Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Common time routines among all ppc machines. |
| 3 | * |
| 4 | * Written by Cort Dougan (cort@cs.nmt.edu) to merge |
| 5 | * Paul Mackerras' version and mine for PReP and Pmac. |
| 6 | * MPC8xx/MBX changes by Dan Malek (dmalek@jlc.net). |
| 7 | * |
| 8 | * First round of bugfixes by Gabriel Paubert (paubert@iram.es) |
| 9 | * to make clock more stable (2.4.0-test5). The only thing |
| 10 | * that this code assumes is that the timebases have been synchronized |
| 11 | * by firmware on SMP and are never stopped (never do sleep |
| 12 | * on SMP then, nap and doze are OK). |
| 13 | * |
| 14 | * TODO (not necessarily in this file): |
| 15 | * - improve precision and reproducibility of timebase frequency |
| 16 | * measurement at boot time. |
| 17 | * - get rid of xtime_lock for gettimeofday (generic kernel problem |
| 18 | * to be implemented on all architectures for SMP scalability and |
| 19 | * eventually implementing gettimeofday without entering the kernel). |
| 20 | * - put all time/clock related variables in a single structure |
| 21 | * to minimize number of cache lines touched by gettimeofday() |
| 22 | * - for astronomical applications: add a new function to get |
| 23 | * non ambiguous timestamps even around leap seconds. This needs |
| 24 | * a new timestamp format and a good name. |
| 25 | * |
| 26 | * |
| 27 | * The following comment is partially obsolete (at least the long wait |
| 28 | * is no more a valid reason): |
| 29 | * Since the MPC8xx has a programmable interrupt timer, I decided to |
| 30 | * use that rather than the decrementer. Two reasons: 1.) the clock |
| 31 | * frequency is low, causing 2.) a long wait in the timer interrupt |
| 32 | * while ((d = get_dec()) == dval) |
| 33 | * loop. The MPC8xx can be driven from a variety of input clocks, |
| 34 | * so a number of assumptions have been made here because the kernel |
| 35 | * parameter HZ is a constant. We assume (correctly, today :-) that |
| 36 | * the MPC8xx on the MBX board is driven from a 32.768 kHz crystal. |
| 37 | * This is then divided by 4, providing a 8192 Hz clock into the PIT. |
| 38 | * Since it is not possible to get a nice 100 Hz clock out of this, without |
| 39 | * creating a software PLL, I have set HZ to 128. -- Dan |
| 40 | * |
| 41 | * 1997-09-10 Updated NTP code according to technical memorandum Jan '96 |
| 42 | * "A Kernel Model for Precision Timekeeping" by Dave Mills |
| 43 | */ |
| 44 | |
| 45 | #include <linux/config.h> |
| 46 | #include <linux/errno.h> |
| 47 | #include <linux/sched.h> |
| 48 | #include <linux/kernel.h> |
| 49 | #include <linux/param.h> |
| 50 | #include <linux/string.h> |
| 51 | #include <linux/mm.h> |
| 52 | #include <linux/module.h> |
| 53 | #include <linux/interrupt.h> |
| 54 | #include <linux/timex.h> |
| 55 | #include <linux/kernel_stat.h> |
| 56 | #include <linux/mc146818rtc.h> |
| 57 | #include <linux/time.h> |
| 58 | #include <linux/init.h> |
| 59 | #include <linux/profile.h> |
| 60 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 61 | #include <asm/io.h> |
| 62 | #include <asm/nvram.h> |
| 63 | #include <asm/cache.h> |
| 64 | #include <asm/8xx_immap.h> |
| 65 | #include <asm/machdep.h> |
| 66 | |
| 67 | #include <asm/time.h> |
| 68 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 69 | unsigned long disarm_decr[NR_CPUS]; |
| 70 | |
| 71 | extern struct timezone sys_tz; |
| 72 | |
| 73 | /* keep track of when we need to update the rtc */ |
| 74 | time_t last_rtc_update; |
| 75 | |
| 76 | /* The decrementer counts down by 128 every 128ns on a 601. */ |
| 77 | #define DECREMENTER_COUNT_601 (1000000000 / HZ) |
| 78 | |
| 79 | unsigned tb_ticks_per_jiffy; |
| 80 | unsigned tb_to_us; |
| 81 | unsigned tb_last_stamp; |
| 82 | unsigned long tb_to_ns_scale; |
| 83 | |
| 84 | extern unsigned long wall_jiffies; |
| 85 | |
john stultz | f326d22 | 2005-07-05 18:54:44 -0700 | [diff] [blame] | 86 | /* used for timezone offset */ |
| 87 | static long timezone_offset; |
| 88 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 89 | DEFINE_SPINLOCK(rtc_lock); |
| 90 | |
| 91 | EXPORT_SYMBOL(rtc_lock); |
| 92 | |
| 93 | /* Timer interrupt helper function */ |
| 94 | static inline int tb_delta(unsigned *jiffy_stamp) { |
| 95 | int delta; |
| 96 | if (__USE_RTC()) { |
| 97 | delta = get_rtcl(); |
| 98 | if (delta < *jiffy_stamp) *jiffy_stamp -= 1000000000; |
| 99 | delta -= *jiffy_stamp; |
| 100 | } else { |
| 101 | delta = get_tbl() - *jiffy_stamp; |
| 102 | } |
| 103 | return delta; |
| 104 | } |
| 105 | |
| 106 | #ifdef CONFIG_SMP |
| 107 | unsigned long profile_pc(struct pt_regs *regs) |
| 108 | { |
| 109 | unsigned long pc = instruction_pointer(regs); |
| 110 | |
| 111 | if (in_lock_functions(pc)) |
| 112 | return regs->link; |
| 113 | |
| 114 | return pc; |
| 115 | } |
| 116 | EXPORT_SYMBOL(profile_pc); |
| 117 | #endif |
| 118 | |
Paul Mackerras | f2783c1 | 2005-10-20 09:23:26 +1000 | [diff] [blame] | 119 | void wakeup_decrementer(void) |
| 120 | { |
| 121 | set_dec(tb_ticks_per_jiffy); |
| 122 | /* No currently-supported powerbook has a 601, |
| 123 | * so use get_tbl, not native |
| 124 | */ |
| 125 | last_jiffy_stamp(0) = tb_last_stamp = get_tbl(); |
| 126 | } |
| 127 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 128 | /* |
| 129 | * timer_interrupt - gets called when the decrementer overflows, |
| 130 | * with interrupts disabled. |
| 131 | * We set it up to overflow again in 1/HZ seconds. |
| 132 | */ |
| 133 | void timer_interrupt(struct pt_regs * regs) |
| 134 | { |
| 135 | int next_dec; |
| 136 | unsigned long cpu = smp_processor_id(); |
| 137 | unsigned jiffy_stamp = last_jiffy_stamp(cpu); |
| 138 | extern void do_IRQ(struct pt_regs *); |
| 139 | |
| 140 | if (atomic_read(&ppc_n_lost_interrupts) != 0) |
| 141 | do_IRQ(regs); |
| 142 | |
| 143 | irq_enter(); |
| 144 | |
| 145 | while ((next_dec = tb_ticks_per_jiffy - tb_delta(&jiffy_stamp)) <= 0) { |
| 146 | jiffy_stamp += tb_ticks_per_jiffy; |
| 147 | |
| 148 | profile_tick(CPU_PROFILING, regs); |
| 149 | update_process_times(user_mode(regs)); |
| 150 | |
| 151 | if (smp_processor_id()) |
| 152 | continue; |
| 153 | |
| 154 | /* We are in an interrupt, no need to save/restore flags */ |
| 155 | write_seqlock(&xtime_lock); |
| 156 | tb_last_stamp = jiffy_stamp; |
| 157 | do_timer(regs); |
| 158 | |
| 159 | /* |
| 160 | * update the rtc when needed, this should be performed on the |
| 161 | * right fraction of a second. Half or full second ? |
| 162 | * Full second works on mk48t59 clocks, others need testing. |
| 163 | * Note that this update is basically only used through |
| 164 | * the adjtimex system calls. Setting the HW clock in |
| 165 | * any other way is a /dev/rtc and userland business. |
| 166 | * This is still wrong by -0.5/+1.5 jiffies because of the |
| 167 | * timer interrupt resolution and possible delay, but here we |
| 168 | * hit a quantization limit which can only be solved by higher |
| 169 | * resolution timers and decoupling time management from timer |
| 170 | * interrupts. This is also wrong on the clocks |
| 171 | * which require being written at the half second boundary. |
| 172 | * We should have an rtc call that only sets the minutes and |
| 173 | * seconds like on Intel to avoid problems with non UTC clocks. |
| 174 | */ |
john stultz | b149ee2 | 2005-09-06 15:17:46 -0700 | [diff] [blame] | 175 | if ( ppc_md.set_rtc_time && ntp_synced() && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 176 | xtime.tv_sec - last_rtc_update >= 659 && |
| 177 | abs((xtime.tv_nsec / 1000) - (1000000-1000000/HZ)) < 500000/HZ && |
| 178 | jiffies - wall_jiffies == 1) { |
john stultz | f326d22 | 2005-07-05 18:54:44 -0700 | [diff] [blame] | 179 | if (ppc_md.set_rtc_time(xtime.tv_sec+1 + timezone_offset) == 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 180 | last_rtc_update = xtime.tv_sec+1; |
| 181 | else |
| 182 | /* Try again one minute later */ |
| 183 | last_rtc_update += 60; |
| 184 | } |
| 185 | write_sequnlock(&xtime_lock); |
| 186 | } |
| 187 | if ( !disarm_decr[smp_processor_id()] ) |
| 188 | set_dec(next_dec); |
| 189 | last_jiffy_stamp(cpu) = jiffy_stamp; |
| 190 | |
| 191 | if (ppc_md.heartbeat && !ppc_md.heartbeat_count--) |
| 192 | ppc_md.heartbeat(); |
| 193 | |
| 194 | irq_exit(); |
| 195 | } |
| 196 | |
| 197 | /* |
| 198 | * This version of gettimeofday has microsecond resolution. |
| 199 | */ |
| 200 | void do_gettimeofday(struct timeval *tv) |
| 201 | { |
| 202 | unsigned long flags; |
| 203 | unsigned long seq; |
| 204 | unsigned delta, lost_ticks, usec, sec; |
| 205 | |
| 206 | do { |
| 207 | seq = read_seqbegin_irqsave(&xtime_lock, flags); |
| 208 | sec = xtime.tv_sec; |
| 209 | usec = (xtime.tv_nsec / 1000); |
| 210 | delta = tb_ticks_since(tb_last_stamp); |
| 211 | #ifdef CONFIG_SMP |
| 212 | /* As long as timebases are not in sync, gettimeofday can only |
| 213 | * have jiffy resolution on SMP. |
| 214 | */ |
| 215 | if (!smp_tb_synchronized) |
| 216 | delta = 0; |
| 217 | #endif /* CONFIG_SMP */ |
| 218 | lost_ticks = jiffies - wall_jiffies; |
| 219 | } while (read_seqretry_irqrestore(&xtime_lock, seq, flags)); |
| 220 | |
| 221 | usec += mulhwu(tb_to_us, tb_ticks_per_jiffy * lost_ticks + delta); |
| 222 | while (usec >= 1000000) { |
| 223 | sec++; |
| 224 | usec -= 1000000; |
| 225 | } |
| 226 | tv->tv_sec = sec; |
| 227 | tv->tv_usec = usec; |
| 228 | } |
| 229 | |
| 230 | EXPORT_SYMBOL(do_gettimeofday); |
| 231 | |
| 232 | int do_settimeofday(struct timespec *tv) |
| 233 | { |
| 234 | time_t wtm_sec, new_sec = tv->tv_sec; |
| 235 | long wtm_nsec, new_nsec = tv->tv_nsec; |
| 236 | unsigned long flags; |
| 237 | int tb_delta; |
| 238 | |
| 239 | if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC) |
| 240 | return -EINVAL; |
| 241 | |
| 242 | write_seqlock_irqsave(&xtime_lock, flags); |
| 243 | /* Updating the RTC is not the job of this code. If the time is |
| 244 | * stepped under NTP, the RTC will be update after STA_UNSYNC |
| 245 | * is cleared. Tool like clock/hwclock either copy the RTC |
| 246 | * to the system time, in which case there is no point in writing |
| 247 | * to the RTC again, or write to the RTC but then they don't call |
| 248 | * settimeofday to perform this operation. Note also that |
| 249 | * we don't touch the decrementer since: |
| 250 | * a) it would lose timer interrupt synchronization on SMP |
| 251 | * (if it is working one day) |
| 252 | * b) it could make one jiffy spuriously shorter or longer |
| 253 | * which would introduce another source of uncertainty potentially |
| 254 | * harmful to relatively short timers. |
| 255 | */ |
| 256 | |
| 257 | /* This works perfectly on SMP only if the tb are in sync but |
| 258 | * guarantees an error < 1 jiffy even if they are off by eons, |
| 259 | * still reasonable when gettimeofday resolution is 1 jiffy. |
| 260 | */ |
| 261 | tb_delta = tb_ticks_since(last_jiffy_stamp(smp_processor_id())); |
| 262 | tb_delta += (jiffies - wall_jiffies) * tb_ticks_per_jiffy; |
| 263 | |
| 264 | new_nsec -= 1000 * mulhwu(tb_to_us, tb_delta); |
| 265 | |
| 266 | wtm_sec = wall_to_monotonic.tv_sec + (xtime.tv_sec - new_sec); |
| 267 | wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - new_nsec); |
| 268 | |
| 269 | set_normalized_timespec(&xtime, new_sec, new_nsec); |
| 270 | set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec); |
| 271 | |
| 272 | /* In case of a large backwards jump in time with NTP, we want the |
| 273 | * clock to be updated as soon as the PLL is again in lock. |
| 274 | */ |
| 275 | last_rtc_update = new_sec - 658; |
| 276 | |
john stultz | b149ee2 | 2005-09-06 15:17:46 -0700 | [diff] [blame] | 277 | ntp_clear(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 278 | write_sequnlock_irqrestore(&xtime_lock, flags); |
| 279 | clock_was_set(); |
| 280 | return 0; |
| 281 | } |
| 282 | |
| 283 | EXPORT_SYMBOL(do_settimeofday); |
| 284 | |
| 285 | /* This function is only called on the boot processor */ |
| 286 | void __init time_init(void) |
| 287 | { |
| 288 | time_t sec, old_sec; |
| 289 | unsigned old_stamp, stamp, elapsed; |
| 290 | |
| 291 | if (ppc_md.time_init != NULL) |
john stultz | f326d22 | 2005-07-05 18:54:44 -0700 | [diff] [blame] | 292 | timezone_offset = ppc_md.time_init(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 293 | |
| 294 | if (__USE_RTC()) { |
| 295 | /* 601 processor: dec counts down by 128 every 128ns */ |
| 296 | tb_ticks_per_jiffy = DECREMENTER_COUNT_601; |
| 297 | /* mulhwu_scale_factor(1000000000, 1000000) is 0x418937 */ |
| 298 | tb_to_us = 0x418937; |
| 299 | } else { |
| 300 | ppc_md.calibrate_decr(); |
| 301 | tb_to_ns_scale = mulhwu(tb_to_us, 1000 << 10); |
| 302 | } |
| 303 | |
| 304 | /* Now that the decrementer is calibrated, it can be used in case the |
| 305 | * clock is stuck, but the fact that we have to handle the 601 |
| 306 | * makes things more complex. Repeatedly read the RTC until the |
| 307 | * next second boundary to try to achieve some precision. If there |
| 308 | * is no RTC, we still need to set tb_last_stamp and |
| 309 | * last_jiffy_stamp(cpu 0) to the current stamp. |
| 310 | */ |
| 311 | stamp = get_native_tbl(); |
| 312 | if (ppc_md.get_rtc_time) { |
| 313 | sec = ppc_md.get_rtc_time(); |
| 314 | elapsed = 0; |
| 315 | do { |
| 316 | old_stamp = stamp; |
| 317 | old_sec = sec; |
| 318 | stamp = get_native_tbl(); |
| 319 | if (__USE_RTC() && stamp < old_stamp) |
| 320 | old_stamp -= 1000000000; |
| 321 | elapsed += stamp - old_stamp; |
| 322 | sec = ppc_md.get_rtc_time(); |
| 323 | } while ( sec == old_sec && elapsed < 2*HZ*tb_ticks_per_jiffy); |
| 324 | if (sec==old_sec) |
| 325 | printk("Warning: real time clock seems stuck!\n"); |
| 326 | xtime.tv_sec = sec; |
| 327 | xtime.tv_nsec = 0; |
| 328 | /* No update now, we just read the time from the RTC ! */ |
| 329 | last_rtc_update = xtime.tv_sec; |
| 330 | } |
| 331 | last_jiffy_stamp(0) = tb_last_stamp = stamp; |
| 332 | |
| 333 | /* Not exact, but the timer interrupt takes care of this */ |
| 334 | set_dec(tb_ticks_per_jiffy); |
| 335 | |
| 336 | /* If platform provided a timezone (pmac), we correct the time */ |
john stultz | f326d22 | 2005-07-05 18:54:44 -0700 | [diff] [blame] | 337 | if (timezone_offset) { |
| 338 | sys_tz.tz_minuteswest = -timezone_offset / 60; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 339 | sys_tz.tz_dsttime = 0; |
john stultz | f326d22 | 2005-07-05 18:54:44 -0700 | [diff] [blame] | 340 | xtime.tv_sec -= timezone_offset; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 341 | } |
| 342 | set_normalized_timespec(&wall_to_monotonic, |
| 343 | -xtime.tv_sec, -xtime.tv_nsec); |
| 344 | } |
| 345 | |
| 346 | #define FEBRUARY 2 |
| 347 | #define STARTOFTIME 1970 |
| 348 | #define SECDAY 86400L |
| 349 | #define SECYR (SECDAY * 365) |
| 350 | |
| 351 | /* |
| 352 | * Note: this is wrong for 2100, but our signed 32-bit time_t will |
| 353 | * have overflowed long before that, so who cares. -- paulus |
| 354 | */ |
| 355 | #define leapyear(year) ((year) % 4 == 0) |
| 356 | #define days_in_year(a) (leapyear(a) ? 366 : 365) |
| 357 | #define days_in_month(a) (month_days[(a) - 1]) |
| 358 | |
| 359 | static int month_days[12] = { |
| 360 | 31, 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31 |
| 361 | }; |
| 362 | |
| 363 | void to_tm(int tim, struct rtc_time * tm) |
| 364 | { |
| 365 | register int i; |
| 366 | register long hms, day, gday; |
| 367 | |
| 368 | gday = day = tim / SECDAY; |
| 369 | hms = tim % SECDAY; |
| 370 | |
| 371 | /* Hours, minutes, seconds are easy */ |
| 372 | tm->tm_hour = hms / 3600; |
| 373 | tm->tm_min = (hms % 3600) / 60; |
| 374 | tm->tm_sec = (hms % 3600) % 60; |
| 375 | |
| 376 | /* Number of years in days */ |
| 377 | for (i = STARTOFTIME; day >= days_in_year(i); i++) |
| 378 | day -= days_in_year(i); |
| 379 | tm->tm_year = i; |
| 380 | |
| 381 | /* Number of months in days left */ |
| 382 | if (leapyear(tm->tm_year)) |
| 383 | days_in_month(FEBRUARY) = 29; |
| 384 | for (i = 1; day >= days_in_month(i); i++) |
| 385 | day -= days_in_month(i); |
| 386 | days_in_month(FEBRUARY) = 28; |
| 387 | tm->tm_mon = i; |
| 388 | |
| 389 | /* Days are what is left over (+1) from all that. */ |
| 390 | tm->tm_mday = day + 1; |
| 391 | |
| 392 | /* |
| 393 | * Determine the day of week. Jan. 1, 1970 was a Thursday. |
| 394 | */ |
| 395 | tm->tm_wday = (gday + 4) % 7; |
| 396 | } |
| 397 | |
| 398 | /* Auxiliary function to compute scaling factors */ |
| 399 | /* Actually the choice of a timebase running at 1/4 the of the bus |
| 400 | * frequency giving resolution of a few tens of nanoseconds is quite nice. |
| 401 | * It makes this computation very precise (27-28 bits typically) which |
| 402 | * is optimistic considering the stability of most processor clock |
| 403 | * oscillators and the precision with which the timebase frequency |
| 404 | * is measured but does not harm. |
| 405 | */ |
| 406 | unsigned mulhwu_scale_factor(unsigned inscale, unsigned outscale) { |
| 407 | unsigned mlt=0, tmp, err; |
| 408 | /* No concern for performance, it's done once: use a stupid |
| 409 | * but safe and compact method to find the multiplier. |
| 410 | */ |
| 411 | for (tmp = 1U<<31; tmp != 0; tmp >>= 1) { |
| 412 | if (mulhwu(inscale, mlt|tmp) < outscale) mlt|=tmp; |
| 413 | } |
| 414 | /* We might still be off by 1 for the best approximation. |
| 415 | * A side effect of this is that if outscale is too large |
| 416 | * the returned value will be zero. |
| 417 | * Many corner cases have been checked and seem to work, |
| 418 | * some might have been forgotten in the test however. |
| 419 | */ |
| 420 | err = inscale*(mlt+1); |
| 421 | if (err <= inscale/2) mlt++; |
| 422 | return mlt; |
| 423 | } |
| 424 | |
| 425 | unsigned long long sched_clock(void) |
| 426 | { |
| 427 | unsigned long lo, hi, hi2; |
| 428 | unsigned long long tb; |
| 429 | |
| 430 | if (!__USE_RTC()) { |
| 431 | do { |
| 432 | hi = get_tbu(); |
| 433 | lo = get_tbl(); |
| 434 | hi2 = get_tbu(); |
| 435 | } while (hi2 != hi); |
| 436 | tb = ((unsigned long long) hi << 32) | lo; |
| 437 | tb = (tb * tb_to_ns_scale) >> 10; |
| 438 | } else { |
| 439 | do { |
| 440 | hi = get_rtcu(); |
| 441 | lo = get_rtcl(); |
| 442 | hi2 = get_rtcu(); |
| 443 | } while (hi2 != hi); |
| 444 | tb = ((unsigned long long) hi) * 1000000000 + lo; |
| 445 | } |
| 446 | return tb; |
| 447 | } |