This commit added the guard timer for hrtimer.
The guard timer uses a small memory footprint, offering two main advantages:
- Reduced branches checking for an empty hrtimer queue, simplifying code implementation and improving the performance.
- Additional health monitoring allows the system to enter a safe state in case of time acquisition errors, and supports custom error handling callback functions.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
If there are no timers in the hrtimer queue, we should cancel the timers to avoid unnecessary timer interruptions. Since we currently do not have a timer cancellation interface, we can achieve cancellation by setting the timer to its maximum value.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
Determining whether a red-black tree node is the left-mode node using `RB_LEFT(hrtimer, node) == NULL` is functionally incorrect. This is because the left node of any leaf node can also be NULL. For example, in the following rbtree:
5
/ \
3 7
\
4
the left node of the right-most node 7 would also be NULL.
To avoid extra performance overhead to find the left-most node when the
rb-tree changed, this commit used `g_cached_first` to cache the
left-most node.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
This commit fixed the overflow check and renamed the period to delay, since the callback return value is the next delay not the period.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
When a hrtimer is removed and then re-enqueued, if the removed hrtimer is the head node and the re-enqueued node is not the head node, the hardware timer still needs to be reset. This patch fixes this issue and simplifies the enqueuing.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
This commit simplified the rbtree in hrtimer and provided better
branchless compare function.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
This commit renamed the hrtimer_is_armed to hrtimer_is_pending, which is
more accurate in the semamtic, and simplify it.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
This commit inlined the `hrtimer_start` to allow the compiler to optimize at least 1 branch in
hrtimer_start.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
The hrtimer subsystem is independent of the OS tick mechanism and tickless
mode. It must always be able to reprogram the hardware timer to achieve
nanosecond-level resolution.
A recent update restricted hardware timer reprogramming to tickless mode
only. As a result, hrtimer no longer functions correctly when the scheduler
is running in tick-based mode.
This change removes the incorrect dependency on tickless mode and restores
proper hrtimer operation.
Signed-off-by: Chengdong Wang <wangchengdong@lixiang.com>
This commit fixed the functional correctness issue in hrtimer_start by
adding HRTIMER_MAX_DELAY fr the HRTIMER_MODE_REL.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
This commit added hrtimer_gettime API to support the query of the
rest of the delay time of the hrtimer in nanoseconds.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
This commit move the hrtimer_gettime to clock_systime_nsec to improve
the code reusability.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
On 32-bit platform with 64-bit pointer, read/write to the
`g_hrtimer_running` is not atomic, we should protect it with the
seqcount to ensure the value we read is correct.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
Since the hrtimer->expired is not monotonic, it can not be used as the
version. This commit added the ownership encoding to ensure the
invariant that the cancelled the timer can not modify the hrtimer again.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
This commit fixed the functional correctness issues in tick mode
and non-tickless alarm mode.
- In tick mode, we do not need reprogram the timer.
- In non-tickless alarm mode, the up_timer_start receive the relative time as the parameter.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
Several code comments in the hrtimer subsystem have become outdated
following recent implementation changes. This patch updates them to
accurately describe the current code behavior and algorithms.
Signed-off-by: Chengdong Wang <wangcd91@gmail.com>
Add support for managing hrtimers using a simple list. This approach
is more memory-efficient, as list nodes use less memory, and it is
preferable to an RB-tree when the number of hrtimers is relatively small.
Signed-off-by: Chengdong Wang <wangchengdong@lixiang.com>
The hrtimer is designed to replace the wdog timer in future, so it
should have the callback function in its API. The evaluation results showed there is no any
performance degradation since the CPU pipeline can hide the latency.
This commit also decoupled the rb-tree implementation with the pending
state checking.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
Separate SMP-specific logic from the hrtimer core to improve
performance and maintainability.
Signed-off-by: Chengdong Wang <wangchengdong@lixiang.com>
This commit renamed the callback type to `hrtimer_entry_t`, aligning with
the `wdentry_t` in wdog..
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
The expired field can not be used to indicate the cancelled state. Because
the `UINT64_MAX` is valid expired time. A periodic hrtimer started with
the hrtimer_start_absolute(hrtimer, period_func, UINT64_MAX) can
not be cancelled via `hrtimer_cancel`.
Signed-off-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
Allow running/armed hrtimer to be restarted to
fix hrtimer bug: #17567
Co-authored-by: ouyangxiangzhen <ouyangxiangzhen@xiaomi.com>
Signed-off-by: Chengdong Wang <wangchengdong@lixiang.com>
Add a safe synchronous hrtimer cancel API, hrtimer_cancel_sync().
If the timer callback is currently executing, this function waits
until the callback has completed and the timer state transitions
to HRTIMER_STATE_INACTIVE.
Signed-off-by: Chengdong Wang <wangchengdong@lixiang.com>
NuttX provides the wdog module to implement timers for the scheduler.
While it is lightweight and efficient, wdog only supports tick-level timers,
typically in milliseconds. Although the tick duration can be configured,
setting it to microsecond or nanosecond resolution is not practical.
Doing so may cause an interrupt storm, where the CPU is constantly
occupied handling tick interrupts.
To address this limitation, a new hrtimer module is introduced.
It coexists with the wdog module, providing both tick-level timers
for wdog and high-resolution timers (nanosecond-level) directly to users.
Signed-off-by: Chengdong Wang <wangchengdong@lixiang.com>
