Skip to content

GitLab

Open
Created by Jeffrey Lee@jleeDeveloper
Report abuse

WIP: Nanosecond resolution monotonic timer

This set of changes aims to upgrade the OS's time keeping abilities, by adding support for a 64 bit, nanosecond-resolution monotonic time source. This includes:

  • A SWI to get the time (OS_ReadMonotonicTime64)
  • The ability to schedule one-shot and repeating events using the 64 bit timer (like OS_CallAfter / OS_CallEvery), via the OS_TimerControl SWI
  • The ability to fine-tune the timer rate (via the OS_TimerControl SWI), to allow the RTC module or similar to apply clock corrections
  • The ability to read the unscaled, 64 bit hardware timer value at its native frequency (via OS_TimerControl 4/5), which could be useful either for code to know roughly how accurate the nanosecond timestamps are, or for a way of reading the time with lower overheads (e.g. for profiling)
  • Definition of the new 'timer' HAL device, which HALs are expected to implement in order to provide the OS with the necessary functionality for the new system to function
    • If the device is missing, the new APIs will not be available and HAL timer 0 will be used for timekeeping, as before
  • If the HAL device is present, then:
    • The old time keeping systems (OS_ReadMonotonicTime, TickerV, OS_CallAfter, etc.) will all be driven by the new timer
      • In particular, OS_ReadMonotonicTime will (normally) be in sync with OS_ReadMonotonicTime64, so there's minimal risk of software getting confused about the current time if it's using a mix of the old and new APIs
    • The kernel will intercept the HAL timer calls and remap the timers so that the numbering starts at 1 (i.e. RISC OS software which talks to HAL timer 1 will actually talk to HAL timer 0)
    • If RISC OS software talks to HAL timer 0 (or uses the HAL counter calls) then it will actually talk to an emulation of timer 0 which reports the value of the new 64 bit timer; this is to ensure compatibility with software which computes high resolution timestamps by combining OS_ReadMonotonicTime with the HAL timer 0 value

See Doc/NanoTime for more in-depth documentation, Dev/NanoTime, and https://www.riscosopen.org/forum/forums/3/topics/11109 for the original forum thread.

Other changed components:

Important TODO items:

  • Official allocation for the new SWIs
  • Better errors & error checking
  • Use chocolate blocks for timernode_t allocation - done
  • Modifications to cope with systems where the timer rate may change? (e.g. Cortex-A9 system timer is affected by CPU speed, but it looks like the HALs can mostly deal with that themselves) - doesn't look like this will be needed
  • Change to nanosecond resolution instead of microsecond - done
  • Optimise HAL devices by making the kernel be responsible for scheduling any regular interrupt that's guard against overflow of the underlying timer device (many platforms only have a 32 bit hardware timer, so regular interrupts or timer reads are needed to protect against overflow causing backwards time travel when the HAL extends the time to 64 bits) - done
Edited by Jeffrey Lee