xref: /freebsd/sys/x86/x86/tsc.c (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 1998-2003 Poul-Henning Kamp
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include "opt_clock.h"
33 
34 #include <sys/param.h>
35 #include <sys/bus.h>
36 #include <sys/cpu.h>
37 #include <sys/limits.h>
38 #include <sys/malloc.h>
39 #include <sys/systm.h>
40 #include <sys/sysctl.h>
41 #include <sys/time.h>
42 #include <sys/timetc.h>
43 #include <sys/kernel.h>
44 #include <sys/power.h>
45 #include <sys/smp.h>
46 #include <sys/vdso.h>
47 #include <machine/clock.h>
48 #include <machine/cputypes.h>
49 #include <machine/md_var.h>
50 #include <machine/specialreg.h>
51 #include <x86/vmware.h>
52 #include <dev/acpica/acpi_hpet.h>
53 #include <contrib/dev/acpica/include/acpi.h>
54 
55 #include "cpufreq_if.h"
56 
57 uint64_t	tsc_freq;
58 int		tsc_is_invariant;
59 int		tsc_perf_stat;
60 
61 static eventhandler_tag tsc_levels_tag, tsc_pre_tag, tsc_post_tag;
62 
63 SYSCTL_INT(_kern_timecounter, OID_AUTO, invariant_tsc, CTLFLAG_RDTUN,
64     &tsc_is_invariant, 0, "Indicates whether the TSC is P-state invariant");
65 
66 #ifdef SMP
67 int	smp_tsc;
68 SYSCTL_INT(_kern_timecounter, OID_AUTO, smp_tsc, CTLFLAG_RDTUN, &smp_tsc, 0,
69     "Indicates whether the TSC is safe to use in SMP mode");
70 
71 int	smp_tsc_adjust = 0;
72 SYSCTL_INT(_kern_timecounter, OID_AUTO, smp_tsc_adjust, CTLFLAG_RDTUN,
73     &smp_tsc_adjust, 0, "Try to adjust TSC on APs to match BSP");
74 #endif
75 
76 static int	tsc_shift = 1;
77 SYSCTL_INT(_kern_timecounter, OID_AUTO, tsc_shift, CTLFLAG_RDTUN,
78     &tsc_shift, 0, "Shift to pre-apply for the maximum TSC frequency");
79 
80 static int	tsc_disabled;
81 SYSCTL_INT(_machdep, OID_AUTO, disable_tsc, CTLFLAG_RDTUN, &tsc_disabled, 0,
82     "Disable x86 Time Stamp Counter");
83 
84 static int	tsc_skip_calibration;
85 SYSCTL_INT(_machdep, OID_AUTO, disable_tsc_calibration, CTLFLAG_RDTUN |
86     CTLFLAG_NOFETCH, &tsc_skip_calibration, 0,
87     "Disable TSC frequency calibration");
88 
89 static void tsc_freq_changed(void *arg, const struct cf_level *level,
90     int status);
91 static void tsc_freq_changing(void *arg, const struct cf_level *level,
92     int *status);
93 static unsigned tsc_get_timecount(struct timecounter *tc);
94 static inline unsigned tsc_get_timecount_low(struct timecounter *tc);
95 static unsigned tsc_get_timecount_lfence(struct timecounter *tc);
96 static unsigned tsc_get_timecount_low_lfence(struct timecounter *tc);
97 static unsigned tsc_get_timecount_mfence(struct timecounter *tc);
98 static unsigned tsc_get_timecount_low_mfence(struct timecounter *tc);
99 static void tsc_levels_changed(void *arg, int unit);
100 static uint32_t x86_tsc_vdso_timehands(struct vdso_timehands *vdso_th,
101     struct timecounter *tc);
102 #ifdef COMPAT_FREEBSD32
103 static uint32_t x86_tsc_vdso_timehands32(struct vdso_timehands32 *vdso_th32,
104     struct timecounter *tc);
105 #endif
106 
107 static struct timecounter tsc_timecounter = {
108 	.tc_get_timecount =		tsc_get_timecount,
109 	.tc_counter_mask =		~0u,
110 	.tc_name =			"TSC",
111 	.tc_quality =			800,	/* adjusted in code */
112 	.tc_fill_vdso_timehands = 	x86_tsc_vdso_timehands,
113 #ifdef COMPAT_FREEBSD32
114 	.tc_fill_vdso_timehands32 = 	x86_tsc_vdso_timehands32,
115 #endif
116 };
117 
118 static void
119 tsc_freq_vmware(void)
120 {
121 	u_int regs[4];
122 
123 	if (hv_high >= 0x40000010) {
124 		do_cpuid(0x40000010, regs);
125 		tsc_freq = regs[0] * 1000;
126 	} else {
127 		vmware_hvcall(VMW_HVCMD_GETHZ, regs);
128 		if (regs[1] != UINT_MAX)
129 			tsc_freq = regs[0] | ((uint64_t)regs[1] << 32);
130 	}
131 	tsc_is_invariant = 1;
132 }
133 
134 /*
135  * Calculate TSC frequency using information from the CPUID leaf 0x15
136  * 'Time Stamp Counter and Nominal Core Crystal Clock'.  It should be
137  * an improvement over the parsing of the CPU model name in
138  * tsc_freq_intel(), when available.
139  */
140 static bool
141 tsc_freq_cpuid(void)
142 {
143 	u_int regs[4];
144 
145 	if (cpu_high < 0x15)
146 		return (false);
147 	do_cpuid(0x15, regs);
148 	if (regs[0] == 0 || regs[1] == 0 || regs[2] == 0)
149 		return (false);
150 	tsc_freq = (uint64_t)regs[2] * regs[1] / regs[0];
151 	return (true);
152 }
153 
154 static void
155 tsc_freq_intel(void)
156 {
157 	char brand[48];
158 	u_int regs[4];
159 	uint64_t freq;
160 	char *p;
161 	u_int i;
162 
163 	/*
164 	 * Intel Processor Identification and the CPUID Instruction
165 	 * Application Note 485.
166 	 * http://www.intel.com/assets/pdf/appnote/241618.pdf
167 	 */
168 	if (cpu_exthigh >= 0x80000004) {
169 		p = brand;
170 		for (i = 0x80000002; i < 0x80000005; i++) {
171 			do_cpuid(i, regs);
172 			memcpy(p, regs, sizeof(regs));
173 			p += sizeof(regs);
174 		}
175 		p = NULL;
176 		for (i = 0; i < sizeof(brand) - 1; i++)
177 			if (brand[i] == 'H' && brand[i + 1] == 'z')
178 				p = brand + i;
179 		if (p != NULL) {
180 			p -= 5;
181 			switch (p[4]) {
182 			case 'M':
183 				i = 1;
184 				break;
185 			case 'G':
186 				i = 1000;
187 				break;
188 			case 'T':
189 				i = 1000000;
190 				break;
191 			default:
192 				return;
193 			}
194 #define	C2D(c)	((c) - '0')
195 			if (p[1] == '.') {
196 				freq = C2D(p[0]) * 1000;
197 				freq += C2D(p[2]) * 100;
198 				freq += C2D(p[3]) * 10;
199 				freq *= i * 1000;
200 			} else {
201 				freq = C2D(p[0]) * 1000;
202 				freq += C2D(p[1]) * 100;
203 				freq += C2D(p[2]) * 10;
204 				freq += C2D(p[3]);
205 				freq *= i * 1000000;
206 			}
207 #undef C2D
208 			tsc_freq = freq;
209 		}
210 	}
211 }
212 
213 static void
214 probe_tsc_freq(void)
215 {
216 	u_int regs[4];
217 	uint64_t tsc1, tsc2;
218 	uint16_t bootflags;
219 
220 	if (cpu_high >= 6) {
221 		do_cpuid(6, regs);
222 		if ((regs[2] & CPUID_PERF_STAT) != 0) {
223 			/*
224 			 * XXX Some emulators expose host CPUID without actual
225 			 * support for these MSRs.  We must test whether they
226 			 * really work.
227 			 */
228 			wrmsr(MSR_MPERF, 0);
229 			wrmsr(MSR_APERF, 0);
230 			DELAY(10);
231 			if (rdmsr(MSR_MPERF) > 0 && rdmsr(MSR_APERF) > 0)
232 				tsc_perf_stat = 1;
233 		}
234 	}
235 
236 	if (vm_guest == VM_GUEST_VMWARE) {
237 		tsc_freq_vmware();
238 		return;
239 	}
240 
241 	switch (cpu_vendor_id) {
242 	case CPU_VENDOR_AMD:
243 		if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
244 		    (vm_guest == VM_GUEST_NO &&
245 		    CPUID_TO_FAMILY(cpu_id) >= 0x10))
246 			tsc_is_invariant = 1;
247 		if (cpu_feature & CPUID_SSE2) {
248 			tsc_timecounter.tc_get_timecount =
249 			    tsc_get_timecount_mfence;
250 		}
251 		break;
252 	case CPU_VENDOR_INTEL:
253 		if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
254 		    (vm_guest == VM_GUEST_NO &&
255 		    ((CPUID_TO_FAMILY(cpu_id) == 0x6 &&
256 		    CPUID_TO_MODEL(cpu_id) >= 0xe) ||
257 		    (CPUID_TO_FAMILY(cpu_id) == 0xf &&
258 		    CPUID_TO_MODEL(cpu_id) >= 0x3))))
259 			tsc_is_invariant = 1;
260 		if (cpu_feature & CPUID_SSE2) {
261 			tsc_timecounter.tc_get_timecount =
262 			    tsc_get_timecount_lfence;
263 		}
264 		break;
265 	case CPU_VENDOR_CENTAUR:
266 		if (vm_guest == VM_GUEST_NO &&
267 		    CPUID_TO_FAMILY(cpu_id) == 0x6 &&
268 		    CPUID_TO_MODEL(cpu_id) >= 0xf &&
269 		    (rdmsr(0x1203) & 0x100000000ULL) == 0)
270 			tsc_is_invariant = 1;
271 		if (cpu_feature & CPUID_SSE2) {
272 			tsc_timecounter.tc_get_timecount =
273 			    tsc_get_timecount_lfence;
274 		}
275 		break;
276 	}
277 
278 	if (!TUNABLE_INT_FETCH("machdep.disable_tsc_calibration",
279 	    &tsc_skip_calibration)) {
280 		/*
281 		 * User did not give the order about calibration.
282 		 * If he did, we do not try to guess.
283 		 *
284 		 * Otherwise, if ACPI FADT reports that the platform
285 		 * is legacy-free and CPUID provides TSC frequency,
286 		 * use it.  The calibration could fail anyway since
287 		 * ISA timer can be absent or power gated.
288 		 */
289 		if (acpi_get_fadt_bootflags(&bootflags) &&
290 		    (bootflags & ACPI_FADT_LEGACY_DEVICES) == 0 &&
291 		    tsc_freq_cpuid()) {
292 			printf("Skipping TSC calibration since no legacy "
293 			    "devices reported by FADT and CPUID works\n");
294 			tsc_skip_calibration = 1;
295 		}
296 	}
297 	if (tsc_skip_calibration) {
298 		if (tsc_freq_cpuid())
299 			;
300 		else if (cpu_vendor_id == CPU_VENDOR_INTEL)
301 			tsc_freq_intel();
302 	} else {
303 		if (bootverbose)
304 			printf("Calibrating TSC clock ... ");
305 		tsc1 = rdtsc();
306 		DELAY(1000000);
307 		tsc2 = rdtsc();
308 		tsc_freq = tsc2 - tsc1;
309 	}
310 	if (bootverbose)
311 		printf("TSC clock: %ju Hz\n", (intmax_t)tsc_freq);
312 }
313 
314 void
315 init_TSC(void)
316 {
317 
318 	if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
319 		return;
320 
321 #ifdef __i386__
322 	/* The TSC is known to be broken on certain CPUs. */
323 	switch (cpu_vendor_id) {
324 	case CPU_VENDOR_AMD:
325 		switch (cpu_id & 0xFF0) {
326 		case 0x500:
327 			/* K5 Model 0 */
328 			return;
329 		}
330 		break;
331 	case CPU_VENDOR_CENTAUR:
332 		switch (cpu_id & 0xff0) {
333 		case 0x540:
334 			/*
335 			 * http://www.centtech.com/c6_data_sheet.pdf
336 			 *
337 			 * I-12 RDTSC may return incoherent values in EDX:EAX
338 			 * I-13 RDTSC hangs when certain event counters are used
339 			 */
340 			return;
341 		}
342 		break;
343 	case CPU_VENDOR_NSC:
344 		switch (cpu_id & 0xff0) {
345 		case 0x540:
346 			if ((cpu_id & CPUID_STEPPING) == 0)
347 				return;
348 			break;
349 		}
350 		break;
351 	}
352 #endif
353 
354 	probe_tsc_freq();
355 
356 	/*
357 	 * Inform CPU accounting about our boot-time clock rate.  This will
358 	 * be updated if someone loads a cpufreq driver after boot that
359 	 * discovers a new max frequency.
360 	 */
361 	if (tsc_freq != 0)
362 		set_cputicker(rdtsc, tsc_freq, !tsc_is_invariant);
363 
364 	if (tsc_is_invariant)
365 		return;
366 
367 	/* Register to find out about changes in CPU frequency. */
368 	tsc_pre_tag = EVENTHANDLER_REGISTER(cpufreq_pre_change,
369 	    tsc_freq_changing, NULL, EVENTHANDLER_PRI_FIRST);
370 	tsc_post_tag = EVENTHANDLER_REGISTER(cpufreq_post_change,
371 	    tsc_freq_changed, NULL, EVENTHANDLER_PRI_FIRST);
372 	tsc_levels_tag = EVENTHANDLER_REGISTER(cpufreq_levels_changed,
373 	    tsc_levels_changed, NULL, EVENTHANDLER_PRI_ANY);
374 }
375 
376 #ifdef SMP
377 
378 /*
379  * RDTSC is not a serializing instruction, and does not drain
380  * instruction stream, so we need to drain the stream before executing
381  * it.  It could be fixed by use of RDTSCP, except the instruction is
382  * not available everywhere.
383  *
384  * Use CPUID for draining in the boot-time SMP constistency test.  The
385  * timecounters use MFENCE for AMD CPUs, and LFENCE for others (Intel
386  * and VIA) when SSE2 is present, and nothing on older machines which
387  * also do not issue RDTSC prematurely.  There, testing for SSE2 and
388  * vendor is too cumbersome, and we learn about TSC presence from CPUID.
389  *
390  * Do not use do_cpuid(), since we do not need CPUID results, which
391  * have to be written into memory with do_cpuid().
392  */
393 #define	TSC_READ(x)							\
394 static void								\
395 tsc_read_##x(void *arg)							\
396 {									\
397 	uint64_t *tsc = arg;						\
398 	u_int cpu = PCPU_GET(cpuid);					\
399 									\
400 	__asm __volatile("cpuid" : : : "eax", "ebx", "ecx", "edx");	\
401 	tsc[cpu * 3 + x] = rdtsc();					\
402 }
403 TSC_READ(0)
404 TSC_READ(1)
405 TSC_READ(2)
406 #undef TSC_READ
407 
408 #define	N	1000
409 
410 static void
411 comp_smp_tsc(void *arg)
412 {
413 	uint64_t *tsc;
414 	int64_t d1, d2;
415 	u_int cpu = PCPU_GET(cpuid);
416 	u_int i, j, size;
417 
418 	size = (mp_maxid + 1) * 3;
419 	for (i = 0, tsc = arg; i < N; i++, tsc += size)
420 		CPU_FOREACH(j) {
421 			if (j == cpu)
422 				continue;
423 			d1 = tsc[cpu * 3 + 1] - tsc[j * 3];
424 			d2 = tsc[cpu * 3 + 2] - tsc[j * 3 + 1];
425 			if (d1 <= 0 || d2 <= 0) {
426 				smp_tsc = 0;
427 				return;
428 			}
429 		}
430 }
431 
432 static void
433 adj_smp_tsc(void *arg)
434 {
435 	uint64_t *tsc;
436 	int64_t d, min, max;
437 	u_int cpu = PCPU_GET(cpuid);
438 	u_int first, i, size;
439 
440 	first = CPU_FIRST();
441 	if (cpu == first)
442 		return;
443 	min = INT64_MIN;
444 	max = INT64_MAX;
445 	size = (mp_maxid + 1) * 3;
446 	for (i = 0, tsc = arg; i < N; i++, tsc += size) {
447 		d = tsc[first * 3] - tsc[cpu * 3 + 1];
448 		if (d > min)
449 			min = d;
450 		d = tsc[first * 3 + 1] - tsc[cpu * 3 + 2];
451 		if (d > min)
452 			min = d;
453 		d = tsc[first * 3 + 1] - tsc[cpu * 3];
454 		if (d < max)
455 			max = d;
456 		d = tsc[first * 3 + 2] - tsc[cpu * 3 + 1];
457 		if (d < max)
458 			max = d;
459 	}
460 	if (min > max)
461 		return;
462 	d = min / 2 + max / 2;
463 	__asm __volatile (
464 		"movl $0x10, %%ecx\n\t"
465 		"rdmsr\n\t"
466 		"addl %%edi, %%eax\n\t"
467 		"adcl %%esi, %%edx\n\t"
468 		"wrmsr\n"
469 		: /* No output */
470 		: "D" ((uint32_t)d), "S" ((uint32_t)(d >> 32))
471 		: "ax", "cx", "dx", "cc"
472 	);
473 }
474 
475 static int
476 test_tsc(int adj_max_count)
477 {
478 	uint64_t *data, *tsc;
479 	u_int i, size, adj;
480 
481 	if ((!smp_tsc && !tsc_is_invariant) || vm_guest)
482 		return (-100);
483 	size = (mp_maxid + 1) * 3;
484 	data = malloc(sizeof(*data) * size * N, M_TEMP, M_WAITOK);
485 	adj = 0;
486 retry:
487 	for (i = 0, tsc = data; i < N; i++, tsc += size)
488 		smp_rendezvous(tsc_read_0, tsc_read_1, tsc_read_2, tsc);
489 	smp_tsc = 1;	/* XXX */
490 	smp_rendezvous(smp_no_rendezvous_barrier, comp_smp_tsc,
491 	    smp_no_rendezvous_barrier, data);
492 	if (!smp_tsc && adj < adj_max_count) {
493 		adj++;
494 		smp_rendezvous(smp_no_rendezvous_barrier, adj_smp_tsc,
495 		    smp_no_rendezvous_barrier, data);
496 		goto retry;
497 	}
498 	free(data, M_TEMP);
499 	if (bootverbose)
500 		printf("SMP: %sed TSC synchronization test%s\n",
501 		    smp_tsc ? "pass" : "fail",
502 		    adj > 0 ? " after adjustment" : "");
503 	if (smp_tsc && tsc_is_invariant) {
504 		switch (cpu_vendor_id) {
505 		case CPU_VENDOR_AMD:
506 			/*
507 			 * Starting with Family 15h processors, TSC clock
508 			 * source is in the north bridge.  Check whether
509 			 * we have a single-socket/multi-core platform.
510 			 * XXX Need more work for complex cases.
511 			 */
512 			if (CPUID_TO_FAMILY(cpu_id) < 0x15 ||
513 			    (amd_feature2 & AMDID2_CMP) == 0 ||
514 			    smp_cpus > (cpu_procinfo2 & AMDID_CMP_CORES) + 1)
515 				break;
516 			return (1000);
517 		case CPU_VENDOR_INTEL:
518 			/*
519 			 * XXX Assume Intel platforms have synchronized TSCs.
520 			 */
521 			return (1000);
522 		}
523 		return (800);
524 	}
525 	return (-100);
526 }
527 
528 #undef N
529 
530 #endif /* SMP */
531 
532 static void
533 init_TSC_tc(void)
534 {
535 	uint64_t max_freq;
536 	int shift;
537 
538 	if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
539 		return;
540 
541 	/*
542 	 * Limit timecounter frequency to fit in an int and prevent it from
543 	 * overflowing too fast.
544 	 */
545 	max_freq = UINT_MAX;
546 
547 	/*
548 	 * We can not use the TSC if we support APM.  Precise timekeeping
549 	 * on an APM'ed machine is at best a fools pursuit, since
550 	 * any and all of the time spent in various SMM code can't
551 	 * be reliably accounted for.  Reading the RTC is your only
552 	 * source of reliable time info.  The i8254 loses too, of course,
553 	 * but we need to have some kind of time...
554 	 * We don't know at this point whether APM is going to be used
555 	 * or not, nor when it might be activated.  Play it safe.
556 	 */
557 	if (power_pm_get_type() == POWER_PM_TYPE_APM) {
558 		tsc_timecounter.tc_quality = -1000;
559 		if (bootverbose)
560 			printf("TSC timecounter disabled: APM enabled.\n");
561 		goto init;
562 	}
563 
564 	/*
565 	 * Intel CPUs without a C-state invariant TSC can stop the TSC
566 	 * in either C2 or C3.  Disable use of C2 and C3 while using
567 	 * the TSC as the timecounter.  The timecounter can be changed
568 	 * to enable C2 and C3.
569 	 *
570 	 * Note that the TSC is used as the cputicker for computing
571 	 * thread runtime regardless of the timecounter setting, so
572 	 * using an alternate timecounter and enabling C2 or C3 can
573 	 * result incorrect runtimes for kernel idle threads (but not
574 	 * for any non-idle threads).
575 	 */
576 	if (cpu_vendor_id == CPU_VENDOR_INTEL &&
577 	    (amd_pminfo & AMDPM_TSC_INVARIANT) == 0) {
578 		tsc_timecounter.tc_flags |= TC_FLAGS_C2STOP;
579 		if (bootverbose)
580 			printf("TSC timecounter disables C2 and C3.\n");
581 	}
582 
583 	/*
584 	 * We can not use the TSC in SMP mode unless the TSCs on all CPUs
585 	 * are synchronized.  If the user is sure that the system has
586 	 * synchronized TSCs, set kern.timecounter.smp_tsc tunable to a
587 	 * non-zero value.  The TSC seems unreliable in virtualized SMP
588 	 * environments, so it is set to a negative quality in those cases.
589 	 */
590 #ifdef SMP
591 	if (mp_ncpus > 1)
592 		tsc_timecounter.tc_quality = test_tsc(smp_tsc_adjust);
593 	else
594 #endif /* SMP */
595 	if (tsc_is_invariant)
596 		tsc_timecounter.tc_quality = 1000;
597 	max_freq >>= tsc_shift;
598 
599 init:
600 	for (shift = 0; shift <= 31 && (tsc_freq >> shift) > max_freq; shift++)
601 		;
602 	if ((cpu_feature & CPUID_SSE2) != 0 && mp_ncpus > 1) {
603 		if (cpu_vendor_id == CPU_VENDOR_AMD) {
604 			tsc_timecounter.tc_get_timecount = shift > 0 ?
605 			    tsc_get_timecount_low_mfence :
606 			    tsc_get_timecount_mfence;
607 		} else {
608 			tsc_timecounter.tc_get_timecount = shift > 0 ?
609 			    tsc_get_timecount_low_lfence :
610 			    tsc_get_timecount_lfence;
611 		}
612 	} else {
613 		tsc_timecounter.tc_get_timecount = shift > 0 ?
614 		    tsc_get_timecount_low : tsc_get_timecount;
615 	}
616 	if (shift > 0) {
617 		tsc_timecounter.tc_name = "TSC-low";
618 		if (bootverbose)
619 			printf("TSC timecounter discards lower %d bit(s)\n",
620 			    shift);
621 	}
622 	if (tsc_freq != 0) {
623 		tsc_timecounter.tc_frequency = tsc_freq >> shift;
624 		tsc_timecounter.tc_priv = (void *)(intptr_t)shift;
625 		tc_init(&tsc_timecounter);
626 	}
627 }
628 SYSINIT(tsc_tc, SI_SUB_SMP, SI_ORDER_ANY, init_TSC_tc, NULL);
629 
630 void
631 resume_TSC(void)
632 {
633 #ifdef SMP
634 	int quality;
635 
636 	/* If TSC was not good on boot, it is unlikely to become good now. */
637 	if (tsc_timecounter.tc_quality < 0)
638 		return;
639 	/* Nothing to do with UP. */
640 	if (mp_ncpus < 2)
641 		return;
642 
643 	/*
644 	 * If TSC was good, a single synchronization should be enough,
645 	 * but honour smp_tsc_adjust if it's set.
646 	 */
647 	quality = test_tsc(MAX(smp_tsc_adjust, 1));
648 	if (quality != tsc_timecounter.tc_quality) {
649 		printf("TSC timecounter quality changed: %d -> %d\n",
650 		    tsc_timecounter.tc_quality, quality);
651 		tsc_timecounter.tc_quality = quality;
652 	}
653 #endif /* SMP */
654 }
655 
656 /*
657  * When cpufreq levels change, find out about the (new) max frequency.  We
658  * use this to update CPU accounting in case it got a lower estimate at boot.
659  */
660 static void
661 tsc_levels_changed(void *arg, int unit)
662 {
663 	device_t cf_dev;
664 	struct cf_level *levels;
665 	int count, error;
666 	uint64_t max_freq;
667 
668 	/* Only use values from the first CPU, assuming all are equal. */
669 	if (unit != 0)
670 		return;
671 
672 	/* Find the appropriate cpufreq device instance. */
673 	cf_dev = devclass_get_device(devclass_find("cpufreq"), unit);
674 	if (cf_dev == NULL) {
675 		printf("tsc_levels_changed() called but no cpufreq device?\n");
676 		return;
677 	}
678 
679 	/* Get settings from the device and find the max frequency. */
680 	count = 64;
681 	levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
682 	if (levels == NULL)
683 		return;
684 	error = CPUFREQ_LEVELS(cf_dev, levels, &count);
685 	if (error == 0 && count != 0) {
686 		max_freq = (uint64_t)levels[0].total_set.freq * 1000000;
687 		set_cputicker(rdtsc, max_freq, 1);
688 	} else
689 		printf("tsc_levels_changed: no max freq found\n");
690 	free(levels, M_TEMP);
691 }
692 
693 /*
694  * If the TSC timecounter is in use, veto the pending change.  It may be
695  * possible in the future to handle a dynamically-changing timecounter rate.
696  */
697 static void
698 tsc_freq_changing(void *arg, const struct cf_level *level, int *status)
699 {
700 
701 	if (*status != 0 || timecounter != &tsc_timecounter)
702 		return;
703 
704 	printf("timecounter TSC must not be in use when "
705 	    "changing frequencies; change denied\n");
706 	*status = EBUSY;
707 }
708 
709 /* Update TSC freq with the value indicated by the caller. */
710 static void
711 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
712 {
713 	uint64_t freq;
714 
715 	/* If there was an error during the transition, don't do anything. */
716 	if (tsc_disabled || status != 0)
717 		return;
718 
719 	/* Total setting for this level gives the new frequency in MHz. */
720 	freq = (uint64_t)level->total_set.freq * 1000000;
721 	atomic_store_rel_64(&tsc_freq, freq);
722 	tsc_timecounter.tc_frequency =
723 	    freq >> (int)(intptr_t)tsc_timecounter.tc_priv;
724 }
725 
726 static int
727 sysctl_machdep_tsc_freq(SYSCTL_HANDLER_ARGS)
728 {
729 	int error;
730 	uint64_t freq;
731 
732 	freq = atomic_load_acq_64(&tsc_freq);
733 	if (freq == 0)
734 		return (EOPNOTSUPP);
735 	error = sysctl_handle_64(oidp, &freq, 0, req);
736 	if (error == 0 && req->newptr != NULL) {
737 		atomic_store_rel_64(&tsc_freq, freq);
738 		atomic_store_rel_64(&tsc_timecounter.tc_frequency,
739 		    freq >> (int)(intptr_t)tsc_timecounter.tc_priv);
740 	}
741 	return (error);
742 }
743 
744 SYSCTL_PROC(_machdep, OID_AUTO, tsc_freq, CTLTYPE_U64 | CTLFLAG_RW,
745     0, 0, sysctl_machdep_tsc_freq, "QU", "Time Stamp Counter frequency");
746 
747 static u_int
748 tsc_get_timecount(struct timecounter *tc __unused)
749 {
750 
751 	return (rdtsc32());
752 }
753 
754 static inline u_int
755 tsc_get_timecount_low(struct timecounter *tc)
756 {
757 	uint32_t rv;
758 
759 	__asm __volatile("rdtsc; shrd %%cl, %%edx, %0"
760 	    : "=a" (rv) : "c" ((int)(intptr_t)tc->tc_priv) : "edx");
761 	return (rv);
762 }
763 
764 static u_int
765 tsc_get_timecount_lfence(struct timecounter *tc __unused)
766 {
767 
768 	lfence();
769 	return (rdtsc32());
770 }
771 
772 static u_int
773 tsc_get_timecount_low_lfence(struct timecounter *tc)
774 {
775 
776 	lfence();
777 	return (tsc_get_timecount_low(tc));
778 }
779 
780 static u_int
781 tsc_get_timecount_mfence(struct timecounter *tc __unused)
782 {
783 
784 	mfence();
785 	return (rdtsc32());
786 }
787 
788 static u_int
789 tsc_get_timecount_low_mfence(struct timecounter *tc)
790 {
791 
792 	mfence();
793 	return (tsc_get_timecount_low(tc));
794 }
795 
796 static uint32_t
797 x86_tsc_vdso_timehands(struct vdso_timehands *vdso_th, struct timecounter *tc)
798 {
799 
800 	vdso_th->th_algo = VDSO_TH_ALGO_X86_TSC;
801 	vdso_th->th_x86_shift = (int)(intptr_t)tc->tc_priv;
802 	vdso_th->th_x86_hpet_idx = 0xffffffff;
803 	bzero(vdso_th->th_res, sizeof(vdso_th->th_res));
804 	return (1);
805 }
806 
807 #ifdef COMPAT_FREEBSD32
808 static uint32_t
809 x86_tsc_vdso_timehands32(struct vdso_timehands32 *vdso_th32,
810     struct timecounter *tc)
811 {
812 
813 	vdso_th32->th_algo = VDSO_TH_ALGO_X86_TSC;
814 	vdso_th32->th_x86_shift = (int)(intptr_t)tc->tc_priv;
815 	vdso_th32->th_x86_hpet_idx = 0xffffffff;
816 	bzero(vdso_th32->th_res, sizeof(vdso_th32->th_res));
817 	return (1);
818 }
819 #endif
820