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