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