The main role of libfev's schedulers is to schedule runnable fibers. libfev implements few strategies to achieve that. They can be split into 2 categories: work sharing and work stealing.
Work sharing schedulers use a single, global queue for runnable fibers, the queue is shared by all workers.
The main advantage of these strategies is that fibers are scheduled fairly (FIFO). However, since the queue is shared, a high contention on the queue is possible.
Note that libfev implements cooperative multitasking, so it is possible that some fibers will get much more CPU time that others.
A bounded MPMC queue is used for the queue implementation. Since the queue is bounded, there may be not enough space for all runnable fibers. In that case, an additional fallback queue is used, which is protected by a global mutex. The fallback queue is based on an intrusive list.
This strategy uses an intrusive linked list for its global queue implementation. The list is protected by a global mutex.
This strategy implements a simple lock-free queue based on Michael Scott's queue.
It needs to allocate memory for each node. A memory pool is used for the allocation. The nodes are never released to the underlying operating system.
If an allocation fails, the whole process is aborted. This is the only strategy that can fail due to failed allocation in user space.
These strategies use a queue per worker for runnable fibers. The scheduling is in most cases independent. A fiber that becomes runnable (for example a newly spawned fiber) is pushed to the current worker's queue. If a worker is out of work, it tries to steal some fibers from other workers before going to sleep.
The main advantage is that a contention of the queues is minimized, and thus these strategies should scale better. However, it is possible that some fibers will be scheduled more often than others, and thus the scheduling is not fair.
A bounded MPMC queue per worker is created. If a queue is full, the fiber is pushed to a global fallback queue based on an intrusive list. The fallback queue is protected by a global mutex.
A bounded SPMC queue per worker is created. Similarly to work-stealing-bounded-mpmc, if a queue is full, the fiber is pushed to a global fallback queue, which is protected by a global mutex.
An intrusive list per worker is created. Each list is protected by its own lock.
The queue's lock strategy can be controlled by FEV_SCHED_STEAL_LOCKING_LOCK option. Currently, a mutex or a spinlock can be used as the queue's lock.