Link
WebDoc
Search

Timing

Table of contents

  1. Circuits
  2. jiffies and xtime
  3. Sub-Tick Resolution
  4. Interrupt
  5. Software Timer

Circuits

Name PrecisionI/O portinterrupt
Real Time Clock (RTC)independent hardware with a small battery. can be accessed by /dev/rtc device file0.12 ~ 500 ms0x70-0x80IRQ8
Time Stamp Counter (TSC)in CPU; increased at each clock signal. can be read by instruction rdtscdepend on CPU clock ticks  
Programmable Interval Timer (PIT)may be adjusted if the computer is synchronized with an external clock.1 ~ 10 ms0x40-0x43IRQ0; to all CPUs
CPU Local Timerat local APIC. can be programmed to issue interrupts at very low freq.Based on bus clock signal vector 239 (0xef); local interrupt
High Precision Event Timer (HPET)can be programmed100 ~ ? nsmemory-map I/O 
ACPI Power Management Timer (ACPI PMT)when the CPU voltage is lowered, the frequency does not change (TSC does!).fixed: 279 nsthe port is determined by BIOS 
  • In a uniprocessor system, all time-keeping activities are triggered by interrupts raised by global timer.
  • In a multi-process system,
    • the general activities are triggered by interrupts raised by global timer (the interrupt handler is timer_interrupt()),
    • while the CPU-specific activities are triggered by local APIC timer (the interrupt handler is apic_timer_interrupt()).

The Linux kernel tries to set the hardware to raise interrupts on IRQ0 (each interrupt is called a tick) and at a frequency of 1000 Hz. In addition, it tries to use the best timer source available in the system.

jiffies and xtime

jiffies is a 32-bit counter that stores the number of elapsed ticks since the system started. It is increased by one when a timer interrupt occurs. It wraps around in appropriate 50 days (2^32 / 1000 / 86400). It is initialized to -300,000, so that it will overflow in 5 minutes and the buggy code will show up.

There is another counter, a 64-bit jiffies_64. The kernel needs Seqlock to protect it in 32-bit architecture.

xtime stores:

  • tv_sec: seconds since January 1, 1970 UTC
  • tv_nsec: nanoseconds whin the last second

It is usually updated once in a tick.

The return values of the syscalls gettimeofday() / time() are calculated depenging on these two variables.

Sub-Tick Resolution

The kenrl can achieve sub-tick resolution by consulting the hardware. This is called time interpolation. In addition, The kernel can delay for a given nubmer of loops. Here is a table showing the resolution:

hardwareinterpolationdelay
HPETHPETHPET
ACPI PMTACPI PMTTSC
TSCTSCTSC
PITPITTight loop

Sub-Tick Resolution is usually used by device drivers.

Interrupt

On every timer interrupt, the kernel:

  • Checks how long the current process has been running. Update its time spent on Kernel mode (stime) and User mode (utime).
  • Profile the kernel code. It fetches the value of eip to discover what the kernel was doing before interrupt. In the long run, the samples accumulated on the hot spots.
  • Check the NMI watchdog. The apic_timer_irqs of the nth entry of the irq_stat array is increased by one on every local APIC timer interrupt. The watchdog chcks if the timer is increased properly. If not, there might be a deadlock when the interrupts are disabled.

Software Timer

The checking for the software timer is done by deferrable functions, so the kernel can only ensure the timer are executed with in a delay of up to a few hundred of milliseconds. Therefore, the software timer is not appropriate for real-time applications.

The Linux kernel use time wheel for dynamic timers. The data structure is a per-CPU variable named tvec_bases. Each element of it is a time wheel. The time wheel resolutions are 255, 2^14, 2^20, 2^26, 2^32 ticks, respectively.

Use nanosleep() syscall to register a software timer and suspend the processor.