disrstate._check(link|exec) modifies repo dir mtime

[Mads Kiilerich]

On 10/16/2016 12:29 PM, Yuya Nishihara wrote:
> On Sat, 1 Oct 2016 11:35:40 -0700, Dorian Taylor wrote:
>> is there a reason why all this testing can’t go on in $REPO/.hg?
> No idea why, but there was an attempt to move check* files to .hg/cache.
>
> https://www.mercurial-scm.org/pipermail/mercurial-devel/2015-October/074885.html
> https://www.mercurial-scm.org/pipermail/mercurial-devel/2015-November/074971.html

About that:
> I don't think this is sufficient. This case can occur, for instance, if
> someone naively copies a repo from one machine to another with a USB
> stick. Or perhaps they've done a naive recursive chmod. Then we're
> permanently stuck thinking our filesystem doesn't have exec bits. If
> our fast-path files don't have the right permissions (for who knows
> what reasons), then we have to do the entire slow path.

I don't see how copying or chmod can trigger wrong measurement of file 
system capabilities. "Wrong" files will be removed again - just like in 
the old slow path. The slow path will pretty much be the existing slow 
path - it will not be slower than before but it will usually not be used.

> Also, putting knowledge of the existence of .hg/cache here is a classic
> layering violation.

Yes, it is a layering violation - a convenient one. I don't know if it 
deserves honorary mentioning as being classic. It just uses .hg/cache if 
it is available. Using .hg for probing properties outside .hg will 
inherently be some kind of violation.

The whole file system capabilities checking (including casing) should 
arguably move into the vfs layer as cached instance properties and 
support for these "capabilities" should not be tied to the platform but 
to the file system.


The implementation proposed in these patches is however more complex 
than the old one and there might be bugs and design flaws ...

/Mads