xref: /freebsd/sys/x86/x86/tsc.c (revision 652a9748855320619e075c4e83aef2f5294412d2)
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/eventhandler.h>
38 #include <sys/limits.h>
39 #include <sys/malloc.h>
40 #include <sys/systm.h>
41 #include <sys/sysctl.h>
42 #include <sys/time.h>
43 #include <sys/timetc.h>
44 #include <sys/kernel.h>
45 #include <sys/power.h>
46 #include <sys/smp.h>
47 #include <sys/vdso.h>
48 #include <machine/clock.h>
49 #include <machine/cputypes.h>
50 #include <machine/md_var.h>
51 #include <machine/specialreg.h>
52 #include <x86/vmware.h>
53 #include <dev/acpica/acpi_hpet.h>
54 #include <contrib/dev/acpica/include/acpi.h>
55 
56 #include "cpufreq_if.h"
57 
58 uint64_t	tsc_freq;
59 int		tsc_is_invariant;
60 int		tsc_perf_stat;
61 
62 static eventhandler_tag tsc_levels_tag, tsc_pre_tag, tsc_post_tag;
63 
64 SYSCTL_INT(_kern_timecounter, OID_AUTO, invariant_tsc, CTLFLAG_RDTUN,
65     &tsc_is_invariant, 0, "Indicates whether the TSC is P-state invariant");
66 
67 #ifdef SMP
68 int	smp_tsc;
69 SYSCTL_INT(_kern_timecounter, OID_AUTO, smp_tsc, CTLFLAG_RDTUN, &smp_tsc, 0,
70     "Indicates whether the TSC is safe to use in SMP mode");
71 
72 int	smp_tsc_adjust = 0;
73 SYSCTL_INT(_kern_timecounter, OID_AUTO, smp_tsc_adjust, CTLFLAG_RDTUN,
74     &smp_tsc_adjust, 0, "Try to adjust TSC on APs to match BSP");
75 #endif
76 
77 static int	tsc_shift = 1;
78 SYSCTL_INT(_kern_timecounter, OID_AUTO, tsc_shift, CTLFLAG_RDTUN,
79     &tsc_shift, 0, "Shift to pre-apply for the maximum TSC frequency");
80 
81 static int	tsc_disabled;
82 SYSCTL_INT(_machdep, OID_AUTO, disable_tsc, CTLFLAG_RDTUN, &tsc_disabled, 0,
83     "Disable x86 Time Stamp Counter");
84 
85 static int	tsc_skip_calibration;
86 SYSCTL_INT(_machdep, OID_AUTO, disable_tsc_calibration, CTLFLAG_RDTUN |
87     CTLFLAG_NOFETCH, &tsc_skip_calibration, 0,
88     "Disable TSC frequency calibration");
89 
90 static void tsc_freq_changed(void *arg, const struct cf_level *level,
91     int status);
92 static void tsc_freq_changing(void *arg, const struct cf_level *level,
93     int *status);
94 static unsigned tsc_get_timecount(struct timecounter *tc);
95 static inline unsigned tsc_get_timecount_low(struct timecounter *tc);
96 static unsigned tsc_get_timecount_lfence(struct timecounter *tc);
97 static unsigned tsc_get_timecount_low_lfence(struct timecounter *tc);
98 static unsigned tsc_get_timecount_mfence(struct timecounter *tc);
99 static unsigned tsc_get_timecount_low_mfence(struct timecounter *tc);
100 static void tsc_levels_changed(void *arg, int unit);
101 static uint32_t x86_tsc_vdso_timehands(struct vdso_timehands *vdso_th,
102     struct timecounter *tc);
103 #ifdef COMPAT_FREEBSD32
104 static uint32_t x86_tsc_vdso_timehands32(struct vdso_timehands32 *vdso_th32,
105     struct timecounter *tc);
106 #endif
107 
108 static struct timecounter tsc_timecounter = {
109 	.tc_get_timecount =		tsc_get_timecount,
110 	.tc_counter_mask =		~0u,
111 	.tc_name =			"TSC",
112 	.tc_quality =			800,	/* adjusted in code */
113 	.tc_fill_vdso_timehands = 	x86_tsc_vdso_timehands,
114 #ifdef COMPAT_FREEBSD32
115 	.tc_fill_vdso_timehands32 = 	x86_tsc_vdso_timehands32,
116 #endif
117 };
118 
119 static void
120 tsc_freq_vmware(void)
121 {
122 	u_int regs[4];
123 
124 	if (hv_high >= 0x40000010) {
125 		do_cpuid(0x40000010, regs);
126 		tsc_freq = regs[0] * 1000;
127 	} else {
128 		vmware_hvcall(VMW_HVCMD_GETHZ, regs);
129 		if (regs[1] != UINT_MAX)
130 			tsc_freq = regs[0] | ((uint64_t)regs[1] << 32);
131 	}
132 	tsc_is_invariant = 1;
133 }
134 
135 /*
136  * Calculate TSC frequency using information from the CPUID leaf 0x15
137  * 'Time Stamp Counter and Nominal Core Crystal Clock'.  If leaf 0x15
138  * is not functional, as it is on Skylake/Kabylake, try 0x16 'Processor
139  * Frequency Information'.  Leaf 0x16 is described in the SDM as
140  * informational only, but if 0x15 did not work, and TSC calibration
141  * is disabled, it is the best we can get at all.  It should still be
142  * an improvement over the parsing of the CPU model name in
143  * tsc_freq_intel(), when available.
144  */
145 static bool
146 tsc_freq_cpuid(uint64_t *res)
147 {
148 	u_int regs[4];
149 
150 	if (cpu_high < 0x15)
151 		return (false);
152 	do_cpuid(0x15, regs);
153 	if (regs[0] != 0 && regs[1] != 0 && regs[2] != 0) {
154 		*res = (uint64_t)regs[2] * regs[1] / regs[0];
155 		return (true);
156 	}
157 
158 	if (cpu_high < 0x16)
159 		return (false);
160 	do_cpuid(0x16, regs);
161 	if (regs[0] != 0) {
162 		*res = (uint64_t)regs[0] * 1000000;
163 		return (true);
164 	}
165 
166 	return (false);
167 }
168 
169 static void
170 tsc_freq_intel(void)
171 {
172 	char brand[48];
173 	u_int regs[4];
174 	uint64_t freq;
175 	char *p;
176 	u_int i;
177 
178 	/*
179 	 * Intel Processor Identification and the CPUID Instruction
180 	 * Application Note 485.
181 	 * http://www.intel.com/assets/pdf/appnote/241618.pdf
182 	 */
183 	if (cpu_exthigh >= 0x80000004) {
184 		p = brand;
185 		for (i = 0x80000002; i < 0x80000005; i++) {
186 			do_cpuid(i, regs);
187 			memcpy(p, regs, sizeof(regs));
188 			p += sizeof(regs);
189 		}
190 		p = NULL;
191 		for (i = 0; i < sizeof(brand) - 1; i++)
192 			if (brand[i] == 'H' && brand[i + 1] == 'z')
193 				p = brand + i;
194 		if (p != NULL) {
195 			p -= 5;
196 			switch (p[4]) {
197 			case 'M':
198 				i = 1;
199 				break;
200 			case 'G':
201 				i = 1000;
202 				break;
203 			case 'T':
204 				i = 1000000;
205 				break;
206 			default:
207 				return;
208 			}
209 #define	C2D(c)	((c) - '0')
210 			if (p[1] == '.') {
211 				freq = C2D(p[0]) * 1000;
212 				freq += C2D(p[2]) * 100;
213 				freq += C2D(p[3]) * 10;
214 				freq *= i * 1000;
215 			} else {
216 				freq = C2D(p[0]) * 1000;
217 				freq += C2D(p[1]) * 100;
218 				freq += C2D(p[2]) * 10;
219 				freq += C2D(p[3]);
220 				freq *= i * 1000000;
221 			}
222 #undef C2D
223 			tsc_freq = freq;
224 		}
225 	}
226 }
227 
228 static void
229 probe_tsc_freq(void)
230 {
231 	uint64_t tmp_freq, tsc1, tsc2;
232 	int no_cpuid_override;
233 	uint16_t bootflags;
234 
235 	if (cpu_power_ecx & CPUID_PERF_STAT) {
236 		/*
237 		 * XXX Some emulators expose host CPUID without actual support
238 		 * for these MSRs.  We must test whether they really work.
239 		 */
240 		wrmsr(MSR_MPERF, 0);
241 		wrmsr(MSR_APERF, 0);
242 		DELAY(10);
243 		if (rdmsr(MSR_MPERF) > 0 && rdmsr(MSR_APERF) > 0)
244 			tsc_perf_stat = 1;
245 	}
246 
247 	if (vm_guest == VM_GUEST_VMWARE) {
248 		tsc_freq_vmware();
249 		return;
250 	}
251 
252 	switch (cpu_vendor_id) {
253 	case CPU_VENDOR_AMD:
254 	case CPU_VENDOR_HYGON:
255 		if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
256 		    (vm_guest == VM_GUEST_NO &&
257 		    CPUID_TO_FAMILY(cpu_id) >= 0x10))
258 			tsc_is_invariant = 1;
259 		if (cpu_feature & CPUID_SSE2) {
260 			tsc_timecounter.tc_get_timecount =
261 			    tsc_get_timecount_mfence;
262 		}
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 		if (cpu_feature & CPUID_SSE2) {
273 			tsc_timecounter.tc_get_timecount =
274 			    tsc_get_timecount_lfence;
275 		}
276 		break;
277 	case CPU_VENDOR_CENTAUR:
278 		if (vm_guest == VM_GUEST_NO &&
279 		    CPUID_TO_FAMILY(cpu_id) == 0x6 &&
280 		    CPUID_TO_MODEL(cpu_id) >= 0xf &&
281 		    (rdmsr(0x1203) & 0x100000000ULL) == 0)
282 			tsc_is_invariant = 1;
283 		if (cpu_feature & CPUID_SSE2) {
284 			tsc_timecounter.tc_get_timecount =
285 			    tsc_get_timecount_lfence;
286 		}
287 		break;
288 	}
289 
290 	if (!TUNABLE_INT_FETCH("machdep.disable_tsc_calibration",
291 	    &tsc_skip_calibration)) {
292 		/*
293 		 * User did not give the order about calibration.
294 		 * If he did, we do not try to guess.
295 		 *
296 		 * Otherwise, if ACPI FADT reports that the platform
297 		 * is legacy-free and CPUID provides TSC frequency,
298 		 * use it.  The calibration could fail anyway since
299 		 * ISA timer can be absent or power gated.
300 		 */
301 		if (acpi_get_fadt_bootflags(&bootflags) &&
302 		    (bootflags & ACPI_FADT_LEGACY_DEVICES) == 0 &&
303 		    tsc_freq_cpuid(&tmp_freq)) {
304 			printf("Skipping TSC calibration since no legacy "
305 			    "devices reported by FADT and CPUID works\n");
306 			tsc_skip_calibration = 1;
307 		}
308 	}
309 	if (tsc_skip_calibration) {
310 		if (tsc_freq_cpuid(&tmp_freq))
311 			tsc_freq = tmp_freq;
312 		else if (cpu_vendor_id == CPU_VENDOR_INTEL)
313 			tsc_freq_intel();
314 	} else {
315 		if (bootverbose)
316 			printf("Calibrating TSC clock ... ");
317 		tsc1 = rdtsc();
318 		DELAY(1000000);
319 		tsc2 = rdtsc();
320 		tsc_freq = tsc2 - tsc1;
321 
322 		/*
323 		 * If the difference between calibrated frequency and
324 		 * the frequency reported by CPUID 0x15/0x16 leafs
325 		 * differ significantly, this probably means that
326 		 * calibration is bogus.  It happens on machines
327 		 * without 8254 timer and with BIOS not properly
328 		 * reporting it in FADT boot flags.
329 		 */
330 		if (tsc_freq_cpuid(&tmp_freq) && qabs(tsc_freq - tmp_freq) >
331 		    uqmin(tsc_freq, tmp_freq)) {
332 			no_cpuid_override = 0;
333 			TUNABLE_INT_FETCH("machdep.disable_tsc_cpuid_override",
334 			    &no_cpuid_override);
335 			if (!no_cpuid_override) {
336 				if (bootverbose) {
337 					printf(
338 	"TSC clock: calibration freq %ju Hz, CPUID freq %ju Hz%s\n",
339 					    (uintmax_t)tsc_freq,
340 					    (uintmax_t)tmp_freq,
341 					    no_cpuid_override ? "" :
342 					    ", doing CPUID override");
343 				}
344 				tsc_freq = tmp_freq;
345 			}
346 		}
347 	}
348 	if (bootverbose)
349 		printf("TSC clock: %ju Hz\n", (intmax_t)tsc_freq);
350 }
351 
352 void
353 init_TSC(void)
354 {
355 
356 	if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
357 		return;
358 
359 #ifdef __i386__
360 	/* The TSC is known to be broken on certain CPUs. */
361 	switch (cpu_vendor_id) {
362 	case CPU_VENDOR_AMD:
363 		switch (cpu_id & 0xFF0) {
364 		case 0x500:
365 			/* K5 Model 0 */
366 			return;
367 		}
368 		break;
369 	case CPU_VENDOR_CENTAUR:
370 		switch (cpu_id & 0xff0) {
371 		case 0x540:
372 			/*
373 			 * http://www.centtech.com/c6_data_sheet.pdf
374 			 *
375 			 * I-12 RDTSC may return incoherent values in EDX:EAX
376 			 * I-13 RDTSC hangs when certain event counters are used
377 			 */
378 			return;
379 		}
380 		break;
381 	case CPU_VENDOR_NSC:
382 		switch (cpu_id & 0xff0) {
383 		case 0x540:
384 			if ((cpu_id & CPUID_STEPPING) == 0)
385 				return;
386 			break;
387 		}
388 		break;
389 	}
390 #endif
391 
392 	probe_tsc_freq();
393 
394 	/*
395 	 * Inform CPU accounting about our boot-time clock rate.  This will
396 	 * be updated if someone loads a cpufreq driver after boot that
397 	 * discovers a new max frequency.
398 	 */
399 	if (tsc_freq != 0)
400 		set_cputicker(rdtsc, tsc_freq, !tsc_is_invariant);
401 
402 	if (tsc_is_invariant)
403 		return;
404 
405 	/* Register to find out about changes in CPU frequency. */
406 	tsc_pre_tag = EVENTHANDLER_REGISTER(cpufreq_pre_change,
407 	    tsc_freq_changing, NULL, EVENTHANDLER_PRI_FIRST);
408 	tsc_post_tag = EVENTHANDLER_REGISTER(cpufreq_post_change,
409 	    tsc_freq_changed, NULL, EVENTHANDLER_PRI_FIRST);
410 	tsc_levels_tag = EVENTHANDLER_REGISTER(cpufreq_levels_changed,
411 	    tsc_levels_changed, NULL, EVENTHANDLER_PRI_ANY);
412 }
413 
414 #ifdef SMP
415 
416 /*
417  * RDTSC is not a serializing instruction, and does not drain
418  * instruction stream, so we need to drain the stream before executing
419  * it.  It could be fixed by use of RDTSCP, except the instruction is
420  * not available everywhere.
421  *
422  * Use CPUID for draining in the boot-time SMP constistency test.  The
423  * timecounters use MFENCE for AMD CPUs, and LFENCE for others (Intel
424  * and VIA) when SSE2 is present, and nothing on older machines which
425  * also do not issue RDTSC prematurely.  There, testing for SSE2 and
426  * vendor is too cumbersome, and we learn about TSC presence from CPUID.
427  *
428  * Do not use do_cpuid(), since we do not need CPUID results, which
429  * have to be written into memory with do_cpuid().
430  */
431 #define	TSC_READ(x)							\
432 static void								\
433 tsc_read_##x(void *arg)							\
434 {									\
435 	uint64_t *tsc = arg;						\
436 	u_int cpu = PCPU_GET(cpuid);					\
437 									\
438 	__asm __volatile("cpuid" : : : "eax", "ebx", "ecx", "edx");	\
439 	tsc[cpu * 3 + x] = rdtsc();					\
440 }
441 TSC_READ(0)
442 TSC_READ(1)
443 TSC_READ(2)
444 #undef TSC_READ
445 
446 #define	N	1000
447 
448 static void
449 comp_smp_tsc(void *arg)
450 {
451 	uint64_t *tsc;
452 	int64_t d1, d2;
453 	u_int cpu = PCPU_GET(cpuid);
454 	u_int i, j, size;
455 
456 	size = (mp_maxid + 1) * 3;
457 	for (i = 0, tsc = arg; i < N; i++, tsc += size)
458 		CPU_FOREACH(j) {
459 			if (j == cpu)
460 				continue;
461 			d1 = tsc[cpu * 3 + 1] - tsc[j * 3];
462 			d2 = tsc[cpu * 3 + 2] - tsc[j * 3 + 1];
463 			if (d1 <= 0 || d2 <= 0) {
464 				smp_tsc = 0;
465 				return;
466 			}
467 		}
468 }
469 
470 static void
471 adj_smp_tsc(void *arg)
472 {
473 	uint64_t *tsc;
474 	int64_t d, min, max;
475 	u_int cpu = PCPU_GET(cpuid);
476 	u_int first, i, size;
477 
478 	first = CPU_FIRST();
479 	if (cpu == first)
480 		return;
481 	min = INT64_MIN;
482 	max = INT64_MAX;
483 	size = (mp_maxid + 1) * 3;
484 	for (i = 0, tsc = arg; i < N; i++, tsc += size) {
485 		d = tsc[first * 3] - tsc[cpu * 3 + 1];
486 		if (d > min)
487 			min = d;
488 		d = tsc[first * 3 + 1] - tsc[cpu * 3 + 2];
489 		if (d > min)
490 			min = d;
491 		d = tsc[first * 3 + 1] - tsc[cpu * 3];
492 		if (d < max)
493 			max = d;
494 		d = tsc[first * 3 + 2] - tsc[cpu * 3 + 1];
495 		if (d < max)
496 			max = d;
497 	}
498 	if (min > max)
499 		return;
500 	d = min / 2 + max / 2;
501 	__asm __volatile (
502 		"movl $0x10, %%ecx\n\t"
503 		"rdmsr\n\t"
504 		"addl %%edi, %%eax\n\t"
505 		"adcl %%esi, %%edx\n\t"
506 		"wrmsr\n"
507 		: /* No output */
508 		: "D" ((uint32_t)d), "S" ((uint32_t)(d >> 32))
509 		: "ax", "cx", "dx", "cc"
510 	);
511 }
512 
513 static int
514 test_tsc(int adj_max_count)
515 {
516 	uint64_t *data, *tsc;
517 	u_int i, size, adj;
518 
519 	if ((!smp_tsc && !tsc_is_invariant) || vm_guest)
520 		return (-100);
521 	size = (mp_maxid + 1) * 3;
522 	data = malloc(sizeof(*data) * size * N, M_TEMP, M_WAITOK);
523 	adj = 0;
524 retry:
525 	for (i = 0, tsc = data; i < N; i++, tsc += size)
526 		smp_rendezvous(tsc_read_0, tsc_read_1, tsc_read_2, tsc);
527 	smp_tsc = 1;	/* XXX */
528 	smp_rendezvous(smp_no_rendezvous_barrier, comp_smp_tsc,
529 	    smp_no_rendezvous_barrier, data);
530 	if (!smp_tsc && adj < adj_max_count) {
531 		adj++;
532 		smp_rendezvous(smp_no_rendezvous_barrier, adj_smp_tsc,
533 		    smp_no_rendezvous_barrier, data);
534 		goto retry;
535 	}
536 	free(data, M_TEMP);
537 	if (bootverbose)
538 		printf("SMP: %sed TSC synchronization test%s\n",
539 		    smp_tsc ? "pass" : "fail",
540 		    adj > 0 ? " after adjustment" : "");
541 	if (smp_tsc && tsc_is_invariant) {
542 		switch (cpu_vendor_id) {
543 		case CPU_VENDOR_AMD:
544 		case CPU_VENDOR_HYGON:
545 			/*
546 			 * Starting with Family 15h processors, TSC clock
547 			 * source is in the north bridge.  Check whether
548 			 * we have a single-socket/multi-core platform.
549 			 * XXX Need more work for complex cases.
550 			 */
551 			if (CPUID_TO_FAMILY(cpu_id) < 0x15 ||
552 			    (amd_feature2 & AMDID2_CMP) == 0 ||
553 			    smp_cpus > (cpu_procinfo2 & AMDID_CMP_CORES) + 1)
554 				break;
555 			return (1000);
556 		case CPU_VENDOR_INTEL:
557 			/*
558 			 * XXX Assume Intel platforms have synchronized TSCs.
559 			 */
560 			return (1000);
561 		}
562 		return (800);
563 	}
564 	return (-100);
565 }
566 
567 #undef N
568 
569 #endif /* SMP */
570 
571 static void
572 init_TSC_tc(void)
573 {
574 	uint64_t max_freq;
575 	int shift;
576 
577 	if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
578 		return;
579 
580 	/*
581 	 * Limit timecounter frequency to fit in an int and prevent it from
582 	 * overflowing too fast.
583 	 */
584 	max_freq = UINT_MAX;
585 
586 	/*
587 	 * We can not use the TSC if we support APM.  Precise timekeeping
588 	 * on an APM'ed machine is at best a fools pursuit, since
589 	 * any and all of the time spent in various SMM code can't
590 	 * be reliably accounted for.  Reading the RTC is your only
591 	 * source of reliable time info.  The i8254 loses too, of course,
592 	 * but we need to have some kind of time...
593 	 * We don't know at this point whether APM is going to be used
594 	 * or not, nor when it might be activated.  Play it safe.
595 	 */
596 	if (power_pm_get_type() == POWER_PM_TYPE_APM) {
597 		tsc_timecounter.tc_quality = -1000;
598 		if (bootverbose)
599 			printf("TSC timecounter disabled: APM enabled.\n");
600 		goto init;
601 	}
602 
603 	/*
604 	 * Intel CPUs without a C-state invariant TSC can stop the TSC
605 	 * in either C2 or C3.  Disable use of C2 and C3 while using
606 	 * the TSC as the timecounter.  The timecounter can be changed
607 	 * to enable C2 and C3.
608 	 *
609 	 * Note that the TSC is used as the cputicker for computing
610 	 * thread runtime regardless of the timecounter setting, so
611 	 * using an alternate timecounter and enabling C2 or C3 can
612 	 * result incorrect runtimes for kernel idle threads (but not
613 	 * for any non-idle threads).
614 	 */
615 	if (cpu_vendor_id == CPU_VENDOR_INTEL &&
616 	    (amd_pminfo & AMDPM_TSC_INVARIANT) == 0) {
617 		tsc_timecounter.tc_flags |= TC_FLAGS_C2STOP;
618 		if (bootverbose)
619 			printf("TSC timecounter disables C2 and C3.\n");
620 	}
621 
622 	/*
623 	 * We can not use the TSC in SMP mode unless the TSCs on all CPUs
624 	 * are synchronized.  If the user is sure that the system has
625 	 * synchronized TSCs, set kern.timecounter.smp_tsc tunable to a
626 	 * non-zero value.  The TSC seems unreliable in virtualized SMP
627 	 * environments, so it is set to a negative quality in those cases.
628 	 */
629 #ifdef SMP
630 	if (mp_ncpus > 1)
631 		tsc_timecounter.tc_quality = test_tsc(smp_tsc_adjust);
632 	else
633 #endif /* SMP */
634 	if (tsc_is_invariant)
635 		tsc_timecounter.tc_quality = 1000;
636 	max_freq >>= tsc_shift;
637 
638 init:
639 	for (shift = 0; shift <= 31 && (tsc_freq >> shift) > max_freq; shift++)
640 		;
641 	if ((cpu_feature & CPUID_SSE2) != 0 && mp_ncpus > 1) {
642 		if (cpu_vendor_id == CPU_VENDOR_AMD ||
643 		    cpu_vendor_id == CPU_VENDOR_HYGON) {
644 			tsc_timecounter.tc_get_timecount = shift > 0 ?
645 			    tsc_get_timecount_low_mfence :
646 			    tsc_get_timecount_mfence;
647 		} else {
648 			tsc_timecounter.tc_get_timecount = shift > 0 ?
649 			    tsc_get_timecount_low_lfence :
650 			    tsc_get_timecount_lfence;
651 		}
652 	} else {
653 		tsc_timecounter.tc_get_timecount = shift > 0 ?
654 		    tsc_get_timecount_low : tsc_get_timecount;
655 	}
656 	if (shift > 0) {
657 		tsc_timecounter.tc_name = "TSC-low";
658 		if (bootverbose)
659 			printf("TSC timecounter discards lower %d bit(s)\n",
660 			    shift);
661 	}
662 	if (tsc_freq != 0) {
663 		tsc_timecounter.tc_frequency = tsc_freq >> shift;
664 		tsc_timecounter.tc_priv = (void *)(intptr_t)shift;
665 		tc_init(&tsc_timecounter);
666 	}
667 }
668 SYSINIT(tsc_tc, SI_SUB_SMP, SI_ORDER_ANY, init_TSC_tc, NULL);
669 
670 void
671 resume_TSC(void)
672 {
673 #ifdef SMP
674 	int quality;
675 
676 	/* If TSC was not good on boot, it is unlikely to become good now. */
677 	if (tsc_timecounter.tc_quality < 0)
678 		return;
679 	/* Nothing to do with UP. */
680 	if (mp_ncpus < 2)
681 		return;
682 
683 	/*
684 	 * If TSC was good, a single synchronization should be enough,
685 	 * but honour smp_tsc_adjust if it's set.
686 	 */
687 	quality = test_tsc(MAX(smp_tsc_adjust, 1));
688 	if (quality != tsc_timecounter.tc_quality) {
689 		printf("TSC timecounter quality changed: %d -> %d\n",
690 		    tsc_timecounter.tc_quality, quality);
691 		tsc_timecounter.tc_quality = quality;
692 	}
693 #endif /* SMP */
694 }
695 
696 /*
697  * When cpufreq levels change, find out about the (new) max frequency.  We
698  * use this to update CPU accounting in case it got a lower estimate at boot.
699  */
700 static void
701 tsc_levels_changed(void *arg, int unit)
702 {
703 	device_t cf_dev;
704 	struct cf_level *levels;
705 	int count, error;
706 	uint64_t max_freq;
707 
708 	/* Only use values from the first CPU, assuming all are equal. */
709 	if (unit != 0)
710 		return;
711 
712 	/* Find the appropriate cpufreq device instance. */
713 	cf_dev = devclass_get_device(devclass_find("cpufreq"), unit);
714 	if (cf_dev == NULL) {
715 		printf("tsc_levels_changed() called but no cpufreq device?\n");
716 		return;
717 	}
718 
719 	/* Get settings from the device and find the max frequency. */
720 	count = 64;
721 	levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
722 	if (levels == NULL)
723 		return;
724 	error = CPUFREQ_LEVELS(cf_dev, levels, &count);
725 	if (error == 0 && count != 0) {
726 		max_freq = (uint64_t)levels[0].total_set.freq * 1000000;
727 		set_cputicker(rdtsc, max_freq, 1);
728 	} else
729 		printf("tsc_levels_changed: no max freq found\n");
730 	free(levels, M_TEMP);
731 }
732 
733 /*
734  * If the TSC timecounter is in use, veto the pending change.  It may be
735  * possible in the future to handle a dynamically-changing timecounter rate.
736  */
737 static void
738 tsc_freq_changing(void *arg, const struct cf_level *level, int *status)
739 {
740 
741 	if (*status != 0 || timecounter != &tsc_timecounter)
742 		return;
743 
744 	printf("timecounter TSC must not be in use when "
745 	    "changing frequencies; change denied\n");
746 	*status = EBUSY;
747 }
748 
749 /* Update TSC freq with the value indicated by the caller. */
750 static void
751 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
752 {
753 	uint64_t freq;
754 
755 	/* If there was an error during the transition, don't do anything. */
756 	if (tsc_disabled || status != 0)
757 		return;
758 
759 	/* Total setting for this level gives the new frequency in MHz. */
760 	freq = (uint64_t)level->total_set.freq * 1000000;
761 	atomic_store_rel_64(&tsc_freq, freq);
762 	tsc_timecounter.tc_frequency =
763 	    freq >> (int)(intptr_t)tsc_timecounter.tc_priv;
764 }
765 
766 static int
767 sysctl_machdep_tsc_freq(SYSCTL_HANDLER_ARGS)
768 {
769 	int error;
770 	uint64_t freq;
771 
772 	freq = atomic_load_acq_64(&tsc_freq);
773 	if (freq == 0)
774 		return (EOPNOTSUPP);
775 	error = sysctl_handle_64(oidp, &freq, 0, req);
776 	if (error == 0 && req->newptr != NULL) {
777 		atomic_store_rel_64(&tsc_freq, freq);
778 		atomic_store_rel_64(&tsc_timecounter.tc_frequency,
779 		    freq >> (int)(intptr_t)tsc_timecounter.tc_priv);
780 	}
781 	return (error);
782 }
783 
784 SYSCTL_PROC(_machdep, OID_AUTO, tsc_freq,
785     CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
786     0, 0, sysctl_machdep_tsc_freq, "QU",
787     "Time Stamp Counter frequency");
788 
789 static u_int
790 tsc_get_timecount(struct timecounter *tc __unused)
791 {
792 
793 	return (rdtsc32());
794 }
795 
796 static inline u_int
797 tsc_get_timecount_low(struct timecounter *tc)
798 {
799 	uint32_t rv;
800 
801 	__asm __volatile("rdtsc; shrd %%cl, %%edx, %0"
802 	    : "=a" (rv) : "c" ((int)(intptr_t)tc->tc_priv) : "edx");
803 	return (rv);
804 }
805 
806 static u_int
807 tsc_get_timecount_lfence(struct timecounter *tc __unused)
808 {
809 
810 	lfence();
811 	return (rdtsc32());
812 }
813 
814 static u_int
815 tsc_get_timecount_low_lfence(struct timecounter *tc)
816 {
817 
818 	lfence();
819 	return (tsc_get_timecount_low(tc));
820 }
821 
822 static u_int
823 tsc_get_timecount_mfence(struct timecounter *tc __unused)
824 {
825 
826 	mfence();
827 	return (rdtsc32());
828 }
829 
830 static u_int
831 tsc_get_timecount_low_mfence(struct timecounter *tc)
832 {
833 
834 	mfence();
835 	return (tsc_get_timecount_low(tc));
836 }
837 
838 static uint32_t
839 x86_tsc_vdso_timehands(struct vdso_timehands *vdso_th, struct timecounter *tc)
840 {
841 
842 	vdso_th->th_algo = VDSO_TH_ALGO_X86_TSC;
843 	vdso_th->th_x86_shift = (int)(intptr_t)tc->tc_priv;
844 	vdso_th->th_x86_hpet_idx = 0xffffffff;
845 	bzero(vdso_th->th_res, sizeof(vdso_th->th_res));
846 	return (1);
847 }
848 
849 #ifdef COMPAT_FREEBSD32
850 static uint32_t
851 x86_tsc_vdso_timehands32(struct vdso_timehands32 *vdso_th32,
852     struct timecounter *tc)
853 {
854 
855 	vdso_th32->th_algo = VDSO_TH_ALGO_X86_TSC;
856 	vdso_th32->th_x86_shift = (int)(intptr_t)tc->tc_priv;
857 	vdso_th32->th_x86_hpet_idx = 0xffffffff;
858 	bzero(vdso_th32->th_res, sizeof(vdso_th32->th_res));
859 	return (1);
860 }
861 #endif
862