diff --git a/arch/sysv_x86-64/thread-entry.asm b/arch/sysv_x86-64/thread-entry.asm
index df61895..6bea7e2 100644
--- a/arch/sysv_x86-64/thread-entry.asm
+++ b/arch/sysv_x86-64/thread-entry.asm
@@ -43,6 +43,11 @@ _cia_start_thread:
     pop rax ; thread_fn
     pop rdi ; ctx
     call rax
+    ; Make return value the first arg and call exit syscall
+    mov rdi, rax
+    mov rax, 60
+    syscall
+    ; Just to be sure...    
     hlt
 .exit:
     ret
diff --git a/cia.c b/cia.c
index ed5852e..edbc8d4 100644
--- a/cia.c
+++ b/cia.c
@@ -1,6 +1,10 @@
 
 #include <cia/def.h>
 
+// Base includes
+#include <cia/def.h>
+#include <cia/internal.h>
+
 // Platform-dependent sources
 #include _CIA_OS_CONF
 
diff --git a/include/cia/internal.h b/include/cia/internal.h
new file mode 100644
index 0000000..da71abc
--- /dev/null
+++ b/include/cia/internal.h
@@ -0,0 +1,23 @@
+
+#pragma once
+
+#include <cia/def.h>
+#include <stdint.h>
+
+struct Cia_CRT_Params typedef Cia_CRT_Params;
+struct Cia_CRT_Params {
+    u64 stack_size;
+    u64 tls_image_size;
+    void *tls_image_base;
+};
+
+struct Cia_TCB typedef Cia_TCB;
+struct Cia_TCB {
+    u64 thread_id;
+    u64 parent_id;
+    u64 pad0[3];
+    u64 stack_canary;
+    u64 pad1[2];
+};
+
+
diff --git a/include/errno.h b/include/errno.h
index 9d99853..dfba625 100644
--- a/include/errno.h
+++ b/include/errno.h
@@ -1,6 +1,8 @@
 
 #pragma once
 
+extern _Thread_local int errno;
+
 // errno-base.h
 #define EPERM     1
 #define ENOENT    2
diff --git a/include/tinyrt.h b/include/tinyrt.h
index 870a0d3..09d85eb 100644
--- a/include/tinyrt.h
+++ b/include/tinyrt.h
@@ -55,7 +55,7 @@ struct _RT_Thread {
     void *handle;
 };
 static _RT_Status _rt_thread_current(_RT_Thread *thread);
-static _RT_Status _rt_thread_create(_RT_Thread *thread, void (*thread_fn)(void *ctx), void *ctx);
+static _RT_Status _rt_thread_create(_RT_Thread *thread, int (*thread_fn)(void *ctx), void *ctx);
 static _RT_Status _rt_thread_join(_RT_Thread *thread);
 static _RT_Status _rt_thread_detach(_RT_Thread *thread);
 static _RT_Status _rt_thread_terminate(_RT_Thread *thread);
diff --git a/loader/loader.c b/loader/loader.c
index 70c3af4..aef90a4 100644
--- a/loader/loader.c
+++ b/loader/loader.c
@@ -15,6 +15,8 @@
 #include <tinyrt.h>
 #include "../os/linux/tinyrt.c"
 
+#include <cia/internal.h>
+
 #include <cia/mem.h>
 #include "../src/cia-mem/util.c"
 #include "../src/cia-mem/allocator.c"
@@ -23,14 +25,6 @@
 
 #include "stack.c"
 
-struct Thread_Control_Block typedef Thread_Control_Block;
-struct Thread_Control_Block {
-    u64 thread_id;
-    u64 pad0[4];
-    u64 stack_canary;
-    u64 pad1[2];
-};
-
 struct Elf_Image typedef Elf_Image;
 struct Elf_Image {
     Cia_Arena arena;
@@ -107,6 +101,7 @@ struct Stage3_Info_Struct {
     Elf_Image *ldso;
     void *stack_base;
     u64 stack_size;
+    void *tls_image;
     u64 tls_size;
 };
 
@@ -356,7 +351,7 @@ static void ld_stage3_entry(u64 has_new_stack, void *ctx) {
     Stage3_Info_Struct *info = ctx;
     _dbg_printf("Stack: %x-%x\n", info->stack_base, (u8 *)info->stack_base+info->stack_size);   
     // Set up the thread control block
-    Thread_Control_Block *tcb = cia_ptr_alignf((u8*)info->stack_base + info->tls_size, 1*MB);
+    Cia_TCB *tcb = cia_ptr_alignf((u8*)info->stack_base + info->tls_size, info->stack_size/2);
     tcb->thread_id = 0;
     tcb->stack_canary = 0x12345678fedcba98;
     // Copy TLS initialization image below TCB
@@ -377,6 +372,7 @@ static void ld_stage3_entry(u64 has_new_stack, void *ctx) {
             for(int i = 0; i < tls_image_size; ++i) {
                 tls_image[i] = tls_image_base[i];
             }
+            info->tls_image = tls_image_base;
         }
     }
     // Set up the thread pointer
@@ -386,6 +382,10 @@ static void ld_stage3_entry(u64 has_new_stack, void *ctx) {
         sys_exit(1);
     }
     _dbg_printf("Entered loader stage 3. Entering main executable\n");
-    void (*crt_entry)() = elf_addr(info->app, ((Elf64_Ehdr *)info->app->base)->e_entry);
-    crt_entry();
+    Cia_CRT_Params params;
+    params.stack_size = info->stack_size/2;
+    params.tls_image_size = info->tls_size;
+    params.tls_image_base = info->tls_image;
+    void (*crt_entry)(Cia_CRT_Params *params) = elf_addr(info->app, ((Elf64_Ehdr *)info->app->base)->e_entry);
+    crt_entry(&params);
 }
diff --git a/os/linux/conf.h b/os/linux/conf.h
index 9091388..3419014 100644
--- a/os/linux/conf.h
+++ b/os/linux/conf.h
@@ -1,6 +1,10 @@
 
 #pragma once
 
+static u64 cia_stack_size;
+static u64 cia_tls_image_size;
+static void *cia_tls_image_base;
+
 #include <linux/futex.h>
 #include <sys/mman.h>
 #include <errno.h>
@@ -9,4 +13,5 @@
 #include "entry.c"
 #include <tinyrt.h>
 #include "tinyrt.c"
+#include "tinyrt-threads.c"
 
diff --git a/os/linux/entry.c b/os/linux/entry.c
index fe63f79..d79f847 100644
--- a/os/linux/entry.c
+++ b/os/linux/entry.c
@@ -14,7 +14,10 @@ void __stack_chk_fail(void) {
 extern int main(int argc, char **argv, char **envp);
 static void _fileapi_init();
 
-void _start() {
+void _start(Cia_CRT_Params *params) {
+    cia_stack_size = params->stack_size;
+    cia_tls_image_size = params->tls_image_size;
+    cia_tls_image_base = params->tls_image_base;
     // Get the envp
     // char **envp = argv + (argc + 1);
     // init(argc, argv, envp);
diff --git a/os/linux/tinyrt-threads.c b/os/linux/tinyrt-threads.c
new file mode 100644
index 0000000..fc03e51
--- /dev/null
+++ b/os/linux/tinyrt-threads.c
@@ -0,0 +1,71 @@
+
+extern i64 _cia_start_thread(
+    u64 flags,
+    void *stack_base,
+    int *parent_tid,
+    int *child_tid,
+    void *tls,
+    int (*thread_fn)(void *ctx),
+    void *ctx
+);
+
+static _RT_Status _rt_thread_current(_RT_Thread *thread) {
+    return _RT_ERROR_NOT_IMPLEMENTED;
+}
+
+static _RT_Status _rt_thread_create(_RT_Thread *thread, int (*thread_fn)(void *ctx), void *ctx) {
+    // Create the memory for stack
+    u64 mmap_prot = PROT_READ|PROT_WRITE;
+    u64 mmap_flags = MAP_PRIVATE|MAP_ANONYMOUS|MAP_NORESERVE;
+    void *stack_base = sys_mmap(0, 2*cia_stack_size, mmap_prot, mmap_flags, -1, 0);
+    if((i64)stack_base < 0) {
+        return _RT_ERROR_GENERIC;
+    }
+    void *stack = (u8*)stack_base + 2*cia_stack_size;
+    // Find the TLS base and initialize the tls
+    Cia_TCB *tcb = (void *)((u64)((u8 *)stack - 1) & ~(cia_stack_size - 1));
+    tcb->stack_canary = 0x12345678deadbeef;
+    u8 *tls_base = (u8 *)tcb - cia_tls_image_size;
+    for(int i = 0; i < cia_tls_image_size; ++i) {
+        tls_base[i] = ((u8 *)cia_tls_image_base)[i];
+    }
+    // Create the new thread
+    u64 flags = 0;
+    // flags |= CLONE_CHILD_CLEARTID;
+    // flags |= CLONE_PARENT_SETTID;
+    flags |= CLONE_FS;
+    flags |= CLONE_FILES;
+    flags |= CLONE_SIGHAND;
+    flags |= CLONE_THREAD;
+    flags |= CLONE_VM;
+    flags |= CLONE_SYSVSEM;
+    int *child_tid = (int *)&tcb->thread_id;
+    int *parent_tid = (int *)&tcb->parent_id;
+    *child_tid = 1;
+    *parent_tid = 0;
+    i64 ret = _cia_start_thread(flags, stack, parent_tid, child_tid, 0, thread_fn, ctx);
+    if(ret < 0) {
+        return _RT_ERROR_GENERIC;
+    }
+    return _RT_STATUS_OK;
+}
+
+static _RT_Status _rt_thread_join(_RT_Thread *thread) {
+    return _RT_ERROR_NOT_IMPLEMENTED;
+}
+
+static _RT_Status _rt_thread_detach(_RT_Thread *thread) {
+    return _RT_ERROR_NOT_IMPLEMENTED;
+}
+
+static _RT_Status _rt_thread_terminate(_RT_Thread *thread) {
+    return _RT_ERROR_NOT_IMPLEMENTED;
+}
+
+static _RT_Status _rt_thread_sleep(u64 time) {
+    return _RT_ERROR_NOT_IMPLEMENTED;
+}
+
+static _RT_Status _rt_thread_get_timer_freq(u64 *freq) {
+    return _RT_ERROR_NOT_IMPLEMENTED;
+}
diff --git a/os/linux/tinyrt.c b/os/linux/tinyrt.c
index b79ea88..92e5027 100644
--- a/os/linux/tinyrt.c
+++ b/os/linux/tinyrt.c
@@ -1,76 +1,10 @@
 
 // See src/tinyrt.h file for the interface this file implements
 
-extern i64 _cia_start_thread(
-    u64 flags,
-    void *stack_base,
-    int *parent_tid,
-    int *child_tid,
-    void *tls,
-    void (*thread_fn)(void *ctx),
-    void *ctx
-);
-
 _Noreturn static void _rt_program_exit(int code) {
     sys_exit(code);
 }
 
-static _RT_Status _rt_thread_current(_RT_Thread *thread) {
-    return _RT_ERROR_NOT_IMPLEMENTED;
-}
-
-static _RT_Status _rt_thread_create(_RT_Thread *thread, void (*thread_fn)(void *ctx), void *ctx) {
-    // Create the memory for stack
-    u64 mmap_prot = PROT_READ|PROT_WRITE;
-    u64 mmap_flags = MAP_PRIVATE|MAP_ANONYMOUS|MAP_NORESERVE;
-    u64 stack_size = 0x10000;
-    void *stack_base = sys_mmap(0, stack_size, mmap_prot, mmap_flags, -1, 0);
-    if((i64)stack_base < 0) {
-        return _RT_ERROR_GENERIC;
-    }
-    void *stack = (u8*)stack_base + stack_size;
-    // Create the new thread
-    u64 flags = 0;
-    // flags |= CLONE_CHILD_CLEARTID;
-    // flags |= CLONE_PARENT_SETTID;
-    flags |= CLONE_FS;
-    flags |= CLONE_FILES;
-    flags |= CLONE_SIGHAND;
-    flags |= CLONE_THREAD;
-    flags |= CLONE_VM;
-    flags |= CLONE_SYSVSEM;
-    int *temp_permanent_storage = stack_base;
-    int *child_tid = &temp_permanent_storage[0];
-    int *parent_tid = &temp_permanent_storage[1];
-    *child_tid = 1;
-    *parent_tid = 0;
-    i64 ret = _cia_start_thread(flags, stack, parent_tid, child_tid, 0, thread_fn, ctx);
-    if(ret < 0) {
-        return _RT_ERROR_GENERIC;
-    }
-    return _RT_STATUS_OK;
-}
-
-static _RT_Status _rt_thread_join(_RT_Thread *thread) {
-    return _RT_ERROR_NOT_IMPLEMENTED;
-}
-
-static _RT_Status _rt_thread_detach(_RT_Thread *thread) {
-    return _RT_ERROR_NOT_IMPLEMENTED;
-}
-
-static _RT_Status _rt_thread_terminate(_RT_Thread *thread) {
-    return _RT_ERROR_NOT_IMPLEMENTED;
-}
-
-static _RT_Status _rt_thread_sleep(u64 time) {
-    return _RT_ERROR_NOT_IMPLEMENTED;
-}
-
-static _RT_Status _rt_thread_get_timer_freq(u64 *freq) {
-    return _RT_ERROR_NOT_IMPLEMENTED;
-}
-
 static _RT_Status _rt_file_std_handles_init() {
     _rt_file_stdin.fd = 0;
     _rt_file_stdin.flags = _RT_FILE_READ;
diff --git a/src/stdlib-thread/thread.c b/src/stdlib-thread/thread.c
index 6065a91..152ab02 100644
--- a/src/stdlib-thread/thread.c
+++ b/src/stdlib-thread/thread.c
@@ -1,6 +1,6 @@
 
 int thrd_create(thrd_t *thr, thrd_start_t func, void *arg) {
-    _RT_Status status = _rt_thread_create(&thr->thread, (void (*)(void *))func, arg);
+    _RT_Status status = _rt_thread_create(&thr->thread, (int (*)(void *))func, arg);
     if(status == _RT_STATUS_OK) {
         return thrd_success;
     }
diff --git a/tests/threaded.c b/tests/threaded.c
index 12dab2b..3ef6952 100644
--- a/tests/threaded.c
+++ b/tests/threaded.c
@@ -57,7 +57,6 @@ int thrd_func(void *arg) {
     print_int(counter);
     print_char('\n');
     cia_mutex_unlock(&g_print_mutex);
-    exit(1);
     return 0;
 }