1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3 * Fast user context implementation of clock_gettime, gettimeofday, and time.
4 *
5 * Copyright (C) 2019 ARM Limited.
6 * Copyright 2006 Andi Kleen, SUSE Labs.
7 * 32 Bit compat layer by Stefani Seibold <stefani@seibold.net>
8 * sponsored by Rohde & Schwarz GmbH & Co. KG Munich/Germany
9 */
10 #ifndef __ASM_VDSO_GETTIMEOFDAY_H
11 #define __ASM_VDSO_GETTIMEOFDAY_H
12
13 #ifndef __ASSEMBLY__
14
15 #include <uapi/linux/time.h>
16 #include <asm/vgtod.h>
17 #include <asm/vvar.h>
18 #include <asm/unistd.h>
19 #include <asm/msr.h>
20 #include <asm/pvclock.h>
21 #include <clocksource/hyperv_timer.h>
22
23 #define __vdso_data (VVAR(_vdso_data))
24 #define __timens_vdso_data (TIMENS(_vdso_data))
25
26 #define VDSO_HAS_TIME 1
27
28 #define VDSO_HAS_CLOCK_GETRES 1
29
30 /*
31 * Declare the memory-mapped vclock data pages. These come from hypervisors.
32 * If we ever reintroduce something like direct access to an MMIO clock like
33 * the HPET again, it will go here as well.
34 *
35 * A load from any of these pages will segfault if the clock in question is
36 * disabled, so appropriate compiler barriers and checks need to be used
37 * to prevent stray loads.
38 *
39 * These declarations MUST NOT be const. The compiler will assume that
40 * an extern const variable has genuinely constant contents, and the
41 * resulting code won't work, since the whole point is that these pages
42 * change over time, possibly while we're accessing them.
43 */
44
45 #ifdef CONFIG_PARAVIRT_CLOCK
46 /*
47 * This is the vCPU 0 pvclock page. We only use pvclock from the vDSO
48 * if the hypervisor tells us that all vCPUs can get valid data from the
49 * vCPU 0 page.
50 */
51 extern struct pvclock_vsyscall_time_info pvclock_page
52 __attribute__((visibility("hidden")));
53 #endif
54
55 #ifdef CONFIG_HYPERV_TIMER
56 extern struct ms_hyperv_tsc_page hvclock_page
57 __attribute__((visibility("hidden")));
58 #endif
59
60 #ifdef CONFIG_TIME_NS
__arch_get_timens_vdso_data(void)61 static __always_inline const struct vdso_data *__arch_get_timens_vdso_data(void)
62 {
63 return __timens_vdso_data;
64 }
65 #endif
66
67 #ifndef BUILD_VDSO32
68
69 static __always_inline
clock_gettime_fallback(clockid_t _clkid,struct __kernel_timespec * _ts)70 long clock_gettime_fallback(clockid_t _clkid, struct __kernel_timespec *_ts)
71 {
72 long ret;
73
74 asm ("syscall" : "=a" (ret), "=m" (*_ts) :
75 "0" (__NR_clock_gettime), "D" (_clkid), "S" (_ts) :
76 "rcx", "r11");
77
78 return ret;
79 }
80
81 static __always_inline
gettimeofday_fallback(struct __kernel_old_timeval * _tv,struct timezone * _tz)82 long gettimeofday_fallback(struct __kernel_old_timeval *_tv,
83 struct timezone *_tz)
84 {
85 long ret;
86
87 asm("syscall" : "=a" (ret) :
88 "0" (__NR_gettimeofday), "D" (_tv), "S" (_tz) : "memory");
89
90 return ret;
91 }
92
93 static __always_inline
clock_getres_fallback(clockid_t _clkid,struct __kernel_timespec * _ts)94 long clock_getres_fallback(clockid_t _clkid, struct __kernel_timespec *_ts)
95 {
96 long ret;
97
98 asm ("syscall" : "=a" (ret), "=m" (*_ts) :
99 "0" (__NR_clock_getres), "D" (_clkid), "S" (_ts) :
100 "rcx", "r11");
101
102 return ret;
103 }
104
105 #else
106
107 static __always_inline
clock_gettime_fallback(clockid_t _clkid,struct __kernel_timespec * _ts)108 long clock_gettime_fallback(clockid_t _clkid, struct __kernel_timespec *_ts)
109 {
110 long ret;
111
112 asm (
113 "mov %%ebx, %%edx \n"
114 "mov %[clock], %%ebx \n"
115 "call __kernel_vsyscall \n"
116 "mov %%edx, %%ebx \n"
117 : "=a" (ret), "=m" (*_ts)
118 : "0" (__NR_clock_gettime64), [clock] "g" (_clkid), "c" (_ts)
119 : "edx");
120
121 return ret;
122 }
123
124 static __always_inline
clock_gettime32_fallback(clockid_t _clkid,struct old_timespec32 * _ts)125 long clock_gettime32_fallback(clockid_t _clkid, struct old_timespec32 *_ts)
126 {
127 long ret;
128
129 asm (
130 "mov %%ebx, %%edx \n"
131 "mov %[clock], %%ebx \n"
132 "call __kernel_vsyscall \n"
133 "mov %%edx, %%ebx \n"
134 : "=a" (ret), "=m" (*_ts)
135 : "0" (__NR_clock_gettime), [clock] "g" (_clkid), "c" (_ts)
136 : "edx");
137
138 return ret;
139 }
140
141 static __always_inline
gettimeofday_fallback(struct __kernel_old_timeval * _tv,struct timezone * _tz)142 long gettimeofday_fallback(struct __kernel_old_timeval *_tv,
143 struct timezone *_tz)
144 {
145 long ret;
146
147 asm(
148 "mov %%ebx, %%edx \n"
149 "mov %2, %%ebx \n"
150 "call __kernel_vsyscall \n"
151 "mov %%edx, %%ebx \n"
152 : "=a" (ret)
153 : "0" (__NR_gettimeofday), "g" (_tv), "c" (_tz)
154 : "memory", "edx");
155
156 return ret;
157 }
158
159 static __always_inline long
clock_getres_fallback(clockid_t _clkid,struct __kernel_timespec * _ts)160 clock_getres_fallback(clockid_t _clkid, struct __kernel_timespec *_ts)
161 {
162 long ret;
163
164 asm (
165 "mov %%ebx, %%edx \n"
166 "mov %[clock], %%ebx \n"
167 "call __kernel_vsyscall \n"
168 "mov %%edx, %%ebx \n"
169 : "=a" (ret), "=m" (*_ts)
170 : "0" (__NR_clock_getres_time64), [clock] "g" (_clkid), "c" (_ts)
171 : "edx");
172
173 return ret;
174 }
175
176 static __always_inline
clock_getres32_fallback(clockid_t _clkid,struct old_timespec32 * _ts)177 long clock_getres32_fallback(clockid_t _clkid, struct old_timespec32 *_ts)
178 {
179 long ret;
180
181 asm (
182 "mov %%ebx, %%edx \n"
183 "mov %[clock], %%ebx \n"
184 "call __kernel_vsyscall \n"
185 "mov %%edx, %%ebx \n"
186 : "=a" (ret), "=m" (*_ts)
187 : "0" (__NR_clock_getres), [clock] "g" (_clkid), "c" (_ts)
188 : "edx");
189
190 return ret;
191 }
192
193 #endif
194
195 #ifdef CONFIG_PARAVIRT_CLOCK
vread_pvclock(void)196 static u64 vread_pvclock(void)
197 {
198 const struct pvclock_vcpu_time_info *pvti = &pvclock_page.pvti;
199 u32 version;
200 u64 ret;
201
202 /*
203 * Note: The kernel and hypervisor must guarantee that cpu ID
204 * number maps 1:1 to per-CPU pvclock time info.
205 *
206 * Because the hypervisor is entirely unaware of guest userspace
207 * preemption, it cannot guarantee that per-CPU pvclock time
208 * info is updated if the underlying CPU changes or that that
209 * version is increased whenever underlying CPU changes.
210 *
211 * On KVM, we are guaranteed that pvti updates for any vCPU are
212 * atomic as seen by *all* vCPUs. This is an even stronger
213 * guarantee than we get with a normal seqlock.
214 *
215 * On Xen, we don't appear to have that guarantee, but Xen still
216 * supplies a valid seqlock using the version field.
217 *
218 * We only do pvclock vdso timing at all if
219 * PVCLOCK_TSC_STABLE_BIT is set, and we interpret that bit to
220 * mean that all vCPUs have matching pvti and that the TSC is
221 * synced, so we can just look at vCPU 0's pvti.
222 */
223
224 do {
225 version = pvclock_read_begin(pvti);
226
227 if (unlikely(!(pvti->flags & PVCLOCK_TSC_STABLE_BIT)))
228 return U64_MAX;
229
230 ret = __pvclock_read_cycles(pvti, rdtsc_ordered());
231 } while (pvclock_read_retry(pvti, version));
232
233 return ret;
234 }
235 #endif
236
237 #ifdef CONFIG_HYPERV_TIMER
vread_hvclock(void)238 static u64 vread_hvclock(void)
239 {
240 return hv_read_tsc_page(&hvclock_page);
241 }
242 #endif
243
__arch_get_hw_counter(s32 clock_mode,const struct vdso_data * vd)244 static inline u64 __arch_get_hw_counter(s32 clock_mode,
245 const struct vdso_data *vd)
246 {
247 if (likely(clock_mode == VDSO_CLOCKMODE_TSC))
248 return (u64)rdtsc_ordered();
249 /*
250 * For any memory-mapped vclock type, we need to make sure that gcc
251 * doesn't cleverly hoist a load before the mode check. Otherwise we
252 * might end up touching the memory-mapped page even if the vclock in
253 * question isn't enabled, which will segfault. Hence the barriers.
254 */
255 #ifdef CONFIG_PARAVIRT_CLOCK
256 if (clock_mode == VDSO_CLOCKMODE_PVCLOCK) {
257 barrier();
258 return vread_pvclock();
259 }
260 #endif
261 #ifdef CONFIG_HYPERV_TIMER
262 if (clock_mode == VDSO_CLOCKMODE_HVCLOCK) {
263 barrier();
264 return vread_hvclock();
265 }
266 #endif
267 return U64_MAX;
268 }
269
__arch_get_vdso_data(void)270 static __always_inline const struct vdso_data *__arch_get_vdso_data(void)
271 {
272 return __vdso_data;
273 }
274
arch_vdso_clocksource_ok(const struct vdso_data * vd)275 static inline bool arch_vdso_clocksource_ok(const struct vdso_data *vd)
276 {
277 return true;
278 }
279 #define vdso_clocksource_ok arch_vdso_clocksource_ok
280
281 /*
282 * Clocksource read value validation to handle PV and HyperV clocksources
283 * which can be invalidated asynchronously and indicate invalidation by
284 * returning U64_MAX, which can be effectively tested by checking for a
285 * negative value after casting it to s64.
286 */
arch_vdso_cycles_ok(u64 cycles)287 static inline bool arch_vdso_cycles_ok(u64 cycles)
288 {
289 return (s64)cycles >= 0;
290 }
291 #define vdso_cycles_ok arch_vdso_cycles_ok
292
293 /*
294 * x86 specific delta calculation.
295 *
296 * The regular implementation assumes that clocksource reads are globally
297 * monotonic. The TSC can be slightly off across sockets which can cause
298 * the regular delta calculation (@cycles - @last) to return a huge time
299 * jump.
300 *
301 * Therefore it needs to be verified that @cycles are greater than
302 * @last. If not then use @last, which is the base time of the current
303 * conversion period.
304 *
305 * This variant also removes the masking of the subtraction because the
306 * clocksource mask of all VDSO capable clocksources on x86 is U64_MAX
307 * which would result in a pointless operation. The compiler cannot
308 * optimize it away as the mask comes from the vdso data and is not compile
309 * time constant.
310 */
311 static __always_inline
vdso_calc_delta(u64 cycles,u64 last,u64 mask,u32 mult)312 u64 vdso_calc_delta(u64 cycles, u64 last, u64 mask, u32 mult)
313 {
314 if (cycles > last)
315 return (cycles - last) * mult;
316 return 0;
317 }
318 #define vdso_calc_delta vdso_calc_delta
319
320 #endif /* !__ASSEMBLY__ */
321
322 #endif /* __ASM_VDSO_GETTIMEOFDAY_H */
323