1 /* 2 * linux/arch/ia64/kernel/time.c 3 * 4 * Copyright (C) 1998-2003 Hewlett-Packard Co 5 * Stephane Eranian <eranian@hpl.hp.com> 6 * David Mosberger <davidm@hpl.hp.com> 7 * Copyright (C) 1999 Don Dugger <don.dugger@intel.com> 8 * Copyright (C) 1999-2000 VA Linux Systems 9 * Copyright (C) 1999-2000 Walt Drummond <drummond@valinux.com> 10 */ 11 12 #include <linux/cpu.h> 13 #include <linux/init.h> 14 #include <linux/kernel.h> 15 #include <linux/module.h> 16 #include <linux/profile.h> 17 #include <linux/sched.h> 18 #include <linux/time.h> 19 #include <linux/interrupt.h> 20 #include <linux/efi.h> 21 #include <linux/timex.h> 22 #include <linux/timekeeper_internal.h> 23 #include <linux/platform_device.h> 24 25 #include <asm/machvec.h> 26 #include <asm/delay.h> 27 #include <asm/hw_irq.h> 28 #include <asm/ptrace.h> 29 #include <asm/sal.h> 30 #include <asm/sections.h> 31 32 #include "fsyscall_gtod_data.h" 33 34 static u64 itc_get_cycles(struct clocksource *cs); 35 36 struct fsyscall_gtod_data_t fsyscall_gtod_data; 37 38 struct itc_jitter_data_t itc_jitter_data; 39 40 volatile int time_keeper_id = 0; /* smp_processor_id() of time-keeper */ 41 42 #ifdef CONFIG_IA64_DEBUG_IRQ 43 44 unsigned long last_cli_ip; 45 EXPORT_SYMBOL(last_cli_ip); 46 47 #endif 48 49 static struct clocksource clocksource_itc = { 50 .name = "itc", 51 .rating = 350, 52 .read = itc_get_cycles, 53 .mask = CLOCKSOURCE_MASK(64), 54 .flags = CLOCK_SOURCE_IS_CONTINUOUS, 55 }; 56 static struct clocksource *itc_clocksource; 57 58 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 59 60 #include <linux/kernel_stat.h> 61 62 extern cputime_t cycle_to_cputime(u64 cyc); 63 64 void vtime_flush(struct task_struct *tsk) 65 { 66 struct thread_info *ti = task_thread_info(tsk); 67 cputime_t delta; 68 69 if (ti->utime) 70 account_user_time(tsk, cputime_to_nsecs(cycle_to_cputime(ti->utime))); 71 72 if (ti->gtime) 73 account_guest_time(tsk, cycle_to_cputime(ti->gtime)); 74 75 if (ti->idle_time) 76 account_idle_time(cputime_to_nsecs(cycle_to_cputime(ti->idle_time))); 77 78 if (ti->stime) { 79 delta = cycle_to_cputime(ti->stime); 80 account_system_index_time(tsk, delta, CPUTIME_SYSTEM); 81 } 82 83 if (ti->hardirq_time) { 84 delta = cycle_to_cputime(ti->hardirq_time); 85 account_system_index_time(tsk, delta, CPUTIME_IRQ); 86 } 87 88 if (ti->softirq_time) { 89 delta = cycle_to_cputime(ti->softirq_time); 90 account_system_index_time(tsk, delta, CPUTIME_SOFTIRQ); 91 } 92 93 ti->utime = 0; 94 ti->gtime = 0; 95 ti->idle_time = 0; 96 ti->stime = 0; 97 ti->hardirq_time = 0; 98 ti->softirq_time = 0; 99 } 100 101 /* 102 * Called from the context switch with interrupts disabled, to charge all 103 * accumulated times to the current process, and to prepare accounting on 104 * the next process. 105 */ 106 void arch_vtime_task_switch(struct task_struct *prev) 107 { 108 struct thread_info *pi = task_thread_info(prev); 109 struct thread_info *ni = task_thread_info(current); 110 111 ni->ac_stamp = pi->ac_stamp; 112 ni->ac_stime = ni->ac_utime = 0; 113 } 114 115 /* 116 * Account time for a transition between system, hard irq or soft irq state. 117 * Note that this function is called with interrupts enabled. 118 */ 119 static __u64 vtime_delta(struct task_struct *tsk) 120 { 121 struct thread_info *ti = task_thread_info(tsk); 122 __u64 now, delta_stime; 123 124 WARN_ON_ONCE(!irqs_disabled()); 125 126 now = ia64_get_itc(); 127 delta_stime = now - ti->ac_stamp; 128 ti->ac_stamp = now; 129 130 return delta_stime; 131 } 132 133 void vtime_account_system(struct task_struct *tsk) 134 { 135 struct thread_info *ti = task_thread_info(tsk); 136 __u64 stime = vtime_delta(tsk); 137 138 if ((tsk->flags & PF_VCPU) && !irq_count()) 139 ti->gtime += stime; 140 else if (hardirq_count()) 141 ti->hardirq_time += stime; 142 else if (in_serving_softirq()) 143 ti->softirq_time += stime; 144 else 145 ti->stime += stime; 146 } 147 EXPORT_SYMBOL_GPL(vtime_account_system); 148 149 void vtime_account_idle(struct task_struct *tsk) 150 { 151 struct thread_info *ti = task_thread_info(tsk); 152 153 ti->idle_time += vtime_delta(tsk); 154 } 155 156 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */ 157 158 static irqreturn_t 159 timer_interrupt (int irq, void *dev_id) 160 { 161 unsigned long new_itm; 162 163 if (cpu_is_offline(smp_processor_id())) { 164 return IRQ_HANDLED; 165 } 166 167 platform_timer_interrupt(irq, dev_id); 168 169 new_itm = local_cpu_data->itm_next; 170 171 if (!time_after(ia64_get_itc(), new_itm)) 172 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n", 173 ia64_get_itc(), new_itm); 174 175 profile_tick(CPU_PROFILING); 176 177 while (1) { 178 update_process_times(user_mode(get_irq_regs())); 179 180 new_itm += local_cpu_data->itm_delta; 181 182 if (smp_processor_id() == time_keeper_id) 183 xtime_update(1); 184 185 local_cpu_data->itm_next = new_itm; 186 187 if (time_after(new_itm, ia64_get_itc())) 188 break; 189 190 /* 191 * Allow IPIs to interrupt the timer loop. 192 */ 193 local_irq_enable(); 194 local_irq_disable(); 195 } 196 197 do { 198 /* 199 * If we're too close to the next clock tick for 200 * comfort, we increase the safety margin by 201 * intentionally dropping the next tick(s). We do NOT 202 * update itm.next because that would force us to call 203 * xtime_update() which in turn would let our clock run 204 * too fast (with the potentially devastating effect 205 * of losing monotony of time). 206 */ 207 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2)) 208 new_itm += local_cpu_data->itm_delta; 209 ia64_set_itm(new_itm); 210 /* double check, in case we got hit by a (slow) PMI: */ 211 } while (time_after_eq(ia64_get_itc(), new_itm)); 212 return IRQ_HANDLED; 213 } 214 215 /* 216 * Encapsulate access to the itm structure for SMP. 217 */ 218 void 219 ia64_cpu_local_tick (void) 220 { 221 int cpu = smp_processor_id(); 222 unsigned long shift = 0, delta; 223 224 /* arrange for the cycle counter to generate a timer interrupt: */ 225 ia64_set_itv(IA64_TIMER_VECTOR); 226 227 delta = local_cpu_data->itm_delta; 228 /* 229 * Stagger the timer tick for each CPU so they don't occur all at (almost) the 230 * same time: 231 */ 232 if (cpu) { 233 unsigned long hi = 1UL << ia64_fls(cpu); 234 shift = (2*(cpu - hi) + 1) * delta/hi/2; 235 } 236 local_cpu_data->itm_next = ia64_get_itc() + delta + shift; 237 ia64_set_itm(local_cpu_data->itm_next); 238 } 239 240 static int nojitter; 241 242 static int __init nojitter_setup(char *str) 243 { 244 nojitter = 1; 245 printk("Jitter checking for ITC timers disabled\n"); 246 return 1; 247 } 248 249 __setup("nojitter", nojitter_setup); 250 251 252 void ia64_init_itm(void) 253 { 254 unsigned long platform_base_freq, itc_freq; 255 struct pal_freq_ratio itc_ratio, proc_ratio; 256 long status, platform_base_drift, itc_drift; 257 258 /* 259 * According to SAL v2.6, we need to use a SAL call to determine the platform base 260 * frequency and then a PAL call to determine the frequency ratio between the ITC 261 * and the base frequency. 262 */ 263 status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM, 264 &platform_base_freq, &platform_base_drift); 265 if (status != 0) { 266 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status)); 267 } else { 268 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio); 269 if (status != 0) 270 printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status); 271 } 272 if (status != 0) { 273 /* invent "random" values */ 274 printk(KERN_ERR 275 "SAL/PAL failed to obtain frequency info---inventing reasonable values\n"); 276 platform_base_freq = 100000000; 277 platform_base_drift = -1; /* no drift info */ 278 itc_ratio.num = 3; 279 itc_ratio.den = 1; 280 } 281 if (platform_base_freq < 40000000) { 282 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n", 283 platform_base_freq); 284 platform_base_freq = 75000000; 285 platform_base_drift = -1; 286 } 287 if (!proc_ratio.den) 288 proc_ratio.den = 1; /* avoid division by zero */ 289 if (!itc_ratio.den) 290 itc_ratio.den = 1; /* avoid division by zero */ 291 292 itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den; 293 294 local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ; 295 printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, " 296 "ITC freq=%lu.%03luMHz", smp_processor_id(), 297 platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000, 298 itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000); 299 300 if (platform_base_drift != -1) { 301 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den; 302 printk("+/-%ldppm\n", itc_drift); 303 } else { 304 itc_drift = -1; 305 printk("\n"); 306 } 307 308 local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den; 309 local_cpu_data->itc_freq = itc_freq; 310 local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC; 311 local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT) 312 + itc_freq/2)/itc_freq; 313 314 if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) { 315 #ifdef CONFIG_SMP 316 /* On IA64 in an SMP configuration ITCs are never accurately synchronized. 317 * Jitter compensation requires a cmpxchg which may limit 318 * the scalability of the syscalls for retrieving time. 319 * The ITC synchronization is usually successful to within a few 320 * ITC ticks but this is not a sure thing. If you need to improve 321 * timer performance in SMP situations then boot the kernel with the 322 * "nojitter" option. However, doing so may result in time fluctuating (maybe 323 * even going backward) if the ITC offsets between the individual CPUs 324 * are too large. 325 */ 326 if (!nojitter) 327 itc_jitter_data.itc_jitter = 1; 328 #endif 329 } else 330 /* 331 * ITC is drifty and we have not synchronized the ITCs in smpboot.c. 332 * ITC values may fluctuate significantly between processors. 333 * Clock should not be used for hrtimers. Mark itc as only 334 * useful for boot and testing. 335 * 336 * Note that jitter compensation is off! There is no point of 337 * synchronizing ITCs since they may be large differentials 338 * that change over time. 339 * 340 * The only way to fix this would be to repeatedly sync the 341 * ITCs. Until that time we have to avoid ITC. 342 */ 343 clocksource_itc.rating = 50; 344 345 /* avoid softlock up message when cpu is unplug and plugged again. */ 346 touch_softlockup_watchdog(); 347 348 /* Setup the CPU local timer tick */ 349 ia64_cpu_local_tick(); 350 351 if (!itc_clocksource) { 352 clocksource_register_hz(&clocksource_itc, 353 local_cpu_data->itc_freq); 354 itc_clocksource = &clocksource_itc; 355 } 356 } 357 358 static u64 itc_get_cycles(struct clocksource *cs) 359 { 360 unsigned long lcycle, now, ret; 361 362 if (!itc_jitter_data.itc_jitter) 363 return get_cycles(); 364 365 lcycle = itc_jitter_data.itc_lastcycle; 366 now = get_cycles(); 367 if (lcycle && time_after(lcycle, now)) 368 return lcycle; 369 370 /* 371 * Keep track of the last timer value returned. 372 * In an SMP environment, you could lose out in contention of 373 * cmpxchg. If so, your cmpxchg returns new value which the 374 * winner of contention updated to. Use the new value instead. 375 */ 376 ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now); 377 if (unlikely(ret != lcycle)) 378 return ret; 379 380 return now; 381 } 382 383 384 static struct irqaction timer_irqaction = { 385 .handler = timer_interrupt, 386 .flags = IRQF_IRQPOLL, 387 .name = "timer" 388 }; 389 390 void read_persistent_clock64(struct timespec64 *ts) 391 { 392 efi_gettimeofday(ts); 393 } 394 395 void __init 396 time_init (void) 397 { 398 register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction); 399 ia64_init_itm(); 400 } 401 402 /* 403 * Generic udelay assumes that if preemption is allowed and the thread 404 * migrates to another CPU, that the ITC values are synchronized across 405 * all CPUs. 406 */ 407 static void 408 ia64_itc_udelay (unsigned long usecs) 409 { 410 unsigned long start = ia64_get_itc(); 411 unsigned long end = start + usecs*local_cpu_data->cyc_per_usec; 412 413 while (time_before(ia64_get_itc(), end)) 414 cpu_relax(); 415 } 416 417 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay; 418 419 void 420 udelay (unsigned long usecs) 421 { 422 (*ia64_udelay)(usecs); 423 } 424 EXPORT_SYMBOL(udelay); 425 426 /* IA64 doesn't cache the timezone */ 427 void update_vsyscall_tz(void) 428 { 429 } 430 431 void update_vsyscall_old(struct timespec *wall, struct timespec *wtm, 432 struct clocksource *c, u32 mult, u64 cycle_last) 433 { 434 write_seqcount_begin(&fsyscall_gtod_data.seq); 435 436 /* copy fsyscall clock data */ 437 fsyscall_gtod_data.clk_mask = c->mask; 438 fsyscall_gtod_data.clk_mult = mult; 439 fsyscall_gtod_data.clk_shift = c->shift; 440 fsyscall_gtod_data.clk_fsys_mmio = c->archdata.fsys_mmio; 441 fsyscall_gtod_data.clk_cycle_last = cycle_last; 442 443 /* copy kernel time structures */ 444 fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec; 445 fsyscall_gtod_data.wall_time.tv_nsec = wall->tv_nsec; 446 fsyscall_gtod_data.monotonic_time.tv_sec = wtm->tv_sec 447 + wall->tv_sec; 448 fsyscall_gtod_data.monotonic_time.tv_nsec = wtm->tv_nsec 449 + wall->tv_nsec; 450 451 /* normalize */ 452 while (fsyscall_gtod_data.monotonic_time.tv_nsec >= NSEC_PER_SEC) { 453 fsyscall_gtod_data.monotonic_time.tv_nsec -= NSEC_PER_SEC; 454 fsyscall_gtod_data.monotonic_time.tv_sec++; 455 } 456 457 write_seqcount_end(&fsyscall_gtod_data.seq); 458 } 459 460