Print this page
patch clock-wakeup-remove
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/os/clock.c
+++ new/usr/src/uts/common/os/clock.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
22 22 /* All Rights Reserved */
23 23
24 24 /*
25 25 * Copyright (c) 1988, 2010, Oracle and/or its affiliates. All rights reserved.
26 26 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27 27 */
28 28
29 29 #include <sys/param.h>
30 30 #include <sys/t_lock.h>
31 31 #include <sys/types.h>
32 32 #include <sys/tuneable.h>
33 33 #include <sys/sysmacros.h>
34 34 #include <sys/systm.h>
35 35 #include <sys/cpuvar.h>
36 36 #include <sys/lgrp.h>
37 37 #include <sys/user.h>
38 38 #include <sys/proc.h>
39 39 #include <sys/callo.h>
40 40 #include <sys/kmem.h>
41 41 #include <sys/var.h>
42 42 #include <sys/cmn_err.h>
43 43 #include <sys/swap.h>
44 44 #include <sys/vmsystm.h>
45 45 #include <sys/class.h>
46 46 #include <sys/time.h>
47 47 #include <sys/debug.h>
48 48 #include <sys/vtrace.h>
49 49 #include <sys/spl.h>
50 50 #include <sys/atomic.h>
51 51 #include <sys/dumphdr.h>
52 52 #include <sys/archsystm.h>
53 53 #include <sys/fs/swapnode.h>
54 54 #include <sys/panic.h>
55 55 #include <sys/disp.h>
56 56 #include <sys/msacct.h>
57 57 #include <sys/mem_cage.h>
58 58
59 59 #include <vm/page.h>
60 60 #include <vm/anon.h>
61 61 #include <vm/rm.h>
62 62 #include <sys/cyclic.h>
63 63 #include <sys/cpupart.h>
64 64 #include <sys/rctl.h>
65 65 #include <sys/task.h>
66 66 #include <sys/sdt.h>
67 67 #include <sys/ddi_periodic.h>
68 68 #include <sys/random.h>
69 69 #include <sys/modctl.h>
70 70
71 71 /*
72 72 * for NTP support
73 73 */
74 74 #include <sys/timex.h>
75 75 #include <sys/inttypes.h>
76 76
77 77 #include <sys/sunddi.h>
78 78 #include <sys/clock_impl.h>
79 79
80 80 /*
81 81 * clock() is called straight from the clock cyclic; see clock_init().
82 82 *
83 83 * Functions:
84 84 * reprime clock
85 85 * maintain date
86 86 * jab the scheduler
87 87 */
88 88
89 89 extern kcondvar_t fsflush_cv;
90 90 extern sysinfo_t sysinfo;
91 91 extern vminfo_t vminfo;
92 92 extern int idleswtch; /* flag set while idle in pswtch() */
93 93 extern hrtime_t volatile devinfo_freeze;
94 94
95 95 /*
96 96 * high-precision avenrun values. These are needed to make the
97 97 * regular avenrun values accurate.
98 98 */
99 99 static uint64_t hp_avenrun[3];
100 100 int avenrun[3]; /* FSCALED average run queue lengths */
101 101 time_t time; /* time in seconds since 1970 - for compatibility only */
102 102
103 103 static struct loadavg_s loadavg;
104 104 /*
105 105 * Phase/frequency-lock loop (PLL/FLL) definitions
106 106 *
107 107 * The following variables are read and set by the ntp_adjtime() system
108 108 * call.
109 109 *
110 110 * time_state shows the state of the system clock, with values defined
111 111 * in the timex.h header file.
112 112 *
113 113 * time_status shows the status of the system clock, with bits defined
114 114 * in the timex.h header file.
115 115 *
116 116 * time_offset is used by the PLL/FLL to adjust the system time in small
117 117 * increments.
118 118 *
119 119 * time_constant determines the bandwidth or "stiffness" of the PLL.
120 120 *
121 121 * time_tolerance determines maximum frequency error or tolerance of the
122 122 * CPU clock oscillator and is a property of the architecture; however,
123 123 * in principle it could change as result of the presence of external
124 124 * discipline signals, for instance.
125 125 *
126 126 * time_precision is usually equal to the kernel tick variable; however,
127 127 * in cases where a precision clock counter or external clock is
128 128 * available, the resolution can be much less than this and depend on
129 129 * whether the external clock is working or not.
130 130 *
131 131 * time_maxerror is initialized by a ntp_adjtime() call and increased by
132 132 * the kernel once each second to reflect the maximum error bound
133 133 * growth.
134 134 *
135 135 * time_esterror is set and read by the ntp_adjtime() call, but
136 136 * otherwise not used by the kernel.
137 137 */
138 138 int32_t time_state = TIME_OK; /* clock state */
139 139 int32_t time_status = STA_UNSYNC; /* clock status bits */
140 140 int32_t time_offset = 0; /* time offset (us) */
141 141 int32_t time_constant = 0; /* pll time constant */
142 142 int32_t time_tolerance = MAXFREQ; /* frequency tolerance (scaled ppm) */
143 143 int32_t time_precision = 1; /* clock precision (us) */
144 144 int32_t time_maxerror = MAXPHASE; /* maximum error (us) */
145 145 int32_t time_esterror = MAXPHASE; /* estimated error (us) */
146 146
147 147 /*
148 148 * The following variables establish the state of the PLL/FLL and the
149 149 * residual time and frequency offset of the local clock. The scale
150 150 * factors are defined in the timex.h header file.
151 151 *
152 152 * time_phase and time_freq are the phase increment and the frequency
153 153 * increment, respectively, of the kernel time variable.
154 154 *
155 155 * time_freq is set via ntp_adjtime() from a value stored in a file when
156 156 * the synchronization daemon is first started. Its value is retrieved
157 157 * via ntp_adjtime() and written to the file about once per hour by the
158 158 * daemon.
159 159 *
160 160 * time_adj is the adjustment added to the value of tick at each timer
161 161 * interrupt and is recomputed from time_phase and time_freq at each
162 162 * seconds rollover.
163 163 *
164 164 * time_reftime is the second's portion of the system time at the last
165 165 * call to ntp_adjtime(). It is used to adjust the time_freq variable
166 166 * and to increase the time_maxerror as the time since last update
167 167 * increases.
168 168 */
169 169 int32_t time_phase = 0; /* phase offset (scaled us) */
170 170 int32_t time_freq = 0; /* frequency offset (scaled ppm) */
171 171 int32_t time_adj = 0; /* tick adjust (scaled 1 / hz) */
172 172 int32_t time_reftime = 0; /* time at last adjustment (s) */
173 173
174 174 /*
175 175 * The scale factors of the following variables are defined in the
176 176 * timex.h header file.
177 177 *
178 178 * pps_time contains the time at each calibration interval, as read by
179 179 * microtime(). pps_count counts the seconds of the calibration
180 180 * interval, the duration of which is nominally pps_shift in powers of
181 181 * two.
182 182 *
183 183 * pps_offset is the time offset produced by the time median filter
184 184 * pps_tf[], while pps_jitter is the dispersion (jitter) measured by
185 185 * this filter.
186 186 *
187 187 * pps_freq is the frequency offset produced by the frequency median
188 188 * filter pps_ff[], while pps_stabil is the dispersion (wander) measured
189 189 * by this filter.
190 190 *
191 191 * pps_usec is latched from a high resolution counter or external clock
192 192 * at pps_time. Here we want the hardware counter contents only, not the
193 193 * contents plus the time_tv.usec as usual.
194 194 *
195 195 * pps_valid counts the number of seconds since the last PPS update. It
196 196 * is used as a watchdog timer to disable the PPS discipline should the
197 197 * PPS signal be lost.
198 198 *
199 199 * pps_glitch counts the number of seconds since the beginning of an
200 200 * offset burst more than tick/2 from current nominal offset. It is used
201 201 * mainly to suppress error bursts due to priority conflicts between the
202 202 * PPS interrupt and timer interrupt.
203 203 *
204 204 * pps_intcnt counts the calibration intervals for use in the interval-
205 205 * adaptation algorithm. It's just too complicated for words.
206 206 */
207 207 struct timeval pps_time; /* kernel time at last interval */
208 208 int32_t pps_tf[] = {0, 0, 0}; /* pps time offset median filter (us) */
209 209 int32_t pps_offset = 0; /* pps time offset (us) */
210 210 int32_t pps_jitter = MAXTIME; /* time dispersion (jitter) (us) */
211 211 int32_t pps_ff[] = {0, 0, 0}; /* pps frequency offset median filter */
212 212 int32_t pps_freq = 0; /* frequency offset (scaled ppm) */
213 213 int32_t pps_stabil = MAXFREQ; /* frequency dispersion (scaled ppm) */
214 214 int32_t pps_usec = 0; /* microsec counter at last interval */
215 215 int32_t pps_valid = PPS_VALID; /* pps signal watchdog counter */
216 216 int32_t pps_glitch = 0; /* pps signal glitch counter */
217 217 int32_t pps_count = 0; /* calibration interval counter (s) */
218 218 int32_t pps_shift = PPS_SHIFT; /* interval duration (s) (shift) */
219 219 int32_t pps_intcnt = 0; /* intervals at current duration */
220 220
221 221 /*
222 222 * PPS signal quality monitors
223 223 *
224 224 * pps_jitcnt counts the seconds that have been discarded because the
225 225 * jitter measured by the time median filter exceeds the limit MAXTIME
226 226 * (100 us).
227 227 *
228 228 * pps_calcnt counts the frequency calibration intervals, which are
229 229 * variable from 4 s to 256 s.
230 230 *
231 231 * pps_errcnt counts the calibration intervals which have been discarded
232 232 * because the wander exceeds the limit MAXFREQ (100 ppm) or where the
233 233 * calibration interval jitter exceeds two ticks.
234 234 *
235 235 * pps_stbcnt counts the calibration intervals that have been discarded
236 236 * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us).
237 237 */
238 238 int32_t pps_jitcnt = 0; /* jitter limit exceeded */
239 239 int32_t pps_calcnt = 0; /* calibration intervals */
240 240 int32_t pps_errcnt = 0; /* calibration errors */
241 241 int32_t pps_stbcnt = 0; /* stability limit exceeded */
242 242
243 243 kcondvar_t lbolt_cv;
244 244
245 245 /*
246 246 * Hybrid lbolt implementation:
247 247 *
248 248 * The service historically provided by the lbolt and lbolt64 variables has
249 249 * been replaced by the ddi_get_lbolt() and ddi_get_lbolt64() routines, and the
250 250 * original symbols removed from the system. The once clock driven variables are
251 251 * now implemented in an event driven fashion, backed by gethrtime() coarsed to
252 252 * the appropriate clock resolution. The default event driven implementation is
253 253 * complemented by a cyclic driven one, active only during periods of intense
254 254 * activity around the DDI lbolt routines, when a lbolt specific cyclic is
255 255 * reprogramed to fire at a clock tick interval to serve consumers of lbolt who
256 256 * rely on the original low cost of consulting a memory position.
257 257 *
258 258 * The implementation uses the number of calls to these routines and the
259 259 * frequency of these to determine when to transition from event to cyclic
260 260 * driven and vice-versa. These values are kept on a per CPU basis for
261 261 * scalability reasons and to prevent CPUs from constantly invalidating a single
262 262 * cache line when modifying a global variable. The transition from event to
263 263 * cyclic mode happens once the thresholds are crossed, and activity on any CPU
264 264 * can cause such transition.
265 265 *
266 266 * The lbolt_hybrid function pointer is called by ddi_get_lbolt() and
267 267 * ddi_get_lbolt64(), and will point to lbolt_event_driven() or
268 268 * lbolt_cyclic_driven() according to the current mode. When the thresholds
269 269 * are exceeded, lbolt_event_driven() will reprogram the lbolt cyclic to
270 270 * fire at a nsec_per_tick interval and increment an internal variable at
271 271 * each firing. lbolt_hybrid will then point to lbolt_cyclic_driven(), which
272 272 * will simply return the value of such variable. lbolt_cyclic() will attempt
273 273 * to shut itself off at each threshold interval (sampling period for calls
274 274 * to the DDI lbolt routines), and return to the event driven mode, but will
275 275 * be prevented from doing so if lbolt_cyclic_driven() is being heavily used.
276 276 *
277 277 * lbolt_bootstrap is used during boot to serve lbolt consumers who don't wait
278 278 * for the cyclic subsystem to be intialized.
279 279 *
280 280 */
281 281 int64_t lbolt_bootstrap(void);
282 282 int64_t lbolt_event_driven(void);
283 283 int64_t lbolt_cyclic_driven(void);
284 284 int64_t (*lbolt_hybrid)(void) = lbolt_bootstrap;
285 285 uint_t lbolt_ev_to_cyclic(caddr_t, caddr_t);
286 286
287 287 /*
288 288 * lbolt's cyclic, installed by clock_init().
289 289 */
290 290 static void lbolt_cyclic(void);
291 291
292 292 /*
293 293 * Tunable to keep lbolt in cyclic driven mode. This will prevent the system
294 294 * from switching back to event driven, once it reaches cyclic mode.
295 295 */
296 296 static boolean_t lbolt_cyc_only = B_FALSE;
297 297
298 298 /*
299 299 * Cache aligned, per CPU structure with lbolt usage statistics.
300 300 */
301 301 static lbolt_cpu_t *lb_cpu;
302 302
303 303 /*
304 304 * Single, cache aligned, structure with all the information required by
305 305 * the lbolt implementation.
306 306 */
307 307 lbolt_info_t *lb_info;
308 308
309 309
310 310 int one_sec = 1; /* turned on once every second */
311 311 static int fsflushcnt; /* counter for t_fsflushr */
312 312 int dosynctodr = 1; /* patchable; enable/disable sync to TOD chip */
313 313 int tod_needsync = 0; /* need to sync tod chip with software time */
314 314 static int tod_broken = 0; /* clock chip doesn't work */
315 315 time_t boot_time = 0; /* Boot time in seconds since 1970 */
316 316 cyclic_id_t clock_cyclic; /* clock()'s cyclic_id */
317 317 cyclic_id_t deadman_cyclic; /* deadman()'s cyclic_id */
318 318
319 319 extern void clock_tick_schedule(int);
320 320
321 321 static int lgrp_ticks; /* counter to schedule lgrp load calcs */
322 322
323 323 /*
324 324 * for tod fault detection
325 325 */
326 326 #define TOD_REF_FREQ ((longlong_t)(NANOSEC))
327 327 #define TOD_STALL_THRESHOLD (TOD_REF_FREQ * 3 / 2)
328 328 #define TOD_JUMP_THRESHOLD (TOD_REF_FREQ / 2)
329 329 #define TOD_FILTER_N 4
330 330 #define TOD_FILTER_SETTLE (4 * TOD_FILTER_N)
331 331 static int tod_faulted = TOD_NOFAULT;
332 332
333 333 static int tod_status_flag = 0; /* used by tod_validate() */
334 334
335 335 static hrtime_t prev_set_tick = 0; /* gethrtime() prior to tod_set() */
336 336 static time_t prev_set_tod = 0; /* tv_sec value passed to tod_set() */
337 337
338 338 /* patchable via /etc/system */
339 339 int tod_validate_enable = 1;
340 340
341 341 /* Diagnose/Limit messages about delay(9F) called from interrupt context */
342 342 int delay_from_interrupt_diagnose = 0;
343 343 volatile uint32_t delay_from_interrupt_msg = 20;
344 344
345 345 /*
346 346 * On non-SPARC systems, TOD validation must be deferred until gethrtime
347 347 * returns non-zero values (after mach_clkinit's execution).
348 348 * On SPARC systems, it must be deferred until after hrtime_base
349 349 * and hres_last_tick are set (in the first invocation of hres_tick).
350 350 * Since in both cases the prerequisites occur before the invocation of
351 351 * tod_get() in clock(), the deferment is lifted there.
352 352 */
353 353 static boolean_t tod_validate_deferred = B_TRUE;
354 354
355 355 /*
356 356 * tod_fault_table[] must be aligned with
357 357 * enum tod_fault_type in systm.h
358 358 */
359 359 static char *tod_fault_table[] = {
360 360 "Reversed", /* TOD_REVERSED */
361 361 "Stalled", /* TOD_STALLED */
362 362 "Jumped", /* TOD_JUMPED */
363 363 "Changed in Clock Rate", /* TOD_RATECHANGED */
364 364 "Is Read-Only" /* TOD_RDONLY */
365 365 /*
366 366 * no strings needed for TOD_NOFAULT
367 367 */
368 368 };
369 369
370 370 /*
371 371 * test hook for tod broken detection in tod_validate
372 372 */
373 373 int tod_unit_test = 0;
374 374 time_t tod_test_injector;
375 375
376 376 #define CLOCK_ADJ_HIST_SIZE 4
377 377
378 378 static int adj_hist_entry;
379 379
380 380 int64_t clock_adj_hist[CLOCK_ADJ_HIST_SIZE];
381 381
382 382 static void calcloadavg(int, uint64_t *);
383 383 static int genloadavg(struct loadavg_s *);
384 384 static void loadavg_update();
385 385
386 386 void (*cmm_clock_callout)() = NULL;
387 387 void (*cpucaps_clock_callout)() = NULL;
388 388
389 389 extern clock_t clock_tick_proc_max;
390 390
391 391 static int64_t deadman_counter = 0;
392 392
393 393 static void
394 394 clock(void)
395 395 {
396 396 kthread_t *t;
397 397 uint_t nrunnable;
398 398 uint_t w_io;
399 399 cpu_t *cp;
400 400 cpupart_t *cpupart;
401 401 extern void set_freemem();
402 402 void (*funcp)();
403 403 int32_t ltemp;
404 404 int64_t lltemp;
405 405 int s;
406 406 int do_lgrp_load;
407 407 int i;
408 408 clock_t now = LBOLT_NO_ACCOUNT; /* current tick */
409 409
410 410 if (panicstr)
411 411 return;
412 412
413 413 /*
414 414 * Make sure that 'freemem' do not drift too far from the truth
415 415 */
416 416 set_freemem();
417 417
418 418
419 419 /*
420 420 * Before the section which is repeated is executed, we do
421 421 * the time delta processing which occurs every clock tick
422 422 *
423 423 * There is additional processing which happens every time
424 424 * the nanosecond counter rolls over which is described
425 425 * below - see the section which begins with : if (one_sec)
426 426 *
427 427 * This section marks the beginning of the precision-kernel
428 428 * code fragment.
429 429 *
430 430 * First, compute the phase adjustment. If the low-order bits
431 431 * (time_phase) of the update overflow, bump the higher order
432 432 * bits (time_update).
433 433 */
434 434 time_phase += time_adj;
435 435 if (time_phase <= -FINEUSEC) {
436 436 ltemp = -time_phase / SCALE_PHASE;
437 437 time_phase += ltemp * SCALE_PHASE;
438 438 s = hr_clock_lock();
439 439 timedelta -= ltemp * (NANOSEC/MICROSEC);
440 440 hr_clock_unlock(s);
441 441 } else if (time_phase >= FINEUSEC) {
442 442 ltemp = time_phase / SCALE_PHASE;
443 443 time_phase -= ltemp * SCALE_PHASE;
444 444 s = hr_clock_lock();
445 445 timedelta += ltemp * (NANOSEC/MICROSEC);
446 446 hr_clock_unlock(s);
447 447 }
448 448
449 449 /*
450 450 * End of precision-kernel code fragment which is processed
451 451 * every timer interrupt.
452 452 *
453 453 * Continue with the interrupt processing as scheduled.
454 454 */
455 455 /*
456 456 * Count the number of runnable threads and the number waiting
457 457 * for some form of I/O to complete -- gets added to
458 458 * sysinfo.waiting. To know the state of the system, must add
459 459 * wait counts from all CPUs. Also add up the per-partition
460 460 * statistics.
461 461 */
462 462 w_io = 0;
463 463 nrunnable = 0;
464 464
465 465 /*
466 466 * keep track of when to update lgrp/part loads
467 467 */
468 468
469 469 do_lgrp_load = 0;
470 470 if (lgrp_ticks++ >= hz / 10) {
471 471 lgrp_ticks = 0;
472 472 do_lgrp_load = 1;
473 473 }
474 474
475 475 if (one_sec) {
476 476 loadavg_update();
477 477 deadman_counter++;
478 478 }
479 479
480 480 /*
481 481 * First count the threads waiting on kpreempt queues in each
482 482 * CPU partition.
483 483 */
484 484
485 485 cpupart = cp_list_head;
486 486 do {
487 487 uint_t cpupart_nrunnable = cpupart->cp_kp_queue.disp_nrunnable;
488 488
489 489 cpupart->cp_updates++;
490 490 nrunnable += cpupart_nrunnable;
491 491 cpupart->cp_nrunnable_cum += cpupart_nrunnable;
492 492 if (one_sec) {
493 493 cpupart->cp_nrunning = 0;
494 494 cpupart->cp_nrunnable = cpupart_nrunnable;
495 495 }
496 496 } while ((cpupart = cpupart->cp_next) != cp_list_head);
497 497
498 498
499 499 /* Now count the per-CPU statistics. */
500 500 cp = cpu_list;
501 501 do {
502 502 uint_t cpu_nrunnable = cp->cpu_disp->disp_nrunnable;
503 503
504 504 nrunnable += cpu_nrunnable;
505 505 cpupart = cp->cpu_part;
506 506 cpupart->cp_nrunnable_cum += cpu_nrunnable;
507 507 if (one_sec) {
508 508 cpupart->cp_nrunnable += cpu_nrunnable;
509 509 /*
510 510 * Update user, system, and idle cpu times.
511 511 */
512 512 cpupart->cp_nrunning++;
513 513 /*
514 514 * w_io is used to update sysinfo.waiting during
515 515 * one_second processing below. Only gather w_io
516 516 * information when we walk the list of cpus if we're
517 517 * going to perform one_second processing.
518 518 */
519 519 w_io += CPU_STATS(cp, sys.iowait);
520 520 }
521 521
522 522 if (one_sec && (cp->cpu_flags & CPU_EXISTS)) {
523 523 int i, load, change;
524 524 hrtime_t intracct, intrused;
525 525 const hrtime_t maxnsec = 1000000000;
526 526 const int precision = 100;
527 527
528 528 /*
529 529 * Estimate interrupt load on this cpu each second.
530 530 * Computes cpu_intrload as %utilization (0-99).
531 531 */
532 532
533 533 /* add up interrupt time from all micro states */
534 534 for (intracct = 0, i = 0; i < NCMSTATES; i++)
535 535 intracct += cp->cpu_intracct[i];
536 536 scalehrtime(&intracct);
537 537
538 538 /* compute nsec used in the past second */
539 539 intrused = intracct - cp->cpu_intrlast;
540 540 cp->cpu_intrlast = intracct;
541 541
542 542 /* limit the value for safety (and the first pass) */
543 543 if (intrused >= maxnsec)
544 544 intrused = maxnsec - 1;
545 545
546 546 /* calculate %time in interrupt */
547 547 load = (precision * intrused) / maxnsec;
548 548 ASSERT(load >= 0 && load < precision);
549 549 change = cp->cpu_intrload - load;
550 550
551 551 /* jump to new max, or decay the old max */
552 552 if (change < 0)
553 553 cp->cpu_intrload = load;
554 554 else if (change > 0)
555 555 cp->cpu_intrload -= (change + 3) / 4;
556 556
557 557 DTRACE_PROBE3(cpu_intrload,
558 558 cpu_t *, cp,
559 559 hrtime_t, intracct,
560 560 hrtime_t, intrused);
561 561 }
562 562
563 563 if (do_lgrp_load &&
564 564 (cp->cpu_flags & CPU_EXISTS)) {
565 565 /*
566 566 * When updating the lgroup's load average,
567 567 * account for the thread running on the CPU.
568 568 * If the CPU is the current one, then we need
569 569 * to account for the underlying thread which
570 570 * got the clock interrupt not the thread that is
571 571 * handling the interrupt and caculating the load
572 572 * average
573 573 */
574 574 t = cp->cpu_thread;
575 575 if (CPU == cp)
576 576 t = t->t_intr;
577 577
578 578 /*
579 579 * Account for the load average for this thread if
580 580 * it isn't the idle thread or it is on the interrupt
581 581 * stack and not the current CPU handling the clock
582 582 * interrupt
583 583 */
584 584 if ((t && t != cp->cpu_idle_thread) || (CPU != cp &&
585 585 CPU_ON_INTR(cp))) {
586 586 if (t->t_lpl == cp->cpu_lpl) {
587 587 /* local thread */
588 588 cpu_nrunnable++;
589 589 } else {
590 590 /*
591 591 * This is a remote thread, charge it
592 592 * against its home lgroup. Note that
593 593 * we notice that a thread is remote
594 594 * only if it's currently executing.
595 595 * This is a reasonable approximation,
596 596 * since queued remote threads are rare.
597 597 * Note also that if we didn't charge
598 598 * it to its home lgroup, remote
599 599 * execution would often make a system
600 600 * appear balanced even though it was
601 601 * not, and thread placement/migration
602 602 * would often not be done correctly.
603 603 */
604 604 lgrp_loadavg(t->t_lpl,
605 605 LGRP_LOADAVG_IN_THREAD_MAX, 0);
606 606 }
607 607 }
608 608 lgrp_loadavg(cp->cpu_lpl,
609 609 cpu_nrunnable * LGRP_LOADAVG_IN_THREAD_MAX, 1);
610 610 }
611 611 } while ((cp = cp->cpu_next) != cpu_list);
612 612
613 613 clock_tick_schedule(one_sec);
614 614
615 615 /*
616 616 * Check for a callout that needs be called from the clock
617 617 * thread to support the membership protocol in a clustered
618 618 * system. Copy the function pointer so that we can reset
619 619 * this to NULL if needed.
620 620 */
621 621 if ((funcp = cmm_clock_callout) != NULL)
622 622 (*funcp)();
623 623
624 624 if ((funcp = cpucaps_clock_callout) != NULL)
625 625 (*funcp)();
626 626
627 627 /*
628 628 * Wakeup the cageout thread waiters once per second.
629 629 */
630 630 if (one_sec)
631 631 kcage_tick();
632 632
633 633 if (one_sec) {
634 634
635 635 int drift, absdrift;
636 636 timestruc_t tod;
637 637 int s;
638 638
639 639 /*
640 640 * Beginning of precision-kernel code fragment executed
641 641 * every second.
642 642 *
643 643 * On rollover of the second the phase adjustment to be
644 644 * used for the next second is calculated. Also, the
645 645 * maximum error is increased by the tolerance. If the
646 646 * PPS frequency discipline code is present, the phase is
647 647 * increased to compensate for the CPU clock oscillator
648 648 * frequency error.
649 649 *
650 650 * On a 32-bit machine and given parameters in the timex.h
651 651 * header file, the maximum phase adjustment is +-512 ms
652 652 * and maximum frequency offset is (a tad less than)
653 653 * +-512 ppm. On a 64-bit machine, you shouldn't need to ask.
654 654 */
655 655 time_maxerror += time_tolerance / SCALE_USEC;
656 656
657 657 /*
658 658 * Leap second processing. If in leap-insert state at
659 659 * the end of the day, the system clock is set back one
660 660 * second; if in leap-delete state, the system clock is
661 661 * set ahead one second. The microtime() routine or
662 662 * external clock driver will insure that reported time
663 663 * is always monotonic. The ugly divides should be
664 664 * replaced.
665 665 */
666 666 switch (time_state) {
667 667
668 668 case TIME_OK:
669 669 if (time_status & STA_INS)
670 670 time_state = TIME_INS;
671 671 else if (time_status & STA_DEL)
672 672 time_state = TIME_DEL;
673 673 break;
674 674
675 675 case TIME_INS:
676 676 if (hrestime.tv_sec % 86400 == 0) {
677 677 s = hr_clock_lock();
678 678 hrestime.tv_sec--;
679 679 hr_clock_unlock(s);
680 680 time_state = TIME_OOP;
681 681 }
682 682 break;
683 683
684 684 case TIME_DEL:
685 685 if ((hrestime.tv_sec + 1) % 86400 == 0) {
686 686 s = hr_clock_lock();
687 687 hrestime.tv_sec++;
688 688 hr_clock_unlock(s);
689 689 time_state = TIME_WAIT;
690 690 }
691 691 break;
692 692
693 693 case TIME_OOP:
694 694 time_state = TIME_WAIT;
695 695 break;
696 696
697 697 case TIME_WAIT:
698 698 if (!(time_status & (STA_INS | STA_DEL)))
699 699 time_state = TIME_OK;
700 700 default:
701 701 break;
702 702 }
703 703
704 704 /*
705 705 * Compute the phase adjustment for the next second. In
706 706 * PLL mode, the offset is reduced by a fixed factor
707 707 * times the time constant. In FLL mode the offset is
708 708 * used directly. In either mode, the maximum phase
709 709 * adjustment for each second is clamped so as to spread
710 710 * the adjustment over not more than the number of
711 711 * seconds between updates.
712 712 */
713 713 if (time_offset == 0)
714 714 time_adj = 0;
715 715 else if (time_offset < 0) {
716 716 lltemp = -time_offset;
717 717 if (!(time_status & STA_FLL)) {
718 718 if ((1 << time_constant) >= SCALE_KG)
719 719 lltemp *= (1 << time_constant) /
720 720 SCALE_KG;
721 721 else
722 722 lltemp = (lltemp / SCALE_KG) >>
723 723 time_constant;
724 724 }
725 725 if (lltemp > (MAXPHASE / MINSEC) * SCALE_UPDATE)
726 726 lltemp = (MAXPHASE / MINSEC) * SCALE_UPDATE;
727 727 time_offset += lltemp;
728 728 time_adj = -(lltemp * SCALE_PHASE) / hz / SCALE_UPDATE;
729 729 } else {
730 730 lltemp = time_offset;
731 731 if (!(time_status & STA_FLL)) {
732 732 if ((1 << time_constant) >= SCALE_KG)
733 733 lltemp *= (1 << time_constant) /
734 734 SCALE_KG;
735 735 else
736 736 lltemp = (lltemp / SCALE_KG) >>
737 737 time_constant;
738 738 }
739 739 if (lltemp > (MAXPHASE / MINSEC) * SCALE_UPDATE)
740 740 lltemp = (MAXPHASE / MINSEC) * SCALE_UPDATE;
741 741 time_offset -= lltemp;
742 742 time_adj = (lltemp * SCALE_PHASE) / hz / SCALE_UPDATE;
743 743 }
744 744
745 745 /*
746 746 * Compute the frequency estimate and additional phase
747 747 * adjustment due to frequency error for the next
748 748 * second. When the PPS signal is engaged, gnaw on the
749 749 * watchdog counter and update the frequency computed by
750 750 * the pll and the PPS signal.
751 751 */
752 752 pps_valid++;
753 753 if (pps_valid == PPS_VALID) {
754 754 pps_jitter = MAXTIME;
755 755 pps_stabil = MAXFREQ;
756 756 time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
757 757 STA_PPSWANDER | STA_PPSERROR);
758 758 }
759 759 lltemp = time_freq + pps_freq;
760 760
761 761 if (lltemp)
762 762 time_adj += (lltemp * SCALE_PHASE) / (SCALE_USEC * hz);
763 763
764 764 /*
765 765 * End of precision kernel-code fragment
766 766 *
767 767 * The section below should be modified if we are planning
768 768 * to use NTP for synchronization.
769 769 *
770 770 * Note: the clock synchronization code now assumes
771 771 * the following:
772 772 * - if dosynctodr is 1, then compute the drift between
773 773 * the tod chip and software time and adjust one or
774 774 * the other depending on the circumstances
775 775 *
776 776 * - if dosynctodr is 0, then the tod chip is independent
777 777 * of the software clock and should not be adjusted,
778 778 * but allowed to free run. this allows NTP to sync.
779 779 * hrestime without any interference from the tod chip.
780 780 */
781 781
782 782 tod_validate_deferred = B_FALSE;
783 783 mutex_enter(&tod_lock);
784 784 tod = tod_get();
785 785 drift = tod.tv_sec - hrestime.tv_sec;
786 786 absdrift = (drift >= 0) ? drift : -drift;
787 787 if (tod_needsync || absdrift > 1) {
788 788 int s;
789 789 if (absdrift > 2) {
790 790 if (!tod_broken && tod_faulted == TOD_NOFAULT) {
791 791 s = hr_clock_lock();
792 792 hrestime = tod;
793 793 membar_enter(); /* hrestime visible */
794 794 timedelta = 0;
795 795 timechanged++;
796 796 tod_needsync = 0;
797 797 hr_clock_unlock(s);
798 798 callout_hrestime();
799 799
800 800 }
801 801 } else {
802 802 if (tod_needsync || !dosynctodr) {
803 803 gethrestime(&tod);
804 804 tod_set(tod);
805 805 s = hr_clock_lock();
806 806 if (timedelta == 0)
807 807 tod_needsync = 0;
808 808 hr_clock_unlock(s);
809 809 } else {
810 810 /*
811 811 * If the drift is 2 seconds on the
812 812 * money, then the TOD is adjusting
813 813 * the clock; record that.
814 814 */
815 815 clock_adj_hist[adj_hist_entry++ %
816 816 CLOCK_ADJ_HIST_SIZE] = now;
817 817 s = hr_clock_lock();
818 818 timedelta = (int64_t)drift*NANOSEC;
819 819 hr_clock_unlock(s);
820 820 }
821 821 }
822 822 }
823 823 one_sec = 0;
824 824 time = gethrestime_sec(); /* for crusty old kmem readers */
825 825 mutex_exit(&tod_lock);
826 826
827 827 /*
828 828 * Some drivers still depend on this... XXX
829 829 */
830 830 cv_broadcast(&lbolt_cv);
831 831
832 832 vminfo.freemem += freemem;
833 833 {
834 834 pgcnt_t maxswap, resv, free;
835 835 pgcnt_t avail =
836 836 MAX((spgcnt_t)(availrmem - swapfs_minfree), 0);
837 837
838 838 maxswap = k_anoninfo.ani_mem_resv +
839 839 k_anoninfo.ani_max +avail;
840 840 /* Update ani_free */
841 841 set_anoninfo();
842 842 free = k_anoninfo.ani_free + avail;
843 843 resv = k_anoninfo.ani_phys_resv +
844 844 k_anoninfo.ani_mem_resv;
845 845
846 846 vminfo.swap_resv += resv;
847 847 /* number of reserved and allocated pages */
848 848 #ifdef DEBUG
849 849 if (maxswap < free)
850 850 cmn_err(CE_WARN, "clock: maxswap < free");
851 851 if (maxswap < resv)
852 852 cmn_err(CE_WARN, "clock: maxswap < resv");
853 853 #endif
854 854 vminfo.swap_alloc += maxswap - free;
855 855 vminfo.swap_avail += maxswap - resv;
856 856 vminfo.swap_free += free;
857 857 }
858 858 vminfo.updates++;
859 859 if (nrunnable) {
860 860 sysinfo.runque += nrunnable;
861 861 sysinfo.runocc++;
862 862 }
863 863 if (nswapped) {
864 864 sysinfo.swpque += nswapped;
865 865 sysinfo.swpocc++;
866 866 }
867 867 sysinfo.waiting += w_io;
868 868 sysinfo.updates++;
869 869
870 870 /*
871 871 * Wake up fsflush to write out DELWRI
872 872 * buffers, dirty pages and other cached
873 873 * administrative data, e.g. inodes.
874 874 */
875 875 if (--fsflushcnt <= 0) {
876 876 fsflushcnt = tune.t_fsflushr;
877 877 cv_signal(&fsflush_cv);
878 878 }
879 879
880 880 vmmeter();
881 881 calcloadavg(genloadavg(&loadavg), hp_avenrun);
882 882 for (i = 0; i < 3; i++)
883 883 /*
884 884 * At the moment avenrun[] can only hold 31
885 885 * bits of load average as it is a signed
886 886 * int in the API. We need to ensure that
887 887 * hp_avenrun[i] >> (16 - FSHIFT) will not be
888 888 * too large. If it is, we put the largest value
889 889 * that we can use into avenrun[i]. This is
890 890 * kludgey, but about all we can do until we
891 891 * avenrun[] is declared as an array of uint64[]
892 892 */
893 893 if (hp_avenrun[i] < ((uint64_t)1<<(31+16-FSHIFT)))
|
↓ open down ↓ |
893 lines elided |
↑ open up ↑ |
894 894 avenrun[i] = (int32_t)(hp_avenrun[i] >>
895 895 (16 - FSHIFT));
896 896 else
897 897 avenrun[i] = 0x7fffffff;
898 898
899 899 cpupart = cp_list_head;
900 900 do {
901 901 calcloadavg(genloadavg(&cpupart->cp_loadavg),
902 902 cpupart->cp_hp_avenrun);
903 903 } while ((cpupart = cpupart->cp_next) != cp_list_head);
904 -
905 - /*
906 - * Wake up the swapper thread if necessary.
907 - */
908 - if (runin ||
909 - (runout && (avefree < desfree || wake_sched_sec))) {
910 - t = &t0;
911 - thread_lock(t);
912 - if (t->t_state == TS_STOPPED) {
913 - runin = runout = 0;
914 - wake_sched_sec = 0;
915 - t->t_whystop = 0;
916 - t->t_whatstop = 0;
917 - t->t_schedflag &= ~TS_ALLSTART;
918 - THREAD_TRANSITION(t);
919 - setfrontdq(t);
920 - }
921 - thread_unlock(t);
922 - }
923 - }
924 -
925 - /*
926 - * Wake up the swapper if any high priority swapped-out threads
927 - * became runable during the last tick.
928 - */
929 - if (wake_sched) {
930 - t = &t0;
931 - thread_lock(t);
932 - if (t->t_state == TS_STOPPED) {
933 - runin = runout = 0;
934 - wake_sched = 0;
935 - t->t_whystop = 0;
936 - t->t_whatstop = 0;
937 - t->t_schedflag &= ~TS_ALLSTART;
938 - THREAD_TRANSITION(t);
939 - setfrontdq(t);
940 - }
941 - thread_unlock(t);
942 904 }
943 905 }
944 906
945 907 void
946 908 clock_init(void)
947 909 {
948 910 cyc_handler_t clk_hdlr, lbolt_hdlr;
949 911 cyc_time_t clk_when, lbolt_when;
950 912 int i, sz;
951 913 intptr_t buf;
952 914
953 915 /*
954 916 * Setup handler and timer for the clock cyclic.
955 917 */
956 918 clk_hdlr.cyh_func = (cyc_func_t)clock;
957 919 clk_hdlr.cyh_level = CY_LOCK_LEVEL;
958 920 clk_hdlr.cyh_arg = NULL;
959 921
960 922 clk_when.cyt_when = 0;
961 923 clk_when.cyt_interval = nsec_per_tick;
962 924
963 925 /*
964 926 * The lbolt cyclic will be reprogramed to fire at a nsec_per_tick
965 927 * interval to satisfy performance needs of the DDI lbolt consumers.
966 928 * It is off by default.
967 929 */
968 930 lbolt_hdlr.cyh_func = (cyc_func_t)lbolt_cyclic;
969 931 lbolt_hdlr.cyh_level = CY_LOCK_LEVEL;
970 932 lbolt_hdlr.cyh_arg = NULL;
971 933
972 934 lbolt_when.cyt_interval = nsec_per_tick;
973 935
974 936 /*
975 937 * Allocate cache line aligned space for the per CPU lbolt data and
976 938 * lbolt info structures, and initialize them with their default
977 939 * values. Note that these structures are also cache line sized.
978 940 */
979 941 sz = sizeof (lbolt_info_t) + CPU_CACHE_COHERENCE_SIZE;
980 942 buf = (intptr_t)kmem_zalloc(sz, KM_SLEEP);
981 943 lb_info = (lbolt_info_t *)P2ROUNDUP(buf, CPU_CACHE_COHERENCE_SIZE);
982 944
983 945 if (hz != HZ_DEFAULT)
984 946 lb_info->lbi_thresh_interval = LBOLT_THRESH_INTERVAL *
985 947 hz/HZ_DEFAULT;
986 948 else
987 949 lb_info->lbi_thresh_interval = LBOLT_THRESH_INTERVAL;
988 950
989 951 lb_info->lbi_thresh_calls = LBOLT_THRESH_CALLS;
990 952
991 953 sz = (sizeof (lbolt_cpu_t) * max_ncpus) + CPU_CACHE_COHERENCE_SIZE;
992 954 buf = (intptr_t)kmem_zalloc(sz, KM_SLEEP);
993 955 lb_cpu = (lbolt_cpu_t *)P2ROUNDUP(buf, CPU_CACHE_COHERENCE_SIZE);
994 956
995 957 for (i = 0; i < max_ncpus; i++)
996 958 lb_cpu[i].lbc_counter = lb_info->lbi_thresh_calls;
997 959
998 960 /*
999 961 * Install the softint used to switch between event and cyclic driven
1000 962 * lbolt. We use a soft interrupt to make sure the context of the
1001 963 * cyclic reprogram call is safe.
1002 964 */
1003 965 lbolt_softint_add();
1004 966
1005 967 /*
1006 968 * Since the hybrid lbolt implementation is based on a hardware counter
1007 969 * that is reset at every hardware reboot and that we'd like to have
1008 970 * the lbolt value starting at zero after both a hardware and a fast
1009 971 * reboot, we calculate the number of clock ticks the system's been up
1010 972 * and store it in the lbi_debug_time field of the lbolt info structure.
1011 973 * The value of this field will be subtracted from lbolt before
1012 974 * returning it.
1013 975 */
1014 976 lb_info->lbi_internal = lb_info->lbi_debug_time =
1015 977 (gethrtime()/nsec_per_tick);
1016 978
1017 979 /*
1018 980 * lbolt_hybrid points at lbolt_bootstrap until now. The LBOLT_* macros
1019 981 * and lbolt_debug_{enter,return} use this value as an indication that
1020 982 * the initializaion above hasn't been completed. Setting lbolt_hybrid
1021 983 * to either lbolt_{cyclic,event}_driven here signals those code paths
1022 984 * that the lbolt related structures can be used.
1023 985 */
1024 986 if (lbolt_cyc_only) {
1025 987 lbolt_when.cyt_when = 0;
1026 988 lbolt_hybrid = lbolt_cyclic_driven;
1027 989 } else {
1028 990 lbolt_when.cyt_when = CY_INFINITY;
1029 991 lbolt_hybrid = lbolt_event_driven;
1030 992 }
1031 993
1032 994 /*
1033 995 * Grab cpu_lock and install all three cyclics.
1034 996 */
1035 997 mutex_enter(&cpu_lock);
1036 998
1037 999 clock_cyclic = cyclic_add(&clk_hdlr, &clk_when);
1038 1000 lb_info->id.lbi_cyclic_id = cyclic_add(&lbolt_hdlr, &lbolt_when);
1039 1001
1040 1002 mutex_exit(&cpu_lock);
1041 1003 }
1042 1004
1043 1005 /*
1044 1006 * Called before calcloadavg to get 10-sec moving loadavg together
1045 1007 */
1046 1008
1047 1009 static int
1048 1010 genloadavg(struct loadavg_s *avgs)
1049 1011 {
1050 1012 int avg;
1051 1013 int spos; /* starting position */
1052 1014 int cpos; /* moving current position */
1053 1015 int i;
1054 1016 int slen;
1055 1017 hrtime_t hr_avg;
1056 1018
1057 1019 /* 10-second snapshot, calculate first positon */
1058 1020 if (avgs->lg_len == 0) {
1059 1021 return (0);
1060 1022 }
1061 1023 slen = avgs->lg_len < S_MOVAVG_SZ ? avgs->lg_len : S_MOVAVG_SZ;
1062 1024
1063 1025 spos = (avgs->lg_cur - 1) >= 0 ? avgs->lg_cur - 1 :
1064 1026 S_LOADAVG_SZ + (avgs->lg_cur - 1);
1065 1027 for (i = hr_avg = 0; i < slen; i++) {
1066 1028 cpos = (spos - i) >= 0 ? spos - i : S_LOADAVG_SZ + (spos - i);
1067 1029 hr_avg += avgs->lg_loads[cpos];
1068 1030 }
1069 1031
1070 1032 hr_avg = hr_avg / slen;
1071 1033 avg = hr_avg / (NANOSEC / LGRP_LOADAVG_IN_THREAD_MAX);
1072 1034
1073 1035 return (avg);
1074 1036 }
1075 1037
1076 1038 /*
1077 1039 * Run every second from clock () to update the loadavg count available to the
1078 1040 * system and cpu-partitions.
1079 1041 *
1080 1042 * This works by sampling the previous usr, sys, wait time elapsed,
1081 1043 * computing a delta, and adding that delta to the elapsed usr, sys,
1082 1044 * wait increase.
1083 1045 */
1084 1046
1085 1047 static void
1086 1048 loadavg_update()
1087 1049 {
1088 1050 cpu_t *cp;
1089 1051 cpupart_t *cpupart;
1090 1052 hrtime_t cpu_total;
1091 1053 int prev;
1092 1054
1093 1055 cp = cpu_list;
1094 1056 loadavg.lg_total = 0;
1095 1057
1096 1058 /*
1097 1059 * first pass totals up per-cpu statistics for system and cpu
1098 1060 * partitions
1099 1061 */
1100 1062
1101 1063 do {
1102 1064 struct loadavg_s *lavg;
1103 1065
1104 1066 lavg = &cp->cpu_loadavg;
1105 1067
1106 1068 cpu_total = cp->cpu_acct[CMS_USER] +
1107 1069 cp->cpu_acct[CMS_SYSTEM] + cp->cpu_waitrq;
1108 1070 /* compute delta against last total */
1109 1071 scalehrtime(&cpu_total);
1110 1072 prev = (lavg->lg_cur - 1) >= 0 ? lavg->lg_cur - 1 :
1111 1073 S_LOADAVG_SZ + (lavg->lg_cur - 1);
1112 1074 if (lavg->lg_loads[prev] <= 0) {
1113 1075 lavg->lg_loads[lavg->lg_cur] = cpu_total;
1114 1076 cpu_total = 0;
1115 1077 } else {
1116 1078 lavg->lg_loads[lavg->lg_cur] = cpu_total;
1117 1079 cpu_total = cpu_total - lavg->lg_loads[prev];
1118 1080 if (cpu_total < 0)
1119 1081 cpu_total = 0;
1120 1082 }
1121 1083
1122 1084 lavg->lg_cur = (lavg->lg_cur + 1) % S_LOADAVG_SZ;
1123 1085 lavg->lg_len = (lavg->lg_len + 1) < S_LOADAVG_SZ ?
1124 1086 lavg->lg_len + 1 : S_LOADAVG_SZ;
1125 1087
1126 1088 loadavg.lg_total += cpu_total;
1127 1089 cp->cpu_part->cp_loadavg.lg_total += cpu_total;
1128 1090
1129 1091 } while ((cp = cp->cpu_next) != cpu_list);
1130 1092
1131 1093 loadavg.lg_loads[loadavg.lg_cur] = loadavg.lg_total;
1132 1094 loadavg.lg_cur = (loadavg.lg_cur + 1) % S_LOADAVG_SZ;
1133 1095 loadavg.lg_len = (loadavg.lg_len + 1) < S_LOADAVG_SZ ?
1134 1096 loadavg.lg_len + 1 : S_LOADAVG_SZ;
1135 1097 /*
1136 1098 * Second pass updates counts
1137 1099 */
1138 1100 cpupart = cp_list_head;
1139 1101
1140 1102 do {
1141 1103 struct loadavg_s *lavg;
1142 1104
1143 1105 lavg = &cpupart->cp_loadavg;
1144 1106 lavg->lg_loads[lavg->lg_cur] = lavg->lg_total;
1145 1107 lavg->lg_total = 0;
1146 1108 lavg->lg_cur = (lavg->lg_cur + 1) % S_LOADAVG_SZ;
1147 1109 lavg->lg_len = (lavg->lg_len + 1) < S_LOADAVG_SZ ?
1148 1110 lavg->lg_len + 1 : S_LOADAVG_SZ;
1149 1111
1150 1112 } while ((cpupart = cpupart->cp_next) != cp_list_head);
1151 1113
1152 1114 }
1153 1115
1154 1116 /*
1155 1117 * clock_update() - local clock update
1156 1118 *
1157 1119 * This routine is called by ntp_adjtime() to update the local clock
1158 1120 * phase and frequency. The implementation is of an
1159 1121 * adaptive-parameter, hybrid phase/frequency-lock loop (PLL/FLL). The
1160 1122 * routine computes new time and frequency offset estimates for each
1161 1123 * call. The PPS signal itself determines the new time offset,
1162 1124 * instead of the calling argument. Presumably, calls to
1163 1125 * ntp_adjtime() occur only when the caller believes the local clock
1164 1126 * is valid within some bound (+-128 ms with NTP). If the caller's
1165 1127 * time is far different than the PPS time, an argument will ensue,
1166 1128 * and it's not clear who will lose.
1167 1129 *
1168 1130 * For uncompensated quartz crystal oscillatores and nominal update
1169 1131 * intervals less than 1024 s, operation should be in phase-lock mode
1170 1132 * (STA_FLL = 0), where the loop is disciplined to phase. For update
1171 1133 * intervals greater than this, operation should be in frequency-lock
1172 1134 * mode (STA_FLL = 1), where the loop is disciplined to frequency.
1173 1135 *
1174 1136 * Note: mutex(&tod_lock) is in effect.
1175 1137 */
1176 1138 void
1177 1139 clock_update(int offset)
1178 1140 {
1179 1141 int ltemp, mtemp, s;
1180 1142
1181 1143 ASSERT(MUTEX_HELD(&tod_lock));
1182 1144
1183 1145 if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
1184 1146 return;
1185 1147 ltemp = offset;
1186 1148 if ((time_status & STA_PPSTIME) && (time_status & STA_PPSSIGNAL))
1187 1149 ltemp = pps_offset;
1188 1150
1189 1151 /*
1190 1152 * Scale the phase adjustment and clamp to the operating range.
1191 1153 */
1192 1154 if (ltemp > MAXPHASE)
1193 1155 time_offset = MAXPHASE * SCALE_UPDATE;
1194 1156 else if (ltemp < -MAXPHASE)
1195 1157 time_offset = -(MAXPHASE * SCALE_UPDATE);
1196 1158 else
1197 1159 time_offset = ltemp * SCALE_UPDATE;
1198 1160
1199 1161 /*
1200 1162 * Select whether the frequency is to be controlled and in which
1201 1163 * mode (PLL or FLL). Clamp to the operating range. Ugly
1202 1164 * multiply/divide should be replaced someday.
1203 1165 */
1204 1166 if (time_status & STA_FREQHOLD || time_reftime == 0)
1205 1167 time_reftime = hrestime.tv_sec;
1206 1168
1207 1169 mtemp = hrestime.tv_sec - time_reftime;
1208 1170 time_reftime = hrestime.tv_sec;
1209 1171
1210 1172 if (time_status & STA_FLL) {
1211 1173 if (mtemp >= MINSEC) {
1212 1174 ltemp = ((time_offset / mtemp) * (SCALE_USEC /
1213 1175 SCALE_UPDATE));
1214 1176 if (ltemp)
1215 1177 time_freq += ltemp / SCALE_KH;
1216 1178 }
1217 1179 } else {
1218 1180 if (mtemp < MAXSEC) {
1219 1181 ltemp *= mtemp;
1220 1182 if (ltemp)
1221 1183 time_freq += (int)(((int64_t)ltemp *
1222 1184 SCALE_USEC) / SCALE_KF)
1223 1185 / (1 << (time_constant * 2));
1224 1186 }
1225 1187 }
1226 1188 if (time_freq > time_tolerance)
1227 1189 time_freq = time_tolerance;
1228 1190 else if (time_freq < -time_tolerance)
1229 1191 time_freq = -time_tolerance;
1230 1192
1231 1193 s = hr_clock_lock();
1232 1194 tod_needsync = 1;
1233 1195 hr_clock_unlock(s);
1234 1196 }
1235 1197
1236 1198 /*
1237 1199 * ddi_hardpps() - discipline CPU clock oscillator to external PPS signal
1238 1200 *
1239 1201 * This routine is called at each PPS interrupt in order to discipline
1240 1202 * the CPU clock oscillator to the PPS signal. It measures the PPS phase
1241 1203 * and leaves it in a handy spot for the clock() routine. It
1242 1204 * integrates successive PPS phase differences and calculates the
1243 1205 * frequency offset. This is used in clock() to discipline the CPU
1244 1206 * clock oscillator so that intrinsic frequency error is cancelled out.
1245 1207 * The code requires the caller to capture the time and hardware counter
1246 1208 * value at the on-time PPS signal transition.
1247 1209 *
1248 1210 * Note that, on some Unix systems, this routine runs at an interrupt
1249 1211 * priority level higher than the timer interrupt routine clock().
1250 1212 * Therefore, the variables used are distinct from the clock()
1251 1213 * variables, except for certain exceptions: The PPS frequency pps_freq
1252 1214 * and phase pps_offset variables are determined by this routine and
1253 1215 * updated atomically. The time_tolerance variable can be considered a
1254 1216 * constant, since it is infrequently changed, and then only when the
1255 1217 * PPS signal is disabled. The watchdog counter pps_valid is updated
1256 1218 * once per second by clock() and is atomically cleared in this
1257 1219 * routine.
1258 1220 *
1259 1221 * tvp is the time of the last tick; usec is a microsecond count since the
1260 1222 * last tick.
1261 1223 *
1262 1224 * Note: In Solaris systems, the tick value is actually given by
1263 1225 * usec_per_tick. This is called from the serial driver cdintr(),
1264 1226 * or equivalent, at a high PIL. Because the kernel keeps a
1265 1227 * highresolution time, the following code can accept either
1266 1228 * the traditional argument pair, or the current highres timestamp
1267 1229 * in tvp and zero in usec.
1268 1230 */
1269 1231 void
1270 1232 ddi_hardpps(struct timeval *tvp, int usec)
1271 1233 {
1272 1234 int u_usec, v_usec, bigtick;
1273 1235 time_t cal_sec;
1274 1236 int cal_usec;
1275 1237
1276 1238 /*
1277 1239 * An occasional glitch can be produced when the PPS interrupt
1278 1240 * occurs in the clock() routine before the time variable is
1279 1241 * updated. Here the offset is discarded when the difference
1280 1242 * between it and the last one is greater than tick/2, but not
1281 1243 * if the interval since the first discard exceeds 30 s.
1282 1244 */
1283 1245 time_status |= STA_PPSSIGNAL;
1284 1246 time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
1285 1247 pps_valid = 0;
1286 1248 u_usec = -tvp->tv_usec;
1287 1249 if (u_usec < -(MICROSEC/2))
1288 1250 u_usec += MICROSEC;
1289 1251 v_usec = pps_offset - u_usec;
1290 1252 if (v_usec < 0)
1291 1253 v_usec = -v_usec;
1292 1254 if (v_usec > (usec_per_tick >> 1)) {
1293 1255 if (pps_glitch > MAXGLITCH) {
1294 1256 pps_glitch = 0;
1295 1257 pps_tf[2] = u_usec;
1296 1258 pps_tf[1] = u_usec;
1297 1259 } else {
1298 1260 pps_glitch++;
1299 1261 u_usec = pps_offset;
1300 1262 }
1301 1263 } else
1302 1264 pps_glitch = 0;
1303 1265
1304 1266 /*
1305 1267 * A three-stage median filter is used to help deglitch the pps
1306 1268 * time. The median sample becomes the time offset estimate; the
1307 1269 * difference between the other two samples becomes the time
1308 1270 * dispersion (jitter) estimate.
1309 1271 */
1310 1272 pps_tf[2] = pps_tf[1];
1311 1273 pps_tf[1] = pps_tf[0];
1312 1274 pps_tf[0] = u_usec;
1313 1275 if (pps_tf[0] > pps_tf[1]) {
1314 1276 if (pps_tf[1] > pps_tf[2]) {
1315 1277 pps_offset = pps_tf[1]; /* 0 1 2 */
1316 1278 v_usec = pps_tf[0] - pps_tf[2];
1317 1279 } else if (pps_tf[2] > pps_tf[0]) {
1318 1280 pps_offset = pps_tf[0]; /* 2 0 1 */
1319 1281 v_usec = pps_tf[2] - pps_tf[1];
1320 1282 } else {
1321 1283 pps_offset = pps_tf[2]; /* 0 2 1 */
1322 1284 v_usec = pps_tf[0] - pps_tf[1];
1323 1285 }
1324 1286 } else {
1325 1287 if (pps_tf[1] < pps_tf[2]) {
1326 1288 pps_offset = pps_tf[1]; /* 2 1 0 */
1327 1289 v_usec = pps_tf[2] - pps_tf[0];
1328 1290 } else if (pps_tf[2] < pps_tf[0]) {
1329 1291 pps_offset = pps_tf[0]; /* 1 0 2 */
1330 1292 v_usec = pps_tf[1] - pps_tf[2];
1331 1293 } else {
1332 1294 pps_offset = pps_tf[2]; /* 1 2 0 */
1333 1295 v_usec = pps_tf[1] - pps_tf[0];
1334 1296 }
1335 1297 }
1336 1298 if (v_usec > MAXTIME)
1337 1299 pps_jitcnt++;
1338 1300 v_usec = (v_usec << PPS_AVG) - pps_jitter;
1339 1301 pps_jitter += v_usec / (1 << PPS_AVG);
1340 1302 if (pps_jitter > (MAXTIME >> 1))
1341 1303 time_status |= STA_PPSJITTER;
1342 1304
1343 1305 /*
1344 1306 * During the calibration interval adjust the starting time when
1345 1307 * the tick overflows. At the end of the interval compute the
1346 1308 * duration of the interval and the difference of the hardware
1347 1309 * counters at the beginning and end of the interval. This code
1348 1310 * is deliciously complicated by the fact valid differences may
1349 1311 * exceed the value of tick when using long calibration
1350 1312 * intervals and small ticks. Note that the counter can be
1351 1313 * greater than tick if caught at just the wrong instant, but
1352 1314 * the values returned and used here are correct.
1353 1315 */
1354 1316 bigtick = (int)usec_per_tick * SCALE_USEC;
1355 1317 pps_usec -= pps_freq;
1356 1318 if (pps_usec >= bigtick)
1357 1319 pps_usec -= bigtick;
1358 1320 if (pps_usec < 0)
1359 1321 pps_usec += bigtick;
1360 1322 pps_time.tv_sec++;
1361 1323 pps_count++;
1362 1324 if (pps_count < (1 << pps_shift))
1363 1325 return;
1364 1326 pps_count = 0;
1365 1327 pps_calcnt++;
1366 1328 u_usec = usec * SCALE_USEC;
1367 1329 v_usec = pps_usec - u_usec;
1368 1330 if (v_usec >= bigtick >> 1)
1369 1331 v_usec -= bigtick;
1370 1332 if (v_usec < -(bigtick >> 1))
1371 1333 v_usec += bigtick;
1372 1334 if (v_usec < 0)
1373 1335 v_usec = -(-v_usec >> pps_shift);
1374 1336 else
1375 1337 v_usec = v_usec >> pps_shift;
1376 1338 pps_usec = u_usec;
1377 1339 cal_sec = tvp->tv_sec;
1378 1340 cal_usec = tvp->tv_usec;
1379 1341 cal_sec -= pps_time.tv_sec;
1380 1342 cal_usec -= pps_time.tv_usec;
1381 1343 if (cal_usec < 0) {
1382 1344 cal_usec += MICROSEC;
1383 1345 cal_sec--;
1384 1346 }
1385 1347 pps_time = *tvp;
1386 1348
1387 1349 /*
1388 1350 * Check for lost interrupts, noise, excessive jitter and
1389 1351 * excessive frequency error. The number of timer ticks during
1390 1352 * the interval may vary +-1 tick. Add to this a margin of one
1391 1353 * tick for the PPS signal jitter and maximum frequency
1392 1354 * deviation. If the limits are exceeded, the calibration
1393 1355 * interval is reset to the minimum and we start over.
1394 1356 */
1395 1357 u_usec = (int)usec_per_tick << 1;
1396 1358 if (!((cal_sec == -1 && cal_usec > (MICROSEC - u_usec)) ||
1397 1359 (cal_sec == 0 && cal_usec < u_usec)) ||
1398 1360 v_usec > time_tolerance || v_usec < -time_tolerance) {
1399 1361 pps_errcnt++;
1400 1362 pps_shift = PPS_SHIFT;
1401 1363 pps_intcnt = 0;
1402 1364 time_status |= STA_PPSERROR;
1403 1365 return;
1404 1366 }
1405 1367
1406 1368 /*
1407 1369 * A three-stage median filter is used to help deglitch the pps
1408 1370 * frequency. The median sample becomes the frequency offset
1409 1371 * estimate; the difference between the other two samples
1410 1372 * becomes the frequency dispersion (stability) estimate.
1411 1373 */
1412 1374 pps_ff[2] = pps_ff[1];
1413 1375 pps_ff[1] = pps_ff[0];
1414 1376 pps_ff[0] = v_usec;
1415 1377 if (pps_ff[0] > pps_ff[1]) {
1416 1378 if (pps_ff[1] > pps_ff[2]) {
1417 1379 u_usec = pps_ff[1]; /* 0 1 2 */
1418 1380 v_usec = pps_ff[0] - pps_ff[2];
1419 1381 } else if (pps_ff[2] > pps_ff[0]) {
1420 1382 u_usec = pps_ff[0]; /* 2 0 1 */
1421 1383 v_usec = pps_ff[2] - pps_ff[1];
1422 1384 } else {
1423 1385 u_usec = pps_ff[2]; /* 0 2 1 */
1424 1386 v_usec = pps_ff[0] - pps_ff[1];
1425 1387 }
1426 1388 } else {
1427 1389 if (pps_ff[1] < pps_ff[2]) {
1428 1390 u_usec = pps_ff[1]; /* 2 1 0 */
1429 1391 v_usec = pps_ff[2] - pps_ff[0];
1430 1392 } else if (pps_ff[2] < pps_ff[0]) {
1431 1393 u_usec = pps_ff[0]; /* 1 0 2 */
1432 1394 v_usec = pps_ff[1] - pps_ff[2];
1433 1395 } else {
1434 1396 u_usec = pps_ff[2]; /* 1 2 0 */
1435 1397 v_usec = pps_ff[1] - pps_ff[0];
1436 1398 }
1437 1399 }
1438 1400
1439 1401 /*
1440 1402 * Here the frequency dispersion (stability) is updated. If it
1441 1403 * is less than one-fourth the maximum (MAXFREQ), the frequency
1442 1404 * offset is updated as well, but clamped to the tolerance. It
1443 1405 * will be processed later by the clock() routine.
1444 1406 */
1445 1407 v_usec = (v_usec >> 1) - pps_stabil;
1446 1408 if (v_usec < 0)
1447 1409 pps_stabil -= -v_usec >> PPS_AVG;
1448 1410 else
1449 1411 pps_stabil += v_usec >> PPS_AVG;
1450 1412 if (pps_stabil > MAXFREQ >> 2) {
1451 1413 pps_stbcnt++;
1452 1414 time_status |= STA_PPSWANDER;
1453 1415 return;
1454 1416 }
1455 1417 if (time_status & STA_PPSFREQ) {
1456 1418 if (u_usec < 0) {
1457 1419 pps_freq -= -u_usec >> PPS_AVG;
1458 1420 if (pps_freq < -time_tolerance)
1459 1421 pps_freq = -time_tolerance;
1460 1422 u_usec = -u_usec;
1461 1423 } else {
1462 1424 pps_freq += u_usec >> PPS_AVG;
1463 1425 if (pps_freq > time_tolerance)
1464 1426 pps_freq = time_tolerance;
1465 1427 }
1466 1428 }
1467 1429
1468 1430 /*
1469 1431 * Here the calibration interval is adjusted. If the maximum
1470 1432 * time difference is greater than tick / 4, reduce the interval
1471 1433 * by half. If this is not the case for four consecutive
1472 1434 * intervals, double the interval.
1473 1435 */
1474 1436 if (u_usec << pps_shift > bigtick >> 2) {
1475 1437 pps_intcnt = 0;
1476 1438 if (pps_shift > PPS_SHIFT)
1477 1439 pps_shift--;
1478 1440 } else if (pps_intcnt >= 4) {
1479 1441 pps_intcnt = 0;
1480 1442 if (pps_shift < PPS_SHIFTMAX)
1481 1443 pps_shift++;
1482 1444 } else
1483 1445 pps_intcnt++;
1484 1446
1485 1447 /*
1486 1448 * If recovering from kmdb, then make sure the tod chip gets resynced.
1487 1449 * If we took an early exit above, then we don't yet have a stable
1488 1450 * calibration signal to lock onto, so don't mark the tod for sync
1489 1451 * until we get all the way here.
1490 1452 */
1491 1453 {
1492 1454 int s = hr_clock_lock();
1493 1455
1494 1456 tod_needsync = 1;
1495 1457 hr_clock_unlock(s);
1496 1458 }
1497 1459 }
1498 1460
1499 1461 /*
1500 1462 * Handle clock tick processing for a thread.
1501 1463 * Check for timer action, enforce CPU rlimit, do profiling etc.
1502 1464 */
1503 1465 void
1504 1466 clock_tick(kthread_t *t, int pending)
1505 1467 {
1506 1468 struct proc *pp;
1507 1469 klwp_id_t lwp;
1508 1470 struct as *as;
1509 1471 clock_t ticks;
1510 1472 int poke = 0; /* notify another CPU */
1511 1473 int user_mode;
1512 1474 size_t rss;
1513 1475 int i, total_usec, usec;
1514 1476 rctl_qty_t secs;
1515 1477
1516 1478 ASSERT(pending > 0);
1517 1479
1518 1480 /* Must be operating on a lwp/thread */
1519 1481 if ((lwp = ttolwp(t)) == NULL) {
1520 1482 panic("clock_tick: no lwp");
1521 1483 /*NOTREACHED*/
1522 1484 }
1523 1485
1524 1486 for (i = 0; i < pending; i++) {
1525 1487 CL_TICK(t); /* Class specific tick processing */
1526 1488 DTRACE_SCHED1(tick, kthread_t *, t);
1527 1489 }
1528 1490
1529 1491 pp = ttoproc(t);
1530 1492
1531 1493 /* pp->p_lock makes sure that the thread does not exit */
1532 1494 ASSERT(MUTEX_HELD(&pp->p_lock));
1533 1495
1534 1496 user_mode = (lwp->lwp_state == LWP_USER);
1535 1497
1536 1498 ticks = (pp->p_utime + pp->p_stime) % hz;
1537 1499 /*
1538 1500 * Update process times. Should use high res clock and state
1539 1501 * changes instead of statistical sampling method. XXX
1540 1502 */
1541 1503 if (user_mode) {
1542 1504 pp->p_utime += pending;
1543 1505 } else {
1544 1506 pp->p_stime += pending;
1545 1507 }
1546 1508
1547 1509 pp->p_ttime += pending;
1548 1510 as = pp->p_as;
1549 1511
1550 1512 /*
1551 1513 * Update user profiling statistics. Get the pc from the
1552 1514 * lwp when the AST happens.
1553 1515 */
1554 1516 if (pp->p_prof.pr_scale) {
1555 1517 atomic_add_32(&lwp->lwp_oweupc, (int32_t)pending);
1556 1518 if (user_mode) {
1557 1519 poke = 1;
1558 1520 aston(t);
1559 1521 }
1560 1522 }
1561 1523
1562 1524 /*
1563 1525 * If CPU was in user state, process lwp-virtual time
1564 1526 * interval timer. The value passed to itimerdecr() has to be
1565 1527 * in microseconds and has to be less than one second. Hence
1566 1528 * this loop.
1567 1529 */
1568 1530 total_usec = usec_per_tick * pending;
1569 1531 while (total_usec > 0) {
1570 1532 usec = MIN(total_usec, (MICROSEC - 1));
1571 1533 if (user_mode &&
1572 1534 timerisset(&lwp->lwp_timer[ITIMER_VIRTUAL].it_value) &&
1573 1535 itimerdecr(&lwp->lwp_timer[ITIMER_VIRTUAL], usec) == 0) {
1574 1536 poke = 1;
1575 1537 sigtoproc(pp, t, SIGVTALRM);
1576 1538 }
1577 1539 total_usec -= usec;
1578 1540 }
1579 1541
1580 1542 /*
1581 1543 * If CPU was in user state, process lwp-profile
1582 1544 * interval timer.
1583 1545 */
1584 1546 total_usec = usec_per_tick * pending;
1585 1547 while (total_usec > 0) {
1586 1548 usec = MIN(total_usec, (MICROSEC - 1));
1587 1549 if (timerisset(&lwp->lwp_timer[ITIMER_PROF].it_value) &&
1588 1550 itimerdecr(&lwp->lwp_timer[ITIMER_PROF], usec) == 0) {
1589 1551 poke = 1;
1590 1552 sigtoproc(pp, t, SIGPROF);
1591 1553 }
1592 1554 total_usec -= usec;
1593 1555 }
1594 1556
1595 1557 /*
1596 1558 * Enforce CPU resource controls:
1597 1559 * (a) process.max-cpu-time resource control
1598 1560 *
1599 1561 * Perform the check only if we have accumulated more a second.
1600 1562 */
1601 1563 if ((ticks + pending) >= hz) {
1602 1564 (void) rctl_test(rctlproc_legacy[RLIMIT_CPU], pp->p_rctls, pp,
1603 1565 (pp->p_utime + pp->p_stime)/hz, RCA_UNSAFE_SIGINFO);
1604 1566 }
1605 1567
1606 1568 /*
1607 1569 * (b) task.max-cpu-time resource control
1608 1570 *
1609 1571 * If we have accumulated enough ticks, increment the task CPU
1610 1572 * time usage and test for the resource limit. This minimizes the
1611 1573 * number of calls to the rct_test(). The task CPU time mutex
1612 1574 * is highly contentious as many processes can be sharing a task.
1613 1575 */
1614 1576 if (pp->p_ttime >= clock_tick_proc_max) {
1615 1577 secs = task_cpu_time_incr(pp->p_task, pp->p_ttime);
1616 1578 pp->p_ttime = 0;
1617 1579 if (secs) {
1618 1580 (void) rctl_test(rc_task_cpu_time, pp->p_task->tk_rctls,
1619 1581 pp, secs, RCA_UNSAFE_SIGINFO);
1620 1582 }
1621 1583 }
1622 1584
1623 1585 /*
1624 1586 * Update memory usage for the currently running process.
1625 1587 */
1626 1588 rss = rm_asrss(as);
1627 1589 PTOU(pp)->u_mem += rss;
1628 1590 if (rss > PTOU(pp)->u_mem_max)
1629 1591 PTOU(pp)->u_mem_max = rss;
1630 1592
1631 1593 /*
1632 1594 * Notify the CPU the thread is running on.
1633 1595 */
1634 1596 if (poke && t->t_cpu != CPU)
1635 1597 poke_cpu(t->t_cpu->cpu_id);
1636 1598 }
1637 1599
1638 1600 void
1639 1601 profil_tick(uintptr_t upc)
1640 1602 {
1641 1603 int ticks;
1642 1604 proc_t *p = ttoproc(curthread);
1643 1605 klwp_t *lwp = ttolwp(curthread);
1644 1606 struct prof *pr = &p->p_prof;
1645 1607
1646 1608 do {
1647 1609 ticks = lwp->lwp_oweupc;
1648 1610 } while (cas32(&lwp->lwp_oweupc, ticks, 0) != ticks);
1649 1611
1650 1612 mutex_enter(&p->p_pflock);
1651 1613 if (pr->pr_scale >= 2 && upc >= pr->pr_off) {
1652 1614 /*
1653 1615 * Old-style profiling
1654 1616 */
1655 1617 uint16_t *slot = pr->pr_base;
1656 1618 uint16_t old, new;
1657 1619 if (pr->pr_scale != 2) {
1658 1620 uintptr_t delta = upc - pr->pr_off;
1659 1621 uintptr_t byteoff = ((delta >> 16) * pr->pr_scale) +
1660 1622 (((delta & 0xffff) * pr->pr_scale) >> 16);
1661 1623 if (byteoff >= (uintptr_t)pr->pr_size) {
1662 1624 mutex_exit(&p->p_pflock);
1663 1625 return;
1664 1626 }
1665 1627 slot += byteoff / sizeof (uint16_t);
1666 1628 }
1667 1629 if (fuword16(slot, &old) < 0 ||
1668 1630 (new = old + ticks) > SHRT_MAX ||
1669 1631 suword16(slot, new) < 0) {
1670 1632 pr->pr_scale = 0;
1671 1633 }
1672 1634 } else if (pr->pr_scale == 1) {
1673 1635 /*
1674 1636 * PC Sampling
1675 1637 */
1676 1638 model_t model = lwp_getdatamodel(lwp);
1677 1639 int result;
1678 1640 #ifdef __lint
1679 1641 model = model;
1680 1642 #endif
1681 1643 while (ticks-- > 0) {
1682 1644 if (pr->pr_samples == pr->pr_size) {
1683 1645 /* buffer full, turn off sampling */
1684 1646 pr->pr_scale = 0;
1685 1647 break;
1686 1648 }
1687 1649 switch (SIZEOF_PTR(model)) {
1688 1650 case sizeof (uint32_t):
1689 1651 result = suword32(pr->pr_base, (uint32_t)upc);
1690 1652 break;
1691 1653 #ifdef _LP64
1692 1654 case sizeof (uint64_t):
1693 1655 result = suword64(pr->pr_base, (uint64_t)upc);
1694 1656 break;
1695 1657 #endif
1696 1658 default:
1697 1659 cmn_err(CE_WARN, "profil_tick: unexpected "
1698 1660 "data model");
1699 1661 result = -1;
1700 1662 break;
1701 1663 }
1702 1664 if (result != 0) {
1703 1665 pr->pr_scale = 0;
1704 1666 break;
1705 1667 }
1706 1668 pr->pr_base = (caddr_t)pr->pr_base + SIZEOF_PTR(model);
1707 1669 pr->pr_samples++;
1708 1670 }
1709 1671 }
1710 1672 mutex_exit(&p->p_pflock);
1711 1673 }
1712 1674
1713 1675 static void
1714 1676 delay_wakeup(void *arg)
1715 1677 {
1716 1678 kthread_t *t = arg;
1717 1679
1718 1680 mutex_enter(&t->t_delay_lock);
1719 1681 cv_signal(&t->t_delay_cv);
1720 1682 mutex_exit(&t->t_delay_lock);
1721 1683 }
1722 1684
1723 1685 /*
1724 1686 * The delay(9F) man page indicates that it can only be called from user or
1725 1687 * kernel context - detect and diagnose bad calls. The following macro will
1726 1688 * produce a limited number of messages identifying bad callers. This is done
1727 1689 * in a macro so that caller() is meaningful. When a bad caller is identified,
1728 1690 * switching to 'drv_usecwait(TICK_TO_USEC(ticks));' may be appropriate.
1729 1691 */
1730 1692 #define DELAY_CONTEXT_CHECK() { \
1731 1693 uint32_t m; \
1732 1694 char *f; \
1733 1695 ulong_t off; \
1734 1696 \
1735 1697 m = delay_from_interrupt_msg; \
1736 1698 if (delay_from_interrupt_diagnose && servicing_interrupt() && \
1737 1699 !panicstr && !devinfo_freeze && \
1738 1700 atomic_cas_32(&delay_from_interrupt_msg, m ? m : 1, m-1)) { \
1739 1701 f = modgetsymname((uintptr_t)caller(), &off); \
1740 1702 cmn_err(CE_WARN, "delay(9F) called from " \
1741 1703 "interrupt context: %s`%s", \
1742 1704 mod_containing_pc(caller()), f ? f : "..."); \
1743 1705 } \
1744 1706 }
1745 1707
1746 1708 /*
1747 1709 * delay_common: common delay code.
1748 1710 */
1749 1711 static void
1750 1712 delay_common(clock_t ticks)
1751 1713 {
1752 1714 kthread_t *t = curthread;
1753 1715 clock_t deadline;
1754 1716 clock_t timeleft;
1755 1717 callout_id_t id;
1756 1718
1757 1719 /* If timeouts aren't running all we can do is spin. */
1758 1720 if (panicstr || devinfo_freeze) {
1759 1721 /* Convert delay(9F) call into drv_usecwait(9F) call. */
1760 1722 if (ticks > 0)
1761 1723 drv_usecwait(TICK_TO_USEC(ticks));
1762 1724 return;
1763 1725 }
1764 1726
1765 1727 deadline = ddi_get_lbolt() + ticks;
1766 1728 while ((timeleft = deadline - ddi_get_lbolt()) > 0) {
1767 1729 mutex_enter(&t->t_delay_lock);
1768 1730 id = timeout_default(delay_wakeup, t, timeleft);
1769 1731 cv_wait(&t->t_delay_cv, &t->t_delay_lock);
1770 1732 mutex_exit(&t->t_delay_lock);
1771 1733 (void) untimeout_default(id, 0);
1772 1734 }
1773 1735 }
1774 1736
1775 1737 /*
1776 1738 * Delay specified number of clock ticks.
1777 1739 */
1778 1740 void
1779 1741 delay(clock_t ticks)
1780 1742 {
1781 1743 DELAY_CONTEXT_CHECK();
1782 1744
1783 1745 delay_common(ticks);
1784 1746 }
1785 1747
1786 1748 /*
1787 1749 * Delay a random number of clock ticks between 1 and ticks.
1788 1750 */
1789 1751 void
1790 1752 delay_random(clock_t ticks)
1791 1753 {
1792 1754 int r;
1793 1755
1794 1756 DELAY_CONTEXT_CHECK();
1795 1757
1796 1758 (void) random_get_pseudo_bytes((void *)&r, sizeof (r));
1797 1759 if (ticks == 0)
1798 1760 ticks = 1;
1799 1761 ticks = (r % ticks) + 1;
1800 1762 delay_common(ticks);
1801 1763 }
1802 1764
1803 1765 /*
1804 1766 * Like delay, but interruptible by a signal.
1805 1767 */
1806 1768 int
1807 1769 delay_sig(clock_t ticks)
1808 1770 {
1809 1771 kthread_t *t = curthread;
1810 1772 clock_t deadline;
1811 1773 clock_t rc;
1812 1774
1813 1775 /* If timeouts aren't running all we can do is spin. */
1814 1776 if (panicstr || devinfo_freeze) {
1815 1777 if (ticks > 0)
1816 1778 drv_usecwait(TICK_TO_USEC(ticks));
1817 1779 return (0);
1818 1780 }
1819 1781
1820 1782 deadline = ddi_get_lbolt() + ticks;
1821 1783 mutex_enter(&t->t_delay_lock);
1822 1784 do {
1823 1785 rc = cv_timedwait_sig(&t->t_delay_cv,
1824 1786 &t->t_delay_lock, deadline);
1825 1787 /* loop until past deadline or signaled */
1826 1788 } while (rc > 0);
1827 1789 mutex_exit(&t->t_delay_lock);
1828 1790 if (rc == 0)
1829 1791 return (EINTR);
1830 1792 return (0);
1831 1793 }
1832 1794
1833 1795
1834 1796 #define SECONDS_PER_DAY 86400
1835 1797
1836 1798 /*
1837 1799 * Initialize the system time based on the TOD chip. approx is used as
1838 1800 * an approximation of time (e.g. from the filesystem) in the event that
1839 1801 * the TOD chip has been cleared or is unresponsive. An approx of -1
1840 1802 * means the filesystem doesn't keep time.
1841 1803 */
1842 1804 void
1843 1805 clkset(time_t approx)
1844 1806 {
1845 1807 timestruc_t ts;
1846 1808 int spl;
1847 1809 int set_clock = 0;
1848 1810
1849 1811 mutex_enter(&tod_lock);
1850 1812 ts = tod_get();
1851 1813
1852 1814 if (ts.tv_sec > 365 * SECONDS_PER_DAY) {
1853 1815 /*
1854 1816 * If the TOD chip is reporting some time after 1971,
1855 1817 * then it probably didn't lose power or become otherwise
1856 1818 * cleared in the recent past; check to assure that
1857 1819 * the time coming from the filesystem isn't in the future
1858 1820 * according to the TOD chip.
1859 1821 */
1860 1822 if (approx != -1 && approx > ts.tv_sec) {
1861 1823 cmn_err(CE_WARN, "Last shutdown is later "
1862 1824 "than time on time-of-day chip; check date.");
1863 1825 }
1864 1826 } else {
1865 1827 /*
1866 1828 * If the TOD chip isn't giving correct time, set it to the
1867 1829 * greater of i) approx and ii) 1987. That way if approx
1868 1830 * is negative or is earlier than 1987, we set the clock
1869 1831 * back to a time when Oliver North, ALF and Dire Straits
1870 1832 * were all on the collective brain: 1987.
1871 1833 */
1872 1834 timestruc_t tmp;
1873 1835 time_t diagnose_date = (1987 - 1970) * 365 * SECONDS_PER_DAY;
1874 1836 ts.tv_sec = (approx > diagnose_date ? approx : diagnose_date);
1875 1837 ts.tv_nsec = 0;
1876 1838
1877 1839 /*
1878 1840 * Attempt to write the new time to the TOD chip. Set spl high
1879 1841 * to avoid getting preempted between the tod_set and tod_get.
1880 1842 */
1881 1843 spl = splhi();
1882 1844 tod_set(ts);
1883 1845 tmp = tod_get();
1884 1846 splx(spl);
1885 1847
1886 1848 if (tmp.tv_sec != ts.tv_sec && tmp.tv_sec != ts.tv_sec + 1) {
1887 1849 tod_broken = 1;
1888 1850 dosynctodr = 0;
1889 1851 cmn_err(CE_WARN, "Time-of-day chip unresponsive.");
1890 1852 } else {
1891 1853 cmn_err(CE_WARN, "Time-of-day chip had "
1892 1854 "incorrect date; check and reset.");
1893 1855 }
1894 1856 set_clock = 1;
1895 1857 }
1896 1858
1897 1859 if (!boot_time) {
1898 1860 boot_time = ts.tv_sec;
1899 1861 set_clock = 1;
1900 1862 }
1901 1863
1902 1864 if (set_clock)
1903 1865 set_hrestime(&ts);
1904 1866
1905 1867 mutex_exit(&tod_lock);
1906 1868 }
1907 1869
1908 1870 int timechanged; /* for testing if the system time has been reset */
1909 1871
1910 1872 void
1911 1873 set_hrestime(timestruc_t *ts)
1912 1874 {
1913 1875 int spl = hr_clock_lock();
1914 1876 hrestime = *ts;
1915 1877 membar_enter(); /* hrestime must be visible before timechanged++ */
1916 1878 timedelta = 0;
1917 1879 timechanged++;
1918 1880 hr_clock_unlock(spl);
1919 1881 callout_hrestime();
1920 1882 }
1921 1883
1922 1884 static uint_t deadman_seconds;
1923 1885 static uint32_t deadman_panics;
1924 1886 static int deadman_enabled = 0;
1925 1887 static int deadman_panic_timers = 1;
1926 1888
1927 1889 static void
1928 1890 deadman(void)
1929 1891 {
1930 1892 if (panicstr) {
1931 1893 /*
1932 1894 * During panic, other CPUs besides the panic
1933 1895 * master continue to handle cyclics and some other
1934 1896 * interrupts. The code below is intended to be
1935 1897 * single threaded, so any CPU other than the master
1936 1898 * must keep out.
1937 1899 */
1938 1900 if (CPU->cpu_id != panic_cpu.cpu_id)
1939 1901 return;
1940 1902
1941 1903 if (!deadman_panic_timers)
1942 1904 return; /* allow all timers to be manually disabled */
1943 1905
1944 1906 /*
1945 1907 * If we are generating a crash dump or syncing filesystems and
1946 1908 * the corresponding timer is set, decrement it and re-enter
1947 1909 * the panic code to abort it and advance to the next state.
1948 1910 * The panic states and triggers are explained in panic.c.
1949 1911 */
1950 1912 if (panic_dump) {
1951 1913 if (dump_timeleft && (--dump_timeleft == 0)) {
1952 1914 panic("panic dump timeout");
1953 1915 /*NOTREACHED*/
1954 1916 }
1955 1917 } else if (panic_sync) {
1956 1918 if (sync_timeleft && (--sync_timeleft == 0)) {
1957 1919 panic("panic sync timeout");
1958 1920 /*NOTREACHED*/
1959 1921 }
1960 1922 }
1961 1923
1962 1924 return;
1963 1925 }
1964 1926
1965 1927 if (deadman_counter != CPU->cpu_deadman_counter) {
1966 1928 CPU->cpu_deadman_counter = deadman_counter;
1967 1929 CPU->cpu_deadman_countdown = deadman_seconds;
1968 1930 return;
1969 1931 }
1970 1932
1971 1933 if (--CPU->cpu_deadman_countdown > 0)
1972 1934 return;
1973 1935
1974 1936 /*
1975 1937 * Regardless of whether or not we actually bring the system down,
1976 1938 * bump the deadman_panics variable.
1977 1939 *
1978 1940 * N.B. deadman_panics is incremented once for each CPU that
1979 1941 * passes through here. It's expected that all the CPUs will
1980 1942 * detect this condition within one second of each other, so
1981 1943 * when deadman_enabled is off, deadman_panics will
1982 1944 * typically be a multiple of the total number of CPUs in
1983 1945 * the system.
1984 1946 */
1985 1947 atomic_add_32(&deadman_panics, 1);
1986 1948
1987 1949 if (!deadman_enabled) {
1988 1950 CPU->cpu_deadman_countdown = deadman_seconds;
1989 1951 return;
1990 1952 }
1991 1953
1992 1954 /*
1993 1955 * If we're here, we want to bring the system down.
1994 1956 */
1995 1957 panic("deadman: timed out after %d seconds of clock "
1996 1958 "inactivity", deadman_seconds);
1997 1959 /*NOTREACHED*/
1998 1960 }
1999 1961
2000 1962 /*ARGSUSED*/
2001 1963 static void
2002 1964 deadman_online(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
2003 1965 {
2004 1966 cpu->cpu_deadman_counter = 0;
2005 1967 cpu->cpu_deadman_countdown = deadman_seconds;
2006 1968
2007 1969 hdlr->cyh_func = (cyc_func_t)deadman;
2008 1970 hdlr->cyh_level = CY_HIGH_LEVEL;
2009 1971 hdlr->cyh_arg = NULL;
2010 1972
2011 1973 /*
2012 1974 * Stagger the CPUs so that they don't all run deadman() at
2013 1975 * the same time. Simplest reason to do this is to make it
2014 1976 * more likely that only one CPU will panic in case of a
2015 1977 * timeout. This is (strictly speaking) an aesthetic, not a
2016 1978 * technical consideration.
2017 1979 */
2018 1980 when->cyt_when = cpu->cpu_id * (NANOSEC / NCPU);
2019 1981 when->cyt_interval = NANOSEC;
2020 1982 }
2021 1983
2022 1984
2023 1985 void
2024 1986 deadman_init(void)
2025 1987 {
2026 1988 cyc_omni_handler_t hdlr;
2027 1989
2028 1990 if (deadman_seconds == 0)
2029 1991 deadman_seconds = snoop_interval / MICROSEC;
2030 1992
2031 1993 if (snooping)
2032 1994 deadman_enabled = 1;
2033 1995
2034 1996 hdlr.cyo_online = deadman_online;
2035 1997 hdlr.cyo_offline = NULL;
2036 1998 hdlr.cyo_arg = NULL;
2037 1999
2038 2000 mutex_enter(&cpu_lock);
2039 2001 deadman_cyclic = cyclic_add_omni(&hdlr);
2040 2002 mutex_exit(&cpu_lock);
2041 2003 }
2042 2004
2043 2005 /*
2044 2006 * tod_fault() is for updating tod validate mechanism state:
2045 2007 * (1) TOD_NOFAULT: for resetting the state to 'normal'.
2046 2008 * currently used for debugging only
2047 2009 * (2) The following four cases detected by tod validate mechanism:
2048 2010 * TOD_REVERSED: current tod value is less than previous value.
2049 2011 * TOD_STALLED: current tod value hasn't advanced.
2050 2012 * TOD_JUMPED: current tod value advanced too far from previous value.
2051 2013 * TOD_RATECHANGED: the ratio between average tod delta and
2052 2014 * average tick delta has changed.
2053 2015 * (3) TOD_RDONLY: when the TOD clock is not writeable e.g. because it is
2054 2016 * a virtual TOD provided by a hypervisor.
2055 2017 */
2056 2018 enum tod_fault_type
2057 2019 tod_fault(enum tod_fault_type ftype, int off)
2058 2020 {
2059 2021 ASSERT(MUTEX_HELD(&tod_lock));
2060 2022
2061 2023 if (tod_faulted != ftype) {
2062 2024 switch (ftype) {
2063 2025 case TOD_NOFAULT:
2064 2026 plat_tod_fault(TOD_NOFAULT);
2065 2027 cmn_err(CE_NOTE, "Restarted tracking "
2066 2028 "Time of Day clock.");
2067 2029 tod_faulted = ftype;
2068 2030 break;
2069 2031 case TOD_REVERSED:
2070 2032 case TOD_JUMPED:
2071 2033 if (tod_faulted == TOD_NOFAULT) {
2072 2034 plat_tod_fault(ftype);
2073 2035 cmn_err(CE_WARN, "Time of Day clock error: "
2074 2036 "reason [%s by 0x%x]. -- "
2075 2037 " Stopped tracking Time Of Day clock.",
2076 2038 tod_fault_table[ftype], off);
2077 2039 tod_faulted = ftype;
2078 2040 }
2079 2041 break;
2080 2042 case TOD_STALLED:
2081 2043 case TOD_RATECHANGED:
2082 2044 if (tod_faulted == TOD_NOFAULT) {
2083 2045 plat_tod_fault(ftype);
2084 2046 cmn_err(CE_WARN, "Time of Day clock error: "
2085 2047 "reason [%s]. -- "
2086 2048 " Stopped tracking Time Of Day clock.",
2087 2049 tod_fault_table[ftype]);
2088 2050 tod_faulted = ftype;
2089 2051 }
2090 2052 break;
2091 2053 case TOD_RDONLY:
2092 2054 if (tod_faulted == TOD_NOFAULT) {
2093 2055 plat_tod_fault(ftype);
2094 2056 cmn_err(CE_NOTE, "!Time of Day clock is "
2095 2057 "Read-Only; set of Date/Time will not "
2096 2058 "persist across reboot.");
2097 2059 tod_faulted = ftype;
2098 2060 }
2099 2061 break;
2100 2062 default:
2101 2063 break;
2102 2064 }
2103 2065 }
2104 2066 return (tod_faulted);
2105 2067 }
2106 2068
2107 2069 /*
2108 2070 * Two functions that allow tod_status_flag to be manipulated by functions
2109 2071 * external to this file.
2110 2072 */
2111 2073
2112 2074 void
2113 2075 tod_status_set(int tod_flag)
2114 2076 {
2115 2077 tod_status_flag |= tod_flag;
2116 2078 }
2117 2079
2118 2080 void
2119 2081 tod_status_clear(int tod_flag)
2120 2082 {
2121 2083 tod_status_flag &= ~tod_flag;
2122 2084 }
2123 2085
2124 2086 /*
2125 2087 * Record a timestamp and the value passed to tod_set(). The next call to
2126 2088 * tod_validate() can use these values, prev_set_tick and prev_set_tod,
2127 2089 * when checking the timestruc_t returned by tod_get(). Ordinarily,
2128 2090 * tod_validate() will use prev_tick and prev_tod for this task but these
2129 2091 * become obsolete, and will be re-assigned with the prev_set_* values,
2130 2092 * in the case when the TOD is re-written.
2131 2093 */
2132 2094 void
2133 2095 tod_set_prev(timestruc_t ts)
2134 2096 {
2135 2097 if ((tod_validate_enable == 0) || (tod_faulted != TOD_NOFAULT) ||
2136 2098 tod_validate_deferred) {
2137 2099 return;
2138 2100 }
2139 2101 prev_set_tick = gethrtime();
2140 2102 /*
2141 2103 * A negative value will be set to zero in utc_to_tod() so we fake
2142 2104 * a zero here in such a case. This would need to change if the
2143 2105 * behavior of utc_to_tod() changes.
2144 2106 */
2145 2107 prev_set_tod = ts.tv_sec < 0 ? 0 : ts.tv_sec;
2146 2108 }
2147 2109
2148 2110 /*
2149 2111 * tod_validate() is used for checking values returned by tod_get().
2150 2112 * Four error cases can be detected by this routine:
2151 2113 * TOD_REVERSED: current tod value is less than previous.
2152 2114 * TOD_STALLED: current tod value hasn't advanced.
2153 2115 * TOD_JUMPED: current tod value advanced too far from previous value.
2154 2116 * TOD_RATECHANGED: the ratio between average tod delta and
2155 2117 * average tick delta has changed.
2156 2118 */
2157 2119 time_t
2158 2120 tod_validate(time_t tod)
2159 2121 {
2160 2122 time_t diff_tod;
2161 2123 hrtime_t diff_tick;
2162 2124
2163 2125 long dtick;
2164 2126 int dtick_delta;
2165 2127
2166 2128 int off = 0;
2167 2129 enum tod_fault_type tod_bad = TOD_NOFAULT;
2168 2130
2169 2131 static int firsttime = 1;
2170 2132
2171 2133 static time_t prev_tod = 0;
2172 2134 static hrtime_t prev_tick = 0;
2173 2135 static long dtick_avg = TOD_REF_FREQ;
2174 2136
2175 2137 int cpr_resume_done = 0;
2176 2138 int dr_resume_done = 0;
2177 2139
2178 2140 hrtime_t tick = gethrtime();
2179 2141
2180 2142 ASSERT(MUTEX_HELD(&tod_lock));
2181 2143
2182 2144 /*
2183 2145 * tod_validate_enable is patchable via /etc/system.
2184 2146 * If TOD is already faulted, or if TOD validation is deferred,
2185 2147 * there is nothing to do.
2186 2148 */
2187 2149 if ((tod_validate_enable == 0) || (tod_faulted != TOD_NOFAULT) ||
2188 2150 tod_validate_deferred) {
2189 2151 return (tod);
2190 2152 }
2191 2153
2192 2154 /*
2193 2155 * If this is the first time through, we just need to save the tod
2194 2156 * we were called with and hrtime so we can use them next time to
2195 2157 * validate tod_get().
2196 2158 */
2197 2159 if (firsttime) {
2198 2160 firsttime = 0;
2199 2161 prev_tod = tod;
2200 2162 prev_tick = tick;
2201 2163 return (tod);
2202 2164 }
2203 2165
2204 2166 /*
2205 2167 * Handle any flags that have been turned on by tod_status_set().
2206 2168 * In the case where a tod_set() is done and then a subsequent
2207 2169 * tod_get() fails (ie, both TOD_SET_DONE and TOD_GET_FAILED are
2208 2170 * true), we treat the TOD_GET_FAILED with precedence by switching
2209 2171 * off the flag, returning tod and leaving TOD_SET_DONE asserted
2210 2172 * until such time as tod_get() completes successfully.
2211 2173 */
2212 2174 if (tod_status_flag & TOD_GET_FAILED) {
2213 2175 /*
2214 2176 * tod_get() has encountered an issue, possibly transitory,
2215 2177 * when reading TOD. We'll just return the incoming tod
2216 2178 * value (which is actually hrestime.tv_sec in this case)
2217 2179 * and when we get a genuine tod, following a successful
2218 2180 * tod_get(), we can validate using prev_tod and prev_tick.
2219 2181 */
2220 2182 tod_status_flag &= ~TOD_GET_FAILED;
2221 2183 return (tod);
2222 2184 } else if (tod_status_flag & TOD_SET_DONE) {
2223 2185 /*
2224 2186 * TOD has been modified. Just before the TOD was written,
2225 2187 * tod_set_prev() saved tod and hrtime; we can now use
2226 2188 * those values, prev_set_tod and prev_set_tick, to validate
2227 2189 * the incoming tod that's just been read.
2228 2190 */
2229 2191 prev_tod = prev_set_tod;
2230 2192 prev_tick = prev_set_tick;
2231 2193 dtick_avg = TOD_REF_FREQ;
2232 2194 tod_status_flag &= ~TOD_SET_DONE;
2233 2195 /*
2234 2196 * If a tod_set() preceded a cpr_suspend() without an
2235 2197 * intervening tod_validate(), we need to ensure that a
2236 2198 * TOD_JUMPED condition is ignored.
2237 2199 * Note this isn't a concern in the case of DR as we've
2238 2200 * just reassigned dtick_avg, above.
2239 2201 */
2240 2202 if (tod_status_flag & TOD_CPR_RESUME_DONE) {
2241 2203 cpr_resume_done = 1;
2242 2204 tod_status_flag &= ~TOD_CPR_RESUME_DONE;
2243 2205 }
2244 2206 } else if (tod_status_flag & TOD_CPR_RESUME_DONE) {
2245 2207 /*
2246 2208 * The system's coming back from a checkpoint resume.
2247 2209 */
2248 2210 cpr_resume_done = 1;
2249 2211 tod_status_flag &= ~TOD_CPR_RESUME_DONE;
2250 2212 /*
2251 2213 * We need to handle the possibility of a CPR suspend
2252 2214 * operation having been initiated whilst a DR event was
2253 2215 * in-flight.
2254 2216 */
2255 2217 if (tod_status_flag & TOD_DR_RESUME_DONE) {
2256 2218 dr_resume_done = 1;
2257 2219 tod_status_flag &= ~TOD_DR_RESUME_DONE;
2258 2220 }
2259 2221 } else if (tod_status_flag & TOD_DR_RESUME_DONE) {
2260 2222 /*
2261 2223 * A Dynamic Reconfiguration event has taken place.
2262 2224 */
2263 2225 dr_resume_done = 1;
2264 2226 tod_status_flag &= ~TOD_DR_RESUME_DONE;
2265 2227 }
2266 2228
2267 2229 /* test hook */
2268 2230 switch (tod_unit_test) {
2269 2231 case 1: /* for testing jumping tod */
2270 2232 tod += tod_test_injector;
2271 2233 tod_unit_test = 0;
2272 2234 break;
2273 2235 case 2: /* for testing stuck tod bit */
2274 2236 tod |= 1 << tod_test_injector;
2275 2237 tod_unit_test = 0;
2276 2238 break;
2277 2239 case 3: /* for testing stalled tod */
2278 2240 tod = prev_tod;
2279 2241 tod_unit_test = 0;
2280 2242 break;
2281 2243 case 4: /* reset tod fault status */
2282 2244 (void) tod_fault(TOD_NOFAULT, 0);
2283 2245 tod_unit_test = 0;
2284 2246 break;
2285 2247 default:
2286 2248 break;
2287 2249 }
2288 2250
2289 2251 diff_tod = tod - prev_tod;
2290 2252 diff_tick = tick - prev_tick;
2291 2253
2292 2254 ASSERT(diff_tick >= 0);
2293 2255
2294 2256 if (diff_tod < 0) {
2295 2257 /* ERROR - tod reversed */
2296 2258 tod_bad = TOD_REVERSED;
2297 2259 off = (int)(prev_tod - tod);
2298 2260 } else if (diff_tod == 0) {
2299 2261 /* tod did not advance */
2300 2262 if (diff_tick > TOD_STALL_THRESHOLD) {
2301 2263 /* ERROR - tod stalled */
2302 2264 tod_bad = TOD_STALLED;
2303 2265 } else {
2304 2266 /*
2305 2267 * Make sure we don't update prev_tick
2306 2268 * so that diff_tick is calculated since
2307 2269 * the first diff_tod == 0
2308 2270 */
2309 2271 return (tod);
2310 2272 }
2311 2273 } else {
2312 2274 /* calculate dtick */
2313 2275 dtick = diff_tick / diff_tod;
2314 2276
2315 2277 /* update dtick averages */
2316 2278 dtick_avg += ((dtick - dtick_avg) / TOD_FILTER_N);
2317 2279
2318 2280 /*
2319 2281 * Calculate dtick_delta as
2320 2282 * variation from reference freq in quartiles
2321 2283 */
2322 2284 dtick_delta = (dtick_avg - TOD_REF_FREQ) /
2323 2285 (TOD_REF_FREQ >> 2);
2324 2286
2325 2287 /*
2326 2288 * Even with a perfectly functioning TOD device,
2327 2289 * when the number of elapsed seconds is low the
2328 2290 * algorithm can calculate a rate that is beyond
2329 2291 * tolerance, causing an error. The algorithm is
2330 2292 * inaccurate when elapsed time is low (less than
2331 2293 * 5 seconds).
2332 2294 */
2333 2295 if (diff_tod > 4) {
2334 2296 if (dtick < TOD_JUMP_THRESHOLD) {
2335 2297 /*
2336 2298 * If we've just done a CPR resume, we detect
2337 2299 * a jump in the TOD but, actually, what's
2338 2300 * happened is that the TOD has been increasing
2339 2301 * whilst the system was suspended and the tick
2340 2302 * count hasn't kept up. We consider the first
2341 2303 * occurrence of this after a resume as normal
2342 2304 * and ignore it; otherwise, in a non-resume
2343 2305 * case, we regard it as a TOD problem.
2344 2306 */
2345 2307 if (!cpr_resume_done) {
2346 2308 /* ERROR - tod jumped */
2347 2309 tod_bad = TOD_JUMPED;
2348 2310 off = (int)diff_tod;
2349 2311 }
2350 2312 }
2351 2313 if (dtick_delta) {
2352 2314 /*
2353 2315 * If we've just done a DR resume, dtick_avg
2354 2316 * can go a bit askew so we reset it and carry
2355 2317 * on; otherwise, the TOD is in error.
2356 2318 */
2357 2319 if (dr_resume_done) {
2358 2320 dtick_avg = TOD_REF_FREQ;
2359 2321 } else {
2360 2322 /* ERROR - change in clock rate */
2361 2323 tod_bad = TOD_RATECHANGED;
2362 2324 }
2363 2325 }
2364 2326 }
2365 2327 }
2366 2328
2367 2329 if (tod_bad != TOD_NOFAULT) {
2368 2330 (void) tod_fault(tod_bad, off);
2369 2331
2370 2332 /*
2371 2333 * Disable dosynctodr since we are going to fault
2372 2334 * the TOD chip anyway here
2373 2335 */
2374 2336 dosynctodr = 0;
2375 2337
2376 2338 /*
2377 2339 * Set tod to the correct value from hrestime
2378 2340 */
2379 2341 tod = hrestime.tv_sec;
2380 2342 }
2381 2343
2382 2344 prev_tod = tod;
2383 2345 prev_tick = tick;
2384 2346 return (tod);
2385 2347 }
2386 2348
2387 2349 static void
2388 2350 calcloadavg(int nrun, uint64_t *hp_ave)
2389 2351 {
2390 2352 static int64_t f[3] = { 135, 27, 9 };
2391 2353 uint_t i;
2392 2354 int64_t q, r;
2393 2355
2394 2356 /*
2395 2357 * Compute load average over the last 1, 5, and 15 minutes
2396 2358 * (60, 300, and 900 seconds). The constants in f[3] are for
2397 2359 * exponential decay:
2398 2360 * (1 - exp(-1/60)) << 13 = 135,
2399 2361 * (1 - exp(-1/300)) << 13 = 27,
2400 2362 * (1 - exp(-1/900)) << 13 = 9.
2401 2363 */
2402 2364
2403 2365 /*
2404 2366 * a little hoop-jumping to avoid integer overflow
2405 2367 */
2406 2368 for (i = 0; i < 3; i++) {
2407 2369 q = (hp_ave[i] >> 16) << 7;
2408 2370 r = (hp_ave[i] & 0xffff) << 7;
2409 2371 hp_ave[i] += ((nrun - q) * f[i] - ((r * f[i]) >> 16)) >> 4;
2410 2372 }
2411 2373 }
2412 2374
2413 2375 /*
2414 2376 * lbolt_hybrid() is used by ddi_get_lbolt() and ddi_get_lbolt64() to
2415 2377 * calculate the value of lbolt according to the current mode. In the event
2416 2378 * driven mode (the default), lbolt is calculated by dividing the current hires
2417 2379 * time by the number of nanoseconds per clock tick. In the cyclic driven mode
2418 2380 * an internal variable is incremented at each firing of the lbolt cyclic
2419 2381 * and returned by lbolt_cyclic_driven().
2420 2382 *
2421 2383 * The system will transition from event to cyclic driven mode when the number
2422 2384 * of calls to lbolt_event_driven() exceeds the (per CPU) threshold within a
2423 2385 * window of time. It does so by reprograming lbolt_cyclic from CY_INFINITY to
2424 2386 * nsec_per_tick. The lbolt cyclic will remain ON while at least one CPU is
2425 2387 * causing enough activity to cross the thresholds.
2426 2388 */
2427 2389 int64_t
2428 2390 lbolt_bootstrap(void)
2429 2391 {
2430 2392 return (0);
2431 2393 }
2432 2394
2433 2395 /* ARGSUSED */
2434 2396 uint_t
2435 2397 lbolt_ev_to_cyclic(caddr_t arg1, caddr_t arg2)
2436 2398 {
2437 2399 hrtime_t ts, exp;
2438 2400 int ret;
2439 2401
2440 2402 ASSERT(lbolt_hybrid != lbolt_cyclic_driven);
2441 2403
2442 2404 kpreempt_disable();
2443 2405
2444 2406 ts = gethrtime();
2445 2407 lb_info->lbi_internal = (ts/nsec_per_tick);
2446 2408
2447 2409 /*
2448 2410 * Align the next expiration to a clock tick boundary.
2449 2411 */
2450 2412 exp = ts + nsec_per_tick - 1;
2451 2413 exp = (exp/nsec_per_tick) * nsec_per_tick;
2452 2414
2453 2415 ret = cyclic_reprogram(lb_info->id.lbi_cyclic_id, exp);
2454 2416 ASSERT(ret);
2455 2417
2456 2418 lbolt_hybrid = lbolt_cyclic_driven;
2457 2419 lb_info->lbi_cyc_deactivate = B_FALSE;
2458 2420 lb_info->lbi_cyc_deac_start = lb_info->lbi_internal;
2459 2421
2460 2422 kpreempt_enable();
2461 2423
2462 2424 ret = atomic_dec_32_nv(&lb_info->lbi_token);
2463 2425 ASSERT(ret == 0);
2464 2426
2465 2427 return (1);
2466 2428 }
2467 2429
2468 2430 int64_t
2469 2431 lbolt_event_driven(void)
2470 2432 {
2471 2433 hrtime_t ts;
2472 2434 int64_t lb;
2473 2435 int ret, cpu = CPU->cpu_seqid;
2474 2436
2475 2437 ts = gethrtime();
2476 2438 ASSERT(ts > 0);
2477 2439
2478 2440 ASSERT(nsec_per_tick > 0);
2479 2441 lb = (ts/nsec_per_tick);
2480 2442
2481 2443 /*
2482 2444 * Switch to cyclic mode if the number of calls to this routine
2483 2445 * has reached the threshold within the interval.
2484 2446 */
2485 2447 if ((lb - lb_cpu[cpu].lbc_cnt_start) < lb_info->lbi_thresh_interval) {
2486 2448
2487 2449 if (--lb_cpu[cpu].lbc_counter == 0) {
2488 2450 /*
2489 2451 * Reached the threshold within the interval, reset
2490 2452 * the usage statistics.
2491 2453 */
2492 2454 lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2493 2455 lb_cpu[cpu].lbc_cnt_start = lb;
2494 2456
2495 2457 /*
2496 2458 * Make sure only one thread reprograms the
2497 2459 * lbolt cyclic and changes the mode.
2498 2460 */
2499 2461 if (panicstr == NULL &&
2500 2462 atomic_cas_32(&lb_info->lbi_token, 0, 1) == 0) {
2501 2463
2502 2464 if (lbolt_hybrid == lbolt_cyclic_driven) {
2503 2465 ret = atomic_dec_32_nv(
2504 2466 &lb_info->lbi_token);
2505 2467 ASSERT(ret == 0);
2506 2468 } else {
2507 2469 lbolt_softint_post();
2508 2470 }
2509 2471 }
2510 2472 }
2511 2473 } else {
2512 2474 /*
2513 2475 * Exceeded the interval, reset the usage statistics.
2514 2476 */
2515 2477 lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2516 2478 lb_cpu[cpu].lbc_cnt_start = lb;
2517 2479 }
2518 2480
2519 2481 ASSERT(lb >= lb_info->lbi_debug_time);
2520 2482
2521 2483 return (lb - lb_info->lbi_debug_time);
2522 2484 }
2523 2485
2524 2486 int64_t
2525 2487 lbolt_cyclic_driven(void)
2526 2488 {
2527 2489 int64_t lb = lb_info->lbi_internal;
2528 2490 int cpu;
2529 2491
2530 2492 /*
2531 2493 * If a CPU has already prevented the lbolt cyclic from deactivating
2532 2494 * itself, don't bother tracking the usage. Otherwise check if we're
2533 2495 * within the interval and how the per CPU counter is doing.
2534 2496 */
2535 2497 if (lb_info->lbi_cyc_deactivate) {
2536 2498 cpu = CPU->cpu_seqid;
2537 2499 if ((lb - lb_cpu[cpu].lbc_cnt_start) <
2538 2500 lb_info->lbi_thresh_interval) {
2539 2501
2540 2502 if (lb_cpu[cpu].lbc_counter == 0)
2541 2503 /*
2542 2504 * Reached the threshold within the interval,
2543 2505 * prevent the lbolt cyclic from turning itself
2544 2506 * off.
2545 2507 */
2546 2508 lb_info->lbi_cyc_deactivate = B_FALSE;
2547 2509 else
2548 2510 lb_cpu[cpu].lbc_counter--;
2549 2511 } else {
2550 2512 /*
2551 2513 * Only reset the usage statistics when we have
2552 2514 * exceeded the interval.
2553 2515 */
2554 2516 lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2555 2517 lb_cpu[cpu].lbc_cnt_start = lb;
2556 2518 }
2557 2519 }
2558 2520
2559 2521 ASSERT(lb >= lb_info->lbi_debug_time);
2560 2522
2561 2523 return (lb - lb_info->lbi_debug_time);
2562 2524 }
2563 2525
2564 2526 /*
2565 2527 * The lbolt_cyclic() routine will fire at a nsec_per_tick interval to satisfy
2566 2528 * performance needs of ddi_get_lbolt() and ddi_get_lbolt64() consumers.
2567 2529 * It is inactive by default, and will be activated when switching from event
2568 2530 * to cyclic driven lbolt. The cyclic will turn itself off unless signaled
2569 2531 * by lbolt_cyclic_driven().
2570 2532 */
2571 2533 static void
2572 2534 lbolt_cyclic(void)
2573 2535 {
2574 2536 int ret;
2575 2537
2576 2538 lb_info->lbi_internal++;
2577 2539
2578 2540 if (!lbolt_cyc_only) {
2579 2541
2580 2542 if (lb_info->lbi_cyc_deactivate) {
2581 2543 /*
2582 2544 * Switching from cyclic to event driven mode.
2583 2545 */
2584 2546 if (panicstr == NULL &&
2585 2547 atomic_cas_32(&lb_info->lbi_token, 0, 1) == 0) {
2586 2548
2587 2549 if (lbolt_hybrid == lbolt_event_driven) {
2588 2550 ret = atomic_dec_32_nv(
2589 2551 &lb_info->lbi_token);
2590 2552 ASSERT(ret == 0);
2591 2553 return;
2592 2554 }
2593 2555
2594 2556 kpreempt_disable();
2595 2557
2596 2558 lbolt_hybrid = lbolt_event_driven;
2597 2559 ret = cyclic_reprogram(
2598 2560 lb_info->id.lbi_cyclic_id,
2599 2561 CY_INFINITY);
2600 2562 ASSERT(ret);
2601 2563
2602 2564 kpreempt_enable();
2603 2565
2604 2566 ret = atomic_dec_32_nv(&lb_info->lbi_token);
2605 2567 ASSERT(ret == 0);
2606 2568 }
2607 2569 }
2608 2570
2609 2571 /*
2610 2572 * The lbolt cyclic should not try to deactivate itself before
2611 2573 * the sampling period has elapsed.
2612 2574 */
2613 2575 if (lb_info->lbi_internal - lb_info->lbi_cyc_deac_start >=
2614 2576 lb_info->lbi_thresh_interval) {
2615 2577 lb_info->lbi_cyc_deactivate = B_TRUE;
2616 2578 lb_info->lbi_cyc_deac_start = lb_info->lbi_internal;
2617 2579 }
2618 2580 }
2619 2581 }
2620 2582
2621 2583 /*
2622 2584 * Since the lbolt service was historically cyclic driven, it must be 'stopped'
2623 2585 * when the system drops into the kernel debugger. lbolt_debug_entry() is
2624 2586 * called by the KDI system claim callbacks to record a hires timestamp at
2625 2587 * debug enter time. lbolt_debug_return() is called by the sistem release
2626 2588 * callbacks to account for the time spent in the debugger. The value is then
2627 2589 * accumulated in the lb_info structure and used by lbolt_event_driven() and
2628 2590 * lbolt_cyclic_driven(), as well as the mdb_get_lbolt() routine.
2629 2591 */
2630 2592 void
2631 2593 lbolt_debug_entry(void)
2632 2594 {
2633 2595 if (lbolt_hybrid != lbolt_bootstrap) {
2634 2596 ASSERT(lb_info != NULL);
2635 2597 lb_info->lbi_debug_ts = gethrtime();
2636 2598 }
2637 2599 }
2638 2600
2639 2601 /*
2640 2602 * Calculate the time spent in the debugger and add it to the lbolt info
2641 2603 * structure. We also update the internal lbolt value in case we were in
2642 2604 * cyclic driven mode going in.
2643 2605 */
2644 2606 void
2645 2607 lbolt_debug_return(void)
2646 2608 {
2647 2609 hrtime_t ts;
2648 2610
2649 2611 if (lbolt_hybrid != lbolt_bootstrap) {
2650 2612 ASSERT(lb_info != NULL);
2651 2613 ASSERT(nsec_per_tick > 0);
2652 2614
2653 2615 ts = gethrtime();
2654 2616 lb_info->lbi_internal = (ts/nsec_per_tick);
2655 2617 lb_info->lbi_debug_time +=
2656 2618 ((ts - lb_info->lbi_debug_ts)/nsec_per_tick);
2657 2619
2658 2620 lb_info->lbi_debug_ts = 0;
2659 2621 }
2660 2622 }
|
↓ open down ↓ |
1709 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX