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/nmi.h> 20 #include <linux/interrupt.h> 21 #include <linux/efi.h> 22 #include <linux/timex.h> 23 #include <linux/timekeeper_internal.h> 24 #include <linux/platform_device.h> 25 #include <linux/sched/cputime.h> 26 27 #include <asm/machvec.h> 28 #include <asm/delay.h> 29 #include <asm/hw_irq.h> 30 #include <asm/ptrace.h> 31 #include <asm/sal.h> 32 #include <asm/sections.h> 33 34 #include "fsyscall_gtod_data.h" 35 36 static u64 itc_get_cycles(struct clocksource *cs); 37 38 struct fsyscall_gtod_data_t fsyscall_gtod_data; 39 40 struct itc_jitter_data_t itc_jitter_data; 41 42 volatile int time_keeper_id = 0; /* smp_processor_id() of time-keeper */ 43 44 #ifdef CONFIG_IA64_DEBUG_IRQ 45 46 unsigned long last_cli_ip; 47 EXPORT_SYMBOL(last_cli_ip); 48 49 #endif 50 51 static struct clocksource clocksource_itc = { 52 .name = "itc", 53 .rating = 350, 54 .read = itc_get_cycles, 55 .mask = CLOCKSOURCE_MASK(64), 56 .flags = CLOCK_SOURCE_IS_CONTINUOUS, 57 }; 58 static struct clocksource *itc_clocksource; 59 60 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 61 62 #include <linux/kernel_stat.h> 63 64 extern u64 cycle_to_nsec(u64 cyc); 65 66 void vtime_flush(struct task_struct *tsk) 67 { 68 struct thread_info *ti = task_thread_info(tsk); 69 u64 delta; 70 71 if (ti->utime) 72 account_user_time(tsk, cycle_to_nsec(ti->utime)); 73 74 if (ti->gtime) 75 account_guest_time(tsk, cycle_to_nsec(ti->gtime)); 76 77 if (ti->idle_time) 78 account_idle_time(cycle_to_nsec(ti->idle_time)); 79 80 if (ti->stime) { 81 delta = cycle_to_nsec(ti->stime); 82 account_system_index_time(tsk, delta, CPUTIME_SYSTEM); 83 } 84 85 if (ti->hardirq_time) { 86 delta = cycle_to_nsec(ti->hardirq_time); 87 account_system_index_time(tsk, delta, CPUTIME_IRQ); 88 } 89 90 if (ti->softirq_time) { 91 delta = cycle_to_nsec(ti->softirq_time); 92 account_system_index_time(tsk, delta, CPUTIME_SOFTIRQ); 93 } 94 95 ti->utime = 0; 96 ti->gtime = 0; 97 ti->idle_time = 0; 98 ti->stime = 0; 99 ti->hardirq_time = 0; 100 ti->softirq_time = 0; 101 } 102 103 /* 104 * Called from the context switch with interrupts disabled, to charge all 105 * accumulated times to the current process, and to prepare accounting on 106 * the next process. 107 */ 108 void arch_vtime_task_switch(struct task_struct *prev) 109 { 110 struct thread_info *pi = task_thread_info(prev); 111 struct thread_info *ni = task_thread_info(current); 112 113 ni->ac_stamp = pi->ac_stamp; 114 ni->ac_stime = ni->ac_utime = 0; 115 } 116 117 /* 118 * Account time for a transition between system, hard irq or soft irq state. 119 * Note that this function is called with interrupts enabled. 120 */ 121 static __u64 vtime_delta(struct task_struct *tsk) 122 { 123 struct thread_info *ti = task_thread_info(tsk); 124 __u64 now, delta_stime; 125 126 WARN_ON_ONCE(!irqs_disabled()); 127 128 now = ia64_get_itc(); 129 delta_stime = now - ti->ac_stamp; 130 ti->ac_stamp = now; 131 132 return delta_stime; 133 } 134 135 void vtime_account_system(struct task_struct *tsk) 136 { 137 struct thread_info *ti = task_thread_info(tsk); 138 __u64 stime = vtime_delta(tsk); 139 140 if ((tsk->flags & PF_VCPU) && !irq_count()) 141 ti->gtime += stime; 142 else if (hardirq_count()) 143 ti->hardirq_time += stime; 144 else if (in_serving_softirq()) 145 ti->softirq_time += stime; 146 else 147 ti->stime += stime; 148 } 149 EXPORT_SYMBOL_GPL(vtime_account_system); 150 151 void vtime_account_idle(struct task_struct *tsk) 152 { 153 struct thread_info *ti = task_thread_info(tsk); 154 155 ti->idle_time += vtime_delta(tsk); 156 } 157 158 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */ 159 160 static irqreturn_t 161 timer_interrupt (int irq, void *dev_id) 162 { 163 unsigned long new_itm; 164 165 if (cpu_is_offline(smp_processor_id())) { 166 return IRQ_HANDLED; 167 } 168 169 platform_timer_interrupt(irq, dev_id); 170 171 new_itm = local_cpu_data->itm_next; 172 173 if (!time_after(ia64_get_itc(), new_itm)) 174 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n", 175 ia64_get_itc(), new_itm); 176 177 profile_tick(CPU_PROFILING); 178 179 while (1) { 180 update_process_times(user_mode(get_irq_regs())); 181 182 new_itm += local_cpu_data->itm_delta; 183 184 if (smp_processor_id() == time_keeper_id) 185 xtime_update(1); 186 187 local_cpu_data->itm_next = new_itm; 188 189 if (time_after(new_itm, ia64_get_itc())) 190 break; 191 192 /* 193 * Allow IPIs to interrupt the timer loop. 194 */ 195 local_irq_enable(); 196 local_irq_disable(); 197 } 198 199 do { 200 /* 201 * If we're too close to the next clock tick for 202 * comfort, we increase the safety margin by 203 * intentionally dropping the next tick(s). We do NOT 204 * update itm.next because that would force us to call 205 * xtime_update() which in turn would let our clock run 206 * too fast (with the potentially devastating effect 207 * of losing monotony of time). 208 */ 209 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2)) 210 new_itm += local_cpu_data->itm_delta; 211 ia64_set_itm(new_itm); 212 /* double check, in case we got hit by a (slow) PMI: */ 213 } while (time_after_eq(ia64_get_itc(), new_itm)); 214 return IRQ_HANDLED; 215 } 216 217 /* 218 * Encapsulate access to the itm structure for SMP. 219 */ 220 void 221 ia64_cpu_local_tick (void) 222 { 223 int cpu = smp_processor_id(); 224 unsigned long shift = 0, delta; 225 226 /* arrange for the cycle counter to generate a timer interrupt: */ 227 ia64_set_itv(IA64_TIMER_VECTOR); 228 229 delta = local_cpu_data->itm_delta; 230 /* 231 * Stagger the timer tick for each CPU so they don't occur all at (almost) the 232 * same time: 233 */ 234 if (cpu) { 235 unsigned long hi = 1UL << ia64_fls(cpu); 236 shift = (2*(cpu - hi) + 1) * delta/hi/2; 237 } 238 local_cpu_data->itm_next = ia64_get_itc() + delta + shift; 239 ia64_set_itm(local_cpu_data->itm_next); 240 } 241 242 static int nojitter; 243 244 static int __init nojitter_setup(char *str) 245 { 246 nojitter = 1; 247 printk("Jitter checking for ITC timers disabled\n"); 248 return 1; 249 } 250 251 __setup("nojitter", nojitter_setup); 252 253 254 void ia64_init_itm(void) 255 { 256 unsigned long platform_base_freq, itc_freq; 257 struct pal_freq_ratio itc_ratio, proc_ratio; 258 long status, platform_base_drift, itc_drift; 259 260 /* 261 * According to SAL v2.6, we need to use a SAL call to determine the platform base 262 * frequency and then a PAL call to determine the frequency ratio between the ITC 263 * and the base frequency. 264 */ 265 status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM, 266 &platform_base_freq, &platform_base_drift); 267 if (status != 0) { 268 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status)); 269 } else { 270 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio); 271 if (status != 0) 272 printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status); 273 } 274 if (status != 0) { 275 /* invent "random" values */ 276 printk(KERN_ERR 277 "SAL/PAL failed to obtain frequency info---inventing reasonable values\n"); 278 platform_base_freq = 100000000; 279 platform_base_drift = -1; /* no drift info */ 280 itc_ratio.num = 3; 281 itc_ratio.den = 1; 282 } 283 if (platform_base_freq < 40000000) { 284 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n", 285 platform_base_freq); 286 platform_base_freq = 75000000; 287 platform_base_drift = -1; 288 } 289 if (!proc_ratio.den) 290 proc_ratio.den = 1; /* avoid division by zero */ 291 if (!itc_ratio.den) 292 itc_ratio.den = 1; /* avoid division by zero */ 293 294 itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den; 295 296 local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ; 297 printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, " 298 "ITC freq=%lu.%03luMHz", smp_processor_id(), 299 platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000, 300 itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000); 301 302 if (platform_base_drift != -1) { 303 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den; 304 printk("+/-%ldppm\n", itc_drift); 305 } else { 306 itc_drift = -1; 307 printk("\n"); 308 } 309 310 local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den; 311 local_cpu_data->itc_freq = itc_freq; 312 local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC; 313 local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT) 314 + itc_freq/2)/itc_freq; 315 316 if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) { 317 #ifdef CONFIG_SMP 318 /* On IA64 in an SMP configuration ITCs are never accurately synchronized. 319 * Jitter compensation requires a cmpxchg which may limit 320 * the scalability of the syscalls for retrieving time. 321 * The ITC synchronization is usually successful to within a few 322 * ITC ticks but this is not a sure thing. If you need to improve 323 * timer performance in SMP situations then boot the kernel with the 324 * "nojitter" option. However, doing so may result in time fluctuating (maybe 325 * even going backward) if the ITC offsets between the individual CPUs 326 * are too large. 327 */ 328 if (!nojitter) 329 itc_jitter_data.itc_jitter = 1; 330 #endif 331 } else 332 /* 333 * ITC is drifty and we have not synchronized the ITCs in smpboot.c. 334 * ITC values may fluctuate significantly between processors. 335 * Clock should not be used for hrtimers. Mark itc as only 336 * useful for boot and testing. 337 * 338 * Note that jitter compensation is off! There is no point of 339 * synchronizing ITCs since they may be large differentials 340 * that change over time. 341 * 342 * The only way to fix this would be to repeatedly sync the 343 * ITCs. Until that time we have to avoid ITC. 344 */ 345 clocksource_itc.rating = 50; 346 347 /* avoid softlock up message when cpu is unplug and plugged again. */ 348 touch_softlockup_watchdog(); 349 350 /* Setup the CPU local timer tick */ 351 ia64_cpu_local_tick(); 352 353 if (!itc_clocksource) { 354 clocksource_register_hz(&clocksource_itc, 355 local_cpu_data->itc_freq); 356 itc_clocksource = &clocksource_itc; 357 } 358 } 359 360 static u64 itc_get_cycles(struct clocksource *cs) 361 { 362 unsigned long lcycle, now, ret; 363 364 if (!itc_jitter_data.itc_jitter) 365 return get_cycles(); 366 367 lcycle = itc_jitter_data.itc_lastcycle; 368 now = get_cycles(); 369 if (lcycle && time_after(lcycle, now)) 370 return lcycle; 371 372 /* 373 * Keep track of the last timer value returned. 374 * In an SMP environment, you could lose out in contention of 375 * cmpxchg. If so, your cmpxchg returns new value which the 376 * winner of contention updated to. Use the new value instead. 377 */ 378 ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now); 379 if (unlikely(ret != lcycle)) 380 return ret; 381 382 return now; 383 } 384 385 386 static struct irqaction timer_irqaction = { 387 .handler = timer_interrupt, 388 .flags = IRQF_IRQPOLL, 389 .name = "timer" 390 }; 391 392 void read_persistent_clock64(struct timespec64 *ts) 393 { 394 efi_gettimeofday(ts); 395 } 396 397 void __init 398 time_init (void) 399 { 400 register_percpu_irq(IA64_TIMER_VECTOR, &timer_irqaction); 401 ia64_init_itm(); 402 } 403 404 /* 405 * Generic udelay assumes that if preemption is allowed and the thread 406 * migrates to another CPU, that the ITC values are synchronized across 407 * all CPUs. 408 */ 409 static void 410 ia64_itc_udelay (unsigned long usecs) 411 { 412 unsigned long start = ia64_get_itc(); 413 unsigned long end = start + usecs*local_cpu_data->cyc_per_usec; 414 415 while (time_before(ia64_get_itc(), end)) 416 cpu_relax(); 417 } 418 419 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay; 420 421 void 422 udelay (unsigned long usecs) 423 { 424 (*ia64_udelay)(usecs); 425 } 426 EXPORT_SYMBOL(udelay); 427 428 /* IA64 doesn't cache the timezone */ 429 void update_vsyscall_tz(void) 430 { 431 } 432 433 void update_vsyscall(struct timekeeper *tk) 434 { 435 write_seqcount_begin(&fsyscall_gtod_data.seq); 436 437 /* copy vsyscall data */ 438 fsyscall_gtod_data.clk_mask = tk->tkr_mono.mask; 439 fsyscall_gtod_data.clk_mult = tk->tkr_mono.mult; 440 fsyscall_gtod_data.clk_shift = tk->tkr_mono.shift; 441 fsyscall_gtod_data.clk_fsys_mmio = tk->tkr_mono.clock->archdata.fsys_mmio; 442 fsyscall_gtod_data.clk_cycle_last = tk->tkr_mono.cycle_last; 443 444 fsyscall_gtod_data.wall_time.sec = tk->xtime_sec; 445 fsyscall_gtod_data.wall_time.snsec = tk->tkr_mono.xtime_nsec; 446 447 fsyscall_gtod_data.monotonic_time.sec = tk->xtime_sec 448 + tk->wall_to_monotonic.tv_sec; 449 fsyscall_gtod_data.monotonic_time.snsec = tk->tkr_mono.xtime_nsec 450 + ((u64)tk->wall_to_monotonic.tv_nsec 451 << tk->tkr_mono.shift); 452 453 /* normalize */ 454 while (fsyscall_gtod_data.monotonic_time.snsec >= 455 (((u64)NSEC_PER_SEC) << tk->tkr_mono.shift)) { 456 fsyscall_gtod_data.monotonic_time.snsec -= 457 ((u64)NSEC_PER_SEC) << tk->tkr_mono.shift; 458 fsyscall_gtod_data.monotonic_time.sec++; 459 } 460 461 write_seqcount_end(&fsyscall_gtod_data.seq); 462 } 463 464