On 11/09/2021 10:05:02-0700, Linus Torvalds wrote:

On Sat, Sep 11, 2021 at 8:59 AM Alexandre Belloni
[off-list ref] wrote:

quoted

The broken down time conversion is similar to what is done
in the time subsystem since v5.14.

By "similar" you mean "identical", no?

Why is the rtc subsystem not just using the generic time64_to_tm()?

Yes, yes, I realize that due to historical mistakes, there's a
duplicate 'struct rtc_time' struct, but it turns out that that is
_identical_ to 'struct tm' except it also has a 'int tm_isdst' at the
end.

So you could literally make a union of the two, pass the 'struct tm'
part down to the generic code, and just do

     rtc_tm->tm_isdst = 0;

at the end.

Rather than have a duplicate copy of that admittedly clever Neri and
Schneider algorithm.

Hmm?

Yes, most of it is historical, I did have a look at removing the copy
but at the time, rtc_time64_to_tm was slightly more efficient because
it knew the time was positive.

The other issue is that struct rtc_time is exposed to userspace while
the kernel struct tm is not and this would tie both struct and if you
look close enough, struct tm has long tm_year and struct rtc_time has
int tm_year which on 32-bit ARM has a different size.

I've been reluctant to change struct tm because I didn't take the time
to check the impact on all the users (IIRC, mainly in filesystems).

-- 
Alexandre Belloni, co-owner and COO, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com