Linux Addressing

The address space in linux is of two kinds

1. Physical Address Space
2. Virtual Address Space

Physical Address Space addresses the contents of the RAM directly.

Virtual Address Space addresses the contents of the RAM not directly, but through a MMU (Memory Management Unit)

The reason why we go with Virtual address space is threefold

1. To provide security between applications. Imagine what would happen if one process could read the memory of another process
2. To provide sandboxing capabilities. Example would be stack of one process remains different from the stack of another process.
3. To provide for encapsulation. The userspace process can just do what it needs to do and not worry about memory management.

If a system is 32 bit , then there are 2³² virtual addresses (4GB) that are possible. Now every process on this 32 bit machine basically thinks its got 4 GB of memory.

Now to add to the complexity :)

Now this 4 GB virtual address space is divided into

1. Userspace virtual address — From 0 GB to 3 GB
2. Kernel virtual address — From 3 GB to 4GB

Now almost every userspace process is going to make a couple of system calls be it to print on the screen , to allocate memory, to take input from user or maybe even to establish a network connection.
So the kernel virtual address space is the mapping of physical address of the kernel on the RAM. This is almost a 1 to 1 mapping except for some conditions which we will talk about later in this article.

Now this kernel virtual address is further divided into

1. LOWMEM — From 3GB to 3GB + 896 MB ( The First 16 MB in LOWMEM is for DMA)
2. HIGHMEM — From 3GB+897M to 4GB

The physical memory is also divided into ZONES

1. ZONE_DMA — From 0MB to 16MB
2. ZONE_NROMAL — From 16MB — 896 MB
3. ZONE_HIGHMEM — From 897MB to remaining physical memory.

So how exactly is this kernel virtual addresses mapped to the physical memory?
Let me show it to you with the help of a picture

Mapping between kernel virtual address space and Physical Address space(Assumption RAM on system is 4GB)

The 128 MB of HIGHMEM is mapped to ZONE_HIGHMEM in the physical memory, this basically means that if we need large amounts of memory in the kernel its given from ZONE_HIGHMEM and if memory is allocated in this region it is not going to be contiguous. ( Like for example a valloc call).
The reason for this is , ZONE_NORMAL and ZONE_DMA are never going to be swapped out and is exclusively reserved for the kernel, whereas ZONE_HIGHMEM is for the userspace process also. ( valloc therefore is slow and results in a lot of thrashing, if you need contiguous memory you need to use kalloc)

Consider an edge case when the RAM present is around 512 MB. In that case ZONE_DMA + ZONE_NORMAL is going to be 512 MB and ZONE_HIGHMEM is going to be 0.