bug#53696: Integer overflow on Guix GC size calculation

[Bengt Richter <bokr@bokr.com>]

Sorry for following up my own post, but maybe it wasn't clear
why I printed (* 17592186042897 (expt 2 20)) in hex ? 

That is the value of [17592186042897 MiB] that you've been discussing.
(expt 2 20) is one MiB

Does that make
> --8<---------------cut here---------------start------------->8---
> $ guile --no-auto-compile -c '(use-modules (ice-9 format))(format #t "~20x\n~20x\n~20d\n" (* 17592186042897 (expt 2 20)) #xa1100000 #xa1100000)';
>     ffffffffa1100000
>             a1100000
>           2702180352
> --8<---------------cut here---------------end--------------->8---
a little clearer?

The discussion seems to be continuing, but no mention of the above.
How come?

Feeling ignored, and top-posting in desperation ;/

CC-ing ludo, who will instantly know where to fix it, if he hasn't already.


On +2022-02-02 13:04:41 +0100, Bengt Richter wrote:
> Hi Maxime, Ekaitz, et al,
> 
> On +2022-02-02 11:05:31 +0100, Maxime Devos wrote:
> > Ekaitz Zarraga schreef op di 01-02-2022 om 14:06 [+0000]:
> > > [17592186042897 MiB] deleting '/gnu/store/wbz6vkiz7cq8c531xvb31lxm28nz332i-ghc-8.10.7'
> > 
> > For comparison, this is about 16 exbibyte.
> > According to <https://en.wikipedia.org/wiki/Byte#Multiple-byte_units>,
> > that's more than the global monthly Internet traffic in 2004.
> > 
> > According to <https://what-if.xkcd.com/31/>, 16 exbibyte would be about
> > 17 million solid-state disks.  Even though this ignores deduplication,
> > this seems rather expensive. 
> > 
> > My guess is that the size of a store item was misrecorded somewhere.
> > 
> > Greetings,
> > Maxime.
> 
> s/misrecorded/mis-defined-in-record/ ?
> Wild guessing follows:
> 
> --8<---------------cut here---------------start------------->8---
> $ guile --no-auto-compile -c '(use-modules (ice-9 format))(format #t "~20x\n~20x\n~20d\n" (* 17592186042897 (expt 2 20)) #xa1100000 #xa1100000)';
>     ffffffffa1100000
>             a1100000
>           2702180352
> --8<---------------cut here---------------end--------------->8---
> 
> It looks to me like a 32-bit unsigned int should have been turned to 64-bit unsigned long or bigint
> but somehow got cast/interpreted as signed, becoming signed 64-bit long,
> which then in turn was seen by the print as 64-bit unsigned long.
> 
> I don't know, but if records are being used, perhaps some slot integer-widening logic
> might be involved? Or a mis-defined int slot that should have been long to accomodate
> big > 31-bit  positive integers?
> 
> Just guessing wildly -- I think I saw something about records and defining their fields
> as fixed C ints or longs.
> 
> -- 
> Regards,
> Bengt Richter
> 
> 
>