Guidelines for Writing DPC Routines
A DpcForIsr or CustomDpc routine must synchronize its access to a physical device, and to any shared state information or resources that the driver maintains, with the driver's other routines that access the same device or memory locations.
If a DpcForIsr or CustomDpc routine shares the device or state with an ISR, it must call KeSynchronizeExecution, supplying the address of a driver-supplied SynchCritSection routine that programs the device or accesses the shared state. For more information, see Using Critical Sections.
If a DpcForIsr or CustomDpc routine shares state or resources, such as an interlocked queue or a timer object, with routines other than an ISR, it must protect the shared state or resources with a driver-initialized executive spin lock. For more information, see Spin Locks.
DpcForIsr and CustomDpc routines run at IRQL = DISPATCH_LEVEL, which restricts the set of support routines they can call.
For example, DpcForIsr and CustomDpc routines can neither access nor allocate pageable memory, and they cannot wait for kernel dispatcher objects to be set to the signaled state. On the other hand, they can acquire and release a driver's executive spin lock with KeAcquireSpinLockAtDpcLevel and KeReleaseSpinLockFromDpcLevel, which run faster than KeAcquireSpinLock and KeReleaseSpinLock.
Although a DPC routine cannot make blocking calls, it can queue a work item to run in a system worker thread that runs at IRQL equal to PASSIVE_LEVEL. The work item can make blocking calls that wait on dispatcher objects. To queue a work item, a DpcForIsr routine typically calls a routine such as IoQueueWorkItem, and a CustomDpc routine typically calls the ExQueueWorkItem routine.
DpcForIsr and CustomDpc routines are typically responsible for starting the next I/O operation on the device.
For lowest-level physical device drivers that use direct I/O, this responsibility can include using a SynchCritSection routine to program the device to transfer more data in order to satisfy the current IRP before the driver calls IoStartNextPacket.
DpcForIsr and CustomDpc routines should run only for brief periods, and should delegate as much processing as possible to worker threads.
While a DPC routine runs on a processor, all threads are prevented from running on the same processor. Other DPC routines that are queued and ready to run can be blocked from executing until the current DPC routine is finished. To avoid degrading system responsiveness, a typical DPC routine should run for no more than 100 microseconds each time it is called. If a task requires longer than 100 microseconds and must execute at IRQL equal to DISPATCH_LEVEL, the DPC routine should end after 100 microseconds and schedule one or more CustomTimerDpc routines to complete the task at a later time. For more information about CustomTimerDpc routines, see Timer Objects and DPCs.
A DPC routine should perform only tasks that must run at DISPATCH_LEVEL, and then delegate any remaining interrupt-related work to threads that run at IRQL = PASSIVE_LEVEL. For example, a DPC routine can queue a work item to run in a system worker thread.
DPC routines that call the KeStallExecutionProcessor routine to delay execution must not specify delays of more than 100 microseconds.
Use the performance analysis tools in the WDK to evaluate the execution times of DPC routines. For an example that uses the Tracelog tool to monitor DPC execution times, see Example 15: Measuring DPC/ISR Time.
If the driver uses DMA and its AdapterControl routine returns KeepObject or DeallocateObjectKeepRegisters (thereby retaining the system DMA controller channel or bus-master adapter for additional transfer operations), the DpcForIsr or CustomDpc routine is responsible for releasing the adapter object or map registers with FreeAdapterChannel or FreeMapRegisters before it completes the current IRP and returns control.
If a lowest-level physical device driver sets up a controller object to synchronize I/O operations through the controller to attached devices, its DpcForIsr or CustomDpc routine is responsible for releasing the controller object using IoFreeController before it completes the current IRP and returns control.
DpcForIsr and CustomDpc routines are generally responsible for logging any device errors that occurred during the processing of a given request, retrying the current request if necessary and possible, and for setting the I/O status block and calling IoCompleteRequest for the current IRP.
If the driver and device support overlapped I/O operations, the driver must follow the rules for handling overlapped I/O operations.
The DpcForIsr or CustomDpc routine of any driver usually completes the I/O processing only for a subset of the public I/O control codes that the driver must support. In particular, the DPC routine completes operations for device control requests with the following characteristics:
Requests that change the state of the physical device
Requests that require the return of inherently volatile information about the physical device