mirror of
https://github.com/apache/nuttx.git
synced 2026-10-11 00:00:24 +00:00
f6ecf80ebb
The documentation grew one page at a time, so the tree follows the
history of who wrote what and not the shape of NuttX. Scheduling is
spread over three places, a driver page can sit above the subsystem
that owns it, and the front page lists everything at the same level.
That is a lot to face when all you want to know is where the scheduler
lives.
This change files every page under the code it describes. It is a move,
not a rewrite: outside the ten pages named below, every page keeps the
text that is already in master, and no page's text is deleted.
What it does:
* Groups the table of contents into nine chapters.
* Moves the OS subsystems under os/: scheduling, memory, drivers,
filesystem, networking, IPC, interrupts, libs, time.
* Renames the platform pages to the names the source tree uses, and
derives their tags from the tree instead of by hand.
* Splits guides/ by subject.
* Adds Documentation/redirects.py, with a rule for every page that left
its old path, so old URLs keep working. The redirect page also carries
a link's #anchor across to the new page.
Ten pages have text that is new or rewritten. Nine of them are the
landing page of a chapter, which has to exist for the new structure:
index the front page
os/index OS Design
os/scheduling/index Scheduling
os/interrupts/index Interrupts
os/ipc/index IPC
os/time/index Time and timers
about/index About
developing/index Developing NuttX
ReleaseNotes/index Release notes
The tenth is os/libs/libbuiltin, the only page here with technical
content: libs/libbuiltin/ had no page at all. Five SVG diagrams come
with these pages, hand-written XML with no editor metadata.
Nothing outside Documentation/ is touched.
How it was checked:
* Sphinx builds with -W: no warnings, and no document left outside a
toctree.
* A script, offered in the PR, proves the narrow claim this rests on.
For every page outside the ten named above it erases what a move
touches -- link target, path, tag line, toctree block, table border --
from the whole old text and the whole new text, and requires the two
to be byte for byte identical. It also requires every sentence of a
deleted page to turn up somewhere, and every page that left its old
path to have a redirect, from a URL that existed, to where its content
went. It exits non-zero and names the page if any of that is not true,
and it tests added pages too, so forgetting to declare one cannot make
it pass.
* An independent audit checked 133 factual claims on these ten pages
against the tree, one shell command per claim: 130 confirmed, 1
refuted and fixed here, 2 not checkable.
* tools/checkpatch.sh is clean over the range.
The diff is large because moving a page changes every link that points
to it. Most of it is pure renames, and board pages that gained one tag
line.
Assisted-by: Claude:claude-opus-5
50 lines
2.3 KiB
ReStructuredText
50 lines
2.3 KiB
ReStructuredText
================
|
|
Time and Timers
|
|
================
|
|
|
|
The system clock, the timers built on it, and what happens when a delay is
|
|
shorter than a clock tick. The code lives in ``sched/clock/``,
|
|
``sched/timer/``, ``sched/wdog/`` and ``sched/hrtimer/``.
|
|
|
|
Three things shape everything on this page.
|
|
|
|
The first is whether the system runs on a periodic tick or
|
|
:doc:`tickless <tickless_os>`. By default a periodic timer interrupt drives
|
|
system time; tickless replaces it with an interval timer programmed for the
|
|
next OS event, so the system can stay asleep in between. It is not a free
|
|
choice: ``CONFIG_SCHED_TICKLESS`` depends on ``CONFIG_ARCH_HAVE_TICKLESS``,
|
|
and a tickless port has additional interfaces to implement, declared in
|
|
``include/nuttx/arch.h``.
|
|
|
|
The second is ``CONFIG_USEC_PER_TICK``, the length of one tick -- and it
|
|
matters in *both* configurations, which is the part that catches people out.
|
|
Without tickless it is the interval at which the hardware interrupts the OS,
|
|
10 ms by default. With tickless there are no such interrupts and it controls
|
|
no timer rate at all, but it still sets the resolution of the time that
|
|
``clock_systime_ticks()`` reports, and of the delays you can ask for from
|
|
watchdog timers and delayed work. Its default simply drops to 100 µs.
|
|
|
|
That makes the tick a trade-off rather than a dial to turn down. The count is
|
|
held in an ``unsigned int`` -- 32 bits on most targets, 16 on some -- so a
|
|
smaller tick buys resolution at the cost of the longest delay that can be
|
|
represented: the 100 µs default reaches about 120 hours. It should also never
|
|
be set below the resolution of the underlying timer.
|
|
|
|
The third is that neither of those bounds you when you need timing finer than
|
|
a tick, and this section holds two different answers. :doc:`Short delays
|
|
<short_time_delays>` covers the counter-intuitive things that happen when a
|
|
requested delay is near or below one tick -- a discussion that applies under
|
|
tickless too, only in different terminology. The high-resolution timer in
|
|
``sched/hrtimer/`` (``CONFIG_HRTIMER``) is the other answer, offering
|
|
nanosecond-level precision; it is described under
|
|
:doc:`System Time and Clock <time_clock>`. Its callbacks run in interrupt
|
|
context, which is the price.
|
|
|
|
.. toctree::
|
|
:maxdepth: 1
|
|
|
|
time_clock.rst
|
|
tickless_os.rst
|
|
short_time_delays.rst
|
|
oneshot_timers_and_cpu_load.rst
|
|
sleep.rst
|