xref: /linux/arch/x86/kernel/kvmclock.c (revision d639d9fa162aadec1ae9980c4dcf6e50bd2f8290)
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/timer.h>
23 #include <asm/x86_init.h>
24 #include <asm/kvmclock.h>
25 
26 static int kvmclock __initdata = 1;
27 static int kvmclock_vsyscall __initdata = 1;
28 static int msr_kvm_system_time __ro_after_init;
29 static int msr_kvm_wall_clock __ro_after_init;
30 static u64 kvm_sched_clock_offset __ro_after_init;
31 
32 static int __init parse_no_kvmclock(char *arg)
33 {
34 	kvmclock = 0;
35 	return 0;
36 }
37 early_param("no-kvmclock", parse_no_kvmclock);
38 
39 static int __init parse_no_kvmclock_vsyscall(char *arg)
40 {
41 	kvmclock_vsyscall = 0;
42 	return 0;
43 }
44 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
45 
46 /* Aligned to page sizes to match what's mapped via vsyscalls to userspace */
47 #define HVC_BOOT_ARRAY_SIZE \
48 	(PAGE_SIZE / sizeof(struct pvclock_vsyscall_time_info))
49 
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);
56 
57 /*
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
61  */
62 static void kvm_get_wallclock(struct timespec64 *now)
63 {
64 	wrmsrq(msr_kvm_wall_clock, slow_virt_to_phys(&wall_clock));
65 	preempt_disable();
66 	pvclock_read_wallclock(&wall_clock, this_cpu_pvti(), now);
67 	preempt_enable();
68 }
69 
70 static int kvm_set_wallclock(const struct timespec64 *now)
71 {
72 	return -ENODEV;
73 }
74 
75 static u64 kvm_clock_read(void)
76 {
77 	u64 ret;
78 
79 	preempt_disable_notrace();
80 	ret = pvclock_clocksource_read_nowd(this_cpu_pvti());
81 	preempt_enable_notrace();
82 	return ret;
83 }
84 
85 static u64 kvm_clock_get_cycles(struct clocksource *cs)
86 {
87 	return kvm_clock_read();
88 }
89 
90 static u64 kvm_clock_get_cycles_snapshot(struct clocksource *cs,
91 					 struct clocksource_hw_snapshot *chs)
92 {
93 	struct pvclock_vcpu_time_info *src;
94 	unsigned version;
95 	u64 ret, tsc;
96 
97 	preempt_disable_notrace();
98 	src = this_cpu_pvti();
99 	do {
100 		version = pvclock_read_begin(src);
101 		tsc = rdtsc_ordered();
102 		ret = __pvclock_read_cycles(src, tsc);
103 	} while (pvclock_read_retry(src, version));
104 	preempt_enable_notrace();
105 
106 	chs->hw_cycles = tsc;
107 	chs->hw_csid = CSID_X86_TSC;
108 	return ret;
109 }
110 
111 static noinstr u64 kvm_sched_clock_read(void)
112 {
113 	return pvclock_clocksource_read_nowd(this_cpu_pvti()) - kvm_sched_clock_offset;
114 }
115 
116 static inline void kvm_sched_clock_init(bool stable)
117 {
118 	if (!stable)
119 		clear_sched_clock_stable();
120 	kvm_sched_clock_offset = kvm_clock_read();
121 	paravirt_set_sched_clock(kvm_sched_clock_read);
122 
123 	pr_info("kvm-clock: using sched offset of %llu cycles",
124 		kvm_sched_clock_offset);
125 
126 	BUILD_BUG_ON(sizeof(kvm_sched_clock_offset) >
127 		sizeof(((struct pvclock_vcpu_time_info *)NULL)->system_time));
128 }
129 
130 /*
131  * If we don't do that, there is the possibility that the guest
132  * will calibrate under heavy load - thus, getting a lower lpj -
133  * and execute the delays themselves without load. This is wrong,
134  * because no delay loop can finish beforehand.
135  * Any heuristics is subject to fail, because ultimately, a large
136  * poll of guests can be running and trouble each other. So we preset
137  * lpj here
138  */
139 static unsigned long kvm_get_tsc_khz(void)
140 {
141 	setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
142 	return pvclock_tsc_khz(this_cpu_pvti());
143 }
144 
145 static void __init kvm_get_preset_lpj(void)
146 {
147 	unsigned long khz;
148 	u64 lpj;
149 
150 	khz = kvm_get_tsc_khz();
151 
152 	lpj = ((u64)khz * 1000);
153 	do_div(lpj, HZ);
154 	preset_lpj = lpj;
155 }
156 
157 bool kvm_check_and_clear_guest_paused(void)
158 {
159 	struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
160 	bool ret = false;
161 
162 	if (!src)
163 		return ret;
164 
165 	if ((src->pvti.flags & PVCLOCK_GUEST_STOPPED) != 0) {
166 		src->pvti.flags &= ~PVCLOCK_GUEST_STOPPED;
167 		pvclock_touch_watchdogs();
168 		ret = true;
169 	}
170 	return ret;
171 }
172 
173 static int kvm_cs_enable(struct clocksource *cs)
174 {
175 	vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK);
176 	return 0;
177 }
178 
179 static struct clocksource kvm_clock = {
180 	.name		= "kvm-clock",
181 	.read		= kvm_clock_get_cycles,
182 	.read_snapshot	= kvm_clock_get_cycles_snapshot,
183 	.rating		= 400,
184 	.mask		= CLOCKSOURCE_MASK(64),
185 	.flags		= CLOCK_SOURCE_IS_CONTINUOUS,
186 	.id		= CSID_X86_KVM_CLK,
187 	.enable		= kvm_cs_enable,
188 };
189 
190 static void kvm_register_clock(char *txt)
191 {
192 	struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
193 	u64 pa;
194 
195 	if (!src)
196 		return;
197 
198 	pa = slow_virt_to_phys(&src->pvti) | 0x01ULL;
199 	wrmsrq(msr_kvm_system_time, pa);
200 	pr_debug("kvm-clock: cpu %d, msr %llx, %s", smp_processor_id(), pa, txt);
201 }
202 
203 static void kvm_save_sched_clock_state(void)
204 {
205 }
206 
207 static void kvm_restore_sched_clock_state(void)
208 {
209 	kvm_register_clock("primary cpu clock, resume");
210 }
211 
212 #ifdef CONFIG_X86_LOCAL_APIC
213 static void kvm_setup_secondary_clock(void)
214 {
215 	kvm_register_clock("secondary cpu clock");
216 }
217 #endif
218 
219 void kvmclock_disable(void)
220 {
221 	if (msr_kvm_system_time)
222 		native_write_msr(msr_kvm_system_time, 0);
223 }
224 
225 static void __init kvmclock_init_mem(void)
226 {
227 	unsigned long ncpus;
228 	unsigned int order;
229 	struct page *p;
230 	int r;
231 
232 	if (HVC_BOOT_ARRAY_SIZE >= num_possible_cpus())
233 		return;
234 
235 	ncpus = num_possible_cpus() - HVC_BOOT_ARRAY_SIZE;
236 	order = get_order(ncpus * sizeof(*hvclock_mem));
237 
238 	p = alloc_pages(GFP_KERNEL, order);
239 	if (!p) {
240 		pr_warn("%s: failed to alloc %d pages", __func__, (1U << order));
241 		return;
242 	}
243 
244 	hvclock_mem = page_address(p);
245 
246 	/*
247 	 * hvclock is shared between the guest and the hypervisor, must
248 	 * be mapped decrypted.
249 	 */
250 	if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) {
251 		r = set_memory_decrypted((unsigned long) hvclock_mem,
252 					 1UL << order);
253 		if (r) {
254 			__free_pages(p, order);
255 			hvclock_mem = NULL;
256 			pr_warn("kvmclock: set_memory_decrypted() failed. Disabling\n");
257 			return;
258 		}
259 	}
260 
261 	memset(hvclock_mem, 0, PAGE_SIZE << order);
262 }
263 
264 static int __init kvm_setup_vsyscall_timeinfo(void)
265 {
266 	if (!kvm_para_available() || !kvmclock || nopv)
267 		return 0;
268 
269 	kvmclock_init_mem();
270 
271 #ifdef CONFIG_X86_64
272 	if (per_cpu(hv_clock_per_cpu, 0) && kvmclock_vsyscall) {
273 		u8 flags;
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.vdso_clock_mode = VDSO_CLOCKMODE_PVCLOCK;
280 	}
281 #endif
282 
283 	return 0;
284 }
285 early_initcall(kvm_setup_vsyscall_timeinfo);
286 
287 static int kvmclock_setup_percpu(unsigned int cpu)
288 {
289 	struct pvclock_vsyscall_time_info *p = per_cpu(hv_clock_per_cpu, cpu);
290 
291 	/*
292 	 * The per cpu area setup replicates CPU0 data to all cpu
293 	 * pointers. So carefully check. CPU0 has been set up in init
294 	 * already.
295 	 */
296 	if (!cpu || (p && p != per_cpu(hv_clock_per_cpu, 0)))
297 		return 0;
298 
299 	/* Use the static page for the first CPUs, allocate otherwise */
300 	if (cpu < HVC_BOOT_ARRAY_SIZE)
301 		p = &hv_clock_boot[cpu];
302 	else if (hvclock_mem)
303 		p = hvclock_mem + cpu - HVC_BOOT_ARRAY_SIZE;
304 	else
305 		return -ENOMEM;
306 
307 	per_cpu(hv_clock_per_cpu, cpu) = p;
308 	return p ? 0 : -ENOMEM;
309 }
310 
311 void __init kvmclock_init(void)
312 {
313 	u8 flags;
314 
315 	if (!kvm_para_available() || !kvmclock)
316 		return;
317 
318 	if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) {
319 		msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW;
320 		msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW;
321 	} else if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)) {
322 		msr_kvm_system_time = MSR_KVM_SYSTEM_TIME;
323 		msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK;
324 	} else {
325 		return;
326 	}
327 
328 	if (cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "kvmclock:setup_percpu",
329 			      kvmclock_setup_percpu, NULL) < 0) {
330 		return;
331 	}
332 
333 	pr_info("kvm-clock: Using msrs %x and %x",
334 		msr_kvm_system_time, msr_kvm_wall_clock);
335 
336 	this_cpu_write(hv_clock_per_cpu, &hv_clock_boot[0]);
337 	kvm_register_clock("primary cpu clock");
338 	pvclock_set_pvti_cpu0_va(hv_clock_boot);
339 
340 	if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT))
341 		pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
342 
343 	flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
344 	kvm_sched_clock_init(flags & PVCLOCK_TSC_STABLE_BIT);
345 
346 	x86_platform.calibrate_tsc = kvm_get_tsc_khz;
347 	x86_platform.calibrate_cpu = kvm_get_tsc_khz;
348 	x86_platform.get_wallclock = kvm_get_wallclock;
349 	x86_platform.set_wallclock = kvm_set_wallclock;
350 #ifdef CONFIG_X86_LOCAL_APIC
351 	x86_cpuinit.early_percpu_clock_init = kvm_setup_secondary_clock;
352 #endif
353 	x86_platform.save_sched_clock_state = kvm_save_sched_clock_state;
354 	x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state;
355 	kvm_get_preset_lpj();
356 
357 	/*
358 	 * X86_FEATURE_NONSTOP_TSC is TSC runs at constant rate
359 	 * with P/T states and does not stop in deep C-states.
360 	 *
361 	 * Invariant TSC exposed by host means kvmclock is not necessary:
362 	 * can use TSC as clocksource.
363 	 *
364 	 */
365 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) &&
366 	    boot_cpu_has(X86_FEATURE_NONSTOP_TSC) &&
367 	    !check_tsc_unstable())
368 		kvm_clock.rating = 299;
369 
370 	clocksource_register_hz(&kvm_clock, NSEC_PER_SEC);
371 	pv_info.name = "KVM";
372 }
373