8 files changed, 101 insertions, 24 deletions

diff --git a/lib/libc/internal/include/__stdio.h b/lib/libc/internal/include/__stdio.h
index 487a0382..7846b139 100644
--- a/lib/libc/internal/include/__stdio.h
+++ b/lib/libc/internal/include/__stdio.h
@@ -31,8 +31,6 @@ struct __FILE {
 	struct __FILE *next;
 };
 
-#define __FILE(__stream) ((struct __FILE *)(__stream))
-
 void __libc_fadd(struct __FILE *f);
 
 #endif
diff --git a/lib/libc/internal/include/__thread.h b/lib/libc/internal/include/__thread.h
index 7639707a..3c68f8dd 100644
--- a/lib/libc/internal/include/__thread.h
+++ b/lib/libc/internal/include/__thread.h
@@ -1,8 +1,50 @@
 #ifndef __LIBC_THREAD_H
 #define __LIBC_THREAD_H
 
+#include <stdatomic.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#define THREAD_STACK_SIZE (1024 * 1024)
+#define THREAD_GUARD_SIZE 8192
+
+enum __thread_state { THREAD_STATE_EXITED, THREAD_STATE_JOINABLE, THREAD_STATE_DETACHED };
+
+struct tls {
+	void *data;
+	size_t length;
+	size_t size;
+	size_t align;
+};
+
 struct __thread_self {
-	int tid;
+	struct __thread_self *self;
+
+	/* Backing mapping for this thread (used by thrd_join/thrd_detach to munmap safely) */
+	void *map_base;
+	size_t map_size;
+
+#ifdef __aarch64__
+	uintptr_t *dtv;
+	uintptr_t canary;
+#endif
+
+	long tid;
+	int res;
+	int errno_v;
+	volatile int state;
+
+#ifdef __x86_64__
+	uintptr_t canary;
+	uintptr_t *dtv;
+#endif
 };
 
+void __libc_tls_copy(void *);
+
+#if defined(__x86_64__)
+long __clone(long (*fn)(void *), void *arg, unsigned long flags, void *child_stack, long *ptid, void *newtls,
+	     long *ctid);
+#endif
+
 #endif
diff --git a/lib/libc/internal/include/atomic.h b/lib/libc/internal/include/atomic.h
index e0593021..d40c1eae 100644
--- a/lib/libc/internal/include/atomic.h
+++ b/lib/libc/internal/include/atomic.h
@@ -1,22 +1,32 @@
 #ifndef __LIBC_ATOMIC_H
 #define __LIBC_ATOMIC_H
 
+#include <libc/futex.h>
+#include <sched.h>
 #include <stdatomic.h>
+#include <unistd.h>
 
 #define LIBC_LOCK(__lock)   __libc_lock(&((__lock)))
-#define LIBC_UNLOCK(__lock) atomic_flag_clear(&((__lock)))
+#define LIBC_UNLOCK(__lock) __libc_unlock(&((__lock)))
 
-static __inline void __libc_lock(volatile atomic_flag *lock)
+__attribute__((__always_inline__)) __inline void __libc_unlock(volatile atomic_flag *lock)
 {
-	while (atomic_flag_test_and_set_explicit(lock, memory_order_acquire)) {
-		unsigned int spins = 1;
-		do {
-			for (unsigned int i = 0; i < spins; i++)
-				__asm__ volatile("pause");
-			if (spins < 64)
-				spins *= 2;
-		} while (atomic_flag_test_and_set_explicit(
-			lock, memory_order_relaxed));
+	atomic_flag_clear_explicit(lock, memory_order_release);
+	__futex_wake((volatile int *)lock, 1);
+}
+
+__attribute__((__always_inline__)) __inline void __libc_lock(volatile atomic_flag *lock)
+{
+	if (atomic_flag_test_and_set_explicit(lock, memory_order_acquire) == 0)
+		return;
+
+	while (1) {
+		__futex_wait((volatile int *)lock, 1);
+
+		if (atomic_flag_test_and_set_explicit(lock, memory_order_acquire) == 0)
+			return;
+
+		sched_yield();
 	}
 }
 
diff --git a/lib/libc/internal/include/libc.h b/lib/libc/internal/include/libc.h
index f06fa6d8..75a06dd2 100644
--- a/lib/libc/internal/include/libc.h
+++ b/lib/libc/internal/include/libc.h
@@ -1,26 +1,20 @@
 #ifndef __LIBC_LIBC_H
 #define __LIBC_LIBC_H
 
+#include <__stdio.h>
 #include <stdatomic.h>
 #include <stddef.h>
 #include <sys/cdefs.h>
 
 #define weak_reference(old, new) extern __typeof(old)((new)) __attribute__((__weak__, __alias__(#old)))
 
-struct tls {
-	struct tls *next;
-	void *data;
-	size_t size;
-	size_t align;
-	size_t length;
-	size_t offset;
-};
-
 struct libc {
-	size_t auxv[32];
+	size_t *auxv;
+
 	struct {
 		void *base;
 		size_t size;
+		size_t align;
 	} tls;
 
 	enum {
@@ -32,6 +26,10 @@ struct libc {
 		volatile atomic_flag malloc;
 		volatile atomic_flag environ;
 	} lock;
+
+	struct __FILE stdin;
+	struct __FILE stdout;
+	struct __FILE stderr;
 };
 
 extern struct libc __libc;
diff --git a/lib/libc/internal/include/libc/fsbase.h b/lib/libc/internal/include/libc/fsbase.h
new file mode 100644
index 00000000..61f4abce
--- /dev/null
+++ b/lib/libc/internal/include/libc/fsbase.h
@@ -0,0 +1,7 @@
+#ifndef __LIBC_FSBASE_H
+#define __LIBC_FSBASE_H
+
+void *rdfsbase(void);
+void wrfsbase(unsigned long);
+
+#endif
diff --git a/lib/libc/internal/include/libc/futex.h b/lib/libc/internal/include/libc/futex.h
new file mode 100644
index 00000000..a00cf0e7
--- /dev/null
+++ b/lib/libc/internal/include/libc/futex.h
@@ -0,0 +1,7 @@
+#ifndef __LIBC_FUTEX_H
+#define __LIBC_FUTEX_H
+
+int __futex_wait(volatile int *, int);
+int __futex_wake(volatile int *, int);
+
+#endif
diff --git a/lib/libc/internal/include/libc/tcb.h b/lib/libc/internal/include/libc/tcb.h
new file mode 100644
index 00000000..1defd41c
--- /dev/null
+++ b/lib/libc/internal/include/libc/tcb.h
@@ -0,0 +1,9 @@
+#ifndef __LIBC_TCB_H
+#define __LIBC_TCB_H
+
+#include <stdint.h>
+
+void *__libc_tcb_get(void);
+void __libc_tcb_set(uint64_t);
+
+#endif
diff --git a/lib/libc/internal/include/libc/thread.h b/lib/libc/internal/include/libc/thread.h
new file mode 100644
index 00000000..6422a0e2
--- /dev/null
+++ b/lib/libc/internal/include/libc/thread.h
@@ -0,0 +1,6 @@
+#ifndef __LIBC_THREAD_H
+#define __LIBC_THREAD_H
+
+void *__libc_thread_self(void);
+
+#endif