1 // SPDX-License-Identifier: GPL-2.0
3 * Core of Xen paravirt_ops implementation.
5 * This file contains the xen_paravirt_ops structure itself, and the
7 * - privileged instructions
12 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
15 #include <linux/cpu.h>
16 #include <linux/kernel.h>
17 #include <linux/init.h>
18 #include <linux/smp.h>
19 #include <linux/preempt.h>
20 #include <linux/hardirq.h>
21 #include <linux/percpu.h>
22 #include <linux/delay.h>
23 #include <linux/start_kernel.h>
24 #include <linux/sched.h>
25 #include <linux/kprobes.h>
26 #include <linux/memblock.h>
27 #include <linux/export.h>
29 #include <linux/page-flags.h>
30 #include <linux/highmem.h>
31 #include <linux/console.h>
32 #include <linux/pci.h>
33 #include <linux/gfp.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/tlbflush.h>
67 #include <asm/reboot.h>
68 #include <asm/stackprotector.h>
69 #include <asm/hypervisor.h>
70 #include <asm/mach_traps.h>
71 #include <asm/mwait.h>
72 #include <asm/pci_x86.h>
74 #ifdef CONFIG_X86_IOPL_IOPERM
75 #include <asm/io_bitmap.h>
79 #include <linux/acpi.h>
81 #include <acpi/pdc_intel.h>
82 #include <acpi/processor.h>
83 #include <xen/interface/platform.h>
89 #include "multicalls.h"
92 #include "../kernel/cpu/cpu.h" /* get_cpu_cap() */
94 void *xen_initial_gdt;
96 static int xen_cpu_up_prepare_pv(unsigned int cpu);
97 static int xen_cpu_dead_pv(unsigned int cpu);
100 struct desc_struct desc[3];
104 * Updating the 3 TLS descriptors in the GDT on every task switch is
105 * surprisingly expensive so we avoid updating them if they haven't
106 * changed. Since Xen writes different descriptors than the one
107 * passed in the update_descriptor hypercall we keep shadow copies to
110 static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
112 static void __init xen_banner(void)
114 unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
115 struct xen_extraversion extra;
116 HYPERVISOR_xen_version(XENVER_extraversion, &extra);
118 pr_info("Booting paravirtualized kernel on %s\n", pv_info.name);
119 printk(KERN_INFO "Xen version: %d.%d%s%s\n",
120 version >> 16, version & 0xffff, extra.extraversion,
121 xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
124 pr_warn("WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! WARNING!\n"
125 "Support for running as 32-bit PV-guest under Xen will soon be removed\n"
126 "from the Linux kernel!\n"
127 "Please use either a 64-bit kernel or switch to HVM or PVH mode!\n"
128 "WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! WARNING!\n");
132 static void __init xen_pv_init_platform(void)
134 populate_extra_pte(fix_to_virt(FIX_PARAVIRT_BOOTMAP));
136 set_fixmap(FIX_PARAVIRT_BOOTMAP, xen_start_info->shared_info);
137 HYPERVISOR_shared_info = (void *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
139 /* xen clock uses per-cpu vcpu_info, need to init it for boot cpu */
140 xen_vcpu_info_reset(0);
142 /* pvclock is in shared info area */
146 static void __init xen_pv_guest_late_init(void)
149 /* Setup shared vcpu info for non-smp configurations */
150 xen_setup_vcpu_info_placement();
154 /* Check if running on Xen version (major, minor) or later */
156 xen_running_on_version_or_later(unsigned int major, unsigned int minor)
158 unsigned int version;
163 version = HYPERVISOR_xen_version(XENVER_version, NULL);
164 if ((((version >> 16) == major) && ((version & 0xffff) >= minor)) ||
165 ((version >> 16) > major))
170 static __read_mostly unsigned int cpuid_leaf5_ecx_val;
171 static __read_mostly unsigned int cpuid_leaf5_edx_val;
173 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
174 unsigned int *cx, unsigned int *dx)
176 unsigned maskebx = ~0;
179 * Mask out inconvenient features, to try and disable as many
180 * unsupported kernel subsystems as possible.
183 case CPUID_MWAIT_LEAF:
184 /* Synthesize the values.. */
187 *cx = cpuid_leaf5_ecx_val;
188 *dx = cpuid_leaf5_edx_val;
192 /* Suppress extended topology stuff */
197 asm(XEN_EMULATE_PREFIX "cpuid"
202 : "0" (*ax), "2" (*cx));
206 STACK_FRAME_NON_STANDARD(xen_cpuid); /* XEN_EMULATE_PREFIX */
208 static bool __init xen_check_mwait(void)
211 struct xen_platform_op op = {
212 .cmd = XENPF_set_processor_pminfo,
213 .u.set_pminfo.id = -1,
214 .u.set_pminfo.type = XEN_PM_PDC,
217 unsigned int ax, bx, cx, dx;
218 unsigned int mwait_mask;
220 /* We need to determine whether it is OK to expose the MWAIT
221 * capability to the kernel to harvest deeper than C3 states from ACPI
222 * _CST using the processor_harvest_xen.c module. For this to work, we
223 * need to gather the MWAIT_LEAF values (which the cstate.c code
224 * checks against). The hypervisor won't expose the MWAIT flag because
225 * it would break backwards compatibility; so we will find out directly
226 * from the hardware and hypercall.
228 if (!xen_initial_domain())
232 * When running under platform earlier than Xen4.2, do not expose
233 * mwait, to avoid the risk of loading native acpi pad driver
235 if (!xen_running_on_version_or_later(4, 2))
241 native_cpuid(&ax, &bx, &cx, &dx);
243 mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
244 (1 << (X86_FEATURE_MWAIT % 32));
246 if ((cx & mwait_mask) != mwait_mask)
249 /* We need to emulate the MWAIT_LEAF and for that we need both
250 * ecx and edx. The hypercall provides only partial information.
253 ax = CPUID_MWAIT_LEAF;
258 native_cpuid(&ax, &bx, &cx, &dx);
260 /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
261 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
263 buf[0] = ACPI_PDC_REVISION_ID;
265 buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
267 set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
269 if ((HYPERVISOR_platform_op(&op) == 0) &&
270 (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
271 cpuid_leaf5_ecx_val = cx;
272 cpuid_leaf5_edx_val = dx;
280 static bool __init xen_check_xsave(void)
282 unsigned int cx, xsave_mask;
286 xsave_mask = (1 << (X86_FEATURE_XSAVE % 32)) |
287 (1 << (X86_FEATURE_OSXSAVE % 32));
289 /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
290 return (cx & xsave_mask) == xsave_mask;
293 static void __init xen_init_capabilities(void)
295 setup_force_cpu_cap(X86_FEATURE_XENPV);
296 setup_clear_cpu_cap(X86_FEATURE_DCA);
297 setup_clear_cpu_cap(X86_FEATURE_APERFMPERF);
298 setup_clear_cpu_cap(X86_FEATURE_MTRR);
299 setup_clear_cpu_cap(X86_FEATURE_ACC);
300 setup_clear_cpu_cap(X86_FEATURE_X2APIC);
301 setup_clear_cpu_cap(X86_FEATURE_SME);
304 * Xen PV would need some work to support PCID: CR3 handling as well
305 * as xen_flush_tlb_others() would need updating.
307 setup_clear_cpu_cap(X86_FEATURE_PCID);
309 if (!xen_initial_domain())
310 setup_clear_cpu_cap(X86_FEATURE_ACPI);
312 if (xen_check_mwait())
313 setup_force_cpu_cap(X86_FEATURE_MWAIT);
315 setup_clear_cpu_cap(X86_FEATURE_MWAIT);
317 if (!xen_check_xsave()) {
318 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
319 setup_clear_cpu_cap(X86_FEATURE_OSXSAVE);
323 static void xen_set_debugreg(int reg, unsigned long val)
325 HYPERVISOR_set_debugreg(reg, val);
328 static unsigned long xen_get_debugreg(int reg)
330 return HYPERVISOR_get_debugreg(reg);
333 static void xen_end_context_switch(struct task_struct *next)
336 paravirt_end_context_switch(next);
339 static unsigned long xen_store_tr(void)
345 * Set the page permissions for a particular virtual address. If the
346 * address is a vmalloc mapping (or other non-linear mapping), then
347 * find the linear mapping of the page and also set its protections to
350 static void set_aliased_prot(void *v, pgprot_t prot)
359 ptep = lookup_address((unsigned long)v, &level);
360 BUG_ON(ptep == NULL);
362 pfn = pte_pfn(*ptep);
363 page = pfn_to_page(pfn);
365 pte = pfn_pte(pfn, prot);
368 * Careful: update_va_mapping() will fail if the virtual address
369 * we're poking isn't populated in the page tables. We don't
370 * need to worry about the direct map (that's always in the page
371 * tables), but we need to be careful about vmap space. In
372 * particular, the top level page table can lazily propagate
373 * entries between processes, so if we've switched mms since we
374 * vmapped the target in the first place, we might not have the
375 * top-level page table entry populated.
377 * We disable preemption because we want the same mm active when
378 * we probe the target and when we issue the hypercall. We'll
379 * have the same nominal mm, but if we're a kernel thread, lazy
380 * mm dropping could change our pgd.
382 * Out of an abundance of caution, this uses __get_user() to fault
383 * in the target address just in case there's some obscure case
384 * in which the target address isn't readable.
389 copy_from_kernel_nofault(&dummy, v, 1);
391 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
394 if (!PageHighMem(page)) {
395 void *av = __va(PFN_PHYS(pfn));
398 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
406 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
408 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
412 * We need to mark the all aliases of the LDT pages RO. We
413 * don't need to call vm_flush_aliases(), though, since that's
414 * only responsible for flushing aliases out the TLBs, not the
415 * page tables, and Xen will flush the TLB for us if needed.
417 * To avoid confusing future readers: none of this is necessary
418 * to load the LDT. The hypervisor only checks this when the
419 * LDT is faulted in due to subsequent descriptor access.
422 for (i = 0; i < entries; i += entries_per_page)
423 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
426 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
428 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
431 for (i = 0; i < entries; i += entries_per_page)
432 set_aliased_prot(ldt + i, PAGE_KERNEL);
435 static void xen_set_ldt(const void *addr, unsigned entries)
437 struct mmuext_op *op;
438 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
440 trace_xen_cpu_set_ldt(addr, entries);
443 op->cmd = MMUEXT_SET_LDT;
444 op->arg1.linear_addr = (unsigned long)addr;
445 op->arg2.nr_ents = entries;
447 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
449 xen_mc_issue(PARAVIRT_LAZY_CPU);
452 static void xen_load_gdt(const struct desc_ptr *dtr)
454 unsigned long va = dtr->address;
455 unsigned int size = dtr->size + 1;
456 unsigned long pfn, mfn;
461 /* @size should be at most GDT_SIZE which is smaller than PAGE_SIZE. */
462 BUG_ON(size > PAGE_SIZE);
463 BUG_ON(va & ~PAGE_MASK);
466 * The GDT is per-cpu and is in the percpu data area.
467 * That can be virtually mapped, so we need to do a
468 * page-walk to get the underlying MFN for the
469 * hypercall. The page can also be in the kernel's
470 * linear range, so we need to RO that mapping too.
472 ptep = lookup_address(va, &level);
473 BUG_ON(ptep == NULL);
475 pfn = pte_pfn(*ptep);
476 mfn = pfn_to_mfn(pfn);
477 virt = __va(PFN_PHYS(pfn));
479 make_lowmem_page_readonly((void *)va);
480 make_lowmem_page_readonly(virt);
482 if (HYPERVISOR_set_gdt(&mfn, size / sizeof(struct desc_struct)))
487 * load_gdt for early boot, when the gdt is only mapped once
489 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
491 unsigned long va = dtr->address;
492 unsigned int size = dtr->size + 1;
493 unsigned long pfn, mfn;
496 /* @size should be at most GDT_SIZE which is smaller than PAGE_SIZE. */
497 BUG_ON(size > PAGE_SIZE);
498 BUG_ON(va & ~PAGE_MASK);
500 pfn = virt_to_pfn(va);
501 mfn = pfn_to_mfn(pfn);
503 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
505 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
508 if (HYPERVISOR_set_gdt(&mfn, size / sizeof(struct desc_struct)))
512 static inline bool desc_equal(const struct desc_struct *d1,
513 const struct desc_struct *d2)
515 return !memcmp(d1, d2, sizeof(*d1));
518 static void load_TLS_descriptor(struct thread_struct *t,
519 unsigned int cpu, unsigned int i)
521 struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
522 struct desc_struct *gdt;
524 struct multicall_space mc;
526 if (desc_equal(shadow, &t->tls_array[i]))
529 *shadow = t->tls_array[i];
531 gdt = get_cpu_gdt_rw(cpu);
532 maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
533 mc = __xen_mc_entry(0);
535 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
538 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
541 * XXX sleazy hack: If we're being called in a lazy-cpu zone
542 * and lazy gs handling is enabled, it means we're in a
543 * context switch, and %gs has just been saved. This means we
544 * can zero it out to prevent faults on exit from the
545 * hypervisor if the next process has no %gs. Either way, it
546 * has been saved, and the new value will get loaded properly.
547 * This will go away as soon as Xen has been modified to not
548 * save/restore %gs for normal hypercalls.
550 * On x86_64, this hack is not used for %gs, because gs points
551 * to KERNEL_GS_BASE (and uses it for PDA references), so we
552 * must not zero %gs on x86_64
554 * For x86_64, we need to zero %fs, otherwise we may get an
555 * exception between the new %fs descriptor being loaded and
556 * %fs being effectively cleared at __switch_to().
558 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
568 load_TLS_descriptor(t, cpu, 0);
569 load_TLS_descriptor(t, cpu, 1);
570 load_TLS_descriptor(t, cpu, 2);
572 xen_mc_issue(PARAVIRT_LAZY_CPU);
576 static void xen_load_gs_index(unsigned int idx)
578 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
583 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
586 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
587 u64 entry = *(u64 *)ptr;
589 trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
594 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
601 void noist_exc_debug(struct pt_regs *regs);
603 DEFINE_IDTENTRY_RAW(xenpv_exc_nmi)
605 /* On Xen PV, NMI doesn't use IST. The C part is the sane as native. */
609 DEFINE_IDTENTRY_RAW(xenpv_exc_debug)
612 * There's no IST on Xen PV, but we still need to dispatch
613 * to the correct handler.
616 noist_exc_debug(regs);
621 struct trap_array_entry {
627 #define TRAP_ENTRY(func, ist_ok) { \
628 .orig = asm_##func, \
629 .xen = xen_asm_##func, \
632 #define TRAP_ENTRY_REDIR(func, ist_ok) { \
633 .orig = asm_##func, \
634 .xen = xen_asm_xenpv_##func, \
637 static struct trap_array_entry trap_array[] = {
638 TRAP_ENTRY_REDIR(exc_debug, true ),
639 TRAP_ENTRY(exc_double_fault, true ),
640 #ifdef CONFIG_X86_MCE
641 TRAP_ENTRY(exc_machine_check, true ),
643 TRAP_ENTRY_REDIR(exc_nmi, true ),
644 TRAP_ENTRY(exc_int3, false ),
645 TRAP_ENTRY(exc_overflow, false ),
646 #ifdef CONFIG_IA32_EMULATION
647 { entry_INT80_compat, xen_entry_INT80_compat, false },
649 TRAP_ENTRY(exc_page_fault, false ),
650 TRAP_ENTRY(exc_divide_error, false ),
651 TRAP_ENTRY(exc_bounds, false ),
652 TRAP_ENTRY(exc_invalid_op, false ),
653 TRAP_ENTRY(exc_device_not_available, false ),
654 TRAP_ENTRY(exc_coproc_segment_overrun, false ),
655 TRAP_ENTRY(exc_invalid_tss, false ),
656 TRAP_ENTRY(exc_segment_not_present, false ),
657 TRAP_ENTRY(exc_stack_segment, false ),
658 TRAP_ENTRY(exc_general_protection, false ),
659 TRAP_ENTRY(exc_spurious_interrupt_bug, false ),
660 TRAP_ENTRY(exc_coprocessor_error, false ),
661 TRAP_ENTRY(exc_alignment_check, false ),
662 TRAP_ENTRY(exc_simd_coprocessor_error, false ),
665 static bool __ref get_trap_addr(void **addr, unsigned int ist)
668 bool ist_okay = false;
671 * Replace trap handler addresses by Xen specific ones.
672 * Check for known traps using IST and whitelist them.
673 * The debugger ones are the only ones we care about.
674 * Xen will handle faults like double_fault, so we should never see
675 * them. Warn if there's an unexpected IST-using fault handler.
677 for (nr = 0; nr < ARRAY_SIZE(trap_array); nr++) {
678 struct trap_array_entry *entry = trap_array + nr;
680 if (*addr == entry->orig) {
682 ist_okay = entry->ist_okay;
687 if (nr == ARRAY_SIZE(trap_array) &&
688 *addr >= (void *)early_idt_handler_array[0] &&
689 *addr < (void *)early_idt_handler_array[NUM_EXCEPTION_VECTORS]) {
690 nr = (*addr - (void *)early_idt_handler_array[0]) /
691 EARLY_IDT_HANDLER_SIZE;
692 *addr = (void *)xen_early_idt_handler_array[nr];
695 if (WARN_ON(ist != 0 && !ist_okay))
702 static int cvt_gate_to_trap(int vector, const gate_desc *val,
703 struct trap_info *info)
707 if (val->bits.type != GATE_TRAP && val->bits.type != GATE_INTERRUPT)
710 info->vector = vector;
712 addr = gate_offset(val);
714 if (!get_trap_addr((void **)&addr, val->bits.ist))
716 #endif /* CONFIG_X86_64 */
717 info->address = addr;
719 info->cs = gate_segment(val);
720 info->flags = val->bits.dpl;
721 /* interrupt gates clear IF */
722 if (val->bits.type == GATE_INTERRUPT)
723 info->flags |= 1 << 2;
728 /* Locations of each CPU's IDT */
729 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
731 /* Set an IDT entry. If the entry is part of the current IDT, then
733 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
735 unsigned long p = (unsigned long)&dt[entrynum];
736 unsigned long start, end;
738 trace_xen_cpu_write_idt_entry(dt, entrynum, g);
742 start = __this_cpu_read(idt_desc.address);
743 end = start + __this_cpu_read(idt_desc.size) + 1;
747 native_write_idt_entry(dt, entrynum, g);
749 if (p >= start && (p + 8) <= end) {
750 struct trap_info info[2];
754 if (cvt_gate_to_trap(entrynum, g, &info[0]))
755 if (HYPERVISOR_set_trap_table(info))
762 static void xen_convert_trap_info(const struct desc_ptr *desc,
763 struct trap_info *traps)
765 unsigned in, out, count;
767 count = (desc->size+1) / sizeof(gate_desc);
770 for (in = out = 0; in < count; in++) {
771 gate_desc *entry = (gate_desc *)(desc->address) + in;
773 if (cvt_gate_to_trap(in, entry, &traps[out]))
776 traps[out].address = 0;
779 void xen_copy_trap_info(struct trap_info *traps)
781 const struct desc_ptr *desc = this_cpu_ptr(&idt_desc);
783 xen_convert_trap_info(desc, traps);
786 /* Load a new IDT into Xen. In principle this can be per-CPU, so we
787 hold a spinlock to protect the static traps[] array (static because
788 it avoids allocation, and saves stack space). */
789 static void xen_load_idt(const struct desc_ptr *desc)
791 static DEFINE_SPINLOCK(lock);
792 static struct trap_info traps[257];
794 trace_xen_cpu_load_idt(desc);
798 memcpy(this_cpu_ptr(&idt_desc), desc, sizeof(idt_desc));
800 xen_convert_trap_info(desc, traps);
803 if (HYPERVISOR_set_trap_table(traps))
809 /* Write a GDT descriptor entry. Ignore LDT descriptors, since
810 they're handled differently. */
811 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
812 const void *desc, int type)
814 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
825 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
828 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
838 * Version of write_gdt_entry for use at early boot-time needed to
839 * update an entry as simply as possible.
841 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
842 const void *desc, int type)
844 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
853 xmaddr_t maddr = virt_to_machine(&dt[entry]);
855 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
856 dt[entry] = *(struct desc_struct *)desc;
862 static void xen_load_sp0(unsigned long sp0)
864 struct multicall_space mcs;
866 mcs = xen_mc_entry(0);
867 MULTI_stack_switch(mcs.mc, __KERNEL_DS, sp0);
868 xen_mc_issue(PARAVIRT_LAZY_CPU);
869 this_cpu_write(cpu_tss_rw.x86_tss.sp0, sp0);
872 #ifdef CONFIG_X86_IOPL_IOPERM
873 static void xen_update_io_bitmap(void)
875 struct physdev_set_iobitmap iobitmap;
876 struct tss_struct *tss = this_cpu_ptr(&cpu_tss_rw);
878 native_tss_update_io_bitmap();
880 iobitmap.bitmap = (uint8_t *)(&tss->x86_tss) +
881 tss->x86_tss.io_bitmap_base;
882 if (tss->x86_tss.io_bitmap_base == IO_BITMAP_OFFSET_INVALID)
883 iobitmap.nr_ports = 0;
885 iobitmap.nr_ports = IO_BITMAP_BITS;
887 HYPERVISOR_physdev_op(PHYSDEVOP_set_iobitmap, &iobitmap);
891 static void xen_io_delay(void)
895 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
897 static unsigned long xen_read_cr0(void)
899 unsigned long cr0 = this_cpu_read(xen_cr0_value);
901 if (unlikely(cr0 == 0)) {
902 cr0 = native_read_cr0();
903 this_cpu_write(xen_cr0_value, cr0);
909 static void xen_write_cr0(unsigned long cr0)
911 struct multicall_space mcs;
913 this_cpu_write(xen_cr0_value, cr0);
915 /* Only pay attention to cr0.TS; everything else is
917 mcs = xen_mc_entry(0);
919 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
921 xen_mc_issue(PARAVIRT_LAZY_CPU);
924 static void xen_write_cr4(unsigned long cr4)
926 cr4 &= ~(X86_CR4_PGE | X86_CR4_PSE | X86_CR4_PCE);
928 native_write_cr4(cr4);
931 static u64 xen_read_msr_safe(unsigned int msr, int *err)
935 if (pmu_msr_read(msr, &val, err))
938 val = native_read_msr_safe(msr, err);
940 case MSR_IA32_APICBASE:
941 val &= ~X2APIC_ENABLE;
947 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
959 case MSR_FS_BASE: which = SEGBASE_FS; goto set;
960 case MSR_KERNEL_GS_BASE: which = SEGBASE_GS_USER; goto set;
961 case MSR_GS_BASE: which = SEGBASE_GS_KERNEL; goto set;
964 base = ((u64)high << 32) | low;
965 if (HYPERVISOR_set_segment_base(which, base) != 0)
973 case MSR_SYSCALL_MASK:
974 case MSR_IA32_SYSENTER_CS:
975 case MSR_IA32_SYSENTER_ESP:
976 case MSR_IA32_SYSENTER_EIP:
977 /* Fast syscall setup is all done in hypercalls, so
978 these are all ignored. Stub them out here to stop
979 Xen console noise. */
983 if (!pmu_msr_write(msr, low, high, &ret))
984 ret = native_write_msr_safe(msr, low, high);
990 static u64 xen_read_msr(unsigned int msr)
993 * This will silently swallow a #GP from RDMSR. It may be worth
998 return xen_read_msr_safe(msr, &err);
1001 static void xen_write_msr(unsigned int msr, unsigned low, unsigned high)
1004 * This will silently swallow a #GP from WRMSR. It may be worth
1007 xen_write_msr_safe(msr, low, high);
1010 /* This is called once we have the cpu_possible_mask */
1011 void __init xen_setup_vcpu_info_placement(void)
1015 for_each_possible_cpu(cpu) {
1016 /* Set up direct vCPU id mapping for PV guests. */
1017 per_cpu(xen_vcpu_id, cpu) = cpu;
1020 * xen_vcpu_setup(cpu) can fail -- in which case it
1021 * falls back to the shared_info version for cpus
1022 * where xen_vcpu_nr(cpu) < MAX_VIRT_CPUS.
1024 * xen_cpu_up_prepare_pv() handles the rest by failing
1027 (void) xen_vcpu_setup(cpu);
1031 * xen_vcpu_setup managed to place the vcpu_info within the
1032 * percpu area for all cpus, so make use of it.
1034 if (xen_have_vcpu_info_placement) {
1035 pv_ops.irq.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1036 pv_ops.irq.restore_fl =
1037 __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1038 pv_ops.irq.irq_disable =
1039 __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1040 pv_ops.irq.irq_enable =
1041 __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1042 pv_ops.mmu.read_cr2 =
1043 __PV_IS_CALLEE_SAVE(xen_read_cr2_direct);
1047 static const struct pv_info xen_info __initconst = {
1048 .shared_kernel_pmd = 0,
1050 #ifdef CONFIG_X86_64
1051 .extra_user_64bit_cs = FLAT_USER_CS64,
1056 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1059 .set_debugreg = xen_set_debugreg,
1060 .get_debugreg = xen_get_debugreg,
1062 .read_cr0 = xen_read_cr0,
1063 .write_cr0 = xen_write_cr0,
1065 .write_cr4 = xen_write_cr4,
1067 .wbinvd = native_wbinvd,
1069 .read_msr = xen_read_msr,
1070 .write_msr = xen_write_msr,
1072 .read_msr_safe = xen_read_msr_safe,
1073 .write_msr_safe = xen_write_msr_safe,
1075 .read_pmc = xen_read_pmc,
1078 #ifdef CONFIG_X86_64
1079 .usergs_sysret64 = xen_sysret64,
1082 .load_tr_desc = paravirt_nop,
1083 .set_ldt = xen_set_ldt,
1084 .load_gdt = xen_load_gdt,
1085 .load_idt = xen_load_idt,
1086 .load_tls = xen_load_tls,
1087 #ifdef CONFIG_X86_64
1088 .load_gs_index = xen_load_gs_index,
1091 .alloc_ldt = xen_alloc_ldt,
1092 .free_ldt = xen_free_ldt,
1094 .store_tr = xen_store_tr,
1096 .write_ldt_entry = xen_write_ldt_entry,
1097 .write_gdt_entry = xen_write_gdt_entry,
1098 .write_idt_entry = xen_write_idt_entry,
1099 .load_sp0 = xen_load_sp0,
1101 #ifdef CONFIG_X86_IOPL_IOPERM
1102 .update_io_bitmap = xen_update_io_bitmap,
1104 .io_delay = xen_io_delay,
1106 /* Xen takes care of %gs when switching to usermode for us */
1107 .swapgs = paravirt_nop,
1109 .start_context_switch = paravirt_start_context_switch,
1110 .end_context_switch = xen_end_context_switch,
1113 static void xen_restart(char *msg)
1115 xen_reboot(SHUTDOWN_reboot);
1118 static void xen_machine_halt(void)
1120 xen_reboot(SHUTDOWN_poweroff);
1123 static void xen_machine_power_off(void)
1127 xen_reboot(SHUTDOWN_poweroff);
1130 static void xen_crash_shutdown(struct pt_regs *regs)
1132 xen_reboot(SHUTDOWN_crash);
1135 static const struct machine_ops xen_machine_ops __initconst = {
1136 .restart = xen_restart,
1137 .halt = xen_machine_halt,
1138 .power_off = xen_machine_power_off,
1139 .shutdown = xen_machine_halt,
1140 .crash_shutdown = xen_crash_shutdown,
1141 .emergency_restart = xen_emergency_restart,
1144 static unsigned char xen_get_nmi_reason(void)
1146 unsigned char reason = 0;
1148 /* Construct a value which looks like it came from port 0x61. */
1149 if (test_bit(_XEN_NMIREASON_io_error,
1150 &HYPERVISOR_shared_info->arch.nmi_reason))
1151 reason |= NMI_REASON_IOCHK;
1152 if (test_bit(_XEN_NMIREASON_pci_serr,
1153 &HYPERVISOR_shared_info->arch.nmi_reason))
1154 reason |= NMI_REASON_SERR;
1159 static void __init xen_boot_params_init_edd(void)
1161 #if IS_ENABLED(CONFIG_EDD)
1162 struct xen_platform_op op;
1163 struct edd_info *edd_info;
1168 edd_info = boot_params.eddbuf;
1169 mbr_signature = boot_params.edd_mbr_sig_buffer;
1171 op.cmd = XENPF_firmware_info;
1173 op.u.firmware_info.type = XEN_FW_DISK_INFO;
1174 for (nr = 0; nr < EDDMAXNR; nr++) {
1175 struct edd_info *info = edd_info + nr;
1177 op.u.firmware_info.index = nr;
1178 info->params.length = sizeof(info->params);
1179 set_xen_guest_handle(op.u.firmware_info.u.disk_info.edd_params,
1181 ret = HYPERVISOR_platform_op(&op);
1185 #define C(x) info->x = op.u.firmware_info.u.disk_info.x
1188 C(interface_support);
1189 C(legacy_max_cylinder);
1191 C(legacy_sectors_per_track);
1194 boot_params.eddbuf_entries = nr;
1196 op.u.firmware_info.type = XEN_FW_DISK_MBR_SIGNATURE;
1197 for (nr = 0; nr < EDD_MBR_SIG_MAX; nr++) {
1198 op.u.firmware_info.index = nr;
1199 ret = HYPERVISOR_platform_op(&op);
1202 mbr_signature[nr] = op.u.firmware_info.u.disk_mbr_signature.mbr_signature;
1204 boot_params.edd_mbr_sig_buf_entries = nr;
1209 * Set up the GDT and segment registers for -fstack-protector. Until
1210 * we do this, we have to be careful not to call any stack-protected
1211 * function, which is most of the kernel.
1213 static void __init xen_setup_gdt(int cpu)
1215 pv_ops.cpu.write_gdt_entry = xen_write_gdt_entry_boot;
1216 pv_ops.cpu.load_gdt = xen_load_gdt_boot;
1218 setup_stack_canary_segment(cpu);
1219 switch_to_new_gdt(cpu);
1221 pv_ops.cpu.write_gdt_entry = xen_write_gdt_entry;
1222 pv_ops.cpu.load_gdt = xen_load_gdt;
1225 static void __init xen_dom0_set_legacy_features(void)
1227 x86_platform.legacy.rtc = 1;
1230 /* First C function to be called on Xen boot */
1231 asmlinkage __visible void __init xen_start_kernel(void)
1233 struct physdev_set_iopl set_iopl;
1234 unsigned long initrd_start = 0;
1237 if (!xen_start_info)
1240 xen_domain_type = XEN_PV_DOMAIN;
1241 xen_start_flags = xen_start_info->flags;
1243 xen_setup_features();
1245 /* Install Xen paravirt ops */
1247 pv_ops.init.patch = paravirt_patch_default;
1248 pv_ops.cpu = xen_cpu_ops;
1252 * Setup xen_vcpu early because it is needed for
1253 * local_irq_disable(), irqs_disabled(), e.g. in printk().
1255 * Don't do the full vcpu_info placement stuff until we have
1256 * the cpu_possible_mask and a non-dummy shared_info.
1258 xen_vcpu_info_reset(0);
1260 x86_platform.get_nmi_reason = xen_get_nmi_reason;
1262 x86_init.resources.memory_setup = xen_memory_setup;
1263 x86_init.irqs.intr_mode_select = x86_init_noop;
1264 x86_init.irqs.intr_mode_init = x86_init_noop;
1265 x86_init.oem.arch_setup = xen_arch_setup;
1266 x86_init.oem.banner = xen_banner;
1267 x86_init.hyper.init_platform = xen_pv_init_platform;
1268 x86_init.hyper.guest_late_init = xen_pv_guest_late_init;
1271 * Set up some pagetable state before starting to set any ptes.
1274 xen_setup_machphys_mapping();
1277 /* Prevent unwanted bits from being set in PTEs. */
1278 __supported_pte_mask &= ~_PAGE_GLOBAL;
1279 __default_kernel_pte_mask &= ~_PAGE_GLOBAL;
1282 * Prevent page tables from being allocated in highmem, even
1283 * if CONFIG_HIGHPTE is enabled.
1285 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1288 xen_build_dynamic_phys_to_machine();
1291 * Set up kernel GDT and segment registers, mainly so that
1292 * -fstack-protector code can be executed.
1296 /* Work out if we support NX */
1297 get_cpu_cap(&boot_cpu_data);
1300 /* Determine virtual and physical address sizes */
1301 get_cpu_address_sizes(&boot_cpu_data);
1303 /* Let's presume PV guests always boot on vCPU with id 0. */
1304 per_cpu(xen_vcpu_id, 0) = 0;
1306 idt_setup_early_handler();
1308 xen_init_capabilities();
1310 #ifdef CONFIG_X86_LOCAL_APIC
1312 * set up the basic apic ops.
1317 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1318 pv_ops.mmu.ptep_modify_prot_start =
1319 xen_ptep_modify_prot_start;
1320 pv_ops.mmu.ptep_modify_prot_commit =
1321 xen_ptep_modify_prot_commit;
1324 machine_ops = xen_machine_ops;
1327 * The only reliable way to retain the initial address of the
1328 * percpu gdt_page is to remember it here, so we can go and
1329 * mark it RW later, when the initial percpu area is freed.
1331 xen_initial_gdt = &per_cpu(gdt_page, 0);
1335 #ifdef CONFIG_ACPI_NUMA
1337 * The pages we from Xen are not related to machine pages, so
1338 * any NUMA information the kernel tries to get from ACPI will
1339 * be meaningless. Prevent it from trying.
1343 WARN_ON(xen_cpuhp_setup(xen_cpu_up_prepare_pv, xen_cpu_dead_pv));
1345 local_irq_disable();
1346 early_boot_irqs_disabled = true;
1348 xen_raw_console_write("mapping kernel into physical memory\n");
1349 xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base,
1350 xen_start_info->nr_pages);
1351 xen_reserve_special_pages();
1353 /* keep using Xen gdt for now; no urgent need to change it */
1355 #ifdef CONFIG_X86_32
1356 pv_info.kernel_rpl = 1;
1357 if (xen_feature(XENFEAT_supervisor_mode_kernel))
1358 pv_info.kernel_rpl = 0;
1360 pv_info.kernel_rpl = 0;
1362 /* set the limit of our address space */
1366 * We used to do this in xen_arch_setup, but that is too late
1367 * on AMD were early_cpu_init (run before ->arch_setup()) calls
1368 * early_amd_init which pokes 0xcf8 port.
1371 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1373 xen_raw_printk("physdev_op failed %d\n", rc);
1375 #ifdef CONFIG_X86_32
1376 /* set up basic CPUID stuff */
1377 cpu_detect(&new_cpu_data);
1378 set_cpu_cap(&new_cpu_data, X86_FEATURE_FPU);
1379 new_cpu_data.x86_capability[CPUID_1_EDX] = cpuid_edx(1);
1382 if (xen_start_info->mod_start) {
1383 if (xen_start_info->flags & SIF_MOD_START_PFN)
1384 initrd_start = PFN_PHYS(xen_start_info->mod_start);
1386 initrd_start = __pa(xen_start_info->mod_start);
1389 /* Poke various useful things into boot_params */
1390 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1391 boot_params.hdr.ramdisk_image = initrd_start;
1392 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1393 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1394 boot_params.hdr.hardware_subarch = X86_SUBARCH_XEN;
1396 if (!xen_initial_domain()) {
1397 add_preferred_console("xenboot", 0, NULL);
1399 x86_init.pci.arch_init = pci_xen_init;
1401 const struct dom0_vga_console_info *info =
1402 (void *)((char *)xen_start_info +
1403 xen_start_info->console.dom0.info_off);
1404 struct xen_platform_op op = {
1405 .cmd = XENPF_firmware_info,
1406 .interface_version = XENPF_INTERFACE_VERSION,
1407 .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
1410 x86_platform.set_legacy_features =
1411 xen_dom0_set_legacy_features;
1412 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1413 xen_start_info->console.domU.mfn = 0;
1414 xen_start_info->console.domU.evtchn = 0;
1416 if (HYPERVISOR_platform_op(&op) == 0)
1417 boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
1419 /* Make sure ACS will be enabled */
1422 xen_acpi_sleep_register();
1424 /* Avoid searching for BIOS MP tables */
1425 x86_init.mpparse.find_smp_config = x86_init_noop;
1426 x86_init.mpparse.get_smp_config = x86_init_uint_noop;
1428 xen_boot_params_init_edd();
1431 if (!boot_params.screen_info.orig_video_isVGA)
1432 add_preferred_console("tty", 0, NULL);
1433 add_preferred_console("hvc", 0, NULL);
1434 if (boot_params.screen_info.orig_video_isVGA)
1435 add_preferred_console("tty", 0, NULL);
1438 /* PCI BIOS service won't work from a PV guest. */
1439 pci_probe &= ~PCI_PROBE_BIOS;
1441 xen_raw_console_write("about to get started...\n");
1443 /* We need this for printk timestamps */
1444 xen_setup_runstate_info(0);
1446 xen_efi_init(&boot_params);
1448 /* Start the world */
1449 #ifdef CONFIG_X86_32
1450 i386_start_kernel();
1452 cr4_init_shadow(); /* 32b kernel does this in i386_start_kernel() */
1453 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1457 static int xen_cpu_up_prepare_pv(unsigned int cpu)
1461 if (per_cpu(xen_vcpu, cpu) == NULL)
1464 xen_setup_timer(cpu);
1466 rc = xen_smp_intr_init(cpu);
1468 WARN(1, "xen_smp_intr_init() for CPU %d failed: %d\n",
1473 rc = xen_smp_intr_init_pv(cpu);
1475 WARN(1, "xen_smp_intr_init_pv() for CPU %d failed: %d\n",
1483 static int xen_cpu_dead_pv(unsigned int cpu)
1485 xen_smp_intr_free(cpu);
1486 xen_smp_intr_free_pv(cpu);
1488 xen_teardown_timer(cpu);
1493 static uint32_t __init xen_platform_pv(void)
1495 if (xen_pv_domain())
1496 return xen_cpuid_base();
1501 const __initconst struct hypervisor_x86 x86_hyper_xen_pv = {
1503 .detect = xen_platform_pv,
1504 .type = X86_HYPER_XEN_PV,
1505 .runtime.pin_vcpu = xen_pin_vcpu,
1506 .ignore_nopv = true,