x86: kvmclock: abstract save/restore sched_clock_state
[~shefty/rdma-dev.git] / arch / x86 / kernel / kvmclock.c
1 /*  KVM paravirtual clock driver. A clocksource implementation
2     Copyright (C) 2008 Glauber de Oliveira Costa, Red Hat Inc.
3
4     This program is free software; you can redistribute it and/or modify
5     it under the terms of the GNU General Public License as published by
6     the Free Software Foundation; either version 2 of the License, or
7     (at your option) any later version.
8
9     This program is distributed in the hope that it will be useful,
10     but WITHOUT ANY WARRANTY; without even the implied warranty of
11     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12     GNU General Public License for more details.
13
14     You should have received a copy of the GNU General Public License
15     along with this program; if not, write to the Free Software
16     Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17 */
18
19 #include <linux/clocksource.h>
20 #include <linux/kvm_para.h>
21 #include <asm/pvclock.h>
22 #include <asm/msr.h>
23 #include <asm/apic.h>
24 #include <linux/percpu.h>
25
26 #include <asm/x86_init.h>
27 #include <asm/reboot.h>
28
29 static int kvmclock = 1;
30 static int msr_kvm_system_time = MSR_KVM_SYSTEM_TIME;
31 static int msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK;
32
33 static int parse_no_kvmclock(char *arg)
34 {
35         kvmclock = 0;
36         return 0;
37 }
38 early_param("no-kvmclock", parse_no_kvmclock);
39
40 /* The hypervisor will put information about time periodically here */
41 static DEFINE_PER_CPU_SHARED_ALIGNED(struct pvclock_vcpu_time_info, hv_clock);
42 static struct pvclock_wall_clock wall_clock;
43
44 /*
45  * The wallclock is the time of day when we booted. Since then, some time may
46  * have elapsed since the hypervisor wrote the data. So we try to account for
47  * that with system time
48  */
49 static unsigned long kvm_get_wallclock(void)
50 {
51         struct pvclock_vcpu_time_info *vcpu_time;
52         struct timespec ts;
53         int low, high;
54
55         low = (int)__pa_symbol(&wall_clock);
56         high = ((u64)__pa_symbol(&wall_clock) >> 32);
57
58         native_write_msr(msr_kvm_wall_clock, low, high);
59
60         vcpu_time = &get_cpu_var(hv_clock);
61         pvclock_read_wallclock(&wall_clock, vcpu_time, &ts);
62         put_cpu_var(hv_clock);
63
64         return ts.tv_sec;
65 }
66
67 static int kvm_set_wallclock(unsigned long now)
68 {
69         return -1;
70 }
71
72 static cycle_t kvm_clock_read(void)
73 {
74         struct pvclock_vcpu_time_info *src;
75         cycle_t ret;
76
77         preempt_disable_notrace();
78         src = &__get_cpu_var(hv_clock);
79         ret = pvclock_clocksource_read(src);
80         preempt_enable_notrace();
81         return ret;
82 }
83
84 static cycle_t kvm_clock_get_cycles(struct clocksource *cs)
85 {
86         return kvm_clock_read();
87 }
88
89 /*
90  * If we don't do that, there is the possibility that the guest
91  * will calibrate under heavy load - thus, getting a lower lpj -
92  * and execute the delays themselves without load. This is wrong,
93  * because no delay loop can finish beforehand.
94  * Any heuristics is subject to fail, because ultimately, a large
95  * poll of guests can be running and trouble each other. So we preset
96  * lpj here
97  */
98 static unsigned long kvm_get_tsc_khz(void)
99 {
100         struct pvclock_vcpu_time_info *src;
101         src = &per_cpu(hv_clock, 0);
102         return pvclock_tsc_khz(src);
103 }
104
105 static void kvm_get_preset_lpj(void)
106 {
107         unsigned long khz;
108         u64 lpj;
109
110         khz = kvm_get_tsc_khz();
111
112         lpj = ((u64)khz * 1000);
113         do_div(lpj, HZ);
114         preset_lpj = lpj;
115 }
116
117 static struct clocksource kvm_clock = {
118         .name = "kvm-clock",
119         .read = kvm_clock_get_cycles,
120         .rating = 400,
121         .mask = CLOCKSOURCE_MASK(64),
122         .flags = CLOCK_SOURCE_IS_CONTINUOUS,
123 };
124
125 int kvm_register_clock(char *txt)
126 {
127         int cpu = smp_processor_id();
128         int low, high, ret;
129
130         low = (int)__pa(&per_cpu(hv_clock, cpu)) | 1;
131         high = ((u64)__pa(&per_cpu(hv_clock, cpu)) >> 32);
132         ret = native_write_msr_safe(msr_kvm_system_time, low, high);
133         printk(KERN_INFO "kvm-clock: cpu %d, msr %x:%x, %s\n",
134                cpu, high, low, txt);
135
136         return ret;
137 }
138
139 static void kvm_save_sched_clock_state(void)
140 {
141 }
142
143 static void kvm_restore_sched_clock_state(void)
144 {
145         kvm_register_clock("primary cpu clock, resume");
146 }
147
148 #ifdef CONFIG_X86_LOCAL_APIC
149 static void __cpuinit kvm_setup_secondary_clock(void)
150 {
151         /*
152          * Now that the first cpu already had this clocksource initialized,
153          * we shouldn't fail.
154          */
155         WARN_ON(kvm_register_clock("secondary cpu clock"));
156 }
157 #endif
158
159 /*
160  * After the clock is registered, the host will keep writing to the
161  * registered memory location. If the guest happens to shutdown, this memory
162  * won't be valid. In cases like kexec, in which you install a new kernel, this
163  * means a random memory location will be kept being written. So before any
164  * kind of shutdown from our side, we unregister the clock by writting anything
165  * that does not have the 'enable' bit set in the msr
166  */
167 #ifdef CONFIG_KEXEC
168 static void kvm_crash_shutdown(struct pt_regs *regs)
169 {
170         native_write_msr(msr_kvm_system_time, 0, 0);
171         kvm_disable_steal_time();
172         native_machine_crash_shutdown(regs);
173 }
174 #endif
175
176 static void kvm_shutdown(void)
177 {
178         native_write_msr(msr_kvm_system_time, 0, 0);
179         kvm_disable_steal_time();
180         native_machine_shutdown();
181 }
182
183 void __init kvmclock_init(void)
184 {
185         if (!kvm_para_available())
186                 return;
187
188         if (kvmclock && kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) {
189                 msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW;
190                 msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW;
191         } else if (!(kvmclock && kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)))
192                 return;
193
194         printk(KERN_INFO "kvm-clock: Using msrs %x and %x",
195                 msr_kvm_system_time, msr_kvm_wall_clock);
196
197         if (kvm_register_clock("boot clock"))
198                 return;
199         pv_time_ops.sched_clock = kvm_clock_read;
200         x86_platform.calibrate_tsc = kvm_get_tsc_khz;
201         x86_platform.get_wallclock = kvm_get_wallclock;
202         x86_platform.set_wallclock = kvm_set_wallclock;
203 #ifdef CONFIG_X86_LOCAL_APIC
204         x86_cpuinit.early_percpu_clock_init =
205                 kvm_setup_secondary_clock;
206 #endif
207         x86_platform.save_sched_clock_state = kvm_save_sched_clock_state;
208         x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state;
209         machine_ops.shutdown  = kvm_shutdown;
210 #ifdef CONFIG_KEXEC
211         machine_ops.crash_shutdown  = kvm_crash_shutdown;
212 #endif
213         kvm_get_preset_lpj();
214         clocksource_register_hz(&kvm_clock, NSEC_PER_SEC);
215         pv_info.paravirt_enabled = 1;
216         pv_info.name = "KVM";
217
218         if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT))
219                 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT);
220 }