feat(thread): futex + TLS + pthread_self substrate

src/start/tls.c

@@ -76,6 +76,17 @@ struct elf64_phdr
     uint64_t p_align;
 };
 
+/*
+ * New-thread TLS geometry, consumed by src/thread/tls.c (todo 44) — see
+ * tcb.h for the full contract. __vlibc_static_tls_size is the EXPORTED
+ * default-visibility object (no attribute!); the other three are internal
+ * and hidden.
+ */
+size_t __vlibc_static_tls_size = 0;
+hidden uintptr_t __vlibc_static_tls_template = 0;
+hidden size_t __vlibc_static_tls_filesz = 0;
+hidden size_t __vlibc_static_tls_align = 16;
+
 /*
  * Round v up to a multiple of a. a must be a power of two, which ELF
  * p_align of a PT_TLS segment always is.
@@ -253,6 +264,13 @@ init_main_tls(char **envp)
         align = 16;
     }
 
+    /* Record the geometry for new-thread TLS blocks (#44/#45) BEFORE the
+     * template address is known: size/filesz/align are PT_TLS facts, the
+     * template address depends on the mmap below. */
+    __vlibc_static_tls_size = memsz;
+    __vlibc_static_tls_filesz = filesz;
+    __vlibc_static_tls_align = align;
+
     /*
      * Single allocation: TLS image at the base, TCB at tp. tp uses exactly
      * p_align (never bumped): the linker bakes tpoff against
@@ -273,6 +291,12 @@ init_main_tls(char **envp)
     tp = base + block;
     image = (char *)base;
 
+    /* The PRIMAL image is the new-thread template: the still-pristine
+     * .tdata at p_vaddr when mapped, else this block (the file-fallback
+     * path leaves no other copy). Never the main thread's live TLS — its
+     * mutations must not leak into threads created later. */
+    __vlibc_static_tls_template = mapped ? vaddr : base;
+
     if (filesz > 0)
     {
         copy_tls_image(image, offset, vaddr, filesz, mapped);