xref: /freebsd/sys/x86/x86/tsc.c (revision 162ae9c834f6d9f9cb443bd62cceb23e0b5fef48)
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(void)
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 		tsc_freq = (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 		tsc_freq = (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 tsc1, tsc2;
232 	uint16_t bootflags;
233 
234 	if (cpu_power_ecx & CPUID_PERF_STAT) {
235 		/*
236 		 * XXX Some emulators expose host CPUID without actual support
237 		 * for these MSRs.  We must test whether they really work.
238 		 */
239 		wrmsr(MSR_MPERF, 0);
240 		wrmsr(MSR_APERF, 0);
241 		DELAY(10);
242 		if (rdmsr(MSR_MPERF) > 0 && rdmsr(MSR_APERF) > 0)
243 			tsc_perf_stat = 1;
244 	}
245 
246 	if (vm_guest == VM_GUEST_VMWARE) {
247 		tsc_freq_vmware();
248 		return;
249 	}
250 
251 	switch (cpu_vendor_id) {
252 	case CPU_VENDOR_AMD:
253 	case CPU_VENDOR_HYGON:
254 		if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
255 		    (vm_guest == VM_GUEST_NO &&
256 		    CPUID_TO_FAMILY(cpu_id) >= 0x10))
257 			tsc_is_invariant = 1;
258 		if (cpu_feature & CPUID_SSE2) {
259 			tsc_timecounter.tc_get_timecount =
260 			    tsc_get_timecount_mfence;
261 		}
262 		break;
263 	case CPU_VENDOR_INTEL:
264 		if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
265 		    (vm_guest == VM_GUEST_NO &&
266 		    ((CPUID_TO_FAMILY(cpu_id) == 0x6 &&
267 		    CPUID_TO_MODEL(cpu_id) >= 0xe) ||
268 		    (CPUID_TO_FAMILY(cpu_id) == 0xf &&
269 		    CPUID_TO_MODEL(cpu_id) >= 0x3))))
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 	case CPU_VENDOR_CENTAUR:
277 		if (vm_guest == VM_GUEST_NO &&
278 		    CPUID_TO_FAMILY(cpu_id) == 0x6 &&
279 		    CPUID_TO_MODEL(cpu_id) >= 0xf &&
280 		    (rdmsr(0x1203) & 0x100000000ULL) == 0)
281 			tsc_is_invariant = 1;
282 		if (cpu_feature & CPUID_SSE2) {
283 			tsc_timecounter.tc_get_timecount =
284 			    tsc_get_timecount_lfence;
285 		}
286 		break;
287 	}
288 
289 	if (!TUNABLE_INT_FETCH("machdep.disable_tsc_calibration",
290 	    &tsc_skip_calibration)) {
291 		/*
292 		 * User did not give the order about calibration.
293 		 * If he did, we do not try to guess.
294 		 *
295 		 * Otherwise, if ACPI FADT reports that the platform
296 		 * is legacy-free and CPUID provides TSC frequency,
297 		 * use it.  The calibration could fail anyway since
298 		 * ISA timer can be absent or power gated.
299 		 */
300 		if (acpi_get_fadt_bootflags(&bootflags) &&
301 		    (bootflags & ACPI_FADT_LEGACY_DEVICES) == 0 &&
302 		    tsc_freq_cpuid()) {
303 			printf("Skipping TSC calibration since no legacy "
304 			    "devices reported by FADT and CPUID works\n");
305 			tsc_skip_calibration = 1;
306 		}
307 	}
308 	if (tsc_skip_calibration) {
309 		if (tsc_freq_cpuid())
310 			;
311 		else if (cpu_vendor_id == CPU_VENDOR_INTEL)
312 			tsc_freq_intel();
313 	} else {
314 		if (bootverbose)
315 			printf("Calibrating TSC clock ... ");
316 		tsc1 = rdtsc();
317 		DELAY(1000000);
318 		tsc2 = rdtsc();
319 		tsc_freq = tsc2 - tsc1;
320 	}
321 	if (bootverbose)
322 		printf("TSC clock: %ju Hz\n", (intmax_t)tsc_freq);
323 }
324 
325 void
326 init_TSC(void)
327 {
328 
329 	if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
330 		return;
331 
332 #ifdef __i386__
333 	/* The TSC is known to be broken on certain CPUs. */
334 	switch (cpu_vendor_id) {
335 	case CPU_VENDOR_AMD:
336 		switch (cpu_id & 0xFF0) {
337 		case 0x500:
338 			/* K5 Model 0 */
339 			return;
340 		}
341 		break;
342 	case CPU_VENDOR_CENTAUR:
343 		switch (cpu_id & 0xff0) {
344 		case 0x540:
345 			/*
346 			 * http://www.centtech.com/c6_data_sheet.pdf
347 			 *
348 			 * I-12 RDTSC may return incoherent values in EDX:EAX
349 			 * I-13 RDTSC hangs when certain event counters are used
350 			 */
351 			return;
352 		}
353 		break;
354 	case CPU_VENDOR_NSC:
355 		switch (cpu_id & 0xff0) {
356 		case 0x540:
357 			if ((cpu_id & CPUID_STEPPING) == 0)
358 				return;
359 			break;
360 		}
361 		break;
362 	}
363 #endif
364 
365 	probe_tsc_freq();
366 
367 	/*
368 	 * Inform CPU accounting about our boot-time clock rate.  This will
369 	 * be updated if someone loads a cpufreq driver after boot that
370 	 * discovers a new max frequency.
371 	 */
372 	if (tsc_freq != 0)
373 		set_cputicker(rdtsc, tsc_freq, !tsc_is_invariant);
374 
375 	if (tsc_is_invariant)
376 		return;
377 
378 	/* Register to find out about changes in CPU frequency. */
379 	tsc_pre_tag = EVENTHANDLER_REGISTER(cpufreq_pre_change,
380 	    tsc_freq_changing, NULL, EVENTHANDLER_PRI_FIRST);
381 	tsc_post_tag = EVENTHANDLER_REGISTER(cpufreq_post_change,
382 	    tsc_freq_changed, NULL, EVENTHANDLER_PRI_FIRST);
383 	tsc_levels_tag = EVENTHANDLER_REGISTER(cpufreq_levels_changed,
384 	    tsc_levels_changed, NULL, EVENTHANDLER_PRI_ANY);
385 }
386 
387 #ifdef SMP
388 
389 /*
390  * RDTSC is not a serializing instruction, and does not drain
391  * instruction stream, so we need to drain the stream before executing
392  * it.  It could be fixed by use of RDTSCP, except the instruction is
393  * not available everywhere.
394  *
395  * Use CPUID for draining in the boot-time SMP constistency test.  The
396  * timecounters use MFENCE for AMD CPUs, and LFENCE for others (Intel
397  * and VIA) when SSE2 is present, and nothing on older machines which
398  * also do not issue RDTSC prematurely.  There, testing for SSE2 and
399  * vendor is too cumbersome, and we learn about TSC presence from CPUID.
400  *
401  * Do not use do_cpuid(), since we do not need CPUID results, which
402  * have to be written into memory with do_cpuid().
403  */
404 #define	TSC_READ(x)							\
405 static void								\
406 tsc_read_##x(void *arg)							\
407 {									\
408 	uint64_t *tsc = arg;						\
409 	u_int cpu = PCPU_GET(cpuid);					\
410 									\
411 	__asm __volatile("cpuid" : : : "eax", "ebx", "ecx", "edx");	\
412 	tsc[cpu * 3 + x] = rdtsc();					\
413 }
414 TSC_READ(0)
415 TSC_READ(1)
416 TSC_READ(2)
417 #undef TSC_READ
418 
419 #define	N	1000
420 
421 static void
422 comp_smp_tsc(void *arg)
423 {
424 	uint64_t *tsc;
425 	int64_t d1, d2;
426 	u_int cpu = PCPU_GET(cpuid);
427 	u_int i, j, size;
428 
429 	size = (mp_maxid + 1) * 3;
430 	for (i = 0, tsc = arg; i < N; i++, tsc += size)
431 		CPU_FOREACH(j) {
432 			if (j == cpu)
433 				continue;
434 			d1 = tsc[cpu * 3 + 1] - tsc[j * 3];
435 			d2 = tsc[cpu * 3 + 2] - tsc[j * 3 + 1];
436 			if (d1 <= 0 || d2 <= 0) {
437 				smp_tsc = 0;
438 				return;
439 			}
440 		}
441 }
442 
443 static void
444 adj_smp_tsc(void *arg)
445 {
446 	uint64_t *tsc;
447 	int64_t d, min, max;
448 	u_int cpu = PCPU_GET(cpuid);
449 	u_int first, i, size;
450 
451 	first = CPU_FIRST();
452 	if (cpu == first)
453 		return;
454 	min = INT64_MIN;
455 	max = INT64_MAX;
456 	size = (mp_maxid + 1) * 3;
457 	for (i = 0, tsc = arg; i < N; i++, tsc += size) {
458 		d = tsc[first * 3] - tsc[cpu * 3 + 1];
459 		if (d > min)
460 			min = d;
461 		d = tsc[first * 3 + 1] - tsc[cpu * 3 + 2];
462 		if (d > min)
463 			min = d;
464 		d = tsc[first * 3 + 1] - tsc[cpu * 3];
465 		if (d < max)
466 			max = d;
467 		d = tsc[first * 3 + 2] - tsc[cpu * 3 + 1];
468 		if (d < max)
469 			max = d;
470 	}
471 	if (min > max)
472 		return;
473 	d = min / 2 + max / 2;
474 	__asm __volatile (
475 		"movl $0x10, %%ecx\n\t"
476 		"rdmsr\n\t"
477 		"addl %%edi, %%eax\n\t"
478 		"adcl %%esi, %%edx\n\t"
479 		"wrmsr\n"
480 		: /* No output */
481 		: "D" ((uint32_t)d), "S" ((uint32_t)(d >> 32))
482 		: "ax", "cx", "dx", "cc"
483 	);
484 }
485 
486 static int
487 test_tsc(int adj_max_count)
488 {
489 	uint64_t *data, *tsc;
490 	u_int i, size, adj;
491 
492 	if ((!smp_tsc && !tsc_is_invariant) || vm_guest)
493 		return (-100);
494 	size = (mp_maxid + 1) * 3;
495 	data = malloc(sizeof(*data) * size * N, M_TEMP, M_WAITOK);
496 	adj = 0;
497 retry:
498 	for (i = 0, tsc = data; i < N; i++, tsc += size)
499 		smp_rendezvous(tsc_read_0, tsc_read_1, tsc_read_2, tsc);
500 	smp_tsc = 1;	/* XXX */
501 	smp_rendezvous(smp_no_rendezvous_barrier, comp_smp_tsc,
502 	    smp_no_rendezvous_barrier, data);
503 	if (!smp_tsc && adj < adj_max_count) {
504 		adj++;
505 		smp_rendezvous(smp_no_rendezvous_barrier, adj_smp_tsc,
506 		    smp_no_rendezvous_barrier, data);
507 		goto retry;
508 	}
509 	free(data, M_TEMP);
510 	if (bootverbose)
511 		printf("SMP: %sed TSC synchronization test%s\n",
512 		    smp_tsc ? "pass" : "fail",
513 		    adj > 0 ? " after adjustment" : "");
514 	if (smp_tsc && tsc_is_invariant) {
515 		switch (cpu_vendor_id) {
516 		case CPU_VENDOR_AMD:
517 		case CPU_VENDOR_HYGON:
518 			/*
519 			 * Starting with Family 15h processors, TSC clock
520 			 * source is in the north bridge.  Check whether
521 			 * we have a single-socket/multi-core platform.
522 			 * XXX Need more work for complex cases.
523 			 */
524 			if (CPUID_TO_FAMILY(cpu_id) < 0x15 ||
525 			    (amd_feature2 & AMDID2_CMP) == 0 ||
526 			    smp_cpus > (cpu_procinfo2 & AMDID_CMP_CORES) + 1)
527 				break;
528 			return (1000);
529 		case CPU_VENDOR_INTEL:
530 			/*
531 			 * XXX Assume Intel platforms have synchronized TSCs.
532 			 */
533 			return (1000);
534 		}
535 		return (800);
536 	}
537 	return (-100);
538 }
539 
540 #undef N
541 
542 #endif /* SMP */
543 
544 static void
545 init_TSC_tc(void)
546 {
547 	uint64_t max_freq;
548 	int shift;
549 
550 	if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
551 		return;
552 
553 	/*
554 	 * Limit timecounter frequency to fit in an int and prevent it from
555 	 * overflowing too fast.
556 	 */
557 	max_freq = UINT_MAX;
558 
559 	/*
560 	 * We can not use the TSC if we support APM.  Precise timekeeping
561 	 * on an APM'ed machine is at best a fools pursuit, since
562 	 * any and all of the time spent in various SMM code can't
563 	 * be reliably accounted for.  Reading the RTC is your only
564 	 * source of reliable time info.  The i8254 loses too, of course,
565 	 * but we need to have some kind of time...
566 	 * We don't know at this point whether APM is going to be used
567 	 * or not, nor when it might be activated.  Play it safe.
568 	 */
569 	if (power_pm_get_type() == POWER_PM_TYPE_APM) {
570 		tsc_timecounter.tc_quality = -1000;
571 		if (bootverbose)
572 			printf("TSC timecounter disabled: APM enabled.\n");
573 		goto init;
574 	}
575 
576 	/*
577 	 * Intel CPUs without a C-state invariant TSC can stop the TSC
578 	 * in either C2 or C3.  Disable use of C2 and C3 while using
579 	 * the TSC as the timecounter.  The timecounter can be changed
580 	 * to enable C2 and C3.
581 	 *
582 	 * Note that the TSC is used as the cputicker for computing
583 	 * thread runtime regardless of the timecounter setting, so
584 	 * using an alternate timecounter and enabling C2 or C3 can
585 	 * result incorrect runtimes for kernel idle threads (but not
586 	 * for any non-idle threads).
587 	 */
588 	if (cpu_vendor_id == CPU_VENDOR_INTEL &&
589 	    (amd_pminfo & AMDPM_TSC_INVARIANT) == 0) {
590 		tsc_timecounter.tc_flags |= TC_FLAGS_C2STOP;
591 		if (bootverbose)
592 			printf("TSC timecounter disables C2 and C3.\n");
593 	}
594 
595 	/*
596 	 * We can not use the TSC in SMP mode unless the TSCs on all CPUs
597 	 * are synchronized.  If the user is sure that the system has
598 	 * synchronized TSCs, set kern.timecounter.smp_tsc tunable to a
599 	 * non-zero value.  The TSC seems unreliable in virtualized SMP
600 	 * environments, so it is set to a negative quality in those cases.
601 	 */
602 #ifdef SMP
603 	if (mp_ncpus > 1)
604 		tsc_timecounter.tc_quality = test_tsc(smp_tsc_adjust);
605 	else
606 #endif /* SMP */
607 	if (tsc_is_invariant)
608 		tsc_timecounter.tc_quality = 1000;
609 	max_freq >>= tsc_shift;
610 
611 init:
612 	for (shift = 0; shift <= 31 && (tsc_freq >> shift) > max_freq; shift++)
613 		;
614 	if ((cpu_feature & CPUID_SSE2) != 0 && mp_ncpus > 1) {
615 		if (cpu_vendor_id == CPU_VENDOR_AMD ||
616 		    cpu_vendor_id == CPU_VENDOR_HYGON) {
617 			tsc_timecounter.tc_get_timecount = shift > 0 ?
618 			    tsc_get_timecount_low_mfence :
619 			    tsc_get_timecount_mfence;
620 		} else {
621 			tsc_timecounter.tc_get_timecount = shift > 0 ?
622 			    tsc_get_timecount_low_lfence :
623 			    tsc_get_timecount_lfence;
624 		}
625 	} else {
626 		tsc_timecounter.tc_get_timecount = shift > 0 ?
627 		    tsc_get_timecount_low : tsc_get_timecount;
628 	}
629 	if (shift > 0) {
630 		tsc_timecounter.tc_name = "TSC-low";
631 		if (bootverbose)
632 			printf("TSC timecounter discards lower %d bit(s)\n",
633 			    shift);
634 	}
635 	if (tsc_freq != 0) {
636 		tsc_timecounter.tc_frequency = tsc_freq >> shift;
637 		tsc_timecounter.tc_priv = (void *)(intptr_t)shift;
638 		tc_init(&tsc_timecounter);
639 	}
640 }
641 SYSINIT(tsc_tc, SI_SUB_SMP, SI_ORDER_ANY, init_TSC_tc, NULL);
642 
643 void
644 resume_TSC(void)
645 {
646 #ifdef SMP
647 	int quality;
648 
649 	/* If TSC was not good on boot, it is unlikely to become good now. */
650 	if (tsc_timecounter.tc_quality < 0)
651 		return;
652 	/* Nothing to do with UP. */
653 	if (mp_ncpus < 2)
654 		return;
655 
656 	/*
657 	 * If TSC was good, a single synchronization should be enough,
658 	 * but honour smp_tsc_adjust if it's set.
659 	 */
660 	quality = test_tsc(MAX(smp_tsc_adjust, 1));
661 	if (quality != tsc_timecounter.tc_quality) {
662 		printf("TSC timecounter quality changed: %d -> %d\n",
663 		    tsc_timecounter.tc_quality, quality);
664 		tsc_timecounter.tc_quality = quality;
665 	}
666 #endif /* SMP */
667 }
668 
669 /*
670  * When cpufreq levels change, find out about the (new) max frequency.  We
671  * use this to update CPU accounting in case it got a lower estimate at boot.
672  */
673 static void
674 tsc_levels_changed(void *arg, int unit)
675 {
676 	device_t cf_dev;
677 	struct cf_level *levels;
678 	int count, error;
679 	uint64_t max_freq;
680 
681 	/* Only use values from the first CPU, assuming all are equal. */
682 	if (unit != 0)
683 		return;
684 
685 	/* Find the appropriate cpufreq device instance. */
686 	cf_dev = devclass_get_device(devclass_find("cpufreq"), unit);
687 	if (cf_dev == NULL) {
688 		printf("tsc_levels_changed() called but no cpufreq device?\n");
689 		return;
690 	}
691 
692 	/* Get settings from the device and find the max frequency. */
693 	count = 64;
694 	levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
695 	if (levels == NULL)
696 		return;
697 	error = CPUFREQ_LEVELS(cf_dev, levels, &count);
698 	if (error == 0 && count != 0) {
699 		max_freq = (uint64_t)levels[0].total_set.freq * 1000000;
700 		set_cputicker(rdtsc, max_freq, 1);
701 	} else
702 		printf("tsc_levels_changed: no max freq found\n");
703 	free(levels, M_TEMP);
704 }
705 
706 /*
707  * If the TSC timecounter is in use, veto the pending change.  It may be
708  * possible in the future to handle a dynamically-changing timecounter rate.
709  */
710 static void
711 tsc_freq_changing(void *arg, const struct cf_level *level, int *status)
712 {
713 
714 	if (*status != 0 || timecounter != &tsc_timecounter)
715 		return;
716 
717 	printf("timecounter TSC must not be in use when "
718 	    "changing frequencies; change denied\n");
719 	*status = EBUSY;
720 }
721 
722 /* Update TSC freq with the value indicated by the caller. */
723 static void
724 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
725 {
726 	uint64_t freq;
727 
728 	/* If there was an error during the transition, don't do anything. */
729 	if (tsc_disabled || status != 0)
730 		return;
731 
732 	/* Total setting for this level gives the new frequency in MHz. */
733 	freq = (uint64_t)level->total_set.freq * 1000000;
734 	atomic_store_rel_64(&tsc_freq, freq);
735 	tsc_timecounter.tc_frequency =
736 	    freq >> (int)(intptr_t)tsc_timecounter.tc_priv;
737 }
738 
739 static int
740 sysctl_machdep_tsc_freq(SYSCTL_HANDLER_ARGS)
741 {
742 	int error;
743 	uint64_t freq;
744 
745 	freq = atomic_load_acq_64(&tsc_freq);
746 	if (freq == 0)
747 		return (EOPNOTSUPP);
748 	error = sysctl_handle_64(oidp, &freq, 0, req);
749 	if (error == 0 && req->newptr != NULL) {
750 		atomic_store_rel_64(&tsc_freq, freq);
751 		atomic_store_rel_64(&tsc_timecounter.tc_frequency,
752 		    freq >> (int)(intptr_t)tsc_timecounter.tc_priv);
753 	}
754 	return (error);
755 }
756 
757 SYSCTL_PROC(_machdep, OID_AUTO, tsc_freq, CTLTYPE_U64 | CTLFLAG_RW,
758     0, 0, sysctl_machdep_tsc_freq, "QU", "Time Stamp Counter frequency");
759 
760 static u_int
761 tsc_get_timecount(struct timecounter *tc __unused)
762 {
763 
764 	return (rdtsc32());
765 }
766 
767 static inline u_int
768 tsc_get_timecount_low(struct timecounter *tc)
769 {
770 	uint32_t rv;
771 
772 	__asm __volatile("rdtsc; shrd %%cl, %%edx, %0"
773 	    : "=a" (rv) : "c" ((int)(intptr_t)tc->tc_priv) : "edx");
774 	return (rv);
775 }
776 
777 static u_int
778 tsc_get_timecount_lfence(struct timecounter *tc __unused)
779 {
780 
781 	lfence();
782 	return (rdtsc32());
783 }
784 
785 static u_int
786 tsc_get_timecount_low_lfence(struct timecounter *tc)
787 {
788 
789 	lfence();
790 	return (tsc_get_timecount_low(tc));
791 }
792 
793 static u_int
794 tsc_get_timecount_mfence(struct timecounter *tc __unused)
795 {
796 
797 	mfence();
798 	return (rdtsc32());
799 }
800 
801 static u_int
802 tsc_get_timecount_low_mfence(struct timecounter *tc)
803 {
804 
805 	mfence();
806 	return (tsc_get_timecount_low(tc));
807 }
808 
809 static uint32_t
810 x86_tsc_vdso_timehands(struct vdso_timehands *vdso_th, struct timecounter *tc)
811 {
812 
813 	vdso_th->th_algo = VDSO_TH_ALGO_X86_TSC;
814 	vdso_th->th_x86_shift = (int)(intptr_t)tc->tc_priv;
815 	vdso_th->th_x86_hpet_idx = 0xffffffff;
816 	bzero(vdso_th->th_res, sizeof(vdso_th->th_res));
817 	return (1);
818 }
819 
820 #ifdef COMPAT_FREEBSD32
821 static uint32_t
822 x86_tsc_vdso_timehands32(struct vdso_timehands32 *vdso_th32,
823     struct timecounter *tc)
824 {
825 
826 	vdso_th32->th_algo = VDSO_TH_ALGO_X86_TSC;
827 	vdso_th32->th_x86_shift = (int)(intptr_t)tc->tc_priv;
828 	vdso_th32->th_x86_hpet_idx = 0xffffffff;
829 	bzero(vdso_th32->th_res, sizeof(vdso_th32->th_res));
830 	return (1);
831 }
832 #endif
833