Kernel Development Learning Pipeline

Overview

What is memory management?

1. Phsyical and virtual memory

2. Allocation

3. Paging

Just scratching the surface!

Virtual memory

We start without any: this is NOMMU mode

Address space isolation

Clone system call

Demo: man 2 clone

Aside: See posix_spawn(3)

Question: How is fork defined?

A: it's just clone

kernel/fork.c

2887 #ifdef __ARCH_WANT_SYS_FORK
2888 SYSCALL_DEFINE0(fork)
2889 {
2890 #ifdef CONFIG_MMU
2891         struct kernel_clone_args args = {
2892                 .exit_signal = SIGCHLD,
2893         };
2894
2895         return kernel_clone(&args);
2896 #else
2897         /* can not support in nommu mode */
2898         return -EINVAL;
2899 #endif
2900 }
2901 #endif

Namespace interlude

clone(CLONE_VM)

1. CLONE_VM: The new process shares the same memory space as the parent.
2. CLONE_FS: Shares file system information.
3. CLONE_FILES: Shares file descriptors.
4. CLONE_SIGHAND: Shares signal handlers.

Discuss namespaces: isolated view of system resources

• Mount Namespace (mnt) – Isolates filesystem mount points.
• Process Namespace (pid) – Provides separate process ID trees.
• Network Namespace (net) – Gives processes their own networking stack.
• User Namespace (user) – Provides separate user and group IDs.
• UTS Namespace (uts) – Isolates hostname and domain name.
• IPC Namespace (ipc) – Isolates inter-process communication mechanisms.
• Cgroup Namespace (cgroup) – Provides separate views of control groups.

Demo: unshare

sudo unshare --pid --fork --mount-proc bash

--pid: Creates a new PID namespace.

--fork: Forks a new process to run the specified program (in this case, bash).

--mount-proc: Mounts a new /proc filesystem for the new PID namespace.

Look at ps auxf inside and outside the new shell

Find the external PID X of the new internal PID 1

Look at both of sudo ls -l /proc/{$X,$$}/ns

NOMMU demo

Q: Can we achieve address space isolation without the MMU?

A: No

What can a nommu do?

1. NOMMU lack of memory isolation demo

Enabling the MMU:

1. Enable the MMU (disable RISCV_M_MODE, enable MMU in menuconfig)

2. Recompile the kernel, init, and run

3. Successful segfault!

Types of kernel addresses in MMU mode

Logical: fixed offset from physical memory

• This is what phys_to_virt does

Virtual: has entry in page tables

Slob/Slab/Slub allocators

1. SLOB

   • Simple list of blocks

   • Slowest, oldest, simplest

   • Uses global slob_lock

2. SLAB

   • Newer, but not the best

   • Still locks

   • Improved performance via caching and per-cpu lists

3. SLUB

   • fastest, newest

   • Lockless fastpath for alloc/deallocs

   • Locks only when crossing CPU boundary

Demo: /proc/slabinfo

To set allocator: add slab_allocator=sl{u,a,o}b to kernel command line at boot

Collectively, these are the "slab allocators" or "slab layer"

Paging, page reclaim, swap

What weighs more: 10G of physical memory of 10G of virtual memory (joke)

Technically: virtual memory requires pagetables so there is some overhead

What's bigger: physical or virtual memory?

Answer: virtual memory. Why?

On 64 bit system each process had maximum address space of size 2^64!

Therefore: we need to swap out pages so physical memory doesn't get full

What is page reclaim

When kernel is under memory pressure:

• Toss File-backed memory pages away

• If pages are dirty, write back to backing file first

• Swap out Anonymous pages

What is swap?

Move memory pages into swap space

• Demo: look at free -mh

Can be a swap file, partition, or compressed RAM

• Demo: see zram with lsblk

Uses least recently used (LRU) algorithm

OOM

Q: What happens when you run out of memory?

Answer: out of memory killer is activated to kill a memory-hogging process

Demo: look at mm/oom_kill.c

Page allocation: buddy system

Pages allocation in power-of-two groups

Use get_free_pages() API

You can go around the slab allocators

Summary

• fork() is a mirage

• Virtual memory is not strictly required

• Virtual memory provides important protections

• The kernel contains multiple memory allocators

• The slab allocators rely on the page allocator on the backend

• There are many aspects to the kernel memory management system beyond this intro

• We are only scratching the surface