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
20 #include <asm/hypervisor.h>
21 #include <asm/mem_encrypt.h>
22 #include <asm/x86_init.h>
23 #include <asm/kvmclock.h>
24
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;
30
parse_no_kvmclock(char * arg)31 static int __init parse_no_kvmclock(char *arg)
32 {
33 kvmclock = 0;
34 return 0;
35 }
36 early_param("no-kvmclock", parse_no_kvmclock);
37
parse_no_kvmclock_vsyscall(char * arg)38 static int __init parse_no_kvmclock_vsyscall(char *arg)
39 {
40 kvmclock_vsyscall = 0;
41 return 0;
42 }
43 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
44
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))
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 */
kvm_get_wallclock(struct timespec64 * now)62 static void kvm_get_wallclock(struct timespec64 *now)
63 {
64 wrmsrl(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
kvm_set_wallclock(const struct timespec64 * now)70 static int kvm_set_wallclock(const struct timespec64 *now)
71 {
72 return -ENODEV;
73 }
74
kvm_clock_read(void)75 static u64 kvm_clock_read(void)
76 {
77 u64 ret;
78
79 preempt_disable_notrace();
80 ret = pvclock_clocksource_read(this_cpu_pvti());
81 preempt_enable_notrace();
82 return ret;
83 }
84
kvm_clock_get_cycles(struct clocksource * cs)85 static u64 kvm_clock_get_cycles(struct clocksource *cs)
86 {
87 return kvm_clock_read();
88 }
89
kvm_sched_clock_read(void)90 static u64 kvm_sched_clock_read(void)
91 {
92 return kvm_clock_read() - kvm_sched_clock_offset;
93 }
94
kvm_sched_clock_init(bool stable)95 static inline void kvm_sched_clock_init(bool stable)
96 {
97 if (!stable)
98 clear_sched_clock_stable();
99 kvm_sched_clock_offset = kvm_clock_read();
100 pv_ops.time.sched_clock = kvm_sched_clock_read;
101
102 pr_info("kvm-clock: using sched offset of %llu cycles",
103 kvm_sched_clock_offset);
104
105 BUILD_BUG_ON(sizeof(kvm_sched_clock_offset) >
106 sizeof(((struct pvclock_vcpu_time_info *)NULL)->system_time));
107 }
108
109 /*
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
116 * lpj here
117 */
kvm_get_tsc_khz(void)118 static unsigned long kvm_get_tsc_khz(void)
119 {
120 setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
121 return pvclock_tsc_khz(this_cpu_pvti());
122 }
123
kvm_get_preset_lpj(void)124 static void __init kvm_get_preset_lpj(void)
125 {
126 unsigned long khz;
127 u64 lpj;
128
129 khz = kvm_get_tsc_khz();
130
131 lpj = ((u64)khz * 1000);
132 do_div(lpj, HZ);
133 preset_lpj = lpj;
134 }
135
kvm_check_and_clear_guest_paused(void)136 bool kvm_check_and_clear_guest_paused(void)
137 {
138 struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
139 bool ret = false;
140
141 if (!src)
142 return ret;
143
144 if ((src->pvti.flags & PVCLOCK_GUEST_STOPPED) != 0) {
145 src->pvti.flags &= ~PVCLOCK_GUEST_STOPPED;
146 pvclock_touch_watchdogs();
147 ret = true;
148 }
149 return ret;
150 }
151
kvm_cs_enable(struct clocksource * cs)152 static int kvm_cs_enable(struct clocksource *cs)
153 {
154 vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK);
155 return 0;
156 }
157
158 struct clocksource kvm_clock = {
159 .name = "kvm-clock",
160 .read = kvm_clock_get_cycles,
161 .rating = 400,
162 .mask = CLOCKSOURCE_MASK(64),
163 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
164 .enable = kvm_cs_enable,
165 };
166 EXPORT_SYMBOL_GPL(kvm_clock);
167
kvm_register_clock(char * txt)168 static void kvm_register_clock(char *txt)
169 {
170 struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
171 u64 pa;
172
173 if (!src)
174 return;
175
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);
179 }
180
kvm_save_sched_clock_state(void)181 static void kvm_save_sched_clock_state(void)
182 {
183 }
184
kvm_restore_sched_clock_state(void)185 static void kvm_restore_sched_clock_state(void)
186 {
187 kvm_register_clock("primary cpu clock, resume");
188 }
189
190 #ifdef CONFIG_X86_LOCAL_APIC
kvm_setup_secondary_clock(void)191 static void kvm_setup_secondary_clock(void)
192 {
193 kvm_register_clock("secondary cpu clock");
194 }
195 #endif
196
kvmclock_disable(void)197 void kvmclock_disable(void)
198 {
199 native_write_msr(msr_kvm_system_time, 0, 0);
200 }
201
kvmclock_init_mem(void)202 static 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 (sev_active()) {
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
kvm_setup_vsyscall_timeinfo(void)241 static int __init kvm_setup_vsyscall_timeinfo(void)
242 {
243 kvmclock_init_mem();
244
245 #ifdef CONFIG_X86_64
246 if (per_cpu(hv_clock_per_cpu, 0) && kvmclock_vsyscall) {
247 u8 flags;
248
249 flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
250 if (!(flags & PVCLOCK_TSC_STABLE_BIT))
251 return 0;
252
253 kvm_clock.vdso_clock_mode = VDSO_CLOCKMODE_PVCLOCK;
254 }
255 #endif
256
257 return 0;
258 }
259 early_initcall(kvm_setup_vsyscall_timeinfo);
260
kvmclock_setup_percpu(unsigned int cpu)261 static int kvmclock_setup_percpu(unsigned int cpu)
262 {
263 struct pvclock_vsyscall_time_info *p = per_cpu(hv_clock_per_cpu, cpu);
264
265 /*
266 * The per cpu area setup replicates CPU0 data to all cpu
267 * pointers. So carefully check. CPU0 has been set up in init
268 * already.
269 */
270 if (!cpu || (p && p != per_cpu(hv_clock_per_cpu, 0)))
271 return 0;
272
273 /* Use the static page for the first CPUs, allocate otherwise */
274 if (cpu < HVC_BOOT_ARRAY_SIZE)
275 p = &hv_clock_boot[cpu];
276 else if (hvclock_mem)
277 p = hvclock_mem + cpu - HVC_BOOT_ARRAY_SIZE;
278 else
279 return -ENOMEM;
280
281 per_cpu(hv_clock_per_cpu, cpu) = p;
282 return p ? 0 : -ENOMEM;
283 }
284
kvmclock_init(void)285 void __init kvmclock_init(void)
286 {
287 u8 flags;
288
289 if (!kvm_para_available() || !kvmclock)
290 return;
291
292 if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) {
293 msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW;
294 msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW;
295 } else if (!kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)) {
296 return;
297 }
298
299 if (cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "kvmclock:setup_percpu",
300 kvmclock_setup_percpu, NULL) < 0) {
301 return;
302 }
303
304 pr_info("kvm-clock: Using msrs %x and %x",
305 msr_kvm_system_time, msr_kvm_wall_clock);
306
307 this_cpu_write(hv_clock_per_cpu, &hv_clock_boot[0]);
308 kvm_register_clock("primary cpu clock");
309 pvclock_set_pvti_cpu0_va(hv_clock_boot);
310
311 if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT))
312 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
313
314 flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
315 kvm_sched_clock_init(flags & PVCLOCK_TSC_STABLE_BIT);
316
317 x86_platform.calibrate_tsc = kvm_get_tsc_khz;
318 x86_platform.calibrate_cpu = kvm_get_tsc_khz;
319 x86_platform.get_wallclock = kvm_get_wallclock;
320 x86_platform.set_wallclock = kvm_set_wallclock;
321 #ifdef CONFIG_X86_LOCAL_APIC
322 x86_cpuinit.early_percpu_clock_init = kvm_setup_secondary_clock;
323 #endif
324 x86_platform.save_sched_clock_state = kvm_save_sched_clock_state;
325 x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state;
326 kvm_get_preset_lpj();
327
328 /*
329 * X86_FEATURE_NONSTOP_TSC is TSC runs at constant rate
330 * with P/T states and does not stop in deep C-states.
331 *
332 * Invariant TSC exposed by host means kvmclock is not necessary:
333 * can use TSC as clocksource.
334 *
335 */
336 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) &&
337 boot_cpu_has(X86_FEATURE_NONSTOP_TSC) &&
338 !check_tsc_unstable())
339 kvm_clock.rating = 299;
340
341 clocksource_register_hz(&kvm_clock, NSEC_PER_SEC);
342 pv_info.name = "KVM";
343 }
344