libfrog: fix overly sleep workqueues
I discovered the following bad behavior in the workqueue code when I
noticed that xfs_scrub was running single-threaded despite having 4
virtual CPUs allocated to the VM. I observed this sequence:
Thread 1 WQ1 WQ2...N
workqueue_create
<start up>
pthread_cond_wait
<start up>
pthread_cond_wait
workqueue_add
next_item == NULL
pthread_cond_signal
workqueue_add
next_item != NULL
<do not pthread_cond_signal>
<receives wakeup>
<run first item>
workqueue_add
next_item != NULL
<do not pthread_cond_signal>
<run second item>
<run third item>
pthread_cond_wait
workqueue_terminate
pthread_cond_broadcast
<receives wakeup>
<nothing to do, exits>
<wakes up again>
<nothing to do, exits>
Notice how threads WQ2...N are completely idle while WQ1 ends up doing
all the work! That wasn't the point of a worker pool! Observe that
thread 1 manages to queue two work items before WQ1 pulls the first item
off the queue. When thread 1 queues the third item, it sees that
next_item is not NULL, so it doesn't wake a worker. If thread 1 queues
all the N work that it has before WQ1 empties the queue, then none of
the other thread get woken up.
Fix this by maintaining a count of the number of active threads, and
using that to wake either the sole idle thread, or all the threads if
there are many that are idle. This dramatically improves startup
behavior of the workqueue and eliminates the collapse case.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
2 files changed
