arch/x86/kernel/kvmclock.c

   1 /*  KVM paravirtual clock driver. A clocksource implementation
   2     Copyright (C) 2008 Glauber de Oliveira Costa, Red Hat Inc.
   3
   4     This program is free software; you can redistribute it and/or modify
   5     it under the terms of the GNU General Public License as published by
   6     the Free Software Foundation; either version 2 of the License, or
   7     (at your option) any later version.
   8
   9     This program is distributed in the hope that it will be useful,
  10     but WITHOUT ANY WARRANTY; without even the implied warranty of
  11     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12     GNU General Public License for more details.
  13
  14     You should have received a copy of the GNU General Public License
  15     along with this program; if not, write to the Free Software
  16     Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17 */
  18
  19 #include <linux/clocksource.h>
  20 #include <linux/kvm_para.h>
  21 #include <asm/pvclock.h>
  22 #include <asm/msr.h>
  23 #include <asm/apic.h>
  24 #include <linux/percpu.h>
  25 #include <linux/hardirq.h>
  26 #include <linux/sched.h>
  27 #include <linux/sched/clock.h>
  28 #include <linux/mm.h>
  29
  30 #include <asm/mem_encrypt.h>
  31 #include <asm/x86_init.h>
  32 #include <asm/reboot.h>
  33 #include <asm/kvmclock.h>
  34
  35 static int kvmclock __ro_after_init = 1;
  36 static int msr_kvm_system_time = MSR_KVM_SYSTEM_TIME;
  37 static int msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK;
  38 static u64 kvm_sched_clock_offset;
  39
  40 static int parse_no_kvmclock(char *arg)
  41 {
  42         kvmclock = 0;
  43         return 0;
  44 }
  45 early_param("no-kvmclock", parse_no_kvmclock);
  46
  47 /* Aligned to page sizes to match whats mapped via vsyscalls to userspace */
  48 #define HV_CLOCK_SIZE   (sizeof(struct pvclock_vsyscall_time_info) * NR_CPUS)
  49 #define WALL_CLOCK_SIZE (sizeof(struct pvclock_wall_clock))
  50
  51 static u8 hv_clock_mem[PAGE_ALIGN(HV_CLOCK_SIZE)] __aligned(PAGE_SIZE);
  52 static u8 wall_clock_mem[PAGE_ALIGN(WALL_CLOCK_SIZE)] __aligned(PAGE_SIZE);
  53
  54 /* The hypervisor will put information about time periodically here */
  55 static struct pvclock_vsyscall_time_info *hv_clock;
  56 static struct pvclock_wall_clock *wall_clock;
  57
  58 /*
  59  * The wallclock is the time of day when we booted. Since then, some time may
  60  * have elapsed since the hypervisor wrote the data. So we try to account for
  61  * that with system time
  62  */
  63 static void kvm_get_wallclock(struct timespec64 *now)
  64 {
  65         struct pvclock_vcpu_time_info *vcpu_time;
  66         int low, high;
  67         int cpu;
  68
  69         low = (int)slow_virt_to_phys(wall_clock);
  70         high = ((u64)slow_virt_to_phys(wall_clock) >> 32);
  71
  72         native_write_msr(msr_kvm_wall_clock, low, high);
  73
  74         cpu = get_cpu();
  75
  76         vcpu_time = &hv_clock[cpu].pvti;
  77         pvclock_read_wallclock(wall_clock, vcpu_time, now);
  78
  79         put_cpu();
  80 }
  81
  82 static int kvm_set_wallclock(const struct timespec64 *now)
  83 {
  84         return -ENODEV;
  85 }
  86
  87 static u64 kvm_clock_read(void)
  88 {
  89         struct pvclock_vcpu_time_info *src;
  90         u64 ret;
  91         int cpu;
  92
  93         preempt_disable_notrace();
  94         cpu = smp_processor_id();
  95         src = &hv_clock[cpu].pvti;
  96         ret = pvclock_clocksource_read(src);
  97         preempt_enable_notrace();
  98         return ret;
  99 }
 100
 101 static u64 kvm_clock_get_cycles(struct clocksource *cs)
 102 {
 103         return kvm_clock_read();
 104 }
 105
 106 static u64 kvm_sched_clock_read(void)
 107 {
 108         return kvm_clock_read() - kvm_sched_clock_offset;
 109 }
 110
 111 static inline void kvm_sched_clock_init(bool stable)
 112 {
 113         if (!stable) {
 114                 pv_time_ops.sched_clock = kvm_clock_read;
 115                 clear_sched_clock_stable();
 116                 return;
 117         }
 118
 119         kvm_sched_clock_offset = kvm_clock_read();
 120         pv_time_ops.sched_clock = kvm_sched_clock_read;
 121
 122         printk(KERN_INFO "kvm-clock: using sched offset of %llu cycles\n",
 123                         kvm_sched_clock_offset);
 124
 125         BUILD_BUG_ON(sizeof(kvm_sched_clock_offset) >
 126                  sizeof(((struct pvclock_vcpu_time_info *)NULL)->system_time));
 127 }
 128
 129 /*
 130  * If we don't do that, there is the possibility that the guest
 131  * will calibrate under heavy load - thus, getting a lower lpj -
 132  * and execute the delays themselves without load. This is wrong,
 133  * because no delay loop can finish beforehand.
 134  * Any heuristics is subject to fail, because ultimately, a large
 135  * poll of guests can be running and trouble each other. So we preset
 136  * lpj here
 137  */
 138 static unsigned long kvm_get_tsc_khz(void)
 139 {
 140         struct pvclock_vcpu_time_info *src;
 141         int cpu;
 142         unsigned long tsc_khz;
 143
 144         cpu = get_cpu();
 145         src = &hv_clock[cpu].pvti;
 146         tsc_khz = pvclock_tsc_khz(src);
 147         put_cpu();
 148         setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
 149         return tsc_khz;
 150 }
 151
 152 static void kvm_get_preset_lpj(void)
 153 {
 154         unsigned long khz;
 155         u64 lpj;
 156
 157         khz = kvm_get_tsc_khz();
 158
 159         lpj = ((u64)khz * 1000);
 160         do_div(lpj, HZ);
 161         preset_lpj = lpj;
 162 }
 163
 164 bool kvm_check_and_clear_guest_paused(void)
 165 {
 166         bool ret = false;
 167         struct pvclock_vcpu_time_info *src;
 168         int cpu = smp_processor_id();
 169
 170         if (!hv_clock)
 171                 return ret;
 172
 173         src = &hv_clock[cpu].pvti;
 174         if ((src->flags & PVCLOCK_GUEST_STOPPED) != 0) {
 175                 src->flags &= ~PVCLOCK_GUEST_STOPPED;
 176                 pvclock_touch_watchdogs();
 177                 ret = true;
 178         }
 179
 180         return ret;
 181 }
 182
 183 struct clocksource kvm_clock = {
 184         .name = "kvm-clock",
 185         .read = kvm_clock_get_cycles,
 186         .rating = 400,
 187         .mask = CLOCKSOURCE_MASK(64),
 188         .flags = CLOCK_SOURCE_IS_CONTINUOUS,
 189 };
 190 EXPORT_SYMBOL_GPL(kvm_clock);
 191
 192 int kvm_register_clock(char *txt)
 193 {
 194         int cpu = smp_processor_id();
 195         int low, high, ret;
 196         struct pvclock_vcpu_time_info *src;
 197
 198         if (!hv_clock)
 199                 return 0;
 200
 201         src = &hv_clock[cpu].pvti;
 202         low = (int)slow_virt_to_phys(src) | 1;
 203         high = ((u64)slow_virt_to_phys(src) >> 32);
 204         ret = native_write_msr_safe(msr_kvm_system_time, low, high);
 205         printk(KERN_INFO "kvm-clock: cpu %d, msr %x:%x, %s\n",
 206                cpu, high, low, txt);
 207
 208         return ret;
 209 }
 210
 211 static void kvm_save_sched_clock_state(void)
 212 {
 213 }
 214
 215 static void kvm_restore_sched_clock_state(void)
 216 {
 217         kvm_register_clock("primary cpu clock, resume");
 218 }
 219
 220 #ifdef CONFIG_X86_LOCAL_APIC
 221 static void kvm_setup_secondary_clock(void)
 222 {
 223         /*
 224          * Now that the first cpu already had this clocksource initialized,
 225          * we shouldn't fail.
 226          */
 227         WARN_ON(kvm_register_clock("secondary cpu clock"));
 228 }
 229 #endif
 230
 231 /*
 232  * After the clock is registered, the host will keep writing to the
 233  * registered memory location. If the guest happens to shutdown, this memory
 234  * won't be valid. In cases like kexec, in which you install a new kernel, this
 235  * means a random memory location will be kept being written. So before any
 236  * kind of shutdown from our side, we unregister the clock by writing anything
 237  * that does not have the 'enable' bit set in the msr
 238  */
 239 #ifdef CONFIG_KEXEC_CORE
 240 static void kvm_crash_shutdown(struct pt_regs *regs)
 241 {
 242         native_write_msr(msr_kvm_system_time, 0, 0);
 243         kvm_disable_steal_time();
 244         native_machine_crash_shutdown(regs);
 245 }
 246 #endif
 247
 248 static void kvm_shutdown(void)
 249 {
 250         native_write_msr(msr_kvm_system_time, 0, 0);
 251         kvm_disable_steal_time();
 252         native_machine_shutdown();
 253 }
 254
 255 void __init kvmclock_init(void)
 256 {
 257         struct pvclock_vcpu_time_info *vcpu_time;
 258         int cpu;
 259         u8 flags;
 260
 261         if (!kvm_para_available())
 262                 return;
 263
 264         if (kvmclock && kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) {
 265                 msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW;
 266                 msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW;
 267         } else if (!(kvmclock && kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)))
 268                 return;
 269
 270         wall_clock = (struct pvclock_wall_clock *)wall_clock_mem;
 271         hv_clock = (struct pvclock_vsyscall_time_info *)hv_clock_mem;
 272
 273         if (kvm_register_clock("primary cpu clock")) {
 274                 hv_clock = NULL;
 275                 wall_clock = NULL;
 276                 return;
 277         }
 278
 279         printk(KERN_INFO "kvm-clock: Using msrs %x and %x",
 280                 msr_kvm_system_time, msr_kvm_wall_clock);
 281
 282         pvclock_set_pvti_cpu0_va(hv_clock);
 283
 284         if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT))
 285                 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
 286
 287         cpu = get_cpu();
 288         vcpu_time = &hv_clock[cpu].pvti;
 289         flags = pvclock_read_flags(vcpu_time);
 290
 291         kvm_sched_clock_init(flags & PVCLOCK_TSC_STABLE_BIT);
 292         put_cpu();
 293
 294         x86_platform.calibrate_tsc = kvm_get_tsc_khz;
 295         x86_platform.calibrate_cpu = kvm_get_tsc_khz;
 296         x86_platform.get_wallclock = kvm_get_wallclock;
 297         x86_platform.set_wallclock = kvm_set_wallclock;
 298 #ifdef CONFIG_X86_LOCAL_APIC
 299         x86_cpuinit.early_percpu_clock_init =
 300                 kvm_setup_secondary_clock;
 301 #endif
 302         x86_platform.save_sched_clock_state = kvm_save_sched_clock_state;
 303         x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state;
 304         machine_ops.shutdown  = kvm_shutdown;
 305 #ifdef CONFIG_KEXEC_CORE
 306         machine_ops.crash_shutdown  = kvm_crash_shutdown;
 307 #endif
 308         kvm_get_preset_lpj();
 309         clocksource_register_hz(&kvm_clock, NSEC_PER_SEC);
 310         pv_info.name = "KVM";
 311 }
 312
 313 int __init kvm_setup_vsyscall_timeinfo(void)
 314 {
 315 #ifdef CONFIG_X86_64
 316         int cpu;
 317         u8 flags;
 318         struct pvclock_vcpu_time_info *vcpu_time;
 319
 320         if (!hv_clock)
 321                 return 0;
 322
 323         cpu = get_cpu();
 324
 325         vcpu_time = &hv_clock[cpu].pvti;
 326         flags = pvclock_read_flags(vcpu_time);
 327
 328         put_cpu();
 329
 330         if (!(flags & PVCLOCK_TSC_STABLE_BIT))
 331                 return 1;
 332
 333         kvm_clock.archdata.vclock_mode = VCLOCK_PVCLOCK;
 334 #endif
 335         return 0;
 336 }