x86/xen: split off enlighten_pv.c
authorVitaly Kuznetsov <vkuznets@redhat.com>
Tue, 14 Mar 2017 17:35:41 +0000 (18:35 +0100)
committerJuergen Gross <jgross@suse.com>
Tue, 2 May 2017 09:04:05 +0000 (11:04 +0200)
Basically, enlighten.c is renamed to enlighten_pv.c and some code moved
out to common enlighten.c.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: Juergen Gross <jgross@suse.com>
Signed-off-by: Juergen Gross <jgross@suse.com>
arch/x86/xen/Makefile
arch/x86/xen/enlighten.c
arch/x86/xen/enlighten_pv.c [new file with mode: 0644]

index 1bca75b..5ca8d3e 100644 (file)
@@ -7,13 +7,13 @@ endif
 
 # Make sure early boot has no stackprotector
 nostackp := $(call cc-option, -fno-stack-protector)
-CFLAGS_enlighten.o             := $(nostackp)
+CFLAGS_enlighten_pv.o          := $(nostackp)
 CFLAGS_mmu.o                   := $(nostackp)
 
 obj-y          := enlighten.o setup.o multicalls.o mmu.o irq.o \
                        time.o xen-asm.o xen-asm_$(BITS).o \
                        grant-table.o suspend.o platform-pci-unplug.o \
-                       p2m.o apic.o pmu.o
+                       p2m.o apic.o pmu.o enlighten_pv.o
 
 obj-$(CONFIG_XEN_PVHVM)                += enlighten_hvm.o
 obj-$(CONFIG_XEN_PVH)                  += enlighten_pvh.o
index 5cd454c..41d324c 100644 (file)
@@ -1,93 +1,16 @@
-/*
- * Core of Xen paravirt_ops implementation.
- *
- * This file contains the xen_paravirt_ops structure itself, and the
- * implementations for:
- * - privileged instructions
- * - interrupt flags
- * - segment operations
- * - booting and setup
- *
- * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
- */
-
 #include <linux/cpu.h>
-#include <linux/kernel.h>
-#include <linux/init.h>
-#include <linux/smp.h>
-#include <linux/preempt.h>
-#include <linux/hardirq.h>
-#include <linux/percpu.h>
-#include <linux/delay.h>
-#include <linux/start_kernel.h>
-#include <linux/sched.h>
-#include <linux/kprobes.h>
-#include <linux/bootmem.h>
-#include <linux/export.h>
-#include <linux/mm.h>
-#include <linux/page-flags.h>
-#include <linux/highmem.h>
-#include <linux/console.h>
-#include <linux/pci.h>
-#include <linux/gfp.h>
-#include <linux/memblock.h>
-#include <linux/edd.h>
-#include <linux/frame.h>
-
 #include <linux/kexec.h>
 
-#include <xen/xen.h>
-#include <xen/events.h>
-#include <xen/interface/xen.h>
-#include <xen/interface/version.h>
-#include <xen/interface/physdev.h>
-#include <xen/interface/vcpu.h>
-#include <xen/interface/memory.h>
-#include <xen/interface/nmi.h>
-#include <xen/interface/xen-mca.h>
 #include <xen/features.h>
 #include <xen/page.h>
-#include <xen/hvc-console.h>
-#include <xen/acpi.h>
 
-#include <asm/paravirt.h>
-#include <asm/apic.h>
-#include <asm/page.h>
-#include <asm/xen/pci.h>
 #include <asm/xen/hypercall.h>
 #include <asm/xen/hypervisor.h>
-#include <asm/xen/cpuid.h>
-#include <asm/fixmap.h>
-#include <asm/processor.h>
-#include <asm/proto.h>
-#include <asm/msr-index.h>
-#include <asm/traps.h>
-#include <asm/setup.h>
-#include <asm/desc.h>
-#include <asm/pgalloc.h>
-#include <asm/pgtable.h>
-#include <asm/tlbflush.h>
-#include <asm/reboot.h>
-#include <asm/stackprotector.h>
-#include <asm/hypervisor.h>
-#include <asm/mach_traps.h>
-#include <asm/mwait.h>
-#include <asm/pci_x86.h>
 #include <asm/cpu.h>
 #include <asm/e820/api.h> 
 
-#ifdef CONFIG_ACPI
-#include <linux/acpi.h>
-#include <asm/acpi.h>
-#include <acpi/pdc_intel.h>
-#include <acpi/processor.h>
-#include <xen/interface/platform.h>
-#endif
-
 #include "xen-ops.h"
-#include "mmu.h"
 #include "smp.h"
-#include "multicalls.h"
 #include "pmu.h"
 
 EXPORT_SYMBOL_GPL(hypercall_page);
@@ -134,14 +57,6 @@ EXPORT_SYMBOL_GPL(xen_start_info);
 
 struct shared_info xen_dummy_shared_info;
 
-void *xen_initial_gdt;
-
-RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
-
-static int xen_cpu_up_prepare_pv(unsigned int cpu);
-static int xen_cpu_up_online(unsigned int cpu);
-static int xen_cpu_dead_pv(unsigned int cpu);
-
 /*
  * Point at some empty memory to start with. We map the real shared_info
  * page as soon as fixmap is up and running.
@@ -163,18 +78,30 @@ struct shared_info *HYPERVISOR_shared_info = &xen_dummy_shared_info;
  */
 int xen_have_vcpu_info_placement = 1;
 
-struct tls_descs {
-       struct desc_struct desc[3];
-};
+static int xen_cpu_up_online(unsigned int cpu)
+{
+       xen_init_lock_cpu(cpu);
+       return 0;
+}
 
-/*
- * Updating the 3 TLS descriptors in the GDT on every task switch is
- * surprisingly expensive so we avoid updating them if they haven't
- * changed.  Since Xen writes different descriptors than the one
- * passed in the update_descriptor hypercall we keep shadow copies to
- * compare against.
- */
-static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
+int xen_cpuhp_setup(int (*cpu_up_prepare_cb)(unsigned int),
+                   int (*cpu_dead_cb)(unsigned int))
+{
+       int rc;
+
+       rc = cpuhp_setup_state_nocalls(CPUHP_XEN_PREPARE,
+                                      "x86/xen/hvm_guest:prepare",
+                                      cpu_up_prepare_cb, cpu_dead_cb);
+       if (rc >= 0) {
+               rc = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
+                                              "x86/xen/hvm_guest:online",
+                                              xen_cpu_up_online, NULL);
+               if (rc < 0)
+                       cpuhp_remove_state_nocalls(CPUHP_XEN_PREPARE);
+       }
+
+       return rc >= 0 ? 0 : rc;
+}
 
 static void clamp_max_cpus(void)
 {
@@ -243,1510 +170,81 @@ void xen_vcpu_setup(int cpu)
        }
 }
 
-/*
- * On restore, set the vcpu placement up again.
- * If it fails, then we're in a bad state, since
- * we can't back out from using it...
- */
-void xen_vcpu_restore(void)
+void xen_reboot(int reason)
 {
+       struct sched_shutdown r = { .reason = reason };
        int cpu;
 
-       for_each_possible_cpu(cpu) {
-               bool other_cpu = (cpu != smp_processor_id());
-               bool is_up = HYPERVISOR_vcpu_op(VCPUOP_is_up, xen_vcpu_nr(cpu),
-                                               NULL);
-
-               if (other_cpu && is_up &&
-                   HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL))
-                       BUG();
-
-               xen_setup_runstate_info(cpu);
-
-               if (xen_have_vcpu_info_placement)
-                       xen_vcpu_setup(cpu);
-
-               if (other_cpu && is_up &&
-                   HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL))
-                       BUG();
-       }
-}
-
-static void __init xen_banner(void)
-{
-       unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
-       struct xen_extraversion extra;
-       HYPERVISOR_xen_version(XENVER_extraversion, &extra);
-
-       pr_info("Booting paravirtualized kernel %son %s\n",
-               xen_feature(XENFEAT_auto_translated_physmap) ?
-                       "with PVH extensions " : "", pv_info.name);
-       printk(KERN_INFO "Xen version: %d.%d%s%s\n",
-              version >> 16, version & 0xffff, extra.extraversion,
-              xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
-}
-/* Check if running on Xen version (major, minor) or later */
-bool
-xen_running_on_version_or_later(unsigned int major, unsigned int minor)
-{
-       unsigned int version;
-
-       if (!xen_domain())
-               return false;
-
-       version = HYPERVISOR_xen_version(XENVER_version, NULL);
-       if ((((version >> 16) == major) && ((version & 0xffff) >= minor)) ||
-               ((version >> 16) > major))
-               return true;
-       return false;
-}
-
-#define CPUID_THERM_POWER_LEAF 6
-#define APERFMPERF_PRESENT 0
-
-static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
-static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
-
-static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
-static __read_mostly unsigned int cpuid_leaf5_ecx_val;
-static __read_mostly unsigned int cpuid_leaf5_edx_val;
-
-static void xen_cpuid(unsigned int *ax, unsigned int *bx,
-                     unsigned int *cx, unsigned int *dx)
-{
-       unsigned maskebx = ~0;
-       unsigned maskecx = ~0;
-       unsigned maskedx = ~0;
-       unsigned setecx = 0;
-       /*
-        * Mask out inconvenient features, to try and disable as many
-        * unsupported kernel subsystems as possible.
-        */
-       switch (*ax) {
-       case 1:
-               maskecx = cpuid_leaf1_ecx_mask;
-               setecx = cpuid_leaf1_ecx_set_mask;
-               maskedx = cpuid_leaf1_edx_mask;
-               break;
-
-       case CPUID_MWAIT_LEAF:
-               /* Synthesize the values.. */
-               *ax = 0;
-               *bx = 0;
-               *cx = cpuid_leaf5_ecx_val;
-               *dx = cpuid_leaf5_edx_val;
-               return;
-
-       case CPUID_THERM_POWER_LEAF:
-               /* Disabling APERFMPERF for kernel usage */
-               maskecx = ~(1 << APERFMPERF_PRESENT);
-               break;
-
-       case 0xb:
-               /* Suppress extended topology stuff */
-               maskebx = 0;
-               break;
-       }
-
-       asm(XEN_EMULATE_PREFIX "cpuid"
-               : "=a" (*ax),
-                 "=b" (*bx),
-                 "=c" (*cx),
-                 "=d" (*dx)
-               : "0" (*ax), "2" (*cx));
-
-       *bx &= maskebx;
-       *cx &= maskecx;
-       *cx |= setecx;
-       *dx &= maskedx;
-}
-STACK_FRAME_NON_STANDARD(xen_cpuid); /* XEN_EMULATE_PREFIX */
-
-static bool __init xen_check_mwait(void)
-{
-#ifdef CONFIG_ACPI
-       struct xen_platform_op op = {
-               .cmd                    = XENPF_set_processor_pminfo,
-               .u.set_pminfo.id        = -1,
-               .u.set_pminfo.type      = XEN_PM_PDC,
-       };
-       uint32_t buf[3];
-       unsigned int ax, bx, cx, dx;
-       unsigned int mwait_mask;
-
-       /* We need to determine whether it is OK to expose the MWAIT
-        * capability to the kernel to harvest deeper than C3 states from ACPI
-        * _CST using the processor_harvest_xen.c module. For this to work, we
-        * need to gather the MWAIT_LEAF values (which the cstate.c code
-        * checks against). The hypervisor won't expose the MWAIT flag because
-        * it would break backwards compatibility; so we will find out directly
-        * from the hardware and hypercall.
-        */
-       if (!xen_initial_domain())
-               return false;
-
-       /*
-        * When running under platform earlier than Xen4.2, do not expose
-        * mwait, to avoid the risk of loading native acpi pad driver
-        */
-       if (!xen_running_on_version_or_later(4, 2))
-               return false;
-
-       ax = 1;
-       cx = 0;
-
-       native_cpuid(&ax, &bx, &cx, &dx);
-
-       mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
-                    (1 << (X86_FEATURE_MWAIT % 32));
-
-       if ((cx & mwait_mask) != mwait_mask)
-               return false;
-
-       /* We need to emulate the MWAIT_LEAF and for that we need both
-        * ecx and edx. The hypercall provides only partial information.
-        */
-
-       ax = CPUID_MWAIT_LEAF;
-       bx = 0;
-       cx = 0;
-       dx = 0;
-
-       native_cpuid(&ax, &bx, &cx, &dx);
-
-       /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
-        * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
-        */
-       buf[0] = ACPI_PDC_REVISION_ID;
-       buf[1] = 1;
-       buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
-
-       set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
-
-       if ((HYPERVISOR_platform_op(&op) == 0) &&
-           (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
-               cpuid_leaf5_ecx_val = cx;
-               cpuid_leaf5_edx_val = dx;
-       }
-       return true;
-#else
-       return false;
-#endif
-}
-static void __init xen_init_cpuid_mask(void)
-{
-       unsigned int ax, bx, cx, dx;
-       unsigned int xsave_mask;
-
-       cpuid_leaf1_edx_mask =
-               ~((1 << X86_FEATURE_MTRR) |  /* disable MTRR */
-                 (1 << X86_FEATURE_ACC));   /* thermal monitoring */
-
-       if (!xen_initial_domain())
-               cpuid_leaf1_edx_mask &=
-                       ~((1 << X86_FEATURE_ACPI));  /* disable ACPI */
-
-       cpuid_leaf1_ecx_mask &= ~(1 << (X86_FEATURE_X2APIC % 32));
-
-       ax = 1;
-       cx = 0;
-       cpuid(1, &ax, &bx, &cx, &dx);
-
-       xsave_mask =
-               (1 << (X86_FEATURE_XSAVE % 32)) |
-               (1 << (X86_FEATURE_OSXSAVE % 32));
-
-       /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
-       if ((cx & xsave_mask) != xsave_mask)
-               cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
-       if (xen_check_mwait())
-               cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
-}
-
-static void xen_set_debugreg(int reg, unsigned long val)
-{
-       HYPERVISOR_set_debugreg(reg, val);
-}
-
-static unsigned long xen_get_debugreg(int reg)
-{
-       return HYPERVISOR_get_debugreg(reg);
-}
-
-static void xen_end_context_switch(struct task_struct *next)
-{
-       xen_mc_flush();
-       paravirt_end_context_switch(next);
-}
-
-static unsigned long xen_store_tr(void)
-{
-       return 0;
-}
-
-/*
- * Set the page permissions for a particular virtual address.  If the
- * address is a vmalloc mapping (or other non-linear mapping), then
- * find the linear mapping of the page and also set its protections to
- * match.
- */
-static void set_aliased_prot(void *v, pgprot_t prot)
-{
-       int level;
-       pte_t *ptep;
-       pte_t pte;
-       unsigned long pfn;
-       struct page *page;
-       unsigned char dummy;
-
-       ptep = lookup_address((unsigned long)v, &level);
-       BUG_ON(ptep == NULL);
-
-       pfn = pte_pfn(*ptep);
-       page = pfn_to_page(pfn);
-
-       pte = pfn_pte(pfn, prot);
-
-       /*
-        * Careful: update_va_mapping() will fail if the virtual address
-        * we're poking isn't populated in the page tables.  We don't
-        * need to worry about the direct map (that's always in the page
-        * tables), but we need to be careful about vmap space.  In
-        * particular, the top level page table can lazily propagate
-        * entries between processes, so if we've switched mms since we
-        * vmapped the target in the first place, we might not have the
-        * top-level page table entry populated.
-        *
-        * We disable preemption because we want the same mm active when
-        * we probe the target and when we issue the hypercall.  We'll
-        * have the same nominal mm, but if we're a kernel thread, lazy
-        * mm dropping could change our pgd.
-        *
-        * Out of an abundance of caution, this uses __get_user() to fault
-        * in the target address just in case there's some obscure case
-        * in which the target address isn't readable.
-        */
-
-       preempt_disable();
-
-       probe_kernel_read(&dummy, v, 1);
+       for_each_online_cpu(cpu)
+               xen_pmu_finish(cpu);
 
-       if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
+       if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
                BUG();
-
-       if (!PageHighMem(page)) {
-               void *av = __va(PFN_PHYS(pfn));
-
-               if (av != v)
-                       if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
-                               BUG();
-       } else
-               kmap_flush_unused();
-
-       preempt_enable();
-}
-
-static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
-{
-       const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
-       int i;
-
-       /*
-        * We need to mark the all aliases of the LDT pages RO.  We
-        * don't need to call vm_flush_aliases(), though, since that's
-        * only responsible for flushing aliases out the TLBs, not the
-        * page tables, and Xen will flush the TLB for us if needed.
-        *
-        * To avoid confusing future readers: none of this is necessary
-        * to load the LDT.  The hypervisor only checks this when the
-        * LDT is faulted in due to subsequent descriptor access.
-        */
-
-       for(i = 0; i < entries; i += entries_per_page)
-               set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
 }
 
-static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
-{
-       const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
-       int i;
-
-       for(i = 0; i < entries; i += entries_per_page)
-               set_aliased_prot(ldt + i, PAGE_KERNEL);
-}
-
-static void xen_set_ldt(const void *addr, unsigned entries)
+void xen_emergency_restart(void)
 {
-       struct mmuext_op *op;
-       struct multicall_space mcs = xen_mc_entry(sizeof(*op));
-
-       trace_xen_cpu_set_ldt(addr, entries);
-
-       op = mcs.args;
-       op->cmd = MMUEXT_SET_LDT;
-       op->arg1.linear_addr = (unsigned long)addr;
-       op->arg2.nr_ents = entries;
-
-       MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
-
-       xen_mc_issue(PARAVIRT_LAZY_CPU);
+       xen_reboot(SHUTDOWN_reboot);
 }
 
-static void xen_load_gdt(const struct desc_ptr *dtr)
+static int
+xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
 {
-       unsigned long va = dtr->address;
-       unsigned int size = dtr->size + 1;
-       unsigned pages = DIV_ROUND_UP(size, PAGE_SIZE);
-       unsigned long frames[pages];
-       int f;
-
-       /*
-        * A GDT can be up to 64k in size, which corresponds to 8192
-        * 8-byte entries, or 16 4k pages..
-        */
-
-       BUG_ON(size > 65536);
-       BUG_ON(va & ~PAGE_MASK);
-
-       for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
-               int level;
-               pte_t *ptep;
-               unsigned long pfn, mfn;
-               void *virt;
-
-               /*
-                * The GDT is per-cpu and is in the percpu data area.
-                * That can be virtually mapped, so we need to do a
-                * page-walk to get the underlying MFN for the
-                * hypercall.  The page can also be in the kernel's
-                * linear range, so we need to RO that mapping too.
-                */
-               ptep = lookup_address(va, &level);
-               BUG_ON(ptep == NULL);
-
-               pfn = pte_pfn(*ptep);
-               mfn = pfn_to_mfn(pfn);
-               virt = __va(PFN_PHYS(pfn));
-
-               frames[f] = mfn;
-
-               make_lowmem_page_readonly((void *)va);
-               make_lowmem_page_readonly(virt);
-       }
-
-       if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
-               BUG();
+       if (!kexec_crash_loaded())
+               xen_reboot(SHUTDOWN_crash);
+       return NOTIFY_DONE;
 }
 
-/*
- * load_gdt for early boot, when the gdt is only mapped once
- */
-static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
-{
-       unsigned long va = dtr->address;
-       unsigned int size = dtr->size + 1;
-       unsigned pages = DIV_ROUND_UP(size, PAGE_SIZE);
-       unsigned long frames[pages];
-       int f;
-
-       /*
-        * A GDT can be up to 64k in size, which corresponds to 8192
-        * 8-byte entries, or 16 4k pages..
-        */
-
-       BUG_ON(size > 65536);
-       BUG_ON(va & ~PAGE_MASK);
-
-       for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
-               pte_t pte;
-               unsigned long pfn, mfn;
-
-               pfn = virt_to_pfn(va);
-               mfn = pfn_to_mfn(pfn);
-
-               pte = pfn_pte(pfn, PAGE_KERNEL_RO);
-
-               if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
-                       BUG();
-
-               frames[f] = mfn;
-       }
-
-       if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
-               BUG();
-}
+static struct notifier_block xen_panic_block = {
+       .notifier_call = xen_panic_event,
+       .priority = INT_MIN
+};
 
-static inline bool desc_equal(const struct desc_struct *d1,
-                             const struct desc_struct *d2)
+int xen_panic_handler_init(void)
 {
-       return d1->a == d2->a && d1->b == d2->b;
+       atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
+       return 0;
 }
 
-static void load_TLS_descriptor(struct thread_struct *t,
-                               unsigned int cpu, unsigned int i)
+void xen_pin_vcpu(int cpu)
 {
-       struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
-       struct desc_struct *gdt;
-       xmaddr_t maddr;
-       struct multicall_space mc;
+       static bool disable_pinning;
+       struct sched_pin_override pin_override;
+       int ret;
 
-       if (desc_equal(shadow, &t->tls_array[i]))
+       if (disable_pinning)
                return;
 
-       *shadow = t->tls_array[i];
-
-       gdt = get_cpu_gdt_rw(cpu);
-       maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
-       mc = __xen_mc_entry(0);
+       pin_override.pcpu = cpu;
+       ret = HYPERVISOR_sched_op(SCHEDOP_pin_override, &pin_override);
 
-       MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
-}
+       /* Ignore errors when removing override. */
+       if (cpu < 0)
+               return;
 
-static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
-{
-       /*
-        * XXX sleazy hack: If we're being called in a lazy-cpu zone
-        * and lazy gs handling is enabled, it means we're in a
-        * context switch, and %gs has just been saved.  This means we
-        * can zero it out to prevent faults on exit from the
-        * hypervisor if the next process has no %gs.  Either way, it
-        * has been saved, and the new value will get loaded properly.
-        * This will go away as soon as Xen has been modified to not
-        * save/restore %gs for normal hypercalls.
-        *
-        * On x86_64, this hack is not used for %gs, because gs points
-        * to KERNEL_GS_BASE (and uses it for PDA references), so we
-        * must not zero %gs on x86_64
-        *
-        * For x86_64, we need to zero %fs, otherwise we may get an
-        * exception between the new %fs descriptor being loaded and
-        * %fs being effectively cleared at __switch_to().
-        */
-       if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
-#ifdef CONFIG_X86_32
-               lazy_load_gs(0);
-#else
-               loadsegment(fs, 0);
-#endif
+       switch (ret) {
+       case -ENOSYS:
+               pr_warn("Unable to pin on physical cpu %d. In case of problems consider vcpu pinning.\n",
+                       cpu);
+               disable_pinning = true;
+               break;
+       case -EPERM:
+               WARN(1, "Trying to pin vcpu without having privilege to do so\n");
+               disable_pinning = true;
+               break;
+       case -EINVAL:
+       case -EBUSY:
+               pr_warn("Physical cpu %d not available for pinning. Check Xen cpu configuration.\n",
+                       cpu);
+               break;
+       case 0:
+               break;
+       default:
+               WARN(1, "rc %d while trying to pin vcpu\n", ret);
+               disable_pinning = true;
        }
-
-       xen_mc_batch();
-
-       load_TLS_descriptor(t, cpu, 0);
-       load_TLS_descriptor(t, cpu, 1);
-       load_TLS_descriptor(t, cpu, 2);
-
-       xen_mc_issue(PARAVIRT_LAZY_CPU);
-}
-
-#ifdef CONFIG_X86_64
-static void xen_load_gs_index(unsigned int idx)
-{
-       if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
-               BUG();
-}
-#endif
-
-static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
-                               const void *ptr)
-{
-       xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
-       u64 entry = *(u64 *)ptr;
-
-       trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
-
-       preempt_disable();
-
-       xen_mc_flush();
-       if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
-               BUG();
-
-       preempt_enable();
 }
 
-static int cvt_gate_to_trap(int vector, const gate_desc *val,
-                           struct trap_info *info)
-{
-       unsigned long addr;
-
-       if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
-               return 0;
-
-       info->vector = vector;
-
-       addr = gate_offset(*val);
-#ifdef CONFIG_X86_64
-       /*
-        * Look for known traps using IST, and substitute them
-        * appropriately.  The debugger ones are the only ones we care
-        * about.  Xen will handle faults like double_fault,
-        * so we should never see them.  Warn if
-        * there's an unexpected IST-using fault handler.
-        */
-       if (addr == (unsigned long)debug)
-               addr = (unsigned long)xen_debug;
-       else if (addr == (unsigned long)int3)
-               addr = (unsigned long)xen_int3;
-       else if (addr == (unsigned long)stack_segment)
-               addr = (unsigned long)xen_stack_segment;
-       else if (addr == (unsigned long)double_fault) {
-               /* Don't need to handle these */
-               return 0;
-#ifdef CONFIG_X86_MCE
-       } else if (addr == (unsigned long)machine_check) {
-               /*
-                * when xen hypervisor inject vMCE to guest,
-                * use native mce handler to handle it
-                */
-               ;
-#endif
-       } else if (addr == (unsigned long)nmi)
-               /*
-                * Use the native version as well.
-                */
-               ;
-       else {
-               /* Some other trap using IST? */
-               if (WARN_ON(val->ist != 0))
-                       return 0;
-       }
-#endif /* CONFIG_X86_64 */
-       info->address = addr;
-
-       info->cs = gate_segment(*val);
-       info->flags = val->dpl;
-       /* interrupt gates clear IF */
-       if (val->type == GATE_INTERRUPT)
-               info->flags |= 1 << 2;
-
-       return 1;
-}
-
-/* Locations of each CPU's IDT */
-static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
-
-/* Set an IDT entry.  If the entry is part of the current IDT, then
-   also update Xen. */
-static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
-{
-       unsigned long p = (unsigned long)&dt[entrynum];
-       unsigned long start, end;
-
-       trace_xen_cpu_write_idt_entry(dt, entrynum, g);
-
-       preempt_disable();
-
-       start = __this_cpu_read(idt_desc.address);
-       end = start + __this_cpu_read(idt_desc.size) + 1;
-
-       xen_mc_flush();
-
-       native_write_idt_entry(dt, entrynum, g);
-
-       if (p >= start && (p + 8) <= end) {
-               struct trap_info info[2];
-
-               info[1].address = 0;
-
-               if (cvt_gate_to_trap(entrynum, g, &info[0]))
-                       if (HYPERVISOR_set_trap_table(info))
-                               BUG();
-       }
-
-       preempt_enable();
-}
-
-static void xen_convert_trap_info(const struct desc_ptr *desc,
-                                 struct trap_info *traps)
-{
-       unsigned in, out, count;
-
-       count = (desc->size+1) / sizeof(gate_desc);
-       BUG_ON(count > 256);
-
-       for (in = out = 0; in < count; in++) {
-               gate_desc *entry = (gate_desc*)(desc->address) + in;
-
-               if (cvt_gate_to_trap(in, entry, &traps[out]))
-                       out++;
-       }
-       traps[out].address = 0;
-}
-
-void xen_copy_trap_info(struct trap_info *traps)
-{
-       const struct desc_ptr *desc = this_cpu_ptr(&idt_desc);
-
-       xen_convert_trap_info(desc, traps);
-}
-
-/* Load a new IDT into Xen.  In principle this can be per-CPU, so we
-   hold a spinlock to protect the static traps[] array (static because
-   it avoids allocation, and saves stack space). */
-static void xen_load_idt(const struct desc_ptr *desc)
-{
-       static DEFINE_SPINLOCK(lock);
-       static struct trap_info traps[257];
-
-       trace_xen_cpu_load_idt(desc);
-
-       spin_lock(&lock);
-
-       memcpy(this_cpu_ptr(&idt_desc), desc, sizeof(idt_desc));
-
-       xen_convert_trap_info(desc, traps);
-
-       xen_mc_flush();
-       if (HYPERVISOR_set_trap_table(traps))
-               BUG();
-
-       spin_unlock(&lock);
-}
-
-/* Write a GDT descriptor entry.  Ignore LDT descriptors, since
-   they're handled differently. */
-static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
-                               const void *desc, int type)
-{
-       trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
-
-       preempt_disable();
-
-       switch (type) {
-       case DESC_LDT:
-       case DESC_TSS:
-               /* ignore */
-               break;
-
-       default: {
-               xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
-
-               xen_mc_flush();
-               if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
-                       BUG();
-       }
-
-       }
-
-       preempt_enable();
-}
-
-/*
- * Version of write_gdt_entry for use at early boot-time needed to
- * update an entry as simply as possible.
- */
-static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
-                                           const void *desc, int type)
-{
-       trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
-
-       switch (type) {
-       case DESC_LDT:
-       case DESC_TSS:
-               /* ignore */
-               break;
-
-       default: {
-               xmaddr_t maddr = virt_to_machine(&dt[entry]);
-
-               if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
-                       dt[entry] = *(struct desc_struct *)desc;
-       }
-
-       }
-}
-
-static void xen_load_sp0(struct tss_struct *tss,
-                        struct thread_struct *thread)
-{
-       struct multicall_space mcs;
-
-       mcs = xen_mc_entry(0);
-       MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
-       xen_mc_issue(PARAVIRT_LAZY_CPU);
-       tss->x86_tss.sp0 = thread->sp0;
-}
-
-void xen_set_iopl_mask(unsigned mask)
-{
-       struct physdev_set_iopl set_iopl;
-
-       /* Force the change at ring 0. */
-       set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
-       HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
-}
-
-static void xen_io_delay(void)
-{
-}
-
-static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
-
-static unsigned long xen_read_cr0(void)
-{
-       unsigned long cr0 = this_cpu_read(xen_cr0_value);
-
-       if (unlikely(cr0 == 0)) {
-               cr0 = native_read_cr0();
-               this_cpu_write(xen_cr0_value, cr0);
-       }
-
-       return cr0;
-}
-
-static void xen_write_cr0(unsigned long cr0)
-{
-       struct multicall_space mcs;
-
-       this_cpu_write(xen_cr0_value, cr0);
-
-       /* Only pay attention to cr0.TS; everything else is
-          ignored. */
-       mcs = xen_mc_entry(0);
-
-       MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
-
-       xen_mc_issue(PARAVIRT_LAZY_CPU);
-}
-
-static void xen_write_cr4(unsigned long cr4)
-{
-       cr4 &= ~(X86_CR4_PGE | X86_CR4_PSE | X86_CR4_PCE);
-
-       native_write_cr4(cr4);
-}
-#ifdef CONFIG_X86_64
-static inline unsigned long xen_read_cr8(void)
-{
-       return 0;
-}
-static inline void xen_write_cr8(unsigned long val)
-{
-       BUG_ON(val);
-}
-#endif
-
-static u64 xen_read_msr_safe(unsigned int msr, int *err)
-{
-       u64 val;
-
-       if (pmu_msr_read(msr, &val, err))
-               return val;
-
-       val = native_read_msr_safe(msr, err);
-       switch (msr) {
-       case MSR_IA32_APICBASE:
-#ifdef CONFIG_X86_X2APIC
-               if (!(cpuid_ecx(1) & (1 << (X86_FEATURE_X2APIC & 31))))
-#endif
-                       val &= ~X2APIC_ENABLE;
-               break;
-       }
-       return val;
-}
-
-static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
-{
-       int ret;
-
-       ret = 0;
-
-       switch (msr) {
-#ifdef CONFIG_X86_64
-               unsigned which;
-               u64 base;
-
-       case MSR_FS_BASE:               which = SEGBASE_FS; goto set;
-       case MSR_KERNEL_GS_BASE:        which = SEGBASE_GS_USER; goto set;
-       case MSR_GS_BASE:               which = SEGBASE_GS_KERNEL; goto set;
-
-       set:
-               base = ((u64)high << 32) | low;
-               if (HYPERVISOR_set_segment_base(which, base) != 0)
-                       ret = -EIO;
-               break;
-#endif
-
-       case MSR_STAR:
-       case MSR_CSTAR:
-       case MSR_LSTAR:
-       case MSR_SYSCALL_MASK:
-       case MSR_IA32_SYSENTER_CS:
-       case MSR_IA32_SYSENTER_ESP:
-       case MSR_IA32_SYSENTER_EIP:
-               /* Fast syscall setup is all done in hypercalls, so
-                  these are all ignored.  Stub them out here to stop
-                  Xen console noise. */
-               break;
-
-       default:
-               if (!pmu_msr_write(msr, low, high, &ret))
-                       ret = native_write_msr_safe(msr, low, high);
-       }
-
-       return ret;
-}
-
-static u64 xen_read_msr(unsigned int msr)
-{
-       /*
-        * This will silently swallow a #GP from RDMSR.  It may be worth
-        * changing that.
-        */
-       int err;
-
-       return xen_read_msr_safe(msr, &err);
-}
-
-static void xen_write_msr(unsigned int msr, unsigned low, unsigned high)
-{
-       /*
-        * This will silently swallow a #GP from WRMSR.  It may be worth
-        * changing that.
-        */
-       xen_write_msr_safe(msr, low, high);
-}
-
-void xen_setup_shared_info(void)
-{
-       if (!xen_feature(XENFEAT_auto_translated_physmap)) {
-               set_fixmap(FIX_PARAVIRT_BOOTMAP,
-                          xen_start_info->shared_info);
-
-               HYPERVISOR_shared_info =
-                       (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
-       } else
-               HYPERVISOR_shared_info =
-                       (struct shared_info *)__va(xen_start_info->shared_info);
-
-#ifndef CONFIG_SMP
-       /* In UP this is as good a place as any to set up shared info */
-       xen_setup_vcpu_info_placement();
-#endif
-
-       xen_setup_mfn_list_list();
-}
-
-/* This is called once we have the cpu_possible_mask */
-void xen_setup_vcpu_info_placement(void)
-{
-       int cpu;
-
-       for_each_possible_cpu(cpu) {
-               /* Set up direct vCPU id mapping for PV guests. */
-               per_cpu(xen_vcpu_id, cpu) = cpu;
-               xen_vcpu_setup(cpu);
-       }
-
-       /*
-        * xen_vcpu_setup managed to place the vcpu_info within the
-        * percpu area for all cpus, so make use of it.
-        */
-       if (xen_have_vcpu_info_placement) {
-               pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
-               pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
-               pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
-               pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
-               pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
-       }
-}
-
-static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
-                         unsigned long addr, unsigned len)
-{
-       char *start, *end, *reloc;
-       unsigned ret;
-
-       start = end = reloc = NULL;
-
-#define SITE(op, x)                                                    \
-       case PARAVIRT_PATCH(op.x):                                      \
-       if (xen_have_vcpu_info_placement) {                             \
-               start = (char *)xen_##x##_direct;                       \
-               end = xen_##x##_direct_end;                             \
-               reloc = xen_##x##_direct_reloc;                         \
-       }                                                               \
-       goto patch_site
-
-       switch (type) {
-               SITE(pv_irq_ops, irq_enable);
-               SITE(pv_irq_ops, irq_disable);
-               SITE(pv_irq_ops, save_fl);
-               SITE(pv_irq_ops, restore_fl);
-#undef SITE
-
-       patch_site:
-               if (start == NULL || (end-start) > len)
-                       goto default_patch;
-
-               ret = paravirt_patch_insns(insnbuf, len, start, end);
-
-               /* Note: because reloc is assigned from something that
-                  appears to be an array, gcc assumes it's non-null,
-                  but doesn't know its relationship with start and
-                  end. */
-               if (reloc > start && reloc < end) {
-                       int reloc_off = reloc - start;
-                       long *relocp = (long *)(insnbuf + reloc_off);
-                       long delta = start - (char *)addr;
-
-                       *relocp += delta;
-               }
-               break;
-
-       default_patch:
-       default:
-               ret = paravirt_patch_default(type, clobbers, insnbuf,
-                                            addr, len);
-               break;
-       }
-
-       return ret;
-}
-
-static const struct pv_info xen_info __initconst = {
-       .shared_kernel_pmd = 0,
-
-#ifdef CONFIG_X86_64
-       .extra_user_64bit_cs = FLAT_USER_CS64,
-#endif
-       .name = "Xen",
-};
-
-static const struct pv_init_ops xen_init_ops __initconst = {
-       .patch = xen_patch,
-};
-
-static const struct pv_cpu_ops xen_cpu_ops __initconst = {
-       .cpuid = xen_cpuid,
-
-       .set_debugreg = xen_set_debugreg,
-       .get_debugreg = xen_get_debugreg,
-
-       .read_cr0 = xen_read_cr0,
-       .write_cr0 = xen_write_cr0,
-
-       .read_cr4 = native_read_cr4,
-       .write_cr4 = xen_write_cr4,
-
-#ifdef CONFIG_X86_64
-       .read_cr8 = xen_read_cr8,
-       .write_cr8 = xen_write_cr8,
-#endif
-
-       .wbinvd = native_wbinvd,
-
-       .read_msr = xen_read_msr,
-       .write_msr = xen_write_msr,
-
-       .read_msr_safe = xen_read_msr_safe,
-       .write_msr_safe = xen_write_msr_safe,
-
-       .read_pmc = xen_read_pmc,
-
-       .iret = xen_iret,
-#ifdef CONFIG_X86_64
-       .usergs_sysret64 = xen_sysret64,
-#endif
-
-       .load_tr_desc = paravirt_nop,
-       .set_ldt = xen_set_ldt,
-       .load_gdt = xen_load_gdt,
-       .load_idt = xen_load_idt,
-       .load_tls = xen_load_tls,
-#ifdef CONFIG_X86_64
-       .load_gs_index = xen_load_gs_index,
-#endif
-
-       .alloc_ldt = xen_alloc_ldt,
-       .free_ldt = xen_free_ldt,
-
-       .store_idt = native_store_idt,
-       .store_tr = xen_store_tr,
-
-       .write_ldt_entry = xen_write_ldt_entry,
-       .write_gdt_entry = xen_write_gdt_entry,
-       .write_idt_entry = xen_write_idt_entry,
-       .load_sp0 = xen_load_sp0,
-
-       .set_iopl_mask = xen_set_iopl_mask,
-       .io_delay = xen_io_delay,
-
-       /* Xen takes care of %gs when switching to usermode for us */
-       .swapgs = paravirt_nop,
-
-       .start_context_switch = paravirt_start_context_switch,
-       .end_context_switch = xen_end_context_switch,
-};
-
-void xen_reboot(int reason)
-{
-       struct sched_shutdown r = { .reason = reason };
-       int cpu;
-
-       for_each_online_cpu(cpu)
-               xen_pmu_finish(cpu);
-
-       if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
-               BUG();
-}
-
-static void xen_restart(char *msg)
-{
-       xen_reboot(SHUTDOWN_reboot);
-}
-
-void xen_emergency_restart(void)
-{
-       xen_reboot(SHUTDOWN_reboot);
-}
-
-static void xen_machine_halt(void)
-{
-       xen_reboot(SHUTDOWN_poweroff);
-}
-
-static void xen_machine_power_off(void)
-{
-       if (pm_power_off)
-               pm_power_off();
-       xen_reboot(SHUTDOWN_poweroff);
-}
-
-static void xen_crash_shutdown(struct pt_regs *regs)
-{
-       xen_reboot(SHUTDOWN_crash);
-}
-
-static int
-xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
-{
-       if (!kexec_crash_loaded())
-               xen_reboot(SHUTDOWN_crash);
-       return NOTIFY_DONE;
-}
-
-static struct notifier_block xen_panic_block = {
-       .notifier_call= xen_panic_event,
-       .priority = INT_MIN
-};
-
-int xen_panic_handler_init(void)
-{
-       atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
-       return 0;
-}
-
-static const struct machine_ops xen_machine_ops __initconst = {
-       .restart = xen_restart,
-       .halt = xen_machine_halt,
-       .power_off = xen_machine_power_off,
-       .shutdown = xen_machine_halt,
-       .crash_shutdown = xen_crash_shutdown,
-       .emergency_restart = xen_emergency_restart,
-};
-
-static unsigned char xen_get_nmi_reason(void)
-{
-       unsigned char reason = 0;
-
-       /* Construct a value which looks like it came from port 0x61. */
-       if (test_bit(_XEN_NMIREASON_io_error,
-                    &HYPERVISOR_shared_info->arch.nmi_reason))
-               reason |= NMI_REASON_IOCHK;
-       if (test_bit(_XEN_NMIREASON_pci_serr,
-                    &HYPERVISOR_shared_info->arch.nmi_reason))
-               reason |= NMI_REASON_SERR;
-
-       return reason;
-}
-
-static void __init xen_boot_params_init_edd(void)
-{
-#if IS_ENABLED(CONFIG_EDD)
-       struct xen_platform_op op;
-       struct edd_info *edd_info;
-       u32 *mbr_signature;
-       unsigned nr;
-       int ret;
-
-       edd_info = boot_params.eddbuf;
-       mbr_signature = boot_params.edd_mbr_sig_buffer;
-
-       op.cmd = XENPF_firmware_info;
-
-       op.u.firmware_info.type = XEN_FW_DISK_INFO;
-       for (nr = 0; nr < EDDMAXNR; nr++) {
-               struct edd_info *info = edd_info + nr;
-
-               op.u.firmware_info.index = nr;
-               info->params.length = sizeof(info->params);
-               set_xen_guest_handle(op.u.firmware_info.u.disk_info.edd_params,
-                                    &info->params);
-               ret = HYPERVISOR_platform_op(&op);
-               if (ret)
-                       break;
-
-#define C(x) info->x = op.u.firmware_info.u.disk_info.x
-               C(device);
-               C(version);
-               C(interface_support);
-               C(legacy_max_cylinder);
-               C(legacy_max_head);
-               C(legacy_sectors_per_track);
-#undef C
-       }
-       boot_params.eddbuf_entries = nr;
-
-       op.u.firmware_info.type = XEN_FW_DISK_MBR_SIGNATURE;
-       for (nr = 0; nr < EDD_MBR_SIG_MAX; nr++) {
-               op.u.firmware_info.index = nr;
-               ret = HYPERVISOR_platform_op(&op);
-               if (ret)
-                       break;
-               mbr_signature[nr] = op.u.firmware_info.u.disk_mbr_signature.mbr_signature;
-       }
-       boot_params.edd_mbr_sig_buf_entries = nr;
-#endif
-}
-
-/*
- * Set up the GDT and segment registers for -fstack-protector.  Until
- * we do this, we have to be careful not to call any stack-protected
- * function, which is most of the kernel.
- */
-static void xen_setup_gdt(int cpu)
-{
-       pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
-       pv_cpu_ops.load_gdt = xen_load_gdt_boot;
-
-       setup_stack_canary_segment(0);
-       switch_to_new_gdt(0);
-
-       pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
-       pv_cpu_ops.load_gdt = xen_load_gdt;
-}
-
-static void __init xen_dom0_set_legacy_features(void)
-{
-       x86_platform.legacy.rtc = 1;
-}
-
-int xen_cpuhp_setup(int (*cpu_up_prepare_cb)(unsigned int),
-                   int (*cpu_dead_cb)(unsigned int))
-{
-       int rc;
-
-       rc = cpuhp_setup_state_nocalls(CPUHP_XEN_PREPARE,
-                                      "x86/xen/hvm_guest:prepare",
-                                      cpu_up_prepare_cb, cpu_dead_cb);
-       if (rc >= 0) {
-               rc = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
-                                              "x86/xen/hvm_guest:online",
-                                              xen_cpu_up_online, NULL);
-               if (rc < 0)
-                       cpuhp_remove_state_nocalls(CPUHP_XEN_PREPARE);
-       }
-
-       return rc >= 0 ? 0 : rc;
-}
-
-/* First C function to be called on Xen boot */
-asmlinkage __visible void __init xen_start_kernel(void)
-{
-       struct physdev_set_iopl set_iopl;
-       unsigned long initrd_start = 0;
-       int rc;
-
-       if (!xen_start_info)
-               return;
-
-       xen_domain_type = XEN_PV_DOMAIN;
-
-       xen_setup_features();
-
-       xen_setup_machphys_mapping();
-
-       /* Install Xen paravirt ops */
-       pv_info = xen_info;
-       pv_init_ops = xen_init_ops;
-       pv_cpu_ops = xen_cpu_ops;
-
-       x86_platform.get_nmi_reason = xen_get_nmi_reason;
-
-       x86_init.resources.memory_setup = xen_memory_setup;
-       x86_init.oem.arch_setup = xen_arch_setup;
-       x86_init.oem.banner = xen_banner;
-
-       xen_init_time_ops();
-
-       /*
-        * Set up some pagetable state before starting to set any ptes.
-        */
-
-       xen_init_mmu_ops();
-
-       /* Prevent unwanted bits from being set in PTEs. */
-       __supported_pte_mask &= ~_PAGE_GLOBAL;
-
-       /*
-        * Prevent page tables from being allocated in highmem, even
-        * if CONFIG_HIGHPTE is enabled.
-        */
-       __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
-
-       /* Work out if we support NX */
-       x86_configure_nx();
-
-       /* Get mfn list */
-       xen_build_dynamic_phys_to_machine();
-
-       /*
-        * Set up kernel GDT and segment registers, mainly so that
-        * -fstack-protector code can be executed.
-        */
-       xen_setup_gdt(0);
-
-       xen_init_irq_ops();
-       xen_init_cpuid_mask();
-
-#ifdef CONFIG_X86_LOCAL_APIC
-       /*
-        * set up the basic apic ops.
-        */
-       xen_init_apic();
-#endif
-
-       if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
-               pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
-               pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
-       }
-
-       machine_ops = xen_machine_ops;
-
-       /*
-        * The only reliable way to retain the initial address of the
-        * percpu gdt_page is to remember it here, so we can go and
-        * mark it RW later, when the initial percpu area is freed.
-        */
-       xen_initial_gdt = &per_cpu(gdt_page, 0);
-
-       xen_smp_init();
-
-#ifdef CONFIG_ACPI_NUMA
-       /*
-        * The pages we from Xen are not related to machine pages, so
-        * any NUMA information the kernel tries to get from ACPI will
-        * be meaningless.  Prevent it from trying.
-        */
-       acpi_numa = -1;
-#endif
-       /* Don't do the full vcpu_info placement stuff until we have a
-          possible map and a non-dummy shared_info. */
-       per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
-
-       WARN_ON(xen_cpuhp_setup(xen_cpu_up_prepare_pv, xen_cpu_dead_pv));
-
-       local_irq_disable();
-       early_boot_irqs_disabled = true;
-
-       xen_raw_console_write("mapping kernel into physical memory\n");
-       xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base,
-                                  xen_start_info->nr_pages);
-       xen_reserve_special_pages();
-
-       /* keep using Xen gdt for now; no urgent need to change it */
-
-#ifdef CONFIG_X86_32
-       pv_info.kernel_rpl = 1;
-       if (xen_feature(XENFEAT_supervisor_mode_kernel))
-               pv_info.kernel_rpl = 0;
-#else
-       pv_info.kernel_rpl = 0;
-#endif
-       /* set the limit of our address space */
-       xen_reserve_top();
-
-       /*
-        * We used to do this in xen_arch_setup, but that is too late
-        * on AMD were early_cpu_init (run before ->arch_setup()) calls
-        * early_amd_init which pokes 0xcf8 port.
-        */
-       set_iopl.iopl = 1;
-       rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
-       if (rc != 0)
-               xen_raw_printk("physdev_op failed %d\n", rc);
-
-#ifdef CONFIG_X86_32
-       /* set up basic CPUID stuff */
-       cpu_detect(&new_cpu_data);
-       set_cpu_cap(&new_cpu_data, X86_FEATURE_FPU);
-       new_cpu_data.x86_capability[CPUID_1_EDX] = cpuid_edx(1);
-#endif
-
-       if (xen_start_info->mod_start) {
-           if (xen_start_info->flags & SIF_MOD_START_PFN)
-               initrd_start = PFN_PHYS(xen_start_info->mod_start);
-           else
-               initrd_start = __pa(xen_start_info->mod_start);
-       }
-
-       /* Poke various useful things into boot_params */
-       boot_params.hdr.type_of_loader = (9 << 4) | 0;
-       boot_params.hdr.ramdisk_image = initrd_start;
-       boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
-       boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
-       boot_params.hdr.hardware_subarch = X86_SUBARCH_XEN;
-
-       if (!xen_initial_domain()) {
-               add_preferred_console("xenboot", 0, NULL);
-               add_preferred_console("tty", 0, NULL);
-               add_preferred_console("hvc", 0, NULL);
-               if (pci_xen)
-                       x86_init.pci.arch_init = pci_xen_init;
-       } else {
-               const struct dom0_vga_console_info *info =
-                       (void *)((char *)xen_start_info +
-                                xen_start_info->console.dom0.info_off);
-               struct xen_platform_op op = {
-                       .cmd = XENPF_firmware_info,
-                       .interface_version = XENPF_INTERFACE_VERSION,
-                       .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
-               };
-
-               x86_platform.set_legacy_features =
-                               xen_dom0_set_legacy_features;
-               xen_init_vga(info, xen_start_info->console.dom0.info_size);
-               xen_start_info->console.domU.mfn = 0;
-               xen_start_info->console.domU.evtchn = 0;
-
-               if (HYPERVISOR_platform_op(&op) == 0)
-                       boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
-
-               /* Make sure ACS will be enabled */
-               pci_request_acs();
-
-               xen_acpi_sleep_register();
-
-               /* Avoid searching for BIOS MP tables */
-               x86_init.mpparse.find_smp_config = x86_init_noop;
-               x86_init.mpparse.get_smp_config = x86_init_uint_noop;
-
-               xen_boot_params_init_edd();
-       }
-#ifdef CONFIG_PCI
-       /* PCI BIOS service won't work from a PV guest. */
-       pci_probe &= ~PCI_PROBE_BIOS;
-#endif
-       xen_raw_console_write("about to get started...\n");
-
-       /* Let's presume PV guests always boot on vCPU with id 0. */
-       per_cpu(xen_vcpu_id, 0) = 0;
-
-       xen_setup_runstate_info(0);
-
-       xen_efi_init();
-
-       /* Start the world */
-#ifdef CONFIG_X86_32
-       i386_start_kernel();
-#else
-       cr4_init_shadow(); /* 32b kernel does this in i386_start_kernel() */
-       x86_64_start_reservations((char *)__pa_symbol(&boot_params));
-#endif
-}
-
-static int xen_cpu_up_prepare_pv(unsigned int cpu)
-{
-       int rc;
-
-       xen_setup_timer(cpu);
-
-       rc = xen_smp_intr_init(cpu);
-       if (rc) {
-               WARN(1, "xen_smp_intr_init() for CPU %d failed: %d\n",
-                    cpu, rc);
-               return rc;
-       }
-       return 0;
-}
-
-static int xen_cpu_dead_pv(unsigned int cpu)
-{
-       xen_smp_intr_free(cpu);
-
-       xen_teardown_timer(cpu);
-
-       return 0;
-}
-
-static int xen_cpu_up_online(unsigned int cpu)
-{
-       xen_init_lock_cpu(cpu);
-       return 0;
-}
-
-static uint32_t __init xen_platform_pv(void)
-{
-       if (xen_pv_domain())
-               return xen_cpuid_base();
-
-       return 0;
-}
-
-static void xen_set_cpu_features(struct cpuinfo_x86 *c)
-{
-       clear_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
-       set_cpu_cap(c, X86_FEATURE_XENPV);
-}
-
-void xen_pin_vcpu(int cpu)
-{
-       static bool disable_pinning;
-       struct sched_pin_override pin_override;
-       int ret;
-
-       if (disable_pinning)
-               return;
-
-       pin_override.pcpu = cpu;
-       ret = HYPERVISOR_sched_op(SCHEDOP_pin_override, &pin_override);
-
-       /* Ignore errors when removing override. */
-       if (cpu < 0)
-               return;
-
-       switch (ret) {
-       case -ENOSYS:
-               pr_warn("Unable to pin on physical cpu %d. In case of problems consider vcpu pinning.\n",
-                       cpu);
-               disable_pinning = true;
-               break;
-       case -EPERM:
-               WARN(1, "Trying to pin vcpu without having privilege to do so\n");
-               disable_pinning = true;
-               break;
-       case -EINVAL:
-       case -EBUSY:
-               pr_warn("Physical cpu %d not available for pinning. Check Xen cpu configuration.\n",
-                       cpu);
-               break;
-       case 0:
-               break;
-       default:
-               WARN(1, "rc %d while trying to pin vcpu\n", ret);
-               disable_pinning = true;
-       }
-}
-
-const struct hypervisor_x86 x86_hyper_xen_pv = {
-       .name                   = "Xen PV",
-       .detect                 = xen_platform_pv,
-       .set_cpu_features       = xen_set_cpu_features,
-       .pin_vcpu               = xen_pin_vcpu,
-};
-EXPORT_SYMBOL(x86_hyper_xen_pv);
-
 #ifdef CONFIG_HOTPLUG_CPU
 void xen_arch_register_cpu(int num)
 {
diff --git a/arch/x86/xen/enlighten_pv.c b/arch/x86/xen/enlighten_pv.c
new file mode 100644 (file)
index 0000000..1cb7947
--- /dev/null
@@ -0,0 +1,1512 @@
+/*
+ * Core of Xen paravirt_ops implementation.
+ *
+ * This file contains the xen_paravirt_ops structure itself, and the
+ * implementations for:
+ * - privileged instructions
+ * - interrupt flags
+ * - segment operations
+ * - booting and setup
+ *
+ * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
+ */
+
+#include <linux/cpu.h>
+#include <linux/kernel.h>
+#include <linux/init.h>
+#include <linux/smp.h>
+#include <linux/preempt.h>
+#include <linux/hardirq.h>
+#include <linux/percpu.h>
+#include <linux/delay.h>
+#include <linux/start_kernel.h>
+#include <linux/sched.h>
+#include <linux/kprobes.h>
+#include <linux/bootmem.h>
+#include <linux/export.h>
+#include <linux/mm.h>
+#include <linux/page-flags.h>
+#include <linux/highmem.h>
+#include <linux/console.h>
+#include <linux/pci.h>
+#include <linux/gfp.h>
+#include <linux/memblock.h>
+#include <linux/edd.h>
+#include <linux/frame.h>
+
+#include <xen/xen.h>
+#include <xen/events.h>
+#include <xen/interface/xen.h>
+#include <xen/interface/version.h>
+#include <xen/interface/physdev.h>
+#include <xen/interface/vcpu.h>
+#include <xen/interface/memory.h>
+#include <xen/interface/nmi.h>
+#include <xen/interface/xen-mca.h>
+#include <xen/features.h>
+#include <xen/page.h>
+#include <xen/hvc-console.h>
+#include <xen/acpi.h>
+
+#include <asm/paravirt.h>
+#include <asm/apic.h>
+#include <asm/page.h>
+#include <asm/xen/pci.h>
+#include <asm/xen/hypercall.h>
+#include <asm/xen/hypervisor.h>
+#include <asm/xen/cpuid.h>
+#include <asm/fixmap.h>
+#include <asm/processor.h>
+#include <asm/proto.h>
+#include <asm/msr-index.h>
+#include <asm/traps.h>
+#include <asm/setup.h>
+#include <asm/desc.h>
+#include <asm/pgalloc.h>
+#include <asm/pgtable.h>
+#include <asm/tlbflush.h>
+#include <asm/reboot.h>
+#include <asm/stackprotector.h>
+#include <asm/hypervisor.h>
+#include <asm/mach_traps.h>
+#include <asm/mwait.h>
+#include <asm/pci_x86.h>
+#include <asm/cpu.h>
+
+#ifdef CONFIG_ACPI
+#include <linux/acpi.h>
+#include <asm/acpi.h>
+#include <acpi/pdc_intel.h>
+#include <acpi/processor.h>
+#include <xen/interface/platform.h>
+#endif
+
+#include "xen-ops.h"
+#include "mmu.h"
+#include "smp.h"
+#include "multicalls.h"
+#include "pmu.h"
+
+void *xen_initial_gdt;
+
+RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
+
+static int xen_cpu_up_prepare_pv(unsigned int cpu);
+static int xen_cpu_dead_pv(unsigned int cpu);
+
+struct tls_descs {
+       struct desc_struct desc[3];
+};
+
+/*
+ * Updating the 3 TLS descriptors in the GDT on every task switch is
+ * surprisingly expensive so we avoid updating them if they haven't
+ * changed.  Since Xen writes different descriptors than the one
+ * passed in the update_descriptor hypercall we keep shadow copies to
+ * compare against.
+ */
+static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
+
+/*
+ * On restore, set the vcpu placement up again.
+ * If it fails, then we're in a bad state, since
+ * we can't back out from using it...
+ */
+void xen_vcpu_restore(void)
+{
+       int cpu;
+
+       for_each_possible_cpu(cpu) {
+               bool other_cpu = (cpu != smp_processor_id());
+               bool is_up = HYPERVISOR_vcpu_op(VCPUOP_is_up, xen_vcpu_nr(cpu),
+                                               NULL);
+
+               if (other_cpu && is_up &&
+                   HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL))
+                       BUG();
+
+               xen_setup_runstate_info(cpu);
+
+               if (xen_have_vcpu_info_placement)
+                       xen_vcpu_setup(cpu);
+
+               if (other_cpu && is_up &&
+                   HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL))
+                       BUG();
+       }
+}
+
+static void __init xen_banner(void)
+{
+       unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
+       struct xen_extraversion extra;
+       HYPERVISOR_xen_version(XENVER_extraversion, &extra);
+
+       pr_info("Booting paravirtualized kernel %son %s\n",
+               xen_feature(XENFEAT_auto_translated_physmap) ?
+                       "with PVH extensions " : "", pv_info.name);
+       printk(KERN_INFO "Xen version: %d.%d%s%s\n",
+              version >> 16, version & 0xffff, extra.extraversion,
+              xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
+}
+/* Check if running on Xen version (major, minor) or later */
+bool
+xen_running_on_version_or_later(unsigned int major, unsigned int minor)
+{
+       unsigned int version;
+
+       if (!xen_domain())
+               return false;
+
+       version = HYPERVISOR_xen_version(XENVER_version, NULL);
+       if ((((version >> 16) == major) && ((version & 0xffff) >= minor)) ||
+               ((version >> 16) > major))
+               return true;
+       return false;
+}
+
+#define CPUID_THERM_POWER_LEAF 6
+#define APERFMPERF_PRESENT 0
+
+static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
+static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
+
+static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
+static __read_mostly unsigned int cpuid_leaf5_ecx_val;
+static __read_mostly unsigned int cpuid_leaf5_edx_val;
+
+static void xen_cpuid(unsigned int *ax, unsigned int *bx,
+                     unsigned int *cx, unsigned int *dx)
+{
+       unsigned maskebx = ~0;
+       unsigned maskecx = ~0;
+       unsigned maskedx = ~0;
+       unsigned setecx = 0;
+       /*
+        * Mask out inconvenient features, to try and disable as many
+        * unsupported kernel subsystems as possible.
+        */
+       switch (*ax) {
+       case 1:
+               maskecx = cpuid_leaf1_ecx_mask;
+               setecx = cpuid_leaf1_ecx_set_mask;
+               maskedx = cpuid_leaf1_edx_mask;
+               break;
+
+       case CPUID_MWAIT_LEAF:
+               /* Synthesize the values.. */
+               *ax = 0;
+               *bx = 0;
+               *cx = cpuid_leaf5_ecx_val;
+               *dx = cpuid_leaf5_edx_val;
+               return;
+
+       case CPUID_THERM_POWER_LEAF:
+               /* Disabling APERFMPERF for kernel usage */
+               maskecx = ~(1 << APERFMPERF_PRESENT);
+               break;
+
+       case 0xb:
+               /* Suppress extended topology stuff */
+               maskebx = 0;
+               break;
+       }
+
+       asm(XEN_EMULATE_PREFIX "cpuid"
+               : "=a" (*ax),
+                 "=b" (*bx),
+                 "=c" (*cx),
+                 "=d" (*dx)
+               : "0" (*ax), "2" (*cx));
+
+       *bx &= maskebx;
+       *cx &= maskecx;
+       *cx |= setecx;
+       *dx &= maskedx;
+}
+STACK_FRAME_NON_STANDARD(xen_cpuid); /* XEN_EMULATE_PREFIX */
+
+static bool __init xen_check_mwait(void)
+{
+#ifdef CONFIG_ACPI
+       struct xen_platform_op op = {
+               .cmd                    = XENPF_set_processor_pminfo,
+               .u.set_pminfo.id        = -1,
+               .u.set_pminfo.type      = XEN_PM_PDC,
+       };
+       uint32_t buf[3];
+       unsigned int ax, bx, cx, dx;
+       unsigned int mwait_mask;
+
+       /* We need to determine whether it is OK to expose the MWAIT
+        * capability to the kernel to harvest deeper than C3 states from ACPI
+        * _CST using the processor_harvest_xen.c module. For this to work, we
+        * need to gather the MWAIT_LEAF values (which the cstate.c code
+        * checks against). The hypervisor won't expose the MWAIT flag because
+        * it would break backwards compatibility; so we will find out directly
+        * from the hardware and hypercall.
+        */
+       if (!xen_initial_domain())
+               return false;
+
+       /*
+        * When running under platform earlier than Xen4.2, do not expose
+        * mwait, to avoid the risk of loading native acpi pad driver
+        */
+       if (!xen_running_on_version_or_later(4, 2))
+               return false;
+
+       ax = 1;
+       cx = 0;
+
+       native_cpuid(&ax, &bx, &cx, &dx);
+
+       mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
+                    (1 << (X86_FEATURE_MWAIT % 32));
+
+       if ((cx & mwait_mask) != mwait_mask)
+               return false;
+
+       /* We need to emulate the MWAIT_LEAF and for that we need both
+        * ecx and edx. The hypercall provides only partial information.
+        */
+
+       ax = CPUID_MWAIT_LEAF;
+       bx = 0;
+       cx = 0;
+       dx = 0;
+
+       native_cpuid(&ax, &bx, &cx, &dx);
+
+       /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
+        * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
+        */
+       buf[0] = ACPI_PDC_REVISION_ID;
+       buf[1] = 1;
+       buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
+
+       set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
+
+       if ((HYPERVISOR_platform_op(&op) == 0) &&
+           (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
+               cpuid_leaf5_ecx_val = cx;
+               cpuid_leaf5_edx_val = dx;
+       }
+       return true;
+#else
+       return false;
+#endif
+}
+static void __init xen_init_cpuid_mask(void)
+{
+       unsigned int ax, bx, cx, dx;
+       unsigned int xsave_mask;
+
+       cpuid_leaf1_edx_mask =
+               ~((1 << X86_FEATURE_MTRR) |  /* disable MTRR */
+                 (1 << X86_FEATURE_ACC));   /* thermal monitoring */
+
+       if (!xen_initial_domain())
+               cpuid_leaf1_edx_mask &=
+                       ~((1 << X86_FEATURE_ACPI));  /* disable ACPI */
+
+       cpuid_leaf1_ecx_mask &= ~(1 << (X86_FEATURE_X2APIC % 32));
+
+       ax = 1;
+       cx = 0;
+       cpuid(1, &ax, &bx, &cx, &dx);
+
+       xsave_mask =
+               (1 << (X86_FEATURE_XSAVE % 32)) |
+               (1 << (X86_FEATURE_OSXSAVE % 32));
+
+       /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
+       if ((cx & xsave_mask) != xsave_mask)
+               cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
+       if (xen_check_mwait())
+               cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
+}
+
+static void xen_set_debugreg(int reg, unsigned long val)
+{
+       HYPERVISOR_set_debugreg(reg, val);
+}
+
+static unsigned long xen_get_debugreg(int reg)
+{
+       return HYPERVISOR_get_debugreg(reg);
+}
+
+static void xen_end_context_switch(struct task_struct *next)
+{
+       xen_mc_flush();
+       paravirt_end_context_switch(next);
+}
+
+static unsigned long xen_store_tr(void)
+{
+       return 0;
+}
+
+/*
+ * Set the page permissions for a particular virtual address.  If the
+ * address is a vmalloc mapping (or other non-linear mapping), then
+ * find the linear mapping of the page and also set its protections to
+ * match.
+ */
+static void set_aliased_prot(void *v, pgprot_t prot)
+{
+       int level;
+       pte_t *ptep;
+       pte_t pte;
+       unsigned long pfn;
+       struct page *page;
+       unsigned char dummy;
+
+       ptep = lookup_address((unsigned long)v, &level);
+       BUG_ON(ptep == NULL);
+
+       pfn = pte_pfn(*ptep);
+       page = pfn_to_page(pfn);
+
+       pte = pfn_pte(pfn, prot);
+
+       /*
+        * Careful: update_va_mapping() will fail if the virtual address
+        * we're poking isn't populated in the page tables.  We don't
+        * need to worry about the direct map (that's always in the page
+        * tables), but we need to be careful about vmap space.  In
+        * particular, the top level page table can lazily propagate
+        * entries between processes, so if we've switched mms since we
+        * vmapped the target in the first place, we might not have the
+        * top-level page table entry populated.
+        *
+        * We disable preemption because we want the same mm active when
+        * we probe the target and when we issue the hypercall.  We'll
+        * have the same nominal mm, but if we're a kernel thread, lazy
+        * mm dropping could change our pgd.
+        *
+        * Out of an abundance of caution, this uses __get_user() to fault
+        * in the target address just in case there's some obscure case
+        * in which the target address isn't readable.
+        */
+
+       preempt_disable();
+
+       probe_kernel_read(&dummy, v, 1);
+
+       if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
+               BUG();
+
+       if (!PageHighMem(page)) {
+               void *av = __va(PFN_PHYS(pfn));
+
+               if (av != v)
+                       if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
+                               BUG();
+       } else
+               kmap_flush_unused();
+
+       preempt_enable();
+}
+
+static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
+{
+       const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
+       int i;
+
+       /*
+        * We need to mark the all aliases of the LDT pages RO.  We
+        * don't need to call vm_flush_aliases(), though, since that's
+        * only responsible for flushing aliases out the TLBs, not the
+        * page tables, and Xen will flush the TLB for us if needed.
+        *
+        * To avoid confusing future readers: none of this is necessary
+        * to load the LDT.  The hypervisor only checks this when the
+        * LDT is faulted in due to subsequent descriptor access.
+        */
+
+       for (i = 0; i < entries; i += entries_per_page)
+               set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
+}
+
+static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
+{
+       const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
+       int i;
+
+       for (i = 0; i < entries; i += entries_per_page)
+               set_aliased_prot(ldt + i, PAGE_KERNEL);
+}
+
+static void xen_set_ldt(const void *addr, unsigned entries)
+{
+       struct mmuext_op *op;
+       struct multicall_space mcs = xen_mc_entry(sizeof(*op));
+
+       trace_xen_cpu_set_ldt(addr, entries);
+
+       op = mcs.args;
+       op->cmd = MMUEXT_SET_LDT;
+       op->arg1.linear_addr = (unsigned long)addr;
+       op->arg2.nr_ents = entries;
+
+       MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
+
+       xen_mc_issue(PARAVIRT_LAZY_CPU);
+}
+
+static void xen_load_gdt(const struct desc_ptr *dtr)
+{
+       unsigned long va = dtr->address;
+       unsigned int size = dtr->size + 1;
+       unsigned pages = DIV_ROUND_UP(size, PAGE_SIZE);
+       unsigned long frames[pages];
+       int f;
+
+       /*
+        * A GDT can be up to 64k in size, which corresponds to 8192
+        * 8-byte entries, or 16 4k pages..
+        */
+
+       BUG_ON(size > 65536);
+       BUG_ON(va & ~PAGE_MASK);
+
+       for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
+               int level;
+               pte_t *ptep;
+               unsigned long pfn, mfn;
+               void *virt;
+
+               /*
+                * The GDT is per-cpu and is in the percpu data area.
+                * That can be virtually mapped, so we need to do a
+                * page-walk to get the underlying MFN for the
+                * hypercall.  The page can also be in the kernel's
+                * linear range, so we need to RO that mapping too.
+                */
+               ptep = lookup_address(va, &level);
+               BUG_ON(ptep == NULL);
+
+               pfn = pte_pfn(*ptep);
+               mfn = pfn_to_mfn(pfn);
+               virt = __va(PFN_PHYS(pfn));
+
+               frames[f] = mfn;
+
+               make_lowmem_page_readonly((void *)va);
+               make_lowmem_page_readonly(virt);
+       }
+
+       if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
+               BUG();
+}
+
+/*
+ * load_gdt for early boot, when the gdt is only mapped once
+ */
+static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
+{
+       unsigned long va = dtr->address;
+       unsigned int size = dtr->size + 1;
+       unsigned pages = DIV_ROUND_UP(size, PAGE_SIZE);
+       unsigned long frames[pages];
+       int f;
+
+       /*
+        * A GDT can be up to 64k in size, which corresponds to 8192
+        * 8-byte entries, or 16 4k pages..
+        */
+
+       BUG_ON(size > 65536);
+       BUG_ON(va & ~PAGE_MASK);
+
+       for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
+               pte_t pte;
+               unsigned long pfn, mfn;
+
+               pfn = virt_to_pfn(va);
+               mfn = pfn_to_mfn(pfn);
+
+               pte = pfn_pte(pfn, PAGE_KERNEL_RO);
+
+               if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
+                       BUG();
+
+               frames[f] = mfn;
+       }
+
+       if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
+               BUG();
+}
+
+static inline bool desc_equal(const struct desc_struct *d1,
+                             const struct desc_struct *d2)
+{
+       return d1->a == d2->a && d1->b == d2->b;
+}
+
+static void load_TLS_descriptor(struct thread_struct *t,
+                               unsigned int cpu, unsigned int i)
+{
+       struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
+       struct desc_struct *gdt;
+       xmaddr_t maddr;
+       struct multicall_space mc;
+
+       if (desc_equal(shadow, &t->tls_array[i]))
+               return;
+
+       *shadow = t->tls_array[i];
+
+       gdt = get_cpu_gdt_rw(cpu);
+       maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
+       mc = __xen_mc_entry(0);
+
+       MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
+}
+
+static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
+{
+       /*
+        * XXX sleazy hack: If we're being called in a lazy-cpu zone
+        * and lazy gs handling is enabled, it means we're in a
+        * context switch, and %gs has just been saved.  This means we
+        * can zero it out to prevent faults on exit from the
+        * hypervisor if the next process has no %gs.  Either way, it
+        * has been saved, and the new value will get loaded properly.
+        * This will go away as soon as Xen has been modified to not
+        * save/restore %gs for normal hypercalls.
+        *
+        * On x86_64, this hack is not used for %gs, because gs points
+        * to KERNEL_GS_BASE (and uses it for PDA references), so we
+        * must not zero %gs on x86_64
+        *
+        * For x86_64, we need to zero %fs, otherwise we may get an
+        * exception between the new %fs descriptor being loaded and
+        * %fs being effectively cleared at __switch_to().
+        */
+       if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
+#ifdef CONFIG_X86_32
+               lazy_load_gs(0);
+#else
+               loadsegment(fs, 0);
+#endif
+       }
+
+       xen_mc_batch();
+
+       load_TLS_descriptor(t, cpu, 0);
+       load_TLS_descriptor(t, cpu, 1);
+       load_TLS_descriptor(t, cpu, 2);
+
+       xen_mc_issue(PARAVIRT_LAZY_CPU);
+}
+
+#ifdef CONFIG_X86_64
+static void xen_load_gs_index(unsigned int idx)
+{
+       if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
+               BUG();
+}
+#endif
+
+static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
+                               const void *ptr)
+{
+       xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
+       u64 entry = *(u64 *)ptr;
+
+       trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
+
+       preempt_disable();
+
+       xen_mc_flush();
+       if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
+               BUG();
+
+       preempt_enable();
+}
+
+static int cvt_gate_to_trap(int vector, const gate_desc *val,
+                           struct trap_info *info)
+{
+       unsigned long addr;
+
+       if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
+               return 0;
+
+       info->vector = vector;
+
+       addr = gate_offset(*val);
+#ifdef CONFIG_X86_64
+       /*
+        * Look for known traps using IST, and substitute them
+        * appropriately.  The debugger ones are the only ones we care
+        * about.  Xen will handle faults like double_fault,
+        * so we should never see them.  Warn if
+        * there's an unexpected IST-using fault handler.
+        */
+       if (addr == (unsigned long)debug)
+               addr = (unsigned long)xen_debug;
+       else if (addr == (unsigned long)int3)
+               addr = (unsigned long)xen_int3;
+       else if (addr == (unsigned long)stack_segment)
+               addr = (unsigned long)xen_stack_segment;
+       else if (addr == (unsigned long)double_fault) {
+               /* Don't need to handle these */
+               return 0;
+#ifdef CONFIG_X86_MCE
+       } else if (addr == (unsigned long)machine_check) {
+               /*
+                * when xen hypervisor inject vMCE to guest,
+                * use native mce handler to handle it
+                */
+               ;
+#endif
+       } else if (addr == (unsigned long)nmi)
+               /*
+                * Use the native version as well.
+                */
+               ;
+       else {
+               /* Some other trap using IST? */
+               if (WARN_ON(val->ist != 0))
+                       return 0;
+       }
+#endif /* CONFIG_X86_64 */
+       info->address = addr;
+
+       info->cs = gate_segment(*val);
+       info->flags = val->dpl;
+       /* interrupt gates clear IF */
+       if (val->type == GATE_INTERRUPT)
+               info->flags |= 1 << 2;
+
+       return 1;
+}
+
+/* Locations of each CPU's IDT */
+static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
+
+/* Set an IDT entry.  If the entry is part of the current IDT, then
+   also update Xen. */
+static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
+{
+       unsigned long p = (unsigned long)&dt[entrynum];
+       unsigned long start, end;
+
+       trace_xen_cpu_write_idt_entry(dt, entrynum, g);
+
+       preempt_disable();
+
+       start = __this_cpu_read(idt_desc.address);
+       end = start + __this_cpu_read(idt_desc.size) + 1;
+
+       xen_mc_flush();
+
+       native_write_idt_entry(dt, entrynum, g);
+
+       if (p >= start && (p + 8) <= end) {
+               struct trap_info info[2];
+
+               info[1].address = 0;
+
+               if (cvt_gate_to_trap(entrynum, g, &info[0]))
+                       if (HYPERVISOR_set_trap_table(info))
+                               BUG();
+       }
+
+       preempt_enable();
+}
+
+static void xen_convert_trap_info(const struct desc_ptr *desc,
+                                 struct trap_info *traps)
+{
+       unsigned in, out, count;
+
+       count = (desc->size+1) / sizeof(gate_desc);
+       BUG_ON(count > 256);
+
+       for (in = out = 0; in < count; in++) {
+               gate_desc *entry = (gate_desc *)(desc->address) + in;
+
+               if (cvt_gate_to_trap(in, entry, &traps[out]))
+                       out++;
+       }
+       traps[out].address = 0;
+}
+
+void xen_copy_trap_info(struct trap_info *traps)
+{
+       const struct desc_ptr *desc = this_cpu_ptr(&idt_desc);
+
+       xen_convert_trap_info(desc, traps);
+}
+
+/* Load a new IDT into Xen.  In principle this can be per-CPU, so we
+   hold a spinlock to protect the static traps[] array (static because
+   it avoids allocation, and saves stack space). */
+static void xen_load_idt(const struct desc_ptr *desc)
+{
+       static DEFINE_SPINLOCK(lock);
+       static struct trap_info traps[257];
+
+       trace_xen_cpu_load_idt(desc);
+
+       spin_lock(&lock);
+
+       memcpy(this_cpu_ptr(&idt_desc), desc, sizeof(idt_desc));
+
+       xen_convert_trap_info(desc, traps);
+
+       xen_mc_flush();
+       if (HYPERVISOR_set_trap_table(traps))
+               BUG();
+
+       spin_unlock(&lock);
+}
+
+/* Write a GDT descriptor entry.  Ignore LDT descriptors, since
+   they're handled differently. */
+static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
+                               const void *desc, int type)
+{
+       trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
+
+       preempt_disable();
+
+       switch (type) {
+       case DESC_LDT:
+       case DESC_TSS:
+               /* ignore */
+               break;
+
+       default: {
+               xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
+
+               xen_mc_flush();
+               if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
+                       BUG();
+       }
+
+       }
+
+       preempt_enable();
+}
+
+/*
+ * Version of write_gdt_entry for use at early boot-time needed to
+ * update an entry as simply as possible.
+ */
+static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
+                                           const void *desc, int type)
+{
+       trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
+
+       switch (type) {
+       case DESC_LDT:
+       case DESC_TSS:
+               /* ignore */
+               break;
+
+       default: {
+               xmaddr_t maddr = virt_to_machine(&dt[entry]);
+
+               if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
+                       dt[entry] = *(struct desc_struct *)desc;
+       }
+
+       }
+}
+
+static void xen_load_sp0(struct tss_struct *tss,
+                        struct thread_struct *thread)
+{
+       struct multicall_space mcs;
+
+       mcs = xen_mc_entry(0);
+       MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
+       xen_mc_issue(PARAVIRT_LAZY_CPU);
+       tss->x86_tss.sp0 = thread->sp0;
+}
+
+void xen_set_iopl_mask(unsigned mask)
+{
+       struct physdev_set_iopl set_iopl;
+
+       /* Force the change at ring 0. */
+       set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
+       HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
+}
+
+static void xen_io_delay(void)
+{
+}
+
+static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
+
+static unsigned long xen_read_cr0(void)
+{
+       unsigned long cr0 = this_cpu_read(xen_cr0_value);
+
+       if (unlikely(cr0 == 0)) {
+               cr0 = native_read_cr0();
+               this_cpu_write(xen_cr0_value, cr0);
+       }
+
+       return cr0;
+}
+
+static void xen_write_cr0(unsigned long cr0)
+{
+       struct multicall_space mcs;
+
+       this_cpu_write(xen_cr0_value, cr0);
+
+       /* Only pay attention to cr0.TS; everything else is
+          ignored. */
+       mcs = xen_mc_entry(0);
+
+       MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
+
+       xen_mc_issue(PARAVIRT_LAZY_CPU);
+}
+
+static void xen_write_cr4(unsigned long cr4)
+{
+       cr4 &= ~(X86_CR4_PGE | X86_CR4_PSE | X86_CR4_PCE);
+
+       native_write_cr4(cr4);
+}
+#ifdef CONFIG_X86_64
+static inline unsigned long xen_read_cr8(void)
+{
+       return 0;
+}
+static inline void xen_write_cr8(unsigned long val)
+{
+       BUG_ON(val);
+}
+#endif
+
+static u64 xen_read_msr_safe(unsigned int msr, int *err)
+{
+       u64 val;
+
+       if (pmu_msr_read(msr, &val, err))
+               return val;
+
+       val = native_read_msr_safe(msr, err);
+       switch (msr) {
+       case MSR_IA32_APICBASE:
+#ifdef CONFIG_X86_X2APIC
+               if (!(cpuid_ecx(1) & (1 << (X86_FEATURE_X2APIC & 31))))
+#endif
+                       val &= ~X2APIC_ENABLE;
+               break;
+       }
+       return val;
+}
+
+static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
+{
+       int ret;
+
+       ret = 0;
+
+       switch (msr) {
+#ifdef CONFIG_X86_64
+               unsigned which;
+               u64 base;
+
+       case MSR_FS_BASE:               which = SEGBASE_FS; goto set;
+       case MSR_KERNEL_GS_BASE:        which = SEGBASE_GS_USER; goto set;
+       case MSR_GS_BASE:               which = SEGBASE_GS_KERNEL; goto set;
+
+       set:
+               base = ((u64)high << 32) | low;
+               if (HYPERVISOR_set_segment_base(which, base) != 0)
+                       ret = -EIO;
+               break;
+#endif
+
+       case MSR_STAR:
+       case MSR_CSTAR:
+       case MSR_LSTAR:
+       case MSR_SYSCALL_MASK:
+       case MSR_IA32_SYSENTER_CS:
+       case MSR_IA32_SYSENTER_ESP:
+       case MSR_IA32_SYSENTER_EIP:
+               /* Fast syscall setup is all done in hypercalls, so
+                  these are all ignored.  Stub them out here to stop
+                  Xen console noise. */
+               break;
+
+       default:
+               if (!pmu_msr_write(msr, low, high, &ret))
+                       ret = native_write_msr_safe(msr, low, high);
+       }
+
+       return ret;
+}
+
+static u64 xen_read_msr(unsigned int msr)
+{
+       /*
+        * This will silently swallow a #GP from RDMSR.  It may be worth
+        * changing that.
+        */
+       int err;
+
+       return xen_read_msr_safe(msr, &err);
+}
+
+static void xen_write_msr(unsigned int msr, unsigned low, unsigned high)
+{
+       /*
+        * This will silently swallow a #GP from WRMSR.  It may be worth
+        * changing that.
+        */
+       xen_write_msr_safe(msr, low, high);
+}
+
+void xen_setup_shared_info(void)
+{
+       if (!xen_feature(XENFEAT_auto_translated_physmap)) {
+               set_fixmap(FIX_PARAVIRT_BOOTMAP,
+                          xen_start_info->shared_info);
+
+               HYPERVISOR_shared_info =
+                       (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
+       } else
+               HYPERVISOR_shared_info =
+                       (struct shared_info *)__va(xen_start_info->shared_info);
+
+#ifndef CONFIG_SMP
+       /* In UP this is as good a place as any to set up shared info */
+       xen_setup_vcpu_info_placement();
+#endif
+
+       xen_setup_mfn_list_list();
+}
+
+/* This is called once we have the cpu_possible_mask */
+void xen_setup_vcpu_info_placement(void)
+{
+       int cpu;
+
+       for_each_possible_cpu(cpu) {
+               /* Set up direct vCPU id mapping for PV guests. */
+               per_cpu(xen_vcpu_id, cpu) = cpu;
+               xen_vcpu_setup(cpu);
+       }
+
+       /*
+        * xen_vcpu_setup managed to place the vcpu_info within the
+        * percpu area for all cpus, so make use of it.
+        */
+       if (xen_have_vcpu_info_placement) {
+               pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
+               pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
+               pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
+               pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
+               pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
+       }
+}
+
+static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
+                         unsigned long addr, unsigned len)
+{
+       char *start, *end, *reloc;
+       unsigned ret;
+
+       start = end = reloc = NULL;
+
+#define SITE(op, x)                                                    \
+       case PARAVIRT_PATCH(op.x):                                      \
+       if (xen_have_vcpu_info_placement) {                             \
+               start = (char *)xen_##x##_direct;                       \
+               end = xen_##x##_direct_end;                             \
+               reloc = xen_##x##_direct_reloc;                         \
+       }                                                               \
+       goto patch_site
+
+       switch (type) {
+               SITE(pv_irq_ops, irq_enable);
+               SITE(pv_irq_ops, irq_disable);
+               SITE(pv_irq_ops, save_fl);
+               SITE(pv_irq_ops, restore_fl);
+#undef SITE
+
+       patch_site:
+               if (start == NULL || (end-start) > len)
+                       goto default_patch;
+
+               ret = paravirt_patch_insns(insnbuf, len, start, end);
+
+               /* Note: because reloc is assigned from something that
+                  appears to be an array, gcc assumes it's non-null,
+                  but doesn't know its relationship with start and
+                  end. */
+               if (reloc > start && reloc < end) {
+                       int reloc_off = reloc - start;
+                       long *relocp = (long *)(insnbuf + reloc_off);
+                       long delta = start - (char *)addr;
+
+                       *relocp += delta;
+               }
+               break;
+
+       default_patch:
+       default:
+               ret = paravirt_patch_default(type, clobbers, insnbuf,
+                                            addr, len);
+               break;
+       }
+
+       return ret;
+}
+
+static const struct pv_info xen_info __initconst = {
+       .shared_kernel_pmd = 0,
+
+#ifdef CONFIG_X86_64
+       .extra_user_64bit_cs = FLAT_USER_CS64,
+#endif
+       .name = "Xen",
+};
+
+static const struct pv_init_ops xen_init_ops __initconst = {
+       .patch = xen_patch,
+};
+
+static const struct pv_cpu_ops xen_cpu_ops __initconst = {
+       .cpuid = xen_cpuid,
+
+       .set_debugreg = xen_set_debugreg,
+       .get_debugreg = xen_get_debugreg,
+
+       .read_cr0 = xen_read_cr0,
+       .write_cr0 = xen_write_cr0,
+
+       .read_cr4 = native_read_cr4,
+       .write_cr4 = xen_write_cr4,
+
+#ifdef CONFIG_X86_64
+       .read_cr8 = xen_read_cr8,
+       .write_cr8 = xen_write_cr8,
+#endif
+
+       .wbinvd = native_wbinvd,
+
+       .read_msr = xen_read_msr,
+       .write_msr = xen_write_msr,
+
+       .read_msr_safe = xen_read_msr_safe,
+       .write_msr_safe = xen_write_msr_safe,
+
+       .read_pmc = xen_read_pmc,
+
+       .iret = xen_iret,
+#ifdef CONFIG_X86_64
+       .usergs_sysret64 = xen_sysret64,
+#endif
+
+       .load_tr_desc = paravirt_nop,
+       .set_ldt = xen_set_ldt,
+       .load_gdt = xen_load_gdt,
+       .load_idt = xen_load_idt,
+       .load_tls = xen_load_tls,
+#ifdef CONFIG_X86_64
+       .load_gs_index = xen_load_gs_index,
+#endif
+
+       .alloc_ldt = xen_alloc_ldt,
+       .free_ldt = xen_free_ldt,
+
+       .store_idt = native_store_idt,
+       .store_tr = xen_store_tr,
+
+       .write_ldt_entry = xen_write_ldt_entry,
+       .write_gdt_entry = xen_write_gdt_entry,
+       .write_idt_entry = xen_write_idt_entry,
+       .load_sp0 = xen_load_sp0,
+
+       .set_iopl_mask = xen_set_iopl_mask,
+       .io_delay = xen_io_delay,
+
+       /* Xen takes care of %gs when switching to usermode for us */
+       .swapgs = paravirt_nop,
+
+       .start_context_switch = paravirt_start_context_switch,
+       .end_context_switch = xen_end_context_switch,
+};
+
+static void xen_restart(char *msg)
+{
+       xen_reboot(SHUTDOWN_reboot);
+}
+
+static void xen_machine_halt(void)
+{
+       xen_reboot(SHUTDOWN_poweroff);
+}
+
+static void xen_machine_power_off(void)
+{
+       if (pm_power_off)
+               pm_power_off();
+       xen_reboot(SHUTDOWN_poweroff);
+}
+
+static void xen_crash_shutdown(struct pt_regs *regs)
+{
+       xen_reboot(SHUTDOWN_crash);
+}
+
+static const struct machine_ops xen_machine_ops __initconst = {
+       .restart = xen_restart,
+       .halt = xen_machine_halt,
+       .power_off = xen_machine_power_off,
+       .shutdown = xen_machine_halt,
+       .crash_shutdown = xen_crash_shutdown,
+       .emergency_restart = xen_emergency_restart,
+};
+
+static unsigned char xen_get_nmi_reason(void)
+{
+       unsigned char reason = 0;
+
+       /* Construct a value which looks like it came from port 0x61. */
+       if (test_bit(_XEN_NMIREASON_io_error,
+                    &HYPERVISOR_shared_info->arch.nmi_reason))
+               reason |= NMI_REASON_IOCHK;
+       if (test_bit(_XEN_NMIREASON_pci_serr,
+                    &HYPERVISOR_shared_info->arch.nmi_reason))
+               reason |= NMI_REASON_SERR;
+
+       return reason;
+}
+
+static void __init xen_boot_params_init_edd(void)
+{
+#if IS_ENABLED(CONFIG_EDD)
+       struct xen_platform_op op;
+       struct edd_info *edd_info;
+       u32 *mbr_signature;
+       unsigned nr;
+       int ret;
+
+       edd_info = boot_params.eddbuf;
+       mbr_signature = boot_params.edd_mbr_sig_buffer;
+
+       op.cmd = XENPF_firmware_info;
+
+       op.u.firmware_info.type = XEN_FW_DISK_INFO;
+       for (nr = 0; nr < EDDMAXNR; nr++) {
+               struct edd_info *info = edd_info + nr;
+
+               op.u.firmware_info.index = nr;
+               info->params.length = sizeof(info->params);
+               set_xen_guest_handle(op.u.firmware_info.u.disk_info.edd_params,
+                                    &info->params);
+               ret = HYPERVISOR_platform_op(&op);
+               if (ret)
+                       break;
+
+#define C(x) info->x = op.u.firmware_info.u.disk_info.x
+               C(device);
+               C(version);
+               C(interface_support);
+               C(legacy_max_cylinder);
+               C(legacy_max_head);
+               C(legacy_sectors_per_track);
+#undef C
+       }
+       boot_params.eddbuf_entries = nr;
+
+       op.u.firmware_info.type = XEN_FW_DISK_MBR_SIGNATURE;
+       for (nr = 0; nr < EDD_MBR_SIG_MAX; nr++) {
+               op.u.firmware_info.index = nr;
+               ret = HYPERVISOR_platform_op(&op);
+               if (ret)
+                       break;
+               mbr_signature[nr] = op.u.firmware_info.u.disk_mbr_signature.mbr_signature;
+       }
+       boot_params.edd_mbr_sig_buf_entries = nr;
+#endif
+}
+
+/*
+ * Set up the GDT and segment registers for -fstack-protector.  Until
+ * we do this, we have to be careful not to call any stack-protected
+ * function, which is most of the kernel.
+ */
+static void xen_setup_gdt(int cpu)
+{
+       pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
+       pv_cpu_ops.load_gdt = xen_load_gdt_boot;
+
+       setup_stack_canary_segment(0);
+       switch_to_new_gdt(0);
+
+       pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
+       pv_cpu_ops.load_gdt = xen_load_gdt;
+}
+
+static void __init xen_dom0_set_legacy_features(void)
+{
+       x86_platform.legacy.rtc = 1;
+}
+
+/* First C function to be called on Xen boot */
+asmlinkage __visible void __init xen_start_kernel(void)
+{
+       struct physdev_set_iopl set_iopl;
+       unsigned long initrd_start = 0;
+       int rc;
+
+       if (!xen_start_info)
+               return;
+
+       xen_domain_type = XEN_PV_DOMAIN;
+
+       xen_setup_features();
+
+       xen_setup_machphys_mapping();
+
+       /* Install Xen paravirt ops */
+       pv_info = xen_info;
+       pv_init_ops = xen_init_ops;
+       pv_cpu_ops = xen_cpu_ops;
+
+       x86_platform.get_nmi_reason = xen_get_nmi_reason;
+
+       x86_init.resources.memory_setup = xen_memory_setup;
+       x86_init.oem.arch_setup = xen_arch_setup;
+       x86_init.oem.banner = xen_banner;
+
+       xen_init_time_ops();
+
+       /*
+        * Set up some pagetable state before starting to set any ptes.
+        */
+
+       xen_init_mmu_ops();
+
+       /* Prevent unwanted bits from being set in PTEs. */
+       __supported_pte_mask &= ~_PAGE_GLOBAL;
+
+       /*
+        * Prevent page tables from being allocated in highmem, even
+        * if CONFIG_HIGHPTE is enabled.
+        */
+       __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
+
+       /* Work out if we support NX */
+       x86_configure_nx();
+
+       /* Get mfn list */
+       xen_build_dynamic_phys_to_machine();
+
+       /*
+        * Set up kernel GDT and segment registers, mainly so that
+        * -fstack-protector code can be executed.
+        */
+       xen_setup_gdt(0);
+
+       xen_init_irq_ops();
+       xen_init_cpuid_mask();
+
+#ifdef CONFIG_X86_LOCAL_APIC
+       /*
+        * set up the basic apic ops.
+        */
+       xen_init_apic();
+#endif
+
+       if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
+               pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
+               pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
+       }
+
+       machine_ops = xen_machine_ops;
+
+       /*
+        * The only reliable way to retain the initial address of the
+        * percpu gdt_page is to remember it here, so we can go and
+        * mark it RW later, when the initial percpu area is freed.
+        */
+       xen_initial_gdt = &per_cpu(gdt_page, 0);
+
+       xen_smp_init();
+
+#ifdef CONFIG_ACPI_NUMA
+       /*
+        * The pages we from Xen are not related to machine pages, so
+        * any NUMA information the kernel tries to get from ACPI will
+        * be meaningless.  Prevent it from trying.
+        */
+       acpi_numa = -1;
+#endif
+       /* Don't do the full vcpu_info placement stuff until we have a
+          possible map and a non-dummy shared_info. */
+       per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
+
+       WARN_ON(xen_cpuhp_setup(xen_cpu_up_prepare_pv, xen_cpu_dead_pv));
+
+       local_irq_disable();
+       early_boot_irqs_disabled = true;
+
+       xen_raw_console_write("mapping kernel into physical memory\n");
+       xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base,
+                                  xen_start_info->nr_pages);
+       xen_reserve_special_pages();
+
+       /* keep using Xen gdt for now; no urgent need to change it */
+
+#ifdef CONFIG_X86_32
+       pv_info.kernel_rpl = 1;
+       if (xen_feature(XENFEAT_supervisor_mode_kernel))
+               pv_info.kernel_rpl = 0;
+#else
+       pv_info.kernel_rpl = 0;
+#endif
+       /* set the limit of our address space */
+       xen_reserve_top();
+
+       /*
+        * We used to do this in xen_arch_setup, but that is too late
+        * on AMD were early_cpu_init (run before ->arch_setup()) calls
+        * early_amd_init which pokes 0xcf8 port.
+        */
+       set_iopl.iopl = 1;
+       rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
+       if (rc != 0)
+               xen_raw_printk("physdev_op failed %d\n", rc);
+
+#ifdef CONFIG_X86_32
+       /* set up basic CPUID stuff */
+       cpu_detect(&new_cpu_data);
+       set_cpu_cap(&new_cpu_data, X86_FEATURE_FPU);
+       new_cpu_data.x86_capability[CPUID_1_EDX] = cpuid_edx(1);
+#endif
+
+       if (xen_start_info->mod_start) {
+           if (xen_start_info->flags & SIF_MOD_START_PFN)
+               initrd_start = PFN_PHYS(xen_start_info->mod_start);
+           else
+               initrd_start = __pa(xen_start_info->mod_start);
+       }
+
+       /* Poke various useful things into boot_params */
+       boot_params.hdr.type_of_loader = (9 << 4) | 0;
+       boot_params.hdr.ramdisk_image = initrd_start;
+       boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
+       boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
+       boot_params.hdr.hardware_subarch = X86_SUBARCH_XEN;
+
+       if (!xen_initial_domain()) {
+               add_preferred_console("xenboot", 0, NULL);
+               add_preferred_console("tty", 0, NULL);
+               add_preferred_console("hvc", 0, NULL);
+               if (pci_xen)
+                       x86_init.pci.arch_init = pci_xen_init;
+       } else {
+               const struct dom0_vga_console_info *info =
+                       (void *)((char *)xen_start_info +
+                                xen_start_info->console.dom0.info_off);
+               struct xen_platform_op op = {
+                       .cmd = XENPF_firmware_info,
+                       .interface_version = XENPF_INTERFACE_VERSION,
+                       .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
+               };
+
+               x86_platform.set_legacy_features =
+                               xen_dom0_set_legacy_features;
+               xen_init_vga(info, xen_start_info->console.dom0.info_size);
+               xen_start_info->console.domU.mfn = 0;
+               xen_start_info->console.domU.evtchn = 0;
+
+               if (HYPERVISOR_platform_op(&op) == 0)
+                       boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
+
+               /* Make sure ACS will be enabled */
+               pci_request_acs();
+
+               xen_acpi_sleep_register();
+
+               /* Avoid searching for BIOS MP tables */
+               x86_init.mpparse.find_smp_config = x86_init_noop;
+               x86_init.mpparse.get_smp_config = x86_init_uint_noop;
+
+               xen_boot_params_init_edd();
+       }
+#ifdef CONFIG_PCI
+       /* PCI BIOS service won't work from a PV guest. */
+       pci_probe &= ~PCI_PROBE_BIOS;
+#endif
+       xen_raw_console_write("about to get started...\n");
+
+       /* Let's presume PV guests always boot on vCPU with id 0. */
+       per_cpu(xen_vcpu_id, 0) = 0;
+
+       xen_setup_runstate_info(0);
+
+       xen_efi_init();
+
+       /* Start the world */
+#ifdef CONFIG_X86_32
+       i386_start_kernel();
+#else
+       cr4_init_shadow(); /* 32b kernel does this in i386_start_kernel() */
+       x86_64_start_reservations((char *)__pa_symbol(&boot_params));
+#endif
+}
+
+static int xen_cpu_up_prepare_pv(unsigned int cpu)
+{
+       int rc;
+
+       xen_setup_timer(cpu);
+
+       rc = xen_smp_intr_init(cpu);
+       if (rc) {
+               WARN(1, "xen_smp_intr_init() for CPU %d failed: %d\n",
+                    cpu, rc);
+               return rc;
+       }
+       return 0;
+}
+
+static int xen_cpu_dead_pv(unsigned int cpu)
+{
+       xen_smp_intr_free(cpu);
+
+       xen_teardown_timer(cpu);
+
+       return 0;
+}
+
+static uint32_t __init xen_platform_pv(void)
+{
+       if (xen_pv_domain())
+               return xen_cpuid_base();
+
+       return 0;
+}
+
+static void xen_set_cpu_features(struct cpuinfo_x86 *c)
+{
+       clear_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
+       set_cpu_cap(c, X86_FEATURE_XENPV);
+}
+
+const struct hypervisor_x86 x86_hyper_xen_pv = {
+       .name                   = "Xen PV",
+       .detect                 = xen_platform_pv,
+       .set_cpu_features       = xen_set_cpu_features,
+       .pin_vcpu               = xen_pin_vcpu,
+};
+EXPORT_SYMBOL(x86_hyper_xen_pv);