Long mode extends general registers to 64 bits (RAX, RBX, RIP, RSP, RFLAGS, etc), and adds eight additional integer registers (R8, R9, ..., R15) plus eight more SSE registers (XMM8 to XMM15) to the CPU. Linear addresses are extended to 64 bit (however, a given CPU may implement less than this) and the physical address space is extended to 52 bits (a given CPU may implement less than this). In essence long mode adds another mode to the CPU.
Long mode does not support hardware task switching or virtual 8086 tasks. In long mode the current CS determines if the code currently running is 64 bit code (true long mode) or 32 bit code (compatibility mode), or even 16-bit protected mode code (still in compatibility mode). Using paging has become mandatory, and segmentation has been stripped down for performance reasons.
The first 64 bit CPUs from both Intel and AMD support 40 bit physical addresses and 48 bit linear addresses.
Segmentation in Long Mode
Segmentation in long mode functions with a flat model with the exception of two registers: FS and GS. Setting the base address for these two segment registers is possible via two specific Model Specific Register (MSR)s, FS.base (C000_0100h) and GS.base (C000_0101h).
Additionally there is a long mode specific instruction called SWAPGS, which swaps the contents of GS.base and another MSR called KernelGSBase (C000_0102h). This instruction is particularly useful for preserving kernel information for a specific logical processor core across context switches. Note: This is an exchange operation.
This feature overview is incomplete. Please see the [Wikipedia article on x86-64] for more information.
How do I detect if the CPU is 64 bits ?
After calling CPUID with EAX=0x80000001, all AMD64 compliant processors have the longmode-capable-bit turned on in the extended feature flags (bit 29) in EDX. There are also other bits required by long mode; you can check them out in the CPUID docs in the AMD general purpose instruction reference (Link dead, the original author probably meant "AMD64 Architecture Programmer’s Manual Volume 3: General Purpose and System Instructions", found here: http://developer.amd.com/resources/developer-guides-manuals/)
How do I enable Long Mode ?
The steps for enabling long mode are:
- Disable paging
- Set the PAE enable bit in CR4
- Load CR3 with the physical address of the PML4
- Enable long mode by setting the EFER.LME flag in MSR 0xC0000080
- Enable paging
Now the CPU will be in compatibility mode, and instructions are still 32-bit. To enter long mode, the D/B bit (bit 22, 2nd 32-bit value) of the GDT code segment must be clear (as it would be for a 16-bit code segment), and the L bit (bit 21, 2nd 32-bit value) of the GDT code segment must be set. Once that is done, the CPU is in 64-bit long mode.
Are there restrictions on 32-bit code running in Legacy Mode ?
x86-64 processors can operate in a legacy mode, they still start in real mode and 16 and 32 bit protected mode is still available (along with the associated Virtual 8086 mode). This means an x86 operating system, even DOS, will still run just fine. The only difference is that physical addresses can be up to 52 bits (or as many bits as implemented by the CPU) when PAE is used.
However, Virtual 8086 Mode does not exist in long/compatibility mode.
If you are running on a multi-processor system, you could send one processor a STARTUP IPI to a real mode memory address (see Intel MultiProcessor specification for more details) that loads a real mode program. The main problem with this approach is that it relies on multiple processors being present in the system.
Entering Long Mode directly
Protected mode must be entered before activating long mode. A minimal protected-mode environment must be established to allow long-mode initialization to take place. This environment must include the following:
- A protected-mode IDT for vectoring interrupts and exceptions to the appropriate handlers while in protected mode.
- The protected-mode interrupt and exception handlers referenced by the IDT.
- Gate descriptors for each handler must be loaded in the IDT.
- --AMD64 docs, volume 2, section 14.4 (Enabling Protected Mode), 24593 Rev. 3.10 February 2005
That being said, we have a thread where Brendan shows how to enable 64-bit long mode with no 32-bit IDT and no 32-bit segments - be assured, however, that any paging-related exception that occurs in long mode before you enable 64-bit IDT will cause the processor to reset due to a triple fault.
Notifying the BIOS
In order for the firmware built into the system to optimize itself for running in Long Mode, AMD recommends that the OS notify the BIOS about the intended target environment that the OS will be running in: 32-bit mode, 64-bit mode, or a mixture of both modes. This can be done by calling the BIOS interrupt 15h from Real Mode with AX set to 0xEC00, and BL set to 1 for 32-bit Protected Mode, 2 for 64-bit Long Mode, or 3 if both modes will be used.
64 bit Environment Models
There are three 64 bit programming models you need to consider: LP64, ILP64, LLP64. Each model has its own pitfalls. The I/L/P stand for Int, Long, Pointer, respectively; the 64 is the number of bits in each.
LP64 means Longs (and Long Longs) and Pointers are 64 bits wide, Ints are 32 bits wide. LLP64 means Long Longs and Pointers are 64 bits wide, Longs and Ints are 32 bit wide. ILP64 means Ints, Longs (and Long Longs) and Pointers are 64 bit wide.
Most *nixes use the LP64 model, Windows uses the LLP64 convention. ILP64 is used very rarely.
This table lists the breakdown of sizes in the various programming models.
Models used by 64bit OSs
The following table lists what some current 64bit OS have as a programming model.
|Windows XP X64||LLP64|
|DEC OSF/1 Alpha||LP64|
|HP UX 11||LP64|
Text Segment Types
Another thing that you must keep in mind, that although the address space (and with it all the pointers) are 64 bit wide, the generated code in the text segment is most likely not. That's because by default gcc compiles to the "mov" instruction which has only 32 bit immediate. This means 64 bit programs are limited to 2G, just as 32 bit mode programs.
If you have ever seen an error message like this:
relocation truncated to fit: R_X86_64_32 against symbol
then your code hit this barrier. For Assembly, you must use the "movabs" instruction instead of "mov", and for gcc you need to select a different text segment model with the "-mcmodel" argument.
|Flag||Text Segment Addressing|
|-mcmodel=small||The program and its symbols must be linked in the lower 2 GB of the address space (this is the default model)|
|-mcmodel=large||This model makes no assumptions about addresses and sizes of sections.|
|-mcmodel=medium||The program is linked in the lower 2 GB of the address space. Small symbols are also placed there. Symbols with sizes larger than -mlarge-data-threshold are put into large data or bss sections and can be located above 2GB.|
|-mcmodel=kernel||The kernel runs in the negative 2 GB of the address space. This model has to be used for Linux kernel code.|
It worth noting that code models are different for architectures, as they are tied with the instruction encoding. For example, AArch64 has a "-mcmodel=tiny" too, which allows 1M addressing, and for AArch64 "-mcmodel=small" has a 4G limitation, not 2G.
From experience, you should not use "-mcmodel=kernel". That is a hack for the Linux kernel on x86_64. It is not available for other architectures, and you should not rely on a hack that was written for a kernel with non-stable API. It might change any time. It is better if you use a proper linker script instead to place your kernel into the negative address range.
- Intel EM64T
- Creating a 64-bit kernel
- Loading a 64-bit kernel
- X86-64 Instruction Encoding
- Setting up long mode