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/topology.h>
18#include <linux/spinlock.h>
19#include <linux/kernel.h>
20#include <linux/smp.h>
21#include <linux/nmi.h>
22#include <asm/tsc.h>
23
24struct tsc_adjust {
25 s64 bootval;
26 s64 adjusted;
27 unsigned long nextcheck;
28 bool warned;
29};
30
31static DEFINE_PER_CPU(struct tsc_adjust, tsc_adjust);
32
33void tsc_verify_tsc_adjust(bool resume)
34{
35 struct tsc_adjust *adj = this_cpu_ptr(&tsc_adjust);
36 s64 curval;
37
38 if (!boot_cpu_has(X86_FEATURE_TSC_ADJUST))
39 return;
40
41 /* Rate limit the MSR check */
42 if (!resume && time_before(jiffies, adj->nextcheck))
43 return;
44
45 adj->nextcheck = jiffies + HZ;
46
47 rdmsrl(MSR_IA32_TSC_ADJUST, curval);
48 if (adj->adjusted == curval)
49 return;
50
51 /* Restore the original value */
52 wrmsrl(MSR_IA32_TSC_ADJUST, adj->adjusted);
53
54 if (!adj->warned || resume) {
55 pr_warn(FW_BUG "TSC ADJUST differs: CPU%u %lld --> %lld. Restoring\n",
56 smp_processor_id(), adj->adjusted, curval);
57 adj->warned = true;
58 }
59}
60
61static void tsc_sanitize_first_cpu(struct tsc_adjust *cur, s64 bootval,
62 unsigned int cpu, bool bootcpu)
63{
64 /*
65 * First online CPU in a package stores the boot value in the
66 * adjustment value. This value might change later via the sync
67 * mechanism. If that fails we still can yell about boot values not
68 * being consistent.
69 *
70 * On the boot cpu we just force set the ADJUST value to 0 if it's
71 * non zero. We don't do that on non boot cpus because physical
72 * hotplug should have set the ADJUST register to a value > 0 so
73 * the TSC is in sync with the already running cpus.
74 *
75 * But we always force positive ADJUST values. Otherwise the TSC
76 * deadline timer creates an interrupt storm. We also have to
77 * prevent values > 0x7FFFFFFF as those wreckage the timer as well.
78 */
79 if ((bootcpu && bootval != 0) || (!bootcpu && bootval < 0) ||
80 (bootval > 0x7FFFFFFF)) {
81 pr_warn(FW_BUG "TSC ADJUST: CPU%u: %lld force to 0\n", cpu,
82 bootval);
83 wrmsrl(MSR_IA32_TSC_ADJUST, 0);
84 bootval = 0;
85 }
86 cur->adjusted = bootval;
87}
88
89#ifndef CONFIG_SMP
90bool __init tsc_store_and_check_tsc_adjust(bool bootcpu)
91{
92 struct tsc_adjust *cur = this_cpu_ptr(&tsc_adjust);
93 s64 bootval;
94
95 if (!boot_cpu_has(X86_FEATURE_TSC_ADJUST))
96 return false;
97
98 rdmsrl(MSR_IA32_TSC_ADJUST, bootval);
99 cur->bootval = bootval;
100 cur->nextcheck = jiffies + HZ;
101 tsc_sanitize_first_cpu(cur, bootval, smp_processor_id(), bootcpu);
102 return false;
103}
104
105#else /* !CONFIG_SMP */
106
107/*
108 * Store and check the TSC ADJUST MSR if available
109 */
110bool tsc_store_and_check_tsc_adjust(bool bootcpu)
111{
112 struct tsc_adjust *ref, *cur = this_cpu_ptr(&tsc_adjust);
113 unsigned int refcpu, cpu = smp_processor_id();
114 struct cpumask *mask;
115 s64 bootval;
116
117 if (!boot_cpu_has(X86_FEATURE_TSC_ADJUST))
118 return false;
119
120 rdmsrl(MSR_IA32_TSC_ADJUST, bootval);
121 cur->bootval = bootval;
122 cur->nextcheck = jiffies + HZ;
123 cur->warned = false;
124
125 /*
126 * Check whether this CPU is the first in a package to come up. In
127 * this case do not check the boot value against another package
128 * because the new package might have been physically hotplugged,
129 * where TSC_ADJUST is expected to be different. When called on the
130 * boot CPU topology_core_cpumask() might not be available yet.
131 */
132 mask = topology_core_cpumask(cpu);
133 refcpu = mask ? cpumask_any_but(mask, cpu) : nr_cpu_ids;
134
135 if (refcpu >= nr_cpu_ids) {
136 tsc_sanitize_first_cpu(cur, bootval, smp_processor_id(),
137 bootcpu);
138 return false;
139 }
140
141 ref = per_cpu_ptr(&tsc_adjust, refcpu);
142 /*
143 * Compare the boot value and complain if it differs in the
144 * package.
145 */
146 if (bootval != ref->bootval) {
147 pr_warn(FW_BUG "TSC ADJUST differs: Reference CPU%u: %lld CPU%u: %lld\n",
148 refcpu, ref->bootval, cpu, bootval);
149 }
150 /*
151 * The TSC_ADJUST values in a package must be the same. If the boot
152 * value on this newly upcoming CPU differs from the adjustment
153 * value of the already online CPU in this package, set it to that
154 * adjusted value.
155 */
156 if (bootval != ref->adjusted) {
157 pr_warn("TSC ADJUST synchronize: Reference CPU%u: %lld CPU%u: %lld\n",
158 refcpu, ref->adjusted, cpu, bootval);
159 cur->adjusted = ref->adjusted;
160 wrmsrl(MSR_IA32_TSC_ADJUST, ref->adjusted);
161 }
162 /*
163 * We have the TSCs forced to be in sync on this package. Skip sync
164 * test:
165 */
166 return true;
167}
168
169/*
170 * Entry/exit counters that make sure that both CPUs
171 * run the measurement code at once:
172 */
173static atomic_t start_count;
174static atomic_t stop_count;
175static atomic_t skip_test;
176static atomic_t test_runs;
177
178/*
179 * We use a raw spinlock in this exceptional case, because
180 * we want to have the fastest, inlined, non-debug version
181 * of a critical section, to be able to prove TSC time-warps:
182 */
183static arch_spinlock_t sync_lock = __ARCH_SPIN_LOCK_UNLOCKED;
184
185static cycles_t last_tsc;
186static cycles_t max_warp;
187static int nr_warps;
188static int random_warps;
189
190/*
191 * TSC-warp measurement loop running on both CPUs. This is not called
192 * if there is no TSC.
193 */
194static cycles_t check_tsc_warp(unsigned int timeout)
195{
196 cycles_t start, now, prev, end, cur_max_warp = 0;
197 int i, cur_warps = 0;
198
199 start = rdtsc_ordered();
200 /*
201 * The measurement runs for 'timeout' msecs:
202 */
203 end = start + (cycles_t) tsc_khz * timeout;
204 now = start;
205
206 for (i = 0; ; i++) {
207 /*
208 * We take the global lock, measure TSC, save the
209 * previous TSC that was measured (possibly on
210 * another CPU) and update the previous TSC timestamp.
211 */
212 arch_spin_lock(&sync_lock);
213 prev = last_tsc;
214 now = rdtsc_ordered();
215 last_tsc = now;
216 arch_spin_unlock(&sync_lock);
217
218 /*
219 * Be nice every now and then (and also check whether
220 * measurement is done [we also insert a 10 million
221 * loops safety exit, so we dont lock up in case the
222 * TSC readout is totally broken]):
223 */
224 if (unlikely(!(i & 7))) {
225 if (now > end || i > 10000000)
226 break;
227 cpu_relax();
228 touch_nmi_watchdog();
229 }
230 /*
231 * Outside the critical section we can now see whether
232 * we saw a time-warp of the TSC going backwards:
233 */
234 if (unlikely(prev > now)) {
235 arch_spin_lock(&sync_lock);
236 max_warp = max(max_warp, prev - now);
237 cur_max_warp = max_warp;
238 /*
239 * Check whether this bounces back and forth. Only
240 * one CPU should observe time going backwards.
241 */
242 if (cur_warps != nr_warps)
243 random_warps++;
244 nr_warps++;
245 cur_warps = nr_warps;
246 arch_spin_unlock(&sync_lock);
247 }
248 }
249 WARN(!(now-start),
250 "Warning: zero tsc calibration delta: %Ld [max: %Ld]\n",
251 now-start, end-start);
252 return cur_max_warp;
253}
254
255/*
256 * If the target CPU coming online doesn't have any of its core-siblings
257 * online, a timeout of 20msec will be used for the TSC-warp measurement
258 * loop. Otherwise a smaller timeout of 2msec will be used, as we have some
259 * information about this socket already (and this information grows as we
260 * have more and more logical-siblings in that socket).
261 *
262 * Ideally we should be able to skip the TSC sync check on the other
263 * core-siblings, if the first logical CPU in a socket passed the sync test.
264 * But as the TSC is per-logical CPU and can potentially be modified wrongly
265 * by the bios, TSC sync test for smaller duration should be able
266 * to catch such errors. Also this will catch the condition where all the
267 * cores in the socket doesn't get reset at the same time.
268 */
269static inline unsigned int loop_timeout(int cpu)
270{
271 return (cpumask_weight(topology_core_cpumask(cpu)) > 1) ? 2 : 20;
272}
273
274/*
275 * Source CPU calls into this - it waits for the freshly booted
276 * target CPU to arrive and then starts the measurement:
277 */
278void check_tsc_sync_source(int cpu)
279{
280 int cpus = 2;
281
282 /*
283 * No need to check if we already know that the TSC is not
284 * synchronized or if we have no TSC.
285 */
286 if (unsynchronized_tsc())
287 return;
288
289 /*
290 * Set the maximum number of test runs to
291 * 1 if the CPU does not provide the TSC_ADJUST MSR
292 * 3 if the MSR is available, so the target can try to adjust
293 */
294 if (!boot_cpu_has(X86_FEATURE_TSC_ADJUST))
295 atomic_set(&test_runs, 1);
296 else
297 atomic_set(&test_runs, 3);
298retry:
299 /*
300 * Wait for the target to start or to skip the test:
301 */
302 while (atomic_read(&start_count) != cpus - 1) {
303 if (atomic_read(&skip_test) > 0) {
304 atomic_set(&skip_test, 0);
305 return;
306 }
307 cpu_relax();
308 }
309
310 /*
311 * Trigger the target to continue into the measurement too:
312 */
313 atomic_inc(&start_count);
314
315 check_tsc_warp(loop_timeout(cpu));
316
317 while (atomic_read(&stop_count) != cpus-1)
318 cpu_relax();
319
320 /*
321 * If the test was successful set the number of runs to zero and
322 * stop. If not, decrement the number of runs an check if we can
323 * retry. In case of random warps no retry is attempted.
324 */
325 if (!nr_warps) {
326 atomic_set(&test_runs, 0);
327
328 pr_debug("TSC synchronization [CPU#%d -> CPU#%d]: passed\n",
329 smp_processor_id(), cpu);
330
331 } else if (atomic_dec_and_test(&test_runs) || random_warps) {
332 /* Force it to 0 if random warps brought us here */
333 atomic_set(&test_runs, 0);
334
335 pr_warning("TSC synchronization [CPU#%d -> CPU#%d]:\n",
336 smp_processor_id(), cpu);
337 pr_warning("Measured %Ld cycles TSC warp between CPUs, "
338 "turning off TSC clock.\n", max_warp);
339 if (random_warps)
340 pr_warning("TSC warped randomly between CPUs\n");
341 mark_tsc_unstable("check_tsc_sync_source failed");
342 }
343
344 /*
345 * Reset it - just in case we boot another CPU later:
346 */
347 atomic_set(&start_count, 0);
348 random_warps = 0;
349 nr_warps = 0;
350 max_warp = 0;
351 last_tsc = 0;
352
353 /*
354 * Let the target continue with the bootup:
355 */
356 atomic_inc(&stop_count);
357
358 /*
359 * Retry, if there is a chance to do so.
360 */
361 if (atomic_read(&test_runs) > 0)
362 goto retry;
363}
364
365/*
366 * Freshly booted CPUs call into this:
367 */
368void check_tsc_sync_target(void)
369{
370 struct tsc_adjust *cur = this_cpu_ptr(&tsc_adjust);
371 unsigned int cpu = smp_processor_id();
372 cycles_t cur_max_warp, gbl_max_warp;
373 int cpus = 2;
374
375 /* Also aborts if there is no TSC. */
376 if (unsynchronized_tsc())
377 return;
378
379 /*
380 * Store, verify and sanitize the TSC adjust register. If
381 * successful skip the test.
382 *
383 * The test is also skipped when the TSC is marked reliable. This
384 * is true for SoCs which have no fallback clocksource. On these
385 * SoCs the TSC is frequency synchronized, but still the TSC ADJUST
386 * register might have been wreckaged by the BIOS..
387 */
388 if (tsc_store_and_check_tsc_adjust(false) || tsc_clocksource_reliable) {
389 atomic_inc(&skip_test);
390 return;
391 }
392
393retry:
394 /*
395 * Register this CPU's participation and wait for the
396 * source CPU to start the measurement:
397 */
398 atomic_inc(&start_count);
399 while (atomic_read(&start_count) != cpus)
400 cpu_relax();
401
402 cur_max_warp = check_tsc_warp(loop_timeout(cpu));
403
404 /*
405 * Store the maximum observed warp value for a potential retry:
406 */
407 gbl_max_warp = max_warp;
408
409 /*
410 * Ok, we are done:
411 */
412 atomic_inc(&stop_count);
413
414 /*
415 * Wait for the source CPU to print stuff:
416 */
417 while (atomic_read(&stop_count) != cpus)
418 cpu_relax();
419
420 /*
421 * Reset it for the next sync test:
422 */
423 atomic_set(&stop_count, 0);
424
425 /*
426 * Check the number of remaining test runs. If not zero, the test
427 * failed and a retry with adjusted TSC is possible. If zero the
428 * test was either successful or failed terminally.
429 */
430 if (!atomic_read(&test_runs))
431 return;
432
433 /*
434 * If the warp value of this CPU is 0, then the other CPU
435 * observed time going backwards so this TSC was ahead and
436 * needs to move backwards.
437 */
438 if (!cur_max_warp)
439 cur_max_warp = -gbl_max_warp;
440
441 /*
442 * Add the result to the previous adjustment value.
443 *
444 * The adjustement value is slightly off by the overhead of the
445 * sync mechanism (observed values are ~200 TSC cycles), but this
446 * really depends on CPU, node distance and frequency. So
447 * compensating for this is hard to get right. Experiments show
448 * that the warp is not longer detectable when the observed warp
449 * value is used. In the worst case the adjustment needs to go
450 * through a 3rd run for fine tuning.
451 */
452 cur->adjusted += cur_max_warp;
453
454 /*
455 * TSC deadline timer stops working or creates an interrupt storm
456 * with adjust values < 0 and > x07ffffff.
457 *
458 * To allow adjust values > 0x7FFFFFFF we need to disable the
459 * deadline timer and use the local APIC timer, but that requires
460 * more intrusive changes and we do not have any useful information
461 * from Intel about the underlying HW wreckage yet.
462 */
463 if (cur->adjusted < 0)
464 cur->adjusted = 0;
465 if (cur->adjusted > 0x7FFFFFFF)
466 cur->adjusted = 0x7FFFFFFF;
467
468 pr_warn("TSC ADJUST compensate: CPU%u observed %lld warp. Adjust: %lld\n",
469 cpu, cur_max_warp, cur->adjusted);
470
471 wrmsrl(MSR_IA32_TSC_ADJUST, cur->adjusted);
472 goto retry;
473
474}
475
476#endif /* CONFIG_SMP */
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}