Softirqs

Softirqs are a mechanism in the Linux kernel for deferring work that cannot be completed safely in the immediate interrupt handler. They represent one of several bottom-half processing mechanisms used to balance low interrupt latency with the need to perform longer tasks in response to hardware events.

In normal operation, when a hardware interrupt fires, the top half runs quickly to acknowledge the interrupt and perform essential work. If more processing is required, the interrupt handler can raise softirqs by marking them as pending. The actual work associated with each softirq type is executed later in a softirq context by the kernel’s softirq processing code (do_softirq). Softirq handlers run with interrupts enabled and must not sleep, making them fast and non-blocking. The system tracks pending softirqs on a per-CPU basis and processes them on each CPU as part of the interrupt return path, or, under heavy load, through a dedicated ksoftirqd thread that can sleep.

Softirqs are a low-level primitive. Tasklets are a higher-level, now largely superseded mechanism built on top of softirqs, while workqueues provide a separate, sleepable path for deferrable work. The use of softirqs aims to keep interrupt handling latency low while still allowing deferred processing to occur without blocking critical paths.

Common softirq types include those used by the networking stack (for receive and transmit processing), timer-related softirqs, and block I/O softirqs, among others used by device drivers. Because softirqs run in a non-sleeping context, long-running tasks should be avoided or migrated to workqueues or other sleepable mechanisms to prevent latency growth or system stalls.

References to the exact softirq vectors and their names vary across kernel versions, but the general concept remains consistent: softirqs provide fast, deferred work handling integrated into the interrupt framework.