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.
4*/
5
6#include <linux/clocksource.h>
7#include <linux/kvm_para.h>
8#include <asm/pvclock.h>
9#include <asm/msr.h>
10#include <asm/apic.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>
16#include <linux/mm.h>
17#include <linux/slab.h>
18#include <linux/set_memory.h>
19#include <linux/cc_platform.h>
20
21#include <asm/hypervisor.h>
22#include <asm/x86_init.h>
23#include <asm/kvmclock.h>
24
25static int kvmclock __initdata = 1;
26static int kvmclock_vsyscall __initdata = 1;
27static int msr_kvm_system_time __ro_after_init;
28static int msr_kvm_wall_clock __ro_after_init;
29static u64 kvm_sched_clock_offset __ro_after_init;
30
31static int __init parse_no_kvmclock(char *arg)
32{
33	kvmclock = 0;
34	return 0;
35}
36early_param("no-kvmclock", parse_no_kvmclock);
37
38static int __init parse_no_kvmclock_vsyscall(char *arg)
39{
40	kvmclock_vsyscall = 0;
41	return 0;
42}
43early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
44
45/* Aligned to page sizes to match what's mapped via vsyscalls to userspace */
46#define HVC_BOOT_ARRAY_SIZE \
47	(PAGE_SIZE / sizeof(struct pvclock_vsyscall_time_info))
48
49static struct pvclock_vsyscall_time_info
50			hv_clock_boot[HVC_BOOT_ARRAY_SIZE] __bss_decrypted __aligned(PAGE_SIZE);
51static struct pvclock_wall_clock wall_clock __bss_decrypted;
52static struct pvclock_vsyscall_time_info *hvclock_mem;
53DEFINE_PER_CPU(struct pvclock_vsyscall_time_info *, hv_clock_per_cpu);
54EXPORT_PER_CPU_SYMBOL_GPL(hv_clock_per_cpu);
55
56/*
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
60 */
61static void kvm_get_wallclock(struct timespec64 *now)
62{
63	wrmsrl(msr_kvm_wall_clock, slow_virt_to_phys(&wall_clock));
64	preempt_disable();
65	pvclock_read_wallclock(&wall_clock, this_cpu_pvti(), now);
66	preempt_enable();
67}
68
69static int kvm_set_wallclock(const struct timespec64 *now)
70{
71	return -ENODEV;
72}
73
74static u64 kvm_clock_read(void)
75{
76	u64 ret;
77
78	preempt_disable_notrace();
79	ret = pvclock_clocksource_read_nowd(this_cpu_pvti());
80	preempt_enable_notrace();
81	return ret;
82}
83
84static u64 kvm_clock_get_cycles(struct clocksource *cs)
85{
86	return kvm_clock_read();
87}
88
89static noinstr u64 kvm_sched_clock_read(void)
90{
91	return pvclock_clocksource_read_nowd(this_cpu_pvti()) - kvm_sched_clock_offset;
92}
93
94static inline void kvm_sched_clock_init(bool stable)
95{
96	if (!stable)
97		clear_sched_clock_stable();
98	kvm_sched_clock_offset = kvm_clock_read();
99	paravirt_set_sched_clock(kvm_sched_clock_read);
100
101	pr_info("kvm-clock: using sched offset of %llu cycles",
102		kvm_sched_clock_offset);
103
104	BUILD_BUG_ON(sizeof(kvm_sched_clock_offset) >
105		sizeof(((struct pvclock_vcpu_time_info *)NULL)->system_time));
106}
107
108/*
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
115 * lpj here
116 */
117static unsigned long kvm_get_tsc_khz(void)
118{
119	setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
120	return pvclock_tsc_khz(this_cpu_pvti());
121}
122
123static void __init kvm_get_preset_lpj(void)
124{
125	unsigned long khz;
126	u64 lpj;
127
128	khz = kvm_get_tsc_khz();
129
130	lpj = ((u64)khz * 1000);
131	do_div(lpj, HZ);
132	preset_lpj = lpj;
133}
134
135bool kvm_check_and_clear_guest_paused(void)
136{
137	struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
138	bool ret = false;
139
140	if (!src)
141		return ret;
142
143	if ((src->pvti.flags & PVCLOCK_GUEST_STOPPED) != 0) {
144		src->pvti.flags &= ~PVCLOCK_GUEST_STOPPED;
145		pvclock_touch_watchdogs();
146		ret = true;
147	}
148	return ret;
149}
150
151static int kvm_cs_enable(struct clocksource *cs)
152{
153	vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK);
154	return 0;
155}
156
157static struct clocksource kvm_clock = {
158	.name	= "kvm-clock",
159	.read	= kvm_clock_get_cycles,
160	.rating	= 400,
161	.mask	= CLOCKSOURCE_MASK(64),
162	.flags	= CLOCK_SOURCE_IS_CONTINUOUS,
163	.id     = CSID_X86_KVM_CLK,
164	.enable	= kvm_cs_enable,
165};
166
167static void kvm_register_clock(char *txt)
168{
169	struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
170	u64 pa;
171
172	if (!src)
173		return;
174
175	pa = slow_virt_to_phys(&src->pvti) | 0x01ULL;
176	wrmsrl(msr_kvm_system_time, pa);
177	pr_debug("kvm-clock: cpu %d, msr %llx, %s", smp_processor_id(), pa, txt);
178}
179
180static void kvm_save_sched_clock_state(void)
181{
182}
183
184static void kvm_restore_sched_clock_state(void)
185{
186	kvm_register_clock("primary cpu clock, resume");
187}
188
189#ifdef CONFIG_X86_LOCAL_APIC
190static void kvm_setup_secondary_clock(void)
191{
192	kvm_register_clock("secondary cpu clock");
193}
194#endif
195
196void kvmclock_disable(void)
197{
198	if (msr_kvm_system_time)
199		native_write_msr(msr_kvm_system_time, 0, 0);
200}
201
202static void __init kvmclock_init_mem(void)
203{
204	unsigned long ncpus;
205	unsigned int order;
206	struct page *p;
207	int r;
208
209	if (HVC_BOOT_ARRAY_SIZE >= num_possible_cpus())
210		return;
211
212	ncpus = num_possible_cpus() - HVC_BOOT_ARRAY_SIZE;
213	order = get_order(ncpus * sizeof(*hvclock_mem));
214
215	p = alloc_pages(GFP_KERNEL, order);
216	if (!p) {
217		pr_warn("%s: failed to alloc %d pages", __func__, (1U << order));
218		return;
219	}
220
221	hvclock_mem = page_address(p);
222
223	/*
224	 * hvclock is shared between the guest and the hypervisor, must
225	 * be mapped decrypted.
226	 */
227	if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) {
228		r = set_memory_decrypted((unsigned long) hvclock_mem,
229					 1UL << order);
230		if (r) {
231			__free_pages(p, order);
232			hvclock_mem = NULL;
233			pr_warn("kvmclock: set_memory_decrypted() failed. Disabling\n");
234			return;
235		}
236	}
237
238	memset(hvclock_mem, 0, PAGE_SIZE << order);
239}
240
241static int __init kvm_setup_vsyscall_timeinfo(void)
242{
243	if (!kvm_para_available() || !kvmclock || nopv)
244		return 0;
245
246	kvmclock_init_mem();
247
248#ifdef CONFIG_X86_64
249	if (per_cpu(hv_clock_per_cpu, 0) && kvmclock_vsyscall) {
250		u8 flags;
251
252		flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
253		if (!(flags & PVCLOCK_TSC_STABLE_BIT))
254			return 0;
255
256		kvm_clock.vdso_clock_mode = VDSO_CLOCKMODE_PVCLOCK;
257	}
258#endif
259
260	return 0;
261}
262early_initcall(kvm_setup_vsyscall_timeinfo);
263
264static int kvmclock_setup_percpu(unsigned int cpu)
265{
266	struct pvclock_vsyscall_time_info *p = per_cpu(hv_clock_per_cpu, cpu);
267
268	/*
269	 * The per cpu area setup replicates CPU0 data to all cpu
270	 * pointers. So carefully check. CPU0 has been set up in init
271	 * already.
272	 */
273	if (!cpu || (p && p != per_cpu(hv_clock_per_cpu, 0)))
274		return 0;
275
276	/* Use the static page for the first CPUs, allocate otherwise */
277	if (cpu < HVC_BOOT_ARRAY_SIZE)
278		p = &hv_clock_boot[cpu];
279	else if (hvclock_mem)
280		p = hvclock_mem + cpu - HVC_BOOT_ARRAY_SIZE;
281	else
282		return -ENOMEM;
283
284	per_cpu(hv_clock_per_cpu, cpu) = p;
285	return p ? 0 : -ENOMEM;
286}
287
288void __init kvmclock_init(void)
289{
290	u8 flags;
291
292	if (!kvm_para_available() || !kvmclock)
293		return;
294
295	if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) {
296		msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW;
297		msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW;
298	} else if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)) {
299		msr_kvm_system_time = MSR_KVM_SYSTEM_TIME;
300		msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK;
301	} else {
302		return;
303	}
304
305	if (cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "kvmclock:setup_percpu",
306			      kvmclock_setup_percpu, NULL) < 0) {
307		return;
308	}
309
310	pr_info("kvm-clock: Using msrs %x and %x",
311		msr_kvm_system_time, msr_kvm_wall_clock);
312
313	this_cpu_write(hv_clock_per_cpu, &hv_clock_boot[0]);
314	kvm_register_clock("primary cpu clock");
315	pvclock_set_pvti_cpu0_va(hv_clock_boot);
316
317	if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT))
318		pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
319
320	flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
321	kvm_sched_clock_init(flags & PVCLOCK_TSC_STABLE_BIT);
322
323	x86_platform.calibrate_tsc = kvm_get_tsc_khz;
324	x86_platform.calibrate_cpu = kvm_get_tsc_khz;
325	x86_platform.get_wallclock = kvm_get_wallclock;
326	x86_platform.set_wallclock = kvm_set_wallclock;
327#ifdef CONFIG_X86_LOCAL_APIC
328	x86_cpuinit.early_percpu_clock_init = kvm_setup_secondary_clock;
329#endif
330	x86_platform.save_sched_clock_state = kvm_save_sched_clock_state;
331	x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state;
332	kvm_get_preset_lpj();
333
334	/*
335	 * X86_FEATURE_NONSTOP_TSC is TSC runs at constant rate
336	 * with P/T states and does not stop in deep C-states.
337	 *
338	 * Invariant TSC exposed by host means kvmclock is not necessary:
339	 * can use TSC as clocksource.
340	 *
341	 */
342	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) &&
343	    boot_cpu_has(X86_FEATURE_NONSTOP_TSC) &&
344	    !check_tsc_unstable())
345		kvm_clock.rating = 299;
346
347	clocksource_register_hz(&kvm_clock, NSEC_PER_SEC);
348	pv_info.name = "KVM";
349}
350