On Tue, Sep 6, 2016 at 6:04 AM, Simon Riggs <si...@2ndquadrant.com> wrote:
> On 1 September 2016 at 21:28, Simon Riggs <si...@2ndquadrant.com> wrote:
>> So the only way to handle multiple locks is to do this roughly the way
>> Rod suggests.
>
> I've just been reading the VACUUM code and it turns out that we
> already use Rod's mechanism internally. So on that basis it seems fine
> to support this as a useful user-level feature. If there is a better
> way of doing it, then that can be added later.
>
> My proposed changes to this patch are these
>
> 1. Rename this WAIT PATIENTLY, which is IMHO a better description of
> what is being requested. Bikeshedding welcome.
>
> 2. Introduce a new API call LockTablePatiently() that returns bool. So
> its usage is similar to ConditionalLockTable(), the only difference is
> you supply some other wait parameters with it. This isolates the
> internal mechanism from the usage, so allows us to more easily support
> any fancy new way of doing this we think of later.
>
> 3. Use LockTablePatiently() within lockcmds.c where appropriate
>
> 4. Replace the code for waiting in VACUUM with the new call to
> LockTablePatiently()
>
> So I see this as 2 patches: 1) new API and make VACUUM use new API, 2)
> Rod's LOCK TABLE patch
>
> First patch attached, requires also lock by Oid. If we agree, Rod,
> please update your patch to match?
Aside from the fact that polling is generally inefficient and wasteful
of system resources, this allows for undetected deadlocks. Consider:
S1: LOCK TABLE A;
S2: LOCK TABLE B;
S1: LOCK TABLE B; -- blocks
S2: LOCK TABLE A PATIENTLY; -- retries forever
Livelock might be possible, too.
I think it would be better to think harder about what would be
required to implement this natively in the lock manager. Suppose we
add a flag to each PROCLOCK which, if true, indicates that the lock
request is low priority. Also, we add a counter to each LOCK
indicating the number of pending low-priority lock requests. When
LockAcquireExtended reaches this code here...
if (lockMethodTable->conflictTab[lockmode] & lock->waitMask)
status = STATUS_FOUND;
else
status = LockCheckConflicts(lockMethodTable, lockmode,
lock, proclock);
...we add an additional check to the upper branch: if the number of
low-priority waiters is not 0, then we walk the wait queue; if all
waiters that conflict with the requested lock mode are low-priority,
then we set status = STATUS_OK. So, new lock requests refuse to queue
behind low-priority lock waiters.
Is that good enough to implement the requested behavior, or do we need
to do more? If we only do what's described above, then a
normal-priority waiter which joins the queue after a low-priority
waiter is already waiting will let the low-priority waiter go first.
That's probably not desirable, but it's pretty easy to fix: the logic
that determines where a new waiter enters the wait queue is in
ProcSleep(), and it wouldn't be too hard to arrange for new
normal-priority waiters to skip over any low-priority waiters that are
at the end of the existing wait queue (but not any that are in the
middle, because if we did that we'd also be skipping over
normal-priority waiters, which we shouldn't).
What more? Well, even after doing those two things, it's still
possible for either the simple deadlock logic in ProcSleep() or the
full deadlock detector to put a low-priority waiter in front of a
normal-priority waiter. However, our typical policy is to advance
waiters in the wait queue as little as possible. In other words, if
the wait queue contains A B C and we will deadlock unless C is moved
up, we will move it ahead of B but not A if that is sufficient to
avoid the deadlock. We will only move it ahead of both B and A if
that is necessary to avoid deadlock. So, low-priority requests won't
be moved up further than needed, which is good.
Still, it is possible to construct scenarios where we won't get
perfect low-priority behavior without more invasive changes. For
example, suppose we make a low-priority request queue-jump over an
intervening waiter to avoid deadlocking against it. Next, a
normal-priority waiter enters the queue. Then, the guy we skipped
aborts. At this point, we could in theory notice that it's possible
to move the low-priority request behind the new normal-priority
waiter. However, I think we shouldn't do that. We certainly can't do
it unconditionally because it might introduce deadlocks. We could
test whether it will introduce a deadlock and do it only if not, but
that's expensive. Instead, I think we should document that a
low-priority request will ordinarily cause the request to be satisfied
only after all conflicting normal-priority lock requests, but that
this is not guaranteed in the case where the wait queue is rearranged
to avoid deadlock. I don't think that limitation ought to be a
tremendous problem for users, and the alternatives are pretty
unappealing.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
