arch/x86/hyperv/hv_init.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * X86 specific Hyper-V initialization code.
   4  *
   5  * Copyright (C) 2016, Microsoft, Inc.
   6  *
   7  * Author : K. Y. Srinivasan <kys@microsoft.com>
   8  */
   9
  10 #include <linux/acpi.h>
  11 #include <linux/efi.h>
  12 #include <linux/types.h>
  13 #include <asm/apic.h>
  14 #include <asm/desc.h>
  15 #include <asm/hypervisor.h>
  16 #include <asm/hyperv-tlfs.h>
  17 #include <asm/mshyperv.h>
  18 #include <asm/idtentry.h>
  19 #include <linux/kexec.h>
  20 #include <linux/version.h>
  21 #include <linux/vmalloc.h>
  22 #include <linux/mm.h>
  23 #include <linux/hyperv.h>
  24 #include <linux/slab.h>
  25 #include <linux/kernel.h>
  26 #include <linux/cpuhotplug.h>
  27 #include <linux/syscore_ops.h>
  28 #include <clocksource/hyperv_timer.h>
  29
  30 int hyperv_init_cpuhp;
  31
  32 void *hv_hypercall_pg;
  33 EXPORT_SYMBOL_GPL(hv_hypercall_pg);
  34
  35 /* Storage to save the hypercall page temporarily for hibernation */
  36 static void *hv_hypercall_pg_saved;
  37
  38 u32 *hv_vp_index;
  39 EXPORT_SYMBOL_GPL(hv_vp_index);
  40
  41 struct hv_vp_assist_page **hv_vp_assist_page;
  42 EXPORT_SYMBOL_GPL(hv_vp_assist_page);
  43
  44 void  __percpu **hyperv_pcpu_input_arg;
  45 EXPORT_SYMBOL_GPL(hyperv_pcpu_input_arg);
  46
  47 u32 hv_max_vp_index;
  48 EXPORT_SYMBOL_GPL(hv_max_vp_index);
  49
  50 void *hv_alloc_hyperv_page(void)
  51 {
  52         BUILD_BUG_ON(PAGE_SIZE != HV_HYP_PAGE_SIZE);
  53
  54         return (void *)__get_free_page(GFP_KERNEL);
  55 }
  56 EXPORT_SYMBOL_GPL(hv_alloc_hyperv_page);
  57
  58 void *hv_alloc_hyperv_zeroed_page(void)
  59 {
  60         BUILD_BUG_ON(PAGE_SIZE != HV_HYP_PAGE_SIZE);
  61
  62         return (void *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
  63 }
  64 EXPORT_SYMBOL_GPL(hv_alloc_hyperv_zeroed_page);
  65
  66 void hv_free_hyperv_page(unsigned long addr)
  67 {
  68         free_page(addr);
  69 }
  70 EXPORT_SYMBOL_GPL(hv_free_hyperv_page);
  71
  72 static int hv_cpu_init(unsigned int cpu)
  73 {
  74         u64 msr_vp_index;
  75         struct hv_vp_assist_page **hvp = &hv_vp_assist_page[smp_processor_id()];
  76         void **input_arg;
  77         struct page *pg;
  78
  79         input_arg = (void **)this_cpu_ptr(hyperv_pcpu_input_arg);
  80         /* hv_cpu_init() can be called with IRQs disabled from hv_resume() */
  81         pg = alloc_page(irqs_disabled() ? GFP_ATOMIC : GFP_KERNEL);
  82         if (unlikely(!pg))
  83                 return -ENOMEM;
  84         *input_arg = page_address(pg);
  85
  86         hv_get_vp_index(msr_vp_index);
  87
  88         hv_vp_index[smp_processor_id()] = msr_vp_index;
  89
  90         if (msr_vp_index > hv_max_vp_index)
  91                 hv_max_vp_index = msr_vp_index;
  92
  93         if (!hv_vp_assist_page)
  94                 return 0;
  95
  96         /*
  97          * The VP ASSIST PAGE is an "overlay" page (see Hyper-V TLFS's Section
  98          * 5.2.1 "GPA Overlay Pages"). Here it must be zeroed out to make sure
  99          * we always write the EOI MSR in hv_apic_eoi_write() *after* the
 100          * EOI optimization is disabled in hv_cpu_die(), otherwise a CPU may
 101          * not be stopped in the case of CPU offlining and the VM will hang.
 102          */
 103         if (!*hvp) {
 104                 *hvp = __vmalloc(PAGE_SIZE, GFP_KERNEL | __GFP_ZERO);
 105         }
 106
 107         if (*hvp) {
 108                 u64 val;
 109
 110                 val = vmalloc_to_pfn(*hvp);
 111                 val = (val << HV_X64_MSR_VP_ASSIST_PAGE_ADDRESS_SHIFT) |
 112                         HV_X64_MSR_VP_ASSIST_PAGE_ENABLE;
 113
 114                 wrmsrl(HV_X64_MSR_VP_ASSIST_PAGE, val);
 115         }
 116
 117         return 0;
 118 }
 119
 120 static void (*hv_reenlightenment_cb)(void);
 121
 122 static void hv_reenlightenment_notify(struct work_struct *dummy)
 123 {
 124         struct hv_tsc_emulation_status emu_status;
 125
 126         rdmsrl(HV_X64_MSR_TSC_EMULATION_STATUS, *(u64 *)&emu_status);
 127
 128         /* Don't issue the callback if TSC accesses are not emulated */
 129         if (hv_reenlightenment_cb && emu_status.inprogress)
 130                 hv_reenlightenment_cb();
 131 }
 132 static DECLARE_DELAYED_WORK(hv_reenlightenment_work, hv_reenlightenment_notify);
 133
 134 void hyperv_stop_tsc_emulation(void)
 135 {
 136         u64 freq;
 137         struct hv_tsc_emulation_status emu_status;
 138
 139         rdmsrl(HV_X64_MSR_TSC_EMULATION_STATUS, *(u64 *)&emu_status);
 140         emu_status.inprogress = 0;
 141         wrmsrl(HV_X64_MSR_TSC_EMULATION_STATUS, *(u64 *)&emu_status);
 142
 143         rdmsrl(HV_X64_MSR_TSC_FREQUENCY, freq);
 144         tsc_khz = div64_u64(freq, 1000);
 145 }
 146 EXPORT_SYMBOL_GPL(hyperv_stop_tsc_emulation);
 147
 148 static inline bool hv_reenlightenment_available(void)
 149 {
 150         /*
 151          * Check for required features and priviliges to make TSC frequency
 152          * change notifications work.
 153          */
 154         return ms_hyperv.features & HV_ACCESS_FREQUENCY_MSRS &&
 155                 ms_hyperv.misc_features & HV_FEATURE_FREQUENCY_MSRS_AVAILABLE &&
 156                 ms_hyperv.features & HV_ACCESS_REENLIGHTENMENT;
 157 }
 158
 159 DEFINE_IDTENTRY_SYSVEC(sysvec_hyperv_reenlightenment)
 160 {
 161         ack_APIC_irq();
 162         inc_irq_stat(irq_hv_reenlightenment_count);
 163         schedule_delayed_work(&hv_reenlightenment_work, HZ/10);
 164 }
 165
 166 void set_hv_tscchange_cb(void (*cb)(void))
 167 {
 168         struct hv_reenlightenment_control re_ctrl = {
 169                 .vector = HYPERV_REENLIGHTENMENT_VECTOR,
 170                 .enabled = 1,
 171                 .target_vp = hv_vp_index[smp_processor_id()]
 172         };
 173         struct hv_tsc_emulation_control emu_ctrl = {.enabled = 1};
 174
 175         if (!hv_reenlightenment_available()) {
 176                 pr_warn("Hyper-V: reenlightenment support is unavailable\n");
 177                 return;
 178         }
 179
 180         hv_reenlightenment_cb = cb;
 181
 182         /* Make sure callback is registered before we write to MSRs */
 183         wmb();
 184
 185         wrmsrl(HV_X64_MSR_REENLIGHTENMENT_CONTROL, *((u64 *)&re_ctrl));
 186         wrmsrl(HV_X64_MSR_TSC_EMULATION_CONTROL, *((u64 *)&emu_ctrl));
 187 }
 188 EXPORT_SYMBOL_GPL(set_hv_tscchange_cb);
 189
 190 void clear_hv_tscchange_cb(void)
 191 {
 192         struct hv_reenlightenment_control re_ctrl;
 193
 194         if (!hv_reenlightenment_available())
 195                 return;
 196
 197         rdmsrl(HV_X64_MSR_REENLIGHTENMENT_CONTROL, *(u64 *)&re_ctrl);
 198         re_ctrl.enabled = 0;
 199         wrmsrl(HV_X64_MSR_REENLIGHTENMENT_CONTROL, *(u64 *)&re_ctrl);
 200
 201         hv_reenlightenment_cb = NULL;
 202 }
 203 EXPORT_SYMBOL_GPL(clear_hv_tscchange_cb);
 204
 205 static int hv_cpu_die(unsigned int cpu)
 206 {
 207         struct hv_reenlightenment_control re_ctrl;
 208         unsigned int new_cpu;
 209         unsigned long flags;
 210         void **input_arg;
 211         void *input_pg = NULL;
 212
 213         local_irq_save(flags);
 214         input_arg = (void **)this_cpu_ptr(hyperv_pcpu_input_arg);
 215         input_pg = *input_arg;
 216         *input_arg = NULL;
 217         local_irq_restore(flags);
 218         free_page((unsigned long)input_pg);
 219
 220         if (hv_vp_assist_page && hv_vp_assist_page[cpu])
 221                 wrmsrl(HV_X64_MSR_VP_ASSIST_PAGE, 0);
 222
 223         if (hv_reenlightenment_cb == NULL)
 224                 return 0;
 225
 226         rdmsrl(HV_X64_MSR_REENLIGHTENMENT_CONTROL, *((u64 *)&re_ctrl));
 227         if (re_ctrl.target_vp == hv_vp_index[cpu]) {
 228                 /*
 229                  * Reassign reenlightenment notifications to some other online
 230                  * CPU or just disable the feature if there are no online CPUs
 231                  * left (happens on hibernation).
 232                  */
 233                 new_cpu = cpumask_any_but(cpu_online_mask, cpu);
 234
 235                 if (new_cpu < nr_cpu_ids)
 236                         re_ctrl.target_vp = hv_vp_index[new_cpu];
 237                 else
 238                         re_ctrl.enabled = 0;
 239
 240                 wrmsrl(HV_X64_MSR_REENLIGHTENMENT_CONTROL, *((u64 *)&re_ctrl));
 241         }
 242
 243         return 0;
 244 }
 245
 246 static int __init hv_pci_init(void)
 247 {
 248         int gen2vm = efi_enabled(EFI_BOOT);
 249
 250         /*
 251          * For Generation-2 VM, we exit from pci_arch_init() by returning 0.
 252          * The purpose is to suppress the harmless warning:
 253          * "PCI: Fatal: No config space access function found"
 254          */
 255         if (gen2vm)
 256                 return 0;
 257
 258         /* For Generation-1 VM, we'll proceed in pci_arch_init().  */
 259         return 1;
 260 }
 261
 262 static int hv_suspend(void)
 263 {
 264         union hv_x64_msr_hypercall_contents hypercall_msr;
 265         int ret;
 266
 267         /*
 268          * Reset the hypercall page as it is going to be invalidated
 269          * accross hibernation. Setting hv_hypercall_pg to NULL ensures
 270          * that any subsequent hypercall operation fails safely instead of
 271          * crashing due to an access of an invalid page. The hypercall page
 272          * pointer is restored on resume.
 273          */
 274         hv_hypercall_pg_saved = hv_hypercall_pg;
 275         hv_hypercall_pg = NULL;
 276
 277         /* Disable the hypercall page in the hypervisor */
 278         rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 279         hypercall_msr.enable = 0;
 280         wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 281
 282         ret = hv_cpu_die(0);
 283         return ret;
 284 }
 285
 286 static void hv_resume(void)
 287 {
 288         union hv_x64_msr_hypercall_contents hypercall_msr;
 289         int ret;
 290
 291         ret = hv_cpu_init(0);
 292         WARN_ON(ret);
 293
 294         /* Re-enable the hypercall page */
 295         rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 296         hypercall_msr.enable = 1;
 297         hypercall_msr.guest_physical_address =
 298                 vmalloc_to_pfn(hv_hypercall_pg_saved);
 299         wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 300
 301         hv_hypercall_pg = hv_hypercall_pg_saved;
 302         hv_hypercall_pg_saved = NULL;
 303
 304         /*
 305          * Reenlightenment notifications are disabled by hv_cpu_die(0),
 306          * reenable them here if hv_reenlightenment_cb was previously set.
 307          */
 308         if (hv_reenlightenment_cb)
 309                 set_hv_tscchange_cb(hv_reenlightenment_cb);
 310 }
 311
 312 /* Note: when the ops are called, only CPU0 is online and IRQs are disabled. */
 313 static struct syscore_ops hv_syscore_ops = {
 314         .suspend        = hv_suspend,
 315         .resume         = hv_resume,
 316 };
 317
 318 static void (* __initdata old_setup_percpu_clockev)(void);
 319
 320 static void __init hv_stimer_setup_percpu_clockev(void)
 321 {
 322         /*
 323          * Ignore any errors in setting up stimer clockevents
 324          * as we can run with the LAPIC timer as a fallback.
 325          */
 326         (void)hv_stimer_alloc();
 327
 328         /*
 329          * Still register the LAPIC timer, because the direct-mode STIMER is
 330          * not supported by old versions of Hyper-V. This also allows users
 331          * to switch to LAPIC timer via /sys, if they want to.
 332          */
 333         if (old_setup_percpu_clockev)
 334                 old_setup_percpu_clockev();
 335 }
 336
 337 /*
 338  * This function is to be invoked early in the boot sequence after the
 339  * hypervisor has been detected.
 340  *
 341  * 1. Setup the hypercall page.
 342  * 2. Register Hyper-V specific clocksource.
 343  * 3. Setup Hyper-V specific APIC entry points.
 344  */
 345 void __init hyperv_init(void)
 346 {
 347         u64 guest_id, required_msrs;
 348         union hv_x64_msr_hypercall_contents hypercall_msr;
 349         int cpuhp, i;
 350
 351         if (x86_hyper_type != X86_HYPER_MS_HYPERV)
 352                 return;
 353
 354         /* Absolutely required MSRs */
 355         required_msrs = HV_MSR_HYPERCALL_AVAILABLE |
 356                 HV_MSR_VP_INDEX_AVAILABLE;
 357
 358         if ((ms_hyperv.features & required_msrs) != required_msrs)
 359                 return;
 360
 361         /*
 362          * Allocate the per-CPU state for the hypercall input arg.
 363          * If this allocation fails, we will not be able to setup
 364          * (per-CPU) hypercall input page and thus this failure is
 365          * fatal on Hyper-V.
 366          */
 367         hyperv_pcpu_input_arg = alloc_percpu(void  *);
 368
 369         BUG_ON(hyperv_pcpu_input_arg == NULL);
 370
 371         /* Allocate percpu VP index */
 372         hv_vp_index = kmalloc_array(num_possible_cpus(), sizeof(*hv_vp_index),
 373                                     GFP_KERNEL);
 374         if (!hv_vp_index)
 375                 return;
 376
 377         for (i = 0; i < num_possible_cpus(); i++)
 378                 hv_vp_index[i] = VP_INVAL;
 379
 380         hv_vp_assist_page = kcalloc(num_possible_cpus(),
 381                                     sizeof(*hv_vp_assist_page), GFP_KERNEL);
 382         if (!hv_vp_assist_page) {
 383                 ms_hyperv.hints &= ~HV_X64_ENLIGHTENED_VMCS_RECOMMENDED;
 384                 goto free_vp_index;
 385         }
 386
 387         cpuhp = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "x86/hyperv_init:online",
 388                                   hv_cpu_init, hv_cpu_die);
 389         if (cpuhp < 0)
 390                 goto free_vp_assist_page;
 391
 392         /*
 393          * Setup the hypercall page and enable hypercalls.
 394          * 1. Register the guest ID
 395          * 2. Enable the hypercall and register the hypercall page
 396          */
 397         guest_id = generate_guest_id(0, LINUX_VERSION_CODE, 0);
 398         wrmsrl(HV_X64_MSR_GUEST_OS_ID, guest_id);
 399
 400         hv_hypercall_pg = __vmalloc_node_range(PAGE_SIZE, 1, VMALLOC_START,
 401                         VMALLOC_END, GFP_KERNEL, PAGE_KERNEL_ROX,
 402                         VM_FLUSH_RESET_PERMS, NUMA_NO_NODE,
 403                         __builtin_return_address(0));
 404         if (hv_hypercall_pg == NULL) {
 405                 wrmsrl(HV_X64_MSR_GUEST_OS_ID, 0);
 406                 goto remove_cpuhp_state;
 407         }
 408
 409         rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 410         hypercall_msr.enable = 1;
 411         hypercall_msr.guest_physical_address = vmalloc_to_pfn(hv_hypercall_pg);
 412         wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 413
 414         /*
 415          * hyperv_init() is called before LAPIC is initialized: see
 416          * apic_intr_mode_init() -> x86_platform.apic_post_init() and
 417          * apic_bsp_setup() -> setup_local_APIC(). The direct-mode STIMER
 418          * depends on LAPIC, so hv_stimer_alloc() should be called from
 419          * x86_init.timers.setup_percpu_clockev.
 420          */
 421         old_setup_percpu_clockev = x86_init.timers.setup_percpu_clockev;
 422         x86_init.timers.setup_percpu_clockev = hv_stimer_setup_percpu_clockev;
 423
 424         hv_apic_init();
 425
 426         x86_init.pci.arch_init = hv_pci_init;
 427
 428         register_syscore_ops(&hv_syscore_ops);
 429
 430         hyperv_init_cpuhp = cpuhp;
 431         return;
 432
 433 remove_cpuhp_state:
 434         cpuhp_remove_state(cpuhp);
 435 free_vp_assist_page:
 436         kfree(hv_vp_assist_page);
 437         hv_vp_assist_page = NULL;
 438 free_vp_index:
 439         kfree(hv_vp_index);
 440         hv_vp_index = NULL;
 441 }
 442
 443 /*
 444  * This routine is called before kexec/kdump, it does the required cleanup.
 445  */
 446 void hyperv_cleanup(void)
 447 {
 448         union hv_x64_msr_hypercall_contents hypercall_msr;
 449
 450         unregister_syscore_ops(&hv_syscore_ops);
 451
 452         /* Reset our OS id */
 453         wrmsrl(HV_X64_MSR_GUEST_OS_ID, 0);
 454
 455         /*
 456          * Reset hypercall page reference before reset the page,
 457          * let hypercall operations fail safely rather than
 458          * panic the kernel for using invalid hypercall page
 459          */
 460         hv_hypercall_pg = NULL;
 461
 462         /* Reset the hypercall page */
 463         hypercall_msr.as_uint64 = 0;
 464         wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 465
 466         /* Reset the TSC page */
 467         hypercall_msr.as_uint64 = 0;
 468         wrmsrl(HV_X64_MSR_REFERENCE_TSC, hypercall_msr.as_uint64);
 469 }
 470 EXPORT_SYMBOL_GPL(hyperv_cleanup);
 471
 472 void hyperv_report_panic(struct pt_regs *regs, long err, bool in_die)
 473 {
 474         static bool panic_reported;
 475         u64 guest_id;
 476
 477         if (in_die && !panic_on_oops)
 478                 return;
 479
 480         /*
 481          * We prefer to report panic on 'die' chain as we have proper
 482          * registers to report, but if we miss it (e.g. on BUG()) we need
 483          * to report it on 'panic'.
 484          */
 485         if (panic_reported)
 486                 return;
 487         panic_reported = true;
 488
 489         rdmsrl(HV_X64_MSR_GUEST_OS_ID, guest_id);
 490
 491         wrmsrl(HV_X64_MSR_CRASH_P0, err);
 492         wrmsrl(HV_X64_MSR_CRASH_P1, guest_id);
 493         wrmsrl(HV_X64_MSR_CRASH_P2, regs->ip);
 494         wrmsrl(HV_X64_MSR_CRASH_P3, regs->ax);
 495         wrmsrl(HV_X64_MSR_CRASH_P4, regs->sp);
 496
 497         /*
 498          * Let Hyper-V know there is crash data available
 499          */
 500         wrmsrl(HV_X64_MSR_CRASH_CTL, HV_CRASH_CTL_CRASH_NOTIFY);
 501 }
 502 EXPORT_SYMBOL_GPL(hyperv_report_panic);
 503
 504 /**
 505  * hyperv_report_panic_msg - report panic message to Hyper-V
 506  * @pa: physical address of the panic page containing the message
 507  * @size: size of the message in the page
 508  */
 509 void hyperv_report_panic_msg(phys_addr_t pa, size_t size)
 510 {
 511         /*
 512          * P3 to contain the physical address of the panic page & P4 to
 513          * contain the size of the panic data in that page. Rest of the
 514          * registers are no-op when the NOTIFY_MSG flag is set.
 515          */
 516         wrmsrl(HV_X64_MSR_CRASH_P0, 0);
 517         wrmsrl(HV_X64_MSR_CRASH_P1, 0);
 518         wrmsrl(HV_X64_MSR_CRASH_P2, 0);
 519         wrmsrl(HV_X64_MSR_CRASH_P3, pa);
 520         wrmsrl(HV_X64_MSR_CRASH_P4, size);
 521
 522         /*
 523          * Let Hyper-V know there is crash data available along with
 524          * the panic message.
 525          */
 526         wrmsrl(HV_X64_MSR_CRASH_CTL,
 527                (HV_CRASH_CTL_CRASH_NOTIFY | HV_CRASH_CTL_CRASH_NOTIFY_MSG));
 528 }
 529 EXPORT_SYMBOL_GPL(hyperv_report_panic_msg);
 530
 531 bool hv_is_hyperv_initialized(void)
 532 {
 533         union hv_x64_msr_hypercall_contents hypercall_msr;
 534
 535         /*
 536          * Ensure that we're really on Hyper-V, and not a KVM or Xen
 537          * emulation of Hyper-V
 538          */
 539         if (x86_hyper_type != X86_HYPER_MS_HYPERV)
 540                 return false;
 541
 542         /*
 543          * Verify that earlier initialization succeeded by checking
 544          * that the hypercall page is setup
 545          */
 546         hypercall_msr.as_uint64 = 0;
 547         rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 548
 549         return hypercall_msr.enable;
 550 }
 551 EXPORT_SYMBOL_GPL(hv_is_hyperv_initialized);
 552
 553 bool hv_is_hibernation_supported(void)
 554 {
 555         return acpi_sleep_state_supported(ACPI_STATE_S4);
 556 }
 557 EXPORT_SYMBOL_GPL(hv_is_hibernation_supported);