2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 1998-2003 Poul-Henning Kamp
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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.
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
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
32 #include "opt_compat.h"
33 #include "opt_clock.h"
35 #include <sys/param.h>
38 #include <sys/limits.h>
39 #include <sys/malloc.h>
40 #include <sys/systm.h>
41 #include <sys/sysctl.h>
43 #include <sys/timetc.h>
44 #include <sys/kernel.h>
45 #include <sys/power.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>
55 #include "cpufreq_if.h"
61 static eventhandler_tag tsc_levels_tag, tsc_pre_tag, tsc_post_tag;
63 SYSCTL_INT(_kern_timecounter, OID_AUTO, invariant_tsc, CTLFLAG_RDTUN,
64 &tsc_is_invariant, 0, "Indicates whether the TSC is P-state invariant");
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");
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");
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");
80 static int tsc_disabled;
81 SYSCTL_INT(_machdep, OID_AUTO, disable_tsc, CTLFLAG_RDTUN, &tsc_disabled, 0,
82 "Disable x86 Time Stamp Counter");
84 static int tsc_skip_calibration;
85 SYSCTL_INT(_machdep, OID_AUTO, disable_tsc_calibration, CTLFLAG_RDTUN,
86 &tsc_skip_calibration, 0, "Disable TSC frequency calibration");
88 static void tsc_freq_changed(void *arg, const struct cf_level *level,
90 static void tsc_freq_changing(void *arg, const struct cf_level *level,
92 static unsigned tsc_get_timecount(struct timecounter *tc);
93 static inline unsigned tsc_get_timecount_low(struct timecounter *tc);
94 static unsigned tsc_get_timecount_lfence(struct timecounter *tc);
95 static unsigned tsc_get_timecount_low_lfence(struct timecounter *tc);
96 static unsigned tsc_get_timecount_mfence(struct timecounter *tc);
97 static unsigned tsc_get_timecount_low_mfence(struct timecounter *tc);
98 static void tsc_levels_changed(void *arg, int unit);
99 static uint32_t x86_tsc_vdso_timehands(struct vdso_timehands *vdso_th,
100 struct timecounter *tc);
101 #ifdef COMPAT_FREEBSD32
102 static uint32_t x86_tsc_vdso_timehands32(struct vdso_timehands32 *vdso_th32,
103 struct timecounter *tc);
106 static struct timecounter tsc_timecounter = {
107 .tc_get_timecount = tsc_get_timecount,
108 .tc_counter_mask = ~0u,
110 .tc_quality = 800, /* adjusted in code */
111 .tc_fill_vdso_timehands = x86_tsc_vdso_timehands,
112 #ifdef COMPAT_FREEBSD32
113 .tc_fill_vdso_timehands32 = x86_tsc_vdso_timehands32,
118 tsc_freq_vmware(void)
122 if (hv_high >= 0x40000010) {
123 do_cpuid(0x40000010, regs);
124 tsc_freq = regs[0] * 1000;
126 vmware_hvcall(VMW_HVCMD_GETHZ, regs);
127 if (regs[1] != UINT_MAX)
128 tsc_freq = regs[0] | ((uint64_t)regs[1] << 32);
130 tsc_is_invariant = 1;
143 * Intel Processor Identification and the CPUID Instruction
144 * Application Note 485.
145 * http://www.intel.com/assets/pdf/appnote/241618.pdf
147 if (cpu_exthigh >= 0x80000004) {
149 for (i = 0x80000002; i < 0x80000005; i++) {
151 memcpy(p, regs, sizeof(regs));
155 for (i = 0; i < sizeof(brand) - 1; i++)
156 if (brand[i] == 'H' && brand[i + 1] == 'z')
173 #define C2D(c) ((c) - '0')
175 freq = C2D(p[0]) * 1000;
176 freq += C2D(p[2]) * 100;
177 freq += C2D(p[3]) * 10;
180 freq = C2D(p[0]) * 1000;
181 freq += C2D(p[1]) * 100;
182 freq += C2D(p[2]) * 10;
200 if ((regs[2] & CPUID_PERF_STAT) != 0) {
202 * XXX Some emulators expose host CPUID without actual
203 * support for these MSRs. We must test whether they
209 if (rdmsr(MSR_MPERF) > 0 && rdmsr(MSR_APERF) > 0)
214 if (vm_guest == VM_GUEST_VMWARE) {
219 switch (cpu_vendor_id) {
221 if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
222 (vm_guest == VM_GUEST_NO &&
223 CPUID_TO_FAMILY(cpu_id) >= 0x10))
224 tsc_is_invariant = 1;
225 if (cpu_feature & CPUID_SSE2) {
226 tsc_timecounter.tc_get_timecount =
227 tsc_get_timecount_mfence;
230 case CPU_VENDOR_INTEL:
231 if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
232 (vm_guest == VM_GUEST_NO &&
233 ((CPUID_TO_FAMILY(cpu_id) == 0x6 &&
234 CPUID_TO_MODEL(cpu_id) >= 0xe) ||
235 (CPUID_TO_FAMILY(cpu_id) == 0xf &&
236 CPUID_TO_MODEL(cpu_id) >= 0x3))))
237 tsc_is_invariant = 1;
238 if (cpu_feature & CPUID_SSE2) {
239 tsc_timecounter.tc_get_timecount =
240 tsc_get_timecount_lfence;
243 case CPU_VENDOR_CENTAUR:
244 if (vm_guest == VM_GUEST_NO &&
245 CPUID_TO_FAMILY(cpu_id) == 0x6 &&
246 CPUID_TO_MODEL(cpu_id) >= 0xf &&
247 (rdmsr(0x1203) & 0x100000000ULL) == 0)
248 tsc_is_invariant = 1;
249 if (cpu_feature & CPUID_SSE2) {
250 tsc_timecounter.tc_get_timecount =
251 tsc_get_timecount_lfence;
256 if (tsc_skip_calibration) {
257 if (cpu_vendor_id == CPU_VENDOR_INTEL)
263 printf("Calibrating TSC clock ... ");
267 tsc_freq = tsc2 - tsc1;
269 printf("TSC clock: %ju Hz\n", (intmax_t)tsc_freq);
276 if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
280 /* The TSC is known to be broken on certain CPUs. */
281 switch (cpu_vendor_id) {
283 switch (cpu_id & 0xFF0) {
289 case CPU_VENDOR_CENTAUR:
290 switch (cpu_id & 0xff0) {
293 * http://www.centtech.com/c6_data_sheet.pdf
295 * I-12 RDTSC may return incoherent values in EDX:EAX
296 * I-13 RDTSC hangs when certain event counters are used
302 switch (cpu_id & 0xff0) {
304 if ((cpu_id & CPUID_STEPPING) == 0)
315 * Inform CPU accounting about our boot-time clock rate. This will
316 * be updated if someone loads a cpufreq driver after boot that
317 * discovers a new max frequency.
320 set_cputicker(rdtsc, tsc_freq, !tsc_is_invariant);
322 if (tsc_is_invariant)
325 /* Register to find out about changes in CPU frequency. */
326 tsc_pre_tag = EVENTHANDLER_REGISTER(cpufreq_pre_change,
327 tsc_freq_changing, NULL, EVENTHANDLER_PRI_FIRST);
328 tsc_post_tag = EVENTHANDLER_REGISTER(cpufreq_post_change,
329 tsc_freq_changed, NULL, EVENTHANDLER_PRI_FIRST);
330 tsc_levels_tag = EVENTHANDLER_REGISTER(cpufreq_levels_changed,
331 tsc_levels_changed, NULL, EVENTHANDLER_PRI_ANY);
337 * RDTSC is not a serializing instruction, and does not drain
338 * instruction stream, so we need to drain the stream before executing
339 * it. It could be fixed by use of RDTSCP, except the instruction is
340 * not available everywhere.
342 * Use CPUID for draining in the boot-time SMP constistency test. The
343 * timecounters use MFENCE for AMD CPUs, and LFENCE for others (Intel
344 * and VIA) when SSE2 is present, and nothing on older machines which
345 * also do not issue RDTSC prematurely. There, testing for SSE2 and
346 * vendor is too cumbersome, and we learn about TSC presence from CPUID.
348 * Do not use do_cpuid(), since we do not need CPUID results, which
349 * have to be written into memory with do_cpuid().
351 #define TSC_READ(x) \
353 tsc_read_##x(void *arg) \
355 uint64_t *tsc = arg; \
356 u_int cpu = PCPU_GET(cpuid); \
358 __asm __volatile("cpuid" : : : "eax", "ebx", "ecx", "edx"); \
359 tsc[cpu * 3 + x] = rdtsc(); \
369 comp_smp_tsc(void *arg)
373 u_int cpu = PCPU_GET(cpuid);
376 size = (mp_maxid + 1) * 3;
377 for (i = 0, tsc = arg; i < N; i++, tsc += size)
381 d1 = tsc[cpu * 3 + 1] - tsc[j * 3];
382 d2 = tsc[cpu * 3 + 2] - tsc[j * 3 + 1];
383 if (d1 <= 0 || d2 <= 0) {
391 adj_smp_tsc(void *arg)
395 u_int cpu = PCPU_GET(cpuid);
396 u_int first, i, size;
403 size = (mp_maxid + 1) * 3;
404 for (i = 0, tsc = arg; i < N; i++, tsc += size) {
405 d = tsc[first * 3] - tsc[cpu * 3 + 1];
408 d = tsc[first * 3 + 1] - tsc[cpu * 3 + 2];
411 d = tsc[first * 3 + 1] - tsc[cpu * 3];
414 d = tsc[first * 3 + 2] - tsc[cpu * 3 + 1];
420 d = min / 2 + max / 2;
422 "movl $0x10, %%ecx\n\t"
424 "addl %%edi, %%eax\n\t"
425 "adcl %%esi, %%edx\n\t"
428 : "D" ((uint32_t)d), "S" ((uint32_t)(d >> 32))
429 : "ax", "cx", "dx", "cc"
436 uint64_t *data, *tsc;
439 if ((!smp_tsc && !tsc_is_invariant) || vm_guest)
441 size = (mp_maxid + 1) * 3;
442 data = malloc(sizeof(*data) * size * N, M_TEMP, M_WAITOK);
445 for (i = 0, tsc = data; i < N; i++, tsc += size)
446 smp_rendezvous(tsc_read_0, tsc_read_1, tsc_read_2, tsc);
447 smp_tsc = 1; /* XXX */
448 smp_rendezvous(smp_no_rendezvous_barrier, comp_smp_tsc,
449 smp_no_rendezvous_barrier, data);
450 if (!smp_tsc && adj < smp_tsc_adjust) {
452 smp_rendezvous(smp_no_rendezvous_barrier, adj_smp_tsc,
453 smp_no_rendezvous_barrier, data);
458 printf("SMP: %sed TSC synchronization test%s\n",
459 smp_tsc ? "pass" : "fail",
460 adj > 0 ? " after adjustment" : "");
461 if (smp_tsc && tsc_is_invariant) {
462 switch (cpu_vendor_id) {
465 * Starting with Family 15h processors, TSC clock
466 * source is in the north bridge. Check whether
467 * we have a single-socket/multi-core platform.
468 * XXX Need more work for complex cases.
470 if (CPUID_TO_FAMILY(cpu_id) < 0x15 ||
471 (amd_feature2 & AMDID2_CMP) == 0 ||
472 smp_cpus > (cpu_procinfo2 & AMDID_CMP_CORES) + 1)
475 case CPU_VENDOR_INTEL:
477 * XXX Assume Intel platforms have synchronized TSCs.
491 * The function is not called, it is provided to avoid linking failure
492 * on uniprocessor kernel.
509 if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
513 * Limit timecounter frequency to fit in an int and prevent it from
514 * overflowing too fast.
519 * We can not use the TSC if we support APM. Precise timekeeping
520 * on an APM'ed machine is at best a fools pursuit, since
521 * any and all of the time spent in various SMM code can't
522 * be reliably accounted for. Reading the RTC is your only
523 * source of reliable time info. The i8254 loses too, of course,
524 * but we need to have some kind of time...
525 * We don't know at this point whether APM is going to be used
526 * or not, nor when it might be activated. Play it safe.
528 if (power_pm_get_type() == POWER_PM_TYPE_APM) {
529 tsc_timecounter.tc_quality = -1000;
531 printf("TSC timecounter disabled: APM enabled.\n");
536 * Intel CPUs without a C-state invariant TSC can stop the TSC
537 * in either C2 or C3. Disable use of C2 and C3 while using
538 * the TSC as the timecounter. The timecounter can be changed
539 * to enable C2 and C3.
541 * Note that the TSC is used as the cputicker for computing
542 * thread runtime regardless of the timecounter setting, so
543 * using an alternate timecounter and enabling C2 or C3 can
544 * result incorrect runtimes for kernel idle threads (but not
545 * for any non-idle threads).
547 if (cpu_vendor_id == CPU_VENDOR_INTEL &&
548 (amd_pminfo & AMDPM_TSC_INVARIANT) == 0) {
549 tsc_timecounter.tc_flags |= TC_FLAGS_C2STOP;
551 printf("TSC timecounter disables C2 and C3.\n");
555 * We can not use the TSC in SMP mode unless the TSCs on all CPUs
556 * are synchronized. If the user is sure that the system has
557 * synchronized TSCs, set kern.timecounter.smp_tsc tunable to a
558 * non-zero value. The TSC seems unreliable in virtualized SMP
559 * environments, so it is set to a negative quality in those cases.
562 tsc_timecounter.tc_quality = test_tsc();
563 else if (tsc_is_invariant)
564 tsc_timecounter.tc_quality = 1000;
565 max_freq >>= tsc_shift;
568 for (shift = 0; shift <= 31 && (tsc_freq >> shift) > max_freq; shift++)
570 if ((cpu_feature & CPUID_SSE2) != 0 && mp_ncpus > 1) {
571 if (cpu_vendor_id == CPU_VENDOR_AMD) {
572 tsc_timecounter.tc_get_timecount = shift > 0 ?
573 tsc_get_timecount_low_mfence :
574 tsc_get_timecount_mfence;
576 tsc_timecounter.tc_get_timecount = shift > 0 ?
577 tsc_get_timecount_low_lfence :
578 tsc_get_timecount_lfence;
581 tsc_timecounter.tc_get_timecount = shift > 0 ?
582 tsc_get_timecount_low : tsc_get_timecount;
585 tsc_timecounter.tc_name = "TSC-low";
587 printf("TSC timecounter discards lower %d bit(s)\n",
591 tsc_timecounter.tc_frequency = tsc_freq >> shift;
592 tsc_timecounter.tc_priv = (void *)(intptr_t)shift;
593 tc_init(&tsc_timecounter);
596 SYSINIT(tsc_tc, SI_SUB_SMP, SI_ORDER_ANY, init_TSC_tc, NULL);
599 * When cpufreq levels change, find out about the (new) max frequency. We
600 * use this to update CPU accounting in case it got a lower estimate at boot.
603 tsc_levels_changed(void *arg, int unit)
606 struct cf_level *levels;
610 /* Only use values from the first CPU, assuming all are equal. */
614 /* Find the appropriate cpufreq device instance. */
615 cf_dev = devclass_get_device(devclass_find("cpufreq"), unit);
616 if (cf_dev == NULL) {
617 printf("tsc_levels_changed() called but no cpufreq device?\n");
621 /* Get settings from the device and find the max frequency. */
623 levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
626 error = CPUFREQ_LEVELS(cf_dev, levels, &count);
627 if (error == 0 && count != 0) {
628 max_freq = (uint64_t)levels[0].total_set.freq * 1000000;
629 set_cputicker(rdtsc, max_freq, 1);
631 printf("tsc_levels_changed: no max freq found\n");
632 free(levels, M_TEMP);
636 * If the TSC timecounter is in use, veto the pending change. It may be
637 * possible in the future to handle a dynamically-changing timecounter rate.
640 tsc_freq_changing(void *arg, const struct cf_level *level, int *status)
643 if (*status != 0 || timecounter != &tsc_timecounter)
646 printf("timecounter TSC must not be in use when "
647 "changing frequencies; change denied\n");
651 /* Update TSC freq with the value indicated by the caller. */
653 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
657 /* If there was an error during the transition, don't do anything. */
658 if (tsc_disabled || status != 0)
661 /* Total setting for this level gives the new frequency in MHz. */
662 freq = (uint64_t)level->total_set.freq * 1000000;
663 atomic_store_rel_64(&tsc_freq, freq);
664 tsc_timecounter.tc_frequency =
665 freq >> (int)(intptr_t)tsc_timecounter.tc_priv;
669 sysctl_machdep_tsc_freq(SYSCTL_HANDLER_ARGS)
674 freq = atomic_load_acq_64(&tsc_freq);
677 error = sysctl_handle_64(oidp, &freq, 0, req);
678 if (error == 0 && req->newptr != NULL) {
679 atomic_store_rel_64(&tsc_freq, freq);
680 atomic_store_rel_64(&tsc_timecounter.tc_frequency,
681 freq >> (int)(intptr_t)tsc_timecounter.tc_priv);
686 SYSCTL_PROC(_machdep, OID_AUTO, tsc_freq, CTLTYPE_U64 | CTLFLAG_RW,
687 0, 0, sysctl_machdep_tsc_freq, "QU", "Time Stamp Counter frequency");
690 tsc_get_timecount(struct timecounter *tc __unused)
697 tsc_get_timecount_low(struct timecounter *tc)
701 __asm __volatile("rdtsc; shrd %%cl, %%edx, %0"
702 : "=a" (rv) : "c" ((int)(intptr_t)tc->tc_priv) : "edx");
707 tsc_get_timecount_lfence(struct timecounter *tc __unused)
715 tsc_get_timecount_low_lfence(struct timecounter *tc)
719 return (tsc_get_timecount_low(tc));
723 tsc_get_timecount_mfence(struct timecounter *tc __unused)
731 tsc_get_timecount_low_mfence(struct timecounter *tc)
735 return (tsc_get_timecount_low(tc));
739 x86_tsc_vdso_timehands(struct vdso_timehands *vdso_th, struct timecounter *tc)
742 vdso_th->th_algo = VDSO_TH_ALGO_X86_TSC;
743 vdso_th->th_x86_shift = (int)(intptr_t)tc->tc_priv;
744 vdso_th->th_x86_hpet_idx = 0xffffffff;
745 bzero(vdso_th->th_res, sizeof(vdso_th->th_res));
749 #ifdef COMPAT_FREEBSD32
751 x86_tsc_vdso_timehands32(struct vdso_timehands32 *vdso_th32,
752 struct timecounter *tc)
755 vdso_th32->th_algo = VDSO_TH_ALGO_X86_TSC;
756 vdso_th32->th_x86_shift = (int)(intptr_t)tc->tc_priv;
757 vdso_th32->th_x86_hpet_idx = 0xffffffff;
758 bzero(vdso_th32->th_res, sizeof(vdso_th32->th_res));