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.
6 #include <linux/clocksource.h>
7 #include <linux/kvm_para.h>
8 #include <asm/pvclock.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>
17 #include <linux/slab.h>
18 #include <linux/set_memory.h>
20 #include <asm/hypervisor.h>
21 #include <asm/mem_encrypt.h>
22 #include <asm/x86_init.h>
23 #include <asm/kvmclock.h>
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;
31 static int __init parse_no_kvmclock(char *arg)
36 early_param("no-kvmclock", parse_no_kvmclock);
38 static int __init parse_no_kvmclock_vsyscall(char *arg)
40 kvmclock_vsyscall = 0;
43 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
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))
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);
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
62 static void kvm_get_wallclock(struct timespec64 *now)
64 wrmsrl(msr_kvm_wall_clock, slow_virt_to_phys(&wall_clock));
66 pvclock_read_wallclock(&wall_clock, this_cpu_pvti(), now);
70 static int kvm_set_wallclock(const struct timespec64 *now)
75 static u64 kvm_clock_read(void)
79 preempt_disable_notrace();
80 ret = pvclock_clocksource_read(this_cpu_pvti());
81 preempt_enable_notrace();
85 static u64 kvm_clock_get_cycles(struct clocksource *cs)
87 return kvm_clock_read();
90 static u64 kvm_sched_clock_read(void)
92 return kvm_clock_read() - kvm_sched_clock_offset;
95 static inline void kvm_sched_clock_init(bool stable)
98 clear_sched_clock_stable();
99 kvm_sched_clock_offset = kvm_clock_read();
100 pv_ops.time.sched_clock = kvm_sched_clock_read;
102 pr_info("kvm-clock: using sched offset of %llu cycles",
103 kvm_sched_clock_offset);
105 BUILD_BUG_ON(sizeof(kvm_sched_clock_offset) >
106 sizeof(((struct pvclock_vcpu_time_info *)NULL)->system_time));
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
118 static unsigned long kvm_get_tsc_khz(void)
120 setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
121 return pvclock_tsc_khz(this_cpu_pvti());
124 static void __init kvm_get_preset_lpj(void)
129 khz = kvm_get_tsc_khz();
131 lpj = ((u64)khz * 1000);
136 bool kvm_check_and_clear_guest_paused(void)
138 struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
144 if ((src->pvti.flags & PVCLOCK_GUEST_STOPPED) != 0) {
145 src->pvti.flags &= ~PVCLOCK_GUEST_STOPPED;
146 pvclock_touch_watchdogs();
152 static int kvm_cs_enable(struct clocksource *cs)
154 vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK);
158 struct clocksource kvm_clock = {
160 .read = kvm_clock_get_cycles,
162 .mask = CLOCKSOURCE_MASK(64),
163 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
164 .enable = kvm_cs_enable,
166 EXPORT_SYMBOL_GPL(kvm_clock);
168 static void kvm_register_clock(char *txt)
170 struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
176 pa = slow_virt_to_phys(&src->pvti) | 0x01ULL;
177 wrmsrl(msr_kvm_system_time, pa);
178 pr_info("kvm-clock: cpu %d, msr %llx, %s", smp_processor_id(), pa, txt);
181 static void kvm_save_sched_clock_state(void)
185 static void kvm_restore_sched_clock_state(void)
187 kvm_register_clock("primary cpu clock, resume");
190 #ifdef CONFIG_X86_LOCAL_APIC
191 static void kvm_setup_secondary_clock(void)
193 kvm_register_clock("secondary cpu clock");
197 void kvmclock_disable(void)
199 if (msr_kvm_system_time)
200 native_write_msr(msr_kvm_system_time, 0, 0);
203 static void __init kvmclock_init_mem(void)
210 if (HVC_BOOT_ARRAY_SIZE >= num_possible_cpus())
213 ncpus = num_possible_cpus() - HVC_BOOT_ARRAY_SIZE;
214 order = get_order(ncpus * sizeof(*hvclock_mem));
216 p = alloc_pages(GFP_KERNEL, order);
218 pr_warn("%s: failed to alloc %d pages", __func__, (1U << order));
222 hvclock_mem = page_address(p);
225 * hvclock is shared between the guest and the hypervisor, must
226 * be mapped decrypted.
229 r = set_memory_decrypted((unsigned long) hvclock_mem,
232 __free_pages(p, order);
234 pr_warn("kvmclock: set_memory_decrypted() failed. Disabling\n");
239 memset(hvclock_mem, 0, PAGE_SIZE << order);
242 static int __init kvm_setup_vsyscall_timeinfo(void)
247 if (per_cpu(hv_clock_per_cpu, 0) && kvmclock_vsyscall) {
250 flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
251 if (!(flags & PVCLOCK_TSC_STABLE_BIT))
254 kvm_clock.vdso_clock_mode = VDSO_CLOCKMODE_PVCLOCK;
260 early_initcall(kvm_setup_vsyscall_timeinfo);
262 static int kvmclock_setup_percpu(unsigned int cpu)
264 struct pvclock_vsyscall_time_info *p = per_cpu(hv_clock_per_cpu, cpu);
267 * The per cpu area setup replicates CPU0 data to all cpu
268 * pointers. So carefully check. CPU0 has been set up in init
271 if (!cpu || (p && p != per_cpu(hv_clock_per_cpu, 0)))
274 /* Use the static page for the first CPUs, allocate otherwise */
275 if (cpu < HVC_BOOT_ARRAY_SIZE)
276 p = &hv_clock_boot[cpu];
277 else if (hvclock_mem)
278 p = hvclock_mem + cpu - HVC_BOOT_ARRAY_SIZE;
282 per_cpu(hv_clock_per_cpu, cpu) = p;
283 return p ? 0 : -ENOMEM;
286 void __init kvmclock_init(void)
290 if (!kvm_para_available() || !kvmclock)
293 if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) {
294 msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW;
295 msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW;
296 } else if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)) {
297 msr_kvm_system_time = MSR_KVM_SYSTEM_TIME;
298 msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK;
303 if (cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "kvmclock:setup_percpu",
304 kvmclock_setup_percpu, NULL) < 0) {
308 pr_info("kvm-clock: Using msrs %x and %x",
309 msr_kvm_system_time, msr_kvm_wall_clock);
311 this_cpu_write(hv_clock_per_cpu, &hv_clock_boot[0]);
312 kvm_register_clock("primary cpu clock");
313 pvclock_set_pvti_cpu0_va(hv_clock_boot);
315 if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT))
316 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
318 flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
319 kvm_sched_clock_init(flags & PVCLOCK_TSC_STABLE_BIT);
321 x86_platform.calibrate_tsc = kvm_get_tsc_khz;
322 x86_platform.calibrate_cpu = kvm_get_tsc_khz;
323 x86_platform.get_wallclock = kvm_get_wallclock;
324 x86_platform.set_wallclock = kvm_set_wallclock;
325 #ifdef CONFIG_X86_LOCAL_APIC
326 x86_cpuinit.early_percpu_clock_init = kvm_setup_secondary_clock;
328 x86_platform.save_sched_clock_state = kvm_save_sched_clock_state;
329 x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state;
330 kvm_get_preset_lpj();
333 * X86_FEATURE_NONSTOP_TSC is TSC runs at constant rate
334 * with P/T states and does not stop in deep C-states.
336 * Invariant TSC exposed by host means kvmclock is not necessary:
337 * can use TSC as clocksource.
340 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) &&
341 boot_cpu_has(X86_FEATURE_NONSTOP_TSC) &&
342 !check_tsc_unstable())
343 kvm_clock.rating = 299;
345 clocksource_register_hz(&kvm_clock, NSEC_PER_SEC);
346 pv_info.name = "KVM";