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