| 1 | /*- |
| 2 | * Copyright (c) 1997, 1998 Poul-Henning Kamp <phk@FreeBSD.org> |
| 3 | * Copyright (c) 1982, 1986, 1991, 1993 |
| 4 | * The Regents of the University of California. All rights reserved. |
| 5 | * (c) UNIX System Laboratories, Inc. |
| 6 | * All or some portions of this file are derived from material licensed |
| 7 | * to the University of California by American Telephone and Telegraph |
| 8 | * Co. or Unix System Laboratories, Inc. and are reproduced herein with |
| 9 | * the permission of UNIX System Laboratories, Inc. |
| 10 | * |
| 11 | * Redistribution and use in source and binary forms, with or without |
| 12 | * modification, are permitted provided that the following conditions |
| 13 | * are met: |
| 14 | * 1. Redistributions of source code must retain the above copyright |
| 15 | * notice, this list of conditions and the following disclaimer. |
| 16 | * 2. Redistributions in binary form must reproduce the above copyright |
| 17 | * notice, this list of conditions and the following disclaimer in the |
| 18 | * documentation and/or other materials provided with the distribution. |
| 19 | * 3. All advertising materials mentioning features or use of this software |
| 20 | * must display the following acknowledgement: |
| 21 | * This product includes software developed by the University of |
| 22 | * California, Berkeley and its contributors. |
| 23 | * 4. Neither the name of the University nor the names of its contributors |
| 24 | * may be used to endorse or promote products derived from this software |
| 25 | * without specific prior written permission. |
| 26 | * |
| 27 | * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND |
| 28 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 29 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 30 | * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE |
| 31 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 32 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
| 33 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
| 34 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
| 35 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
| 36 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 37 | * SUCH DAMAGE. |
| 38 | * |
| 39 | * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 |
| 40 | * $FreeBSD: src/sys/kern/kern_clock.c,v 1.105.2.10 2002/10/17 13:19:40 maxim Exp $ |
| 41 | * $DragonFly: src/sys/kern/kern_clock.c,v 1.7 2003/07/10 04:47:54 dillon Exp $ |
| 42 | */ |
| 43 | |
| 44 | #include "opt_ntp.h" |
| 45 | |
| 46 | #include <sys/param.h> |
| 47 | #include <sys/systm.h> |
| 48 | #include <sys/dkstat.h> |
| 49 | #include <sys/callout.h> |
| 50 | #include <sys/kernel.h> |
| 51 | #include <sys/proc.h> |
| 52 | #include <sys/malloc.h> |
| 53 | #include <sys/resourcevar.h> |
| 54 | #include <sys/signalvar.h> |
| 55 | #include <sys/timex.h> |
| 56 | #include <sys/timepps.h> |
| 57 | #include <vm/vm.h> |
| 58 | #include <sys/lock.h> |
| 59 | #include <vm/pmap.h> |
| 60 | #include <vm/vm_map.h> |
| 61 | #include <sys/sysctl.h> |
| 62 | |
| 63 | #include <machine/cpu.h> |
| 64 | #include <machine/limits.h> |
| 65 | #include <machine/smp.h> |
| 66 | |
| 67 | #ifdef GPROF |
| 68 | #include <sys/gmon.h> |
| 69 | #endif |
| 70 | |
| 71 | #ifdef DEVICE_POLLING |
| 72 | extern void init_device_poll(void); |
| 73 | extern void hardclock_device_poll(void); |
| 74 | #endif /* DEVICE_POLLING */ |
| 75 | |
| 76 | /* |
| 77 | * Number of timecounters used to implement stable storage |
| 78 | */ |
| 79 | #ifndef NTIMECOUNTER |
| 80 | #define NTIMECOUNTER 5 |
| 81 | #endif |
| 82 | |
| 83 | static MALLOC_DEFINE(M_TIMECOUNTER, "timecounter", |
| 84 | "Timecounter stable storage"); |
| 85 | |
| 86 | static void initclocks __P((void *dummy)); |
| 87 | SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) |
| 88 | |
| 89 | static void tco_forward __P((int force)); |
| 90 | static void tco_setscales __P((struct timecounter *tc)); |
| 91 | static __inline unsigned tco_delta __P((struct timecounter *tc)); |
| 92 | |
| 93 | /* Some of these don't belong here, but it's easiest to concentrate them. */ |
| 94 | long cp_time[CPUSTATES]; |
| 95 | |
| 96 | SYSCTL_OPAQUE(_kern, OID_AUTO, cp_time, CTLFLAG_RD, &cp_time, sizeof(cp_time), |
| 97 | "LU", "CPU time statistics"); |
| 98 | |
| 99 | long tk_cancc; |
| 100 | long tk_nin; |
| 101 | long tk_nout; |
| 102 | long tk_rawcc; |
| 103 | |
| 104 | time_t time_second; |
| 105 | |
| 106 | struct timeval boottime; |
| 107 | SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD, |
| 108 | &boottime, timeval, "System boottime"); |
| 109 | |
| 110 | /* |
| 111 | * Which update policy to use. |
| 112 | * 0 - every tick, bad hardware may fail with "calcru negative..." |
| 113 | * 1 - more resistent to the above hardware, but less efficient. |
| 114 | */ |
| 115 | static int tco_method; |
| 116 | |
| 117 | /* |
| 118 | * Implement a dummy timecounter which we can use until we get a real one |
| 119 | * in the air. This allows the console and other early stuff to use |
| 120 | * timeservices. |
| 121 | */ |
| 122 | |
| 123 | static unsigned |
| 124 | dummy_get_timecount(struct timecounter *tc) |
| 125 | { |
| 126 | static unsigned now; |
| 127 | return (++now); |
| 128 | } |
| 129 | |
| 130 | static struct timecounter dummy_timecounter = { |
| 131 | dummy_get_timecount, |
| 132 | 0, |
| 133 | ~0u, |
| 134 | 1000000, |
| 135 | "dummy" |
| 136 | }; |
| 137 | |
| 138 | struct timecounter *timecounter = &dummy_timecounter; |
| 139 | |
| 140 | /* |
| 141 | * Clock handling routines. |
| 142 | * |
| 143 | * This code is written to operate with two timers that run independently of |
| 144 | * each other. |
| 145 | * |
| 146 | * The main timer, running hz times per second, is used to trigger interval |
| 147 | * timers, timeouts and rescheduling as needed. |
| 148 | * |
| 149 | * The second timer handles kernel and user profiling, |
| 150 | * and does resource use estimation. If the second timer is programmable, |
| 151 | * it is randomized to avoid aliasing between the two clocks. For example, |
| 152 | * the randomization prevents an adversary from always giving up the cpu |
| 153 | * just before its quantum expires. Otherwise, it would never accumulate |
| 154 | * cpu ticks. The mean frequency of the second timer is stathz. |
| 155 | * |
| 156 | * If no second timer exists, stathz will be zero; in this case we drive |
| 157 | * profiling and statistics off the main clock. This WILL NOT be accurate; |
| 158 | * do not do it unless absolutely necessary. |
| 159 | * |
| 160 | * The statistics clock may (or may not) be run at a higher rate while |
| 161 | * profiling. This profile clock runs at profhz. We require that profhz |
| 162 | * be an integral multiple of stathz. |
| 163 | * |
| 164 | * If the statistics clock is running fast, it must be divided by the ratio |
| 165 | * profhz/stathz for statistics. (For profiling, every tick counts.) |
| 166 | * |
| 167 | * Time-of-day is maintained using a "timecounter", which may or may |
| 168 | * not be related to the hardware generating the above mentioned |
| 169 | * interrupts. |
| 170 | */ |
| 171 | |
| 172 | int stathz; |
| 173 | int profhz; |
| 174 | static int profprocs; |
| 175 | int ticks; |
| 176 | static int psdiv, pscnt; /* prof => stat divider */ |
| 177 | int psratio; /* ratio: prof / stat */ |
| 178 | |
| 179 | /* |
| 180 | * Initialize clock frequencies and start both clocks running. |
| 181 | */ |
| 182 | /* ARGSUSED*/ |
| 183 | static void |
| 184 | initclocks(dummy) |
| 185 | void *dummy; |
| 186 | { |
| 187 | register int i; |
| 188 | |
| 189 | /* |
| 190 | * Set divisors to 1 (normal case) and let the machine-specific |
| 191 | * code do its bit. |
| 192 | */ |
| 193 | psdiv = pscnt = 1; |
| 194 | cpu_initclocks(); |
| 195 | |
| 196 | #ifdef DEVICE_POLLING |
| 197 | init_device_poll(); |
| 198 | #endif |
| 199 | |
| 200 | /* |
| 201 | * Compute profhz/stathz, and fix profhz if needed. |
| 202 | */ |
| 203 | i = stathz ? stathz : hz; |
| 204 | if (profhz == 0) |
| 205 | profhz = i; |
| 206 | psratio = profhz / i; |
| 207 | } |
| 208 | |
| 209 | /* |
| 210 | * The real-time timer, interrupting hz times per second. This is implemented |
| 211 | * as a FAST interrupt so it is in the context of the thread it interrupted, |
| 212 | * and not in an interrupt thread. YYY needs help. |
| 213 | */ |
| 214 | void |
| 215 | hardclock(frame) |
| 216 | register struct clockframe *frame; |
| 217 | { |
| 218 | register struct proc *p; |
| 219 | |
| 220 | p = curproc; |
| 221 | if (p) { |
| 222 | register struct pstats *pstats; |
| 223 | |
| 224 | /* |
| 225 | * Run current process's virtual and profile time, as needed. |
| 226 | */ |
| 227 | pstats = p->p_stats; |
| 228 | if (CLKF_USERMODE(frame) && |
| 229 | timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && |
| 230 | itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) |
| 231 | psignal(p, SIGVTALRM); |
| 232 | if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) && |
| 233 | itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) |
| 234 | psignal(p, SIGPROF); |
| 235 | } |
| 236 | |
| 237 | #if defined(SMP) && defined(BETTER_CLOCK) |
| 238 | forward_hardclock(pscnt); |
| 239 | #endif |
| 240 | |
| 241 | /* |
| 242 | * If no separate statistics clock is available, run it from here. |
| 243 | */ |
| 244 | if (stathz == 0) |
| 245 | statclock(frame); |
| 246 | |
| 247 | tco_forward(0); |
| 248 | ticks++; |
| 249 | |
| 250 | #ifdef DEVICE_POLLING |
| 251 | hardclock_device_poll(); /* this is very short and quick */ |
| 252 | #endif /* DEVICE_POLLING */ |
| 253 | |
| 254 | /* |
| 255 | * Process callouts at a very low cpu priority, so we don't keep the |
| 256 | * relatively high clock interrupt priority any longer than necessary. |
| 257 | */ |
| 258 | if (TAILQ_FIRST(&callwheel[ticks & callwheelmask]) != NULL) { |
| 259 | setsoftclock(); |
| 260 | } else if (softticks + 1 == ticks) { |
| 261 | ++softticks; |
| 262 | } |
| 263 | } |
| 264 | |
| 265 | /* |
| 266 | * Compute number of ticks in the specified amount of time. |
| 267 | */ |
| 268 | int |
| 269 | tvtohz(tv) |
| 270 | struct timeval *tv; |
| 271 | { |
| 272 | register unsigned long ticks; |
| 273 | register long sec, usec; |
| 274 | |
| 275 | /* |
| 276 | * If the number of usecs in the whole seconds part of the time |
| 277 | * difference fits in a long, then the total number of usecs will |
| 278 | * fit in an unsigned long. Compute the total and convert it to |
| 279 | * ticks, rounding up and adding 1 to allow for the current tick |
| 280 | * to expire. Rounding also depends on unsigned long arithmetic |
| 281 | * to avoid overflow. |
| 282 | * |
| 283 | * Otherwise, if the number of ticks in the whole seconds part of |
| 284 | * the time difference fits in a long, then convert the parts to |
| 285 | * ticks separately and add, using similar rounding methods and |
| 286 | * overflow avoidance. This method would work in the previous |
| 287 | * case but it is slightly slower and assumes that hz is integral. |
| 288 | * |
| 289 | * Otherwise, round the time difference down to the maximum |
| 290 | * representable value. |
| 291 | * |
| 292 | * If ints have 32 bits, then the maximum value for any timeout in |
| 293 | * 10ms ticks is 248 days. |
| 294 | */ |
| 295 | sec = tv->tv_sec; |
| 296 | usec = tv->tv_usec; |
| 297 | if (usec < 0) { |
| 298 | sec--; |
| 299 | usec += 1000000; |
| 300 | } |
| 301 | if (sec < 0) { |
| 302 | #ifdef DIAGNOSTIC |
| 303 | if (usec > 0) { |
| 304 | sec++; |
| 305 | usec -= 1000000; |
| 306 | } |
| 307 | printf("tvotohz: negative time difference %ld sec %ld usec\n", |
| 308 | sec, usec); |
| 309 | #endif |
| 310 | ticks = 1; |
| 311 | } else if (sec <= LONG_MAX / 1000000) |
| 312 | ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) |
| 313 | / tick + 1; |
| 314 | else if (sec <= LONG_MAX / hz) |
| 315 | ticks = sec * hz |
| 316 | + ((unsigned long)usec + (tick - 1)) / tick + 1; |
| 317 | else |
| 318 | ticks = LONG_MAX; |
| 319 | if (ticks > INT_MAX) |
| 320 | ticks = INT_MAX; |
| 321 | return ((int)ticks); |
| 322 | } |
| 323 | |
| 324 | /* |
| 325 | * Start profiling on a process. |
| 326 | * |
| 327 | * Kernel profiling passes proc0 which never exits and hence |
| 328 | * keeps the profile clock running constantly. |
| 329 | */ |
| 330 | void |
| 331 | startprofclock(p) |
| 332 | register struct proc *p; |
| 333 | { |
| 334 | int s; |
| 335 | |
| 336 | if ((p->p_flag & P_PROFIL) == 0) { |
| 337 | p->p_flag |= P_PROFIL; |
| 338 | if (++profprocs == 1 && stathz != 0) { |
| 339 | s = splstatclock(); |
| 340 | psdiv = pscnt = psratio; |
| 341 | setstatclockrate(profhz); |
| 342 | splx(s); |
| 343 | } |
| 344 | } |
| 345 | } |
| 346 | |
| 347 | /* |
| 348 | * Stop profiling on a process. |
| 349 | */ |
| 350 | void |
| 351 | stopprofclock(p) |
| 352 | register struct proc *p; |
| 353 | { |
| 354 | int s; |
| 355 | |
| 356 | if (p->p_flag & P_PROFIL) { |
| 357 | p->p_flag &= ~P_PROFIL; |
| 358 | if (--profprocs == 0 && stathz != 0) { |
| 359 | s = splstatclock(); |
| 360 | psdiv = pscnt = 1; |
| 361 | setstatclockrate(stathz); |
| 362 | splx(s); |
| 363 | } |
| 364 | } |
| 365 | } |
| 366 | |
| 367 | /* |
| 368 | * Statistics clock. Grab profile sample, and if divider reaches 0, |
| 369 | * do process and kernel statistics. Most of the statistics are only |
| 370 | * used by user-level statistics programs. The main exceptions are |
| 371 | * p->p_uticks, p->p_sticks, p->p_iticks, and p->p_estcpu. |
| 372 | * |
| 373 | * The statclock should be called from an exclusive, fast interrupt, |
| 374 | * so the context should be the thread/process that got interrupted and |
| 375 | * not an interrupt thread. |
| 376 | */ |
| 377 | void |
| 378 | statclock(frame) |
| 379 | register struct clockframe *frame; |
| 380 | { |
| 381 | #ifdef GPROF |
| 382 | register struct gmonparam *g; |
| 383 | int i; |
| 384 | #endif |
| 385 | thread_t td; |
| 386 | struct pstats *pstats; |
| 387 | long rss; |
| 388 | struct rusage *ru; |
| 389 | struct vmspace *vm; |
| 390 | struct proc *p; |
| 391 | int bump; |
| 392 | struct timeval tv; |
| 393 | struct timeval *stv; |
| 394 | |
| 395 | /* |
| 396 | * How big was our timeslice relative to the last time |
| 397 | */ |
| 398 | microuptime(&tv); |
| 399 | stv = &mycpu->gd_stattv; |
| 400 | if (stv->tv_sec == 0) { |
| 401 | bump = 1; |
| 402 | } else { |
| 403 | bump = tv.tv_usec - stv->tv_usec + |
| 404 | (tv.tv_sec - stv->tv_sec) * 1000000; |
| 405 | if (bump < 0) |
| 406 | bump = 0; |
| 407 | if (bump > 1000000) |
| 408 | bump = 1000000; |
| 409 | } |
| 410 | *stv = tv; |
| 411 | |
| 412 | td = curthread; |
| 413 | p = td->td_proc; |
| 414 | |
| 415 | if (CLKF_USERMODE(frame)) { |
| 416 | /* |
| 417 | * Came from userland, handle user time and deal with |
| 418 | * possible process. |
| 419 | */ |
| 420 | if (p && (p->p_flag & P_PROFIL)) |
| 421 | addupc_intr(p, CLKF_PC(frame), 1); |
| 422 | #if defined(SMP) && defined(BETTER_CLOCK) |
| 423 | if (stathz != 0) |
| 424 | forward_statclock(pscnt); |
| 425 | #endif |
| 426 | td->td_uticks += bump; |
| 427 | if (--pscnt > 0) |
| 428 | return; |
| 429 | |
| 430 | /* |
| 431 | * Charge the time as appropriate |
| 432 | */ |
| 433 | if (p && p->p_nice > NZERO) |
| 434 | ++cp_time[CP_NICE]; |
| 435 | else |
| 436 | ++cp_time[CP_USER]; |
| 437 | } else { |
| 438 | #ifdef GPROF |
| 439 | /* |
| 440 | * Kernel statistics are just like addupc_intr, only easier. |
| 441 | */ |
| 442 | g = &_gmonparam; |
| 443 | if (g->state == GMON_PROF_ON) { |
| 444 | i = CLKF_PC(frame) - g->lowpc; |
| 445 | if (i < g->textsize) { |
| 446 | i /= HISTFRACTION * sizeof(*g->kcount); |
| 447 | g->kcount[i]++; |
| 448 | } |
| 449 | } |
| 450 | #endif |
| 451 | #if defined(SMP) && defined(BETTER_CLOCK) |
| 452 | if (stathz != 0) |
| 453 | forward_statclock(pscnt); |
| 454 | #endif |
| 455 | /* |
| 456 | * Came from kernel mode, so we were: |
| 457 | * - handling an interrupt, |
| 458 | * - doing syscall or trap work on behalf of the current |
| 459 | * user process, or |
| 460 | * - spinning in the idle loop. |
| 461 | * Whichever it is, charge the time as appropriate. |
| 462 | * Note that we charge interrupts to the current process, |
| 463 | * regardless of whether they are ``for'' that process, |
| 464 | * so that we know how much of its real time was spent |
| 465 | * in ``non-process'' (i.e., interrupt) work. |
| 466 | */ |
| 467 | if (CLKF_INTR(frame)) |
| 468 | td->td_iticks += bump; |
| 469 | else |
| 470 | td->td_sticks += bump; |
| 471 | |
| 472 | if (--pscnt > 0) |
| 473 | return; |
| 474 | |
| 475 | if (CLKF_INTR(frame)) { |
| 476 | cp_time[CP_INTR]++; |
| 477 | } else { |
| 478 | if (td == &mycpu->gd_idlethread) |
| 479 | ++cp_time[CP_IDLE]; |
| 480 | else |
| 481 | ++cp_time[CP_SYS]; |
| 482 | } |
| 483 | } |
| 484 | pscnt = psdiv; |
| 485 | |
| 486 | if (p != NULL) { |
| 487 | schedclock(p); |
| 488 | |
| 489 | /* Update resource usage integrals and maximums. */ |
| 490 | if ((pstats = p->p_stats) != NULL && |
| 491 | (ru = &pstats->p_ru) != NULL && |
| 492 | (vm = p->p_vmspace) != NULL) { |
| 493 | ru->ru_ixrss += pgtok(vm->vm_tsize); |
| 494 | ru->ru_idrss += pgtok(vm->vm_dsize); |
| 495 | ru->ru_isrss += pgtok(vm->vm_ssize); |
| 496 | rss = pgtok(vmspace_resident_count(vm)); |
| 497 | if (ru->ru_maxrss < rss) |
| 498 | ru->ru_maxrss = rss; |
| 499 | } |
| 500 | } |
| 501 | } |
| 502 | |
| 503 | /* |
| 504 | * Return information about system clocks. |
| 505 | */ |
| 506 | static int |
| 507 | sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) |
| 508 | { |
| 509 | struct clockinfo clkinfo; |
| 510 | /* |
| 511 | * Construct clockinfo structure. |
| 512 | */ |
| 513 | clkinfo.hz = hz; |
| 514 | clkinfo.tick = tick; |
| 515 | clkinfo.tickadj = tickadj; |
| 516 | clkinfo.profhz = profhz; |
| 517 | clkinfo.stathz = stathz ? stathz : hz; |
| 518 | return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); |
| 519 | } |
| 520 | |
| 521 | SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, |
| 522 | 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); |
| 523 | |
| 524 | static __inline unsigned |
| 525 | tco_delta(struct timecounter *tc) |
| 526 | { |
| 527 | |
| 528 | return ((tc->tc_get_timecount(tc) - tc->tc_offset_count) & |
| 529 | tc->tc_counter_mask); |
| 530 | } |
| 531 | |
| 532 | /* |
| 533 | * We have eight functions for looking at the clock, four for |
| 534 | * microseconds and four for nanoseconds. For each there is fast |
| 535 | * but less precise version "get{nano|micro}[up]time" which will |
| 536 | * return a time which is up to 1/HZ previous to the call, whereas |
| 537 | * the raw version "{nano|micro}[up]time" will return a timestamp |
| 538 | * which is as precise as possible. The "up" variants return the |
| 539 | * time relative to system boot, these are well suited for time |
| 540 | * interval measurements. |
| 541 | */ |
| 542 | |
| 543 | void |
| 544 | getmicrotime(struct timeval *tvp) |
| 545 | { |
| 546 | struct timecounter *tc; |
| 547 | |
| 548 | if (!tco_method) { |
| 549 | tc = timecounter; |
| 550 | *tvp = tc->tc_microtime; |
| 551 | } else { |
| 552 | microtime(tvp); |
| 553 | } |
| 554 | } |
| 555 | |
| 556 | void |
| 557 | getnanotime(struct timespec *tsp) |
| 558 | { |
| 559 | struct timecounter *tc; |
| 560 | |
| 561 | if (!tco_method) { |
| 562 | tc = timecounter; |
| 563 | *tsp = tc->tc_nanotime; |
| 564 | } else { |
| 565 | nanotime(tsp); |
| 566 | } |
| 567 | } |
| 568 | |
| 569 | void |
| 570 | microtime(struct timeval *tv) |
| 571 | { |
| 572 | struct timecounter *tc; |
| 573 | |
| 574 | tc = timecounter; |
| 575 | tv->tv_sec = tc->tc_offset_sec; |
| 576 | tv->tv_usec = tc->tc_offset_micro; |
| 577 | tv->tv_usec += ((u_int64_t)tco_delta(tc) * tc->tc_scale_micro) >> 32; |
| 578 | tv->tv_usec += boottime.tv_usec; |
| 579 | tv->tv_sec += boottime.tv_sec; |
| 580 | while (tv->tv_usec < 0) { |
| 581 | tv->tv_usec += 1000000; |
| 582 | if (tv->tv_sec > 0) |
| 583 | tv->tv_sec--; |
| 584 | } |
| 585 | while (tv->tv_usec >= 1000000) { |
| 586 | tv->tv_usec -= 1000000; |
| 587 | tv->tv_sec++; |
| 588 | } |
| 589 | } |
| 590 | |
| 591 | void |
| 592 | nanotime(struct timespec *ts) |
| 593 | { |
| 594 | unsigned count; |
| 595 | u_int64_t delta; |
| 596 | struct timecounter *tc; |
| 597 | |
| 598 | tc = timecounter; |
| 599 | ts->tv_sec = tc->tc_offset_sec; |
| 600 | count = tco_delta(tc); |
| 601 | delta = tc->tc_offset_nano; |
| 602 | delta += ((u_int64_t)count * tc->tc_scale_nano_f); |
| 603 | delta >>= 32; |
| 604 | delta += ((u_int64_t)count * tc->tc_scale_nano_i); |
| 605 | delta += boottime.tv_usec * 1000; |
| 606 | ts->tv_sec += boottime.tv_sec; |
| 607 | while (delta < 0) { |
| 608 | delta += 1000000000; |
| 609 | if (ts->tv_sec > 0) |
| 610 | ts->tv_sec--; |
| 611 | } |
| 612 | while (delta >= 1000000000) { |
| 613 | delta -= 1000000000; |
| 614 | ts->tv_sec++; |
| 615 | } |
| 616 | ts->tv_nsec = delta; |
| 617 | } |
| 618 | |
| 619 | void |
| 620 | getmicrouptime(struct timeval *tvp) |
| 621 | { |
| 622 | struct timecounter *tc; |
| 623 | |
| 624 | if (!tco_method) { |
| 625 | tc = timecounter; |
| 626 | tvp->tv_sec = tc->tc_offset_sec; |
| 627 | tvp->tv_usec = tc->tc_offset_micro; |
| 628 | } else { |
| 629 | microuptime(tvp); |
| 630 | } |
| 631 | } |
| 632 | |
| 633 | void |
| 634 | getnanouptime(struct timespec *tsp) |
| 635 | { |
| 636 | struct timecounter *tc; |
| 637 | |
| 638 | if (!tco_method) { |
| 639 | tc = timecounter; |
| 640 | tsp->tv_sec = tc->tc_offset_sec; |
| 641 | tsp->tv_nsec = tc->tc_offset_nano >> 32; |
| 642 | } else { |
| 643 | nanouptime(tsp); |
| 644 | } |
| 645 | } |
| 646 | |
| 647 | void |
| 648 | microuptime(struct timeval *tv) |
| 649 | { |
| 650 | struct timecounter *tc; |
| 651 | |
| 652 | tc = timecounter; |
| 653 | tv->tv_sec = tc->tc_offset_sec; |
| 654 | tv->tv_usec = tc->tc_offset_micro; |
| 655 | tv->tv_usec += ((u_int64_t)tco_delta(tc) * tc->tc_scale_micro) >> 32; |
| 656 | while (tv->tv_usec < 0) { |
| 657 | tv->tv_usec += 1000000; |
| 658 | if (tv->tv_sec > 0) |
| 659 | tv->tv_sec--; |
| 660 | } |
| 661 | while (tv->tv_usec >= 1000000) { |
| 662 | tv->tv_usec -= 1000000; |
| 663 | tv->tv_sec++; |
| 664 | } |
| 665 | } |
| 666 | |
| 667 | void |
| 668 | nanouptime(struct timespec *ts) |
| 669 | { |
| 670 | unsigned count; |
| 671 | u_int64_t delta; |
| 672 | struct timecounter *tc; |
| 673 | |
| 674 | tc = timecounter; |
| 675 | ts->tv_sec = tc->tc_offset_sec; |
| 676 | count = tco_delta(tc); |
| 677 | delta = tc->tc_offset_nano; |
| 678 | delta += ((u_int64_t)count * tc->tc_scale_nano_f); |
| 679 | delta >>= 32; |
| 680 | delta += ((u_int64_t)count * tc->tc_scale_nano_i); |
| 681 | while (delta < 0) { |
| 682 | delta += 1000000000; |
| 683 | if (ts->tv_sec > 0) |
| 684 | ts->tv_sec--; |
| 685 | } |
| 686 | while (delta >= 1000000000) { |
| 687 | delta -= 1000000000; |
| 688 | ts->tv_sec++; |
| 689 | } |
| 690 | ts->tv_nsec = delta; |
| 691 | } |
| 692 | |
| 693 | static void |
| 694 | tco_setscales(struct timecounter *tc) |
| 695 | { |
| 696 | u_int64_t scale; |
| 697 | |
| 698 | scale = 1000000000LL << 32; |
| 699 | scale += tc->tc_adjustment; |
| 700 | scale /= tc->tc_tweak->tc_frequency; |
| 701 | tc->tc_scale_micro = scale / 1000; |
| 702 | tc->tc_scale_nano_f = scale & 0xffffffff; |
| 703 | tc->tc_scale_nano_i = scale >> 32; |
| 704 | } |
| 705 | |
| 706 | void |
| 707 | update_timecounter(struct timecounter *tc) |
| 708 | { |
| 709 | tco_setscales(tc); |
| 710 | } |
| 711 | |
| 712 | void |
| 713 | init_timecounter(struct timecounter *tc) |
| 714 | { |
| 715 | struct timespec ts1; |
| 716 | struct timecounter *t1, *t2, *t3; |
| 717 | unsigned u; |
| 718 | int i; |
| 719 | |
| 720 | u = tc->tc_frequency / tc->tc_counter_mask; |
| 721 | if (u > hz) { |
| 722 | printf("Timecounter \"%s\" frequency %lu Hz" |
| 723 | " -- Insufficient hz, needs at least %u\n", |
| 724 | tc->tc_name, (u_long) tc->tc_frequency, u); |
| 725 | return; |
| 726 | } |
| 727 | |
| 728 | tc->tc_adjustment = 0; |
| 729 | tc->tc_tweak = tc; |
| 730 | tco_setscales(tc); |
| 731 | tc->tc_offset_count = tc->tc_get_timecount(tc); |
| 732 | if (timecounter == &dummy_timecounter) |
| 733 | tc->tc_avail = tc; |
| 734 | else { |
| 735 | tc->tc_avail = timecounter->tc_tweak->tc_avail; |
| 736 | timecounter->tc_tweak->tc_avail = tc; |
| 737 | } |
| 738 | MALLOC(t1, struct timecounter *, sizeof *t1, M_TIMECOUNTER, M_WAITOK); |
| 739 | tc->tc_other = t1; |
| 740 | *t1 = *tc; |
| 741 | t2 = t1; |
| 742 | for (i = 1; i < NTIMECOUNTER; i++) { |
| 743 | MALLOC(t3, struct timecounter *, sizeof *t3, |
| 744 | M_TIMECOUNTER, M_WAITOK); |
| 745 | *t3 = *tc; |
| 746 | t3->tc_other = t2; |
| 747 | t2 = t3; |
| 748 | } |
| 749 | t1->tc_other = t3; |
| 750 | tc = t1; |
| 751 | |
| 752 | printf("Timecounter \"%s\" frequency %lu Hz\n", |
| 753 | tc->tc_name, (u_long)tc->tc_frequency); |
| 754 | |
| 755 | /* XXX: For now always start using the counter. */ |
| 756 | tc->tc_offset_count = tc->tc_get_timecount(tc); |
| 757 | nanouptime(&ts1); |
| 758 | tc->tc_offset_nano = (u_int64_t)ts1.tv_nsec << 32; |
| 759 | tc->tc_offset_micro = ts1.tv_nsec / 1000; |
| 760 | tc->tc_offset_sec = ts1.tv_sec; |
| 761 | timecounter = tc; |
| 762 | } |
| 763 | |
| 764 | void |
| 765 | set_timecounter(struct timespec *ts) |
| 766 | { |
| 767 | struct timespec ts2; |
| 768 | |
| 769 | nanouptime(&ts2); |
| 770 | boottime.tv_sec = ts->tv_sec - ts2.tv_sec; |
| 771 | boottime.tv_usec = (ts->tv_nsec - ts2.tv_nsec) / 1000; |
| 772 | if (boottime.tv_usec < 0) { |
| 773 | boottime.tv_usec += 1000000; |
| 774 | boottime.tv_sec--; |
| 775 | } |
| 776 | /* fiddle all the little crinkly bits around the fiords... */ |
| 777 | tco_forward(1); |
| 778 | } |
| 779 | |
| 780 | static void |
| 781 | switch_timecounter(struct timecounter *newtc) |
| 782 | { |
| 783 | int s; |
| 784 | struct timecounter *tc; |
| 785 | struct timespec ts; |
| 786 | |
| 787 | s = splclock(); |
| 788 | tc = timecounter; |
| 789 | if (newtc->tc_tweak == tc->tc_tweak) { |
| 790 | splx(s); |
| 791 | return; |
| 792 | } |
| 793 | newtc = newtc->tc_tweak->tc_other; |
| 794 | nanouptime(&ts); |
| 795 | newtc->tc_offset_sec = ts.tv_sec; |
| 796 | newtc->tc_offset_nano = (u_int64_t)ts.tv_nsec << 32; |
| 797 | newtc->tc_offset_micro = ts.tv_nsec / 1000; |
| 798 | newtc->tc_offset_count = newtc->tc_get_timecount(newtc); |
| 799 | tco_setscales(newtc); |
| 800 | timecounter = newtc; |
| 801 | splx(s); |
| 802 | } |
| 803 | |
| 804 | static struct timecounter * |
| 805 | sync_other_counter(void) |
| 806 | { |
| 807 | struct timecounter *tc, *tcn, *tco; |
| 808 | unsigned delta; |
| 809 | |
| 810 | tco = timecounter; |
| 811 | tc = tco->tc_other; |
| 812 | tcn = tc->tc_other; |
| 813 | *tc = *tco; |
| 814 | tc->tc_other = tcn; |
| 815 | delta = tco_delta(tc); |
| 816 | tc->tc_offset_count += delta; |
| 817 | tc->tc_offset_count &= tc->tc_counter_mask; |
| 818 | tc->tc_offset_nano += (u_int64_t)delta * tc->tc_scale_nano_f; |
| 819 | tc->tc_offset_nano += (u_int64_t)delta * tc->tc_scale_nano_i << 32; |
| 820 | return (tc); |
| 821 | } |
| 822 | |
| 823 | static void |
| 824 | tco_forward(int force) |
| 825 | { |
| 826 | struct timecounter *tc, *tco; |
| 827 | struct timeval tvt; |
| 828 | |
| 829 | tco = timecounter; |
| 830 | tc = sync_other_counter(); |
| 831 | /* |
| 832 | * We may be inducing a tiny error here, the tc_poll_pps() may |
| 833 | * process a latched count which happens after the tco_delta() |
| 834 | * in sync_other_counter(), which would extend the previous |
| 835 | * counters parameters into the domain of this new one. |
| 836 | * Since the timewindow is very small for this, the error is |
| 837 | * going to be only a few weenieseconds (as Dave Mills would |
| 838 | * say), so lets just not talk more about it, OK ? |
| 839 | */ |
| 840 | if (tco->tc_poll_pps) |
| 841 | tco->tc_poll_pps(tco); |
| 842 | if (timedelta != 0) { |
| 843 | tvt = boottime; |
| 844 | tvt.tv_usec += tickdelta; |
| 845 | if (tvt.tv_usec >= 1000000) { |
| 846 | tvt.tv_sec++; |
| 847 | tvt.tv_usec -= 1000000; |
| 848 | } else if (tvt.tv_usec < 0) { |
| 849 | tvt.tv_sec--; |
| 850 | tvt.tv_usec += 1000000; |
| 851 | } |
| 852 | boottime = tvt; |
| 853 | timedelta -= tickdelta; |
| 854 | } |
| 855 | |
| 856 | while (tc->tc_offset_nano >= 1000000000ULL << 32) { |
| 857 | tc->tc_offset_nano -= 1000000000ULL << 32; |
| 858 | tc->tc_offset_sec++; |
| 859 | ntp_update_second(tc); /* XXX only needed if xntpd runs */ |
| 860 | tco_setscales(tc); |
| 861 | force++; |
| 862 | } |
| 863 | |
| 864 | if (tco_method && !force) |
| 865 | return; |
| 866 | |
| 867 | tc->tc_offset_micro = (tc->tc_offset_nano / 1000) >> 32; |
| 868 | |
| 869 | /* Figure out the wall-clock time */ |
| 870 | tc->tc_nanotime.tv_sec = tc->tc_offset_sec + boottime.tv_sec; |
| 871 | tc->tc_nanotime.tv_nsec = |
| 872 | (tc->tc_offset_nano >> 32) + boottime.tv_usec * 1000; |
| 873 | tc->tc_microtime.tv_usec = tc->tc_offset_micro + boottime.tv_usec; |
| 874 | while (tc->tc_nanotime.tv_nsec >= 1000000000) { |
| 875 | tc->tc_nanotime.tv_nsec -= 1000000000; |
| 876 | tc->tc_microtime.tv_usec -= 1000000; |
| 877 | tc->tc_nanotime.tv_sec++; |
| 878 | } |
| 879 | time_second = tc->tc_microtime.tv_sec = tc->tc_nanotime.tv_sec; |
| 880 | |
| 881 | timecounter = tc; |
| 882 | } |
| 883 | |
| 884 | SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, ""); |
| 885 | |
| 886 | SYSCTL_INT(_kern_timecounter, OID_AUTO, method, CTLFLAG_RW, &tco_method, 0, |
| 887 | "This variable determines the method used for updating timecounters. " |
| 888 | "If the default algorithm (0) fails with \"calcru negative...\" messages " |
| 889 | "try the alternate algorithm (1) which handles bad hardware better." |
| 890 | |
| 891 | ); |
| 892 | |
| 893 | static int |
| 894 | sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS) |
| 895 | { |
| 896 | char newname[32]; |
| 897 | struct timecounter *newtc, *tc; |
| 898 | int error; |
| 899 | |
| 900 | tc = timecounter->tc_tweak; |
| 901 | strncpy(newname, tc->tc_name, sizeof(newname)); |
| 902 | error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req); |
| 903 | if (error == 0 && req->newptr != NULL && |
| 904 | strcmp(newname, tc->tc_name) != 0) { |
| 905 | for (newtc = tc->tc_avail; newtc != tc; |
| 906 | newtc = newtc->tc_avail) { |
| 907 | if (strcmp(newname, newtc->tc_name) == 0) { |
| 908 | /* Warm up new timecounter. */ |
| 909 | (void)newtc->tc_get_timecount(newtc); |
| 910 | |
| 911 | switch_timecounter(newtc); |
| 912 | return (0); |
| 913 | } |
| 914 | } |
| 915 | return (EINVAL); |
| 916 | } |
| 917 | return (error); |
| 918 | } |
| 919 | |
| 920 | SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW, |
| 921 | 0, 0, sysctl_kern_timecounter_hardware, "A", ""); |
| 922 | |
| 923 | |
| 924 | int |
| 925 | pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) |
| 926 | { |
| 927 | pps_params_t *app; |
| 928 | struct pps_fetch_args *fapi; |
| 929 | #ifdef PPS_SYNC |
| 930 | struct pps_kcbind_args *kapi; |
| 931 | #endif |
| 932 | |
| 933 | switch (cmd) { |
| 934 | case PPS_IOC_CREATE: |
| 935 | return (0); |
| 936 | case PPS_IOC_DESTROY: |
| 937 | return (0); |
| 938 | case PPS_IOC_SETPARAMS: |
| 939 | app = (pps_params_t *)data; |
| 940 | if (app->mode & ~pps->ppscap) |
| 941 | return (EINVAL); |
| 942 | pps->ppsparam = *app; |
| 943 | return (0); |
| 944 | case PPS_IOC_GETPARAMS: |
| 945 | app = (pps_params_t *)data; |
| 946 | *app = pps->ppsparam; |
| 947 | app->api_version = PPS_API_VERS_1; |
| 948 | return (0); |
| 949 | case PPS_IOC_GETCAP: |
| 950 | *(int*)data = pps->ppscap; |
| 951 | return (0); |
| 952 | case PPS_IOC_FETCH: |
| 953 | fapi = (struct pps_fetch_args *)data; |
| 954 | if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC) |
| 955 | return (EINVAL); |
| 956 | if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) |
| 957 | return (EOPNOTSUPP); |
| 958 | pps->ppsinfo.current_mode = pps->ppsparam.mode; |
| 959 | fapi->pps_info_buf = pps->ppsinfo; |
| 960 | return (0); |
| 961 | case PPS_IOC_KCBIND: |
| 962 | #ifdef PPS_SYNC |
| 963 | kapi = (struct pps_kcbind_args *)data; |
| 964 | /* XXX Only root should be able to do this */ |
| 965 | if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC) |
| 966 | return (EINVAL); |
| 967 | if (kapi->kernel_consumer != PPS_KC_HARDPPS) |
| 968 | return (EINVAL); |
| 969 | if (kapi->edge & ~pps->ppscap) |
| 970 | return (EINVAL); |
| 971 | pps->kcmode = kapi->edge; |
| 972 | return (0); |
| 973 | #else |
| 974 | return (EOPNOTSUPP); |
| 975 | #endif |
| 976 | default: |
| 977 | return (ENOTTY); |
| 978 | } |
| 979 | } |
| 980 | |
| 981 | void |
| 982 | pps_init(struct pps_state *pps) |
| 983 | { |
| 984 | pps->ppscap |= PPS_TSFMT_TSPEC; |
| 985 | if (pps->ppscap & PPS_CAPTUREASSERT) |
| 986 | pps->ppscap |= PPS_OFFSETASSERT; |
| 987 | if (pps->ppscap & PPS_CAPTURECLEAR) |
| 988 | pps->ppscap |= PPS_OFFSETCLEAR; |
| 989 | } |
| 990 | |
| 991 | void |
| 992 | pps_event(struct pps_state *pps, struct timecounter *tc, unsigned count, int event) |
| 993 | { |
| 994 | struct timespec ts, *tsp, *osp; |
| 995 | u_int64_t delta; |
| 996 | unsigned tcount, *pcount; |
| 997 | int foff, fhard; |
| 998 | pps_seq_t *pseq; |
| 999 | |
| 1000 | /* Things would be easier with arrays... */ |
| 1001 | if (event == PPS_CAPTUREASSERT) { |
| 1002 | tsp = &pps->ppsinfo.assert_timestamp; |
| 1003 | osp = &pps->ppsparam.assert_offset; |
| 1004 | foff = pps->ppsparam.mode & PPS_OFFSETASSERT; |
| 1005 | fhard = pps->kcmode & PPS_CAPTUREASSERT; |
| 1006 | pcount = &pps->ppscount[0]; |
| 1007 | pseq = &pps->ppsinfo.assert_sequence; |
| 1008 | } else { |
| 1009 | tsp = &pps->ppsinfo.clear_timestamp; |
| 1010 | osp = &pps->ppsparam.clear_offset; |
| 1011 | foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; |
| 1012 | fhard = pps->kcmode & PPS_CAPTURECLEAR; |
| 1013 | pcount = &pps->ppscount[1]; |
| 1014 | pseq = &pps->ppsinfo.clear_sequence; |
| 1015 | } |
| 1016 | |
| 1017 | /* The timecounter changed: bail */ |
| 1018 | if (!pps->ppstc || |
| 1019 | pps->ppstc->tc_name != tc->tc_name || |
| 1020 | tc->tc_name != timecounter->tc_name) { |
| 1021 | pps->ppstc = tc; |
| 1022 | *pcount = count; |
| 1023 | return; |
| 1024 | } |
| 1025 | |
| 1026 | /* Nothing really happened */ |
| 1027 | if (*pcount == count) |
| 1028 | return; |
| 1029 | |
| 1030 | *pcount = count; |
| 1031 | |
| 1032 | /* Convert the count to timespec */ |
| 1033 | ts.tv_sec = tc->tc_offset_sec; |
| 1034 | tcount = count - tc->tc_offset_count; |
| 1035 | tcount &= tc->tc_counter_mask; |
| 1036 | delta = tc->tc_offset_nano; |
| 1037 | delta += ((u_int64_t)tcount * tc->tc_scale_nano_f); |
| 1038 | delta >>= 32; |
| 1039 | delta += ((u_int64_t)tcount * tc->tc_scale_nano_i); |
| 1040 | delta += boottime.tv_usec * 1000; |
| 1041 | ts.tv_sec += boottime.tv_sec; |
| 1042 | while (delta >= 1000000000) { |
| 1043 | delta -= 1000000000; |
| 1044 | ts.tv_sec++; |
| 1045 | } |
| 1046 | ts.tv_nsec = delta; |
| 1047 | |
| 1048 | (*pseq)++; |
| 1049 | *tsp = ts; |
| 1050 | |
| 1051 | if (foff) { |
| 1052 | timespecadd(tsp, osp); |
| 1053 | if (tsp->tv_nsec < 0) { |
| 1054 | tsp->tv_nsec += 1000000000; |
| 1055 | tsp->tv_sec -= 1; |
| 1056 | } |
| 1057 | } |
| 1058 | #ifdef PPS_SYNC |
| 1059 | if (fhard) { |
| 1060 | /* magic, at its best... */ |
| 1061 | tcount = count - pps->ppscount[2]; |
| 1062 | pps->ppscount[2] = count; |
| 1063 | tcount &= tc->tc_counter_mask; |
| 1064 | delta = ((u_int64_t)tcount * tc->tc_tweak->tc_scale_nano_f); |
| 1065 | delta >>= 32; |
| 1066 | delta += ((u_int64_t)tcount * tc->tc_tweak->tc_scale_nano_i); |
| 1067 | hardpps(tsp, delta); |
| 1068 | } |
| 1069 | #endif |
| 1070 | } |