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/reboot.h>
  24 #include <asm/kvmclock.h>
  25
  26 static int kvmclock __initdata = 1;
  27 static int kvmclock_vsyscall __initdata = 1;
  28 static int msr_kvm_system_time __ro_after_init = MSR_KVM_SYSTEM_TIME;
  29 static int msr_kvm_wall_clock __ro_after_init = MSR_KVM_WALL_CLOCK;
  30 static u64 kvm_sched_clock_offset __ro_after_init;
  31
  32 static int __init parse_no_kvmclock(char *arg)
  33 {
  34         kvmclock = 0;
  35         return 0;
  36 }
  37 early_param("no-kvmclock", parse_no_kvmclock);
  38
  39 static int __init parse_no_kvmclock_vsyscall(char *arg)
  40 {
  41         kvmclock_vsyscall = 0;
  42         return 0;
  43 }
  44 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
  45
  46 /* Aligned to page sizes to match whats mapped via vsyscalls to userspace */
  47 #define HVC_BOOT_ARRAY_SIZE \
  48         (PAGE_SIZE / sizeof(struct pvclock_vsyscall_time_info))
  49
  50 static struct pvclock_vsyscall_time_info
  51                         hv_clock_boot[HVC_BOOT_ARRAY_SIZE] __bss_decrypted __aligned(PAGE_SIZE);
  52 static struct pvclock_wall_clock wall_clock __bss_decrypted;
  53 static DEFINE_PER_CPU(struct pvclock_vsyscall_time_info *, hv_clock_per_cpu);
  54 static struct pvclock_vsyscall_time_info *hvclock_mem;
  55
  56 static inline struct pvclock_vcpu_time_info *this_cpu_pvti(void)
  57 {
  58         return &this_cpu_read(hv_clock_per_cpu)->pvti;
  59 }
  60
  61 static inline struct pvclock_vsyscall_time_info *this_cpu_hvclock(void)
  62 {
  63         return this_cpu_read(hv_clock_per_cpu);
  64 }
  65
  66 /*
  67  * The wallclock is the time of day when we booted. Since then, some time may
  68  * have elapsed since the hypervisor wrote the data. So we try to account for
  69  * that with system time
  70  */
  71 static void kvm_get_wallclock(struct timespec64 *now)
  72 {
  73         wrmsrl(msr_kvm_wall_clock, slow_virt_to_phys(&wall_clock));
  74         preempt_disable();
  75         pvclock_read_wallclock(&wall_clock, this_cpu_pvti(), now);
  76         preempt_enable();
  77 }
  78
  79 static int kvm_set_wallclock(const struct timespec64 *now)
  80 {
  81         return -ENODEV;
  82 }
  83
  84 static u64 kvm_clock_read(void)
  85 {
  86         u64 ret;
  87
  88         preempt_disable_notrace();
  89         ret = pvclock_clocksource_read(this_cpu_pvti());
  90         preempt_enable_notrace();
  91         return ret;
  92 }
  93
  94 static u64 kvm_clock_get_cycles(struct clocksource *cs)
  95 {
  96         return kvm_clock_read();
  97 }
  98
  99 static u64 kvm_sched_clock_read(void)
 100 {
 101         return kvm_clock_read() - kvm_sched_clock_offset;
 102 }
 103
 104 static inline void kvm_sched_clock_init(bool stable)
 105 {
 106         if (!stable)
 107                 clear_sched_clock_stable();
 108         kvm_sched_clock_offset = kvm_clock_read();
 109         pv_ops.time.sched_clock = kvm_sched_clock_read;
 110
 111         pr_info("kvm-clock: using sched offset of %llu cycles",
 112                 kvm_sched_clock_offset);
 113
 114         BUILD_BUG_ON(sizeof(kvm_sched_clock_offset) >
 115                 sizeof(((struct pvclock_vcpu_time_info *)NULL)->system_time));
 116 }
 117
 118 /*
 119  * If we don't do that, there is the possibility that the guest
 120  * will calibrate under heavy load - thus, getting a lower lpj -
 121  * and execute the delays themselves without load. This is wrong,
 122  * because no delay loop can finish beforehand.
 123  * Any heuristics is subject to fail, because ultimately, a large
 124  * poll of guests can be running and trouble each other. So we preset
 125  * lpj here
 126  */
 127 static unsigned long kvm_get_tsc_khz(void)
 128 {
 129         setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
 130         return pvclock_tsc_khz(this_cpu_pvti());
 131 }
 132
 133 static void __init kvm_get_preset_lpj(void)
 134 {
 135         unsigned long khz;
 136         u64 lpj;
 137
 138         khz = kvm_get_tsc_khz();
 139
 140         lpj = ((u64)khz * 1000);
 141         do_div(lpj, HZ);
 142         preset_lpj = lpj;
 143 }
 144
 145 bool kvm_check_and_clear_guest_paused(void)
 146 {
 147         struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
 148         bool ret = false;
 149
 150         if (!src)
 151                 return ret;
 152
 153         if ((src->pvti.flags & PVCLOCK_GUEST_STOPPED) != 0) {
 154                 src->pvti.flags &= ~PVCLOCK_GUEST_STOPPED;
 155                 pvclock_touch_watchdogs();
 156                 ret = true;
 157         }
 158         return ret;
 159 }
 160
 161 static int kvm_cs_enable(struct clocksource *cs)
 162 {
 163         vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK);
 164         return 0;
 165 }
 166
 167 struct clocksource kvm_clock = {
 168         .name   = "kvm-clock",
 169         .read   = kvm_clock_get_cycles,
 170         .rating = 400,
 171         .mask   = CLOCKSOURCE_MASK(64),
 172         .flags  = CLOCK_SOURCE_IS_CONTINUOUS,
 173         .enable = kvm_cs_enable,
 174 };
 175 EXPORT_SYMBOL_GPL(kvm_clock);
 176
 177 static void kvm_register_clock(char *txt)
 178 {
 179         struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
 180         u64 pa;
 181
 182         if (!src)
 183                 return;
 184
 185         pa = slow_virt_to_phys(&src->pvti) | 0x01ULL;
 186         wrmsrl(msr_kvm_system_time, pa);
 187         pr_info("kvm-clock: cpu %d, msr %llx, %s", smp_processor_id(), pa, txt);
 188 }
 189
 190 static void kvm_save_sched_clock_state(void)
 191 {
 192 }
 193
 194 static void kvm_restore_sched_clock_state(void)
 195 {
 196         kvm_register_clock("primary cpu clock, resume");
 197 }
 198
 199 #ifdef CONFIG_X86_LOCAL_APIC
 200 static void kvm_setup_secondary_clock(void)
 201 {
 202         kvm_register_clock("secondary cpu clock");
 203 }
 204 #endif
 205
 206 /*
 207  * After the clock is registered, the host will keep writing to the
 208  * registered memory location. If the guest happens to shutdown, this memory
 209  * won't be valid. In cases like kexec, in which you install a new kernel, this
 210  * means a random memory location will be kept being written. So before any
 211  * kind of shutdown from our side, we unregister the clock by writing anything
 212  * that does not have the 'enable' bit set in the msr
 213  */
 214 #ifdef CONFIG_KEXEC_CORE
 215 static void kvm_crash_shutdown(struct pt_regs *regs)
 216 {
 217         native_write_msr(msr_kvm_system_time, 0, 0);
 218         kvm_disable_steal_time();
 219         native_machine_crash_shutdown(regs);
 220 }
 221 #endif
 222
 223 static void kvm_shutdown(void)
 224 {
 225         native_write_msr(msr_kvm_system_time, 0, 0);
 226         kvm_disable_steal_time();
 227         native_machine_shutdown();
 228 }
 229
 230 static void __init kvmclock_init_mem(void)
 231 {
 232         unsigned long ncpus;
 233         unsigned int order;
 234         struct page *p;
 235         int r;
 236
 237         if (HVC_BOOT_ARRAY_SIZE >= num_possible_cpus())
 238                 return;
 239
 240         ncpus = num_possible_cpus() - HVC_BOOT_ARRAY_SIZE;
 241         order = get_order(ncpus * sizeof(*hvclock_mem));
 242
 243         p = alloc_pages(GFP_KERNEL, order);
 244         if (!p) {
 245                 pr_warn("%s: failed to alloc %d pages", __func__, (1U << order));
 246                 return;
 247         }
 248
 249         hvclock_mem = page_address(p);
 250
 251         /*
 252          * hvclock is shared between the guest and the hypervisor, must
 253          * be mapped decrypted.
 254          */
 255         if (sev_active()) {
 256                 r = set_memory_decrypted((unsigned long) hvclock_mem,
 257                                          1UL << order);
 258                 if (r) {
 259                         __free_pages(p, order);
 260                         hvclock_mem = NULL;
 261                         pr_warn("kvmclock: set_memory_decrypted() failed. Disabling\n");
 262                         return;
 263                 }
 264         }
 265
 266         memset(hvclock_mem, 0, PAGE_SIZE << order);
 267 }
 268
 269 static int __init kvm_setup_vsyscall_timeinfo(void)
 270 {
 271 #ifdef CONFIG_X86_64
 272         u8 flags;
 273
 274         if (!per_cpu(hv_clock_per_cpu, 0) || !kvmclock_vsyscall)
 275                 return 0;
 276
 277         flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
 278         if (!(flags & PVCLOCK_TSC_STABLE_BIT))
 279                 return 0;
 280
 281         kvm_clock.vdso_clock_mode = VDSO_CLOCKMODE_PVCLOCK;
 282 #endif
 283
 284         kvmclock_init_mem();
 285
 286         return 0;
 287 }
 288 early_initcall(kvm_setup_vsyscall_timeinfo);
 289
 290 static int kvmclock_setup_percpu(unsigned int cpu)
 291 {
 292         struct pvclock_vsyscall_time_info *p = per_cpu(hv_clock_per_cpu, cpu);
 293
 294         /*
 295          * The per cpu area setup replicates CPU0 data to all cpu
 296          * pointers. So carefully check. CPU0 has been set up in init
 297          * already.
 298          */
 299         if (!cpu || (p && p != per_cpu(hv_clock_per_cpu, 0)))
 300                 return 0;
 301
 302         /* Use the static page for the first CPUs, allocate otherwise */
 303         if (cpu < HVC_BOOT_ARRAY_SIZE)
 304                 p = &hv_clock_boot[cpu];
 305         else if (hvclock_mem)
 306                 p = hvclock_mem + cpu - HVC_BOOT_ARRAY_SIZE;
 307         else
 308                 return -ENOMEM;
 309
 310         per_cpu(hv_clock_per_cpu, cpu) = p;
 311         return p ? 0 : -ENOMEM;
 312 }
 313
 314 void __init kvmclock_init(void)
 315 {
 316         u8 flags;
 317
 318         if (!kvm_para_available() || !kvmclock)
 319                 return;
 320
 321         if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) {
 322                 msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW;
 323                 msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW;
 324         } else if (!kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)) {
 325                 return;
 326         }
 327
 328         if (cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "kvmclock:setup_percpu",
 329                               kvmclock_setup_percpu, NULL) < 0) {
 330                 return;
 331         }
 332
 333         pr_info("kvm-clock: Using msrs %x and %x",
 334                 msr_kvm_system_time, msr_kvm_wall_clock);
 335
 336         this_cpu_write(hv_clock_per_cpu, &hv_clock_boot[0]);
 337         kvm_register_clock("primary cpu clock");
 338         pvclock_set_pvti_cpu0_va(hv_clock_boot);
 339
 340         if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT))
 341                 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
 342
 343         flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
 344         kvm_sched_clock_init(flags & PVCLOCK_TSC_STABLE_BIT);
 345
 346         x86_platform.calibrate_tsc = kvm_get_tsc_khz;
 347         x86_platform.calibrate_cpu = kvm_get_tsc_khz;
 348         x86_platform.get_wallclock = kvm_get_wallclock;
 349         x86_platform.set_wallclock = kvm_set_wallclock;
 350 #ifdef CONFIG_X86_LOCAL_APIC
 351         x86_cpuinit.early_percpu_clock_init = kvm_setup_secondary_clock;
 352 #endif
 353         x86_platform.save_sched_clock_state = kvm_save_sched_clock_state;
 354         x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state;
 355         machine_ops.shutdown  = kvm_shutdown;
 356 #ifdef CONFIG_KEXEC_CORE
 357         machine_ops.crash_shutdown  = kvm_crash_shutdown;
 358 #endif
 359         kvm_get_preset_lpj();
 360
 361         /*
 362          * X86_FEATURE_NONSTOP_TSC is TSC runs at constant rate
 363          * with P/T states and does not stop in deep C-states.
 364          *
 365          * Invariant TSC exposed by host means kvmclock is not necessary:
 366          * can use TSC as clocksource.
 367          *
 368          */
 369         if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) &&
 370             boot_cpu_has(X86_FEATURE_NONSTOP_TSC) &&
 371             !check_tsc_unstable())
 372                 kvm_clock.rating = 299;
 373
 374         clocksource_register_hz(&kvm_clock, NSEC_PER_SEC);
 375         pv_info.name = "KVM";
 376 }