hook write

Author: loongs-zhang

core/src/syscall/unix/write.rs

@@ -11,9 +11,38 @@ trait WriteSyscall {
     ) -> ssize_t;
 }
 
-impl_syscall!(WriteSyscallFacade, IoUringWriteSyscall, NioWriteSyscall, RawWriteSyscall,
-    write(fd: c_int, buf: *const c_void, len: size_t) -> ssize_t
-);
+//在最顶层对stdout/stderr/重入写入做早期旁路：直接调用原始系统调用，
+//跳过整个facade链(WriteSyscallFacade/IoUring/NIO)，最小化每次info!()
+//调用write()时的函数调用开销。在QEMU等慢速平台上，每个额外的函数调用
+//可能耗时0.5-1ms，累积的开销会导致协程在10ms抢占窗口内无法完成工作。
+// Early bypass at the top-level dispatcher for stdout/stderr/re-entrant writes:
+// call the raw syscall directly, skipping the entire facade chain
+// (WriteSyscallFacade/IoUring/NIO). This minimizes function call overhead
+// per info!() → write() invocation. On slow platforms (QEMU), each extra
+// function call can cost 0.5-1ms, and cumulative overhead prevents coroutines
+// from completing work within the 10ms preemption window.
+#[must_use]
+pub extern "C" fn write(
+    fn_ptr: Option<&extern "C" fn(c_int, *const c_void, size_t) -> ssize_t>,
+    fd: c_int,
+    buf: *const c_void,
+    len: size_t,
+) -> ssize_t {
+    cfg_if::cfg_if! {
+        if #[cfg(all(target_os = "linux", feature = "io_uring"))] {
+            static CHAIN: once_cell::sync::Lazy<
+                WriteSyscallFacade<IoUringWriteSyscall<NioWriteSyscall<RawWriteSyscall>>>
+            > = once_cell::sync::Lazy::new(Default::default);
+        } else {
+            static CHAIN: once_cell::sync::Lazy<WriteSyscallFacade<NioWriteSyscall<RawWriteSyscall>>> =
+                once_cell::sync::Lazy::new(Default::default);
+        }
+    }
+    if fd == libc::STDOUT_FILENO || fd == libc::STDERR_FILENO {
+        return RawWriteSyscall::default().write(fn_ptr, fd, buf, len);
+    }
+    CHAIN.write(fn_ptr, fd, buf, len)
+}
 
 impl_facade!(WriteSyscallFacade, WriteSyscall,
     write(fd: c_int, buf: *const c_void, len: size_t) -> ssize_t

hook/src/syscall/unix.rs

@@ -62,6 +62,7 @@ impl_hook!(PREADV, preadv(fd: c_int, iov: *const iovec, iovcnt: c_int, offset: o
 impl_hook!(RECVMSG, recvmsg(fd: c_int, msg: *mut msghdr, flags: c_int) -> ssize_t);
 impl_hook!(SEND, send(fd: c_int, buf: *const c_void, len: size_t, flags: c_int) -> ssize_t);
 impl_hook!(SENDTO, sendto(fd: c_int, buf: *const c_void, len: size_t, flags: c_int, addr: *const sockaddr, addrlen: socklen_t) -> ssize_t);
+impl_hook!(WRITE, write(fd: c_int, buf: *const c_void, count: size_t) -> ssize_t);
 impl_hook!(PWRITE, pwrite(fd: c_int, buf: *const c_void, count: size_t, offset: off_t) -> ssize_t);
 impl_hook!(WRITEV, writev(fd: c_int, iov: *const iovec, iovcnt: c_int) -> ssize_t);
 impl_hook!(PWRITEV, pwritev(fd: c_int, iov: *const iovec, iovcnt: c_int, offset: off_t) -> ssize_t);
@@ -82,7 +83,6 @@ impl_hook!(RENAMEAT2, renameat2(olddirfd: c_int, oldpath: *const c_char, newdirf
 // NOTE: unhook poll due to mio's poller
 // impl_hook!(POLL, poll(fds: *mut pollfd, nfds: nfds_t, timeout: c_int) -> c_int);
 
-// NOTE: unhook write/pthread_mutex_lock/pthread_mutex_unlock due to stack overflow or bug
-// impl_hook!(WRITE, write(fd: c_int, buf: *const c_void, count: size_t) -> ssize_t);
+// NOTE: unhook pthread_mutex_lock/pthread_mutex_unlock due to stack overflow or bug
 // impl_hook!(PTHREAD_MUTEX_LOCK, pthread_mutex_lock(lock: *mut pthread_mutex_t) -> c_int);
 // impl_hook!(PTHREAD_MUTEX_UNLOCK, pthread_mutex_unlock(lock: *mut pthread_mutex_t) -> c_int);