Unwinding a failed character-device registration

ID: unwinding-a-failed-character-device-registration

Unwinding a failed character-device registration releases resources acquired before an initialization failure while preserving the lifetime of anything that concurrent users can still access. In Linux, reserving device numbers, publishing the character-device object and creating userspace-visible names are distinct steps. The cleanup path must know which steps actually succeeded.
Consider an original abstract initialization sequence: reserve a number range, initialize a character-device object, publish it, then finish a later setup step. The following table tracks successful acquisitions; it is not a complete driver implementation.
State reachedRange reserved?Character device live?Required response to a later failure
Before reservation succeedsNoNoDo not release a range that was not acquired.
Reservation succeeded, publication not reachedYesNoRelease that reserved range and other private resources already acquired.
Publication succeededYesYesStop new access, unwind the reservation, and preserve storage needed by existing users.
alloc_chrdev_region reserves a range and reports failure through its return value. cdev_init initializes the object with its operations; cdev_add publishes it and makes it live immediately. unregister_chrdev_region releases a successful number reservation. Pair each release with the acquisition that actually succeeded, using the same first number and count.[ref]
For a concrete bookkeeping example, suppose a successful reservation contains four interfaces starting at returned pair . Failure before publication requires releasing that four-number range. Releasing only three leaves part of the reservation behind; releasing a guessed starting major or an unacquired range is a different error. The returned major from dynamic character-device number allocation must be retained rather than reconstructed from a pathname.
Publication changes the lifetime problem. cdev_del prevents new opens, but existing open file descriptors can still invoke their file operations. Therefore removing the device from lookup is not a proof that its backing state can immediately be freed. Character-device lifetime after deletion must account for the references and callbacks already in progress.[ref]
The combined cdev_device_add helper has an especially important failure contract: callers must assume userspace could open the device and call operations even if the helper returns failure. A blanket error handler that immediately frees all backing state would violate that contract.[ref] Design ownership and reference handling around the documented publication boundary, then release each owned resource once. The general lesson is that reverse acquisition order must be combined with concurrency-aware lifetime management.

New to topics? Read the docs here!