arch/i386/kernel/vmitime.c

   1 /*
   2  * VMI paravirtual timer support routines.
   3  *
   4  * Copyright (C) 2005, VMware, Inc.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License as published by
   8  * the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful, but
  12  * WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  14  * NON INFRINGEMENT.  See the GNU General Public License for more
  15  * details.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * along with this program; if not, write to the Free Software
  19  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  * Send feedback to dhecht@vmware.com
  22  *
  23  */
  24
  25 /*
  26  * Portions of this code from arch/i386/kernel/timers/timer_tsc.c.
  27  * Portions of the CONFIG_NO_IDLE_HZ code from arch/s390/kernel/time.c.
  28  * See comments there for proper credits.
  29  */
  30
  31 #include <linux/spinlock.h>
  32 #include <linux/init.h>
  33 #include <linux/errno.h>
  34 #include <linux/jiffies.h>
  35 #include <linux/interrupt.h>
  36 #include <linux/kernel_stat.h>
  37 #include <linux/rcupdate.h>
  38 #include <linux/clocksource.h>
  39
  40 #include <asm/timer.h>
  41 #include <asm/io.h>
  42 #include <asm/apic.h>
  43 #include <asm/div64.h>
  44 #include <asm/timer.h>
  45 #include <asm/desc.h>
  46
  47 #include <asm/vmi.h>
  48 #include <asm/vmi_time.h>
  49
  50 #include <mach_timer.h>
  51 #include <io_ports.h>
  52
  53 #ifdef CONFIG_X86_LOCAL_APIC
  54 #define VMI_ALARM_WIRING VMI_ALARM_WIRED_LVTT
  55 #else
  56 #define VMI_ALARM_WIRING VMI_ALARM_WIRED_IRQ0
  57 #endif
  58
  59 /* Cached VMI operations */
  60 struct vmi_timer_ops vmi_timer_ops;
  61
  62 #ifdef CONFIG_NO_IDLE_HZ
  63
  64 /* /proc/sys/kernel/hz_timer state. */
  65 int sysctl_hz_timer;
  66
  67 /* Some stats */
  68 static DEFINE_PER_CPU(unsigned long, vmi_idle_no_hz_irqs);
  69 static DEFINE_PER_CPU(unsigned long, vmi_idle_no_hz_jiffies);
  70 static DEFINE_PER_CPU(unsigned long, idle_start_jiffies);
  71
  72 #endif /* CONFIG_NO_IDLE_HZ */
  73
  74 /* Number of alarms per second. By default this is CONFIG_VMI_ALARM_HZ. */
  75 static int alarm_hz = CONFIG_VMI_ALARM_HZ;
  76
  77 /* Cache of the value get_cycle_frequency / HZ. */
  78 static signed long long cycles_per_jiffy;
  79
  80 /* Cache of the value get_cycle_frequency / alarm_hz. */
  81 static signed long long cycles_per_alarm;
  82
  83 /* The number of cycles accounted for by the 'jiffies'/'xtime' count.
  84  * Protected by xtime_lock. */
  85 static unsigned long long real_cycles_accounted_system;
  86
  87 /* The number of cycles accounted for by update_process_times(), per cpu. */
  88 static DEFINE_PER_CPU(unsigned long long, process_times_cycles_accounted_cpu);
  89
  90 /* The number of stolen cycles accounted, per cpu. */
  91 static DEFINE_PER_CPU(unsigned long long, stolen_cycles_accounted_cpu);
  92
  93 /* Clock source. */
  94 static cycle_t read_real_cycles(void)
  95 {
  96         return vmi_timer_ops.get_cycle_counter(VMI_CYCLES_REAL);
  97 }
  98
  99 static cycle_t read_available_cycles(void)
 100 {
 101         return vmi_timer_ops.get_cycle_counter(VMI_CYCLES_AVAILABLE);
 102 }
 103
 104 #if 0
 105 static cycle_t read_stolen_cycles(void)
 106 {
 107         return vmi_timer_ops.get_cycle_counter(VMI_CYCLES_STOLEN);
 108 }
 109 #endif  /*  0  */
 110
 111 static struct clocksource clocksource_vmi = {
 112         .name                   = "vmi-timer",
 113         .rating                 = 450,
 114         .read                   = read_real_cycles,
 115         .mask                   = CLOCKSOURCE_MASK(64),
 116         .mult                   = 0, /* to be set */
 117         .shift                  = 22,
 118         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS,
 119 };
 120
 121
 122 /* Timer interrupt handler. */
 123 static irqreturn_t vmi_timer_interrupt(int irq, void *dev_id);
 124
 125 static struct irqaction vmi_timer_irq  = {
 126         vmi_timer_interrupt,
 127         SA_INTERRUPT,
 128         CPU_MASK_NONE,
 129         "VMI-alarm",
 130         NULL,
 131         NULL
 132 };
 133
 134 /* Alarm rate */
 135 static int __init vmi_timer_alarm_rate_setup(char* str)
 136 {
 137         int alarm_rate;
 138         if (get_option(&str, &alarm_rate) == 1 && alarm_rate > 0) {
 139                 alarm_hz = alarm_rate;
 140                 printk(KERN_WARNING "VMI timer alarm HZ set to %d\n", alarm_hz);
 141         }
 142         return 1;
 143 }
 144 __setup("vmi_timer_alarm_hz=", vmi_timer_alarm_rate_setup);
 145
 146
 147 /* Initialization */
 148 static void vmi_get_wallclock_ts(struct timespec *ts)
 149 {
 150         unsigned long long wallclock;
 151         wallclock = vmi_timer_ops.get_wallclock(); // nsec units
 152         ts->tv_nsec = do_div(wallclock, 1000000000);
 153         ts->tv_sec = wallclock;
 154 }
 155
 156 static void update_xtime_from_wallclock(void)
 157 {
 158         struct timespec ts;
 159         vmi_get_wallclock_ts(&ts);
 160         do_settimeofday(&ts);
 161 }
 162
 163 unsigned long vmi_get_wallclock(void)
 164 {
 165         struct timespec ts;
 166         vmi_get_wallclock_ts(&ts);
 167         return ts.tv_sec;
 168 }
 169
 170 int vmi_set_wallclock(unsigned long now)
 171 {
 172         return -1;
 173 }
 174
 175 unsigned long long vmi_sched_clock(void)
 176 {
 177         return read_available_cycles();
 178 }
 179
 180 void __init vmi_time_init(void)
 181 {
 182         unsigned long long cycles_per_sec, cycles_per_msec;
 183         unsigned long flags;
 184
 185         local_irq_save(flags);
 186         setup_irq(0, &vmi_timer_irq);
 187 #ifdef CONFIG_X86_LOCAL_APIC
 188         set_intr_gate(LOCAL_TIMER_VECTOR, apic_vmi_timer_interrupt);
 189 #endif
 190
 191         no_sync_cmos_clock = 1;
 192
 193         vmi_get_wallclock_ts(&xtime);
 194         set_normalized_timespec(&wall_to_monotonic,
 195                 -xtime.tv_sec, -xtime.tv_nsec);
 196
 197         real_cycles_accounted_system = read_real_cycles();
 198         update_xtime_from_wallclock();
 199         per_cpu(process_times_cycles_accounted_cpu, 0) = read_available_cycles();
 200
 201         cycles_per_sec = vmi_timer_ops.get_cycle_frequency();
 202
 203         cycles_per_jiffy = cycles_per_sec;
 204         (void)do_div(cycles_per_jiffy, HZ);
 205         cycles_per_alarm = cycles_per_sec;
 206         (void)do_div(cycles_per_alarm, alarm_hz);
 207         cycles_per_msec = cycles_per_sec;
 208         (void)do_div(cycles_per_msec, 1000);
 209         cpu_khz = cycles_per_msec;
 210
 211         printk(KERN_WARNING "VMI timer cycles/sec = %llu ; cycles/jiffy = %llu ;"
 212                "cycles/alarm = %llu\n", cycles_per_sec, cycles_per_jiffy,
 213                cycles_per_alarm);
 214
 215         clocksource_vmi.mult = clocksource_khz2mult(cycles_per_msec,
 216                                                     clocksource_vmi.shift);
 217         if (clocksource_register(&clocksource_vmi))
 218                 printk(KERN_WARNING "Error registering VMITIME clocksource.");
 219
 220         /* Disable PIT. */
 221         outb_p(0x3a, PIT_MODE); /* binary, mode 5, LSB/MSB, ch 0 */
 222
 223         /* schedule the alarm. do this in phase with process_times_cycles_accounted_cpu
 224          * reduce the latency calling update_process_times. */
 225         vmi_timer_ops.set_alarm(
 226                       VMI_ALARM_WIRED_IRQ0 | VMI_ALARM_IS_PERIODIC | VMI_CYCLES_AVAILABLE,
 227                       per_cpu(process_times_cycles_accounted_cpu, 0) + cycles_per_alarm,
 228                       cycles_per_alarm);
 229
 230         local_irq_restore(flags);
 231 }
 232
 233 #ifdef CONFIG_X86_LOCAL_APIC
 234
 235 void __init vmi_timer_setup_boot_alarm(void)
 236 {
 237         local_irq_disable();
 238
 239         /* Route the interrupt to the correct vector. */
 240         apic_write_around(APIC_LVTT, LOCAL_TIMER_VECTOR);
 241
 242         /* Cancel the IRQ0 wired alarm, and setup the LVTT alarm. */
 243         vmi_timer_ops.cancel_alarm(VMI_CYCLES_AVAILABLE);
 244         vmi_timer_ops.set_alarm(
 245                       VMI_ALARM_WIRED_LVTT | VMI_ALARM_IS_PERIODIC | VMI_CYCLES_AVAILABLE,
 246                       per_cpu(process_times_cycles_accounted_cpu, 0) + cycles_per_alarm,
 247                       cycles_per_alarm);
 248         local_irq_enable();
 249 }
 250
 251 /* Initialize the time accounting variables for an AP on an SMP system.
 252  * Also, set the local alarm for the AP. */
 253 void __init vmi_timer_setup_secondary_alarm(void)
 254 {
 255         int cpu = smp_processor_id();
 256
 257         /* Route the interrupt to the correct vector. */
 258         apic_write_around(APIC_LVTT, LOCAL_TIMER_VECTOR);
 259
 260         per_cpu(process_times_cycles_accounted_cpu, cpu) = read_available_cycles();
 261
 262         vmi_timer_ops.set_alarm(
 263                       VMI_ALARM_WIRED_LVTT | VMI_ALARM_IS_PERIODIC | VMI_CYCLES_AVAILABLE,
 264                       per_cpu(process_times_cycles_accounted_cpu, cpu) + cycles_per_alarm,
 265                       cycles_per_alarm);
 266 }
 267
 268 #endif
 269
 270 /* Update system wide (real) time accounting (e.g. jiffies, xtime). */
 271 static void vmi_account_real_cycles(unsigned long long cur_real_cycles)
 272 {
 273         long long cycles_not_accounted;
 274
 275         write_seqlock(&xtime_lock);
 276
 277         cycles_not_accounted = cur_real_cycles - real_cycles_accounted_system;
 278         while (cycles_not_accounted >= cycles_per_jiffy) {
 279                 /* systems wide jiffies. */
 280                 do_timer(1);
 281
 282                 cycles_not_accounted -= cycles_per_jiffy;
 283                 real_cycles_accounted_system += cycles_per_jiffy;
 284         }
 285
 286         write_sequnlock(&xtime_lock);
 287 }
 288
 289 /* Update per-cpu process times. */
 290 static void vmi_account_process_times_cycles(struct pt_regs *regs, int cpu,
 291                                              unsigned long long cur_process_times_cycles)
 292 {
 293         long long cycles_not_accounted;
 294         cycles_not_accounted = cur_process_times_cycles -
 295                 per_cpu(process_times_cycles_accounted_cpu, cpu);
 296
 297         while (cycles_not_accounted >= cycles_per_jiffy) {
 298                 /* Account time to the current process.  This includes
 299                  * calling into the scheduler to decrement the timeslice
 300                  * and possibly reschedule.*/
 301                 update_process_times(user_mode(regs));
 302                 /* XXX handle /proc/profile multiplier.  */
 303                 profile_tick(CPU_PROFILING);
 304
 305                 cycles_not_accounted -= cycles_per_jiffy;
 306                 per_cpu(process_times_cycles_accounted_cpu, cpu) += cycles_per_jiffy;
 307         }
 308 }
 309
 310 #ifdef CONFIG_NO_IDLE_HZ
 311 /* Update per-cpu idle times.  Used when a no-hz halt is ended. */
 312 static void vmi_account_no_hz_idle_cycles(int cpu,
 313                                           unsigned long long cur_process_times_cycles)
 314 {
 315         long long cycles_not_accounted;
 316         unsigned long no_idle_hz_jiffies = 0;
 317
 318         cycles_not_accounted = cur_process_times_cycles -
 319                 per_cpu(process_times_cycles_accounted_cpu, cpu);
 320
 321         while (cycles_not_accounted >= cycles_per_jiffy) {
 322                 no_idle_hz_jiffies++;
 323                 cycles_not_accounted -= cycles_per_jiffy;
 324                 per_cpu(process_times_cycles_accounted_cpu, cpu) += cycles_per_jiffy;
 325         }
 326         /* Account time to the idle process. */
 327         account_steal_time(idle_task(cpu), jiffies_to_cputime(no_idle_hz_jiffies));
 328 }
 329 #endif
 330
 331 /* Update per-cpu stolen time. */
 332 static void vmi_account_stolen_cycles(int cpu,
 333                                       unsigned long long cur_real_cycles,
 334                                       unsigned long long cur_avail_cycles)
 335 {
 336         long long stolen_cycles_not_accounted;
 337         unsigned long stolen_jiffies = 0;
 338
 339         if (cur_real_cycles < cur_avail_cycles)
 340                 return;
 341
 342         stolen_cycles_not_accounted = cur_real_cycles - cur_avail_cycles -
 343                 per_cpu(stolen_cycles_accounted_cpu, cpu);
 344
 345         while (stolen_cycles_not_accounted >= cycles_per_jiffy) {
 346                 stolen_jiffies++;
 347                 stolen_cycles_not_accounted -= cycles_per_jiffy;
 348                 per_cpu(stolen_cycles_accounted_cpu, cpu) += cycles_per_jiffy;
 349         }
 350         /* HACK: pass NULL to force time onto cpustat->steal. */
 351         account_steal_time(NULL, jiffies_to_cputime(stolen_jiffies));
 352 }
 353
 354 /* Body of either IRQ0 interrupt handler (UP no local-APIC) or
 355  * local-APIC LVTT interrupt handler (UP & local-APIC or SMP). */
 356 static void vmi_local_timer_interrupt(int cpu)
 357 {
 358         unsigned long long cur_real_cycles, cur_process_times_cycles;
 359
 360         cur_real_cycles = read_real_cycles();
 361         cur_process_times_cycles = read_available_cycles();
 362         /* Update system wide (real) time state (xtime, jiffies). */
 363         vmi_account_real_cycles(cur_real_cycles);
 364         /* Update per-cpu process times. */
 365         vmi_account_process_times_cycles(get_irq_regs(), cpu, cur_process_times_cycles);
 366         /* Update time stolen from this cpu by the hypervisor. */
 367         vmi_account_stolen_cycles(cpu, cur_real_cycles, cur_process_times_cycles);
 368 }
 369
 370 #ifdef CONFIG_NO_IDLE_HZ
 371
 372 /* Must be called only from idle loop, with interrupts disabled. */
 373 int vmi_stop_hz_timer(void)
 374 {
 375         /* Note that cpu_set, cpu_clear are (SMP safe) atomic on x86. */
 376
 377         unsigned long seq, next;
 378         unsigned long long real_cycles_expiry;
 379         int cpu = smp_processor_id();
 380
 381         BUG_ON(!irqs_disabled());
 382         if (sysctl_hz_timer != 0)
 383                 return 0;
 384
 385         cpu_set(cpu, nohz_cpu_mask);
 386         smp_mb();
 387
 388         if (rcu_needs_cpu(cpu) || local_softirq_pending() ||
 389             (next = next_timer_interrupt(),
 390              time_before_eq(next, jiffies + HZ/CONFIG_VMI_ALARM_HZ))) {
 391                 cpu_clear(cpu, nohz_cpu_mask);
 392                 return 0;
 393         }
 394
 395         /* Convert jiffies to the real cycle counter. */
 396         do {
 397                 seq = read_seqbegin(&xtime_lock);
 398                 real_cycles_expiry = real_cycles_accounted_system +
 399                         (long)(next - jiffies) * cycles_per_jiffy;
 400         } while (read_seqretry(&xtime_lock, seq));
 401
 402         /* This cpu is going idle. Disable the periodic alarm. */
 403         vmi_timer_ops.cancel_alarm(VMI_CYCLES_AVAILABLE);
 404         per_cpu(idle_start_jiffies, cpu) = jiffies;
 405         /* Set the real time alarm to expire at the next event. */
 406         vmi_timer_ops.set_alarm(
 407                 VMI_ALARM_WIRING | VMI_ALARM_IS_ONESHOT | VMI_CYCLES_REAL,
 408                 real_cycles_expiry, 0);
 409         return 1;
 410 }
 411
 412 static void vmi_reenable_hz_timer(int cpu)
 413 {
 414         /* For /proc/vmi/info idle_hz stat. */
 415         per_cpu(vmi_idle_no_hz_jiffies, cpu) += jiffies - per_cpu(idle_start_jiffies, cpu);
 416         per_cpu(vmi_idle_no_hz_irqs, cpu)++;
 417
 418         /* Don't bother explicitly cancelling the one-shot alarm -- at
 419          * worse we will receive a spurious timer interrupt. */
 420         vmi_timer_ops.set_alarm(
 421                       VMI_ALARM_WIRING | VMI_ALARM_IS_PERIODIC | VMI_CYCLES_AVAILABLE,
 422                       per_cpu(process_times_cycles_accounted_cpu, cpu) + cycles_per_alarm,
 423                       cycles_per_alarm);
 424         /* Indicate this cpu is no longer nohz idle. */
 425         cpu_clear(cpu, nohz_cpu_mask);
 426 }
 427
 428 /* Called from interrupt handlers when (local) HZ timer is disabled. */
 429 void vmi_account_time_restart_hz_timer(void)
 430 {
 431         unsigned long long cur_real_cycles, cur_process_times_cycles;
 432         int cpu = smp_processor_id();
 433
 434         BUG_ON(!irqs_disabled());
 435         /* Account the time during which the HZ timer was disabled. */
 436         cur_real_cycles = read_real_cycles();
 437         cur_process_times_cycles = read_available_cycles();
 438         /* Update system wide (real) time state (xtime, jiffies). */
 439         vmi_account_real_cycles(cur_real_cycles);
 440         /* Update per-cpu idle times. */
 441         vmi_account_no_hz_idle_cycles(cpu, cur_process_times_cycles);
 442         /* Update time stolen from this cpu by the hypervisor. */
 443         vmi_account_stolen_cycles(cpu, cur_real_cycles, cur_process_times_cycles);
 444         /* Reenable the hz timer. */
 445         vmi_reenable_hz_timer(cpu);
 446 }
 447
 448 #endif /* CONFIG_NO_IDLE_HZ */
 449
 450 /* UP (and no local-APIC) VMI-timer alarm interrupt handler.
 451  * Handler for IRQ0. Not used when SMP or X86_LOCAL_APIC after
 452  * APIC setup and setup_boot_vmi_alarm() is called.  */
 453 static irqreturn_t vmi_timer_interrupt(int irq, void *dev_id)
 454 {
 455         vmi_local_timer_interrupt(smp_processor_id());
 456         return IRQ_HANDLED;
 457 }
 458
 459 #ifdef CONFIG_X86_LOCAL_APIC
 460
 461 /* SMP VMI-timer alarm interrupt handler. Handler for LVTT vector.
 462  * Also used in UP when CONFIG_X86_LOCAL_APIC.
 463  * The wrapper code is from arch/i386/kernel/apic.c#smp_apic_timer_interrupt. */
 464 void smp_apic_vmi_timer_interrupt(struct pt_regs *regs)
 465 {
 466         struct pt_regs *old_regs = set_irq_regs(regs);
 467         int cpu = smp_processor_id();
 468
 469         /*
 470          * the NMI deadlock-detector uses this.
 471          */
 472         per_cpu(irq_stat,cpu).apic_timer_irqs++;
 473
 474         /*
 475          * NOTE! We'd better ACK the irq immediately,
 476          * because timer handling can be slow.
 477          */
 478         ack_APIC_irq();
 479
 480         /*
 481          * update_process_times() expects us to have done irq_enter().
 482          * Besides, if we don't timer interrupts ignore the global
 483          * interrupt lock, which is the WrongThing (tm) to do.
 484          */
 485         irq_enter();
 486         vmi_local_timer_interrupt(cpu);
 487         irq_exit();
 488         set_irq_regs(old_regs);
 489 }
 490
 491 #endif  /* CONFIG_X86_LOCAL_APIC */