arch/x86/kernel/cpu/mshyperv.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * HyperV  Detection code.
   4  *
   5  * Copyright (C) 2010, Novell, Inc.
   6  * Author : K. Y. Srinivasan <ksrinivasan@novell.com>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/time.h>
  11 #include <linux/clocksource.h>
  12 #include <linux/init.h>
  13 #include <linux/export.h>
  14 #include <linux/hardirq.h>
  15 #include <linux/efi.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/irq.h>
  18 #include <linux/kexec.h>
  19 #include <linux/i8253.h>
  20 #include <linux/random.h>
  21 #include <asm/processor.h>
  22 #include <asm/hypervisor.h>
  23 #include <asm/hyperv-tlfs.h>
  24 #include <asm/mshyperv.h>
  25 #include <asm/desc.h>
  26 #include <asm/idtentry.h>
  27 #include <asm/irq_regs.h>
  28 #include <asm/i8259.h>
  29 #include <asm/apic.h>
  30 #include <asm/timer.h>
  31 #include <asm/reboot.h>
  32 #include <asm/nmi.h>
  33 #include <clocksource/hyperv_timer.h>
  34
  35 struct ms_hyperv_info ms_hyperv;
  36 EXPORT_SYMBOL_GPL(ms_hyperv);
  37
  38 #if IS_ENABLED(CONFIG_HYPERV)
  39 static void (*vmbus_handler)(void);
  40 static void (*hv_stimer0_handler)(void);
  41 static void (*hv_kexec_handler)(void);
  42 static void (*hv_crash_handler)(struct pt_regs *regs);
  43
  44 DEFINE_IDTENTRY_SYSVEC(sysvec_hyperv_callback)
  45 {
  46         struct pt_regs *old_regs = set_irq_regs(regs);
  47
  48         inc_irq_stat(irq_hv_callback_count);
  49         if (vmbus_handler)
  50                 vmbus_handler();
  51
  52         if (ms_hyperv.hints & HV_DEPRECATING_AEOI_RECOMMENDED)
  53                 ack_APIC_irq();
  54
  55         set_irq_regs(old_regs);
  56 }
  57
  58 int hv_setup_vmbus_irq(int irq, void (*handler)(void))
  59 {
  60         /*
  61          * The 'irq' argument is ignored on x86/x64 because a hard-coded
  62          * interrupt vector is used for Hyper-V interrupts.
  63          */
  64         vmbus_handler = handler;
  65         return 0;
  66 }
  67
  68 void hv_remove_vmbus_irq(void)
  69 {
  70         /* We have no way to deallocate the interrupt gate */
  71         vmbus_handler = NULL;
  72 }
  73 EXPORT_SYMBOL_GPL(hv_setup_vmbus_irq);
  74 EXPORT_SYMBOL_GPL(hv_remove_vmbus_irq);
  75
  76 /*
  77  * Routines to do per-architecture handling of stimer0
  78  * interrupts when in Direct Mode
  79  */
  80 DEFINE_IDTENTRY_SYSVEC(sysvec_hyperv_stimer0)
  81 {
  82         struct pt_regs *old_regs = set_irq_regs(regs);
  83
  84         inc_irq_stat(hyperv_stimer0_count);
  85         if (hv_stimer0_handler)
  86                 hv_stimer0_handler();
  87         add_interrupt_randomness(HYPERV_STIMER0_VECTOR, 0);
  88         ack_APIC_irq();
  89
  90         set_irq_regs(old_regs);
  91 }
  92
  93 int hv_setup_stimer0_irq(int *irq, int *vector, void (*handler)(void))
  94 {
  95         *vector = HYPERV_STIMER0_VECTOR;
  96         *irq = -1;   /* Unused on x86/x64 */
  97         hv_stimer0_handler = handler;
  98         return 0;
  99 }
 100 EXPORT_SYMBOL_GPL(hv_setup_stimer0_irq);
 101
 102 void hv_remove_stimer0_irq(int irq)
 103 {
 104         /* We have no way to deallocate the interrupt gate */
 105         hv_stimer0_handler = NULL;
 106 }
 107 EXPORT_SYMBOL_GPL(hv_remove_stimer0_irq);
 108
 109 void hv_setup_kexec_handler(void (*handler)(void))
 110 {
 111         hv_kexec_handler = handler;
 112 }
 113 EXPORT_SYMBOL_GPL(hv_setup_kexec_handler);
 114
 115 void hv_remove_kexec_handler(void)
 116 {
 117         hv_kexec_handler = NULL;
 118 }
 119 EXPORT_SYMBOL_GPL(hv_remove_kexec_handler);
 120
 121 void hv_setup_crash_handler(void (*handler)(struct pt_regs *regs))
 122 {
 123         hv_crash_handler = handler;
 124 }
 125 EXPORT_SYMBOL_GPL(hv_setup_crash_handler);
 126
 127 void hv_remove_crash_handler(void)
 128 {
 129         hv_crash_handler = NULL;
 130 }
 131 EXPORT_SYMBOL_GPL(hv_remove_crash_handler);
 132
 133 #ifdef CONFIG_KEXEC_CORE
 134 static void hv_machine_shutdown(void)
 135 {
 136         if (kexec_in_progress && hv_kexec_handler)
 137                 hv_kexec_handler();
 138
 139         /*
 140          * Call hv_cpu_die() on all the CPUs, otherwise later the hypervisor
 141          * corrupts the old VP Assist Pages and can crash the kexec kernel.
 142          */
 143         if (kexec_in_progress && hyperv_init_cpuhp > 0)
 144                 cpuhp_remove_state(hyperv_init_cpuhp);
 145
 146         /* The function calls stop_other_cpus(). */
 147         native_machine_shutdown();
 148
 149         /* Disable the hypercall page when there is only 1 active CPU. */
 150         if (kexec_in_progress)
 151                 hyperv_cleanup();
 152 }
 153
 154 static void hv_machine_crash_shutdown(struct pt_regs *regs)
 155 {
 156         if (hv_crash_handler)
 157                 hv_crash_handler(regs);
 158
 159         /* The function calls crash_smp_send_stop(). */
 160         native_machine_crash_shutdown(regs);
 161
 162         /* Disable the hypercall page when there is only 1 active CPU. */
 163         hyperv_cleanup();
 164 }
 165 #endif /* CONFIG_KEXEC_CORE */
 166 #endif /* CONFIG_HYPERV */
 167
 168 static uint32_t  __init ms_hyperv_platform(void)
 169 {
 170         u32 eax;
 171         u32 hyp_signature[3];
 172
 173         if (!boot_cpu_has(X86_FEATURE_HYPERVISOR))
 174                 return 0;
 175
 176         cpuid(HYPERV_CPUID_VENDOR_AND_MAX_FUNCTIONS,
 177               &eax, &hyp_signature[0], &hyp_signature[1], &hyp_signature[2]);
 178
 179         if (eax >= HYPERV_CPUID_MIN &&
 180             eax <= HYPERV_CPUID_MAX &&
 181             !memcmp("Microsoft Hv", hyp_signature, 12))
 182                 return HYPERV_CPUID_VENDOR_AND_MAX_FUNCTIONS;
 183
 184         return 0;
 185 }
 186
 187 static unsigned char hv_get_nmi_reason(void)
 188 {
 189         return 0;
 190 }
 191
 192 #ifdef CONFIG_X86_LOCAL_APIC
 193 /*
 194  * Prior to WS2016 Debug-VM sends NMIs to all CPUs which makes
 195  * it dificult to process CHANNELMSG_UNLOAD in case of crash. Handle
 196  * unknown NMI on the first CPU which gets it.
 197  */
 198 static int hv_nmi_unknown(unsigned int val, struct pt_regs *regs)
 199 {
 200         static atomic_t nmi_cpu = ATOMIC_INIT(-1);
 201
 202         if (!unknown_nmi_panic)
 203                 return NMI_DONE;
 204
 205         if (atomic_cmpxchg(&nmi_cpu, -1, raw_smp_processor_id()) != -1)
 206                 return NMI_HANDLED;
 207
 208         return NMI_DONE;
 209 }
 210 #endif
 211
 212 static unsigned long hv_get_tsc_khz(void)
 213 {
 214         unsigned long freq;
 215
 216         rdmsrl(HV_X64_MSR_TSC_FREQUENCY, freq);
 217
 218         return freq / 1000;
 219 }
 220
 221 #if defined(CONFIG_SMP) && IS_ENABLED(CONFIG_HYPERV)
 222 static void __init hv_smp_prepare_boot_cpu(void)
 223 {
 224         native_smp_prepare_boot_cpu();
 225 #if defined(CONFIG_X86_64) && defined(CONFIG_PARAVIRT_SPINLOCKS)
 226         hv_init_spinlocks();
 227 #endif
 228 }
 229 #endif
 230
 231 static void __init ms_hyperv_init_platform(void)
 232 {
 233         int hv_host_info_eax;
 234         int hv_host_info_ebx;
 235         int hv_host_info_ecx;
 236         int hv_host_info_edx;
 237
 238 #ifdef CONFIG_PARAVIRT
 239         pv_info.name = "Hyper-V";
 240 #endif
 241
 242         /*
 243          * Extract the features and hints
 244          */
 245         ms_hyperv.features = cpuid_eax(HYPERV_CPUID_FEATURES);
 246         ms_hyperv.misc_features = cpuid_edx(HYPERV_CPUID_FEATURES);
 247         ms_hyperv.hints    = cpuid_eax(HYPERV_CPUID_ENLIGHTMENT_INFO);
 248
 249         pr_info("Hyper-V: features 0x%x, hints 0x%x, misc 0x%x\n",
 250                 ms_hyperv.features, ms_hyperv.hints, ms_hyperv.misc_features);
 251
 252         ms_hyperv.max_vp_index = cpuid_eax(HYPERV_CPUID_IMPLEMENT_LIMITS);
 253         ms_hyperv.max_lp_index = cpuid_ebx(HYPERV_CPUID_IMPLEMENT_LIMITS);
 254
 255         pr_debug("Hyper-V: max %u virtual processors, %u logical processors\n",
 256                  ms_hyperv.max_vp_index, ms_hyperv.max_lp_index);
 257
 258         /*
 259          * Extract host information.
 260          */
 261         if (cpuid_eax(HYPERV_CPUID_VENDOR_AND_MAX_FUNCTIONS) >=
 262             HYPERV_CPUID_VERSION) {
 263                 hv_host_info_eax = cpuid_eax(HYPERV_CPUID_VERSION);
 264                 hv_host_info_ebx = cpuid_ebx(HYPERV_CPUID_VERSION);
 265                 hv_host_info_ecx = cpuid_ecx(HYPERV_CPUID_VERSION);
 266                 hv_host_info_edx = cpuid_edx(HYPERV_CPUID_VERSION);
 267
 268                 pr_info("Hyper-V Host Build:%d-%d.%d-%d-%d.%d\n",
 269                         hv_host_info_eax, hv_host_info_ebx >> 16,
 270                         hv_host_info_ebx & 0xFFFF, hv_host_info_ecx,
 271                         hv_host_info_edx >> 24, hv_host_info_edx & 0xFFFFFF);
 272         }
 273
 274         if (ms_hyperv.features & HV_ACCESS_FREQUENCY_MSRS &&
 275             ms_hyperv.misc_features & HV_FEATURE_FREQUENCY_MSRS_AVAILABLE) {
 276                 x86_platform.calibrate_tsc = hv_get_tsc_khz;
 277                 x86_platform.calibrate_cpu = hv_get_tsc_khz;
 278         }
 279
 280         if (ms_hyperv.hints & HV_X64_ENLIGHTENED_VMCS_RECOMMENDED) {
 281                 ms_hyperv.nested_features =
 282                         cpuid_eax(HYPERV_CPUID_NESTED_FEATURES);
 283         }
 284
 285         /*
 286          * Hyper-V expects to get crash register data or kmsg when
 287          * crash enlightment is available and system crashes. Set
 288          * crash_kexec_post_notifiers to be true to make sure that
 289          * calling crash enlightment interface before running kdump
 290          * kernel.
 291          */
 292         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE)
 293                 crash_kexec_post_notifiers = true;
 294
 295 #ifdef CONFIG_X86_LOCAL_APIC
 296         if (ms_hyperv.features & HV_ACCESS_FREQUENCY_MSRS &&
 297             ms_hyperv.misc_features & HV_FEATURE_FREQUENCY_MSRS_AVAILABLE) {
 298                 /*
 299                  * Get the APIC frequency.
 300                  */
 301                 u64     hv_lapic_frequency;
 302
 303                 rdmsrl(HV_X64_MSR_APIC_FREQUENCY, hv_lapic_frequency);
 304                 hv_lapic_frequency = div_u64(hv_lapic_frequency, HZ);
 305                 lapic_timer_period = hv_lapic_frequency;
 306                 pr_info("Hyper-V: LAPIC Timer Frequency: %#x\n",
 307                         lapic_timer_period);
 308         }
 309
 310         register_nmi_handler(NMI_UNKNOWN, hv_nmi_unknown, NMI_FLAG_FIRST,
 311                              "hv_nmi_unknown");
 312 #endif
 313
 314 #ifdef CONFIG_X86_IO_APIC
 315         no_timer_check = 1;
 316 #endif
 317
 318 #if IS_ENABLED(CONFIG_HYPERV) && defined(CONFIG_KEXEC_CORE)
 319         machine_ops.shutdown = hv_machine_shutdown;
 320         machine_ops.crash_shutdown = hv_machine_crash_shutdown;
 321 #endif
 322         if (ms_hyperv.features & HV_ACCESS_TSC_INVARIANT) {
 323                 wrmsrl(HV_X64_MSR_TSC_INVARIANT_CONTROL, 0x1);
 324                 setup_force_cpu_cap(X86_FEATURE_TSC_RELIABLE);
 325         } else {
 326                 mark_tsc_unstable("running on Hyper-V");
 327         }
 328
 329         /*
 330          * Generation 2 instances don't support reading the NMI status from
 331          * 0x61 port.
 332          */
 333         if (efi_enabled(EFI_BOOT))
 334                 x86_platform.get_nmi_reason = hv_get_nmi_reason;
 335
 336         /*
 337          * Hyper-V VMs have a PIT emulation quirk such that zeroing the
 338          * counter register during PIT shutdown restarts the PIT. So it
 339          * continues to interrupt @18.2 HZ. Setting i8253_clear_counter
 340          * to false tells pit_shutdown() not to zero the counter so that
 341          * the PIT really is shutdown. Generation 2 VMs don't have a PIT,
 342          * and setting this value has no effect.
 343          */
 344         i8253_clear_counter_on_shutdown = false;
 345
 346 #if IS_ENABLED(CONFIG_HYPERV)
 347         /*
 348          * Setup the hook to get control post apic initialization.
 349          */
 350         x86_platform.apic_post_init = hyperv_init;
 351         hyperv_setup_mmu_ops();
 352         /* Setup the IDT for hypervisor callback */
 353         alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_hyperv_callback);
 354
 355         /* Setup the IDT for reenlightenment notifications */
 356         if (ms_hyperv.features & HV_ACCESS_REENLIGHTENMENT) {
 357                 alloc_intr_gate(HYPERV_REENLIGHTENMENT_VECTOR,
 358                                 asm_sysvec_hyperv_reenlightenment);
 359         }
 360
 361         /* Setup the IDT for stimer0 */
 362         if (ms_hyperv.misc_features & HV_STIMER_DIRECT_MODE_AVAILABLE) {
 363                 alloc_intr_gate(HYPERV_STIMER0_VECTOR,
 364                                 asm_sysvec_hyperv_stimer0);
 365         }
 366
 367 # ifdef CONFIG_SMP
 368         smp_ops.smp_prepare_boot_cpu = hv_smp_prepare_boot_cpu;
 369 # endif
 370
 371         /*
 372          * Hyper-V doesn't provide irq remapping for IO-APIC. To enable x2apic,
 373          * set x2apic destination mode to physcial mode when x2apic is available
 374          * and Hyper-V IOMMU driver makes sure cpus assigned with IO-APIC irqs
 375          * have 8-bit APIC id.
 376          */
 377 # ifdef CONFIG_X86_X2APIC
 378         if (x2apic_supported())
 379                 x2apic_phys = 1;
 380 # endif
 381
 382         /* Register Hyper-V specific clocksource */
 383         hv_init_clocksource();
 384 #endif
 385 }
 386
 387 static bool __init ms_hyperv_x2apic_available(void)
 388 {
 389         return x2apic_supported();
 390 }
 391
 392 /*
 393  * If ms_hyperv_msi_ext_dest_id() returns true, hyperv_prepare_irq_remapping()
 394  * returns -ENODEV and the Hyper-V IOMMU driver is not used; instead, the
 395  * generic support of the 15-bit APIC ID is used: see __irq_msi_compose_msg().
 396  *
 397  * Note: for a VM on Hyper-V, the I/O-APIC is the only device which
 398  * (logically) generates MSIs directly to the system APIC irq domain.
 399  * There is no HPET, and PCI MSI/MSI-X interrupts are remapped by the
 400  * pci-hyperv host bridge.
 401  */
 402 static bool __init ms_hyperv_msi_ext_dest_id(void)
 403 {
 404         u32 eax;
 405
 406         eax = cpuid_eax(HYPERV_CPUID_VIRT_STACK_INTERFACE);
 407         if (eax != HYPERV_VS_INTERFACE_EAX_SIGNATURE)
 408                 return false;
 409
 410         eax = cpuid_eax(HYPERV_CPUID_VIRT_STACK_PROPERTIES);
 411         return eax & HYPERV_VS_PROPERTIES_EAX_EXTENDED_IOAPIC_RTE;
 412 }
 413
 414 const __initconst struct hypervisor_x86 x86_hyper_ms_hyperv = {
 415         .name                   = "Microsoft Hyper-V",
 416         .detect                 = ms_hyperv_platform,
 417         .type                   = X86_HYPER_MS_HYPERV,
 418         .init.x2apic_available  = ms_hyperv_x2apic_available,
 419         .init.msi_ext_dest_id   = ms_hyperv_msi_ext_dest_id,
 420         .init.init_platform     = ms_hyperv_init_platform,
 421 };