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_clock.h"
34 #include <sys/param.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>
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>
54 #include <contrib/dev/acpica/include/acpi.h>
56 #include "cpufreq_if.h"
62 static eventhandler_tag tsc_levels_tag, tsc_pre_tag, tsc_post_tag;
64 SYSCTL_INT(_kern_timecounter, OID_AUTO, invariant_tsc, CTLFLAG_RDTUN,
65 &tsc_is_invariant, 0, "Indicates whether the TSC is P-state invariant");
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");
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");
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");
81 static int tsc_disabled;
82 SYSCTL_INT(_machdep, OID_AUTO, disable_tsc, CTLFLAG_RDTUN, &tsc_disabled, 0,
83 "Disable x86 Time Stamp Counter");
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");
90 static void tsc_freq_changed(void *arg, const struct cf_level *level,
92 static void tsc_freq_changing(void *arg, const struct cf_level *level,
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);
108 static struct timecounter tsc_timecounter = {
109 .tc_get_timecount = tsc_get_timecount,
110 .tc_counter_mask = ~0u,
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,
120 tsc_freq_vmware(void)
124 if (hv_high >= 0x40000010) {
125 do_cpuid(0x40000010, regs);
126 tsc_freq = regs[0] * 1000;
128 vmware_hvcall(VMW_HVCMD_GETHZ, regs);
129 if (regs[1] != UINT_MAX)
130 tsc_freq = regs[0] | ((uint64_t)regs[1] << 32);
132 tsc_is_invariant = 1;
136 * Calculate TSC frequency using information from the CPUID leaf 0x15
137 * 'Time Stamp Counter and Nominal Core Crystal Clock'. It should be
138 * an improvement over the parsing of the CPU model name in
139 * tsc_freq_intel(), when available.
148 do_cpuid(0x15, regs);
149 if (regs[0] == 0 || regs[1] == 0 || regs[2] == 0)
151 tsc_freq = (uint64_t)regs[2] * regs[1] / regs[0];
165 * Intel Processor Identification and the CPUID Instruction
166 * Application Note 485.
167 * http://www.intel.com/assets/pdf/appnote/241618.pdf
169 if (cpu_exthigh >= 0x80000004) {
171 for (i = 0x80000002; i < 0x80000005; i++) {
173 memcpy(p, regs, sizeof(regs));
177 for (i = 0; i < sizeof(brand) - 1; i++)
178 if (brand[i] == 'H' && brand[i + 1] == 'z')
195 #define C2D(c) ((c) - '0')
197 freq = C2D(p[0]) * 1000;
198 freq += C2D(p[2]) * 100;
199 freq += C2D(p[3]) * 10;
202 freq = C2D(p[0]) * 1000;
203 freq += C2D(p[1]) * 100;
204 freq += C2D(p[2]) * 10;
223 if ((regs[2] & CPUID_PERF_STAT) != 0) {
225 * XXX Some emulators expose host CPUID without actual
226 * support for these MSRs. We must test whether they
232 if (rdmsr(MSR_MPERF) > 0 && rdmsr(MSR_APERF) > 0)
237 if (vm_guest == VM_GUEST_VMWARE) {
242 switch (cpu_vendor_id) {
244 if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
245 (vm_guest == VM_GUEST_NO &&
246 CPUID_TO_FAMILY(cpu_id) >= 0x10))
247 tsc_is_invariant = 1;
248 if (cpu_feature & CPUID_SSE2) {
249 tsc_timecounter.tc_get_timecount =
250 tsc_get_timecount_mfence;
253 case CPU_VENDOR_INTEL:
254 if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
255 (vm_guest == VM_GUEST_NO &&
256 ((CPUID_TO_FAMILY(cpu_id) == 0x6 &&
257 CPUID_TO_MODEL(cpu_id) >= 0xe) ||
258 (CPUID_TO_FAMILY(cpu_id) == 0xf &&
259 CPUID_TO_MODEL(cpu_id) >= 0x3))))
260 tsc_is_invariant = 1;
261 if (cpu_feature & CPUID_SSE2) {
262 tsc_timecounter.tc_get_timecount =
263 tsc_get_timecount_lfence;
266 case CPU_VENDOR_CENTAUR:
267 if (vm_guest == VM_GUEST_NO &&
268 CPUID_TO_FAMILY(cpu_id) == 0x6 &&
269 CPUID_TO_MODEL(cpu_id) >= 0xf &&
270 (rdmsr(0x1203) & 0x100000000ULL) == 0)
271 tsc_is_invariant = 1;
272 if (cpu_feature & CPUID_SSE2) {
273 tsc_timecounter.tc_get_timecount =
274 tsc_get_timecount_lfence;
279 if (!TUNABLE_INT_FETCH("machdep.disable_tsc_calibration",
280 &tsc_skip_calibration)) {
282 * User did not give the order about calibration.
283 * If he did, we do not try to guess.
285 * Otherwise, if ACPI FADT reports that the platform
286 * is legacy-free and CPUID provides TSC frequency,
287 * use it. The calibration could fail anyway since
288 * ISA timer can be absent or power gated.
290 if (acpi_get_fadt_bootflags(&bootflags) &&
291 (bootflags & ACPI_FADT_LEGACY_DEVICES) == 0 &&
293 printf("Skipping TSC calibration since no legacy "
294 "devices reported by FADT and CPUID works\n");
295 tsc_skip_calibration = 1;
298 if (tsc_skip_calibration) {
299 if (tsc_freq_cpuid())
301 else if (cpu_vendor_id == CPU_VENDOR_INTEL)
305 printf("Calibrating TSC clock ... ");
309 tsc_freq = tsc2 - tsc1;
312 printf("TSC clock: %ju Hz\n", (intmax_t)tsc_freq);
319 if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
323 /* The TSC is known to be broken on certain CPUs. */
324 switch (cpu_vendor_id) {
326 switch (cpu_id & 0xFF0) {
332 case CPU_VENDOR_CENTAUR:
333 switch (cpu_id & 0xff0) {
336 * http://www.centtech.com/c6_data_sheet.pdf
338 * I-12 RDTSC may return incoherent values in EDX:EAX
339 * I-13 RDTSC hangs when certain event counters are used
345 switch (cpu_id & 0xff0) {
347 if ((cpu_id & CPUID_STEPPING) == 0)
358 * Inform CPU accounting about our boot-time clock rate. This will
359 * be updated if someone loads a cpufreq driver after boot that
360 * discovers a new max frequency.
363 set_cputicker(rdtsc, tsc_freq, !tsc_is_invariant);
365 if (tsc_is_invariant)
368 /* Register to find out about changes in CPU frequency. */
369 tsc_pre_tag = EVENTHANDLER_REGISTER(cpufreq_pre_change,
370 tsc_freq_changing, NULL, EVENTHANDLER_PRI_FIRST);
371 tsc_post_tag = EVENTHANDLER_REGISTER(cpufreq_post_change,
372 tsc_freq_changed, NULL, EVENTHANDLER_PRI_FIRST);
373 tsc_levels_tag = EVENTHANDLER_REGISTER(cpufreq_levels_changed,
374 tsc_levels_changed, NULL, EVENTHANDLER_PRI_ANY);
380 * RDTSC is not a serializing instruction, and does not drain
381 * instruction stream, so we need to drain the stream before executing
382 * it. It could be fixed by use of RDTSCP, except the instruction is
383 * not available everywhere.
385 * Use CPUID for draining in the boot-time SMP constistency test. The
386 * timecounters use MFENCE for AMD CPUs, and LFENCE for others (Intel
387 * and VIA) when SSE2 is present, and nothing on older machines which
388 * also do not issue RDTSC prematurely. There, testing for SSE2 and
389 * vendor is too cumbersome, and we learn about TSC presence from CPUID.
391 * Do not use do_cpuid(), since we do not need CPUID results, which
392 * have to be written into memory with do_cpuid().
394 #define TSC_READ(x) \
396 tsc_read_##x(void *arg) \
398 uint64_t *tsc = arg; \
399 u_int cpu = PCPU_GET(cpuid); \
401 __asm __volatile("cpuid" : : : "eax", "ebx", "ecx", "edx"); \
402 tsc[cpu * 3 + x] = rdtsc(); \
412 comp_smp_tsc(void *arg)
416 u_int cpu = PCPU_GET(cpuid);
419 size = (mp_maxid + 1) * 3;
420 for (i = 0, tsc = arg; i < N; i++, tsc += size)
424 d1 = tsc[cpu * 3 + 1] - tsc[j * 3];
425 d2 = tsc[cpu * 3 + 2] - tsc[j * 3 + 1];
426 if (d1 <= 0 || d2 <= 0) {
434 adj_smp_tsc(void *arg)
438 u_int cpu = PCPU_GET(cpuid);
439 u_int first, i, size;
446 size = (mp_maxid + 1) * 3;
447 for (i = 0, tsc = arg; i < N; i++, tsc += size) {
448 d = tsc[first * 3] - tsc[cpu * 3 + 1];
451 d = tsc[first * 3 + 1] - tsc[cpu * 3 + 2];
454 d = tsc[first * 3 + 1] - tsc[cpu * 3];
457 d = tsc[first * 3 + 2] - tsc[cpu * 3 + 1];
463 d = min / 2 + max / 2;
465 "movl $0x10, %%ecx\n\t"
467 "addl %%edi, %%eax\n\t"
468 "adcl %%esi, %%edx\n\t"
471 : "D" ((uint32_t)d), "S" ((uint32_t)(d >> 32))
472 : "ax", "cx", "dx", "cc"
477 test_tsc(int adj_max_count)
479 uint64_t *data, *tsc;
482 if ((!smp_tsc && !tsc_is_invariant) || vm_guest)
484 size = (mp_maxid + 1) * 3;
485 data = malloc(sizeof(*data) * size * N, M_TEMP, M_WAITOK);
488 for (i = 0, tsc = data; i < N; i++, tsc += size)
489 smp_rendezvous(tsc_read_0, tsc_read_1, tsc_read_2, tsc);
490 smp_tsc = 1; /* XXX */
491 smp_rendezvous(smp_no_rendezvous_barrier, comp_smp_tsc,
492 smp_no_rendezvous_barrier, data);
493 if (!smp_tsc && adj < adj_max_count) {
495 smp_rendezvous(smp_no_rendezvous_barrier, adj_smp_tsc,
496 smp_no_rendezvous_barrier, data);
501 printf("SMP: %sed TSC synchronization test%s\n",
502 smp_tsc ? "pass" : "fail",
503 adj > 0 ? " after adjustment" : "");
504 if (smp_tsc && tsc_is_invariant) {
505 switch (cpu_vendor_id) {
508 * Starting with Family 15h processors, TSC clock
509 * source is in the north bridge. Check whether
510 * we have a single-socket/multi-core platform.
511 * XXX Need more work for complex cases.
513 if (CPUID_TO_FAMILY(cpu_id) < 0x15 ||
514 (amd_feature2 & AMDID2_CMP) == 0 ||
515 smp_cpus > (cpu_procinfo2 & AMDID_CMP_CORES) + 1)
518 case CPU_VENDOR_INTEL:
520 * XXX Assume Intel platforms have synchronized TSCs.
539 if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
543 * Limit timecounter frequency to fit in an int and prevent it from
544 * overflowing too fast.
549 * We can not use the TSC if we support APM. Precise timekeeping
550 * on an APM'ed machine is at best a fools pursuit, since
551 * any and all of the time spent in various SMM code can't
552 * be reliably accounted for. Reading the RTC is your only
553 * source of reliable time info. The i8254 loses too, of course,
554 * but we need to have some kind of time...
555 * We don't know at this point whether APM is going to be used
556 * or not, nor when it might be activated. Play it safe.
558 if (power_pm_get_type() == POWER_PM_TYPE_APM) {
559 tsc_timecounter.tc_quality = -1000;
561 printf("TSC timecounter disabled: APM enabled.\n");
566 * Intel CPUs without a C-state invariant TSC can stop the TSC
567 * in either C2 or C3. Disable use of C2 and C3 while using
568 * the TSC as the timecounter. The timecounter can be changed
569 * to enable C2 and C3.
571 * Note that the TSC is used as the cputicker for computing
572 * thread runtime regardless of the timecounter setting, so
573 * using an alternate timecounter and enabling C2 or C3 can
574 * result incorrect runtimes for kernel idle threads (but not
575 * for any non-idle threads).
577 if (cpu_vendor_id == CPU_VENDOR_INTEL &&
578 (amd_pminfo & AMDPM_TSC_INVARIANT) == 0) {
579 tsc_timecounter.tc_flags |= TC_FLAGS_C2STOP;
581 printf("TSC timecounter disables C2 and C3.\n");
585 * We can not use the TSC in SMP mode unless the TSCs on all CPUs
586 * are synchronized. If the user is sure that the system has
587 * synchronized TSCs, set kern.timecounter.smp_tsc tunable to a
588 * non-zero value. The TSC seems unreliable in virtualized SMP
589 * environments, so it is set to a negative quality in those cases.
593 tsc_timecounter.tc_quality = test_tsc(smp_tsc_adjust);
596 if (tsc_is_invariant)
597 tsc_timecounter.tc_quality = 1000;
598 max_freq >>= tsc_shift;
601 for (shift = 0; shift <= 31 && (tsc_freq >> shift) > max_freq; shift++)
603 if ((cpu_feature & CPUID_SSE2) != 0 && mp_ncpus > 1) {
604 if (cpu_vendor_id == CPU_VENDOR_AMD) {
605 tsc_timecounter.tc_get_timecount = shift > 0 ?
606 tsc_get_timecount_low_mfence :
607 tsc_get_timecount_mfence;
609 tsc_timecounter.tc_get_timecount = shift > 0 ?
610 tsc_get_timecount_low_lfence :
611 tsc_get_timecount_lfence;
614 tsc_timecounter.tc_get_timecount = shift > 0 ?
615 tsc_get_timecount_low : tsc_get_timecount;
618 tsc_timecounter.tc_name = "TSC-low";
620 printf("TSC timecounter discards lower %d bit(s)\n",
624 tsc_timecounter.tc_frequency = tsc_freq >> shift;
625 tsc_timecounter.tc_priv = (void *)(intptr_t)shift;
626 tc_init(&tsc_timecounter);
629 SYSINIT(tsc_tc, SI_SUB_SMP, SI_ORDER_ANY, init_TSC_tc, NULL);
637 /* If TSC was not good on boot, it is unlikely to become good now. */
638 if (tsc_timecounter.tc_quality < 0)
640 /* Nothing to do with UP. */
645 * If TSC was good, a single synchronization should be enough,
646 * but honour smp_tsc_adjust if it's set.
648 quality = test_tsc(MAX(smp_tsc_adjust, 1));
649 if (quality != tsc_timecounter.tc_quality) {
650 printf("TSC timecounter quality changed: %d -> %d\n",
651 tsc_timecounter.tc_quality, quality);
652 tsc_timecounter.tc_quality = quality;
658 * When cpufreq levels change, find out about the (new) max frequency. We
659 * use this to update CPU accounting in case it got a lower estimate at boot.
662 tsc_levels_changed(void *arg, int unit)
665 struct cf_level *levels;
669 /* Only use values from the first CPU, assuming all are equal. */
673 /* Find the appropriate cpufreq device instance. */
674 cf_dev = devclass_get_device(devclass_find("cpufreq"), unit);
675 if (cf_dev == NULL) {
676 printf("tsc_levels_changed() called but no cpufreq device?\n");
680 /* Get settings from the device and find the max frequency. */
682 levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
685 error = CPUFREQ_LEVELS(cf_dev, levels, &count);
686 if (error == 0 && count != 0) {
687 max_freq = (uint64_t)levels[0].total_set.freq * 1000000;
688 set_cputicker(rdtsc, max_freq, 1);
690 printf("tsc_levels_changed: no max freq found\n");
691 free(levels, M_TEMP);
695 * If the TSC timecounter is in use, veto the pending change. It may be
696 * possible in the future to handle a dynamically-changing timecounter rate.
699 tsc_freq_changing(void *arg, const struct cf_level *level, int *status)
702 if (*status != 0 || timecounter != &tsc_timecounter)
705 printf("timecounter TSC must not be in use when "
706 "changing frequencies; change denied\n");
710 /* Update TSC freq with the value indicated by the caller. */
712 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
716 /* If there was an error during the transition, don't do anything. */
717 if (tsc_disabled || status != 0)
720 /* Total setting for this level gives the new frequency in MHz. */
721 freq = (uint64_t)level->total_set.freq * 1000000;
722 atomic_store_rel_64(&tsc_freq, freq);
723 tsc_timecounter.tc_frequency =
724 freq >> (int)(intptr_t)tsc_timecounter.tc_priv;
728 sysctl_machdep_tsc_freq(SYSCTL_HANDLER_ARGS)
733 freq = atomic_load_acq_64(&tsc_freq);
736 error = sysctl_handle_64(oidp, &freq, 0, req);
737 if (error == 0 && req->newptr != NULL) {
738 atomic_store_rel_64(&tsc_freq, freq);
739 atomic_store_rel_64(&tsc_timecounter.tc_frequency,
740 freq >> (int)(intptr_t)tsc_timecounter.tc_priv);
745 SYSCTL_PROC(_machdep, OID_AUTO, tsc_freq, CTLTYPE_U64 | CTLFLAG_RW,
746 0, 0, sysctl_machdep_tsc_freq, "QU", "Time Stamp Counter frequency");
749 tsc_get_timecount(struct timecounter *tc __unused)
756 tsc_get_timecount_low(struct timecounter *tc)
760 __asm __volatile("rdtsc; shrd %%cl, %%edx, %0"
761 : "=a" (rv) : "c" ((int)(intptr_t)tc->tc_priv) : "edx");
766 tsc_get_timecount_lfence(struct timecounter *tc __unused)
774 tsc_get_timecount_low_lfence(struct timecounter *tc)
778 return (tsc_get_timecount_low(tc));
782 tsc_get_timecount_mfence(struct timecounter *tc __unused)
790 tsc_get_timecount_low_mfence(struct timecounter *tc)
794 return (tsc_get_timecount_low(tc));
798 x86_tsc_vdso_timehands(struct vdso_timehands *vdso_th, struct timecounter *tc)
801 vdso_th->th_algo = VDSO_TH_ALGO_X86_TSC;
802 vdso_th->th_x86_shift = (int)(intptr_t)tc->tc_priv;
803 vdso_th->th_x86_hpet_idx = 0xffffffff;
804 bzero(vdso_th->th_res, sizeof(vdso_th->th_res));
808 #ifdef COMPAT_FREEBSD32
810 x86_tsc_vdso_timehands32(struct vdso_timehands32 *vdso_th32,
811 struct timecounter *tc)
814 vdso_th32->th_algo = VDSO_TH_ALGO_X86_TSC;
815 vdso_th32->th_x86_shift = (int)(intptr_t)tc->tc_priv;
816 vdso_th32->th_x86_hpet_idx = 0xffffffff;
817 bzero(vdso_th32->th_res, sizeof(vdso_th32->th_res));