Noticed that my sigtimedwait timeout patch got into the kernel, so polled signal I/O should now
work much better.
The question on why the timeout is calculated with an +1 for non-zero timeouts is still open.
AFAICT is is not needed as timespec_to_jiffies() does a correct rounding. The effect now is
timeout sleeping
0 0
1ns 2 jiffies
1 jiffies 2 jiffies
2 jiffies 3 jiffies
3 jiffies 4 jiffies
...
If the "+1" is taken out then the timeout scale becomes the expected one, starting at 1 jiffie, not
2.
--
Henirk Nordstrom
Henrik Nordstrom wrote 22 September 2000:
> As I mentioned earlier sigtimedwait with a zero timeout (0,0) should not
> block, but it currently does for 10msec (one jiffie). This is a
> performance problem for applications using polled signal queues. SUSV2
> says specifically for this case "returns immediately with an error".
>
> Attached is a new version of my patch. The previous version messed up
> the signal mask if the signal queue was empty and a zero timeout was
> selected.
>
> It is still waiting one more jiffie than what is indicated by the
> timeout value if other than zero, caused by the following code fragment:
>
> timeout = (timespec_to_jiffies(&ts)
> + (ts.tv_sec || ts.tv_nsec));
>
> Does anyone have any clue on why this +1 is there? I think this should
> also go away to only read
>
> timeout = timespec_to_jiffies(&ts);
>
> --
> Henrik Nordstrom
[patch deleted]