bug#19581: 'sleep' return value is incorrectly rounded

bug#19581: 'sleep' return value is incorrectly rounded

[Ludovic Courtès]

I just noticed that ‘sleep’ essentially always rounds down its return
value, which makes it unreliable, as in this example:

--8<---------------cut here---------------start------------->8---
$ time guile -c '(sigaction SIGINT +) (call-with-new-thread (lambda () (let loop () (kill (getpid) SIGINT) (loop)))) (let loop ((n 3)) (when (> n 0) (loop (sleep (pk "s" n)))))'

;;; ("s" 3)

;;; ("s" 2)

;;; ("s" 1)

real	0m0.039s
user	0m0.052s
sys	0m0.017s
--8<---------------cut here---------------end--------------->8---

Here ‘loop’ is meant to assure we sleep for roughly 3 seconds, but
because of the incorrect rounding, we end up not sleeping at all.

Ludo’.

bug#19581: 'sleep' return value is incorrectly rounded

[Andy Wingo]

Hi :)

On Tue 13 Jan 2015 13:58, ludo@gnu.org (Ludovic Courtès) writes:

> I just noticed that ‘sleep’ essentially always rounds down its return
> value, which makes it unreliable, as in this example:
>
> $ time guile -c '(sigaction SIGINT +) (call-with-new-thread (lambda () (let loop () (kill (getpid) SIGINT) (loop)))) (let loop ((n 3)) (when (> n 0) (loop (sleep (pk "s" n)))))'
>
> ;;; ("s" 3)
>
> ;;; ("s" 2)
>
> ;;; ("s" 1)
>
> real	0m0.039s
> user	0m0.052s
> sys	0m0.017s
>
> Here ‘loop’ is meant to assure we sleep for roughly 3 seconds, but
> because of the incorrect rounding, we end up not sleeping at all.

Hummmmmmmmmm.

I don't think we can change anything in 2.0.  Right?

In master, how about (1) we allow sleep to take any real number, in
seconds; (2) sleep returns an inexact real number.  Then we can
deprecate usleep and "standardize" on sleeping in units of seconds.
Internally we refactor scm_std_usleep/* to take a 64-bit double for the
number of seconds to sleep and return a double for the time remaining.

WDYT?

Andy

bug#19581: 'sleep' return value is incorrectly rounded

[Ludovic Courtès]

Hello,

Andy Wingo <wingo@pobox.com> skribis:

> On Tue 13 Jan 2015 13:58, ludo@gnu.org (Ludovic Courtès) writes:
>
>> I just noticed that ‘sleep’ essentially always rounds down its return
>> value, which makes it unreliable, as in this example:
>>
>> $ time guile -c '(sigaction SIGINT +) (call-with-new-thread (lambda () (let loop () (kill (getpid) SIGINT) (loop)))) (let loop ((n 3)) (when (> n 0) (loop (sleep (pk "s" n)))))'
>>
>> ;;; ("s" 3)
>>
>> ;;; ("s" 2)
>>
>> ;;; ("s" 1)
>>
>> real	0m0.039s
>> user	0m0.052s
>> sys	0m0.017s
>>
>> Here ‘loop’ is meant to assure we sleep for roughly 3 seconds, but
>> because of the incorrect rounding, we end up not sleeping at all.
>
> Hummmmmmmmmm.
>
> I don't think we can change anything in 2.0.  Right?
>
> In master, how about (1) we allow sleep to take any real number, in
> seconds; (2) sleep returns an inexact real number.  Then we can
> deprecate usleep and "standardize" on sleeping in units of seconds.
> Internally we refactor scm_std_usleep/* to take a 64-bit double for the
> number of seconds to sleep and return a double for the time remaining.

Sounds like a plan.  There could be programs that will break if ‘sleep’
gives them an inexact instead of an integer.  Maybe we can assume this
is an unlikely situation, though.

Thoughts?

Ludo’.