1 /*
2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
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 *
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
34 * Copyright (c) 1997, 1998 Poul-Henning Kamp <phk@FreeBSD.org>
35 * Copyright (c) 1982, 1986, 1991, 1993
36 * The Regents of the University of California. All rights reserved.
37 * (c) UNIX System Laboratories, Inc.
38 * All or some portions of this file are derived from material licensed
39 * to the University of California by American Telephone and Telegraph
40 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
41 * the permission of UNIX System Laboratories, Inc.
42 *
43 * Redistribution and use in source and binary forms, with or without
44 * modification, are permitted provided that the following conditions
45 * are met:
46 * 1. Redistributions of source code must retain the above copyright
47 * notice, this list of conditions and the following disclaimer.
48 * 2. Redistributions in binary form must reproduce the above copyright
49 * notice, this list of conditions and the following disclaimer in the
50 * documentation and/or other materials provided with the distribution.
51 * 3. Neither the name of the University nor the names of its contributors
52 * may be used to endorse or promote products derived from this software
53 * without specific prior written permission.
54 *
55 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
56 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
57 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
58 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
59 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
60 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
61 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
62 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
63 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
64 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
65 * SUCH DAMAGE.
66 *
67 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94
68 * $FreeBSD: src/sys/kern/kern_clock.c,v 1.105.2.10 2002/10/17 13:19:40 maxim Exp $
69 */
70
71 #include "opt_ntp.h"
72 #include "opt_pctrack.h"
73
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/callout.h>
77 #include <sys/kernel.h>
78 #include <sys/kinfo.h>
79 #include <sys/proc.h>
80 #include <sys/malloc.h>
81 #include <sys/resource.h>
82 #include <sys/resourcevar.h>
83 #include <sys/signalvar.h>
84 #include <sys/caps.h>
85 #include <sys/timex.h>
86 #include <sys/timepps.h>
87 #include <sys/upmap.h>
88 #include <sys/lock.h>
89 #include <sys/sysctl.h>
90 #include <sys/kcollect.h>
91 #include <sys/exislock.h>
92 #include <sys/exislock2.h>
93
94 #include <vm/vm.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_extern.h>
98
99 #include <sys/thread2.h>
100 #include <sys/spinlock2.h>
101
102 #include <machine/cpu.h>
103 #include <machine/limits.h>
104 #include <machine/smp.h>
105 #include <machine/cpufunc.h>
106 #include <machine/specialreg.h>
107 #include <machine/clock.h>
108
109 #ifdef DEBUG_PCTRACK
110 static void do_pctrack(struct intrframe *frame, int which);
111 #endif
112
113 static void initclocks (void *dummy);
114 SYSINIT(clocks, SI_BOOT2_CLOCKS, SI_ORDER_FIRST, initclocks, NULL);
115
116 /*
117 * Some of these don't belong here, but it's easiest to concentrate them.
118 * Note that cpu_time counts in microseconds, but most userland programs
119 * just compare relative times against the total by delta.
120 */
121 struct kinfo_cputime cputime_percpu[MAXCPU];
122 #ifdef DEBUG_PCTRACK
123 struct kinfo_pcheader cputime_pcheader = { PCTRACK_SIZE, PCTRACK_ARYSIZE };
124 struct kinfo_pctrack cputime_pctrack[MAXCPU][PCTRACK_SIZE];
125 #endif
126
127 __read_mostly static int sniff_enable = 1;
128 __read_mostly static int sniff_target = -1;
129 __read_mostly static int clock_debug2 = 0;
130 SYSCTL_INT(_kern, OID_AUTO, sniff_enable, CTLFLAG_RW, &sniff_enable, 0 , "");
131 SYSCTL_INT(_kern, OID_AUTO, sniff_target, CTLFLAG_RW, &sniff_target, 0 , "");
132 SYSCTL_INT(_debug, OID_AUTO, clock_debug2, CTLFLAG_RW, &clock_debug2, 0 , "");
133
134 __read_mostly long pseudo_ticks = 1; /* existential timed locks */
135
136 static int
sysctl_cputime(SYSCTL_HANDLER_ARGS)137 sysctl_cputime(SYSCTL_HANDLER_ARGS)
138 {
139 int cpu, error = 0;
140 int root_error;
141 size_t size = sizeof(struct kinfo_cputime);
142 struct kinfo_cputime tmp;
143
144 /*
145 * NOTE: For security reasons, only root can sniff %rip
146 */
147 root_error = caps_priv_check_self(SYSCAP_RESTRICTEDROOT);
148
149 for (cpu = 0; cpu < ncpus; ++cpu) {
150 tmp = cputime_percpu[cpu];
151 if (root_error == 0) {
152 tmp.cp_sample_pc =
153 (int64_t)globaldata_find(cpu)->gd_sample_pc;
154 tmp.cp_sample_sp =
155 (int64_t)globaldata_find(cpu)->gd_sample_sp;
156 }
157 if ((error = SYSCTL_OUT(req, &tmp, size)) != 0)
158 break;
159 }
160
161 if (root_error == 0) {
162 if (sniff_enable) {
163 int n = sniff_target;
164 if (n < 0)
165 smp_sniff();
166 else if (n < ncpus)
167 cpu_sniff(n);
168 }
169 }
170
171 return (error);
172 }
173 SYSCTL_PROC(_kern, OID_AUTO, cputime, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
174 sysctl_cputime, "S,kinfo_cputime", "CPU time statistics");
175
176 static int
sysctl_cp_time(SYSCTL_HANDLER_ARGS)177 sysctl_cp_time(SYSCTL_HANDLER_ARGS)
178 {
179 long cpu_states[CPUSTATES] = {0};
180 int cpu, error = 0;
181 size_t size = sizeof(cpu_states);
182
183 for (cpu = 0; cpu < ncpus; ++cpu) {
184 cpu_states[CP_USER] += cputime_percpu[cpu].cp_user;
185 cpu_states[CP_NICE] += cputime_percpu[cpu].cp_nice;
186 cpu_states[CP_SYS] += cputime_percpu[cpu].cp_sys;
187 cpu_states[CP_INTR] += cputime_percpu[cpu].cp_intr;
188 cpu_states[CP_IDLE] += cputime_percpu[cpu].cp_idle;
189 }
190
191 error = SYSCTL_OUT(req, cpu_states, size);
192
193 return (error);
194 }
195
196 SYSCTL_PROC(_kern, OID_AUTO, cp_time, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
197 sysctl_cp_time, "LU", "CPU time statistics");
198
199 static int
sysctl_cp_times(SYSCTL_HANDLER_ARGS)200 sysctl_cp_times(SYSCTL_HANDLER_ARGS)
201 {
202 long cpu_states[CPUSTATES] = {0};
203 int cpu, error;
204 size_t size = sizeof(cpu_states);
205
206 for (error = 0, cpu = 0; error == 0 && cpu < ncpus; ++cpu) {
207 cpu_states[CP_USER] = cputime_percpu[cpu].cp_user;
208 cpu_states[CP_NICE] = cputime_percpu[cpu].cp_nice;
209 cpu_states[CP_SYS] = cputime_percpu[cpu].cp_sys;
210 cpu_states[CP_INTR] = cputime_percpu[cpu].cp_intr;
211 cpu_states[CP_IDLE] = cputime_percpu[cpu].cp_idle;
212 error = SYSCTL_OUT(req, cpu_states, size);
213 }
214
215 return (error);
216 }
217
218 SYSCTL_PROC(_kern, OID_AUTO, cp_times, (CTLTYPE_LONG|CTLFLAG_RD), 0, 0,
219 sysctl_cp_times, "LU", "per-CPU time statistics");
220
221 /*
222 * boottime is used to calculate the 'real' uptime. Do not confuse this with
223 * microuptime(). microtime() is not drift compensated. The real uptime
224 * with compensation is nanotime() - bootime. boottime is recalculated
225 * whenever the real time is set based on the compensated elapsed time
226 * in seconds (gd->gd_time_seconds).
227 *
228 * The gd_time_seconds and gd_cpuclock_base fields remain fairly monotonic.
229 * Slight adjustments to gd_cpuclock_base are made to phase-lock it to
230 * the real time.
231 *
232 * WARNING! time_second can backstep on time corrections. Also, unlike
233 * time_second, time_uptime is not a "real" time_t (seconds
234 * since the Epoch) but seconds since booting.
235 */
236 __read_mostly struct timespec boottime; /* boot time (realtime) for ref only */
237 __read_mostly struct timespec ticktime0;/* updated every tick */
238 __read_mostly struct timespec ticktime2;/* updated every tick */
239 __read_mostly int ticktime_update;
240 __read_mostly time_t time_second; /* read-only 'passive' rt in seconds */
241 __read_mostly time_t time_uptime; /* read-only 'passive' ut in seconds */
242
243 /*
244 * basetime is used to calculate the compensated real time of day. The
245 * basetime can be modified on a per-tick basis by the adjtime(),
246 * ntp_adjtime(), and sysctl-based time correction APIs.
247 *
248 * Note that frequency corrections can also be made by adjusting
249 * gd_cpuclock_base.
250 *
251 * basetime is a tail-chasing FIFO, updated only by cpu #0. The FIFO is
252 * used on both SMP and UP systems to avoid MP races between cpu's and
253 * interrupt races on UP systems.
254 */
255 struct hardtime {
256 __uint32_t time_second;
257 sysclock_t cpuclock_base;
258 };
259
260 #define BASETIME_ARYSIZE 16
261 #define BASETIME_ARYMASK (BASETIME_ARYSIZE - 1)
262 static struct timespec basetime[BASETIME_ARYSIZE];
263 static struct hardtime hardtime[BASETIME_ARYSIZE];
264 static volatile int basetime_index;
265
266 static int
sysctl_get_basetime(SYSCTL_HANDLER_ARGS)267 sysctl_get_basetime(SYSCTL_HANDLER_ARGS)
268 {
269 struct timespec *bt;
270 int error;
271 int index;
272
273 /*
274 * Because basetime data and index may be updated by another cpu,
275 * a load fence is required to ensure that the data we read has
276 * not been speculatively read relative to a possibly updated index.
277 */
278 index = basetime_index;
279 cpu_lfence();
280 bt = &basetime[index];
281 error = SYSCTL_OUT(req, bt, sizeof(*bt));
282 return (error);
283 }
284
285 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
286 &boottime, timespec, "System boottime");
287 SYSCTL_PROC(_kern, OID_AUTO, basetime, CTLTYPE_STRUCT|CTLFLAG_RD, 0, 0,
288 sysctl_get_basetime, "S,timespec", "System basetime");
289
290 static void hardclock(systimer_t info, int, struct intrframe *frame);
291 static void statclock(systimer_t info, int, struct intrframe *frame);
292 static void schedclock(systimer_t info, int, struct intrframe *frame);
293 static void getnanotime_nbt(struct timespec *nbt, struct timespec *tsp);
294
295 /*
296 * Use __read_mostly for ticks and sched_ticks because these variables are
297 * used all over the kernel and only updated once per tick.
298 */
299 __read_mostly sbintime_t sbticks; /* system master ticks at hz (64bit) */
300 __read_mostly int ticks; /* system master ticks at hz */
301 __read_mostly int sched_ticks; /* global schedule clock ticks */
302 __read_mostly int clocks_running; /* tsleep/timeout clocks operational */
303 int64_t nsec_adj; /* ntpd per-tick adjustment in nsec << 32 */
304 int64_t nsec_acc; /* accumulator */
305
306 /* NTPD time correction fields */
307 int64_t ntp_tick_permanent; /* per-tick adjustment in nsec << 32 */
308 int64_t ntp_tick_acc; /* accumulator for per-tick adjustment */
309 int64_t ntp_delta; /* one-time correction in nsec */
310 int64_t ntp_big_delta = 1000000000;
311 int32_t ntp_tick_delta; /* current adjustment rate */
312 int32_t ntp_default_tick_delta; /* adjustment rate for ntp_delta */
313 time_t ntp_leap_second; /* time of next leap second */
314 int ntp_leap_insert; /* whether to insert or remove a second */
315 struct spinlock ntp_spin;
316
317 /*
318 * Finish initializing clock frequencies and start all clocks running.
319 */
320 /* ARGSUSED*/
321 static void
initclocks(void * dummy)322 initclocks(void *dummy)
323 {
324 /*psratio = profhz / stathz;*/
325 spin_init(&ntp_spin, "ntp");
326 initclocks_pcpu();
327 clocks_running = 1;
328 if (kpmap) {
329 kpmap->tsc_freq = tsc_frequency;
330 kpmap->tick_freq = hz;
331 }
332 }
333
334 /*
335 * Called on a per-cpu basis from the idle thread bootstrap on each cpu
336 * during SMP initialization.
337 *
338 * This routine is called concurrently during low-level SMP initialization
339 * and may not block in any way. Meaning, among other things, we can't
340 * acquire any tokens.
341 */
342 void
initclocks_pcpu(void)343 initclocks_pcpu(void)
344 {
345 struct globaldata *gd = mycpu;
346
347 crit_enter();
348 if (gd->gd_cpuid == 0) {
349 gd->gd_time_seconds = 1;
350 gd->gd_cpuclock_base = sys_cputimer->count();
351 hardtime[0].time_second = gd->gd_time_seconds;
352 hardtime[0].cpuclock_base = gd->gd_cpuclock_base;
353 } else {
354 gd->gd_time_seconds = globaldata_find(0)->gd_time_seconds;
355 gd->gd_cpuclock_base = globaldata_find(0)->gd_cpuclock_base;
356 }
357
358 systimer_intr_enable();
359
360 crit_exit();
361 }
362
363 /*
364 * Called on a 10-second interval after the system is operational.
365 * Return the collection data for USERPCT and install the data for
366 * SYSTPCT and IDLEPCT.
367 */
368 static
369 uint64_t
collect_cputime_callback(int n)370 collect_cputime_callback(int n)
371 {
372 static long cpu_base[CPUSTATES];
373 long cpu_states[CPUSTATES];
374 long total;
375 long acc;
376 long lsb;
377
378 bzero(cpu_states, sizeof(cpu_states));
379 for (n = 0; n < ncpus; ++n) {
380 cpu_states[CP_USER] += cputime_percpu[n].cp_user;
381 cpu_states[CP_NICE] += cputime_percpu[n].cp_nice;
382 cpu_states[CP_SYS] += cputime_percpu[n].cp_sys;
383 cpu_states[CP_INTR] += cputime_percpu[n].cp_intr;
384 cpu_states[CP_IDLE] += cputime_percpu[n].cp_idle;
385 }
386
387 acc = 0;
388 for (n = 0; n < CPUSTATES; ++n) {
389 total = cpu_states[n] - cpu_base[n];
390 cpu_base[n] = cpu_states[n];
391 cpu_states[n] = total;
392 acc += total;
393 }
394 if (acc == 0) /* prevent degenerate divide by 0 */
395 acc = 1;
396 lsb = acc / (10000 * 2);
397 kcollect_setvalue(KCOLLECT_SYSTPCT,
398 (cpu_states[CP_SYS] + lsb) * 10000 / acc);
399 kcollect_setvalue(KCOLLECT_IDLEPCT,
400 (cpu_states[CP_IDLE] + lsb) * 10000 / acc);
401 kcollect_setvalue(KCOLLECT_INTRPCT,
402 (cpu_states[CP_INTR] + lsb) * 10000 / acc);
403 return((cpu_states[CP_USER] + cpu_states[CP_NICE] + lsb) * 10000 / acc);
404 }
405
406 /*
407 * This routine is called on just the BSP, just after SMP initialization
408 * completes to * finish initializing any clocks that might contend/block
409 * (e.g. like on a token). We can't do this in initclocks_pcpu() because
410 * that function is called from the idle thread bootstrap for each cpu and
411 * not allowed to block at all.
412 */
413 static
414 void
initclocks_other(void * dummy)415 initclocks_other(void *dummy)
416 {
417 struct globaldata *ogd = mycpu;
418 struct globaldata *gd;
419 int n;
420
421 for (n = 0; n < ncpus; ++n) {
422 lwkt_setcpu_self(globaldata_find(n));
423 gd = mycpu;
424
425 /*
426 * Use a non-queued periodic systimer to prevent multiple
427 * ticks from building up if the sysclock jumps forward
428 * (8254 gets reset). The sysclock will never jump backwards.
429 * Our time sync is based on the actual sysclock, not the
430 * ticks count.
431 *
432 * Install statclock before hardclock to prevent statclock
433 * from misinterpreting gd_flags for tick assignment when
434 * they overlap. Also offset the statclock by half of
435 * its interval to try to avoid being coincident with
436 * callouts.
437 */
438 systimer_init_periodic_flags(&gd->gd_statclock, statclock,
439 NULL, stathz,
440 SYSTF_MSSYNC | SYSTF_FIRST |
441 SYSTF_OFFSET50 | SYSTF_OFFSETCPU);
442 systimer_init_periodic_flags(&gd->gd_hardclock, hardclock,
443 NULL, hz,
444 SYSTF_MSSYNC | SYSTF_OFFSETCPU);
445 }
446 lwkt_setcpu_self(ogd);
447
448 /*
449 * Regular data collection
450 */
451 kcollect_register(KCOLLECT_USERPCT, "user", collect_cputime_callback,
452 KCOLLECT_SCALE(KCOLLECT_USERPCT_FORMAT, 0));
453 kcollect_register(KCOLLECT_SYSTPCT, "syst", NULL,
454 KCOLLECT_SCALE(KCOLLECT_SYSTPCT_FORMAT, 0));
455 kcollect_register(KCOLLECT_IDLEPCT, "idle", NULL,
456 KCOLLECT_SCALE(KCOLLECT_IDLEPCT_FORMAT, 0));
457 }
458 SYSINIT(clocks2, SI_BOOT2_POST_SMP, SI_ORDER_ANY, initclocks_other, NULL);
459
460 /*
461 * This method is called on just the BSP, after all the usched implementations
462 * are initialized. This avoids races between usched initialization functions
463 * and usched_schedulerclock().
464 */
465 static
466 void
initclocks_usched(void * dummy)467 initclocks_usched(void *dummy)
468 {
469 struct globaldata *ogd = mycpu;
470 struct globaldata *gd;
471 int n;
472
473 for (n = 0; n < ncpus; ++n) {
474 lwkt_setcpu_self(globaldata_find(n));
475 gd = mycpu;
476
477 /* XXX correct the frequency for scheduler / estcpu tests */
478 systimer_init_periodic_flags(&gd->gd_schedclock, schedclock,
479 NULL, ESTCPUFREQ,
480 SYSTF_MSSYNC | SYSTF_OFFSETCPU);
481 }
482 lwkt_setcpu_self(ogd);
483 }
484 SYSINIT(clocks3, SI_BOOT2_USCHED, SI_ORDER_ANY, initclocks_usched, NULL);
485
486 /*
487 * This sets the current real time of day. Timespecs are in seconds and
488 * nanoseconds. We do not mess with gd_time_seconds and gd_cpuclock_base,
489 * instead we adjust basetime so basetime + gd_* results in the current
490 * time of day. This way the gd_* fields are guaranteed to represent
491 * a monotonically increasing 'uptime' value.
492 *
493 * When set_timeofday() is called from userland, the system call forces it
494 * onto cpu #0 since only cpu #0 can update basetime_index.
495 */
496 void
set_timeofday(struct timespec * ts)497 set_timeofday(struct timespec *ts)
498 {
499 struct timespec *nbt;
500 int ni;
501
502 /*
503 * XXX SMP / non-atomic basetime updates
504 */
505 crit_enter();
506 ni = (basetime_index + 1) & BASETIME_ARYMASK;
507 cpu_lfence();
508 nbt = &basetime[ni];
509 nanouptime(nbt);
510 nbt->tv_sec = ts->tv_sec - nbt->tv_sec;
511 nbt->tv_nsec = ts->tv_nsec - nbt->tv_nsec;
512 if (nbt->tv_nsec < 0) {
513 nbt->tv_nsec += 1000000000;
514 --nbt->tv_sec;
515 }
516
517 /*
518 * Note that basetime diverges from boottime as the clock drift is
519 * compensated for, so we cannot do away with boottime. When setting
520 * the absolute time of day the drift is 0 (for an instant) and we
521 * can simply assign boottime to basetime.
522 *
523 * Note that nanouptime() is based on gd_time_seconds which is drift
524 * compensated up to a point (it is guaranteed to remain monotonically
525 * increasing). gd_time_seconds is thus our best uptime guess and
526 * suitable for use in the boottime calculation. It is already taken
527 * into account in the basetime calculation above.
528 */
529 spin_lock(&ntp_spin);
530 boottime.tv_sec = nbt->tv_sec;
531 ntp_delta = 0;
532
533 /*
534 * We now have a new basetime, make sure all other cpus have it,
535 * then update the index.
536 */
537 cpu_sfence();
538 basetime_index = ni;
539 spin_unlock(&ntp_spin);
540
541 crit_exit();
542 }
543
544 /*
545 * Each cpu has its own hardclock, but we only increment ticks and softticks
546 * on cpu #0.
547 *
548 * NOTE! systimer! the MP lock might not be held here. We can only safely
549 * manipulate objects owned by the current cpu.
550 */
551 static void
hardclock(systimer_t info,int in_ipi,struct intrframe * frame)552 hardclock(systimer_t info, int in_ipi, struct intrframe *frame)
553 {
554 sysclock_t cputicks;
555 struct proc *p;
556 struct globaldata *gd = mycpu;
557
558 if ((gd->gd_reqflags & RQF_IPIQ) == 0 && lwkt_need_ipiq_process(gd)) {
559 /* Defer to doreti on passive IPIQ processing */
560 need_ipiq();
561 }
562
563 /*
564 * We update the compensation base to calculate fine-grained time
565 * from the sys_cputimer on a per-cpu basis in order to avoid
566 * having to mess around with locks. sys_cputimer is assumed to
567 * be consistent across all cpus. CPU N copies the base state from
568 * CPU 0 using the same FIFO trick that we use for basetime (so we
569 * don't catch a CPU 0 update in the middle).
570 *
571 * Note that we never allow info->time (aka gd->gd_hardclock.time)
572 * to reverse index gd_cpuclock_base, but that it is possible for
573 * it to temporarily get behind in the seconds if something in the
574 * system locks interrupts for a long period of time. Since periodic
575 * timers count events, though everything should resynch again
576 * immediately.
577 */
578 if (gd->gd_cpuid == 0) {
579 int ni;
580
581 cputicks = info->time - gd->gd_cpuclock_base;
582 if (cputicks >= sys_cputimer->freq) {
583 cputicks /= sys_cputimer->freq;
584 if (cputicks != 0 && cputicks != 1)
585 kprintf("Warning: hardclock missed > 1 sec\n");
586 gd->gd_time_seconds += cputicks;
587 gd->gd_cpuclock_base += sys_cputimer->freq * cputicks;
588 /* uncorrected monotonic 1-sec gran */
589 time_uptime += cputicks;
590 }
591 ni = (basetime_index + 1) & BASETIME_ARYMASK;
592 hardtime[ni].time_second = gd->gd_time_seconds;
593 hardtime[ni].cpuclock_base = gd->gd_cpuclock_base;
594 } else {
595 int ni;
596
597 ni = basetime_index;
598 cpu_lfence();
599 gd->gd_time_seconds = hardtime[ni].time_second;
600 gd->gd_cpuclock_base = hardtime[ni].cpuclock_base;
601 }
602
603 /*
604 * The system-wide ticks counter and NTP related timedelta/tickdelta
605 * adjustments only occur on cpu #0. NTP adjustments are accomplished
606 * by updating basetime.
607 */
608 if (gd->gd_cpuid == 0) {
609 struct timespec *nbt;
610 struct timespec nts;
611 int leap;
612 int ni;
613
614 /*
615 * Update system-wide ticks
616 */
617 ++ticks;
618 ++sbticks;
619
620 /*
621 * Update system-wide ticktime for getnanotime() and getmicrotime()
622 */
623 nanotime(&nts);
624 atomic_add_int_nonlocked(&ticktime_update, 1);
625 cpu_sfence();
626 if (ticktime_update & 2)
627 ticktime2 = nts;
628 else
629 ticktime0 = nts;
630 cpu_sfence();
631 atomic_add_int_nonlocked(&ticktime_update, 1);
632
633 #if 0
634 if (tco->tc_poll_pps)
635 tco->tc_poll_pps(tco);
636 #endif
637
638 /*
639 * Calculate the new basetime index. We are in a critical section
640 * on cpu #0 and can safely play with basetime_index. Start
641 * with the current basetime and then make adjustments.
642 */
643 ni = (basetime_index + 1) & BASETIME_ARYMASK;
644 nbt = &basetime[ni];
645 *nbt = basetime[basetime_index];
646
647 /*
648 * ntp adjustments only occur on cpu 0 and are protected by
649 * ntp_spin. This spinlock virtually never conflicts.
650 */
651 spin_lock(&ntp_spin);
652
653 /*
654 * Apply adjtime corrections. (adjtime() API)
655 *
656 * adjtime() only runs on cpu #0 so our critical section is
657 * sufficient to access these variables.
658 */
659 if (ntp_delta != 0) {
660 nbt->tv_nsec += ntp_tick_delta;
661 ntp_delta -= ntp_tick_delta;
662 if ((ntp_delta > 0 && ntp_delta < ntp_tick_delta) ||
663 (ntp_delta < 0 && ntp_delta > ntp_tick_delta)) {
664 ntp_tick_delta = ntp_delta;
665 }
666 }
667
668 /*
669 * Apply permanent frequency corrections. (sysctl API)
670 */
671 if (ntp_tick_permanent != 0) {
672 ntp_tick_acc += ntp_tick_permanent;
673 if (ntp_tick_acc >= (1LL << 32)) {
674 nbt->tv_nsec += ntp_tick_acc >> 32;
675 ntp_tick_acc -= (ntp_tick_acc >> 32) << 32;
676 } else if (ntp_tick_acc <= -(1LL << 32)) {
677 /* Negate ntp_tick_acc to avoid shifting the sign bit. */
678 nbt->tv_nsec -= (-ntp_tick_acc) >> 32;
679 ntp_tick_acc += ((-ntp_tick_acc) >> 32) << 32;
680 }
681 }
682
683 if (nbt->tv_nsec >= 1000000000) {
684 nbt->tv_sec++;
685 nbt->tv_nsec -= 1000000000;
686 } else if (nbt->tv_nsec < 0) {
687 nbt->tv_sec--;
688 nbt->tv_nsec += 1000000000;
689 }
690
691 /*
692 * Another per-tick compensation. (for ntp_adjtime() API)
693 */
694 if (nsec_adj != 0) {
695 nsec_acc += nsec_adj;
696 if (nsec_acc >= 0x100000000LL) {
697 nbt->tv_nsec += nsec_acc >> 32;
698 nsec_acc = (nsec_acc & 0xFFFFFFFFLL);
699 } else if (nsec_acc <= -0x100000000LL) {
700 nbt->tv_nsec -= -nsec_acc >> 32;
701 nsec_acc = -(-nsec_acc & 0xFFFFFFFFLL);
702 }
703 if (nbt->tv_nsec >= 1000000000) {
704 nbt->tv_nsec -= 1000000000;
705 ++nbt->tv_sec;
706 } else if (nbt->tv_nsec < 0) {
707 nbt->tv_nsec += 1000000000;
708 --nbt->tv_sec;
709 }
710 }
711 spin_unlock(&ntp_spin);
712
713 /************************************************************
714 * LEAP SECOND CORRECTION *
715 ************************************************************
716 *
717 * Taking into account all the corrections made above, figure
718 * out the new real time. If the seconds field has changed
719 * then apply any pending leap-second corrections.
720 */
721 getnanotime_nbt(nbt, &nts);
722
723 if (time_second != nts.tv_sec) {
724 /*
725 * Apply leap second (sysctl API). Adjust nts for changes
726 * so we do not have to call getnanotime_nbt again.
727 */
728 if (ntp_leap_second) {
729 if (ntp_leap_second == nts.tv_sec) {
730 if (ntp_leap_insert) {
731 nbt->tv_sec++;
732 nts.tv_sec++;
733 } else {
734 nbt->tv_sec--;
735 nts.tv_sec--;
736 }
737 ntp_leap_second--;
738 }
739 }
740
741 /*
742 * Apply leap second (ntp_adjtime() API), calculate a new
743 * nsec_adj field. ntp_update_second() returns nsec_adj
744 * as a per-second value but we need it as a per-tick value.
745 */
746 leap = ntp_update_second(time_second, &nsec_adj);
747 nsec_adj /= hz;
748 nbt->tv_sec += leap;
749 nts.tv_sec += leap;
750
751 /*
752 * Update the time_second 'approximate time' global.
753 */
754 time_second = nts.tv_sec;
755
756 /*
757 * Clear the IPC hint for the currently running thread once
758 * per second, allowing us to disconnect the hint from a
759 * thread which may no longer care.
760 */
761 curthread->td_wakefromcpu = -1;
762 }
763
764 /*
765 * Finally, our new basetime is ready to go live!
766 */
767 cpu_sfence();
768 basetime_index = ni;
769
770 /*
771 * Update kpmap on each tick. TS updates are integrated with
772 * fences and upticks allowing userland to read the data
773 * deterministically.
774 */
775 if (kpmap) {
776 int w;
777
778 w = (kpmap->upticks + 1) & 1;
779 getnanouptime(&kpmap->ts_uptime[w]);
780 getnanotime(&kpmap->ts_realtime[w]);
781 cpu_sfence();
782 ++kpmap->upticks;
783 cpu_sfence();
784 }
785
786 /*
787 * Handle exislock pseudo_ticks. We make things as simple as
788 * possible for the critical path arming code by adding a little
789 * complication here.
790 *
791 * When we find that all cores have been armed, we increment
792 * pseudo_ticks and disarm all the cores.
793 */
794 {
795 globaldata_t gd;
796 int n;
797
798 for (n = 0; n < ncpus; ++n) {
799 gd = globaldata_find(n);
800 if (gd->gd_exisarmed == 0)
801 break;
802 }
803
804 if (n == ncpus) {
805 for (n = 0; n < ncpus; ++n) {
806 gd = globaldata_find(n);
807 gd->gd_exisarmed = 0;
808 }
809 ++pseudo_ticks;
810 }
811 }
812 }
813
814 /*
815 * lwkt thread scheduler fair queueing
816 */
817 lwkt_schedulerclock(curthread);
818
819 /*
820 * Cycle the existential lock system on odd ticks in order to re-arm
821 * our cpu (in case the cpu is idle or nobody is using any exis locks).
822 */
823 if (ticks & 1) {
824 exis_hold_gd(gd);
825 exis_drop_gd(gd);
826 }
827
828 /*
829 * softticks are handled for all cpus
830 */
831 hardclock_softtick(gd);
832
833 /*
834 * Rollup accumulated vmstats, copy-back for critical path checks.
835 */
836 vmstats_rollup_cpu(gd);
837 vfscache_rollup_cpu(gd);
838 mycpu->gd_vmstats = vmstats;
839
840 /*
841 * ITimer handling is per-tick, per-cpu.
842 *
843 * We must acquire the per-process token in order for ksignal()
844 * to be non-blocking. For the moment this requires an AST fault,
845 * the ksignal() cannot be safely issued from this hard interrupt.
846 *
847 * XXX Even the trytoken here isn't right, and itimer operation in
848 * a multi threaded environment is going to be weird at the
849 * very least.
850 */
851 if ((p = curproc) != NULL && lwkt_trytoken(&p->p_token)) {
852 crit_enter_hard();
853 if (p->p_upmap)
854 ++p->p_upmap->runticks;
855
856 if (frame && CLKF_USERMODE(frame) &&
857 timevalisset(&p->p_timer[ITIMER_VIRTUAL].it_value) &&
858 itimerdecr(&p->p_timer[ITIMER_VIRTUAL], ustick) == 0) {
859 p->p_flags |= P_SIGVTALRM;
860 need_user_resched();
861 }
862 if (timevalisset(&p->p_timer[ITIMER_PROF].it_value) &&
863 itimerdecr(&p->p_timer[ITIMER_PROF], ustick) == 0) {
864 p->p_flags |= P_SIGPROF;
865 need_user_resched();
866 }
867 crit_exit_hard();
868 lwkt_reltoken(&p->p_token);
869 }
870 setdelayed();
871 }
872
873 /*
874 * The statistics clock typically runs at a 125Hz rate, and is intended
875 * to be frequency offset from the hardclock (typ 100Hz). It is per-cpu.
876 *
877 * NOTE! systimer! the MP lock might not be held here. We can only safely
878 * manipulate objects owned by the current cpu.
879 *
880 * The stats clock is responsible for grabbing a profiling sample.
881 * Most of the statistics are only used by user-level statistics programs.
882 * The main exceptions are p->p_uticks, p->p_sticks, p->p_iticks, and
883 * p->p_estcpu.
884 *
885 * Like the other clocks, the stat clock is called from what is effectively
886 * a fast interrupt, so the context should be the thread/process that got
887 * interrupted.
888 */
889 static void
statclock(systimer_t info,int in_ipi,struct intrframe * frame)890 statclock(systimer_t info, int in_ipi, struct intrframe *frame)
891 {
892 globaldata_t gd = mycpu;
893 thread_t td;
894 struct proc *p;
895 int bump;
896 sysclock_t cv;
897 sysclock_t scv;
898
899 /*
900 * How big was our timeslice relative to the last time? Calculate
901 * in microseconds.
902 *
903 * NOTE: Use of microuptime() is typically MPSAFE, but usually not
904 * during early boot. Just use the systimer count to be nice
905 * to e.g. qemu. The systimer has a better chance of being
906 * MPSAFE at early boot.
907 */
908 cv = sys_cputimer->count();
909 scv = gd->statint.gd_statcv;
910 if (scv == 0) {
911 bump = 1;
912 } else {
913 bump = muldivu64(sys_cputimer->freq64_usec,
914 (cv - scv), 1L << 32);
915 if (bump < 0)
916 bump = 0;
917 if (bump > 1000000)
918 bump = 1000000;
919 }
920 gd->statint.gd_statcv = cv;
921
922 #if 0
923 stv = &gd->gd_stattv;
924 if (stv->tv_sec == 0) {
925 bump = 1;
926 } else {
927 bump = tv.tv_usec - stv->tv_usec +
928 (tv.tv_sec - stv->tv_sec) * 1000000;
929 if (bump < 0)
930 bump = 0;
931 if (bump > 1000000)
932 bump = 1000000;
933 }
934 *stv = tv;
935 #endif
936
937 td = curthread;
938 p = td->td_proc;
939
940 /*
941 * If this is an interrupt thread used for the clock interrupt, adjust
942 * td to the thread it is preempting. If a frame is available, it will
943 * be related to the thread being preempted.
944 */
945 if ((td->td_flags & TDF_CLKTHREAD) && td->td_preempted)
946 td = td->td_preempted;
947
948 if (frame && CLKF_USERMODE(frame)) {
949 /*
950 * Came from userland, handle user time and deal with
951 * possible process.
952 */
953 if (p && (p->p_flags & P_PROFIL))
954 addupc_intr(p, CLKF_PC(frame), 1);
955 td->td_uticks += bump;
956
957 /*
958 * Charge the time as appropriate
959 */
960 if (p && p->p_nice > NZERO)
961 cpu_time.cp_nice += bump;
962 else
963 cpu_time.cp_user += bump;
964 } else {
965 int intr_nest = gd->gd_intr_nesting_level;
966
967 if (in_ipi) {
968 /*
969 * IPI processing code will bump gd_intr_nesting_level
970 * up by one, which breaks following CLKF_INTR testing,
971 * so we subtract it by one here.
972 */
973 --intr_nest;
974 }
975
976 /*
977 * Came from kernel mode, so we were:
978 * - handling an interrupt,
979 * - doing syscall or trap work on behalf of the current
980 * user process, or
981 * - spinning in the idle loop.
982 * Whichever it is, charge the time as appropriate.
983 * Note that we charge interrupts to the current process,
984 * regardless of whether they are ``for'' that process,
985 * so that we know how much of its real time was spent
986 * in ``non-process'' (i.e., interrupt) work.
987 *
988 * XXX assume system if frame is NULL. A NULL frame
989 * can occur if ipi processing is done from a crit_exit().
990 */
991 if ((frame && CLKF_INTR(intr_nest)) ||
992 cpu_interrupt_running(td)) {
993 /*
994 * If we interrupted an interrupt thread, well,
995 * count it as interrupt time.
996 */
997 td->td_iticks += bump;
998 #ifdef DEBUG_PCTRACK
999 if (frame)
1000 do_pctrack(frame, PCTRACK_INT);
1001 #endif
1002 cpu_time.cp_intr += bump;
1003 } else if (gd->gd_flags & GDF_VIRTUSER) {
1004 /*
1005 * The vkernel doesn't do a good job providing trap
1006 * frames that we can test. If the GDF_VIRTUSER
1007 * flag is set we probably interrupted user mode.
1008 */
1009 td->td_uticks += bump;
1010
1011 /*
1012 * Charge the time as appropriate
1013 */
1014 if (p && p->p_nice > NZERO)
1015 cpu_time.cp_nice += bump;
1016 else
1017 cpu_time.cp_user += bump;
1018 } else {
1019 if (clock_debug2 > 0) {
1020 --clock_debug2;
1021 kprintf("statclock preempt %s (%p %p)\n", td->td_comm, td, &gd->gd_idlethread);
1022 }
1023 td->td_sticks += bump;
1024 if (td == &gd->gd_idlethread) {
1025 /*
1026 * We want to count token contention as
1027 * system time. When token contention occurs
1028 * the cpu may only be outside its critical
1029 * section while switching through the idle
1030 * thread. In this situation, various flags
1031 * will be set in gd_reqflags.
1032 *
1033 * INTPEND is not necessarily useful because
1034 * it will be set if the clock interrupt
1035 * happens to be on an interrupt thread, the
1036 * cpu_interrupt_running() call does a better
1037 * job so we've already handled it.
1038 */
1039 if (gd->gd_reqflags &
1040 (RQF_IDLECHECK_WK_MASK & ~RQF_INTPEND)) {
1041 cpu_time.cp_sys += bump;
1042 } else {
1043 cpu_time.cp_idle += bump;
1044 }
1045 } else {
1046 /*
1047 * System thread was running.
1048 */
1049 #ifdef DEBUG_PCTRACK
1050 if (frame)
1051 do_pctrack(frame, PCTRACK_SYS);
1052 #endif
1053 cpu_time.cp_sys += bump;
1054 }
1055 }
1056 }
1057 }
1058
1059 #ifdef DEBUG_PCTRACK
1060 /*
1061 * Sample the PC when in the kernel or in an interrupt. User code can
1062 * retrieve the information and generate a histogram or other output.
1063 */
1064
1065 static void
do_pctrack(struct intrframe * frame,int which)1066 do_pctrack(struct intrframe *frame, int which)
1067 {
1068 struct kinfo_pctrack *pctrack;
1069
1070 pctrack = &cputime_pctrack[mycpu->gd_cpuid][which];
1071 pctrack->pc_array[pctrack->pc_index & PCTRACK_ARYMASK] =
1072 (void *)CLKF_PC(frame);
1073 ++pctrack->pc_index;
1074 }
1075
1076 static int
sysctl_pctrack(SYSCTL_HANDLER_ARGS)1077 sysctl_pctrack(SYSCTL_HANDLER_ARGS)
1078 {
1079 struct kinfo_pcheader head;
1080 int error;
1081 int cpu;
1082 int ntrack;
1083
1084 head.pc_ntrack = PCTRACK_SIZE;
1085 head.pc_arysize = PCTRACK_ARYSIZE;
1086
1087 if ((error = SYSCTL_OUT(req, &head, sizeof(head))) != 0)
1088 return (error);
1089
1090 for (cpu = 0; cpu < ncpus; ++cpu) {
1091 for (ntrack = 0; ntrack < PCTRACK_SIZE; ++ntrack) {
1092 error = SYSCTL_OUT(req, &cputime_pctrack[cpu][ntrack],
1093 sizeof(struct kinfo_pctrack));
1094 if (error)
1095 break;
1096 }
1097 if (error)
1098 break;
1099 }
1100 return (error);
1101 }
1102 SYSCTL_PROC(_kern, OID_AUTO, pctrack, (CTLTYPE_OPAQUE|CTLFLAG_RD), 0, 0,
1103 sysctl_pctrack, "S,kinfo_pcheader", "CPU PC tracking");
1104
1105 #endif
1106
1107 /*
1108 * The scheduler clock typically runs at a 50Hz rate. NOTE! systimer,
1109 * the MP lock might not be held. We can safely manipulate parts of curproc
1110 * but that's about it.
1111 *
1112 * Each cpu has its own scheduler clock.
1113 */
1114 static void
schedclock(systimer_t info,int in_ipi __unused,struct intrframe * frame)1115 schedclock(systimer_t info, int in_ipi __unused, struct intrframe *frame)
1116 {
1117 struct lwp *lp;
1118 struct rusage *ru;
1119 struct vmspace *vm;
1120 long rss;
1121
1122 if ((lp = lwkt_preempted_proc()) != NULL) {
1123 /*
1124 * Account for cpu time used and hit the scheduler. Note
1125 * that this call MUST BE MP SAFE, and the BGL IS NOT HELD
1126 * HERE.
1127 */
1128 ++lp->lwp_cpticks;
1129 usched_schedulerclock(lp, info->periodic, info->time);
1130 } else {
1131 usched_schedulerclock(NULL, info->periodic, info->time);
1132 }
1133 if ((lp = curthread->td_lwp) != NULL) {
1134 /*
1135 * Update resource usage integrals and maximums.
1136 */
1137 if ((ru = &lp->lwp_proc->p_ru) &&
1138 (vm = lp->lwp_proc->p_vmspace) != NULL) {
1139 ru->ru_ixrss += pgtok(btoc(vm->vm_tsize));
1140 ru->ru_idrss += pgtok(btoc(vm->vm_dsize));
1141 ru->ru_isrss += pgtok(btoc(vm->vm_ssize));
1142 if (lwkt_trytoken(&vm->vm_map.token)) {
1143 rss = pgtok(vmspace_resident_count(vm));
1144 if (ru->ru_maxrss < rss)
1145 ru->ru_maxrss = rss;
1146 lwkt_reltoken(&vm->vm_map.token);
1147 }
1148 }
1149 }
1150 /* Increment the global sched_ticks */
1151 if (mycpu->gd_cpuid == 0)
1152 ++sched_ticks;
1153 }
1154
1155 /*
1156 * Compute number of ticks for the specified amount of time. The
1157 * return value is intended to be used in a clock interrupt timed
1158 * operation and guaranteed to meet or exceed the requested time.
1159 * If the representation overflows, return INT_MAX. The minimum return
1160 * value is 1 ticks and the function will average the calculation up.
1161 * If any value greater then 0 microseconds is supplied, a value
1162 * of at least 2 will be returned to ensure that a near-term clock
1163 * interrupt does not cause the timeout to occur (degenerately) early.
1164 *
1165 * Note that limit checks must take into account microseconds, which is
1166 * done simply by using the smaller signed long maximum instead of
1167 * the unsigned long maximum.
1168 *
1169 * If ints have 32 bits, then the maximum value for any timeout in
1170 * 10ms ticks is 248 days.
1171 */
1172 int
tvtohz_high(struct timeval * tv)1173 tvtohz_high(struct timeval *tv)
1174 {
1175 int ticks;
1176 long sec, usec;
1177
1178 sec = tv->tv_sec;
1179 usec = tv->tv_usec;
1180 if (usec < 0) {
1181 sec--;
1182 usec += 1000000;
1183 }
1184 if (sec < 0) {
1185 #ifdef DIAGNOSTIC
1186 if (usec > 0) {
1187 sec++;
1188 usec -= 1000000;
1189 }
1190 kprintf("tvtohz_high: negative time difference "
1191 "%ld sec %ld usec\n",
1192 sec, usec);
1193 #endif
1194 ticks = 1;
1195 } else if (sec <= INT_MAX / hz) {
1196 ticks = (int)(sec * hz + howmany((u_long)usec, ustick)) + 1;
1197 } else {
1198 ticks = INT_MAX;
1199 }
1200 return (ticks);
1201 }
1202
1203 int
tstohz_high(struct timespec * ts)1204 tstohz_high(struct timespec *ts)
1205 {
1206 int ticks;
1207 long sec, nsec;
1208
1209 sec = ts->tv_sec;
1210 nsec = ts->tv_nsec;
1211 if (nsec < 0) {
1212 sec--;
1213 nsec += 1000000000;
1214 }
1215 if (sec < 0) {
1216 #ifdef DIAGNOSTIC
1217 if (nsec > 0) {
1218 sec++;
1219 nsec -= 1000000000;
1220 }
1221 kprintf("tstohz_high: negative time difference "
1222 "%ld sec %ld nsec\n",
1223 sec, nsec);
1224 #endif
1225 ticks = 1;
1226 } else if (sec <= INT_MAX / hz) {
1227 ticks = (int)(sec * hz + howmany((u_long)nsec, nstick)) + 1;
1228 } else {
1229 ticks = INT_MAX;
1230 }
1231 return (ticks);
1232 }
1233
1234
1235 /*
1236 * Compute number of ticks for the specified amount of time, erroring on
1237 * the side of it being too low to ensure that sleeping the returned number
1238 * of ticks will not result in a late return.
1239 *
1240 * The supplied timeval may not be negative and should be normalized. A
1241 * return value of 0 is possible if the timeval converts to less then
1242 * 1 tick.
1243 *
1244 * If ints have 32 bits, then the maximum value for any timeout in
1245 * 10ms ticks is 248 days.
1246 */
1247 int
tvtohz_low(struct timeval * tv)1248 tvtohz_low(struct timeval *tv)
1249 {
1250 int ticks;
1251 long sec;
1252
1253 sec = tv->tv_sec;
1254 if (sec <= INT_MAX / hz)
1255 ticks = (int)(sec * hz + (u_long)tv->tv_usec / ustick);
1256 else
1257 ticks = INT_MAX;
1258 return (ticks);
1259 }
1260
1261 int
tstohz_low(struct timespec * ts)1262 tstohz_low(struct timespec *ts)
1263 {
1264 int ticks;
1265 long sec;
1266
1267 sec = ts->tv_sec;
1268 if (sec <= INT_MAX / hz)
1269 ticks = (int)(sec * hz + (u_long)ts->tv_nsec / nstick);
1270 else
1271 ticks = INT_MAX;
1272 return (ticks);
1273 }
1274
1275 /*
1276 * Start profiling on a process.
1277 *
1278 * Caller must hold p->p_token();
1279 *
1280 * Kernel profiling passes proc0 which never exits and hence
1281 * keeps the profile clock running constantly.
1282 */
1283 void
startprofclock(struct proc * p)1284 startprofclock(struct proc *p)
1285 {
1286 if ((p->p_flags & P_PROFIL) == 0) {
1287 p->p_flags |= P_PROFIL;
1288 #if 0 /* XXX */
1289 if (++profprocs == 1 && stathz != 0) {
1290 crit_enter();
1291 psdiv = psratio;
1292 setstatclockrate(profhz);
1293 crit_exit();
1294 }
1295 #endif
1296 }
1297 }
1298
1299 /*
1300 * Stop profiling on a process.
1301 *
1302 * caller must hold p->p_token
1303 */
1304 void
stopprofclock(struct proc * p)1305 stopprofclock(struct proc *p)
1306 {
1307 if (p->p_flags & P_PROFIL) {
1308 p->p_flags &= ~P_PROFIL;
1309 #if 0 /* XXX */
1310 if (--profprocs == 0 && stathz != 0) {
1311 crit_enter();
1312 psdiv = 1;
1313 setstatclockrate(stathz);
1314 crit_exit();
1315 }
1316 #endif
1317 }
1318 }
1319
1320 /*
1321 * Return information about system clocks.
1322 */
1323 static int
sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)1324 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
1325 {
1326 struct kinfo_clockinfo clkinfo;
1327 /*
1328 * Construct clockinfo structure.
1329 */
1330 clkinfo.ci_hz = hz;
1331 clkinfo.ci_tick = ustick;
1332 clkinfo.ci_tickadj = ntp_default_tick_delta / 1000;
1333 clkinfo.ci_profhz = profhz;
1334 clkinfo.ci_stathz = stathz ? stathz : hz;
1335 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
1336 }
1337
1338 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
1339 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
1340
1341 /*
1342 * We have eight functions for looking at the clock, four for
1343 * microseconds and four for nanoseconds. For each there is fast
1344 * but less precise version "get{nano|micro}[up]time" which will
1345 * return a time which is up to 1/HZ previous to the call, whereas
1346 * the raw version "{nano|micro}[up]time" will return a timestamp
1347 * which is as precise as possible. The "up" variants return the
1348 * time relative to system boot, these are well suited for time
1349 * interval measurements.
1350 *
1351 * Each cpu independently maintains the current time of day, so all
1352 * we need to do to protect ourselves from changes is to do a loop
1353 * check on the seconds field changing out from under us.
1354 *
1355 * The system timer maintains a 32 bit count and due to various issues
1356 * it is possible for the calculated delta to occasionally exceed
1357 * sys_cputimer->freq. If this occurs the sys_cputimer->freq64_nsec
1358 * multiplication can easily overflow, so we deal with the case. For
1359 * uniformity we deal with the case in the usec case too.
1360 *
1361 * All the [get][micro,nano][time,uptime]() routines are MPSAFE.
1362 *
1363 * NEW CODE (!)
1364 *
1365 * cpu 0 now maintains global ticktimes and an update counter. The
1366 * getnanotime() and getmicrotime() routines use these globals.
1367 */
1368 void
getmicrouptime(struct timeval * tvp)1369 getmicrouptime(struct timeval *tvp)
1370 {
1371 struct globaldata *gd = mycpu;
1372 sysclock_t delta;
1373
1374 do {
1375 tvp->tv_sec = gd->gd_time_seconds;
1376 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1377 } while (tvp->tv_sec != gd->gd_time_seconds);
1378
1379 if (delta >= sys_cputimer->freq) {
1380 tvp->tv_sec += delta / sys_cputimer->freq;
1381 delta %= sys_cputimer->freq;
1382 }
1383 tvp->tv_usec = muldivu64(sys_cputimer->freq64_usec, delta, 1L << 32);
1384 if (tvp->tv_usec >= 1000000) {
1385 tvp->tv_usec -= 1000000;
1386 ++tvp->tv_sec;
1387 }
1388 }
1389
1390 void
getnanouptime(struct timespec * tsp)1391 getnanouptime(struct timespec *tsp)
1392 {
1393 struct globaldata *gd = mycpu;
1394 sysclock_t delta;
1395
1396 do {
1397 tsp->tv_sec = gd->gd_time_seconds;
1398 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1399 } while (tsp->tv_sec != gd->gd_time_seconds);
1400
1401 if (delta >= sys_cputimer->freq) {
1402 tsp->tv_sec += delta / sys_cputimer->freq;
1403 delta %= sys_cputimer->freq;
1404 }
1405 tsp->tv_nsec = muldivu64(sys_cputimer->freq64_nsec, delta, 1L << 32);
1406 }
1407
1408 void
microuptime(struct timeval * tvp)1409 microuptime(struct timeval *tvp)
1410 {
1411 struct globaldata *gd = mycpu;
1412 sysclock_t delta;
1413
1414 do {
1415 tvp->tv_sec = gd->gd_time_seconds;
1416 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1417 } while (tvp->tv_sec != gd->gd_time_seconds);
1418
1419 if (delta >= sys_cputimer->freq) {
1420 tvp->tv_sec += delta / sys_cputimer->freq;
1421 delta %= sys_cputimer->freq;
1422 }
1423 tvp->tv_usec = muldivu64(sys_cputimer->freq64_usec, delta, 1L << 32);
1424 }
1425
1426 void
nanouptime(struct timespec * tsp)1427 nanouptime(struct timespec *tsp)
1428 {
1429 struct globaldata *gd = mycpu;
1430 sysclock_t delta;
1431
1432 do {
1433 tsp->tv_sec = gd->gd_time_seconds;
1434 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1435 } while (tsp->tv_sec != gd->gd_time_seconds);
1436
1437 if (delta >= sys_cputimer->freq) {
1438 tsp->tv_sec += delta / sys_cputimer->freq;
1439 delta %= sys_cputimer->freq;
1440 }
1441 tsp->tv_nsec = muldivu64(sys_cputimer->freq64_nsec, delta, 1L << 32);
1442 }
1443
1444 /*
1445 * realtime routines
1446 */
1447 void
getmicrotime(struct timeval * tvp)1448 getmicrotime(struct timeval *tvp)
1449 {
1450 struct timespec ts;
1451 int counter;
1452
1453 do {
1454 counter = *(volatile int *)&ticktime_update;
1455 cpu_lfence();
1456 switch(counter & 3) {
1457 case 0: /* ticktime2 completed update */
1458 ts = ticktime2;
1459 break;
1460 case 1: /* ticktime0 update in progress */
1461 ts = ticktime2;
1462 break;
1463 case 2: /* ticktime0 completed update */
1464 ts = ticktime0;
1465 break;
1466 case 3: /* ticktime2 update in progress */
1467 ts = ticktime0;
1468 break;
1469 }
1470 cpu_lfence();
1471 } while (counter != *(volatile int *)&ticktime_update);
1472 tvp->tv_sec = ts.tv_sec;
1473 tvp->tv_usec = ts.tv_nsec / 1000;
1474 }
1475
1476 void
getnanotime(struct timespec * tsp)1477 getnanotime(struct timespec *tsp)
1478 {
1479 struct timespec ts;
1480 int counter;
1481
1482 do {
1483 counter = *(volatile int *)&ticktime_update;
1484 cpu_lfence();
1485 switch(counter & 3) {
1486 case 0: /* ticktime2 completed update */
1487 ts = ticktime2;
1488 break;
1489 case 1: /* ticktime0 update in progress */
1490 ts = ticktime2;
1491 break;
1492 case 2: /* ticktime0 completed update */
1493 ts = ticktime0;
1494 break;
1495 case 3: /* ticktime2 update in progress */
1496 ts = ticktime0;
1497 break;
1498 }
1499 cpu_lfence();
1500 } while (counter != *(volatile int *)&ticktime_update);
1501 *tsp = ts;
1502 }
1503
1504 static void
getnanotime_nbt(struct timespec * nbt,struct timespec * tsp)1505 getnanotime_nbt(struct timespec *nbt, struct timespec *tsp)
1506 {
1507 struct globaldata *gd = mycpu;
1508 sysclock_t delta;
1509
1510 do {
1511 tsp->tv_sec = gd->gd_time_seconds;
1512 delta = gd->gd_hardclock.time - gd->gd_cpuclock_base;
1513 } while (tsp->tv_sec != gd->gd_time_seconds);
1514
1515 if (delta >= sys_cputimer->freq) {
1516 tsp->tv_sec += delta / sys_cputimer->freq;
1517 delta %= sys_cputimer->freq;
1518 }
1519 tsp->tv_nsec = muldivu64(sys_cputimer->freq64_nsec, delta, 1L << 32);
1520
1521 tsp->tv_sec += nbt->tv_sec;
1522 tsp->tv_nsec += nbt->tv_nsec;
1523 while (tsp->tv_nsec >= 1000000000) {
1524 tsp->tv_nsec -= 1000000000;
1525 ++tsp->tv_sec;
1526 }
1527 }
1528
1529
1530 void
microtime(struct timeval * tvp)1531 microtime(struct timeval *tvp)
1532 {
1533 struct globaldata *gd = mycpu;
1534 struct timespec *bt;
1535 sysclock_t delta;
1536
1537 do {
1538 tvp->tv_sec = gd->gd_time_seconds;
1539 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1540 } while (tvp->tv_sec != gd->gd_time_seconds);
1541
1542 if (delta >= sys_cputimer->freq) {
1543 tvp->tv_sec += delta / sys_cputimer->freq;
1544 delta %= sys_cputimer->freq;
1545 }
1546 tvp->tv_usec = muldivu64(sys_cputimer->freq64_usec, delta, 1L << 32);
1547
1548 bt = &basetime[basetime_index];
1549 cpu_lfence();
1550 tvp->tv_sec += bt->tv_sec;
1551 tvp->tv_usec += bt->tv_nsec / 1000;
1552 while (tvp->tv_usec >= 1000000) {
1553 tvp->tv_usec -= 1000000;
1554 ++tvp->tv_sec;
1555 }
1556 }
1557
1558 void
nanotime(struct timespec * tsp)1559 nanotime(struct timespec *tsp)
1560 {
1561 struct globaldata *gd = mycpu;
1562 struct timespec *bt;
1563 sysclock_t delta;
1564
1565 do {
1566 tsp->tv_sec = gd->gd_time_seconds;
1567 delta = sys_cputimer->count() - gd->gd_cpuclock_base;
1568 } while (tsp->tv_sec != gd->gd_time_seconds);
1569
1570 if (delta >= sys_cputimer->freq) {
1571 tsp->tv_sec += delta / sys_cputimer->freq;
1572 delta %= sys_cputimer->freq;
1573 }
1574 tsp->tv_nsec = muldivu64(sys_cputimer->freq64_nsec, delta, 1L << 32);
1575
1576 bt = &basetime[basetime_index];
1577 cpu_lfence();
1578 tsp->tv_sec += bt->tv_sec;
1579 tsp->tv_nsec += bt->tv_nsec;
1580 while (tsp->tv_nsec >= 1000000000) {
1581 tsp->tv_nsec -= 1000000000;
1582 ++tsp->tv_sec;
1583 }
1584 }
1585
1586 /*
1587 * Get an approximate time_t. It does not have to be accurate. This
1588 * function is called only from KTR and can be called with the system in
1589 * any state so do not use a critical section or other complex operation
1590 * here.
1591 *
1592 * NOTE: This is not exactly synchronized with real time. To do that we
1593 * would have to do what microtime does and check for a nanoseconds
1594 * overflow.
1595 */
1596 time_t
get_approximate_time_t(void)1597 get_approximate_time_t(void)
1598 {
1599 struct globaldata *gd = mycpu;
1600 struct timespec *bt;
1601
1602 bt = &basetime[basetime_index];
1603 return(gd->gd_time_seconds + bt->tv_sec);
1604 }
1605
1606 static int
pps_fetch_timeout(struct timespec * timeout,struct pps_state * pps)1607 pps_fetch_timeout(struct timespec *timeout, struct pps_state *pps)
1608 {
1609 int to, err;
1610 pps_seq_t *ap, *cp;
1611 pps_seq_t a, c;
1612
1613 to = INT_MAX;
1614 if (timeout->tv_sec > -1)
1615 to = tstohz_low(timeout);
1616
1617 ap = &pps->ppsinfo.assert_sequence;
1618 cp = &pps->ppsinfo.clear_sequence;
1619 a = atomic_load_acq_int(ap);
1620 c = atomic_load_acq_int(cp);
1621
1622 while (a == atomic_load_acq_int(ap) && c == atomic_load_acq_int(cp)) {
1623 err = tsleep(pps, PCATCH, "ppsfch", to);
1624 if (err == EWOULDBLOCK) {
1625 if (timeout->tv_sec < 0)
1626 continue;
1627 return (ETIMEDOUT);
1628 }
1629 if (err != 0)
1630 return (err);
1631 }
1632
1633 return (0);
1634 }
1635
1636 int
pps_ioctl(u_long cmd,caddr_t data,struct pps_state * pps)1637 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1638 {
1639 pps_params_t *app;
1640 struct pps_fetch_args *fapi;
1641 #ifdef PPS_SYNC
1642 struct pps_kcbind_args *kapi;
1643 #endif
1644 int err;
1645
1646 switch (cmd) {
1647 case PPS_IOC_CREATE:
1648 return (0);
1649 case PPS_IOC_DESTROY:
1650 return (0);
1651 case PPS_IOC_SETPARAMS:
1652 app = (pps_params_t *)data;
1653 if (app->mode & ~pps->ppscap)
1654 return (EINVAL);
1655 pps->ppsparam = *app;
1656 return (0);
1657 case PPS_IOC_GETPARAMS:
1658 app = (pps_params_t *)data;
1659 *app = pps->ppsparam;
1660 app->api_version = PPS_API_VERS_1;
1661 return (0);
1662 case PPS_IOC_GETCAP:
1663 *(int*)data = pps->ppscap;
1664 return (0);
1665 case PPS_IOC_FETCH:
1666 fapi = (struct pps_fetch_args *)data;
1667 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1668 return (EINVAL);
1669 if (fapi->timeout.tv_sec != 0 || fapi->timeout.tv_nsec != 0) {
1670 err = pps_fetch_timeout(&fapi->timeout, pps);
1671 if (err != 0)
1672 return (err);
1673 }
1674 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1675 fapi->pps_info_buf = pps->ppsinfo;
1676 return (0);
1677 case PPS_IOC_KCBIND:
1678 #ifdef PPS_SYNC
1679 kapi = (struct pps_kcbind_args *)data;
1680 /* XXX Only root should be able to do this */
1681 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1682 return (EINVAL);
1683 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1684 return (EINVAL);
1685 if (kapi->edge & ~pps->ppscap)
1686 return (EINVAL);
1687 pps->kcmode = kapi->edge;
1688 return (0);
1689 #else
1690 return (EOPNOTSUPP);
1691 #endif
1692 default:
1693 return (ENOTTY);
1694 }
1695 }
1696
1697 void
pps_init(struct pps_state * pps)1698 pps_init(struct pps_state *pps)
1699 {
1700 pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1701 if (pps->ppscap & PPS_CAPTUREASSERT)
1702 pps->ppscap |= PPS_OFFSETASSERT;
1703 if (pps->ppscap & PPS_CAPTURECLEAR)
1704 pps->ppscap |= PPS_OFFSETCLEAR;
1705 }
1706
1707 void
pps_event(struct pps_state * pps,sysclock_t count,int event)1708 pps_event(struct pps_state *pps, sysclock_t count, int event)
1709 {
1710 struct globaldata *gd;
1711 struct timespec *tsp;
1712 struct timespec *osp;
1713 struct timespec *bt;
1714 struct timespec ts;
1715 sysclock_t *pcount;
1716 #ifdef PPS_SYNC
1717 sysclock_t tcount;
1718 #endif
1719 sysclock_t delta;
1720 pps_seq_t *pseq;
1721 int foff;
1722 #ifdef PPS_SYNC
1723 int fhard;
1724 #endif
1725 int ni;
1726
1727 gd = mycpu;
1728
1729 /* Things would be easier with arrays... */
1730 if (event == PPS_CAPTUREASSERT) {
1731 tsp = &pps->ppsinfo.assert_timestamp;
1732 osp = &pps->ppsparam.assert_offset;
1733 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1734 #ifdef PPS_SYNC
1735 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1736 #endif
1737 pcount = &pps->ppscount[0];
1738 pseq = &pps->ppsinfo.assert_sequence;
1739 } else {
1740 tsp = &pps->ppsinfo.clear_timestamp;
1741 osp = &pps->ppsparam.clear_offset;
1742 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1743 #ifdef PPS_SYNC
1744 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1745 #endif
1746 pcount = &pps->ppscount[1];
1747 pseq = &pps->ppsinfo.clear_sequence;
1748 }
1749
1750 /* Nothing really happened */
1751 if (*pcount == count)
1752 return;
1753
1754 *pcount = count;
1755
1756 do {
1757 ts.tv_sec = gd->gd_time_seconds;
1758 delta = count - gd->gd_cpuclock_base;
1759 } while (ts.tv_sec != gd->gd_time_seconds);
1760
1761 if (delta >= sys_cputimer->freq) {
1762 ts.tv_sec += delta / sys_cputimer->freq;
1763 delta %= sys_cputimer->freq;
1764 }
1765 ts.tv_nsec = muldivu64(sys_cputimer->freq64_nsec, delta, 1L << 32);
1766 ni = basetime_index;
1767 cpu_lfence();
1768 bt = &basetime[ni];
1769 ts.tv_sec += bt->tv_sec;
1770 ts.tv_nsec += bt->tv_nsec;
1771 while (ts.tv_nsec >= 1000000000) {
1772 ts.tv_nsec -= 1000000000;
1773 ++ts.tv_sec;
1774 }
1775
1776 atomic_add_rel_int(pseq, 1);
1777 *tsp = ts;
1778
1779 if (foff) {
1780 timespecadd(tsp, osp, tsp);
1781 if (tsp->tv_nsec < 0) {
1782 tsp->tv_nsec += 1000000000;
1783 tsp->tv_sec -= 1;
1784 }
1785 }
1786 #ifdef PPS_SYNC
1787 if (fhard) {
1788 /* magic, at its best... */
1789 tcount = count - pps->ppscount[2];
1790 pps->ppscount[2] = count;
1791 if (tcount >= sys_cputimer->freq) {
1792 delta = (1000000000 * (tcount / sys_cputimer->freq) +
1793 sys_cputimer->freq64_nsec *
1794 (tcount % sys_cputimer->freq)) >> 32;
1795 } else {
1796 delta = muldivu64(sys_cputimer->freq64_nsec,
1797 tcount, 1L << 32);
1798 }
1799 hardpps(tsp, delta);
1800 }
1801 #endif
1802 wakeup(pps);
1803 }
1804
1805 /*
1806 * Return the tsc target value for a delay of (ns).
1807 *
1808 * Returns -1 if the TSC is not supported.
1809 */
1810 tsc_uclock_t
tsc_get_target(int ns)1811 tsc_get_target(int ns)
1812 {
1813 #if defined(_RDTSC_SUPPORTED_)
1814 if (cpu_feature & CPUID_TSC) {
1815 return (rdtsc() + tsc_frequency * ns / (int64_t)1000000000);
1816 }
1817 #endif
1818 return(-1);
1819 }
1820
1821 /*
1822 * Compare the tsc against the passed target
1823 *
1824 * Returns +1 if the target has been reached
1825 * Returns 0 if the target has not yet been reached
1826 * Returns -1 if the TSC is not supported.
1827 *
1828 * Typical use: while (tsc_test_target(target) == 0) { ...poll... }
1829 */
1830 int
tsc_test_target(int64_t target)1831 tsc_test_target(int64_t target)
1832 {
1833 #if defined(_RDTSC_SUPPORTED_)
1834 if (cpu_feature & CPUID_TSC) {
1835 if ((int64_t)(target - rdtsc()) <= 0)
1836 return(1);
1837 return(0);
1838 }
1839 #endif
1840 return(-1);
1841 }
1842
1843 /*
1844 * Delay the specified number of nanoseconds using the tsc. This function
1845 * returns immediately if the TSC is not supported. At least one cpu_pause()
1846 * will be issued.
1847 */
1848 void
tsc_delay(int ns)1849 tsc_delay(int ns)
1850 {
1851 int64_t clk;
1852
1853 clk = tsc_get_target(ns);
1854 cpu_pause();
1855 cpu_pause();
1856 while (tsc_test_target(clk) == 0) {
1857 cpu_pause();
1858 cpu_pause();
1859 cpu_pause();
1860 cpu_pause();
1861 }
1862 }
1863