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/events.h>
37 #include <xen/interface/xen.h>
38 #include <xen/interface/version.h>
39 #include <xen/interface/physdev.h>
40 #include <xen/interface/vcpu.h>
41 #include <xen/interface/memory.h>
42 #include <xen/interface/xen-mca.h>
43 #include <xen/features.h>
46 #include <xen/hvc-console.h>
49 #include <asm/paravirt.h>
52 #include <asm/xen/pci.h>
53 #include <asm/xen/hypercall.h>
54 #include <asm/xen/hypervisor.h>
55 #include <asm/fixmap.h>
56 #include <asm/processor.h>
57 #include <asm/proto.h>
58 #include <asm/msr-index.h>
59 #include <asm/traps.h>
60 #include <asm/setup.h>
62 #include <asm/pgalloc.h>
63 #include <asm/pgtable.h>
64 #include <asm/tlbflush.h>
65 #include <asm/reboot.h>
66 #include <asm/stackprotector.h>
67 #include <asm/hypervisor.h>
68 #include <asm/mwait.h>
69 #include <asm/pci_x86.h>
72 #include <linux/acpi.h>
74 #include <acpi/pdc_intel.h>
75 #include <acpi/processor.h>
76 #include <xen/interface/platform.h>
82 #include "multicalls.h"
84 #include <xen/events.h>
86 EXPORT_SYMBOL_GPL(hypercall_page);
88 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
89 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
91 enum xen_domain_type xen_domain_type = XEN_NATIVE;
92 EXPORT_SYMBOL_GPL(xen_domain_type);
94 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
95 EXPORT_SYMBOL(machine_to_phys_mapping);
96 unsigned long machine_to_phys_nr;
97 EXPORT_SYMBOL(machine_to_phys_nr);
99 struct start_info *xen_start_info;
100 EXPORT_SYMBOL_GPL(xen_start_info);
102 struct shared_info xen_dummy_shared_info;
104 void *xen_initial_gdt;
106 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
107 __read_mostly int xen_have_vector_callback;
108 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
111 * Point at some empty memory to start with. We map the real shared_info
112 * page as soon as fixmap is up and running.
114 struct shared_info *HYPERVISOR_shared_info = &xen_dummy_shared_info;
117 * Flag to determine whether vcpu info placement is available on all
118 * VCPUs. We assume it is to start with, and then set it to zero on
119 * the first failure. This is because it can succeed on some VCPUs
120 * and not others, since it can involve hypervisor memory allocation,
121 * or because the guest failed to guarantee all the appropriate
122 * constraints on all VCPUs (ie buffer can't cross a page boundary).
124 * Note that any particular CPU may be using a placed vcpu structure,
125 * but we can only optimise if the all are.
127 * 0: not available, 1: available
129 static int have_vcpu_info_placement = 1;
132 struct desc_struct desc[3];
136 * Updating the 3 TLS descriptors in the GDT on every task switch is
137 * surprisingly expensive so we avoid updating them if they haven't
138 * changed. Since Xen writes different descriptors than the one
139 * passed in the update_descriptor hypercall we keep shadow copies to
142 static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
144 static void clamp_max_cpus(void)
147 if (setup_max_cpus > MAX_VIRT_CPUS)
148 setup_max_cpus = MAX_VIRT_CPUS;
152 static void xen_vcpu_setup(int cpu)
154 struct vcpu_register_vcpu_info info;
156 struct vcpu_info *vcpup;
158 BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
160 if (cpu < MAX_VIRT_CPUS)
161 per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
163 if (!have_vcpu_info_placement) {
164 if (cpu >= MAX_VIRT_CPUS)
169 vcpup = &per_cpu(xen_vcpu_info, cpu);
170 info.mfn = arbitrary_virt_to_mfn(vcpup);
171 info.offset = offset_in_page(vcpup);
173 /* Check to see if the hypervisor will put the vcpu_info
174 structure where we want it, which allows direct access via
175 a percpu-variable. */
176 err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
179 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
180 have_vcpu_info_placement = 0;
183 /* This cpu is using the registered vcpu info, even if
184 later ones fail to. */
185 per_cpu(xen_vcpu, cpu) = vcpup;
190 * On restore, set the vcpu placement up again.
191 * If it fails, then we're in a bad state, since
192 * we can't back out from using it...
194 void xen_vcpu_restore(void)
198 for_each_online_cpu(cpu) {
199 bool other_cpu = (cpu != smp_processor_id());
202 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
205 xen_setup_runstate_info(cpu);
207 if (have_vcpu_info_placement)
211 HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
216 static void __init xen_banner(void)
218 unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
219 struct xen_extraversion extra;
220 HYPERVISOR_xen_version(XENVER_extraversion, &extra);
222 printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
224 printk(KERN_INFO "Xen version: %d.%d%s%s\n",
225 version >> 16, version & 0xffff, extra.extraversion,
226 xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
229 #define CPUID_THERM_POWER_LEAF 6
230 #define APERFMPERF_PRESENT 0
232 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
233 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
235 static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
236 static __read_mostly unsigned int cpuid_leaf5_ecx_val;
237 static __read_mostly unsigned int cpuid_leaf5_edx_val;
239 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
240 unsigned int *cx, unsigned int *dx)
242 unsigned maskebx = ~0;
243 unsigned maskecx = ~0;
244 unsigned maskedx = ~0;
247 * Mask out inconvenient features, to try and disable as many
248 * unsupported kernel subsystems as possible.
252 maskecx = cpuid_leaf1_ecx_mask;
253 setecx = cpuid_leaf1_ecx_set_mask;
254 maskedx = cpuid_leaf1_edx_mask;
257 case CPUID_MWAIT_LEAF:
258 /* Synthesize the values.. */
261 *cx = cpuid_leaf5_ecx_val;
262 *dx = cpuid_leaf5_edx_val;
265 case CPUID_THERM_POWER_LEAF:
266 /* Disabling APERFMPERF for kernel usage */
267 maskecx = ~(1 << APERFMPERF_PRESENT);
271 /* Suppress extended topology stuff */
276 asm(XEN_EMULATE_PREFIX "cpuid"
281 : "0" (*ax), "2" (*cx));
290 static bool __init xen_check_mwait(void)
292 #if defined(CONFIG_ACPI) && !defined(CONFIG_ACPI_PROCESSOR_AGGREGATOR) && \
293 !defined(CONFIG_ACPI_PROCESSOR_AGGREGATOR_MODULE)
294 struct xen_platform_op op = {
295 .cmd = XENPF_set_processor_pminfo,
296 .u.set_pminfo.id = -1,
297 .u.set_pminfo.type = XEN_PM_PDC,
300 unsigned int ax, bx, cx, dx;
301 unsigned int mwait_mask;
303 /* We need to determine whether it is OK to expose the MWAIT
304 * capability to the kernel to harvest deeper than C3 states from ACPI
305 * _CST using the processor_harvest_xen.c module. For this to work, we
306 * need to gather the MWAIT_LEAF values (which the cstate.c code
307 * checks against). The hypervisor won't expose the MWAIT flag because
308 * it would break backwards compatibility; so we will find out directly
309 * from the hardware and hypercall.
311 if (!xen_initial_domain())
317 native_cpuid(&ax, &bx, &cx, &dx);
319 mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
320 (1 << (X86_FEATURE_MWAIT % 32));
322 if ((cx & mwait_mask) != mwait_mask)
325 /* We need to emulate the MWAIT_LEAF and for that we need both
326 * ecx and edx. The hypercall provides only partial information.
329 ax = CPUID_MWAIT_LEAF;
334 native_cpuid(&ax, &bx, &cx, &dx);
336 /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
337 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
339 buf[0] = ACPI_PDC_REVISION_ID;
341 buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
343 set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
345 if ((HYPERVISOR_dom0_op(&op) == 0) &&
346 (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
347 cpuid_leaf5_ecx_val = cx;
348 cpuid_leaf5_edx_val = dx;
355 static void __init xen_init_cpuid_mask(void)
357 unsigned int ax, bx, cx, dx;
358 unsigned int xsave_mask;
360 cpuid_leaf1_edx_mask =
361 ~((1 << X86_FEATURE_MTRR) | /* disable MTRR */
362 (1 << X86_FEATURE_ACC)); /* thermal monitoring */
364 if (!xen_initial_domain())
365 cpuid_leaf1_edx_mask &=
366 ~((1 << X86_FEATURE_APIC) | /* disable local APIC */
367 (1 << X86_FEATURE_ACPI)); /* disable ACPI */
370 xen_cpuid(&ax, &bx, &cx, &dx);
373 (1 << (X86_FEATURE_XSAVE % 32)) |
374 (1 << (X86_FEATURE_OSXSAVE % 32));
376 /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
377 if ((cx & xsave_mask) != xsave_mask)
378 cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
379 if (xen_check_mwait())
380 cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
383 static void xen_set_debugreg(int reg, unsigned long val)
385 HYPERVISOR_set_debugreg(reg, val);
388 static unsigned long xen_get_debugreg(int reg)
390 return HYPERVISOR_get_debugreg(reg);
393 static void xen_end_context_switch(struct task_struct *next)
396 paravirt_end_context_switch(next);
399 static unsigned long xen_store_tr(void)
405 * Set the page permissions for a particular virtual address. If the
406 * address is a vmalloc mapping (or other non-linear mapping), then
407 * find the linear mapping of the page and also set its protections to
410 static void set_aliased_prot(void *v, pgprot_t prot)
418 ptep = lookup_address((unsigned long)v, &level);
419 BUG_ON(ptep == NULL);
421 pfn = pte_pfn(*ptep);
422 page = pfn_to_page(pfn);
424 pte = pfn_pte(pfn, prot);
426 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
429 if (!PageHighMem(page)) {
430 void *av = __va(PFN_PHYS(pfn));
433 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
439 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
441 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
444 for(i = 0; i < entries; i += entries_per_page)
445 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
448 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
450 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
453 for(i = 0; i < entries; i += entries_per_page)
454 set_aliased_prot(ldt + i, PAGE_KERNEL);
457 static void xen_set_ldt(const void *addr, unsigned entries)
459 struct mmuext_op *op;
460 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
462 trace_xen_cpu_set_ldt(addr, entries);
465 op->cmd = MMUEXT_SET_LDT;
466 op->arg1.linear_addr = (unsigned long)addr;
467 op->arg2.nr_ents = entries;
469 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
471 xen_mc_issue(PARAVIRT_LAZY_CPU);
474 static void xen_load_gdt(const struct desc_ptr *dtr)
476 unsigned long va = dtr->address;
477 unsigned int size = dtr->size + 1;
478 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
479 unsigned long frames[pages];
483 * A GDT can be up to 64k in size, which corresponds to 8192
484 * 8-byte entries, or 16 4k pages..
487 BUG_ON(size > 65536);
488 BUG_ON(va & ~PAGE_MASK);
490 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
493 unsigned long pfn, mfn;
497 * The GDT is per-cpu and is in the percpu data area.
498 * That can be virtually mapped, so we need to do a
499 * page-walk to get the underlying MFN for the
500 * hypercall. The page can also be in the kernel's
501 * linear range, so we need to RO that mapping too.
503 ptep = lookup_address(va, &level);
504 BUG_ON(ptep == NULL);
506 pfn = pte_pfn(*ptep);
507 mfn = pfn_to_mfn(pfn);
508 virt = __va(PFN_PHYS(pfn));
512 make_lowmem_page_readonly((void *)va);
513 make_lowmem_page_readonly(virt);
516 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
521 * load_gdt for early boot, when the gdt is only mapped once
523 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
525 unsigned long va = dtr->address;
526 unsigned int size = dtr->size + 1;
527 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
528 unsigned long frames[pages];
532 * A GDT can be up to 64k in size, which corresponds to 8192
533 * 8-byte entries, or 16 4k pages..
536 BUG_ON(size > 65536);
537 BUG_ON(va & ~PAGE_MASK);
539 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
541 unsigned long pfn, mfn;
543 pfn = virt_to_pfn(va);
544 mfn = pfn_to_mfn(pfn);
546 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
548 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
554 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
558 static inline bool desc_equal(const struct desc_struct *d1,
559 const struct desc_struct *d2)
561 return d1->a == d2->a && d1->b == d2->b;
564 static void load_TLS_descriptor(struct thread_struct *t,
565 unsigned int cpu, unsigned int i)
567 struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
568 struct desc_struct *gdt;
570 struct multicall_space mc;
572 if (desc_equal(shadow, &t->tls_array[i]))
575 *shadow = t->tls_array[i];
577 gdt = get_cpu_gdt_table(cpu);
578 maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
579 mc = __xen_mc_entry(0);
581 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
584 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
587 * XXX sleazy hack: If we're being called in a lazy-cpu zone
588 * and lazy gs handling is enabled, it means we're in a
589 * context switch, and %gs has just been saved. This means we
590 * can zero it out to prevent faults on exit from the
591 * hypervisor if the next process has no %gs. Either way, it
592 * has been saved, and the new value will get loaded properly.
593 * This will go away as soon as Xen has been modified to not
594 * save/restore %gs for normal hypercalls.
596 * On x86_64, this hack is not used for %gs, because gs points
597 * to KERNEL_GS_BASE (and uses it for PDA references), so we
598 * must not zero %gs on x86_64
600 * For x86_64, we need to zero %fs, otherwise we may get an
601 * exception between the new %fs descriptor being loaded and
602 * %fs being effectively cleared at __switch_to().
604 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
614 load_TLS_descriptor(t, cpu, 0);
615 load_TLS_descriptor(t, cpu, 1);
616 load_TLS_descriptor(t, cpu, 2);
618 xen_mc_issue(PARAVIRT_LAZY_CPU);
622 static void xen_load_gs_index(unsigned int idx)
624 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
629 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
632 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
633 u64 entry = *(u64 *)ptr;
635 trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
640 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
646 static int cvt_gate_to_trap(int vector, const gate_desc *val,
647 struct trap_info *info)
651 if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
654 info->vector = vector;
656 addr = gate_offset(*val);
659 * Look for known traps using IST, and substitute them
660 * appropriately. The debugger ones are the only ones we care
661 * about. Xen will handle faults like double_fault,
662 * so we should never see them. Warn if
663 * there's an unexpected IST-using fault handler.
665 if (addr == (unsigned long)debug)
666 addr = (unsigned long)xen_debug;
667 else if (addr == (unsigned long)int3)
668 addr = (unsigned long)xen_int3;
669 else if (addr == (unsigned long)stack_segment)
670 addr = (unsigned long)xen_stack_segment;
671 else if (addr == (unsigned long)double_fault ||
672 addr == (unsigned long)nmi) {
673 /* Don't need to handle these */
675 #ifdef CONFIG_X86_MCE
676 } else if (addr == (unsigned long)machine_check) {
678 * when xen hypervisor inject vMCE to guest,
679 * use native mce handler to handle it
684 /* Some other trap using IST? */
685 if (WARN_ON(val->ist != 0))
688 #endif /* CONFIG_X86_64 */
689 info->address = addr;
691 info->cs = gate_segment(*val);
692 info->flags = val->dpl;
693 /* interrupt gates clear IF */
694 if (val->type == GATE_INTERRUPT)
695 info->flags |= 1 << 2;
700 /* Locations of each CPU's IDT */
701 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
703 /* Set an IDT entry. If the entry is part of the current IDT, then
705 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
707 unsigned long p = (unsigned long)&dt[entrynum];
708 unsigned long start, end;
710 trace_xen_cpu_write_idt_entry(dt, entrynum, g);
714 start = __this_cpu_read(idt_desc.address);
715 end = start + __this_cpu_read(idt_desc.size) + 1;
719 native_write_idt_entry(dt, entrynum, g);
721 if (p >= start && (p + 8) <= end) {
722 struct trap_info info[2];
726 if (cvt_gate_to_trap(entrynum, g, &info[0]))
727 if (HYPERVISOR_set_trap_table(info))
734 static void xen_convert_trap_info(const struct desc_ptr *desc,
735 struct trap_info *traps)
737 unsigned in, out, count;
739 count = (desc->size+1) / sizeof(gate_desc);
742 for (in = out = 0; in < count; in++) {
743 gate_desc *entry = (gate_desc*)(desc->address) + in;
745 if (cvt_gate_to_trap(in, entry, &traps[out]))
748 traps[out].address = 0;
751 void xen_copy_trap_info(struct trap_info *traps)
753 const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
755 xen_convert_trap_info(desc, traps);
758 /* Load a new IDT into Xen. In principle this can be per-CPU, so we
759 hold a spinlock to protect the static traps[] array (static because
760 it avoids allocation, and saves stack space). */
761 static void xen_load_idt(const struct desc_ptr *desc)
763 static DEFINE_SPINLOCK(lock);
764 static struct trap_info traps[257];
766 trace_xen_cpu_load_idt(desc);
770 __get_cpu_var(idt_desc) = *desc;
772 xen_convert_trap_info(desc, traps);
775 if (HYPERVISOR_set_trap_table(traps))
781 /* Write a GDT descriptor entry. Ignore LDT descriptors, since
782 they're handled differently. */
783 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
784 const void *desc, int type)
786 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
797 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
800 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
810 * Version of write_gdt_entry for use at early boot-time needed to
811 * update an entry as simply as possible.
813 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
814 const void *desc, int type)
816 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
825 xmaddr_t maddr = virt_to_machine(&dt[entry]);
827 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
828 dt[entry] = *(struct desc_struct *)desc;
834 static void xen_load_sp0(struct tss_struct *tss,
835 struct thread_struct *thread)
837 struct multicall_space mcs;
839 mcs = xen_mc_entry(0);
840 MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
841 xen_mc_issue(PARAVIRT_LAZY_CPU);
844 static void xen_set_iopl_mask(unsigned mask)
846 struct physdev_set_iopl set_iopl;
848 /* Force the change at ring 0. */
849 set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
850 HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
853 static void xen_io_delay(void)
857 #ifdef CONFIG_X86_LOCAL_APIC
858 static unsigned long xen_set_apic_id(unsigned int x)
863 static unsigned int xen_get_apic_id(unsigned long x)
865 return ((x)>>24) & 0xFFu;
867 static u32 xen_apic_read(u32 reg)
869 struct xen_platform_op op = {
870 .cmd = XENPF_get_cpuinfo,
871 .interface_version = XENPF_INTERFACE_VERSION,
872 .u.pcpu_info.xen_cpuid = 0,
876 /* Shouldn't need this as APIC is turned off for PV, and we only
877 * get called on the bootup processor. But just in case. */
878 if (!xen_initial_domain() || smp_processor_id())
887 ret = HYPERVISOR_dom0_op(&op);
891 return op.u.pcpu_info.apic_id << 24;
894 static void xen_apic_write(u32 reg, u32 val)
896 /* Warn to see if there's any stray references */
900 static u64 xen_apic_icr_read(void)
905 static void xen_apic_icr_write(u32 low, u32 id)
907 /* Warn to see if there's any stray references */
911 static void xen_apic_wait_icr_idle(void)
916 static u32 xen_safe_apic_wait_icr_idle(void)
921 static void set_xen_basic_apic_ops(void)
923 apic->read = xen_apic_read;
924 apic->write = xen_apic_write;
925 apic->icr_read = xen_apic_icr_read;
926 apic->icr_write = xen_apic_icr_write;
927 apic->wait_icr_idle = xen_apic_wait_icr_idle;
928 apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
929 apic->set_apic_id = xen_set_apic_id;
930 apic->get_apic_id = xen_get_apic_id;
933 apic->send_IPI_allbutself = xen_send_IPI_allbutself;
934 apic->send_IPI_mask_allbutself = xen_send_IPI_mask_allbutself;
935 apic->send_IPI_mask = xen_send_IPI_mask;
936 apic->send_IPI_all = xen_send_IPI_all;
937 apic->send_IPI_self = xen_send_IPI_self;
943 static void xen_clts(void)
945 struct multicall_space mcs;
947 mcs = xen_mc_entry(0);
949 MULTI_fpu_taskswitch(mcs.mc, 0);
951 xen_mc_issue(PARAVIRT_LAZY_CPU);
954 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
956 static unsigned long xen_read_cr0(void)
958 unsigned long cr0 = this_cpu_read(xen_cr0_value);
960 if (unlikely(cr0 == 0)) {
961 cr0 = native_read_cr0();
962 this_cpu_write(xen_cr0_value, cr0);
968 static void xen_write_cr0(unsigned long cr0)
970 struct multicall_space mcs;
972 this_cpu_write(xen_cr0_value, cr0);
974 /* Only pay attention to cr0.TS; everything else is
976 mcs = xen_mc_entry(0);
978 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
980 xen_mc_issue(PARAVIRT_LAZY_CPU);
983 static void xen_write_cr4(unsigned long cr4)
988 native_write_cr4(cr4);
991 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
1002 case MSR_FS_BASE: which = SEGBASE_FS; goto set;
1003 case MSR_KERNEL_GS_BASE: which = SEGBASE_GS_USER; goto set;
1004 case MSR_GS_BASE: which = SEGBASE_GS_KERNEL; goto set;
1007 base = ((u64)high << 32) | low;
1008 if (HYPERVISOR_set_segment_base(which, base) != 0)
1016 case MSR_SYSCALL_MASK:
1017 case MSR_IA32_SYSENTER_CS:
1018 case MSR_IA32_SYSENTER_ESP:
1019 case MSR_IA32_SYSENTER_EIP:
1020 /* Fast syscall setup is all done in hypercalls, so
1021 these are all ignored. Stub them out here to stop
1022 Xen console noise. */
1025 case MSR_IA32_CR_PAT:
1026 if (smp_processor_id() == 0)
1027 xen_set_pat(((u64)high << 32) | low);
1031 ret = native_write_msr_safe(msr, low, high);
1037 void xen_setup_shared_info(void)
1039 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1040 set_fixmap(FIX_PARAVIRT_BOOTMAP,
1041 xen_start_info->shared_info);
1043 HYPERVISOR_shared_info =
1044 (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
1046 HYPERVISOR_shared_info =
1047 (struct shared_info *)__va(xen_start_info->shared_info);
1050 /* In UP this is as good a place as any to set up shared info */
1051 xen_setup_vcpu_info_placement();
1054 xen_setup_mfn_list_list();
1057 /* This is called once we have the cpu_possible_mask */
1058 void xen_setup_vcpu_info_placement(void)
1062 for_each_possible_cpu(cpu)
1063 xen_vcpu_setup(cpu);
1065 /* xen_vcpu_setup managed to place the vcpu_info within the
1066 percpu area for all cpus, so make use of it */
1067 if (have_vcpu_info_placement) {
1068 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1069 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1070 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1071 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1072 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1076 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1077 unsigned long addr, unsigned len)
1079 char *start, *end, *reloc;
1082 start = end = reloc = NULL;
1084 #define SITE(op, x) \
1085 case PARAVIRT_PATCH(op.x): \
1086 if (have_vcpu_info_placement) { \
1087 start = (char *)xen_##x##_direct; \
1088 end = xen_##x##_direct_end; \
1089 reloc = xen_##x##_direct_reloc; \
1094 SITE(pv_irq_ops, irq_enable);
1095 SITE(pv_irq_ops, irq_disable);
1096 SITE(pv_irq_ops, save_fl);
1097 SITE(pv_irq_ops, restore_fl);
1101 if (start == NULL || (end-start) > len)
1104 ret = paravirt_patch_insns(insnbuf, len, start, end);
1106 /* Note: because reloc is assigned from something that
1107 appears to be an array, gcc assumes it's non-null,
1108 but doesn't know its relationship with start and
1110 if (reloc > start && reloc < end) {
1111 int reloc_off = reloc - start;
1112 long *relocp = (long *)(insnbuf + reloc_off);
1113 long delta = start - (char *)addr;
1121 ret = paravirt_patch_default(type, clobbers, insnbuf,
1129 static const struct pv_info xen_info __initconst = {
1130 .paravirt_enabled = 1,
1131 .shared_kernel_pmd = 0,
1133 #ifdef CONFIG_X86_64
1134 .extra_user_64bit_cs = FLAT_USER_CS64,
1140 static const struct pv_init_ops xen_init_ops __initconst = {
1144 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1147 .set_debugreg = xen_set_debugreg,
1148 .get_debugreg = xen_get_debugreg,
1152 .read_cr0 = xen_read_cr0,
1153 .write_cr0 = xen_write_cr0,
1155 .read_cr4 = native_read_cr4,
1156 .read_cr4_safe = native_read_cr4_safe,
1157 .write_cr4 = xen_write_cr4,
1159 .wbinvd = native_wbinvd,
1161 .read_msr = native_read_msr_safe,
1162 .write_msr = xen_write_msr_safe,
1164 .read_tsc = native_read_tsc,
1165 .read_pmc = native_read_pmc,
1168 .irq_enable_sysexit = xen_sysexit,
1169 #ifdef CONFIG_X86_64
1170 .usergs_sysret32 = xen_sysret32,
1171 .usergs_sysret64 = xen_sysret64,
1174 .load_tr_desc = paravirt_nop,
1175 .set_ldt = xen_set_ldt,
1176 .load_gdt = xen_load_gdt,
1177 .load_idt = xen_load_idt,
1178 .load_tls = xen_load_tls,
1179 #ifdef CONFIG_X86_64
1180 .load_gs_index = xen_load_gs_index,
1183 .alloc_ldt = xen_alloc_ldt,
1184 .free_ldt = xen_free_ldt,
1186 .store_gdt = native_store_gdt,
1187 .store_idt = native_store_idt,
1188 .store_tr = xen_store_tr,
1190 .write_ldt_entry = xen_write_ldt_entry,
1191 .write_gdt_entry = xen_write_gdt_entry,
1192 .write_idt_entry = xen_write_idt_entry,
1193 .load_sp0 = xen_load_sp0,
1195 .set_iopl_mask = xen_set_iopl_mask,
1196 .io_delay = xen_io_delay,
1198 /* Xen takes care of %gs when switching to usermode for us */
1199 .swapgs = paravirt_nop,
1201 .start_context_switch = paravirt_start_context_switch,
1202 .end_context_switch = xen_end_context_switch,
1205 static const struct pv_apic_ops xen_apic_ops __initconst = {
1206 #ifdef CONFIG_X86_LOCAL_APIC
1207 .startup_ipi_hook = paravirt_nop,
1211 static void xen_reboot(int reason)
1213 struct sched_shutdown r = { .reason = reason };
1215 if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1219 static void xen_restart(char *msg)
1221 xen_reboot(SHUTDOWN_reboot);
1224 static void xen_emergency_restart(void)
1226 xen_reboot(SHUTDOWN_reboot);
1229 static void xen_machine_halt(void)
1231 xen_reboot(SHUTDOWN_poweroff);
1234 static void xen_machine_power_off(void)
1238 xen_reboot(SHUTDOWN_poweroff);
1241 static void xen_crash_shutdown(struct pt_regs *regs)
1243 xen_reboot(SHUTDOWN_crash);
1247 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1249 xen_reboot(SHUTDOWN_crash);
1253 static struct notifier_block xen_panic_block = {
1254 .notifier_call= xen_panic_event,
1257 int xen_panic_handler_init(void)
1259 atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1263 static const struct machine_ops xen_machine_ops __initconst = {
1264 .restart = xen_restart,
1265 .halt = xen_machine_halt,
1266 .power_off = xen_machine_power_off,
1267 .shutdown = xen_machine_halt,
1268 .crash_shutdown = xen_crash_shutdown,
1269 .emergency_restart = xen_emergency_restart,
1273 * Set up the GDT and segment registers for -fstack-protector. Until
1274 * we do this, we have to be careful not to call any stack-protected
1275 * function, which is most of the kernel.
1277 static void __init xen_setup_stackprotector(void)
1279 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1280 pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1282 setup_stack_canary_segment(0);
1283 switch_to_new_gdt(0);
1285 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1286 pv_cpu_ops.load_gdt = xen_load_gdt;
1289 /* First C function to be called on Xen boot */
1290 asmlinkage void __init xen_start_kernel(void)
1292 struct physdev_set_iopl set_iopl;
1295 if (!xen_start_info)
1298 xen_domain_type = XEN_PV_DOMAIN;
1300 xen_setup_machphys_mapping();
1302 /* Install Xen paravirt ops */
1304 pv_init_ops = xen_init_ops;
1305 pv_cpu_ops = xen_cpu_ops;
1306 pv_apic_ops = xen_apic_ops;
1308 x86_init.resources.memory_setup = xen_memory_setup;
1309 x86_init.oem.arch_setup = xen_arch_setup;
1310 x86_init.oem.banner = xen_banner;
1312 xen_init_time_ops();
1315 * Set up some pagetable state before starting to set any ptes.
1320 /* Prevent unwanted bits from being set in PTEs. */
1321 __supported_pte_mask &= ~_PAGE_GLOBAL;
1323 if (!xen_initial_domain())
1325 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1327 __supported_pte_mask |= _PAGE_IOMAP;
1330 * Prevent page tables from being allocated in highmem, even
1331 * if CONFIG_HIGHPTE is enabled.
1333 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1335 /* Work out if we support NX */
1338 xen_setup_features();
1341 if (!xen_feature(XENFEAT_auto_translated_physmap))
1342 xen_build_dynamic_phys_to_machine();
1345 * Set up kernel GDT and segment registers, mainly so that
1346 * -fstack-protector code can be executed.
1348 xen_setup_stackprotector();
1351 xen_init_cpuid_mask();
1353 #ifdef CONFIG_X86_LOCAL_APIC
1355 * set up the basic apic ops.
1357 set_xen_basic_apic_ops();
1360 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1361 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1362 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1365 machine_ops = xen_machine_ops;
1368 * The only reliable way to retain the initial address of the
1369 * percpu gdt_page is to remember it here, so we can go and
1370 * mark it RW later, when the initial percpu area is freed.
1372 xen_initial_gdt = &per_cpu(gdt_page, 0);
1376 #ifdef CONFIG_ACPI_NUMA
1378 * The pages we from Xen are not related to machine pages, so
1379 * any NUMA information the kernel tries to get from ACPI will
1380 * be meaningless. Prevent it from trying.
1385 /* Don't do the full vcpu_info placement stuff until we have a
1386 possible map and a non-dummy shared_info. */
1387 per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1389 local_irq_disable();
1390 early_boot_irqs_disabled = true;
1392 xen_raw_console_write("mapping kernel into physical memory\n");
1393 xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base, xen_start_info->nr_pages);
1395 /* Allocate and initialize top and mid mfn levels for p2m structure */
1396 xen_build_mfn_list_list();
1398 /* keep using Xen gdt for now; no urgent need to change it */
1400 #ifdef CONFIG_X86_32
1401 pv_info.kernel_rpl = 1;
1402 if (xen_feature(XENFEAT_supervisor_mode_kernel))
1403 pv_info.kernel_rpl = 0;
1405 pv_info.kernel_rpl = 0;
1407 /* set the limit of our address space */
1410 /* We used to do this in xen_arch_setup, but that is too late on AMD
1411 * were early_cpu_init (run before ->arch_setup()) calls early_amd_init
1412 * which pokes 0xcf8 port.
1415 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1417 xen_raw_printk("physdev_op failed %d\n", rc);
1419 #ifdef CONFIG_X86_32
1420 /* set up basic CPUID stuff */
1421 cpu_detect(&new_cpu_data);
1422 new_cpu_data.hard_math = 1;
1423 new_cpu_data.wp_works_ok = 1;
1424 new_cpu_data.x86_capability[0] = cpuid_edx(1);
1427 /* Poke various useful things into boot_params */
1428 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1429 boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1430 ? __pa(xen_start_info->mod_start) : 0;
1431 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1432 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1434 if (!xen_initial_domain()) {
1435 add_preferred_console("xenboot", 0, NULL);
1436 add_preferred_console("tty", 0, NULL);
1437 add_preferred_console("hvc", 0, NULL);
1439 x86_init.pci.arch_init = pci_xen_init;
1441 const struct dom0_vga_console_info *info =
1442 (void *)((char *)xen_start_info +
1443 xen_start_info->console.dom0.info_off);
1444 struct xen_platform_op op = {
1445 .cmd = XENPF_firmware_info,
1446 .interface_version = XENPF_INTERFACE_VERSION,
1447 .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
1450 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1451 xen_start_info->console.domU.mfn = 0;
1452 xen_start_info->console.domU.evtchn = 0;
1454 if (HYPERVISOR_dom0_op(&op) == 0)
1455 boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
1459 /* Make sure ACS will be enabled */
1462 xen_acpi_sleep_register();
1464 /* Avoid searching for BIOS MP tables */
1465 x86_init.mpparse.find_smp_config = x86_init_noop;
1466 x86_init.mpparse.get_smp_config = x86_init_uint_noop;
1469 /* PCI BIOS service won't work from a PV guest. */
1470 pci_probe &= ~PCI_PROBE_BIOS;
1472 xen_raw_console_write("about to get started...\n");
1474 xen_setup_runstate_info(0);
1476 /* Start the world */
1477 #ifdef CONFIG_X86_32
1478 i386_start_kernel();
1480 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1484 void __ref xen_hvm_init_shared_info(void)
1487 struct xen_add_to_physmap xatp;
1488 static struct shared_info *shared_info_page = 0;
1490 if (!shared_info_page)
1491 shared_info_page = (struct shared_info *)
1492 extend_brk(PAGE_SIZE, PAGE_SIZE);
1493 xatp.domid = DOMID_SELF;
1495 xatp.space = XENMAPSPACE_shared_info;
1496 xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT;
1497 if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1500 HYPERVISOR_shared_info = (struct shared_info *)shared_info_page;
1502 /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1503 * page, we use it in the event channel upcall and in some pvclock
1504 * related functions. We don't need the vcpu_info placement
1505 * optimizations because we don't use any pv_mmu or pv_irq op on
1507 * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is
1508 * online but xen_hvm_init_shared_info is run at resume time too and
1509 * in that case multiple vcpus might be online. */
1510 for_each_online_cpu(cpu) {
1511 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1515 #ifdef CONFIG_XEN_PVHVM
1516 static void __init init_hvm_pv_info(void)
1519 uint32_t eax, ebx, ecx, edx, pages, msr, base;
1522 base = xen_cpuid_base();
1523 cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1526 minor = eax & 0xffff;
1527 printk(KERN_INFO "Xen version %d.%d.\n", major, minor);
1529 cpuid(base + 2, &pages, &msr, &ecx, &edx);
1531 pfn = __pa(hypercall_page);
1532 wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1534 xen_setup_features();
1536 pv_info.name = "Xen HVM";
1538 xen_domain_type = XEN_HVM_DOMAIN;
1541 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self,
1542 unsigned long action, void *hcpu)
1544 int cpu = (long)hcpu;
1546 case CPU_UP_PREPARE:
1547 xen_vcpu_setup(cpu);
1548 if (xen_have_vector_callback)
1549 xen_init_lock_cpu(cpu);
1557 static struct notifier_block xen_hvm_cpu_notifier __cpuinitdata = {
1558 .notifier_call = xen_hvm_cpu_notify,
1561 static void __init xen_hvm_guest_init(void)
1565 xen_hvm_init_shared_info();
1567 if (xen_feature(XENFEAT_hvm_callback_vector))
1568 xen_have_vector_callback = 1;
1570 register_cpu_notifier(&xen_hvm_cpu_notifier);
1571 xen_unplug_emulated_devices();
1572 x86_init.irqs.intr_init = xen_init_IRQ;
1573 xen_hvm_init_time_ops();
1574 xen_hvm_init_mmu_ops();
1577 static bool __init xen_hvm_platform(void)
1579 if (xen_pv_domain())
1582 if (!xen_cpuid_base())
1588 bool xen_hvm_need_lapic(void)
1590 if (xen_pv_domain())
1592 if (!xen_hvm_domain())
1594 if (xen_feature(XENFEAT_hvm_pirqs) && xen_have_vector_callback)
1598 EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1600 const struct hypervisor_x86 x86_hyper_xen_hvm __refconst = {
1602 .detect = xen_hvm_platform,
1603 .init_platform = xen_hvm_guest_init,
1605 EXPORT_SYMBOL(x86_hyper_xen_hvm);