Use smallvec for dtor lists. (#93)

Use `smallvec` for `at_exit` and `at_thread_exit` dtor lists. This helps
us follow the POSIX rule that at least 32 functions be registerable, and
helps origin run in settings where thread destructors are registered
from within a malloc implementation.

Cargo.toml

@@ -40,6 +40,7 @@ errno = { version = "0.3.3", default-features = false, optional = true }
 core = { version = "1.0.0", optional = true, package = "rustc-std-workspace-core" }
 alloc = { version = "1.0.0", optional = true, package = "rustc-std-workspace-alloc" }
 compiler_builtins = { version = "0.1.101", optional = true }
+smallvec = { version = "1.11.1", features = ["const_new"] }
 
 [target.'cfg(not(target_arch = "arm"))'.dependencies.unwinding]
 version = "0.2.0"

src/program.rs

@@ -31,8 +31,6 @@
 use crate::thread::{initialize_main_thread, initialize_startup_thread_info};
 #[cfg(feature = "alloc")]
 use alloc::boxed::Box;
-#[cfg(all(feature = "alloc", feature = "origin-program"))]
-use alloc::vec::Vec;
 #[cfg(not(feature = "origin-program"))]
 use core::ptr::null_mut;
 use linux_raw_sys::ctypes::c_int;
@@ -219,8 +217,14 @@ unsafe fn init_runtime(mem: *mut usize, envp: *mut *mut u8) {
 }
 
 /// Functions registered with [`at_exit`].
+///
+/// [POSIX guarantees] at least 32 handlers can be registered, so use a
+/// `SmallVec` to ensure we can register that many without allocating.
+///
+/// [POSIX guarantees]: https://pubs.opengroup.org/onlinepubs/9699919799/functions/atexit.html
 #[cfg(all(feature = "alloc", feature = "origin-program"))]
-static DTORS: Mutex<Vec<Box<dyn FnOnce() + Send>>> = Mutex::new(Vec::new());
+static DTORS: Mutex<smallvec::SmallVec<[Box<dyn FnOnce() + Send>; 32]>> =
+    Mutex::new(smallvec::SmallVec::new_const());
 
 /// Register a function to be called when [`exit`] is called.
 ///

src/thread/linux_raw.rs

@@ -2,7 +2,6 @@
 
 use crate::arch::{clone, munmap_and_exit_thread, set_thread_pointer, thread_pointer, TLS_OFFSET};
 use alloc::boxed::Box;
-use alloc::vec::Vec;
 use core::any::Any;
 use core::cmp::max;
 use core::ffi::c_void;
@@ -63,7 +62,9 @@ struct ThreadData {
     stack_size: usize,
     guard_size: usize,
     map_size: usize,
-    dtors: Vec<Box<dyn FnOnce()>>,
+
+    // Support a few dtors before using dynamic allocation.
+    dtors: smallvec::SmallVec<[Box<dyn FnOnce()>; 4]>,
 }
 
 // Values for `ThreadData::detached`.
@@ -87,7 +88,7 @@ impl ThreadData {
             stack_size,
             guard_size,
             map_size,
-            dtors: Vec::new(),
+            dtors: smallvec::SmallVec::new(),
         }
     }
 }