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