netbsd: Simplify weak lookup.

Remove LazyPtr. Avoid constructing an invalid pointer as a sentinel to
indicate that the pointer is uninitialized. Now, a null pointer means
it is uninitialized, and a non-null pointer means it is initialized.
This is less questionable from a safety perspective, and should also be
more efficient.

Reduce duplication between the "getrandom is available" and the
fallback case.

src/lazy.rs

@@ -1,10 +1,7 @@
 //! Helpers built around pointer-sized atomics.
 #![cfg(target_has_atomic = "ptr")]
 #![allow(dead_code)]
-use core::{
-    ffi::c_void,
-    sync::atomic::{AtomicPtr, AtomicUsize, Ordering},
-};
+use core::sync::atomic::{AtomicUsize, Ordering};
 
 // This structure represents a lazily initialized static usize value. Useful
 // when it is preferable to just rerun initialization instead of locking.
@@ -67,58 +64,3 @@ impl LazyBool {
         self.0.unsync_init(|| usize::from(init())) != 0
     }
 }
-
-// This structure represents a lazily initialized static pointer value.
-///
-/// It's intended to be used for weak linking of a C function that may
-/// or may not be present at runtime.
-///
-/// Based off of the DlsymWeak struct in libstd:
-/// https://github.com/rust-lang/rust/blob/1.61.0/library/std/src/sys/unix/weak.rs#L84
-/// except that the caller must manually cast self.ptr() to a function pointer.
-pub struct LazyPtr {
-    addr: AtomicPtr<c_void>,
-}
-
-impl LazyPtr {
-    /// A non-null pointer value which indicates we are uninitialized.
-    ///
-    /// This constant should ideally not be a valid pointer. However,
-    /// if by chance initialization function passed to the `unsync_init`
-    /// method does return UNINIT, there will not be undefined behavior.
-    /// The initialization function will just be called each time `get()`
-    /// is called. This would be inefficient, but correct.
-    const UNINIT: *mut c_void = !0usize as *mut c_void;
-
-    /// Construct new `LazyPtr` in uninitialized state.
-    pub const fn new() -> Self {
-        Self {
-            addr: AtomicPtr::new(Self::UNINIT),
-        }
-    }
-
-    // Runs the init() function at most once, returning the value of some run of
-    // init(). Multiple callers can run their init() functions in parallel.
-    // init() should always return the same value, if it succeeds.
-    pub fn unsync_init(&self, init: impl Fn() -> *mut c_void) -> *mut c_void {
-        #[cold]
-        fn do_init(this: &LazyPtr, init: impl Fn() -> *mut c_void) -> *mut c_void {
-            let addr = init();
-            this.addr.store(addr, Ordering::Release);
-            addr
-        }
-
-        // Despite having only a single atomic variable (self.addr), we still
-        // cannot always use Ordering::Relaxed, as we need to make sure a
-        // successful call to `init` is "ordered before" any data read through
-        // the returned pointer (which occurs when the function is called).
-        // Our implementation mirrors that of the one in libstd, meaning that
-        // the use of non-Relaxed operations is probably unnecessary.
-        let val = self.addr.load(Ordering::Acquire);
-        if val != Self::UNINIT {
-            val
-        } else {
-            do_init(self, init)
-        }
-    }
-}

src/netbsd.rs

@@ -1,15 +1,31 @@
 //! Implementation for NetBSD
-use crate::{lazy::LazyPtr, util_libc::sys_fill_exact, Error};
-use core::{ffi::c_void, mem::MaybeUninit, ptr};
+use crate::{util_libc::sys_fill_exact, Error};
+use core::{
+    cmp,
+    ffi::c_void,
+    mem::MaybeUninit,
+    ptr,
+    sync::atomic::{AtomicPtr, Ordering},
+};
+
+unsafe extern "C" fn polyfill_using_kern_arand(
+    buf: *mut u8,
+    buflen: libc::size_t,
+    flags: libc::c_uint,
+) -> libc::ssize_t {
+    debug_assert_eq!(flags, 0);
 
-fn kern_arnd(buf: &mut [MaybeUninit<u8>]) -> libc::ssize_t {
     static MIB: [libc::c_int; 2] = [libc::CTL_KERN, libc::KERN_ARND];
-    let mut len = buf.len();
+
+    // NetBSD will only return up to 256 bytes at a time, and
+    // older NetBSD kernels will fail on longer buffers.
+    let mut len = cmp::min(buflen, 256);
+
     let ret = unsafe {
         libc::sysctl(
             MIB.as_ptr(),
             MIB.len() as libc::c_uint,
-            buf.as_mut_ptr().cast::<c_void>(),
+            buf.cast::<c_void>(),
             &mut len,
             ptr::null(),
             0,
@@ -25,27 +41,35 @@ fn kern_arnd(buf: &mut [MaybeUninit<u8>]) -> libc::ssize_t {
 type GetRandomFn = unsafe extern "C" fn(*mut u8, libc::size_t, libc::c_uint) -> libc::ssize_t;
 
 // getrandom(2) was introduced in NetBSD 10.0
-static GETRANDOM: LazyPtr = LazyPtr::new();
+static GETRANDOM: AtomicPtr<c_void> = AtomicPtr::new(ptr::null_mut());
 
-fn dlsym_getrandom() -> *mut c_void {
+#[cold]
+fn init() -> *mut c_void {
     static NAME: &[u8] = b"getrandom\0";
     let name_ptr = NAME.as_ptr().cast::<libc::c_char>();
-    unsafe { libc::dlsym(libc::RTLD_DEFAULT, name_ptr) }
+    let ptr = match unsafe { libc::dlsym(libc::RTLD_DEFAULT, name_ptr) } {
+        Some(ptr) => ptr,
+        None => unsafe {
+            core::mem::transmute::<GetRandomFn, *mut c_void>(polyfill_using_kern_arand)
+        },
+    };
+    GETRANDOM.store(ptr, Ordering::Release);
+    ptr
 }
 
 pub fn getrandom_inner(dest: &mut [MaybeUninit<u8>]) -> Result<(), Error> {
-    let fptr = GETRANDOM.unsync_init(dlsym_getrandom);
-    if !fptr.is_null() {
-        let func: GetRandomFn = unsafe { core::mem::transmute(fptr) };
-        return sys_fill_exact(dest, |buf| unsafe {
-            func(buf.as_mut_ptr().cast::<u8>(), buf.len(), 0)
-        });
-    }
-
-    // NetBSD will only return up to 256 bytes at a time, and
-    // older NetBSD kernels will fail on longer buffers.
-    for chunk in dest.chunks_mut(256) {
-        sys_fill_exact(chunk, kern_arnd)?
-    }
-    Ok(())
+    // Despite being only a single atomic variable, we still cannot always use
+    // Ordering::Relaxed, as we need to make sure a successful call to `init`
+    // is "ordered before" any data read through the returned pointer (which
+    // occurs when the function is called). Our implementation mirrors that of
+    // the one in libstd, meaning that the use of non-Relaxed operations is
+    // probably unnecessary.
+    let fptr = match GETRANDOM.load(Ordering::Acquire) {
+        Some(fptr) => fptr,
+        None => init(),
+    };
+    let fptr = unsafe { core::mem::transmute::<*mut c_void, GetRandomFn>(fptr) };
+    sys_fill_exact(dest, |buf| unsafe {
+        fptr(buf.as_mut_ptr().cast::<u8>(), buf.len(), 0)
+    })
 }