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 struct pvclock_vsyscall_time_info
*hvclock_mem
;
53 DEFINE_PER_CPU(struct pvclock_vsyscall_time_info
*, hv_clock_per_cpu
);
54 EXPORT_PER_CPU_SYMBOL_GPL(hv_clock_per_cpu
);
57 * The wallclock is the time of day when we booted. Since then, some time may
58 * have elapsed since the hypervisor wrote the data. So we try to account for
59 * that with system time
61 static void kvm_get_wallclock(struct timespec64
*now
)
63 wrmsrl(msr_kvm_wall_clock
, slow_virt_to_phys(&wall_clock
));
65 pvclock_read_wallclock(&wall_clock
, this_cpu_pvti(), now
);
69 static int kvm_set_wallclock(const struct timespec64
*now
)
74 static u64
kvm_clock_read(void)
78 preempt_disable_notrace();
79 ret
= pvclock_clocksource_read(this_cpu_pvti());
80 preempt_enable_notrace();
84 static u64
kvm_clock_get_cycles(struct clocksource
*cs
)
86 return kvm_clock_read();
89 static u64
kvm_sched_clock_read(void)
91 return kvm_clock_read() - kvm_sched_clock_offset
;
94 static inline void kvm_sched_clock_init(bool stable
)
97 clear_sched_clock_stable();
98 kvm_sched_clock_offset
= kvm_clock_read();
99 paravirt_set_sched_clock(kvm_sched_clock_read
);
101 pr_info("kvm-clock: using sched offset of %llu cycles",
102 kvm_sched_clock_offset
);
104 BUILD_BUG_ON(sizeof(kvm_sched_clock_offset
) >
105 sizeof(((struct pvclock_vcpu_time_info
*)NULL
)->system_time
));
109 * If we don't do that, there is the possibility that the guest
110 * will calibrate under heavy load - thus, getting a lower lpj -
111 * and execute the delays themselves without load. This is wrong,
112 * because no delay loop can finish beforehand.
113 * Any heuristics is subject to fail, because ultimately, a large
114 * poll of guests can be running and trouble each other. So we preset
117 static unsigned long kvm_get_tsc_khz(void)
119 setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ
);
120 return pvclock_tsc_khz(this_cpu_pvti());
123 static void __init
kvm_get_preset_lpj(void)
128 khz
= kvm_get_tsc_khz();
130 lpj
= ((u64
)khz
* 1000);
135 bool kvm_check_and_clear_guest_paused(void)
137 struct pvclock_vsyscall_time_info
*src
= this_cpu_hvclock();
143 if ((src
->pvti
.flags
& PVCLOCK_GUEST_STOPPED
) != 0) {
144 src
->pvti
.flags
&= ~PVCLOCK_GUEST_STOPPED
;
145 pvclock_touch_watchdogs();
151 static int kvm_cs_enable(struct clocksource
*cs
)
153 vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK
);
157 struct clocksource kvm_clock
= {
159 .read
= kvm_clock_get_cycles
,
161 .mask
= CLOCKSOURCE_MASK(64),
162 .flags
= CLOCK_SOURCE_IS_CONTINUOUS
,
163 .enable
= kvm_cs_enable
,
165 EXPORT_SYMBOL_GPL(kvm_clock
);
167 static void kvm_register_clock(char *txt
)
169 struct pvclock_vsyscall_time_info
*src
= this_cpu_hvclock();
175 pa
= slow_virt_to_phys(&src
->pvti
) | 0x01ULL
;
176 wrmsrl(msr_kvm_system_time
, pa
);
177 pr_info("kvm-clock: cpu %d, msr %llx, %s", smp_processor_id(), pa
, txt
);
180 static void kvm_save_sched_clock_state(void)
184 static void kvm_restore_sched_clock_state(void)
186 kvm_register_clock("primary cpu clock, resume");
189 #ifdef CONFIG_X86_LOCAL_APIC
190 static void kvm_setup_secondary_clock(void)
192 kvm_register_clock("secondary cpu clock");
196 void kvmclock_disable(void)
198 native_write_msr(msr_kvm_system_time
, 0, 0);
201 static void __init
kvmclock_init_mem(void)
208 if (HVC_BOOT_ARRAY_SIZE
>= num_possible_cpus())
211 ncpus
= num_possible_cpus() - HVC_BOOT_ARRAY_SIZE
;
212 order
= get_order(ncpus
* sizeof(*hvclock_mem
));
214 p
= alloc_pages(GFP_KERNEL
, order
);
216 pr_warn("%s: failed to alloc %d pages", __func__
, (1U << order
));
220 hvclock_mem
= page_address(p
);
223 * hvclock is shared between the guest and the hypervisor, must
224 * be mapped decrypted.
227 r
= set_memory_decrypted((unsigned long) hvclock_mem
,
230 __free_pages(p
, order
);
232 pr_warn("kvmclock: set_memory_decrypted() failed. Disabling\n");
237 memset(hvclock_mem
, 0, PAGE_SIZE
<< order
);
240 static int __init
kvm_setup_vsyscall_timeinfo(void)
245 if (per_cpu(hv_clock_per_cpu
, 0) && kvmclock_vsyscall
) {
248 flags
= pvclock_read_flags(&hv_clock_boot
[0].pvti
);
249 if (!(flags
& PVCLOCK_TSC_STABLE_BIT
))
252 kvm_clock
.vdso_clock_mode
= VDSO_CLOCKMODE_PVCLOCK
;
258 early_initcall(kvm_setup_vsyscall_timeinfo
);
260 static int kvmclock_setup_percpu(unsigned int cpu
)
262 struct pvclock_vsyscall_time_info
*p
= per_cpu(hv_clock_per_cpu
, cpu
);
265 * The per cpu area setup replicates CPU0 data to all cpu
266 * pointers. So carefully check. CPU0 has been set up in init
269 if (!cpu
|| (p
&& p
!= per_cpu(hv_clock_per_cpu
, 0)))
272 /* Use the static page for the first CPUs, allocate otherwise */
273 if (cpu
< HVC_BOOT_ARRAY_SIZE
)
274 p
= &hv_clock_boot
[cpu
];
275 else if (hvclock_mem
)
276 p
= hvclock_mem
+ cpu
- HVC_BOOT_ARRAY_SIZE
;
280 per_cpu(hv_clock_per_cpu
, cpu
) = p
;
281 return p
? 0 : -ENOMEM
;
284 void __init
kvmclock_init(void)
288 if (!kvm_para_available() || !kvmclock
)
291 if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2
)) {
292 msr_kvm_system_time
= MSR_KVM_SYSTEM_TIME_NEW
;
293 msr_kvm_wall_clock
= MSR_KVM_WALL_CLOCK_NEW
;
294 } else if (!kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE
)) {
298 if (cpuhp_setup_state(CPUHP_BP_PREPARE_DYN
, "kvmclock:setup_percpu",
299 kvmclock_setup_percpu
, NULL
) < 0) {
303 pr_info("kvm-clock: Using msrs %x and %x",
304 msr_kvm_system_time
, msr_kvm_wall_clock
);
306 this_cpu_write(hv_clock_per_cpu
, &hv_clock_boot
[0]);
307 kvm_register_clock("primary cpu clock");
308 pvclock_set_pvti_cpu0_va(hv_clock_boot
);
310 if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT
))
311 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT
);
313 flags
= pvclock_read_flags(&hv_clock_boot
[0].pvti
);
314 kvm_sched_clock_init(flags
& PVCLOCK_TSC_STABLE_BIT
);
316 x86_platform
.calibrate_tsc
= kvm_get_tsc_khz
;
317 x86_platform
.calibrate_cpu
= kvm_get_tsc_khz
;
318 x86_platform
.get_wallclock
= kvm_get_wallclock
;
319 x86_platform
.set_wallclock
= kvm_set_wallclock
;
320 #ifdef CONFIG_X86_LOCAL_APIC
321 x86_cpuinit
.early_percpu_clock_init
= kvm_setup_secondary_clock
;
323 x86_platform
.save_sched_clock_state
= kvm_save_sched_clock_state
;
324 x86_platform
.restore_sched_clock_state
= kvm_restore_sched_clock_state
;
325 kvm_get_preset_lpj();
328 * X86_FEATURE_NONSTOP_TSC is TSC runs at constant rate
329 * with P/T states and does not stop in deep C-states.
331 * Invariant TSC exposed by host means kvmclock is not necessary:
332 * can use TSC as clocksource.
335 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC
) &&
336 boot_cpu_has(X86_FEATURE_NONSTOP_TSC
) &&
337 !check_tsc_unstable())
338 kvm_clock
.rating
= 299;
340 clocksource_register_hz(&kvm_clock
, NSEC_PER_SEC
);
341 pv_info
.name
= "KVM";