1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * linux/arch/ia64/kernel/time.c 4 * 5 * Copyright (C) 1998-2003 Hewlett-Packard Co 6 * Stephane Eranian <eranian@hpl.hp.com> 7 * David Mosberger <davidm@hpl.hp.com> 8 * Copyright (C) 1999 Don Dugger <don.dugger@intel.com> 9 * Copyright (C) 1999-2000 VA Linux Systems 10 * Copyright (C) 1999-2000 Walt Drummond <drummond@valinux.com> 11 */ 12 13 #include <linux/cpu.h> 14 #include <linux/init.h> 15 #include <linux/kernel.h> 16 #include <linux/module.h> 17 #include <linux/profile.h> 18 #include <linux/sched.h> 19 #include <linux/time.h> 20 #include <linux/nmi.h> 21 #include <linux/interrupt.h> 22 #include <linux/efi.h> 23 #include <linux/timex.h> 24 #include <linux/timekeeper_internal.h> 25 #include <linux/platform_device.h> 26 #include <linux/sched/cputime.h> 27 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 #include "irq.h" 36 37 static u64 itc_get_cycles(struct clocksource *cs); 38 39 struct fsyscall_gtod_data_t fsyscall_gtod_data; 40 41 struct itc_jitter_data_t itc_jitter_data; 42 43 volatile int time_keeper_id = 0; /* smp_processor_id() of time-keeper */ 44 45 #ifdef CONFIG_IA64_DEBUG_IRQ 46 47 unsigned long last_cli_ip; 48 EXPORT_SYMBOL(last_cli_ip); 49 50 #endif 51 52 static struct clocksource clocksource_itc = { 53 .name = "itc", 54 .rating = 350, 55 .read = itc_get_cycles, 56 .mask = CLOCKSOURCE_MASK(64), 57 .flags = CLOCK_SOURCE_IS_CONTINUOUS, 58 }; 59 static struct clocksource *itc_clocksource; 60 61 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 62 63 #include <linux/kernel_stat.h> 64 65 extern u64 cycle_to_nsec(u64 cyc); 66 67 void vtime_flush(struct task_struct *tsk) 68 { 69 struct thread_info *ti = task_thread_info(tsk); 70 u64 delta; 71 72 if (ti->utime) 73 account_user_time(tsk, cycle_to_nsec(ti->utime)); 74 75 if (ti->gtime) 76 account_guest_time(tsk, cycle_to_nsec(ti->gtime)); 77 78 if (ti->idle_time) 79 account_idle_time(cycle_to_nsec(ti->idle_time)); 80 81 if (ti->stime) { 82 delta = cycle_to_nsec(ti->stime); 83 account_system_index_time(tsk, delta, CPUTIME_SYSTEM); 84 } 85 86 if (ti->hardirq_time) { 87 delta = cycle_to_nsec(ti->hardirq_time); 88 account_system_index_time(tsk, delta, CPUTIME_IRQ); 89 } 90 91 if (ti->softirq_time) { 92 delta = cycle_to_nsec(ti->softirq_time); 93 account_system_index_time(tsk, delta, CPUTIME_SOFTIRQ); 94 } 95 96 ti->utime = 0; 97 ti->gtime = 0; 98 ti->idle_time = 0; 99 ti->stime = 0; 100 ti->hardirq_time = 0; 101 ti->softirq_time = 0; 102 } 103 104 /* 105 * Called from the context switch with interrupts disabled, to charge all 106 * accumulated times to the current process, and to prepare accounting on 107 * the next process. 108 */ 109 void arch_vtime_task_switch(struct task_struct *prev) 110 { 111 struct thread_info *pi = task_thread_info(prev); 112 struct thread_info *ni = task_thread_info(current); 113 114 ni->ac_stamp = pi->ac_stamp; 115 ni->ac_stime = ni->ac_utime = 0; 116 } 117 118 /* 119 * Account time for a transition between system, hard irq or soft irq state. 120 * Note that this function is called with interrupts enabled. 121 */ 122 static __u64 vtime_delta(struct task_struct *tsk) 123 { 124 struct thread_info *ti = task_thread_info(tsk); 125 __u64 now, delta_stime; 126 127 WARN_ON_ONCE(!irqs_disabled()); 128 129 now = ia64_get_itc(); 130 delta_stime = now - ti->ac_stamp; 131 ti->ac_stamp = now; 132 133 return delta_stime; 134 } 135 136 void vtime_account_kernel(struct task_struct *tsk) 137 { 138 struct thread_info *ti = task_thread_info(tsk); 139 __u64 stime = vtime_delta(tsk); 140 141 if (tsk->flags & PF_VCPU) 142 ti->gtime += stime; 143 else 144 ti->stime += stime; 145 } 146 EXPORT_SYMBOL_GPL(vtime_account_kernel); 147 148 void vtime_account_idle(struct task_struct *tsk) 149 { 150 struct thread_info *ti = task_thread_info(tsk); 151 152 ti->idle_time += vtime_delta(tsk); 153 } 154 155 void vtime_account_softirq(struct task_struct *tsk) 156 { 157 struct thread_info *ti = task_thread_info(tsk); 158 159 ti->softirq_time += vtime_delta(tsk); 160 } 161 162 void vtime_account_hardirq(struct task_struct *tsk) 163 { 164 struct thread_info *ti = task_thread_info(tsk); 165 166 ti->hardirq_time += vtime_delta(tsk); 167 } 168 169 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */ 170 171 static irqreturn_t 172 timer_interrupt (int irq, void *dev_id) 173 { 174 unsigned long cur_itm, new_itm, ticks; 175 176 if (cpu_is_offline(smp_processor_id())) { 177 return IRQ_HANDLED; 178 } 179 180 new_itm = local_cpu_data->itm_next; 181 cur_itm = ia64_get_itc(); 182 183 if (!time_after(cur_itm, new_itm)) { 184 printk(KERN_ERR "Oops: timer tick before it's due (itc=%lx,itm=%lx)\n", 185 cur_itm, new_itm); 186 ticks = 1; 187 } else { 188 ticks = DIV_ROUND_UP(cur_itm - new_itm, 189 local_cpu_data->itm_delta); 190 new_itm += ticks * local_cpu_data->itm_delta; 191 } 192 193 if (smp_processor_id() != time_keeper_id) 194 ticks = 0; 195 196 legacy_timer_tick(ticks); 197 198 do { 199 /* 200 * If we're too close to the next clock tick for 201 * comfort, we increase the safety margin by 202 * intentionally dropping the next tick(s). We do NOT 203 * update itm.next because that would force us to call 204 * xtime_update() which in turn would let our clock run 205 * too fast (with the potentially devastating effect 206 * of losing monotony of time). 207 */ 208 while (!time_after(new_itm, ia64_get_itc() + local_cpu_data->itm_delta/2)) 209 new_itm += local_cpu_data->itm_delta; 210 ia64_set_itm(new_itm); 211 /* double check, in case we got hit by a (slow) PMI: */ 212 } while (time_after_eq(ia64_get_itc(), new_itm)); 213 return IRQ_HANDLED; 214 } 215 216 /* 217 * Encapsulate access to the itm structure for SMP. 218 */ 219 void 220 ia64_cpu_local_tick (void) 221 { 222 int cpu = smp_processor_id(); 223 unsigned long shift = 0, delta; 224 225 /* arrange for the cycle counter to generate a timer interrupt: */ 226 ia64_set_itv(IA64_TIMER_VECTOR); 227 228 delta = local_cpu_data->itm_delta; 229 /* 230 * Stagger the timer tick for each CPU so they don't occur all at (almost) the 231 * same time: 232 */ 233 if (cpu) { 234 unsigned long hi = 1UL << ia64_fls(cpu); 235 shift = (2*(cpu - hi) + 1) * delta/hi/2; 236 } 237 local_cpu_data->itm_next = ia64_get_itc() + delta + shift; 238 ia64_set_itm(local_cpu_data->itm_next); 239 } 240 241 static int nojitter; 242 243 static int __init nojitter_setup(char *str) 244 { 245 nojitter = 1; 246 printk("Jitter checking for ITC timers disabled\n"); 247 return 1; 248 } 249 250 __setup("nojitter", nojitter_setup); 251 252 253 void ia64_init_itm(void) 254 { 255 unsigned long platform_base_freq, itc_freq; 256 struct pal_freq_ratio itc_ratio, proc_ratio; 257 long status, platform_base_drift, itc_drift; 258 259 /* 260 * According to SAL v2.6, we need to use a SAL call to determine the platform base 261 * frequency and then a PAL call to determine the frequency ratio between the ITC 262 * and the base frequency. 263 */ 264 status = ia64_sal_freq_base(SAL_FREQ_BASE_PLATFORM, 265 &platform_base_freq, &platform_base_drift); 266 if (status != 0) { 267 printk(KERN_ERR "SAL_FREQ_BASE_PLATFORM failed: %s\n", ia64_sal_strerror(status)); 268 } else { 269 status = ia64_pal_freq_ratios(&proc_ratio, NULL, &itc_ratio); 270 if (status != 0) 271 printk(KERN_ERR "PAL_FREQ_RATIOS failed with status=%ld\n", status); 272 } 273 if (status != 0) { 274 /* invent "random" values */ 275 printk(KERN_ERR 276 "SAL/PAL failed to obtain frequency info---inventing reasonable values\n"); 277 platform_base_freq = 100000000; 278 platform_base_drift = -1; /* no drift info */ 279 itc_ratio.num = 3; 280 itc_ratio.den = 1; 281 } 282 if (platform_base_freq < 40000000) { 283 printk(KERN_ERR "Platform base frequency %lu bogus---resetting to 75MHz!\n", 284 platform_base_freq); 285 platform_base_freq = 75000000; 286 platform_base_drift = -1; 287 } 288 if (!proc_ratio.den) 289 proc_ratio.den = 1; /* avoid division by zero */ 290 if (!itc_ratio.den) 291 itc_ratio.den = 1; /* avoid division by zero */ 292 293 itc_freq = (platform_base_freq*itc_ratio.num)/itc_ratio.den; 294 295 local_cpu_data->itm_delta = (itc_freq + HZ/2) / HZ; 296 printk(KERN_DEBUG "CPU %d: base freq=%lu.%03luMHz, ITC ratio=%u/%u, " 297 "ITC freq=%lu.%03luMHz", smp_processor_id(), 298 platform_base_freq / 1000000, (platform_base_freq / 1000) % 1000, 299 itc_ratio.num, itc_ratio.den, itc_freq / 1000000, (itc_freq / 1000) % 1000); 300 301 if (platform_base_drift != -1) { 302 itc_drift = platform_base_drift*itc_ratio.num/itc_ratio.den; 303 printk("+/-%ldppm\n", itc_drift); 304 } else { 305 itc_drift = -1; 306 printk("\n"); 307 } 308 309 local_cpu_data->proc_freq = (platform_base_freq*proc_ratio.num)/proc_ratio.den; 310 local_cpu_data->itc_freq = itc_freq; 311 local_cpu_data->cyc_per_usec = (itc_freq + USEC_PER_SEC/2) / USEC_PER_SEC; 312 local_cpu_data->nsec_per_cyc = ((NSEC_PER_SEC<<IA64_NSEC_PER_CYC_SHIFT) 313 + itc_freq/2)/itc_freq; 314 315 if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) { 316 #ifdef CONFIG_SMP 317 /* On IA64 in an SMP configuration ITCs are never accurately synchronized. 318 * Jitter compensation requires a cmpxchg which may limit 319 * the scalability of the syscalls for retrieving time. 320 * The ITC synchronization is usually successful to within a few 321 * ITC ticks but this is not a sure thing. If you need to improve 322 * timer performance in SMP situations then boot the kernel with the 323 * "nojitter" option. However, doing so may result in time fluctuating (maybe 324 * even going backward) if the ITC offsets between the individual CPUs 325 * are too large. 326 */ 327 if (!nojitter) 328 itc_jitter_data.itc_jitter = 1; 329 #endif 330 } else 331 /* 332 * ITC is drifty and we have not synchronized the ITCs in smpboot.c. 333 * ITC values may fluctuate significantly between processors. 334 * Clock should not be used for hrtimers. Mark itc as only 335 * useful for boot and testing. 336 * 337 * Note that jitter compensation is off! There is no point of 338 * synchronizing ITCs since they may be large differentials 339 * that change over time. 340 * 341 * The only way to fix this would be to repeatedly sync the 342 * ITCs. Until that time we have to avoid ITC. 343 */ 344 clocksource_itc.rating = 50; 345 346 /* avoid softlock up message when cpu is unplug and plugged again. */ 347 touch_softlockup_watchdog(); 348 349 /* Setup the CPU local timer tick */ 350 ia64_cpu_local_tick(); 351 352 if (!itc_clocksource) { 353 clocksource_register_hz(&clocksource_itc, 354 local_cpu_data->itc_freq); 355 itc_clocksource = &clocksource_itc; 356 } 357 } 358 359 static u64 itc_get_cycles(struct clocksource *cs) 360 { 361 unsigned long lcycle, now, ret; 362 363 if (!itc_jitter_data.itc_jitter) 364 return get_cycles(); 365 366 lcycle = itc_jitter_data.itc_lastcycle; 367 now = get_cycles(); 368 if (lcycle && time_after(lcycle, now)) 369 return lcycle; 370 371 /* 372 * Keep track of the last timer value returned. 373 * In an SMP environment, you could lose out in contention of 374 * cmpxchg. If so, your cmpxchg returns new value which the 375 * winner of contention updated to. Use the new value instead. 376 */ 377 ret = cmpxchg(&itc_jitter_data.itc_lastcycle, lcycle, now); 378 if (unlikely(ret != lcycle)) 379 return ret; 380 381 return now; 382 } 383 384 void read_persistent_clock64(struct timespec64 *ts) 385 { 386 efi_gettimeofday(ts); 387 } 388 389 void __init 390 time_init (void) 391 { 392 register_percpu_irq(IA64_TIMER_VECTOR, timer_interrupt, IRQF_IRQPOLL, 393 "timer"); 394 ia64_init_itm(); 395 } 396 397 /* 398 * Generic udelay assumes that if preemption is allowed and the thread 399 * migrates to another CPU, that the ITC values are synchronized across 400 * all CPUs. 401 */ 402 static void 403 ia64_itc_udelay (unsigned long usecs) 404 { 405 unsigned long start = ia64_get_itc(); 406 unsigned long end = start + usecs*local_cpu_data->cyc_per_usec; 407 408 while (time_before(ia64_get_itc(), end)) 409 cpu_relax(); 410 } 411 412 void (*ia64_udelay)(unsigned long usecs) = &ia64_itc_udelay; 413 414 void 415 udelay (unsigned long usecs) 416 { 417 (*ia64_udelay)(usecs); 418 } 419 EXPORT_SYMBOL(udelay); 420 421 /* IA64 doesn't cache the timezone */ 422 void update_vsyscall_tz(void) 423 { 424 } 425 426 void update_vsyscall(struct timekeeper *tk) 427 { 428 write_seqcount_begin(&fsyscall_gtod_data.seq); 429 430 /* copy vsyscall data */ 431 fsyscall_gtod_data.clk_mask = tk->tkr_mono.mask; 432 fsyscall_gtod_data.clk_mult = tk->tkr_mono.mult; 433 fsyscall_gtod_data.clk_shift = tk->tkr_mono.shift; 434 fsyscall_gtod_data.clk_fsys_mmio = tk->tkr_mono.clock->archdata.fsys_mmio; 435 fsyscall_gtod_data.clk_cycle_last = tk->tkr_mono.cycle_last; 436 437 fsyscall_gtod_data.wall_time.sec = tk->xtime_sec; 438 fsyscall_gtod_data.wall_time.snsec = tk->tkr_mono.xtime_nsec; 439 440 fsyscall_gtod_data.monotonic_time.sec = tk->xtime_sec 441 + tk->wall_to_monotonic.tv_sec; 442 fsyscall_gtod_data.monotonic_time.snsec = tk->tkr_mono.xtime_nsec 443 + ((u64)tk->wall_to_monotonic.tv_nsec 444 << tk->tkr_mono.shift); 445 446 /* normalize */ 447 while (fsyscall_gtod_data.monotonic_time.snsec >= 448 (((u64)NSEC_PER_SEC) << tk->tkr_mono.shift)) { 449 fsyscall_gtod_data.monotonic_time.snsec -= 450 ((u64)NSEC_PER_SEC) << tk->tkr_mono.shift; 451 fsyscall_gtod_data.monotonic_time.sec++; 452 } 453 454 write_seqcount_end(&fsyscall_gtod_data.seq); 455 } 456 457