5 files changed, 130 insertions, 1 deletions
diff --git a/lib/libc/internal/arch/aarch64/thread_self.c b/lib/libc/internal/arch/aarch64/thread_self.c
new file mode 100644
index 00000000..543b07ab
--- /dev/null
+++ b/lib/libc/internal/arch/aarch64/thread_self.c
@@ -0,0 +1,8 @@
+#include <__thread.h>
+
+inline struct __thread_self *__thread_self(void)
+{
+	struct __thread_self *self;
+	__asm__("mrs %0, tpidr_el0" : "=r"(self));
+	return self;
+}
diff --git a/lib/libc/internal/arch/x86_64/Kbuild b/lib/libc/internal/arch/x86_64/Kbuild
index 830a1c5b..49e4284f 100644
--- a/lib/libc/internal/arch/x86_64/Kbuild
+++ b/lib/libc/internal/arch/x86_64/Kbuild
@@ -1,5 +1,7 @@
 obj-y += _start.o
+obj-y += clone.o
 obj-y += fenv.o
 obj-y += longjmp.o
 obj-y += setjmp.o
 obj-y += sigsetjmp.o
+obj-y += tcb.o
diff --git a/lib/libc/internal/arch/x86_64/_start.c b/lib/libc/internal/arch/x86_64/_start.c
index f3f4fef2..e8846bfa 100644
--- a/lib/libc/internal/arch/x86_64/_start.c
+++ b/lib/libc/internal/arch/x86_64/_start.c
@@ -4,5 +4,5 @@
 __dead __naked void _start(void)
 {
 	__asm__ __volatile__("mov %rsp, %rdi\n"
-			     "call __init\n");
+			     "call __libc_init\n");
 }
diff --git a/lib/libc/internal/arch/x86_64/clone.s b/lib/libc/internal/arch/x86_64/clone.s
new file mode 100644
index 00000000..80f512b1
--- /dev/null
+++ b/lib/libc/internal/arch/x86_64/clone.s
@@ -0,0 +1,106 @@
+/*
+ * x86_64 clone syscall trampoline
+ *
+ * Goal: provide a reliable way to run a C function in the child after the
+ * clone(2) syscall, without depending on the parent's stack frame (which is
+ * invalid in the child because the child starts with a new stack pointer).
+ *
+ * API (internal):
+ *
+ *   long __clone(long (*fn)(void *), void *arg,
+ *               unsigned long flags, void *child_stack,
+ *               int *ptid, void *newtls, int *ctid);
+ *
+ * This uses the raw Linux x86_64 clone syscall:
+ *   long clone(unsigned long flags, void *child_stack,
+ *              int *ptid, int *ctid, unsigned long newtls);
+ *
+ * Behavior:
+ * - In the parent, returns child TID (> 0) or -errno (< 0).
+ * - In the child, calls fn(arg) and then exits the thread/process via
+ *   SYS_exit (does not return).
+ *
+ * Notes:
+ * - We explicitly place (fn,arg) onto the child's new stack before issuing
+ *   the syscall. On success in the child, we recover them from %rsp and call.
+ * - This avoids relying on any locals from the parent frame.
+ * - The stack is aligned to 16 bytes before the call into C.
+ */
+
+	.text
+	.globl __clone
+	.type  __clone, @function
+
+__clone:
+	/*
+	 * SysV AMD64 calling convention:
+	 *   fn          in %rdi
+	 *   arg         in %rsi
+	 *   flags       in %rdx
+	 *   child_stack in %rcx
+	 *   ptid        in %r8
+	 *   newtls      in %r9
+	 *   ctid        at 8(%rsp)
+	 */
+	movq 8(%rsp), %r10          /* r10 = ctid (7th arg from caller stack) */
+
+	/* r11 = child_stack (we'll use it as our working "new sp") */
+	movq %rcx, %r11
+
+	/*
+	 * Ensure 16-byte alignment for the eventual C call:
+	 * We'll build the child's stack as:
+	 *   [low]  fn (8)
+	 *          arg (8)
+	 *   [high]  <-- %rsp in child before call site
+	 *
+	 * We align first, then reserve 16 bytes.
+	 */
+	andq $-16, %r11
+	subq $16, %r11
+
+	/* Store fn and arg onto the child's stack */
+	movq %rdi, 0(%r11)          /* fn  */
+	movq %rsi, 8(%r11)          /* arg */
+
+	/*
+	 * Prepare registers for Linux x86_64 clone syscall:
+	 *   rax = SYS_clone
+	 *   rdi = flags
+	 *   rsi = child_stack
+	 *   rdx = ptid
+	 *   r10 = ctid
+	 *   r8  = newtls
+	 */
+	movq %rdx, %rdi             /* flags */
+	movq %r11, %rsi             /* child_stack (new sp) */
+	movq %r8,  %rdx             /* ptid */
+	movq %r9,  %r8              /* newtls */
+	/* r10 already holds ctid */
+
+	movq $56, %rax              /* SYS_clone on x86_64 */
+	syscall
+
+	/* Parent: rax = child tid (>0) or -errno (<0) */
+	testq %rax, %rax
+	jnz 1f
+
+	/* Child: rax == 0 */
+
+	/* %rsp is already the child_stack we provided. Recover fn/arg from it. */
+	movq 8(%rsp), %rdi          /* first arg to fn = arg */
+	call *0(%rsp)               /* call fn(arg) */
+
+	/*
+	 * If fn returns, exit the thread/process.
+	 * Use SYS_exit (60). Return value of fn is in %rax; pass as status in %rdi.
+	 */
+	movq %rax, %rdi
+	movq $60, %rax              /* SYS_exit */
+	syscall
+	hlt                         /* should not reach */
+
+1:
+	ret
+
+	.size __clone, .-__clone
diff --git a/lib/libc/internal/arch/x86_64/tcb.c b/lib/libc/internal/arch/x86_64/tcb.c
new file mode 100644
index 00000000..6e1f0c15
--- /dev/null
+++ b/lib/libc/internal/arch/x86_64/tcb.c
@@ -0,0 +1,13 @@
+#include <libc/tcb.h>
+
+__attribute__((__always_inline__)) void __libc_tcb_set(uint64_t tcb)
+{
+	__asm__ volatile("wrfsbase %0" ::"r"(tcb));
+}
+
+__attribute__((__always_inline__)) void *__libc_tcb_get(void)
+{
+	void *tcb;
+	__asm__ volatile("rdfsbase %0" : "=r"(tcb));
+	return tcb;
+}