arch/x86/kernel/kvmclock.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*  KVM paravirtual clock driver. A clocksource implementation
   3     Copyright (C) 2008 Glauber de Oliveira Costa, Red Hat Inc.
   4 */
   5
   6 #include <linux/clocksource.h>
   7 #include <linux/kvm_para.h>
   8 #include <asm/pvclock.h>
   9 #include <asm/msr.h>
  10 #include <asm/apic.h>
  11 #include <linux/percpu.h>
  12 #include <linux/hardirq.h>
  13 #include <linux/cpuhotplug.h>
  14 #include <linux/sched.h>
  15 #include <linux/sched/clock.h>
  16 #include <linux/mm.h>
  17 #include <linux/slab.h>
  18 #include <linux/set_memory.h>
  19
  20 #include <asm/hypervisor.h>
  21 #include <asm/mem_encrypt.h>
  22 #include <asm/x86_init.h>
  23 #include <asm/kvmclock.h>
  24
  25 static int kvmclock __initdata = 1;
  26 static int kvmclock_vsyscall __initdata = 1;
  27 static int msr_kvm_system_time __ro_after_init = MSR_KVM_SYSTEM_TIME;
  28 static int msr_kvm_wall_clock __ro_after_init = MSR_KVM_WALL_CLOCK;
  29 static u64 kvm_sched_clock_offset __ro_after_init;
  30
  31 static int __init parse_no_kvmclock(char *arg)
  32 {
  33         kvmclock = 0;
  34         return 0;
  35 }
  36 early_param("no-kvmclock", parse_no_kvmclock);
  37
  38 static int __init parse_no_kvmclock_vsyscall(char *arg)
  39 {
  40         kvmclock_vsyscall = 0;
  41         return 0;
  42 }
  43 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
  44
  45 /* Aligned to page sizes to match whats mapped via vsyscalls to userspace */
  46 #define HVC_BOOT_ARRAY_SIZE \
  47         (PAGE_SIZE / sizeof(struct pvclock_vsyscall_time_info))
  48
  49 static struct pvclock_vsyscall_time_info
  50                         hv_clock_boot[HVC_BOOT_ARRAY_SIZE] __bss_decrypted __aligned(PAGE_SIZE);
  51 static struct pvclock_wall_clock wall_clock __bss_decrypted;
  52 static DEFINE_PER_CPU(struct pvclock_vsyscall_time_info *, hv_clock_per_cpu);
  53 static struct pvclock_vsyscall_time_info *hvclock_mem;
  54
  55 static inline struct pvclock_vcpu_time_info *this_cpu_pvti(void)
  56 {
  57         return &this_cpu_read(hv_clock_per_cpu)->pvti;
  58 }
  59
  60 static inline struct pvclock_vsyscall_time_info *this_cpu_hvclock(void)
  61 {
  62         return this_cpu_read(hv_clock_per_cpu);
  63 }
  64
  65 /*
  66  * The wallclock is the time of day when we booted. Since then, some time may
  67  * have elapsed since the hypervisor wrote the data. So we try to account for
  68  * that with system time
  69  */
  70 static void kvm_get_wallclock(struct timespec64 *now)
  71 {
  72         wrmsrl(msr_kvm_wall_clock, slow_virt_to_phys(&wall_clock));
  73         preempt_disable();
  74         pvclock_read_wallclock(&wall_clock, this_cpu_pvti(), now);
  75         preempt_enable();
  76 }
  77
  78 static int kvm_set_wallclock(const struct timespec64 *now)
  79 {
  80         return -ENODEV;
  81 }
  82
  83 static u64 kvm_clock_read(void)
  84 {
  85         u64 ret;
  86
  87         preempt_disable_notrace();
  88         ret = pvclock_clocksource_read(this_cpu_pvti());
  89         preempt_enable_notrace();
  90         return ret;
  91 }
  92
  93 static u64 kvm_clock_get_cycles(struct clocksource *cs)
  94 {
  95         return kvm_clock_read();
  96 }
  97
  98 static u64 kvm_sched_clock_read(void)
  99 {
 100         return kvm_clock_read() - kvm_sched_clock_offset;
 101 }
 102
 103 static inline void kvm_sched_clock_init(bool stable)
 104 {
 105         if (!stable)
 106                 clear_sched_clock_stable();
 107         kvm_sched_clock_offset = kvm_clock_read();
 108         paravirt_set_sched_clock(kvm_sched_clock_read);
 109
 110         pr_info("kvm-clock: using sched offset of %llu cycles",
 111                 kvm_sched_clock_offset);
 112
 113         BUILD_BUG_ON(sizeof(kvm_sched_clock_offset) >
 114                 sizeof(((struct pvclock_vcpu_time_info *)NULL)->system_time));
 115 }
 116
 117 /*
 118  * If we don't do that, there is the possibility that the guest
 119  * will calibrate under heavy load - thus, getting a lower lpj -
 120  * and execute the delays themselves without load. This is wrong,
 121  * because no delay loop can finish beforehand.
 122  * Any heuristics is subject to fail, because ultimately, a large
 123  * poll of guests can be running and trouble each other. So we preset
 124  * lpj here
 125  */
 126 static unsigned long kvm_get_tsc_khz(void)
 127 {
 128         setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
 129         return pvclock_tsc_khz(this_cpu_pvti());
 130 }
 131
 132 static void __init kvm_get_preset_lpj(void)
 133 {
 134         unsigned long khz;
 135         u64 lpj;
 136
 137         khz = kvm_get_tsc_khz();
 138
 139         lpj = ((u64)khz * 1000);
 140         do_div(lpj, HZ);
 141         preset_lpj = lpj;
 142 }
 143
 144 bool kvm_check_and_clear_guest_paused(void)
 145 {
 146         struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
 147         bool ret = false;
 148
 149         if (!src)
 150                 return ret;
 151
 152         if ((src->pvti.flags & PVCLOCK_GUEST_STOPPED) != 0) {
 153                 src->pvti.flags &= ~PVCLOCK_GUEST_STOPPED;
 154                 pvclock_touch_watchdogs();
 155                 ret = true;
 156         }
 157         return ret;
 158 }
 159
 160 static int kvm_cs_enable(struct clocksource *cs)
 161 {
 162         vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK);
 163         return 0;
 164 }
 165
 166 struct clocksource kvm_clock = {
 167         .name   = "kvm-clock",
 168         .read   = kvm_clock_get_cycles,
 169         .rating = 400,
 170         .mask   = CLOCKSOURCE_MASK(64),
 171         .flags  = CLOCK_SOURCE_IS_CONTINUOUS,
 172         .enable = kvm_cs_enable,
 173 };
 174 EXPORT_SYMBOL_GPL(kvm_clock);
 175
 176 static void kvm_register_clock(char *txt)
 177 {
 178         struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
 179         u64 pa;
 180
 181         if (!src)
 182                 return;
 183
 184         pa = slow_virt_to_phys(&src->pvti) | 0x01ULL;
 185         wrmsrl(msr_kvm_system_time, pa);
 186         pr_info("kvm-clock: cpu %d, msr %llx, %s", smp_processor_id(), pa, txt);
 187 }
 188
 189 static void kvm_save_sched_clock_state(void)
 190 {
 191 }
 192
 193 static void kvm_restore_sched_clock_state(void)
 194 {
 195         kvm_register_clock("primary cpu clock, resume");
 196 }
 197
 198 #ifdef CONFIG_X86_LOCAL_APIC
 199 static void kvm_setup_secondary_clock(void)
 200 {
 201         kvm_register_clock("secondary cpu clock");
 202 }
 203 #endif
 204
 205 void kvmclock_disable(void)
 206 {
 207         native_write_msr(msr_kvm_system_time, 0, 0);
 208 }
 209
 210 static void __init kvmclock_init_mem(void)
 211 {
 212         unsigned long ncpus;
 213         unsigned int order;
 214         struct page *p;
 215         int r;
 216
 217         if (HVC_BOOT_ARRAY_SIZE >= num_possible_cpus())
 218                 return;
 219
 220         ncpus = num_possible_cpus() - HVC_BOOT_ARRAY_SIZE;
 221         order = get_order(ncpus * sizeof(*hvclock_mem));
 222
 223         p = alloc_pages(GFP_KERNEL, order);
 224         if (!p) {
 225                 pr_warn("%s: failed to alloc %d pages", __func__, (1U << order));
 226                 return;
 227         }
 228
 229         hvclock_mem = page_address(p);
 230
 231         /*
 232          * hvclock is shared between the guest and the hypervisor, must
 233          * be mapped decrypted.
 234          */
 235         if (sev_active()) {
 236                 r = set_memory_decrypted((unsigned long) hvclock_mem,
 237                                          1UL << order);
 238                 if (r) {
 239                         __free_pages(p, order);
 240                         hvclock_mem = NULL;
 241                         pr_warn("kvmclock: set_memory_decrypted() failed. Disabling\n");
 242                         return;
 243                 }
 244         }
 245
 246         memset(hvclock_mem, 0, PAGE_SIZE << order);
 247 }
 248
 249 static int __init kvm_setup_vsyscall_timeinfo(void)
 250 {
 251         kvmclock_init_mem();
 252
 253 #ifdef CONFIG_X86_64
 254         if (per_cpu(hv_clock_per_cpu, 0) && kvmclock_vsyscall) {
 255                 u8 flags;
 256
 257                 flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
 258                 if (!(flags & PVCLOCK_TSC_STABLE_BIT))
 259                         return 0;
 260
 261                 kvm_clock.vdso_clock_mode = VDSO_CLOCKMODE_PVCLOCK;
 262         }
 263 #endif
 264
 265         return 0;
 266 }
 267 early_initcall(kvm_setup_vsyscall_timeinfo);
 268
 269 static int kvmclock_setup_percpu(unsigned int cpu)
 270 {
 271         struct pvclock_vsyscall_time_info *p = per_cpu(hv_clock_per_cpu, cpu);
 272
 273         /*
 274          * The per cpu area setup replicates CPU0 data to all cpu
 275          * pointers. So carefully check. CPU0 has been set up in init
 276          * already.
 277          */
 278         if (!cpu || (p && p != per_cpu(hv_clock_per_cpu, 0)))
 279                 return 0;
 280
 281         /* Use the static page for the first CPUs, allocate otherwise */
 282         if (cpu < HVC_BOOT_ARRAY_SIZE)
 283                 p = &hv_clock_boot[cpu];
 284         else if (hvclock_mem)
 285                 p = hvclock_mem + cpu - HVC_BOOT_ARRAY_SIZE;
 286         else
 287                 return -ENOMEM;
 288
 289         per_cpu(hv_clock_per_cpu, cpu) = p;
 290         return p ? 0 : -ENOMEM;
 291 }
 292
 293 void __init kvmclock_init(void)
 294 {
 295         u8 flags;
 296
 297         if (!kvm_para_available() || !kvmclock)
 298                 return;
 299
 300         if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) {
 301                 msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW;
 302                 msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW;
 303         } else if (!kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)) {
 304                 return;
 305         }
 306
 307         if (cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "kvmclock:setup_percpu",
 308                               kvmclock_setup_percpu, NULL) < 0) {
 309                 return;
 310         }
 311
 312         pr_info("kvm-clock: Using msrs %x and %x",
 313                 msr_kvm_system_time, msr_kvm_wall_clock);
 314
 315         this_cpu_write(hv_clock_per_cpu, &hv_clock_boot[0]);
 316         kvm_register_clock("primary cpu clock");
 317         pvclock_set_pvti_cpu0_va(hv_clock_boot);
 318
 319         if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT))
 320                 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
 321
 322         flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
 323         kvm_sched_clock_init(flags & PVCLOCK_TSC_STABLE_BIT);
 324
 325         x86_platform.calibrate_tsc = kvm_get_tsc_khz;
 326         x86_platform.calibrate_cpu = kvm_get_tsc_khz;
 327         x86_platform.get_wallclock = kvm_get_wallclock;
 328         x86_platform.set_wallclock = kvm_set_wallclock;
 329 #ifdef CONFIG_X86_LOCAL_APIC
 330         x86_cpuinit.early_percpu_clock_init = kvm_setup_secondary_clock;
 331 #endif
 332         x86_platform.save_sched_clock_state = kvm_save_sched_clock_state;
 333         x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state;
 334         kvm_get_preset_lpj();
 335
 336         /*
 337          * X86_FEATURE_NONSTOP_TSC is TSC runs at constant rate
 338          * with P/T states and does not stop in deep C-states.
 339          *
 340          * Invariant TSC exposed by host means kvmclock is not necessary:
 341          * can use TSC as clocksource.
 342          *
 343          */
 344         if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) &&
 345             boot_cpu_has(X86_FEATURE_NONSTOP_TSC) &&
 346             !check_tsc_unstable())
 347                 kvm_clock.rating = 299;
 348
 349         clocksource_register_hz(&kvm_clock, NSEC_PER_SEC);
 350         pv_info.name = "KVM";
 351 }