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/export.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>
35 #include <linux/frame.h>
38 #include <xen/events.h>
39 #include <xen/interface/xen.h>
40 #include <xen/interface/version.h>
41 #include <xen/interface/physdev.h>
42 #include <xen/interface/vcpu.h>
43 #include <xen/interface/memory.h>
44 #include <xen/interface/nmi.h>
45 #include <xen/interface/xen-mca.h>
46 #include <xen/features.h>
48 #include <xen/hvc-console.h>
51 #include <asm/paravirt.h>
54 #include <asm/xen/pci.h>
55 #include <asm/xen/hypercall.h>
56 #include <asm/xen/hypervisor.h>
57 #include <asm/xen/cpuid.h>
58 #include <asm/fixmap.h>
59 #include <asm/processor.h>
60 #include <asm/proto.h>
61 #include <asm/msr-index.h>
62 #include <asm/traps.h>
63 #include <asm/setup.h>
65 #include <asm/pgalloc.h>
66 #include <asm/pgtable.h>
67 #include <asm/tlbflush.h>
68 #include <asm/reboot.h>
69 #include <asm/stackprotector.h>
70 #include <asm/hypervisor.h>
71 #include <asm/mach_traps.h>
72 #include <asm/mwait.h>
73 #include <asm/pci_x86.h>
77 #include <linux/acpi.h>
79 #include <acpi/pdc_intel.h>
80 #include <acpi/processor.h>
81 #include <xen/interface/platform.h>
87 #include "multicalls.h"
90 void *xen_initial_gdt;
92 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
94 static int xen_cpu_up_prepare_pv(unsigned int cpu);
95 static int xen_cpu_dead_pv(unsigned int cpu);
98 struct desc_struct desc[3];
102 * Updating the 3 TLS descriptors in the GDT on every task switch is
103 * surprisingly expensive so we avoid updating them if they haven't
104 * changed. Since Xen writes different descriptors than the one
105 * passed in the update_descriptor hypercall we keep shadow copies to
108 static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
111 * On restore, set the vcpu placement up again.
112 * If it fails, then we're in a bad state, since
113 * we can't back out from using it...
115 void xen_vcpu_restore(void)
119 for_each_possible_cpu(cpu) {
120 bool other_cpu = (cpu != smp_processor_id());
121 bool is_up = HYPERVISOR_vcpu_op(VCPUOP_is_up, xen_vcpu_nr(cpu),
124 if (other_cpu && is_up &&
125 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL))
128 xen_setup_runstate_info(cpu);
130 if (xen_have_vcpu_info_placement)
133 if (other_cpu && is_up &&
134 HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL))
139 static void __init xen_banner(void)
141 unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
142 struct xen_extraversion extra;
143 HYPERVISOR_xen_version(XENVER_extraversion, &extra);
145 pr_info("Booting paravirtualized kernel %son %s\n",
146 xen_feature(XENFEAT_auto_translated_physmap) ?
147 "with PVH extensions " : "", pv_info.name);
148 printk(KERN_INFO "Xen version: %d.%d%s%s\n",
149 version >> 16, version & 0xffff, extra.extraversion,
150 xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
152 /* Check if running on Xen version (major, minor) or later */
154 xen_running_on_version_or_later(unsigned int major, unsigned int minor)
156 unsigned int version;
161 version = HYPERVISOR_xen_version(XENVER_version, NULL);
162 if ((((version >> 16) == major) && ((version & 0xffff) >= minor)) ||
163 ((version >> 16) > major))
168 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
169 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
171 static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
172 static __read_mostly unsigned int cpuid_leaf5_ecx_val;
173 static __read_mostly unsigned int cpuid_leaf5_edx_val;
175 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
176 unsigned int *cx, unsigned int *dx)
178 unsigned maskebx = ~0;
179 unsigned maskecx = ~0;
180 unsigned maskedx = ~0;
183 * Mask out inconvenient features, to try and disable as many
184 * unsupported kernel subsystems as possible.
188 maskecx = cpuid_leaf1_ecx_mask;
189 setecx = cpuid_leaf1_ecx_set_mask;
190 maskedx = cpuid_leaf1_edx_mask;
193 case CPUID_MWAIT_LEAF:
194 /* Synthesize the values.. */
197 *cx = cpuid_leaf5_ecx_val;
198 *dx = cpuid_leaf5_edx_val;
202 /* Suppress extended topology stuff */
207 asm(XEN_EMULATE_PREFIX "cpuid"
212 : "0" (*ax), "2" (*cx));
219 STACK_FRAME_NON_STANDARD(xen_cpuid); /* XEN_EMULATE_PREFIX */
221 static bool __init xen_check_mwait(void)
224 struct xen_platform_op op = {
225 .cmd = XENPF_set_processor_pminfo,
226 .u.set_pminfo.id = -1,
227 .u.set_pminfo.type = XEN_PM_PDC,
230 unsigned int ax, bx, cx, dx;
231 unsigned int mwait_mask;
233 /* We need to determine whether it is OK to expose the MWAIT
234 * capability to the kernel to harvest deeper than C3 states from ACPI
235 * _CST using the processor_harvest_xen.c module. For this to work, we
236 * need to gather the MWAIT_LEAF values (which the cstate.c code
237 * checks against). The hypervisor won't expose the MWAIT flag because
238 * it would break backwards compatibility; so we will find out directly
239 * from the hardware and hypercall.
241 if (!xen_initial_domain())
245 * When running under platform earlier than Xen4.2, do not expose
246 * mwait, to avoid the risk of loading native acpi pad driver
248 if (!xen_running_on_version_or_later(4, 2))
254 native_cpuid(&ax, &bx, &cx, &dx);
256 mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
257 (1 << (X86_FEATURE_MWAIT % 32));
259 if ((cx & mwait_mask) != mwait_mask)
262 /* We need to emulate the MWAIT_LEAF and for that we need both
263 * ecx and edx. The hypercall provides only partial information.
266 ax = CPUID_MWAIT_LEAF;
271 native_cpuid(&ax, &bx, &cx, &dx);
273 /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
274 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
276 buf[0] = ACPI_PDC_REVISION_ID;
278 buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
280 set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
282 if ((HYPERVISOR_platform_op(&op) == 0) &&
283 (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
284 cpuid_leaf5_ecx_val = cx;
285 cpuid_leaf5_edx_val = dx;
292 static void __init xen_init_cpuid_mask(void)
294 unsigned int ax, bx, cx, dx;
295 unsigned int xsave_mask;
297 if (!xen_initial_domain())
298 cpuid_leaf1_edx_mask &=
299 ~((1 << X86_FEATURE_ACPI)); /* disable ACPI */
301 cpuid_leaf1_ecx_mask &= ~(1 << (X86_FEATURE_X2APIC % 32));
305 cpuid(1, &ax, &bx, &cx, &dx);
308 (1 << (X86_FEATURE_XSAVE % 32)) |
309 (1 << (X86_FEATURE_OSXSAVE % 32));
311 /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
312 if ((cx & xsave_mask) != xsave_mask)
313 cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
314 if (xen_check_mwait())
315 cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
318 static void __init xen_init_capabilities(void)
320 setup_clear_cpu_cap(X86_BUG_SYSRET_SS_ATTRS);
321 setup_force_cpu_cap(X86_FEATURE_XENPV);
322 setup_clear_cpu_cap(X86_FEATURE_DCA);
323 setup_clear_cpu_cap(X86_FEATURE_APERFMPERF);
324 setup_clear_cpu_cap(X86_FEATURE_MTRR);
325 setup_clear_cpu_cap(X86_FEATURE_ACC);
328 static void xen_set_debugreg(int reg, unsigned long val)
330 HYPERVISOR_set_debugreg(reg, val);
333 static unsigned long xen_get_debugreg(int reg)
335 return HYPERVISOR_get_debugreg(reg);
338 static void xen_end_context_switch(struct task_struct *next)
341 paravirt_end_context_switch(next);
344 static unsigned long xen_store_tr(void)
350 * Set the page permissions for a particular virtual address. If the
351 * address is a vmalloc mapping (or other non-linear mapping), then
352 * find the linear mapping of the page and also set its protections to
355 static void set_aliased_prot(void *v, pgprot_t prot)
364 ptep = lookup_address((unsigned long)v, &level);
365 BUG_ON(ptep == NULL);
367 pfn = pte_pfn(*ptep);
368 page = pfn_to_page(pfn);
370 pte = pfn_pte(pfn, prot);
373 * Careful: update_va_mapping() will fail if the virtual address
374 * we're poking isn't populated in the page tables. We don't
375 * need to worry about the direct map (that's always in the page
376 * tables), but we need to be careful about vmap space. In
377 * particular, the top level page table can lazily propagate
378 * entries between processes, so if we've switched mms since we
379 * vmapped the target in the first place, we might not have the
380 * top-level page table entry populated.
382 * We disable preemption because we want the same mm active when
383 * we probe the target and when we issue the hypercall. We'll
384 * have the same nominal mm, but if we're a kernel thread, lazy
385 * mm dropping could change our pgd.
387 * Out of an abundance of caution, this uses __get_user() to fault
388 * in the target address just in case there's some obscure case
389 * in which the target address isn't readable.
394 probe_kernel_read(&dummy, v, 1);
396 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
399 if (!PageHighMem(page)) {
400 void *av = __va(PFN_PHYS(pfn));
403 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
411 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
413 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
417 * We need to mark the all aliases of the LDT pages RO. We
418 * don't need to call vm_flush_aliases(), though, since that's
419 * only responsible for flushing aliases out the TLBs, not the
420 * page tables, and Xen will flush the TLB for us if needed.
422 * To avoid confusing future readers: none of this is necessary
423 * to load the LDT. The hypervisor only checks this when the
424 * LDT is faulted in due to subsequent descriptor access.
427 for (i = 0; i < entries; i += entries_per_page)
428 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
431 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
433 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
436 for (i = 0; i < entries; i += entries_per_page)
437 set_aliased_prot(ldt + i, PAGE_KERNEL);
440 static void xen_set_ldt(const void *addr, unsigned entries)
442 struct mmuext_op *op;
443 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
445 trace_xen_cpu_set_ldt(addr, entries);
448 op->cmd = MMUEXT_SET_LDT;
449 op->arg1.linear_addr = (unsigned long)addr;
450 op->arg2.nr_ents = entries;
452 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
454 xen_mc_issue(PARAVIRT_LAZY_CPU);
457 static void xen_load_gdt(const struct desc_ptr *dtr)
459 unsigned long va = dtr->address;
460 unsigned int size = dtr->size + 1;
461 unsigned pages = DIV_ROUND_UP(size, PAGE_SIZE);
462 unsigned long frames[pages];
466 * A GDT can be up to 64k in size, which corresponds to 8192
467 * 8-byte entries, or 16 4k pages..
470 BUG_ON(size > 65536);
471 BUG_ON(va & ~PAGE_MASK);
473 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
476 unsigned long pfn, mfn;
480 * The GDT is per-cpu and is in the percpu data area.
481 * That can be virtually mapped, so we need to do a
482 * page-walk to get the underlying MFN for the
483 * hypercall. The page can also be in the kernel's
484 * linear range, so we need to RO that mapping too.
486 ptep = lookup_address(va, &level);
487 BUG_ON(ptep == NULL);
489 pfn = pte_pfn(*ptep);
490 mfn = pfn_to_mfn(pfn);
491 virt = __va(PFN_PHYS(pfn));
495 make_lowmem_page_readonly((void *)va);
496 make_lowmem_page_readonly(virt);
499 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
504 * load_gdt for early boot, when the gdt is only mapped once
506 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
508 unsigned long va = dtr->address;
509 unsigned int size = dtr->size + 1;
510 unsigned pages = DIV_ROUND_UP(size, PAGE_SIZE);
511 unsigned long frames[pages];
515 * A GDT can be up to 64k in size, which corresponds to 8192
516 * 8-byte entries, or 16 4k pages..
519 BUG_ON(size > 65536);
520 BUG_ON(va & ~PAGE_MASK);
522 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
524 unsigned long pfn, mfn;
526 pfn = virt_to_pfn(va);
527 mfn = pfn_to_mfn(pfn);
529 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
531 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
537 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
541 static inline bool desc_equal(const struct desc_struct *d1,
542 const struct desc_struct *d2)
544 return d1->a == d2->a && d1->b == d2->b;
547 static void load_TLS_descriptor(struct thread_struct *t,
548 unsigned int cpu, unsigned int i)
550 struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
551 struct desc_struct *gdt;
553 struct multicall_space mc;
555 if (desc_equal(shadow, &t->tls_array[i]))
558 *shadow = t->tls_array[i];
560 gdt = get_cpu_gdt_rw(cpu);
561 maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
562 mc = __xen_mc_entry(0);
564 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
567 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
570 * XXX sleazy hack: If we're being called in a lazy-cpu zone
571 * and lazy gs handling is enabled, it means we're in a
572 * context switch, and %gs has just been saved. This means we
573 * can zero it out to prevent faults on exit from the
574 * hypervisor if the next process has no %gs. Either way, it
575 * has been saved, and the new value will get loaded properly.
576 * This will go away as soon as Xen has been modified to not
577 * save/restore %gs for normal hypercalls.
579 * On x86_64, this hack is not used for %gs, because gs points
580 * to KERNEL_GS_BASE (and uses it for PDA references), so we
581 * must not zero %gs on x86_64
583 * For x86_64, we need to zero %fs, otherwise we may get an
584 * exception between the new %fs descriptor being loaded and
585 * %fs being effectively cleared at __switch_to().
587 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
597 load_TLS_descriptor(t, cpu, 0);
598 load_TLS_descriptor(t, cpu, 1);
599 load_TLS_descriptor(t, cpu, 2);
601 xen_mc_issue(PARAVIRT_LAZY_CPU);
605 static void xen_load_gs_index(unsigned int idx)
607 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
612 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
615 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
616 u64 entry = *(u64 *)ptr;
618 trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
623 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
629 static int cvt_gate_to_trap(int vector, const gate_desc *val,
630 struct trap_info *info)
634 if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
637 info->vector = vector;
639 addr = gate_offset(*val);
642 * Look for known traps using IST, and substitute them
643 * appropriately. The debugger ones are the only ones we care
644 * about. Xen will handle faults like double_fault,
645 * so we should never see them. Warn if
646 * there's an unexpected IST-using fault handler.
648 if (addr == (unsigned long)debug)
649 addr = (unsigned long)xen_debug;
650 else if (addr == (unsigned long)int3)
651 addr = (unsigned long)xen_int3;
652 else if (addr == (unsigned long)stack_segment)
653 addr = (unsigned long)xen_stack_segment;
654 else if (addr == (unsigned long)double_fault) {
655 /* Don't need to handle these */
657 #ifdef CONFIG_X86_MCE
658 } else if (addr == (unsigned long)machine_check) {
660 * when xen hypervisor inject vMCE to guest,
661 * use native mce handler to handle it
665 } else if (addr == (unsigned long)nmi)
667 * Use the native version as well.
671 /* Some other trap using IST? */
672 if (WARN_ON(val->ist != 0))
675 #endif /* CONFIG_X86_64 */
676 info->address = addr;
678 info->cs = gate_segment(*val);
679 info->flags = val->dpl;
680 /* interrupt gates clear IF */
681 if (val->type == GATE_INTERRUPT)
682 info->flags |= 1 << 2;
687 /* Locations of each CPU's IDT */
688 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
690 /* Set an IDT entry. If the entry is part of the current IDT, then
692 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
694 unsigned long p = (unsigned long)&dt[entrynum];
695 unsigned long start, end;
697 trace_xen_cpu_write_idt_entry(dt, entrynum, g);
701 start = __this_cpu_read(idt_desc.address);
702 end = start + __this_cpu_read(idt_desc.size) + 1;
706 native_write_idt_entry(dt, entrynum, g);
708 if (p >= start && (p + 8) <= end) {
709 struct trap_info info[2];
713 if (cvt_gate_to_trap(entrynum, g, &info[0]))
714 if (HYPERVISOR_set_trap_table(info))
721 static void xen_convert_trap_info(const struct desc_ptr *desc,
722 struct trap_info *traps)
724 unsigned in, out, count;
726 count = (desc->size+1) / sizeof(gate_desc);
729 for (in = out = 0; in < count; in++) {
730 gate_desc *entry = (gate_desc *)(desc->address) + in;
732 if (cvt_gate_to_trap(in, entry, &traps[out]))
735 traps[out].address = 0;
738 void xen_copy_trap_info(struct trap_info *traps)
740 const struct desc_ptr *desc = this_cpu_ptr(&idt_desc);
742 xen_convert_trap_info(desc, traps);
745 /* Load a new IDT into Xen. In principle this can be per-CPU, so we
746 hold a spinlock to protect the static traps[] array (static because
747 it avoids allocation, and saves stack space). */
748 static void xen_load_idt(const struct desc_ptr *desc)
750 static DEFINE_SPINLOCK(lock);
751 static struct trap_info traps[257];
753 trace_xen_cpu_load_idt(desc);
757 memcpy(this_cpu_ptr(&idt_desc), desc, sizeof(idt_desc));
759 xen_convert_trap_info(desc, traps);
762 if (HYPERVISOR_set_trap_table(traps))
768 /* Write a GDT descriptor entry. Ignore LDT descriptors, since
769 they're handled differently. */
770 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
771 const void *desc, int type)
773 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
784 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
787 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
797 * Version of write_gdt_entry for use at early boot-time needed to
798 * update an entry as simply as possible.
800 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
801 const void *desc, int type)
803 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
812 xmaddr_t maddr = virt_to_machine(&dt[entry]);
814 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
815 dt[entry] = *(struct desc_struct *)desc;
821 static void xen_load_sp0(struct tss_struct *tss,
822 struct thread_struct *thread)
824 struct multicall_space mcs;
826 mcs = xen_mc_entry(0);
827 MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
828 xen_mc_issue(PARAVIRT_LAZY_CPU);
829 tss->x86_tss.sp0 = thread->sp0;
832 void xen_set_iopl_mask(unsigned mask)
834 struct physdev_set_iopl set_iopl;
836 /* Force the change at ring 0. */
837 set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
838 HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
841 static void xen_io_delay(void)
845 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
847 static unsigned long xen_read_cr0(void)
849 unsigned long cr0 = this_cpu_read(xen_cr0_value);
851 if (unlikely(cr0 == 0)) {
852 cr0 = native_read_cr0();
853 this_cpu_write(xen_cr0_value, cr0);
859 static void xen_write_cr0(unsigned long cr0)
861 struct multicall_space mcs;
863 this_cpu_write(xen_cr0_value, cr0);
865 /* Only pay attention to cr0.TS; everything else is
867 mcs = xen_mc_entry(0);
869 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
871 xen_mc_issue(PARAVIRT_LAZY_CPU);
874 static void xen_write_cr4(unsigned long cr4)
876 cr4 &= ~(X86_CR4_PGE | X86_CR4_PSE | X86_CR4_PCE);
878 native_write_cr4(cr4);
881 static inline unsigned long xen_read_cr8(void)
885 static inline void xen_write_cr8(unsigned long val)
891 static u64 xen_read_msr_safe(unsigned int msr, int *err)
895 if (pmu_msr_read(msr, &val, err))
898 val = native_read_msr_safe(msr, err);
900 case MSR_IA32_APICBASE:
901 #ifdef CONFIG_X86_X2APIC
902 if (!(cpuid_ecx(1) & (1 << (X86_FEATURE_X2APIC & 31))))
904 val &= ~X2APIC_ENABLE;
910 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
921 case MSR_FS_BASE: which = SEGBASE_FS; goto set;
922 case MSR_KERNEL_GS_BASE: which = SEGBASE_GS_USER; goto set;
923 case MSR_GS_BASE: which = SEGBASE_GS_KERNEL; goto set;
926 base = ((u64)high << 32) | low;
927 if (HYPERVISOR_set_segment_base(which, base) != 0)
935 case MSR_SYSCALL_MASK:
936 case MSR_IA32_SYSENTER_CS:
937 case MSR_IA32_SYSENTER_ESP:
938 case MSR_IA32_SYSENTER_EIP:
939 /* Fast syscall setup is all done in hypercalls, so
940 these are all ignored. Stub them out here to stop
941 Xen console noise. */
945 if (!pmu_msr_write(msr, low, high, &ret))
946 ret = native_write_msr_safe(msr, low, high);
952 static u64 xen_read_msr(unsigned int msr)
955 * This will silently swallow a #GP from RDMSR. It may be worth
960 return xen_read_msr_safe(msr, &err);
963 static void xen_write_msr(unsigned int msr, unsigned low, unsigned high)
966 * This will silently swallow a #GP from WRMSR. It may be worth
969 xen_write_msr_safe(msr, low, high);
972 void xen_setup_shared_info(void)
974 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
975 set_fixmap(FIX_PARAVIRT_BOOTMAP,
976 xen_start_info->shared_info);
978 HYPERVISOR_shared_info =
979 (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
981 HYPERVISOR_shared_info =
982 (struct shared_info *)__va(xen_start_info->shared_info);
985 /* In UP this is as good a place as any to set up shared info */
986 xen_setup_vcpu_info_placement();
989 xen_setup_mfn_list_list();
992 /* This is called once we have the cpu_possible_mask */
993 void xen_setup_vcpu_info_placement(void)
997 for_each_possible_cpu(cpu) {
998 /* Set up direct vCPU id mapping for PV guests. */
999 per_cpu(xen_vcpu_id, cpu) = cpu;
1000 xen_vcpu_setup(cpu);
1004 * xen_vcpu_setup managed to place the vcpu_info within the
1005 * percpu area for all cpus, so make use of it.
1007 if (xen_have_vcpu_info_placement) {
1008 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1009 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1010 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1011 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1012 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1016 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1017 unsigned long addr, unsigned len)
1019 char *start, *end, *reloc;
1022 start = end = reloc = NULL;
1024 #define SITE(op, x) \
1025 case PARAVIRT_PATCH(op.x): \
1026 if (xen_have_vcpu_info_placement) { \
1027 start = (char *)xen_##x##_direct; \
1028 end = xen_##x##_direct_end; \
1029 reloc = xen_##x##_direct_reloc; \
1034 SITE(pv_irq_ops, irq_enable);
1035 SITE(pv_irq_ops, irq_disable);
1036 SITE(pv_irq_ops, save_fl);
1037 SITE(pv_irq_ops, restore_fl);
1041 if (start == NULL || (end-start) > len)
1044 ret = paravirt_patch_insns(insnbuf, len, start, end);
1046 /* Note: because reloc is assigned from something that
1047 appears to be an array, gcc assumes it's non-null,
1048 but doesn't know its relationship with start and
1050 if (reloc > start && reloc < end) {
1051 int reloc_off = reloc - start;
1052 long *relocp = (long *)(insnbuf + reloc_off);
1053 long delta = start - (char *)addr;
1061 ret = paravirt_patch_default(type, clobbers, insnbuf,
1069 static const struct pv_info xen_info __initconst = {
1070 .shared_kernel_pmd = 0,
1072 #ifdef CONFIG_X86_64
1073 .extra_user_64bit_cs = FLAT_USER_CS64,
1078 static const struct pv_init_ops xen_init_ops __initconst = {
1082 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1085 .set_debugreg = xen_set_debugreg,
1086 .get_debugreg = xen_get_debugreg,
1088 .read_cr0 = xen_read_cr0,
1089 .write_cr0 = xen_write_cr0,
1091 .read_cr4 = native_read_cr4,
1092 .write_cr4 = xen_write_cr4,
1094 #ifdef CONFIG_X86_64
1095 .read_cr8 = xen_read_cr8,
1096 .write_cr8 = xen_write_cr8,
1099 .wbinvd = native_wbinvd,
1101 .read_msr = xen_read_msr,
1102 .write_msr = xen_write_msr,
1104 .read_msr_safe = xen_read_msr_safe,
1105 .write_msr_safe = xen_write_msr_safe,
1107 .read_pmc = xen_read_pmc,
1110 #ifdef CONFIG_X86_64
1111 .usergs_sysret64 = xen_sysret64,
1114 .load_tr_desc = paravirt_nop,
1115 .set_ldt = xen_set_ldt,
1116 .load_gdt = xen_load_gdt,
1117 .load_idt = xen_load_idt,
1118 .load_tls = xen_load_tls,
1119 #ifdef CONFIG_X86_64
1120 .load_gs_index = xen_load_gs_index,
1123 .alloc_ldt = xen_alloc_ldt,
1124 .free_ldt = xen_free_ldt,
1126 .store_idt = native_store_idt,
1127 .store_tr = xen_store_tr,
1129 .write_ldt_entry = xen_write_ldt_entry,
1130 .write_gdt_entry = xen_write_gdt_entry,
1131 .write_idt_entry = xen_write_idt_entry,
1132 .load_sp0 = xen_load_sp0,
1134 .set_iopl_mask = xen_set_iopl_mask,
1135 .io_delay = xen_io_delay,
1137 /* Xen takes care of %gs when switching to usermode for us */
1138 .swapgs = paravirt_nop,
1140 .start_context_switch = paravirt_start_context_switch,
1141 .end_context_switch = xen_end_context_switch,
1144 static void xen_restart(char *msg)
1146 xen_reboot(SHUTDOWN_reboot);
1149 static void xen_machine_halt(void)
1151 xen_reboot(SHUTDOWN_poweroff);
1154 static void xen_machine_power_off(void)
1158 xen_reboot(SHUTDOWN_poweroff);
1161 static void xen_crash_shutdown(struct pt_regs *regs)
1163 xen_reboot(SHUTDOWN_crash);
1166 static const struct machine_ops xen_machine_ops __initconst = {
1167 .restart = xen_restart,
1168 .halt = xen_machine_halt,
1169 .power_off = xen_machine_power_off,
1170 .shutdown = xen_machine_halt,
1171 .crash_shutdown = xen_crash_shutdown,
1172 .emergency_restart = xen_emergency_restart,
1175 static unsigned char xen_get_nmi_reason(void)
1177 unsigned char reason = 0;
1179 /* Construct a value which looks like it came from port 0x61. */
1180 if (test_bit(_XEN_NMIREASON_io_error,
1181 &HYPERVISOR_shared_info->arch.nmi_reason))
1182 reason |= NMI_REASON_IOCHK;
1183 if (test_bit(_XEN_NMIREASON_pci_serr,
1184 &HYPERVISOR_shared_info->arch.nmi_reason))
1185 reason |= NMI_REASON_SERR;
1190 static void __init xen_boot_params_init_edd(void)
1192 #if IS_ENABLED(CONFIG_EDD)
1193 struct xen_platform_op op;
1194 struct edd_info *edd_info;
1199 edd_info = boot_params.eddbuf;
1200 mbr_signature = boot_params.edd_mbr_sig_buffer;
1202 op.cmd = XENPF_firmware_info;
1204 op.u.firmware_info.type = XEN_FW_DISK_INFO;
1205 for (nr = 0; nr < EDDMAXNR; nr++) {
1206 struct edd_info *info = edd_info + nr;
1208 op.u.firmware_info.index = nr;
1209 info->params.length = sizeof(info->params);
1210 set_xen_guest_handle(op.u.firmware_info.u.disk_info.edd_params,
1212 ret = HYPERVISOR_platform_op(&op);
1216 #define C(x) info->x = op.u.firmware_info.u.disk_info.x
1219 C(interface_support);
1220 C(legacy_max_cylinder);
1222 C(legacy_sectors_per_track);
1225 boot_params.eddbuf_entries = nr;
1227 op.u.firmware_info.type = XEN_FW_DISK_MBR_SIGNATURE;
1228 for (nr = 0; nr < EDD_MBR_SIG_MAX; nr++) {
1229 op.u.firmware_info.index = nr;
1230 ret = HYPERVISOR_platform_op(&op);
1233 mbr_signature[nr] = op.u.firmware_info.u.disk_mbr_signature.mbr_signature;
1235 boot_params.edd_mbr_sig_buf_entries = nr;
1240 * Set up the GDT and segment registers for -fstack-protector. Until
1241 * we do this, we have to be careful not to call any stack-protected
1242 * function, which is most of the kernel.
1244 static void xen_setup_gdt(int cpu)
1246 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1247 pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1249 setup_stack_canary_segment(0);
1250 switch_to_new_gdt(0);
1252 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1253 pv_cpu_ops.load_gdt = xen_load_gdt;
1256 static void __init xen_dom0_set_legacy_features(void)
1258 x86_platform.legacy.rtc = 1;
1261 /* First C function to be called on Xen boot */
1262 asmlinkage __visible void __init xen_start_kernel(void)
1264 struct physdev_set_iopl set_iopl;
1265 unsigned long initrd_start = 0;
1268 if (!xen_start_info)
1271 xen_domain_type = XEN_PV_DOMAIN;
1273 xen_setup_features();
1275 xen_setup_machphys_mapping();
1277 /* Install Xen paravirt ops */
1279 pv_init_ops = xen_init_ops;
1280 pv_cpu_ops = xen_cpu_ops;
1282 x86_platform.get_nmi_reason = xen_get_nmi_reason;
1284 x86_init.resources.memory_setup = xen_memory_setup;
1285 x86_init.oem.arch_setup = xen_arch_setup;
1286 x86_init.oem.banner = xen_banner;
1288 xen_init_time_ops();
1291 * Set up some pagetable state before starting to set any ptes.
1296 /* Prevent unwanted bits from being set in PTEs. */
1297 __supported_pte_mask &= ~_PAGE_GLOBAL;
1300 * Prevent page tables from being allocated in highmem, even
1301 * if CONFIG_HIGHPTE is enabled.
1303 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1305 /* Work out if we support NX */
1309 xen_build_dynamic_phys_to_machine();
1312 * Set up kernel GDT and segment registers, mainly so that
1313 * -fstack-protector code can be executed.
1318 xen_init_cpuid_mask();
1319 xen_init_capabilities();
1321 #ifdef CONFIG_X86_LOCAL_APIC
1323 * set up the basic apic ops.
1328 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1329 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1330 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1333 machine_ops = xen_machine_ops;
1336 * The only reliable way to retain the initial address of the
1337 * percpu gdt_page is to remember it here, so we can go and
1338 * mark it RW later, when the initial percpu area is freed.
1340 xen_initial_gdt = &per_cpu(gdt_page, 0);
1344 #ifdef CONFIG_ACPI_NUMA
1346 * The pages we from Xen are not related to machine pages, so
1347 * any NUMA information the kernel tries to get from ACPI will
1348 * be meaningless. Prevent it from trying.
1352 /* Don't do the full vcpu_info placement stuff until we have a
1353 possible map and a non-dummy shared_info. */
1354 per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1356 WARN_ON(xen_cpuhp_setup(xen_cpu_up_prepare_pv, xen_cpu_dead_pv));
1358 local_irq_disable();
1359 early_boot_irqs_disabled = true;
1361 xen_raw_console_write("mapping kernel into physical memory\n");
1362 xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base,
1363 xen_start_info->nr_pages);
1364 xen_reserve_special_pages();
1366 /* keep using Xen gdt for now; no urgent need to change it */
1368 #ifdef CONFIG_X86_32
1369 pv_info.kernel_rpl = 1;
1370 if (xen_feature(XENFEAT_supervisor_mode_kernel))
1371 pv_info.kernel_rpl = 0;
1373 pv_info.kernel_rpl = 0;
1375 /* set the limit of our address space */
1379 * We used to do this in xen_arch_setup, but that is too late
1380 * on AMD were early_cpu_init (run before ->arch_setup()) calls
1381 * early_amd_init which pokes 0xcf8 port.
1384 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1386 xen_raw_printk("physdev_op failed %d\n", rc);
1388 #ifdef CONFIG_X86_32
1389 /* set up basic CPUID stuff */
1390 cpu_detect(&new_cpu_data);
1391 set_cpu_cap(&new_cpu_data, X86_FEATURE_FPU);
1392 new_cpu_data.x86_capability[CPUID_1_EDX] = cpuid_edx(1);
1395 if (xen_start_info->mod_start) {
1396 if (xen_start_info->flags & SIF_MOD_START_PFN)
1397 initrd_start = PFN_PHYS(xen_start_info->mod_start);
1399 initrd_start = __pa(xen_start_info->mod_start);
1402 /* Poke various useful things into boot_params */
1403 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1404 boot_params.hdr.ramdisk_image = initrd_start;
1405 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1406 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1407 boot_params.hdr.hardware_subarch = X86_SUBARCH_XEN;
1409 if (!xen_initial_domain()) {
1410 add_preferred_console("xenboot", 0, NULL);
1411 add_preferred_console("tty", 0, NULL);
1412 add_preferred_console("hvc", 0, NULL);
1414 x86_init.pci.arch_init = pci_xen_init;
1416 const struct dom0_vga_console_info *info =
1417 (void *)((char *)xen_start_info +
1418 xen_start_info->console.dom0.info_off);
1419 struct xen_platform_op op = {
1420 .cmd = XENPF_firmware_info,
1421 .interface_version = XENPF_INTERFACE_VERSION,
1422 .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
1425 x86_platform.set_legacy_features =
1426 xen_dom0_set_legacy_features;
1427 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1428 xen_start_info->console.domU.mfn = 0;
1429 xen_start_info->console.domU.evtchn = 0;
1431 if (HYPERVISOR_platform_op(&op) == 0)
1432 boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
1434 /* Make sure ACS will be enabled */
1437 xen_acpi_sleep_register();
1439 /* Avoid searching for BIOS MP tables */
1440 x86_init.mpparse.find_smp_config = x86_init_noop;
1441 x86_init.mpparse.get_smp_config = x86_init_uint_noop;
1443 xen_boot_params_init_edd();
1446 /* PCI BIOS service won't work from a PV guest. */
1447 pci_probe &= ~PCI_PROBE_BIOS;
1449 xen_raw_console_write("about to get started...\n");
1451 /* Let's presume PV guests always boot on vCPU with id 0. */
1452 per_cpu(xen_vcpu_id, 0) = 0;
1454 xen_setup_runstate_info(0);
1458 /* Start the world */
1459 #ifdef CONFIG_X86_32
1460 i386_start_kernel();
1462 cr4_init_shadow(); /* 32b kernel does this in i386_start_kernel() */
1463 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1467 static int xen_cpu_up_prepare_pv(unsigned int cpu)
1471 xen_setup_timer(cpu);
1473 rc = xen_smp_intr_init(cpu);
1475 WARN(1, "xen_smp_intr_init() for CPU %d failed: %d\n",
1480 rc = xen_smp_intr_init_pv(cpu);
1482 WARN(1, "xen_smp_intr_init_pv() for CPU %d failed: %d\n",
1490 static int xen_cpu_dead_pv(unsigned int cpu)
1492 xen_smp_intr_free(cpu);
1493 xen_smp_intr_free_pv(cpu);
1495 xen_teardown_timer(cpu);
1500 static uint32_t __init xen_platform_pv(void)
1502 if (xen_pv_domain())
1503 return xen_cpuid_base();
1508 const struct hypervisor_x86 x86_hyper_xen_pv = {
1510 .detect = xen_platform_pv,
1511 .pin_vcpu = xen_pin_vcpu,
1513 EXPORT_SYMBOL(x86_hyper_xen_pv);