Following the pattern of several other multi-threaded UNIX® kernels, FreeBSD deals with interrupt handlers by giving them their own thread context. Providing a context for interrupt handlers allows them to block on locks. To help avoid latency, however, interrupt threads run at real-time kernel priority. Thus, interrupt handlers should not execute for very long to avoid starving other kernel threads. In addition, since multiple handlers may share an interrupt thread, interrupt handlers should not sleep or use a sleepable lock to avoid starving another interrupt handler.

The interrupt threads currently in FreeBSD are referred to as heavyweight interrupt threads. They are called this because switching to an interrupt thread involves a full context switch. In the initial implementation, the kernel was not preemptive and thus interrupts that interrupted a kernel thread would have to wait until the kernel thread blocked or returned to userland before they would have an opportunity to run.

To deal with the latency problems, the kernel in FreeBSD has been made preemptive. Currently, we only preempt a kernel thread when we release a sleep mutex or when an interrupt comes in. However, the plan is to make the FreeBSD kernel fully preemptive as described below.

Not all interrupt handlers execute in a thread context. Instead, some handlers execute directly in primary interrupt context. These interrupt handlers are currently misnamed “fast” interrupt handlers since the INTR_FAST flag used in earlier versions of the kernel is used to mark these handlers. The only interrupts which currently use these types of interrupt handlers are clock interrupts and serial I/O device interrupts. Since these handlers do not have their own context, they may not acquire blocking locks and thus may only use spin mutexes.

Finally, there is one optional optimization that can be added in MD code called lightweight context switches. Since an interrupt thread executes in a kernel context, it can borrow the vmspace of any process. Thus, in a lightweight context switch, the switch to the interrupt thread does not switch vmspaces but borrows the vmspace of the interrupted thread. In order to ensure that the vmspace of the interrupted thread does not disappear out from under us, the interrupted thread is not allowed to execute until the interrupt thread is no longer borrowing its vmspace. This can happen when the interrupt thread either blocks or finishes. If an interrupt thread blocks, then it will use its own context when it is made runnable again. Thus, it can release the interrupted thread.

The cons of this optimization are that they are very machine specific and complex and thus only worth the effort if their is a large performance improvement. At this point it is probably too early to tell, and in fact, will probably hurt performance as almost all interrupt handlers will immediately block on Giant and require a thread fix-up when they block. Also, an alternative method of interrupt handling has been proposed by Mike Smith that works like so:

Fitting light weight context switches into this scheme might prove rather complicated. Since we may want to change to this scheme at some point in the future, it is probably best to defer work on light weight context switches until we have settled on the final interrupt handling architecture and determined how light weight context switches might or might not fit into it.

Simply put, a thread migrates when it moves from one CPU to another. In a non-preemptive kernel this can only happen at well-defined points such as when calling msleep or returning to userland. However, in the preemptive kernel, an interrupt can force a preemption and possible migration at any time. This can have negative affects on per-CPU data since with the exception of curthread and curpcb the data can change whenever you migrate. Since you can potentially migrate at any time this renders unprotected per-CPU data access rather useless. Thus it is desirable to be able to disable migration for sections of code that need per-CPU data to be stable.

Critical sections currently prevent migration since they do not allow context switches. However, this may be too strong of a requirement to enforce in some cases since a critical section also effectively blocks interrupt threads on the current processor. As a result, another API has been provided to allow the current thread to indicate that if it preempted it should not migrate to another CPU.

This API is known as thread pinning and is provided by the scheduler. The API consists of two functions: sched_pin and sched_unpin. These functions manage a per-thread nesting count td_pinned. A thread is pinned when its nesting count is greater than zero and a thread starts off unpinned with a nesting count of zero. Each scheduler implementation is required to ensure that pinned threads are only executed on the CPU that they were executing on when the sched_pin was first called. Since the nesting count is only written to by the thread itself and is only read by other threads when the pinned thread is not executing but while sched_lock is held, then td_pinned does not need any locking. The sched_pin function increments the nesting count and sched_unpin decrements the nesting count. Note that these functions only operate on the current thread and bind the current thread to the CPU it is executing on at the time. To bind an arbitrary thread to a specific CPU, the sched_bind and sched_unbind functions should be used instead.

The timeout kernel facility permits kernel services to register functions for execution as part of the softclock software interrupt. Events are scheduled based on a desired number of clock ticks, and callbacks to the consumer-provided function will occur at approximately the right time.

The global list of pending timeout events is protected by a global spin mutex, callout_lock; all access to the timeout list must be performed with this mutex held. When softclock is woken up, it scans the list of pending timeouts for those that should fire. In order to avoid lock order reversal, the softclock thread will release the callout_lock mutex when invoking the provided timeout callback function. If the CALLOUT_MPSAFE flag was not set during registration, then Giant will be grabbed before invoking the callout, and then released afterwards. The callout_lock mutex will be re-grabbed before proceeding. The softclock code is careful to leave the list in a consistent state while releasing the mutex. If DIAGNOSTIC is enabled, then the time taken to execute each function is measured, and a warning is generated if it exceeds a threshold.