[pacman-dev] Disk space checking branch "complete"

[Dan McGee]

On Wed, Nov 17, 2010 at 12:13 PM, Nezmer <git at nezmer.info> wrote:
> On Tue, Nov 16, 2010 at 06:06:32PM -0600, Dan McGee wrote:
>> On Tue, Nov 16, 2010 at 5:36 PM, Allan McRae <allan at archlinux.org> wrote:
>> > On 17/11/10 03:25, Nezmer wrote:
>> >>
>> >> On Tue, Nov 16, 2010 at 09:03:37PM +1000, Allan McRae wrote:
>> >>>
>> >>> On 16/11/10 18:02, Bryce Gibson wrote:
>> >>>>
>> >>>> Hi Allan,
>> >>>> I just wanted to ask, it looks like your patches will make a sync fail
>> >>>> if it
>> >>>> finds there's not enough space, is that correct?
>> >>>> Because I'd suggest a warning may be more appropriate, especially for
>> >>>> use
>> >>>> cases like compressed filesystems.
>> >>>> Cheers
>> >>>> Bryce
>> >>>> ps. I noticed that space checking can be completely disabled via
>> >>>> pacman.conf, which may be seen as a suitable solution for people in
>> >>>> these
>> >>>> types of situation.
>> >>>
>> >>> Disabling checking in pacman.conf is what I would recommend in these
>> >>> sort of cases.  These are the sort of things we will only find out
>> >>> after we get some widespread usage of the feature.  I would be
>> >>> interested in what the size calculations actually do on compressed
>> >>> filesystems.
>> >>>
>> >>> Allan
>> >>>
>> >>>
>> >>
>> >> On the topic of FS compression(also FS caching and delayed allocation).
>> >>
>> >> The real issue I had was the calculation of install size when building
>> >> the package. The calculated size was always *very* small. I addressed
>> >> this issue with 2 patches. The 1st one (adding sync before running du)
>> >> didn't fix anything. The 2nd patch worked pretty well.
>> >>
>> >>
>> >> http://gitorious.org/pacman-bsd/pacman-bsd/commit/8b367d4441ba85b5548285d987afcfc84c4fcb3e
>> >>
>> >> http://gitorious.org/pacman-bsd/pacman-bsd/commit/cf3341fe591293a493bc925e0433a1f696b37d90
>> >
>> > I think that initial "sync" is not needed because du does one anyway. The
>> > second one looks interesting so I have bookmarked it to look at later.
>>
>> We've putzed around with this a few times, haven't we? This basically
>> "reverts" this one:
>>
>> commit 149839c5391e9a93465f86dbb8d095a0150d755d
>> Author: Xavier Chantry <shiningxc at gmail.com>
>> Date:   Mon May 26 23:46:01 2008 +0200
>>
>>     du -b is not available on BSD, use du -k instead.
>>
>>     This fixes FS#10459.
>>
>>     There is apparently no portable ways to get the apparent size of a file,
>>     like du -b does. So the best compromise seems to get the block size in kB,
>>     and then convert that to byte so that we keep compatibility.
>>
>>     Signed-off-by: Xavier Chantry <shiningxc at gmail.com>
>>     Signed-off-by: Dan McGee <dan at archlinux.org>
>>
>> There are three or so file sizes that matter:
>> 1) du -s : what we do now, sums up actual taken space so would be more
>> accurate with many < 4K files.
>> 2) du -sb: what we did before and what is proposed, sums up apparent
>> size, so does not necessarily best represent installed size (either
>> sparse files or many <4K files would throw the number off). The 4K
>> assumption also may not always hold...
>> 3) du --tell-me-how-many-blocks-but-not-compressed : what seems like
>> perhaps the ideal? I'm not sure, but this would basically
>>     for file in tree:
>>         total += ceil(filesize to 4K)
>>
>> 1 is the most portable; 2 we need different flags all over the place;
>> 3 we definitely don't seem to be able to use exiting tools but this
>> would not be an awful one to write.
>>
>> -Dan
>>
>
> Is this 3 ?
>
> install_size=0
> for f in `find $pkgdir -type f`; do
>  (( s=$(${SIZECMD} ${f})/1024 ))
>  (( $(${SIZECMD} ${f})%1024 > 0 )) && (( s++ ))
>  (( $s < 4 )) && (( s=4 ))
>  (( install_size+=${s} ))
> done
>
> Unfortunately, this takes around 39 seconds in a 13085 files package.

Sure, but we can write a 25 line C program to do the same thing in
about 1% of the time.

-Dan