tsc.c revision 276724
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 276724 2015-01-05 20:44:44Z jhb $");
29
30#include "opt_compat.h"
31#include "opt_clock.h"
32
33#include <sys/param.h>
34#include <sys/bus.h>
35#include <sys/cpu.h>
36#include <sys/limits.h>
37#include <sys/malloc.h>
38#include <sys/systm.h>
39#include <sys/sysctl.h>
40#include <sys/time.h>
41#include <sys/timetc.h>
42#include <sys/kernel.h>
43#include <sys/power.h>
44#include <sys/smp.h>
45#include <sys/vdso.h>
46#include <machine/clock.h>
47#include <machine/cputypes.h>
48#include <machine/md_var.h>
49#include <machine/specialreg.h>
50#include <x86/vmware.h>
51
52#include "cpufreq_if.h"
53
54uint64_t	tsc_freq;
55int		tsc_is_invariant;
56int		tsc_perf_stat;
57
58static eventhandler_tag tsc_levels_tag, tsc_pre_tag, tsc_post_tag;
59
60SYSCTL_INT(_kern_timecounter, OID_AUTO, invariant_tsc, CTLFLAG_RDTUN,
61    &tsc_is_invariant, 0, "Indicates whether the TSC is P-state invariant");
62
63#ifdef SMP
64int	smp_tsc;
65SYSCTL_INT(_kern_timecounter, OID_AUTO, smp_tsc, CTLFLAG_RDTUN, &smp_tsc, 0,
66    "Indicates whether the TSC is safe to use in SMP mode");
67
68int	smp_tsc_adjust = 0;
69SYSCTL_INT(_kern_timecounter, OID_AUTO, smp_tsc_adjust, CTLFLAG_RDTUN,
70    &smp_tsc_adjust, 0, "Try to adjust TSC on APs to match BSP");
71#endif
72
73static int	tsc_shift = 1;
74SYSCTL_INT(_kern_timecounter, OID_AUTO, tsc_shift, CTLFLAG_RDTUN,
75    &tsc_shift, 0, "Shift to pre-apply for the maximum TSC frequency");
76
77static int	tsc_disabled;
78SYSCTL_INT(_machdep, OID_AUTO, disable_tsc, CTLFLAG_RDTUN, &tsc_disabled, 0,
79    "Disable x86 Time Stamp Counter");
80
81static int	tsc_skip_calibration;
82SYSCTL_INT(_machdep, OID_AUTO, disable_tsc_calibration, CTLFLAG_RDTUN,
83    &tsc_skip_calibration, 0, "Disable TSC frequency calibration");
84
85static void tsc_freq_changed(void *arg, const struct cf_level *level,
86    int status);
87static void tsc_freq_changing(void *arg, const struct cf_level *level,
88    int *status);
89static unsigned tsc_get_timecount(struct timecounter *tc);
90static inline unsigned tsc_get_timecount_low(struct timecounter *tc);
91static unsigned tsc_get_timecount_lfence(struct timecounter *tc);
92static unsigned tsc_get_timecount_low_lfence(struct timecounter *tc);
93static unsigned tsc_get_timecount_mfence(struct timecounter *tc);
94static unsigned tsc_get_timecount_low_mfence(struct timecounter *tc);
95static void tsc_levels_changed(void *arg, int unit);
96
97static struct timecounter tsc_timecounter = {
98	tsc_get_timecount,	/* get_timecount */
99	0,			/* no poll_pps */
100	~0u,			/* counter_mask */
101	0,			/* frequency */
102	"TSC",			/* name */
103	800,			/* quality (adjusted in code) */
104};
105
106static void
107tsc_freq_vmware(void)
108{
109	u_int regs[4];
110
111	if (hv_high >= 0x40000010) {
112		do_cpuid(0x40000010, regs);
113		tsc_freq = regs[0] * 1000;
114	} else {
115		vmware_hvcall(VMW_HVCMD_GETHZ, regs);
116		if (regs[1] != UINT_MAX)
117			tsc_freq = regs[0] | ((uint64_t)regs[1] << 32);
118	}
119	tsc_is_invariant = 1;
120}
121
122static void
123tsc_freq_intel(void)
124{
125	char brand[48];
126	u_int regs[4];
127	uint64_t freq;
128	char *p;
129	u_int i;
130
131	/*
132	 * Intel Processor Identification and the CPUID Instruction
133	 * Application Note 485.
134	 * http://www.intel.com/assets/pdf/appnote/241618.pdf
135	 */
136	if (cpu_exthigh >= 0x80000004) {
137		p = brand;
138		for (i = 0x80000002; i < 0x80000005; i++) {
139			do_cpuid(i, regs);
140			memcpy(p, regs, sizeof(regs));
141			p += sizeof(regs);
142		}
143		p = NULL;
144		for (i = 0; i < sizeof(brand) - 1; i++)
145			if (brand[i] == 'H' && brand[i + 1] == 'z')
146				p = brand + i;
147		if (p != NULL) {
148			p -= 5;
149			switch (p[4]) {
150			case 'M':
151				i = 1;
152				break;
153			case 'G':
154				i = 1000;
155				break;
156			case 'T':
157				i = 1000000;
158				break;
159			default:
160				return;
161			}
162#define	C2D(c)	((c) - '0')
163			if (p[1] == '.') {
164				freq = C2D(p[0]) * 1000;
165				freq += C2D(p[2]) * 100;
166				freq += C2D(p[3]) * 10;
167				freq *= i * 1000;
168			} else {
169				freq = C2D(p[0]) * 1000;
170				freq += C2D(p[1]) * 100;
171				freq += C2D(p[2]) * 10;
172				freq += C2D(p[3]);
173				freq *= i * 1000000;
174			}
175#undef C2D
176			tsc_freq = freq;
177		}
178	}
179}
180
181static void
182probe_tsc_freq(void)
183{
184	u_int regs[4];
185	uint64_t tsc1, tsc2;
186
187	if (cpu_high >= 6) {
188		do_cpuid(6, regs);
189		if ((regs[2] & CPUID_PERF_STAT) != 0) {
190			/*
191			 * XXX Some emulators expose host CPUID without actual
192			 * support for these MSRs.  We must test whether they
193			 * really work.
194			 */
195			wrmsr(MSR_MPERF, 0);
196			wrmsr(MSR_APERF, 0);
197			DELAY(10);
198			if (rdmsr(MSR_MPERF) > 0 && rdmsr(MSR_APERF) > 0)
199				tsc_perf_stat = 1;
200		}
201	}
202
203	if (vm_guest == VM_GUEST_VMWARE) {
204		tsc_freq_vmware();
205		return;
206	}
207
208	switch (cpu_vendor_id) {
209	case CPU_VENDOR_AMD:
210		if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
211		    (vm_guest == VM_GUEST_NO &&
212		    CPUID_TO_FAMILY(cpu_id) >= 0x10))
213			tsc_is_invariant = 1;
214		if (cpu_feature & CPUID_SSE2) {
215			tsc_timecounter.tc_get_timecount =
216			    tsc_get_timecount_mfence;
217		}
218		break;
219	case CPU_VENDOR_INTEL:
220		if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
221		    (vm_guest == VM_GUEST_NO &&
222		    ((CPUID_TO_FAMILY(cpu_id) == 0x6 &&
223		    CPUID_TO_MODEL(cpu_id) >= 0xe) ||
224		    (CPUID_TO_FAMILY(cpu_id) == 0xf &&
225		    CPUID_TO_MODEL(cpu_id) >= 0x3))))
226			tsc_is_invariant = 1;
227		if (cpu_feature & CPUID_SSE2) {
228			tsc_timecounter.tc_get_timecount =
229			    tsc_get_timecount_lfence;
230		}
231		break;
232	case CPU_VENDOR_CENTAUR:
233		if (vm_guest == VM_GUEST_NO &&
234		    CPUID_TO_FAMILY(cpu_id) == 0x6 &&
235		    CPUID_TO_MODEL(cpu_id) >= 0xf &&
236		    (rdmsr(0x1203) & 0x100000000ULL) == 0)
237			tsc_is_invariant = 1;
238		if (cpu_feature & CPUID_SSE2) {
239			tsc_timecounter.tc_get_timecount =
240			    tsc_get_timecount_lfence;
241		}
242		break;
243	}
244
245	if (tsc_skip_calibration) {
246		if (cpu_vendor_id == CPU_VENDOR_INTEL)
247			tsc_freq_intel();
248		return;
249	}
250
251	if (bootverbose)
252	        printf("Calibrating TSC clock ... ");
253	tsc1 = rdtsc();
254	DELAY(1000000);
255	tsc2 = rdtsc();
256	tsc_freq = tsc2 - tsc1;
257	if (bootverbose)
258		printf("TSC clock: %ju Hz\n", (intmax_t)tsc_freq);
259}
260
261void
262init_TSC(void)
263{
264
265	if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
266		return;
267
268#ifdef __i386__
269	/* The TSC is known to be broken on certain CPUs. */
270	switch (cpu_vendor_id) {
271	case CPU_VENDOR_AMD:
272		switch (cpu_id & 0xFF0) {
273		case 0x500:
274			/* K5 Model 0 */
275			return;
276		}
277		break;
278	case CPU_VENDOR_CENTAUR:
279		switch (cpu_id & 0xff0) {
280		case 0x540:
281			/*
282			 * http://www.centtech.com/c6_data_sheet.pdf
283			 *
284			 * I-12 RDTSC may return incoherent values in EDX:EAX
285			 * I-13 RDTSC hangs when certain event counters are used
286			 */
287			return;
288		}
289		break;
290	case CPU_VENDOR_NSC:
291		switch (cpu_id & 0xff0) {
292		case 0x540:
293			if ((cpu_id & CPUID_STEPPING) == 0)
294				return;
295			break;
296		}
297		break;
298	}
299#endif
300
301	probe_tsc_freq();
302
303	/*
304	 * Inform CPU accounting about our boot-time clock rate.  This will
305	 * be updated if someone loads a cpufreq driver after boot that
306	 * discovers a new max frequency.
307	 */
308	if (tsc_freq != 0)
309		set_cputicker(rdtsc, tsc_freq, !tsc_is_invariant);
310
311	if (tsc_is_invariant)
312		return;
313
314	/* Register to find out about changes in CPU frequency. */
315	tsc_pre_tag = EVENTHANDLER_REGISTER(cpufreq_pre_change,
316	    tsc_freq_changing, NULL, EVENTHANDLER_PRI_FIRST);
317	tsc_post_tag = EVENTHANDLER_REGISTER(cpufreq_post_change,
318	    tsc_freq_changed, NULL, EVENTHANDLER_PRI_FIRST);
319	tsc_levels_tag = EVENTHANDLER_REGISTER(cpufreq_levels_changed,
320	    tsc_levels_changed, NULL, EVENTHANDLER_PRI_ANY);
321}
322
323#ifdef SMP
324
325/*
326 * RDTSC is not a serializing instruction, and does not drain
327 * instruction stream, so we need to drain the stream before executing
328 * it.  It could be fixed by use of RDTSCP, except the instruction is
329 * not available everywhere.
330 *
331 * Use CPUID for draining in the boot-time SMP constistency test.  The
332 * timecounters use MFENCE for AMD CPUs, and LFENCE for others (Intel
333 * and VIA) when SSE2 is present, and nothing on older machines which
334 * also do not issue RDTSC prematurely.  There, testing for SSE2 and
335 * vendor is too cumbersome, and we learn about TSC presence from CPUID.
336 *
337 * Do not use do_cpuid(), since we do not need CPUID results, which
338 * have to be written into memory with do_cpuid().
339 */
340#define	TSC_READ(x)							\
341static void								\
342tsc_read_##x(void *arg)							\
343{									\
344	uint64_t *tsc = arg;						\
345	u_int cpu = PCPU_GET(cpuid);					\
346									\
347	__asm __volatile("cpuid" : : : "eax", "ebx", "ecx", "edx");	\
348	tsc[cpu * 3 + x] = rdtsc();					\
349}
350TSC_READ(0)
351TSC_READ(1)
352TSC_READ(2)
353#undef TSC_READ
354
355#define	N	1000
356
357static void
358comp_smp_tsc(void *arg)
359{
360	uint64_t *tsc;
361	int64_t d1, d2;
362	u_int cpu = PCPU_GET(cpuid);
363	u_int i, j, size;
364
365	size = (mp_maxid + 1) * 3;
366	for (i = 0, tsc = arg; i < N; i++, tsc += size)
367		CPU_FOREACH(j) {
368			if (j == cpu)
369				continue;
370			d1 = tsc[cpu * 3 + 1] - tsc[j * 3];
371			d2 = tsc[cpu * 3 + 2] - tsc[j * 3 + 1];
372			if (d1 <= 0 || d2 <= 0) {
373				smp_tsc = 0;
374				return;
375			}
376		}
377}
378
379static void
380adj_smp_tsc(void *arg)
381{
382	uint64_t *tsc;
383	int64_t d, min, max;
384	u_int cpu = PCPU_GET(cpuid);
385	u_int first, i, size;
386
387	first = CPU_FIRST();
388	if (cpu == first)
389		return;
390	min = INT64_MIN;
391	max = INT64_MAX;
392	size = (mp_maxid + 1) * 3;
393	for (i = 0, tsc = arg; i < N; i++, tsc += size) {
394		d = tsc[first * 3] - tsc[cpu * 3 + 1];
395		if (d > min)
396			min = d;
397		d = tsc[first * 3 + 1] - tsc[cpu * 3 + 2];
398		if (d > min)
399			min = d;
400		d = tsc[first * 3 + 1] - tsc[cpu * 3];
401		if (d < max)
402			max = d;
403		d = tsc[first * 3 + 2] - tsc[cpu * 3 + 1];
404		if (d < max)
405			max = d;
406	}
407	if (min > max)
408		return;
409	d = min / 2 + max / 2;
410	__asm __volatile (
411		"movl $0x10, %%ecx\n\t"
412		"rdmsr\n\t"
413		"addl %%edi, %%eax\n\t"
414		"adcl %%esi, %%edx\n\t"
415		"wrmsr\n"
416		: /* No output */
417		: "D" ((uint32_t)d), "S" ((uint32_t)(d >> 32))
418		: "ax", "cx", "dx", "cc"
419	);
420}
421
422static int
423test_tsc(void)
424{
425	uint64_t *data, *tsc;
426	u_int i, size, adj;
427
428	if ((!smp_tsc && !tsc_is_invariant) || vm_guest)
429		return (-100);
430	size = (mp_maxid + 1) * 3;
431	data = malloc(sizeof(*data) * size * N, M_TEMP, M_WAITOK);
432	adj = 0;
433retry:
434	for (i = 0, tsc = data; i < N; i++, tsc += size)
435		smp_rendezvous(tsc_read_0, tsc_read_1, tsc_read_2, tsc);
436	smp_tsc = 1;	/* XXX */
437	smp_rendezvous(smp_no_rendevous_barrier, comp_smp_tsc,
438	    smp_no_rendevous_barrier, data);
439	if (!smp_tsc && adj < smp_tsc_adjust) {
440		adj++;
441		smp_rendezvous(smp_no_rendevous_barrier, adj_smp_tsc,
442		    smp_no_rendevous_barrier, data);
443		goto retry;
444	}
445	free(data, M_TEMP);
446	if (bootverbose)
447		printf("SMP: %sed TSC synchronization test%s\n",
448		    smp_tsc ? "pass" : "fail",
449		    adj > 0 ? " after adjustment" : "");
450	if (smp_tsc && tsc_is_invariant) {
451		switch (cpu_vendor_id) {
452		case CPU_VENDOR_AMD:
453			/*
454			 * Starting with Family 15h processors, TSC clock
455			 * source is in the north bridge.  Check whether
456			 * we have a single-socket/multi-core platform.
457			 * XXX Need more work for complex cases.
458			 */
459			if (CPUID_TO_FAMILY(cpu_id) < 0x15 ||
460			    (amd_feature2 & AMDID2_CMP) == 0 ||
461			    smp_cpus > (cpu_procinfo2 & AMDID_CMP_CORES) + 1)
462				break;
463			return (1000);
464		case CPU_VENDOR_INTEL:
465			/*
466			 * XXX Assume Intel platforms have synchronized TSCs.
467			 */
468			return (1000);
469		}
470		return (800);
471	}
472	return (-100);
473}
474
475#undef N
476
477#else
478
479/*
480 * The function is not called, it is provided to avoid linking failure
481 * on uniprocessor kernel.
482 */
483static int
484test_tsc(void)
485{
486
487	return (0);
488}
489
490#endif /* SMP */
491
492static void
493init_TSC_tc(void)
494{
495	uint64_t max_freq;
496	int shift;
497
498	if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
499		return;
500
501	/*
502	 * Limit timecounter frequency to fit in an int and prevent it from
503	 * overflowing too fast.
504	 */
505	max_freq = UINT_MAX;
506
507	/*
508	 * We can not use the TSC if we support APM.  Precise timekeeping
509	 * on an APM'ed machine is at best a fools pursuit, since
510	 * any and all of the time spent in various SMM code can't
511	 * be reliably accounted for.  Reading the RTC is your only
512	 * source of reliable time info.  The i8254 loses too, of course,
513	 * but we need to have some kind of time...
514	 * We don't know at this point whether APM is going to be used
515	 * or not, nor when it might be activated.  Play it safe.
516	 */
517	if (power_pm_get_type() == POWER_PM_TYPE_APM) {
518		tsc_timecounter.tc_quality = -1000;
519		if (bootverbose)
520			printf("TSC timecounter disabled: APM enabled.\n");
521		goto init;
522	}
523
524	/*
525	 * We cannot use the TSC if it stops incrementing while idle.
526	 * Intel CPUs without a C-state invariant TSC can stop the TSC
527	 * in either C2 or C3.
528	 */
529	if (cpu_deepest_sleep >= 2 && cpu_vendor_id == CPU_VENDOR_INTEL &&
530	    (amd_pminfo & AMDPM_TSC_INVARIANT) == 0) {
531		tsc_timecounter.tc_quality = -1000;
532		tsc_timecounter.tc_flags |= TC_FLAGS_C2STOP;
533		if (bootverbose)
534			printf("TSC timecounter disabled: C2/C3 may halt it.\n");
535		goto init;
536	}
537
538	/*
539	 * We can not use the TSC in SMP mode unless the TSCs on all CPUs
540	 * are synchronized.  If the user is sure that the system has
541	 * synchronized TSCs, set kern.timecounter.smp_tsc tunable to a
542	 * non-zero value.  The TSC seems unreliable in virtualized SMP
543	 * environments, so it is set to a negative quality in those cases.
544	 */
545	if (mp_ncpus > 1)
546		tsc_timecounter.tc_quality = test_tsc();
547	else if (tsc_is_invariant)
548		tsc_timecounter.tc_quality = 1000;
549	max_freq >>= tsc_shift;
550
551init:
552	for (shift = 0; shift <= 31 && (tsc_freq >> shift) > max_freq; shift++)
553		;
554	if ((cpu_feature & CPUID_SSE2) != 0 && mp_ncpus > 1) {
555		if (cpu_vendor_id == CPU_VENDOR_AMD) {
556			tsc_timecounter.tc_get_timecount = shift > 0 ?
557			    tsc_get_timecount_low_mfence :
558			    tsc_get_timecount_mfence;
559		} else {
560			tsc_timecounter.tc_get_timecount = shift > 0 ?
561			    tsc_get_timecount_low_lfence :
562			    tsc_get_timecount_lfence;
563		}
564	} else {
565		tsc_timecounter.tc_get_timecount = shift > 0 ?
566		    tsc_get_timecount_low : tsc_get_timecount;
567	}
568	if (shift > 0) {
569		tsc_timecounter.tc_name = "TSC-low";
570		if (bootverbose)
571			printf("TSC timecounter discards lower %d bit(s)\n",
572			    shift);
573	}
574	if (tsc_freq != 0) {
575		tsc_timecounter.tc_frequency = tsc_freq >> shift;
576		tsc_timecounter.tc_priv = (void *)(intptr_t)shift;
577		tc_init(&tsc_timecounter);
578	}
579}
580SYSINIT(tsc_tc, SI_SUB_SMP, SI_ORDER_ANY, init_TSC_tc, NULL);
581
582/*
583 * When cpufreq levels change, find out about the (new) max frequency.  We
584 * use this to update CPU accounting in case it got a lower estimate at boot.
585 */
586static void
587tsc_levels_changed(void *arg, int unit)
588{
589	device_t cf_dev;
590	struct cf_level *levels;
591	int count, error;
592	uint64_t max_freq;
593
594	/* Only use values from the first CPU, assuming all are equal. */
595	if (unit != 0)
596		return;
597
598	/* Find the appropriate cpufreq device instance. */
599	cf_dev = devclass_get_device(devclass_find("cpufreq"), unit);
600	if (cf_dev == NULL) {
601		printf("tsc_levels_changed() called but no cpufreq device?\n");
602		return;
603	}
604
605	/* Get settings from the device and find the max frequency. */
606	count = 64;
607	levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
608	if (levels == NULL)
609		return;
610	error = CPUFREQ_LEVELS(cf_dev, levels, &count);
611	if (error == 0 && count != 0) {
612		max_freq = (uint64_t)levels[0].total_set.freq * 1000000;
613		set_cputicker(rdtsc, max_freq, 1);
614	} else
615		printf("tsc_levels_changed: no max freq found\n");
616	free(levels, M_TEMP);
617}
618
619/*
620 * If the TSC timecounter is in use, veto the pending change.  It may be
621 * possible in the future to handle a dynamically-changing timecounter rate.
622 */
623static void
624tsc_freq_changing(void *arg, const struct cf_level *level, int *status)
625{
626
627	if (*status != 0 || timecounter != &tsc_timecounter)
628		return;
629
630	printf("timecounter TSC must not be in use when "
631	    "changing frequencies; change denied\n");
632	*status = EBUSY;
633}
634
635/* Update TSC freq with the value indicated by the caller. */
636static void
637tsc_freq_changed(void *arg, const struct cf_level *level, int status)
638{
639	uint64_t freq;
640
641	/* If there was an error during the transition, don't do anything. */
642	if (tsc_disabled || status != 0)
643		return;
644
645	/* Total setting for this level gives the new frequency in MHz. */
646	freq = (uint64_t)level->total_set.freq * 1000000;
647	atomic_store_rel_64(&tsc_freq, freq);
648	tsc_timecounter.tc_frequency =
649	    freq >> (int)(intptr_t)tsc_timecounter.tc_priv;
650}
651
652static int
653sysctl_machdep_tsc_freq(SYSCTL_HANDLER_ARGS)
654{
655	int error;
656	uint64_t freq;
657
658	freq = atomic_load_acq_64(&tsc_freq);
659	if (freq == 0)
660		return (EOPNOTSUPP);
661	error = sysctl_handle_64(oidp, &freq, 0, req);
662	if (error == 0 && req->newptr != NULL) {
663		atomic_store_rel_64(&tsc_freq, freq);
664		atomic_store_rel_64(&tsc_timecounter.tc_frequency,
665		    freq >> (int)(intptr_t)tsc_timecounter.tc_priv);
666	}
667	return (error);
668}
669
670SYSCTL_PROC(_machdep, OID_AUTO, tsc_freq, CTLTYPE_U64 | CTLFLAG_RW,
671    0, 0, sysctl_machdep_tsc_freq, "QU", "Time Stamp Counter frequency");
672
673static u_int
674tsc_get_timecount(struct timecounter *tc __unused)
675{
676
677	return (rdtsc32());
678}
679
680static inline u_int
681tsc_get_timecount_low(struct timecounter *tc)
682{
683	uint32_t rv;
684
685	__asm __volatile("rdtsc; shrd %%cl, %%edx, %0"
686	    : "=a" (rv) : "c" ((int)(intptr_t)tc->tc_priv) : "edx");
687	return (rv);
688}
689
690static u_int
691tsc_get_timecount_lfence(struct timecounter *tc __unused)
692{
693
694	lfence();
695	return (rdtsc32());
696}
697
698static u_int
699tsc_get_timecount_low_lfence(struct timecounter *tc)
700{
701
702	lfence();
703	return (tsc_get_timecount_low(tc));
704}
705
706static u_int
707tsc_get_timecount_mfence(struct timecounter *tc __unused)
708{
709
710	mfence();
711	return (rdtsc32());
712}
713
714static u_int
715tsc_get_timecount_low_mfence(struct timecounter *tc)
716{
717
718	mfence();
719	return (tsc_get_timecount_low(tc));
720}
721
722uint32_t
723cpu_fill_vdso_timehands(struct vdso_timehands *vdso_th)
724{
725
726	vdso_th->th_x86_shift = (int)(intptr_t)timecounter->tc_priv;
727	bzero(vdso_th->th_res, sizeof(vdso_th->th_res));
728	return (timecounter == &tsc_timecounter);
729}
730
731#ifdef COMPAT_FREEBSD32
732uint32_t
733cpu_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
734{
735
736	vdso_th32->th_x86_shift = (int)(intptr_t)timecounter->tc_priv;
737	bzero(vdso_th32->th_res, sizeof(vdso_th32->th_res));
738	return (timecounter == &tsc_timecounter);
739}
740#endif
741