Re: Interruptible sleeps (was Re: CommitFest 2009-07: Yay, Kevin! Thanks, reviewers!)

On 03/09/10 21:50, Tom Lane wrote:
> Heikki Linnakangas<heikki(dot)linnakangas(at)enterprisedb(dot)com> writes:
>> On 03/09/10 21:16, Tom Lane wrote:
>>> It's probably not too unreasonable to assume that pid_t assignment is
>>> atomic. But I'm still thinking that we have bigger problems than that
>>> if there are really cases where SetLatch can execute at approximately
>>> the same time as a latch owner is coming or going.
>
>> I don't see how to avoid it. A walsender, or any process really, can
>> exit at any time. It can make the latch inaccessible to others before it
>> exits to minimize the window, but it's always going to be possible that
>> another process is just about to call SetLatch when you exit.
>
> Well, in that case what we need to do is presume that the latch object
> has a continuing existence but the owner/receiver can come and go.
> I would suggest that InitLatch needs to initialize the object into a
> valid but unowned state; there is *no* deinitialize operation; and
> there are AcquireLatch and ReleaseLatch operations to become owner
> or stop being owner.

I think we have just a terminology issue. What you're describing is
exactly how it works now, if you just s/InitLatch/AcquireLatch. At the
moment there's no need for an initialization function other than the
InitLatch/AcquireLatch that associates the latch with the current
process. I can add one for the sake of future-proofing, and to have
better-defined behavior for setting a latch that has not been owned by
anyone yet, but it's not strictly necessary.

> We also need to define the semantics of SetLatch
> on an unowned latch --- does this set a signal condition that will be
> available to the next owner?

At the moment, no. Perhaps that would be useful, separating the Init and
Acquire operations is needed to make that sane.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com