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