xref: /illumos-gate/usr/src/lib/commpage/common/cp_main.c (revision c3937c08b16d32677a7dc955a81325f2e7a61b5d)
1 /*
2  * This file and its contents are supplied under the terms of the
3  * Common Development and Distribution License ("CDDL"), version 1.0.
4  * You may only use this file in accordance with the terms of version
5  * 1.0 of the CDDL.
6  *
7  * A full copy of the text of the CDDL should have accompanied this
8  * source.  A copy of the CDDL is also available via the Internet at
9  * http://www.illumos.org/license/CDDL.
10  */
11 
12 /*
13  * Copyright 2019 Joyent, Inc.
14  */
15 
16 #include <sys/comm_page.h>
17 #include <sys/tsc.h>
18 
19 
20 /*
21  * Interrogate if querying the clock via the comm page is possible.
22  */
23 int
24 __cp_can_gettime(comm_page_t *cp)
25 {
26 	switch (cp->cp_tsc_type) {
27 	case TSC_TSCP:
28 	case TSC_RDTSC_MFENCE:
29 	case TSC_RDTSC_LFENCE:
30 	case TSC_RDTSC_CPUID:
31 		return (1);
32 	default:
33 		break;
34 	}
35 	return (0);
36 }
37 
38 #ifdef __amd64
39 
40 /*
41  * The functions used for calculating time (both monotonic and wall-clock) are
42  * implemented in assembly on amd64.  This is primarily for stack conservation.
43  */
44 
45 #else /* i386 below */
46 
47 /*
48  * ASM-defined functions.
49  */
50 extern hrtime_t __cp_tsc_read(comm_page_t *);
51 extern hrtime_t __cp_gethrtime_fasttrap();
52 
53 /*
54  * These are cloned from TSC and time related code in the kernel.  The should
55  * be kept in sync in the case that the source values are changed.
56  */
57 #define	NSEC_SHIFT	5
58 #define	ADJ_SHIFT	4
59 #define	NANOSEC		1000000000LL
60 
61 #define	TSC_CONVERT_AND_ADD(tsc, hrt, scale) do {		\
62 	uint32_t *_l = (uint32_t *)&(tsc);			\
63 	uint64_t sc = (uint32_t)(scale);			\
64 	(hrt) += (uint64_t)(_l[1] * sc) << NSEC_SHIFT;		\
65 	(hrt) += (uint64_t)(_l[0] * sc) >> (32 - NSEC_SHIFT);	\
66 } while (0)
67 
68 /*
69  * Userspace version of tsc_gethrtime.
70  * See: uts/i86pc/os/timestamp.c
71  */
72 hrtime_t
73 __cp_gethrtime(comm_page_t *cp)
74 {
75 	uint32_t old_hres_lock;
76 	hrtime_t tsc, hrt, tsc_last;
77 
78 	/*
79 	 * Several precautions must be taken when collecting the data necessary
80 	 * to perform an accurate gethrtime calculation.
81 	 *
82 	 * While much of the TSC state stored in the comm page is unchanging
83 	 * after boot, portions of it are periodically updated during OS ticks.
84 	 * Changes to hres_lock during the course of the copy indicates a
85 	 * potentially inconsistent snapshot, necessitating a loop.
86 	 *
87 	 * Even more complicated is the handling for TSCs which require sync
88 	 * offsets between different CPUs.  Since userspace lacks the luxury of
89 	 * disabling interrupts, a validation loop checking for CPU migrations
90 	 * is used.  Pathological scheduling could, in theory, "outwit"
91 	 * this check.  Such a possibility is considered an acceptable risk.
92 	 *
93 	 */
94 	do {
95 		old_hres_lock = cp->cp_hres_lock;
96 		tsc_last = cp->cp_tsc_last;
97 		hrt = cp->cp_tsc_hrtime_base;
98 		tsc = __cp_tsc_read(cp);
99 
100 		/*
101 		 * A TSC reading of 0 indicates the special case of an error
102 		 * bail-out.  Rely on the fasttrap to supply an hrtime value.
103 		 */
104 		if (tsc == 0) {
105 			return (__cp_gethrtime_fasttrap());
106 		}
107 	} while ((old_hres_lock & ~1) != cp->cp_hres_lock);
108 
109 	if (tsc >= tsc_last) {
110 		tsc -= tsc_last;
111 	} else if (tsc >= tsc_last - (2 * cp->cp_tsc_max_delta)) {
112 		tsc = 0;
113 	} else if (tsc > cp->cp_tsc_resume_cap) {
114 		tsc = cp->cp_tsc_resume_cap;
115 	}
116 	TSC_CONVERT_AND_ADD(tsc, hrt, cp->cp_nsec_scale);
117 
118 	return (hrt);
119 }
120 
121 /*
122  * Userspace version of pc_gethrestime.
123  * See: uts/i86pc/os/machdep.c
124  */
125 int
126 __cp_clock_gettime_realtime(comm_page_t *cp, timespec_t *tsp)
127 {
128 	int lock_prev, nslt;
129 	timespec_t now;
130 	int64_t hres_adj;
131 
132 loop:
133 	lock_prev = cp->cp_hres_lock;
134 	now.tv_sec = cp->cp_hrestime[0];
135 	now.tv_nsec = cp->cp_hrestime[1];
136 	nslt = (int)(__cp_gethrtime(cp) - cp->cp_hres_last_tick);
137 	hres_adj = cp->cp_hrestime_adj;
138 	if (nslt < 0) {
139 		/*
140 		 * Tick came between sampling hrtime and hres_last_tick;
141 		 */
142 		goto loop;
143 	}
144 	now.tv_nsec += nslt;
145 
146 	/*
147 	 * Apply hres_adj skew, if needed.
148 	 */
149 	if (hres_adj > 0) {
150 		nslt = (nslt >> ADJ_SHIFT);
151 		if (nslt > hres_adj)
152 			nslt = (int)hres_adj;
153 		now.tv_nsec += nslt;
154 	} else if (hres_adj < 0) {
155 		nslt = -(nslt >> ADJ_SHIFT);
156 		if (nslt < hres_adj)
157 			nslt = (int)hres_adj;
158 		now.tv_nsec += nslt;
159 	}
160 
161 	/*
162 	 * Rope in tv_nsec from any excessive adjustments.
163 	 */
164 	while ((unsigned long)now.tv_nsec >= NANOSEC) {
165 		now.tv_nsec -= NANOSEC;
166 		now.tv_sec++;
167 	}
168 
169 	if ((cp->cp_hres_lock & ~1) != lock_prev)
170 		goto loop;
171 
172 	*tsp = now;
173 	return (0);
174 }
175 
176 /*
177  * The __cp_clock_gettime_monotonic function expects that hrt2ts be present
178  * when the code is finally linked.
179  * (The amd64 version has no such requirement.)
180  */
181 extern void hrt2ts(hrtime_t, timespec_t *);
182 
183 int
184 __cp_clock_gettime_monotonic(comm_page_t *cp, timespec_t *tsp)
185 {
186 	hrtime_t hrt;
187 
188 	hrt = __cp_gethrtime(cp);
189 	hrt2ts(hrt, tsp);
190 	return (0);
191 }
192 
193 #endif /* __amd64 */
194