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/cpuhotplug.h>
  27 #include <linux/sched.h>
  28 #include <linux/sched/clock.h>
  29 #include <linux/mm.h>
  30 #include <linux/slab.h>
  31 #include <linux/set_memory.h>
  32
  33 #include <asm/hypervisor.h>
  34 #include <asm/mem_encrypt.h>
  35 #include <asm/x86_init.h>
  36 #include <asm/reboot.h>
  37 #include <asm/kvmclock.h>
  38
  39 static int kvmclock __initdata = 1;
  40 static int kvmclock_vsyscall __initdata = 1;
  41 static int msr_kvm_system_time __ro_after_init = MSR_KVM_SYSTEM_TIME;
  42 static int msr_kvm_wall_clock __ro_after_init = MSR_KVM_WALL_CLOCK;
  43 static u64 kvm_sched_clock_offset __ro_after_init;
  44
  45 static int __init parse_no_kvmclock(char *arg)
  46 {
  47         kvmclock = 0;
  48         return 0;
  49 }
  50 early_param("no-kvmclock", parse_no_kvmclock);
  51
  52 static int __init parse_no_kvmclock_vsyscall(char *arg)
  53 {
  54         kvmclock_vsyscall = 0;
  55         return 0;
  56 }
  57 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
  58
  59 /* Aligned to page sizes to match whats mapped via vsyscalls to userspace */
  60 #define HV_CLOCK_SIZE   (sizeof(struct pvclock_vsyscall_time_info) * NR_CPUS)
  61 #define HVC_BOOT_ARRAY_SIZE \
  62         (PAGE_SIZE / sizeof(struct pvclock_vsyscall_time_info))
  63
  64 static struct pvclock_vsyscall_time_info
  65                         hv_clock_boot[HVC_BOOT_ARRAY_SIZE] __bss_decrypted __aligned(PAGE_SIZE);
  66 static struct pvclock_wall_clock wall_clock __bss_decrypted;
  67 static struct pvclock_vsyscall_time_info *hvclock_mem;
  68 DEFINE_PER_CPU(struct pvclock_vsyscall_time_info *, hv_clock_per_cpu);
  69 EXPORT_PER_CPU_SYMBOL_GPL(hv_clock_per_cpu);
  70
  71 /*
  72  * The wallclock is the time of day when we booted. Since then, some time may
  73  * have elapsed since the hypervisor wrote the data. So we try to account for
  74  * that with system time
  75  */
  76 static void kvm_get_wallclock(struct timespec64 *now)
  77 {
  78         wrmsrl(msr_kvm_wall_clock, slow_virt_to_phys(&wall_clock));
  79         preempt_disable();
  80         pvclock_read_wallclock(&wall_clock, this_cpu_pvti(), now);
  81         preempt_enable();
  82 }
  83
  84 static int kvm_set_wallclock(const struct timespec64 *now)
  85 {
  86         return -ENODEV;
  87 }
  88
  89 static u64 kvm_clock_read(void)
  90 {
  91         u64 ret;
  92
  93         preempt_disable_notrace();
  94         ret = pvclock_clocksource_read(this_cpu_pvti());
  95         preempt_enable_notrace();
  96         return ret;
  97 }
  98
  99 static u64 kvm_clock_get_cycles(struct clocksource *cs)
 100 {
 101         return kvm_clock_read();
 102 }
 103
 104 static u64 kvm_sched_clock_read(void)
 105 {
 106         return kvm_clock_read() - kvm_sched_clock_offset;
 107 }
 108
 109 static inline void kvm_sched_clock_init(bool stable)
 110 {
 111         if (!stable)
 112                 clear_sched_clock_stable();
 113         kvm_sched_clock_offset = kvm_clock_read();
 114         pv_time_ops.sched_clock = kvm_sched_clock_read;
 115
 116         pr_info("kvm-clock: using sched offset of %llu cycles",
 117                 kvm_sched_clock_offset);
 118
 119         BUILD_BUG_ON(sizeof(kvm_sched_clock_offset) >
 120                 sizeof(((struct pvclock_vcpu_time_info *)NULL)->system_time));
 121 }
 122
 123 /*
 124  * If we don't do that, there is the possibility that the guest
 125  * will calibrate under heavy load - thus, getting a lower lpj -
 126  * and execute the delays themselves without load. This is wrong,
 127  * because no delay loop can finish beforehand.
 128  * Any heuristics is subject to fail, because ultimately, a large
 129  * poll of guests can be running and trouble each other. So we preset
 130  * lpj here
 131  */
 132 static unsigned long kvm_get_tsc_khz(void)
 133 {
 134         setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
 135         return pvclock_tsc_khz(this_cpu_pvti());
 136 }
 137
 138 static void __init kvm_get_preset_lpj(void)
 139 {
 140         unsigned long khz;
 141         u64 lpj;
 142
 143         khz = kvm_get_tsc_khz();
 144
 145         lpj = ((u64)khz * 1000);
 146         do_div(lpj, HZ);
 147         preset_lpj = lpj;
 148 }
 149
 150 bool kvm_check_and_clear_guest_paused(void)
 151 {
 152         struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
 153         bool ret = false;
 154
 155         if (!src)
 156                 return ret;
 157
 158         if ((src->pvti.flags & PVCLOCK_GUEST_STOPPED) != 0) {
 159                 src->pvti.flags &= ~PVCLOCK_GUEST_STOPPED;
 160                 pvclock_touch_watchdogs();
 161                 ret = true;
 162         }
 163         return ret;
 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 };
 173 EXPORT_SYMBOL_GPL(kvm_clock);
 174
 175 static void kvm_register_clock(char *txt)
 176 {
 177         struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
 178         u64 pa;
 179
 180         if (!src)
 181                 return;
 182
 183         pa = slow_virt_to_phys(&src->pvti) | 0x01ULL;
 184         wrmsrl(msr_kvm_system_time, pa);
 185         pr_info("kvm-clock: cpu %d, msr %llx, %s", smp_processor_id(), pa, txt);
 186 }
 187
 188 static void kvm_save_sched_clock_state(void)
 189 {
 190 }
 191
 192 static void kvm_restore_sched_clock_state(void)
 193 {
 194         kvm_register_clock("primary cpu clock, resume");
 195 }
 196
 197 #ifdef CONFIG_X86_LOCAL_APIC
 198 static void kvm_setup_secondary_clock(void)
 199 {
 200         kvm_register_clock("secondary cpu clock");
 201 }
 202 #endif
 203
 204 /*
 205  * After the clock is registered, the host will keep writing to the
 206  * registered memory location. If the guest happens to shutdown, this memory
 207  * won't be valid. In cases like kexec, in which you install a new kernel, this
 208  * means a random memory location will be kept being written. So before any
 209  * kind of shutdown from our side, we unregister the clock by writing anything
 210  * that does not have the 'enable' bit set in the msr
 211  */
 212 #ifdef CONFIG_KEXEC_CORE
 213 static void kvm_crash_shutdown(struct pt_regs *regs)
 214 {
 215         native_write_msr(msr_kvm_system_time, 0, 0);
 216         kvm_disable_steal_time();
 217         native_machine_crash_shutdown(regs);
 218 }
 219 #endif
 220
 221 static void kvm_shutdown(void)
 222 {
 223         native_write_msr(msr_kvm_system_time, 0, 0);
 224         kvm_disable_steal_time();
 225         native_machine_shutdown();
 226 }
 227
 228 static void __init kvmclock_init_mem(void)
 229 {
 230         unsigned long ncpus;
 231         unsigned int order;
 232         struct page *p;
 233         int r;
 234
 235         if (HVC_BOOT_ARRAY_SIZE >= num_possible_cpus())
 236                 return;
 237
 238         ncpus = num_possible_cpus() - HVC_BOOT_ARRAY_SIZE;
 239         order = get_order(ncpus * sizeof(*hvclock_mem));
 240
 241         p = alloc_pages(GFP_KERNEL, order);
 242         if (!p) {
 243                 pr_warn("%s: failed to alloc %d pages", __func__, (1U << order));
 244                 return;
 245         }
 246
 247         hvclock_mem = page_address(p);
 248
 249         /*
 250          * hvclock is shared between the guest and the hypervisor, must
 251          * be mapped decrypted.
 252          */
 253         if (sev_active()) {
 254                 r = set_memory_decrypted((unsigned long) hvclock_mem,
 255                                          1UL << order);
 256                 if (r) {
 257                         __free_pages(p, order);
 258                         hvclock_mem = NULL;
 259                         pr_warn("kvmclock: set_memory_decrypted() failed. Disabling\n");
 260                         return;
 261                 }
 262         }
 263
 264         memset(hvclock_mem, 0, PAGE_SIZE << order);
 265 }
 266
 267 static int __init kvm_setup_vsyscall_timeinfo(void)
 268 {
 269 #ifdef CONFIG_X86_64
 270         u8 flags;
 271
 272         if (!per_cpu(hv_clock_per_cpu, 0) || !kvmclock_vsyscall)
 273                 return 0;
 274
 275         flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
 276         if (!(flags & PVCLOCK_TSC_STABLE_BIT))
 277                 return 0;
 278
 279         kvm_clock.archdata.vclock_mode = VCLOCK_PVCLOCK;
 280 #endif
 281
 282         kvmclock_init_mem();
 283
 284         return 0;
 285 }
 286 early_initcall(kvm_setup_vsyscall_timeinfo);
 287
 288 static int kvmclock_setup_percpu(unsigned int cpu)
 289 {
 290         struct pvclock_vsyscall_time_info *p = per_cpu(hv_clock_per_cpu, cpu);
 291
 292         /*
 293          * The per cpu area setup replicates CPU0 data to all cpu
 294          * pointers. So carefully check. CPU0 has been set up in init
 295          * already.
 296          */
 297         if (!cpu || (p && p != per_cpu(hv_clock_per_cpu, 0)))
 298                 return 0;
 299
 300         /* Use the static page for the first CPUs, allocate otherwise */
 301         if (cpu < HVC_BOOT_ARRAY_SIZE)
 302                 p = &hv_clock_boot[cpu];
 303         else if (hvclock_mem)
 304                 p = hvclock_mem + cpu - HVC_BOOT_ARRAY_SIZE;
 305         else
 306                 return -ENOMEM;
 307
 308         per_cpu(hv_clock_per_cpu, cpu) = p;
 309         return p ? 0 : -ENOMEM;
 310 }
 311
 312 void __init kvmclock_init(void)
 313 {
 314         u8 flags;
 315
 316         if (!kvm_para_available() || !kvmclock)
 317                 return;
 318
 319         if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) {
 320                 msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW;
 321                 msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW;
 322         } else if (!kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)) {
 323                 return;
 324         }
 325
 326         if (cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "kvmclock:setup_percpu",
 327                               kvmclock_setup_percpu, NULL) < 0) {
 328                 return;
 329         }
 330
 331         pr_info("kvm-clock: Using msrs %x and %x",
 332                 msr_kvm_system_time, msr_kvm_wall_clock);
 333
 334         this_cpu_write(hv_clock_per_cpu, &hv_clock_boot[0]);
 335         kvm_register_clock("primary cpu clock");
 336         pvclock_set_pvti_cpu0_va(hv_clock_boot);
 337
 338         if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT))
 339                 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
 340
 341         flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
 342         kvm_sched_clock_init(flags & PVCLOCK_TSC_STABLE_BIT);
 343
 344         x86_platform.calibrate_tsc = kvm_get_tsc_khz;
 345         x86_platform.calibrate_cpu = kvm_get_tsc_khz;
 346         x86_platform.get_wallclock = kvm_get_wallclock;
 347         x86_platform.set_wallclock = kvm_set_wallclock;
 348 #ifdef CONFIG_X86_LOCAL_APIC
 349         x86_cpuinit.early_percpu_clock_init = kvm_setup_secondary_clock;
 350 #endif
 351         x86_platform.save_sched_clock_state = kvm_save_sched_clock_state;
 352         x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state;
 353         machine_ops.shutdown  = kvm_shutdown;
 354 #ifdef CONFIG_KEXEC_CORE
 355         machine_ops.crash_shutdown  = kvm_crash_shutdown;
 356 #endif
 357         kvm_get_preset_lpj();
 358         clocksource_register_hz(&kvm_clock, NSEC_PER_SEC);
 359         pv_info.name = "KVM";
 360 }