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