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 cpuid_leaf1_edx_mask =
298 ~((1 << X86_FEATURE_MTRR) | /* disable MTRR */
299 (1 << X86_FEATURE_ACC)); /* thermal monitoring */
301 if (!xen_initial_domain())
302 cpuid_leaf1_edx_mask &=
303 ~((1 << X86_FEATURE_ACPI)); /* disable ACPI */
305 cpuid_leaf1_ecx_mask &= ~(1 << (X86_FEATURE_X2APIC % 32));
309 cpuid(1, &ax, &bx, &cx, &dx);
312 (1 << (X86_FEATURE_XSAVE % 32)) |
313 (1 << (X86_FEATURE_OSXSAVE % 32));
315 /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
316 if ((cx & xsave_mask) != xsave_mask)
317 cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
318 if (xen_check_mwait())
319 cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
322 static void __init xen_init_capabilities(void)
324 setup_clear_cpu_cap(X86_BUG_SYSRET_SS_ATTRS);
325 setup_force_cpu_cap(X86_FEATURE_XENPV);
326 setup_clear_cpu_cap(X86_FEATURE_DCA);
327 setup_clear_cpu_cap(X86_FEATURE_APERFMPERF);
330 static void xen_set_debugreg(int reg, unsigned long val)
332 HYPERVISOR_set_debugreg(reg, val);
335 static unsigned long xen_get_debugreg(int reg)
337 return HYPERVISOR_get_debugreg(reg);
340 static void xen_end_context_switch(struct task_struct *next)
343 paravirt_end_context_switch(next);
346 static unsigned long xen_store_tr(void)
352 * Set the page permissions for a particular virtual address. If the
353 * address is a vmalloc mapping (or other non-linear mapping), then
354 * find the linear mapping of the page and also set its protections to
357 static void set_aliased_prot(void *v, pgprot_t prot)
366 ptep = lookup_address((unsigned long)v, &level);
367 BUG_ON(ptep == NULL);
369 pfn = pte_pfn(*ptep);
370 page = pfn_to_page(pfn);
372 pte = pfn_pte(pfn, prot);
375 * Careful: update_va_mapping() will fail if the virtual address
376 * we're poking isn't populated in the page tables. We don't
377 * need to worry about the direct map (that's always in the page
378 * tables), but we need to be careful about vmap space. In
379 * particular, the top level page table can lazily propagate
380 * entries between processes, so if we've switched mms since we
381 * vmapped the target in the first place, we might not have the
382 * top-level page table entry populated.
384 * We disable preemption because we want the same mm active when
385 * we probe the target and when we issue the hypercall. We'll
386 * have the same nominal mm, but if we're a kernel thread, lazy
387 * mm dropping could change our pgd.
389 * Out of an abundance of caution, this uses __get_user() to fault
390 * in the target address just in case there's some obscure case
391 * in which the target address isn't readable.
396 probe_kernel_read(&dummy, v, 1);
398 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
401 if (!PageHighMem(page)) {
402 void *av = __va(PFN_PHYS(pfn));
405 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
413 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
415 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
419 * We need to mark the all aliases of the LDT pages RO. We
420 * don't need to call vm_flush_aliases(), though, since that's
421 * only responsible for flushing aliases out the TLBs, not the
422 * page tables, and Xen will flush the TLB for us if needed.
424 * To avoid confusing future readers: none of this is necessary
425 * to load the LDT. The hypervisor only checks this when the
426 * LDT is faulted in due to subsequent descriptor access.
429 for (i = 0; i < entries; i += entries_per_page)
430 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
433 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
435 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
438 for (i = 0; i < entries; i += entries_per_page)
439 set_aliased_prot(ldt + i, PAGE_KERNEL);
442 static void xen_set_ldt(const void *addr, unsigned entries)
444 struct mmuext_op *op;
445 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
447 trace_xen_cpu_set_ldt(addr, entries);
450 op->cmd = MMUEXT_SET_LDT;
451 op->arg1.linear_addr = (unsigned long)addr;
452 op->arg2.nr_ents = entries;
454 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
456 xen_mc_issue(PARAVIRT_LAZY_CPU);
459 static void xen_load_gdt(const struct desc_ptr *dtr)
461 unsigned long va = dtr->address;
462 unsigned int size = dtr->size + 1;
463 unsigned pages = DIV_ROUND_UP(size, PAGE_SIZE);
464 unsigned long frames[pages];
468 * A GDT can be up to 64k in size, which corresponds to 8192
469 * 8-byte entries, or 16 4k pages..
472 BUG_ON(size > 65536);
473 BUG_ON(va & ~PAGE_MASK);
475 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
478 unsigned long pfn, mfn;
482 * The GDT is per-cpu and is in the percpu data area.
483 * That can be virtually mapped, so we need to do a
484 * page-walk to get the underlying MFN for the
485 * hypercall. The page can also be in the kernel's
486 * linear range, so we need to RO that mapping too.
488 ptep = lookup_address(va, &level);
489 BUG_ON(ptep == NULL);
491 pfn = pte_pfn(*ptep);
492 mfn = pfn_to_mfn(pfn);
493 virt = __va(PFN_PHYS(pfn));
497 make_lowmem_page_readonly((void *)va);
498 make_lowmem_page_readonly(virt);
501 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
506 * load_gdt for early boot, when the gdt is only mapped once
508 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
510 unsigned long va = dtr->address;
511 unsigned int size = dtr->size + 1;
512 unsigned pages = DIV_ROUND_UP(size, PAGE_SIZE);
513 unsigned long frames[pages];
517 * A GDT can be up to 64k in size, which corresponds to 8192
518 * 8-byte entries, or 16 4k pages..
521 BUG_ON(size > 65536);
522 BUG_ON(va & ~PAGE_MASK);
524 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
526 unsigned long pfn, mfn;
528 pfn = virt_to_pfn(va);
529 mfn = pfn_to_mfn(pfn);
531 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
533 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
539 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
543 static inline bool desc_equal(const struct desc_struct *d1,
544 const struct desc_struct *d2)
546 return d1->a == d2->a && d1->b == d2->b;
549 static void load_TLS_descriptor(struct thread_struct *t,
550 unsigned int cpu, unsigned int i)
552 struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
553 struct desc_struct *gdt;
555 struct multicall_space mc;
557 if (desc_equal(shadow, &t->tls_array[i]))
560 *shadow = t->tls_array[i];
562 gdt = get_cpu_gdt_rw(cpu);
563 maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
564 mc = __xen_mc_entry(0);
566 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
569 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
572 * XXX sleazy hack: If we're being called in a lazy-cpu zone
573 * and lazy gs handling is enabled, it means we're in a
574 * context switch, and %gs has just been saved. This means we
575 * can zero it out to prevent faults on exit from the
576 * hypervisor if the next process has no %gs. Either way, it
577 * has been saved, and the new value will get loaded properly.
578 * This will go away as soon as Xen has been modified to not
579 * save/restore %gs for normal hypercalls.
581 * On x86_64, this hack is not used for %gs, because gs points
582 * to KERNEL_GS_BASE (and uses it for PDA references), so we
583 * must not zero %gs on x86_64
585 * For x86_64, we need to zero %fs, otherwise we may get an
586 * exception between the new %fs descriptor being loaded and
587 * %fs being effectively cleared at __switch_to().
589 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
599 load_TLS_descriptor(t, cpu, 0);
600 load_TLS_descriptor(t, cpu, 1);
601 load_TLS_descriptor(t, cpu, 2);
603 xen_mc_issue(PARAVIRT_LAZY_CPU);
607 static void xen_load_gs_index(unsigned int idx)
609 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
614 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
617 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
618 u64 entry = *(u64 *)ptr;
620 trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
625 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
631 static int cvt_gate_to_trap(int vector, const gate_desc *val,
632 struct trap_info *info)
636 if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
639 info->vector = vector;
641 addr = gate_offset(*val);
644 * Look for known traps using IST, and substitute them
645 * appropriately. The debugger ones are the only ones we care
646 * about. Xen will handle faults like double_fault,
647 * so we should never see them. Warn if
648 * there's an unexpected IST-using fault handler.
650 if (addr == (unsigned long)debug)
651 addr = (unsigned long)xen_debug;
652 else if (addr == (unsigned long)int3)
653 addr = (unsigned long)xen_int3;
654 else if (addr == (unsigned long)stack_segment)
655 addr = (unsigned long)xen_stack_segment;
656 else if (addr == (unsigned long)double_fault) {
657 /* Don't need to handle these */
659 #ifdef CONFIG_X86_MCE
660 } else if (addr == (unsigned long)machine_check) {
662 * when xen hypervisor inject vMCE to guest,
663 * use native mce handler to handle it
667 } else if (addr == (unsigned long)nmi)
669 * Use the native version as well.
673 /* Some other trap using IST? */
674 if (WARN_ON(val->ist != 0))
677 #endif /* CONFIG_X86_64 */
678 info->address = addr;
680 info->cs = gate_segment(*val);
681 info->flags = val->dpl;
682 /* interrupt gates clear IF */
683 if (val->type == GATE_INTERRUPT)
684 info->flags |= 1 << 2;
689 /* Locations of each CPU's IDT */
690 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
692 /* Set an IDT entry. If the entry is part of the current IDT, then
694 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
696 unsigned long p = (unsigned long)&dt[entrynum];
697 unsigned long start, end;
699 trace_xen_cpu_write_idt_entry(dt, entrynum, g);
703 start = __this_cpu_read(idt_desc.address);
704 end = start + __this_cpu_read(idt_desc.size) + 1;
708 native_write_idt_entry(dt, entrynum, g);
710 if (p >= start && (p + 8) <= end) {
711 struct trap_info info[2];
715 if (cvt_gate_to_trap(entrynum, g, &info[0]))
716 if (HYPERVISOR_set_trap_table(info))
723 static void xen_convert_trap_info(const struct desc_ptr *desc,
724 struct trap_info *traps)
726 unsigned in, out, count;
728 count = (desc->size+1) / sizeof(gate_desc);
731 for (in = out = 0; in < count; in++) {
732 gate_desc *entry = (gate_desc *)(desc->address) + in;
734 if (cvt_gate_to_trap(in, entry, &traps[out]))
737 traps[out].address = 0;
740 void xen_copy_trap_info(struct trap_info *traps)
742 const struct desc_ptr *desc = this_cpu_ptr(&idt_desc);
744 xen_convert_trap_info(desc, traps);
747 /* Load a new IDT into Xen. In principle this can be per-CPU, so we
748 hold a spinlock to protect the static traps[] array (static because
749 it avoids allocation, and saves stack space). */
750 static void xen_load_idt(const struct desc_ptr *desc)
752 static DEFINE_SPINLOCK(lock);
753 static struct trap_info traps[257];
755 trace_xen_cpu_load_idt(desc);
759 memcpy(this_cpu_ptr(&idt_desc), desc, sizeof(idt_desc));
761 xen_convert_trap_info(desc, traps);
764 if (HYPERVISOR_set_trap_table(traps))
770 /* Write a GDT descriptor entry. Ignore LDT descriptors, since
771 they're handled differently. */
772 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
773 const void *desc, int type)
775 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
786 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
789 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
799 * Version of write_gdt_entry for use at early boot-time needed to
800 * update an entry as simply as possible.
802 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
803 const void *desc, int type)
805 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
814 xmaddr_t maddr = virt_to_machine(&dt[entry]);
816 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
817 dt[entry] = *(struct desc_struct *)desc;
823 static void xen_load_sp0(struct tss_struct *tss,
824 struct thread_struct *thread)
826 struct multicall_space mcs;
828 mcs = xen_mc_entry(0);
829 MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
830 xen_mc_issue(PARAVIRT_LAZY_CPU);
831 tss->x86_tss.sp0 = thread->sp0;
834 void xen_set_iopl_mask(unsigned mask)
836 struct physdev_set_iopl set_iopl;
838 /* Force the change at ring 0. */
839 set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
840 HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
843 static void xen_io_delay(void)
847 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
849 static unsigned long xen_read_cr0(void)
851 unsigned long cr0 = this_cpu_read(xen_cr0_value);
853 if (unlikely(cr0 == 0)) {
854 cr0 = native_read_cr0();
855 this_cpu_write(xen_cr0_value, cr0);
861 static void xen_write_cr0(unsigned long cr0)
863 struct multicall_space mcs;
865 this_cpu_write(xen_cr0_value, cr0);
867 /* Only pay attention to cr0.TS; everything else is
869 mcs = xen_mc_entry(0);
871 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
873 xen_mc_issue(PARAVIRT_LAZY_CPU);
876 static void xen_write_cr4(unsigned long cr4)
878 cr4 &= ~(X86_CR4_PGE | X86_CR4_PSE | X86_CR4_PCE);
880 native_write_cr4(cr4);
883 static inline unsigned long xen_read_cr8(void)
887 static inline void xen_write_cr8(unsigned long val)
893 static u64 xen_read_msr_safe(unsigned int msr, int *err)
897 if (pmu_msr_read(msr, &val, err))
900 val = native_read_msr_safe(msr, err);
902 case MSR_IA32_APICBASE:
903 #ifdef CONFIG_X86_X2APIC
904 if (!(cpuid_ecx(1) & (1 << (X86_FEATURE_X2APIC & 31))))
906 val &= ~X2APIC_ENABLE;
912 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
923 case MSR_FS_BASE: which = SEGBASE_FS; goto set;
924 case MSR_KERNEL_GS_BASE: which = SEGBASE_GS_USER; goto set;
925 case MSR_GS_BASE: which = SEGBASE_GS_KERNEL; goto set;
928 base = ((u64)high << 32) | low;
929 if (HYPERVISOR_set_segment_base(which, base) != 0)
937 case MSR_SYSCALL_MASK:
938 case MSR_IA32_SYSENTER_CS:
939 case MSR_IA32_SYSENTER_ESP:
940 case MSR_IA32_SYSENTER_EIP:
941 /* Fast syscall setup is all done in hypercalls, so
942 these are all ignored. Stub them out here to stop
943 Xen console noise. */
947 if (!pmu_msr_write(msr, low, high, &ret))
948 ret = native_write_msr_safe(msr, low, high);
954 static u64 xen_read_msr(unsigned int msr)
957 * This will silently swallow a #GP from RDMSR. It may be worth
962 return xen_read_msr_safe(msr, &err);
965 static void xen_write_msr(unsigned int msr, unsigned low, unsigned high)
968 * This will silently swallow a #GP from WRMSR. It may be worth
971 xen_write_msr_safe(msr, low, high);
974 void xen_setup_shared_info(void)
976 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
977 set_fixmap(FIX_PARAVIRT_BOOTMAP,
978 xen_start_info->shared_info);
980 HYPERVISOR_shared_info =
981 (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
983 HYPERVISOR_shared_info =
984 (struct shared_info *)__va(xen_start_info->shared_info);
987 /* In UP this is as good a place as any to set up shared info */
988 xen_setup_vcpu_info_placement();
991 xen_setup_mfn_list_list();
994 /* This is called once we have the cpu_possible_mask */
995 void xen_setup_vcpu_info_placement(void)
999 for_each_possible_cpu(cpu) {
1000 /* Set up direct vCPU id mapping for PV guests. */
1001 per_cpu(xen_vcpu_id, cpu) = cpu;
1002 xen_vcpu_setup(cpu);
1006 * xen_vcpu_setup managed to place the vcpu_info within the
1007 * percpu area for all cpus, so make use of it.
1009 if (xen_have_vcpu_info_placement) {
1010 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1011 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1012 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1013 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1014 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1018 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1019 unsigned long addr, unsigned len)
1021 char *start, *end, *reloc;
1024 start = end = reloc = NULL;
1026 #define SITE(op, x) \
1027 case PARAVIRT_PATCH(op.x): \
1028 if (xen_have_vcpu_info_placement) { \
1029 start = (char *)xen_##x##_direct; \
1030 end = xen_##x##_direct_end; \
1031 reloc = xen_##x##_direct_reloc; \
1036 SITE(pv_irq_ops, irq_enable);
1037 SITE(pv_irq_ops, irq_disable);
1038 SITE(pv_irq_ops, save_fl);
1039 SITE(pv_irq_ops, restore_fl);
1043 if (start == NULL || (end-start) > len)
1046 ret = paravirt_patch_insns(insnbuf, len, start, end);
1048 /* Note: because reloc is assigned from something that
1049 appears to be an array, gcc assumes it's non-null,
1050 but doesn't know its relationship with start and
1052 if (reloc > start && reloc < end) {
1053 int reloc_off = reloc - start;
1054 long *relocp = (long *)(insnbuf + reloc_off);
1055 long delta = start - (char *)addr;
1063 ret = paravirt_patch_default(type, clobbers, insnbuf,
1071 static const struct pv_info xen_info __initconst = {
1072 .shared_kernel_pmd = 0,
1074 #ifdef CONFIG_X86_64
1075 .extra_user_64bit_cs = FLAT_USER_CS64,
1080 static const struct pv_init_ops xen_init_ops __initconst = {
1084 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1087 .set_debugreg = xen_set_debugreg,
1088 .get_debugreg = xen_get_debugreg,
1090 .read_cr0 = xen_read_cr0,
1091 .write_cr0 = xen_write_cr0,
1093 .read_cr4 = native_read_cr4,
1094 .write_cr4 = xen_write_cr4,
1096 #ifdef CONFIG_X86_64
1097 .read_cr8 = xen_read_cr8,
1098 .write_cr8 = xen_write_cr8,
1101 .wbinvd = native_wbinvd,
1103 .read_msr = xen_read_msr,
1104 .write_msr = xen_write_msr,
1106 .read_msr_safe = xen_read_msr_safe,
1107 .write_msr_safe = xen_write_msr_safe,
1109 .read_pmc = xen_read_pmc,
1112 #ifdef CONFIG_X86_64
1113 .usergs_sysret64 = xen_sysret64,
1116 .load_tr_desc = paravirt_nop,
1117 .set_ldt = xen_set_ldt,
1118 .load_gdt = xen_load_gdt,
1119 .load_idt = xen_load_idt,
1120 .load_tls = xen_load_tls,
1121 #ifdef CONFIG_X86_64
1122 .load_gs_index = xen_load_gs_index,
1125 .alloc_ldt = xen_alloc_ldt,
1126 .free_ldt = xen_free_ldt,
1128 .store_idt = native_store_idt,
1129 .store_tr = xen_store_tr,
1131 .write_ldt_entry = xen_write_ldt_entry,
1132 .write_gdt_entry = xen_write_gdt_entry,
1133 .write_idt_entry = xen_write_idt_entry,
1134 .load_sp0 = xen_load_sp0,
1136 .set_iopl_mask = xen_set_iopl_mask,
1137 .io_delay = xen_io_delay,
1139 /* Xen takes care of %gs when switching to usermode for us */
1140 .swapgs = paravirt_nop,
1142 .start_context_switch = paravirt_start_context_switch,
1143 .end_context_switch = xen_end_context_switch,
1146 static void xen_restart(char *msg)
1148 xen_reboot(SHUTDOWN_reboot);
1151 static void xen_machine_halt(void)
1153 xen_reboot(SHUTDOWN_poweroff);
1156 static void xen_machine_power_off(void)
1160 xen_reboot(SHUTDOWN_poweroff);
1163 static void xen_crash_shutdown(struct pt_regs *regs)
1165 xen_reboot(SHUTDOWN_crash);
1168 static const struct machine_ops xen_machine_ops __initconst = {
1169 .restart = xen_restart,
1170 .halt = xen_machine_halt,
1171 .power_off = xen_machine_power_off,
1172 .shutdown = xen_machine_halt,
1173 .crash_shutdown = xen_crash_shutdown,
1174 .emergency_restart = xen_emergency_restart,
1177 static unsigned char xen_get_nmi_reason(void)
1179 unsigned char reason = 0;
1181 /* Construct a value which looks like it came from port 0x61. */
1182 if (test_bit(_XEN_NMIREASON_io_error,
1183 &HYPERVISOR_shared_info->arch.nmi_reason))
1184 reason |= NMI_REASON_IOCHK;
1185 if (test_bit(_XEN_NMIREASON_pci_serr,
1186 &HYPERVISOR_shared_info->arch.nmi_reason))
1187 reason |= NMI_REASON_SERR;
1192 static void __init xen_boot_params_init_edd(void)
1194 #if IS_ENABLED(CONFIG_EDD)
1195 struct xen_platform_op op;
1196 struct edd_info *edd_info;
1201 edd_info = boot_params.eddbuf;
1202 mbr_signature = boot_params.edd_mbr_sig_buffer;
1204 op.cmd = XENPF_firmware_info;
1206 op.u.firmware_info.type = XEN_FW_DISK_INFO;
1207 for (nr = 0; nr < EDDMAXNR; nr++) {
1208 struct edd_info *info = edd_info + nr;
1210 op.u.firmware_info.index = nr;
1211 info->params.length = sizeof(info->params);
1212 set_xen_guest_handle(op.u.firmware_info.u.disk_info.edd_params,
1214 ret = HYPERVISOR_platform_op(&op);
1218 #define C(x) info->x = op.u.firmware_info.u.disk_info.x
1221 C(interface_support);
1222 C(legacy_max_cylinder);
1224 C(legacy_sectors_per_track);
1227 boot_params.eddbuf_entries = nr;
1229 op.u.firmware_info.type = XEN_FW_DISK_MBR_SIGNATURE;
1230 for (nr = 0; nr < EDD_MBR_SIG_MAX; nr++) {
1231 op.u.firmware_info.index = nr;
1232 ret = HYPERVISOR_platform_op(&op);
1235 mbr_signature[nr] = op.u.firmware_info.u.disk_mbr_signature.mbr_signature;
1237 boot_params.edd_mbr_sig_buf_entries = nr;
1242 * Set up the GDT and segment registers for -fstack-protector. Until
1243 * we do this, we have to be careful not to call any stack-protected
1244 * function, which is most of the kernel.
1246 static void xen_setup_gdt(int cpu)
1248 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1249 pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1251 setup_stack_canary_segment(0);
1252 switch_to_new_gdt(0);
1254 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1255 pv_cpu_ops.load_gdt = xen_load_gdt;
1258 static void __init xen_dom0_set_legacy_features(void)
1260 x86_platform.legacy.rtc = 1;
1263 /* First C function to be called on Xen boot */
1264 asmlinkage __visible void __init xen_start_kernel(void)
1266 struct physdev_set_iopl set_iopl;
1267 unsigned long initrd_start = 0;
1270 if (!xen_start_info)
1273 xen_domain_type = XEN_PV_DOMAIN;
1275 xen_setup_features();
1277 xen_setup_machphys_mapping();
1279 /* Install Xen paravirt ops */
1281 pv_init_ops = xen_init_ops;
1282 pv_cpu_ops = xen_cpu_ops;
1284 x86_platform.get_nmi_reason = xen_get_nmi_reason;
1286 x86_init.resources.memory_setup = xen_memory_setup;
1287 x86_init.oem.arch_setup = xen_arch_setup;
1288 x86_init.oem.banner = xen_banner;
1290 xen_init_time_ops();
1293 * Set up some pagetable state before starting to set any ptes.
1298 /* Prevent unwanted bits from being set in PTEs. */
1299 __supported_pte_mask &= ~_PAGE_GLOBAL;
1302 * Prevent page tables from being allocated in highmem, even
1303 * if CONFIG_HIGHPTE is enabled.
1305 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1307 /* Work out if we support NX */
1311 xen_build_dynamic_phys_to_machine();
1314 * Set up kernel GDT and segment registers, mainly so that
1315 * -fstack-protector code can be executed.
1320 xen_init_cpuid_mask();
1321 xen_init_capabilities();
1323 #ifdef CONFIG_X86_LOCAL_APIC
1325 * set up the basic apic ops.
1330 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1331 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1332 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1335 machine_ops = xen_machine_ops;
1338 * The only reliable way to retain the initial address of the
1339 * percpu gdt_page is to remember it here, so we can go and
1340 * mark it RW later, when the initial percpu area is freed.
1342 xen_initial_gdt = &per_cpu(gdt_page, 0);
1346 #ifdef CONFIG_ACPI_NUMA
1348 * The pages we from Xen are not related to machine pages, so
1349 * any NUMA information the kernel tries to get from ACPI will
1350 * be meaningless. Prevent it from trying.
1354 /* Don't do the full vcpu_info placement stuff until we have a
1355 possible map and a non-dummy shared_info. */
1356 per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1358 WARN_ON(xen_cpuhp_setup(xen_cpu_up_prepare_pv, xen_cpu_dead_pv));
1360 local_irq_disable();
1361 early_boot_irqs_disabled = true;
1363 xen_raw_console_write("mapping kernel into physical memory\n");
1364 xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base,
1365 xen_start_info->nr_pages);
1366 xen_reserve_special_pages();
1368 /* keep using Xen gdt for now; no urgent need to change it */
1370 #ifdef CONFIG_X86_32
1371 pv_info.kernel_rpl = 1;
1372 if (xen_feature(XENFEAT_supervisor_mode_kernel))
1373 pv_info.kernel_rpl = 0;
1375 pv_info.kernel_rpl = 0;
1377 /* set the limit of our address space */
1381 * We used to do this in xen_arch_setup, but that is too late
1382 * on AMD were early_cpu_init (run before ->arch_setup()) calls
1383 * early_amd_init which pokes 0xcf8 port.
1386 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1388 xen_raw_printk("physdev_op failed %d\n", rc);
1390 #ifdef CONFIG_X86_32
1391 /* set up basic CPUID stuff */
1392 cpu_detect(&new_cpu_data);
1393 set_cpu_cap(&new_cpu_data, X86_FEATURE_FPU);
1394 new_cpu_data.x86_capability[CPUID_1_EDX] = cpuid_edx(1);
1397 if (xen_start_info->mod_start) {
1398 if (xen_start_info->flags & SIF_MOD_START_PFN)
1399 initrd_start = PFN_PHYS(xen_start_info->mod_start);
1401 initrd_start = __pa(xen_start_info->mod_start);
1404 /* Poke various useful things into boot_params */
1405 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1406 boot_params.hdr.ramdisk_image = initrd_start;
1407 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1408 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1409 boot_params.hdr.hardware_subarch = X86_SUBARCH_XEN;
1411 if (!xen_initial_domain()) {
1412 add_preferred_console("xenboot", 0, NULL);
1413 add_preferred_console("tty", 0, NULL);
1414 add_preferred_console("hvc", 0, NULL);
1416 x86_init.pci.arch_init = pci_xen_init;
1418 const struct dom0_vga_console_info *info =
1419 (void *)((char *)xen_start_info +
1420 xen_start_info->console.dom0.info_off);
1421 struct xen_platform_op op = {
1422 .cmd = XENPF_firmware_info,
1423 .interface_version = XENPF_INTERFACE_VERSION,
1424 .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
1427 x86_platform.set_legacy_features =
1428 xen_dom0_set_legacy_features;
1429 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1430 xen_start_info->console.domU.mfn = 0;
1431 xen_start_info->console.domU.evtchn = 0;
1433 if (HYPERVISOR_platform_op(&op) == 0)
1434 boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
1436 /* Make sure ACS will be enabled */
1439 xen_acpi_sleep_register();
1441 /* Avoid searching for BIOS MP tables */
1442 x86_init.mpparse.find_smp_config = x86_init_noop;
1443 x86_init.mpparse.get_smp_config = x86_init_uint_noop;
1445 xen_boot_params_init_edd();
1448 /* PCI BIOS service won't work from a PV guest. */
1449 pci_probe &= ~PCI_PROBE_BIOS;
1451 xen_raw_console_write("about to get started...\n");
1453 /* Let's presume PV guests always boot on vCPU with id 0. */
1454 per_cpu(xen_vcpu_id, 0) = 0;
1456 xen_setup_runstate_info(0);
1460 /* Start the world */
1461 #ifdef CONFIG_X86_32
1462 i386_start_kernel();
1464 cr4_init_shadow(); /* 32b kernel does this in i386_start_kernel() */
1465 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1469 static int xen_cpu_up_prepare_pv(unsigned int cpu)
1473 xen_setup_timer(cpu);
1475 rc = xen_smp_intr_init(cpu);
1477 WARN(1, "xen_smp_intr_init() for CPU %d failed: %d\n",
1482 rc = xen_smp_intr_init_pv(cpu);
1484 WARN(1, "xen_smp_intr_init_pv() for CPU %d failed: %d\n",
1492 static int xen_cpu_dead_pv(unsigned int cpu)
1494 xen_smp_intr_free(cpu);
1495 xen_smp_intr_free_pv(cpu);
1497 xen_teardown_timer(cpu);
1502 static uint32_t __init xen_platform_pv(void)
1504 if (xen_pv_domain())
1505 return xen_cpuid_base();
1510 const struct hypervisor_x86 x86_hyper_xen_pv = {
1512 .detect = xen_platform_pv,
1513 .pin_vcpu = xen_pin_vcpu,
1515 EXPORT_SYMBOL(x86_hyper_xen_pv);