Linux Memory Map and Address Space

32-bit processors

• Address space limited to 32 bits, i.e. cannot directly address more than 4GB

Virtual addresses space for every processes (and for the kernel): - [ 0x00000000 - 0xc0000000 - 1 ] -> user mode memory: 3GB - [ 0xc0000000 - 0xffffffff ] -> kernel mode memory: 1GB

Kernel is mapped in every process Virtual Address space, so that switching in/out of kernel is fast (uses TLB).

So, if total system memory is <= 1GB: kernel mode can permanently map all physical memory!

But if total system memory is > 1GB: - Physical addresses 0 - 869MB are always mapped in the (kernel mode) virtual addresses from [ 0xc0000000 - +869MB ] (ZONE_NORMAL) Here goes kernel data frequently accessed. These is also known as Low Memory

• Physical addresses above 869MB are mapped on demand into (kernel mode) virtual [ -128MB - 0xffffffff ] (ZONE_HIGHMEM) This is mapped when need to access processes memory, page tables, etc... This is also known as High Memory

References: - https://www.kernel.org/doc/Documentation/vm/highmem.txt

See also /proc/PID/maps and /proc/PID/smaps.

64-bit processors

Virtual address spaces are 64-bits wide, but only 48 bits are used.

• Bits [63:48] must be extension of bit 47, so either all 0 or all 1 (0xffff) So all adresses where this range of bits is not as that are invalid: 0008000000000000 - ffff7fffffffffff

0000000000000000 - 00007fffffffffff (=47 bits) user space

---

ffff800000000000 - ffff87ffffffffff (=43 bits) guard hole, reserved for hypervisor ffff880000000000 - ffffc7ffffffffff (=64 TB) direct mapping of all phys. memory ... ffffffffff600000 - ffffffffff600fff (=4 kB) legacy vsyscall ABI ffffffffffe00000 - ffffffffffffffff (=2 MB) unused hole