xref: /linux/arch/x86/include/asm/vdso/gettimeofday.h (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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/unistd.h>
18 #include <asm/msr.h>
19 #include <asm/pvclock.h>
20 #include <clocksource/hyperv_timer.h>
21 
22 extern struct vdso_data vvar_page
23 	__attribute__((visibility("hidden")));
24 
25 extern struct vdso_data timens_page
26 	__attribute__((visibility("hidden")));
27 
28 #define VDSO_HAS_TIME 1
29 
30 #define VDSO_HAS_CLOCK_GETRES 1
31 
32 /*
33  * Declare the memory-mapped vclock data pages.  These come from hypervisors.
34  * If we ever reintroduce something like direct access to an MMIO clock like
35  * the HPET again, it will go here as well.
36  *
37  * A load from any of these pages will segfault if the clock in question is
38  * disabled, so appropriate compiler barriers and checks need to be used
39  * to prevent stray loads.
40  *
41  * These declarations MUST NOT be const.  The compiler will assume that
42  * an extern const variable has genuinely constant contents, and the
43  * resulting code won't work, since the whole point is that these pages
44  * change over time, possibly while we're accessing them.
45  */
46 
47 #ifdef CONFIG_PARAVIRT_CLOCK
48 /*
49  * This is the vCPU 0 pvclock page.  We only use pvclock from the vDSO
50  * if the hypervisor tells us that all vCPUs can get valid data from the
51  * vCPU 0 page.
52  */
53 extern struct pvclock_vsyscall_time_info pvclock_page
54 	__attribute__((visibility("hidden")));
55 #endif
56 
57 #ifdef CONFIG_HYPERV_TIMER
58 extern struct ms_hyperv_tsc_page hvclock_page
59 	__attribute__((visibility("hidden")));
60 #endif
61 
62 #ifdef CONFIG_TIME_NS
63 static __always_inline
64 const struct vdso_data *__arch_get_timens_vdso_data(const struct vdso_data *vd)
65 {
66 	return &timens_page;
67 }
68 #endif
69 
70 #ifndef BUILD_VDSO32
71 
72 static __always_inline
73 long clock_gettime_fallback(clockid_t _clkid, struct __kernel_timespec *_ts)
74 {
75 	long ret;
76 
77 	asm ("syscall" : "=a" (ret), "=m" (*_ts) :
78 	     "0" (__NR_clock_gettime), "D" (_clkid), "S" (_ts) :
79 	     "rcx", "r11");
80 
81 	return ret;
82 }
83 
84 static __always_inline
85 long gettimeofday_fallback(struct __kernel_old_timeval *_tv,
86 			   struct timezone *_tz)
87 {
88 	long ret;
89 
90 	asm("syscall" : "=a" (ret) :
91 	    "0" (__NR_gettimeofday), "D" (_tv), "S" (_tz) : "memory");
92 
93 	return ret;
94 }
95 
96 static __always_inline
97 long clock_getres_fallback(clockid_t _clkid, struct __kernel_timespec *_ts)
98 {
99 	long ret;
100 
101 	asm ("syscall" : "=a" (ret), "=m" (*_ts) :
102 	     "0" (__NR_clock_getres), "D" (_clkid), "S" (_ts) :
103 	     "rcx", "r11");
104 
105 	return ret;
106 }
107 
108 #else
109 
110 static __always_inline
111 long clock_gettime_fallback(clockid_t _clkid, struct __kernel_timespec *_ts)
112 {
113 	long ret;
114 
115 	asm (
116 		"mov %%ebx, %%edx \n"
117 		"mov %[clock], %%ebx \n"
118 		"call __kernel_vsyscall \n"
119 		"mov %%edx, %%ebx \n"
120 		: "=a" (ret), "=m" (*_ts)
121 		: "0" (__NR_clock_gettime64), [clock] "g" (_clkid), "c" (_ts)
122 		: "edx");
123 
124 	return ret;
125 }
126 
127 static __always_inline
128 long clock_gettime32_fallback(clockid_t _clkid, struct old_timespec32 *_ts)
129 {
130 	long ret;
131 
132 	asm (
133 		"mov %%ebx, %%edx \n"
134 		"mov %[clock], %%ebx \n"
135 		"call __kernel_vsyscall \n"
136 		"mov %%edx, %%ebx \n"
137 		: "=a" (ret), "=m" (*_ts)
138 		: "0" (__NR_clock_gettime), [clock] "g" (_clkid), "c" (_ts)
139 		: "edx");
140 
141 	return ret;
142 }
143 
144 static __always_inline
145 long gettimeofday_fallback(struct __kernel_old_timeval *_tv,
146 			   struct timezone *_tz)
147 {
148 	long ret;
149 
150 	asm(
151 		"mov %%ebx, %%edx \n"
152 		"mov %2, %%ebx \n"
153 		"call __kernel_vsyscall \n"
154 		"mov %%edx, %%ebx \n"
155 		: "=a" (ret)
156 		: "0" (__NR_gettimeofday), "g" (_tv), "c" (_tz)
157 		: "memory", "edx");
158 
159 	return ret;
160 }
161 
162 static __always_inline long
163 clock_getres_fallback(clockid_t _clkid, struct __kernel_timespec *_ts)
164 {
165 	long ret;
166 
167 	asm (
168 		"mov %%ebx, %%edx \n"
169 		"mov %[clock], %%ebx \n"
170 		"call __kernel_vsyscall \n"
171 		"mov %%edx, %%ebx \n"
172 		: "=a" (ret), "=m" (*_ts)
173 		: "0" (__NR_clock_getres_time64), [clock] "g" (_clkid), "c" (_ts)
174 		: "edx");
175 
176 	return ret;
177 }
178 
179 static __always_inline
180 long clock_getres32_fallback(clockid_t _clkid, struct old_timespec32 *_ts)
181 {
182 	long ret;
183 
184 	asm (
185 		"mov %%ebx, %%edx \n"
186 		"mov %[clock], %%ebx \n"
187 		"call __kernel_vsyscall \n"
188 		"mov %%edx, %%ebx \n"
189 		: "=a" (ret), "=m" (*_ts)
190 		: "0" (__NR_clock_getres), [clock] "g" (_clkid), "c" (_ts)
191 		: "edx");
192 
193 	return ret;
194 }
195 
196 #endif
197 
198 #ifdef CONFIG_PARAVIRT_CLOCK
199 static u64 vread_pvclock(void)
200 {
201 	const struct pvclock_vcpu_time_info *pvti = &pvclock_page.pvti;
202 	u32 version;
203 	u64 ret;
204 
205 	/*
206 	 * Note: The kernel and hypervisor must guarantee that cpu ID
207 	 * number maps 1:1 to per-CPU pvclock time info.
208 	 *
209 	 * Because the hypervisor is entirely unaware of guest userspace
210 	 * preemption, it cannot guarantee that per-CPU pvclock time
211 	 * info is updated if the underlying CPU changes or that that
212 	 * version is increased whenever underlying CPU changes.
213 	 *
214 	 * On KVM, we are guaranteed that pvti updates for any vCPU are
215 	 * atomic as seen by *all* vCPUs.  This is an even stronger
216 	 * guarantee than we get with a normal seqlock.
217 	 *
218 	 * On Xen, we don't appear to have that guarantee, but Xen still
219 	 * supplies a valid seqlock using the version field.
220 	 *
221 	 * We only do pvclock vdso timing at all if
222 	 * PVCLOCK_TSC_STABLE_BIT is set, and we interpret that bit to
223 	 * mean that all vCPUs have matching pvti and that the TSC is
224 	 * synced, so we can just look at vCPU 0's pvti.
225 	 */
226 
227 	do {
228 		version = pvclock_read_begin(pvti);
229 
230 		if (unlikely(!(pvti->flags & PVCLOCK_TSC_STABLE_BIT)))
231 			return U64_MAX;
232 
233 		ret = __pvclock_read_cycles(pvti, rdtsc_ordered());
234 	} while (pvclock_read_retry(pvti, version));
235 
236 	return ret & S64_MAX;
237 }
238 #endif
239 
240 #ifdef CONFIG_HYPERV_TIMER
241 static u64 vread_hvclock(void)
242 {
243 	u64 tsc, time;
244 
245 	if (hv_read_tsc_page_tsc(&hvclock_page, &tsc, &time))
246 		return time & S64_MAX;
247 
248 	return U64_MAX;
249 }
250 #endif
251 
252 static inline u64 __arch_get_hw_counter(s32 clock_mode,
253 					const struct vdso_data *vd)
254 {
255 	if (likely(clock_mode == VDSO_CLOCKMODE_TSC))
256 		return (u64)rdtsc_ordered() & S64_MAX;
257 	/*
258 	 * For any memory-mapped vclock type, we need to make sure that gcc
259 	 * doesn't cleverly hoist a load before the mode check.  Otherwise we
260 	 * might end up touching the memory-mapped page even if the vclock in
261 	 * question isn't enabled, which will segfault.  Hence the barriers.
262 	 */
263 #ifdef CONFIG_PARAVIRT_CLOCK
264 	if (clock_mode == VDSO_CLOCKMODE_PVCLOCK) {
265 		barrier();
266 		return vread_pvclock();
267 	}
268 #endif
269 #ifdef CONFIG_HYPERV_TIMER
270 	if (clock_mode == VDSO_CLOCKMODE_HVCLOCK) {
271 		barrier();
272 		return vread_hvclock();
273 	}
274 #endif
275 	return U64_MAX;
276 }
277 
278 static __always_inline const struct vdso_data *__arch_get_vdso_data(void)
279 {
280 	return &vvar_page;
281 }
282 
283 static inline bool arch_vdso_clocksource_ok(const struct vdso_data *vd)
284 {
285 	return true;
286 }
287 #define vdso_clocksource_ok arch_vdso_clocksource_ok
288 
289 /*
290  * Clocksource read value validation to handle PV and HyperV clocksources
291  * which can be invalidated asynchronously and indicate invalidation by
292  * returning U64_MAX, which can be effectively tested by checking for a
293  * negative value after casting it to s64.
294  *
295  * This effectively forces a S64_MAX mask on the calculations, unlike the
296  * U64_MAX mask normally used by x86 clocksources.
297  */
298 static inline bool arch_vdso_cycles_ok(u64 cycles)
299 {
300 	return (s64)cycles >= 0;
301 }
302 #define vdso_cycles_ok arch_vdso_cycles_ok
303 
304 /*
305  * x86 specific calculation of nanoseconds for the current cycle count
306  *
307  * The regular implementation assumes that clocksource reads are globally
308  * monotonic. The TSC can be slightly off across sockets which can cause
309  * the regular delta calculation (@cycles - @last) to return a huge time
310  * jump.
311  *
312  * Therefore it needs to be verified that @cycles are greater than
313  * @vd->cycles_last. If not then use @vd->cycles_last, which is the base
314  * time of the current conversion period.
315  *
316  * This variant also uses a custom mask because while the clocksource mask of
317  * all the VDSO capable clocksources on x86 is U64_MAX, the above code uses
318  * U64_MASK as an exception value, additionally arch_vdso_cycles_ok() above
319  * declares everything with the MSB/Sign-bit set as invalid. Therefore the
320  * effective mask is S64_MAX.
321  */
322 static __always_inline u64 vdso_calc_ns(const struct vdso_data *vd, u64 cycles, u64 base)
323 {
324 	u64 delta = cycles - vd->cycle_last;
325 
326 	/*
327 	 * Negative motion and deltas which can cause multiplication
328 	 * overflow require special treatment. This check covers both as
329 	 * negative motion is guaranteed to be greater than @vd::max_cycles
330 	 * due to unsigned comparison.
331 	 *
332 	 * Due to the MSB/Sign-bit being used as invalid marker (see
333 	 * arch_vdso_cycles_ok() above), the effective mask is S64_MAX, but that
334 	 * case is also unlikely and will also take the unlikely path here.
335 	 */
336 	if (unlikely(delta > vd->max_cycles)) {
337 		/*
338 		 * Due to the above mentioned TSC wobbles, filter out
339 		 * negative motion.  Per the above masking, the effective
340 		 * sign bit is now bit 62.
341 		 */
342 		if (delta & (1ULL << 62))
343 			return base >> vd->shift;
344 
345 		/* Handle multiplication overflow gracefully */
346 		return mul_u64_u32_add_u64_shr(delta & S64_MAX, vd->mult, base, vd->shift);
347 	}
348 
349 	return ((delta * vd->mult) + base) >> vd->shift;
350 }
351 #define vdso_calc_ns vdso_calc_ns
352 
353 #endif /* !__ASSEMBLY__ */
354 
355 #endif /* __ASM_VDSO_GETTIMEOFDAY_H */
356