huojianchaoyy opened a new pull request, #21654: URL: https://github.com/apache/nuttx/pull/21654
*Note: Please adhere to [Contributing Guidelines](https://github.com/apache/nuttx/blob/master/CONTRIBUTING.md).* ## Summary This PR updates the critical section monitor (`sched/sched/sched_critmonitor.c`) and the cpuload accounting built on it: - **sched: divide nxsched_switch_critmon into two steps: suspend and resume**: after suspend and resume were unified into `nxsched_switch_critmon()`, it is no longer possible to guarantee that both `from` and `to` are valid at the call point, which changed the semantics of `nxsched_switch_context()`. Split it back into `nxsched_suspend_critmon()` and `nxsched_resume_critmon()`, and update `Documentation/os/interrupts/critical_sections.rst`. - **sched: add running time record tracepoint**: new `CONFIG_SCHED_INSTRUMENTATION_THREADTIME` records a `NOTE_DUMP_THREADTIME` note with the run time of the thread that is switched out. - **note: add runtime record threshold control**: new `CONFIG_SCHED_INSTRUMENTATION_THREAD_RUNTIME_DURATION` (us) filters out runs shorter than the threshold. - **sched: thread time recording ignores idle threads**. - **sched: Optimize threadtime statistics performance**: compare against a threshold converted to perf counts once, and only divide when a note is recorded. - **sched: optimize critmon_cpuload performance**: with `CONFIG_SCHED_CPULOAD_CRITMONITOR`, account the cpuload in raw perf counts instead of dividing on every context switch. The decay threshold is converted to perf counts once. - **sched/critmonitor: fix cpuload tick unit in nxsched_update_critmon**: `nxsched_update_critmon()` also passes raw perf counts, so both paths use the same unit. - **sched: fix critmonitor print format error**: `clock_t` is 64-bit, so the `CHECK_*` messages print it with `%ju` and a `uintmax_t` cast instead of `PRIu32`. - **sysinfo: Modify the types of time-related variables to meet MISRA-C2012 10.3.**: explicit casts in `sysinfo()`. The decayed cpuload total is accumulated in `clock_t`, as perf counts can exceed 32 bits. ## Impact - New Kconfig options, both off by default: - `SCHED_INSTRUMENTATION_THREADTIME`, depends on `SCHED_INSTRUMENTATION_DUMP` and `SCHED_CRITMONITOR`. It uses the thread run time, which is only tracked with `SCHED_CRITMONITOR_MAXTIME_THREAD >= 0`; with `-1` no note is recorded. - `SCHED_INSTRUMENTATION_THREAD_RUNTIME_DURATION`, default 0. - `include/nuttx/sched_note.h`: new note type `NOTE_DUMP_THREADTIME`, appended after `NOTE_DUMP_COUNTER` so existing values do not change, plus `struct note_threadtime_s` and `sched_note_threadtime()`. The noteram text dump does not decode the new type and skips it like other unknown types. - `nxsched_switch_critmon()` is replaced by `nxsched_suspend_critmon()` and `nxsched_resume_critmon()`. Both are internal to `sched/`; the only caller is `nxsched_switch_context()`. - With `CONFIG_SCHED_CPULOAD_CRITMONITOR`, the `total` and `active` fields of `struct cpuload_s` are in perf counts instead of system ticks. `ps`, `/proc/cpuload`, `/proc/<pid>/loadavg` and `sysinfo()` use the ratio, so the reported load is unchanged. Other cpuload backends are not affected. - Without `CONFIG_SCHED_CRITMONITOR`, the only changes are the casts in `sysinfo()` and the wider accumulator in the cpuload decay. No defconfig changes. ## Testing Host: Ubuntu 22.04.5 LTS x86_64, gcc 13.4.0, CMake 3.23 + Ninja. Every commit of the series was built in three configurations: - `sim:nsh` (UP) and `sim:smp` (4 CPUs), with: ``` CONFIG_SCHED_CRITMONITOR=y CONFIG_SCHED_CRITMONITOR_MAXTIME_THREAD=2000000000 CONFIG_SCHED_CRITMONITOR_MAXTIME_PREEMPTION=2000000000 CONFIG_SCHED_CRITMONITOR_MAXTIME_CSECTION=2000000000 CONFIG_SCHED_CRITMONITOR_MAXTIME_BUSYWAIT=2000000000 # sim:smp only CONFIG_SCHED_CPULOAD_CRITMONITOR=y CONFIG_SCHED_INSTRUMENTATION=y CONFIG_SCHED_INSTRUMENTATION_DUMP=y CONFIG_SCHED_INSTRUMENTATION_THREADTIME=y CONFIG_SCHED_INSTRUMENTATION_THREAD_RUNTIME_DURATION=100 CONFIG_DRIVERS_NOTE=y CONFIG_DRIVERS_NOTERAM=y ``` - `sim:nsh` with `CONFIG_SCHED_CRITMONITOR_MAXTIME_THREAD=-1` and `CONFIG_SCHED_INSTRUMENTATION_THREADTIME=y`. `./tools/checkpatch.sh -c -m -g <base>..HEAD` passes. ostest passes on the last commit, on both sim:nsh and sim:smp: <details> <summary>sim:nsh</summary> ``` user_main: Exiting ostest_main: Exiting with status 0 nsh> ps TID PID PPID PRI POLICY TYPE NPX STATE EVENT SIGMASK STACK CPU COMMAND 0 0 9 0 FIFO Kthread - Ready 0000000000000000 0069584 99.9% Idle_Task 1 0 9 224 FIFO Kthread - Waiting Semaphore 0000000000000000 0067448 0.0% sim_loop_wq 0x73ae4be003f0 0x73ae4be00478 2 0 9 224 FIFO Kthread - Waiting Semaphore 0000000000000000 0067472 0.0% hpwork 0x40193800 0x40193888 4 4 0 100 FIFO Task - Waiting Semaphore 0000000000000000 0067504 0.0% init_main 7 7 4 100 FIFO Task - Running 0000000000000000 0067496 0.0% sh nsh> cat /proc/cpuload 0.1% ``` </details> <details> <summary>sim:smp</summary> ``` user_main: Exiting ostest_main: Exiting with status 0 nsh> ps TID PID PPID CPU PRI POLICY TYPE NPX STATE EVENT SIGMASK STACK USED FILLED CPU COMMAND 0 0 10 0 0 FIFO Kthread - Assigned 0000000000000000 0066512 0005544 8.3% 22.1% CPU0 IDLE 1 0 10 1 0 FIFO Kthread - Running 0000000000000000 0066504 0009448 14.2% 22.1% CPU1 IDLE 2 0 10 2 0 FIFO Kthread - Running 0000000000000000 0066512 0009736 14.6% 27.6% CPU2 IDLE 3 0 10 3 0 FIFO Kthread - Running 0000000000000000 0066504 0009944 14.9% 28.1% CPU3 IDLE 4 0 10 --- 224 FIFO Kthread - Waiting Semaphore 0000000000000000 0067448 0000840 1.2% 0.0% sim_loop_wq 0x731792a30fe0 0x731792a31068 5 0 10 --- 224 FIFO Kthread - Waiting Semaphore 0000000000000000 0067472 0001640 2.4% 0.0% hpwork 0x400a0080 0x400a0108 7 7 0 --- 100 FIFO Task - Waiting Semaphore 0000000000000000 0067504 0002728 4.0% 0.0% init_main 8 8 7 0 100 FIFO Task - Running 0000000000000000 0067496 0004088 6.0% 0.0% sh nsh> cat /proc/cpuload 0.0% ``` </details> The threadtime note was checked under GDB on sim:nsh, with breakpoints on the threshold check in `nxsched_suspend_critmon()` and on `sched_note_event_ip()`. Only non-idle threads that ran longer than 100 us record a `NOTE_DUMP_THREADTIME` note (`elapsed` is in perf counts, ns on sim): ``` suspend#1: pid=0 elapsed=2976489 threshold=100000 record=False suspend#2: pid=3 elapsed=497038 threshold=100000 record=True -> NOTE_DUMP_THREADTIME recorded, elapsed_us=497 suspend#3: pid=1 elapsed=51160 threshold=100000 record=False suspend#4: pid=2 elapsed=4727 threshold=100000 record=False suspend#5: pid=4 elapsed=234640 threshold=100000 record=True -> NOTE_DUMP_THREADTIME recorded, elapsed_us=234 ``` -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
