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>
34 #include <linux/edd.h>
37 #include <xen/events.h>
38 #include <xen/interface/xen.h>
39 #include <xen/interface/version.h>
40 #include <xen/interface/physdev.h>
41 #include <xen/interface/vcpu.h>
42 #include <xen/interface/memory.h>
43 #include <xen/interface/xen-mca.h>
44 #include <xen/features.h>
47 #include <xen/hvc-console.h>
50 #include <asm/paravirt.h>
53 #include <asm/xen/pci.h>
54 #include <asm/xen/hypercall.h>
55 #include <asm/xen/hypervisor.h>
56 #include <asm/fixmap.h>
57 #include <asm/processor.h>
58 #include <asm/proto.h>
59 #include <asm/msr-index.h>
60 #include <asm/traps.h>
61 #include <asm/setup.h>
63 #include <asm/pgalloc.h>
64 #include <asm/pgtable.h>
65 #include <asm/tlbflush.h>
66 #include <asm/reboot.h>
67 #include <asm/stackprotector.h>
68 #include <asm/hypervisor.h>
69 #include <asm/mwait.h>
70 #include <asm/pci_x86.h>
74 #include <linux/acpi.h>
76 #include <acpi/pdc_intel.h>
77 #include <acpi/processor.h>
78 #include <xen/interface/platform.h>
84 #include "multicalls.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_possible_cpu(cpu) {
199 bool other_cpu = (cpu != smp_processor_id());
200 bool is_up = HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL);
202 if (other_cpu && is_up &&
203 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
206 xen_setup_runstate_info(cpu);
208 if (have_vcpu_info_placement)
211 if (other_cpu && is_up &&
212 HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
217 static void __init xen_banner(void)
219 unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
220 struct xen_extraversion extra;
221 HYPERVISOR_xen_version(XENVER_extraversion, &extra);
223 printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
225 printk(KERN_INFO "Xen version: %d.%d%s%s\n",
226 version >> 16, version & 0xffff, extra.extraversion,
227 xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
229 /* Check if running on Xen version (major, minor) or later */
231 xen_running_on_version_or_later(unsigned int major, unsigned int minor)
233 unsigned int version;
238 version = HYPERVISOR_xen_version(XENVER_version, NULL);
239 if ((((version >> 16) == major) && ((version & 0xffff) >= minor)) ||
240 ((version >> 16) > major))
245 #define CPUID_THERM_POWER_LEAF 6
246 #define APERFMPERF_PRESENT 0
248 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
249 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
251 static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
252 static __read_mostly unsigned int cpuid_leaf5_ecx_val;
253 static __read_mostly unsigned int cpuid_leaf5_edx_val;
255 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
256 unsigned int *cx, unsigned int *dx)
258 unsigned maskebx = ~0;
259 unsigned maskecx = ~0;
260 unsigned maskedx = ~0;
263 * Mask out inconvenient features, to try and disable as many
264 * unsupported kernel subsystems as possible.
268 maskecx = cpuid_leaf1_ecx_mask;
269 setecx = cpuid_leaf1_ecx_set_mask;
270 maskedx = cpuid_leaf1_edx_mask;
273 case CPUID_MWAIT_LEAF:
274 /* Synthesize the values.. */
277 *cx = cpuid_leaf5_ecx_val;
278 *dx = cpuid_leaf5_edx_val;
281 case CPUID_THERM_POWER_LEAF:
282 /* Disabling APERFMPERF for kernel usage */
283 maskecx = ~(1 << APERFMPERF_PRESENT);
287 /* Suppress extended topology stuff */
292 asm(XEN_EMULATE_PREFIX "cpuid"
297 : "0" (*ax), "2" (*cx));
306 static bool __init xen_check_mwait(void)
309 struct xen_platform_op op = {
310 .cmd = XENPF_set_processor_pminfo,
311 .u.set_pminfo.id = -1,
312 .u.set_pminfo.type = XEN_PM_PDC,
315 unsigned int ax, bx, cx, dx;
316 unsigned int mwait_mask;
318 /* We need to determine whether it is OK to expose the MWAIT
319 * capability to the kernel to harvest deeper than C3 states from ACPI
320 * _CST using the processor_harvest_xen.c module. For this to work, we
321 * need to gather the MWAIT_LEAF values (which the cstate.c code
322 * checks against). The hypervisor won't expose the MWAIT flag because
323 * it would break backwards compatibility; so we will find out directly
324 * from the hardware and hypercall.
326 if (!xen_initial_domain())
330 * When running under platform earlier than Xen4.2, do not expose
331 * mwait, to avoid the risk of loading native acpi pad driver
333 if (!xen_running_on_version_or_later(4, 2))
339 native_cpuid(&ax, &bx, &cx, &dx);
341 mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
342 (1 << (X86_FEATURE_MWAIT % 32));
344 if ((cx & mwait_mask) != mwait_mask)
347 /* We need to emulate the MWAIT_LEAF and for that we need both
348 * ecx and edx. The hypercall provides only partial information.
351 ax = CPUID_MWAIT_LEAF;
356 native_cpuid(&ax, &bx, &cx, &dx);
358 /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
359 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
361 buf[0] = ACPI_PDC_REVISION_ID;
363 buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
365 set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
367 if ((HYPERVISOR_dom0_op(&op) == 0) &&
368 (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
369 cpuid_leaf5_ecx_val = cx;
370 cpuid_leaf5_edx_val = dx;
377 static void __init xen_init_cpuid_mask(void)
379 unsigned int ax, bx, cx, dx;
380 unsigned int xsave_mask;
382 cpuid_leaf1_edx_mask =
383 ~((1 << X86_FEATURE_MTRR) | /* disable MTRR */
384 (1 << X86_FEATURE_ACC)); /* thermal monitoring */
386 if (!xen_initial_domain())
387 cpuid_leaf1_edx_mask &=
388 ~((1 << X86_FEATURE_APIC) | /* disable local APIC */
389 (1 << X86_FEATURE_ACPI)); /* disable ACPI */
392 xen_cpuid(&ax, &bx, &cx, &dx);
395 (1 << (X86_FEATURE_XSAVE % 32)) |
396 (1 << (X86_FEATURE_OSXSAVE % 32));
398 /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
399 if ((cx & xsave_mask) != xsave_mask)
400 cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
401 if (xen_check_mwait())
402 cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
405 static void xen_set_debugreg(int reg, unsigned long val)
407 HYPERVISOR_set_debugreg(reg, val);
410 static unsigned long xen_get_debugreg(int reg)
412 return HYPERVISOR_get_debugreg(reg);
415 static void xen_end_context_switch(struct task_struct *next)
418 paravirt_end_context_switch(next);
421 static unsigned long xen_store_tr(void)
427 * Set the page permissions for a particular virtual address. If the
428 * address is a vmalloc mapping (or other non-linear mapping), then
429 * find the linear mapping of the page and also set its protections to
432 static void set_aliased_prot(void *v, pgprot_t prot)
440 ptep = lookup_address((unsigned long)v, &level);
441 BUG_ON(ptep == NULL);
443 pfn = pte_pfn(*ptep);
444 page = pfn_to_page(pfn);
446 pte = pfn_pte(pfn, prot);
448 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
451 if (!PageHighMem(page)) {
452 void *av = __va(PFN_PHYS(pfn));
455 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
461 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
463 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
466 for(i = 0; i < entries; i += entries_per_page)
467 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
470 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
472 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
475 for(i = 0; i < entries; i += entries_per_page)
476 set_aliased_prot(ldt + i, PAGE_KERNEL);
479 static void xen_set_ldt(const void *addr, unsigned entries)
481 struct mmuext_op *op;
482 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
484 trace_xen_cpu_set_ldt(addr, entries);
487 op->cmd = MMUEXT_SET_LDT;
488 op->arg1.linear_addr = (unsigned long)addr;
489 op->arg2.nr_ents = entries;
491 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
493 xen_mc_issue(PARAVIRT_LAZY_CPU);
496 static void xen_load_gdt(const struct desc_ptr *dtr)
498 unsigned long va = dtr->address;
499 unsigned int size = dtr->size + 1;
500 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
501 unsigned long frames[pages];
505 * A GDT can be up to 64k in size, which corresponds to 8192
506 * 8-byte entries, or 16 4k pages..
509 BUG_ON(size > 65536);
510 BUG_ON(va & ~PAGE_MASK);
512 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
515 unsigned long pfn, mfn;
519 * The GDT is per-cpu and is in the percpu data area.
520 * That can be virtually mapped, so we need to do a
521 * page-walk to get the underlying MFN for the
522 * hypercall. The page can also be in the kernel's
523 * linear range, so we need to RO that mapping too.
525 ptep = lookup_address(va, &level);
526 BUG_ON(ptep == NULL);
528 pfn = pte_pfn(*ptep);
529 mfn = pfn_to_mfn(pfn);
530 virt = __va(PFN_PHYS(pfn));
534 make_lowmem_page_readonly((void *)va);
535 make_lowmem_page_readonly(virt);
538 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
543 * load_gdt for early boot, when the gdt is only mapped once
545 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
547 unsigned long va = dtr->address;
548 unsigned int size = dtr->size + 1;
549 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
550 unsigned long frames[pages];
554 * A GDT can be up to 64k in size, which corresponds to 8192
555 * 8-byte entries, or 16 4k pages..
558 BUG_ON(size > 65536);
559 BUG_ON(va & ~PAGE_MASK);
561 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
563 unsigned long pfn, mfn;
565 pfn = virt_to_pfn(va);
566 mfn = pfn_to_mfn(pfn);
568 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
570 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
576 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
580 static inline bool desc_equal(const struct desc_struct *d1,
581 const struct desc_struct *d2)
583 return d1->a == d2->a && d1->b == d2->b;
586 static void load_TLS_descriptor(struct thread_struct *t,
587 unsigned int cpu, unsigned int i)
589 struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
590 struct desc_struct *gdt;
592 struct multicall_space mc;
594 if (desc_equal(shadow, &t->tls_array[i]))
597 *shadow = t->tls_array[i];
599 gdt = get_cpu_gdt_table(cpu);
600 maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
601 mc = __xen_mc_entry(0);
603 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
606 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
609 * XXX sleazy hack: If we're being called in a lazy-cpu zone
610 * and lazy gs handling is enabled, it means we're in a
611 * context switch, and %gs has just been saved. This means we
612 * can zero it out to prevent faults on exit from the
613 * hypervisor if the next process has no %gs. Either way, it
614 * has been saved, and the new value will get loaded properly.
615 * This will go away as soon as Xen has been modified to not
616 * save/restore %gs for normal hypercalls.
618 * On x86_64, this hack is not used for %gs, because gs points
619 * to KERNEL_GS_BASE (and uses it for PDA references), so we
620 * must not zero %gs on x86_64
622 * For x86_64, we need to zero %fs, otherwise we may get an
623 * exception between the new %fs descriptor being loaded and
624 * %fs being effectively cleared at __switch_to().
626 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
636 load_TLS_descriptor(t, cpu, 0);
637 load_TLS_descriptor(t, cpu, 1);
638 load_TLS_descriptor(t, cpu, 2);
640 xen_mc_issue(PARAVIRT_LAZY_CPU);
644 static void xen_load_gs_index(unsigned int idx)
646 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
651 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
654 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
655 u64 entry = *(u64 *)ptr;
657 trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
662 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
668 static int cvt_gate_to_trap(int vector, const gate_desc *val,
669 struct trap_info *info)
673 if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
676 info->vector = vector;
678 addr = gate_offset(*val);
681 * Look for known traps using IST, and substitute them
682 * appropriately. The debugger ones are the only ones we care
683 * about. Xen will handle faults like double_fault,
684 * so we should never see them. Warn if
685 * there's an unexpected IST-using fault handler.
687 if (addr == (unsigned long)debug)
688 addr = (unsigned long)xen_debug;
689 else if (addr == (unsigned long)int3)
690 addr = (unsigned long)xen_int3;
691 else if (addr == (unsigned long)stack_segment)
692 addr = (unsigned long)xen_stack_segment;
693 else if (addr == (unsigned long)double_fault ||
694 addr == (unsigned long)nmi) {
695 /* Don't need to handle these */
697 #ifdef CONFIG_X86_MCE
698 } else if (addr == (unsigned long)machine_check) {
700 * when xen hypervisor inject vMCE to guest,
701 * use native mce handler to handle it
706 /* Some other trap using IST? */
707 if (WARN_ON(val->ist != 0))
710 #endif /* CONFIG_X86_64 */
711 info->address = addr;
713 info->cs = gate_segment(*val);
714 info->flags = val->dpl;
715 /* interrupt gates clear IF */
716 if (val->type == GATE_INTERRUPT)
717 info->flags |= 1 << 2;
722 /* Locations of each CPU's IDT */
723 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
725 /* Set an IDT entry. If the entry is part of the current IDT, then
727 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
729 unsigned long p = (unsigned long)&dt[entrynum];
730 unsigned long start, end;
732 trace_xen_cpu_write_idt_entry(dt, entrynum, g);
736 start = __this_cpu_read(idt_desc.address);
737 end = start + __this_cpu_read(idt_desc.size) + 1;
741 native_write_idt_entry(dt, entrynum, g);
743 if (p >= start && (p + 8) <= end) {
744 struct trap_info info[2];
748 if (cvt_gate_to_trap(entrynum, g, &info[0]))
749 if (HYPERVISOR_set_trap_table(info))
756 static void xen_convert_trap_info(const struct desc_ptr *desc,
757 struct trap_info *traps)
759 unsigned in, out, count;
761 count = (desc->size+1) / sizeof(gate_desc);
764 for (in = out = 0; in < count; in++) {
765 gate_desc *entry = (gate_desc*)(desc->address) + in;
767 if (cvt_gate_to_trap(in, entry, &traps[out]))
770 traps[out].address = 0;
773 void xen_copy_trap_info(struct trap_info *traps)
775 const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
777 xen_convert_trap_info(desc, traps);
780 /* Load a new IDT into Xen. In principle this can be per-CPU, so we
781 hold a spinlock to protect the static traps[] array (static because
782 it avoids allocation, and saves stack space). */
783 static void xen_load_idt(const struct desc_ptr *desc)
785 static DEFINE_SPINLOCK(lock);
786 static struct trap_info traps[257];
788 trace_xen_cpu_load_idt(desc);
792 __get_cpu_var(idt_desc) = *desc;
794 xen_convert_trap_info(desc, traps);
797 if (HYPERVISOR_set_trap_table(traps))
803 /* Write a GDT descriptor entry. Ignore LDT descriptors, since
804 they're handled differently. */
805 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
806 const void *desc, int type)
808 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
819 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
822 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
832 * Version of write_gdt_entry for use at early boot-time needed to
833 * update an entry as simply as possible.
835 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
836 const void *desc, int type)
838 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
847 xmaddr_t maddr = virt_to_machine(&dt[entry]);
849 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
850 dt[entry] = *(struct desc_struct *)desc;
856 static void xen_load_sp0(struct tss_struct *tss,
857 struct thread_struct *thread)
859 struct multicall_space mcs;
861 mcs = xen_mc_entry(0);
862 MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
863 xen_mc_issue(PARAVIRT_LAZY_CPU);
866 static void xen_set_iopl_mask(unsigned mask)
868 struct physdev_set_iopl set_iopl;
870 /* Force the change at ring 0. */
871 set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
872 HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
875 static void xen_io_delay(void)
879 #ifdef CONFIG_X86_LOCAL_APIC
880 static unsigned long xen_set_apic_id(unsigned int x)
885 static unsigned int xen_get_apic_id(unsigned long x)
887 return ((x)>>24) & 0xFFu;
889 static u32 xen_apic_read(u32 reg)
891 struct xen_platform_op op = {
892 .cmd = XENPF_get_cpuinfo,
893 .interface_version = XENPF_INTERFACE_VERSION,
894 .u.pcpu_info.xen_cpuid = 0,
898 /* Shouldn't need this as APIC is turned off for PV, and we only
899 * get called on the bootup processor. But just in case. */
900 if (!xen_initial_domain() || smp_processor_id())
909 ret = HYPERVISOR_dom0_op(&op);
913 return op.u.pcpu_info.apic_id << 24;
916 static void xen_apic_write(u32 reg, u32 val)
918 /* Warn to see if there's any stray references */
922 static u64 xen_apic_icr_read(void)
927 static void xen_apic_icr_write(u32 low, u32 id)
929 /* Warn to see if there's any stray references */
933 static void xen_apic_wait_icr_idle(void)
938 static u32 xen_safe_apic_wait_icr_idle(void)
943 static void set_xen_basic_apic_ops(void)
945 apic->read = xen_apic_read;
946 apic->write = xen_apic_write;
947 apic->icr_read = xen_apic_icr_read;
948 apic->icr_write = xen_apic_icr_write;
949 apic->wait_icr_idle = xen_apic_wait_icr_idle;
950 apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
951 apic->set_apic_id = xen_set_apic_id;
952 apic->get_apic_id = xen_get_apic_id;
955 apic->send_IPI_allbutself = xen_send_IPI_allbutself;
956 apic->send_IPI_mask_allbutself = xen_send_IPI_mask_allbutself;
957 apic->send_IPI_mask = xen_send_IPI_mask;
958 apic->send_IPI_all = xen_send_IPI_all;
959 apic->send_IPI_self = xen_send_IPI_self;
965 static void xen_clts(void)
967 struct multicall_space mcs;
969 mcs = xen_mc_entry(0);
971 MULTI_fpu_taskswitch(mcs.mc, 0);
973 xen_mc_issue(PARAVIRT_LAZY_CPU);
976 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
978 static unsigned long xen_read_cr0(void)
980 unsigned long cr0 = this_cpu_read(xen_cr0_value);
982 if (unlikely(cr0 == 0)) {
983 cr0 = native_read_cr0();
984 this_cpu_write(xen_cr0_value, cr0);
990 static void xen_write_cr0(unsigned long cr0)
992 struct multicall_space mcs;
994 this_cpu_write(xen_cr0_value, cr0);
996 /* Only pay attention to cr0.TS; everything else is
998 mcs = xen_mc_entry(0);
1000 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
1002 xen_mc_issue(PARAVIRT_LAZY_CPU);
1005 static void xen_write_cr4(unsigned long cr4)
1007 cr4 &= ~X86_CR4_PGE;
1008 cr4 &= ~X86_CR4_PSE;
1010 native_write_cr4(cr4);
1012 #ifdef CONFIG_X86_64
1013 static inline unsigned long xen_read_cr8(void)
1017 static inline void xen_write_cr8(unsigned long val)
1022 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
1029 #ifdef CONFIG_X86_64
1033 case MSR_FS_BASE: which = SEGBASE_FS; goto set;
1034 case MSR_KERNEL_GS_BASE: which = SEGBASE_GS_USER; goto set;
1035 case MSR_GS_BASE: which = SEGBASE_GS_KERNEL; goto set;
1038 base = ((u64)high << 32) | low;
1039 if (HYPERVISOR_set_segment_base(which, base) != 0)
1047 case MSR_SYSCALL_MASK:
1048 case MSR_IA32_SYSENTER_CS:
1049 case MSR_IA32_SYSENTER_ESP:
1050 case MSR_IA32_SYSENTER_EIP:
1051 /* Fast syscall setup is all done in hypercalls, so
1052 these are all ignored. Stub them out here to stop
1053 Xen console noise. */
1056 case MSR_IA32_CR_PAT:
1057 if (smp_processor_id() == 0)
1058 xen_set_pat(((u64)high << 32) | low);
1062 ret = native_write_msr_safe(msr, low, high);
1068 void xen_setup_shared_info(void)
1070 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1071 set_fixmap(FIX_PARAVIRT_BOOTMAP,
1072 xen_start_info->shared_info);
1074 HYPERVISOR_shared_info =
1075 (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
1077 HYPERVISOR_shared_info =
1078 (struct shared_info *)__va(xen_start_info->shared_info);
1081 /* In UP this is as good a place as any to set up shared info */
1082 xen_setup_vcpu_info_placement();
1085 xen_setup_mfn_list_list();
1088 /* This is called once we have the cpu_possible_mask */
1089 void xen_setup_vcpu_info_placement(void)
1093 for_each_possible_cpu(cpu)
1094 xen_vcpu_setup(cpu);
1096 /* xen_vcpu_setup managed to place the vcpu_info within the
1097 percpu area for all cpus, so make use of it */
1098 if (have_vcpu_info_placement) {
1099 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1100 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1101 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1102 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1103 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1107 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1108 unsigned long addr, unsigned len)
1110 char *start, *end, *reloc;
1113 start = end = reloc = NULL;
1115 #define SITE(op, x) \
1116 case PARAVIRT_PATCH(op.x): \
1117 if (have_vcpu_info_placement) { \
1118 start = (char *)xen_##x##_direct; \
1119 end = xen_##x##_direct_end; \
1120 reloc = xen_##x##_direct_reloc; \
1125 SITE(pv_irq_ops, irq_enable);
1126 SITE(pv_irq_ops, irq_disable);
1127 SITE(pv_irq_ops, save_fl);
1128 SITE(pv_irq_ops, restore_fl);
1132 if (start == NULL || (end-start) > len)
1135 ret = paravirt_patch_insns(insnbuf, len, start, end);
1137 /* Note: because reloc is assigned from something that
1138 appears to be an array, gcc assumes it's non-null,
1139 but doesn't know its relationship with start and
1141 if (reloc > start && reloc < end) {
1142 int reloc_off = reloc - start;
1143 long *relocp = (long *)(insnbuf + reloc_off);
1144 long delta = start - (char *)addr;
1152 ret = paravirt_patch_default(type, clobbers, insnbuf,
1160 static const struct pv_info xen_info __initconst = {
1161 .paravirt_enabled = 1,
1162 .shared_kernel_pmd = 0,
1164 #ifdef CONFIG_X86_64
1165 .extra_user_64bit_cs = FLAT_USER_CS64,
1171 static const struct pv_init_ops xen_init_ops __initconst = {
1175 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1178 .set_debugreg = xen_set_debugreg,
1179 .get_debugreg = xen_get_debugreg,
1183 .read_cr0 = xen_read_cr0,
1184 .write_cr0 = xen_write_cr0,
1186 .read_cr4 = native_read_cr4,
1187 .read_cr4_safe = native_read_cr4_safe,
1188 .write_cr4 = xen_write_cr4,
1190 #ifdef CONFIG_X86_64
1191 .read_cr8 = xen_read_cr8,
1192 .write_cr8 = xen_write_cr8,
1195 .wbinvd = native_wbinvd,
1197 .read_msr = native_read_msr_safe,
1198 .write_msr = xen_write_msr_safe,
1200 .read_tsc = native_read_tsc,
1201 .read_pmc = native_read_pmc,
1203 .read_tscp = native_read_tscp,
1206 .irq_enable_sysexit = xen_sysexit,
1207 #ifdef CONFIG_X86_64
1208 .usergs_sysret32 = xen_sysret32,
1209 .usergs_sysret64 = xen_sysret64,
1212 .load_tr_desc = paravirt_nop,
1213 .set_ldt = xen_set_ldt,
1214 .load_gdt = xen_load_gdt,
1215 .load_idt = xen_load_idt,
1216 .load_tls = xen_load_tls,
1217 #ifdef CONFIG_X86_64
1218 .load_gs_index = xen_load_gs_index,
1221 .alloc_ldt = xen_alloc_ldt,
1222 .free_ldt = xen_free_ldt,
1224 .store_gdt = native_store_gdt,
1225 .store_idt = native_store_idt,
1226 .store_tr = xen_store_tr,
1228 .write_ldt_entry = xen_write_ldt_entry,
1229 .write_gdt_entry = xen_write_gdt_entry,
1230 .write_idt_entry = xen_write_idt_entry,
1231 .load_sp0 = xen_load_sp0,
1233 .set_iopl_mask = xen_set_iopl_mask,
1234 .io_delay = xen_io_delay,
1236 /* Xen takes care of %gs when switching to usermode for us */
1237 .swapgs = paravirt_nop,
1239 .start_context_switch = paravirt_start_context_switch,
1240 .end_context_switch = xen_end_context_switch,
1243 static const struct pv_apic_ops xen_apic_ops __initconst = {
1244 #ifdef CONFIG_X86_LOCAL_APIC
1245 .startup_ipi_hook = paravirt_nop,
1249 static void xen_reboot(int reason)
1251 struct sched_shutdown r = { .reason = reason };
1253 if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1257 static void xen_restart(char *msg)
1259 xen_reboot(SHUTDOWN_reboot);
1262 static void xen_emergency_restart(void)
1264 xen_reboot(SHUTDOWN_reboot);
1267 static void xen_machine_halt(void)
1269 xen_reboot(SHUTDOWN_poweroff);
1272 static void xen_machine_power_off(void)
1276 xen_reboot(SHUTDOWN_poweroff);
1279 static void xen_crash_shutdown(struct pt_regs *regs)
1281 xen_reboot(SHUTDOWN_crash);
1285 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1287 xen_reboot(SHUTDOWN_crash);
1291 static struct notifier_block xen_panic_block = {
1292 .notifier_call= xen_panic_event,
1295 int xen_panic_handler_init(void)
1297 atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1301 static const struct machine_ops xen_machine_ops __initconst = {
1302 .restart = xen_restart,
1303 .halt = xen_machine_halt,
1304 .power_off = xen_machine_power_off,
1305 .shutdown = xen_machine_halt,
1306 .crash_shutdown = xen_crash_shutdown,
1307 .emergency_restart = xen_emergency_restart,
1310 static void __init xen_boot_params_init_edd(void)
1312 #if IS_ENABLED(CONFIG_EDD)
1313 struct xen_platform_op op;
1314 struct edd_info *edd_info;
1319 edd_info = boot_params.eddbuf;
1320 mbr_signature = boot_params.edd_mbr_sig_buffer;
1322 op.cmd = XENPF_firmware_info;
1324 op.u.firmware_info.type = XEN_FW_DISK_INFO;
1325 for (nr = 0; nr < EDDMAXNR; nr++) {
1326 struct edd_info *info = edd_info + nr;
1328 op.u.firmware_info.index = nr;
1329 info->params.length = sizeof(info->params);
1330 set_xen_guest_handle(op.u.firmware_info.u.disk_info.edd_params,
1332 ret = HYPERVISOR_dom0_op(&op);
1336 #define C(x) info->x = op.u.firmware_info.u.disk_info.x
1339 C(interface_support);
1340 C(legacy_max_cylinder);
1342 C(legacy_sectors_per_track);
1345 boot_params.eddbuf_entries = nr;
1347 op.u.firmware_info.type = XEN_FW_DISK_MBR_SIGNATURE;
1348 for (nr = 0; nr < EDD_MBR_SIG_MAX; nr++) {
1349 op.u.firmware_info.index = nr;
1350 ret = HYPERVISOR_dom0_op(&op);
1353 mbr_signature[nr] = op.u.firmware_info.u.disk_mbr_signature.mbr_signature;
1355 boot_params.edd_mbr_sig_buf_entries = nr;
1360 * Set up the GDT and segment registers for -fstack-protector. Until
1361 * we do this, we have to be careful not to call any stack-protected
1362 * function, which is most of the kernel.
1364 static void __init xen_setup_stackprotector(void)
1366 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1367 pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1369 setup_stack_canary_segment(0);
1370 switch_to_new_gdt(0);
1372 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1373 pv_cpu_ops.load_gdt = xen_load_gdt;
1376 /* First C function to be called on Xen boot */
1377 asmlinkage void __init xen_start_kernel(void)
1379 struct physdev_set_iopl set_iopl;
1382 if (!xen_start_info)
1385 xen_domain_type = XEN_PV_DOMAIN;
1387 xen_setup_machphys_mapping();
1389 /* Install Xen paravirt ops */
1391 pv_init_ops = xen_init_ops;
1392 pv_cpu_ops = xen_cpu_ops;
1393 pv_apic_ops = xen_apic_ops;
1395 x86_init.resources.memory_setup = xen_memory_setup;
1396 x86_init.oem.arch_setup = xen_arch_setup;
1397 x86_init.oem.banner = xen_banner;
1399 xen_init_time_ops();
1402 * Set up some pagetable state before starting to set any ptes.
1407 /* Prevent unwanted bits from being set in PTEs. */
1408 __supported_pte_mask &= ~_PAGE_GLOBAL;
1410 if (!xen_initial_domain())
1412 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1414 __supported_pte_mask |= _PAGE_IOMAP;
1417 * Prevent page tables from being allocated in highmem, even
1418 * if CONFIG_HIGHPTE is enabled.
1420 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1422 /* Work out if we support NX */
1425 xen_setup_features();
1428 if (!xen_feature(XENFEAT_auto_translated_physmap))
1429 xen_build_dynamic_phys_to_machine();
1432 * Set up kernel GDT and segment registers, mainly so that
1433 * -fstack-protector code can be executed.
1435 xen_setup_stackprotector();
1438 xen_init_cpuid_mask();
1440 #ifdef CONFIG_X86_LOCAL_APIC
1442 * set up the basic apic ops.
1444 set_xen_basic_apic_ops();
1447 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1448 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1449 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1452 machine_ops = xen_machine_ops;
1455 * The only reliable way to retain the initial address of the
1456 * percpu gdt_page is to remember it here, so we can go and
1457 * mark it RW later, when the initial percpu area is freed.
1459 xen_initial_gdt = &per_cpu(gdt_page, 0);
1463 #ifdef CONFIG_ACPI_NUMA
1465 * The pages we from Xen are not related to machine pages, so
1466 * any NUMA information the kernel tries to get from ACPI will
1467 * be meaningless. Prevent it from trying.
1471 #ifdef CONFIG_X86_PAT
1473 * For right now disable the PAT. We should remove this once
1474 * git commit 8eaffa67b43e99ae581622c5133e20b0f48bcef1
1475 * (xen/pat: Disable PAT support for now) is reverted.
1479 /* Don't do the full vcpu_info placement stuff until we have a
1480 possible map and a non-dummy shared_info. */
1481 per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1483 local_irq_disable();
1484 early_boot_irqs_disabled = true;
1486 xen_raw_console_write("mapping kernel into physical memory\n");
1487 xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base, xen_start_info->nr_pages);
1489 /* Allocate and initialize top and mid mfn levels for p2m structure */
1490 xen_build_mfn_list_list();
1492 /* keep using Xen gdt for now; no urgent need to change it */
1494 #ifdef CONFIG_X86_32
1495 pv_info.kernel_rpl = 1;
1496 if (xen_feature(XENFEAT_supervisor_mode_kernel))
1497 pv_info.kernel_rpl = 0;
1499 pv_info.kernel_rpl = 0;
1501 /* set the limit of our address space */
1504 /* We used to do this in xen_arch_setup, but that is too late on AMD
1505 * were early_cpu_init (run before ->arch_setup()) calls early_amd_init
1506 * which pokes 0xcf8 port.
1509 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1511 xen_raw_printk("physdev_op failed %d\n", rc);
1513 #ifdef CONFIG_X86_32
1514 /* set up basic CPUID stuff */
1515 cpu_detect(&new_cpu_data);
1516 new_cpu_data.hard_math = 1;
1517 new_cpu_data.wp_works_ok = 1;
1518 new_cpu_data.x86_capability[0] = cpuid_edx(1);
1521 /* Poke various useful things into boot_params */
1522 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1523 boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1524 ? __pa(xen_start_info->mod_start) : 0;
1525 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1526 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1528 if (!xen_initial_domain()) {
1529 add_preferred_console("xenboot", 0, NULL);
1530 add_preferred_console("tty", 0, NULL);
1531 add_preferred_console("hvc", 0, NULL);
1533 x86_init.pci.arch_init = pci_xen_init;
1535 const struct dom0_vga_console_info *info =
1536 (void *)((char *)xen_start_info +
1537 xen_start_info->console.dom0.info_off);
1538 struct xen_platform_op op = {
1539 .cmd = XENPF_firmware_info,
1540 .interface_version = XENPF_INTERFACE_VERSION,
1541 .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
1544 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1545 xen_start_info->console.domU.mfn = 0;
1546 xen_start_info->console.domU.evtchn = 0;
1548 if (HYPERVISOR_dom0_op(&op) == 0)
1549 boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
1553 /* Make sure ACS will be enabled */
1556 xen_acpi_sleep_register();
1558 /* Avoid searching for BIOS MP tables */
1559 x86_init.mpparse.find_smp_config = x86_init_noop;
1560 x86_init.mpparse.get_smp_config = x86_init_uint_noop;
1562 xen_boot_params_init_edd();
1565 /* PCI BIOS service won't work from a PV guest. */
1566 pci_probe &= ~PCI_PROBE_BIOS;
1568 xen_raw_console_write("about to get started...\n");
1570 xen_setup_runstate_info(0);
1572 /* Start the world */
1573 #ifdef CONFIG_X86_32
1574 i386_start_kernel();
1576 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1580 void __ref xen_hvm_init_shared_info(void)
1583 struct xen_add_to_physmap xatp;
1584 static struct shared_info *shared_info_page = 0;
1586 if (!shared_info_page)
1587 shared_info_page = (struct shared_info *)
1588 extend_brk(PAGE_SIZE, PAGE_SIZE);
1589 xatp.domid = DOMID_SELF;
1591 xatp.space = XENMAPSPACE_shared_info;
1592 xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT;
1593 if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1596 HYPERVISOR_shared_info = (struct shared_info *)shared_info_page;
1598 /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1599 * page, we use it in the event channel upcall and in some pvclock
1600 * related functions. We don't need the vcpu_info placement
1601 * optimizations because we don't use any pv_mmu or pv_irq op on
1603 * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is
1604 * online but xen_hvm_init_shared_info is run at resume time too and
1605 * in that case multiple vcpus might be online. */
1606 for_each_online_cpu(cpu) {
1607 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1611 #ifdef CONFIG_XEN_PVHVM
1612 static void __init init_hvm_pv_info(void)
1615 uint32_t eax, ebx, ecx, edx, pages, msr, base;
1618 base = xen_cpuid_base();
1619 cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1622 minor = eax & 0xffff;
1623 printk(KERN_INFO "Xen version %d.%d.\n", major, minor);
1625 cpuid(base + 2, &pages, &msr, &ecx, &edx);
1627 pfn = __pa(hypercall_page);
1628 wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1630 xen_setup_features();
1632 pv_info.name = "Xen HVM";
1634 xen_domain_type = XEN_HVM_DOMAIN;
1637 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self,
1638 unsigned long action, void *hcpu)
1640 int cpu = (long)hcpu;
1642 case CPU_UP_PREPARE:
1643 xen_vcpu_setup(cpu);
1644 if (xen_have_vector_callback) {
1645 xen_init_lock_cpu(cpu);
1646 if (xen_feature(XENFEAT_hvm_safe_pvclock))
1647 xen_setup_timer(cpu);
1656 static struct notifier_block xen_hvm_cpu_notifier __cpuinitdata = {
1657 .notifier_call = xen_hvm_cpu_notify,
1660 static void __init xen_hvm_guest_init(void)
1664 xen_hvm_init_shared_info();
1666 if (xen_feature(XENFEAT_hvm_callback_vector))
1667 xen_have_vector_callback = 1;
1669 register_cpu_notifier(&xen_hvm_cpu_notifier);
1670 xen_unplug_emulated_devices();
1671 x86_init.irqs.intr_init = xen_init_IRQ;
1672 xen_hvm_init_time_ops();
1673 xen_hvm_init_mmu_ops();
1676 static bool __init xen_hvm_platform(void)
1678 if (xen_pv_domain())
1681 if (!xen_cpuid_base())
1687 bool xen_hvm_need_lapic(void)
1689 if (xen_pv_domain())
1691 if (!xen_hvm_domain())
1693 if (xen_feature(XENFEAT_hvm_pirqs) && xen_have_vector_callback)
1697 EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1699 const struct hypervisor_x86 x86_hyper_xen_hvm __refconst = {
1701 .detect = xen_hvm_platform,
1702 .init_platform = xen_hvm_guest_init,
1703 .x2apic_available = xen_x2apic_para_available,
1705 EXPORT_SYMBOL(x86_hyper_xen_hvm);