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>
65 #include <asm/pci_x86.h>
69 #include "multicalls.h"
71 EXPORT_SYMBOL_GPL(hypercall_page);
73 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
74 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
76 enum xen_domain_type xen_domain_type = XEN_NATIVE;
77 EXPORT_SYMBOL_GPL(xen_domain_type);
79 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
80 EXPORT_SYMBOL(machine_to_phys_mapping);
81 unsigned long machine_to_phys_nr;
82 EXPORT_SYMBOL(machine_to_phys_nr);
84 struct start_info *xen_start_info;
85 EXPORT_SYMBOL_GPL(xen_start_info);
87 struct shared_info xen_dummy_shared_info;
89 void *xen_initial_gdt;
91 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
92 __read_mostly int xen_have_vector_callback;
93 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
96 * Point at some empty memory to start with. We map the real shared_info
97 * page as soon as fixmap is up and running.
99 struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info;
102 * Flag to determine whether vcpu info placement is available on all
103 * VCPUs. We assume it is to start with, and then set it to zero on
104 * the first failure. This is because it can succeed on some VCPUs
105 * and not others, since it can involve hypervisor memory allocation,
106 * or because the guest failed to guarantee all the appropriate
107 * constraints on all VCPUs (ie buffer can't cross a page boundary).
109 * Note that any particular CPU may be using a placed vcpu structure,
110 * but we can only optimise if the all are.
112 * 0: not available, 1: available
114 static int have_vcpu_info_placement = 1;
116 static void clamp_max_cpus(void)
119 if (setup_max_cpus > MAX_VIRT_CPUS)
120 setup_max_cpus = MAX_VIRT_CPUS;
124 static void xen_vcpu_setup(int cpu)
126 struct vcpu_register_vcpu_info info;
128 struct vcpu_info *vcpup;
130 BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
132 if (cpu < MAX_VIRT_CPUS)
133 per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
135 if (!have_vcpu_info_placement) {
136 if (cpu >= MAX_VIRT_CPUS)
141 vcpup = &per_cpu(xen_vcpu_info, cpu);
142 info.mfn = arbitrary_virt_to_mfn(vcpup);
143 info.offset = offset_in_page(vcpup);
145 /* Check to see if the hypervisor will put the vcpu_info
146 structure where we want it, which allows direct access via
147 a percpu-variable. */
148 err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
151 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
152 have_vcpu_info_placement = 0;
155 /* This cpu is using the registered vcpu info, even if
156 later ones fail to. */
157 per_cpu(xen_vcpu, cpu) = vcpup;
162 * On restore, set the vcpu placement up again.
163 * If it fails, then we're in a bad state, since
164 * we can't back out from using it...
166 void xen_vcpu_restore(void)
170 for_each_online_cpu(cpu) {
171 bool other_cpu = (cpu != smp_processor_id());
174 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
177 xen_setup_runstate_info(cpu);
179 if (have_vcpu_info_placement)
183 HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
188 static void __init xen_banner(void)
190 unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
191 struct xen_extraversion extra;
192 HYPERVISOR_xen_version(XENVER_extraversion, &extra);
194 printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
196 printk(KERN_INFO "Xen version: %d.%d%s%s\n",
197 version >> 16, version & 0xffff, extra.extraversion,
198 xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
201 #define CPUID_THERM_POWER_LEAF 6
202 #define APERFMPERF_PRESENT 0
204 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
205 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
207 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
208 unsigned int *cx, unsigned int *dx)
210 unsigned maskebx = ~0;
211 unsigned maskecx = ~0;
212 unsigned maskedx = ~0;
215 * Mask out inconvenient features, to try and disable as many
216 * unsupported kernel subsystems as possible.
220 maskecx = cpuid_leaf1_ecx_mask;
221 maskedx = cpuid_leaf1_edx_mask;
224 case CPUID_THERM_POWER_LEAF:
225 /* Disabling APERFMPERF for kernel usage */
226 maskecx = ~(1 << APERFMPERF_PRESENT);
230 /* Suppress extended topology stuff */
235 asm(XEN_EMULATE_PREFIX "cpuid"
240 : "0" (*ax), "2" (*cx));
247 static void __init xen_init_cpuid_mask(void)
249 unsigned int ax, bx, cx, dx;
250 unsigned int xsave_mask;
252 cpuid_leaf1_edx_mask =
253 ~((1 << X86_FEATURE_MCE) | /* disable MCE */
254 (1 << X86_FEATURE_MCA) | /* disable MCA */
255 (1 << X86_FEATURE_MTRR) | /* disable MTRR */
256 (1 << X86_FEATURE_ACC)); /* thermal monitoring */
258 if (!xen_initial_domain())
259 cpuid_leaf1_edx_mask &=
260 ~((1 << X86_FEATURE_APIC) | /* disable local APIC */
261 (1 << X86_FEATURE_ACPI)); /* disable ACPI */
264 xen_cpuid(&ax, &bx, &cx, &dx);
267 (1 << (X86_FEATURE_XSAVE % 32)) |
268 (1 << (X86_FEATURE_OSXSAVE % 32));
270 /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
271 if ((cx & xsave_mask) != xsave_mask)
272 cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
275 static void xen_set_debugreg(int reg, unsigned long val)
277 HYPERVISOR_set_debugreg(reg, val);
280 static unsigned long xen_get_debugreg(int reg)
282 return HYPERVISOR_get_debugreg(reg);
285 static void xen_end_context_switch(struct task_struct *next)
288 paravirt_end_context_switch(next);
291 static unsigned long xen_store_tr(void)
297 * Set the page permissions for a particular virtual address. If the
298 * address is a vmalloc mapping (or other non-linear mapping), then
299 * find the linear mapping of the page and also set its protections to
302 static void set_aliased_prot(void *v, pgprot_t prot)
310 ptep = lookup_address((unsigned long)v, &level);
311 BUG_ON(ptep == NULL);
313 pfn = pte_pfn(*ptep);
314 page = pfn_to_page(pfn);
316 pte = pfn_pte(pfn, prot);
318 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
321 if (!PageHighMem(page)) {
322 void *av = __va(PFN_PHYS(pfn));
325 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
331 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
333 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
336 for(i = 0; i < entries; i += entries_per_page)
337 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
340 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
342 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
345 for(i = 0; i < entries; i += entries_per_page)
346 set_aliased_prot(ldt + i, PAGE_KERNEL);
349 static void xen_set_ldt(const void *addr, unsigned entries)
351 struct mmuext_op *op;
352 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
354 trace_xen_cpu_set_ldt(addr, entries);
357 op->cmd = MMUEXT_SET_LDT;
358 op->arg1.linear_addr = (unsigned long)addr;
359 op->arg2.nr_ents = entries;
361 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
363 xen_mc_issue(PARAVIRT_LAZY_CPU);
366 static void xen_load_gdt(const struct desc_ptr *dtr)
368 unsigned long va = dtr->address;
369 unsigned int size = dtr->size + 1;
370 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
371 unsigned long frames[pages];
375 * A GDT can be up to 64k in size, which corresponds to 8192
376 * 8-byte entries, or 16 4k pages..
379 BUG_ON(size > 65536);
380 BUG_ON(va & ~PAGE_MASK);
382 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
385 unsigned long pfn, mfn;
389 * The GDT is per-cpu and is in the percpu data area.
390 * That can be virtually mapped, so we need to do a
391 * page-walk to get the underlying MFN for the
392 * hypercall. The page can also be in the kernel's
393 * linear range, so we need to RO that mapping too.
395 ptep = lookup_address(va, &level);
396 BUG_ON(ptep == NULL);
398 pfn = pte_pfn(*ptep);
399 mfn = pfn_to_mfn(pfn);
400 virt = __va(PFN_PHYS(pfn));
404 make_lowmem_page_readonly((void *)va);
405 make_lowmem_page_readonly(virt);
408 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
413 * load_gdt for early boot, when the gdt is only mapped once
415 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
417 unsigned long va = dtr->address;
418 unsigned int size = dtr->size + 1;
419 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
420 unsigned long frames[pages];
424 * A GDT can be up to 64k in size, which corresponds to 8192
425 * 8-byte entries, or 16 4k pages..
428 BUG_ON(size > 65536);
429 BUG_ON(va & ~PAGE_MASK);
431 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
433 unsigned long pfn, mfn;
435 pfn = virt_to_pfn(va);
436 mfn = pfn_to_mfn(pfn);
438 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
440 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
446 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
450 static void load_TLS_descriptor(struct thread_struct *t,
451 unsigned int cpu, unsigned int i)
453 struct desc_struct *gdt = get_cpu_gdt_table(cpu);
454 xmaddr_t maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
455 struct multicall_space mc = __xen_mc_entry(0);
457 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
460 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
463 * XXX sleazy hack: If we're being called in a lazy-cpu zone
464 * and lazy gs handling is enabled, it means we're in a
465 * context switch, and %gs has just been saved. This means we
466 * can zero it out to prevent faults on exit from the
467 * hypervisor if the next process has no %gs. Either way, it
468 * has been saved, and the new value will get loaded properly.
469 * This will go away as soon as Xen has been modified to not
470 * save/restore %gs for normal hypercalls.
472 * On x86_64, this hack is not used for %gs, because gs points
473 * to KERNEL_GS_BASE (and uses it for PDA references), so we
474 * must not zero %gs on x86_64
476 * For x86_64, we need to zero %fs, otherwise we may get an
477 * exception between the new %fs descriptor being loaded and
478 * %fs being effectively cleared at __switch_to().
480 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
490 load_TLS_descriptor(t, cpu, 0);
491 load_TLS_descriptor(t, cpu, 1);
492 load_TLS_descriptor(t, cpu, 2);
494 xen_mc_issue(PARAVIRT_LAZY_CPU);
498 static void xen_load_gs_index(unsigned int idx)
500 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
505 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
508 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
509 u64 entry = *(u64 *)ptr;
511 trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
516 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
522 static int cvt_gate_to_trap(int vector, const gate_desc *val,
523 struct trap_info *info)
527 if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
530 info->vector = vector;
532 addr = gate_offset(*val);
535 * Look for known traps using IST, and substitute them
536 * appropriately. The debugger ones are the only ones we care
537 * about. Xen will handle faults like double_fault and
538 * machine_check, so we should never see them. Warn if
539 * there's an unexpected IST-using fault handler.
541 if (addr == (unsigned long)debug)
542 addr = (unsigned long)xen_debug;
543 else if (addr == (unsigned long)int3)
544 addr = (unsigned long)xen_int3;
545 else if (addr == (unsigned long)stack_segment)
546 addr = (unsigned long)xen_stack_segment;
547 else if (addr == (unsigned long)double_fault ||
548 addr == (unsigned long)nmi) {
549 /* Don't need to handle these */
551 #ifdef CONFIG_X86_MCE
552 } else if (addr == (unsigned long)machine_check) {
556 /* Some other trap using IST? */
557 if (WARN_ON(val->ist != 0))
560 #endif /* CONFIG_X86_64 */
561 info->address = addr;
563 info->cs = gate_segment(*val);
564 info->flags = val->dpl;
565 /* interrupt gates clear IF */
566 if (val->type == GATE_INTERRUPT)
567 info->flags |= 1 << 2;
572 /* Locations of each CPU's IDT */
573 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
575 /* Set an IDT entry. If the entry is part of the current IDT, then
577 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
579 unsigned long p = (unsigned long)&dt[entrynum];
580 unsigned long start, end;
582 trace_xen_cpu_write_idt_entry(dt, entrynum, g);
586 start = __this_cpu_read(idt_desc.address);
587 end = start + __this_cpu_read(idt_desc.size) + 1;
591 native_write_idt_entry(dt, entrynum, g);
593 if (p >= start && (p + 8) <= end) {
594 struct trap_info info[2];
598 if (cvt_gate_to_trap(entrynum, g, &info[0]))
599 if (HYPERVISOR_set_trap_table(info))
606 static void xen_convert_trap_info(const struct desc_ptr *desc,
607 struct trap_info *traps)
609 unsigned in, out, count;
611 count = (desc->size+1) / sizeof(gate_desc);
614 for (in = out = 0; in < count; in++) {
615 gate_desc *entry = (gate_desc*)(desc->address) + in;
617 if (cvt_gate_to_trap(in, entry, &traps[out]))
620 traps[out].address = 0;
623 void xen_copy_trap_info(struct trap_info *traps)
625 const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
627 xen_convert_trap_info(desc, traps);
630 /* Load a new IDT into Xen. In principle this can be per-CPU, so we
631 hold a spinlock to protect the static traps[] array (static because
632 it avoids allocation, and saves stack space). */
633 static void xen_load_idt(const struct desc_ptr *desc)
635 static DEFINE_SPINLOCK(lock);
636 static struct trap_info traps[257];
638 trace_xen_cpu_load_idt(desc);
642 __get_cpu_var(idt_desc) = *desc;
644 xen_convert_trap_info(desc, traps);
647 if (HYPERVISOR_set_trap_table(traps))
653 /* Write a GDT descriptor entry. Ignore LDT descriptors, since
654 they're handled differently. */
655 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
656 const void *desc, int type)
658 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
669 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
672 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
682 * Version of write_gdt_entry for use at early boot-time needed to
683 * update an entry as simply as possible.
685 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
686 const void *desc, int type)
688 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
697 xmaddr_t maddr = virt_to_machine(&dt[entry]);
699 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
700 dt[entry] = *(struct desc_struct *)desc;
706 static void xen_load_sp0(struct tss_struct *tss,
707 struct thread_struct *thread)
709 struct multicall_space mcs;
711 mcs = xen_mc_entry(0);
712 MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
713 xen_mc_issue(PARAVIRT_LAZY_CPU);
716 static void xen_set_iopl_mask(unsigned mask)
718 struct physdev_set_iopl set_iopl;
720 /* Force the change at ring 0. */
721 set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
722 HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
725 static void xen_io_delay(void)
729 #ifdef CONFIG_X86_LOCAL_APIC
730 static u32 xen_apic_read(u32 reg)
735 static void xen_apic_write(u32 reg, u32 val)
737 /* Warn to see if there's any stray references */
741 static u64 xen_apic_icr_read(void)
746 static void xen_apic_icr_write(u32 low, u32 id)
748 /* Warn to see if there's any stray references */
752 static void xen_apic_wait_icr_idle(void)
757 static u32 xen_safe_apic_wait_icr_idle(void)
762 static void set_xen_basic_apic_ops(void)
764 apic->read = xen_apic_read;
765 apic->write = xen_apic_write;
766 apic->icr_read = xen_apic_icr_read;
767 apic->icr_write = xen_apic_icr_write;
768 apic->wait_icr_idle = xen_apic_wait_icr_idle;
769 apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
774 static void xen_clts(void)
776 struct multicall_space mcs;
778 mcs = xen_mc_entry(0);
780 MULTI_fpu_taskswitch(mcs.mc, 0);
782 xen_mc_issue(PARAVIRT_LAZY_CPU);
785 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
787 static unsigned long xen_read_cr0(void)
789 unsigned long cr0 = percpu_read(xen_cr0_value);
791 if (unlikely(cr0 == 0)) {
792 cr0 = native_read_cr0();
793 percpu_write(xen_cr0_value, cr0);
799 static void xen_write_cr0(unsigned long cr0)
801 struct multicall_space mcs;
803 percpu_write(xen_cr0_value, cr0);
805 /* Only pay attention to cr0.TS; everything else is
807 mcs = xen_mc_entry(0);
809 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
811 xen_mc_issue(PARAVIRT_LAZY_CPU);
814 static void xen_write_cr4(unsigned long cr4)
819 native_write_cr4(cr4);
822 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
833 case MSR_FS_BASE: which = SEGBASE_FS; goto set;
834 case MSR_KERNEL_GS_BASE: which = SEGBASE_GS_USER; goto set;
835 case MSR_GS_BASE: which = SEGBASE_GS_KERNEL; goto set;
838 base = ((u64)high << 32) | low;
839 if (HYPERVISOR_set_segment_base(which, base) != 0)
847 case MSR_SYSCALL_MASK:
848 case MSR_IA32_SYSENTER_CS:
849 case MSR_IA32_SYSENTER_ESP:
850 case MSR_IA32_SYSENTER_EIP:
851 /* Fast syscall setup is all done in hypercalls, so
852 these are all ignored. Stub them out here to stop
853 Xen console noise. */
856 case MSR_IA32_CR_PAT:
857 if (smp_processor_id() == 0)
858 xen_set_pat(((u64)high << 32) | low);
862 ret = native_write_msr_safe(msr, low, high);
868 void xen_setup_shared_info(void)
870 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
871 set_fixmap(FIX_PARAVIRT_BOOTMAP,
872 xen_start_info->shared_info);
874 HYPERVISOR_shared_info =
875 (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
877 HYPERVISOR_shared_info =
878 (struct shared_info *)__va(xen_start_info->shared_info);
881 /* In UP this is as good a place as any to set up shared info */
882 xen_setup_vcpu_info_placement();
885 xen_setup_mfn_list_list();
888 /* This is called once we have the cpu_possible_map */
889 void xen_setup_vcpu_info_placement(void)
893 for_each_possible_cpu(cpu)
896 /* xen_vcpu_setup managed to place the vcpu_info within the
897 percpu area for all cpus, so make use of it */
898 if (have_vcpu_info_placement) {
899 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
900 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
901 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
902 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
903 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
907 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
908 unsigned long addr, unsigned len)
910 char *start, *end, *reloc;
913 start = end = reloc = NULL;
915 #define SITE(op, x) \
916 case PARAVIRT_PATCH(op.x): \
917 if (have_vcpu_info_placement) { \
918 start = (char *)xen_##x##_direct; \
919 end = xen_##x##_direct_end; \
920 reloc = xen_##x##_direct_reloc; \
925 SITE(pv_irq_ops, irq_enable);
926 SITE(pv_irq_ops, irq_disable);
927 SITE(pv_irq_ops, save_fl);
928 SITE(pv_irq_ops, restore_fl);
932 if (start == NULL || (end-start) > len)
935 ret = paravirt_patch_insns(insnbuf, len, start, end);
937 /* Note: because reloc is assigned from something that
938 appears to be an array, gcc assumes it's non-null,
939 but doesn't know its relationship with start and
941 if (reloc > start && reloc < end) {
942 int reloc_off = reloc - start;
943 long *relocp = (long *)(insnbuf + reloc_off);
944 long delta = start - (char *)addr;
952 ret = paravirt_patch_default(type, clobbers, insnbuf,
960 static const struct pv_info xen_info __initconst = {
961 .paravirt_enabled = 1,
962 .shared_kernel_pmd = 0,
965 .extra_user_64bit_cs = FLAT_USER_CS64,
971 static const struct pv_init_ops xen_init_ops __initconst = {
975 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
978 .set_debugreg = xen_set_debugreg,
979 .get_debugreg = xen_get_debugreg,
983 .read_cr0 = xen_read_cr0,
984 .write_cr0 = xen_write_cr0,
986 .read_cr4 = native_read_cr4,
987 .read_cr4_safe = native_read_cr4_safe,
988 .write_cr4 = xen_write_cr4,
990 .wbinvd = native_wbinvd,
992 .read_msr = native_read_msr_safe,
993 .rdmsr_regs = native_rdmsr_safe_regs,
994 .write_msr = xen_write_msr_safe,
995 .wrmsr_regs = native_wrmsr_safe_regs,
997 .read_tsc = native_read_tsc,
998 .read_pmc = native_read_pmc,
1001 .irq_enable_sysexit = xen_sysexit,
1002 #ifdef CONFIG_X86_64
1003 .usergs_sysret32 = xen_sysret32,
1004 .usergs_sysret64 = xen_sysret64,
1007 .load_tr_desc = paravirt_nop,
1008 .set_ldt = xen_set_ldt,
1009 .load_gdt = xen_load_gdt,
1010 .load_idt = xen_load_idt,
1011 .load_tls = xen_load_tls,
1012 #ifdef CONFIG_X86_64
1013 .load_gs_index = xen_load_gs_index,
1016 .alloc_ldt = xen_alloc_ldt,
1017 .free_ldt = xen_free_ldt,
1019 .store_gdt = native_store_gdt,
1020 .store_idt = native_store_idt,
1021 .store_tr = xen_store_tr,
1023 .write_ldt_entry = xen_write_ldt_entry,
1024 .write_gdt_entry = xen_write_gdt_entry,
1025 .write_idt_entry = xen_write_idt_entry,
1026 .load_sp0 = xen_load_sp0,
1028 .set_iopl_mask = xen_set_iopl_mask,
1029 .io_delay = xen_io_delay,
1031 /* Xen takes care of %gs when switching to usermode for us */
1032 .swapgs = paravirt_nop,
1034 .start_context_switch = paravirt_start_context_switch,
1035 .end_context_switch = xen_end_context_switch,
1038 static const struct pv_apic_ops xen_apic_ops __initconst = {
1039 #ifdef CONFIG_X86_LOCAL_APIC
1040 .startup_ipi_hook = paravirt_nop,
1044 static void xen_reboot(int reason)
1046 struct sched_shutdown r = { .reason = reason };
1048 if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1052 static void xen_restart(char *msg)
1054 xen_reboot(SHUTDOWN_reboot);
1057 static void xen_emergency_restart(void)
1059 xen_reboot(SHUTDOWN_reboot);
1062 static void xen_machine_halt(void)
1064 xen_reboot(SHUTDOWN_poweroff);
1067 static void xen_machine_power_off(void)
1071 xen_reboot(SHUTDOWN_poweroff);
1074 static void xen_crash_shutdown(struct pt_regs *regs)
1076 xen_reboot(SHUTDOWN_crash);
1080 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1082 xen_reboot(SHUTDOWN_crash);
1086 static struct notifier_block xen_panic_block = {
1087 .notifier_call= xen_panic_event,
1090 int xen_panic_handler_init(void)
1092 atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1096 static const struct machine_ops xen_machine_ops __initconst = {
1097 .restart = xen_restart,
1098 .halt = xen_machine_halt,
1099 .power_off = xen_machine_power_off,
1100 .shutdown = xen_machine_halt,
1101 .crash_shutdown = xen_crash_shutdown,
1102 .emergency_restart = xen_emergency_restart,
1106 * Set up the GDT and segment registers for -fstack-protector. Until
1107 * we do this, we have to be careful not to call any stack-protected
1108 * function, which is most of the kernel.
1110 static void __init xen_setup_stackprotector(void)
1112 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1113 pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1115 setup_stack_canary_segment(0);
1116 switch_to_new_gdt(0);
1118 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1119 pv_cpu_ops.load_gdt = xen_load_gdt;
1122 /* First C function to be called on Xen boot */
1123 asmlinkage void __init xen_start_kernel(void)
1125 struct physdev_set_iopl set_iopl;
1129 if (!xen_start_info)
1132 xen_domain_type = XEN_PV_DOMAIN;
1134 xen_setup_machphys_mapping();
1136 /* Install Xen paravirt ops */
1138 pv_init_ops = xen_init_ops;
1139 pv_cpu_ops = xen_cpu_ops;
1140 pv_apic_ops = xen_apic_ops;
1142 x86_init.resources.memory_setup = xen_memory_setup;
1143 x86_init.oem.arch_setup = xen_arch_setup;
1144 x86_init.oem.banner = xen_banner;
1146 xen_init_time_ops();
1149 * Set up some pagetable state before starting to set any ptes.
1154 /* Prevent unwanted bits from being set in PTEs. */
1155 __supported_pte_mask &= ~_PAGE_GLOBAL;
1156 if (!xen_initial_domain())
1157 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1159 __supported_pte_mask |= _PAGE_IOMAP;
1162 * Prevent page tables from being allocated in highmem, even
1163 * if CONFIG_HIGHPTE is enabled.
1165 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1167 /* Work out if we support NX */
1170 xen_setup_features();
1173 if (!xen_feature(XENFEAT_auto_translated_physmap))
1174 xen_build_dynamic_phys_to_machine();
1177 * Set up kernel GDT and segment registers, mainly so that
1178 * -fstack-protector code can be executed.
1180 xen_setup_stackprotector();
1183 xen_init_cpuid_mask();
1185 #ifdef CONFIG_X86_LOCAL_APIC
1187 * set up the basic apic ops.
1189 set_xen_basic_apic_ops();
1192 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1193 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1194 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1197 machine_ops = xen_machine_ops;
1200 * The only reliable way to retain the initial address of the
1201 * percpu gdt_page is to remember it here, so we can go and
1202 * mark it RW later, when the initial percpu area is freed.
1204 xen_initial_gdt = &per_cpu(gdt_page, 0);
1208 #ifdef CONFIG_ACPI_NUMA
1210 * The pages we from Xen are not related to machine pages, so
1211 * any NUMA information the kernel tries to get from ACPI will
1212 * be meaningless. Prevent it from trying.
1217 pgd = (pgd_t *)xen_start_info->pt_base;
1219 if (!xen_initial_domain())
1220 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1222 __supported_pte_mask |= _PAGE_IOMAP;
1223 /* Don't do the full vcpu_info placement stuff until we have a
1224 possible map and a non-dummy shared_info. */
1225 per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1227 local_irq_disable();
1228 early_boot_irqs_disabled = true;
1232 xen_raw_console_write("mapping kernel into physical memory\n");
1233 pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
1234 xen_ident_map_ISA();
1236 /* Allocate and initialize top and mid mfn levels for p2m structure */
1237 xen_build_mfn_list_list();
1239 /* keep using Xen gdt for now; no urgent need to change it */
1241 #ifdef CONFIG_X86_32
1242 pv_info.kernel_rpl = 1;
1243 if (xen_feature(XENFEAT_supervisor_mode_kernel))
1244 pv_info.kernel_rpl = 0;
1246 pv_info.kernel_rpl = 0;
1248 /* set the limit of our address space */
1251 /* We used to do this in xen_arch_setup, but that is too late on AMD
1252 * were early_cpu_init (run before ->arch_setup()) calls early_amd_init
1253 * which pokes 0xcf8 port.
1256 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1258 xen_raw_printk("physdev_op failed %d\n", rc);
1260 #ifdef CONFIG_X86_32
1261 /* set up basic CPUID stuff */
1262 cpu_detect(&new_cpu_data);
1263 new_cpu_data.hard_math = 1;
1264 new_cpu_data.wp_works_ok = 1;
1265 new_cpu_data.x86_capability[0] = cpuid_edx(1);
1268 /* Poke various useful things into boot_params */
1269 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1270 boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1271 ? __pa(xen_start_info->mod_start) : 0;
1272 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1273 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1275 if (!xen_initial_domain()) {
1276 add_preferred_console("xenboot", 0, NULL);
1277 add_preferred_console("tty", 0, NULL);
1278 add_preferred_console("hvc", 0, NULL);
1280 x86_init.pci.arch_init = pci_xen_init;
1282 const struct dom0_vga_console_info *info =
1283 (void *)((char *)xen_start_info +
1284 xen_start_info->console.dom0.info_off);
1286 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1287 xen_start_info->console.domU.mfn = 0;
1288 xen_start_info->console.domU.evtchn = 0;
1290 /* Make sure ACS will be enabled */
1294 /* PCI BIOS service won't work from a PV guest. */
1295 pci_probe &= ~PCI_PROBE_BIOS;
1297 xen_raw_console_write("about to get started...\n");
1299 xen_setup_runstate_info(0);
1301 /* Start the world */
1302 #ifdef CONFIG_X86_32
1303 i386_start_kernel();
1305 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1309 static int init_hvm_pv_info(int *major, int *minor)
1311 uint32_t eax, ebx, ecx, edx, pages, msr, base;
1314 base = xen_cpuid_base();
1315 cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1318 *minor = eax & 0xffff;
1319 printk(KERN_INFO "Xen version %d.%d.\n", *major, *minor);
1321 cpuid(base + 2, &pages, &msr, &ecx, &edx);
1323 pfn = __pa(hypercall_page);
1324 wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1326 xen_setup_features();
1328 pv_info.name = "Xen HVM";
1330 xen_domain_type = XEN_HVM_DOMAIN;
1335 void __ref xen_hvm_init_shared_info(void)
1338 struct xen_add_to_physmap xatp;
1339 static struct shared_info *shared_info_page = 0;
1341 if (!shared_info_page)
1342 shared_info_page = (struct shared_info *)
1343 extend_brk(PAGE_SIZE, PAGE_SIZE);
1344 xatp.domid = DOMID_SELF;
1346 xatp.space = XENMAPSPACE_shared_info;
1347 xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT;
1348 if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1351 HYPERVISOR_shared_info = (struct shared_info *)shared_info_page;
1353 /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1354 * page, we use it in the event channel upcall and in some pvclock
1355 * related functions. We don't need the vcpu_info placement
1356 * optimizations because we don't use any pv_mmu or pv_irq op on
1358 * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is
1359 * online but xen_hvm_init_shared_info is run at resume time too and
1360 * in that case multiple vcpus might be online. */
1361 for_each_online_cpu(cpu) {
1362 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1366 #ifdef CONFIG_XEN_PVHVM
1367 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self,
1368 unsigned long action, void *hcpu)
1370 int cpu = (long)hcpu;
1372 case CPU_UP_PREPARE:
1373 xen_vcpu_setup(cpu);
1374 if (xen_have_vector_callback)
1375 xen_init_lock_cpu(cpu);
1383 static struct notifier_block xen_hvm_cpu_notifier __cpuinitdata = {
1384 .notifier_call = xen_hvm_cpu_notify,
1387 static void __init xen_hvm_guest_init(void)
1392 r = init_hvm_pv_info(&major, &minor);
1396 xen_hvm_init_shared_info();
1398 if (xen_feature(XENFEAT_hvm_callback_vector))
1399 xen_have_vector_callback = 1;
1401 register_cpu_notifier(&xen_hvm_cpu_notifier);
1402 xen_unplug_emulated_devices();
1403 x86_init.irqs.intr_init = xen_init_IRQ;
1404 xen_hvm_init_time_ops();
1405 xen_hvm_init_mmu_ops();
1408 static bool __init xen_hvm_platform(void)
1410 if (xen_pv_domain())
1413 if (!xen_cpuid_base())
1419 bool xen_hvm_need_lapic(void)
1421 if (xen_pv_domain())
1423 if (!xen_hvm_domain())
1425 if (xen_feature(XENFEAT_hvm_pirqs) && xen_have_vector_callback)
1429 EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1431 const struct hypervisor_x86 x86_hyper_xen_hvm __refconst = {
1433 .detect = xen_hvm_platform,
1434 .init_platform = xen_hvm_guest_init,
1436 EXPORT_SYMBOL(x86_hyper_xen_hvm);