diff options
author | Tom Lane <tgl@sss.pgh.pa.us> | 2025-09-25 17:02:15 -0400 |
---|---|---|
committer | Tom Lane <tgl@sss.pgh.pa.us> | 2025-09-25 17:02:15 -0400 |
commit | 02c4bc88302a750a2aed20e6f17885bda872d228 (patch) | |
tree | e1febbf58e2e18134beece3ec0382b9bd9a93e3b | |
parent | e849bd551c323a384f2b14d20a1b7bfaa6127ed7 (diff) |
Try to avoid floating-point roundoff error in pg_sleep().
I noticed the surprising behavior that pg_sleep(0.001) will sleep
for 2ms not the expected 1ms. Apparently the float8 calculation of
time-to-sleep is managing to produce something a hair over 1, which
ceil() rounds up to 2, and then WaitLatch() faithfully waits 2ms.
It could be that this works as-expected for some ranges of current
timestamp but not others, which would account for not having seen
it before. In any case, let's try to avoid it by removing the
float arithmetic in the delay calculation. We're stuck with the
declared input type being float8, but we can convert that to integer
microseconds right away, and then work strictly with integral values.
There might still be roundoff surprises for certain input values,
but at least the behavior won't be time-varying.
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Robert Haas <robertmhaas@gmail.com>
Reviewed-by: Nathan Bossart <nathandbossart@gmail.com>
Discussion: https://postgr.es/m/3879137.1758825752@sss.pgh.pa.us
-rw-r--r-- | src/backend/utils/adt/misc.c | 29 |
1 files changed, 20 insertions, 9 deletions
diff --git a/src/backend/utils/adt/misc.c b/src/backend/utils/adt/misc.c index 6c5e3438447..2c5a7ee9ddc 100644 --- a/src/backend/utils/adt/misc.c +++ b/src/backend/utils/adt/misc.c @@ -369,7 +369,20 @@ Datum pg_sleep(PG_FUNCTION_ARGS) { float8 secs = PG_GETARG_FLOAT8(0); - float8 endtime; + int64 usecs; + TimestampTz endtime; + + /* + * Convert the delay to int64 microseconds, rounding up any fraction, and + * silently limiting it to PG_INT64_MAX/2 microseconds (about 150K years) + * to ensure the computation of endtime won't overflow. Historically + * we've treated NaN as "no wait", not an error, so keep that behavior. + */ + if (isnan(secs) || secs <= 0.0) + PG_RETURN_VOID(); + secs *= USECS_PER_SEC; /* we assume overflow will produce +Inf */ + secs = ceil(secs); /* round up any fractional microsecond */ + usecs = (int64) Min(secs, (float8) (PG_INT64_MAX / 2)); /* * We sleep using WaitLatch, to ensure that we'll wake up promptly if an @@ -383,22 +396,20 @@ pg_sleep(PG_FUNCTION_ARGS) * less than the specified time when WaitLatch is terminated early by a * non-query-canceling signal such as SIGHUP. */ -#define GetNowFloat() ((float8) GetCurrentTimestamp() / 1000000.0) - - endtime = GetNowFloat() + secs; + endtime = GetCurrentTimestamp() + usecs; for (;;) { - float8 delay; + TimestampTz delay; long delay_ms; CHECK_FOR_INTERRUPTS(); - delay = endtime - GetNowFloat(); - if (delay >= 600.0) + delay = endtime - GetCurrentTimestamp(); + if (delay >= 600 * USECS_PER_SEC) delay_ms = 600000; - else if (delay > 0.0) - delay_ms = (long) ceil(delay * 1000.0); + else if (delay > 0) + delay_ms = (long) ((delay + 999) / 1000); else break; |