Dr. Dobb's is part of the Informa Tech Division of Informa PLC

This site is operated by a business or businesses owned by Informa PLC and all copyright resides with them. Informa PLC's registered office is 5 Howick Place, London SW1P 1WG. Registered in England and Wales. Number 8860726.

Channels ▼


Direct Port I/O and Windows NT

MAY96: Direct Port I/O and Windows NT

Direct Port I/O and Windows NT

Undocumented features for direct control of hardware devices

Dale Roberts

Dale works with data acquisition and control software at the Vestibular Laboratory of the Johns Hopkins University School of Medicine. He can be reached at [email protected].

Port I/O instructions allow all 80x86 CPUs to communicate with other hardware devices in the system. For low-level, direct control of a hardware device, the C functions _inp() and _outp() (implemented using the 80x86 processor's IN and OUT instructions) let you read from or write to an I/O port. However, inserting _inp() or _outp() in a Windows NT application gives you a privileged-instruction exception message and the option of terminating or debugging the offending app. If you attempt port I/O from a 16-bit DOS app in an NT console window, the I/O is either ignored or emulated by NT's virtual device drivers--you don't get an exception, but you don't get the direct I/O either.

This isn't a bug; NT is supposed to work this way. The NT architects decided that it would be too risky to allow applications to directly access the system hardware. With unrestricted I/O access, an application could turn off all interrupts, take over the system, and trash the display or the hard drive. A buggy program could unintentionally do the same. NT's architecture requires that all hardware be accessed via kernel-mode device drivers--special, trusted pieces of software that essentially become part of the operating system when loaded. These device drivers have complete access to the entire system memory, all hardware devices, and all privileged processor instructions. In contrast, applications run in user mode, where they have restricted access to memory--and where the CPU can't execute certain privileged operating-system instructions, including I/O instructions.

The restriction on I/O port access is both a blessing and a curse. On one hand, it makes NT exceptionally stable. Generally, application programmers can write and crash and debug programs all day long without shaking NT. Several applications can run without adversely affecting one another. On the other hand, I/O restrictions prevent you from communicating directly and quickly with the hardware without taking the relatively large amount of time required for a call to a device driver. Whenever you want to communicate with a device driver, you must send a request through NT's I/O subsystem. This can take thousands of processor-clock cycles. A port I/O instruction would take about 30 clock cycles.

Why would you ever need to put I/O instructions in user-mode code? When writing a device driver, it might make things easier if you could write a quick program to interact with the device, sprinkling printf()s and getchar()s among port I/O instructions so that you could verify that you are driving the device correctly before you put the code into an actual device driver and chance a system lockup. Or you may want to write a portion of a driver in a user-mode DLL (as with video drivers, for instance) to achieve a desired level of performance. One of my favorite uses of I/O is for using an oscilloscope to debug programs and time sections of code. To do this, you need to set and clear a bit in a digital output port and monitor the voltage on a scope.

Since direct, user-mode port I/O in NT seems so useful, you'd think there would be an accepted way to achieve it. A quick look through the sample source code in the Windows NT Device Driver Kit (DDK) reveals a program called "PORTIO." Initially, I thought this would provide direct port I/O from an app. However, PORTIO is merely an example showing how to use Win32 DeviceIoControl() calls to a kernel-mode device driver, which implements the actual I/O. Using PORTIO, each I/O operation requires a costly, time-consuming call to the device driver. This was useless for my oscilloscope timings. I needed a better way.

Accomplishing I/O Protection in NT

To figure out how to grant I/O access to a user-mode app, you have to understand how I/O protection is implemented in Windows NT. NT does not actually implement the I/O protection on its own. Since the CPU can trap attempted I/O port accesses, NT depends on this 80x86 feature. The first mechanism that must be understood is the privilege-level system used by the 80x86 processors. Four privilege levels are defined by the processor--0, 1, 2, and 3--and the CPU always operates at one of these levels. The most privileged level is 0; the least

privileged, 3. NT uses only levels 0 and 3. Privilege level 0 is used for the full-access kernel mode, and 3 for the more-restrictive user mode. The current privilege level (CPL) of the processor is stored in the two least-significant bits of the CS (code segment) register.

Rather than statically defining which privilege levels can have I/O access, the CPU defines an I/O privilege level (IOPL) value, which is compared against the CPL to determine if I/O instructions can be used freely. The IOPL is stored in two bits of the processor's EFLAGS register. Any process with a CPL that is numerically greater than the IOPL must go through the I/O protection mechanism when attempting port I/O access. Because the IOPL cannot be less than 0, programs running at privilege level 0 (like kernel-mode device drivers) will always have direct port I/O access. NT sets the IOPL to 0. User-mode code always has a CPL of 3, which is larger than the IOPL. Therefore, user-mode port I/O access attempts must go through the protection mechanism.

Determining if CPL>IOPL is the first step in the protection mechanism. I/O protection is not all-or-nothing. The processor uses a flexible mechanism that allows the operating system to grant direct access to any subset of I/O ports on a task-by-task basis.

The CPU accomplishes this by using a bitmask array, where each bit corresponds to an I/O port. If the bit is a 1, access is disallowed and an exception occurs whenever access to the corresponding port is attempted. If the bit is a 0, direct and unhampered access is granted to that particular port. The I/O address space of the 80x86 processors encompasses 65,536 8-bit ports. The bitmask array is 8192 (0x2000) bytes long, since the bitmask array is packed so that each byte holds eight bits of the array. There is even flexibility in how much of the bitmask array must be provided. You can provide anywhere from 0 to the full 8192 bytes of the table. The table always starts from I/O address 0, but you can choose not to provide the bitmask for upper I/O addresses. Any part of the bitmask that you do not provide is assumed to be 1, and therefore access is not granted to those ports.

The bitmask array, called the I/O Permission bit Map (IOPM), is stored in the Task State Segment (TSS) structure in main memory, which is contained in a special segment referenced by the segment selector in the processor's Task Register (TR). The location of the IOPM within the TSS is flexible. The offset of the IOPM within the TSS is stored in a 2-byte integer at location 0x66 in the TSS; see Figure 1.

NT TSS Specifics

The 80x86 TSS was designed so that each task in the system could have its own TSS. In NT, however, the TSS is not fully used. The TR, which points to the TSS segment descriptor, is never modified. Each process uses the same copy of the TSS, so each process uses the same copy of the IOPM.

In NT, the default IOPM offset points beyond the end of the TSS. This effectively denies access by user-mode processes to all I/O ports. To grant access to I/O ports for user-mode processes, you must modify the IOPM offset so that it is within the TSS, or extend the TSS so that the original default offset falls within the TSS.

The Video-Port Routines

Since I didn't want to reinvent the wheel, I looked through the NT DDK documentation to see if there was a facility to deal with user-mode I/O access. In the Kernel Mode Driver Reference Manual, I came across the video-driver support routines VideoPortMapMemory() and VideoPortSetTrappedEmulatorPorts(). The former grants direct access of memory and I/O ports to user-mode portions of video drivers, presumably for performance. Source-code examples in the DDK show user-mode portions of the VGA video drivers using the IN and OUT port I/O instructions. The latter video-port function grants full-screen DOS-mode programs direct access to a subset of the VGA I/O ports. The description given in the DDK documentation for this second routine even makes reference to the IOPM and notes that it is shared by all of the virtual DOS machines (more accurately, it is shared across all NT processes).

The video-port routines suggest that there is a mechanism within NT for allowing user-mode access to I/O ports. Initially, I tried to use the video routines in a kernel-mode driver to grant I/O access to my user-mode test program, but this turned out to be complicated. The kernel-mode device driver has to pretend that it is a video driver to use these routines. Video header files must be included, the video library must be linked to, and video initialization routines must be called.

The presence of these two routines demonstrates why user-mode I/O can be useful, and their descriptions in the DDK documentation are enlightening. But the functions are intended to be used only with video drivers. Using the video routines with a nonvideo driver was messy, so I dropped this as an option.

Delving Further

The video-port functions are the only documented method for enabling direct-port I/O. Since I found them difficult to use, I decided to create my own. I first tried increasing the size of the TSS so that the default IOPM offset would land within the TSS; see Figure 2. I had to modify the TSS segment descriptor in the global descriptor table (GDT) directly and change the default segment size of 0x20AB to 0x20AB+0xF00 to allow access to the first 0xF00 I/O ports. The processor's TR then had to be reloaded for the change in the TSS descriptor to take effect. It isn't a good idea to extend segments in a haphazard fashion, because all memory must be accounted for by the 80x86 paging system. A page fault could occur during a reference to the IOPM, which would crash the system. But because the physical page size is 4 KB and I did not extend the TSS beyond the end of a physical page, there was no trouble. Since there were only zeros beyond the original end of the TSS, increasing its size granted universal I/O access across all applications. The TOTALIO device driver in Listing One illustrates this.

At first, this may seem like the best possible method to grant I/O access--you set it once and don't need to grant access to each process individually. However, this method is dangerous and unrestrictive. It would allow, for instance, DOS programs to directly access the video registers, even if they were not running in full-screen mode. It would allow DOS disk utilities to access the hard drive directly and wreak havoc on NTFS partitions. NT device drivers keep information on the state of the devices they control, and TOTALIO would allow applications to completely violate this arrangement. As soon as you start up a DOS program, or any other program with port I/O, you risk trashing the whole system.

Granting Access to a Single Process

Since TOTALIO was risky, I looked for a method that would allow a kernel-mode driver to grant I/O access to a single process.

Using a debugger, I examined the NT TSS and found a block of 0xFFs extending from offset 0x88 up to the end of the TSS. I assumed that in NT, the block of 0xFFs was where the IOPM was intended to sit, even though the default IOPM offset points beyond this area. There were 0x2004 bytes of 0xFF. The extra four bytes are present because the 80x86 requires at least one extra byte of 0xFF at the end of the IOPM. The 80x86 requires the extra byte because it always accesses two bytes of the IOPM at a time.

I moved the IOPM offset to point to the start of the 0xFFs, as in Figure 3. I zeroed a few bytes of the IOPM and tried to access ports. Nothing happened. My application still caused exceptions. The kernel-mode device-driver fragment in Listing Two illustrates this attempt.

What was wrong? A visual inspection of a memory dump of an NT process structure showed that NT stores the IOPM offset in a location of its own, within the process structure. The actual IOPM offset in the TSS is loaded from the value in the process structure whenever a process gains control, so changing the TSS directly is of no use. To change the IOPM base address, the value in the process structure must be changed. Once the IOPM offset in the process structure is changed, user-mode I/O access is granted to that process for all ports whose corresponding IOPM access bit is 0. Listing Three illustrates direct modification of the process structure.

Yet Another Way

Early on, I ran across some kernel-mode function names in the NTOSKRNL library (which contains kernel-mode device-driver support routines) that weren't documented in the DDK. Among these functions were Ke386SetIoAccessMap(), Ke386QueryIoAccessMap(), and Ke386IoSetAccessProcess(). From their names, these functions sounded like they might do what I needed, but because they were not documented, I initially had difficulty getting them to work. Only after I completely understood the 80x86 I/O protection mechanism and had my own implementation working, did I have the knowledge to go back and decipher them.

Ke386SetIoAccessMap() takes two arguments: an integer that must be set to 1 in order for the function to work, and a buffer pointer. It copies a supplied I/O access bitmap of length 0x2000 from the buffer into the TSS at offset 0x88. Ke386QueryIoAccessMap() takes the same arguments but does the opposite, copying the current IOPM from the TSS into a buffer of length 0x2000. If the integer argument is set to 0, the set function copies 0xFFs to the IOPM, and the query function copies 0xFFs to the user's buffer.

Ke386IoSetAccessProcess() takes two arguments: a pointer to a process structure obtained from a call to PsGetCurrentProcess(), and an integer that must be set to 1 to grant I/O access, or to 0 to remove I/O access. When the integer argument is 0, the function disables I/O access by setting the IOPM offset of the passed process to point beyond the end of the TSS. When the integer argument is 1, the function enables I/O access by setting the IOPM offset of the passed process to point to the start of the IOPM at offset 0x88 in the TSS.

Using set and query together, it is possible to read, modify, and write back the IOPM, adding access to the desired ports by setting their respective permission bits to zero. Ke386IoSetAccessProcess() then enables the IOPM lookup for the desired process. The kernel-mode device driver in Listing Four, GIVEIO.C, sets the IOPM to 0s to allow full user-mode access to all I/O ports. Listing Five, a user-mode test application called TESTIO.C, uses direct port I/O to exercise the PC's internal speaker.

Direct--for Real?

Once a user-mode process is given permission to access an I/O port, the I/O access proceeds without any further help from the device driver. The purpose of the device driver is to modify the IOPM and the process's copy of the IOPM offset. Once that's done, the application's I/O port accesses proceed unhindered. In fact, the device driver could be unloaded once the IOPM is modified, and the application could still do direct I/O. Listing Five illustrates this by opening and closing the GIVEIO driver, giving the application I/O access, before it performs the port I/O.

I/O Timing

Using port I/O from an application isn't a free ride. There's overhead in the protection mechanism, so the 80x86 IN and OUT instructions take longer in user mode, where CPL > IOPL. The number of processor-clock cycles it takes to execute the IN and OUT instructions varies depending on the CPU mode. In so-called real mode (plain-vanilla, nonextended DOS), an OUT instruction takes 16 processor-clock cycles to execute on a 486; in virtual-8086 mode (a DOS program running in a Windows DOS box or an NT console window), it takes 29 cycles. In protected mode, the execution time depends on whether CPL > IOPL. In the context of NT, this means that it depends on whether a process is executing in kernel mode or user mode. In kernel mode, an OUT instruction takes a mere ten cycles. In user mode it takes a whopping 30 cycles! So the execution time of a "direct" I/O operation is in fact three times longer for a user-mode process, but it is still tiny compared to a device-driver call, which might take on the order of 6000 to 12,000 clocks (somewhere in the 100-200ms range on my 486). The extra time taken when CPL > IOPL, and when the processor is in virtual 8086 mode, is the time it takes the processor to check the bits in the IOPM.

Careful with that Axe, Eugene!

Pardon the Pink Floyd reference, but it seems appropriate to provide warnings about this potentially dangerous tool.

With I/O access knowledge, you may be tempted to start using it for everything, but remember that I/O protection exists in NT for good reasons. I/O protection helps give the operating system its seemingly bullet-proof stability by forcing all access to a device to go through a single, controlled channel. Frivolous use of user-mode port I/O would tend to erode NT's stability. Circumventing an existing kernel-mode device driver is a bad idea. Device drivers maintain information about the state of the devices they control. Bypassing a driver and accessing hardware directly may cause the driver and the hardware to get out of sync, with unpredictable results. Imagine the chaos that would result if every

application tried to directly access the network card.

User-mode I/O may be useful for developing device drivers. It might serve as a development tool for quickly testing new hardware. Direct I/O from user-mode processes should find very little use in software that is distributed to end users. It should never occur in an application. If you are accessing a device, you should be doing it from a device driver. User-mode port I/O might occasionally be useful in user-mode portions of a device driver to achieve better overall performance for the driver.

Having ruled out its use in applications, it is likely that even most device drivers would not benefit from user-mode port I/O. Although it may be tempting to use it in every device driver, just to squeeze out that last bit of performance, most devices would not become appreciably faster by using user-mode port I/O. In many devices, the time delays perceived by the user are not in the calls to the device driver, but in the action of the device. The user isn't usually waiting for the device-driver call itself to complete, but rather for the disk drive to spin, the read/write head to move, or the paper to feed through the printer. User-mode I/O should only be used if there is a definite bottleneck in port I/O access from applications, and then, only if direct user-mode I/O access would improve the driver by making a noticeable and significant difference to the user. Even Microsoft uses this technique sparingly. The only device driver in the system's DRIVERS directory that references the three undocumented routines is the VIDEOPRT.SYS driver, which contains the VideoPort...() functions.

If I/O access is done in the user-mode section of a driver, kernel mode may still be needed for, among other things, servicing interrupts and controlling DMA. User-mode port I/O does not remove the necessity of writing kernel-mode device drivers.

If you decide that you want to use port I/O in the user-mode portion of your device driver, your kernel-mode driver should modify only the IOPM permission bits that correspond to the I/O ports required by the user-mode portion of the driver. You should use the Ke386QueryIoAccessMap() to get the current IOPM, zero each of the permission bits required by the driver, then use the Ke386SetIoAccessMap() routine to write the IOPM back. If and when your driver is unloaded, it should set each permission bit back to 1. Only one IOPM is used by all processes in the system, including, possibly, the video driver. For this reason it is important that the IOPM is not simply written with 0xFFs when access is no longer needed. Of course, the usual device-driver rules given in the DDK manual for allocating I/O ports and keeping track of them in the registry would still apply.

Is system security and integrity violated by user-mode port I/O access? No, because a device driver is still required to grant the I/O access to the application, so it is not possible for an app to gain access to I/O ports on its own. The granting of I/O access may be done on a per-process basis, and the kernel-mode device driver that grants I/O access could be modified to grant access only to those processes that it trusts. Only a user running with administrator privileges can load device drivers, so in general, a user-mode application cannot load a device driver and grant itself I/O access unless the administrator is running it. Granting I/O access to a user-mode process is not directly related to NT's security system and does not make any attempt to foil it.

Granting I/O access to a process is a very specific action. It does not enable the use of the other protected 80x86 instructions, such as STI (enable interrupts) and CLI (disable interrupts). These, and the other privileged instructions, can be executed only by a kernel-mode driver.


The technique described here is specific to 80x86-compatible CPUs. Still, NT runs on several other platforms, including the DEC Alpha, MIPS, and PowerPC. Although this specific implementation is not portable to those processors, there shouldn't be any reason why the same effect could not be achieved on them. None of the other processors have I/O instructions; all hardware is memory mapped. Since any physical memory can be mapped into a user-mode process's memory space (see the MAPMEM example program in the DDK), it should be possible to make any hardware accessible to a user-mode process.

None of the techniques presented here are documented by Microsoft, so portability across releases of NT, even on the same processor platform, could be problematic. It is not likely that the whole mechanism would be removed, since the video drivers rely on it. But the names and functionality of specific undocumented routines could change, or the routines could go away altogether and perhaps become embedded in the video-port library.

It appears that the undocumented functions were added to NT to increase video-driver performance. On the 80x86 platform, increasing video performance required allowing access to some of the video I/O ports in user mode. Since this is the only use of the mechanism, and since it is documented indirectly through the VideoPort...() routines, there was no reason for the underlying Ke386...() routines to be documented.

Another reason Microsoft may have chosen not to document this mechanism is that it is not fully implemented. Currently, the IOPM is shared by all user-mode processes in the system. To be safer and more useful, the system should maintain a separate IOPM for each process. One way to do this would be to save the IOPM (or a pointer to one) in the process structure and copy it into the TSS each time a process changes. But copying 8192 bytes would add a large amount of overhead to a task switch. Another way to give each process its own IOPM would be to give each process its own TSS. NT's process structure could be stored in the TSS, since Intel reserves only the first 0x68 bytes of the TSS for the processor's use and allows the rest to be used by the operating system. Switching the TSS for each process requires reloading the TR. The LTR instruction, which loads the 80x86 task register, has only a tiny overhead of 20 clocks on a 486. Each TSS could have its own segment descriptor in the GDT. Or, since segment descriptors are a commodity, the segment descriptor for the TSS could be modified for each process switch. In any case, keeping the entire IOPM would require an overhead of 8 KB for each process. An alternative would be to store only as much of the IOPM as is needed, and to only create a new TSS and store the IOPM for processes that require user-mode I/O access. Most I/O devices exist below the 0x400 port address, so this would require only 0x80 (128) bytes of storage. To achieve full generality, NT could just save as much of the IOPM as is necessary to map the

highest I/O address that needs to be accessed.


The availability of direct port I/O in user-mode processes opens new doors for NT programmers. Hopefully this technique will prove useful in dealing with hardware devices in an 80x86 NT environment.

Figure 1: The segment selector in the processor's TR points to the segment descriptor in the GDT, which defines the location and size of the TSS in memory. The IOPM is stored as part of the TSS. Its offset is stored in a 2-byte integer at location 0x66 in the TSS.

Figure 2: The default TSS size is 0x20AB. We need to extend it to 0x2FAB so that the IOPM offset falls within the TSS.

Figure 3: NT places the IOPM at offset 0x88 in the TSS. We need to modify the IOPM offset to point to this area.

Listing One

TOTALIO.SYS -- by Dale Roberts
Compile: Use DDK BUILD facility
Purpose: Give direct port I/O access to the whole system. This driver grants
total system-wide I/O access to all applications. Very dangerous, but useful 
for short tests.  Note that no test application is required. Just use control 
panel or "net start totalio" to start the device driver.  When the driver is
stopped, total I/O is removed.  Because no Win32 app needs to communicate with
the driver, we don't have to create a device object. So we have a tiny driver 
here. Since we can safely extend the TSS only to the end of the physical memory
page in which it lies, the I/O access is granted only up to port 0xf00. 
Accesses beyond this port address will still generate exceptions.
#include <ntddk.h>
/* Make sure our structure is packed properly, on byte boundary, not
 * on the default doubleword boundary. */
#pragma pack(push,1)
/* Structures for manipulating the GDT register and a GDT segment
 * descriptor entry.  Documented in Intel processor handbooks. */
typedef struct {
    unsigned short  limit;
    GDTENT  *base;
typedef struct {
    unsigned limit : 16;
    unsigned baselo : 16;
    unsigned basemid : 8;
    unsigned type : 4;
    unsigned system : 1;
    unsigned dpl : 2;
    unsigned present : 1;
    unsigned limithi : 4;
    unsigned available : 1;
    unsigned zero : 1;
    unsigned size : 1;
    unsigned granularity : 1;
    unsigned basehi : 8;
#pragma pack(pop)
/* This is the lowest level for setting the TSS segment descriptor limit field.
 * We get the selector ID from the STR instruction, index into the GDT, and 
 * poke in the new limit.  In order for the new limit to take effect, we must 
 * then read the task segment selector back into the task register (TR).
void SetTSSLimit(int size)
    GDTREG gdtreg;
    GDTENT *g;
    short TaskSeg;
    _asm cli;                           // don't get interrupted!
    _asm sgdt gdtreg;                   // get GDT address
    _asm str TaskSeg;                   // get TSS selector index
    g = gdtreg.base + (TaskSeg >> 3);   // get ptr to TSS descriptor
    g->limit = size;                    // modify TSS segment limit
//  MUST set selector type field to 9, to indicate the task is
// NOT BUSY.  Otherwise the LTR instruction causes a fault.
    g->type = 9;                        // mark TSS as "not busy"
//  We must do a load of the Task register, else the processor
// never sees the new TSS selector limit.
    _asm ltr TaskSeg;                   // reload task register (TR)
    _asm sti;                           // let interrupts continue
/* This routine gives total I/O access across the whole system. It does this
 * by modifying the limit of the TSS segment by direct modification of the TSS
 * descriptor entry in the GDT. This descriptor is set up just once at system 
 * init time. Once we modify it, it stays untouched across all processes.
void GiveTotalIO(void)
    SetTSSLimit(0x20ab + 0xf00);
/* This returns the TSS segment to its normal size of 0x20ab, which
 * is two less than the default I/O map base address of 0x20ad. */
void RemoveTotalIO(void)
/****** Release all memory 'n' stuff. *******/
    IN  PDRIVER_OBJECT  DriverObject
/****** Entry routine.  Set everything up. *****/
NTSTATUS DriverEntry(
    IN PDRIVER_OBJECT DriverObject,
    DriverObject->DriverUnload = TotalIOdrvUnload;
    return STATUS_SUCCESS;

Listing Two

This code fragment illustrates the unsuccessful attempt to directly modify
the IOPM base address. This code would appear in a kernel-mode device driver.
Refer to the GIVEIO.C listing for a complete device driver example.
/* Make sure our structure is packed properly, on byte boundary, not
 * on the default doubleword boundary. */
#pragma pack(push,1)
/* Structure of a GDT (global descriptor table) entry; from processor manual.*/
typedef struct {
    unsigned limit : 16;
    unsigned baselo : 16;
    unsigned basemid : 8;
    unsigned type : 4;
    unsigned system : 1;
    unsigned dpl : 2;
    unsigned present : 1;
    unsigned limithi : 4;
    unsigned available : 1;
    unsigned zero : 1;
    unsigned size : 1;
    unsigned granularity : 1;
    unsigned basehi : 8;
/* Structure of the 48 bits of the GDT register that are stored
 * by the SGDT instruction. */
typedef struct {
    unsigned short  limit;
    GDTENT  *base;
#pragma pack(pop)
/* This code demonstrates the brute force approach to modifying the IOPM base.
 * The IOPM base is stored as a two byte integer at offset 0x66 within the TSS,
 * as documented in the processor manual. In Windows NT, the IOPM is stored 
 * within the TSS starting at offset 0x88, and going for 0x2004 bytes. This is
 * not documented anywhere, and was determined by inspection. The code here 
 * puts some 0's into the IOPM so that we can try to access some I/O ports, 
 * then modifies the IOPM base address. This code is unsuccessful because NT 
 * overwrites the IOPM base on each process switch. */
void GiveIO()
    GDTREG gdtreg;
    GDTENT *g;
    short TaskSeg;
    char *TSSbase;
    int i;
    _asm str TaskSeg;                   // get the TSS selector
    _asm sgdt gdtreg;                   // get the GDT address
    g = gdtreg.base + (TaskSeg >> 3);   // get the TSS descriptor
                                        // get the TSS address
    TSSbase = (PVOID)(g->baselo | (g->basemid << 16) 
                       | (g->basehi << 24));
    for(i=0; i < 16; ++i)               // poke some 0's into the
        TSSbase[0x88 + i] = 0;          //   IOPM
    *((USHORT *)(TSSbase + 0x66)) = 0x88;

Listing Three

/* From inpection of the TSS we know that NT's default IOPM offset is 0x20AD.
 * From an inspection of a dump of a process structure, we can find the bytes 
 * 'AD 20' at offset 0x30.  This is where NT stores the IOPM offset for each 
 * process, so that I/O access can be granted on a process-by-process basis.  
 * This portion of the process structure is not documented in the DDK.
 * This kernel mode driver fragment illustrates the brute force
 * method of poking the IOPM base into the process structure. */
void GiveIO()
    char *CurProc;
    CurProc = IoGetCurrentProcess();
    *((USHORT *)(CurProc + 0x30)) = 0x88;

Listing Four

GIVEIO.SYS -- by Dale Roberts
Compile:    Use DDK BUILD facility
Purpose:    Give direct port I/O access to a user mode process.
#include <ntddk.h>
#include <mondebug.h>
/* The name of our device driver. */
#define DEVICE_NAME_STRING  L"giveio"
/* This is the "structure" of the IOPM.  It is just a simple character array 
 * of length 0x2000. This holds 8K * 8 bits -> 64K bits of the IOPM, which 
 * maps the entire 64K I/O space of the x86 processor.  Any 0 bits will give
 * access to the corresponding port for user mode processes. Any 1
 * bits will disallow I/O access to the corresponding port. */
#define IOPM_SIZE   0x2000
/* This will hold simply an array of 0's which will be copied into our actual 
 * IOPM in the TSS by Ke386SetIoAccessMap(). The memory is allocated at 
 * driver load time.  */
IOPM *IOPM_local = 0;
/* These are the two undocumented calls that we will use to give the calling 
 * process I/O access. Ke386IoSetAccessMap() copies the passed map to the TSS.
 *  Ke386IoSetAccessProcess() adjusts the IOPM offset pointer so that the newly
 * copied map is actually used.  Otherwise, the IOPM offset points beyond the 
 * end of the TSS segment limit, causing any I/O access by the user-mode 
 * process to generate an exception. */
void Ke386SetIoAccessMap(int, IOPM *);
void Ke386QueryIoAccessMap(int, IOPM *);
void Ke386IoSetAccessProcess(PEPROCESS, int);
/***** Release any allocated objects. ******/
VOID GiveioUnload(IN PDRIVER_OBJECT DriverObject)
    WCHAR DOSNameBuffer[] = L"\\DosDevices\\" DEVICE_NAME_STRING;
        MmFreeNonCachedMemory(IOPM_local, sizeof(IOPM));
    RtlInitUnicodeString(&uniDOSString, DOSNameBuffer);
    IoDeleteSymbolicLink (&uniDOSString);
 Set the IOPM (I/O permission map) of the calling process so that it is given
full I/O access. Our IOPM_local[] array is all zeros, so IOPM will be all 0s.
If OnFlag is 1, process is given I/O access. If it is 0, access is removed.
VOID SetIOPermissionMap(int OnFlag)
    Ke386IoSetAccessProcess(PsGetCurrentProcess(), OnFlag);
    Ke386SetIoAccessMap(1, IOPM_local);
void GiveIO(void)
Service handler for a CreateFile() user mode call. This routine is entered in
the driver object function call table by DriverEntry(). When the user-mode 
application calls CreateFile(), this routine gets called while still in the 
context of the user-mode application, but with the CPL (the processor's Current
Privelege Level) set to 0. This allows us to do kernel-mode operations. 
GiveIO() is called to give the calling process I/O access. All the user-mode 
app needs do to obtain I/O access is open this device with CreateFile(). No
other operations are required.
NTSTATUS GiveioCreateDispatch(
    IN  PDEVICE_OBJECT  DeviceObject,
    IN  PIRP            Irp
    GiveIO();           // give the calling process I/O access
    Irp->IoStatus.Information = 0;
    Irp->IoStatus.Status = STATUS_SUCCESS;
    IoCompleteRequest(Irp, IO_NO_INCREMENT);
    return STATUS_SUCCESS;
Driver Entry routine. This routine is called only once after the driver is 
initially loaded into memory. It allocates everything necessary for the 
driver's operation. In our case, it allocates memory for our IOPM array, and 
creates a device which user-mode applications can open. It also creates a 
symbolic link to the device driver. This allows a user-mode application to 
access our driver using the \\.\giveio notation.
NTSTATUS DriverEntry(
    IN PDRIVER_OBJECT DriverObject,
    PDEVICE_OBJECT deviceObject;
    NTSTATUS status;
    WCHAR NameBuffer[] = L"\\Device\\" DEVICE_NAME_STRING;
    WCHAR DOSNameBuffer[] = L"\\DosDevices\\" DEVICE_NAME_STRING;
    UNICODE_STRING uniNameString, uniDOSString;
    //  Allocate a buffer for the local IOPM and zero it.
    IOPM_local = MmAllocateNonCachedMemory(sizeof(IOPM));
    if(IOPM_local == 0)
    RtlZeroMemory(IOPM_local, sizeof(IOPM));
    //  Set up device driver name and device object.
    RtlInitUnicodeString(&uniNameString, NameBuffer);
    RtlInitUnicodeString(&uniDOSString, DOSNameBuffer);
    status = IoCreateDevice(DriverObject, 0, &uniNameString,
                FILE_DEVICE_UNKNOWN, 0, FALSE, &deviceObject);
        return status;
    status = IoCreateSymbolicLink (&uniDOSString, &uniNameString);
    if (!NT_SUCCESS(status))
        return status;
    //  Initialize the Driver Object with driver's entry points.
    // All we require are the Create and Unload operations.
    DriverObject->MajorFunction[IRP_MJ_CREATE] = GiveioCreateDispatch;
    DriverObject->DriverUnload = GiveioUnload;
    return STATUS_SUCCESS;

Listing Five

TSTIO.EXE -- by Dale Roberts
Compile:    cl -DWIN32 tstio.c
Purpose:    Test the GIVEIO device driver by doing some direct
            port I/O.  We access the PC's internal speaker.
#include <stdio.h>
#include <windows.h>
#include <math.h>
#include <conio.h>
typedef struct {
    short int pitch;
    short int duration;
/* Table of notes. Given in half steps. Communication from "other side."  */
NOTE notes[] = {{14, 500}, {16, 500}, {12, 500}, {0, 500}, {7, 1000}};
/***** Set PC's speaker frequency in Hz.  The speaker is controlled by an
 ***** Intel 8253/8254 timer at I/O port addresses 0x40-0x43. *****/
void setfreq(int hz)
    hz = 1193180 / hz;                      // clocked at 1.19MHz
    _outp(0x43, 0xb6);                      // timer 2, square wave
    _outp(0x42, hz);
    _outp(0x42, hz >> 8);
Pass a note, in half steps relative to 400 Hz.  The 12 step scale is an 
exponential thing. Speaker control is at port 0x61. Setting lowest two bits 
enables timer 2 of the 8253/8254 timer and turns on the speaker.
void playnote(NOTE note)
    _outp(0x61, _inp(0x61) | 0x03);         // start speaker going
    setfreq((int)(400 * pow(2, note.pitch / 12.0)));
    _outp(0x61, _inp(0x61) & ~0x03);        // stop that racket!
  Open and close the GIVEIO device.  This should give us direct I/O
access.  Then try it out by playin' our tune.
int main()
    int i;
    HANDLE h;
    h = CreateFile("\\\\.\\giveio", GENERIC_READ, 0, NULL,
        printf("Couldn't access giveio device\n");
        return -1;
    for(i=0; i < sizeof(notes)/sizeof(int); ++i)
    return 0;

Related Reading

More Insights

Currently we allow the following HTML tags in comments:

Single tags

These tags can be used alone and don't need an ending tag.

<br> Defines a single line break

<hr> Defines a horizontal line

Matching tags

These require an ending tag - e.g. <i>italic text</i>

<a> Defines an anchor

<b> Defines bold text

<big> Defines big text

<blockquote> Defines a long quotation

<caption> Defines a table caption

<cite> Defines a citation

<code> Defines computer code text

<em> Defines emphasized text

<fieldset> Defines a border around elements in a form

<h1> This is heading 1

<h2> This is heading 2

<h3> This is heading 3

<h4> This is heading 4

<h5> This is heading 5

<h6> This is heading 6

<i> Defines italic text

<p> Defines a paragraph

<pre> Defines preformatted text

<q> Defines a short quotation

<samp> Defines sample computer code text

<small> Defines small text

<span> Defines a section in a document

<s> Defines strikethrough text

<strike> Defines strikethrough text

<strong> Defines strong text

<sub> Defines subscripted text

<sup> Defines superscripted text

<u> Defines underlined text

Dr. Dobb's encourages readers to engage in spirited, healthy debate, including taking us to task. However, Dr. Dobb's moderates all comments posted to our site, and reserves the right to modify or remove any content that it determines to be derogatory, offensive, inflammatory, vulgar, irrelevant/off-topic, racist or obvious marketing or spam. Dr. Dobb's further reserves the right to disable the profile of any commenter participating in said activities.

Disqus Tips To upload an avatar photo, first complete your Disqus profile. | View the list of supported HTML tags you can use to style comments. | Please read our commenting policy.