2 * Copyright (c) 1998-2003 Poul-Henning Kamp
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
30 #include "opt_compat.h"
31 #include "opt_clock.h"
33 #include <sys/param.h>
36 #include <sys/limits.h>
37 #include <sys/malloc.h>
38 #include <sys/systm.h>
39 #include <sys/sysctl.h>
41 #include <sys/timetc.h>
42 #include <sys/kernel.h>
43 #include <sys/power.h>
46 #include <machine/clock.h>
47 #include <machine/cputypes.h>
48 #include <machine/md_var.h>
49 #include <machine/specialreg.h>
51 #include "cpufreq_if.h"
57 static eventhandler_tag tsc_levels_tag, tsc_pre_tag, tsc_post_tag;
59 SYSCTL_INT(_kern_timecounter, OID_AUTO, invariant_tsc, CTLFLAG_RDTUN,
60 &tsc_is_invariant, 0, "Indicates whether the TSC is P-state invariant");
61 TUNABLE_INT("kern.timecounter.invariant_tsc", &tsc_is_invariant);
65 SYSCTL_INT(_kern_timecounter, OID_AUTO, smp_tsc, CTLFLAG_RDTUN, &smp_tsc, 0,
66 "Indicates whether the TSC is safe to use in SMP mode");
67 TUNABLE_INT("kern.timecounter.smp_tsc", &smp_tsc);
70 static int tsc_disabled;
71 SYSCTL_INT(_machdep, OID_AUTO, disable_tsc, CTLFLAG_RDTUN, &tsc_disabled, 0,
72 "Disable x86 Time Stamp Counter");
73 TUNABLE_INT("machdep.disable_tsc", &tsc_disabled);
75 static int tsc_skip_calibration;
76 SYSCTL_INT(_machdep, OID_AUTO, disable_tsc_calibration, CTLFLAG_RDTUN,
77 &tsc_skip_calibration, 0, "Disable TSC frequency calibration");
78 TUNABLE_INT("machdep.disable_tsc_calibration", &tsc_skip_calibration);
80 static void tsc_freq_changed(void *arg, const struct cf_level *level,
82 static void tsc_freq_changing(void *arg, const struct cf_level *level,
84 static unsigned tsc_get_timecount(struct timecounter *tc);
85 static unsigned tsc_get_timecount_low(struct timecounter *tc);
86 static void tsc_levels_changed(void *arg, int unit);
88 static struct timecounter tsc_timecounter = {
89 tsc_get_timecount, /* get_timecount */
91 ~0u, /* counter_mask */
94 800, /* quality (adjusted in code) */
97 #define VMW_HVMAGIC 0x564d5868
98 #define VMW_HVPORT 0x5658
99 #define VMW_HVCMD_GETVERSION 10
100 #define VMW_HVCMD_GETHZ 45
103 vmware_hvcall(u_int cmd, u_int *p)
106 __asm __volatile("inl %w3, %0"
107 : "=a" (p[0]), "=b" (p[1]), "=c" (p[2]), "=d" (p[3])
108 : "0" (VMW_HVMAGIC), "1" (UINT_MAX), "2" (cmd), "3" (VMW_HVPORT)
113 tsc_freq_vmware(void)
122 * [RFC] CPUID usage for interaction between Hypervisors and Linux.
123 * http://lkml.org/lkml/2008/10/1/246
125 * KB1009458: Mechanisms to determine if software is running in
126 * a VMware virtual machine
127 * http://kb.vmware.com/kb/1009458
130 if ((cpu_feature2 & CPUID2_HV) != 0) {
131 do_cpuid(0x40000000, regs);
133 for (i = 1, p = hv_sig; i < 4; i++, p += sizeof(regs) / 4)
134 memcpy(p, ®s[i], sizeof(regs[i]));
138 * HV vendor ID string
139 * ------------+--------------
141 * Microsoft "Microsoft Hv"
142 * VMware "VMwareVMware"
145 printf("Hypervisor: Origin = \"%s\"\n", hv_sig);
147 if (strncmp(hv_sig, "VMwareVMware", 12) != 0)
150 p = getenv("smbios.system.serial");
153 if (strncmp(p, "VMware-", 7) != 0 &&
154 strncmp(p, "VMW", 3) != 0) {
159 vmware_hvcall(VMW_HVCMD_GETVERSION, regs);
160 if (regs[1] != VMW_HVMAGIC)
163 if (hv_high >= 0x40000010) {
164 do_cpuid(0x40000010, regs);
165 tsc_freq = regs[0] * 1000;
167 vmware_hvcall(VMW_HVCMD_GETHZ, regs);
168 if (regs[1] != UINT_MAX)
169 tsc_freq = regs[0] | ((uint64_t)regs[1] << 32);
171 tsc_is_invariant = 1;
185 * Intel Processor Identification and the CPUID Instruction
186 * Application Note 485.
187 * http://www.intel.com/assets/pdf/appnote/241618.pdf
189 if (cpu_exthigh >= 0x80000004) {
191 for (i = 0x80000002; i < 0x80000005; i++) {
193 memcpy(p, regs, sizeof(regs));
197 for (i = 0; i < sizeof(brand) - 1; i++)
198 if (brand[i] == 'H' && brand[i + 1] == 'z')
215 #define C2D(c) ((c) - '0')
217 freq = C2D(p[0]) * 1000;
218 freq += C2D(p[2]) * 100;
219 freq += C2D(p[3]) * 10;
222 freq = C2D(p[0]) * 1000;
223 freq += C2D(p[1]) * 100;
224 freq += C2D(p[2]) * 10;
242 if ((regs[2] & CPUID_PERF_STAT) != 0) {
244 * XXX Some emulators expose host CPUID without actual
245 * support for these MSRs. We must test whether they
251 if (rdmsr(MSR_MPERF) > 0 && rdmsr(MSR_APERF) > 0)
256 if (tsc_freq_vmware())
259 switch (cpu_vendor_id) {
261 if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
262 (vm_guest == VM_GUEST_NO &&
263 CPUID_TO_FAMILY(cpu_id) >= 0x10))
264 tsc_is_invariant = 1;
266 case CPU_VENDOR_INTEL:
267 if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 ||
268 (vm_guest == VM_GUEST_NO &&
269 ((CPUID_TO_FAMILY(cpu_id) == 0x6 &&
270 CPUID_TO_MODEL(cpu_id) >= 0xe) ||
271 (CPUID_TO_FAMILY(cpu_id) == 0xf &&
272 CPUID_TO_MODEL(cpu_id) >= 0x3))))
273 tsc_is_invariant = 1;
275 case CPU_VENDOR_CENTAUR:
276 if (vm_guest == VM_GUEST_NO &&
277 CPUID_TO_FAMILY(cpu_id) == 0x6 &&
278 CPUID_TO_MODEL(cpu_id) >= 0xf &&
279 (rdmsr(0x1203) & 0x100000000ULL) == 0)
280 tsc_is_invariant = 1;
284 if (tsc_skip_calibration) {
285 if (cpu_vendor_id == CPU_VENDOR_INTEL)
291 printf("Calibrating TSC clock ... ");
295 tsc_freq = tsc2 - tsc1;
297 printf("TSC clock: %ju Hz\n", (intmax_t)tsc_freq);
304 if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
310 * Inform CPU accounting about our boot-time clock rate. This will
311 * be updated if someone loads a cpufreq driver after boot that
312 * discovers a new max frequency.
315 set_cputicker(rdtsc, tsc_freq, !tsc_is_invariant);
317 if (tsc_is_invariant)
320 /* Register to find out about changes in CPU frequency. */
321 tsc_pre_tag = EVENTHANDLER_REGISTER(cpufreq_pre_change,
322 tsc_freq_changing, NULL, EVENTHANDLER_PRI_FIRST);
323 tsc_post_tag = EVENTHANDLER_REGISTER(cpufreq_post_change,
324 tsc_freq_changed, NULL, EVENTHANDLER_PRI_FIRST);
325 tsc_levels_tag = EVENTHANDLER_REGISTER(cpufreq_levels_changed,
326 tsc_levels_changed, NULL, EVENTHANDLER_PRI_ANY);
331 /* rmb is required here because rdtsc is not a serializing instruction. */
332 #define TSC_READ(x) \
334 tsc_read_##x(void *arg) \
336 uint32_t *tsc = arg; \
337 u_int cpu = PCPU_GET(cpuid); \
340 tsc[cpu * 3 + x] = rdtsc32(); \
350 comp_smp_tsc(void *arg)
354 u_int cpu = PCPU_GET(cpuid);
357 size = (mp_maxid + 1) * 3;
358 for (i = 0, tsc = arg; i < N; i++, tsc += size)
362 d1 = tsc[cpu * 3 + 1] - tsc[j * 3];
363 d2 = tsc[cpu * 3 + 2] - tsc[j * 3 + 1];
364 if (d1 <= 0 || d2 <= 0) {
374 uint32_t *data, *tsc;
377 if (!smp_tsc && !tsc_is_invariant)
379 size = (mp_maxid + 1) * 3;
380 data = malloc(sizeof(*data) * size * N, M_TEMP, M_WAITOK);
381 for (i = 0, tsc = data; i < N; i++, tsc += size)
382 smp_rendezvous(tsc_read_0, tsc_read_1, tsc_read_2, tsc);
383 smp_tsc = 1; /* XXX */
384 smp_rendezvous(smp_no_rendevous_barrier, comp_smp_tsc,
385 smp_no_rendevous_barrier, data);
388 printf("SMP: %sed TSC synchronization test\n",
389 smp_tsc ? "pass" : "fail");
390 if (smp_tsc && tsc_is_invariant) {
391 switch (cpu_vendor_id) {
394 * Starting with Family 15h processors, TSC clock
395 * source is in the north bridge. Check whether
396 * we have a single-socket/multi-core platform.
397 * XXX Need more work for complex cases.
399 if (CPUID_TO_FAMILY(cpu_id) < 0x15 ||
400 (amd_feature2 & AMDID2_CMP) == 0 ||
401 smp_cpus > (cpu_procinfo2 & AMDID_CMP_CORES) + 1)
404 case CPU_VENDOR_INTEL:
406 * XXX Assume Intel platforms have synchronized TSCs.
425 if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled)
429 * Limit timecounter frequency to fit in an int and prevent it from
430 * overflowing too fast.
435 * We can not use the TSC if we support APM. Precise timekeeping
436 * on an APM'ed machine is at best a fools pursuit, since
437 * any and all of the time spent in various SMM code can't
438 * be reliably accounted for. Reading the RTC is your only
439 * source of reliable time info. The i8254 loses too, of course,
440 * but we need to have some kind of time...
441 * We don't know at this point whether APM is going to be used
442 * or not, nor when it might be activated. Play it safe.
444 if (power_pm_get_type() == POWER_PM_TYPE_APM) {
445 tsc_timecounter.tc_quality = -1000;
447 printf("TSC timecounter disabled: APM enabled.\n");
452 * We cannot use the TSC if it stops incrementing in deep sleep.
453 * Currently only Intel CPUs are known for this problem unless
454 * the invariant TSC bit is set.
456 if (cpu_can_deep_sleep && cpu_vendor_id == CPU_VENDOR_INTEL &&
457 (amd_pminfo & AMDPM_TSC_INVARIANT) == 0) {
458 tsc_timecounter.tc_quality = -1000;
459 tsc_timecounter.tc_flags |= TC_FLAGS_C3STOP;
461 printf("TSC timecounter disabled: C3 enabled.\n");
467 * We can not use the TSC in SMP mode unless the TSCs on all CPUs are
468 * synchronized. If the user is sure that the system has synchronized
469 * TSCs, set kern.timecounter.smp_tsc tunable to a non-zero value.
470 * We also limit the frequency even lower to avoid "temporal anomalies"
471 * as much as possible. The TSC seems unreliable in virtualized SMP
472 * environments, so it is set to a negative quality in those cases.
476 tsc_timecounter.tc_quality = -100;
478 tsc_timecounter.tc_quality = test_smp_tsc();
483 if (tsc_is_invariant)
484 tsc_timecounter.tc_quality = 1000;
487 for (shift = 0; shift < 31 && (tsc_freq >> shift) > max_freq; shift++)
490 tsc_timecounter.tc_get_timecount = tsc_get_timecount_low;
491 tsc_timecounter.tc_name = "TSC-low";
493 printf("TSC timecounter discards lower %d bit(s)\n",
497 tsc_timecounter.tc_frequency = tsc_freq >> shift;
498 tsc_timecounter.tc_priv = (void *)(intptr_t)shift;
499 tc_init(&tsc_timecounter);
502 SYSINIT(tsc_tc, SI_SUB_SMP, SI_ORDER_ANY, init_TSC_tc, NULL);
505 * When cpufreq levels change, find out about the (new) max frequency. We
506 * use this to update CPU accounting in case it got a lower estimate at boot.
509 tsc_levels_changed(void *arg, int unit)
512 struct cf_level *levels;
516 /* Only use values from the first CPU, assuming all are equal. */
520 /* Find the appropriate cpufreq device instance. */
521 cf_dev = devclass_get_device(devclass_find("cpufreq"), unit);
522 if (cf_dev == NULL) {
523 printf("tsc_levels_changed() called but no cpufreq device?\n");
527 /* Get settings from the device and find the max frequency. */
529 levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
532 error = CPUFREQ_LEVELS(cf_dev, levels, &count);
533 if (error == 0 && count != 0) {
534 max_freq = (uint64_t)levels[0].total_set.freq * 1000000;
535 set_cputicker(rdtsc, max_freq, 1);
537 printf("tsc_levels_changed: no max freq found\n");
538 free(levels, M_TEMP);
542 * If the TSC timecounter is in use, veto the pending change. It may be
543 * possible in the future to handle a dynamically-changing timecounter rate.
546 tsc_freq_changing(void *arg, const struct cf_level *level, int *status)
549 if (*status != 0 || timecounter != &tsc_timecounter)
552 printf("timecounter TSC must not be in use when "
553 "changing frequencies; change denied\n");
557 /* Update TSC freq with the value indicated by the caller. */
559 tsc_freq_changed(void *arg, const struct cf_level *level, int status)
563 /* If there was an error during the transition, don't do anything. */
564 if (tsc_disabled || status != 0)
567 /* Total setting for this level gives the new frequency in MHz. */
568 freq = (uint64_t)level->total_set.freq * 1000000;
569 atomic_store_rel_64(&tsc_freq, freq);
570 tsc_timecounter.tc_frequency =
571 freq >> (int)(intptr_t)tsc_timecounter.tc_priv;
575 sysctl_machdep_tsc_freq(SYSCTL_HANDLER_ARGS)
580 freq = atomic_load_acq_64(&tsc_freq);
583 error = sysctl_handle_64(oidp, &freq, 0, req);
584 if (error == 0 && req->newptr != NULL) {
585 atomic_store_rel_64(&tsc_freq, freq);
586 atomic_store_rel_64(&tsc_timecounter.tc_frequency,
587 freq >> (int)(intptr_t)tsc_timecounter.tc_priv);
592 SYSCTL_PROC(_machdep, OID_AUTO, tsc_freq, CTLTYPE_U64 | CTLFLAG_RW,
593 0, 0, sysctl_machdep_tsc_freq, "QU", "Time Stamp Counter frequency");
596 tsc_get_timecount(struct timecounter *tc __unused)
603 tsc_get_timecount_low(struct timecounter *tc)
607 __asm __volatile("rdtsc; shrd %%cl, %%edx, %0"
608 : "=a" (rv) : "c" ((int)(intptr_t)tc->tc_priv) : "edx");
613 cpu_fill_vdso_timehands(struct vdso_timehands *vdso_th)
616 vdso_th->th_x86_shift = (int)(intptr_t)timecounter->tc_priv;
617 bzero(vdso_th->th_res, sizeof(vdso_th->th_res));
618 return (timecounter == &tsc_timecounter);
621 #ifdef COMPAT_FREEBSD32
623 cpu_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
626 vdso_th32->th_x86_shift = (int)(intptr_t)timecounter->tc_priv;
627 bzero(vdso_th32->th_res, sizeof(vdso_th32->th_res));
628 return (timecounter == &tsc_timecounter);