On Mon, Jun 8, 2015 at 1:51 PM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:
> Robert Haas wrote:
>
>> After studying this, I think it's a bug. See this comment:
>>
>> * Normal child backends can only be launched when we are in PM_RUN or
>> * PM_HOT_STANDBY state. (We also allow launch of normal
>> * child backends in PM_WAIT_BACKUP state, but only for superusers.)
>> * In other states we handle connection requests by launching "dead_end"
>> * child processes, which will simply send the client an error message and
>> * quit. (We track these in the BackendList so that we can know when they
>> * are all gone; this is important because they're still connected to shared
>> * memory, and would interfere with an attempt to destroy the shmem segment,
>> * possibly leading to SHMALL failure when we try to make a new one.)
>> * In PM_WAIT_DEAD_END state we are waiting for all the dead_end children
>> * to drain out of the system, and therefore stop accepting connection
>> * requests at all until the last existing child has quit (which hopefully
>> * will not be very long).
>>
>> That comment seems to imply that, at the very least, all backends with
>> shared-memory access need to be part of BackendList. But really, I'm
>> unclear why we'd ever want to exit without all background workers
>> having shut down, so maybe they should all be in there. Isn't that
>> sorta the point of this facility?
>>
>> Alvaro, any thoughts?
>
> As I recall, my thinking was:
>
> * anything connected to shmem that crashes could leave things in bad
> state (for example locks improperly held), whereas things not connected
> to shmem could crash without it being a problem for the rest of the
> system and thus do not require a full restart cycle. This stuff is
> detected with the PMChildSlot thingy; therefore all bgworkers with shmem
> access should have one of these.
>
> * I don't recall offhand what the criteria is for stuff to be in
> postmaster's BackendList. According to the comment on top of struct
> Backend, bgworkers connected to shmem should be on that list, even if
> they did not have the BGWORKER_BACKEND_DATABASE_CONNECTION flag on
> registration. So that would be a bug.
>
> Does this help you any?
Well, it sounds like we at least need to think about making it
consistently depend on shmem-access rather than database-connection.
I'm less sure what to do with workers that have not even got shmem
access.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
