Line data Source code
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/reboot.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 = MSR_KVM_SYSTEM_TIME;
29 : static int msr_kvm_wall_clock __ro_after_init = MSR_KVM_WALL_CLOCK;
30 : static u64 kvm_sched_clock_offset __ro_after_init;
31 :
32 0 : static int __init parse_no_kvmclock(char *arg)
33 : {
34 0 : kvmclock = 0;
35 0 : return 0;
36 : }
37 : early_param("no-kvmclock", parse_no_kvmclock);
38 :
39 0 : static int __init parse_no_kvmclock_vsyscall(char *arg)
40 : {
41 0 : kvmclock_vsyscall = 0;
42 0 : return 0;
43 : }
44 : early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
45 :
46 : /* Aligned to page sizes to match whats 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 DEFINE_PER_CPU(struct pvclock_vsyscall_time_info *, hv_clock_per_cpu);
54 : static struct pvclock_vsyscall_time_info *hvclock_mem;
55 :
56 651876 : static inline struct pvclock_vcpu_time_info *this_cpu_pvti(void)
57 : {
58 1305396 : return &this_cpu_read(hv_clock_per_cpu)->pvti;
59 : }
60 :
61 4 : static inline struct pvclock_vsyscall_time_info *this_cpu_hvclock(void)
62 : {
63 8 : return this_cpu_read(hv_clock_per_cpu);
64 : }
65 :
66 : /*
67 : * The wallclock is the time of day when we booted. Since then, some time may
68 : * have elapsed since the hypervisor wrote the data. So we try to account for
69 : * that with system time
70 : */
71 1 : static void kvm_get_wallclock(struct timespec64 *now)
72 : {
73 1 : wrmsrl(msr_kvm_wall_clock, slow_virt_to_phys(&wall_clock));
74 1 : preempt_disable();
75 1 : pvclock_read_wallclock(&wall_clock, this_cpu_pvti(), now);
76 1 : preempt_enable();
77 1 : }
78 :
79 0 : static int kvm_set_wallclock(const struct timespec64 *now)
80 : {
81 0 : return -ENODEV;
82 : }
83 :
84 651217 : static u64 kvm_clock_read(void)
85 : {
86 651217 : u64 ret;
87 :
88 651217 : preempt_disable_notrace();
89 651872 : ret = pvclock_clocksource_read(this_cpu_pvti());
90 662203 : preempt_enable_notrace();
91 662958 : return ret;
92 : }
93 :
94 389153 : static u64 kvm_clock_get_cycles(struct clocksource *cs)
95 : {
96 389153 : return kvm_clock_read();
97 : }
98 :
99 264445 : static u64 kvm_sched_clock_read(void)
100 : {
101 264445 : return kvm_clock_read() - kvm_sched_clock_offset;
102 : }
103 :
104 1 : static inline void kvm_sched_clock_init(bool stable)
105 : {
106 1 : if (!stable)
107 0 : clear_sched_clock_stable();
108 1 : kvm_sched_clock_offset = kvm_clock_read();
109 1 : pv_ops.time.sched_clock = kvm_sched_clock_read;
110 :
111 1 : pr_info("kvm-clock: using sched offset of %llu cycles",
112 : kvm_sched_clock_offset);
113 :
114 1 : BUILD_BUG_ON(sizeof(kvm_sched_clock_offset) >
115 : sizeof(((struct pvclock_vcpu_time_info *)NULL)->system_time));
116 1 : }
117 :
118 : /*
119 : * If we don't do that, there is the possibility that the guest
120 : * will calibrate under heavy load - thus, getting a lower lpj -
121 : * and execute the delays themselves without load. This is wrong,
122 : * because no delay loop can finish beforehand.
123 : * Any heuristics is subject to fail, because ultimately, a large
124 : * poll of guests can be running and trouble each other. So we preset
125 : * lpj here
126 : */
127 3 : static unsigned long kvm_get_tsc_khz(void)
128 : {
129 3 : setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
130 3 : return pvclock_tsc_khz(this_cpu_pvti());
131 : }
132 :
133 1 : static void __init kvm_get_preset_lpj(void)
134 : {
135 1 : unsigned long khz;
136 1 : u64 lpj;
137 :
138 1 : khz = kvm_get_tsc_khz();
139 :
140 1 : lpj = ((u64)khz * 1000);
141 1 : do_div(lpj, HZ);
142 1 : preset_lpj = lpj;
143 1 : }
144 :
145 0 : bool kvm_check_and_clear_guest_paused(void)
146 : {
147 0 : struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
148 0 : bool ret = false;
149 :
150 0 : if (!src)
151 : return ret;
152 :
153 0 : if ((src->pvti.flags & PVCLOCK_GUEST_STOPPED) != 0) {
154 0 : src->pvti.flags &= ~PVCLOCK_GUEST_STOPPED;
155 0 : pvclock_touch_watchdogs();
156 0 : ret = true;
157 : }
158 : return ret;
159 : }
160 :
161 1 : static int kvm_cs_enable(struct clocksource *cs)
162 : {
163 1 : vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK);
164 1 : return 0;
165 : }
166 :
167 : struct clocksource kvm_clock = {
168 : .name = "kvm-clock",
169 : .read = kvm_clock_get_cycles,
170 : .rating = 400,
171 : .mask = CLOCKSOURCE_MASK(64),
172 : .flags = CLOCK_SOURCE_IS_CONTINUOUS,
173 : .enable = kvm_cs_enable,
174 : };
175 : EXPORT_SYMBOL_GPL(kvm_clock);
176 :
177 4 : static void kvm_register_clock(char *txt)
178 : {
179 4 : struct pvclock_vsyscall_time_info *src = this_cpu_hvclock();
180 4 : u64 pa;
181 :
182 4 : if (!src)
183 : return;
184 :
185 4 : pa = slow_virt_to_phys(&src->pvti) | 0x01ULL;
186 4 : wrmsrl(msr_kvm_system_time, pa);
187 4 : pr_info("kvm-clock: cpu %d, msr %llx, %s", smp_processor_id(), pa, txt);
188 : }
189 :
190 0 : static void kvm_save_sched_clock_state(void)
191 : {
192 0 : }
193 :
194 0 : static void kvm_restore_sched_clock_state(void)
195 : {
196 0 : kvm_register_clock("primary cpu clock, resume");
197 0 : }
198 :
199 : #ifdef CONFIG_X86_LOCAL_APIC
200 3 : static void kvm_setup_secondary_clock(void)
201 : {
202 3 : kvm_register_clock("secondary cpu clock");
203 3 : }
204 : #endif
205 :
206 : /*
207 : * After the clock is registered, the host will keep writing to the
208 : * registered memory location. If the guest happens to shutdown, this memory
209 : * won't be valid. In cases like kexec, in which you install a new kernel, this
210 : * means a random memory location will be kept being written. So before any
211 : * kind of shutdown from our side, we unregister the clock by writing anything
212 : * that does not have the 'enable' bit set in the msr
213 : */
214 : #ifdef CONFIG_KEXEC_CORE
215 : static void kvm_crash_shutdown(struct pt_regs *regs)
216 : {
217 : native_write_msr(msr_kvm_system_time, 0, 0);
218 : kvm_disable_steal_time();
219 : native_machine_crash_shutdown(regs);
220 : }
221 : #endif
222 :
223 0 : static void kvm_shutdown(void)
224 : {
225 0 : native_write_msr(msr_kvm_system_time, 0, 0);
226 0 : kvm_disable_steal_time();
227 0 : native_machine_shutdown();
228 0 : }
229 :
230 1 : static void __init kvmclock_init_mem(void)
231 : {
232 1 : unsigned long ncpus;
233 1 : unsigned int order;
234 1 : struct page *p;
235 1 : int r;
236 :
237 1 : if (HVC_BOOT_ARRAY_SIZE >= num_possible_cpus())
238 : return;
239 :
240 0 : ncpus = num_possible_cpus() - HVC_BOOT_ARRAY_SIZE;
241 0 : order = get_order(ncpus * sizeof(*hvclock_mem));
242 :
243 0 : p = alloc_pages(GFP_KERNEL, order);
244 0 : if (!p) {
245 0 : pr_warn("%s: failed to alloc %d pages", __func__, (1U << order));
246 0 : return;
247 : }
248 :
249 0 : hvclock_mem = page_address(p);
250 :
251 : /*
252 : * hvclock is shared between the guest and the hypervisor, must
253 : * be mapped decrypted.
254 : */
255 0 : if (sev_active()) {
256 : r = set_memory_decrypted((unsigned long) hvclock_mem,
257 : 1UL << order);
258 : if (r) {
259 : __free_pages(p, order);
260 : hvclock_mem = NULL;
261 : pr_warn("kvmclock: set_memory_decrypted() failed. Disabling\n");
262 : return;
263 : }
264 : }
265 :
266 0 : memset(hvclock_mem, 0, PAGE_SIZE << order);
267 : }
268 :
269 1 : static int __init kvm_setup_vsyscall_timeinfo(void)
270 : {
271 1 : kvmclock_init_mem();
272 :
273 : #ifdef CONFIG_X86_64
274 1 : if (per_cpu(hv_clock_per_cpu, 0) && kvmclock_vsyscall) {
275 1 : u8 flags;
276 :
277 1 : flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
278 1 : if (!(flags & PVCLOCK_TSC_STABLE_BIT))
279 : return 0;
280 :
281 1 : kvm_clock.vdso_clock_mode = VDSO_CLOCKMODE_PVCLOCK;
282 : }
283 : #endif
284 :
285 : return 0;
286 : }
287 : early_initcall(kvm_setup_vsyscall_timeinfo);
288 :
289 3 : static int kvmclock_setup_percpu(unsigned int cpu)
290 : {
291 3 : struct pvclock_vsyscall_time_info *p = per_cpu(hv_clock_per_cpu, cpu);
292 :
293 : /*
294 : * The per cpu area setup replicates CPU0 data to all cpu
295 : * pointers. So carefully check. CPU0 has been set up in init
296 : * already.
297 : */
298 3 : if (!cpu || (p && p != per_cpu(hv_clock_per_cpu, 0)))
299 : return 0;
300 :
301 : /* Use the static page for the first CPUs, allocate otherwise */
302 3 : if (cpu < HVC_BOOT_ARRAY_SIZE)
303 3 : p = &hv_clock_boot[cpu];
304 0 : else if (hvclock_mem)
305 0 : p = hvclock_mem + cpu - HVC_BOOT_ARRAY_SIZE;
306 : else
307 : return -ENOMEM;
308 :
309 3 : per_cpu(hv_clock_per_cpu, cpu) = p;
310 3 : return p ? 0 : -ENOMEM;
311 : }
312 :
313 1 : void __init kvmclock_init(void)
314 : {
315 1 : u8 flags;
316 :
317 1 : if (!kvm_para_available() || !kvmclock)
318 : return;
319 :
320 1 : if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) {
321 1 : msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW;
322 1 : msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW;
323 0 : } else if (!kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)) {
324 : return;
325 : }
326 :
327 1 : if (cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "kvmclock:setup_percpu",
328 : kvmclock_setup_percpu, NULL) < 0) {
329 : return;
330 : }
331 :
332 1 : pr_info("kvm-clock: Using msrs %x and %x",
333 : msr_kvm_system_time, msr_kvm_wall_clock);
334 :
335 1 : this_cpu_write(hv_clock_per_cpu, &hv_clock_boot[0]);
336 1 : kvm_register_clock("primary cpu clock");
337 1 : pvclock_set_pvti_cpu0_va(hv_clock_boot);
338 :
339 1 : if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT))
340 1 : pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
341 :
342 1 : flags = pvclock_read_flags(&hv_clock_boot[0].pvti);
343 1 : kvm_sched_clock_init(flags & PVCLOCK_TSC_STABLE_BIT);
344 :
345 1 : x86_platform.calibrate_tsc = kvm_get_tsc_khz;
346 1 : x86_platform.calibrate_cpu = kvm_get_tsc_khz;
347 1 : x86_platform.get_wallclock = kvm_get_wallclock;
348 1 : x86_platform.set_wallclock = kvm_set_wallclock;
349 : #ifdef CONFIG_X86_LOCAL_APIC
350 1 : x86_cpuinit.early_percpu_clock_init = kvm_setup_secondary_clock;
351 : #endif
352 1 : x86_platform.save_sched_clock_state = kvm_save_sched_clock_state;
353 1 : x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state;
354 1 : machine_ops.shutdown = kvm_shutdown;
355 : #ifdef CONFIG_KEXEC_CORE
356 : machine_ops.crash_shutdown = kvm_crash_shutdown;
357 : #endif
358 1 : kvm_get_preset_lpj();
359 :
360 : /*
361 : * X86_FEATURE_NONSTOP_TSC is TSC runs at constant rate
362 : * with P/T states and does not stop in deep C-states.
363 : *
364 : * Invariant TSC exposed by host means kvmclock is not necessary:
365 : * can use TSC as clocksource.
366 : *
367 : */
368 1 : if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) &&
369 1 : boot_cpu_has(X86_FEATURE_NONSTOP_TSC) &&
370 0 : !check_tsc_unstable())
371 0 : kvm_clock.rating = 299;
372 :
373 1 : clocksource_register_hz(&kvm_clock, NSEC_PER_SEC);
374 1 : pv_info.name = "KVM";
375 : }
|
