Running a Multithreaded Executable on Linux

An ELF binary is divided into segments (runtime view via program headers) and sections (link-time view via section headers). The kernel only cares about segments when loading.

execve replaces the process image. The kernel maps ELF segments, sets up the initial stack with argc/argv/envp/auxv, then jumps to the dynamic linker's entry point.

Each VMA (vm_area_struct) tracks permissions, file backing, and CoW state. ASLR randomises base addresses. /proc/<pid>/maps lists all VMAs live.

POSIX threads share address space, file descriptors, and signal disposition but have private stacks, TLS, and scheduling attributes. All are Linux tasks under the hood.

On x86-64 the syscall instruction saves registers, switches stacks to the per-CPU kernel stack, and jumps to entry_SYSCALL_64.

Linux uses a single task_struct for both processes and threads. Threads in a process share one mm_struct but each has its own kernel stack.

CFS tracks vruntime per-task and always picks the leftmost node of the red-black tree. Threads compete for slices within a cgroup hierarchy.

The kernel maintains a 4- (or 5-) level page table per mm_struct. A page fault is the kernel's opportunity to demand-page, CoW-break, or extend the stack.

pthread_mutex_lock calls futex(FUTEX_WAIT) only when contended — uncontended lock/unlock stays entirely in user space.

The scheduler timer interrupt fires via the Local APIC (~250 Hz default). The MMU's CR3 register holds the physical address of the top-level page table — reloaded on every context switch.