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 EXPORT_SYMBOL_GPL(hypercall_page);
86 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
87 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
89 enum xen_domain_type xen_domain_type = XEN_NATIVE;
90 EXPORT_SYMBOL_GPL(xen_domain_type);
92 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
93 EXPORT_SYMBOL(machine_to_phys_mapping);
94 unsigned long machine_to_phys_nr;
95 EXPORT_SYMBOL(machine_to_phys_nr);
97 struct start_info *xen_start_info;
98 EXPORT_SYMBOL_GPL(xen_start_info);
100 struct shared_info xen_dummy_shared_info;
102 void *xen_initial_gdt;
104 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
105 __read_mostly int xen_have_vector_callback;
106 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
109 * Point at some empty memory to start with. We map the real shared_info
110 * page as soon as fixmap is up and running.
112 struct shared_info *HYPERVISOR_shared_info = &xen_dummy_shared_info;
115 * Flag to determine whether vcpu info placement is available on all
116 * VCPUs. We assume it is to start with, and then set it to zero on
117 * the first failure. This is because it can succeed on some VCPUs
118 * and not others, since it can involve hypervisor memory allocation,
119 * or because the guest failed to guarantee all the appropriate
120 * constraints on all VCPUs (ie buffer can't cross a page boundary).
122 * Note that any particular CPU may be using a placed vcpu structure,
123 * but we can only optimise if the all are.
125 * 0: not available, 1: available
127 static int have_vcpu_info_placement = 1;
130 struct desc_struct desc[3];
134 * Updating the 3 TLS descriptors in the GDT on every task switch is
135 * surprisingly expensive so we avoid updating them if they haven't
136 * changed. Since Xen writes different descriptors than the one
137 * passed in the update_descriptor hypercall we keep shadow copies to
140 static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
142 static void clamp_max_cpus(void)
145 if (setup_max_cpus > MAX_VIRT_CPUS)
146 setup_max_cpus = MAX_VIRT_CPUS;
150 static void xen_vcpu_setup(int cpu)
152 struct vcpu_register_vcpu_info info;
154 struct vcpu_info *vcpup;
156 BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
158 if (cpu < MAX_VIRT_CPUS)
159 per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
161 if (!have_vcpu_info_placement) {
162 if (cpu >= MAX_VIRT_CPUS)
167 vcpup = &per_cpu(xen_vcpu_info, cpu);
168 info.mfn = arbitrary_virt_to_mfn(vcpup);
169 info.offset = offset_in_page(vcpup);
171 /* Check to see if the hypervisor will put the vcpu_info
172 structure where we want it, which allows direct access via
173 a percpu-variable. */
174 err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
177 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
178 have_vcpu_info_placement = 0;
181 /* This cpu is using the registered vcpu info, even if
182 later ones fail to. */
183 per_cpu(xen_vcpu, cpu) = vcpup;
188 * On restore, set the vcpu placement up again.
189 * If it fails, then we're in a bad state, since
190 * we can't back out from using it...
192 void xen_vcpu_restore(void)
196 for_each_online_cpu(cpu) {
197 bool other_cpu = (cpu != smp_processor_id());
200 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
203 xen_setup_runstate_info(cpu);
205 if (have_vcpu_info_placement)
209 HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
214 static void __init xen_banner(void)
216 unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
217 struct xen_extraversion extra;
218 HYPERVISOR_xen_version(XENVER_extraversion, &extra);
220 printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
222 printk(KERN_INFO "Xen version: %d.%d%s%s\n",
223 version >> 16, version & 0xffff, extra.extraversion,
224 xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
226 /* Check if running on Xen version (major, minor) or later */
228 xen_running_on_version_or_later(unsigned int major, unsigned int minor)
230 unsigned int version;
235 version = HYPERVISOR_xen_version(XENVER_version, NULL);
236 if ((((version >> 16) == major) && ((version & 0xffff) >= minor)) ||
237 ((version >> 16) > major))
242 #define CPUID_THERM_POWER_LEAF 6
243 #define APERFMPERF_PRESENT 0
245 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
246 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
248 static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
249 static __read_mostly unsigned int cpuid_leaf5_ecx_val;
250 static __read_mostly unsigned int cpuid_leaf5_edx_val;
252 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
253 unsigned int *cx, unsigned int *dx)
255 unsigned maskebx = ~0;
256 unsigned maskecx = ~0;
257 unsigned maskedx = ~0;
260 * Mask out inconvenient features, to try and disable as many
261 * unsupported kernel subsystems as possible.
265 maskecx = cpuid_leaf1_ecx_mask;
266 setecx = cpuid_leaf1_ecx_set_mask;
267 maskedx = cpuid_leaf1_edx_mask;
270 case CPUID_MWAIT_LEAF:
271 /* Synthesize the values.. */
274 *cx = cpuid_leaf5_ecx_val;
275 *dx = cpuid_leaf5_edx_val;
278 case CPUID_THERM_POWER_LEAF:
279 /* Disabling APERFMPERF for kernel usage */
280 maskecx = ~(1 << APERFMPERF_PRESENT);
284 /* Suppress extended topology stuff */
289 asm(XEN_EMULATE_PREFIX "cpuid"
294 : "0" (*ax), "2" (*cx));
303 static bool __init xen_check_mwait(void)
306 struct xen_platform_op op = {
307 .cmd = XENPF_set_processor_pminfo,
308 .u.set_pminfo.id = -1,
309 .u.set_pminfo.type = XEN_PM_PDC,
312 unsigned int ax, bx, cx, dx;
313 unsigned int mwait_mask;
315 /* We need to determine whether it is OK to expose the MWAIT
316 * capability to the kernel to harvest deeper than C3 states from ACPI
317 * _CST using the processor_harvest_xen.c module. For this to work, we
318 * need to gather the MWAIT_LEAF values (which the cstate.c code
319 * checks against). The hypervisor won't expose the MWAIT flag because
320 * it would break backwards compatibility; so we will find out directly
321 * from the hardware and hypercall.
323 if (!xen_initial_domain())
327 * When running under platform earlier than Xen4.2, do not expose
328 * mwait, to avoid the risk of loading native acpi pad driver
330 if (!xen_running_on_version_or_later(4, 2))
336 native_cpuid(&ax, &bx, &cx, &dx);
338 mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
339 (1 << (X86_FEATURE_MWAIT % 32));
341 if ((cx & mwait_mask) != mwait_mask)
344 /* We need to emulate the MWAIT_LEAF and for that we need both
345 * ecx and edx. The hypercall provides only partial information.
348 ax = CPUID_MWAIT_LEAF;
353 native_cpuid(&ax, &bx, &cx, &dx);
355 /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
356 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
358 buf[0] = ACPI_PDC_REVISION_ID;
360 buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
362 set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
364 if ((HYPERVISOR_dom0_op(&op) == 0) &&
365 (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
366 cpuid_leaf5_ecx_val = cx;
367 cpuid_leaf5_edx_val = dx;
374 static void __init xen_init_cpuid_mask(void)
376 unsigned int ax, bx, cx, dx;
377 unsigned int xsave_mask;
379 cpuid_leaf1_edx_mask =
380 ~((1 << X86_FEATURE_MTRR) | /* disable MTRR */
381 (1 << X86_FEATURE_ACC)); /* thermal monitoring */
383 if (!xen_initial_domain())
384 cpuid_leaf1_edx_mask &=
385 ~((1 << X86_FEATURE_APIC) | /* disable local APIC */
386 (1 << X86_FEATURE_ACPI)); /* disable ACPI */
389 xen_cpuid(&ax, &bx, &cx, &dx);
392 (1 << (X86_FEATURE_XSAVE % 32)) |
393 (1 << (X86_FEATURE_OSXSAVE % 32));
395 /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
396 if ((cx & xsave_mask) != xsave_mask)
397 cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
398 if (xen_check_mwait())
399 cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
402 static void xen_set_debugreg(int reg, unsigned long val)
404 HYPERVISOR_set_debugreg(reg, val);
407 static unsigned long xen_get_debugreg(int reg)
409 return HYPERVISOR_get_debugreg(reg);
412 static void xen_end_context_switch(struct task_struct *next)
415 paravirt_end_context_switch(next);
418 static unsigned long xen_store_tr(void)
424 * Set the page permissions for a particular virtual address. If the
425 * address is a vmalloc mapping (or other non-linear mapping), then
426 * find the linear mapping of the page and also set its protections to
429 static void set_aliased_prot(void *v, pgprot_t prot)
437 ptep = lookup_address((unsigned long)v, &level);
438 BUG_ON(ptep == NULL);
440 pfn = pte_pfn(*ptep);
441 page = pfn_to_page(pfn);
443 pte = pfn_pte(pfn, prot);
445 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
448 if (!PageHighMem(page)) {
449 void *av = __va(PFN_PHYS(pfn));
452 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
458 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
460 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
463 for(i = 0; i < entries; i += entries_per_page)
464 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
467 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
469 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
472 for(i = 0; i < entries; i += entries_per_page)
473 set_aliased_prot(ldt + i, PAGE_KERNEL);
476 static void xen_set_ldt(const void *addr, unsigned entries)
478 struct mmuext_op *op;
479 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
481 trace_xen_cpu_set_ldt(addr, entries);
484 op->cmd = MMUEXT_SET_LDT;
485 op->arg1.linear_addr = (unsigned long)addr;
486 op->arg2.nr_ents = entries;
488 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
490 xen_mc_issue(PARAVIRT_LAZY_CPU);
493 static void xen_load_gdt(const struct desc_ptr *dtr)
495 unsigned long va = dtr->address;
496 unsigned int size = dtr->size + 1;
497 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
498 unsigned long frames[pages];
502 * A GDT can be up to 64k in size, which corresponds to 8192
503 * 8-byte entries, or 16 4k pages..
506 BUG_ON(size > 65536);
507 BUG_ON(va & ~PAGE_MASK);
509 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
512 unsigned long pfn, mfn;
516 * The GDT is per-cpu and is in the percpu data area.
517 * That can be virtually mapped, so we need to do a
518 * page-walk to get the underlying MFN for the
519 * hypercall. The page can also be in the kernel's
520 * linear range, so we need to RO that mapping too.
522 ptep = lookup_address(va, &level);
523 BUG_ON(ptep == NULL);
525 pfn = pte_pfn(*ptep);
526 mfn = pfn_to_mfn(pfn);
527 virt = __va(PFN_PHYS(pfn));
531 make_lowmem_page_readonly((void *)va);
532 make_lowmem_page_readonly(virt);
535 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
540 * load_gdt for early boot, when the gdt is only mapped once
542 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
544 unsigned long va = dtr->address;
545 unsigned int size = dtr->size + 1;
546 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
547 unsigned long frames[pages];
551 * A GDT can be up to 64k in size, which corresponds to 8192
552 * 8-byte entries, or 16 4k pages..
555 BUG_ON(size > 65536);
556 BUG_ON(va & ~PAGE_MASK);
558 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
560 unsigned long pfn, mfn;
562 pfn = virt_to_pfn(va);
563 mfn = pfn_to_mfn(pfn);
565 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
567 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
573 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
577 static inline bool desc_equal(const struct desc_struct *d1,
578 const struct desc_struct *d2)
580 return d1->a == d2->a && d1->b == d2->b;
583 static void load_TLS_descriptor(struct thread_struct *t,
584 unsigned int cpu, unsigned int i)
586 struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
587 struct desc_struct *gdt;
589 struct multicall_space mc;
591 if (desc_equal(shadow, &t->tls_array[i]))
594 *shadow = t->tls_array[i];
596 gdt = get_cpu_gdt_table(cpu);
597 maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
598 mc = __xen_mc_entry(0);
600 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
603 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
606 * XXX sleazy hack: If we're being called in a lazy-cpu zone
607 * and lazy gs handling is enabled, it means we're in a
608 * context switch, and %gs has just been saved. This means we
609 * can zero it out to prevent faults on exit from the
610 * hypervisor if the next process has no %gs. Either way, it
611 * has been saved, and the new value will get loaded properly.
612 * This will go away as soon as Xen has been modified to not
613 * save/restore %gs for normal hypercalls.
615 * On x86_64, this hack is not used for %gs, because gs points
616 * to KERNEL_GS_BASE (and uses it for PDA references), so we
617 * must not zero %gs on x86_64
619 * For x86_64, we need to zero %fs, otherwise we may get an
620 * exception between the new %fs descriptor being loaded and
621 * %fs being effectively cleared at __switch_to().
623 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
633 load_TLS_descriptor(t, cpu, 0);
634 load_TLS_descriptor(t, cpu, 1);
635 load_TLS_descriptor(t, cpu, 2);
637 xen_mc_issue(PARAVIRT_LAZY_CPU);
641 static void xen_load_gs_index(unsigned int idx)
643 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
648 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
651 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
652 u64 entry = *(u64 *)ptr;
654 trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
659 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
665 static int cvt_gate_to_trap(int vector, const gate_desc *val,
666 struct trap_info *info)
670 if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
673 info->vector = vector;
675 addr = gate_offset(*val);
678 * Look for known traps using IST, and substitute them
679 * appropriately. The debugger ones are the only ones we care
680 * about. Xen will handle faults like double_fault,
681 * so we should never see them. Warn if
682 * there's an unexpected IST-using fault handler.
684 if (addr == (unsigned long)debug)
685 addr = (unsigned long)xen_debug;
686 else if (addr == (unsigned long)int3)
687 addr = (unsigned long)xen_int3;
688 else if (addr == (unsigned long)stack_segment)
689 addr = (unsigned long)xen_stack_segment;
690 else if (addr == (unsigned long)double_fault ||
691 addr == (unsigned long)nmi) {
692 /* Don't need to handle these */
694 #ifdef CONFIG_X86_MCE
695 } else if (addr == (unsigned long)machine_check) {
697 * when xen hypervisor inject vMCE to guest,
698 * use native mce handler to handle it
703 /* Some other trap using IST? */
704 if (WARN_ON(val->ist != 0))
707 #endif /* CONFIG_X86_64 */
708 info->address = addr;
710 info->cs = gate_segment(*val);
711 info->flags = val->dpl;
712 /* interrupt gates clear IF */
713 if (val->type == GATE_INTERRUPT)
714 info->flags |= 1 << 2;
719 /* Locations of each CPU's IDT */
720 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
722 /* Set an IDT entry. If the entry is part of the current IDT, then
724 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
726 unsigned long p = (unsigned long)&dt[entrynum];
727 unsigned long start, end;
729 trace_xen_cpu_write_idt_entry(dt, entrynum, g);
733 start = __this_cpu_read(idt_desc.address);
734 end = start + __this_cpu_read(idt_desc.size) + 1;
738 native_write_idt_entry(dt, entrynum, g);
740 if (p >= start && (p + 8) <= end) {
741 struct trap_info info[2];
745 if (cvt_gate_to_trap(entrynum, g, &info[0]))
746 if (HYPERVISOR_set_trap_table(info))
753 static void xen_convert_trap_info(const struct desc_ptr *desc,
754 struct trap_info *traps)
756 unsigned in, out, count;
758 count = (desc->size+1) / sizeof(gate_desc);
761 for (in = out = 0; in < count; in++) {
762 gate_desc *entry = (gate_desc*)(desc->address) + in;
764 if (cvt_gate_to_trap(in, entry, &traps[out]))
767 traps[out].address = 0;
770 void xen_copy_trap_info(struct trap_info *traps)
772 const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
774 xen_convert_trap_info(desc, traps);
777 /* Load a new IDT into Xen. In principle this can be per-CPU, so we
778 hold a spinlock to protect the static traps[] array (static because
779 it avoids allocation, and saves stack space). */
780 static void xen_load_idt(const struct desc_ptr *desc)
782 static DEFINE_SPINLOCK(lock);
783 static struct trap_info traps[257];
785 trace_xen_cpu_load_idt(desc);
789 __get_cpu_var(idt_desc) = *desc;
791 xen_convert_trap_info(desc, traps);
794 if (HYPERVISOR_set_trap_table(traps))
800 /* Write a GDT descriptor entry. Ignore LDT descriptors, since
801 they're handled differently. */
802 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
803 const void *desc, int type)
805 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
816 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
819 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
829 * Version of write_gdt_entry for use at early boot-time needed to
830 * update an entry as simply as possible.
832 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
833 const void *desc, int type)
835 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
844 xmaddr_t maddr = virt_to_machine(&dt[entry]);
846 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
847 dt[entry] = *(struct desc_struct *)desc;
853 static void xen_load_sp0(struct tss_struct *tss,
854 struct thread_struct *thread)
856 struct multicall_space mcs;
858 mcs = xen_mc_entry(0);
859 MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
860 xen_mc_issue(PARAVIRT_LAZY_CPU);
863 static void xen_set_iopl_mask(unsigned mask)
865 struct physdev_set_iopl set_iopl;
867 /* Force the change at ring 0. */
868 set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
869 HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
872 static void xen_io_delay(void)
876 #ifdef CONFIG_X86_LOCAL_APIC
877 static unsigned long xen_set_apic_id(unsigned int x)
882 static unsigned int xen_get_apic_id(unsigned long x)
884 return ((x)>>24) & 0xFFu;
886 static u32 xen_apic_read(u32 reg)
888 struct xen_platform_op op = {
889 .cmd = XENPF_get_cpuinfo,
890 .interface_version = XENPF_INTERFACE_VERSION,
891 .u.pcpu_info.xen_cpuid = 0,
895 /* Shouldn't need this as APIC is turned off for PV, and we only
896 * get called on the bootup processor. But just in case. */
897 if (!xen_initial_domain() || smp_processor_id())
906 ret = HYPERVISOR_dom0_op(&op);
910 return op.u.pcpu_info.apic_id << 24;
913 static void xen_apic_write(u32 reg, u32 val)
915 /* Warn to see if there's any stray references */
919 static u64 xen_apic_icr_read(void)
924 static void xen_apic_icr_write(u32 low, u32 id)
926 /* Warn to see if there's any stray references */
930 static void xen_apic_wait_icr_idle(void)
935 static u32 xen_safe_apic_wait_icr_idle(void)
940 static void set_xen_basic_apic_ops(void)
942 apic->read = xen_apic_read;
943 apic->write = xen_apic_write;
944 apic->icr_read = xen_apic_icr_read;
945 apic->icr_write = xen_apic_icr_write;
946 apic->wait_icr_idle = xen_apic_wait_icr_idle;
947 apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
948 apic->set_apic_id = xen_set_apic_id;
949 apic->get_apic_id = xen_get_apic_id;
952 apic->send_IPI_allbutself = xen_send_IPI_allbutself;
953 apic->send_IPI_mask_allbutself = xen_send_IPI_mask_allbutself;
954 apic->send_IPI_mask = xen_send_IPI_mask;
955 apic->send_IPI_all = xen_send_IPI_all;
956 apic->send_IPI_self = xen_send_IPI_self;
962 static void xen_clts(void)
964 struct multicall_space mcs;
966 mcs = xen_mc_entry(0);
968 MULTI_fpu_taskswitch(mcs.mc, 0);
970 xen_mc_issue(PARAVIRT_LAZY_CPU);
973 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
975 static unsigned long xen_read_cr0(void)
977 unsigned long cr0 = this_cpu_read(xen_cr0_value);
979 if (unlikely(cr0 == 0)) {
980 cr0 = native_read_cr0();
981 this_cpu_write(xen_cr0_value, cr0);
987 static void xen_write_cr0(unsigned long cr0)
989 struct multicall_space mcs;
991 this_cpu_write(xen_cr0_value, cr0);
993 /* Only pay attention to cr0.TS; everything else is
995 mcs = xen_mc_entry(0);
997 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
999 xen_mc_issue(PARAVIRT_LAZY_CPU);
1002 static void xen_write_cr4(unsigned long cr4)
1004 cr4 &= ~X86_CR4_PGE;
1005 cr4 &= ~X86_CR4_PSE;
1007 native_write_cr4(cr4);
1009 #ifdef CONFIG_X86_64
1010 static inline unsigned long xen_read_cr8(void)
1014 static inline void xen_write_cr8(unsigned long val)
1019 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
1026 #ifdef CONFIG_X86_64
1030 case MSR_FS_BASE: which = SEGBASE_FS; goto set;
1031 case MSR_KERNEL_GS_BASE: which = SEGBASE_GS_USER; goto set;
1032 case MSR_GS_BASE: which = SEGBASE_GS_KERNEL; goto set;
1035 base = ((u64)high << 32) | low;
1036 if (HYPERVISOR_set_segment_base(which, base) != 0)
1044 case MSR_SYSCALL_MASK:
1045 case MSR_IA32_SYSENTER_CS:
1046 case MSR_IA32_SYSENTER_ESP:
1047 case MSR_IA32_SYSENTER_EIP:
1048 /* Fast syscall setup is all done in hypercalls, so
1049 these are all ignored. Stub them out here to stop
1050 Xen console noise. */
1053 case MSR_IA32_CR_PAT:
1054 if (smp_processor_id() == 0)
1055 xen_set_pat(((u64)high << 32) | low);
1059 ret = native_write_msr_safe(msr, low, high);
1065 void xen_setup_shared_info(void)
1067 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1068 set_fixmap(FIX_PARAVIRT_BOOTMAP,
1069 xen_start_info->shared_info);
1071 HYPERVISOR_shared_info =
1072 (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
1074 HYPERVISOR_shared_info =
1075 (struct shared_info *)__va(xen_start_info->shared_info);
1078 /* In UP this is as good a place as any to set up shared info */
1079 xen_setup_vcpu_info_placement();
1082 xen_setup_mfn_list_list();
1085 /* This is called once we have the cpu_possible_mask */
1086 void xen_setup_vcpu_info_placement(void)
1090 for_each_possible_cpu(cpu)
1091 xen_vcpu_setup(cpu);
1093 /* xen_vcpu_setup managed to place the vcpu_info within the
1094 percpu area for all cpus, so make use of it */
1095 if (have_vcpu_info_placement) {
1096 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1097 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1098 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1099 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1100 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1104 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1105 unsigned long addr, unsigned len)
1107 char *start, *end, *reloc;
1110 start = end = reloc = NULL;
1112 #define SITE(op, x) \
1113 case PARAVIRT_PATCH(op.x): \
1114 if (have_vcpu_info_placement) { \
1115 start = (char *)xen_##x##_direct; \
1116 end = xen_##x##_direct_end; \
1117 reloc = xen_##x##_direct_reloc; \
1122 SITE(pv_irq_ops, irq_enable);
1123 SITE(pv_irq_ops, irq_disable);
1124 SITE(pv_irq_ops, save_fl);
1125 SITE(pv_irq_ops, restore_fl);
1129 if (start == NULL || (end-start) > len)
1132 ret = paravirt_patch_insns(insnbuf, len, start, end);
1134 /* Note: because reloc is assigned from something that
1135 appears to be an array, gcc assumes it's non-null,
1136 but doesn't know its relationship with start and
1138 if (reloc > start && reloc < end) {
1139 int reloc_off = reloc - start;
1140 long *relocp = (long *)(insnbuf + reloc_off);
1141 long delta = start - (char *)addr;
1149 ret = paravirt_patch_default(type, clobbers, insnbuf,
1157 static const struct pv_info xen_info __initconst = {
1158 .paravirt_enabled = 1,
1159 .shared_kernel_pmd = 0,
1161 #ifdef CONFIG_X86_64
1162 .extra_user_64bit_cs = FLAT_USER_CS64,
1168 static const struct pv_init_ops xen_init_ops __initconst = {
1172 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1175 .set_debugreg = xen_set_debugreg,
1176 .get_debugreg = xen_get_debugreg,
1180 .read_cr0 = xen_read_cr0,
1181 .write_cr0 = xen_write_cr0,
1183 .read_cr4 = native_read_cr4,
1184 .read_cr4_safe = native_read_cr4_safe,
1185 .write_cr4 = xen_write_cr4,
1187 #ifdef CONFIG_X86_64
1188 .read_cr8 = xen_read_cr8,
1189 .write_cr8 = xen_write_cr8,
1192 .wbinvd = native_wbinvd,
1194 .read_msr = native_read_msr_safe,
1195 .write_msr = xen_write_msr_safe,
1197 .read_tsc = native_read_tsc,
1198 .read_pmc = native_read_pmc,
1200 .read_tscp = native_read_tscp,
1203 .irq_enable_sysexit = xen_sysexit,
1204 #ifdef CONFIG_X86_64
1205 .usergs_sysret32 = xen_sysret32,
1206 .usergs_sysret64 = xen_sysret64,
1209 .load_tr_desc = paravirt_nop,
1210 .set_ldt = xen_set_ldt,
1211 .load_gdt = xen_load_gdt,
1212 .load_idt = xen_load_idt,
1213 .load_tls = xen_load_tls,
1214 #ifdef CONFIG_X86_64
1215 .load_gs_index = xen_load_gs_index,
1218 .alloc_ldt = xen_alloc_ldt,
1219 .free_ldt = xen_free_ldt,
1221 .store_gdt = native_store_gdt,
1222 .store_idt = native_store_idt,
1223 .store_tr = xen_store_tr,
1225 .write_ldt_entry = xen_write_ldt_entry,
1226 .write_gdt_entry = xen_write_gdt_entry,
1227 .write_idt_entry = xen_write_idt_entry,
1228 .load_sp0 = xen_load_sp0,
1230 .set_iopl_mask = xen_set_iopl_mask,
1231 .io_delay = xen_io_delay,
1233 /* Xen takes care of %gs when switching to usermode for us */
1234 .swapgs = paravirt_nop,
1236 .start_context_switch = paravirt_start_context_switch,
1237 .end_context_switch = xen_end_context_switch,
1240 static const struct pv_apic_ops xen_apic_ops __initconst = {
1241 #ifdef CONFIG_X86_LOCAL_APIC
1242 .startup_ipi_hook = paravirt_nop,
1246 static void xen_reboot(int reason)
1248 struct sched_shutdown r = { .reason = reason };
1250 if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1254 static void xen_restart(char *msg)
1256 xen_reboot(SHUTDOWN_reboot);
1259 static void xen_emergency_restart(void)
1261 xen_reboot(SHUTDOWN_reboot);
1264 static void xen_machine_halt(void)
1266 xen_reboot(SHUTDOWN_poweroff);
1269 static void xen_machine_power_off(void)
1273 xen_reboot(SHUTDOWN_poweroff);
1276 static void xen_crash_shutdown(struct pt_regs *regs)
1278 xen_reboot(SHUTDOWN_crash);
1282 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1284 xen_reboot(SHUTDOWN_crash);
1288 static struct notifier_block xen_panic_block = {
1289 .notifier_call= xen_panic_event,
1292 int xen_panic_handler_init(void)
1294 atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1298 static const struct machine_ops xen_machine_ops __initconst = {
1299 .restart = xen_restart,
1300 .halt = xen_machine_halt,
1301 .power_off = xen_machine_power_off,
1302 .shutdown = xen_machine_halt,
1303 .crash_shutdown = xen_crash_shutdown,
1304 .emergency_restart = xen_emergency_restart,
1308 * Set up the GDT and segment registers for -fstack-protector. Until
1309 * we do this, we have to be careful not to call any stack-protected
1310 * function, which is most of the kernel.
1312 static void __init xen_setup_stackprotector(void)
1314 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1315 pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1317 setup_stack_canary_segment(0);
1318 switch_to_new_gdt(0);
1320 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1321 pv_cpu_ops.load_gdt = xen_load_gdt;
1324 /* First C function to be called on Xen boot */
1325 asmlinkage void __init xen_start_kernel(void)
1327 struct physdev_set_iopl set_iopl;
1330 if (!xen_start_info)
1333 xen_domain_type = XEN_PV_DOMAIN;
1335 xen_setup_machphys_mapping();
1337 /* Install Xen paravirt ops */
1339 pv_init_ops = xen_init_ops;
1340 pv_cpu_ops = xen_cpu_ops;
1341 pv_apic_ops = xen_apic_ops;
1343 x86_init.resources.memory_setup = xen_memory_setup;
1344 x86_init.oem.arch_setup = xen_arch_setup;
1345 x86_init.oem.banner = xen_banner;
1347 xen_init_time_ops();
1350 * Set up some pagetable state before starting to set any ptes.
1355 /* Prevent unwanted bits from being set in PTEs. */
1356 __supported_pte_mask &= ~_PAGE_GLOBAL;
1358 if (!xen_initial_domain())
1360 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1362 __supported_pte_mask |= _PAGE_IOMAP;
1365 * Prevent page tables from being allocated in highmem, even
1366 * if CONFIG_HIGHPTE is enabled.
1368 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1370 /* Work out if we support NX */
1373 xen_setup_features();
1376 if (!xen_feature(XENFEAT_auto_translated_physmap))
1377 xen_build_dynamic_phys_to_machine();
1380 * Set up kernel GDT and segment registers, mainly so that
1381 * -fstack-protector code can be executed.
1383 xen_setup_stackprotector();
1386 xen_init_cpuid_mask();
1388 #ifdef CONFIG_X86_LOCAL_APIC
1390 * set up the basic apic ops.
1392 set_xen_basic_apic_ops();
1395 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1396 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1397 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1400 machine_ops = xen_machine_ops;
1403 * The only reliable way to retain the initial address of the
1404 * percpu gdt_page is to remember it here, so we can go and
1405 * mark it RW later, when the initial percpu area is freed.
1407 xen_initial_gdt = &per_cpu(gdt_page, 0);
1411 #ifdef CONFIG_ACPI_NUMA
1413 * The pages we from Xen are not related to machine pages, so
1414 * any NUMA information the kernel tries to get from ACPI will
1415 * be meaningless. Prevent it from trying.
1420 /* Don't do the full vcpu_info placement stuff until we have a
1421 possible map and a non-dummy shared_info. */
1422 per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1424 local_irq_disable();
1425 early_boot_irqs_disabled = true;
1427 xen_raw_console_write("mapping kernel into physical memory\n");
1428 xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base, xen_start_info->nr_pages);
1430 /* Allocate and initialize top and mid mfn levels for p2m structure */
1431 xen_build_mfn_list_list();
1433 /* keep using Xen gdt for now; no urgent need to change it */
1435 #ifdef CONFIG_X86_32
1436 pv_info.kernel_rpl = 1;
1437 if (xen_feature(XENFEAT_supervisor_mode_kernel))
1438 pv_info.kernel_rpl = 0;
1440 pv_info.kernel_rpl = 0;
1442 /* set the limit of our address space */
1445 /* We used to do this in xen_arch_setup, but that is too late on AMD
1446 * were early_cpu_init (run before ->arch_setup()) calls early_amd_init
1447 * which pokes 0xcf8 port.
1450 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1452 xen_raw_printk("physdev_op failed %d\n", rc);
1454 #ifdef CONFIG_X86_32
1455 /* set up basic CPUID stuff */
1456 cpu_detect(&new_cpu_data);
1457 new_cpu_data.hard_math = 1;
1458 new_cpu_data.wp_works_ok = 1;
1459 new_cpu_data.x86_capability[0] = cpuid_edx(1);
1462 /* Poke various useful things into boot_params */
1463 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1464 boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1465 ? __pa(xen_start_info->mod_start) : 0;
1466 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1467 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1469 if (!xen_initial_domain()) {
1470 add_preferred_console("xenboot", 0, NULL);
1471 add_preferred_console("tty", 0, NULL);
1472 add_preferred_console("hvc", 0, NULL);
1474 x86_init.pci.arch_init = pci_xen_init;
1476 const struct dom0_vga_console_info *info =
1477 (void *)((char *)xen_start_info +
1478 xen_start_info->console.dom0.info_off);
1479 struct xen_platform_op op = {
1480 .cmd = XENPF_firmware_info,
1481 .interface_version = XENPF_INTERFACE_VERSION,
1482 .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
1485 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1486 xen_start_info->console.domU.mfn = 0;
1487 xen_start_info->console.domU.evtchn = 0;
1489 if (HYPERVISOR_dom0_op(&op) == 0)
1490 boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
1494 /* Make sure ACS will be enabled */
1497 xen_acpi_sleep_register();
1499 /* Avoid searching for BIOS MP tables */
1500 x86_init.mpparse.find_smp_config = x86_init_noop;
1501 x86_init.mpparse.get_smp_config = x86_init_uint_noop;
1504 /* PCI BIOS service won't work from a PV guest. */
1505 pci_probe &= ~PCI_PROBE_BIOS;
1507 xen_raw_console_write("about to get started...\n");
1509 xen_setup_runstate_info(0);
1511 /* Start the world */
1512 #ifdef CONFIG_X86_32
1513 i386_start_kernel();
1515 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1519 #ifdef CONFIG_XEN_PVHVM
1520 #define HVM_SHARED_INFO_ADDR 0xFE700000UL
1521 static struct shared_info *xen_hvm_shared_info;
1522 static unsigned long xen_hvm_sip_phys;
1523 static int xen_major, xen_minor;
1525 static void xen_hvm_connect_shared_info(unsigned long pfn)
1527 struct xen_add_to_physmap xatp;
1529 xatp.domid = DOMID_SELF;
1531 xatp.space = XENMAPSPACE_shared_info;
1533 if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1537 static void __init xen_hvm_set_shared_info(struct shared_info *sip)
1541 HYPERVISOR_shared_info = sip;
1543 /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1544 * page, we use it in the event channel upcall and in some pvclock
1545 * related functions. We don't need the vcpu_info placement
1546 * optimizations because we don't use any pv_mmu or pv_irq op on
1548 for_each_online_cpu(cpu)
1549 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1552 /* Reconnect the shared_info pfn to a (new) mfn */
1553 void xen_hvm_resume_shared_info(void)
1555 xen_hvm_connect_shared_info(xen_hvm_sip_phys >> PAGE_SHIFT);
1558 /* Xen tools prior to Xen 4 do not provide a E820_Reserved area for guest usage.
1559 * On these old tools the shared info page will be placed in E820_Ram.
1560 * Xen 4 provides a E820_Reserved area at 0xFC000000, and this code expects
1561 * that nothing is mapped up to HVM_SHARED_INFO_ADDR.
1562 * Xen 4.3+ provides an explicit 1MB area at HVM_SHARED_INFO_ADDR which is used
1563 * here for the shared info page. */
1564 static void __init xen_hvm_init_shared_info(void)
1566 if (xen_major < 4) {
1567 xen_hvm_shared_info = extend_brk(PAGE_SIZE, PAGE_SIZE);
1568 xen_hvm_sip_phys = __pa(xen_hvm_shared_info);
1570 xen_hvm_sip_phys = HVM_SHARED_INFO_ADDR;
1571 set_fixmap(FIX_PARAVIRT_BOOTMAP, xen_hvm_sip_phys);
1572 xen_hvm_shared_info =
1573 (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
1575 xen_hvm_connect_shared_info(xen_hvm_sip_phys >> PAGE_SHIFT);
1576 xen_hvm_set_shared_info(xen_hvm_shared_info);
1579 static void __init init_hvm_pv_info(void)
1581 uint32_t ecx, edx, pages, msr, base;
1584 base = xen_cpuid_base();
1585 cpuid(base + 2, &pages, &msr, &ecx, &edx);
1587 pfn = __pa(hypercall_page);
1588 wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1590 xen_setup_features();
1592 pv_info.name = "Xen HVM";
1594 xen_domain_type = XEN_HVM_DOMAIN;
1597 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self,
1598 unsigned long action, void *hcpu)
1600 int cpu = (long)hcpu;
1602 case CPU_UP_PREPARE:
1603 xen_vcpu_setup(cpu);
1604 if (xen_have_vector_callback)
1605 xen_init_lock_cpu(cpu);
1613 static struct notifier_block xen_hvm_cpu_notifier __cpuinitdata = {
1614 .notifier_call = xen_hvm_cpu_notify,
1617 static void __init xen_hvm_guest_init(void)
1621 xen_hvm_init_shared_info();
1623 if (xen_feature(XENFEAT_hvm_callback_vector))
1624 xen_have_vector_callback = 1;
1626 register_cpu_notifier(&xen_hvm_cpu_notifier);
1627 xen_unplug_emulated_devices();
1628 x86_init.irqs.intr_init = xen_init_IRQ;
1629 xen_hvm_init_time_ops();
1630 xen_hvm_init_mmu_ops();
1633 static bool __init xen_hvm_platform(void)
1635 uint32_t eax, ebx, ecx, edx, base;
1637 if (xen_pv_domain())
1640 base = xen_cpuid_base();
1644 cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1646 xen_major = eax >> 16;
1647 xen_minor = eax & 0xffff;
1649 printk(KERN_INFO "Xen version %d.%d.\n", xen_major, xen_minor);
1654 bool xen_hvm_need_lapic(void)
1656 if (xen_pv_domain())
1658 if (!xen_hvm_domain())
1660 if (xen_feature(XENFEAT_hvm_pirqs) && xen_have_vector_callback)
1664 EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1666 const struct hypervisor_x86 x86_hyper_xen_hvm __refconst = {
1668 .detect = xen_hvm_platform,
1669 .init_platform = xen_hvm_guest_init,
1671 EXPORT_SYMBOL(x86_hyper_xen_hvm);