2 * Core of Xen paravirt_ops implementation.
4 * This file contains the xen_paravirt_ops structure itself, and the
6 * - privileged instructions
11 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
14 #include <linux/cpu.h>
15 #include <linux/kernel.h>
16 #include <linux/init.h>
17 #include <linux/smp.h>
18 #include <linux/preempt.h>
19 #include <linux/hardirq.h>
20 #include <linux/percpu.h>
21 #include <linux/delay.h>
22 #include <linux/start_kernel.h>
23 #include <linux/sched.h>
24 #include <linux/kprobes.h>
25 #include <linux/bootmem.h>
26 #include <linux/module.h>
28 #include <linux/page-flags.h>
29 #include <linux/highmem.h>
30 #include <linux/console.h>
31 #include <linux/pci.h>
32 #include <linux/gfp.h>
33 #include <linux/memblock.h>
36 #include <xen/interface/xen.h>
37 #include <xen/interface/version.h>
38 #include <xen/interface/physdev.h>
39 #include <xen/interface/vcpu.h>
40 #include <xen/interface/memory.h>
41 #include <xen/features.h>
44 #include <xen/hvc-console.h>
46 #include <asm/paravirt.h>
49 #include <asm/xen/pci.h>
50 #include <asm/xen/hypercall.h>
51 #include <asm/xen/hypervisor.h>
52 #include <asm/fixmap.h>
53 #include <asm/processor.h>
54 #include <asm/proto.h>
55 #include <asm/msr-index.h>
56 #include <asm/traps.h>
57 #include <asm/setup.h>
59 #include <asm/pgalloc.h>
60 #include <asm/pgtable.h>
61 #include <asm/tlbflush.h>
62 #include <asm/reboot.h>
63 #include <asm/stackprotector.h>
64 #include <asm/hypervisor.h>
68 #include "multicalls.h"
70 EXPORT_SYMBOL_GPL(hypercall_page);
72 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
73 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
75 enum xen_domain_type xen_domain_type = XEN_NATIVE;
76 EXPORT_SYMBOL_GPL(xen_domain_type);
78 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
79 EXPORT_SYMBOL(machine_to_phys_mapping);
80 unsigned int machine_to_phys_order;
81 EXPORT_SYMBOL(machine_to_phys_order);
83 struct start_info *xen_start_info;
84 EXPORT_SYMBOL_GPL(xen_start_info);
86 struct shared_info xen_dummy_shared_info;
88 void *xen_initial_gdt;
90 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
91 __read_mostly int xen_have_vector_callback;
92 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
95 * Point at some empty memory to start with. We map the real shared_info
96 * page as soon as fixmap is up and running.
98 struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info;
101 * Flag to determine whether vcpu info placement is available on all
102 * VCPUs. We assume it is to start with, and then set it to zero on
103 * the first failure. This is because it can succeed on some VCPUs
104 * and not others, since it can involve hypervisor memory allocation,
105 * or because the guest failed to guarantee all the appropriate
106 * constraints on all VCPUs (ie buffer can't cross a page boundary).
108 * Note that any particular CPU may be using a placed vcpu structure,
109 * but we can only optimise if the all are.
111 * 0: not available, 1: available
113 static int have_vcpu_info_placement = 1;
115 static void clamp_max_cpus(void)
118 if (setup_max_cpus > MAX_VIRT_CPUS)
119 setup_max_cpus = MAX_VIRT_CPUS;
123 static void xen_vcpu_setup(int cpu)
125 struct vcpu_register_vcpu_info info;
127 struct vcpu_info *vcpup;
129 BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
131 if (cpu < MAX_VIRT_CPUS)
132 per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
134 if (!have_vcpu_info_placement) {
135 if (cpu >= MAX_VIRT_CPUS)
140 vcpup = &per_cpu(xen_vcpu_info, cpu);
141 info.mfn = arbitrary_virt_to_mfn(vcpup);
142 info.offset = offset_in_page(vcpup);
144 /* Check to see if the hypervisor will put the vcpu_info
145 structure where we want it, which allows direct access via
146 a percpu-variable. */
147 err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
150 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
151 have_vcpu_info_placement = 0;
154 /* This cpu is using the registered vcpu info, even if
155 later ones fail to. */
156 per_cpu(xen_vcpu, cpu) = vcpup;
161 * On restore, set the vcpu placement up again.
162 * If it fails, then we're in a bad state, since
163 * we can't back out from using it...
165 void xen_vcpu_restore(void)
169 for_each_online_cpu(cpu) {
170 bool other_cpu = (cpu != smp_processor_id());
173 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
176 xen_setup_runstate_info(cpu);
178 if (have_vcpu_info_placement)
182 HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
187 static void __init xen_banner(void)
189 unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
190 struct xen_extraversion extra;
191 HYPERVISOR_xen_version(XENVER_extraversion, &extra);
193 printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
195 printk(KERN_INFO "Xen version: %d.%d%s%s\n",
196 version >> 16, version & 0xffff, extra.extraversion,
197 xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
200 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
201 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
203 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
204 unsigned int *cx, unsigned int *dx)
206 unsigned maskebx = ~0;
207 unsigned maskecx = ~0;
208 unsigned maskedx = ~0;
211 * Mask out inconvenient features, to try and disable as many
212 * unsupported kernel subsystems as possible.
216 maskecx = cpuid_leaf1_ecx_mask;
217 maskedx = cpuid_leaf1_edx_mask;
221 /* Suppress extended topology stuff */
226 asm(XEN_EMULATE_PREFIX "cpuid"
231 : "0" (*ax), "2" (*cx));
238 static __init void xen_init_cpuid_mask(void)
240 unsigned int ax, bx, cx, dx;
242 cpuid_leaf1_edx_mask =
243 ~((1 << X86_FEATURE_MCE) | /* disable MCE */
244 (1 << X86_FEATURE_MCA) | /* disable MCA */
245 (1 << X86_FEATURE_MTRR) | /* disable MTRR */
246 (1 << X86_FEATURE_ACC)); /* thermal monitoring */
248 if (!xen_initial_domain())
249 cpuid_leaf1_edx_mask &=
250 ~((1 << X86_FEATURE_APIC) | /* disable local APIC */
251 (1 << X86_FEATURE_ACPI)); /* disable ACPI */
255 xen_cpuid(&ax, &bx, &cx, &dx);
257 /* cpuid claims we support xsave; try enabling it to see what happens */
258 if (cx & (1 << (X86_FEATURE_XSAVE % 32))) {
261 set_in_cr4(X86_CR4_OSXSAVE);
265 if ((cr4 & X86_CR4_OSXSAVE) == 0)
266 cpuid_leaf1_ecx_mask &= ~(1 << (X86_FEATURE_XSAVE % 32));
268 clear_in_cr4(X86_CR4_OSXSAVE);
272 static void xen_set_debugreg(int reg, unsigned long val)
274 HYPERVISOR_set_debugreg(reg, val);
277 static unsigned long xen_get_debugreg(int reg)
279 return HYPERVISOR_get_debugreg(reg);
282 static void xen_end_context_switch(struct task_struct *next)
285 paravirt_end_context_switch(next);
288 static unsigned long xen_store_tr(void)
294 * Set the page permissions for a particular virtual address. If the
295 * address is a vmalloc mapping (or other non-linear mapping), then
296 * find the linear mapping of the page and also set its protections to
299 static void set_aliased_prot(void *v, pgprot_t prot)
307 ptep = lookup_address((unsigned long)v, &level);
308 BUG_ON(ptep == NULL);
310 pfn = pte_pfn(*ptep);
311 page = pfn_to_page(pfn);
313 pte = pfn_pte(pfn, prot);
315 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
318 if (!PageHighMem(page)) {
319 void *av = __va(PFN_PHYS(pfn));
322 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
328 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
330 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
333 for(i = 0; i < entries; i += entries_per_page)
334 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
337 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
339 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
342 for(i = 0; i < entries; i += entries_per_page)
343 set_aliased_prot(ldt + i, PAGE_KERNEL);
346 static void xen_set_ldt(const void *addr, unsigned entries)
348 struct mmuext_op *op;
349 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
352 op->cmd = MMUEXT_SET_LDT;
353 op->arg1.linear_addr = (unsigned long)addr;
354 op->arg2.nr_ents = entries;
356 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
358 xen_mc_issue(PARAVIRT_LAZY_CPU);
362 static void xen_load_user_cs_desc(int cpu, struct mm_struct *mm)
367 struct desc_struct user_cs;
369 gdt = &get_cpu_gdt_table(cpu)[GDT_ENTRY_DEFAULT_USER_CS];
370 mgdt = virt_to_machine(gdt);
372 user_cs = mm->context.user_cs;
373 descriptor = (u64) user_cs.a | ((u64) user_cs.b) << 32;
375 HYPERVISOR_update_descriptor(mgdt.maddr, descriptor);
377 #endif /*CONFIG_X86_32*/
379 static void xen_load_gdt(const struct desc_ptr *dtr)
381 unsigned long va = dtr->address;
382 unsigned int size = dtr->size + 1;
383 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
384 unsigned long frames[pages];
388 * A GDT can be up to 64k in size, which corresponds to 8192
389 * 8-byte entries, or 16 4k pages..
392 BUG_ON(size > 65536);
393 BUG_ON(va & ~PAGE_MASK);
395 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
398 unsigned long pfn, mfn;
402 * The GDT is per-cpu and is in the percpu data area.
403 * That can be virtually mapped, so we need to do a
404 * page-walk to get the underlying MFN for the
405 * hypercall. The page can also be in the kernel's
406 * linear range, so we need to RO that mapping too.
408 ptep = lookup_address(va, &level);
409 BUG_ON(ptep == NULL);
411 pfn = pte_pfn(*ptep);
412 mfn = pfn_to_mfn(pfn);
413 virt = __va(PFN_PHYS(pfn));
417 make_lowmem_page_readonly((void *)va);
418 make_lowmem_page_readonly(virt);
421 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
426 * load_gdt for early boot, when the gdt is only mapped once
428 static __init void xen_load_gdt_boot(const struct desc_ptr *dtr)
430 unsigned long va = dtr->address;
431 unsigned int size = dtr->size + 1;
432 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
433 unsigned long frames[pages];
437 * A GDT can be up to 64k in size, which corresponds to 8192
438 * 8-byte entries, or 16 4k pages..
441 BUG_ON(size > 65536);
442 BUG_ON(va & ~PAGE_MASK);
444 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
446 unsigned long pfn, mfn;
448 pfn = virt_to_pfn(va);
449 mfn = pfn_to_mfn(pfn);
451 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
453 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
459 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
463 static void load_TLS_descriptor(struct thread_struct *t,
464 unsigned int cpu, unsigned int i)
466 struct desc_struct *gdt = get_cpu_gdt_table(cpu);
467 xmaddr_t maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
468 struct multicall_space mc = __xen_mc_entry(0);
470 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
473 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
476 * XXX sleazy hack: If we're being called in a lazy-cpu zone
477 * and lazy gs handling is enabled, it means we're in a
478 * context switch, and %gs has just been saved. This means we
479 * can zero it out to prevent faults on exit from the
480 * hypervisor if the next process has no %gs. Either way, it
481 * has been saved, and the new value will get loaded properly.
482 * This will go away as soon as Xen has been modified to not
483 * save/restore %gs for normal hypercalls.
485 * On x86_64, this hack is not used for %gs, because gs points
486 * to KERNEL_GS_BASE (and uses it for PDA references), so we
487 * must not zero %gs on x86_64
489 * For x86_64, we need to zero %fs, otherwise we may get an
490 * exception between the new %fs descriptor being loaded and
491 * %fs being effectively cleared at __switch_to().
493 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
503 load_TLS_descriptor(t, cpu, 0);
504 load_TLS_descriptor(t, cpu, 1);
505 load_TLS_descriptor(t, cpu, 2);
507 xen_mc_issue(PARAVIRT_LAZY_CPU);
511 static void xen_load_gs_index(unsigned int idx)
513 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
518 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
521 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
522 u64 entry = *(u64 *)ptr;
527 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
533 static int cvt_gate_to_trap(int vector, const gate_desc *val,
534 struct trap_info *info)
538 if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
541 info->vector = vector;
543 addr = gate_offset(*val);
546 * Look for known traps using IST, and substitute them
547 * appropriately. The debugger ones are the only ones we care
548 * about. Xen will handle faults like double_fault and
549 * machine_check, so we should never see them. Warn if
550 * there's an unexpected IST-using fault handler.
552 if (addr == (unsigned long)debug)
553 addr = (unsigned long)xen_debug;
554 else if (addr == (unsigned long)int3)
555 addr = (unsigned long)xen_int3;
556 else if (addr == (unsigned long)stack_segment)
557 addr = (unsigned long)xen_stack_segment;
558 else if (addr == (unsigned long)double_fault ||
559 addr == (unsigned long)nmi) {
560 /* Don't need to handle these */
562 #ifdef CONFIG_X86_MCE
563 } else if (addr == (unsigned long)machine_check) {
567 /* Some other trap using IST? */
568 if (WARN_ON(val->ist != 0))
571 #endif /* CONFIG_X86_64 */
572 info->address = addr;
574 info->cs = gate_segment(*val);
575 info->flags = val->dpl;
576 /* interrupt gates clear IF */
577 if (val->type == GATE_INTERRUPT)
578 info->flags |= 1 << 2;
583 /* Locations of each CPU's IDT */
584 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
586 /* Set an IDT entry. If the entry is part of the current IDT, then
588 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
590 unsigned long p = (unsigned long)&dt[entrynum];
591 unsigned long start, end;
595 start = __this_cpu_read(idt_desc.address);
596 end = start + __this_cpu_read(idt_desc.size) + 1;
600 native_write_idt_entry(dt, entrynum, g);
602 if (p >= start && (p + 8) <= end) {
603 struct trap_info info[2];
607 if (cvt_gate_to_trap(entrynum, g, &info[0]))
608 if (HYPERVISOR_set_trap_table(info))
615 static void xen_convert_trap_info(const struct desc_ptr *desc,
616 struct trap_info *traps)
618 unsigned in, out, count;
620 count = (desc->size+1) / sizeof(gate_desc);
623 for (in = out = 0; in < count; in++) {
624 gate_desc *entry = (gate_desc*)(desc->address) + in;
626 if (cvt_gate_to_trap(in, entry, &traps[out]))
629 traps[out].address = 0;
632 void xen_copy_trap_info(struct trap_info *traps)
634 const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
636 xen_convert_trap_info(desc, traps);
639 /* Load a new IDT into Xen. In principle this can be per-CPU, so we
640 hold a spinlock to protect the static traps[] array (static because
641 it avoids allocation, and saves stack space). */
642 static void xen_load_idt(const struct desc_ptr *desc)
644 static DEFINE_SPINLOCK(lock);
645 static struct trap_info traps[257];
649 __get_cpu_var(idt_desc) = *desc;
651 xen_convert_trap_info(desc, traps);
654 if (HYPERVISOR_set_trap_table(traps))
660 /* Write a GDT descriptor entry. Ignore LDT descriptors, since
661 they're handled differently. */
662 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
663 const void *desc, int type)
674 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
677 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
687 * Version of write_gdt_entry for use at early boot-time needed to
688 * update an entry as simply as possible.
690 static __init void xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
691 const void *desc, int type)
700 xmaddr_t maddr = virt_to_machine(&dt[entry]);
702 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
703 dt[entry] = *(struct desc_struct *)desc;
709 static void xen_load_sp0(struct tss_struct *tss,
710 struct thread_struct *thread)
712 struct multicall_space mcs = xen_mc_entry(0);
713 MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
714 xen_mc_issue(PARAVIRT_LAZY_CPU);
717 static void xen_set_iopl_mask(unsigned mask)
719 struct physdev_set_iopl set_iopl;
721 /* Force the change at ring 0. */
722 set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
723 HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
726 static void xen_io_delay(void)
730 #ifdef CONFIG_X86_LOCAL_APIC
731 static u32 xen_apic_read(u32 reg)
736 static void xen_apic_write(u32 reg, u32 val)
738 /* Warn to see if there's any stray references */
742 static u64 xen_apic_icr_read(void)
747 static void xen_apic_icr_write(u32 low, u32 id)
749 /* Warn to see if there's any stray references */
753 static void xen_apic_wait_icr_idle(void)
758 static u32 xen_safe_apic_wait_icr_idle(void)
763 static void set_xen_basic_apic_ops(void)
765 apic->read = xen_apic_read;
766 apic->write = xen_apic_write;
767 apic->icr_read = xen_apic_icr_read;
768 apic->icr_write = xen_apic_icr_write;
769 apic->wait_icr_idle = xen_apic_wait_icr_idle;
770 apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
775 static void xen_clts(void)
777 struct multicall_space mcs;
779 mcs = xen_mc_entry(0);
781 MULTI_fpu_taskswitch(mcs.mc, 0);
783 xen_mc_issue(PARAVIRT_LAZY_CPU);
786 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
788 static unsigned long xen_read_cr0(void)
790 unsigned long cr0 = percpu_read(xen_cr0_value);
792 if (unlikely(cr0 == 0)) {
793 cr0 = native_read_cr0();
794 percpu_write(xen_cr0_value, cr0);
800 static void xen_write_cr0(unsigned long cr0)
802 struct multicall_space mcs;
804 percpu_write(xen_cr0_value, cr0);
806 /* Only pay attention to cr0.TS; everything else is
808 mcs = xen_mc_entry(0);
810 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
812 xen_mc_issue(PARAVIRT_LAZY_CPU);
815 static void xen_write_cr4(unsigned long cr4)
819 cr4 &= ~X86_CR4_OSXSAVE;
821 native_write_cr4(cr4);
824 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
835 case MSR_FS_BASE: which = SEGBASE_FS; goto set;
836 case MSR_KERNEL_GS_BASE: which = SEGBASE_GS_USER; goto set;
837 case MSR_GS_BASE: which = SEGBASE_GS_KERNEL; goto set;
840 base = ((u64)high << 32) | low;
841 if (HYPERVISOR_set_segment_base(which, base) != 0)
849 case MSR_SYSCALL_MASK:
850 case MSR_IA32_SYSENTER_CS:
851 case MSR_IA32_SYSENTER_ESP:
852 case MSR_IA32_SYSENTER_EIP:
853 /* Fast syscall setup is all done in hypercalls, so
854 these are all ignored. Stub them out here to stop
855 Xen console noise. */
858 case MSR_IA32_CR_PAT:
859 if (smp_processor_id() == 0)
860 xen_set_pat(((u64)high << 32) | low);
864 ret = native_write_msr_safe(msr, low, high);
870 void xen_setup_shared_info(void)
872 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
873 set_fixmap(FIX_PARAVIRT_BOOTMAP,
874 xen_start_info->shared_info);
876 HYPERVISOR_shared_info =
877 (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
879 HYPERVISOR_shared_info =
880 (struct shared_info *)__va(xen_start_info->shared_info);
883 /* In UP this is as good a place as any to set up shared info */
884 xen_setup_vcpu_info_placement();
887 xen_setup_mfn_list_list();
890 /* This is called once we have the cpu_possible_map */
891 void xen_setup_vcpu_info_placement(void)
895 for_each_possible_cpu(cpu)
898 /* xen_vcpu_setup managed to place the vcpu_info within the
899 percpu area for all cpus, so make use of it */
900 if (have_vcpu_info_placement) {
901 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
902 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
903 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
904 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
905 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
909 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
910 unsigned long addr, unsigned len)
912 char *start, *end, *reloc;
915 start = end = reloc = NULL;
917 #define SITE(op, x) \
918 case PARAVIRT_PATCH(op.x): \
919 if (have_vcpu_info_placement) { \
920 start = (char *)xen_##x##_direct; \
921 end = xen_##x##_direct_end; \
922 reloc = xen_##x##_direct_reloc; \
927 SITE(pv_irq_ops, irq_enable);
928 SITE(pv_irq_ops, irq_disable);
929 SITE(pv_irq_ops, save_fl);
930 SITE(pv_irq_ops, restore_fl);
934 if (start == NULL || (end-start) > len)
937 ret = paravirt_patch_insns(insnbuf, len, start, end);
939 /* Note: because reloc is assigned from something that
940 appears to be an array, gcc assumes it's non-null,
941 but doesn't know its relationship with start and
943 if (reloc > start && reloc < end) {
944 int reloc_off = reloc - start;
945 long *relocp = (long *)(insnbuf + reloc_off);
946 long delta = start - (char *)addr;
954 ret = paravirt_patch_default(type, clobbers, insnbuf,
962 static const struct pv_info xen_info __initdata = {
963 .paravirt_enabled = 1,
964 .shared_kernel_pmd = 0,
969 static const struct pv_init_ops xen_init_ops __initdata = {
973 static const struct pv_cpu_ops xen_cpu_ops __initdata = {
976 .set_debugreg = xen_set_debugreg,
977 .get_debugreg = xen_get_debugreg,
981 .read_cr0 = xen_read_cr0,
982 .write_cr0 = xen_write_cr0,
984 .read_cr4 = native_read_cr4,
985 .read_cr4_safe = native_read_cr4_safe,
986 .write_cr4 = xen_write_cr4,
988 .wbinvd = native_wbinvd,
990 .read_msr = native_read_msr_safe,
991 .write_msr = xen_write_msr_safe,
992 .read_tsc = native_read_tsc,
993 .read_pmc = native_read_pmc,
996 .irq_enable_sysexit = xen_sysexit,
998 .usergs_sysret32 = xen_sysret32,
999 .usergs_sysret64 = xen_sysret64,
1002 .load_tr_desc = paravirt_nop,
1003 .set_ldt = xen_set_ldt,
1004 #ifdef CONFIG_X86_32
1005 .load_user_cs_desc = xen_load_user_cs_desc,
1006 #endif /*CONFIG_X86_32*/
1007 .load_gdt = xen_load_gdt,
1008 .load_idt = xen_load_idt,
1009 .load_tls = xen_load_tls,
1010 #ifdef CONFIG_X86_64
1011 .load_gs_index = xen_load_gs_index,
1014 .alloc_ldt = xen_alloc_ldt,
1015 .free_ldt = xen_free_ldt,
1017 .store_gdt = native_store_gdt,
1018 .store_idt = native_store_idt,
1019 .store_tr = xen_store_tr,
1021 .write_ldt_entry = xen_write_ldt_entry,
1022 .write_gdt_entry = xen_write_gdt_entry,
1023 .write_idt_entry = xen_write_idt_entry,
1024 .load_sp0 = xen_load_sp0,
1026 .set_iopl_mask = xen_set_iopl_mask,
1027 .io_delay = xen_io_delay,
1029 /* Xen takes care of %gs when switching to usermode for us */
1030 .swapgs = paravirt_nop,
1032 .start_context_switch = paravirt_start_context_switch,
1033 .end_context_switch = xen_end_context_switch,
1036 static const struct pv_apic_ops xen_apic_ops __initdata = {
1037 #ifdef CONFIG_X86_LOCAL_APIC
1038 .startup_ipi_hook = paravirt_nop,
1042 static void xen_reboot(int reason)
1044 struct sched_shutdown r = { .reason = reason };
1046 if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1050 static void xen_restart(char *msg)
1052 xen_reboot(SHUTDOWN_reboot);
1055 static void xen_emergency_restart(void)
1057 xen_reboot(SHUTDOWN_reboot);
1060 static void xen_machine_halt(void)
1062 xen_reboot(SHUTDOWN_poweroff);
1065 static void xen_crash_shutdown(struct pt_regs *regs)
1067 xen_reboot(SHUTDOWN_crash);
1071 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1073 xen_reboot(SHUTDOWN_crash);
1077 static struct notifier_block xen_panic_block = {
1078 .notifier_call= xen_panic_event,
1081 int xen_panic_handler_init(void)
1083 atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1087 static const struct machine_ops __initdata xen_machine_ops = {
1088 .restart = xen_restart,
1089 .halt = xen_machine_halt,
1090 .power_off = xen_machine_halt,
1091 .shutdown = xen_machine_halt,
1092 .crash_shutdown = xen_crash_shutdown,
1093 .emergency_restart = xen_emergency_restart,
1097 * Set up the GDT and segment registers for -fstack-protector. Until
1098 * we do this, we have to be careful not to call any stack-protected
1099 * function, which is most of the kernel.
1101 static void __init xen_setup_stackprotector(void)
1103 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1104 pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1106 setup_stack_canary_segment(0);
1107 switch_to_new_gdt(0);
1109 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1110 pv_cpu_ops.load_gdt = xen_load_gdt;
1113 /* First C function to be called on Xen boot */
1114 asmlinkage void __init xen_start_kernel(void)
1116 struct physdev_set_iopl set_iopl;
1120 if (!xen_start_info)
1123 xen_domain_type = XEN_PV_DOMAIN;
1125 xen_setup_machphys_mapping();
1127 /* Install Xen paravirt ops */
1129 pv_init_ops = xen_init_ops;
1130 pv_cpu_ops = xen_cpu_ops;
1131 pv_apic_ops = xen_apic_ops;
1133 x86_init.resources.memory_setup = xen_memory_setup;
1134 x86_init.oem.arch_setup = xen_arch_setup;
1135 x86_init.oem.banner = xen_banner;
1137 xen_init_time_ops();
1140 * Set up some pagetable state before starting to set any ptes.
1145 /* Prevent unwanted bits from being set in PTEs. */
1146 __supported_pte_mask &= ~_PAGE_GLOBAL;
1147 if (!xen_initial_domain())
1148 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1150 __supported_pte_mask |= _PAGE_IOMAP;
1153 * Prevent page tables from being allocated in highmem, even
1154 * if CONFIG_HIGHPTE is enabled.
1156 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1158 /* Work out if we support NX */
1161 xen_setup_features();
1164 if (!xen_feature(XENFEAT_auto_translated_physmap))
1165 xen_build_dynamic_phys_to_machine();
1168 * Set up kernel GDT and segment registers, mainly so that
1169 * -fstack-protector code can be executed.
1171 xen_setup_stackprotector();
1174 xen_init_cpuid_mask();
1176 #ifdef CONFIG_X86_LOCAL_APIC
1178 * set up the basic apic ops.
1180 set_xen_basic_apic_ops();
1183 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1184 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1185 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1188 machine_ops = xen_machine_ops;
1191 * The only reliable way to retain the initial address of the
1192 * percpu gdt_page is to remember it here, so we can go and
1193 * mark it RW later, when the initial percpu area is freed.
1195 xen_initial_gdt = &per_cpu(gdt_page, 0);
1199 #ifdef CONFIG_ACPI_NUMA
1201 * The pages we from Xen are not related to machine pages, so
1202 * any NUMA information the kernel tries to get from ACPI will
1203 * be meaningless. Prevent it from trying.
1208 pgd = (pgd_t *)xen_start_info->pt_base;
1210 if (!xen_initial_domain())
1211 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1213 __supported_pte_mask |= _PAGE_IOMAP;
1214 /* Don't do the full vcpu_info placement stuff until we have a
1215 possible map and a non-dummy shared_info. */
1216 per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1218 local_irq_disable();
1219 early_boot_irqs_disabled = true;
1223 xen_raw_console_write("mapping kernel into physical memory\n");
1224 pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
1225 xen_ident_map_ISA();
1227 /* Allocate and initialize top and mid mfn levels for p2m structure */
1228 xen_build_mfn_list_list();
1230 /* keep using Xen gdt for now; no urgent need to change it */
1232 #ifdef CONFIG_X86_32
1233 pv_info.kernel_rpl = 1;
1234 if (xen_feature(XENFEAT_supervisor_mode_kernel))
1235 pv_info.kernel_rpl = 0;
1237 pv_info.kernel_rpl = 0;
1239 /* set the limit of our address space */
1242 /* We used to do this in xen_arch_setup, but that is too late on AMD
1243 * were early_cpu_init (run before ->arch_setup()) calls early_amd_init
1244 * which pokes 0xcf8 port.
1247 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1249 xen_raw_printk("physdev_op failed %d\n", rc);
1251 #ifdef CONFIG_X86_32
1252 /* set up basic CPUID stuff */
1253 cpu_detect(&new_cpu_data);
1254 new_cpu_data.hard_math = 1;
1255 new_cpu_data.wp_works_ok = 1;
1256 new_cpu_data.x86_capability[0] = cpuid_edx(1);
1259 /* Poke various useful things into boot_params */
1260 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1261 boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1262 ? __pa(xen_start_info->mod_start) : 0;
1263 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1264 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1266 if (!xen_initial_domain()) {
1267 add_preferred_console("xenboot", 0, NULL);
1268 add_preferred_console("tty", 0, NULL);
1269 add_preferred_console("hvc", 0, NULL);
1271 x86_init.pci.arch_init = pci_xen_init;
1273 /* Make sure ACS will be enabled */
1278 xen_raw_console_write("about to get started...\n");
1280 xen_setup_runstate_info(0);
1282 /* Start the world */
1283 #ifdef CONFIG_X86_32
1284 i386_start_kernel();
1286 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1290 static int init_hvm_pv_info(int *major, int *minor)
1292 uint32_t eax, ebx, ecx, edx, pages, msr, base;
1295 base = xen_cpuid_base();
1296 cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1299 *minor = eax & 0xffff;
1300 printk(KERN_INFO "Xen version %d.%d.\n", *major, *minor);
1302 cpuid(base + 2, &pages, &msr, &ecx, &edx);
1304 pfn = __pa(hypercall_page);
1305 wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1307 xen_setup_features();
1310 pv_info.kernel_rpl = 0;
1312 xen_domain_type = XEN_HVM_DOMAIN;
1317 void xen_hvm_init_shared_info(void)
1320 struct xen_add_to_physmap xatp;
1321 static struct shared_info *shared_info_page = 0;
1323 if (!shared_info_page)
1324 shared_info_page = (struct shared_info *)
1325 extend_brk(PAGE_SIZE, PAGE_SIZE);
1326 xatp.domid = DOMID_SELF;
1328 xatp.space = XENMAPSPACE_shared_info;
1329 xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT;
1330 if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1333 HYPERVISOR_shared_info = (struct shared_info *)shared_info_page;
1335 /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1336 * page, we use it in the event channel upcall and in some pvclock
1337 * related functions. We don't need the vcpu_info placement
1338 * optimizations because we don't use any pv_mmu or pv_irq op on
1340 * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is
1341 * online but xen_hvm_init_shared_info is run at resume time too and
1342 * in that case multiple vcpus might be online. */
1343 for_each_online_cpu(cpu) {
1344 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1348 #ifdef CONFIG_XEN_PVHVM
1349 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self,
1350 unsigned long action, void *hcpu)
1352 int cpu = (long)hcpu;
1354 case CPU_UP_PREPARE:
1355 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1363 static struct notifier_block __cpuinitdata xen_hvm_cpu_notifier = {
1364 .notifier_call = xen_hvm_cpu_notify,
1367 static void __init xen_hvm_guest_init(void)
1372 r = init_hvm_pv_info(&major, &minor);
1376 xen_hvm_init_shared_info();
1378 if (xen_feature(XENFEAT_hvm_callback_vector))
1379 xen_have_vector_callback = 1;
1380 register_cpu_notifier(&xen_hvm_cpu_notifier);
1381 xen_unplug_emulated_devices();
1382 have_vcpu_info_placement = 0;
1383 x86_init.irqs.intr_init = xen_init_IRQ;
1384 xen_hvm_init_time_ops();
1385 xen_hvm_init_mmu_ops();
1388 static bool __init xen_hvm_platform(void)
1390 if (xen_pv_domain())
1393 if (!xen_cpuid_base())
1399 bool xen_hvm_need_lapic(void)
1401 if (xen_pv_domain())
1403 if (!xen_hvm_domain())
1405 if (xen_feature(XENFEAT_hvm_pirqs) && xen_have_vector_callback)
1409 EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1411 const __refconst struct hypervisor_x86 x86_hyper_xen_hvm = {
1413 .detect = xen_hvm_platform,
1414 .init_platform = xen_hvm_guest_init,
1416 EXPORT_SYMBOL(x86_hyper_xen_hvm);