1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright 2006 Andi Kleen, SUSE Labs.
4  *
5  * Fast user context implementation of clock_gettime, gettimeofday, and time.
6  *
7  * 32 Bit compat layer by Stefani Seibold <stefani@seibold.net>
8  *  sponsored by Rohde & Schwarz GmbH & Co. KG Munich/Germany
9  *
10  * The code should have no internal unresolved relocations.
11  * Check with readelf after changing.
12  */
13 
14 #include <uapi/linux/time.h>
15 #include <asm/vgtod.h>
16 #include <asm/vvar.h>
17 #include <asm/unistd.h>
18 #include <asm/msr.h>
19 #include <asm/pvclock.h>
20 #include <asm/mshyperv.h>
21 #include <linux/math64.h>
22 #include <linux/time.h>
23 #include <linux/kernel.h>
24 
25 #define gtod (&VVAR(vsyscall_gtod_data))
26 
27 extern int __vdso_clock_gettime(clockid_t clock, struct timespec *ts);
28 extern int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz);
29 extern time_t __vdso_time(time_t *t);
30 
31 #ifdef CONFIG_PARAVIRT_CLOCK
32 extern u8 pvclock_page[PAGE_SIZE]
33 	__attribute__((visibility("hidden")));
34 #endif
35 
36 #ifdef CONFIG_HYPERV_TSCPAGE
37 extern u8 hvclock_page[PAGE_SIZE]
38 	__attribute__((visibility("hidden")));
39 #endif
40 
41 #ifndef BUILD_VDSO32
42 
43 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
44 {
45 	long ret;
46 	asm ("syscall" : "=a" (ret), "=m" (*ts) :
47 	     "0" (__NR_clock_gettime), "D" (clock), "S" (ts) :
48 	     "rcx", "r11");
49 	return ret;
50 }
51 
52 #else
53 
54 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
55 {
56 	long ret;
57 
58 	asm (
59 		"mov %%ebx, %%edx \n"
60 		"mov %[clock], %%ebx \n"
61 		"call __kernel_vsyscall \n"
62 		"mov %%edx, %%ebx \n"
63 		: "=a" (ret), "=m" (*ts)
64 		: "0" (__NR_clock_gettime), [clock] "g" (clock), "c" (ts)
65 		: "edx");
66 	return ret;
67 }
68 
69 #endif
70 
71 #ifdef CONFIG_PARAVIRT_CLOCK
72 static notrace const struct pvclock_vsyscall_time_info *get_pvti0(void)
73 {
74 	return (const struct pvclock_vsyscall_time_info *)&pvclock_page;
75 }
76 
77 static notrace u64 vread_pvclock(void)
78 {
79 	const struct pvclock_vcpu_time_info *pvti = &get_pvti0()->pvti;
80 	u32 version;
81 	u64 ret;
82 
83 	/*
84 	 * Note: The kernel and hypervisor must guarantee that cpu ID
85 	 * number maps 1:1 to per-CPU pvclock time info.
86 	 *
87 	 * Because the hypervisor is entirely unaware of guest userspace
88 	 * preemption, it cannot guarantee that per-CPU pvclock time
89 	 * info is updated if the underlying CPU changes or that that
90 	 * version is increased whenever underlying CPU changes.
91 	 *
92 	 * On KVM, we are guaranteed that pvti updates for any vCPU are
93 	 * atomic as seen by *all* vCPUs.  This is an even stronger
94 	 * guarantee than we get with a normal seqlock.
95 	 *
96 	 * On Xen, we don't appear to have that guarantee, but Xen still
97 	 * supplies a valid seqlock using the version field.
98 	 *
99 	 * We only do pvclock vdso timing at all if
100 	 * PVCLOCK_TSC_STABLE_BIT is set, and we interpret that bit to
101 	 * mean that all vCPUs have matching pvti and that the TSC is
102 	 * synced, so we can just look at vCPU 0's pvti.
103 	 */
104 
105 	do {
106 		version = pvclock_read_begin(pvti);
107 
108 		if (unlikely(!(pvti->flags & PVCLOCK_TSC_STABLE_BIT)))
109 			return U64_MAX;
110 
111 		ret = __pvclock_read_cycles(pvti, rdtsc_ordered());
112 	} while (pvclock_read_retry(pvti, version));
113 
114 	return ret;
115 }
116 #endif
117 #ifdef CONFIG_HYPERV_TSCPAGE
118 static notrace u64 vread_hvclock(void)
119 {
120 	const struct ms_hyperv_tsc_page *tsc_pg =
121 		(const struct ms_hyperv_tsc_page *)&hvclock_page;
122 
123 	return hv_read_tsc_page(tsc_pg);
124 }
125 #endif
126 
127 notrace static inline u64 vgetcyc(int mode)
128 {
129 	if (mode == VCLOCK_TSC)
130 		return (u64)rdtsc_ordered();
131 
132 	/*
133 	 * For any memory-mapped vclock type, we need to make sure that gcc
134 	 * doesn't cleverly hoist a load before the mode check.  Otherwise we
135 	 * might end up touching the memory-mapped page even if the vclock in
136 	 * question isn't enabled, which will segfault.  Hence the barriers.
137 	 */
138 #ifdef CONFIG_PARAVIRT_CLOCK
139 	if (mode == VCLOCK_PVCLOCK) {
140 		barrier();
141 		return vread_pvclock();
142 	}
143 #endif
144 #ifdef CONFIG_HYPERV_TSCPAGE
145 	if (mode == VCLOCK_HVCLOCK) {
146 		barrier();
147 		return vread_hvclock();
148 	}
149 #endif
150 	return U64_MAX;
151 }
152 
153 notrace static int do_hres(clockid_t clk, struct timespec *ts)
154 {
155 	struct vgtod_ts *base = &gtod->basetime[clk];
156 	u64 cycles, last, sec, ns;
157 	unsigned int seq;
158 
159 	do {
160 		seq = gtod_read_begin(gtod);
161 		cycles = vgetcyc(gtod->vclock_mode);
162 		ns = base->nsec;
163 		last = gtod->cycle_last;
164 		if (unlikely((s64)cycles < 0))
165 			return vdso_fallback_gettime(clk, ts);
166 		if (cycles > last)
167 			ns += (cycles - last) * gtod->mult;
168 		ns >>= gtod->shift;
169 		sec = base->sec;
170 	} while (unlikely(gtod_read_retry(gtod, seq)));
171 
172 	/*
173 	 * Do this outside the loop: a race inside the loop could result
174 	 * in __iter_div_u64_rem() being extremely slow.
175 	 */
176 	ts->tv_sec = sec + __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
177 	ts->tv_nsec = ns;
178 
179 	return 0;
180 }
181 
182 notrace static void do_coarse(clockid_t clk, struct timespec *ts)
183 {
184 	struct vgtod_ts *base = &gtod->basetime[clk];
185 	unsigned int seq;
186 
187 	do {
188 		seq = gtod_read_begin(gtod);
189 		ts->tv_sec = base->sec;
190 		ts->tv_nsec = base->nsec;
191 	} while (unlikely(gtod_read_retry(gtod, seq)));
192 }
193 
194 notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
195 {
196 	unsigned int msk;
197 
198 	/* Sort out negative (CPU/FD) and invalid clocks */
199 	if (unlikely((unsigned int) clock >= MAX_CLOCKS))
200 		return vdso_fallback_gettime(clock, ts);
201 
202 	/*
203 	 * Convert the clockid to a bitmask and use it to check which
204 	 * clocks are handled in the VDSO directly.
205 	 */
206 	msk = 1U << clock;
207 	if (likely(msk & VGTOD_HRES)) {
208 		return do_hres(clock, ts);
209 	} else if (msk & VGTOD_COARSE) {
210 		do_coarse(clock, ts);
211 		return 0;
212 	}
213 	return vdso_fallback_gettime(clock, ts);
214 }
215 
216 int clock_gettime(clockid_t, struct timespec *)
217 	__attribute__((weak, alias("__vdso_clock_gettime")));
218 
219 notrace int __vdso_gettimeofday(struct timeval *tv, struct timezone *tz)
220 {
221 	if (likely(tv != NULL)) {
222 		struct timespec *ts = (struct timespec *) tv;
223 
224 		do_hres(CLOCK_REALTIME, ts);
225 		tv->tv_usec /= 1000;
226 	}
227 	if (unlikely(tz != NULL)) {
228 		tz->tz_minuteswest = gtod->tz_minuteswest;
229 		tz->tz_dsttime = gtod->tz_dsttime;
230 	}
231 
232 	return 0;
233 }
234 int gettimeofday(struct timeval *, struct timezone *)
235 	__attribute__((weak, alias("__vdso_gettimeofday")));
236 
237 /*
238  * This will break when the xtime seconds get inaccurate, but that is
239  * unlikely
240  */
241 notrace time_t __vdso_time(time_t *t)
242 {
243 	/* This is atomic on x86 so we don't need any locks. */
244 	time_t result = READ_ONCE(gtod->basetime[CLOCK_REALTIME].sec);
245 
246 	if (t)
247 		*t = result;
248 	return result;
249 }
250 time_t time(time_t *t)
251 	__attribute__((weak, alias("__vdso_time")));
252