1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  * Copyright 2016 Joyent, Inc.
26  */
27 
28 #include <sys/types.h>
29 #include <sys/clock.h>
30 #include <sys/panic.h>
31 #include <sys/atomic.h>
32 #include <sys/hypervisor.h>
33 
34 #include <sys/archsystm.h>
35 
36 /*
37  * On the hypervisor, we have a virtualized system time based upon the
38  * information provided for each VCPU, which is updated every time it is
39  * scheduled onto a real CPU.  Thus, none of the traditional code in
40  * i86pc/os/timestamp.c applies, our gethrtime() implementation is run through
41  * the PSM, and there is no scaling step to apply.
42  *
43  * However, the platform does not guarantee monotonicity; thus we have to fake
44  * this up, which is a deeply unpleasant thing to have to do.
45  *
46  * Note that the virtualized interface still relies on the current TSC to
47  * calculate the time in nanoseconds since the VCPU was scheduled, and is thus
48  * subject to all the problems with that.  For the most part, the hypervisor is
49  * supposed to deal with them.
50  *
51  * Another wrinkle involves suspend/resume/migration.  If we come back and time
52  * is apparently less, we may have resumed on a different machine or on the
53  * same machine after a reboot.  In this case we need to maintain an addend to
54  * ensure time continues reasonably.  Otherwise we could end up taking a very
55  * long time to expire cyclics in the heap.  Thus we have two functions:
56  *
57  * xpv_getsystime()
58  *
59  *	The unadulterated system time from the hypervisor.  This is only to be
60  *	used when programming the hypervisor (setting a timer or calculating
61  *	the TOD).
62  *
63  * xpv_gethrtime()
64  *
65  *	This is the monotonic hrtime counter to be used by everything else such
66  *	as the cyclic subsystem.  We should never pass an hrtime directly into
67  *	a hypervisor interface, as hrtime_addend may well be non-zero.
68  */
69 
70 int hrtime_fake_mt = 1;
71 static volatile hrtime_t hrtime_last;
72 static hrtime_t hrtime_suspend_time;
73 static hrtime_t hrtime_addend;
74 
75 volatile uint32_t hres_lock;
76 hrtime_t hres_last_tick;
77 int64_t hrestime_adj;
78 volatile timestruc_t hrestime;
79 
80 /*
81  * These functions are used in DTrace probe context, and must be removed from
82  * fbt consideration.  Currently fbt ignores all weak symbols, so this will
83  * achieve that.
84  */
85 #pragma weak xpv_gethrtime = dtrace_xpv_gethrtime
86 #pragma weak xpv_getsystime = dtrace_xpv_getsystime
87 #pragma weak dtrace_gethrtime = dtrace_xpv_gethrtime
88 #pragma weak tsc_read = dtrace_xpv_gethrtime
89 
90 hrtime_t
dtrace_xpv_getsystime(void)91 dtrace_xpv_getsystime(void)
92 {
93 	vcpu_time_info_t *src;
94 	vcpu_time_info_t __vti, *dst = &__vti;
95 	uint64_t tsc_delta;
96 	uint64_t tsc;
97 	hrtime_t result;
98 	uint32_t stamp;
99 
100 	src = &CPU->cpu_m.mcpu_vcpu_info->time;
101 
102 	/*
103 	 * Loop until version has not been changed during our update, and a Xen
104 	 * update is not under way (lowest bit is set).
105 	 */
106 	do {
107 		dst->version = src->version;
108 		stamp = CPU->cpu_m.mcpu_istamp;
109 
110 		membar_consumer();
111 
112 		dst->tsc_timestamp = src->tsc_timestamp;
113 		dst->system_time = src->system_time;
114 		dst->tsc_to_system_mul = src->tsc_to_system_mul;
115 		dst->tsc_shift = src->tsc_shift;
116 
117 		/*
118 		 * Note that this use of the -actual- TSC register
119 		 * should probably be the SOLE one in the system on this
120 		 * paravirtualized platform.
121 		 */
122 		tsc = __rdtsc_insn();
123 		tsc_delta = tsc - dst->tsc_timestamp;
124 
125 		membar_consumer();
126 
127 	} while (((src->version & 1) | (dst->version ^ src->version)) ||
128 	    CPU->cpu_m.mcpu_istamp != stamp);
129 
130 	if (dst->tsc_shift >= 0)
131 		tsc_delta <<= dst->tsc_shift;
132 	else if (dst->tsc_shift < 0)
133 		tsc_delta >>= -dst->tsc_shift;
134 
135 	result = dst->system_time +
136 	    ((uint64_t)(tsc_delta * (uint64_t)dst->tsc_to_system_mul) >> 32);
137 
138 	return (result);
139 }
140 
141 hrtime_t
dtrace_xpv_gethrtime(void)142 dtrace_xpv_gethrtime(void)
143 {
144 	hrtime_t result = xpv_getsystime() + hrtime_addend;
145 
146 	if (hrtime_fake_mt) {
147 		hrtime_t last;
148 		do {
149 			last = hrtime_last;
150 			if (result < last)
151 				result = last + 1;
152 		} while (atomic_cas_64((volatile uint64_t *)&hrtime_last,
153 		    last, result) != last);
154 	}
155 
156 	return (result);
157 }
158 
159 void
xpv_time_suspend(void)160 xpv_time_suspend(void)
161 {
162 	hrtime_suspend_time = xpv_getsystime();
163 }
164 
165 void
xpv_time_resume(void)166 xpv_time_resume(void)
167 {
168 	hrtime_t delta = xpv_getsystime() - hrtime_suspend_time;
169 
170 	if (delta < 0)
171 		hrtime_addend += -delta;
172 }
173