Oh, wow! For us, setting `TZ` is a speedup, but perhaps that's unique to
unique situation. (/etc is a FUSE mount in our setting, which is much
slower.)
Let me know if it would be helpful for me to do something else
here---though I have to confess that gnulib's time code is a little
baffling to me!
Cheers,
Michael
On 2024-11-26 at 10:01:56 PM, Pádraig Brady wrote:
> On 26/11/2024 03:34, Michael Greenberg wrote:
>> If TZ is unset, the date utility will call newfstat("/etc/localtime")
>> each time it reads a time using date --date or date -f. If reading from
>> /etc is expensive, you can experience significant slowdowns.
>>
>> I didn't write a new test---you can observe the phenomenon by running
>> the following:
>>
>> $ cat dates # 4 isn't enough for slowness, but enough to make the trace
>> clear
>> Mon Nov 25 09:14:54 PM EST 2024
>> Mon Nov 25 09:15:01 PM EST 2024
>> Mon Nov 25 09:15:01 PM EST 2024
>> Mon Nov 25 09:15:01 PM EST 2024
>> $ strace date -f dates +%a
>> ... see 4 calls to newfstatat on /etc/localtime
>> $ strace coreutils/src/date -f dates +%a # use the patched version
>> ... see 1 call to newfstatat on /usr/share/zoneinfo/YOURTIMEZONEHERE ...
>
> Indeed:
>
> $ yes 1 | head -n100 | strace -e newfstatat -c date -f- > /dev/null
> % time seconds usecs/call calls errors syscall
> ------ ----------- ----------- --------- --------- ----------------
> 100.00 0.000071 0 100 newfstatat
>
> Also if specifying TZ in the date, it gets worse!
>
> $ yes 'TZ="UTC" 1' | head -n100 | strace -e newfstatat -c date -f- >
> /dev/null
> % time seconds usecs/call calls errors syscall
> ------ ----------- ----------- --------- --------- ----------------
> 100.00 0.005097 5 999 newfstatat
>
>> It might be possible to achieve the same effect by patching gnulib, but
>> I haven't dug into it.
>
> Thanks for the info. It does seem like we could improve here.
> I won't be able to dig into this, this week at least.
>
> thanks,
> Pádraig.
>
>>
>> Cheers,
>> Michael
>>
>> ---
>> src/date.c | 8 ++++++++
>> 1 file changed, 8 insertions(+)
>>
>> diff --git a/src/date.c b/src/date.c
>> index ee32403ef..45b111c4e 100644
>> --- a/src/date.c
>> +++ b/src/date.c
>> @@ -575,6 +575,14 @@ main (int argc, char **argv)
>> char *format_copy = adjust_resolution (format);
>> char const *format_res = format_copy ? format_copy : format;
>> char const *tzstring = getenv ("TZ");
>> + /* If TZ is unset, set it to the system timezone. This caches the
>> + value, so batch_convert won't have to do it over and over
>> + again. */
>> + if (tzstring == NULL) {
>> + tzset();
>> + tzstring = tzname[0];
>> + setenv("TZ", tzstring, 0);
>> + }
>> timezone_t tz = tzalloc (tzstring);
>>
>> if (batch_file != nullptr)
