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/mwait.h>
66 #include <asm/pci_x86.h>
69 #include <linux/acpi.h>
71 #include <acpi/pdc_intel.h>
72 #include <acpi/processor.h>
73 #include <xen/interface/platform.h>
78 #include "multicalls.h"
80 EXPORT_SYMBOL_GPL(hypercall_page);
82 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
83 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
85 enum xen_domain_type xen_domain_type = XEN_NATIVE;
86 EXPORT_SYMBOL_GPL(xen_domain_type);
88 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
89 EXPORT_SYMBOL(machine_to_phys_mapping);
90 unsigned long machine_to_phys_nr;
91 EXPORT_SYMBOL(machine_to_phys_nr);
93 struct start_info *xen_start_info;
94 EXPORT_SYMBOL_GPL(xen_start_info);
96 struct shared_info xen_dummy_shared_info;
98 void *xen_initial_gdt;
100 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
101 __read_mostly int xen_have_vector_callback;
102 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
105 * Point at some empty memory to start with. We map the real shared_info
106 * page as soon as fixmap is up and running.
108 struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info;
111 * Flag to determine whether vcpu info placement is available on all
112 * VCPUs. We assume it is to start with, and then set it to zero on
113 * the first failure. This is because it can succeed on some VCPUs
114 * and not others, since it can involve hypervisor memory allocation,
115 * or because the guest failed to guarantee all the appropriate
116 * constraints on all VCPUs (ie buffer can't cross a page boundary).
118 * Note that any particular CPU may be using a placed vcpu structure,
119 * but we can only optimise if the all are.
121 * 0: not available, 1: available
123 static int have_vcpu_info_placement = 1;
125 static void clamp_max_cpus(void)
128 if (setup_max_cpus > MAX_VIRT_CPUS)
129 setup_max_cpus = MAX_VIRT_CPUS;
133 static void xen_vcpu_setup(int cpu)
135 struct vcpu_register_vcpu_info info;
137 struct vcpu_info *vcpup;
139 BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
141 if (cpu < MAX_VIRT_CPUS)
142 per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
144 if (!have_vcpu_info_placement) {
145 if (cpu >= MAX_VIRT_CPUS)
150 vcpup = &per_cpu(xen_vcpu_info, cpu);
151 info.mfn = arbitrary_virt_to_mfn(vcpup);
152 info.offset = offset_in_page(vcpup);
154 /* Check to see if the hypervisor will put the vcpu_info
155 structure where we want it, which allows direct access via
156 a percpu-variable. */
157 err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
160 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
161 have_vcpu_info_placement = 0;
164 /* This cpu is using the registered vcpu info, even if
165 later ones fail to. */
166 per_cpu(xen_vcpu, cpu) = vcpup;
171 * On restore, set the vcpu placement up again.
172 * If it fails, then we're in a bad state, since
173 * we can't back out from using it...
175 void xen_vcpu_restore(void)
179 for_each_online_cpu(cpu) {
180 bool other_cpu = (cpu != smp_processor_id());
183 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
186 xen_setup_runstate_info(cpu);
188 if (have_vcpu_info_placement)
192 HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
197 static void __init xen_banner(void)
199 unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
200 struct xen_extraversion extra;
201 HYPERVISOR_xen_version(XENVER_extraversion, &extra);
203 printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
205 printk(KERN_INFO "Xen version: %d.%d%s%s\n",
206 version >> 16, version & 0xffff, extra.extraversion,
207 xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
210 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
211 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
213 static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
214 static __read_mostly unsigned int cpuid_leaf5_ecx_val;
215 static __read_mostly unsigned int cpuid_leaf5_edx_val;
217 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
218 unsigned int *cx, unsigned int *dx)
220 unsigned maskebx = ~0;
221 unsigned maskecx = ~0;
222 unsigned maskedx = ~0;
225 * Mask out inconvenient features, to try and disable as many
226 * unsupported kernel subsystems as possible.
230 maskecx = cpuid_leaf1_ecx_mask;
231 setecx = cpuid_leaf1_ecx_set_mask;
232 maskedx = cpuid_leaf1_edx_mask;
235 case CPUID_MWAIT_LEAF:
236 /* Synthesize the values.. */
239 *cx = cpuid_leaf5_ecx_val;
240 *dx = cpuid_leaf5_edx_val;
244 /* Suppress extended topology stuff */
249 asm(XEN_EMULATE_PREFIX "cpuid"
254 : "0" (*ax), "2" (*cx));
263 static bool __init xen_check_mwait(void)
265 #if defined(CONFIG_ACPI) && !defined(CONFIG_ACPI_PROCESSOR_AGGREGATOR) && \
266 !defined(CONFIG_ACPI_PROCESSOR_AGGREGATOR_MODULE)
267 struct xen_platform_op op = {
268 .cmd = XENPF_set_processor_pminfo,
269 .u.set_pminfo.id = -1,
270 .u.set_pminfo.type = XEN_PM_PDC,
273 unsigned int ax, bx, cx, dx;
274 unsigned int mwait_mask;
276 /* We need to determine whether it is OK to expose the MWAIT
277 * capability to the kernel to harvest deeper than C3 states from ACPI
278 * _CST using the processor_harvest_xen.c module. For this to work, we
279 * need to gather the MWAIT_LEAF values (which the cstate.c code
280 * checks against). The hypervisor won't expose the MWAIT flag because
281 * it would break backwards compatibility; so we will find out directly
282 * from the hardware and hypercall.
284 if (!xen_initial_domain())
290 native_cpuid(&ax, &bx, &cx, &dx);
292 mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
293 (1 << (X86_FEATURE_MWAIT % 32));
295 if ((cx & mwait_mask) != mwait_mask)
298 /* We need to emulate the MWAIT_LEAF and for that we need both
299 * ecx and edx. The hypercall provides only partial information.
302 ax = CPUID_MWAIT_LEAF;
307 native_cpuid(&ax, &bx, &cx, &dx);
309 /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
310 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
312 buf[0] = ACPI_PDC_REVISION_ID;
314 buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
316 set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
318 if ((HYPERVISOR_dom0_op(&op) == 0) &&
319 (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
320 cpuid_leaf5_ecx_val = cx;
321 cpuid_leaf5_edx_val = dx;
328 static void __init xen_init_cpuid_mask(void)
330 unsigned int ax, bx, cx, dx;
331 unsigned int xsave_mask;
333 cpuid_leaf1_edx_mask =
334 ~((1 << X86_FEATURE_MCE) | /* disable MCE */
335 (1 << X86_FEATURE_MCA) | /* disable MCA */
336 (1 << X86_FEATURE_MTRR) | /* disable MTRR */
337 (1 << X86_FEATURE_ACC)); /* thermal monitoring */
339 if (!xen_initial_domain())
340 cpuid_leaf1_edx_mask &=
341 ~((1 << X86_FEATURE_APIC) | /* disable local APIC */
342 (1 << X86_FEATURE_ACPI)); /* disable ACPI */
345 xen_cpuid(&ax, &bx, &cx, &dx);
348 (1 << (X86_FEATURE_XSAVE % 32)) |
349 (1 << (X86_FEATURE_OSXSAVE % 32));
351 /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
352 if ((cx & xsave_mask) != xsave_mask)
353 cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
354 if (xen_check_mwait())
355 cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
358 static void xen_set_debugreg(int reg, unsigned long val)
360 HYPERVISOR_set_debugreg(reg, val);
363 static unsigned long xen_get_debugreg(int reg)
365 return HYPERVISOR_get_debugreg(reg);
368 static void xen_end_context_switch(struct task_struct *next)
371 paravirt_end_context_switch(next);
374 static unsigned long xen_store_tr(void)
380 * Set the page permissions for a particular virtual address. If the
381 * address is a vmalloc mapping (or other non-linear mapping), then
382 * find the linear mapping of the page and also set its protections to
385 static void set_aliased_prot(void *v, pgprot_t prot)
393 ptep = lookup_address((unsigned long)v, &level);
394 BUG_ON(ptep == NULL);
396 pfn = pte_pfn(*ptep);
397 page = pfn_to_page(pfn);
399 pte = pfn_pte(pfn, prot);
401 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
404 if (!PageHighMem(page)) {
405 void *av = __va(PFN_PHYS(pfn));
408 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
414 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
416 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
419 for(i = 0; i < entries; i += entries_per_page)
420 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
423 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
425 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
428 for(i = 0; i < entries; i += entries_per_page)
429 set_aliased_prot(ldt + i, PAGE_KERNEL);
432 static void xen_set_ldt(const void *addr, unsigned entries)
434 struct mmuext_op *op;
435 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
437 trace_xen_cpu_set_ldt(addr, entries);
440 op->cmd = MMUEXT_SET_LDT;
441 op->arg1.linear_addr = (unsigned long)addr;
442 op->arg2.nr_ents = entries;
444 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
446 xen_mc_issue(PARAVIRT_LAZY_CPU);
449 static void xen_load_gdt(const struct desc_ptr *dtr)
451 unsigned long va = dtr->address;
452 unsigned int size = dtr->size + 1;
453 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
454 unsigned long frames[pages];
458 * A GDT can be up to 64k in size, which corresponds to 8192
459 * 8-byte entries, or 16 4k pages..
462 BUG_ON(size > 65536);
463 BUG_ON(va & ~PAGE_MASK);
465 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
468 unsigned long pfn, mfn;
472 * The GDT is per-cpu and is in the percpu data area.
473 * That can be virtually mapped, so we need to do a
474 * page-walk to get the underlying MFN for the
475 * hypercall. The page can also be in the kernel's
476 * linear range, so we need to RO that mapping too.
478 ptep = lookup_address(va, &level);
479 BUG_ON(ptep == NULL);
481 pfn = pte_pfn(*ptep);
482 mfn = pfn_to_mfn(pfn);
483 virt = __va(PFN_PHYS(pfn));
487 make_lowmem_page_readonly((void *)va);
488 make_lowmem_page_readonly(virt);
491 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
496 * load_gdt for early boot, when the gdt is only mapped once
498 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
500 unsigned long va = dtr->address;
501 unsigned int size = dtr->size + 1;
502 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
503 unsigned long frames[pages];
507 * A GDT can be up to 64k in size, which corresponds to 8192
508 * 8-byte entries, or 16 4k pages..
511 BUG_ON(size > 65536);
512 BUG_ON(va & ~PAGE_MASK);
514 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
516 unsigned long pfn, mfn;
518 pfn = virt_to_pfn(va);
519 mfn = pfn_to_mfn(pfn);
521 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
523 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
529 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
533 static void load_TLS_descriptor(struct thread_struct *t,
534 unsigned int cpu, unsigned int i)
536 struct desc_struct *gdt = get_cpu_gdt_table(cpu);
537 xmaddr_t maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
538 struct multicall_space mc = __xen_mc_entry(0);
540 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
543 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
546 * XXX sleazy hack: If we're being called in a lazy-cpu zone
547 * and lazy gs handling is enabled, it means we're in a
548 * context switch, and %gs has just been saved. This means we
549 * can zero it out to prevent faults on exit from the
550 * hypervisor if the next process has no %gs. Either way, it
551 * has been saved, and the new value will get loaded properly.
552 * This will go away as soon as Xen has been modified to not
553 * save/restore %gs for normal hypercalls.
555 * On x86_64, this hack is not used for %gs, because gs points
556 * to KERNEL_GS_BASE (and uses it for PDA references), so we
557 * must not zero %gs on x86_64
559 * For x86_64, we need to zero %fs, otherwise we may get an
560 * exception between the new %fs descriptor being loaded and
561 * %fs being effectively cleared at __switch_to().
563 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
573 load_TLS_descriptor(t, cpu, 0);
574 load_TLS_descriptor(t, cpu, 1);
575 load_TLS_descriptor(t, cpu, 2);
577 xen_mc_issue(PARAVIRT_LAZY_CPU);
581 static void xen_load_gs_index(unsigned int idx)
583 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
588 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
591 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
592 u64 entry = *(u64 *)ptr;
594 trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
599 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
605 static int cvt_gate_to_trap(int vector, const gate_desc *val,
606 struct trap_info *info)
610 if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
613 info->vector = vector;
615 addr = gate_offset(*val);
618 * Look for known traps using IST, and substitute them
619 * appropriately. The debugger ones are the only ones we care
620 * about. Xen will handle faults like double_fault and
621 * machine_check, so we should never see them. Warn if
622 * there's an unexpected IST-using fault handler.
624 if (addr == (unsigned long)debug)
625 addr = (unsigned long)xen_debug;
626 else if (addr == (unsigned long)int3)
627 addr = (unsigned long)xen_int3;
628 else if (addr == (unsigned long)stack_segment)
629 addr = (unsigned long)xen_stack_segment;
630 else if (addr == (unsigned long)double_fault ||
631 addr == (unsigned long)nmi) {
632 /* Don't need to handle these */
634 #ifdef CONFIG_X86_MCE
635 } else if (addr == (unsigned long)machine_check) {
639 /* Some other trap using IST? */
640 if (WARN_ON(val->ist != 0))
643 #endif /* CONFIG_X86_64 */
644 info->address = addr;
646 info->cs = gate_segment(*val);
647 info->flags = val->dpl;
648 /* interrupt gates clear IF */
649 if (val->type == GATE_INTERRUPT)
650 info->flags |= 1 << 2;
655 /* Locations of each CPU's IDT */
656 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
658 /* Set an IDT entry. If the entry is part of the current IDT, then
660 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
662 unsigned long p = (unsigned long)&dt[entrynum];
663 unsigned long start, end;
665 trace_xen_cpu_write_idt_entry(dt, entrynum, g);
669 start = __this_cpu_read(idt_desc.address);
670 end = start + __this_cpu_read(idt_desc.size) + 1;
674 native_write_idt_entry(dt, entrynum, g);
676 if (p >= start && (p + 8) <= end) {
677 struct trap_info info[2];
681 if (cvt_gate_to_trap(entrynum, g, &info[0]))
682 if (HYPERVISOR_set_trap_table(info))
689 static void xen_convert_trap_info(const struct desc_ptr *desc,
690 struct trap_info *traps)
692 unsigned in, out, count;
694 count = (desc->size+1) / sizeof(gate_desc);
697 for (in = out = 0; in < count; in++) {
698 gate_desc *entry = (gate_desc*)(desc->address) + in;
700 if (cvt_gate_to_trap(in, entry, &traps[out]))
703 traps[out].address = 0;
706 void xen_copy_trap_info(struct trap_info *traps)
708 const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
710 xen_convert_trap_info(desc, traps);
713 /* Load a new IDT into Xen. In principle this can be per-CPU, so we
714 hold a spinlock to protect the static traps[] array (static because
715 it avoids allocation, and saves stack space). */
716 static void xen_load_idt(const struct desc_ptr *desc)
718 static DEFINE_SPINLOCK(lock);
719 static struct trap_info traps[257];
721 trace_xen_cpu_load_idt(desc);
725 __get_cpu_var(idt_desc) = *desc;
727 xen_convert_trap_info(desc, traps);
730 if (HYPERVISOR_set_trap_table(traps))
736 /* Write a GDT descriptor entry. Ignore LDT descriptors, since
737 they're handled differently. */
738 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
739 const void *desc, int type)
741 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
752 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
755 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
765 * Version of write_gdt_entry for use at early boot-time needed to
766 * update an entry as simply as possible.
768 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
769 const void *desc, int type)
771 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
780 xmaddr_t maddr = virt_to_machine(&dt[entry]);
782 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
783 dt[entry] = *(struct desc_struct *)desc;
789 static void xen_load_sp0(struct tss_struct *tss,
790 struct thread_struct *thread)
792 struct multicall_space mcs;
794 mcs = xen_mc_entry(0);
795 MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
796 xen_mc_issue(PARAVIRT_LAZY_CPU);
799 static void xen_set_iopl_mask(unsigned mask)
801 struct physdev_set_iopl set_iopl;
803 /* Force the change at ring 0. */
804 set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
805 HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
808 static void xen_io_delay(void)
812 #ifdef CONFIG_X86_LOCAL_APIC
813 static unsigned long xen_set_apic_id(unsigned int x)
818 static unsigned int xen_get_apic_id(unsigned long x)
820 return ((x)>>24) & 0xFFu;
822 static u32 xen_apic_read(u32 reg)
824 struct xen_platform_op op = {
825 .cmd = XENPF_get_cpuinfo,
826 .interface_version = XENPF_INTERFACE_VERSION,
827 .u.pcpu_info.xen_cpuid = 0,
831 /* Shouldn't need this as APIC is turned off for PV, and we only
832 * get called on the bootup processor. But just in case. */
833 if (!xen_initial_domain() || smp_processor_id())
842 ret = HYPERVISOR_dom0_op(&op);
846 return op.u.pcpu_info.apic_id << 24;
849 static void xen_apic_write(u32 reg, u32 val)
851 /* Warn to see if there's any stray references */
855 static u64 xen_apic_icr_read(void)
860 static void xen_apic_icr_write(u32 low, u32 id)
862 /* Warn to see if there's any stray references */
866 static void xen_apic_wait_icr_idle(void)
871 static u32 xen_safe_apic_wait_icr_idle(void)
876 static void set_xen_basic_apic_ops(void)
878 apic->read = xen_apic_read;
879 apic->write = xen_apic_write;
880 apic->icr_read = xen_apic_icr_read;
881 apic->icr_write = xen_apic_icr_write;
882 apic->wait_icr_idle = xen_apic_wait_icr_idle;
883 apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
884 apic->set_apic_id = xen_set_apic_id;
885 apic->get_apic_id = xen_get_apic_id;
890 static void xen_clts(void)
892 struct multicall_space mcs;
894 mcs = xen_mc_entry(0);
896 MULTI_fpu_taskswitch(mcs.mc, 0);
898 xen_mc_issue(PARAVIRT_LAZY_CPU);
901 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
903 static unsigned long xen_read_cr0(void)
905 unsigned long cr0 = this_cpu_read(xen_cr0_value);
907 if (unlikely(cr0 == 0)) {
908 cr0 = native_read_cr0();
909 this_cpu_write(xen_cr0_value, cr0);
915 static void xen_write_cr0(unsigned long cr0)
917 struct multicall_space mcs;
919 this_cpu_write(xen_cr0_value, cr0);
921 /* Only pay attention to cr0.TS; everything else is
923 mcs = xen_mc_entry(0);
925 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
927 xen_mc_issue(PARAVIRT_LAZY_CPU);
930 static void xen_write_cr4(unsigned long cr4)
935 native_write_cr4(cr4);
938 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
949 case MSR_FS_BASE: which = SEGBASE_FS; goto set;
950 case MSR_KERNEL_GS_BASE: which = SEGBASE_GS_USER; goto set;
951 case MSR_GS_BASE: which = SEGBASE_GS_KERNEL; goto set;
954 base = ((u64)high << 32) | low;
955 if (HYPERVISOR_set_segment_base(which, base) != 0)
963 case MSR_SYSCALL_MASK:
964 case MSR_IA32_SYSENTER_CS:
965 case MSR_IA32_SYSENTER_ESP:
966 case MSR_IA32_SYSENTER_EIP:
967 /* Fast syscall setup is all done in hypercalls, so
968 these are all ignored. Stub them out here to stop
969 Xen console noise. */
972 case MSR_IA32_CR_PAT:
973 if (smp_processor_id() == 0)
974 xen_set_pat(((u64)high << 32) | low);
978 ret = native_write_msr_safe(msr, low, high);
984 void xen_setup_shared_info(void)
986 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
987 set_fixmap(FIX_PARAVIRT_BOOTMAP,
988 xen_start_info->shared_info);
990 HYPERVISOR_shared_info =
991 (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
993 HYPERVISOR_shared_info =
994 (struct shared_info *)__va(xen_start_info->shared_info);
997 /* In UP this is as good a place as any to set up shared info */
998 xen_setup_vcpu_info_placement();
1001 xen_setup_mfn_list_list();
1004 /* This is called once we have the cpu_possible_mask */
1005 void xen_setup_vcpu_info_placement(void)
1009 for_each_possible_cpu(cpu)
1010 xen_vcpu_setup(cpu);
1012 /* xen_vcpu_setup managed to place the vcpu_info within the
1013 percpu area for all cpus, so make use of it */
1014 if (have_vcpu_info_placement) {
1015 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1016 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1017 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1018 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1019 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1023 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1024 unsigned long addr, unsigned len)
1026 char *start, *end, *reloc;
1029 start = end = reloc = NULL;
1031 #define SITE(op, x) \
1032 case PARAVIRT_PATCH(op.x): \
1033 if (have_vcpu_info_placement) { \
1034 start = (char *)xen_##x##_direct; \
1035 end = xen_##x##_direct_end; \
1036 reloc = xen_##x##_direct_reloc; \
1041 SITE(pv_irq_ops, irq_enable);
1042 SITE(pv_irq_ops, irq_disable);
1043 SITE(pv_irq_ops, save_fl);
1044 SITE(pv_irq_ops, restore_fl);
1048 if (start == NULL || (end-start) > len)
1051 ret = paravirt_patch_insns(insnbuf, len, start, end);
1053 /* Note: because reloc is assigned from something that
1054 appears to be an array, gcc assumes it's non-null,
1055 but doesn't know its relationship with start and
1057 if (reloc > start && reloc < end) {
1058 int reloc_off = reloc - start;
1059 long *relocp = (long *)(insnbuf + reloc_off);
1060 long delta = start - (char *)addr;
1068 ret = paravirt_patch_default(type, clobbers, insnbuf,
1076 static const struct pv_info xen_info __initconst = {
1077 .paravirt_enabled = 1,
1078 .shared_kernel_pmd = 0,
1080 #ifdef CONFIG_X86_64
1081 .extra_user_64bit_cs = FLAT_USER_CS64,
1087 static const struct pv_init_ops xen_init_ops __initconst = {
1091 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1094 .set_debugreg = xen_set_debugreg,
1095 .get_debugreg = xen_get_debugreg,
1099 .read_cr0 = xen_read_cr0,
1100 .write_cr0 = xen_write_cr0,
1102 .read_cr4 = native_read_cr4,
1103 .read_cr4_safe = native_read_cr4_safe,
1104 .write_cr4 = xen_write_cr4,
1106 .wbinvd = native_wbinvd,
1108 .read_msr = native_read_msr_safe,
1109 .write_msr = xen_write_msr_safe,
1110 .read_tsc = native_read_tsc,
1111 .read_pmc = native_read_pmc,
1114 .irq_enable_sysexit = xen_sysexit,
1115 #ifdef CONFIG_X86_64
1116 .usergs_sysret32 = xen_sysret32,
1117 .usergs_sysret64 = xen_sysret64,
1120 .load_tr_desc = paravirt_nop,
1121 .set_ldt = xen_set_ldt,
1122 .load_gdt = xen_load_gdt,
1123 .load_idt = xen_load_idt,
1124 .load_tls = xen_load_tls,
1125 #ifdef CONFIG_X86_64
1126 .load_gs_index = xen_load_gs_index,
1129 .alloc_ldt = xen_alloc_ldt,
1130 .free_ldt = xen_free_ldt,
1132 .store_gdt = native_store_gdt,
1133 .store_idt = native_store_idt,
1134 .store_tr = xen_store_tr,
1136 .write_ldt_entry = xen_write_ldt_entry,
1137 .write_gdt_entry = xen_write_gdt_entry,
1138 .write_idt_entry = xen_write_idt_entry,
1139 .load_sp0 = xen_load_sp0,
1141 .set_iopl_mask = xen_set_iopl_mask,
1142 .io_delay = xen_io_delay,
1144 /* Xen takes care of %gs when switching to usermode for us */
1145 .swapgs = paravirt_nop,
1147 .start_context_switch = paravirt_start_context_switch,
1148 .end_context_switch = xen_end_context_switch,
1151 static const struct pv_apic_ops xen_apic_ops __initconst = {
1152 #ifdef CONFIG_X86_LOCAL_APIC
1153 .startup_ipi_hook = paravirt_nop,
1157 static void xen_reboot(int reason)
1159 struct sched_shutdown r = { .reason = reason };
1161 if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1165 static void xen_restart(char *msg)
1167 xen_reboot(SHUTDOWN_reboot);
1170 static void xen_emergency_restart(void)
1172 xen_reboot(SHUTDOWN_reboot);
1175 static void xen_machine_halt(void)
1177 xen_reboot(SHUTDOWN_poweroff);
1180 static void xen_machine_power_off(void)
1184 xen_reboot(SHUTDOWN_poweroff);
1187 static void xen_crash_shutdown(struct pt_regs *regs)
1189 xen_reboot(SHUTDOWN_crash);
1193 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1195 xen_reboot(SHUTDOWN_crash);
1199 static struct notifier_block xen_panic_block = {
1200 .notifier_call= xen_panic_event,
1203 int xen_panic_handler_init(void)
1205 atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1209 static const struct machine_ops xen_machine_ops __initconst = {
1210 .restart = xen_restart,
1211 .halt = xen_machine_halt,
1212 .power_off = xen_machine_power_off,
1213 .shutdown = xen_machine_halt,
1214 .crash_shutdown = xen_crash_shutdown,
1215 .emergency_restart = xen_emergency_restart,
1219 * Set up the GDT and segment registers for -fstack-protector. Until
1220 * we do this, we have to be careful not to call any stack-protected
1221 * function, which is most of the kernel.
1223 static void __init xen_setup_stackprotector(void)
1225 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1226 pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1228 setup_stack_canary_segment(0);
1229 switch_to_new_gdt(0);
1231 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1232 pv_cpu_ops.load_gdt = xen_load_gdt;
1235 /* First C function to be called on Xen boot */
1236 asmlinkage void __init xen_start_kernel(void)
1238 struct physdev_set_iopl set_iopl;
1242 if (!xen_start_info)
1245 xen_domain_type = XEN_PV_DOMAIN;
1247 xen_setup_machphys_mapping();
1249 /* Install Xen paravirt ops */
1251 pv_init_ops = xen_init_ops;
1252 pv_cpu_ops = xen_cpu_ops;
1253 pv_apic_ops = xen_apic_ops;
1255 x86_init.resources.memory_setup = xen_memory_setup;
1256 x86_init.oem.arch_setup = xen_arch_setup;
1257 x86_init.oem.banner = xen_banner;
1259 xen_init_time_ops();
1262 * Set up some pagetable state before starting to set any ptes.
1267 /* Prevent unwanted bits from being set in PTEs. */
1268 __supported_pte_mask &= ~_PAGE_GLOBAL;
1270 if (!xen_initial_domain())
1272 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1274 __supported_pte_mask |= _PAGE_IOMAP;
1277 * Prevent page tables from being allocated in highmem, even
1278 * if CONFIG_HIGHPTE is enabled.
1280 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1282 /* Work out if we support NX */
1285 xen_setup_features();
1288 if (!xen_feature(XENFEAT_auto_translated_physmap))
1289 xen_build_dynamic_phys_to_machine();
1292 * Set up kernel GDT and segment registers, mainly so that
1293 * -fstack-protector code can be executed.
1295 xen_setup_stackprotector();
1298 xen_init_cpuid_mask();
1300 #ifdef CONFIG_X86_LOCAL_APIC
1302 * set up the basic apic ops.
1304 set_xen_basic_apic_ops();
1307 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1308 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1309 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1312 machine_ops = xen_machine_ops;
1315 * The only reliable way to retain the initial address of the
1316 * percpu gdt_page is to remember it here, so we can go and
1317 * mark it RW later, when the initial percpu area is freed.
1319 xen_initial_gdt = &per_cpu(gdt_page, 0);
1323 #ifdef CONFIG_ACPI_NUMA
1325 * The pages we from Xen are not related to machine pages, so
1326 * any NUMA information the kernel tries to get from ACPI will
1327 * be meaningless. Prevent it from trying.
1332 pgd = (pgd_t *)xen_start_info->pt_base;
1334 /* Don't do the full vcpu_info placement stuff until we have a
1335 possible map and a non-dummy shared_info. */
1336 per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1338 local_irq_disable();
1339 early_boot_irqs_disabled = true;
1341 xen_raw_console_write("mapping kernel into physical memory\n");
1342 pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
1343 xen_ident_map_ISA();
1345 /* Allocate and initialize top and mid mfn levels for p2m structure */
1346 xen_build_mfn_list_list();
1348 /* keep using Xen gdt for now; no urgent need to change it */
1350 #ifdef CONFIG_X86_32
1351 pv_info.kernel_rpl = 1;
1352 if (xen_feature(XENFEAT_supervisor_mode_kernel))
1353 pv_info.kernel_rpl = 0;
1355 pv_info.kernel_rpl = 0;
1357 /* set the limit of our address space */
1360 /* We used to do this in xen_arch_setup, but that is too late on AMD
1361 * were early_cpu_init (run before ->arch_setup()) calls early_amd_init
1362 * which pokes 0xcf8 port.
1365 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1367 xen_raw_printk("physdev_op failed %d\n", rc);
1369 #ifdef CONFIG_X86_32
1370 /* set up basic CPUID stuff */
1371 cpu_detect(&new_cpu_data);
1372 new_cpu_data.hard_math = 1;
1373 new_cpu_data.wp_works_ok = 1;
1374 new_cpu_data.x86_capability[0] = cpuid_edx(1);
1377 /* Poke various useful things into boot_params */
1378 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1379 boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1380 ? __pa(xen_start_info->mod_start) : 0;
1381 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1382 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1384 if (!xen_initial_domain()) {
1385 add_preferred_console("xenboot", 0, NULL);
1386 add_preferred_console("tty", 0, NULL);
1387 add_preferred_console("hvc", 0, NULL);
1389 x86_init.pci.arch_init = pci_xen_init;
1391 const struct dom0_vga_console_info *info =
1392 (void *)((char *)xen_start_info +
1393 xen_start_info->console.dom0.info_off);
1395 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1396 xen_start_info->console.domU.mfn = 0;
1397 xen_start_info->console.domU.evtchn = 0;
1399 /* Make sure ACS will be enabled */
1403 /* PCI BIOS service won't work from a PV guest. */
1404 pci_probe &= ~PCI_PROBE_BIOS;
1406 xen_raw_console_write("about to get started...\n");
1408 xen_setup_runstate_info(0);
1410 /* Start the world */
1411 #ifdef CONFIG_X86_32
1412 i386_start_kernel();
1414 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1418 static int init_hvm_pv_info(int *major, int *minor)
1420 uint32_t eax, ebx, ecx, edx, pages, msr, base;
1423 base = xen_cpuid_base();
1424 cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1427 *minor = eax & 0xffff;
1428 printk(KERN_INFO "Xen version %d.%d.\n", *major, *minor);
1430 cpuid(base + 2, &pages, &msr, &ecx, &edx);
1432 pfn = __pa(hypercall_page);
1433 wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1435 xen_setup_features();
1437 pv_info.name = "Xen HVM";
1439 xen_domain_type = XEN_HVM_DOMAIN;
1444 void __ref xen_hvm_init_shared_info(void)
1447 struct xen_add_to_physmap xatp;
1448 static struct shared_info *shared_info_page = 0;
1450 if (!shared_info_page)
1451 shared_info_page = (struct shared_info *)
1452 extend_brk(PAGE_SIZE, PAGE_SIZE);
1453 xatp.domid = DOMID_SELF;
1455 xatp.space = XENMAPSPACE_shared_info;
1456 xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT;
1457 if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1460 HYPERVISOR_shared_info = (struct shared_info *)shared_info_page;
1462 /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1463 * page, we use it in the event channel upcall and in some pvclock
1464 * related functions. We don't need the vcpu_info placement
1465 * optimizations because we don't use any pv_mmu or pv_irq op on
1467 * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is
1468 * online but xen_hvm_init_shared_info is run at resume time too and
1469 * in that case multiple vcpus might be online. */
1470 for_each_online_cpu(cpu) {
1471 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1475 #ifdef CONFIG_XEN_PVHVM
1476 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self,
1477 unsigned long action, void *hcpu)
1479 int cpu = (long)hcpu;
1481 case CPU_UP_PREPARE:
1482 xen_vcpu_setup(cpu);
1483 if (xen_have_vector_callback)
1484 xen_init_lock_cpu(cpu);
1492 static struct notifier_block xen_hvm_cpu_notifier __cpuinitdata = {
1493 .notifier_call = xen_hvm_cpu_notify,
1496 static void __init xen_hvm_guest_init(void)
1501 r = init_hvm_pv_info(&major, &minor);
1505 xen_hvm_init_shared_info();
1507 if (xen_feature(XENFEAT_hvm_callback_vector))
1508 xen_have_vector_callback = 1;
1510 register_cpu_notifier(&xen_hvm_cpu_notifier);
1511 xen_unplug_emulated_devices();
1512 x86_init.irqs.intr_init = xen_init_IRQ;
1513 xen_hvm_init_time_ops();
1514 xen_hvm_init_mmu_ops();
1517 static bool __init xen_hvm_platform(void)
1519 if (xen_pv_domain())
1522 if (!xen_cpuid_base())
1528 bool xen_hvm_need_lapic(void)
1530 if (xen_pv_domain())
1532 if (!xen_hvm_domain())
1534 if (xen_feature(XENFEAT_hvm_pirqs) && xen_have_vector_callback)
1538 EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1540 const struct hypervisor_x86 x86_hyper_xen_hvm __refconst = {
1542 .detect = xen_hvm_platform,
1543 .init_platform = xen_hvm_guest_init,
1545 EXPORT_SYMBOL(x86_hyper_xen_hvm);