6.3.2 DispatchCleanup Functionality

Any NT driver in which IRPs can be held for an indefinite interval should have a DispatchCleanup routine. If IRPs might be queued when a user-mode handle for a file object that represents the underlying physical or logical device is closed, the highest-level driver holding IRPs bound for that device must have a DispatchCleanup routine.

An NT driver's DispatchCleanup routine is called to cancel all IRPs that are currently queued to the target device object for the file object specified in the driver's I/O stack location of the cleanup IRP.

In general, a DispatchCleanup routine must do the following:

·Reset the driver's Cancel entry point, if any, to NULL in every IRP currently in the device queue (or in the driver's internal queue of IRPs) for the target device object and given file object with IoSetCancelRoutine.

·Cancel every IRP currently in the queue for the target device object and the file object specified in the driver's I/O stack location of the cleanup IRP.

·Complete the input cleanup IRP and return STATUS_SUCCESS.

The DispatchCleanup routine must set up the I/O status block as follows in each IRP queued to the target device object for the given file object, before it completes each IRP to be cancelled:

The DispatchCleanup routine specifies IO_NO_INCREMENT when it calls IoCompleteRequest with each cancelled IRP. After it has cancelled all queued IRPs for the given file object, the DispatchCleanup routine completes the cleanup request itself with STATUS_SUCCESS and a PriorityBoost of IO_NO_INCREMENT.

If a driver manages its own internal queue of IRPs, rather than using the device queue associated with its device object, the driver should set up an interlocked queue that is protected with an executive spin lock. Otherwise, another driver routine (running on another processor in an SMP machine) might dequeue an IRP that should be cancelled by the DispatchCleanup routine. To remove an IRP from such a queue and cancel it, the DispatchCleanup routine calls an ExInterlocked..List support routine with a driver-supplied executive spin lock.

Any area in the device extension of the target device object that is shared between the DispatchCleanup routine and another driver routine also must be protected by a spin lock.

Any NT driver must synchronize its access to cancelable IRPs with the I/O Manager and with its own Cancel routine(s), if any.

Consider the following guidelines for managing spin lock usage within a DispatchCleanup routine that must synchronize with one or more Cancel routines:

·If the DispatchCleanup routine must acquire both the system cancel spin lock and one or more of the driver's own spin locks, it should call IoAcquireCancelSpinLock before it acquires its own spin lock(s).

Such a DispatchCleanup routine must release any nested spin lock acquisitions in inverse order: that is, it must release its own spin lock(s) before it calls IoReleaseCancelSpinLock.

·If a driver uses the I/O-manager-supplied device queue in the device object, then the DispatchCleanup routine must call IoAcquireCancelSpinLock before it calls IoSetCancelRoutine to reset the Cancel entry point in an IRP to NULL.

If a driver manages its own queue(s) of IRPs, then the DispatchCleanup routine does not need to acquire the cancel spin lock before calling IoSetCancelRoutine.

·The DispatchCleanup routine must release every spin lock it is holding before it calls IoCompleteRequest with an IRP, including each IRP that it cancels.

For more information about Cancel routines, see also Chapter 12.

Keep the following points in mind when implementing a DispatchCleanup routine:

·This routine must walk the queue of IRPs for the target device object and cancel every outstanding request for the given file object before it completes the input cleanup IRP, which is sent when the handle for the file object that represents the device in user mode is closed.

·The queue should be protected by a spin lock to prevent IRPs that should be cancelled from being dequeued by any other driver routine. The device queue associated with a device object comes with a spin lock. If a driver sets up its own queue for IRPs, it should set up an interlocked queue.

·This routine must acquire the system cancel spin lock with IoAcquireCancelSpinLock before it calls IoSetCancelRoutine to reset its Cancel entry point to NULL in each queued IRP if the driver has a Cancel routine. The exception to this rule is if the driver manages its own queue(s) of IRPs and synchronizes access to them with driver-supplied spin locks; in that case, the driver does not need to acquire the cancel spin lock before calling IoSetCancelRoutine.

·This routine must set each queued IRP's I/O status block with Status set to STATUS_CANCELLED and Information set to zero before it completes each such IRP.

·This routine must release all spin locks it is holding before calling IoCompleteRequest with an IRP.

·This routine should set the PriorityBoost to IO_NO_INCREMENT when it completes each cancelled IRP and the cleanup IRP.