this before IRQs get reenabled. Fortunately, it is
925 apply_paravirt(__start_parainstructions, __stop_parainstructions);
934 void __init vmi_init(void)
941 check_vmi_rom(vmi_rom);
943 /* In case probing for or validating the ROM failed, basil */
947 reserve_top_address(-vmi_rom->virtual_top);
949 local_irq_save(flags);
952 #ifdef CONFIG_X86_IO_APIC
953 /* This is virtual hardware; timer routing is wired correctly */
956 local_irq_restore(flags & X86_EFLAGS_IF);
959 static int __init parse_vmi(char *arg)
964 if (!strcmp(arg, "disable_pge")) {
965 clear_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability);
967 } else if (!strcmp(arg, "disable_pse")) {
968 clear_bit(X86_FEATURE_PSE, boot_cpu_data.x86_capability);
970 } else if (!strcmp(arg, "disable_sep")) {
971 clear_bit(X86_FEATURE_SEP, boot_cpu_data.x86_capability);
973 } else if (!strcmp(arg, "disable_tsc")) {
974 clear_bit(X86_FEATURE_TSC, boot_cpu_data.x86_capability);
976 } else if (!strcmp(arg, "disable_mtrr")) {
977 clear_bit(X86_FEATURE_MTRR, boot_cpu_data.x86_capability);
979 } else if (!strcmp(arg, "disable_timer")) {
980 disable_vmi_timer = 1;
982 } else if (!strcmp(arg, "disable_noidle"))
987 early_param("vmi", parse_vmi);