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 = MSR_KVM_SYSTEM_TIME;
28 static int msr_kvm_wall_clock __ro_after_init = MSR_KVM_WALL_CLOCK;
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 HVC_BOOT_ARRAY_SIZE \
47 (PAGE_SIZE / sizeof(struct pvclock_vsyscall_time_info))
49 static struct pvclock_vsyscall_time_info
50 hv_clock_boot[HVC_BOOT_ARRAY_SIZE] __bss_decrypted __aligned(PAGE_SIZE);
51 static struct pvclock_wall_clock wall_clock __bss_decrypted;
52 static DEFINE_PER_CPU(struct pvclock_vsyscall_time_info *, hv_clock_per_cpu);
53 static struct pvclock_vsyscall_time_info *hvclock_mem;
55 static inline struct pvclock_vcpu_time_info *this_cpu_pvti(void)
57 return &this_cpu_read(hv_clock_per_cpu)->pvti;
60 static inline struct pvclock_vsyscall_time_info *this_cpu_hvclock(void)
62 return this_cpu_read(hv_clock_per_cpu);
66 * The wallclock is the time of day when we booted. Since then, some time may
67 * have elapsed since the hypervisor wrote the data. So we try to account for
68 * that with system time
70 static void kvm_get_wallclock(struct timespec64 *now)
72 wrmsrl(msr_kvm_wall_clock, slow_virt_to_phys(&wall_clock));
74 pvclock_read_wallclock(&wall_clock, this_cpu_pvti(), now);
78 static int kvm_set_wallclock(const struct timespec64 *now)
83 static u64 kvm_clock_read(void)
87 preempt_disable_notrace();
88 ret = pvclock_clocksource_read(this_cpu_pvti());
89 preempt_enable_notrace();
93 static u64 kvm_clock_get_cycles(struct clocksource *cs)
95 return kvm_clock_read();
98 static u64 kvm_sched_clock_read(void)
100 return kvm_clock_read() - kvm_sched_clock_offset;
103 static inline void kvm_sched_clock_init(bool stable)
106 clear_sched_clock_stable();
107 kvm_sched_clock_offset = kvm_clock_read();
108 paravirt_set_sched_clock(kvm_sched_clock_read);
110 pr_info("kvm-clock: using sched offset of %llu cycles",
111 kvm_sched_clock_offset);
113 BUILD_BUG_ON(sizeof(kvm_sched_clock_offset) >
114 sizeof(((struct pvclock_vcpu_time_info *)NULL)->system_time));
118 * If we don't do that, there is the possibility that the guest
119 * will calibrate under heavy load - thus, getting a lower lpj -
120 * and execute the delays themselves without load. This is wrong,
121 * because no delay loop can finish beforehand.
122 * Any heuristics is subject to fail, because ultimately, a large
123 * poll of guests can be running and trouble each other. So we preset
126 static unsigned long kvm_get_tsc_khz(void)
128 setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
129 return pvclock_tsc_khz(this_cpu_pvti());
132 static void __init kvm_get_preset_lpj(void)
137 khz = kvm_get_tsc_khz();
139 lpj = ((u64)khz * 1000);
144 bool kvm_check_and_clear_guest_paused(void)
146 struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
152 if ((src->pvti.flags & PVCLOCK_GUEST_STOPPED) != 0) {
153 src->pvti.flags &= ~PVCLOCK_GUEST_STOPPED;
154 pvclock_touch_watchdogs();
160 static int kvm_cs_enable(struct clocksource *cs)
162 vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK);
166 struct clocksource kvm_clock = {
168 .read = kvm_clock_get_cycles,
170 .mask = CLOCKSOURCE_MASK(64),
171 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
172 .enable = kvm_cs_enable,
174 EXPORT_SYMBOL_GPL(kvm_clock);
176 static void kvm_register_clock(char *txt)
178 struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
184 pa = slow_virt_to_phys(&src->pvti) | 0x01ULL;
185 wrmsrl(msr_kvm_system_time, pa);
186 pr_info("kvm-clock: cpu %d, msr %llx, %s", smp_processor_id(), pa, txt);
189 static void kvm_save_sched_clock_state(void)
193 static void kvm_restore_sched_clock_state(void)
195 kvm_register_clock("primary cpu clock, resume");
198 #ifdef CONFIG_X86_LOCAL_APIC
199 static void kvm_setup_secondary_clock(void)
201 kvm_register_clock("secondary cpu clock");
205 void kvmclock_disable(void)
207 native_write_msr(msr_kvm_system_time, 0, 0);
210 static void __init kvmclock_init_mem(void)
217 if (HVC_BOOT_ARRAY_SIZE >= num_possible_cpus())
220 ncpus = num_possible_cpus() - HVC_BOOT_ARRAY_SIZE;
221 order = get_order(ncpus * sizeof(*hvclock_mem));
223 p = alloc_pages(GFP_KERNEL, order);
225 pr_warn("%s: failed to alloc %d pages", __func__, (1U << order));
229 hvclock_mem = page_address(p);
232 * hvclock is shared between the guest and the hypervisor, must
233 * be mapped decrypted.
236 r = set_memory_decrypted((unsigned long) hvclock_mem,
239 __free_pages(p, order);
241 pr_warn("kvmclock: set_memory_decrypted() failed. Disabling\n");
246 memset(hvclock_mem, 0, PAGE_SIZE << order);
249 static int __init kvm_setup_vsyscall_timeinfo(void)
254 if (per_cpu(hv_clock_per_cpu, 0) && kvmclock_vsyscall) {
257 flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
258 if (!(flags & PVCLOCK_TSC_STABLE_BIT))
261 kvm_clock.vdso_clock_mode = VDSO_CLOCKMODE_PVCLOCK;
267 early_initcall(kvm_setup_vsyscall_timeinfo);
269 static int kvmclock_setup_percpu(unsigned int cpu)
271 struct pvclock_vsyscall_time_info *p = per_cpu(hv_clock_per_cpu, cpu);
274 * The per cpu area setup replicates CPU0 data to all cpu
275 * pointers. So carefully check. CPU0 has been set up in init
278 if (!cpu || (p && p != per_cpu(hv_clock_per_cpu, 0)))
281 /* Use the static page for the first CPUs, allocate otherwise */
282 if (cpu < HVC_BOOT_ARRAY_SIZE)
283 p = &hv_clock_boot[cpu];
284 else if (hvclock_mem)
285 p = hvclock_mem + cpu - HVC_BOOT_ARRAY_SIZE;
289 per_cpu(hv_clock_per_cpu, cpu) = p;
290 return p ? 0 : -ENOMEM;
293 void __init kvmclock_init(void)
297 if (!kvm_para_available() || !kvmclock)
300 if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) {
301 msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW;
302 msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW;
303 } else if (!kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)) {
307 if (cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "kvmclock:setup_percpu",
308 kvmclock_setup_percpu, NULL) < 0) {
312 pr_info("kvm-clock: Using msrs %x and %x",
313 msr_kvm_system_time, msr_kvm_wall_clock);
315 this_cpu_write(hv_clock_per_cpu, &hv_clock_boot[0]);
316 kvm_register_clock("primary cpu clock");
317 pvclock_set_pvti_cpu0_va(hv_clock_boot);
319 if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT))
320 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
322 flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
323 kvm_sched_clock_init(flags & PVCLOCK_TSC_STABLE_BIT);
325 x86_platform.calibrate_tsc = kvm_get_tsc_khz;
326 x86_platform.calibrate_cpu = kvm_get_tsc_khz;
327 x86_platform.get_wallclock = kvm_get_wallclock;
328 x86_platform.set_wallclock = kvm_set_wallclock;
329 #ifdef CONFIG_X86_LOCAL_APIC
330 x86_cpuinit.early_percpu_clock_init = kvm_setup_secondary_clock;
332 x86_platform.save_sched_clock_state = kvm_save_sched_clock_state;
333 x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state;
334 kvm_get_preset_lpj();
337 * X86_FEATURE_NONSTOP_TSC is TSC runs at constant rate
338 * with P/T states and does not stop in deep C-states.
340 * Invariant TSC exposed by host means kvmclock is not necessary:
341 * can use TSC as clocksource.
344 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) &&
345 boot_cpu_has(X86_FEATURE_NONSTOP_TSC) &&
346 !check_tsc_unstable())
347 kvm_clock.rating = 299;
349 clocksource_register_hz(&kvm_clock, NSEC_PER_SEC);
350 pv_info.name = "KVM";