On Fri, 14 Feb 2025 11:37:40 +0000
Gary Guo [off-list ref] wrote:

On Fri, 14 Feb 2025 13:05:30 +0900 (JST)
FUJITA Tomonori [off-list ref] wrote:

quoted

On Sun, 9 Feb 2025 16:20:48 +0000
Gary Guo [off-list ref] wrote:

quoted

quoted

+fn might_sleep(loc: &Location<'_>) {
+    // SAFETY: FFI call.
+    unsafe {
+        crate::bindings::__might_sleep_precision(
+            loc.file().as_ptr().cast(),
+            loc.file().len() as i32,
+            loc.line() as i32,
+        )
+    }
+}  

One last Q: why isn't `might_sleep` marked as `track_caller` and then
have `Location::caller` be called internally?

It would make the API same as the C macro.  

Equivalent to the C side __might_sleep(), not might_sleep(). To avoid
confusion, it might be better to change the name of this function.

The reason why __might_sleep() is used instead of might_sleep() is
might_sleep() can't always be called. It was discussed in v2:

https://lore.kernel.org/all/ZwPT7HZvG1aYONkQ@boqun-archlinux/ (local)

I don't follow. `__might_sleep` or `might_sleep` wouldn't make a
difference here, given that this function may actually sleep.

Yeah, it doesn't matter here. If I understand correctly, the
discussion is about whether might_sleep() itself should be unsafe
considering the case where it is called from other functions. I simply
chose uncontroversial __might_sleep().

After reviewing the code again, I realized that I made a mistake;
__might_sleep() should only be executed when CONFIG_DEBUG_ATOMIC_SLEEP
is enabled. I also think that it is confusing that might_sleep() calls
C's __might_sleep().

How about implementing the equivalent to might_sleep()?

/// Annotation for functions that can sleep.
///
/// Equivalent to the C side [`might_sleep()`], this function serves as
/// a debugging aid and a potential scheduling point.
///
/// This function can only be used in a nonatomic context.
#[track_caller]
fn might_sleep() {
    #[cfg(CONFIG_DEBUG_ATOMIC_SLEEP)]
    {
        let loc = core::panic::Location::caller();
	// SAFETY: FFI call.
	unsafe {
	    crate::bindings::__might_sleep_precision(
	        loc.file().as_ptr().cast(),
	        loc.file().len() as i32,
                loc.line() as i32,
	    )
	}
    }
    // SAFETY: FFI call.
    unsafe { crate::bindings::might_resched() }
}

quoted

quoted

Also -- perhaps this function can be public (though I guess you'd need
to put it in a new module).  

Wouldn't it be better to keep it private until actual users appear?

I'll make the above public if you think that is the better approach.

C's might_sleep() is defined in linux/kernel.h but kernel/kernel.rs
isn't a good choice, I guess. kernel/sched.rs or other options?