tsc.c revision 223426
1/*-
2 * Copyright (c) 1998-2003 Poul-Henning Kamp
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 *    notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 *    notice, this list of conditions and the following disclaimer in the
12 *    documentation and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24 * SUCH DAMAGE.
25 */
26
27#include <sys/cdefs.h>
28__FBSDID("$FreeBSD: head/sys/x86/x86/tsc.c 223426 2011-06-22 16:40:45Z jkim $");
29
30#include "opt_clock.h"
31
32#include <sys/param.h>
33#include <sys/bus.h>
34#include <sys/cpu.h>
35#include <sys/limits.h>
36#include <sys/malloc.h>
37#include <sys/systm.h>
38#include <sys/sysctl.h>
39#include <sys/time.h>
40#include <sys/timetc.h>
41#include <sys/kernel.h>
42#include <sys/power.h>
43#include <sys/smp.h>
44#include <machine/clock.h>
45#include <machine/cputypes.h>
46#include <machine/md_var.h>
47#include <machine/specialreg.h>
48
49#include "cpufreq_if.h"
50
51uint64_t	tsc_freq;
52int		tsc_is_invariant;
53int		tsc_perf_stat;
54
55static eventhandler_tag tsc_levels_tag, tsc_pre_tag, tsc_post_tag;
56
57SYSCTL_INT(_kern_timecounter, OID_AUTO, invariant_tsc, CTLFLAG_RDTUN,
58    &tsc_is_invariant, 0, "Indicates whether the TSC is P-state invariant");
59TUNABLE_INT("kern.timecounter.invariant_tsc", &tsc_is_invariant);
60
61#ifdef SMP
62static int	smp_tsc;
63SYSCTL_INT(_kern_timecounter, OID_AUTO, smp_tsc, CTLFLAG_RDTUN, &smp_tsc, 0,
64    "Indicates whether the TSC is safe to use in SMP mode");
65TUNABLE_INT("kern.timecounter.smp_tsc", &smp_tsc);
66#endif
67
68static int	tsc_disabled;
69SYSCTL_INT(_machdep, OID_AUTO, disable_tsc, CTLFLAG_RDTUN, &tsc_disabled, 0,
70    "Disable x86 Time Stamp Counter");
71TUNABLE_INT("machdep.disable_tsc", &tsc_disabled);
72
73static int	tsc_skip_calibration;
74SYSCTL_INT(_machdep, OID_AUTO, disable_tsc_calibration, CTLFLAG_RDTUN,
75    &tsc_skip_calibration, 0, "Disable TSC frequency calibration");
76TUNABLE_INT("machdep.disable_tsc_calibration", &tsc_skip_calibration);
77
78static void tsc_freq_changed(void *arg, const struct cf_level *level,
79    int status);
80static void tsc_freq_changing(void *arg, const struct cf_level *level,
81    int *status);
82static unsigned tsc_get_timecount(struct timecounter *tc);
83static unsigned tsc_get_timecount_low(struct timecounter *tc);
84static void tsc_levels_changed(void *arg, int unit);
85
86static struct timecounter tsc_timecounter = {
87	tsc_get_timecount,	/* get_timecount */
88	0,			/* no poll_pps */
89	~0u,			/* counter_mask */
90	0,			/* frequency */
91	"TSC",			/* name */
92	800,			/* quality (adjusted in code) */
93};
94
95#define	VMW_HVMAGIC		0x564d5868
96#define	VMW_HVPORT		0x5658
97#define	VMW_HVCMD_GETVERSION	10
98#define	VMW_HVCMD_GETHZ		45
99
100static __inline void
101vmware_hvcall(u_int cmd, u_int *p)
102{
103
104	__asm __volatile("inl %w3, %0"
105	: "=a" (p[0]), "=b" (p[1]), "=c" (p[2]), "=d" (p[3])
106	: "0" (VMW_HVMAGIC), "1" (UINT_MAX), "2" (cmd), "3" (VMW_HVPORT)
107	: "memory");
108}
109
110static int
111tsc_freq_vmware(void)
112{
113	char hv_sig[13];
114	u_int regs[4];
115	char *p;
116	u_int hv_high;
117	int i;
118
119	/*
120	 * [RFC] CPUID usage for interaction between Hypervisors and Linux.
121	 * http://lkml.org/lkml/2008/10/1/246
122	 *
123	 * KB1009458: Mechanisms to determine if software is running in
124	 * a VMware virtual machine
125	 * http://kb.vmware.com/kb/1009458
126	 */
127	hv_high = 0;
128	if ((cpu_feature2 & CPUID2_HV) != 0) {
129		do_cpuid(0x40000000, regs);
130		hv_high = regs[0];
131		for (i = 1, p = hv_sig; i < 4; i++, p += sizeof(regs) / 4)
132			memcpy(p, &regs[i], sizeof(regs[i]));
133		*p = '\0';
134		if (bootverbose) {
135			/*
136			 * HV vendor	ID string
137			 * ------------+--------------
138			 * KVM		"KVMKVMKVM"
139			 * Microsoft	"Microsoft Hv"
140			 * VMware	"VMwareVMware"
141			 * Xen		"XenVMMXenVMM"
142			 */
143			printf("Hypervisor: Origin = \"%s\"\n", hv_sig);
144		}
145		if (strncmp(hv_sig, "VMwareVMware", 12) != 0)
146			return (0);
147	} else {
148		p = getenv("smbios.system.serial");
149		if (p == NULL)
150			return (0);
151		if (strncmp(p, "VMware-", 7) != 0 &&
152		    strncmp(p, "VMW", 3) != 0) {
153			freeenv(p);
154			return (0);
155		}
156		freeenv(p);
157		vmware_hvcall(VMW_HVCMD_GETVERSION, regs);
158		if (regs[1] != VMW_HVMAGIC)
159			return (0);
160	}
161	if (hv_high >= 0x40000010) {
162		do_cpuid(0x40000010, regs);
163		tsc_freq = regs[0] * 1000;
164	} else {
165		vmware_hvcall(VMW_HVCMD_GETHZ, regs);
166		if (regs[1] != UINT_MAX)
167			tsc_freq = regs[0] | ((uint64_t)regs[1] << 32);
168	}
169	tsc_is_invariant = 1;
170	return (1);
171}
172
173static void
174tsc_freq_intel(void)
175{
176	char brand[48];
177	u_int regs[4];
178	uint64_t freq;
179	char *p;
180	u_int i;
181
182	/*
183	 * Intel Processor Identification and the CPUID Instruction
184	 * Application Note 485.
185	 * http://www.intel.com/assets/pdf/appnote/241618.pdf
186	 */
187	if (cpu_exthigh >= 0x80000004) {
188		p = brand;
189		for (i = 0x80000002; i < 0x80000005; i++) {
190			do_cpuid(i, regs);
191			memcpy(p, regs, sizeof(regs));
192			p += sizeof(regs);
193		}
194		p = NULL;
195		for (i = 0; i < sizeof(brand) - 1; i++)
196			if (brand[i] == 'H' && brand[i + 1] == 'z')
197				p = brand + i;
198		if (p != NULL) {
199			p -= 5;
200			switch (p[4]) {
201			case 'M':
202				i = 1;
203				break;
204			case 'G':
205				i = 1000;
206				break;
207			case 'T':
208				i = 1000000;
209				break;
210			default:
211				return;
212			}
213#define	C2D(c)	((c) - '0')
214			if (p[1] == '.') {
215				freq = C2D(p[0]) * 1000;
216				freq += C2D(p[2]) * 100;
217				freq += C2D(p[3]) * 10;
218				freq *= i * 1000;
219			} else {
220				freq = C2D(p[0]) * 1000;
221				freq += C2D(p[1]) * 100;
222				freq += C2D(p[2]) * 10;
223				freq += C2D(p[3]);
224				freq *= i * 1000000;
225			}
226#undef C2D
227			tsc_freq = freq;
228		}
229	}
230}
231
232static void
233probe_tsc_freq(void)
234{
235	u_int regs[4];
236	uint64_t tsc1, tsc2;
237
238	if (cpu_high >= 6) {
239		do_cpuid(6, regs);
240		if ((regs[2] & CPUID_PERF_STAT) != 0) {
241			/*
242			 * XXX Some emulators expose host CPUID without actual
243			 * support for these MSRs.  We must test whether they
244			 * really work.
245			 */
246			wrmsr(MSR_MPERF, 0);
247			wrmsr(MSR_APERF, 0);
248			DELAY(10);
249			if (rdmsr(MSR_MPERF) > 0 && rdmsr(MSR_APERF) > 0)
250				tsc_perf_stat = 1;
251		}
252	}
253
254	if (tsc_freq_vmware())
255		return;
256
257	switch (cpu_vendor_id) {
258	case CPU_VENDOR_AMD:
259		if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
260		    (vm_guest == VM_GUEST_NO &&
261		    CPUID_TO_FAMILY(cpu_id) >= 0x10))
262			tsc_is_invariant = 1;
263		break;
264	case CPU_VENDOR_INTEL:
265		if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
266		    (vm_guest == VM_GUEST_NO &&
267		    ((CPUID_TO_FAMILY(cpu_id) == 0x6 &&
268		    CPUID_TO_MODEL(cpu_id) >= 0xe) ||
269		    (CPUID_TO_FAMILY(cpu_id) == 0xf &&
270		    CPUID_TO_MODEL(cpu_id) >= 0x3))))
271			tsc_is_invariant = 1;
272		break;
273	case CPU_VENDOR_CENTAUR:
274		if (vm_guest == VM_GUEST_NO &&
275		    CPUID_TO_FAMILY(cpu_id) == 0x6 &&
276		    CPUID_TO_MODEL(cpu_id) >= 0xf &&
277		    (rdmsr(0x1203) & 0x100000000ULL) == 0)
278			tsc_is_invariant = 1;
279		break;
280	}
281
282	if (tsc_skip_calibration) {
283		if (cpu_vendor_id == CPU_VENDOR_INTEL)
284			tsc_freq_intel();
285		return;
286	}
287
288	if (bootverbose)
289	        printf("Calibrating TSC clock ... ");
290	tsc1 = rdtsc();
291	DELAY(1000000);
292	tsc2 = rdtsc();
293	tsc_freq = tsc2 - tsc1;
294	if (bootverbose)
295		printf("TSC clock: %ju Hz\n", (intmax_t)tsc_freq);
296}
297
298void
299init_TSC(void)
300{
301
302	if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
303		return;
304
305	probe_tsc_freq();
306
307	/*
308	 * Inform CPU accounting about our boot-time clock rate.  This will
309	 * be updated if someone loads a cpufreq driver after boot that
310	 * discovers a new max frequency.
311	 */
312	if (tsc_freq != 0)
313		set_cputicker(rdtsc, tsc_freq, !tsc_is_invariant);
314
315	if (tsc_is_invariant)
316		return;
317
318	/* Register to find out about changes in CPU frequency. */
319	tsc_pre_tag = EVENTHANDLER_REGISTER(cpufreq_pre_change,
320	    tsc_freq_changing, NULL, EVENTHANDLER_PRI_FIRST);
321	tsc_post_tag = EVENTHANDLER_REGISTER(cpufreq_post_change,
322	    tsc_freq_changed, NULL, EVENTHANDLER_PRI_FIRST);
323	tsc_levels_tag = EVENTHANDLER_REGISTER(cpufreq_levels_changed,
324	    tsc_levels_changed, NULL, EVENTHANDLER_PRI_ANY);
325}
326
327#ifdef SMP
328
329#define	TSC_READ(x)			\
330static void				\
331tsc_read_##x(void *arg)			\
332{					\
333	uint32_t *tsc = arg;		\
334	u_int cpu = PCPU_GET(cpuid);	\
335					\
336	tsc[cpu * 3 + x] = rdtsc32();	\
337}
338TSC_READ(0)
339TSC_READ(1)
340TSC_READ(2)
341#undef TSC_READ
342
343#define	N	1000
344
345static void
346comp_smp_tsc(void *arg)
347{
348	uint32_t *tsc;
349	int32_t d1, d2;
350	u_int cpu = PCPU_GET(cpuid);
351	u_int i, j, size;
352
353	size = (mp_maxid + 1) * 3;
354	for (i = 0, tsc = arg; i < N; i++, tsc += size)
355		CPU_FOREACH(j) {
356			if (j == cpu)
357				continue;
358			d1 = tsc[cpu * 3 + 1] - tsc[j * 3];
359			d2 = tsc[cpu * 3 + 2] - tsc[j * 3 + 1];
360			if (d1 <= 0 || d2 <= 0) {
361				smp_tsc = 0;
362				return;
363			}
364		}
365}
366
367static int
368test_smp_tsc(void)
369{
370	uint32_t *data, *tsc;
371	u_int i, size;
372
373	if (!smp_tsc && !tsc_is_invariant)
374		return (-100);
375	size = (mp_maxid + 1) * 3;
376	data = malloc(sizeof(*data) * size * N, M_TEMP, M_WAITOK);
377	for (i = 0, tsc = data; i < N; i++, tsc += size)
378		smp_rendezvous(tsc_read_0, tsc_read_1, tsc_read_2, tsc);
379	smp_tsc = 1;	/* XXX */
380	smp_rendezvous(smp_no_rendevous_barrier, comp_smp_tsc,
381	    smp_no_rendevous_barrier, data);
382	free(data, M_TEMP);
383	if (bootverbose)
384		printf("SMP: %sed TSC synchronization test\n",
385		    smp_tsc ? "pass" : "fail");
386	if (smp_tsc && tsc_is_invariant) {
387		switch (cpu_vendor_id) {
388		case CPU_VENDOR_AMD:
389			/*
390			 * Starting with Family 15h processors, TSC clock
391			 * source is in the north bridge.  Check whether
392			 * we have a single-socket/multi-core platform.
393			 * XXX Need more work for complex cases.
394			 */
395			if (CPUID_TO_FAMILY(cpu_id) < 0x15 ||
396			    (amd_feature2 & AMDID2_CMP) == 0 ||
397			    smp_cpus > (cpu_procinfo2 & AMDID_CMP_CORES) + 1)
398				break;
399			return (1000);
400		case CPU_VENDOR_INTEL:
401			/*
402			 * XXX Assume Intel platforms have synchronized TSCs.
403			 */
404			return (1000);
405		}
406		return (800);
407	}
408	return (-100);
409}
410
411#undef N
412
413#endif /* SMP */
414
415static void
416init_TSC_tc(void)
417{
418	uint64_t max_freq;
419	int shift;
420
421	if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
422		return;
423
424	/*
425	 * Limit timecounter frequency to fit in an int and prevent it from
426	 * overflowing too fast.
427	 */
428	max_freq = UINT_MAX;
429
430	/*
431	 * We can not use the TSC if we support APM.  Precise timekeeping
432	 * on an APM'ed machine is at best a fools pursuit, since
433	 * any and all of the time spent in various SMM code can't
434	 * be reliably accounted for.  Reading the RTC is your only
435	 * source of reliable time info.  The i8254 loses too, of course,
436	 * but we need to have some kind of time...
437	 * We don't know at this point whether APM is going to be used
438	 * or not, nor when it might be activated.  Play it safe.
439	 */
440	if (power_pm_get_type() == POWER_PM_TYPE_APM) {
441		tsc_timecounter.tc_quality = -1000;
442		if (bootverbose)
443			printf("TSC timecounter disabled: APM enabled.\n");
444		goto init;
445	}
446
447	/*
448	 * We cannot use the TSC if it stops incrementing in deep sleep.
449	 * Currently only Intel CPUs are known for this problem unless
450	 * the invariant TSC bit is set.
451	 */
452	if (cpu_can_deep_sleep && cpu_vendor_id == CPU_VENDOR_INTEL &&
453	    (amd_pminfo & AMDPM_TSC_INVARIANT) == 0) {
454		tsc_timecounter.tc_quality = -1000;
455		if (bootverbose)
456			printf("TSC timecounter disabled: C3 enabled.\n");
457		goto init;
458	}
459
460#ifdef SMP
461	/*
462	 * We can not use the TSC in SMP mode unless the TSCs on all CPUs are
463	 * synchronized.  If the user is sure that the system has synchronized
464	 * TSCs, set kern.timecounter.smp_tsc tunable to a non-zero value.
465	 * We also limit the frequency even lower to avoid "temporal anomalies"
466	 * as much as possible.
467	 */
468	if (smp_cpus > 1) {
469		tsc_timecounter.tc_quality = test_smp_tsc();
470		max_freq >>= 8;
471	} else
472#endif
473	if (tsc_is_invariant)
474		tsc_timecounter.tc_quality = 1000;
475
476init:
477	for (shift = 0; shift < 31 && (tsc_freq >> shift) > max_freq; shift++)
478		;
479	if (shift > 0) {
480		tsc_timecounter.tc_get_timecount = tsc_get_timecount_low;
481		tsc_timecounter.tc_name = "TSC-low";
482		if (bootverbose)
483			printf("TSC timecounter discards lower %d bit(s)\n",
484			    shift);
485	}
486	if (tsc_freq != 0) {
487		tsc_timecounter.tc_frequency = tsc_freq >> shift;
488		tsc_timecounter.tc_priv = (void *)(intptr_t)shift;
489		tc_init(&tsc_timecounter);
490	}
491}
492SYSINIT(tsc_tc, SI_SUB_SMP, SI_ORDER_ANY, init_TSC_tc, NULL);
493
494/*
495 * When cpufreq levels change, find out about the (new) max frequency.  We
496 * use this to update CPU accounting in case it got a lower estimate at boot.
497 */
498static void
499tsc_levels_changed(void *arg, int unit)
500{
501	device_t cf_dev;
502	struct cf_level *levels;
503	int count, error;
504	uint64_t max_freq;
505
506	/* Only use values from the first CPU, assuming all are equal. */
507	if (unit != 0)
508		return;
509
510	/* Find the appropriate cpufreq device instance. */
511	cf_dev = devclass_get_device(devclass_find("cpufreq"), unit);
512	if (cf_dev == NULL) {
513		printf("tsc_levels_changed() called but no cpufreq device?\n");
514		return;
515	}
516
517	/* Get settings from the device and find the max frequency. */
518	count = 64;
519	levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
520	if (levels == NULL)
521		return;
522	error = CPUFREQ_LEVELS(cf_dev, levels, &count);
523	if (error == 0 && count != 0) {
524		max_freq = (uint64_t)levels[0].total_set.freq * 1000000;
525		set_cputicker(rdtsc, max_freq, 1);
526	} else
527		printf("tsc_levels_changed: no max freq found\n");
528	free(levels, M_TEMP);
529}
530
531/*
532 * If the TSC timecounter is in use, veto the pending change.  It may be
533 * possible in the future to handle a dynamically-changing timecounter rate.
534 */
535static void
536tsc_freq_changing(void *arg, const struct cf_level *level, int *status)
537{
538
539	if (*status != 0 || timecounter != &tsc_timecounter)
540		return;
541
542	printf("timecounter TSC must not be in use when "
543	    "changing frequencies; change denied\n");
544	*status = EBUSY;
545}
546
547/* Update TSC freq with the value indicated by the caller. */
548static void
549tsc_freq_changed(void *arg, const struct cf_level *level, int status)
550{
551	uint64_t freq;
552
553	/* If there was an error during the transition, don't do anything. */
554	if (tsc_disabled || status != 0)
555		return;
556
557	/* Total setting for this level gives the new frequency in MHz. */
558	freq = (uint64_t)level->total_set.freq * 1000000;
559	atomic_store_rel_64(&tsc_freq, freq);
560	tsc_timecounter.tc_frequency =
561	    freq >> (int)(intptr_t)tsc_timecounter.tc_priv;
562}
563
564static int
565sysctl_machdep_tsc_freq(SYSCTL_HANDLER_ARGS)
566{
567	int error;
568	uint64_t freq;
569
570	freq = atomic_load_acq_64(&tsc_freq);
571	if (freq == 0)
572		return (EOPNOTSUPP);
573	error = sysctl_handle_64(oidp, &freq, 0, req);
574	if (error == 0 && req->newptr != NULL) {
575		atomic_store_rel_64(&tsc_freq, freq);
576		atomic_store_rel_64(&tsc_timecounter.tc_frequency,
577		    freq >> (int)(intptr_t)tsc_timecounter.tc_priv);
578	}
579	return (error);
580}
581
582SYSCTL_PROC(_machdep, OID_AUTO, tsc_freq, CTLTYPE_U64 | CTLFLAG_RW,
583    0, 0, sysctl_machdep_tsc_freq, "QU", "Time Stamp Counter frequency");
584
585static u_int
586tsc_get_timecount(struct timecounter *tc __unused)
587{
588
589	return (rdtsc32());
590}
591
592static u_int
593tsc_get_timecount_low(struct timecounter *tc)
594{
595	uint32_t rv;
596
597	__asm __volatile("rdtsc; shrd %%cl, %%edx, %0"
598	: "=a" (rv) : "c" ((int)(intptr_t)tc->tc_priv) : "edx");
599	return (rv);
600}
601