Ntoskrnl.exe

ntoskrnl.exe, also known as the kernel image, is an executable file that contains the kernel and executive layers of the Microsoft Windows NT kernel, and is responsible for hardware abstraction, process handling, and memory management. In addition to the kernel and executive layers, it contains the cache manager, security reference monitor, memory manager, scheduler, and blue screen of death.

Overview

x86 versions of ntoskrnl.exe depend on bootvid.dll, hal.dll and kdcom.dll. However, it is not a native application thus it is not linked against ntdll.dll. Instead, ntoskrnl.exe has its own entry point KiSystemStartup that calls the architecture-independent kernel initialization function. Because it requires a static copy of the C Runtime objects, the executable is usually about 10 MB in size.
In Windows XP and earlier, the Windows installation source ships four kernel image files to support uniprocessor systems, symmetric multiprocessor (SMP) systems, CPUs with PAE, and CPUs without PAE. Windows setup decides whether the system is uniprocessor or multiprocessor, then, installs both the PAE and non-PAE variants of the kernel image for the decided kind. On a multiprocessor system, setup installs ntkrnlmp.exe and ntkrpamp.exe but renames them to ntoskrnl.exe and ntkrnlpa.exe respectively.
Starting with Windows Vista, Microsoft began unifying the kernel images as multi-core CPUs took to the market and PAE became mandatory.
The kernel's code uses prefixes to indicate their associated subsystem.
An example is IoCreateDevice and ObReferenceObjectByHandle. Both functions have different prefix names to differentiate critical managers within the kernel code: Io being used for functions and Ob for Object Manager functions.
Variations of these prefixes exist for internal functions that are not being exported by the kernel, such as adding an i after the first letter or appending p to the full prefix.
The following table lists some known prefixes.

Export Prefix	Internal Prefix	Meaning
`Arb`	Arbp	Plug-and-play resource arbiter
`Cc`	Ccp	File system cache
`Cm`	Cmp	Configuration Manager, the kernel mode side of Windows Registry
`Dbg`	Dbg	Debugging aid functions, such as a software break point
`Dbgk`	Dbgk	A set of debugging functions that are being exposed to user mode through NTDLL.dll
`Ex`	Exp	Windows executive, an "outer layer" of `ntoskrnl.exe`
`FsRtl`	FsRtlp	File system runtime library
`Io`	Iop	I/O manager
`Kd`	Kdp	Kernel debugging facilities, used by WinDbg
`Ke`	Ki	Core kernel routines
`Ke`	Kx	Interrupt handling, semaphores, spinlocks, multithreading and context switching related functions
`Ke`	Ks	Kernel streaming
`Ldr`	Ldrp	NT's PE Executables loader, NTLDR
`Lpc`	Lpcp	Local Procedure Call, an internal, undocumented, interprocess or user/kernel message passing mechanism
`Lsa`	Lsap	Local Security Authority
`Mm`	Mi	Memory management
`Nls`	Nls	Native Language Support
`Ob`	Obp	Object Manager
`Po`	Pop	Plug-and-play and power management
`Ps`	Psp	Process and thread management
`Rtl`	Rtlp	Runtime library, i.e., many utility functions that can be used by native applications, yet don't directly involve kernel support
`Se`	Sep	Security Manager, access token for the Win32 API
`Vf`	Vi	Driver Verifier
`Zw/Nt`		`Nt` or `Zw` are system calls declared in `ntdll.dll` and `ntoskrnl.exe`. When called from `ntdll.dll` in user mode, these groups are almost exactly the same; they trap into kernel mode and call the equivalent function in `ntoskrnl.exe` via the SSDT. When calling the functions directly in `ntoskrnl.exe`, the `Zw` variants ensure kernel mode, whereas the `Nt` variants do not.

Initialization

When control is transferred to the kernel, it receives an information record, known as the Loader Parameter Block, from the bootloader. The block contains information about the hardware, the path to the Windows Registry file, kernel parameters, path of the files loaded by the bootloader. The definition of this structure can be retrieved by using the kernel debugger.
In the x86 architecture, the kernel receives the system already in protected mode, with the GDT, IDT and TSS ready..
The main entry point of ntoskrnl.exe first runs on the boot processor, where it initializes some systemwide kernel components and creates a system thread. After enabling interrupts, the boot processor enters an idle loop. The system thread then starts up any remaining cores and initializes the rest of the system.

Construction

Interrupt handling

Modern operating systems use interrupts instead of I/O port polling to wait for information from devices.
In the x86 architecture, interrupts are handled through the Interrupt Dispatch Table. When a device triggers an interrupt and the interrupt flag in the FLAGS register is set, the processor's hardware looks for an interrupt handler in the table entry corresponding to the interrupt number to which in turn has been translated from IRQ by PIC chips, or in more modern hardware, APIC. Interrupt handlers usually save some subset of the state of registers before handling it and restore them back to their original values when done.
The interrupt table contains handlers for hardware interrupts, software interrupts, and exceptions. For some IA-32 versions of the kernel, one example of such a software interrupt handler is in its IDT table entry 2E₁₆, used in assembly language as INT 2EH for system calls, which are Microsoft internal usage calls, leading to kernel system calls. In the real implementation the entry points to an internal subroutine named KiSystemService. For newer versions, different mechanisms making use of SYSENTER instruction and in x86-64 SYSCALL instruction are used instead.
One notable feature of NT's interrupt handling is that interrupts are usually conditionally masked based on their priority, instead of disabling all IRQs via the interrupt flag. This permits various kernel components to carry on critical operations without necessarily blocking services of peripherals and other devices.

Memory manager

The entire physical memory address range is broken into many small blocks also called pages, 4KB in size each, and mapped to virtual addresses. A few of the properties of each block are stored in structures called page table entries, which are managed by the OS and accessed by the processor's hardware. Page tables are organized into a tree structure, and the physical page number of the top-level table is stored in control register 3.
Microsoft Windows divides virtual address space into two regions. The lower part, starting at zero, is instantiated separately for each process and is accessible from both user and kernel mode. Application programs run in processes and supply code that runs in user mode.
The upper part is accessible only from kernel mode, and with some exceptions, is instantiated just once, system-wide. ntoskrnl.exe is mapped into this region, as are several other kernel mode components. This region also contains data used by kernel mode code, such as the kernel mode heaps and the file system cache.

Arch	MmHighestUserAddress	MmSystemRangeStart
x86	`0x7fffffff`	`0x80000000`
ARM	`0x7fffffff`	`0x80000000`
x86-64	`0x000007ff'ffffffff` `0x00007fff'ffffffff`	`0xffff8000'00000000`

Registry

Windows Registry is a repository for configuration and settings information for the operating system and for other software, such as applications. It can be thought of as a filesystem optimized for small files. However, it is not accessed through file system-like semantics, but rather through a specialized set of APIs, implemented in kernel mode and exposed to user mode.
The registry is stored on disk as several different files called "hives." One, the System hive, is loaded early in the boot sequence and provides configuration information required at that time. Additional registry hives, providing software-specific and user-specific data, are loaded during later phases of system initialization and during user login, respectively.

Drivers

The list of drivers to be loaded from the disk are retrieved from the Services key of the current control set's key in the SYSTEM registry hive. That key stores device drivers, kernel processes and user processes. They are all collectively called "services" and are all stored mixed on the same place.
During initialization or upon driver load request, the kernel traverses that tree looking for services tagged as kernel services.