After looking over the state machine in xact.c, I'm thinking of removing
the TBLOCK_SUBENDABORT_ALL and TBLOCK_SUBENDABORT_RELEASE states in
favor of having the ROLLBACK command mark the whole transaction state
stack similarly to what is now done for COMMIT. In detail this would
require adding a TBLOCK_ABORT_PENDING state to use at the top level,
and ROLLBACK would act thus:
* For each subtransaction level: if it's in SUBABORT state (ie, already
aborted) then shift it to SUBENDABORT state (giving permission to pop it
from the stack); otherwise mark it SUBABORT_PENDING.
* At the outer level: if it's in ABORT state then shift to ENDABORT,
otherwise mark it ABORT_PENDING.
In CommitTransactionCommand we would have the behaviors:
ABORT:SUBABORT: do nothing (same as now)
SUBENDABORT: cleanup & pop recursively examine parent
SUBABORT_PENDING: abort subtransaction cleanup & pop recursively examine parent
ENDABORT: cleanup go to DEFAULT state
ABORT_PENDING: abort transaction cleanup go to DEFAULT state
I'm also toying with handling ROLLBACK TO by marking all the levels
above the target as SUBENDABORT or SUBABORT_PENDING, and then marking
the target level with one of two new states, TBLOCK_SUBRESTART or
TBLOCK_SUBABORT_RESTART (the latter if it was already SUBABORT).
These would have the behaviors
TBLOCK_SUBRESTART: abort subtransaction cleanup & pop start new subtransaction with same name
TBLOCK_SUBABORT_RESTART: cleanup & pop start new subtransaction with same name
This isn't any fewer states than we have now, but the states seem much
more clearly organized to me --- in particular, other than the RESTART
states there's full symmetry between outer-level and subtransaction
states. Also, this ensures that the planned state transitions are fully
marked out on the state stack before we start to do anything, which
I think is going to be more robust. AbortOutOfAnyTransaction is a bit of
a kluge and I don't really want to depend on it to implement ROLLBACK.
Comments?
regards, tom lane
