Loading...
1/*
2 * check TSC synchronization.
3 *
4 * Copyright (C) 2006, Red Hat, Inc., Ingo Molnar
5 *
6 * We check whether all boot CPUs have their TSC's synchronized,
7 * print a warning if not and turn off the TSC clock-source.
8 *
9 * The warp-check is point-to-point between two CPUs, the CPU
10 * initiating the bootup is the 'source CPU', the freshly booting
11 * CPU is the 'target CPU'.
12 *
13 * Only two CPUs may participate - they can enter in any order.
14 * ( The serial nature of the boot logic and the CPU hotplug lock
15 * protects against more than 2 CPUs entering this code. )
16 */
17#include <linux/spinlock.h>
18#include <linux/kernel.h>
19#include <linux/init.h>
20#include <linux/smp.h>
21#include <linux/nmi.h>
22#include <asm/tsc.h>
23
24/*
25 * Entry/exit counters that make sure that both CPUs
26 * run the measurement code at once:
27 */
28static __cpuinitdata atomic_t start_count;
29static __cpuinitdata atomic_t stop_count;
30
31/*
32 * We use a raw spinlock in this exceptional case, because
33 * we want to have the fastest, inlined, non-debug version
34 * of a critical section, to be able to prove TSC time-warps:
35 */
36static __cpuinitdata arch_spinlock_t sync_lock = __ARCH_SPIN_LOCK_UNLOCKED;
37
38static __cpuinitdata cycles_t last_tsc;
39static __cpuinitdata cycles_t max_warp;
40static __cpuinitdata int nr_warps;
41
42/*
43 * TSC-warp measurement loop running on both CPUs:
44 */
45static __cpuinit void check_tsc_warp(unsigned int timeout)
46{
47 cycles_t start, now, prev, end;
48 int i;
49
50 rdtsc_barrier();
51 start = get_cycles();
52 rdtsc_barrier();
53 /*
54 * The measurement runs for 'timeout' msecs:
55 */
56 end = start + (cycles_t) tsc_khz * timeout;
57 now = start;
58
59 for (i = 0; ; i++) {
60 /*
61 * We take the global lock, measure TSC, save the
62 * previous TSC that was measured (possibly on
63 * another CPU) and update the previous TSC timestamp.
64 */
65 arch_spin_lock(&sync_lock);
66 prev = last_tsc;
67 rdtsc_barrier();
68 now = get_cycles();
69 rdtsc_barrier();
70 last_tsc = now;
71 arch_spin_unlock(&sync_lock);
72
73 /*
74 * Be nice every now and then (and also check whether
75 * measurement is done [we also insert a 10 million
76 * loops safety exit, so we dont lock up in case the
77 * TSC readout is totally broken]):
78 */
79 if (unlikely(!(i & 7))) {
80 if (now > end || i > 10000000)
81 break;
82 cpu_relax();
83 touch_nmi_watchdog();
84 }
85 /*
86 * Outside the critical section we can now see whether
87 * we saw a time-warp of the TSC going backwards:
88 */
89 if (unlikely(prev > now)) {
90 arch_spin_lock(&sync_lock);
91 max_warp = max(max_warp, prev - now);
92 nr_warps++;
93 arch_spin_unlock(&sync_lock);
94 }
95 }
96 WARN(!(now-start),
97 "Warning: zero tsc calibration delta: %Ld [max: %Ld]\n",
98 now-start, end-start);
99}
100
101/*
102 * If the target CPU coming online doesn't have any of its core-siblings
103 * online, a timeout of 20msec will be used for the TSC-warp measurement
104 * loop. Otherwise a smaller timeout of 2msec will be used, as we have some
105 * information about this socket already (and this information grows as we
106 * have more and more logical-siblings in that socket).
107 *
108 * Ideally we should be able to skip the TSC sync check on the other
109 * core-siblings, if the first logical CPU in a socket passed the sync test.
110 * But as the TSC is per-logical CPU and can potentially be modified wrongly
111 * by the bios, TSC sync test for smaller duration should be able
112 * to catch such errors. Also this will catch the condition where all the
113 * cores in the socket doesn't get reset at the same time.
114 */
115static inline unsigned int loop_timeout(int cpu)
116{
117 return (cpumask_weight(cpu_core_mask(cpu)) > 1) ? 2 : 20;
118}
119
120/*
121 * Source CPU calls into this - it waits for the freshly booted
122 * target CPU to arrive and then starts the measurement:
123 */
124void __cpuinit check_tsc_sync_source(int cpu)
125{
126 int cpus = 2;
127
128 /*
129 * No need to check if we already know that the TSC is not
130 * synchronized:
131 */
132 if (unsynchronized_tsc())
133 return;
134
135 if (tsc_clocksource_reliable) {
136 if (cpu == (nr_cpu_ids-1) || system_state != SYSTEM_BOOTING)
137 pr_info(
138 "Skipped synchronization checks as TSC is reliable.\n");
139 return;
140 }
141
142 /*
143 * Reset it - in case this is a second bootup:
144 */
145 atomic_set(&stop_count, 0);
146
147 /*
148 * Wait for the target to arrive:
149 */
150 while (atomic_read(&start_count) != cpus-1)
151 cpu_relax();
152 /*
153 * Trigger the target to continue into the measurement too:
154 */
155 atomic_inc(&start_count);
156
157 check_tsc_warp(loop_timeout(cpu));
158
159 while (atomic_read(&stop_count) != cpus-1)
160 cpu_relax();
161
162 if (nr_warps) {
163 pr_warning("TSC synchronization [CPU#%d -> CPU#%d]:\n",
164 smp_processor_id(), cpu);
165 pr_warning("Measured %Ld cycles TSC warp between CPUs, "
166 "turning off TSC clock.\n", max_warp);
167 mark_tsc_unstable("check_tsc_sync_source failed");
168 } else {
169 pr_debug("TSC synchronization [CPU#%d -> CPU#%d]: passed\n",
170 smp_processor_id(), cpu);
171 }
172
173 /*
174 * Reset it - just in case we boot another CPU later:
175 */
176 atomic_set(&start_count, 0);
177 nr_warps = 0;
178 max_warp = 0;
179 last_tsc = 0;
180
181 /*
182 * Let the target continue with the bootup:
183 */
184 atomic_inc(&stop_count);
185}
186
187/*
188 * Freshly booted CPUs call into this:
189 */
190void __cpuinit check_tsc_sync_target(void)
191{
192 int cpus = 2;
193
194 if (unsynchronized_tsc() || tsc_clocksource_reliable)
195 return;
196
197 /*
198 * Register this CPU's participation and wait for the
199 * source CPU to start the measurement:
200 */
201 atomic_inc(&start_count);
202 while (atomic_read(&start_count) != cpus)
203 cpu_relax();
204
205 check_tsc_warp(loop_timeout(smp_processor_id()));
206
207 /*
208 * Ok, we are done:
209 */
210 atomic_inc(&stop_count);
211
212 /*
213 * Wait for the source CPU to print stuff:
214 */
215 while (atomic_read(&stop_count) != cpus)
216 cpu_relax();
217}
1/*
2 * check TSC synchronization.
3 *
4 * Copyright (C) 2006, Red Hat, Inc., Ingo Molnar
5 *
6 * We check whether all boot CPUs have their TSC's synchronized,
7 * print a warning if not and turn off the TSC clock-source.
8 *
9 * The warp-check is point-to-point between two CPUs, the CPU
10 * initiating the bootup is the 'source CPU', the freshly booting
11 * CPU is the 'target CPU'.
12 *
13 * Only two CPUs may participate - they can enter in any order.
14 * ( The serial nature of the boot logic and the CPU hotplug lock
15 * protects against more than 2 CPUs entering this code. )
16 */
17#include <linux/spinlock.h>
18#include <linux/kernel.h>
19#include <linux/smp.h>
20#include <linux/nmi.h>
21#include <asm/tsc.h>
22
23/*
24 * Entry/exit counters that make sure that both CPUs
25 * run the measurement code at once:
26 */
27static atomic_t start_count;
28static atomic_t stop_count;
29
30/*
31 * We use a raw spinlock in this exceptional case, because
32 * we want to have the fastest, inlined, non-debug version
33 * of a critical section, to be able to prove TSC time-warps:
34 */
35static arch_spinlock_t sync_lock = __ARCH_SPIN_LOCK_UNLOCKED;
36
37static cycles_t last_tsc;
38static cycles_t max_warp;
39static int nr_warps;
40
41/*
42 * TSC-warp measurement loop running on both CPUs. This is not called
43 * if there is no TSC.
44 */
45static void check_tsc_warp(unsigned int timeout)
46{
47 cycles_t start, now, prev, end;
48 int i;
49
50 start = rdtsc_ordered();
51 /*
52 * The measurement runs for 'timeout' msecs:
53 */
54 end = start + (cycles_t) tsc_khz * timeout;
55 now = start;
56
57 for (i = 0; ; i++) {
58 /*
59 * We take the global lock, measure TSC, save the
60 * previous TSC that was measured (possibly on
61 * another CPU) and update the previous TSC timestamp.
62 */
63 arch_spin_lock(&sync_lock);
64 prev = last_tsc;
65 now = rdtsc_ordered();
66 last_tsc = now;
67 arch_spin_unlock(&sync_lock);
68
69 /*
70 * Be nice every now and then (and also check whether
71 * measurement is done [we also insert a 10 million
72 * loops safety exit, so we dont lock up in case the
73 * TSC readout is totally broken]):
74 */
75 if (unlikely(!(i & 7))) {
76 if (now > end || i > 10000000)
77 break;
78 cpu_relax();
79 touch_nmi_watchdog();
80 }
81 /*
82 * Outside the critical section we can now see whether
83 * we saw a time-warp of the TSC going backwards:
84 */
85 if (unlikely(prev > now)) {
86 arch_spin_lock(&sync_lock);
87 max_warp = max(max_warp, prev - now);
88 nr_warps++;
89 arch_spin_unlock(&sync_lock);
90 }
91 }
92 WARN(!(now-start),
93 "Warning: zero tsc calibration delta: %Ld [max: %Ld]\n",
94 now-start, end-start);
95}
96
97/*
98 * If the target CPU coming online doesn't have any of its core-siblings
99 * online, a timeout of 20msec will be used for the TSC-warp measurement
100 * loop. Otherwise a smaller timeout of 2msec will be used, as we have some
101 * information about this socket already (and this information grows as we
102 * have more and more logical-siblings in that socket).
103 *
104 * Ideally we should be able to skip the TSC sync check on the other
105 * core-siblings, if the first logical CPU in a socket passed the sync test.
106 * But as the TSC is per-logical CPU and can potentially be modified wrongly
107 * by the bios, TSC sync test for smaller duration should be able
108 * to catch such errors. Also this will catch the condition where all the
109 * cores in the socket doesn't get reset at the same time.
110 */
111static inline unsigned int loop_timeout(int cpu)
112{
113 return (cpumask_weight(topology_core_cpumask(cpu)) > 1) ? 2 : 20;
114}
115
116/*
117 * Source CPU calls into this - it waits for the freshly booted
118 * target CPU to arrive and then starts the measurement:
119 */
120void check_tsc_sync_source(int cpu)
121{
122 int cpus = 2;
123
124 /*
125 * No need to check if we already know that the TSC is not
126 * synchronized or if we have no TSC.
127 */
128 if (unsynchronized_tsc())
129 return;
130
131 if (tsc_clocksource_reliable) {
132 if (cpu == (nr_cpu_ids-1) || system_state != SYSTEM_BOOTING)
133 pr_info(
134 "Skipped synchronization checks as TSC is reliable.\n");
135 return;
136 }
137
138 /*
139 * Reset it - in case this is a second bootup:
140 */
141 atomic_set(&stop_count, 0);
142
143 /*
144 * Wait for the target to arrive:
145 */
146 while (atomic_read(&start_count) != cpus-1)
147 cpu_relax();
148 /*
149 * Trigger the target to continue into the measurement too:
150 */
151 atomic_inc(&start_count);
152
153 check_tsc_warp(loop_timeout(cpu));
154
155 while (atomic_read(&stop_count) != cpus-1)
156 cpu_relax();
157
158 if (nr_warps) {
159 pr_warning("TSC synchronization [CPU#%d -> CPU#%d]:\n",
160 smp_processor_id(), cpu);
161 pr_warning("Measured %Ld cycles TSC warp between CPUs, "
162 "turning off TSC clock.\n", max_warp);
163 mark_tsc_unstable("check_tsc_sync_source failed");
164 } else {
165 pr_debug("TSC synchronization [CPU#%d -> CPU#%d]: passed\n",
166 smp_processor_id(), cpu);
167 }
168
169 /*
170 * Reset it - just in case we boot another CPU later:
171 */
172 atomic_set(&start_count, 0);
173 nr_warps = 0;
174 max_warp = 0;
175 last_tsc = 0;
176
177 /*
178 * Let the target continue with the bootup:
179 */
180 atomic_inc(&stop_count);
181}
182
183/*
184 * Freshly booted CPUs call into this:
185 */
186void check_tsc_sync_target(void)
187{
188 int cpus = 2;
189
190 /* Also aborts if there is no TSC. */
191 if (unsynchronized_tsc() || tsc_clocksource_reliable)
192 return;
193
194 /*
195 * Register this CPU's participation and wait for the
196 * source CPU to start the measurement:
197 */
198 atomic_inc(&start_count);
199 while (atomic_read(&start_count) != cpus)
200 cpu_relax();
201
202 check_tsc_warp(loop_timeout(smp_processor_id()));
203
204 /*
205 * Ok, we are done:
206 */
207 atomic_inc(&stop_count);
208
209 /*
210 * Wait for the source CPU to print stuff:
211 */
212 while (atomic_read(&stop_count) != cpus)
213 cpu_relax();
214}