On Wed, Feb 21, 2024 at 07:08:03AM +0900, Michael Paquier wrote:
> Well, both you and Andrey are asking for it now, so let's do it. The
> implementation is simple:
> - Store in InjectionPointSharedState an array of wait_counts and an
> array of names. There is only one condition variable.
> - When a point wants to wait, it takes the spinlock and looks within
> the array of names until it finds a free slot, adds its name into the
> array to reserve a wait counter at the same position, releases the
> spinlock. Then it loops on the condition variable for an update of
> the counter it has reserved. It is possible to make something more
> efficient, but at a small size it would not really matter.
> - The wakeup takes a point name in argument, acquires the spinlock,
> and checks if it can find the point into the array, pinpoints the
> location of the counter to update and updates it. Then it broadcasts
> the change.
> - The wait loop checks its counter, leaves its loop, cancels the
> sleep, takes the spinlock to unregister from the array, and leaves.
>
> I would just hardcode the number of points that can wait, say 5 of
> them tracked in shmem? Does that look like what you are looking at?
I was looking at that, and it proves to work OK, so you can do stuff
like waits and wakeups for multiple processes in a controlled manner.
The attached patch authorizes up to 32 waiters. I have switched
things so as the information reported in pg_stat_activity is the name
of the injection point itself.
Comments and ideas are welcome.
--
Michael