diff --git a/include/sys/sched.h b/include/sys/sched.h
new file mode 100644
index 0000000..72394ee
--- /dev/null
+++ b/include/sys/sched.h
@@ -0,0 +1,5 @@
+#pragma once
+#include "sys/thread.h"
+
+void sched_add(struct thread *t);
+void sched_init(void);
diff --git a/include/sys/thread.h b/include/sys/thread.h
index 68379b8..ebc3e5b 100644
--- a/include/sys/thread.h
+++ b/include/sys/thread.h
@@ -3,6 +3,7 @@
 #include "sys/fs/vfs.h"
 
 #define THREAD_KERNEL       (1 << 31)
+#define THREAD_STOPPED      (1 << 2)
 
 #if defined(ARCH_AMD64)
 typedef uint64_t *mm_space_t;
@@ -22,6 +23,8 @@ struct thread {
     mm_space_t space;
 
     // TODO: maybe __sched_thread
+    int cpu;
+    struct thread *prev;
     struct thread *next;
 };
 
diff --git a/sys/amd64/kernel.c b/sys/amd64/kernel.c
index bc497db..df2f783 100644
--- a/sys/amd64/kernel.c
+++ b/sys/amd64/kernel.c
@@ -17,6 +17,7 @@
 #include "sys/fs/vfs.h"
 #include "sys/blk/ram.h"
 #include "sys/fs/tar.h"
+#include "sys/sched.h"
 #include "sys/tty.h"
 
 static multiboot_info_t *multiboot_info;
@@ -37,7 +38,6 @@ void kernel_main(struct amd64_loader_data *data) {
     amd64_mm_init(data);
     amd64_acpi_init();
 
-    extern void sched_init(void);
     sched_init();
     amd64_apic_init();
     rtc_init();
diff --git a/sys/amd64/sys/sched.c b/sys/amd64/sys/sched.c
index b541dd1..76e32ab 100644
--- a/sys/amd64/sys/sched.c
+++ b/sys/amd64/sys/sched.c
@@ -126,6 +126,12 @@ void init_func(void *arg) {
     }
 }
 
+static uint32_t sched_alloc_pid(void) {
+    // The first pid allocation will be 1 (init process)
+    static uint32_t last_pid = 1;
+    return last_pid++;
+}
+
 void sched_add_to(int cpu, struct thread *t) {
     t->next = NULL;
 
@@ -135,6 +141,14 @@ void sched_add_to(int cpu, struct thread *t) {
     } else {
         sched_queue_heads[cpu] = t;
     }
+    t->prev = sched_queue_tails[cpu];
+    t->cpu = cpu;
+
+    if (!(t->flags & THREAD_KERNEL)) {
+        t->pid = sched_alloc_pid();
+    } else {
+        t->pid = 0;
+    }
     sched_queue_tails[cpu] = t;
     ++sched_queue_sizes[cpu];
     spin_release(&sched_lock);
@@ -187,12 +201,52 @@ void sched_init(void) {
 int sched(void) {
     struct thread *from = get_cpu()->thread;
     struct thread *to;
+    int cpu = get_cpu()->processor_id;
 
     spin_lock(&sched_lock);
+    // Cleanup stopped tasks
+    // TODO: this won't be needed if I call sched_remove directly
+    to = sched_queue_heads[cpu];
+    while (to) {
+        if (to->flags & THREAD_STOPPED) {
+            // So we don't remove [idle]
+            _assert(to != sched_queue_heads[cpu]);
+            kdebug("(cpu%d) Removing %u from queue\n", cpu, to->pid);
+
+            if (to->pid == 1) {
+                panic("init process was killed\n");
+            }
+
+            struct thread *prev, *next;
+            prev = to->prev;
+            next = to->next;
+
+            _assert(prev);
+
+            prev->next = next;
+            if (next) {
+                next->prev = prev;
+            } else {
+                sched_queue_tails[cpu] = prev;
+            }
+
+            // TODO: free the task?
+            if (to == from) {
+                // Removing current thread
+                from = NULL;
+            }
+
+            if (!next) {
+                break;
+            }
+        }
+        to = to->next;
+    }
+
     if (from && from->next) {
         to = from->next;
     } else {
-        to = sched_queue_heads[get_cpu()->processor_id];
+        to = sched_queue_heads[cpu];
     }
     spin_release(&sched_lock);
 
diff --git a/sys/amd64/syscall.c b/sys/amd64/syscall.c
index 6ada799..4f1b338 100644
--- a/sys/amd64/syscall.c
+++ b/sys/amd64/syscall.c
@@ -7,6 +7,7 @@
 
 static ssize_t sys_read(int fd, void *buf, size_t lim);
 static ssize_t sys_write(int fd, const void *buf, size_t lim);
+static void sys_exit(int status);
 
 uint8_t irq1_key = 0;
 
@@ -16,7 +17,16 @@ intptr_t amd64_syscall(uintptr_t rdi, uintptr_t rsi, uintptr_t rdx, uintptr_t rc
         return sys_read((int) rdi, (void *) rsi, (size_t) rdx);
     case 1:
         return sys_write((int) rdi, (const void *) rsi, (size_t) rdx);
+    case 60:
+        sys_exit((int) rdi);
+        // TODO: call sched_remove and reschedule a new task so that
+        // this one does not return
+        asm volatile ("sti; hlt");
+        panic("This should not happen\n");
+        // WTF?
+        return -1;
     default:
+        kdebug("unknown syscall: %u\n", rax);
         return -1;
     }
     return 0;
@@ -38,6 +48,13 @@ static ssize_t sys_read(int fd, void *buf, size_t lim) {
     return vfs_read(&thr->ioctx, of, buf, lim);
 }
 
+static void sys_exit(int status) {
+    struct thread *thr = get_cpu()->thread;
+    _assert(thr);
+    kdebug("%u exited with code %d\n", thr->pid, status);
+    thr->flags |= THREAD_STOPPED;
+}
+
 static ssize_t sys_write(int fd, const void *buf, size_t lim) {
     if (fd >= 4 || fd < 0) {
         return -EBADF;
diff --git a/usr/Makefile b/usr/Makefile
index 464b76f..22da67b 100644
--- a/usr/Makefile
+++ b/usr/Makefile
@@ -12,6 +12,7 @@ usr_CFLAGS=-ffreestanding \
 		   -mno-sse \
 		   -mno-sse2 \
 		   -Ilibc/include \
+		   -O0 \
 		   -I$(S)/include
 usr_LDFLAGS=-nostdlib \
 			-lgcc \
diff --git a/usr/init.c b/usr/init.c
index 250cbf5..47759c5 100644
--- a/usr/init.c
+++ b/usr/init.c
@@ -8,10 +8,17 @@ int main(int argc, char **argv) {
         ssize_t sz = read(STDIN_FILENO, &c, 1);
 
         if (sz >= 0) {
+            if (c == '\n') {
+                break;
+            }
+
             write(STDOUT_FILENO, "Char: ", 6);
             write(STDOUT_FILENO, &c, 1);
             c = '\n';
             write(STDOUT_FILENO, &c, 1);
         }
     }
+
+    write(STDOUT_FILENO, "Enter\n", 6);
+    return 0;
 }
diff --git a/usr/libc/crt0.S b/usr/libc/crt0.S
index 7c0d376..b998d51 100644
--- a/usr/libc/crt0.S
+++ b/usr/libc/crt0.S
@@ -16,9 +16,8 @@ _start:
     // Ignore %rdi/%rsi for now
     xorq %rdi, %rdi
 
+    // exit(main(argc, argv))
     call main
-
-    // I haven't implemented exit() yet
-    xorq %rdi, %rdi
-    jmp *%rdi
+    movq %rax, %rdi
+    call exit
 .size _start, . - _start
diff --git a/usr/libc/include/bits/syscall.h b/usr/libc/include/bits/syscall.h
index 1284043..71341d8 100644
--- a/usr/libc/include/bits/syscall.h
+++ b/usr/libc/include/bits/syscall.h
@@ -3,3 +3,4 @@
 
 ssize_t read(int fd, void *buf, size_t count);
 ssize_t write(int fd, const void *buf, size_t count);
+__attribute__((noreturn)) void exit(int code);
diff --git a/usr/libc/syscall.c b/usr/libc/syscall.c
index baa0b61..e05c21b 100644
--- a/usr/libc/syscall.c
+++ b/usr/libc/syscall.c
@@ -3,12 +3,27 @@
 #define ASM_REGISTER(name) \
         register uint64_t name asm (#name)
 
-#define ASM_SYSCALL3(r0, r1, r2, r3)        ({ \
+#define ASM_SYSCALL1(r0, r1)                ({ \
+        ASM_REGISTER(rdi) = (uint64_t) (r1); \
+        /*
+         * Should be the last one because memory accesses for arguments
+         * fuck up %rax
+         */ \
         ASM_REGISTER(rax) = (uint64_t) (r0); \
+        asm volatile ("syscall":::"memory"); \
+        rax; \
+    })
+
+#define ASM_SYSCALL3(r0, r1, r2, r3)        ({ \
         ASM_REGISTER(rdi) = (uint64_t) (r1); \
         ASM_REGISTER(rsi) = (uint64_t) (r2); \
         ASM_REGISTER(rdx) = (uint64_t) (r3); \
-        asm volatile ("syscall":::"memory"); \
+        /*
+         * Should be the last one because memory accesses for arguments
+         * fuck up %rax
+         */ \
+        ASM_REGISTER(rax) = (uint64_t) (r0); \
+        asm volatile ("syscall":::"memory", "rax", "rdi", "rsi", "rdx"); \
         rax; \
     })
 
@@ -19,3 +34,8 @@ ssize_t write(int fd, const void *buf, size_t count) {
 ssize_t read(int fd, void *buf, size_t count) {
     return (ssize_t) ASM_SYSCALL3(0, fd, buf, count);
 }
+
+__attribute__((noreturn)) void exit(int code) {
+    (void) ASM_SYSCALL1(60, code);
+    while (1);
+}