[Previous] [Next]

Many devices signal completion of I/O operations by asynchronously interrupting the processor. In this section, I'll discuss how you configure your driver for interrupt handling and how you service interrupts when they occur.

You configure an interrupt resource in your StartDevice function by calling IoConnectInterrupt using parameters that you can simply extract from a CmResourceTypeInterrupt descriptor. Your driver and device need to be entirely ready to work correctly when you call IoConnectInterrupt—you might even have to service an interrupt before the function returns—so you normally make the call near the end of the configuration process. Some devices have a hardware feature that allows you to prevent them from interrupting. If your device has such a feature, disable interrupts before calling IoConnectInterrupt and enable them afterward. The extraction and configuration code for an interrupt would look like this:

1 2 3 4 5

typedef struct _DEVICE_EXTENSION {   ...   PKINTERRUPT InterruptObject;   ...} DEVICE_EXTENSION, *PDEVICE_EXTENSION; ULONG vector;             // interrupt vector KIRQL irql;               // interrupt level KINTERRUPT_MODE mode;     // latching mode KAFFINITY affinity;       // processor affinity BOOLEAN irqshare;         // shared interrupt? ... for (ULONG i = 0; i < nres; ++i, ++resource)   {   switch (resource->Type)     {   case CmResourceTypeInterrupt:    irql = (KIRQL) resource->u.Interrupt.Level;    vector = resource->u.Interrupt.Vector;    affinity = resource->u.Interrupt.Affinity; mode = (resource->Flags == CM_RESOURCE_INTERRUPT_LATCHED)       ? Latched : LevelSensitive; irqshare = resource->ShareDisposition == CmResourceShareShared;    break;   ...   } ... status = IoConnectInterrupt(&pdx->InterruptObject,   (PKSERVICE_ROUTINE) OnInterrupt, (PVOID) pdx, NULL,   vector, irql, irql, mode, irqshare, affinity, FALSE);

1. The Level parameter specifies the interrupt request level (IRQL) for this interrupt.
2. The Vector parameter specifies the hardware interrupt vector for this interrupt. We don't care what this number is, since we're just going to act as a conduit between the PnP Manager and IoConnectInterrupt. All that matters is that the HAL understand what the number means.
3. Affinity is a bit mask that indicates which CPUs will be allowed to handle this interrupt.
4. We need to tell IoConnectInterrupt whether our interrupt is edge-triggered or level-triggered. If the resource Flags are CM_RESOURCE_INTERRUPT_LATCHED, we have an edge-triggered interrupt. Otherwise, we have a level-triggered interrupt.
5. Use this statement to discover whether your interrupt is shared.

In the call to IoConnectInterrupt at the end of this sequence, we will simply regurgitate the values we pulled out of the interrupt resource descriptor. The first argument (&pdx->InterruptObject) indicates where to store the result of the connection operation—namely, a pointer to a kernel interrupt object that describes your interrupt. The second argument (OnInterrupt) is the name of your interrupt service routine; I'll discuss ISRs a bit further on in this chapter. The third argument (pdx) is a context value that will be passed as an argument to the ISR each time your device interrupts. I'll have more to say about this context parameter later as well in "Selecting an Appropriate Context Argument."

The fifth and sixth arguments (vector and irql) specify the interrupt vector number and interrupt request level, respectively, for the interrupt you're connecting. The eighth argument (mode) is either Latched or LevelSensitive to indicate whether the interrupt is edge-triggered or level-triggered. The ninth argument is TRUE if your interrupt is shared with other devices and FALSE otherwise. The tenth argument (affinity) is the processor affinity mask for this interrupt. The eleventh and final argument indicates whether the operating system needs to save the floating-point context when the device interrupts. Since you're not allowed to do floating-point calculations in an ISR on an x86 platform, a portable driver would always set this flag to FALSE.

I haven't yet described two other arguments to IoConnectInterrupt. These become important when your device uses more than one interrupt. In such a case, you would create spin locks for your interrupts and initialize them by calling KeInitializeSpinLock. You would also calculate the largest IRQL needed by any of your interrupts before connecting any of them. In each call to IoConnectInterrupt, you'd specify the address of the appropriate spin lock for the fourth argument (which is NULL in my example) and you'd specify the maximum IRQL for the seventh argument (which is irql in my example). This seventh argument indicates the IRQL used for synchronizing the interrupts, which you should make the maximum of all your interrupt IRQLs so that you're troubled by only one of your interrupts at a time.

If, however, your device uses only a single interrupt, you won't need a special spin lock (because the I/O Manager automatically allocates one for you) and the synchronization level for your interrupt will be the same as the interrupt IRQL.

When your device generates an interrupt, the HAL selects a CPU to service the interrupt based on the CPU affinity mask you specified. It raises that CPU's IRQL to the appropriate synchronization level and claims the spin lock associated with your interrupt object. Then it calls your ISR, which would have the following skeletal form:

BOOLEAN OnInterrupt(PKINTERRUPT InterruptObject, PVOID Context)   {   if (<device not interrupting>)     return FALSE;   <handle interrupt>   return TRUE;   }

Windows NT's interrupt-handling mechanism assumes that hardware interrupts can be shared by many devices. Thus your first job in the ISR is to determine whether your device is interrupting at the present moment. If not, you return FALSE right away so that the HAL can send the interrupt to another device driver. If yes, you clear the interrupt at the device level and return TRUE. Whether the HAL then calls other drivers' ISRs depends on whether the device interrupt is edge-triggered or level-triggered and on other platform details.

Your main job in the ISR is to service your hardware to clear the interrupt. I'll have some general things to say about this job, but the details pretty much depend on how your hardware works. Once you've performed this major task, you return TRUE to indicate to the HAL that you've serviced a device interrupt.

ISRs execute at an IRQL higher than DISPATCH_LEVEL. All code and data used in an ISR must therefore be in nonpaged memory. Furthermore, the set of kernel-mode functions that an ISR can call is very limited.

Since an ISR executes at elevated IRQL, it freezes out other activities on its CPU that require the same or a lower IRQL. For best system performance, therefore, your ISR should execute as quickly as possible. Basically, do the minimum amount of work required to service your hardware and return. If there is additional work to do (such as completing an IRP), schedule a DPC to handle that work.

Despite the admonition you usually receive to do the smallest amount of work possible in your ISR, you don't want to carry that idea to an extreme. For example, if you're dealing with a device that interrupts to signal its readiness for the next output byte, go ahead and send the next byte directly from your ISR. It's fundamentally silly to schedule a DPC just to transfer a single byte. Remember that the end user wants you to service your hardware (or else he or she wouldn't have the hardware installed on the computer), and you are entitled to your fair share of system resources to provide that service.

But don't go crazy calculating pi to a thousand decimal places in your ISR, either (unless your device requires you to do something that ridiculous, and it probably doesn't). Good sense should tell you what the right balance of work between an ISR and a DPC routine should be.

In the call to IoConnectInterrupt, the third argument is an arbitrary context value that eventually shows up as the second argument to your ISR. You want to choose this argument so as to allow your ISR to execute as rapidly as possible; the address of your device object or of your device extension would be a good choice. The device extension is where you'll be storing items—such as your device's base port address—that you'll use in testing whether your device is currently asserting an interrupt. To illustrate, suppose that your device, which is I/O-mapped, has a status port at its base address and that the low-order bit of the status value indicates whether the device is currently trying to interrupt. If you adopt my suggestion, the first few lines of your ISR would read like this:

BOOLEAN OnInterrupt(PKINTERRUPT InterruptObject, PDEVICE_EXTENSION pdx)   {   UCHAR devstatus = READ_PORT_UCHAR(pdx->portbase);   if (!(devstatus & 1))     return FALSE;   <etc.>   }

The fully optimized code for this function will require only a few instructions to read the status port and test the low-order bit.

If you elect to use the device extension as your context argument, be sure to supply a cast when you call IoConnectInterrupt:

IoConnectInterrupt(..., (PKSERVICE_ROUTINE) OnInterrupt, ...);

If you omit the cast, the compiler will generate an exceptionally obscure error mes-sage because the second argument to your OnInterrupt routine (a PDEVICE_EXTENSION) won't match the prototype of the function pointer argument to IoConnect-Interrupt, which demands a PVOID.

As a general rule, the ISR shares data and hardware resources with other parts of the driver. Anytime you hear the word share, you should immediately start thinking about synchronization problems. For example, a standard UART (universal asynchronous receiver-transmitter) device has a data port the driver uses for reading and writing data. You'd expect a serial port driver's ISR to access this port from time to time. Changing the baud rate also entails setting a control flag called the divisor latch, performing two single-byte write operations to this same data port, and then clearing the divisor latch. If the UART were to interrupt in the middle of changing the baud rate, you can see that a data byte intended to be transmitted could easily end up in the baud-rate divisor register or that a byte intended for the divisor register could end up being transmitted as data.

The system guards the ISR with a spin lock and with a relatively high IRQL—the device IRQL (DIRQL). To simplify the mechanics of obtaining the same spin lock and raising IRQL to the same level as an interrupt, the system provides this service function:

BOOLEAN result = KeSynchronizeExecution(InterruptObject,   SynchRoutine, Context);

where InterruptObject (PKINTERRUPT) is a pointer to the interrupt object describing the interrupt we're trying to synchronize with, SynchRoutine (PKSYNCHRONIZE_ROUTINE) is the address of a callback function in our driver, and Context (PVOID) is an arbitrary context parameter to be sent to the SynchRoutine as an argument. We use the generic term synch critical section routine to describe a subroutine that we call by means of KeSynchronizeExecution. The synch critical section routine has the following prototype:

BOOLEAN SynchRoutine(PVOID Context);

That is, it receives a single argument and returns a BOOLEAN result. When it gets control, the current CPU is running at the synchronization IRQL that the original call to IoConnectInterrupt specified, and it owns the spin lock associated with the interrupt. Consequently, interrupts from the device are temporarily blocked out, and the SynchRoutine can freely access data and hardware resources that it shares with the ISR.

KeSynchronizeExecution returns whatever value SynchRoutine returns, by the way. This gives you a way of providing a little bit—actually 8 bits, since BOOLEAN is declared as an unsigned character—of feedback from SynchRoutine to whatever calls KeSynchronizeExecution.

Completely servicing a device interrupt often requires you to perform operations that aren't legal inside an ISR or that are too expensive to carry out at the elevated IRQL of an ISR. To avoid these problems, the designers of Windows NT provided the deferred procedure call mechanism. The DPC is a general-purpose mechanism, but you use it most often in connection with interrupt handling. In the most common scenario, your ISR decides that the current request is complete and requests a DPC . Later on, the kernel calls your DPC routine at DISPATCH_LEVEL. While restrictions on what service routines you can call and on paging still apply, fewer restrictions apply because you're now running at a lower IRQL than inside the ISR. In particular, it's legal to call routines like IoCompleteRequest or IoStartNextPacket that are logically necessary at the end of an I/O operation.

Every device object gets a DPC object "for free." That is, the DEVICE_OBJECT has a DPC object—named, prosaically enough, Dpc—built in. You need to initialize this built-in DPC object shortly after you create your device object:

NTSTATUS AddDevice(...)   {   PDEVICE_OBJECT fdo;   IoCreateDevice(..., &fdo);   IoInitializeDpcRequest(fdo, DpcForIsr);   ...   }

IoInitializeDpcRequest is a macro in WDM.H that initializes the device object's built-in DPC object. The second argument is the address of the DPC procedure that I'll show you presently.

With your initialized DPC object in place, your ISR can request a DPC by using the following macro:

BOOLEAN OnInterrupt(...)   {   ...   IoRequestDpc(pdx->DeviceObject, NULL, NULL);   ...   }

This call to IoRequestDpc places your device object's DPC object in a systemwide queue, as illustrated in Figure 7-5.

Figure 7-5. Processing DPC requests.

The two NULL parameters are context values that don't really have a good use in this particular situation. Later on, when no other activity is occurring at DISPATCH_LEVEL, the kernel removes your DPC object from the queue and calls your DPC routine, which has the following prototype:

VOID DpcForIsr(PKDPC Dpc, PDEVICE_OBJECT fdo, PIRP junk, PVOID Context)   {   }

What you do inside the DPC routine depends in great measure on how your device works. A likely task would be to complete the current IRP and release the next IRP from the queue. If you use the "standard model" for IRP queuing, the code would be as follows:

VOID DpcForIsr(...)   {   PIRP Irp = fdo->CurrentIrp;   IoStartNextPacket(fdo, TRUE);   IoCompleteRequest(Irp, <boost value>);   }

The TRUE argument to IoStartNextPacket indicates that the next IRP is cancellable—meaning that the original call to IoStartPacket specified a cancel routine—and causes IoStartNextPacket to acquire and release the global cancel spin lock around its access to the device queue and CurrentIrp.

In this code fragment, we rely on the fact that the I/O Manager sets the device object's CurrentIrp field to point to the IRP it sends to our StartIo routine. The IRP we want to complete is the one that's the CurrentIrp when we commence the DPC routine. It's customary to call IoStartNextPacket before IoCompleteRequest so that we can get our device busy with a new request before we start the potentially long process of completing the current IRP.

If you use the DEVQUEUE object presented in the previous chapter for IRP queuing, the code would be similar:

VOID DpcForIsr(...)   {   PDEVICE_EXTENSION pdx = ...;   PIRP Irp = GetCurrentIrp(&pdx->dqRead);   StartNextPacket(&pdx->dqRead, fdo);   IoCompleteRequest(Irp, <boost value>);   }

I've glossed over two fairly important details and a minor one about DPCs until now. The first important detail is implicit in the fact that you have a DPC object that gets put onto a queue by IoRequestDpc. If your device generates an additional interrupt before the DPC routine actually runs, and if your ISR requests another DPC, the kernel will simply ignore the second request. In other words, your DPC object will be on the queue one time no matter how many DPCs are requested by successive invocations of your ISR, and the kernel will call your callback routine only one time. During that one invocation, your DPC routine needs to accomplish all the work related to all the interrupts that have occurred since the last DPC.

As soon as the DPC dispatcher dequeues your DPC object, it's possible for something to queue it again, even while your DPC routine executes. This won't cause you any grief if the object happens to be queued on the same CPU both times. The second important detail about DPC processing, therefore, has to do with CPU affinity. Normally, the kernel queues a DPC object for handling on the same processor that requests the DPC—for example, the processor that just handled an interrupt and called IoRequestDpc. As soon as the DPC dispatcher dequeues the DPC object and calls your callback routine on one CPU, it's theoretically possible for your device to interrupt on a different CPU, which might end up requesting a DPC that could execute simultaneously on that different CPU. Whether simultaneous execution of your DPC routine poses a problem or not depends, obviously, on the details of your coding.

You can avoid the potential problems that might come from having your DPC routine simultaneously active on multiple CPUs in several ways. One way, which isn't the best, is to designate a particular CPU for running your DPC by calling KeSetTargetProcessorDpc. Also, you could theoretically restrict the CPU affinity of your interrupt when you first connect it; if you never queue the DPC except from your ISR, you'll never be executing the DPC on any different CPU. The real reason you're able to specify the CPU affinity of a DPC or an interrupt, however, is to improve performance by allowing the code and data accessed during your DPC or ISR routines to remain in a cache.

You can also use a spin lock or other synchronization primitive to prevent interference between two instances of your DPC routine. Be careful of using a spin lock here: you often need to coordinate the hypothetical multiple instances of your DPC routine with your ISR, and an ISR runs at too high an IRQL to use an ordinary spin lock. An interlocked list—that is, one you manipulate by using support functions in the same family as ExInterlockedInsertHeadList—might help you, since (so long as you never explicitly acquire the same spin lock that you use to guard the list) you can use the list at any IRQL. Interlocked forms of the bitwise OR and AND operators also might help by allowing you to manage a bit mask (such as a mask indicating recent interrupt conditions) that controls what your DPC routine is supposed to accomplish; you can cobble these functions together with the help of InterlockedCompareExchange.

Most simply, you can just make sure that your device won't interrupt in between the time you request a DPC and the time your DPC routine finishes its work. ("Yo, hardware guys, stop flooding me with interrupts!")

The third DPC detail, which I consider less crucial than the two I've just explained, concerns the importance of the DPC. By calling KeSetImportanceDpc, you can designate one of three importance levels for your DPC:

• MediumImportance is the default and indicates that the DPC should be queued after all currently queued DPCs. If the DPC is queued to another processor, that other processor won't necessarily be interrupted right away to service the DPC. If it's queued to the current processor, the kernel will request a DPC interrupt as soon as possible to begin servicing DPCs.
• HighImportance causes the DPC to be queued first. If two or more high importance DPCs get requested at about the same time, the last one queued gets serviced first.
• LowImportance causes the DPC to be queued last. In addition, the kernel won't necessarily request a DPC interrupt for whatever processor is destined to service the DPC.

The net effect of a DPC's importance level is to influence, but not necessarily control, how soon the DPC occurs. Even a DPC that has low importance might trigger a DPC interrupt on another CPU if that other CPU reaches some threshold for queued DPCs or if DPCs haven't been getting processed fast enough on it. If your device is capable of interrupting again before your DPC routine runs, changing your DPC to low importance will increase the likelihood that you'll have multiple work items to perform. If your DPC has an affinity for some CPU other than the one that requests the DPC, choosing high importance for your DPC will increase the likelihood that your ISR will still be active when your DPC routine begins to run. But neither of these possibilities is a certainty; conversely, altering or not altering your importance can't prevent either of them from happening.

You can create other DPC objects besides the one named Dpc in a device object. Simply reserve storage—in your device extension or some other persistent place that isn't paged—for a KDPC object, and initialize it:

typedef struct _DEVICE_EXTENSION {   ...   KDPC CustomDpc;   ... }; KeInitializeDpc(&pdx->CustomDpc, (PKDEFERRED_ROUTINE) DpcRoutine, fdo);

In the call to KeInitializeDpc, the second argument is the address of a DPC routine in nonpaged memory, and the third argument is an arbitrary context parameter that will be sent to the DPC routine as its second argument.

To request a deferred call to a custom DPC routine, call KeInsertQueueDpc:

BOOLEAN inserted = KeInsertQueueDpc(&pdx->CustomDpc, arg1, arg2);

Here, arg1 and arg2 are arbitrary context pointers that will be passed to the custom DPC routine. The return value is FALSE if the DPC object was already in a processor queue and TRUE otherwise.

Also, you can also remove a DPC object from a processor queue by calling KeRemoveQueueDpc.

I wrote the PCI42 sample driver (available on the companion disc) to illustrate how to write the various different driver routines that a typical interrupt-driven, non-DMA device might use. The method used to handle such a device is often called programmed I/O (PIO) because program intervention is required to transfer each unit of data.

PCI42 is a dumbed-down driver for the S5933 PCI chip set from Applied Micro Circuits Corporation (AMCC). The S5933 acts as a matchmaker between the PCI bus and an add-on device that implements the actual function of a device. The S5933 is very flexible. In particular, you can program nonvolatile RAM so as to initialize the PCI configuration space for your device in any desired way. PCI42 uses the S5933 in its factory default state, however.

To grossly oversimplify matters, a WDM driver communicates with the add-on device connected to an S5933 either by doing DMA (which I'll discuss in the next major section of this chapter) or by sending and receiving data through a set of mailbox registers. PCI42 will be using one byte in one of the mailbox registers to transfer data one byte at a time.

The AMCC development kit for the S5933 (part number S5933DK1) includes two breadboard cards and an ISA (Industry Standard Architecture) interface card that connects to the S5933 development board via a ribbon cable. The ISA card allows you to access the S5933 from the add-on device side in order to provide software simulation of the add-on function. One component of the PCI42 sample is a driver (S5933DK1.SYS) for the ISA card that exports an interface for use by test programs.

Hardware people will snicker at the simplicity of the way PCI42 manages the device. The advantage of using such a trivial example is that you'll be able to see each step in the process of handling an I/O operation unfold at human speed. So chortle right back if your social dynamics allow it.

The StartDevice function in PCI42 handles a port resource and an interrupt resource. The port resource describes a collection of sixteen 32-bit operation registers in I/O space, and the interrupt resource describes the host manifestation of the device's INTA# interrupt capability. At the end of StartDevice, we have the following device-specific code:

NTSTATUS StartDevice(...)   {   ...   ResetDevice(pdx);   status = IoConnectInterrupt(...);   KeSynchronizeExecution(pdx->InterruptObject,     (PKSYNCHRONIZE_ROUTINE) SetupDevice, pdx);   return STATUS_SUCCESS;   }

That is, we invoke a helper routine (ResetDevice) to reset the hardware. One of the tasks for ResetDevice is to prevent the device from generating any interrupts, insofar as that's possible. Then we call IoConnectInterrupt to connect the device interrupt to our ISR. Even before IoConnectInterrupt returns, it's possible for our device to generate an interrupt, so everything about our driver and the hardware has to be ready to go beforehand. After connecting the interrupt, we invoke another helper routine named SetupDevice to program the device to act the way we want it to. We must synchronize this step with our ISR because it uses the same hardware registers as our ISR would use, and we don't want any possibility of sending the device inconsistent instructions. The SetupDevice call is the last step in PCI42's StartDevice because—contrary to what I told you in Chapter 2, "Basic Structure of a WDM Driver"—PCI42 hasn't registered any device interfaces and therefore has none to enable at this point.

ResetDevice is highly device-specific and reads as follows:

1 2 3

VOID ResetDevice(PDEVICE_EXTENSION pdx)   {   PAGED_CODE(); WRITE_PORT_ULONG((PULONG) (pdx->portbase + MCSR), MCSR_RESET);   LARGE_INTEGER timeout;   timeout.QuadPart = -10 * 10000; // i.e., 10 milliseconds KeDelayExecutionThread(KernelMode, FALSE, &timeout);   WRITE_PORT_ULONG((PULONG) (pdx->portbase + MCSR), 0); WRITE_PORT_ULONG((PULONG) (pdx->portbase + INTCSR),     INTCSR_INTERRUPT_MASK);   }

1. The S5933 has a master control/status register (MCSR) that controls bus-mastering DMA transfers and other actions. Asserting four of these bits resets different features of the device. I defined the constant MCSR_RESET to be a mask containing all four of these reset flags. This and other manifest constants for S5933 features are in the S5933.H file that's part of the PCI42 project.
2. Three of the reset flags pertain to features internal to the S5933 and take effect immediately. Setting the fourth flag to 1 asserts a reset signal for the add-on function. To deassert the add-on reset, you have to explicitly reset this flag to 0. In general, you want to give the hardware a little bit of time to recognize a reset pulse. KeDelayExecutionThread, which I discussed in Chapter 4, "Synchronization," puts this thread to sleep for about 10 milliseconds. You can raise or lower this constant if your hardware has different requirements, but don't forget that the timeout will never be less than the granularity of the system clock. Since we're blocking our thread, we need to be running at PASSIVE_LEVEL in a nonarbitrary thread context. Those conditions are met because our ultimate caller is the PnP Manager, which has sent us an IRP_MN_START_DEVICE in the full expectation that we'd be blocking the system thread we happen to be in.
3. The last step in resetting the device is to clear any pending interrupts. The S5933 has six interrupt flags in an interrupt control/status register (INTCSR). Writing 1 bits in these six positions clears all pending interrupts. (If we write back a mask value that has a 0 bit in one of the interrupt flag positions, the state of that interrupt is not affected. This kind of flag bit is called read/write-clear or just R/WC.) Other bits in the INTCSR enable interrupts of various kinds. By writing 0 bits in those locations, we're disabling the device to the maximum extent possible.

Our SetupDevice function is quite simple:

VOID SetupDevice(PDEVICE_EXTENSION pdx)   {   WRITE_PORT_ULONG((PULONG) (pdx->portbase + INTCSR),     INTCSR_IMBI_ENABLE     | (INTCSR_MB1 << INTCSR_IMBI_REG_SELECT_SHIFT)     | (INTCSR_BYTE0 << INTCSR_IMBI_BYTE_SELECT_SHIFT)     );   }

This function reprograms the INTCSR to specify that we want an interrupt to occur when there's a change to byte 0 of inbound mailbox register 1. We could have specified other interrupt conditions for this chip, including the emptying of a particular byte of a specified outbound mailbox register, the completion of a read DMA transfer, and the completion of a write DMA transfer.

PCI42's StartIo routine follows the pattern we've already studied:

1 2

VOID StartIo(IN PDEVICE_OBJECT fdo, IN PIRP Irp)   {   PDEVICE_EXTENSION pdx = (PDEVICE_EXTENSION) fdo->DeviceExtension;   PIO_STACK_LOCATION stack = IoGetCurrentIrpStackLocation(Irp);   NTSTATUS status = IoAcquireRemoveLock(&pdx->RemoveLock, Irp);   if (!NT_SUCCESS(status))     {     CompleteRequest(Irp, status, 0);     return;     }   if (!stack->Parameters.Read.Length)     {     StartNextPacket(&pdx->dqReadWrite, fdo);     CompleteRequest(Irp, STATUS_SUCCESS, 0);     return;     } pdx->buffer = (PUCHAR) Irp->AssociatedIrp.SystemBuffer;   pdx->nbytes = stack->Parameters.Read.Length;   pdx->numxfer = 0; KeSynchronizeExecution(pdx->InterruptObject,      (PKSYNCHRONIZE_ROUTINE) TransferFirst, pdx);   }

1. Here, we save parameters in the device extension to describe the ongoing progress of the input operation we're about to undertake. PCI42 uses the DO_BUFFERED_IO method, which isn't typical but helps make this driver simple enough to be used as an example.
2. Since our interrupt is connected, our device can interrupt at any time. The ISR will want to transfer data bytes when interrupts happen, but we want to be sure that the ISR is never confused about which data buffer to use or about the number of bytes we're trying to read. To restrain our ISR's eagerness, we put a flag in the device extension named activerequest that's ordinarily FALSE. Now is the time to set that flag to TRUE. As usual when dealing with a shared resource, we need to synchronize the setting of the flag with the code in the ISR that tests it, and we therefore need to invoke a SynchCritSection routine as I previously discussed. It might also happen that a data byte is already available, in which case the first interrupt will never happen. TransferFirst is a helper routine that checks for this eventuality and reads the first byte. The add-on function has ways of knowing that we emptied the mailbox, so it will presumably send the next byte in due course. Here's TransferFirst:

VOID TransferFirst(PDEVICE_EXTENSION pdx)   {   pdx->activerequest = TRUE;   ULONG mbef = READ_PORT_ULONG((PULONG) (pdx->portbase + MBEF));   if (!(mbef & MBEF_IN1_0))     return;   *pdx->buffer = READ_PORT_UCHAR(pdx->portbase + IMB1);   ++pdx->buffer;   ++pdx->numxfer;   if (—pdx->nbytes == 0)     {     pdx->activerequest = FALSE;     IoRequestDpc(pdx->DeviceObject, NULL, NULL);     }   }

The S5933 has a mailbox empty/full register (MBEF) whose bits indicate the current status of each byte of each mailbox register. Here, we check whether the register byte we're using for input (inbound mailbox register 1, byte 0) is presently unread. If so, we read it. That might exhaust the transfer count. We already have a subroutine (DpcForIsr) that knows what to do with a complete request, so we request a DPC if this first byte turns out to satisfy the request. (Recall that we're executing at DIRQL under protection of an interrupt spin lock because we've been invoked as a SynchCritSection routine, so we can't just complete the IRP right now.)

In normal operation with PCI42, the S5933 interrupts when a new data byte arrives in mailbox 1. The following ISR then gains control:

1 2 3 4 5 6 7 8

BOOLEAN OnInterrupt(PKINTERRUPT InterruptObject, PDEVICE_EXTENSION pdx)   { ULONG intcsr = READ_PORT_ULONG((PULONG) (pdx->portbase + INTCSR));   if (!(intcsr & INTCSR_INTERRUPT_PENDING))     return FALSE;   BOOLEAN dpc = FALSE; PIRP Irp = GetCurrentIrp(&pdx->dqReadWrite);   if (!Irp     || AreRequestsBeingAborted(&pdx->dqReadWrite)     || Irp->Cancel)     {     pdx->nbytes = 0;     dpc = Irp != NULL;     } while (intcsr & INTCSR_INTERRUPT_PENDING)     {    if (intcsr & INTCSR_IMBI)       {       if (pdx->nbytes && pdx->activerequest)         {         *pdx->buffer = READ_PORT_UCHAR(pdx->portbase + IMB1);         ++pdx->buffer;         ++pdx->numxfer;         if (!—pdx->nbytes)           dpc = TRUE;         }       } WRITE_PORT_ULONG((PULONG) (pdx->portbase + INTCSR), intcsr);          intcsr = READ_PORT_ULONG((PULONG) (pdx->portbase + INTCSR));     } if (dpc)     {     pdx->activerequest = FALSE;     IoRequestDpc(pdx->DeviceObject, NULL, NULL);     }   return TRUE;   }

1. Our first task is to discover whether our own device is trying to interrupt now. We read the S5933's INTCSR and test a bit (INTCSR_INTERRUPT_PENDING) that summarizes all pending causes of interrupts. If this bit is clear, we return immediately. The reason I chose to use the device extension pointer as the context argument to this routine—back when I called IoConnectInterrupt—should now be clear: we need immediate access to this structure to get the base port address.
2. When we use a DEVQUEUE, we rely on the queue object to keep track of the current IRP. This interrupt might be one that we don't expect because we're not currently servicing any IRP. In that case, we still have to clear the interrupt but shouldn't do anything else.
3. It's also possible that a Plug and Play or power event has occurred that will cause any new IRPs to be rejected by the dispatch routine. The DEVQUEUE's AreRequestsBeingAborted function tells us that fact so that we can abort the current request right now. Aborting an active request is a reasonable thing to do with a device such as this that proceeds byte by byte. Similarly it's a good idea to check whether the IRP has been cancelled if it will take a long time to finish the IRP. If your device interrupts only when it's done with a long transfer, you could leave this test out of your ISR.
4. We're now embarking on a loop that will terminate when all of our device's current interrupts have been cleared. At the end of the loop, we'll reread the INTCSR to determine whether any more interrupt conditions have arisen. If so, we'll repeat the loop. We're not being greedy with CPU time here—we want to avoid letting interrupts cascade into the system because servicing an interrupt is by itself relatively expensive.
5. If the S5933 has interrupted because of a mailbox event, we'll read a new data byte from the mailbox into the I/O buffer for the current IRP. If you were to look in the MBEF register immediately after the read, you'd see that the bit corresponding to inbound mailbox register 1, byte 0, gets cleared by the read. Note that we needn't test the MBEF to determine whether our byte has actually changed because we programmed the device to interrupt only upon a change to that single byte.
6. Writing the INTCSR with its previous contents has the effect of clearing the six R/WC interrupt bits, not changing a few read-only bits, and preserving the original setting of all read/write control bits.
7. Here, we read the INTCSR to determine whether additional interrupt conditions have arisen. If so, we'll repeat this loop to service them.
8. As we progressed through the preceding code, we set the Boolean dpc variable to TRUE if a DPC is now appropriate to complete the current IRP.

The DPC routine for PCI42 is as follows:

VOID DpcForIsr(PKDPC Dpc, PDEVICE_OBJECT fdo, PIRP junk, PVOID Context)   {   PDEVICE_EXTENSION pdx = (PDEVICE_EXTENSION) fdo->DeviceExtension;   NTSTATUS status = STATUS_SUCCESS;   PIRP Irp = GetCurrentIrp(&pdx->dqReadWrite);   ULONG info = pdx->numxfer;   StartNextPacket(&pdx->dqReadWrite, fdo);   CompleteRequest(Irp, status, info);   }

If you want to examine PCI42 in operation, you need to do several things. First obtain and install an S5933DK1 development board, including the ISA add-in interface card. Use the Add Hardware wizard to install the S5933DK1.SYS driver and the PCI42.SYS driver. (I found that Windows 98 initially identified the development board as a nonworking sound card and that I had to remove it in the Device Manager before I could install PCI42 as its driver. Windows 2000 handled the board normally, but I did encounter an annoying setup freeze when trying to upgrade from one release candidate to another during the beta phase.)

Then run both the ADDONSIM and TEST programs, which are in the PCI42 directory tree on the companion disc. ADDONSIM writes a data value to the mailbox via the ISA interface. TEST reads a data byte from PCI42. Determining the value of the data byte is left as an exercise for you.