Robert Haas <robertmhaas@gmail.com> writes:
> ... When I say I want my handler to be
> fired in 3 s, I don't mean that I want it to be fired when the system
> time is 3 seconds greater than it is right now. I mean I want it to be
> fired in 3 actual seconds, regardless of what dumb thing the system
> clock may choose to do.
That would be lovely, certainly. But aren't you moving the goalposts
rather far? I don't think we make any promises about such things
today, so why has the issue suddenly gotten more pressing? In particular,
why do you think Nitin's patch is proof against this? Seems to me it's
probably got *more* failure cases, not fewer, if the system clock is
acting funny.
BTW, one could imagine addressing this concern by having timeout.c work
with CLOCK_MONOTONIC instead of the regular wall clock. But I fear
we'd have to drop enable_timeout_at(), for lack of ability to translate
between CLOCK_MONOTONIC timestamps and those used by anybody else.
Also get_timeout_start_time/get_timeout_finish_time would become
problematic. Maybe we only really care about deltas, so the more
restrictive API would be workable, but it seems like a nontrivial
amount of work.
On the whole, in these days of NTP, I'm not sure I care to spend
large amounts of effort on dealing with a bogus system clock.
regards, tom lane