In kernel 2.4.0-test11-pre3 the waitqueues and wakeup mechanisms were redesigned to remove the TASK_EXCLUSIVE bit from task_struct.state.  This was a good move, as that bit was inappropriate and led to various scenarios where wakeups could be lost.

There are still some shortcomings in the current design:
 

• For non-exclusive tasks, waitqueues are always FIFO.  If a designer wishes to use LIFO he can't.
• For exclusive tasks, waitqueues are also always FIFO.  This is a significant drawback,  because we always choose to wake the task which has the least amount of data in the CPU's caches.

The approach taken in thes patches is to split the waitqueues by putting an extra list_head into the waitqueue structure.  So the waitqueue ring has two halves: exclusive tasks on the right, non-exclusive on the left. This adds another eight bytes to struct page, which takes it over 64 bytes.  That is a big problem.

This provides a good deal of symmetry, leading to the following API:

add_wait_queue_lifo(wait_queue_head_t *q, wait_queue_t * wait);

Adds a task non-exclusively to the wait queue.  It will receive LIFO semantics on wakeup.

Adds a task non-exclusively to the wait queue.  It will receive FIFO semantics on wakeup.

Adds a task exclusively to the wait queue.  It will receive LIFO semantics on wakeup.

Adds a task exclusively to the wait queue.  It will receive FIFO semantics on wakeup.

#define add_wait_queue add_wait_queue_fifo
#define add_wait_queue_exclusive add_wait_queue_exclusive_fifo

This patch also cleans up a lot of cruft in wait.h and sched.c

Patches:

2.4.0-test11-pre3-task_exclusive.patch